2 * Performance events core code:
4 * Copyright (C) 2008 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 <pzijlstr@redhat.com>
7 * Copyright © 2009 Paul Mackerras, IBM Corp. <paulus@au1.ibm.com>
9 * For licensing details see kernel-base/COPYING
14 #include <linux/cpu.h>
15 #include <linux/smp.h>
16 #include <linux/idr.h>
17 #include <linux/file.h>
18 #include <linux/poll.h>
19 #include <linux/slab.h>
20 #include <linux/hash.h>
21 #include <linux/tick.h>
22 #include <linux/sysfs.h>
23 #include <linux/dcache.h>
24 #include <linux/percpu.h>
25 #include <linux/ptrace.h>
26 #include <linux/reboot.h>
27 #include <linux/vmstat.h>
28 #include <linux/device.h>
29 #include <linux/export.h>
30 #include <linux/vmalloc.h>
31 #include <linux/hardirq.h>
32 #include <linux/rculist.h>
33 #include <linux/uaccess.h>
34 #include <linux/syscalls.h>
35 #include <linux/anon_inodes.h>
36 #include <linux/kernel_stat.h>
37 #include <linux/perf_event.h>
38 #include <linux/ftrace_event.h>
39 #include <linux/hw_breakpoint.h>
40 #include <linux/mm_types.h>
41 #include <linux/cgroup.h>
45 #include <asm/irq_regs.h>
47 struct remote_function_call
{
48 struct task_struct
*p
;
49 int (*func
)(void *info
);
54 static void remote_function(void *data
)
56 struct remote_function_call
*tfc
= data
;
57 struct task_struct
*p
= tfc
->p
;
61 if (task_cpu(p
) != smp_processor_id() || !task_curr(p
))
65 tfc
->ret
= tfc
->func(tfc
->info
);
69 * task_function_call - call a function on the cpu on which a task runs
70 * @p: the task to evaluate
71 * @func: the function to be called
72 * @info: the function call argument
74 * Calls the function @func when the task is currently running. This might
75 * be on the current CPU, which just calls the function directly
77 * returns: @func return value, or
78 * -ESRCH - when the process isn't running
79 * -EAGAIN - when the process moved away
82 task_function_call(struct task_struct
*p
, int (*func
) (void *info
), void *info
)
84 struct remote_function_call data
= {
88 .ret
= -ESRCH
, /* No such (running) process */
92 smp_call_function_single(task_cpu(p
), remote_function
, &data
, 1);
98 * cpu_function_call - call a function on the cpu
99 * @func: the function to be called
100 * @info: the function call argument
102 * Calls the function @func on the remote cpu.
104 * returns: @func return value or -ENXIO when the cpu is offline
106 static int cpu_function_call(int cpu
, int (*func
) (void *info
), void *info
)
108 struct remote_function_call data
= {
112 .ret
= -ENXIO
, /* No such CPU */
115 smp_call_function_single(cpu
, remote_function
, &data
, 1);
120 #define PERF_FLAG_ALL (PERF_FLAG_FD_NO_GROUP |\
121 PERF_FLAG_FD_OUTPUT |\
122 PERF_FLAG_PID_CGROUP)
125 * branch priv levels that need permission checks
127 #define PERF_SAMPLE_BRANCH_PERM_PLM \
128 (PERF_SAMPLE_BRANCH_KERNEL |\
129 PERF_SAMPLE_BRANCH_HV)
132 EVENT_FLEXIBLE
= 0x1,
134 EVENT_ALL
= EVENT_FLEXIBLE
| EVENT_PINNED
,
138 * perf_sched_events : >0 events exist
139 * perf_cgroup_events: >0 per-cpu cgroup events exist on this cpu
141 struct static_key_deferred perf_sched_events __read_mostly
;
142 static DEFINE_PER_CPU(atomic_t
, perf_cgroup_events
);
143 static DEFINE_PER_CPU(atomic_t
, perf_branch_stack_events
);
145 static atomic_t nr_mmap_events __read_mostly
;
146 static atomic_t nr_comm_events __read_mostly
;
147 static atomic_t nr_task_events __read_mostly
;
148 static atomic_t nr_freq_events __read_mostly
;
150 static LIST_HEAD(pmus
);
151 static DEFINE_MUTEX(pmus_lock
);
152 static struct srcu_struct pmus_srcu
;
155 * perf event paranoia level:
156 * -1 - not paranoid at all
157 * 0 - disallow raw tracepoint access for unpriv
158 * 1 - disallow cpu events for unpriv
159 * 2 - disallow kernel profiling for unpriv
161 int sysctl_perf_event_paranoid __read_mostly
= 1;
163 /* Minimum for 512 kiB + 1 user control page */
164 int sysctl_perf_event_mlock __read_mostly
= 512 + (PAGE_SIZE
/ 1024); /* 'free' kiB per user */
167 * max perf event sample rate
169 #define DEFAULT_MAX_SAMPLE_RATE 100000
170 #define DEFAULT_SAMPLE_PERIOD_NS (NSEC_PER_SEC / DEFAULT_MAX_SAMPLE_RATE)
171 #define DEFAULT_CPU_TIME_MAX_PERCENT 25
173 int sysctl_perf_event_sample_rate __read_mostly
= DEFAULT_MAX_SAMPLE_RATE
;
175 static int max_samples_per_tick __read_mostly
= DIV_ROUND_UP(DEFAULT_MAX_SAMPLE_RATE
, HZ
);
176 static int perf_sample_period_ns __read_mostly
= DEFAULT_SAMPLE_PERIOD_NS
;
178 static atomic_t perf_sample_allowed_ns __read_mostly
=
179 ATOMIC_INIT( DEFAULT_SAMPLE_PERIOD_NS
* DEFAULT_CPU_TIME_MAX_PERCENT
/ 100);
181 void update_perf_cpu_limits(void)
183 u64 tmp
= perf_sample_period_ns
;
185 tmp
*= sysctl_perf_cpu_time_max_percent
;
187 atomic_set(&perf_sample_allowed_ns
, tmp
);
190 static int perf_rotate_context(struct perf_cpu_context
*cpuctx
);
192 int perf_proc_update_handler(struct ctl_table
*table
, int write
,
193 void __user
*buffer
, size_t *lenp
,
196 int ret
= proc_dointvec_minmax(table
, write
, buffer
, lenp
, ppos
);
201 max_samples_per_tick
= DIV_ROUND_UP(sysctl_perf_event_sample_rate
, HZ
);
202 perf_sample_period_ns
= NSEC_PER_SEC
/ sysctl_perf_event_sample_rate
;
203 update_perf_cpu_limits();
208 int sysctl_perf_cpu_time_max_percent __read_mostly
= DEFAULT_CPU_TIME_MAX_PERCENT
;
210 int perf_cpu_time_max_percent_handler(struct ctl_table
*table
, int write
,
211 void __user
*buffer
, size_t *lenp
,
214 int ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
219 update_perf_cpu_limits();
225 * perf samples are done in some very critical code paths (NMIs).
226 * If they take too much CPU time, the system can lock up and not
227 * get any real work done. This will drop the sample rate when
228 * we detect that events are taking too long.
230 #define NR_ACCUMULATED_SAMPLES 128
231 DEFINE_PER_CPU(u64
, running_sample_length
);
233 void perf_sample_event_took(u64 sample_len_ns
)
235 u64 avg_local_sample_len
;
236 u64 local_samples_len
;
238 if (atomic_read(&perf_sample_allowed_ns
) == 0)
241 /* decay the counter by 1 average sample */
242 local_samples_len
= __get_cpu_var(running_sample_length
);
243 local_samples_len
-= local_samples_len
/NR_ACCUMULATED_SAMPLES
;
244 local_samples_len
+= sample_len_ns
;
245 __get_cpu_var(running_sample_length
) = local_samples_len
;
248 * note: this will be biased artifically low until we have
249 * seen NR_ACCUMULATED_SAMPLES. Doing it this way keeps us
250 * from having to maintain a count.
252 avg_local_sample_len
= local_samples_len
/NR_ACCUMULATED_SAMPLES
;
254 if (avg_local_sample_len
<= atomic_read(&perf_sample_allowed_ns
))
257 if (max_samples_per_tick
<= 1)
260 max_samples_per_tick
= DIV_ROUND_UP(max_samples_per_tick
, 2);
261 sysctl_perf_event_sample_rate
= max_samples_per_tick
* HZ
;
262 perf_sample_period_ns
= NSEC_PER_SEC
/ sysctl_perf_event_sample_rate
;
264 printk_ratelimited(KERN_WARNING
265 "perf samples too long (%lld > %d), lowering "
266 "kernel.perf_event_max_sample_rate to %d\n",
267 avg_local_sample_len
,
268 atomic_read(&perf_sample_allowed_ns
),
269 sysctl_perf_event_sample_rate
);
271 update_perf_cpu_limits();
274 static atomic64_t perf_event_id
;
276 static void cpu_ctx_sched_out(struct perf_cpu_context
*cpuctx
,
277 enum event_type_t event_type
);
279 static void cpu_ctx_sched_in(struct perf_cpu_context
*cpuctx
,
280 enum event_type_t event_type
,
281 struct task_struct
*task
);
283 static void update_context_time(struct perf_event_context
*ctx
);
284 static u64
perf_event_time(struct perf_event
*event
);
286 void __weak
perf_event_print_debug(void) { }
288 extern __weak
const char *perf_pmu_name(void)
293 static inline u64
perf_clock(void)
295 return local_clock();
298 static inline struct perf_cpu_context
*
299 __get_cpu_context(struct perf_event_context
*ctx
)
301 return this_cpu_ptr(ctx
->pmu
->pmu_cpu_context
);
304 static void perf_ctx_lock(struct perf_cpu_context
*cpuctx
,
305 struct perf_event_context
*ctx
)
307 raw_spin_lock(&cpuctx
->ctx
.lock
);
309 raw_spin_lock(&ctx
->lock
);
312 static void perf_ctx_unlock(struct perf_cpu_context
*cpuctx
,
313 struct perf_event_context
*ctx
)
316 raw_spin_unlock(&ctx
->lock
);
317 raw_spin_unlock(&cpuctx
->ctx
.lock
);
320 #ifdef CONFIG_CGROUP_PERF
323 * perf_cgroup_info keeps track of time_enabled for a cgroup.
324 * This is a per-cpu dynamically allocated data structure.
326 struct perf_cgroup_info
{
332 struct cgroup_subsys_state css
;
333 struct perf_cgroup_info __percpu
*info
;
337 * Must ensure cgroup is pinned (css_get) before calling
338 * this function. In other words, we cannot call this function
339 * if there is no cgroup event for the current CPU context.
341 static inline struct perf_cgroup
*
342 perf_cgroup_from_task(struct task_struct
*task
)
344 return container_of(task_css(task
, perf_subsys_id
),
345 struct perf_cgroup
, css
);
349 perf_cgroup_match(struct perf_event
*event
)
351 struct perf_event_context
*ctx
= event
->ctx
;
352 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
354 /* @event doesn't care about cgroup */
358 /* wants specific cgroup scope but @cpuctx isn't associated with any */
363 * Cgroup scoping is recursive. An event enabled for a cgroup is
364 * also enabled for all its descendant cgroups. If @cpuctx's
365 * cgroup is a descendant of @event's (the test covers identity
366 * case), it's a match.
368 return cgroup_is_descendant(cpuctx
->cgrp
->css
.cgroup
,
369 event
->cgrp
->css
.cgroup
);
372 static inline bool perf_tryget_cgroup(struct perf_event
*event
)
374 return css_tryget(&event
->cgrp
->css
);
377 static inline void perf_put_cgroup(struct perf_event
*event
)
379 css_put(&event
->cgrp
->css
);
382 static inline void perf_detach_cgroup(struct perf_event
*event
)
384 perf_put_cgroup(event
);
388 static inline int is_cgroup_event(struct perf_event
*event
)
390 return event
->cgrp
!= NULL
;
393 static inline u64
perf_cgroup_event_time(struct perf_event
*event
)
395 struct perf_cgroup_info
*t
;
397 t
= per_cpu_ptr(event
->cgrp
->info
, event
->cpu
);
401 static inline void __update_cgrp_time(struct perf_cgroup
*cgrp
)
403 struct perf_cgroup_info
*info
;
408 info
= this_cpu_ptr(cgrp
->info
);
410 info
->time
+= now
- info
->timestamp
;
411 info
->timestamp
= now
;
414 static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context
*cpuctx
)
416 struct perf_cgroup
*cgrp_out
= cpuctx
->cgrp
;
418 __update_cgrp_time(cgrp_out
);
421 static inline void update_cgrp_time_from_event(struct perf_event
*event
)
423 struct perf_cgroup
*cgrp
;
426 * ensure we access cgroup data only when needed and
427 * when we know the cgroup is pinned (css_get)
429 if (!is_cgroup_event(event
))
432 cgrp
= perf_cgroup_from_task(current
);
434 * Do not update time when cgroup is not active
436 if (cgrp
== event
->cgrp
)
437 __update_cgrp_time(event
->cgrp
);
441 perf_cgroup_set_timestamp(struct task_struct
*task
,
442 struct perf_event_context
*ctx
)
444 struct perf_cgroup
*cgrp
;
445 struct perf_cgroup_info
*info
;
448 * ctx->lock held by caller
449 * ensure we do not access cgroup data
450 * unless we have the cgroup pinned (css_get)
452 if (!task
|| !ctx
->nr_cgroups
)
455 cgrp
= perf_cgroup_from_task(task
);
456 info
= this_cpu_ptr(cgrp
->info
);
457 info
->timestamp
= ctx
->timestamp
;
460 #define PERF_CGROUP_SWOUT 0x1 /* cgroup switch out every event */
461 #define PERF_CGROUP_SWIN 0x2 /* cgroup switch in events based on task */
464 * reschedule events based on the cgroup constraint of task.
466 * mode SWOUT : schedule out everything
467 * mode SWIN : schedule in based on cgroup for next
469 void perf_cgroup_switch(struct task_struct
*task
, int mode
)
471 struct perf_cpu_context
*cpuctx
;
476 * disable interrupts to avoid geting nr_cgroup
477 * changes via __perf_event_disable(). Also
480 local_irq_save(flags
);
483 * we reschedule only in the presence of cgroup
484 * constrained events.
488 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
489 cpuctx
= this_cpu_ptr(pmu
->pmu_cpu_context
);
490 if (cpuctx
->unique_pmu
!= pmu
)
491 continue; /* ensure we process each cpuctx once */
494 * perf_cgroup_events says at least one
495 * context on this CPU has cgroup events.
497 * ctx->nr_cgroups reports the number of cgroup
498 * events for a context.
500 if (cpuctx
->ctx
.nr_cgroups
> 0) {
501 perf_ctx_lock(cpuctx
, cpuctx
->task_ctx
);
502 perf_pmu_disable(cpuctx
->ctx
.pmu
);
504 if (mode
& PERF_CGROUP_SWOUT
) {
505 cpu_ctx_sched_out(cpuctx
, EVENT_ALL
);
507 * must not be done before ctxswout due
508 * to event_filter_match() in event_sched_out()
513 if (mode
& PERF_CGROUP_SWIN
) {
514 WARN_ON_ONCE(cpuctx
->cgrp
);
516 * set cgrp before ctxsw in to allow
517 * event_filter_match() to not have to pass
520 cpuctx
->cgrp
= perf_cgroup_from_task(task
);
521 cpu_ctx_sched_in(cpuctx
, EVENT_ALL
, task
);
523 perf_pmu_enable(cpuctx
->ctx
.pmu
);
524 perf_ctx_unlock(cpuctx
, cpuctx
->task_ctx
);
530 local_irq_restore(flags
);
533 static inline void perf_cgroup_sched_out(struct task_struct
*task
,
534 struct task_struct
*next
)
536 struct perf_cgroup
*cgrp1
;
537 struct perf_cgroup
*cgrp2
= NULL
;
540 * we come here when we know perf_cgroup_events > 0
542 cgrp1
= perf_cgroup_from_task(task
);
545 * next is NULL when called from perf_event_enable_on_exec()
546 * that will systematically cause a cgroup_switch()
549 cgrp2
= perf_cgroup_from_task(next
);
552 * only schedule out current cgroup events if we know
553 * that we are switching to a different cgroup. Otherwise,
554 * do no touch the cgroup events.
557 perf_cgroup_switch(task
, PERF_CGROUP_SWOUT
);
560 static inline void perf_cgroup_sched_in(struct task_struct
*prev
,
561 struct task_struct
*task
)
563 struct perf_cgroup
*cgrp1
;
564 struct perf_cgroup
*cgrp2
= NULL
;
567 * we come here when we know perf_cgroup_events > 0
569 cgrp1
= perf_cgroup_from_task(task
);
571 /* prev can never be NULL */
572 cgrp2
= perf_cgroup_from_task(prev
);
575 * only need to schedule in cgroup events if we are changing
576 * cgroup during ctxsw. Cgroup events were not scheduled
577 * out of ctxsw out if that was not the case.
580 perf_cgroup_switch(task
, PERF_CGROUP_SWIN
);
583 static inline int perf_cgroup_connect(int fd
, struct perf_event
*event
,
584 struct perf_event_attr
*attr
,
585 struct perf_event
*group_leader
)
587 struct perf_cgroup
*cgrp
;
588 struct cgroup_subsys_state
*css
;
589 struct fd f
= fdget(fd
);
597 css
= css_from_dir(f
.file
->f_dentry
, &perf_subsys
);
603 cgrp
= container_of(css
, struct perf_cgroup
, css
);
606 /* must be done before we fput() the file */
607 if (!perf_tryget_cgroup(event
)) {
614 * all events in a group must monitor
615 * the same cgroup because a task belongs
616 * to only one perf cgroup at a time
618 if (group_leader
&& group_leader
->cgrp
!= cgrp
) {
619 perf_detach_cgroup(event
);
629 perf_cgroup_set_shadow_time(struct perf_event
*event
, u64 now
)
631 struct perf_cgroup_info
*t
;
632 t
= per_cpu_ptr(event
->cgrp
->info
, event
->cpu
);
633 event
->shadow_ctx_time
= now
- t
->timestamp
;
637 perf_cgroup_defer_enabled(struct perf_event
*event
)
640 * when the current task's perf cgroup does not match
641 * the event's, we need to remember to call the
642 * perf_mark_enable() function the first time a task with
643 * a matching perf cgroup is scheduled in.
645 if (is_cgroup_event(event
) && !perf_cgroup_match(event
))
646 event
->cgrp_defer_enabled
= 1;
650 perf_cgroup_mark_enabled(struct perf_event
*event
,
651 struct perf_event_context
*ctx
)
653 struct perf_event
*sub
;
654 u64 tstamp
= perf_event_time(event
);
656 if (!event
->cgrp_defer_enabled
)
659 event
->cgrp_defer_enabled
= 0;
661 event
->tstamp_enabled
= tstamp
- event
->total_time_enabled
;
662 list_for_each_entry(sub
, &event
->sibling_list
, group_entry
) {
663 if (sub
->state
>= PERF_EVENT_STATE_INACTIVE
) {
664 sub
->tstamp_enabled
= tstamp
- sub
->total_time_enabled
;
665 sub
->cgrp_defer_enabled
= 0;
669 #else /* !CONFIG_CGROUP_PERF */
672 perf_cgroup_match(struct perf_event
*event
)
677 static inline void perf_detach_cgroup(struct perf_event
*event
)
680 static inline int is_cgroup_event(struct perf_event
*event
)
685 static inline u64
perf_cgroup_event_cgrp_time(struct perf_event
*event
)
690 static inline void update_cgrp_time_from_event(struct perf_event
*event
)
694 static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context
*cpuctx
)
698 static inline void perf_cgroup_sched_out(struct task_struct
*task
,
699 struct task_struct
*next
)
703 static inline void perf_cgroup_sched_in(struct task_struct
*prev
,
704 struct task_struct
*task
)
708 static inline int perf_cgroup_connect(pid_t pid
, struct perf_event
*event
,
709 struct perf_event_attr
*attr
,
710 struct perf_event
*group_leader
)
716 perf_cgroup_set_timestamp(struct task_struct
*task
,
717 struct perf_event_context
*ctx
)
722 perf_cgroup_switch(struct task_struct
*task
, struct task_struct
*next
)
727 perf_cgroup_set_shadow_time(struct perf_event
*event
, u64 now
)
731 static inline u64
perf_cgroup_event_time(struct perf_event
*event
)
737 perf_cgroup_defer_enabled(struct perf_event
*event
)
742 perf_cgroup_mark_enabled(struct perf_event
*event
,
743 struct perf_event_context
*ctx
)
749 * set default to be dependent on timer tick just
752 #define PERF_CPU_HRTIMER (1000 / HZ)
754 * function must be called with interrupts disbled
756 static enum hrtimer_restart
perf_cpu_hrtimer_handler(struct hrtimer
*hr
)
758 struct perf_cpu_context
*cpuctx
;
759 enum hrtimer_restart ret
= HRTIMER_NORESTART
;
762 WARN_ON(!irqs_disabled());
764 cpuctx
= container_of(hr
, struct perf_cpu_context
, hrtimer
);
766 rotations
= perf_rotate_context(cpuctx
);
769 * arm timer if needed
772 hrtimer_forward_now(hr
, cpuctx
->hrtimer_interval
);
773 ret
= HRTIMER_RESTART
;
779 /* CPU is going down */
780 void perf_cpu_hrtimer_cancel(int cpu
)
782 struct perf_cpu_context
*cpuctx
;
786 if (WARN_ON(cpu
!= smp_processor_id()))
789 local_irq_save(flags
);
793 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
794 cpuctx
= this_cpu_ptr(pmu
->pmu_cpu_context
);
796 if (pmu
->task_ctx_nr
== perf_sw_context
)
799 hrtimer_cancel(&cpuctx
->hrtimer
);
804 local_irq_restore(flags
);
807 static void __perf_cpu_hrtimer_init(struct perf_cpu_context
*cpuctx
, int cpu
)
809 struct hrtimer
*hr
= &cpuctx
->hrtimer
;
810 struct pmu
*pmu
= cpuctx
->ctx
.pmu
;
813 /* no multiplexing needed for SW PMU */
814 if (pmu
->task_ctx_nr
== perf_sw_context
)
818 * check default is sane, if not set then force to
819 * default interval (1/tick)
821 timer
= pmu
->hrtimer_interval_ms
;
823 timer
= pmu
->hrtimer_interval_ms
= PERF_CPU_HRTIMER
;
825 cpuctx
->hrtimer_interval
= ns_to_ktime(NSEC_PER_MSEC
* timer
);
827 hrtimer_init(hr
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL_PINNED
);
828 hr
->function
= perf_cpu_hrtimer_handler
;
831 static void perf_cpu_hrtimer_restart(struct perf_cpu_context
*cpuctx
)
833 struct hrtimer
*hr
= &cpuctx
->hrtimer
;
834 struct pmu
*pmu
= cpuctx
->ctx
.pmu
;
837 if (pmu
->task_ctx_nr
== perf_sw_context
)
840 if (hrtimer_active(hr
))
843 if (!hrtimer_callback_running(hr
))
844 __hrtimer_start_range_ns(hr
, cpuctx
->hrtimer_interval
,
845 0, HRTIMER_MODE_REL_PINNED
, 0);
848 void perf_pmu_disable(struct pmu
*pmu
)
850 int *count
= this_cpu_ptr(pmu
->pmu_disable_count
);
852 pmu
->pmu_disable(pmu
);
855 void perf_pmu_enable(struct pmu
*pmu
)
857 int *count
= this_cpu_ptr(pmu
->pmu_disable_count
);
859 pmu
->pmu_enable(pmu
);
862 static DEFINE_PER_CPU(struct list_head
, rotation_list
);
865 * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
866 * because they're strictly cpu affine and rotate_start is called with IRQs
867 * disabled, while rotate_context is called from IRQ context.
869 static void perf_pmu_rotate_start(struct pmu
*pmu
)
871 struct perf_cpu_context
*cpuctx
= this_cpu_ptr(pmu
->pmu_cpu_context
);
872 struct list_head
*head
= &__get_cpu_var(rotation_list
);
874 WARN_ON(!irqs_disabled());
876 if (list_empty(&cpuctx
->rotation_list
))
877 list_add(&cpuctx
->rotation_list
, head
);
880 static void get_ctx(struct perf_event_context
*ctx
)
882 WARN_ON(!atomic_inc_not_zero(&ctx
->refcount
));
885 static void put_ctx(struct perf_event_context
*ctx
)
887 if (atomic_dec_and_test(&ctx
->refcount
)) {
889 put_ctx(ctx
->parent_ctx
);
891 put_task_struct(ctx
->task
);
892 kfree_rcu(ctx
, rcu_head
);
896 static void unclone_ctx(struct perf_event_context
*ctx
)
898 if (ctx
->parent_ctx
) {
899 put_ctx(ctx
->parent_ctx
);
900 ctx
->parent_ctx
= NULL
;
904 static u32
perf_event_pid(struct perf_event
*event
, struct task_struct
*p
)
907 * only top level events have the pid namespace they were created in
910 event
= event
->parent
;
912 return task_tgid_nr_ns(p
, event
->ns
);
915 static u32
perf_event_tid(struct perf_event
*event
, struct task_struct
*p
)
918 * only top level events have the pid namespace they were created in
921 event
= event
->parent
;
923 return task_pid_nr_ns(p
, event
->ns
);
927 * If we inherit events we want to return the parent event id
930 static u64
primary_event_id(struct perf_event
*event
)
935 id
= event
->parent
->id
;
941 * Get the perf_event_context for a task and lock it.
942 * This has to cope with with the fact that until it is locked,
943 * the context could get moved to another task.
945 static struct perf_event_context
*
946 perf_lock_task_context(struct task_struct
*task
, int ctxn
, unsigned long *flags
)
948 struct perf_event_context
*ctx
;
952 * One of the few rules of preemptible RCU is that one cannot do
953 * rcu_read_unlock() while holding a scheduler (or nested) lock when
954 * part of the read side critical section was preemptible -- see
955 * rcu_read_unlock_special().
957 * Since ctx->lock nests under rq->lock we must ensure the entire read
958 * side critical section is non-preemptible.
962 ctx
= rcu_dereference(task
->perf_event_ctxp
[ctxn
]);
965 * If this context is a clone of another, it might
966 * get swapped for another underneath us by
967 * perf_event_task_sched_out, though the
968 * rcu_read_lock() protects us from any context
969 * getting freed. Lock the context and check if it
970 * got swapped before we could get the lock, and retry
971 * if so. If we locked the right context, then it
972 * can't get swapped on us any more.
974 raw_spin_lock_irqsave(&ctx
->lock
, *flags
);
975 if (ctx
!= rcu_dereference(task
->perf_event_ctxp
[ctxn
])) {
976 raw_spin_unlock_irqrestore(&ctx
->lock
, *flags
);
982 if (!atomic_inc_not_zero(&ctx
->refcount
)) {
983 raw_spin_unlock_irqrestore(&ctx
->lock
, *flags
);
993 * Get the context for a task and increment its pin_count so it
994 * can't get swapped to another task. This also increments its
995 * reference count so that the context can't get freed.
997 static struct perf_event_context
*
998 perf_pin_task_context(struct task_struct
*task
, int ctxn
)
1000 struct perf_event_context
*ctx
;
1001 unsigned long flags
;
1003 ctx
= perf_lock_task_context(task
, ctxn
, &flags
);
1006 raw_spin_unlock_irqrestore(&ctx
->lock
, flags
);
1011 static void perf_unpin_context(struct perf_event_context
*ctx
)
1013 unsigned long flags
;
1015 raw_spin_lock_irqsave(&ctx
->lock
, flags
);
1017 raw_spin_unlock_irqrestore(&ctx
->lock
, flags
);
1021 * Update the record of the current time in a context.
1023 static void update_context_time(struct perf_event_context
*ctx
)
1025 u64 now
= perf_clock();
1027 ctx
->time
+= now
- ctx
->timestamp
;
1028 ctx
->timestamp
= now
;
1031 static u64
perf_event_time(struct perf_event
*event
)
1033 struct perf_event_context
*ctx
= event
->ctx
;
1035 if (is_cgroup_event(event
))
1036 return perf_cgroup_event_time(event
);
1038 return ctx
? ctx
->time
: 0;
1042 * Update the total_time_enabled and total_time_running fields for a event.
1043 * The caller of this function needs to hold the ctx->lock.
1045 static void update_event_times(struct perf_event
*event
)
1047 struct perf_event_context
*ctx
= event
->ctx
;
1050 if (event
->state
< PERF_EVENT_STATE_INACTIVE
||
1051 event
->group_leader
->state
< PERF_EVENT_STATE_INACTIVE
)
1054 * in cgroup mode, time_enabled represents
1055 * the time the event was enabled AND active
1056 * tasks were in the monitored cgroup. This is
1057 * independent of the activity of the context as
1058 * there may be a mix of cgroup and non-cgroup events.
1060 * That is why we treat cgroup events differently
1063 if (is_cgroup_event(event
))
1064 run_end
= perf_cgroup_event_time(event
);
1065 else if (ctx
->is_active
)
1066 run_end
= ctx
->time
;
1068 run_end
= event
->tstamp_stopped
;
1070 event
->total_time_enabled
= run_end
- event
->tstamp_enabled
;
1072 if (event
->state
== PERF_EVENT_STATE_INACTIVE
)
1073 run_end
= event
->tstamp_stopped
;
1075 run_end
= perf_event_time(event
);
1077 event
->total_time_running
= run_end
- event
->tstamp_running
;
1082 * Update total_time_enabled and total_time_running for all events in a group.
1084 static void update_group_times(struct perf_event
*leader
)
1086 struct perf_event
*event
;
1088 update_event_times(leader
);
1089 list_for_each_entry(event
, &leader
->sibling_list
, group_entry
)
1090 update_event_times(event
);
1093 static struct list_head
*
1094 ctx_group_list(struct perf_event
*event
, struct perf_event_context
*ctx
)
1096 if (event
->attr
.pinned
)
1097 return &ctx
->pinned_groups
;
1099 return &ctx
->flexible_groups
;
1103 * Add a event from the lists for its context.
1104 * Must be called with ctx->mutex and ctx->lock held.
1107 list_add_event(struct perf_event
*event
, struct perf_event_context
*ctx
)
1109 WARN_ON_ONCE(event
->attach_state
& PERF_ATTACH_CONTEXT
);
1110 event
->attach_state
|= PERF_ATTACH_CONTEXT
;
1113 * If we're a stand alone event or group leader, we go to the context
1114 * list, group events are kept attached to the group so that
1115 * perf_group_detach can, at all times, locate all siblings.
1117 if (event
->group_leader
== event
) {
1118 struct list_head
*list
;
1120 if (is_software_event(event
))
1121 event
->group_flags
|= PERF_GROUP_SOFTWARE
;
1123 list
= ctx_group_list(event
, ctx
);
1124 list_add_tail(&event
->group_entry
, list
);
1127 if (is_cgroup_event(event
))
1130 if (has_branch_stack(event
))
1131 ctx
->nr_branch_stack
++;
1133 list_add_rcu(&event
->event_entry
, &ctx
->event_list
);
1134 if (!ctx
->nr_events
)
1135 perf_pmu_rotate_start(ctx
->pmu
);
1137 if (event
->attr
.inherit_stat
)
1142 * Initialize event state based on the perf_event_attr::disabled.
1144 static inline void perf_event__state_init(struct perf_event
*event
)
1146 event
->state
= event
->attr
.disabled
? PERF_EVENT_STATE_OFF
:
1147 PERF_EVENT_STATE_INACTIVE
;
1151 * Called at perf_event creation and when events are attached/detached from a
1154 static void perf_event__read_size(struct perf_event
*event
)
1156 int entry
= sizeof(u64
); /* value */
1160 if (event
->attr
.read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
1161 size
+= sizeof(u64
);
1163 if (event
->attr
.read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
1164 size
+= sizeof(u64
);
1166 if (event
->attr
.read_format
& PERF_FORMAT_ID
)
1167 entry
+= sizeof(u64
);
1169 if (event
->attr
.read_format
& PERF_FORMAT_GROUP
) {
1170 nr
+= event
->group_leader
->nr_siblings
;
1171 size
+= sizeof(u64
);
1175 event
->read_size
= size
;
1178 static void perf_event__header_size(struct perf_event
*event
)
1180 struct perf_sample_data
*data
;
1181 u64 sample_type
= event
->attr
.sample_type
;
1184 perf_event__read_size(event
);
1186 if (sample_type
& PERF_SAMPLE_IP
)
1187 size
+= sizeof(data
->ip
);
1189 if (sample_type
& PERF_SAMPLE_ADDR
)
1190 size
+= sizeof(data
->addr
);
1192 if (sample_type
& PERF_SAMPLE_PERIOD
)
1193 size
+= sizeof(data
->period
);
1195 if (sample_type
& PERF_SAMPLE_WEIGHT
)
1196 size
+= sizeof(data
->weight
);
1198 if (sample_type
& PERF_SAMPLE_READ
)
1199 size
+= event
->read_size
;
1201 if (sample_type
& PERF_SAMPLE_DATA_SRC
)
1202 size
+= sizeof(data
->data_src
.val
);
1204 event
->header_size
= size
;
1207 static void perf_event__id_header_size(struct perf_event
*event
)
1209 struct perf_sample_data
*data
;
1210 u64 sample_type
= event
->attr
.sample_type
;
1213 if (sample_type
& PERF_SAMPLE_TID
)
1214 size
+= sizeof(data
->tid_entry
);
1216 if (sample_type
& PERF_SAMPLE_TIME
)
1217 size
+= sizeof(data
->time
);
1219 if (sample_type
& PERF_SAMPLE_IDENTIFIER
)
1220 size
+= sizeof(data
->id
);
1222 if (sample_type
& PERF_SAMPLE_ID
)
1223 size
+= sizeof(data
->id
);
1225 if (sample_type
& PERF_SAMPLE_STREAM_ID
)
1226 size
+= sizeof(data
->stream_id
);
1228 if (sample_type
& PERF_SAMPLE_CPU
)
1229 size
+= sizeof(data
->cpu_entry
);
1231 event
->id_header_size
= size
;
1234 static void perf_group_attach(struct perf_event
*event
)
1236 struct perf_event
*group_leader
= event
->group_leader
, *pos
;
1239 * We can have double attach due to group movement in perf_event_open.
1241 if (event
->attach_state
& PERF_ATTACH_GROUP
)
1244 event
->attach_state
|= PERF_ATTACH_GROUP
;
1246 if (group_leader
== event
)
1249 if (group_leader
->group_flags
& PERF_GROUP_SOFTWARE
&&
1250 !is_software_event(event
))
1251 group_leader
->group_flags
&= ~PERF_GROUP_SOFTWARE
;
1253 list_add_tail(&event
->group_entry
, &group_leader
->sibling_list
);
1254 group_leader
->nr_siblings
++;
1256 perf_event__header_size(group_leader
);
1258 list_for_each_entry(pos
, &group_leader
->sibling_list
, group_entry
)
1259 perf_event__header_size(pos
);
1263 * Remove a event from the lists for its context.
1264 * Must be called with ctx->mutex and ctx->lock held.
1267 list_del_event(struct perf_event
*event
, struct perf_event_context
*ctx
)
1269 struct perf_cpu_context
*cpuctx
;
1271 * We can have double detach due to exit/hot-unplug + close.
1273 if (!(event
->attach_state
& PERF_ATTACH_CONTEXT
))
1276 event
->attach_state
&= ~PERF_ATTACH_CONTEXT
;
1278 if (is_cgroup_event(event
)) {
1280 cpuctx
= __get_cpu_context(ctx
);
1282 * if there are no more cgroup events
1283 * then cler cgrp to avoid stale pointer
1284 * in update_cgrp_time_from_cpuctx()
1286 if (!ctx
->nr_cgroups
)
1287 cpuctx
->cgrp
= NULL
;
1290 if (has_branch_stack(event
))
1291 ctx
->nr_branch_stack
--;
1294 if (event
->attr
.inherit_stat
)
1297 list_del_rcu(&event
->event_entry
);
1299 if (event
->group_leader
== event
)
1300 list_del_init(&event
->group_entry
);
1302 update_group_times(event
);
1305 * If event was in error state, then keep it
1306 * that way, otherwise bogus counts will be
1307 * returned on read(). The only way to get out
1308 * of error state is by explicit re-enabling
1311 if (event
->state
> PERF_EVENT_STATE_OFF
)
1312 event
->state
= PERF_EVENT_STATE_OFF
;
1315 static void perf_group_detach(struct perf_event
*event
)
1317 struct perf_event
*sibling
, *tmp
;
1318 struct list_head
*list
= NULL
;
1321 * We can have double detach due to exit/hot-unplug + close.
1323 if (!(event
->attach_state
& PERF_ATTACH_GROUP
))
1326 event
->attach_state
&= ~PERF_ATTACH_GROUP
;
1329 * If this is a sibling, remove it from its group.
1331 if (event
->group_leader
!= event
) {
1332 list_del_init(&event
->group_entry
);
1333 event
->group_leader
->nr_siblings
--;
1337 if (!list_empty(&event
->group_entry
))
1338 list
= &event
->group_entry
;
1341 * If this was a group event with sibling events then
1342 * upgrade the siblings to singleton events by adding them
1343 * to whatever list we are on.
1345 list_for_each_entry_safe(sibling
, tmp
, &event
->sibling_list
, group_entry
) {
1347 list_move_tail(&sibling
->group_entry
, list
);
1348 sibling
->group_leader
= sibling
;
1350 /* Inherit group flags from the previous leader */
1351 sibling
->group_flags
= event
->group_flags
;
1355 perf_event__header_size(event
->group_leader
);
1357 list_for_each_entry(tmp
, &event
->group_leader
->sibling_list
, group_entry
)
1358 perf_event__header_size(tmp
);
1362 event_filter_match(struct perf_event
*event
)
1364 return (event
->cpu
== -1 || event
->cpu
== smp_processor_id())
1365 && perf_cgroup_match(event
);
1369 event_sched_out(struct perf_event
*event
,
1370 struct perf_cpu_context
*cpuctx
,
1371 struct perf_event_context
*ctx
)
1373 u64 tstamp
= perf_event_time(event
);
1376 * An event which could not be activated because of
1377 * filter mismatch still needs to have its timings
1378 * maintained, otherwise bogus information is return
1379 * via read() for time_enabled, time_running:
1381 if (event
->state
== PERF_EVENT_STATE_INACTIVE
1382 && !event_filter_match(event
)) {
1383 delta
= tstamp
- event
->tstamp_stopped
;
1384 event
->tstamp_running
+= delta
;
1385 event
->tstamp_stopped
= tstamp
;
1388 if (event
->state
!= PERF_EVENT_STATE_ACTIVE
)
1391 event
->state
= PERF_EVENT_STATE_INACTIVE
;
1392 if (event
->pending_disable
) {
1393 event
->pending_disable
= 0;
1394 event
->state
= PERF_EVENT_STATE_OFF
;
1396 event
->tstamp_stopped
= tstamp
;
1397 event
->pmu
->del(event
, 0);
1400 if (!is_software_event(event
))
1401 cpuctx
->active_oncpu
--;
1403 if (event
->attr
.freq
&& event
->attr
.sample_freq
)
1405 if (event
->attr
.exclusive
|| !cpuctx
->active_oncpu
)
1406 cpuctx
->exclusive
= 0;
1410 group_sched_out(struct perf_event
*group_event
,
1411 struct perf_cpu_context
*cpuctx
,
1412 struct perf_event_context
*ctx
)
1414 struct perf_event
*event
;
1415 int state
= group_event
->state
;
1417 event_sched_out(group_event
, cpuctx
, ctx
);
1420 * Schedule out siblings (if any):
1422 list_for_each_entry(event
, &group_event
->sibling_list
, group_entry
)
1423 event_sched_out(event
, cpuctx
, ctx
);
1425 if (state
== PERF_EVENT_STATE_ACTIVE
&& group_event
->attr
.exclusive
)
1426 cpuctx
->exclusive
= 0;
1429 struct remove_event
{
1430 struct perf_event
*event
;
1435 * Cross CPU call to remove a performance event
1437 * We disable the event on the hardware level first. After that we
1438 * remove it from the context list.
1440 static int __perf_remove_from_context(void *info
)
1442 struct remove_event
*re
= info
;
1443 struct perf_event
*event
= re
->event
;
1444 struct perf_event_context
*ctx
= event
->ctx
;
1445 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
1447 raw_spin_lock(&ctx
->lock
);
1448 event_sched_out(event
, cpuctx
, ctx
);
1449 if (re
->detach_group
)
1450 perf_group_detach(event
);
1451 list_del_event(event
, ctx
);
1452 if (!ctx
->nr_events
&& cpuctx
->task_ctx
== ctx
) {
1454 cpuctx
->task_ctx
= NULL
;
1456 raw_spin_unlock(&ctx
->lock
);
1463 * Remove the event from a task's (or a CPU's) list of events.
1465 * CPU events are removed with a smp call. For task events we only
1466 * call when the task is on a CPU.
1468 * If event->ctx is a cloned context, callers must make sure that
1469 * every task struct that event->ctx->task could possibly point to
1470 * remains valid. This is OK when called from perf_release since
1471 * that only calls us on the top-level context, which can't be a clone.
1472 * When called from perf_event_exit_task, it's OK because the
1473 * context has been detached from its task.
1475 static void perf_remove_from_context(struct perf_event
*event
, bool detach_group
)
1477 struct perf_event_context
*ctx
= event
->ctx
;
1478 struct task_struct
*task
= ctx
->task
;
1479 struct remove_event re
= {
1481 .detach_group
= detach_group
,
1484 lockdep_assert_held(&ctx
->mutex
);
1488 * Per cpu events are removed via an smp call and
1489 * the removal is always successful.
1491 cpu_function_call(event
->cpu
, __perf_remove_from_context
, &re
);
1496 if (!task_function_call(task
, __perf_remove_from_context
, &re
))
1499 raw_spin_lock_irq(&ctx
->lock
);
1501 * If we failed to find a running task, but find the context active now
1502 * that we've acquired the ctx->lock, retry.
1504 if (ctx
->is_active
) {
1505 raw_spin_unlock_irq(&ctx
->lock
);
1510 * Since the task isn't running, its safe to remove the event, us
1511 * holding the ctx->lock ensures the task won't get scheduled in.
1514 perf_group_detach(event
);
1515 list_del_event(event
, ctx
);
1516 raw_spin_unlock_irq(&ctx
->lock
);
1520 * Cross CPU call to disable a performance event
1522 int __perf_event_disable(void *info
)
1524 struct perf_event
*event
= info
;
1525 struct perf_event_context
*ctx
= event
->ctx
;
1526 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
1529 * If this is a per-task event, need to check whether this
1530 * event's task is the current task on this cpu.
1532 * Can trigger due to concurrent perf_event_context_sched_out()
1533 * flipping contexts around.
1535 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
)
1538 raw_spin_lock(&ctx
->lock
);
1541 * If the event is on, turn it off.
1542 * If it is in error state, leave it in error state.
1544 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
) {
1545 update_context_time(ctx
);
1546 update_cgrp_time_from_event(event
);
1547 update_group_times(event
);
1548 if (event
== event
->group_leader
)
1549 group_sched_out(event
, cpuctx
, ctx
);
1551 event_sched_out(event
, cpuctx
, ctx
);
1552 event
->state
= PERF_EVENT_STATE_OFF
;
1555 raw_spin_unlock(&ctx
->lock
);
1563 * If event->ctx is a cloned context, callers must make sure that
1564 * every task struct that event->ctx->task could possibly point to
1565 * remains valid. This condition is satisifed when called through
1566 * perf_event_for_each_child or perf_event_for_each because they
1567 * hold the top-level event's child_mutex, so any descendant that
1568 * goes to exit will block in sync_child_event.
1569 * When called from perf_pending_event it's OK because event->ctx
1570 * is the current context on this CPU and preemption is disabled,
1571 * hence we can't get into perf_event_task_sched_out for this context.
1573 void perf_event_disable(struct perf_event
*event
)
1575 struct perf_event_context
*ctx
= event
->ctx
;
1576 struct task_struct
*task
= ctx
->task
;
1580 * Disable the event on the cpu that it's on
1582 cpu_function_call(event
->cpu
, __perf_event_disable
, event
);
1587 if (!task_function_call(task
, __perf_event_disable
, event
))
1590 raw_spin_lock_irq(&ctx
->lock
);
1592 * If the event is still active, we need to retry the cross-call.
1594 if (event
->state
== PERF_EVENT_STATE_ACTIVE
) {
1595 raw_spin_unlock_irq(&ctx
->lock
);
1597 * Reload the task pointer, it might have been changed by
1598 * a concurrent perf_event_context_sched_out().
1605 * Since we have the lock this context can't be scheduled
1606 * in, so we can change the state safely.
1608 if (event
->state
== PERF_EVENT_STATE_INACTIVE
) {
1609 update_group_times(event
);
1610 event
->state
= PERF_EVENT_STATE_OFF
;
1612 raw_spin_unlock_irq(&ctx
->lock
);
1614 EXPORT_SYMBOL_GPL(perf_event_disable
);
1616 static void perf_set_shadow_time(struct perf_event
*event
,
1617 struct perf_event_context
*ctx
,
1621 * use the correct time source for the time snapshot
1623 * We could get by without this by leveraging the
1624 * fact that to get to this function, the caller
1625 * has most likely already called update_context_time()
1626 * and update_cgrp_time_xx() and thus both timestamp
1627 * are identical (or very close). Given that tstamp is,
1628 * already adjusted for cgroup, we could say that:
1629 * tstamp - ctx->timestamp
1631 * tstamp - cgrp->timestamp.
1633 * Then, in perf_output_read(), the calculation would
1634 * work with no changes because:
1635 * - event is guaranteed scheduled in
1636 * - no scheduled out in between
1637 * - thus the timestamp would be the same
1639 * But this is a bit hairy.
1641 * So instead, we have an explicit cgroup call to remain
1642 * within the time time source all along. We believe it
1643 * is cleaner and simpler to understand.
1645 if (is_cgroup_event(event
))
1646 perf_cgroup_set_shadow_time(event
, tstamp
);
1648 event
->shadow_ctx_time
= tstamp
- ctx
->timestamp
;
1651 #define MAX_INTERRUPTS (~0ULL)
1653 static void perf_log_throttle(struct perf_event
*event
, int enable
);
1656 event_sched_in(struct perf_event
*event
,
1657 struct perf_cpu_context
*cpuctx
,
1658 struct perf_event_context
*ctx
)
1660 u64 tstamp
= perf_event_time(event
);
1662 if (event
->state
<= PERF_EVENT_STATE_OFF
)
1665 event
->state
= PERF_EVENT_STATE_ACTIVE
;
1666 event
->oncpu
= smp_processor_id();
1669 * Unthrottle events, since we scheduled we might have missed several
1670 * ticks already, also for a heavily scheduling task there is little
1671 * guarantee it'll get a tick in a timely manner.
1673 if (unlikely(event
->hw
.interrupts
== MAX_INTERRUPTS
)) {
1674 perf_log_throttle(event
, 1);
1675 event
->hw
.interrupts
= 0;
1679 * The new state must be visible before we turn it on in the hardware:
1683 if (event
->pmu
->add(event
, PERF_EF_START
)) {
1684 event
->state
= PERF_EVENT_STATE_INACTIVE
;
1689 event
->tstamp_running
+= tstamp
- event
->tstamp_stopped
;
1691 perf_set_shadow_time(event
, ctx
, tstamp
);
1693 if (!is_software_event(event
))
1694 cpuctx
->active_oncpu
++;
1696 if (event
->attr
.freq
&& event
->attr
.sample_freq
)
1699 if (event
->attr
.exclusive
)
1700 cpuctx
->exclusive
= 1;
1706 group_sched_in(struct perf_event
*group_event
,
1707 struct perf_cpu_context
*cpuctx
,
1708 struct perf_event_context
*ctx
)
1710 struct perf_event
*event
, *partial_group
= NULL
;
1711 struct pmu
*pmu
= group_event
->pmu
;
1712 u64 now
= ctx
->time
;
1713 bool simulate
= false;
1715 if (group_event
->state
== PERF_EVENT_STATE_OFF
)
1718 pmu
->start_txn(pmu
);
1720 if (event_sched_in(group_event
, cpuctx
, ctx
)) {
1721 pmu
->cancel_txn(pmu
);
1722 perf_cpu_hrtimer_restart(cpuctx
);
1727 * Schedule in siblings as one group (if any):
1729 list_for_each_entry(event
, &group_event
->sibling_list
, group_entry
) {
1730 if (event_sched_in(event
, cpuctx
, ctx
)) {
1731 partial_group
= event
;
1736 if (!pmu
->commit_txn(pmu
))
1741 * Groups can be scheduled in as one unit only, so undo any
1742 * partial group before returning:
1743 * The events up to the failed event are scheduled out normally,
1744 * tstamp_stopped will be updated.
1746 * The failed events and the remaining siblings need to have
1747 * their timings updated as if they had gone thru event_sched_in()
1748 * and event_sched_out(). This is required to get consistent timings
1749 * across the group. This also takes care of the case where the group
1750 * could never be scheduled by ensuring tstamp_stopped is set to mark
1751 * the time the event was actually stopped, such that time delta
1752 * calculation in update_event_times() is correct.
1754 list_for_each_entry(event
, &group_event
->sibling_list
, group_entry
) {
1755 if (event
== partial_group
)
1759 event
->tstamp_running
+= now
- event
->tstamp_stopped
;
1760 event
->tstamp_stopped
= now
;
1762 event_sched_out(event
, cpuctx
, ctx
);
1765 event_sched_out(group_event
, cpuctx
, ctx
);
1767 pmu
->cancel_txn(pmu
);
1769 perf_cpu_hrtimer_restart(cpuctx
);
1775 * Work out whether we can put this event group on the CPU now.
1777 static int group_can_go_on(struct perf_event
*event
,
1778 struct perf_cpu_context
*cpuctx
,
1782 * Groups consisting entirely of software events can always go on.
1784 if (event
->group_flags
& PERF_GROUP_SOFTWARE
)
1787 * If an exclusive group is already on, no other hardware
1790 if (cpuctx
->exclusive
)
1793 * If this group is exclusive and there are already
1794 * events on the CPU, it can't go on.
1796 if (event
->attr
.exclusive
&& cpuctx
->active_oncpu
)
1799 * Otherwise, try to add it if all previous groups were able
1805 static void add_event_to_ctx(struct perf_event
*event
,
1806 struct perf_event_context
*ctx
)
1808 u64 tstamp
= perf_event_time(event
);
1810 list_add_event(event
, ctx
);
1811 perf_group_attach(event
);
1812 event
->tstamp_enabled
= tstamp
;
1813 event
->tstamp_running
= tstamp
;
1814 event
->tstamp_stopped
= tstamp
;
1817 static void task_ctx_sched_out(struct perf_event_context
*ctx
);
1819 ctx_sched_in(struct perf_event_context
*ctx
,
1820 struct perf_cpu_context
*cpuctx
,
1821 enum event_type_t event_type
,
1822 struct task_struct
*task
);
1824 static void perf_event_sched_in(struct perf_cpu_context
*cpuctx
,
1825 struct perf_event_context
*ctx
,
1826 struct task_struct
*task
)
1828 cpu_ctx_sched_in(cpuctx
, EVENT_PINNED
, task
);
1830 ctx_sched_in(ctx
, cpuctx
, EVENT_PINNED
, task
);
1831 cpu_ctx_sched_in(cpuctx
, EVENT_FLEXIBLE
, task
);
1833 ctx_sched_in(ctx
, cpuctx
, EVENT_FLEXIBLE
, task
);
1837 * Cross CPU call to install and enable a performance event
1839 * Must be called with ctx->mutex held
1841 static int __perf_install_in_context(void *info
)
1843 struct perf_event
*event
= info
;
1844 struct perf_event_context
*ctx
= event
->ctx
;
1845 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
1846 struct perf_event_context
*task_ctx
= cpuctx
->task_ctx
;
1847 struct task_struct
*task
= current
;
1849 perf_ctx_lock(cpuctx
, task_ctx
);
1850 perf_pmu_disable(cpuctx
->ctx
.pmu
);
1853 * If there was an active task_ctx schedule it out.
1856 task_ctx_sched_out(task_ctx
);
1859 * If the context we're installing events in is not the
1860 * active task_ctx, flip them.
1862 if (ctx
->task
&& task_ctx
!= ctx
) {
1864 raw_spin_unlock(&task_ctx
->lock
);
1865 raw_spin_lock(&ctx
->lock
);
1870 cpuctx
->task_ctx
= task_ctx
;
1871 task
= task_ctx
->task
;
1874 cpu_ctx_sched_out(cpuctx
, EVENT_ALL
);
1876 update_context_time(ctx
);
1878 * update cgrp time only if current cgrp
1879 * matches event->cgrp. Must be done before
1880 * calling add_event_to_ctx()
1882 update_cgrp_time_from_event(event
);
1884 add_event_to_ctx(event
, ctx
);
1887 * Schedule everything back in
1889 perf_event_sched_in(cpuctx
, task_ctx
, task
);
1891 perf_pmu_enable(cpuctx
->ctx
.pmu
);
1892 perf_ctx_unlock(cpuctx
, task_ctx
);
1898 * Attach a performance event to a context
1900 * First we add the event to the list with the hardware enable bit
1901 * in event->hw_config cleared.
1903 * If the event is attached to a task which is on a CPU we use a smp
1904 * call to enable it in the task context. The task might have been
1905 * scheduled away, but we check this in the smp call again.
1908 perf_install_in_context(struct perf_event_context
*ctx
,
1909 struct perf_event
*event
,
1912 struct task_struct
*task
= ctx
->task
;
1914 lockdep_assert_held(&ctx
->mutex
);
1917 if (event
->cpu
!= -1)
1922 * Per cpu events are installed via an smp call and
1923 * the install is always successful.
1925 cpu_function_call(cpu
, __perf_install_in_context
, event
);
1930 if (!task_function_call(task
, __perf_install_in_context
, event
))
1933 raw_spin_lock_irq(&ctx
->lock
);
1935 * If we failed to find a running task, but find the context active now
1936 * that we've acquired the ctx->lock, retry.
1938 if (ctx
->is_active
) {
1939 raw_spin_unlock_irq(&ctx
->lock
);
1944 * Since the task isn't running, its safe to add the event, us holding
1945 * the ctx->lock ensures the task won't get scheduled in.
1947 add_event_to_ctx(event
, ctx
);
1948 raw_spin_unlock_irq(&ctx
->lock
);
1952 * Put a event into inactive state and update time fields.
1953 * Enabling the leader of a group effectively enables all
1954 * the group members that aren't explicitly disabled, so we
1955 * have to update their ->tstamp_enabled also.
1956 * Note: this works for group members as well as group leaders
1957 * since the non-leader members' sibling_lists will be empty.
1959 static void __perf_event_mark_enabled(struct perf_event
*event
)
1961 struct perf_event
*sub
;
1962 u64 tstamp
= perf_event_time(event
);
1964 event
->state
= PERF_EVENT_STATE_INACTIVE
;
1965 event
->tstamp_enabled
= tstamp
- event
->total_time_enabled
;
1966 list_for_each_entry(sub
, &event
->sibling_list
, group_entry
) {
1967 if (sub
->state
>= PERF_EVENT_STATE_INACTIVE
)
1968 sub
->tstamp_enabled
= tstamp
- sub
->total_time_enabled
;
1973 * Cross CPU call to enable a performance event
1975 static int __perf_event_enable(void *info
)
1977 struct perf_event
*event
= info
;
1978 struct perf_event_context
*ctx
= event
->ctx
;
1979 struct perf_event
*leader
= event
->group_leader
;
1980 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
1984 * There's a time window between 'ctx->is_active' check
1985 * in perf_event_enable function and this place having:
1987 * - ctx->lock unlocked
1989 * where the task could be killed and 'ctx' deactivated
1990 * by perf_event_exit_task.
1992 if (!ctx
->is_active
)
1995 raw_spin_lock(&ctx
->lock
);
1996 update_context_time(ctx
);
1998 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
)
2002 * set current task's cgroup time reference point
2004 perf_cgroup_set_timestamp(current
, ctx
);
2006 __perf_event_mark_enabled(event
);
2008 if (!event_filter_match(event
)) {
2009 if (is_cgroup_event(event
))
2010 perf_cgroup_defer_enabled(event
);
2015 * If the event is in a group and isn't the group leader,
2016 * then don't put it on unless the group is on.
2018 if (leader
!= event
&& leader
->state
!= PERF_EVENT_STATE_ACTIVE
)
2021 if (!group_can_go_on(event
, cpuctx
, 1)) {
2024 if (event
== leader
)
2025 err
= group_sched_in(event
, cpuctx
, ctx
);
2027 err
= event_sched_in(event
, cpuctx
, ctx
);
2032 * If this event can't go on and it's part of a
2033 * group, then the whole group has to come off.
2035 if (leader
!= event
) {
2036 group_sched_out(leader
, cpuctx
, ctx
);
2037 perf_cpu_hrtimer_restart(cpuctx
);
2039 if (leader
->attr
.pinned
) {
2040 update_group_times(leader
);
2041 leader
->state
= PERF_EVENT_STATE_ERROR
;
2046 raw_spin_unlock(&ctx
->lock
);
2054 * If event->ctx is a cloned context, callers must make sure that
2055 * every task struct that event->ctx->task could possibly point to
2056 * remains valid. This condition is satisfied when called through
2057 * perf_event_for_each_child or perf_event_for_each as described
2058 * for perf_event_disable.
2060 void perf_event_enable(struct perf_event
*event
)
2062 struct perf_event_context
*ctx
= event
->ctx
;
2063 struct task_struct
*task
= ctx
->task
;
2067 * Enable the event on the cpu that it's on
2069 cpu_function_call(event
->cpu
, __perf_event_enable
, event
);
2073 raw_spin_lock_irq(&ctx
->lock
);
2074 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
)
2078 * If the event is in error state, clear that first.
2079 * That way, if we see the event in error state below, we
2080 * know that it has gone back into error state, as distinct
2081 * from the task having been scheduled away before the
2082 * cross-call arrived.
2084 if (event
->state
== PERF_EVENT_STATE_ERROR
)
2085 event
->state
= PERF_EVENT_STATE_OFF
;
2088 if (!ctx
->is_active
) {
2089 __perf_event_mark_enabled(event
);
2093 raw_spin_unlock_irq(&ctx
->lock
);
2095 if (!task_function_call(task
, __perf_event_enable
, event
))
2098 raw_spin_lock_irq(&ctx
->lock
);
2101 * If the context is active and the event is still off,
2102 * we need to retry the cross-call.
2104 if (ctx
->is_active
&& event
->state
== PERF_EVENT_STATE_OFF
) {
2106 * task could have been flipped by a concurrent
2107 * perf_event_context_sched_out()
2114 raw_spin_unlock_irq(&ctx
->lock
);
2116 EXPORT_SYMBOL_GPL(perf_event_enable
);
2118 int perf_event_refresh(struct perf_event
*event
, int refresh
)
2121 * not supported on inherited events
2123 if (event
->attr
.inherit
|| !is_sampling_event(event
))
2126 atomic_add(refresh
, &event
->event_limit
);
2127 perf_event_enable(event
);
2131 EXPORT_SYMBOL_GPL(perf_event_refresh
);
2133 static void ctx_sched_out(struct perf_event_context
*ctx
,
2134 struct perf_cpu_context
*cpuctx
,
2135 enum event_type_t event_type
)
2137 struct perf_event
*event
;
2138 int is_active
= ctx
->is_active
;
2140 ctx
->is_active
&= ~event_type
;
2141 if (likely(!ctx
->nr_events
))
2144 update_context_time(ctx
);
2145 update_cgrp_time_from_cpuctx(cpuctx
);
2146 if (!ctx
->nr_active
)
2149 perf_pmu_disable(ctx
->pmu
);
2150 if ((is_active
& EVENT_PINNED
) && (event_type
& EVENT_PINNED
)) {
2151 list_for_each_entry(event
, &ctx
->pinned_groups
, group_entry
)
2152 group_sched_out(event
, cpuctx
, ctx
);
2155 if ((is_active
& EVENT_FLEXIBLE
) && (event_type
& EVENT_FLEXIBLE
)) {
2156 list_for_each_entry(event
, &ctx
->flexible_groups
, group_entry
)
2157 group_sched_out(event
, cpuctx
, ctx
);
2159 perf_pmu_enable(ctx
->pmu
);
2163 * Test whether two contexts are equivalent, i.e. whether they
2164 * have both been cloned from the same version of the same context
2165 * and they both have the same number of enabled events.
2166 * If the number of enabled events is the same, then the set
2167 * of enabled events should be the same, because these are both
2168 * inherited contexts, therefore we can't access individual events
2169 * in them directly with an fd; we can only enable/disable all
2170 * events via prctl, or enable/disable all events in a family
2171 * via ioctl, which will have the same effect on both contexts.
2173 static int context_equiv(struct perf_event_context
*ctx1
,
2174 struct perf_event_context
*ctx2
)
2176 return ctx1
->parent_ctx
&& ctx1
->parent_ctx
== ctx2
->parent_ctx
2177 && ctx1
->parent_gen
== ctx2
->parent_gen
2178 && !ctx1
->pin_count
&& !ctx2
->pin_count
;
2181 static void __perf_event_sync_stat(struct perf_event
*event
,
2182 struct perf_event
*next_event
)
2186 if (!event
->attr
.inherit_stat
)
2190 * Update the event value, we cannot use perf_event_read()
2191 * because we're in the middle of a context switch and have IRQs
2192 * disabled, which upsets smp_call_function_single(), however
2193 * we know the event must be on the current CPU, therefore we
2194 * don't need to use it.
2196 switch (event
->state
) {
2197 case PERF_EVENT_STATE_ACTIVE
:
2198 event
->pmu
->read(event
);
2201 case PERF_EVENT_STATE_INACTIVE
:
2202 update_event_times(event
);
2210 * In order to keep per-task stats reliable we need to flip the event
2211 * values when we flip the contexts.
2213 value
= local64_read(&next_event
->count
);
2214 value
= local64_xchg(&event
->count
, value
);
2215 local64_set(&next_event
->count
, value
);
2217 swap(event
->total_time_enabled
, next_event
->total_time_enabled
);
2218 swap(event
->total_time_running
, next_event
->total_time_running
);
2221 * Since we swizzled the values, update the user visible data too.
2223 perf_event_update_userpage(event
);
2224 perf_event_update_userpage(next_event
);
2227 static void perf_event_sync_stat(struct perf_event_context
*ctx
,
2228 struct perf_event_context
*next_ctx
)
2230 struct perf_event
*event
, *next_event
;
2235 update_context_time(ctx
);
2237 event
= list_first_entry(&ctx
->event_list
,
2238 struct perf_event
, event_entry
);
2240 next_event
= list_first_entry(&next_ctx
->event_list
,
2241 struct perf_event
, event_entry
);
2243 while (&event
->event_entry
!= &ctx
->event_list
&&
2244 &next_event
->event_entry
!= &next_ctx
->event_list
) {
2246 __perf_event_sync_stat(event
, next_event
);
2248 event
= list_next_entry(event
, event_entry
);
2249 next_event
= list_next_entry(next_event
, event_entry
);
2253 static void perf_event_context_sched_out(struct task_struct
*task
, int ctxn
,
2254 struct task_struct
*next
)
2256 struct perf_event_context
*ctx
= task
->perf_event_ctxp
[ctxn
];
2257 struct perf_event_context
*next_ctx
;
2258 struct perf_event_context
*parent
;
2259 struct perf_cpu_context
*cpuctx
;
2265 cpuctx
= __get_cpu_context(ctx
);
2266 if (!cpuctx
->task_ctx
)
2270 parent
= rcu_dereference(ctx
->parent_ctx
);
2271 next_ctx
= next
->perf_event_ctxp
[ctxn
];
2272 if (parent
&& next_ctx
&&
2273 rcu_dereference(next_ctx
->parent_ctx
) == parent
) {
2275 * Looks like the two contexts are clones, so we might be
2276 * able to optimize the context switch. We lock both
2277 * contexts and check that they are clones under the
2278 * lock (including re-checking that neither has been
2279 * uncloned in the meantime). It doesn't matter which
2280 * order we take the locks because no other cpu could
2281 * be trying to lock both of these tasks.
2283 raw_spin_lock(&ctx
->lock
);
2284 raw_spin_lock_nested(&next_ctx
->lock
, SINGLE_DEPTH_NESTING
);
2285 if (context_equiv(ctx
, next_ctx
)) {
2287 * XXX do we need a memory barrier of sorts
2288 * wrt to rcu_dereference() of perf_event_ctxp
2290 task
->perf_event_ctxp
[ctxn
] = next_ctx
;
2291 next
->perf_event_ctxp
[ctxn
] = ctx
;
2293 next_ctx
->task
= task
;
2296 perf_event_sync_stat(ctx
, next_ctx
);
2298 raw_spin_unlock(&next_ctx
->lock
);
2299 raw_spin_unlock(&ctx
->lock
);
2304 raw_spin_lock(&ctx
->lock
);
2305 ctx_sched_out(ctx
, cpuctx
, EVENT_ALL
);
2306 cpuctx
->task_ctx
= NULL
;
2307 raw_spin_unlock(&ctx
->lock
);
2311 #define for_each_task_context_nr(ctxn) \
2312 for ((ctxn) = 0; (ctxn) < perf_nr_task_contexts; (ctxn)++)
2315 * Called from scheduler to remove the events of the current task,
2316 * with interrupts disabled.
2318 * We stop each event and update the event value in event->count.
2320 * This does not protect us against NMI, but disable()
2321 * sets the disabled bit in the control field of event _before_
2322 * accessing the event control register. If a NMI hits, then it will
2323 * not restart the event.
2325 void __perf_event_task_sched_out(struct task_struct
*task
,
2326 struct task_struct
*next
)
2330 for_each_task_context_nr(ctxn
)
2331 perf_event_context_sched_out(task
, ctxn
, next
);
2334 * if cgroup events exist on this CPU, then we need
2335 * to check if we have to switch out PMU state.
2336 * cgroup event are system-wide mode only
2338 if (atomic_read(&__get_cpu_var(perf_cgroup_events
)))
2339 perf_cgroup_sched_out(task
, next
);
2342 static void task_ctx_sched_out(struct perf_event_context
*ctx
)
2344 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
2346 if (!cpuctx
->task_ctx
)
2349 if (WARN_ON_ONCE(ctx
!= cpuctx
->task_ctx
))
2352 ctx_sched_out(ctx
, cpuctx
, EVENT_ALL
);
2353 cpuctx
->task_ctx
= NULL
;
2357 * Called with IRQs disabled
2359 static void cpu_ctx_sched_out(struct perf_cpu_context
*cpuctx
,
2360 enum event_type_t event_type
)
2362 ctx_sched_out(&cpuctx
->ctx
, cpuctx
, event_type
);
2366 ctx_pinned_sched_in(struct perf_event_context
*ctx
,
2367 struct perf_cpu_context
*cpuctx
)
2369 struct perf_event
*event
;
2371 list_for_each_entry(event
, &ctx
->pinned_groups
, group_entry
) {
2372 if (event
->state
<= PERF_EVENT_STATE_OFF
)
2374 if (!event_filter_match(event
))
2377 /* may need to reset tstamp_enabled */
2378 if (is_cgroup_event(event
))
2379 perf_cgroup_mark_enabled(event
, ctx
);
2381 if (group_can_go_on(event
, cpuctx
, 1))
2382 group_sched_in(event
, cpuctx
, ctx
);
2385 * If this pinned group hasn't been scheduled,
2386 * put it in error state.
2388 if (event
->state
== PERF_EVENT_STATE_INACTIVE
) {
2389 update_group_times(event
);
2390 event
->state
= PERF_EVENT_STATE_ERROR
;
2396 ctx_flexible_sched_in(struct perf_event_context
*ctx
,
2397 struct perf_cpu_context
*cpuctx
)
2399 struct perf_event
*event
;
2402 list_for_each_entry(event
, &ctx
->flexible_groups
, group_entry
) {
2403 /* Ignore events in OFF or ERROR state */
2404 if (event
->state
<= PERF_EVENT_STATE_OFF
)
2407 * Listen to the 'cpu' scheduling filter constraint
2410 if (!event_filter_match(event
))
2413 /* may need to reset tstamp_enabled */
2414 if (is_cgroup_event(event
))
2415 perf_cgroup_mark_enabled(event
, ctx
);
2417 if (group_can_go_on(event
, cpuctx
, can_add_hw
)) {
2418 if (group_sched_in(event
, cpuctx
, ctx
))
2425 ctx_sched_in(struct perf_event_context
*ctx
,
2426 struct perf_cpu_context
*cpuctx
,
2427 enum event_type_t event_type
,
2428 struct task_struct
*task
)
2431 int is_active
= ctx
->is_active
;
2433 ctx
->is_active
|= event_type
;
2434 if (likely(!ctx
->nr_events
))
2438 ctx
->timestamp
= now
;
2439 perf_cgroup_set_timestamp(task
, ctx
);
2441 * First go through the list and put on any pinned groups
2442 * in order to give them the best chance of going on.
2444 if (!(is_active
& EVENT_PINNED
) && (event_type
& EVENT_PINNED
))
2445 ctx_pinned_sched_in(ctx
, cpuctx
);
2447 /* Then walk through the lower prio flexible groups */
2448 if (!(is_active
& EVENT_FLEXIBLE
) && (event_type
& EVENT_FLEXIBLE
))
2449 ctx_flexible_sched_in(ctx
, cpuctx
);
2452 static void cpu_ctx_sched_in(struct perf_cpu_context
*cpuctx
,
2453 enum event_type_t event_type
,
2454 struct task_struct
*task
)
2456 struct perf_event_context
*ctx
= &cpuctx
->ctx
;
2458 ctx_sched_in(ctx
, cpuctx
, event_type
, task
);
2461 static void perf_event_context_sched_in(struct perf_event_context
*ctx
,
2462 struct task_struct
*task
)
2464 struct perf_cpu_context
*cpuctx
;
2466 cpuctx
= __get_cpu_context(ctx
);
2467 if (cpuctx
->task_ctx
== ctx
)
2470 perf_ctx_lock(cpuctx
, ctx
);
2471 perf_pmu_disable(ctx
->pmu
);
2473 * We want to keep the following priority order:
2474 * cpu pinned (that don't need to move), task pinned,
2475 * cpu flexible, task flexible.
2477 cpu_ctx_sched_out(cpuctx
, EVENT_FLEXIBLE
);
2480 cpuctx
->task_ctx
= ctx
;
2482 perf_event_sched_in(cpuctx
, cpuctx
->task_ctx
, task
);
2484 perf_pmu_enable(ctx
->pmu
);
2485 perf_ctx_unlock(cpuctx
, ctx
);
2488 * Since these rotations are per-cpu, we need to ensure the
2489 * cpu-context we got scheduled on is actually rotating.
2491 perf_pmu_rotate_start(ctx
->pmu
);
2495 * When sampling the branck stack in system-wide, it may be necessary
2496 * to flush the stack on context switch. This happens when the branch
2497 * stack does not tag its entries with the pid of the current task.
2498 * Otherwise it becomes impossible to associate a branch entry with a
2499 * task. This ambiguity is more likely to appear when the branch stack
2500 * supports priv level filtering and the user sets it to monitor only
2501 * at the user level (which could be a useful measurement in system-wide
2502 * mode). In that case, the risk is high of having a branch stack with
2503 * branch from multiple tasks. Flushing may mean dropping the existing
2504 * entries or stashing them somewhere in the PMU specific code layer.
2506 * This function provides the context switch callback to the lower code
2507 * layer. It is invoked ONLY when there is at least one system-wide context
2508 * with at least one active event using taken branch sampling.
2510 static void perf_branch_stack_sched_in(struct task_struct
*prev
,
2511 struct task_struct
*task
)
2513 struct perf_cpu_context
*cpuctx
;
2515 unsigned long flags
;
2517 /* no need to flush branch stack if not changing task */
2521 local_irq_save(flags
);
2525 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
2526 cpuctx
= this_cpu_ptr(pmu
->pmu_cpu_context
);
2529 * check if the context has at least one
2530 * event using PERF_SAMPLE_BRANCH_STACK
2532 if (cpuctx
->ctx
.nr_branch_stack
> 0
2533 && pmu
->flush_branch_stack
) {
2535 pmu
= cpuctx
->ctx
.pmu
;
2537 perf_ctx_lock(cpuctx
, cpuctx
->task_ctx
);
2539 perf_pmu_disable(pmu
);
2541 pmu
->flush_branch_stack();
2543 perf_pmu_enable(pmu
);
2545 perf_ctx_unlock(cpuctx
, cpuctx
->task_ctx
);
2551 local_irq_restore(flags
);
2555 * Called from scheduler to add the events of the current task
2556 * with interrupts disabled.
2558 * We restore the event value and then enable it.
2560 * This does not protect us against NMI, but enable()
2561 * sets the enabled bit in the control field of event _before_
2562 * accessing the event control register. If a NMI hits, then it will
2563 * keep the event running.
2565 void __perf_event_task_sched_in(struct task_struct
*prev
,
2566 struct task_struct
*task
)
2568 struct perf_event_context
*ctx
;
2571 for_each_task_context_nr(ctxn
) {
2572 ctx
= task
->perf_event_ctxp
[ctxn
];
2576 perf_event_context_sched_in(ctx
, task
);
2579 * if cgroup events exist on this CPU, then we need
2580 * to check if we have to switch in PMU state.
2581 * cgroup event are system-wide mode only
2583 if (atomic_read(&__get_cpu_var(perf_cgroup_events
)))
2584 perf_cgroup_sched_in(prev
, task
);
2586 /* check for system-wide branch_stack events */
2587 if (atomic_read(&__get_cpu_var(perf_branch_stack_events
)))
2588 perf_branch_stack_sched_in(prev
, task
);
2591 static u64
perf_calculate_period(struct perf_event
*event
, u64 nsec
, u64 count
)
2593 u64 frequency
= event
->attr
.sample_freq
;
2594 u64 sec
= NSEC_PER_SEC
;
2595 u64 divisor
, dividend
;
2597 int count_fls
, nsec_fls
, frequency_fls
, sec_fls
;
2599 count_fls
= fls64(count
);
2600 nsec_fls
= fls64(nsec
);
2601 frequency_fls
= fls64(frequency
);
2605 * We got @count in @nsec, with a target of sample_freq HZ
2606 * the target period becomes:
2609 * period = -------------------
2610 * @nsec * sample_freq
2615 * Reduce accuracy by one bit such that @a and @b converge
2616 * to a similar magnitude.
2618 #define REDUCE_FLS(a, b) \
2620 if (a##_fls > b##_fls) { \
2630 * Reduce accuracy until either term fits in a u64, then proceed with
2631 * the other, so that finally we can do a u64/u64 division.
2633 while (count_fls
+ sec_fls
> 64 && nsec_fls
+ frequency_fls
> 64) {
2634 REDUCE_FLS(nsec
, frequency
);
2635 REDUCE_FLS(sec
, count
);
2638 if (count_fls
+ sec_fls
> 64) {
2639 divisor
= nsec
* frequency
;
2641 while (count_fls
+ sec_fls
> 64) {
2642 REDUCE_FLS(count
, sec
);
2646 dividend
= count
* sec
;
2648 dividend
= count
* sec
;
2650 while (nsec_fls
+ frequency_fls
> 64) {
2651 REDUCE_FLS(nsec
, frequency
);
2655 divisor
= nsec
* frequency
;
2661 return div64_u64(dividend
, divisor
);
2664 static DEFINE_PER_CPU(int, perf_throttled_count
);
2665 static DEFINE_PER_CPU(u64
, perf_throttled_seq
);
2667 static void perf_adjust_period(struct perf_event
*event
, u64 nsec
, u64 count
, bool disable
)
2669 struct hw_perf_event
*hwc
= &event
->hw
;
2670 s64 period
, sample_period
;
2673 period
= perf_calculate_period(event
, nsec
, count
);
2675 delta
= (s64
)(period
- hwc
->sample_period
);
2676 delta
= (delta
+ 7) / 8; /* low pass filter */
2678 sample_period
= hwc
->sample_period
+ delta
;
2683 hwc
->sample_period
= sample_period
;
2685 if (local64_read(&hwc
->period_left
) > 8*sample_period
) {
2687 event
->pmu
->stop(event
, PERF_EF_UPDATE
);
2689 local64_set(&hwc
->period_left
, 0);
2692 event
->pmu
->start(event
, PERF_EF_RELOAD
);
2697 * combine freq adjustment with unthrottling to avoid two passes over the
2698 * events. At the same time, make sure, having freq events does not change
2699 * the rate of unthrottling as that would introduce bias.
2701 static void perf_adjust_freq_unthr_context(struct perf_event_context
*ctx
,
2704 struct perf_event
*event
;
2705 struct hw_perf_event
*hwc
;
2706 u64 now
, period
= TICK_NSEC
;
2710 * only need to iterate over all events iff:
2711 * - context have events in frequency mode (needs freq adjust)
2712 * - there are events to unthrottle on this cpu
2714 if (!(ctx
->nr_freq
|| needs_unthr
))
2717 raw_spin_lock(&ctx
->lock
);
2718 perf_pmu_disable(ctx
->pmu
);
2720 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
2721 if (event
->state
!= PERF_EVENT_STATE_ACTIVE
)
2724 if (!event_filter_match(event
))
2729 if (hwc
->interrupts
== MAX_INTERRUPTS
) {
2730 hwc
->interrupts
= 0;
2731 perf_log_throttle(event
, 1);
2732 event
->pmu
->start(event
, 0);
2735 if (!event
->attr
.freq
|| !event
->attr
.sample_freq
)
2739 * stop the event and update event->count
2741 event
->pmu
->stop(event
, PERF_EF_UPDATE
);
2743 now
= local64_read(&event
->count
);
2744 delta
= now
- hwc
->freq_count_stamp
;
2745 hwc
->freq_count_stamp
= now
;
2749 * reload only if value has changed
2750 * we have stopped the event so tell that
2751 * to perf_adjust_period() to avoid stopping it
2755 perf_adjust_period(event
, period
, delta
, false);
2757 event
->pmu
->start(event
, delta
> 0 ? PERF_EF_RELOAD
: 0);
2760 perf_pmu_enable(ctx
->pmu
);
2761 raw_spin_unlock(&ctx
->lock
);
2765 * Round-robin a context's events:
2767 static void rotate_ctx(struct perf_event_context
*ctx
)
2770 * Rotate the first entry last of non-pinned groups. Rotation might be
2771 * disabled by the inheritance code.
2773 if (!ctx
->rotate_disable
)
2774 list_rotate_left(&ctx
->flexible_groups
);
2778 * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
2779 * because they're strictly cpu affine and rotate_start is called with IRQs
2780 * disabled, while rotate_context is called from IRQ context.
2782 static int perf_rotate_context(struct perf_cpu_context
*cpuctx
)
2784 struct perf_event_context
*ctx
= NULL
;
2785 int rotate
= 0, remove
= 1;
2787 if (cpuctx
->ctx
.nr_events
) {
2789 if (cpuctx
->ctx
.nr_events
!= cpuctx
->ctx
.nr_active
)
2793 ctx
= cpuctx
->task_ctx
;
2794 if (ctx
&& ctx
->nr_events
) {
2796 if (ctx
->nr_events
!= ctx
->nr_active
)
2803 perf_ctx_lock(cpuctx
, cpuctx
->task_ctx
);
2804 perf_pmu_disable(cpuctx
->ctx
.pmu
);
2806 cpu_ctx_sched_out(cpuctx
, EVENT_FLEXIBLE
);
2808 ctx_sched_out(ctx
, cpuctx
, EVENT_FLEXIBLE
);
2810 rotate_ctx(&cpuctx
->ctx
);
2814 perf_event_sched_in(cpuctx
, ctx
, current
);
2816 perf_pmu_enable(cpuctx
->ctx
.pmu
);
2817 perf_ctx_unlock(cpuctx
, cpuctx
->task_ctx
);
2820 list_del_init(&cpuctx
->rotation_list
);
2825 #ifdef CONFIG_NO_HZ_FULL
2826 bool perf_event_can_stop_tick(void)
2828 if (atomic_read(&nr_freq_events
) ||
2829 __this_cpu_read(perf_throttled_count
))
2836 void perf_event_task_tick(void)
2838 struct list_head
*head
= &__get_cpu_var(rotation_list
);
2839 struct perf_cpu_context
*cpuctx
, *tmp
;
2840 struct perf_event_context
*ctx
;
2843 WARN_ON(!irqs_disabled());
2845 __this_cpu_inc(perf_throttled_seq
);
2846 throttled
= __this_cpu_xchg(perf_throttled_count
, 0);
2848 list_for_each_entry_safe(cpuctx
, tmp
, head
, rotation_list
) {
2850 perf_adjust_freq_unthr_context(ctx
, throttled
);
2852 ctx
= cpuctx
->task_ctx
;
2854 perf_adjust_freq_unthr_context(ctx
, throttled
);
2858 static int event_enable_on_exec(struct perf_event
*event
,
2859 struct perf_event_context
*ctx
)
2861 if (!event
->attr
.enable_on_exec
)
2864 event
->attr
.enable_on_exec
= 0;
2865 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
)
2868 __perf_event_mark_enabled(event
);
2874 * Enable all of a task's events that have been marked enable-on-exec.
2875 * This expects task == current.
2877 static void perf_event_enable_on_exec(struct perf_event_context
*ctx
)
2879 struct perf_event
*event
;
2880 unsigned long flags
;
2884 local_irq_save(flags
);
2885 if (!ctx
|| !ctx
->nr_events
)
2889 * We must ctxsw out cgroup events to avoid conflict
2890 * when invoking perf_task_event_sched_in() later on
2891 * in this function. Otherwise we end up trying to
2892 * ctxswin cgroup events which are already scheduled
2895 perf_cgroup_sched_out(current
, NULL
);
2897 raw_spin_lock(&ctx
->lock
);
2898 task_ctx_sched_out(ctx
);
2900 list_for_each_entry(event
, &ctx
->event_list
, event_entry
) {
2901 ret
= event_enable_on_exec(event
, ctx
);
2907 * Unclone this context if we enabled any event.
2912 raw_spin_unlock(&ctx
->lock
);
2915 * Also calls ctxswin for cgroup events, if any:
2917 perf_event_context_sched_in(ctx
, ctx
->task
);
2919 local_irq_restore(flags
);
2923 * Cross CPU call to read the hardware event
2925 static void __perf_event_read(void *info
)
2927 struct perf_event
*event
= info
;
2928 struct perf_event_context
*ctx
= event
->ctx
;
2929 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
2932 * If this is a task context, we need to check whether it is
2933 * the current task context of this cpu. If not it has been
2934 * scheduled out before the smp call arrived. In that case
2935 * event->count would have been updated to a recent sample
2936 * when the event was scheduled out.
2938 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
)
2941 raw_spin_lock(&ctx
->lock
);
2942 if (ctx
->is_active
) {
2943 update_context_time(ctx
);
2944 update_cgrp_time_from_event(event
);
2946 update_event_times(event
);
2947 if (event
->state
== PERF_EVENT_STATE_ACTIVE
)
2948 event
->pmu
->read(event
);
2949 raw_spin_unlock(&ctx
->lock
);
2952 static inline u64
perf_event_count(struct perf_event
*event
)
2954 return local64_read(&event
->count
) + atomic64_read(&event
->child_count
);
2957 static u64
perf_event_read(struct perf_event
*event
)
2960 * If event is enabled and currently active on a CPU, update the
2961 * value in the event structure:
2963 if (event
->state
== PERF_EVENT_STATE_ACTIVE
) {
2964 smp_call_function_single(event
->oncpu
,
2965 __perf_event_read
, event
, 1);
2966 } else if (event
->state
== PERF_EVENT_STATE_INACTIVE
) {
2967 struct perf_event_context
*ctx
= event
->ctx
;
2968 unsigned long flags
;
2970 raw_spin_lock_irqsave(&ctx
->lock
, flags
);
2972 * may read while context is not active
2973 * (e.g., thread is blocked), in that case
2974 * we cannot update context time
2976 if (ctx
->is_active
) {
2977 update_context_time(ctx
);
2978 update_cgrp_time_from_event(event
);
2980 update_event_times(event
);
2981 raw_spin_unlock_irqrestore(&ctx
->lock
, flags
);
2984 return perf_event_count(event
);
2988 * Initialize the perf_event context in a task_struct:
2990 static void __perf_event_init_context(struct perf_event_context
*ctx
)
2992 raw_spin_lock_init(&ctx
->lock
);
2993 mutex_init(&ctx
->mutex
);
2994 INIT_LIST_HEAD(&ctx
->pinned_groups
);
2995 INIT_LIST_HEAD(&ctx
->flexible_groups
);
2996 INIT_LIST_HEAD(&ctx
->event_list
);
2997 atomic_set(&ctx
->refcount
, 1);
3000 static struct perf_event_context
*
3001 alloc_perf_context(struct pmu
*pmu
, struct task_struct
*task
)
3003 struct perf_event_context
*ctx
;
3005 ctx
= kzalloc(sizeof(struct perf_event_context
), GFP_KERNEL
);
3009 __perf_event_init_context(ctx
);
3012 get_task_struct(task
);
3019 static struct task_struct
*
3020 find_lively_task_by_vpid(pid_t vpid
)
3022 struct task_struct
*task
;
3029 task
= find_task_by_vpid(vpid
);
3031 get_task_struct(task
);
3035 return ERR_PTR(-ESRCH
);
3037 /* Reuse ptrace permission checks for now. */
3039 if (!ptrace_may_access(task
, PTRACE_MODE_READ
))
3044 put_task_struct(task
);
3045 return ERR_PTR(err
);
3050 * Returns a matching context with refcount and pincount.
3052 static struct perf_event_context
*
3053 find_get_context(struct pmu
*pmu
, struct task_struct
*task
, int cpu
)
3055 struct perf_event_context
*ctx
;
3056 struct perf_cpu_context
*cpuctx
;
3057 unsigned long flags
;
3061 /* Must be root to operate on a CPU event: */
3062 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN
))
3063 return ERR_PTR(-EACCES
);
3066 * We could be clever and allow to attach a event to an
3067 * offline CPU and activate it when the CPU comes up, but
3070 if (!cpu_online(cpu
))
3071 return ERR_PTR(-ENODEV
);
3073 cpuctx
= per_cpu_ptr(pmu
->pmu_cpu_context
, cpu
);
3082 ctxn
= pmu
->task_ctx_nr
;
3087 ctx
= perf_lock_task_context(task
, ctxn
, &flags
);
3091 raw_spin_unlock_irqrestore(&ctx
->lock
, flags
);
3093 ctx
= alloc_perf_context(pmu
, task
);
3099 mutex_lock(&task
->perf_event_mutex
);
3101 * If it has already passed perf_event_exit_task().
3102 * we must see PF_EXITING, it takes this mutex too.
3104 if (task
->flags
& PF_EXITING
)
3106 else if (task
->perf_event_ctxp
[ctxn
])
3111 rcu_assign_pointer(task
->perf_event_ctxp
[ctxn
], ctx
);
3113 mutex_unlock(&task
->perf_event_mutex
);
3115 if (unlikely(err
)) {
3127 return ERR_PTR(err
);
3130 static void perf_event_free_filter(struct perf_event
*event
);
3132 static void free_event_rcu(struct rcu_head
*head
)
3134 struct perf_event
*event
;
3136 event
= container_of(head
, struct perf_event
, rcu_head
);
3138 put_pid_ns(event
->ns
);
3139 perf_event_free_filter(event
);
3143 static void ring_buffer_put(struct ring_buffer
*rb
);
3144 static void ring_buffer_detach(struct perf_event
*event
, struct ring_buffer
*rb
);
3146 static void unaccount_event_cpu(struct perf_event
*event
, int cpu
)
3151 if (has_branch_stack(event
)) {
3152 if (!(event
->attach_state
& PERF_ATTACH_TASK
))
3153 atomic_dec(&per_cpu(perf_branch_stack_events
, cpu
));
3155 if (is_cgroup_event(event
))
3156 atomic_dec(&per_cpu(perf_cgroup_events
, cpu
));
3159 static void unaccount_event(struct perf_event
*event
)
3164 if (event
->attach_state
& PERF_ATTACH_TASK
)
3165 static_key_slow_dec_deferred(&perf_sched_events
);
3166 if (event
->attr
.mmap
|| event
->attr
.mmap_data
)
3167 atomic_dec(&nr_mmap_events
);
3168 if (event
->attr
.comm
)
3169 atomic_dec(&nr_comm_events
);
3170 if (event
->attr
.task
)
3171 atomic_dec(&nr_task_events
);
3172 if (event
->attr
.freq
)
3173 atomic_dec(&nr_freq_events
);
3174 if (is_cgroup_event(event
))
3175 static_key_slow_dec_deferred(&perf_sched_events
);
3176 if (has_branch_stack(event
))
3177 static_key_slow_dec_deferred(&perf_sched_events
);
3179 unaccount_event_cpu(event
, event
->cpu
);
3182 static void __free_event(struct perf_event
*event
)
3184 if (!event
->parent
) {
3185 if (event
->attr
.sample_type
& PERF_SAMPLE_CALLCHAIN
)
3186 put_callchain_buffers();
3190 event
->destroy(event
);
3193 put_ctx(event
->ctx
);
3195 call_rcu(&event
->rcu_head
, free_event_rcu
);
3197 static void free_event(struct perf_event
*event
)
3199 irq_work_sync(&event
->pending
);
3201 unaccount_event(event
);
3204 struct ring_buffer
*rb
;
3207 * Can happen when we close an event with re-directed output.
3209 * Since we have a 0 refcount, perf_mmap_close() will skip
3210 * over us; possibly making our ring_buffer_put() the last.
3212 mutex_lock(&event
->mmap_mutex
);
3215 rcu_assign_pointer(event
->rb
, NULL
);
3216 ring_buffer_detach(event
, rb
);
3217 ring_buffer_put(rb
); /* could be last */
3219 mutex_unlock(&event
->mmap_mutex
);
3222 if (is_cgroup_event(event
))
3223 perf_detach_cgroup(event
);
3226 __free_event(event
);
3229 int perf_event_release_kernel(struct perf_event
*event
)
3231 struct perf_event_context
*ctx
= event
->ctx
;
3233 WARN_ON_ONCE(ctx
->parent_ctx
);
3235 * There are two ways this annotation is useful:
3237 * 1) there is a lock recursion from perf_event_exit_task
3238 * see the comment there.
3240 * 2) there is a lock-inversion with mmap_sem through
3241 * perf_event_read_group(), which takes faults while
3242 * holding ctx->mutex, however this is called after
3243 * the last filedesc died, so there is no possibility
3244 * to trigger the AB-BA case.
3246 mutex_lock_nested(&ctx
->mutex
, SINGLE_DEPTH_NESTING
);
3247 perf_remove_from_context(event
, true);
3248 mutex_unlock(&ctx
->mutex
);
3254 EXPORT_SYMBOL_GPL(perf_event_release_kernel
);
3257 * Called when the last reference to the file is gone.
3259 static void put_event(struct perf_event
*event
)
3261 struct task_struct
*owner
;
3263 if (!atomic_long_dec_and_test(&event
->refcount
))
3267 owner
= ACCESS_ONCE(event
->owner
);
3269 * Matches the smp_wmb() in perf_event_exit_task(). If we observe
3270 * !owner it means the list deletion is complete and we can indeed
3271 * free this event, otherwise we need to serialize on
3272 * owner->perf_event_mutex.
3274 smp_read_barrier_depends();
3277 * Since delayed_put_task_struct() also drops the last
3278 * task reference we can safely take a new reference
3279 * while holding the rcu_read_lock().
3281 get_task_struct(owner
);
3286 mutex_lock(&owner
->perf_event_mutex
);
3288 * We have to re-check the event->owner field, if it is cleared
3289 * we raced with perf_event_exit_task(), acquiring the mutex
3290 * ensured they're done, and we can proceed with freeing the
3294 list_del_init(&event
->owner_entry
);
3295 mutex_unlock(&owner
->perf_event_mutex
);
3296 put_task_struct(owner
);
3299 perf_event_release_kernel(event
);
3302 static int perf_release(struct inode
*inode
, struct file
*file
)
3304 put_event(file
->private_data
);
3308 u64
perf_event_read_value(struct perf_event
*event
, u64
*enabled
, u64
*running
)
3310 struct perf_event
*child
;
3316 mutex_lock(&event
->child_mutex
);
3317 total
+= perf_event_read(event
);
3318 *enabled
+= event
->total_time_enabled
+
3319 atomic64_read(&event
->child_total_time_enabled
);
3320 *running
+= event
->total_time_running
+
3321 atomic64_read(&event
->child_total_time_running
);
3323 list_for_each_entry(child
, &event
->child_list
, child_list
) {
3324 total
+= perf_event_read(child
);
3325 *enabled
+= child
->total_time_enabled
;
3326 *running
+= child
->total_time_running
;
3328 mutex_unlock(&event
->child_mutex
);
3332 EXPORT_SYMBOL_GPL(perf_event_read_value
);
3334 static int perf_event_read_group(struct perf_event
*event
,
3335 u64 read_format
, char __user
*buf
)
3337 struct perf_event
*leader
= event
->group_leader
, *sub
;
3338 int n
= 0, size
= 0, ret
= -EFAULT
;
3339 struct perf_event_context
*ctx
= leader
->ctx
;
3341 u64 count
, enabled
, running
;
3343 mutex_lock(&ctx
->mutex
);
3344 count
= perf_event_read_value(leader
, &enabled
, &running
);
3346 values
[n
++] = 1 + leader
->nr_siblings
;
3347 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
3348 values
[n
++] = enabled
;
3349 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
3350 values
[n
++] = running
;
3351 values
[n
++] = count
;
3352 if (read_format
& PERF_FORMAT_ID
)
3353 values
[n
++] = primary_event_id(leader
);
3355 size
= n
* sizeof(u64
);
3357 if (copy_to_user(buf
, values
, size
))
3362 list_for_each_entry(sub
, &leader
->sibling_list
, group_entry
) {
3365 values
[n
++] = perf_event_read_value(sub
, &enabled
, &running
);
3366 if (read_format
& PERF_FORMAT_ID
)
3367 values
[n
++] = primary_event_id(sub
);
3369 size
= n
* sizeof(u64
);
3371 if (copy_to_user(buf
+ ret
, values
, size
)) {
3379 mutex_unlock(&ctx
->mutex
);
3384 static int perf_event_read_one(struct perf_event
*event
,
3385 u64 read_format
, char __user
*buf
)
3387 u64 enabled
, running
;
3391 values
[n
++] = perf_event_read_value(event
, &enabled
, &running
);
3392 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
3393 values
[n
++] = enabled
;
3394 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
3395 values
[n
++] = running
;
3396 if (read_format
& PERF_FORMAT_ID
)
3397 values
[n
++] = primary_event_id(event
);
3399 if (copy_to_user(buf
, values
, n
* sizeof(u64
)))
3402 return n
* sizeof(u64
);
3406 * Read the performance event - simple non blocking version for now
3409 perf_read_hw(struct perf_event
*event
, char __user
*buf
, size_t count
)
3411 u64 read_format
= event
->attr
.read_format
;
3415 * Return end-of-file for a read on a event that is in
3416 * error state (i.e. because it was pinned but it couldn't be
3417 * scheduled on to the CPU at some point).
3419 if (event
->state
== PERF_EVENT_STATE_ERROR
)
3422 if (count
< event
->read_size
)
3425 WARN_ON_ONCE(event
->ctx
->parent_ctx
);
3426 if (read_format
& PERF_FORMAT_GROUP
)
3427 ret
= perf_event_read_group(event
, read_format
, buf
);
3429 ret
= perf_event_read_one(event
, read_format
, buf
);
3435 perf_read(struct file
*file
, char __user
*buf
, size_t count
, loff_t
*ppos
)
3437 struct perf_event
*event
= file
->private_data
;
3439 return perf_read_hw(event
, buf
, count
);
3442 static unsigned int perf_poll(struct file
*file
, poll_table
*wait
)
3444 struct perf_event
*event
= file
->private_data
;
3445 struct ring_buffer
*rb
;
3446 unsigned int events
= POLL_HUP
;
3449 * Pin the event->rb by taking event->mmap_mutex; otherwise
3450 * perf_event_set_output() can swizzle our rb and make us miss wakeups.
3452 mutex_lock(&event
->mmap_mutex
);
3455 events
= atomic_xchg(&rb
->poll
, 0);
3456 mutex_unlock(&event
->mmap_mutex
);
3458 poll_wait(file
, &event
->waitq
, wait
);
3463 static void perf_event_reset(struct perf_event
*event
)
3465 (void)perf_event_read(event
);
3466 local64_set(&event
->count
, 0);
3467 perf_event_update_userpage(event
);
3471 * Holding the top-level event's child_mutex means that any
3472 * descendant process that has inherited this event will block
3473 * in sync_child_event if it goes to exit, thus satisfying the
3474 * task existence requirements of perf_event_enable/disable.
3476 static void perf_event_for_each_child(struct perf_event
*event
,
3477 void (*func
)(struct perf_event
*))
3479 struct perf_event
*child
;
3481 WARN_ON_ONCE(event
->ctx
->parent_ctx
);
3482 mutex_lock(&event
->child_mutex
);
3484 list_for_each_entry(child
, &event
->child_list
, child_list
)
3486 mutex_unlock(&event
->child_mutex
);
3489 static void perf_event_for_each(struct perf_event
*event
,
3490 void (*func
)(struct perf_event
*))
3492 struct perf_event_context
*ctx
= event
->ctx
;
3493 struct perf_event
*sibling
;
3495 WARN_ON_ONCE(ctx
->parent_ctx
);
3496 mutex_lock(&ctx
->mutex
);
3497 event
= event
->group_leader
;
3499 perf_event_for_each_child(event
, func
);
3500 list_for_each_entry(sibling
, &event
->sibling_list
, group_entry
)
3501 perf_event_for_each_child(sibling
, func
);
3502 mutex_unlock(&ctx
->mutex
);
3505 static int perf_event_period(struct perf_event
*event
, u64 __user
*arg
)
3507 struct perf_event_context
*ctx
= event
->ctx
;
3511 if (!is_sampling_event(event
))
3514 if (copy_from_user(&value
, arg
, sizeof(value
)))
3520 raw_spin_lock_irq(&ctx
->lock
);
3521 if (event
->attr
.freq
) {
3522 if (value
> sysctl_perf_event_sample_rate
) {
3527 event
->attr
.sample_freq
= value
;
3529 event
->attr
.sample_period
= value
;
3530 event
->hw
.sample_period
= value
;
3533 raw_spin_unlock_irq(&ctx
->lock
);
3538 static const struct file_operations perf_fops
;
3540 static inline int perf_fget_light(int fd
, struct fd
*p
)
3542 struct fd f
= fdget(fd
);
3546 if (f
.file
->f_op
!= &perf_fops
) {
3554 static int perf_event_set_output(struct perf_event
*event
,
3555 struct perf_event
*output_event
);
3556 static int perf_event_set_filter(struct perf_event
*event
, void __user
*arg
);
3558 static long perf_ioctl(struct file
*file
, unsigned int cmd
, unsigned long arg
)
3560 struct perf_event
*event
= file
->private_data
;
3561 void (*func
)(struct perf_event
*);
3565 case PERF_EVENT_IOC_ENABLE
:
3566 func
= perf_event_enable
;
3568 case PERF_EVENT_IOC_DISABLE
:
3569 func
= perf_event_disable
;
3571 case PERF_EVENT_IOC_RESET
:
3572 func
= perf_event_reset
;
3575 case PERF_EVENT_IOC_REFRESH
:
3576 return perf_event_refresh(event
, arg
);
3578 case PERF_EVENT_IOC_PERIOD
:
3579 return perf_event_period(event
, (u64 __user
*)arg
);
3581 case PERF_EVENT_IOC_ID
:
3583 u64 id
= primary_event_id(event
);
3585 if (copy_to_user((void __user
*)arg
, &id
, sizeof(id
)))
3590 case PERF_EVENT_IOC_SET_OUTPUT
:
3594 struct perf_event
*output_event
;
3596 ret
= perf_fget_light(arg
, &output
);
3599 output_event
= output
.file
->private_data
;
3600 ret
= perf_event_set_output(event
, output_event
);
3603 ret
= perf_event_set_output(event
, NULL
);
3608 case PERF_EVENT_IOC_SET_FILTER
:
3609 return perf_event_set_filter(event
, (void __user
*)arg
);
3615 if (flags
& PERF_IOC_FLAG_GROUP
)
3616 perf_event_for_each(event
, func
);
3618 perf_event_for_each_child(event
, func
);
3623 int perf_event_task_enable(void)
3625 struct perf_event
*event
;
3627 mutex_lock(¤t
->perf_event_mutex
);
3628 list_for_each_entry(event
, ¤t
->perf_event_list
, owner_entry
)
3629 perf_event_for_each_child(event
, perf_event_enable
);
3630 mutex_unlock(¤t
->perf_event_mutex
);
3635 int perf_event_task_disable(void)
3637 struct perf_event
*event
;
3639 mutex_lock(¤t
->perf_event_mutex
);
3640 list_for_each_entry(event
, ¤t
->perf_event_list
, owner_entry
)
3641 perf_event_for_each_child(event
, perf_event_disable
);
3642 mutex_unlock(¤t
->perf_event_mutex
);
3647 static int perf_event_index(struct perf_event
*event
)
3649 if (event
->hw
.state
& PERF_HES_STOPPED
)
3652 if (event
->state
!= PERF_EVENT_STATE_ACTIVE
)
3655 return event
->pmu
->event_idx(event
);
3658 static void calc_timer_values(struct perf_event
*event
,
3665 *now
= perf_clock();
3666 ctx_time
= event
->shadow_ctx_time
+ *now
;
3667 *enabled
= ctx_time
- event
->tstamp_enabled
;
3668 *running
= ctx_time
- event
->tstamp_running
;
3671 static void perf_event_init_userpage(struct perf_event
*event
)
3673 struct perf_event_mmap_page
*userpg
;
3674 struct ring_buffer
*rb
;
3677 rb
= rcu_dereference(event
->rb
);
3681 userpg
= rb
->user_page
;
3683 /* Allow new userspace to detect that bit 0 is deprecated */
3684 userpg
->cap_bit0_is_deprecated
= 1;
3685 userpg
->size
= offsetof(struct perf_event_mmap_page
, __reserved
);
3691 void __weak
arch_perf_update_userpage(struct perf_event_mmap_page
*userpg
, u64 now
)
3696 * Callers need to ensure there can be no nesting of this function, otherwise
3697 * the seqlock logic goes bad. We can not serialize this because the arch
3698 * code calls this from NMI context.
3700 void perf_event_update_userpage(struct perf_event
*event
)
3702 struct perf_event_mmap_page
*userpg
;
3703 struct ring_buffer
*rb
;
3704 u64 enabled
, running
, now
;
3707 rb
= rcu_dereference(event
->rb
);
3712 * compute total_time_enabled, total_time_running
3713 * based on snapshot values taken when the event
3714 * was last scheduled in.
3716 * we cannot simply called update_context_time()
3717 * because of locking issue as we can be called in
3720 calc_timer_values(event
, &now
, &enabled
, &running
);
3722 userpg
= rb
->user_page
;
3724 * Disable preemption so as to not let the corresponding user-space
3725 * spin too long if we get preempted.
3730 userpg
->index
= perf_event_index(event
);
3731 userpg
->offset
= perf_event_count(event
);
3733 userpg
->offset
-= local64_read(&event
->hw
.prev_count
);
3735 userpg
->time_enabled
= enabled
+
3736 atomic64_read(&event
->child_total_time_enabled
);
3738 userpg
->time_running
= running
+
3739 atomic64_read(&event
->child_total_time_running
);
3741 arch_perf_update_userpage(userpg
, now
);
3750 static int perf_mmap_fault(struct vm_area_struct
*vma
, struct vm_fault
*vmf
)
3752 struct perf_event
*event
= vma
->vm_file
->private_data
;
3753 struct ring_buffer
*rb
;
3754 int ret
= VM_FAULT_SIGBUS
;
3756 if (vmf
->flags
& FAULT_FLAG_MKWRITE
) {
3757 if (vmf
->pgoff
== 0)
3763 rb
= rcu_dereference(event
->rb
);
3767 if (vmf
->pgoff
&& (vmf
->flags
& FAULT_FLAG_WRITE
))
3770 vmf
->page
= perf_mmap_to_page(rb
, vmf
->pgoff
);
3774 get_page(vmf
->page
);
3775 vmf
->page
->mapping
= vma
->vm_file
->f_mapping
;
3776 vmf
->page
->index
= vmf
->pgoff
;
3785 static void ring_buffer_attach(struct perf_event
*event
,
3786 struct ring_buffer
*rb
)
3788 unsigned long flags
;
3790 if (!list_empty(&event
->rb_entry
))
3793 spin_lock_irqsave(&rb
->event_lock
, flags
);
3794 if (list_empty(&event
->rb_entry
))
3795 list_add(&event
->rb_entry
, &rb
->event_list
);
3796 spin_unlock_irqrestore(&rb
->event_lock
, flags
);
3799 static void ring_buffer_detach(struct perf_event
*event
, struct ring_buffer
*rb
)
3801 unsigned long flags
;
3803 if (list_empty(&event
->rb_entry
))
3806 spin_lock_irqsave(&rb
->event_lock
, flags
);
3807 list_del_init(&event
->rb_entry
);
3808 wake_up_all(&event
->waitq
);
3809 spin_unlock_irqrestore(&rb
->event_lock
, flags
);
3812 static void ring_buffer_wakeup(struct perf_event
*event
)
3814 struct ring_buffer
*rb
;
3817 rb
= rcu_dereference(event
->rb
);
3819 list_for_each_entry_rcu(event
, &rb
->event_list
, rb_entry
)
3820 wake_up_all(&event
->waitq
);
3825 static void rb_free_rcu(struct rcu_head
*rcu_head
)
3827 struct ring_buffer
*rb
;
3829 rb
= container_of(rcu_head
, struct ring_buffer
, rcu_head
);
3833 static struct ring_buffer
*ring_buffer_get(struct perf_event
*event
)
3835 struct ring_buffer
*rb
;
3838 rb
= rcu_dereference(event
->rb
);
3840 if (!atomic_inc_not_zero(&rb
->refcount
))
3848 static void ring_buffer_put(struct ring_buffer
*rb
)
3850 if (!atomic_dec_and_test(&rb
->refcount
))
3853 WARN_ON_ONCE(!list_empty(&rb
->event_list
));
3855 call_rcu(&rb
->rcu_head
, rb_free_rcu
);
3858 static void perf_mmap_open(struct vm_area_struct
*vma
)
3860 struct perf_event
*event
= vma
->vm_file
->private_data
;
3862 atomic_inc(&event
->mmap_count
);
3863 atomic_inc(&event
->rb
->mmap_count
);
3867 * A buffer can be mmap()ed multiple times; either directly through the same
3868 * event, or through other events by use of perf_event_set_output().
3870 * In order to undo the VM accounting done by perf_mmap() we need to destroy
3871 * the buffer here, where we still have a VM context. This means we need
3872 * to detach all events redirecting to us.
3874 static void perf_mmap_close(struct vm_area_struct
*vma
)
3876 struct perf_event
*event
= vma
->vm_file
->private_data
;
3878 struct ring_buffer
*rb
= event
->rb
;
3879 struct user_struct
*mmap_user
= rb
->mmap_user
;
3880 int mmap_locked
= rb
->mmap_locked
;
3881 unsigned long size
= perf_data_size(rb
);
3883 atomic_dec(&rb
->mmap_count
);
3885 if (!atomic_dec_and_mutex_lock(&event
->mmap_count
, &event
->mmap_mutex
))
3888 /* Detach current event from the buffer. */
3889 rcu_assign_pointer(event
->rb
, NULL
);
3890 ring_buffer_detach(event
, rb
);
3891 mutex_unlock(&event
->mmap_mutex
);
3893 /* If there's still other mmap()s of this buffer, we're done. */
3894 if (atomic_read(&rb
->mmap_count
)) {
3895 ring_buffer_put(rb
); /* can't be last */
3900 * No other mmap()s, detach from all other events that might redirect
3901 * into the now unreachable buffer. Somewhat complicated by the
3902 * fact that rb::event_lock otherwise nests inside mmap_mutex.
3906 list_for_each_entry_rcu(event
, &rb
->event_list
, rb_entry
) {
3907 if (!atomic_long_inc_not_zero(&event
->refcount
)) {
3909 * This event is en-route to free_event() which will
3910 * detach it and remove it from the list.
3916 mutex_lock(&event
->mmap_mutex
);
3918 * Check we didn't race with perf_event_set_output() which can
3919 * swizzle the rb from under us while we were waiting to
3920 * acquire mmap_mutex.
3922 * If we find a different rb; ignore this event, a next
3923 * iteration will no longer find it on the list. We have to
3924 * still restart the iteration to make sure we're not now
3925 * iterating the wrong list.
3927 if (event
->rb
== rb
) {
3928 rcu_assign_pointer(event
->rb
, NULL
);
3929 ring_buffer_detach(event
, rb
);
3930 ring_buffer_put(rb
); /* can't be last, we still have one */
3932 mutex_unlock(&event
->mmap_mutex
);
3936 * Restart the iteration; either we're on the wrong list or
3937 * destroyed its integrity by doing a deletion.
3944 * It could be there's still a few 0-ref events on the list; they'll
3945 * get cleaned up by free_event() -- they'll also still have their
3946 * ref on the rb and will free it whenever they are done with it.
3948 * Aside from that, this buffer is 'fully' detached and unmapped,
3949 * undo the VM accounting.
3952 atomic_long_sub((size
>> PAGE_SHIFT
) + 1, &mmap_user
->locked_vm
);
3953 vma
->vm_mm
->pinned_vm
-= mmap_locked
;
3954 free_uid(mmap_user
);
3956 ring_buffer_put(rb
); /* could be last */
3959 static const struct vm_operations_struct perf_mmap_vmops
= {
3960 .open
= perf_mmap_open
,
3961 .close
= perf_mmap_close
,
3962 .fault
= perf_mmap_fault
,
3963 .page_mkwrite
= perf_mmap_fault
,
3966 static int perf_mmap(struct file
*file
, struct vm_area_struct
*vma
)
3968 struct perf_event
*event
= file
->private_data
;
3969 unsigned long user_locked
, user_lock_limit
;
3970 struct user_struct
*user
= current_user();
3971 unsigned long locked
, lock_limit
;
3972 struct ring_buffer
*rb
;
3973 unsigned long vma_size
;
3974 unsigned long nr_pages
;
3975 long user_extra
, extra
;
3976 int ret
= 0, flags
= 0;
3979 * Don't allow mmap() of inherited per-task counters. This would
3980 * create a performance issue due to all children writing to the
3983 if (event
->cpu
== -1 && event
->attr
.inherit
)
3986 if (!(vma
->vm_flags
& VM_SHARED
))
3989 vma_size
= vma
->vm_end
- vma
->vm_start
;
3990 nr_pages
= (vma_size
/ PAGE_SIZE
) - 1;
3993 * If we have rb pages ensure they're a power-of-two number, so we
3994 * can do bitmasks instead of modulo.
3996 if (nr_pages
!= 0 && !is_power_of_2(nr_pages
))
3999 if (vma_size
!= PAGE_SIZE
* (1 + nr_pages
))
4002 if (vma
->vm_pgoff
!= 0)
4005 WARN_ON_ONCE(event
->ctx
->parent_ctx
);
4007 mutex_lock(&event
->mmap_mutex
);
4009 if (event
->rb
->nr_pages
!= nr_pages
) {
4014 if (!atomic_inc_not_zero(&event
->rb
->mmap_count
)) {
4016 * Raced against perf_mmap_close() through
4017 * perf_event_set_output(). Try again, hope for better
4020 mutex_unlock(&event
->mmap_mutex
);
4027 user_extra
= nr_pages
+ 1;
4028 user_lock_limit
= sysctl_perf_event_mlock
>> (PAGE_SHIFT
- 10);
4031 * Increase the limit linearly with more CPUs:
4033 user_lock_limit
*= num_online_cpus();
4035 user_locked
= atomic_long_read(&user
->locked_vm
) + user_extra
;
4038 if (user_locked
> user_lock_limit
)
4039 extra
= user_locked
- user_lock_limit
;
4041 lock_limit
= rlimit(RLIMIT_MEMLOCK
);
4042 lock_limit
>>= PAGE_SHIFT
;
4043 locked
= vma
->vm_mm
->pinned_vm
+ extra
;
4045 if ((locked
> lock_limit
) && perf_paranoid_tracepoint_raw() &&
4046 !capable(CAP_IPC_LOCK
)) {
4053 if (vma
->vm_flags
& VM_WRITE
)
4054 flags
|= RING_BUFFER_WRITABLE
;
4056 rb
= rb_alloc(nr_pages
,
4057 event
->attr
.watermark
? event
->attr
.wakeup_watermark
: 0,
4065 atomic_set(&rb
->mmap_count
, 1);
4066 rb
->mmap_locked
= extra
;
4067 rb
->mmap_user
= get_current_user();
4069 atomic_long_add(user_extra
, &user
->locked_vm
);
4070 vma
->vm_mm
->pinned_vm
+= extra
;
4072 ring_buffer_attach(event
, rb
);
4073 rcu_assign_pointer(event
->rb
, rb
);
4075 perf_event_init_userpage(event
);
4076 perf_event_update_userpage(event
);
4080 atomic_inc(&event
->mmap_count
);
4081 mutex_unlock(&event
->mmap_mutex
);
4084 * Since pinned accounting is per vm we cannot allow fork() to copy our
4087 vma
->vm_flags
|= VM_DONTCOPY
| VM_DONTEXPAND
| VM_DONTDUMP
;
4088 vma
->vm_ops
= &perf_mmap_vmops
;
4093 static int perf_fasync(int fd
, struct file
*filp
, int on
)
4095 struct inode
*inode
= file_inode(filp
);
4096 struct perf_event
*event
= filp
->private_data
;
4099 mutex_lock(&inode
->i_mutex
);
4100 retval
= fasync_helper(fd
, filp
, on
, &event
->fasync
);
4101 mutex_unlock(&inode
->i_mutex
);
4109 static const struct file_operations perf_fops
= {
4110 .llseek
= no_llseek
,
4111 .release
= perf_release
,
4114 .unlocked_ioctl
= perf_ioctl
,
4115 .compat_ioctl
= perf_ioctl
,
4117 .fasync
= perf_fasync
,
4123 * If there's data, ensure we set the poll() state and publish everything
4124 * to user-space before waking everybody up.
4127 void perf_event_wakeup(struct perf_event
*event
)
4129 ring_buffer_wakeup(event
);
4131 if (event
->pending_kill
) {
4132 kill_fasync(&event
->fasync
, SIGIO
, event
->pending_kill
);
4133 event
->pending_kill
= 0;
4137 static void perf_pending_event(struct irq_work
*entry
)
4139 struct perf_event
*event
= container_of(entry
,
4140 struct perf_event
, pending
);
4142 if (event
->pending_disable
) {
4143 event
->pending_disable
= 0;
4144 __perf_event_disable(event
);
4147 if (event
->pending_wakeup
) {
4148 event
->pending_wakeup
= 0;
4149 perf_event_wakeup(event
);
4154 * We assume there is only KVM supporting the callbacks.
4155 * Later on, we might change it to a list if there is
4156 * another virtualization implementation supporting the callbacks.
4158 struct perf_guest_info_callbacks
*perf_guest_cbs
;
4160 int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks
*cbs
)
4162 perf_guest_cbs
= cbs
;
4165 EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks
);
4167 int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks
*cbs
)
4169 perf_guest_cbs
= NULL
;
4172 EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks
);
4175 perf_output_sample_regs(struct perf_output_handle
*handle
,
4176 struct pt_regs
*regs
, u64 mask
)
4180 for_each_set_bit(bit
, (const unsigned long *) &mask
,
4181 sizeof(mask
) * BITS_PER_BYTE
) {
4184 val
= perf_reg_value(regs
, bit
);
4185 perf_output_put(handle
, val
);
4189 static void perf_sample_regs_user(struct perf_regs_user
*regs_user
,
4190 struct pt_regs
*regs
)
4192 if (!user_mode(regs
)) {
4194 regs
= task_pt_regs(current
);
4200 regs_user
->regs
= regs
;
4201 regs_user
->abi
= perf_reg_abi(current
);
4206 * Get remaining task size from user stack pointer.
4208 * It'd be better to take stack vma map and limit this more
4209 * precisly, but there's no way to get it safely under interrupt,
4210 * so using TASK_SIZE as limit.
4212 static u64
perf_ustack_task_size(struct pt_regs
*regs
)
4214 unsigned long addr
= perf_user_stack_pointer(regs
);
4216 if (!addr
|| addr
>= TASK_SIZE
)
4219 return TASK_SIZE
- addr
;
4223 perf_sample_ustack_size(u16 stack_size
, u16 header_size
,
4224 struct pt_regs
*regs
)
4228 /* No regs, no stack pointer, no dump. */
4233 * Check if we fit in with the requested stack size into the:
4235 * If we don't, we limit the size to the TASK_SIZE.
4237 * - remaining sample size
4238 * If we don't, we customize the stack size to
4239 * fit in to the remaining sample size.
4242 task_size
= min((u64
) USHRT_MAX
, perf_ustack_task_size(regs
));
4243 stack_size
= min(stack_size
, (u16
) task_size
);
4245 /* Current header size plus static size and dynamic size. */
4246 header_size
+= 2 * sizeof(u64
);
4248 /* Do we fit in with the current stack dump size? */
4249 if ((u16
) (header_size
+ stack_size
) < header_size
) {
4251 * If we overflow the maximum size for the sample,
4252 * we customize the stack dump size to fit in.
4254 stack_size
= USHRT_MAX
- header_size
- sizeof(u64
);
4255 stack_size
= round_up(stack_size
, sizeof(u64
));
4262 perf_output_sample_ustack(struct perf_output_handle
*handle
, u64 dump_size
,
4263 struct pt_regs
*regs
)
4265 /* Case of a kernel thread, nothing to dump */
4268 perf_output_put(handle
, size
);
4277 * - the size requested by user or the best one we can fit
4278 * in to the sample max size
4280 * - user stack dump data
4282 * - the actual dumped size
4286 perf_output_put(handle
, dump_size
);
4289 sp
= perf_user_stack_pointer(regs
);
4290 rem
= __output_copy_user(handle
, (void *) sp
, dump_size
);
4291 dyn_size
= dump_size
- rem
;
4293 perf_output_skip(handle
, rem
);
4296 perf_output_put(handle
, dyn_size
);
4300 static void __perf_event_header__init_id(struct perf_event_header
*header
,
4301 struct perf_sample_data
*data
,
4302 struct perf_event
*event
)
4304 u64 sample_type
= event
->attr
.sample_type
;
4306 data
->type
= sample_type
;
4307 header
->size
+= event
->id_header_size
;
4309 if (sample_type
& PERF_SAMPLE_TID
) {
4310 /* namespace issues */
4311 data
->tid_entry
.pid
= perf_event_pid(event
, current
);
4312 data
->tid_entry
.tid
= perf_event_tid(event
, current
);
4315 if (sample_type
& PERF_SAMPLE_TIME
)
4316 data
->time
= perf_clock();
4318 if (sample_type
& (PERF_SAMPLE_ID
| PERF_SAMPLE_IDENTIFIER
))
4319 data
->id
= primary_event_id(event
);
4321 if (sample_type
& PERF_SAMPLE_STREAM_ID
)
4322 data
->stream_id
= event
->id
;
4324 if (sample_type
& PERF_SAMPLE_CPU
) {
4325 data
->cpu_entry
.cpu
= raw_smp_processor_id();
4326 data
->cpu_entry
.reserved
= 0;
4330 void perf_event_header__init_id(struct perf_event_header
*header
,
4331 struct perf_sample_data
*data
,
4332 struct perf_event
*event
)
4334 if (event
->attr
.sample_id_all
)
4335 __perf_event_header__init_id(header
, data
, event
);
4338 static void __perf_event__output_id_sample(struct perf_output_handle
*handle
,
4339 struct perf_sample_data
*data
)
4341 u64 sample_type
= data
->type
;
4343 if (sample_type
& PERF_SAMPLE_TID
)
4344 perf_output_put(handle
, data
->tid_entry
);
4346 if (sample_type
& PERF_SAMPLE_TIME
)
4347 perf_output_put(handle
, data
->time
);
4349 if (sample_type
& PERF_SAMPLE_ID
)
4350 perf_output_put(handle
, data
->id
);
4352 if (sample_type
& PERF_SAMPLE_STREAM_ID
)
4353 perf_output_put(handle
, data
->stream_id
);
4355 if (sample_type
& PERF_SAMPLE_CPU
)
4356 perf_output_put(handle
, data
->cpu_entry
);
4358 if (sample_type
& PERF_SAMPLE_IDENTIFIER
)
4359 perf_output_put(handle
, data
->id
);
4362 void perf_event__output_id_sample(struct perf_event
*event
,
4363 struct perf_output_handle
*handle
,
4364 struct perf_sample_data
*sample
)
4366 if (event
->attr
.sample_id_all
)
4367 __perf_event__output_id_sample(handle
, sample
);
4370 static void perf_output_read_one(struct perf_output_handle
*handle
,
4371 struct perf_event
*event
,
4372 u64 enabled
, u64 running
)
4374 u64 read_format
= event
->attr
.read_format
;
4378 values
[n
++] = perf_event_count(event
);
4379 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
) {
4380 values
[n
++] = enabled
+
4381 atomic64_read(&event
->child_total_time_enabled
);
4383 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
) {
4384 values
[n
++] = running
+
4385 atomic64_read(&event
->child_total_time_running
);
4387 if (read_format
& PERF_FORMAT_ID
)
4388 values
[n
++] = primary_event_id(event
);
4390 __output_copy(handle
, values
, n
* sizeof(u64
));
4394 * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
4396 static void perf_output_read_group(struct perf_output_handle
*handle
,
4397 struct perf_event
*event
,
4398 u64 enabled
, u64 running
)
4400 struct perf_event
*leader
= event
->group_leader
, *sub
;
4401 u64 read_format
= event
->attr
.read_format
;
4405 values
[n
++] = 1 + leader
->nr_siblings
;
4407 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
4408 values
[n
++] = enabled
;
4410 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
4411 values
[n
++] = running
;
4413 if (leader
!= event
)
4414 leader
->pmu
->read(leader
);
4416 values
[n
++] = perf_event_count(leader
);
4417 if (read_format
& PERF_FORMAT_ID
)
4418 values
[n
++] = primary_event_id(leader
);
4420 __output_copy(handle
, values
, n
* sizeof(u64
));
4422 list_for_each_entry(sub
, &leader
->sibling_list
, group_entry
) {
4425 if ((sub
!= event
) &&
4426 (sub
->state
== PERF_EVENT_STATE_ACTIVE
))
4427 sub
->pmu
->read(sub
);
4429 values
[n
++] = perf_event_count(sub
);
4430 if (read_format
& PERF_FORMAT_ID
)
4431 values
[n
++] = primary_event_id(sub
);
4433 __output_copy(handle
, values
, n
* sizeof(u64
));
4437 #define PERF_FORMAT_TOTAL_TIMES (PERF_FORMAT_TOTAL_TIME_ENABLED|\
4438 PERF_FORMAT_TOTAL_TIME_RUNNING)
4440 static void perf_output_read(struct perf_output_handle
*handle
,
4441 struct perf_event
*event
)
4443 u64 enabled
= 0, running
= 0, now
;
4444 u64 read_format
= event
->attr
.read_format
;
4447 * compute total_time_enabled, total_time_running
4448 * based on snapshot values taken when the event
4449 * was last scheduled in.
4451 * we cannot simply called update_context_time()
4452 * because of locking issue as we are called in
4455 if (read_format
& PERF_FORMAT_TOTAL_TIMES
)
4456 calc_timer_values(event
, &now
, &enabled
, &running
);
4458 if (event
->attr
.read_format
& PERF_FORMAT_GROUP
)
4459 perf_output_read_group(handle
, event
, enabled
, running
);
4461 perf_output_read_one(handle
, event
, enabled
, running
);
4464 void perf_output_sample(struct perf_output_handle
*handle
,
4465 struct perf_event_header
*header
,
4466 struct perf_sample_data
*data
,
4467 struct perf_event
*event
)
4469 u64 sample_type
= data
->type
;
4471 perf_output_put(handle
, *header
);
4473 if (sample_type
& PERF_SAMPLE_IDENTIFIER
)
4474 perf_output_put(handle
, data
->id
);
4476 if (sample_type
& PERF_SAMPLE_IP
)
4477 perf_output_put(handle
, data
->ip
);
4479 if (sample_type
& PERF_SAMPLE_TID
)
4480 perf_output_put(handle
, data
->tid_entry
);
4482 if (sample_type
& PERF_SAMPLE_TIME
)
4483 perf_output_put(handle
, data
->time
);
4485 if (sample_type
& PERF_SAMPLE_ADDR
)
4486 perf_output_put(handle
, data
->addr
);
4488 if (sample_type
& PERF_SAMPLE_ID
)
4489 perf_output_put(handle
, data
->id
);
4491 if (sample_type
& PERF_SAMPLE_STREAM_ID
)
4492 perf_output_put(handle
, data
->stream_id
);
4494 if (sample_type
& PERF_SAMPLE_CPU
)
4495 perf_output_put(handle
, data
->cpu_entry
);
4497 if (sample_type
& PERF_SAMPLE_PERIOD
)
4498 perf_output_put(handle
, data
->period
);
4500 if (sample_type
& PERF_SAMPLE_READ
)
4501 perf_output_read(handle
, event
);
4503 if (sample_type
& PERF_SAMPLE_CALLCHAIN
) {
4504 if (data
->callchain
) {
4507 if (data
->callchain
)
4508 size
+= data
->callchain
->nr
;
4510 size
*= sizeof(u64
);
4512 __output_copy(handle
, data
->callchain
, size
);
4515 perf_output_put(handle
, nr
);
4519 if (sample_type
& PERF_SAMPLE_RAW
) {
4521 perf_output_put(handle
, data
->raw
->size
);
4522 __output_copy(handle
, data
->raw
->data
,
4529 .size
= sizeof(u32
),
4532 perf_output_put(handle
, raw
);
4536 if (sample_type
& PERF_SAMPLE_BRANCH_STACK
) {
4537 if (data
->br_stack
) {
4540 size
= data
->br_stack
->nr
4541 * sizeof(struct perf_branch_entry
);
4543 perf_output_put(handle
, data
->br_stack
->nr
);
4544 perf_output_copy(handle
, data
->br_stack
->entries
, size
);
4547 * we always store at least the value of nr
4550 perf_output_put(handle
, nr
);
4554 if (sample_type
& PERF_SAMPLE_REGS_USER
) {
4555 u64 abi
= data
->regs_user
.abi
;
4558 * If there are no regs to dump, notice it through
4559 * first u64 being zero (PERF_SAMPLE_REGS_ABI_NONE).
4561 perf_output_put(handle
, abi
);
4564 u64 mask
= event
->attr
.sample_regs_user
;
4565 perf_output_sample_regs(handle
,
4566 data
->regs_user
.regs
,
4571 if (sample_type
& PERF_SAMPLE_STACK_USER
) {
4572 perf_output_sample_ustack(handle
,
4573 data
->stack_user_size
,
4574 data
->regs_user
.regs
);
4577 if (sample_type
& PERF_SAMPLE_WEIGHT
)
4578 perf_output_put(handle
, data
->weight
);
4580 if (sample_type
& PERF_SAMPLE_DATA_SRC
)
4581 perf_output_put(handle
, data
->data_src
.val
);
4583 if (!event
->attr
.watermark
) {
4584 int wakeup_events
= event
->attr
.wakeup_events
;
4586 if (wakeup_events
) {
4587 struct ring_buffer
*rb
= handle
->rb
;
4588 int events
= local_inc_return(&rb
->events
);
4590 if (events
>= wakeup_events
) {
4591 local_sub(wakeup_events
, &rb
->events
);
4592 local_inc(&rb
->wakeup
);
4598 void perf_prepare_sample(struct perf_event_header
*header
,
4599 struct perf_sample_data
*data
,
4600 struct perf_event
*event
,
4601 struct pt_regs
*regs
)
4603 u64 sample_type
= event
->attr
.sample_type
;
4605 header
->type
= PERF_RECORD_SAMPLE
;
4606 header
->size
= sizeof(*header
) + event
->header_size
;
4609 header
->misc
|= perf_misc_flags(regs
);
4611 __perf_event_header__init_id(header
, data
, event
);
4613 if (sample_type
& PERF_SAMPLE_IP
)
4614 data
->ip
= perf_instruction_pointer(regs
);
4616 if (sample_type
& PERF_SAMPLE_CALLCHAIN
) {
4619 data
->callchain
= perf_callchain(event
, regs
);
4621 if (data
->callchain
)
4622 size
+= data
->callchain
->nr
;
4624 header
->size
+= size
* sizeof(u64
);
4627 if (sample_type
& PERF_SAMPLE_RAW
) {
4628 int size
= sizeof(u32
);
4631 size
+= data
->raw
->size
;
4633 size
+= sizeof(u32
);
4635 WARN_ON_ONCE(size
& (sizeof(u64
)-1));
4636 header
->size
+= size
;
4639 if (sample_type
& PERF_SAMPLE_BRANCH_STACK
) {
4640 int size
= sizeof(u64
); /* nr */
4641 if (data
->br_stack
) {
4642 size
+= data
->br_stack
->nr
4643 * sizeof(struct perf_branch_entry
);
4645 header
->size
+= size
;
4648 if (sample_type
& PERF_SAMPLE_REGS_USER
) {
4649 /* regs dump ABI info */
4650 int size
= sizeof(u64
);
4652 perf_sample_regs_user(&data
->regs_user
, regs
);
4654 if (data
->regs_user
.regs
) {
4655 u64 mask
= event
->attr
.sample_regs_user
;
4656 size
+= hweight64(mask
) * sizeof(u64
);
4659 header
->size
+= size
;
4662 if (sample_type
& PERF_SAMPLE_STACK_USER
) {
4664 * Either we need PERF_SAMPLE_STACK_USER bit to be allways
4665 * processed as the last one or have additional check added
4666 * in case new sample type is added, because we could eat
4667 * up the rest of the sample size.
4669 struct perf_regs_user
*uregs
= &data
->regs_user
;
4670 u16 stack_size
= event
->attr
.sample_stack_user
;
4671 u16 size
= sizeof(u64
);
4674 perf_sample_regs_user(uregs
, regs
);
4676 stack_size
= perf_sample_ustack_size(stack_size
, header
->size
,
4680 * If there is something to dump, add space for the dump
4681 * itself and for the field that tells the dynamic size,
4682 * which is how many have been actually dumped.
4685 size
+= sizeof(u64
) + stack_size
;
4687 data
->stack_user_size
= stack_size
;
4688 header
->size
+= size
;
4692 static void perf_event_output(struct perf_event
*event
,
4693 struct perf_sample_data
*data
,
4694 struct pt_regs
*regs
)
4696 struct perf_output_handle handle
;
4697 struct perf_event_header header
;
4699 /* protect the callchain buffers */
4702 perf_prepare_sample(&header
, data
, event
, regs
);
4704 if (perf_output_begin(&handle
, event
, header
.size
))
4707 perf_output_sample(&handle
, &header
, data
, event
);
4709 perf_output_end(&handle
);
4719 struct perf_read_event
{
4720 struct perf_event_header header
;
4727 perf_event_read_event(struct perf_event
*event
,
4728 struct task_struct
*task
)
4730 struct perf_output_handle handle
;
4731 struct perf_sample_data sample
;
4732 struct perf_read_event read_event
= {
4734 .type
= PERF_RECORD_READ
,
4736 .size
= sizeof(read_event
) + event
->read_size
,
4738 .pid
= perf_event_pid(event
, task
),
4739 .tid
= perf_event_tid(event
, task
),
4743 perf_event_header__init_id(&read_event
.header
, &sample
, event
);
4744 ret
= perf_output_begin(&handle
, event
, read_event
.header
.size
);
4748 perf_output_put(&handle
, read_event
);
4749 perf_output_read(&handle
, event
);
4750 perf_event__output_id_sample(event
, &handle
, &sample
);
4752 perf_output_end(&handle
);
4755 typedef void (perf_event_aux_output_cb
)(struct perf_event
*event
, void *data
);
4758 perf_event_aux_ctx(struct perf_event_context
*ctx
,
4759 perf_event_aux_output_cb output
,
4762 struct perf_event
*event
;
4764 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
4765 if (event
->state
< PERF_EVENT_STATE_INACTIVE
)
4767 if (!event_filter_match(event
))
4769 output(event
, data
);
4774 perf_event_aux(perf_event_aux_output_cb output
, void *data
,
4775 struct perf_event_context
*task_ctx
)
4777 struct perf_cpu_context
*cpuctx
;
4778 struct perf_event_context
*ctx
;
4783 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
4784 cpuctx
= get_cpu_ptr(pmu
->pmu_cpu_context
);
4785 if (cpuctx
->unique_pmu
!= pmu
)
4787 perf_event_aux_ctx(&cpuctx
->ctx
, output
, data
);
4790 ctxn
= pmu
->task_ctx_nr
;
4793 ctx
= rcu_dereference(current
->perf_event_ctxp
[ctxn
]);
4795 perf_event_aux_ctx(ctx
, output
, data
);
4797 put_cpu_ptr(pmu
->pmu_cpu_context
);
4802 perf_event_aux_ctx(task_ctx
, output
, data
);
4809 * task tracking -- fork/exit
4811 * enabled by: attr.comm | attr.mmap | attr.mmap2 | attr.mmap_data | attr.task
4814 struct perf_task_event
{
4815 struct task_struct
*task
;
4816 struct perf_event_context
*task_ctx
;
4819 struct perf_event_header header
;
4829 static int perf_event_task_match(struct perf_event
*event
)
4831 return event
->attr
.comm
|| event
->attr
.mmap
||
4832 event
->attr
.mmap2
|| event
->attr
.mmap_data
||
4836 static void perf_event_task_output(struct perf_event
*event
,
4839 struct perf_task_event
*task_event
= data
;
4840 struct perf_output_handle handle
;
4841 struct perf_sample_data sample
;
4842 struct task_struct
*task
= task_event
->task
;
4843 int ret
, size
= task_event
->event_id
.header
.size
;
4845 if (!perf_event_task_match(event
))
4848 perf_event_header__init_id(&task_event
->event_id
.header
, &sample
, event
);
4850 ret
= perf_output_begin(&handle
, event
,
4851 task_event
->event_id
.header
.size
);
4855 task_event
->event_id
.pid
= perf_event_pid(event
, task
);
4856 task_event
->event_id
.ppid
= perf_event_pid(event
, current
);
4858 task_event
->event_id
.tid
= perf_event_tid(event
, task
);
4859 task_event
->event_id
.ptid
= perf_event_tid(event
, current
);
4861 perf_output_put(&handle
, task_event
->event_id
);
4863 perf_event__output_id_sample(event
, &handle
, &sample
);
4865 perf_output_end(&handle
);
4867 task_event
->event_id
.header
.size
= size
;
4870 static void perf_event_task(struct task_struct
*task
,
4871 struct perf_event_context
*task_ctx
,
4874 struct perf_task_event task_event
;
4876 if (!atomic_read(&nr_comm_events
) &&
4877 !atomic_read(&nr_mmap_events
) &&
4878 !atomic_read(&nr_task_events
))
4881 task_event
= (struct perf_task_event
){
4883 .task_ctx
= task_ctx
,
4886 .type
= new ? PERF_RECORD_FORK
: PERF_RECORD_EXIT
,
4888 .size
= sizeof(task_event
.event_id
),
4894 .time
= perf_clock(),
4898 perf_event_aux(perf_event_task_output
,
4903 void perf_event_fork(struct task_struct
*task
)
4905 perf_event_task(task
, NULL
, 1);
4912 struct perf_comm_event
{
4913 struct task_struct
*task
;
4918 struct perf_event_header header
;
4925 static int perf_event_comm_match(struct perf_event
*event
)
4927 return event
->attr
.comm
;
4930 static void perf_event_comm_output(struct perf_event
*event
,
4933 struct perf_comm_event
*comm_event
= data
;
4934 struct perf_output_handle handle
;
4935 struct perf_sample_data sample
;
4936 int size
= comm_event
->event_id
.header
.size
;
4939 if (!perf_event_comm_match(event
))
4942 perf_event_header__init_id(&comm_event
->event_id
.header
, &sample
, event
);
4943 ret
= perf_output_begin(&handle
, event
,
4944 comm_event
->event_id
.header
.size
);
4949 comm_event
->event_id
.pid
= perf_event_pid(event
, comm_event
->task
);
4950 comm_event
->event_id
.tid
= perf_event_tid(event
, comm_event
->task
);
4952 perf_output_put(&handle
, comm_event
->event_id
);
4953 __output_copy(&handle
, comm_event
->comm
,
4954 comm_event
->comm_size
);
4956 perf_event__output_id_sample(event
, &handle
, &sample
);
4958 perf_output_end(&handle
);
4960 comm_event
->event_id
.header
.size
= size
;
4963 static void perf_event_comm_event(struct perf_comm_event
*comm_event
)
4965 char comm
[TASK_COMM_LEN
];
4968 memset(comm
, 0, sizeof(comm
));
4969 strlcpy(comm
, comm_event
->task
->comm
, sizeof(comm
));
4970 size
= ALIGN(strlen(comm
)+1, sizeof(u64
));
4972 comm_event
->comm
= comm
;
4973 comm_event
->comm_size
= size
;
4975 comm_event
->event_id
.header
.size
= sizeof(comm_event
->event_id
) + size
;
4977 perf_event_aux(perf_event_comm_output
,
4982 void perf_event_comm(struct task_struct
*task
)
4984 struct perf_comm_event comm_event
;
4985 struct perf_event_context
*ctx
;
4989 for_each_task_context_nr(ctxn
) {
4990 ctx
= task
->perf_event_ctxp
[ctxn
];
4994 perf_event_enable_on_exec(ctx
);
4998 if (!atomic_read(&nr_comm_events
))
5001 comm_event
= (struct perf_comm_event
){
5007 .type
= PERF_RECORD_COMM
,
5016 perf_event_comm_event(&comm_event
);
5023 struct perf_mmap_event
{
5024 struct vm_area_struct
*vma
;
5026 const char *file_name
;
5033 struct perf_event_header header
;
5043 static int perf_event_mmap_match(struct perf_event
*event
,
5046 struct perf_mmap_event
*mmap_event
= data
;
5047 struct vm_area_struct
*vma
= mmap_event
->vma
;
5048 int executable
= vma
->vm_flags
& VM_EXEC
;
5050 return (!executable
&& event
->attr
.mmap_data
) ||
5051 (executable
&& (event
->attr
.mmap
|| event
->attr
.mmap2
));
5054 static void perf_event_mmap_output(struct perf_event
*event
,
5057 struct perf_mmap_event
*mmap_event
= data
;
5058 struct perf_output_handle handle
;
5059 struct perf_sample_data sample
;
5060 int size
= mmap_event
->event_id
.header
.size
;
5063 if (!perf_event_mmap_match(event
, data
))
5066 if (event
->attr
.mmap2
) {
5067 mmap_event
->event_id
.header
.type
= PERF_RECORD_MMAP2
;
5068 mmap_event
->event_id
.header
.size
+= sizeof(mmap_event
->maj
);
5069 mmap_event
->event_id
.header
.size
+= sizeof(mmap_event
->min
);
5070 mmap_event
->event_id
.header
.size
+= sizeof(mmap_event
->ino
);
5071 mmap_event
->event_id
.header
.size
+= sizeof(mmap_event
->ino_generation
);
5074 perf_event_header__init_id(&mmap_event
->event_id
.header
, &sample
, event
);
5075 ret
= perf_output_begin(&handle
, event
,
5076 mmap_event
->event_id
.header
.size
);
5080 mmap_event
->event_id
.pid
= perf_event_pid(event
, current
);
5081 mmap_event
->event_id
.tid
= perf_event_tid(event
, current
);
5083 perf_output_put(&handle
, mmap_event
->event_id
);
5085 if (event
->attr
.mmap2
) {
5086 perf_output_put(&handle
, mmap_event
->maj
);
5087 perf_output_put(&handle
, mmap_event
->min
);
5088 perf_output_put(&handle
, mmap_event
->ino
);
5089 perf_output_put(&handle
, mmap_event
->ino_generation
);
5092 __output_copy(&handle
, mmap_event
->file_name
,
5093 mmap_event
->file_size
);
5095 perf_event__output_id_sample(event
, &handle
, &sample
);
5097 perf_output_end(&handle
);
5099 mmap_event
->event_id
.header
.size
= size
;
5102 static void perf_event_mmap_event(struct perf_mmap_event
*mmap_event
)
5104 struct vm_area_struct
*vma
= mmap_event
->vma
;
5105 struct file
*file
= vma
->vm_file
;
5106 int maj
= 0, min
= 0;
5107 u64 ino
= 0, gen
= 0;
5113 memset(tmp
, 0, sizeof(tmp
));
5116 struct inode
*inode
;
5119 * d_path works from the end of the rb backwards, so we
5120 * need to add enough zero bytes after the string to handle
5121 * the 64bit alignment we do later.
5123 buf
= kzalloc(PATH_MAX
+ sizeof(u64
), GFP_KERNEL
);
5125 name
= strncpy(tmp
, "//enomem", sizeof(tmp
));
5128 name
= d_path(&file
->f_path
, buf
, PATH_MAX
);
5130 name
= strncpy(tmp
, "//toolong", sizeof(tmp
));
5133 inode
= file_inode(vma
->vm_file
);
5134 dev
= inode
->i_sb
->s_dev
;
5136 gen
= inode
->i_generation
;
5141 if (arch_vma_name(mmap_event
->vma
)) {
5142 name
= strncpy(tmp
, arch_vma_name(mmap_event
->vma
),
5144 tmp
[sizeof(tmp
) - 1] = '\0';
5149 name
= strncpy(tmp
, "[vdso]", sizeof(tmp
));
5151 } else if (vma
->vm_start
<= vma
->vm_mm
->start_brk
&&
5152 vma
->vm_end
>= vma
->vm_mm
->brk
) {
5153 name
= strncpy(tmp
, "[heap]", sizeof(tmp
));
5155 } else if (vma
->vm_start
<= vma
->vm_mm
->start_stack
&&
5156 vma
->vm_end
>= vma
->vm_mm
->start_stack
) {
5157 name
= strncpy(tmp
, "[stack]", sizeof(tmp
));
5161 name
= strncpy(tmp
, "//anon", sizeof(tmp
));
5166 size
= ALIGN(strlen(name
)+1, sizeof(u64
));
5168 mmap_event
->file_name
= name
;
5169 mmap_event
->file_size
= size
;
5170 mmap_event
->maj
= maj
;
5171 mmap_event
->min
= min
;
5172 mmap_event
->ino
= ino
;
5173 mmap_event
->ino_generation
= gen
;
5175 if (!(vma
->vm_flags
& VM_EXEC
))
5176 mmap_event
->event_id
.header
.misc
|= PERF_RECORD_MISC_MMAP_DATA
;
5178 mmap_event
->event_id
.header
.size
= sizeof(mmap_event
->event_id
) + size
;
5180 perf_event_aux(perf_event_mmap_output
,
5187 void perf_event_mmap(struct vm_area_struct
*vma
)
5189 struct perf_mmap_event mmap_event
;
5191 if (!atomic_read(&nr_mmap_events
))
5194 mmap_event
= (struct perf_mmap_event
){
5200 .type
= PERF_RECORD_MMAP
,
5201 .misc
= PERF_RECORD_MISC_USER
,
5206 .start
= vma
->vm_start
,
5207 .len
= vma
->vm_end
- vma
->vm_start
,
5208 .pgoff
= (u64
)vma
->vm_pgoff
<< PAGE_SHIFT
,
5210 /* .maj (attr_mmap2 only) */
5211 /* .min (attr_mmap2 only) */
5212 /* .ino (attr_mmap2 only) */
5213 /* .ino_generation (attr_mmap2 only) */
5216 perf_event_mmap_event(&mmap_event
);
5220 * IRQ throttle logging
5223 static void perf_log_throttle(struct perf_event
*event
, int enable
)
5225 struct perf_output_handle handle
;
5226 struct perf_sample_data sample
;
5230 struct perf_event_header header
;
5234 } throttle_event
= {
5236 .type
= PERF_RECORD_THROTTLE
,
5238 .size
= sizeof(throttle_event
),
5240 .time
= perf_clock(),
5241 .id
= primary_event_id(event
),
5242 .stream_id
= event
->id
,
5246 throttle_event
.header
.type
= PERF_RECORD_UNTHROTTLE
;
5248 perf_event_header__init_id(&throttle_event
.header
, &sample
, event
);
5250 ret
= perf_output_begin(&handle
, event
,
5251 throttle_event
.header
.size
);
5255 perf_output_put(&handle
, throttle_event
);
5256 perf_event__output_id_sample(event
, &handle
, &sample
);
5257 perf_output_end(&handle
);
5261 * Generic event overflow handling, sampling.
5264 static int __perf_event_overflow(struct perf_event
*event
,
5265 int throttle
, struct perf_sample_data
*data
,
5266 struct pt_regs
*regs
)
5268 int events
= atomic_read(&event
->event_limit
);
5269 struct hw_perf_event
*hwc
= &event
->hw
;
5274 * Non-sampling counters might still use the PMI to fold short
5275 * hardware counters, ignore those.
5277 if (unlikely(!is_sampling_event(event
)))
5280 seq
= __this_cpu_read(perf_throttled_seq
);
5281 if (seq
!= hwc
->interrupts_seq
) {
5282 hwc
->interrupts_seq
= seq
;
5283 hwc
->interrupts
= 1;
5286 if (unlikely(throttle
5287 && hwc
->interrupts
>= max_samples_per_tick
)) {
5288 __this_cpu_inc(perf_throttled_count
);
5289 hwc
->interrupts
= MAX_INTERRUPTS
;
5290 perf_log_throttle(event
, 0);
5291 tick_nohz_full_kick();
5296 if (event
->attr
.freq
) {
5297 u64 now
= perf_clock();
5298 s64 delta
= now
- hwc
->freq_time_stamp
;
5300 hwc
->freq_time_stamp
= now
;
5302 if (delta
> 0 && delta
< 2*TICK_NSEC
)
5303 perf_adjust_period(event
, delta
, hwc
->last_period
, true);
5307 * XXX event_limit might not quite work as expected on inherited
5311 event
->pending_kill
= POLL_IN
;
5312 if (events
&& atomic_dec_and_test(&event
->event_limit
)) {
5314 event
->pending_kill
= POLL_HUP
;
5315 event
->pending_disable
= 1;
5316 irq_work_queue(&event
->pending
);
5319 if (event
->overflow_handler
)
5320 event
->overflow_handler(event
, data
, regs
);
5322 perf_event_output(event
, data
, regs
);
5324 if (event
->fasync
&& event
->pending_kill
) {
5325 event
->pending_wakeup
= 1;
5326 irq_work_queue(&event
->pending
);
5332 int perf_event_overflow(struct perf_event
*event
,
5333 struct perf_sample_data
*data
,
5334 struct pt_regs
*regs
)
5336 return __perf_event_overflow(event
, 1, data
, regs
);
5340 * Generic software event infrastructure
5343 struct swevent_htable
{
5344 struct swevent_hlist
*swevent_hlist
;
5345 struct mutex hlist_mutex
;
5348 /* Recursion avoidance in each contexts */
5349 int recursion
[PERF_NR_CONTEXTS
];
5351 /* Keeps track of cpu being initialized/exited */
5355 static DEFINE_PER_CPU(struct swevent_htable
, swevent_htable
);
5358 * We directly increment event->count and keep a second value in
5359 * event->hw.period_left to count intervals. This period event
5360 * is kept in the range [-sample_period, 0] so that we can use the
5364 u64
perf_swevent_set_period(struct perf_event
*event
)
5366 struct hw_perf_event
*hwc
= &event
->hw
;
5367 u64 period
= hwc
->last_period
;
5371 hwc
->last_period
= hwc
->sample_period
;
5374 old
= val
= local64_read(&hwc
->period_left
);
5378 nr
= div64_u64(period
+ val
, period
);
5379 offset
= nr
* period
;
5381 if (local64_cmpxchg(&hwc
->period_left
, old
, val
) != old
)
5387 static void perf_swevent_overflow(struct perf_event
*event
, u64 overflow
,
5388 struct perf_sample_data
*data
,
5389 struct pt_regs
*regs
)
5391 struct hw_perf_event
*hwc
= &event
->hw
;
5395 overflow
= perf_swevent_set_period(event
);
5397 if (hwc
->interrupts
== MAX_INTERRUPTS
)
5400 for (; overflow
; overflow
--) {
5401 if (__perf_event_overflow(event
, throttle
,
5404 * We inhibit the overflow from happening when
5405 * hwc->interrupts == MAX_INTERRUPTS.
5413 static void perf_swevent_event(struct perf_event
*event
, u64 nr
,
5414 struct perf_sample_data
*data
,
5415 struct pt_regs
*regs
)
5417 struct hw_perf_event
*hwc
= &event
->hw
;
5419 local64_add(nr
, &event
->count
);
5424 if (!is_sampling_event(event
))
5427 if ((event
->attr
.sample_type
& PERF_SAMPLE_PERIOD
) && !event
->attr
.freq
) {
5429 return perf_swevent_overflow(event
, 1, data
, regs
);
5431 data
->period
= event
->hw
.last_period
;
5433 if (nr
== 1 && hwc
->sample_period
== 1 && !event
->attr
.freq
)
5434 return perf_swevent_overflow(event
, 1, data
, regs
);
5436 if (local64_add_negative(nr
, &hwc
->period_left
))
5439 perf_swevent_overflow(event
, 0, data
, regs
);
5442 static int perf_exclude_event(struct perf_event
*event
,
5443 struct pt_regs
*regs
)
5445 if (event
->hw
.state
& PERF_HES_STOPPED
)
5449 if (event
->attr
.exclude_user
&& user_mode(regs
))
5452 if (event
->attr
.exclude_kernel
&& !user_mode(regs
))
5459 static int perf_swevent_match(struct perf_event
*event
,
5460 enum perf_type_id type
,
5462 struct perf_sample_data
*data
,
5463 struct pt_regs
*regs
)
5465 if (event
->attr
.type
!= type
)
5468 if (event
->attr
.config
!= event_id
)
5471 if (perf_exclude_event(event
, regs
))
5477 static inline u64
swevent_hash(u64 type
, u32 event_id
)
5479 u64 val
= event_id
| (type
<< 32);
5481 return hash_64(val
, SWEVENT_HLIST_BITS
);
5484 static inline struct hlist_head
*
5485 __find_swevent_head(struct swevent_hlist
*hlist
, u64 type
, u32 event_id
)
5487 u64 hash
= swevent_hash(type
, event_id
);
5489 return &hlist
->heads
[hash
];
5492 /* For the read side: events when they trigger */
5493 static inline struct hlist_head
*
5494 find_swevent_head_rcu(struct swevent_htable
*swhash
, u64 type
, u32 event_id
)
5496 struct swevent_hlist
*hlist
;
5498 hlist
= rcu_dereference(swhash
->swevent_hlist
);
5502 return __find_swevent_head(hlist
, type
, event_id
);
5505 /* For the event head insertion and removal in the hlist */
5506 static inline struct hlist_head
*
5507 find_swevent_head(struct swevent_htable
*swhash
, struct perf_event
*event
)
5509 struct swevent_hlist
*hlist
;
5510 u32 event_id
= event
->attr
.config
;
5511 u64 type
= event
->attr
.type
;
5514 * Event scheduling is always serialized against hlist allocation
5515 * and release. Which makes the protected version suitable here.
5516 * The context lock guarantees that.
5518 hlist
= rcu_dereference_protected(swhash
->swevent_hlist
,
5519 lockdep_is_held(&event
->ctx
->lock
));
5523 return __find_swevent_head(hlist
, type
, event_id
);
5526 static void do_perf_sw_event(enum perf_type_id type
, u32 event_id
,
5528 struct perf_sample_data
*data
,
5529 struct pt_regs
*regs
)
5531 struct swevent_htable
*swhash
= &__get_cpu_var(swevent_htable
);
5532 struct perf_event
*event
;
5533 struct hlist_head
*head
;
5536 head
= find_swevent_head_rcu(swhash
, type
, event_id
);
5540 hlist_for_each_entry_rcu(event
, head
, hlist_entry
) {
5541 if (perf_swevent_match(event
, type
, event_id
, data
, regs
))
5542 perf_swevent_event(event
, nr
, data
, regs
);
5548 int perf_swevent_get_recursion_context(void)
5550 struct swevent_htable
*swhash
= &__get_cpu_var(swevent_htable
);
5552 return get_recursion_context(swhash
->recursion
);
5554 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context
);
5556 inline void perf_swevent_put_recursion_context(int rctx
)
5558 struct swevent_htable
*swhash
= &__get_cpu_var(swevent_htable
);
5560 put_recursion_context(swhash
->recursion
, rctx
);
5563 void __perf_sw_event(u32 event_id
, u64 nr
, struct pt_regs
*regs
, u64 addr
)
5565 struct perf_sample_data data
;
5568 preempt_disable_notrace();
5569 rctx
= perf_swevent_get_recursion_context();
5573 perf_sample_data_init(&data
, addr
, 0);
5575 do_perf_sw_event(PERF_TYPE_SOFTWARE
, event_id
, nr
, &data
, regs
);
5577 perf_swevent_put_recursion_context(rctx
);
5578 preempt_enable_notrace();
5581 static void perf_swevent_read(struct perf_event
*event
)
5585 static int perf_swevent_add(struct perf_event
*event
, int flags
)
5587 struct swevent_htable
*swhash
= &__get_cpu_var(swevent_htable
);
5588 struct hw_perf_event
*hwc
= &event
->hw
;
5589 struct hlist_head
*head
;
5591 if (is_sampling_event(event
)) {
5592 hwc
->last_period
= hwc
->sample_period
;
5593 perf_swevent_set_period(event
);
5596 hwc
->state
= !(flags
& PERF_EF_START
);
5598 head
= find_swevent_head(swhash
, event
);
5601 * We can race with cpu hotplug code. Do not
5602 * WARN if the cpu just got unplugged.
5604 WARN_ON_ONCE(swhash
->online
);
5608 hlist_add_head_rcu(&event
->hlist_entry
, head
);
5613 static void perf_swevent_del(struct perf_event
*event
, int flags
)
5615 hlist_del_rcu(&event
->hlist_entry
);
5618 static void perf_swevent_start(struct perf_event
*event
, int flags
)
5620 event
->hw
.state
= 0;
5623 static void perf_swevent_stop(struct perf_event
*event
, int flags
)
5625 event
->hw
.state
= PERF_HES_STOPPED
;
5628 /* Deref the hlist from the update side */
5629 static inline struct swevent_hlist
*
5630 swevent_hlist_deref(struct swevent_htable
*swhash
)
5632 return rcu_dereference_protected(swhash
->swevent_hlist
,
5633 lockdep_is_held(&swhash
->hlist_mutex
));
5636 static void swevent_hlist_release(struct swevent_htable
*swhash
)
5638 struct swevent_hlist
*hlist
= swevent_hlist_deref(swhash
);
5643 rcu_assign_pointer(swhash
->swevent_hlist
, NULL
);
5644 kfree_rcu(hlist
, rcu_head
);
5647 static void swevent_hlist_put_cpu(struct perf_event
*event
, int cpu
)
5649 struct swevent_htable
*swhash
= &per_cpu(swevent_htable
, cpu
);
5651 mutex_lock(&swhash
->hlist_mutex
);
5653 if (!--swhash
->hlist_refcount
)
5654 swevent_hlist_release(swhash
);
5656 mutex_unlock(&swhash
->hlist_mutex
);
5659 static void swevent_hlist_put(struct perf_event
*event
)
5663 if (event
->cpu
!= -1) {
5664 swevent_hlist_put_cpu(event
, event
->cpu
);
5668 for_each_possible_cpu(cpu
)
5669 swevent_hlist_put_cpu(event
, cpu
);
5672 static int swevent_hlist_get_cpu(struct perf_event
*event
, int cpu
)
5674 struct swevent_htable
*swhash
= &per_cpu(swevent_htable
, cpu
);
5677 mutex_lock(&swhash
->hlist_mutex
);
5679 if (!swevent_hlist_deref(swhash
) && cpu_online(cpu
)) {
5680 struct swevent_hlist
*hlist
;
5682 hlist
= kzalloc(sizeof(*hlist
), GFP_KERNEL
);
5687 rcu_assign_pointer(swhash
->swevent_hlist
, hlist
);
5689 swhash
->hlist_refcount
++;
5691 mutex_unlock(&swhash
->hlist_mutex
);
5696 static int swevent_hlist_get(struct perf_event
*event
)
5699 int cpu
, failed_cpu
;
5701 if (event
->cpu
!= -1)
5702 return swevent_hlist_get_cpu(event
, event
->cpu
);
5705 for_each_possible_cpu(cpu
) {
5706 err
= swevent_hlist_get_cpu(event
, cpu
);
5716 for_each_possible_cpu(cpu
) {
5717 if (cpu
== failed_cpu
)
5719 swevent_hlist_put_cpu(event
, cpu
);
5726 struct static_key perf_swevent_enabled
[PERF_COUNT_SW_MAX
];
5728 static void sw_perf_event_destroy(struct perf_event
*event
)
5730 u64 event_id
= event
->attr
.config
;
5732 WARN_ON(event
->parent
);
5734 static_key_slow_dec(&perf_swevent_enabled
[event_id
]);
5735 swevent_hlist_put(event
);
5738 static int perf_swevent_init(struct perf_event
*event
)
5740 u64 event_id
= event
->attr
.config
;
5742 if (event
->attr
.type
!= PERF_TYPE_SOFTWARE
)
5746 * no branch sampling for software events
5748 if (has_branch_stack(event
))
5752 case PERF_COUNT_SW_CPU_CLOCK
:
5753 case PERF_COUNT_SW_TASK_CLOCK
:
5760 if (event_id
>= PERF_COUNT_SW_MAX
)
5763 if (!event
->parent
) {
5766 err
= swevent_hlist_get(event
);
5770 static_key_slow_inc(&perf_swevent_enabled
[event_id
]);
5771 event
->destroy
= sw_perf_event_destroy
;
5777 static int perf_swevent_event_idx(struct perf_event
*event
)
5782 static struct pmu perf_swevent
= {
5783 .task_ctx_nr
= perf_sw_context
,
5785 .event_init
= perf_swevent_init
,
5786 .add
= perf_swevent_add
,
5787 .del
= perf_swevent_del
,
5788 .start
= perf_swevent_start
,
5789 .stop
= perf_swevent_stop
,
5790 .read
= perf_swevent_read
,
5792 .event_idx
= perf_swevent_event_idx
,
5795 #ifdef CONFIG_EVENT_TRACING
5797 static int perf_tp_filter_match(struct perf_event
*event
,
5798 struct perf_sample_data
*data
)
5800 void *record
= data
->raw
->data
;
5802 if (likely(!event
->filter
) || filter_match_preds(event
->filter
, record
))
5807 static int perf_tp_event_match(struct perf_event
*event
,
5808 struct perf_sample_data
*data
,
5809 struct pt_regs
*regs
)
5811 if (event
->hw
.state
& PERF_HES_STOPPED
)
5814 * All tracepoints are from kernel-space.
5816 if (event
->attr
.exclude_kernel
)
5819 if (!perf_tp_filter_match(event
, data
))
5825 void perf_tp_event(u64 addr
, u64 count
, void *record
, int entry_size
,
5826 struct pt_regs
*regs
, struct hlist_head
*head
, int rctx
,
5827 struct task_struct
*task
)
5829 struct perf_sample_data data
;
5830 struct perf_event
*event
;
5832 struct perf_raw_record raw
= {
5837 perf_sample_data_init(&data
, addr
, 0);
5840 hlist_for_each_entry_rcu(event
, head
, hlist_entry
) {
5841 if (perf_tp_event_match(event
, &data
, regs
))
5842 perf_swevent_event(event
, count
, &data
, regs
);
5846 * If we got specified a target task, also iterate its context and
5847 * deliver this event there too.
5849 if (task
&& task
!= current
) {
5850 struct perf_event_context
*ctx
;
5851 struct trace_entry
*entry
= record
;
5854 ctx
= rcu_dereference(task
->perf_event_ctxp
[perf_sw_context
]);
5858 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
5859 if (event
->attr
.type
!= PERF_TYPE_TRACEPOINT
)
5861 if (event
->attr
.config
!= entry
->type
)
5863 if (perf_tp_event_match(event
, &data
, regs
))
5864 perf_swevent_event(event
, count
, &data
, regs
);
5870 perf_swevent_put_recursion_context(rctx
);
5872 EXPORT_SYMBOL_GPL(perf_tp_event
);
5874 static void tp_perf_event_destroy(struct perf_event
*event
)
5876 perf_trace_destroy(event
);
5879 static int perf_tp_event_init(struct perf_event
*event
)
5883 if (event
->attr
.type
!= PERF_TYPE_TRACEPOINT
)
5887 * no branch sampling for tracepoint events
5889 if (has_branch_stack(event
))
5892 err
= perf_trace_init(event
);
5896 event
->destroy
= tp_perf_event_destroy
;
5901 static struct pmu perf_tracepoint
= {
5902 .task_ctx_nr
= perf_sw_context
,
5904 .event_init
= perf_tp_event_init
,
5905 .add
= perf_trace_add
,
5906 .del
= perf_trace_del
,
5907 .start
= perf_swevent_start
,
5908 .stop
= perf_swevent_stop
,
5909 .read
= perf_swevent_read
,
5911 .event_idx
= perf_swevent_event_idx
,
5914 static inline void perf_tp_register(void)
5916 perf_pmu_register(&perf_tracepoint
, "tracepoint", PERF_TYPE_TRACEPOINT
);
5919 static int perf_event_set_filter(struct perf_event
*event
, void __user
*arg
)
5924 if (event
->attr
.type
!= PERF_TYPE_TRACEPOINT
)
5927 filter_str
= strndup_user(arg
, PAGE_SIZE
);
5928 if (IS_ERR(filter_str
))
5929 return PTR_ERR(filter_str
);
5931 ret
= ftrace_profile_set_filter(event
, event
->attr
.config
, filter_str
);
5937 static void perf_event_free_filter(struct perf_event
*event
)
5939 ftrace_profile_free_filter(event
);
5944 static inline void perf_tp_register(void)
5948 static int perf_event_set_filter(struct perf_event
*event
, void __user
*arg
)
5953 static void perf_event_free_filter(struct perf_event
*event
)
5957 #endif /* CONFIG_EVENT_TRACING */
5959 #ifdef CONFIG_HAVE_HW_BREAKPOINT
5960 void perf_bp_event(struct perf_event
*bp
, void *data
)
5962 struct perf_sample_data sample
;
5963 struct pt_regs
*regs
= data
;
5965 perf_sample_data_init(&sample
, bp
->attr
.bp_addr
, 0);
5967 if (!bp
->hw
.state
&& !perf_exclude_event(bp
, regs
))
5968 perf_swevent_event(bp
, 1, &sample
, regs
);
5973 * hrtimer based swevent callback
5976 static enum hrtimer_restart
perf_swevent_hrtimer(struct hrtimer
*hrtimer
)
5978 enum hrtimer_restart ret
= HRTIMER_RESTART
;
5979 struct perf_sample_data data
;
5980 struct pt_regs
*regs
;
5981 struct perf_event
*event
;
5984 event
= container_of(hrtimer
, struct perf_event
, hw
.hrtimer
);
5986 if (event
->state
!= PERF_EVENT_STATE_ACTIVE
)
5987 return HRTIMER_NORESTART
;
5989 event
->pmu
->read(event
);
5991 perf_sample_data_init(&data
, 0, event
->hw
.last_period
);
5992 regs
= get_irq_regs();
5994 if (regs
&& !perf_exclude_event(event
, regs
)) {
5995 if (!(event
->attr
.exclude_idle
&& is_idle_task(current
)))
5996 if (__perf_event_overflow(event
, 1, &data
, regs
))
5997 ret
= HRTIMER_NORESTART
;
6000 period
= max_t(u64
, 10000, event
->hw
.sample_period
);
6001 hrtimer_forward_now(hrtimer
, ns_to_ktime(period
));
6006 static void perf_swevent_start_hrtimer(struct perf_event
*event
)
6008 struct hw_perf_event
*hwc
= &event
->hw
;
6011 if (!is_sampling_event(event
))
6014 period
= local64_read(&hwc
->period_left
);
6019 local64_set(&hwc
->period_left
, 0);
6021 period
= max_t(u64
, 10000, hwc
->sample_period
);
6023 __hrtimer_start_range_ns(&hwc
->hrtimer
,
6024 ns_to_ktime(period
), 0,
6025 HRTIMER_MODE_REL_PINNED
, 0);
6028 static void perf_swevent_cancel_hrtimer(struct perf_event
*event
)
6030 struct hw_perf_event
*hwc
= &event
->hw
;
6032 if (is_sampling_event(event
)) {
6033 ktime_t remaining
= hrtimer_get_remaining(&hwc
->hrtimer
);
6034 local64_set(&hwc
->period_left
, ktime_to_ns(remaining
));
6036 hrtimer_cancel(&hwc
->hrtimer
);
6040 static void perf_swevent_init_hrtimer(struct perf_event
*event
)
6042 struct hw_perf_event
*hwc
= &event
->hw
;
6044 if (!is_sampling_event(event
))
6047 hrtimer_init(&hwc
->hrtimer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
6048 hwc
->hrtimer
.function
= perf_swevent_hrtimer
;
6051 * Since hrtimers have a fixed rate, we can do a static freq->period
6052 * mapping and avoid the whole period adjust feedback stuff.
6054 if (event
->attr
.freq
) {
6055 long freq
= event
->attr
.sample_freq
;
6057 event
->attr
.sample_period
= NSEC_PER_SEC
/ freq
;
6058 hwc
->sample_period
= event
->attr
.sample_period
;
6059 local64_set(&hwc
->period_left
, hwc
->sample_period
);
6060 hwc
->last_period
= hwc
->sample_period
;
6061 event
->attr
.freq
= 0;
6066 * Software event: cpu wall time clock
6069 static void cpu_clock_event_update(struct perf_event
*event
)
6074 now
= local_clock();
6075 prev
= local64_xchg(&event
->hw
.prev_count
, now
);
6076 local64_add(now
- prev
, &event
->count
);
6079 static void cpu_clock_event_start(struct perf_event
*event
, int flags
)
6081 local64_set(&event
->hw
.prev_count
, local_clock());
6082 perf_swevent_start_hrtimer(event
);
6085 static void cpu_clock_event_stop(struct perf_event
*event
, int flags
)
6087 perf_swevent_cancel_hrtimer(event
);
6088 cpu_clock_event_update(event
);
6091 static int cpu_clock_event_add(struct perf_event
*event
, int flags
)
6093 if (flags
& PERF_EF_START
)
6094 cpu_clock_event_start(event
, flags
);
6099 static void cpu_clock_event_del(struct perf_event
*event
, int flags
)
6101 cpu_clock_event_stop(event
, flags
);
6104 static void cpu_clock_event_read(struct perf_event
*event
)
6106 cpu_clock_event_update(event
);
6109 static int cpu_clock_event_init(struct perf_event
*event
)
6111 if (event
->attr
.type
!= PERF_TYPE_SOFTWARE
)
6114 if (event
->attr
.config
!= PERF_COUNT_SW_CPU_CLOCK
)
6118 * no branch sampling for software events
6120 if (has_branch_stack(event
))
6123 perf_swevent_init_hrtimer(event
);
6128 static struct pmu perf_cpu_clock
= {
6129 .task_ctx_nr
= perf_sw_context
,
6131 .event_init
= cpu_clock_event_init
,
6132 .add
= cpu_clock_event_add
,
6133 .del
= cpu_clock_event_del
,
6134 .start
= cpu_clock_event_start
,
6135 .stop
= cpu_clock_event_stop
,
6136 .read
= cpu_clock_event_read
,
6138 .event_idx
= perf_swevent_event_idx
,
6142 * Software event: task time clock
6145 static void task_clock_event_update(struct perf_event
*event
, u64 now
)
6150 prev
= local64_xchg(&event
->hw
.prev_count
, now
);
6152 local64_add(delta
, &event
->count
);
6155 static void task_clock_event_start(struct perf_event
*event
, int flags
)
6157 local64_set(&event
->hw
.prev_count
, event
->ctx
->time
);
6158 perf_swevent_start_hrtimer(event
);
6161 static void task_clock_event_stop(struct perf_event
*event
, int flags
)
6163 perf_swevent_cancel_hrtimer(event
);
6164 task_clock_event_update(event
, event
->ctx
->time
);
6167 static int task_clock_event_add(struct perf_event
*event
, int flags
)
6169 if (flags
& PERF_EF_START
)
6170 task_clock_event_start(event
, flags
);
6175 static void task_clock_event_del(struct perf_event
*event
, int flags
)
6177 task_clock_event_stop(event
, PERF_EF_UPDATE
);
6180 static void task_clock_event_read(struct perf_event
*event
)
6182 u64 now
= perf_clock();
6183 u64 delta
= now
- event
->ctx
->timestamp
;
6184 u64 time
= event
->ctx
->time
+ delta
;
6186 task_clock_event_update(event
, time
);
6189 static int task_clock_event_init(struct perf_event
*event
)
6191 if (event
->attr
.type
!= PERF_TYPE_SOFTWARE
)
6194 if (event
->attr
.config
!= PERF_COUNT_SW_TASK_CLOCK
)
6198 * no branch sampling for software events
6200 if (has_branch_stack(event
))
6203 perf_swevent_init_hrtimer(event
);
6208 static struct pmu perf_task_clock
= {
6209 .task_ctx_nr
= perf_sw_context
,
6211 .event_init
= task_clock_event_init
,
6212 .add
= task_clock_event_add
,
6213 .del
= task_clock_event_del
,
6214 .start
= task_clock_event_start
,
6215 .stop
= task_clock_event_stop
,
6216 .read
= task_clock_event_read
,
6218 .event_idx
= perf_swevent_event_idx
,
6221 static void perf_pmu_nop_void(struct pmu
*pmu
)
6225 static int perf_pmu_nop_int(struct pmu
*pmu
)
6230 static void perf_pmu_start_txn(struct pmu
*pmu
)
6232 perf_pmu_disable(pmu
);
6235 static int perf_pmu_commit_txn(struct pmu
*pmu
)
6237 perf_pmu_enable(pmu
);
6241 static void perf_pmu_cancel_txn(struct pmu
*pmu
)
6243 perf_pmu_enable(pmu
);
6246 static int perf_event_idx_default(struct perf_event
*event
)
6248 return event
->hw
.idx
+ 1;
6252 * Ensures all contexts with the same task_ctx_nr have the same
6253 * pmu_cpu_context too.
6255 static void *find_pmu_context(int ctxn
)
6262 list_for_each_entry(pmu
, &pmus
, entry
) {
6263 if (pmu
->task_ctx_nr
== ctxn
)
6264 return pmu
->pmu_cpu_context
;
6270 static void update_pmu_context(struct pmu
*pmu
, struct pmu
*old_pmu
)
6274 for_each_possible_cpu(cpu
) {
6275 struct perf_cpu_context
*cpuctx
;
6277 cpuctx
= per_cpu_ptr(pmu
->pmu_cpu_context
, cpu
);
6279 if (cpuctx
->unique_pmu
== old_pmu
)
6280 cpuctx
->unique_pmu
= pmu
;
6284 static void free_pmu_context(struct pmu
*pmu
)
6288 mutex_lock(&pmus_lock
);
6290 * Like a real lame refcount.
6292 list_for_each_entry(i
, &pmus
, entry
) {
6293 if (i
->pmu_cpu_context
== pmu
->pmu_cpu_context
) {
6294 update_pmu_context(i
, pmu
);
6299 free_percpu(pmu
->pmu_cpu_context
);
6301 mutex_unlock(&pmus_lock
);
6303 static struct idr pmu_idr
;
6306 type_show(struct device
*dev
, struct device_attribute
*attr
, char *page
)
6308 struct pmu
*pmu
= dev_get_drvdata(dev
);
6310 return snprintf(page
, PAGE_SIZE
-1, "%d\n", pmu
->type
);
6314 perf_event_mux_interval_ms_show(struct device
*dev
,
6315 struct device_attribute
*attr
,
6318 struct pmu
*pmu
= dev_get_drvdata(dev
);
6320 return snprintf(page
, PAGE_SIZE
-1, "%d\n", pmu
->hrtimer_interval_ms
);
6324 perf_event_mux_interval_ms_store(struct device
*dev
,
6325 struct device_attribute
*attr
,
6326 const char *buf
, size_t count
)
6328 struct pmu
*pmu
= dev_get_drvdata(dev
);
6329 int timer
, cpu
, ret
;
6331 ret
= kstrtoint(buf
, 0, &timer
);
6338 /* same value, noting to do */
6339 if (timer
== pmu
->hrtimer_interval_ms
)
6342 pmu
->hrtimer_interval_ms
= timer
;
6344 /* update all cpuctx for this PMU */
6345 for_each_possible_cpu(cpu
) {
6346 struct perf_cpu_context
*cpuctx
;
6347 cpuctx
= per_cpu_ptr(pmu
->pmu_cpu_context
, cpu
);
6348 cpuctx
->hrtimer_interval
= ns_to_ktime(NSEC_PER_MSEC
* timer
);
6350 if (hrtimer_active(&cpuctx
->hrtimer
))
6351 hrtimer_forward_now(&cpuctx
->hrtimer
, cpuctx
->hrtimer_interval
);
6357 static struct device_attribute pmu_dev_attrs
[] = {
6359 __ATTR_RW(perf_event_mux_interval_ms
),
6363 static int pmu_bus_running
;
6364 static struct bus_type pmu_bus
= {
6365 .name
= "event_source",
6366 .dev_attrs
= pmu_dev_attrs
,
6369 static void pmu_dev_release(struct device
*dev
)
6374 static int pmu_dev_alloc(struct pmu
*pmu
)
6378 pmu
->dev
= kzalloc(sizeof(struct device
), GFP_KERNEL
);
6382 pmu
->dev
->groups
= pmu
->attr_groups
;
6383 device_initialize(pmu
->dev
);
6384 ret
= dev_set_name(pmu
->dev
, "%s", pmu
->name
);
6388 dev_set_drvdata(pmu
->dev
, pmu
);
6389 pmu
->dev
->bus
= &pmu_bus
;
6390 pmu
->dev
->release
= pmu_dev_release
;
6391 ret
= device_add(pmu
->dev
);
6399 put_device(pmu
->dev
);
6403 static struct lock_class_key cpuctx_mutex
;
6404 static struct lock_class_key cpuctx_lock
;
6406 int perf_pmu_register(struct pmu
*pmu
, const char *name
, int type
)
6410 mutex_lock(&pmus_lock
);
6412 pmu
->pmu_disable_count
= alloc_percpu(int);
6413 if (!pmu
->pmu_disable_count
)
6422 type
= idr_alloc(&pmu_idr
, pmu
, PERF_TYPE_MAX
, 0, GFP_KERNEL
);
6430 if (pmu_bus_running
) {
6431 ret
= pmu_dev_alloc(pmu
);
6437 pmu
->pmu_cpu_context
= find_pmu_context(pmu
->task_ctx_nr
);
6438 if (pmu
->pmu_cpu_context
)
6439 goto got_cpu_context
;
6442 pmu
->pmu_cpu_context
= alloc_percpu(struct perf_cpu_context
);
6443 if (!pmu
->pmu_cpu_context
)
6446 for_each_possible_cpu(cpu
) {
6447 struct perf_cpu_context
*cpuctx
;
6449 cpuctx
= per_cpu_ptr(pmu
->pmu_cpu_context
, cpu
);
6450 __perf_event_init_context(&cpuctx
->ctx
);
6451 lockdep_set_class(&cpuctx
->ctx
.mutex
, &cpuctx_mutex
);
6452 lockdep_set_class(&cpuctx
->ctx
.lock
, &cpuctx_lock
);
6453 cpuctx
->ctx
.type
= cpu_context
;
6454 cpuctx
->ctx
.pmu
= pmu
;
6456 __perf_cpu_hrtimer_init(cpuctx
, cpu
);
6458 INIT_LIST_HEAD(&cpuctx
->rotation_list
);
6459 cpuctx
->unique_pmu
= pmu
;
6463 if (!pmu
->start_txn
) {
6464 if (pmu
->pmu_enable
) {
6466 * If we have pmu_enable/pmu_disable calls, install
6467 * transaction stubs that use that to try and batch
6468 * hardware accesses.
6470 pmu
->start_txn
= perf_pmu_start_txn
;
6471 pmu
->commit_txn
= perf_pmu_commit_txn
;
6472 pmu
->cancel_txn
= perf_pmu_cancel_txn
;
6474 pmu
->start_txn
= perf_pmu_nop_void
;
6475 pmu
->commit_txn
= perf_pmu_nop_int
;
6476 pmu
->cancel_txn
= perf_pmu_nop_void
;
6480 if (!pmu
->pmu_enable
) {
6481 pmu
->pmu_enable
= perf_pmu_nop_void
;
6482 pmu
->pmu_disable
= perf_pmu_nop_void
;
6485 if (!pmu
->event_idx
)
6486 pmu
->event_idx
= perf_event_idx_default
;
6488 list_add_rcu(&pmu
->entry
, &pmus
);
6491 mutex_unlock(&pmus_lock
);
6496 device_del(pmu
->dev
);
6497 put_device(pmu
->dev
);
6500 if (pmu
->type
>= PERF_TYPE_MAX
)
6501 idr_remove(&pmu_idr
, pmu
->type
);
6504 free_percpu(pmu
->pmu_disable_count
);
6508 void perf_pmu_unregister(struct pmu
*pmu
)
6510 mutex_lock(&pmus_lock
);
6511 list_del_rcu(&pmu
->entry
);
6512 mutex_unlock(&pmus_lock
);
6515 * We dereference the pmu list under both SRCU and regular RCU, so
6516 * synchronize against both of those.
6518 synchronize_srcu(&pmus_srcu
);
6521 free_percpu(pmu
->pmu_disable_count
);
6522 if (pmu
->type
>= PERF_TYPE_MAX
)
6523 idr_remove(&pmu_idr
, pmu
->type
);
6524 device_del(pmu
->dev
);
6525 put_device(pmu
->dev
);
6526 free_pmu_context(pmu
);
6529 struct pmu
*perf_init_event(struct perf_event
*event
)
6531 struct pmu
*pmu
= NULL
;
6535 idx
= srcu_read_lock(&pmus_srcu
);
6538 pmu
= idr_find(&pmu_idr
, event
->attr
.type
);
6542 ret
= pmu
->event_init(event
);
6548 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
6550 ret
= pmu
->event_init(event
);
6554 if (ret
!= -ENOENT
) {
6559 pmu
= ERR_PTR(-ENOENT
);
6561 srcu_read_unlock(&pmus_srcu
, idx
);
6566 static void account_event_cpu(struct perf_event
*event
, int cpu
)
6571 if (has_branch_stack(event
)) {
6572 if (!(event
->attach_state
& PERF_ATTACH_TASK
))
6573 atomic_inc(&per_cpu(perf_branch_stack_events
, cpu
));
6575 if (is_cgroup_event(event
))
6576 atomic_inc(&per_cpu(perf_cgroup_events
, cpu
));
6579 static void account_event(struct perf_event
*event
)
6584 if (event
->attach_state
& PERF_ATTACH_TASK
)
6585 static_key_slow_inc(&perf_sched_events
.key
);
6586 if (event
->attr
.mmap
|| event
->attr
.mmap_data
)
6587 atomic_inc(&nr_mmap_events
);
6588 if (event
->attr
.comm
)
6589 atomic_inc(&nr_comm_events
);
6590 if (event
->attr
.task
)
6591 atomic_inc(&nr_task_events
);
6592 if (event
->attr
.freq
) {
6593 if (atomic_inc_return(&nr_freq_events
) == 1)
6594 tick_nohz_full_kick_all();
6596 if (has_branch_stack(event
))
6597 static_key_slow_inc(&perf_sched_events
.key
);
6598 if (is_cgroup_event(event
))
6599 static_key_slow_inc(&perf_sched_events
.key
);
6601 account_event_cpu(event
, event
->cpu
);
6605 * Allocate and initialize a event structure
6607 static struct perf_event
*
6608 perf_event_alloc(struct perf_event_attr
*attr
, int cpu
,
6609 struct task_struct
*task
,
6610 struct perf_event
*group_leader
,
6611 struct perf_event
*parent_event
,
6612 perf_overflow_handler_t overflow_handler
,
6616 struct perf_event
*event
;
6617 struct hw_perf_event
*hwc
;
6620 if ((unsigned)cpu
>= nr_cpu_ids
) {
6621 if (!task
|| cpu
!= -1)
6622 return ERR_PTR(-EINVAL
);
6625 event
= kzalloc(sizeof(*event
), GFP_KERNEL
);
6627 return ERR_PTR(-ENOMEM
);
6630 * Single events are their own group leaders, with an
6631 * empty sibling list:
6634 group_leader
= event
;
6636 mutex_init(&event
->child_mutex
);
6637 INIT_LIST_HEAD(&event
->child_list
);
6639 INIT_LIST_HEAD(&event
->group_entry
);
6640 INIT_LIST_HEAD(&event
->event_entry
);
6641 INIT_LIST_HEAD(&event
->sibling_list
);
6642 INIT_LIST_HEAD(&event
->rb_entry
);
6644 init_waitqueue_head(&event
->waitq
);
6645 init_irq_work(&event
->pending
, perf_pending_event
);
6647 mutex_init(&event
->mmap_mutex
);
6649 atomic_long_set(&event
->refcount
, 1);
6651 event
->attr
= *attr
;
6652 event
->group_leader
= group_leader
;
6656 event
->parent
= parent_event
;
6658 event
->ns
= get_pid_ns(task_active_pid_ns(current
));
6659 event
->id
= atomic64_inc_return(&perf_event_id
);
6661 event
->state
= PERF_EVENT_STATE_INACTIVE
;
6664 event
->attach_state
= PERF_ATTACH_TASK
;
6666 if (attr
->type
== PERF_TYPE_TRACEPOINT
)
6667 event
->hw
.tp_target
= task
;
6668 #ifdef CONFIG_HAVE_HW_BREAKPOINT
6670 * hw_breakpoint is a bit difficult here..
6672 else if (attr
->type
== PERF_TYPE_BREAKPOINT
)
6673 event
->hw
.bp_target
= task
;
6677 if (!overflow_handler
&& parent_event
) {
6678 overflow_handler
= parent_event
->overflow_handler
;
6679 context
= parent_event
->overflow_handler_context
;
6682 event
->overflow_handler
= overflow_handler
;
6683 event
->overflow_handler_context
= context
;
6685 perf_event__state_init(event
);
6690 hwc
->sample_period
= attr
->sample_period
;
6691 if (attr
->freq
&& attr
->sample_freq
)
6692 hwc
->sample_period
= 1;
6693 hwc
->last_period
= hwc
->sample_period
;
6695 local64_set(&hwc
->period_left
, hwc
->sample_period
);
6698 * we currently do not support PERF_FORMAT_GROUP on inherited events
6700 if (attr
->inherit
&& (attr
->read_format
& PERF_FORMAT_GROUP
))
6703 pmu
= perf_init_event(event
);
6706 else if (IS_ERR(pmu
)) {
6711 if (!event
->parent
) {
6712 if (event
->attr
.sample_type
& PERF_SAMPLE_CALLCHAIN
) {
6713 err
= get_callchain_buffers();
6723 event
->destroy(event
);
6726 put_pid_ns(event
->ns
);
6729 return ERR_PTR(err
);
6732 static int perf_copy_attr(struct perf_event_attr __user
*uattr
,
6733 struct perf_event_attr
*attr
)
6738 if (!access_ok(VERIFY_WRITE
, uattr
, PERF_ATTR_SIZE_VER0
))
6742 * zero the full structure, so that a short copy will be nice.
6744 memset(attr
, 0, sizeof(*attr
));
6746 ret
= get_user(size
, &uattr
->size
);
6750 if (size
> PAGE_SIZE
) /* silly large */
6753 if (!size
) /* abi compat */
6754 size
= PERF_ATTR_SIZE_VER0
;
6756 if (size
< PERF_ATTR_SIZE_VER0
)
6760 * If we're handed a bigger struct than we know of,
6761 * ensure all the unknown bits are 0 - i.e. new
6762 * user-space does not rely on any kernel feature
6763 * extensions we dont know about yet.
6765 if (size
> sizeof(*attr
)) {
6766 unsigned char __user
*addr
;
6767 unsigned char __user
*end
;
6770 addr
= (void __user
*)uattr
+ sizeof(*attr
);
6771 end
= (void __user
*)uattr
+ size
;
6773 for (; addr
< end
; addr
++) {
6774 ret
= get_user(val
, addr
);
6780 size
= sizeof(*attr
);
6783 ret
= copy_from_user(attr
, uattr
, size
);
6787 /* disabled for now */
6791 if (attr
->__reserved_1
)
6794 if (attr
->sample_type
& ~(PERF_SAMPLE_MAX
-1))
6797 if (attr
->read_format
& ~(PERF_FORMAT_MAX
-1))
6800 if (attr
->sample_type
& PERF_SAMPLE_BRANCH_STACK
) {
6801 u64 mask
= attr
->branch_sample_type
;
6803 /* only using defined bits */
6804 if (mask
& ~(PERF_SAMPLE_BRANCH_MAX
-1))
6807 /* at least one branch bit must be set */
6808 if (!(mask
& ~PERF_SAMPLE_BRANCH_PLM_ALL
))
6811 /* propagate priv level, when not set for branch */
6812 if (!(mask
& PERF_SAMPLE_BRANCH_PLM_ALL
)) {
6814 /* exclude_kernel checked on syscall entry */
6815 if (!attr
->exclude_kernel
)
6816 mask
|= PERF_SAMPLE_BRANCH_KERNEL
;
6818 if (!attr
->exclude_user
)
6819 mask
|= PERF_SAMPLE_BRANCH_USER
;
6821 if (!attr
->exclude_hv
)
6822 mask
|= PERF_SAMPLE_BRANCH_HV
;
6824 * adjust user setting (for HW filter setup)
6826 attr
->branch_sample_type
= mask
;
6828 /* privileged levels capture (kernel, hv): check permissions */
6829 if ((mask
& PERF_SAMPLE_BRANCH_PERM_PLM
)
6830 && perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN
))
6834 if (attr
->sample_type
& PERF_SAMPLE_REGS_USER
) {
6835 ret
= perf_reg_validate(attr
->sample_regs_user
);
6840 if (attr
->sample_type
& PERF_SAMPLE_STACK_USER
) {
6841 if (!arch_perf_have_user_stack_dump())
6845 * We have __u32 type for the size, but so far
6846 * we can only use __u16 as maximum due to the
6847 * __u16 sample size limit.
6849 if (attr
->sample_stack_user
>= USHRT_MAX
)
6851 else if (!IS_ALIGNED(attr
->sample_stack_user
, sizeof(u64
)))
6859 put_user(sizeof(*attr
), &uattr
->size
);
6865 perf_event_set_output(struct perf_event
*event
, struct perf_event
*output_event
)
6867 struct ring_buffer
*rb
= NULL
, *old_rb
= NULL
;
6873 /* don't allow circular references */
6874 if (event
== output_event
)
6878 * Don't allow cross-cpu buffers
6880 if (output_event
->cpu
!= event
->cpu
)
6884 * If its not a per-cpu rb, it must be the same task.
6886 if (output_event
->cpu
== -1 && output_event
->ctx
!= event
->ctx
)
6890 mutex_lock(&event
->mmap_mutex
);
6891 /* Can't redirect output if we've got an active mmap() */
6892 if (atomic_read(&event
->mmap_count
))
6898 /* get the rb we want to redirect to */
6899 rb
= ring_buffer_get(output_event
);
6905 ring_buffer_detach(event
, old_rb
);
6908 ring_buffer_attach(event
, rb
);
6910 rcu_assign_pointer(event
->rb
, rb
);
6913 ring_buffer_put(old_rb
);
6915 * Since we detached before setting the new rb, so that we
6916 * could attach the new rb, we could have missed a wakeup.
6919 wake_up_all(&event
->waitq
);
6924 mutex_unlock(&event
->mmap_mutex
);
6931 * sys_perf_event_open - open a performance event, associate it to a task/cpu
6933 * @attr_uptr: event_id type attributes for monitoring/sampling
6936 * @group_fd: group leader event fd
6938 SYSCALL_DEFINE5(perf_event_open
,
6939 struct perf_event_attr __user
*, attr_uptr
,
6940 pid_t
, pid
, int, cpu
, int, group_fd
, unsigned long, flags
)
6942 struct perf_event
*group_leader
= NULL
, *output_event
= NULL
;
6943 struct perf_event
*event
, *sibling
;
6944 struct perf_event_attr attr
;
6945 struct perf_event_context
*ctx
;
6946 struct file
*event_file
= NULL
;
6947 struct fd group
= {NULL
, 0};
6948 struct task_struct
*task
= NULL
;
6954 /* for future expandability... */
6955 if (flags
& ~PERF_FLAG_ALL
)
6958 err
= perf_copy_attr(attr_uptr
, &attr
);
6962 if (!attr
.exclude_kernel
) {
6963 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN
))
6968 if (attr
.sample_freq
> sysctl_perf_event_sample_rate
)
6971 if (attr
.sample_period
& (1ULL << 63))
6976 * In cgroup mode, the pid argument is used to pass the fd
6977 * opened to the cgroup directory in cgroupfs. The cpu argument
6978 * designates the cpu on which to monitor threads from that
6981 if ((flags
& PERF_FLAG_PID_CGROUP
) && (pid
== -1 || cpu
== -1))
6984 event_fd
= get_unused_fd();
6988 if (group_fd
!= -1) {
6989 err
= perf_fget_light(group_fd
, &group
);
6992 group_leader
= group
.file
->private_data
;
6993 if (flags
& PERF_FLAG_FD_OUTPUT
)
6994 output_event
= group_leader
;
6995 if (flags
& PERF_FLAG_FD_NO_GROUP
)
6996 group_leader
= NULL
;
6999 if (pid
!= -1 && !(flags
& PERF_FLAG_PID_CGROUP
)) {
7000 task
= find_lively_task_by_vpid(pid
);
7002 err
= PTR_ERR(task
);
7009 event
= perf_event_alloc(&attr
, cpu
, task
, group_leader
, NULL
,
7011 if (IS_ERR(event
)) {
7012 err
= PTR_ERR(event
);
7016 if (flags
& PERF_FLAG_PID_CGROUP
) {
7017 err
= perf_cgroup_connect(pid
, event
, &attr
, group_leader
);
7019 __free_event(event
);
7024 account_event(event
);
7027 * Special case software events and allow them to be part of
7028 * any hardware group.
7033 (is_software_event(event
) != is_software_event(group_leader
))) {
7034 if (is_software_event(event
)) {
7036 * If event and group_leader are not both a software
7037 * event, and event is, then group leader is not.
7039 * Allow the addition of software events to !software
7040 * groups, this is safe because software events never
7043 pmu
= group_leader
->pmu
;
7044 } else if (is_software_event(group_leader
) &&
7045 (group_leader
->group_flags
& PERF_GROUP_SOFTWARE
)) {
7047 * In case the group is a pure software group, and we
7048 * try to add a hardware event, move the whole group to
7049 * the hardware context.
7056 * Get the target context (task or percpu):
7058 ctx
= find_get_context(pmu
, task
, event
->cpu
);
7065 put_task_struct(task
);
7070 * Look up the group leader (we will attach this event to it):
7076 * Do not allow a recursive hierarchy (this new sibling
7077 * becoming part of another group-sibling):
7079 if (group_leader
->group_leader
!= group_leader
)
7082 * Do not allow to attach to a group in a different
7083 * task or CPU context:
7086 if (group_leader
->ctx
->type
!= ctx
->type
)
7089 if (group_leader
->ctx
!= ctx
)
7094 * Only a group leader can be exclusive or pinned
7096 if (attr
.exclusive
|| attr
.pinned
)
7101 err
= perf_event_set_output(event
, output_event
);
7106 event_file
= anon_inode_getfile("[perf_event]", &perf_fops
, event
, O_RDWR
);
7107 if (IS_ERR(event_file
)) {
7108 err
= PTR_ERR(event_file
);
7113 struct perf_event_context
*gctx
= group_leader
->ctx
;
7115 mutex_lock(&gctx
->mutex
);
7116 perf_remove_from_context(group_leader
, false);
7119 * Removing from the context ends up with disabled
7120 * event. What we want here is event in the initial
7121 * startup state, ready to be add into new context.
7123 perf_event__state_init(group_leader
);
7124 list_for_each_entry(sibling
, &group_leader
->sibling_list
,
7126 perf_remove_from_context(sibling
, false);
7127 perf_event__state_init(sibling
);
7130 mutex_unlock(&gctx
->mutex
);
7134 WARN_ON_ONCE(ctx
->parent_ctx
);
7135 mutex_lock(&ctx
->mutex
);
7139 perf_install_in_context(ctx
, group_leader
, event
->cpu
);
7141 list_for_each_entry(sibling
, &group_leader
->sibling_list
,
7143 perf_install_in_context(ctx
, sibling
, event
->cpu
);
7148 perf_install_in_context(ctx
, event
, event
->cpu
);
7150 perf_unpin_context(ctx
);
7151 mutex_unlock(&ctx
->mutex
);
7155 event
->owner
= current
;
7157 mutex_lock(¤t
->perf_event_mutex
);
7158 list_add_tail(&event
->owner_entry
, ¤t
->perf_event_list
);
7159 mutex_unlock(¤t
->perf_event_mutex
);
7162 * Precalculate sample_data sizes
7164 perf_event__header_size(event
);
7165 perf_event__id_header_size(event
);
7168 * Drop the reference on the group_event after placing the
7169 * new event on the sibling_list. This ensures destruction
7170 * of the group leader will find the pointer to itself in
7171 * perf_group_detach().
7174 fd_install(event_fd
, event_file
);
7178 perf_unpin_context(ctx
);
7185 put_task_struct(task
);
7189 put_unused_fd(event_fd
);
7194 * perf_event_create_kernel_counter
7196 * @attr: attributes of the counter to create
7197 * @cpu: cpu in which the counter is bound
7198 * @task: task to profile (NULL for percpu)
7201 perf_event_create_kernel_counter(struct perf_event_attr
*attr
, int cpu
,
7202 struct task_struct
*task
,
7203 perf_overflow_handler_t overflow_handler
,
7206 struct perf_event_context
*ctx
;
7207 struct perf_event
*event
;
7211 * Get the target context (task or percpu):
7214 event
= perf_event_alloc(attr
, cpu
, task
, NULL
, NULL
,
7215 overflow_handler
, context
);
7216 if (IS_ERR(event
)) {
7217 err
= PTR_ERR(event
);
7221 account_event(event
);
7223 ctx
= find_get_context(event
->pmu
, task
, cpu
);
7229 WARN_ON_ONCE(ctx
->parent_ctx
);
7230 mutex_lock(&ctx
->mutex
);
7231 perf_install_in_context(ctx
, event
, cpu
);
7233 perf_unpin_context(ctx
);
7234 mutex_unlock(&ctx
->mutex
);
7241 return ERR_PTR(err
);
7243 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter
);
7245 void perf_pmu_migrate_context(struct pmu
*pmu
, int src_cpu
, int dst_cpu
)
7247 struct perf_event_context
*src_ctx
;
7248 struct perf_event_context
*dst_ctx
;
7249 struct perf_event
*event
, *tmp
;
7252 src_ctx
= &per_cpu_ptr(pmu
->pmu_cpu_context
, src_cpu
)->ctx
;
7253 dst_ctx
= &per_cpu_ptr(pmu
->pmu_cpu_context
, dst_cpu
)->ctx
;
7255 mutex_lock(&src_ctx
->mutex
);
7256 list_for_each_entry_safe(event
, tmp
, &src_ctx
->event_list
,
7258 perf_remove_from_context(event
, false);
7259 unaccount_event_cpu(event
, src_cpu
);
7261 list_add(&event
->migrate_entry
, &events
);
7263 mutex_unlock(&src_ctx
->mutex
);
7267 mutex_lock(&dst_ctx
->mutex
);
7268 list_for_each_entry_safe(event
, tmp
, &events
, migrate_entry
) {
7269 list_del(&event
->migrate_entry
);
7270 if (event
->state
>= PERF_EVENT_STATE_OFF
)
7271 event
->state
= PERF_EVENT_STATE_INACTIVE
;
7272 account_event_cpu(event
, dst_cpu
);
7273 perf_install_in_context(dst_ctx
, event
, dst_cpu
);
7276 mutex_unlock(&dst_ctx
->mutex
);
7278 EXPORT_SYMBOL_GPL(perf_pmu_migrate_context
);
7280 static void sync_child_event(struct perf_event
*child_event
,
7281 struct task_struct
*child
)
7283 struct perf_event
*parent_event
= child_event
->parent
;
7286 if (child_event
->attr
.inherit_stat
)
7287 perf_event_read_event(child_event
, child
);
7289 child_val
= perf_event_count(child_event
);
7292 * Add back the child's count to the parent's count:
7294 atomic64_add(child_val
, &parent_event
->child_count
);
7295 atomic64_add(child_event
->total_time_enabled
,
7296 &parent_event
->child_total_time_enabled
);
7297 atomic64_add(child_event
->total_time_running
,
7298 &parent_event
->child_total_time_running
);
7301 * Remove this event from the parent's list
7303 WARN_ON_ONCE(parent_event
->ctx
->parent_ctx
);
7304 mutex_lock(&parent_event
->child_mutex
);
7305 list_del_init(&child_event
->child_list
);
7306 mutex_unlock(&parent_event
->child_mutex
);
7309 * Release the parent event, if this was the last
7312 put_event(parent_event
);
7316 __perf_event_exit_task(struct perf_event
*child_event
,
7317 struct perf_event_context
*child_ctx
,
7318 struct task_struct
*child
)
7320 perf_remove_from_context(child_event
, !!child_event
->parent
);
7323 * It can happen that the parent exits first, and has events
7324 * that are still around due to the child reference. These
7325 * events need to be zapped.
7327 if (child_event
->parent
) {
7328 sync_child_event(child_event
, child
);
7329 free_event(child_event
);
7333 static void perf_event_exit_task_context(struct task_struct
*child
, int ctxn
)
7335 struct perf_event
*child_event
, *tmp
;
7336 struct perf_event_context
*child_ctx
;
7337 unsigned long flags
;
7339 if (likely(!child
->perf_event_ctxp
[ctxn
])) {
7340 perf_event_task(child
, NULL
, 0);
7344 local_irq_save(flags
);
7346 * We can't reschedule here because interrupts are disabled,
7347 * and either child is current or it is a task that can't be
7348 * scheduled, so we are now safe from rescheduling changing
7351 child_ctx
= rcu_dereference_raw(child
->perf_event_ctxp
[ctxn
]);
7354 * Take the context lock here so that if find_get_context is
7355 * reading child->perf_event_ctxp, we wait until it has
7356 * incremented the context's refcount before we do put_ctx below.
7358 raw_spin_lock(&child_ctx
->lock
);
7359 task_ctx_sched_out(child_ctx
);
7360 child
->perf_event_ctxp
[ctxn
] = NULL
;
7362 * If this context is a clone; unclone it so it can't get
7363 * swapped to another process while we're removing all
7364 * the events from it.
7366 unclone_ctx(child_ctx
);
7367 update_context_time(child_ctx
);
7368 raw_spin_unlock_irqrestore(&child_ctx
->lock
, flags
);
7371 * Report the task dead after unscheduling the events so that we
7372 * won't get any samples after PERF_RECORD_EXIT. We can however still
7373 * get a few PERF_RECORD_READ events.
7375 perf_event_task(child
, child_ctx
, 0);
7378 * We can recurse on the same lock type through:
7380 * __perf_event_exit_task()
7381 * sync_child_event()
7383 * mutex_lock(&ctx->mutex)
7385 * But since its the parent context it won't be the same instance.
7387 mutex_lock(&child_ctx
->mutex
);
7390 list_for_each_entry_safe(child_event
, tmp
, &child_ctx
->pinned_groups
,
7392 __perf_event_exit_task(child_event
, child_ctx
, child
);
7394 list_for_each_entry_safe(child_event
, tmp
, &child_ctx
->flexible_groups
,
7396 __perf_event_exit_task(child_event
, child_ctx
, child
);
7399 * If the last event was a group event, it will have appended all
7400 * its siblings to the list, but we obtained 'tmp' before that which
7401 * will still point to the list head terminating the iteration.
7403 if (!list_empty(&child_ctx
->pinned_groups
) ||
7404 !list_empty(&child_ctx
->flexible_groups
))
7407 mutex_unlock(&child_ctx
->mutex
);
7413 * When a child task exits, feed back event values to parent events.
7415 void perf_event_exit_task(struct task_struct
*child
)
7417 struct perf_event
*event
, *tmp
;
7420 mutex_lock(&child
->perf_event_mutex
);
7421 list_for_each_entry_safe(event
, tmp
, &child
->perf_event_list
,
7423 list_del_init(&event
->owner_entry
);
7426 * Ensure the list deletion is visible before we clear
7427 * the owner, closes a race against perf_release() where
7428 * we need to serialize on the owner->perf_event_mutex.
7431 event
->owner
= NULL
;
7433 mutex_unlock(&child
->perf_event_mutex
);
7435 for_each_task_context_nr(ctxn
)
7436 perf_event_exit_task_context(child
, ctxn
);
7439 static void perf_free_event(struct perf_event
*event
,
7440 struct perf_event_context
*ctx
)
7442 struct perf_event
*parent
= event
->parent
;
7444 if (WARN_ON_ONCE(!parent
))
7447 mutex_lock(&parent
->child_mutex
);
7448 list_del_init(&event
->child_list
);
7449 mutex_unlock(&parent
->child_mutex
);
7453 perf_group_detach(event
);
7454 list_del_event(event
, ctx
);
7459 * free an unexposed, unused context as created by inheritance by
7460 * perf_event_init_task below, used by fork() in case of fail.
7462 void perf_event_free_task(struct task_struct
*task
)
7464 struct perf_event_context
*ctx
;
7465 struct perf_event
*event
, *tmp
;
7468 for_each_task_context_nr(ctxn
) {
7469 ctx
= task
->perf_event_ctxp
[ctxn
];
7473 mutex_lock(&ctx
->mutex
);
7475 list_for_each_entry_safe(event
, tmp
, &ctx
->pinned_groups
,
7477 perf_free_event(event
, ctx
);
7479 list_for_each_entry_safe(event
, tmp
, &ctx
->flexible_groups
,
7481 perf_free_event(event
, ctx
);
7483 if (!list_empty(&ctx
->pinned_groups
) ||
7484 !list_empty(&ctx
->flexible_groups
))
7487 mutex_unlock(&ctx
->mutex
);
7493 void perf_event_delayed_put(struct task_struct
*task
)
7497 for_each_task_context_nr(ctxn
)
7498 WARN_ON_ONCE(task
->perf_event_ctxp
[ctxn
]);
7502 * inherit a event from parent task to child task:
7504 static struct perf_event
*
7505 inherit_event(struct perf_event
*parent_event
,
7506 struct task_struct
*parent
,
7507 struct perf_event_context
*parent_ctx
,
7508 struct task_struct
*child
,
7509 struct perf_event
*group_leader
,
7510 struct perf_event_context
*child_ctx
)
7512 struct perf_event
*child_event
;
7513 unsigned long flags
;
7516 * Instead of creating recursive hierarchies of events,
7517 * we link inherited events back to the original parent,
7518 * which has a filp for sure, which we use as the reference
7521 if (parent_event
->parent
)
7522 parent_event
= parent_event
->parent
;
7524 child_event
= perf_event_alloc(&parent_event
->attr
,
7527 group_leader
, parent_event
,
7529 if (IS_ERR(child_event
))
7532 if (!atomic_long_inc_not_zero(&parent_event
->refcount
)) {
7533 free_event(child_event
);
7540 * Make the child state follow the state of the parent event,
7541 * not its attr.disabled bit. We hold the parent's mutex,
7542 * so we won't race with perf_event_{en, dis}able_family.
7544 if (parent_event
->state
>= PERF_EVENT_STATE_INACTIVE
)
7545 child_event
->state
= PERF_EVENT_STATE_INACTIVE
;
7547 child_event
->state
= PERF_EVENT_STATE_OFF
;
7549 if (parent_event
->attr
.freq
) {
7550 u64 sample_period
= parent_event
->hw
.sample_period
;
7551 struct hw_perf_event
*hwc
= &child_event
->hw
;
7553 hwc
->sample_period
= sample_period
;
7554 hwc
->last_period
= sample_period
;
7556 local64_set(&hwc
->period_left
, sample_period
);
7559 child_event
->ctx
= child_ctx
;
7560 child_event
->overflow_handler
= parent_event
->overflow_handler
;
7561 child_event
->overflow_handler_context
7562 = parent_event
->overflow_handler_context
;
7565 * Precalculate sample_data sizes
7567 perf_event__header_size(child_event
);
7568 perf_event__id_header_size(child_event
);
7571 * Link it up in the child's context:
7573 raw_spin_lock_irqsave(&child_ctx
->lock
, flags
);
7574 add_event_to_ctx(child_event
, child_ctx
);
7575 raw_spin_unlock_irqrestore(&child_ctx
->lock
, flags
);
7578 * Link this into the parent event's child list
7580 WARN_ON_ONCE(parent_event
->ctx
->parent_ctx
);
7581 mutex_lock(&parent_event
->child_mutex
);
7582 list_add_tail(&child_event
->child_list
, &parent_event
->child_list
);
7583 mutex_unlock(&parent_event
->child_mutex
);
7588 static int inherit_group(struct perf_event
*parent_event
,
7589 struct task_struct
*parent
,
7590 struct perf_event_context
*parent_ctx
,
7591 struct task_struct
*child
,
7592 struct perf_event_context
*child_ctx
)
7594 struct perf_event
*leader
;
7595 struct perf_event
*sub
;
7596 struct perf_event
*child_ctr
;
7598 leader
= inherit_event(parent_event
, parent
, parent_ctx
,
7599 child
, NULL
, child_ctx
);
7601 return PTR_ERR(leader
);
7602 list_for_each_entry(sub
, &parent_event
->sibling_list
, group_entry
) {
7603 child_ctr
= inherit_event(sub
, parent
, parent_ctx
,
7604 child
, leader
, child_ctx
);
7605 if (IS_ERR(child_ctr
))
7606 return PTR_ERR(child_ctr
);
7612 inherit_task_group(struct perf_event
*event
, struct task_struct
*parent
,
7613 struct perf_event_context
*parent_ctx
,
7614 struct task_struct
*child
, int ctxn
,
7618 struct perf_event_context
*child_ctx
;
7620 if (!event
->attr
.inherit
) {
7625 child_ctx
= child
->perf_event_ctxp
[ctxn
];
7628 * This is executed from the parent task context, so
7629 * inherit events that have been marked for cloning.
7630 * First allocate and initialize a context for the
7634 child_ctx
= alloc_perf_context(parent_ctx
->pmu
, child
);
7638 child
->perf_event_ctxp
[ctxn
] = child_ctx
;
7641 ret
= inherit_group(event
, parent
, parent_ctx
,
7651 * Initialize the perf_event context in task_struct
7653 int perf_event_init_context(struct task_struct
*child
, int ctxn
)
7655 struct perf_event_context
*child_ctx
, *parent_ctx
;
7656 struct perf_event_context
*cloned_ctx
;
7657 struct perf_event
*event
;
7658 struct task_struct
*parent
= current
;
7659 int inherited_all
= 1;
7660 unsigned long flags
;
7663 if (likely(!parent
->perf_event_ctxp
[ctxn
]))
7667 * If the parent's context is a clone, pin it so it won't get
7670 parent_ctx
= perf_pin_task_context(parent
, ctxn
);
7673 * No need to check if parent_ctx != NULL here; since we saw
7674 * it non-NULL earlier, the only reason for it to become NULL
7675 * is if we exit, and since we're currently in the middle of
7676 * a fork we can't be exiting at the same time.
7680 * Lock the parent list. No need to lock the child - not PID
7681 * hashed yet and not running, so nobody can access it.
7683 mutex_lock(&parent_ctx
->mutex
);
7686 * We dont have to disable NMIs - we are only looking at
7687 * the list, not manipulating it:
7689 list_for_each_entry(event
, &parent_ctx
->pinned_groups
, group_entry
) {
7690 ret
= inherit_task_group(event
, parent
, parent_ctx
,
7691 child
, ctxn
, &inherited_all
);
7697 * We can't hold ctx->lock when iterating the ->flexible_group list due
7698 * to allocations, but we need to prevent rotation because
7699 * rotate_ctx() will change the list from interrupt context.
7701 raw_spin_lock_irqsave(&parent_ctx
->lock
, flags
);
7702 parent_ctx
->rotate_disable
= 1;
7703 raw_spin_unlock_irqrestore(&parent_ctx
->lock
, flags
);
7705 list_for_each_entry(event
, &parent_ctx
->flexible_groups
, group_entry
) {
7706 ret
= inherit_task_group(event
, parent
, parent_ctx
,
7707 child
, ctxn
, &inherited_all
);
7712 raw_spin_lock_irqsave(&parent_ctx
->lock
, flags
);
7713 parent_ctx
->rotate_disable
= 0;
7715 child_ctx
= child
->perf_event_ctxp
[ctxn
];
7717 if (child_ctx
&& inherited_all
) {
7719 * Mark the child context as a clone of the parent
7720 * context, or of whatever the parent is a clone of.
7722 * Note that if the parent is a clone, the holding of
7723 * parent_ctx->lock avoids it from being uncloned.
7725 cloned_ctx
= parent_ctx
->parent_ctx
;
7727 child_ctx
->parent_ctx
= cloned_ctx
;
7728 child_ctx
->parent_gen
= parent_ctx
->parent_gen
;
7730 child_ctx
->parent_ctx
= parent_ctx
;
7731 child_ctx
->parent_gen
= parent_ctx
->generation
;
7733 get_ctx(child_ctx
->parent_ctx
);
7736 raw_spin_unlock_irqrestore(&parent_ctx
->lock
, flags
);
7737 mutex_unlock(&parent_ctx
->mutex
);
7739 perf_unpin_context(parent_ctx
);
7740 put_ctx(parent_ctx
);
7746 * Initialize the perf_event context in task_struct
7748 int perf_event_init_task(struct task_struct
*child
)
7752 memset(child
->perf_event_ctxp
, 0, sizeof(child
->perf_event_ctxp
));
7753 mutex_init(&child
->perf_event_mutex
);
7754 INIT_LIST_HEAD(&child
->perf_event_list
);
7756 for_each_task_context_nr(ctxn
) {
7757 ret
= perf_event_init_context(child
, ctxn
);
7765 static void __init
perf_event_init_all_cpus(void)
7767 struct swevent_htable
*swhash
;
7770 for_each_possible_cpu(cpu
) {
7771 swhash
= &per_cpu(swevent_htable
, cpu
);
7772 mutex_init(&swhash
->hlist_mutex
);
7773 INIT_LIST_HEAD(&per_cpu(rotation_list
, cpu
));
7777 static void perf_event_init_cpu(int cpu
)
7779 struct swevent_htable
*swhash
= &per_cpu(swevent_htable
, cpu
);
7781 mutex_lock(&swhash
->hlist_mutex
);
7782 swhash
->online
= true;
7783 if (swhash
->hlist_refcount
> 0) {
7784 struct swevent_hlist
*hlist
;
7786 hlist
= kzalloc_node(sizeof(*hlist
), GFP_KERNEL
, cpu_to_node(cpu
));
7788 rcu_assign_pointer(swhash
->swevent_hlist
, hlist
);
7790 mutex_unlock(&swhash
->hlist_mutex
);
7793 #if defined CONFIG_HOTPLUG_CPU || defined CONFIG_KEXEC
7794 static void perf_pmu_rotate_stop(struct pmu
*pmu
)
7796 struct perf_cpu_context
*cpuctx
= this_cpu_ptr(pmu
->pmu_cpu_context
);
7798 WARN_ON(!irqs_disabled());
7800 list_del_init(&cpuctx
->rotation_list
);
7803 static void __perf_event_exit_context(void *__info
)
7805 struct remove_event re
= { .detach_group
= false };
7806 struct perf_event_context
*ctx
= __info
;
7808 perf_pmu_rotate_stop(ctx
->pmu
);
7811 list_for_each_entry_rcu(re
.event
, &ctx
->event_list
, event_entry
)
7812 __perf_remove_from_context(&re
);
7816 static void perf_event_exit_cpu_context(int cpu
)
7818 struct perf_event_context
*ctx
;
7822 idx
= srcu_read_lock(&pmus_srcu
);
7823 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
7824 ctx
= &per_cpu_ptr(pmu
->pmu_cpu_context
, cpu
)->ctx
;
7826 mutex_lock(&ctx
->mutex
);
7827 smp_call_function_single(cpu
, __perf_event_exit_context
, ctx
, 1);
7828 mutex_unlock(&ctx
->mutex
);
7830 srcu_read_unlock(&pmus_srcu
, idx
);
7833 static void perf_event_exit_cpu(int cpu
)
7835 struct swevent_htable
*swhash
= &per_cpu(swevent_htable
, cpu
);
7837 perf_event_exit_cpu_context(cpu
);
7839 mutex_lock(&swhash
->hlist_mutex
);
7840 swhash
->online
= false;
7841 swevent_hlist_release(swhash
);
7842 mutex_unlock(&swhash
->hlist_mutex
);
7845 static inline void perf_event_exit_cpu(int cpu
) { }
7849 perf_reboot(struct notifier_block
*notifier
, unsigned long val
, void *v
)
7853 for_each_online_cpu(cpu
)
7854 perf_event_exit_cpu(cpu
);
7860 * Run the perf reboot notifier at the very last possible moment so that
7861 * the generic watchdog code runs as long as possible.
7863 static struct notifier_block perf_reboot_notifier
= {
7864 .notifier_call
= perf_reboot
,
7865 .priority
= INT_MIN
,
7869 perf_cpu_notify(struct notifier_block
*self
, unsigned long action
, void *hcpu
)
7871 unsigned int cpu
= (long)hcpu
;
7873 switch (action
& ~CPU_TASKS_FROZEN
) {
7875 case CPU_UP_PREPARE
:
7876 case CPU_DOWN_FAILED
:
7877 perf_event_init_cpu(cpu
);
7880 case CPU_UP_CANCELED
:
7881 case CPU_DOWN_PREPARE
:
7882 perf_event_exit_cpu(cpu
);
7891 void __init
perf_event_init(void)
7897 perf_event_init_all_cpus();
7898 init_srcu_struct(&pmus_srcu
);
7899 perf_pmu_register(&perf_swevent
, "software", PERF_TYPE_SOFTWARE
);
7900 perf_pmu_register(&perf_cpu_clock
, NULL
, -1);
7901 perf_pmu_register(&perf_task_clock
, NULL
, -1);
7903 perf_cpu_notifier(perf_cpu_notify
);
7904 register_reboot_notifier(&perf_reboot_notifier
);
7906 ret
= init_hw_breakpoint();
7907 WARN(ret
, "hw_breakpoint initialization failed with: %d", ret
);
7909 /* do not patch jump label more than once per second */
7910 jump_label_rate_limit(&perf_sched_events
, HZ
);
7913 * Build time assertion that we keep the data_head at the intended
7914 * location. IOW, validation we got the __reserved[] size right.
7916 BUILD_BUG_ON((offsetof(struct perf_event_mmap_page
, data_head
))
7920 static int __init
perf_event_sysfs_init(void)
7925 mutex_lock(&pmus_lock
);
7927 ret
= bus_register(&pmu_bus
);
7931 list_for_each_entry(pmu
, &pmus
, entry
) {
7932 if (!pmu
->name
|| pmu
->type
< 0)
7935 ret
= pmu_dev_alloc(pmu
);
7936 WARN(ret
, "Failed to register pmu: %s, reason %d\n", pmu
->name
, ret
);
7938 pmu_bus_running
= 1;
7942 mutex_unlock(&pmus_lock
);
7946 device_initcall(perf_event_sysfs_init
);
7948 #ifdef CONFIG_CGROUP_PERF
7949 static struct cgroup_subsys_state
*
7950 perf_cgroup_css_alloc(struct cgroup_subsys_state
*parent_css
)
7952 struct perf_cgroup
*jc
;
7954 jc
= kzalloc(sizeof(*jc
), GFP_KERNEL
);
7956 return ERR_PTR(-ENOMEM
);
7958 jc
->info
= alloc_percpu(struct perf_cgroup_info
);
7961 return ERR_PTR(-ENOMEM
);
7967 static void perf_cgroup_css_free(struct cgroup_subsys_state
*css
)
7969 struct perf_cgroup
*jc
= container_of(css
, struct perf_cgroup
, css
);
7971 free_percpu(jc
->info
);
7975 static int __perf_cgroup_move(void *info
)
7977 struct task_struct
*task
= info
;
7978 perf_cgroup_switch(task
, PERF_CGROUP_SWOUT
| PERF_CGROUP_SWIN
);
7982 static void perf_cgroup_attach(struct cgroup_subsys_state
*css
,
7983 struct cgroup_taskset
*tset
)
7985 struct task_struct
*task
;
7987 cgroup_taskset_for_each(task
, css
, tset
)
7988 task_function_call(task
, __perf_cgroup_move
, task
);
7991 static void perf_cgroup_exit(struct cgroup_subsys_state
*css
,
7992 struct cgroup_subsys_state
*old_css
,
7993 struct task_struct
*task
)
7996 * cgroup_exit() is called in the copy_process() failure path.
7997 * Ignore this case since the task hasn't ran yet, this avoids
7998 * trying to poke a half freed task state from generic code.
8000 if (!(task
->flags
& PF_EXITING
))
8003 task_function_call(task
, __perf_cgroup_move
, task
);
8006 struct cgroup_subsys perf_subsys
= {
8007 .name
= "perf_event",
8008 .subsys_id
= perf_subsys_id
,
8009 .css_alloc
= perf_cgroup_css_alloc
,
8010 .css_free
= perf_cgroup_css_free
,
8011 .exit
= perf_cgroup_exit
,
8012 .attach
= perf_cgroup_attach
,
8014 #endif /* CONFIG_CGROUP_PERF */