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(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;
1430 * Cross CPU call to remove a performance event
1432 * We disable the event on the hardware level first. After that we
1433 * remove it from the context list.
1435 static int __perf_remove_from_context(void *info
)
1437 struct perf_event
*event
= info
;
1438 struct perf_event_context
*ctx
= event
->ctx
;
1439 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
1441 raw_spin_lock(&ctx
->lock
);
1442 event_sched_out(event
, cpuctx
, ctx
);
1443 list_del_event(event
, ctx
);
1444 if (!ctx
->nr_events
&& cpuctx
->task_ctx
== ctx
) {
1446 cpuctx
->task_ctx
= NULL
;
1448 raw_spin_unlock(&ctx
->lock
);
1455 * Remove the event from a task's (or a CPU's) list of events.
1457 * CPU events are removed with a smp call. For task events we only
1458 * call when the task is on a CPU.
1460 * If event->ctx is a cloned context, callers must make sure that
1461 * every task struct that event->ctx->task could possibly point to
1462 * remains valid. This is OK when called from perf_release since
1463 * that only calls us on the top-level context, which can't be a clone.
1464 * When called from perf_event_exit_task, it's OK because the
1465 * context has been detached from its task.
1467 static void perf_remove_from_context(struct perf_event
*event
)
1469 struct perf_event_context
*ctx
= event
->ctx
;
1470 struct task_struct
*task
= ctx
->task
;
1472 lockdep_assert_held(&ctx
->mutex
);
1476 * Per cpu events are removed via an smp call and
1477 * the removal is always successful.
1479 cpu_function_call(event
->cpu
, __perf_remove_from_context
, event
);
1484 if (!task_function_call(task
, __perf_remove_from_context
, event
))
1487 raw_spin_lock_irq(&ctx
->lock
);
1489 * If we failed to find a running task, but find the context active now
1490 * that we've acquired the ctx->lock, retry.
1492 if (ctx
->is_active
) {
1493 raw_spin_unlock_irq(&ctx
->lock
);
1498 * Since the task isn't running, its safe to remove the event, us
1499 * holding the ctx->lock ensures the task won't get scheduled in.
1501 list_del_event(event
, ctx
);
1502 raw_spin_unlock_irq(&ctx
->lock
);
1506 * Cross CPU call to disable a performance event
1508 int __perf_event_disable(void *info
)
1510 struct perf_event
*event
= info
;
1511 struct perf_event_context
*ctx
= event
->ctx
;
1512 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
1515 * If this is a per-task event, need to check whether this
1516 * event's task is the current task on this cpu.
1518 * Can trigger due to concurrent perf_event_context_sched_out()
1519 * flipping contexts around.
1521 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
)
1524 raw_spin_lock(&ctx
->lock
);
1527 * If the event is on, turn it off.
1528 * If it is in error state, leave it in error state.
1530 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
) {
1531 update_context_time(ctx
);
1532 update_cgrp_time_from_event(event
);
1533 update_group_times(event
);
1534 if (event
== event
->group_leader
)
1535 group_sched_out(event
, cpuctx
, ctx
);
1537 event_sched_out(event
, cpuctx
, ctx
);
1538 event
->state
= PERF_EVENT_STATE_OFF
;
1541 raw_spin_unlock(&ctx
->lock
);
1549 * If event->ctx is a cloned context, callers must make sure that
1550 * every task struct that event->ctx->task could possibly point to
1551 * remains valid. This condition is satisifed when called through
1552 * perf_event_for_each_child or perf_event_for_each because they
1553 * hold the top-level event's child_mutex, so any descendant that
1554 * goes to exit will block in sync_child_event.
1555 * When called from perf_pending_event it's OK because event->ctx
1556 * is the current context on this CPU and preemption is disabled,
1557 * hence we can't get into perf_event_task_sched_out for this context.
1559 void perf_event_disable(struct perf_event
*event
)
1561 struct perf_event_context
*ctx
= event
->ctx
;
1562 struct task_struct
*task
= ctx
->task
;
1566 * Disable the event on the cpu that it's on
1568 cpu_function_call(event
->cpu
, __perf_event_disable
, event
);
1573 if (!task_function_call(task
, __perf_event_disable
, event
))
1576 raw_spin_lock_irq(&ctx
->lock
);
1578 * If the event is still active, we need to retry the cross-call.
1580 if (event
->state
== PERF_EVENT_STATE_ACTIVE
) {
1581 raw_spin_unlock_irq(&ctx
->lock
);
1583 * Reload the task pointer, it might have been changed by
1584 * a concurrent perf_event_context_sched_out().
1591 * Since we have the lock this context can't be scheduled
1592 * in, so we can change the state safely.
1594 if (event
->state
== PERF_EVENT_STATE_INACTIVE
) {
1595 update_group_times(event
);
1596 event
->state
= PERF_EVENT_STATE_OFF
;
1598 raw_spin_unlock_irq(&ctx
->lock
);
1600 EXPORT_SYMBOL_GPL(perf_event_disable
);
1602 static void perf_set_shadow_time(struct perf_event
*event
,
1603 struct perf_event_context
*ctx
,
1607 * use the correct time source for the time snapshot
1609 * We could get by without this by leveraging the
1610 * fact that to get to this function, the caller
1611 * has most likely already called update_context_time()
1612 * and update_cgrp_time_xx() and thus both timestamp
1613 * are identical (or very close). Given that tstamp is,
1614 * already adjusted for cgroup, we could say that:
1615 * tstamp - ctx->timestamp
1617 * tstamp - cgrp->timestamp.
1619 * Then, in perf_output_read(), the calculation would
1620 * work with no changes because:
1621 * - event is guaranteed scheduled in
1622 * - no scheduled out in between
1623 * - thus the timestamp would be the same
1625 * But this is a bit hairy.
1627 * So instead, we have an explicit cgroup call to remain
1628 * within the time time source all along. We believe it
1629 * is cleaner and simpler to understand.
1631 if (is_cgroup_event(event
))
1632 perf_cgroup_set_shadow_time(event
, tstamp
);
1634 event
->shadow_ctx_time
= tstamp
- ctx
->timestamp
;
1637 #define MAX_INTERRUPTS (~0ULL)
1639 static void perf_log_throttle(struct perf_event
*event
, int enable
);
1642 event_sched_in(struct perf_event
*event
,
1643 struct perf_cpu_context
*cpuctx
,
1644 struct perf_event_context
*ctx
)
1646 u64 tstamp
= perf_event_time(event
);
1648 if (event
->state
<= PERF_EVENT_STATE_OFF
)
1651 event
->state
= PERF_EVENT_STATE_ACTIVE
;
1652 event
->oncpu
= smp_processor_id();
1655 * Unthrottle events, since we scheduled we might have missed several
1656 * ticks already, also for a heavily scheduling task there is little
1657 * guarantee it'll get a tick in a timely manner.
1659 if (unlikely(event
->hw
.interrupts
== MAX_INTERRUPTS
)) {
1660 perf_log_throttle(event
, 1);
1661 event
->hw
.interrupts
= 0;
1665 * The new state must be visible before we turn it on in the hardware:
1669 if (event
->pmu
->add(event
, PERF_EF_START
)) {
1670 event
->state
= PERF_EVENT_STATE_INACTIVE
;
1675 event
->tstamp_running
+= tstamp
- event
->tstamp_stopped
;
1677 perf_set_shadow_time(event
, ctx
, tstamp
);
1679 if (!is_software_event(event
))
1680 cpuctx
->active_oncpu
++;
1682 if (event
->attr
.freq
&& event
->attr
.sample_freq
)
1685 if (event
->attr
.exclusive
)
1686 cpuctx
->exclusive
= 1;
1692 group_sched_in(struct perf_event
*group_event
,
1693 struct perf_cpu_context
*cpuctx
,
1694 struct perf_event_context
*ctx
)
1696 struct perf_event
*event
, *partial_group
= NULL
;
1697 struct pmu
*pmu
= group_event
->pmu
;
1698 u64 now
= ctx
->time
;
1699 bool simulate
= false;
1701 if (group_event
->state
== PERF_EVENT_STATE_OFF
)
1704 pmu
->start_txn(pmu
);
1706 if (event_sched_in(group_event
, cpuctx
, ctx
)) {
1707 pmu
->cancel_txn(pmu
);
1708 perf_cpu_hrtimer_restart(cpuctx
);
1713 * Schedule in siblings as one group (if any):
1715 list_for_each_entry(event
, &group_event
->sibling_list
, group_entry
) {
1716 if (event_sched_in(event
, cpuctx
, ctx
)) {
1717 partial_group
= event
;
1722 if (!pmu
->commit_txn(pmu
))
1727 * Groups can be scheduled in as one unit only, so undo any
1728 * partial group before returning:
1729 * The events up to the failed event are scheduled out normally,
1730 * tstamp_stopped will be updated.
1732 * The failed events and the remaining siblings need to have
1733 * their timings updated as if they had gone thru event_sched_in()
1734 * and event_sched_out(). This is required to get consistent timings
1735 * across the group. This also takes care of the case where the group
1736 * could never be scheduled by ensuring tstamp_stopped is set to mark
1737 * the time the event was actually stopped, such that time delta
1738 * calculation in update_event_times() is correct.
1740 list_for_each_entry(event
, &group_event
->sibling_list
, group_entry
) {
1741 if (event
== partial_group
)
1745 event
->tstamp_running
+= now
- event
->tstamp_stopped
;
1746 event
->tstamp_stopped
= now
;
1748 event_sched_out(event
, cpuctx
, ctx
);
1751 event_sched_out(group_event
, cpuctx
, ctx
);
1753 pmu
->cancel_txn(pmu
);
1755 perf_cpu_hrtimer_restart(cpuctx
);
1761 * Work out whether we can put this event group on the CPU now.
1763 static int group_can_go_on(struct perf_event
*event
,
1764 struct perf_cpu_context
*cpuctx
,
1768 * Groups consisting entirely of software events can always go on.
1770 if (event
->group_flags
& PERF_GROUP_SOFTWARE
)
1773 * If an exclusive group is already on, no other hardware
1776 if (cpuctx
->exclusive
)
1779 * If this group is exclusive and there are already
1780 * events on the CPU, it can't go on.
1782 if (event
->attr
.exclusive
&& cpuctx
->active_oncpu
)
1785 * Otherwise, try to add it if all previous groups were able
1791 static void add_event_to_ctx(struct perf_event
*event
,
1792 struct perf_event_context
*ctx
)
1794 u64 tstamp
= perf_event_time(event
);
1796 list_add_event(event
, ctx
);
1797 perf_group_attach(event
);
1798 event
->tstamp_enabled
= tstamp
;
1799 event
->tstamp_running
= tstamp
;
1800 event
->tstamp_stopped
= tstamp
;
1803 static void task_ctx_sched_out(struct perf_event_context
*ctx
);
1805 ctx_sched_in(struct perf_event_context
*ctx
,
1806 struct perf_cpu_context
*cpuctx
,
1807 enum event_type_t event_type
,
1808 struct task_struct
*task
);
1810 static void perf_event_sched_in(struct perf_cpu_context
*cpuctx
,
1811 struct perf_event_context
*ctx
,
1812 struct task_struct
*task
)
1814 cpu_ctx_sched_in(cpuctx
, EVENT_PINNED
, task
);
1816 ctx_sched_in(ctx
, cpuctx
, EVENT_PINNED
, task
);
1817 cpu_ctx_sched_in(cpuctx
, EVENT_FLEXIBLE
, task
);
1819 ctx_sched_in(ctx
, cpuctx
, EVENT_FLEXIBLE
, task
);
1823 * Cross CPU call to install and enable a performance event
1825 * Must be called with ctx->mutex held
1827 static int __perf_install_in_context(void *info
)
1829 struct perf_event
*event
= info
;
1830 struct perf_event_context
*ctx
= event
->ctx
;
1831 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
1832 struct perf_event_context
*task_ctx
= cpuctx
->task_ctx
;
1833 struct task_struct
*task
= current
;
1835 perf_ctx_lock(cpuctx
, task_ctx
);
1836 perf_pmu_disable(cpuctx
->ctx
.pmu
);
1839 * If there was an active task_ctx schedule it out.
1842 task_ctx_sched_out(task_ctx
);
1845 * If the context we're installing events in is not the
1846 * active task_ctx, flip them.
1848 if (ctx
->task
&& task_ctx
!= ctx
) {
1850 raw_spin_unlock(&task_ctx
->lock
);
1851 raw_spin_lock(&ctx
->lock
);
1856 cpuctx
->task_ctx
= task_ctx
;
1857 task
= task_ctx
->task
;
1860 cpu_ctx_sched_out(cpuctx
, EVENT_ALL
);
1862 update_context_time(ctx
);
1864 * update cgrp time only if current cgrp
1865 * matches event->cgrp. Must be done before
1866 * calling add_event_to_ctx()
1868 update_cgrp_time_from_event(event
);
1870 add_event_to_ctx(event
, ctx
);
1873 * Schedule everything back in
1875 perf_event_sched_in(cpuctx
, task_ctx
, task
);
1877 perf_pmu_enable(cpuctx
->ctx
.pmu
);
1878 perf_ctx_unlock(cpuctx
, task_ctx
);
1884 * Attach a performance event to a context
1886 * First we add the event to the list with the hardware enable bit
1887 * in event->hw_config cleared.
1889 * If the event is attached to a task which is on a CPU we use a smp
1890 * call to enable it in the task context. The task might have been
1891 * scheduled away, but we check this in the smp call again.
1894 perf_install_in_context(struct perf_event_context
*ctx
,
1895 struct perf_event
*event
,
1898 struct task_struct
*task
= ctx
->task
;
1900 lockdep_assert_held(&ctx
->mutex
);
1903 if (event
->cpu
!= -1)
1908 * Per cpu events are installed via an smp call and
1909 * the install is always successful.
1911 cpu_function_call(cpu
, __perf_install_in_context
, event
);
1916 if (!task_function_call(task
, __perf_install_in_context
, event
))
1919 raw_spin_lock_irq(&ctx
->lock
);
1921 * If we failed to find a running task, but find the context active now
1922 * that we've acquired the ctx->lock, retry.
1924 if (ctx
->is_active
) {
1925 raw_spin_unlock_irq(&ctx
->lock
);
1930 * Since the task isn't running, its safe to add the event, us holding
1931 * the ctx->lock ensures the task won't get scheduled in.
1933 add_event_to_ctx(event
, ctx
);
1934 raw_spin_unlock_irq(&ctx
->lock
);
1938 * Put a event into inactive state and update time fields.
1939 * Enabling the leader of a group effectively enables all
1940 * the group members that aren't explicitly disabled, so we
1941 * have to update their ->tstamp_enabled also.
1942 * Note: this works for group members as well as group leaders
1943 * since the non-leader members' sibling_lists will be empty.
1945 static void __perf_event_mark_enabled(struct perf_event
*event
)
1947 struct perf_event
*sub
;
1948 u64 tstamp
= perf_event_time(event
);
1950 event
->state
= PERF_EVENT_STATE_INACTIVE
;
1951 event
->tstamp_enabled
= tstamp
- event
->total_time_enabled
;
1952 list_for_each_entry(sub
, &event
->sibling_list
, group_entry
) {
1953 if (sub
->state
>= PERF_EVENT_STATE_INACTIVE
)
1954 sub
->tstamp_enabled
= tstamp
- sub
->total_time_enabled
;
1959 * Cross CPU call to enable a performance event
1961 static int __perf_event_enable(void *info
)
1963 struct perf_event
*event
= info
;
1964 struct perf_event_context
*ctx
= event
->ctx
;
1965 struct perf_event
*leader
= event
->group_leader
;
1966 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
1970 * There's a time window between 'ctx->is_active' check
1971 * in perf_event_enable function and this place having:
1973 * - ctx->lock unlocked
1975 * where the task could be killed and 'ctx' deactivated
1976 * by perf_event_exit_task.
1978 if (!ctx
->is_active
)
1981 raw_spin_lock(&ctx
->lock
);
1982 update_context_time(ctx
);
1984 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
)
1988 * set current task's cgroup time reference point
1990 perf_cgroup_set_timestamp(current
, ctx
);
1992 __perf_event_mark_enabled(event
);
1994 if (!event_filter_match(event
)) {
1995 if (is_cgroup_event(event
))
1996 perf_cgroup_defer_enabled(event
);
2001 * If the event is in a group and isn't the group leader,
2002 * then don't put it on unless the group is on.
2004 if (leader
!= event
&& leader
->state
!= PERF_EVENT_STATE_ACTIVE
)
2007 if (!group_can_go_on(event
, cpuctx
, 1)) {
2010 if (event
== leader
)
2011 err
= group_sched_in(event
, cpuctx
, ctx
);
2013 err
= event_sched_in(event
, cpuctx
, ctx
);
2018 * If this event can't go on and it's part of a
2019 * group, then the whole group has to come off.
2021 if (leader
!= event
) {
2022 group_sched_out(leader
, cpuctx
, ctx
);
2023 perf_cpu_hrtimer_restart(cpuctx
);
2025 if (leader
->attr
.pinned
) {
2026 update_group_times(leader
);
2027 leader
->state
= PERF_EVENT_STATE_ERROR
;
2032 raw_spin_unlock(&ctx
->lock
);
2040 * If event->ctx is a cloned context, callers must make sure that
2041 * every task struct that event->ctx->task could possibly point to
2042 * remains valid. This condition is satisfied when called through
2043 * perf_event_for_each_child or perf_event_for_each as described
2044 * for perf_event_disable.
2046 void perf_event_enable(struct perf_event
*event
)
2048 struct perf_event_context
*ctx
= event
->ctx
;
2049 struct task_struct
*task
= ctx
->task
;
2053 * Enable the event on the cpu that it's on
2055 cpu_function_call(event
->cpu
, __perf_event_enable
, event
);
2059 raw_spin_lock_irq(&ctx
->lock
);
2060 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
)
2064 * If the event is in error state, clear that first.
2065 * That way, if we see the event in error state below, we
2066 * know that it has gone back into error state, as distinct
2067 * from the task having been scheduled away before the
2068 * cross-call arrived.
2070 if (event
->state
== PERF_EVENT_STATE_ERROR
)
2071 event
->state
= PERF_EVENT_STATE_OFF
;
2074 if (!ctx
->is_active
) {
2075 __perf_event_mark_enabled(event
);
2079 raw_spin_unlock_irq(&ctx
->lock
);
2081 if (!task_function_call(task
, __perf_event_enable
, event
))
2084 raw_spin_lock_irq(&ctx
->lock
);
2087 * If the context is active and the event is still off,
2088 * we need to retry the cross-call.
2090 if (ctx
->is_active
&& event
->state
== PERF_EVENT_STATE_OFF
) {
2092 * task could have been flipped by a concurrent
2093 * perf_event_context_sched_out()
2100 raw_spin_unlock_irq(&ctx
->lock
);
2102 EXPORT_SYMBOL_GPL(perf_event_enable
);
2104 int perf_event_refresh(struct perf_event
*event
, int refresh
)
2107 * not supported on inherited events
2109 if (event
->attr
.inherit
|| !is_sampling_event(event
))
2112 atomic_add(refresh
, &event
->event_limit
);
2113 perf_event_enable(event
);
2117 EXPORT_SYMBOL_GPL(perf_event_refresh
);
2119 static void ctx_sched_out(struct perf_event_context
*ctx
,
2120 struct perf_cpu_context
*cpuctx
,
2121 enum event_type_t event_type
)
2123 struct perf_event
*event
;
2124 int is_active
= ctx
->is_active
;
2126 ctx
->is_active
&= ~event_type
;
2127 if (likely(!ctx
->nr_events
))
2130 update_context_time(ctx
);
2131 update_cgrp_time_from_cpuctx(cpuctx
);
2132 if (!ctx
->nr_active
)
2135 perf_pmu_disable(ctx
->pmu
);
2136 if ((is_active
& EVENT_PINNED
) && (event_type
& EVENT_PINNED
)) {
2137 list_for_each_entry(event
, &ctx
->pinned_groups
, group_entry
)
2138 group_sched_out(event
, cpuctx
, ctx
);
2141 if ((is_active
& EVENT_FLEXIBLE
) && (event_type
& EVENT_FLEXIBLE
)) {
2142 list_for_each_entry(event
, &ctx
->flexible_groups
, group_entry
)
2143 group_sched_out(event
, cpuctx
, ctx
);
2145 perf_pmu_enable(ctx
->pmu
);
2149 * Test whether two contexts are equivalent, i.e. whether they
2150 * have both been cloned from the same version of the same context
2151 * and they both have the same number of enabled events.
2152 * If the number of enabled events is the same, then the set
2153 * of enabled events should be the same, because these are both
2154 * inherited contexts, therefore we can't access individual events
2155 * in them directly with an fd; we can only enable/disable all
2156 * events via prctl, or enable/disable all events in a family
2157 * via ioctl, which will have the same effect on both contexts.
2159 static int context_equiv(struct perf_event_context
*ctx1
,
2160 struct perf_event_context
*ctx2
)
2162 return ctx1
->parent_ctx
&& ctx1
->parent_ctx
== ctx2
->parent_ctx
2163 && ctx1
->parent_gen
== ctx2
->parent_gen
2164 && !ctx1
->pin_count
&& !ctx2
->pin_count
;
2167 static void __perf_event_sync_stat(struct perf_event
*event
,
2168 struct perf_event
*next_event
)
2172 if (!event
->attr
.inherit_stat
)
2176 * Update the event value, we cannot use perf_event_read()
2177 * because we're in the middle of a context switch and have IRQs
2178 * disabled, which upsets smp_call_function_single(), however
2179 * we know the event must be on the current CPU, therefore we
2180 * don't need to use it.
2182 switch (event
->state
) {
2183 case PERF_EVENT_STATE_ACTIVE
:
2184 event
->pmu
->read(event
);
2187 case PERF_EVENT_STATE_INACTIVE
:
2188 update_event_times(event
);
2196 * In order to keep per-task stats reliable we need to flip the event
2197 * values when we flip the contexts.
2199 value
= local64_read(&next_event
->count
);
2200 value
= local64_xchg(&event
->count
, value
);
2201 local64_set(&next_event
->count
, value
);
2203 swap(event
->total_time_enabled
, next_event
->total_time_enabled
);
2204 swap(event
->total_time_running
, next_event
->total_time_running
);
2207 * Since we swizzled the values, update the user visible data too.
2209 perf_event_update_userpage(event
);
2210 perf_event_update_userpage(next_event
);
2213 #define list_next_entry(pos, member) \
2214 list_entry(pos->member.next, typeof(*pos), member)
2216 static void perf_event_sync_stat(struct perf_event_context
*ctx
,
2217 struct perf_event_context
*next_ctx
)
2219 struct perf_event
*event
, *next_event
;
2224 update_context_time(ctx
);
2226 event
= list_first_entry(&ctx
->event_list
,
2227 struct perf_event
, event_entry
);
2229 next_event
= list_first_entry(&next_ctx
->event_list
,
2230 struct perf_event
, event_entry
);
2232 while (&event
->event_entry
!= &ctx
->event_list
&&
2233 &next_event
->event_entry
!= &next_ctx
->event_list
) {
2235 __perf_event_sync_stat(event
, next_event
);
2237 event
= list_next_entry(event
, event_entry
);
2238 next_event
= list_next_entry(next_event
, event_entry
);
2242 static void perf_event_context_sched_out(struct task_struct
*task
, int ctxn
,
2243 struct task_struct
*next
)
2245 struct perf_event_context
*ctx
= task
->perf_event_ctxp
[ctxn
];
2246 struct perf_event_context
*next_ctx
;
2247 struct perf_event_context
*parent
;
2248 struct perf_cpu_context
*cpuctx
;
2254 cpuctx
= __get_cpu_context(ctx
);
2255 if (!cpuctx
->task_ctx
)
2259 parent
= rcu_dereference(ctx
->parent_ctx
);
2260 next_ctx
= next
->perf_event_ctxp
[ctxn
];
2261 if (parent
&& next_ctx
&&
2262 rcu_dereference(next_ctx
->parent_ctx
) == parent
) {
2264 * Looks like the two contexts are clones, so we might be
2265 * able to optimize the context switch. We lock both
2266 * contexts and check that they are clones under the
2267 * lock (including re-checking that neither has been
2268 * uncloned in the meantime). It doesn't matter which
2269 * order we take the locks because no other cpu could
2270 * be trying to lock both of these tasks.
2272 raw_spin_lock(&ctx
->lock
);
2273 raw_spin_lock_nested(&next_ctx
->lock
, SINGLE_DEPTH_NESTING
);
2274 if (context_equiv(ctx
, next_ctx
)) {
2276 * XXX do we need a memory barrier of sorts
2277 * wrt to rcu_dereference() of perf_event_ctxp
2279 task
->perf_event_ctxp
[ctxn
] = next_ctx
;
2280 next
->perf_event_ctxp
[ctxn
] = ctx
;
2282 next_ctx
->task
= task
;
2285 perf_event_sync_stat(ctx
, next_ctx
);
2287 raw_spin_unlock(&next_ctx
->lock
);
2288 raw_spin_unlock(&ctx
->lock
);
2293 raw_spin_lock(&ctx
->lock
);
2294 ctx_sched_out(ctx
, cpuctx
, EVENT_ALL
);
2295 cpuctx
->task_ctx
= NULL
;
2296 raw_spin_unlock(&ctx
->lock
);
2300 #define for_each_task_context_nr(ctxn) \
2301 for ((ctxn) = 0; (ctxn) < perf_nr_task_contexts; (ctxn)++)
2304 * Called from scheduler to remove the events of the current task,
2305 * with interrupts disabled.
2307 * We stop each event and update the event value in event->count.
2309 * This does not protect us against NMI, but disable()
2310 * sets the disabled bit in the control field of event _before_
2311 * accessing the event control register. If a NMI hits, then it will
2312 * not restart the event.
2314 void __perf_event_task_sched_out(struct task_struct
*task
,
2315 struct task_struct
*next
)
2319 for_each_task_context_nr(ctxn
)
2320 perf_event_context_sched_out(task
, ctxn
, next
);
2323 * if cgroup events exist on this CPU, then we need
2324 * to check if we have to switch out PMU state.
2325 * cgroup event are system-wide mode only
2327 if (atomic_read(&__get_cpu_var(perf_cgroup_events
)))
2328 perf_cgroup_sched_out(task
, next
);
2331 static void task_ctx_sched_out(struct perf_event_context
*ctx
)
2333 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
2335 if (!cpuctx
->task_ctx
)
2338 if (WARN_ON_ONCE(ctx
!= cpuctx
->task_ctx
))
2341 ctx_sched_out(ctx
, cpuctx
, EVENT_ALL
);
2342 cpuctx
->task_ctx
= NULL
;
2346 * Called with IRQs disabled
2348 static void cpu_ctx_sched_out(struct perf_cpu_context
*cpuctx
,
2349 enum event_type_t event_type
)
2351 ctx_sched_out(&cpuctx
->ctx
, cpuctx
, event_type
);
2355 ctx_pinned_sched_in(struct perf_event_context
*ctx
,
2356 struct perf_cpu_context
*cpuctx
)
2358 struct perf_event
*event
;
2360 list_for_each_entry(event
, &ctx
->pinned_groups
, group_entry
) {
2361 if (event
->state
<= PERF_EVENT_STATE_OFF
)
2363 if (!event_filter_match(event
))
2366 /* may need to reset tstamp_enabled */
2367 if (is_cgroup_event(event
))
2368 perf_cgroup_mark_enabled(event
, ctx
);
2370 if (group_can_go_on(event
, cpuctx
, 1))
2371 group_sched_in(event
, cpuctx
, ctx
);
2374 * If this pinned group hasn't been scheduled,
2375 * put it in error state.
2377 if (event
->state
== PERF_EVENT_STATE_INACTIVE
) {
2378 update_group_times(event
);
2379 event
->state
= PERF_EVENT_STATE_ERROR
;
2385 ctx_flexible_sched_in(struct perf_event_context
*ctx
,
2386 struct perf_cpu_context
*cpuctx
)
2388 struct perf_event
*event
;
2391 list_for_each_entry(event
, &ctx
->flexible_groups
, group_entry
) {
2392 /* Ignore events in OFF or ERROR state */
2393 if (event
->state
<= PERF_EVENT_STATE_OFF
)
2396 * Listen to the 'cpu' scheduling filter constraint
2399 if (!event_filter_match(event
))
2402 /* may need to reset tstamp_enabled */
2403 if (is_cgroup_event(event
))
2404 perf_cgroup_mark_enabled(event
, ctx
);
2406 if (group_can_go_on(event
, cpuctx
, can_add_hw
)) {
2407 if (group_sched_in(event
, cpuctx
, ctx
))
2414 ctx_sched_in(struct perf_event_context
*ctx
,
2415 struct perf_cpu_context
*cpuctx
,
2416 enum event_type_t event_type
,
2417 struct task_struct
*task
)
2420 int is_active
= ctx
->is_active
;
2422 ctx
->is_active
|= event_type
;
2423 if (likely(!ctx
->nr_events
))
2427 ctx
->timestamp
= now
;
2428 perf_cgroup_set_timestamp(task
, ctx
);
2430 * First go through the list and put on any pinned groups
2431 * in order to give them the best chance of going on.
2433 if (!(is_active
& EVENT_PINNED
) && (event_type
& EVENT_PINNED
))
2434 ctx_pinned_sched_in(ctx
, cpuctx
);
2436 /* Then walk through the lower prio flexible groups */
2437 if (!(is_active
& EVENT_FLEXIBLE
) && (event_type
& EVENT_FLEXIBLE
))
2438 ctx_flexible_sched_in(ctx
, cpuctx
);
2441 static void cpu_ctx_sched_in(struct perf_cpu_context
*cpuctx
,
2442 enum event_type_t event_type
,
2443 struct task_struct
*task
)
2445 struct perf_event_context
*ctx
= &cpuctx
->ctx
;
2447 ctx_sched_in(ctx
, cpuctx
, event_type
, task
);
2450 static void perf_event_context_sched_in(struct perf_event_context
*ctx
,
2451 struct task_struct
*task
)
2453 struct perf_cpu_context
*cpuctx
;
2455 cpuctx
= __get_cpu_context(ctx
);
2456 if (cpuctx
->task_ctx
== ctx
)
2459 perf_ctx_lock(cpuctx
, ctx
);
2460 perf_pmu_disable(ctx
->pmu
);
2462 * We want to keep the following priority order:
2463 * cpu pinned (that don't need to move), task pinned,
2464 * cpu flexible, task flexible.
2466 cpu_ctx_sched_out(cpuctx
, EVENT_FLEXIBLE
);
2469 cpuctx
->task_ctx
= ctx
;
2471 perf_event_sched_in(cpuctx
, cpuctx
->task_ctx
, task
);
2473 perf_pmu_enable(ctx
->pmu
);
2474 perf_ctx_unlock(cpuctx
, ctx
);
2477 * Since these rotations are per-cpu, we need to ensure the
2478 * cpu-context we got scheduled on is actually rotating.
2480 perf_pmu_rotate_start(ctx
->pmu
);
2484 * When sampling the branck stack in system-wide, it may be necessary
2485 * to flush the stack on context switch. This happens when the branch
2486 * stack does not tag its entries with the pid of the current task.
2487 * Otherwise it becomes impossible to associate a branch entry with a
2488 * task. This ambiguity is more likely to appear when the branch stack
2489 * supports priv level filtering and the user sets it to monitor only
2490 * at the user level (which could be a useful measurement in system-wide
2491 * mode). In that case, the risk is high of having a branch stack with
2492 * branch from multiple tasks. Flushing may mean dropping the existing
2493 * entries or stashing them somewhere in the PMU specific code layer.
2495 * This function provides the context switch callback to the lower code
2496 * layer. It is invoked ONLY when there is at least one system-wide context
2497 * with at least one active event using taken branch sampling.
2499 static void perf_branch_stack_sched_in(struct task_struct
*prev
,
2500 struct task_struct
*task
)
2502 struct perf_cpu_context
*cpuctx
;
2504 unsigned long flags
;
2506 /* no need to flush branch stack if not changing task */
2510 local_irq_save(flags
);
2514 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
2515 cpuctx
= this_cpu_ptr(pmu
->pmu_cpu_context
);
2518 * check if the context has at least one
2519 * event using PERF_SAMPLE_BRANCH_STACK
2521 if (cpuctx
->ctx
.nr_branch_stack
> 0
2522 && pmu
->flush_branch_stack
) {
2524 pmu
= cpuctx
->ctx
.pmu
;
2526 perf_ctx_lock(cpuctx
, cpuctx
->task_ctx
);
2528 perf_pmu_disable(pmu
);
2530 pmu
->flush_branch_stack();
2532 perf_pmu_enable(pmu
);
2534 perf_ctx_unlock(cpuctx
, cpuctx
->task_ctx
);
2540 local_irq_restore(flags
);
2544 * Called from scheduler to add the events of the current task
2545 * with interrupts disabled.
2547 * We restore the event value and then enable it.
2549 * This does not protect us against NMI, but enable()
2550 * sets the enabled bit in the control field of event _before_
2551 * accessing the event control register. If a NMI hits, then it will
2552 * keep the event running.
2554 void __perf_event_task_sched_in(struct task_struct
*prev
,
2555 struct task_struct
*task
)
2557 struct perf_event_context
*ctx
;
2560 for_each_task_context_nr(ctxn
) {
2561 ctx
= task
->perf_event_ctxp
[ctxn
];
2565 perf_event_context_sched_in(ctx
, task
);
2568 * if cgroup events exist on this CPU, then we need
2569 * to check if we have to switch in PMU state.
2570 * cgroup event are system-wide mode only
2572 if (atomic_read(&__get_cpu_var(perf_cgroup_events
)))
2573 perf_cgroup_sched_in(prev
, task
);
2575 /* check for system-wide branch_stack events */
2576 if (atomic_read(&__get_cpu_var(perf_branch_stack_events
)))
2577 perf_branch_stack_sched_in(prev
, task
);
2580 static u64
perf_calculate_period(struct perf_event
*event
, u64 nsec
, u64 count
)
2582 u64 frequency
= event
->attr
.sample_freq
;
2583 u64 sec
= NSEC_PER_SEC
;
2584 u64 divisor
, dividend
;
2586 int count_fls
, nsec_fls
, frequency_fls
, sec_fls
;
2588 count_fls
= fls64(count
);
2589 nsec_fls
= fls64(nsec
);
2590 frequency_fls
= fls64(frequency
);
2594 * We got @count in @nsec, with a target of sample_freq HZ
2595 * the target period becomes:
2598 * period = -------------------
2599 * @nsec * sample_freq
2604 * Reduce accuracy by one bit such that @a and @b converge
2605 * to a similar magnitude.
2607 #define REDUCE_FLS(a, b) \
2609 if (a##_fls > b##_fls) { \
2619 * Reduce accuracy until either term fits in a u64, then proceed with
2620 * the other, so that finally we can do a u64/u64 division.
2622 while (count_fls
+ sec_fls
> 64 && nsec_fls
+ frequency_fls
> 64) {
2623 REDUCE_FLS(nsec
, frequency
);
2624 REDUCE_FLS(sec
, count
);
2627 if (count_fls
+ sec_fls
> 64) {
2628 divisor
= nsec
* frequency
;
2630 while (count_fls
+ sec_fls
> 64) {
2631 REDUCE_FLS(count
, sec
);
2635 dividend
= count
* sec
;
2637 dividend
= count
* sec
;
2639 while (nsec_fls
+ frequency_fls
> 64) {
2640 REDUCE_FLS(nsec
, frequency
);
2644 divisor
= nsec
* frequency
;
2650 return div64_u64(dividend
, divisor
);
2653 static DEFINE_PER_CPU(int, perf_throttled_count
);
2654 static DEFINE_PER_CPU(u64
, perf_throttled_seq
);
2656 static void perf_adjust_period(struct perf_event
*event
, u64 nsec
, u64 count
, bool disable
)
2658 struct hw_perf_event
*hwc
= &event
->hw
;
2659 s64 period
, sample_period
;
2662 period
= perf_calculate_period(event
, nsec
, count
);
2664 delta
= (s64
)(period
- hwc
->sample_period
);
2665 delta
= (delta
+ 7) / 8; /* low pass filter */
2667 sample_period
= hwc
->sample_period
+ delta
;
2672 hwc
->sample_period
= sample_period
;
2674 if (local64_read(&hwc
->period_left
) > 8*sample_period
) {
2676 event
->pmu
->stop(event
, PERF_EF_UPDATE
);
2678 local64_set(&hwc
->period_left
, 0);
2681 event
->pmu
->start(event
, PERF_EF_RELOAD
);
2686 * combine freq adjustment with unthrottling to avoid two passes over the
2687 * events. At the same time, make sure, having freq events does not change
2688 * the rate of unthrottling as that would introduce bias.
2690 static void perf_adjust_freq_unthr_context(struct perf_event_context
*ctx
,
2693 struct perf_event
*event
;
2694 struct hw_perf_event
*hwc
;
2695 u64 now
, period
= TICK_NSEC
;
2699 * only need to iterate over all events iff:
2700 * - context have events in frequency mode (needs freq adjust)
2701 * - there are events to unthrottle on this cpu
2703 if (!(ctx
->nr_freq
|| needs_unthr
))
2706 raw_spin_lock(&ctx
->lock
);
2707 perf_pmu_disable(ctx
->pmu
);
2709 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
2710 if (event
->state
!= PERF_EVENT_STATE_ACTIVE
)
2713 if (!event_filter_match(event
))
2718 if (hwc
->interrupts
== MAX_INTERRUPTS
) {
2719 hwc
->interrupts
= 0;
2720 perf_log_throttle(event
, 1);
2721 event
->pmu
->start(event
, 0);
2724 if (!event
->attr
.freq
|| !event
->attr
.sample_freq
)
2728 * stop the event and update event->count
2730 event
->pmu
->stop(event
, PERF_EF_UPDATE
);
2732 now
= local64_read(&event
->count
);
2733 delta
= now
- hwc
->freq_count_stamp
;
2734 hwc
->freq_count_stamp
= now
;
2738 * reload only if value has changed
2739 * we have stopped the event so tell that
2740 * to perf_adjust_period() to avoid stopping it
2744 perf_adjust_period(event
, period
, delta
, false);
2746 event
->pmu
->start(event
, delta
> 0 ? PERF_EF_RELOAD
: 0);
2749 perf_pmu_enable(ctx
->pmu
);
2750 raw_spin_unlock(&ctx
->lock
);
2754 * Round-robin a context's events:
2756 static void rotate_ctx(struct perf_event_context
*ctx
)
2759 * Rotate the first entry last of non-pinned groups. Rotation might be
2760 * disabled by the inheritance code.
2762 if (!ctx
->rotate_disable
)
2763 list_rotate_left(&ctx
->flexible_groups
);
2767 * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
2768 * because they're strictly cpu affine and rotate_start is called with IRQs
2769 * disabled, while rotate_context is called from IRQ context.
2771 static int perf_rotate_context(struct perf_cpu_context
*cpuctx
)
2773 struct perf_event_context
*ctx
= NULL
;
2774 int rotate
= 0, remove
= 1;
2776 if (cpuctx
->ctx
.nr_events
) {
2778 if (cpuctx
->ctx
.nr_events
!= cpuctx
->ctx
.nr_active
)
2782 ctx
= cpuctx
->task_ctx
;
2783 if (ctx
&& ctx
->nr_events
) {
2785 if (ctx
->nr_events
!= ctx
->nr_active
)
2792 perf_ctx_lock(cpuctx
, cpuctx
->task_ctx
);
2793 perf_pmu_disable(cpuctx
->ctx
.pmu
);
2795 cpu_ctx_sched_out(cpuctx
, EVENT_FLEXIBLE
);
2797 ctx_sched_out(ctx
, cpuctx
, EVENT_FLEXIBLE
);
2799 rotate_ctx(&cpuctx
->ctx
);
2803 perf_event_sched_in(cpuctx
, ctx
, current
);
2805 perf_pmu_enable(cpuctx
->ctx
.pmu
);
2806 perf_ctx_unlock(cpuctx
, cpuctx
->task_ctx
);
2809 list_del_init(&cpuctx
->rotation_list
);
2814 #ifdef CONFIG_NO_HZ_FULL
2815 bool perf_event_can_stop_tick(void)
2817 if (atomic_read(&nr_freq_events
) ||
2818 __this_cpu_read(perf_throttled_count
))
2825 void perf_event_task_tick(void)
2827 struct list_head
*head
= &__get_cpu_var(rotation_list
);
2828 struct perf_cpu_context
*cpuctx
, *tmp
;
2829 struct perf_event_context
*ctx
;
2832 WARN_ON(!irqs_disabled());
2834 __this_cpu_inc(perf_throttled_seq
);
2835 throttled
= __this_cpu_xchg(perf_throttled_count
, 0);
2837 list_for_each_entry_safe(cpuctx
, tmp
, head
, rotation_list
) {
2839 perf_adjust_freq_unthr_context(ctx
, throttled
);
2841 ctx
= cpuctx
->task_ctx
;
2843 perf_adjust_freq_unthr_context(ctx
, throttled
);
2847 static int event_enable_on_exec(struct perf_event
*event
,
2848 struct perf_event_context
*ctx
)
2850 if (!event
->attr
.enable_on_exec
)
2853 event
->attr
.enable_on_exec
= 0;
2854 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
)
2857 __perf_event_mark_enabled(event
);
2863 * Enable all of a task's events that have been marked enable-on-exec.
2864 * This expects task == current.
2866 static void perf_event_enable_on_exec(struct perf_event_context
*ctx
)
2868 struct perf_event
*event
;
2869 unsigned long flags
;
2873 local_irq_save(flags
);
2874 if (!ctx
|| !ctx
->nr_events
)
2878 * We must ctxsw out cgroup events to avoid conflict
2879 * when invoking perf_task_event_sched_in() later on
2880 * in this function. Otherwise we end up trying to
2881 * ctxswin cgroup events which are already scheduled
2884 perf_cgroup_sched_out(current
, NULL
);
2886 raw_spin_lock(&ctx
->lock
);
2887 task_ctx_sched_out(ctx
);
2889 list_for_each_entry(event
, &ctx
->event_list
, event_entry
) {
2890 ret
= event_enable_on_exec(event
, ctx
);
2896 * Unclone this context if we enabled any event.
2901 raw_spin_unlock(&ctx
->lock
);
2904 * Also calls ctxswin for cgroup events, if any:
2906 perf_event_context_sched_in(ctx
, ctx
->task
);
2908 local_irq_restore(flags
);
2912 * Cross CPU call to read the hardware event
2914 static void __perf_event_read(void *info
)
2916 struct perf_event
*event
= info
;
2917 struct perf_event_context
*ctx
= event
->ctx
;
2918 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
2921 * If this is a task context, we need to check whether it is
2922 * the current task context of this cpu. If not it has been
2923 * scheduled out before the smp call arrived. In that case
2924 * event->count would have been updated to a recent sample
2925 * when the event was scheduled out.
2927 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
)
2930 raw_spin_lock(&ctx
->lock
);
2931 if (ctx
->is_active
) {
2932 update_context_time(ctx
);
2933 update_cgrp_time_from_event(event
);
2935 update_event_times(event
);
2936 if (event
->state
== PERF_EVENT_STATE_ACTIVE
)
2937 event
->pmu
->read(event
);
2938 raw_spin_unlock(&ctx
->lock
);
2941 static inline u64
perf_event_count(struct perf_event
*event
)
2943 return local64_read(&event
->count
) + atomic64_read(&event
->child_count
);
2946 static u64
perf_event_read(struct perf_event
*event
)
2949 * If event is enabled and currently active on a CPU, update the
2950 * value in the event structure:
2952 if (event
->state
== PERF_EVENT_STATE_ACTIVE
) {
2953 smp_call_function_single(event
->oncpu
,
2954 __perf_event_read
, event
, 1);
2955 } else if (event
->state
== PERF_EVENT_STATE_INACTIVE
) {
2956 struct perf_event_context
*ctx
= event
->ctx
;
2957 unsigned long flags
;
2959 raw_spin_lock_irqsave(&ctx
->lock
, flags
);
2961 * may read while context is not active
2962 * (e.g., thread is blocked), in that case
2963 * we cannot update context time
2965 if (ctx
->is_active
) {
2966 update_context_time(ctx
);
2967 update_cgrp_time_from_event(event
);
2969 update_event_times(event
);
2970 raw_spin_unlock_irqrestore(&ctx
->lock
, flags
);
2973 return perf_event_count(event
);
2977 * Initialize the perf_event context in a task_struct:
2979 static void __perf_event_init_context(struct perf_event_context
*ctx
)
2981 raw_spin_lock_init(&ctx
->lock
);
2982 mutex_init(&ctx
->mutex
);
2983 INIT_LIST_HEAD(&ctx
->pinned_groups
);
2984 INIT_LIST_HEAD(&ctx
->flexible_groups
);
2985 INIT_LIST_HEAD(&ctx
->event_list
);
2986 atomic_set(&ctx
->refcount
, 1);
2989 static struct perf_event_context
*
2990 alloc_perf_context(struct pmu
*pmu
, struct task_struct
*task
)
2992 struct perf_event_context
*ctx
;
2994 ctx
= kzalloc(sizeof(struct perf_event_context
), GFP_KERNEL
);
2998 __perf_event_init_context(ctx
);
3001 get_task_struct(task
);
3008 static struct task_struct
*
3009 find_lively_task_by_vpid(pid_t vpid
)
3011 struct task_struct
*task
;
3018 task
= find_task_by_vpid(vpid
);
3020 get_task_struct(task
);
3024 return ERR_PTR(-ESRCH
);
3026 /* Reuse ptrace permission checks for now. */
3028 if (!ptrace_may_access(task
, PTRACE_MODE_READ
))
3033 put_task_struct(task
);
3034 return ERR_PTR(err
);
3039 * Returns a matching context with refcount and pincount.
3041 static struct perf_event_context
*
3042 find_get_context(struct pmu
*pmu
, struct task_struct
*task
, int cpu
)
3044 struct perf_event_context
*ctx
;
3045 struct perf_cpu_context
*cpuctx
;
3046 unsigned long flags
;
3050 /* Must be root to operate on a CPU event: */
3051 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN
))
3052 return ERR_PTR(-EACCES
);
3055 * We could be clever and allow to attach a event to an
3056 * offline CPU and activate it when the CPU comes up, but
3059 if (!cpu_online(cpu
))
3060 return ERR_PTR(-ENODEV
);
3062 cpuctx
= per_cpu_ptr(pmu
->pmu_cpu_context
, cpu
);
3071 ctxn
= pmu
->task_ctx_nr
;
3076 ctx
= perf_lock_task_context(task
, ctxn
, &flags
);
3080 raw_spin_unlock_irqrestore(&ctx
->lock
, flags
);
3082 ctx
= alloc_perf_context(pmu
, task
);
3088 mutex_lock(&task
->perf_event_mutex
);
3090 * If it has already passed perf_event_exit_task().
3091 * we must see PF_EXITING, it takes this mutex too.
3093 if (task
->flags
& PF_EXITING
)
3095 else if (task
->perf_event_ctxp
[ctxn
])
3100 rcu_assign_pointer(task
->perf_event_ctxp
[ctxn
], ctx
);
3102 mutex_unlock(&task
->perf_event_mutex
);
3104 if (unlikely(err
)) {
3116 return ERR_PTR(err
);
3119 static void perf_event_free_filter(struct perf_event
*event
);
3121 static void free_event_rcu(struct rcu_head
*head
)
3123 struct perf_event
*event
;
3125 event
= container_of(head
, struct perf_event
, rcu_head
);
3127 put_pid_ns(event
->ns
);
3128 perf_event_free_filter(event
);
3132 static void ring_buffer_put(struct ring_buffer
*rb
);
3133 static void ring_buffer_detach(struct perf_event
*event
, struct ring_buffer
*rb
);
3135 static void unaccount_event_cpu(struct perf_event
*event
, int cpu
)
3140 if (has_branch_stack(event
)) {
3141 if (!(event
->attach_state
& PERF_ATTACH_TASK
))
3142 atomic_dec(&per_cpu(perf_branch_stack_events
, cpu
));
3144 if (is_cgroup_event(event
))
3145 atomic_dec(&per_cpu(perf_cgroup_events
, cpu
));
3148 static void unaccount_event(struct perf_event
*event
)
3153 if (event
->attach_state
& PERF_ATTACH_TASK
)
3154 static_key_slow_dec_deferred(&perf_sched_events
);
3155 if (event
->attr
.mmap
|| event
->attr
.mmap_data
)
3156 atomic_dec(&nr_mmap_events
);
3157 if (event
->attr
.comm
)
3158 atomic_dec(&nr_comm_events
);
3159 if (event
->attr
.task
)
3160 atomic_dec(&nr_task_events
);
3161 if (event
->attr
.freq
)
3162 atomic_dec(&nr_freq_events
);
3163 if (is_cgroup_event(event
))
3164 static_key_slow_dec_deferred(&perf_sched_events
);
3165 if (has_branch_stack(event
))
3166 static_key_slow_dec_deferred(&perf_sched_events
);
3168 unaccount_event_cpu(event
, event
->cpu
);
3171 static void __free_event(struct perf_event
*event
)
3173 if (!event
->parent
) {
3174 if (event
->attr
.sample_type
& PERF_SAMPLE_CALLCHAIN
)
3175 put_callchain_buffers();
3179 event
->destroy(event
);
3182 put_ctx(event
->ctx
);
3184 call_rcu(&event
->rcu_head
, free_event_rcu
);
3186 static void free_event(struct perf_event
*event
)
3188 irq_work_sync(&event
->pending
);
3190 unaccount_event(event
);
3193 struct ring_buffer
*rb
;
3196 * Can happen when we close an event with re-directed output.
3198 * Since we have a 0 refcount, perf_mmap_close() will skip
3199 * over us; possibly making our ring_buffer_put() the last.
3201 mutex_lock(&event
->mmap_mutex
);
3204 rcu_assign_pointer(event
->rb
, NULL
);
3205 ring_buffer_detach(event
, rb
);
3206 ring_buffer_put(rb
); /* could be last */
3208 mutex_unlock(&event
->mmap_mutex
);
3211 if (is_cgroup_event(event
))
3212 perf_detach_cgroup(event
);
3215 __free_event(event
);
3218 int perf_event_release_kernel(struct perf_event
*event
)
3220 struct perf_event_context
*ctx
= event
->ctx
;
3222 WARN_ON_ONCE(ctx
->parent_ctx
);
3224 * There are two ways this annotation is useful:
3226 * 1) there is a lock recursion from perf_event_exit_task
3227 * see the comment there.
3229 * 2) there is a lock-inversion with mmap_sem through
3230 * perf_event_read_group(), which takes faults while
3231 * holding ctx->mutex, however this is called after
3232 * the last filedesc died, so there is no possibility
3233 * to trigger the AB-BA case.
3235 mutex_lock_nested(&ctx
->mutex
, SINGLE_DEPTH_NESTING
);
3236 raw_spin_lock_irq(&ctx
->lock
);
3237 perf_group_detach(event
);
3238 raw_spin_unlock_irq(&ctx
->lock
);
3239 perf_remove_from_context(event
);
3240 mutex_unlock(&ctx
->mutex
);
3246 EXPORT_SYMBOL_GPL(perf_event_release_kernel
);
3249 * Called when the last reference to the file is gone.
3251 static void put_event(struct perf_event
*event
)
3253 struct task_struct
*owner
;
3255 if (!atomic_long_dec_and_test(&event
->refcount
))
3259 owner
= ACCESS_ONCE(event
->owner
);
3261 * Matches the smp_wmb() in perf_event_exit_task(). If we observe
3262 * !owner it means the list deletion is complete and we can indeed
3263 * free this event, otherwise we need to serialize on
3264 * owner->perf_event_mutex.
3266 smp_read_barrier_depends();
3269 * Since delayed_put_task_struct() also drops the last
3270 * task reference we can safely take a new reference
3271 * while holding the rcu_read_lock().
3273 get_task_struct(owner
);
3278 mutex_lock(&owner
->perf_event_mutex
);
3280 * We have to re-check the event->owner field, if it is cleared
3281 * we raced with perf_event_exit_task(), acquiring the mutex
3282 * ensured they're done, and we can proceed with freeing the
3286 list_del_init(&event
->owner_entry
);
3287 mutex_unlock(&owner
->perf_event_mutex
);
3288 put_task_struct(owner
);
3291 perf_event_release_kernel(event
);
3294 static int perf_release(struct inode
*inode
, struct file
*file
)
3296 put_event(file
->private_data
);
3300 u64
perf_event_read_value(struct perf_event
*event
, u64
*enabled
, u64
*running
)
3302 struct perf_event
*child
;
3308 mutex_lock(&event
->child_mutex
);
3309 total
+= perf_event_read(event
);
3310 *enabled
+= event
->total_time_enabled
+
3311 atomic64_read(&event
->child_total_time_enabled
);
3312 *running
+= event
->total_time_running
+
3313 atomic64_read(&event
->child_total_time_running
);
3315 list_for_each_entry(child
, &event
->child_list
, child_list
) {
3316 total
+= perf_event_read(child
);
3317 *enabled
+= child
->total_time_enabled
;
3318 *running
+= child
->total_time_running
;
3320 mutex_unlock(&event
->child_mutex
);
3324 EXPORT_SYMBOL_GPL(perf_event_read_value
);
3326 static int perf_event_read_group(struct perf_event
*event
,
3327 u64 read_format
, char __user
*buf
)
3329 struct perf_event
*leader
= event
->group_leader
, *sub
;
3330 int n
= 0, size
= 0, ret
= -EFAULT
;
3331 struct perf_event_context
*ctx
= leader
->ctx
;
3333 u64 count
, enabled
, running
;
3335 mutex_lock(&ctx
->mutex
);
3336 count
= perf_event_read_value(leader
, &enabled
, &running
);
3338 values
[n
++] = 1 + leader
->nr_siblings
;
3339 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
3340 values
[n
++] = enabled
;
3341 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
3342 values
[n
++] = running
;
3343 values
[n
++] = count
;
3344 if (read_format
& PERF_FORMAT_ID
)
3345 values
[n
++] = primary_event_id(leader
);
3347 size
= n
* sizeof(u64
);
3349 if (copy_to_user(buf
, values
, size
))
3354 list_for_each_entry(sub
, &leader
->sibling_list
, group_entry
) {
3357 values
[n
++] = perf_event_read_value(sub
, &enabled
, &running
);
3358 if (read_format
& PERF_FORMAT_ID
)
3359 values
[n
++] = primary_event_id(sub
);
3361 size
= n
* sizeof(u64
);
3363 if (copy_to_user(buf
+ ret
, values
, size
)) {
3371 mutex_unlock(&ctx
->mutex
);
3376 static int perf_event_read_one(struct perf_event
*event
,
3377 u64 read_format
, char __user
*buf
)
3379 u64 enabled
, running
;
3383 values
[n
++] = perf_event_read_value(event
, &enabled
, &running
);
3384 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
3385 values
[n
++] = enabled
;
3386 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
3387 values
[n
++] = running
;
3388 if (read_format
& PERF_FORMAT_ID
)
3389 values
[n
++] = primary_event_id(event
);
3391 if (copy_to_user(buf
, values
, n
* sizeof(u64
)))
3394 return n
* sizeof(u64
);
3398 * Read the performance event - simple non blocking version for now
3401 perf_read_hw(struct perf_event
*event
, char __user
*buf
, size_t count
)
3403 u64 read_format
= event
->attr
.read_format
;
3407 * Return end-of-file for a read on a event that is in
3408 * error state (i.e. because it was pinned but it couldn't be
3409 * scheduled on to the CPU at some point).
3411 if (event
->state
== PERF_EVENT_STATE_ERROR
)
3414 if (count
< event
->read_size
)
3417 WARN_ON_ONCE(event
->ctx
->parent_ctx
);
3418 if (read_format
& PERF_FORMAT_GROUP
)
3419 ret
= perf_event_read_group(event
, read_format
, buf
);
3421 ret
= perf_event_read_one(event
, read_format
, buf
);
3427 perf_read(struct file
*file
, char __user
*buf
, size_t count
, loff_t
*ppos
)
3429 struct perf_event
*event
= file
->private_data
;
3431 return perf_read_hw(event
, buf
, count
);
3434 static unsigned int perf_poll(struct file
*file
, poll_table
*wait
)
3436 struct perf_event
*event
= file
->private_data
;
3437 struct ring_buffer
*rb
;
3438 unsigned int events
= POLL_HUP
;
3441 * Pin the event->rb by taking event->mmap_mutex; otherwise
3442 * perf_event_set_output() can swizzle our rb and make us miss wakeups.
3444 mutex_lock(&event
->mmap_mutex
);
3447 events
= atomic_xchg(&rb
->poll
, 0);
3448 mutex_unlock(&event
->mmap_mutex
);
3450 poll_wait(file
, &event
->waitq
, wait
);
3455 static void perf_event_reset(struct perf_event
*event
)
3457 (void)perf_event_read(event
);
3458 local64_set(&event
->count
, 0);
3459 perf_event_update_userpage(event
);
3463 * Holding the top-level event's child_mutex means that any
3464 * descendant process that has inherited this event will block
3465 * in sync_child_event if it goes to exit, thus satisfying the
3466 * task existence requirements of perf_event_enable/disable.
3468 static void perf_event_for_each_child(struct perf_event
*event
,
3469 void (*func
)(struct perf_event
*))
3471 struct perf_event
*child
;
3473 WARN_ON_ONCE(event
->ctx
->parent_ctx
);
3474 mutex_lock(&event
->child_mutex
);
3476 list_for_each_entry(child
, &event
->child_list
, child_list
)
3478 mutex_unlock(&event
->child_mutex
);
3481 static void perf_event_for_each(struct perf_event
*event
,
3482 void (*func
)(struct perf_event
*))
3484 struct perf_event_context
*ctx
= event
->ctx
;
3485 struct perf_event
*sibling
;
3487 WARN_ON_ONCE(ctx
->parent_ctx
);
3488 mutex_lock(&ctx
->mutex
);
3489 event
= event
->group_leader
;
3491 perf_event_for_each_child(event
, func
);
3492 list_for_each_entry(sibling
, &event
->sibling_list
, group_entry
)
3493 perf_event_for_each_child(sibling
, func
);
3494 mutex_unlock(&ctx
->mutex
);
3497 static int perf_event_period(struct perf_event
*event
, u64 __user
*arg
)
3499 struct perf_event_context
*ctx
= event
->ctx
;
3503 if (!is_sampling_event(event
))
3506 if (copy_from_user(&value
, arg
, sizeof(value
)))
3512 raw_spin_lock_irq(&ctx
->lock
);
3513 if (event
->attr
.freq
) {
3514 if (value
> sysctl_perf_event_sample_rate
) {
3519 event
->attr
.sample_freq
= value
;
3521 event
->attr
.sample_period
= value
;
3522 event
->hw
.sample_period
= value
;
3525 raw_spin_unlock_irq(&ctx
->lock
);
3530 static const struct file_operations perf_fops
;
3532 static inline int perf_fget_light(int fd
, struct fd
*p
)
3534 struct fd f
= fdget(fd
);
3538 if (f
.file
->f_op
!= &perf_fops
) {
3546 static int perf_event_set_output(struct perf_event
*event
,
3547 struct perf_event
*output_event
);
3548 static int perf_event_set_filter(struct perf_event
*event
, void __user
*arg
);
3550 static long perf_ioctl(struct file
*file
, unsigned int cmd
, unsigned long arg
)
3552 struct perf_event
*event
= file
->private_data
;
3553 void (*func
)(struct perf_event
*);
3557 case PERF_EVENT_IOC_ENABLE
:
3558 func
= perf_event_enable
;
3560 case PERF_EVENT_IOC_DISABLE
:
3561 func
= perf_event_disable
;
3563 case PERF_EVENT_IOC_RESET
:
3564 func
= perf_event_reset
;
3567 case PERF_EVENT_IOC_REFRESH
:
3568 return perf_event_refresh(event
, arg
);
3570 case PERF_EVENT_IOC_PERIOD
:
3571 return perf_event_period(event
, (u64 __user
*)arg
);
3573 case PERF_EVENT_IOC_ID
:
3575 u64 id
= primary_event_id(event
);
3577 if (copy_to_user((void __user
*)arg
, &id
, sizeof(id
)))
3582 case PERF_EVENT_IOC_SET_OUTPUT
:
3586 struct perf_event
*output_event
;
3588 ret
= perf_fget_light(arg
, &output
);
3591 output_event
= output
.file
->private_data
;
3592 ret
= perf_event_set_output(event
, output_event
);
3595 ret
= perf_event_set_output(event
, NULL
);
3600 case PERF_EVENT_IOC_SET_FILTER
:
3601 return perf_event_set_filter(event
, (void __user
*)arg
);
3607 if (flags
& PERF_IOC_FLAG_GROUP
)
3608 perf_event_for_each(event
, func
);
3610 perf_event_for_each_child(event
, func
);
3615 int perf_event_task_enable(void)
3617 struct perf_event
*event
;
3619 mutex_lock(¤t
->perf_event_mutex
);
3620 list_for_each_entry(event
, ¤t
->perf_event_list
, owner_entry
)
3621 perf_event_for_each_child(event
, perf_event_enable
);
3622 mutex_unlock(¤t
->perf_event_mutex
);
3627 int perf_event_task_disable(void)
3629 struct perf_event
*event
;
3631 mutex_lock(¤t
->perf_event_mutex
);
3632 list_for_each_entry(event
, ¤t
->perf_event_list
, owner_entry
)
3633 perf_event_for_each_child(event
, perf_event_disable
);
3634 mutex_unlock(¤t
->perf_event_mutex
);
3639 static int perf_event_index(struct perf_event
*event
)
3641 if (event
->hw
.state
& PERF_HES_STOPPED
)
3644 if (event
->state
!= PERF_EVENT_STATE_ACTIVE
)
3647 return event
->pmu
->event_idx(event
);
3650 static void calc_timer_values(struct perf_event
*event
,
3657 *now
= perf_clock();
3658 ctx_time
= event
->shadow_ctx_time
+ *now
;
3659 *enabled
= ctx_time
- event
->tstamp_enabled
;
3660 *running
= ctx_time
- event
->tstamp_running
;
3663 static void perf_event_init_userpage(struct perf_event
*event
)
3665 struct perf_event_mmap_page
*userpg
;
3666 struct ring_buffer
*rb
;
3669 rb
= rcu_dereference(event
->rb
);
3673 userpg
= rb
->user_page
;
3675 /* Allow new userspace to detect that bit 0 is deprecated */
3676 userpg
->cap_bit0_is_deprecated
= 1;
3677 userpg
->size
= offsetof(struct perf_event_mmap_page
, __reserved
);
3683 void __weak
arch_perf_update_userpage(struct perf_event_mmap_page
*userpg
, u64 now
)
3688 * Callers need to ensure there can be no nesting of this function, otherwise
3689 * the seqlock logic goes bad. We can not serialize this because the arch
3690 * code calls this from NMI context.
3692 void perf_event_update_userpage(struct perf_event
*event
)
3694 struct perf_event_mmap_page
*userpg
;
3695 struct ring_buffer
*rb
;
3696 u64 enabled
, running
, now
;
3699 rb
= rcu_dereference(event
->rb
);
3704 * compute total_time_enabled, total_time_running
3705 * based on snapshot values taken when the event
3706 * was last scheduled in.
3708 * we cannot simply called update_context_time()
3709 * because of locking issue as we can be called in
3712 calc_timer_values(event
, &now
, &enabled
, &running
);
3714 userpg
= rb
->user_page
;
3716 * Disable preemption so as to not let the corresponding user-space
3717 * spin too long if we get preempted.
3722 userpg
->index
= perf_event_index(event
);
3723 userpg
->offset
= perf_event_count(event
);
3725 userpg
->offset
-= local64_read(&event
->hw
.prev_count
);
3727 userpg
->time_enabled
= enabled
+
3728 atomic64_read(&event
->child_total_time_enabled
);
3730 userpg
->time_running
= running
+
3731 atomic64_read(&event
->child_total_time_running
);
3733 arch_perf_update_userpage(userpg
, now
);
3742 static int perf_mmap_fault(struct vm_area_struct
*vma
, struct vm_fault
*vmf
)
3744 struct perf_event
*event
= vma
->vm_file
->private_data
;
3745 struct ring_buffer
*rb
;
3746 int ret
= VM_FAULT_SIGBUS
;
3748 if (vmf
->flags
& FAULT_FLAG_MKWRITE
) {
3749 if (vmf
->pgoff
== 0)
3755 rb
= rcu_dereference(event
->rb
);
3759 if (vmf
->pgoff
&& (vmf
->flags
& FAULT_FLAG_WRITE
))
3762 vmf
->page
= perf_mmap_to_page(rb
, vmf
->pgoff
);
3766 get_page(vmf
->page
);
3767 vmf
->page
->mapping
= vma
->vm_file
->f_mapping
;
3768 vmf
->page
->index
= vmf
->pgoff
;
3777 static void ring_buffer_attach(struct perf_event
*event
,
3778 struct ring_buffer
*rb
)
3780 unsigned long flags
;
3782 if (!list_empty(&event
->rb_entry
))
3785 spin_lock_irqsave(&rb
->event_lock
, flags
);
3786 if (list_empty(&event
->rb_entry
))
3787 list_add(&event
->rb_entry
, &rb
->event_list
);
3788 spin_unlock_irqrestore(&rb
->event_lock
, flags
);
3791 static void ring_buffer_detach(struct perf_event
*event
, struct ring_buffer
*rb
)
3793 unsigned long flags
;
3795 if (list_empty(&event
->rb_entry
))
3798 spin_lock_irqsave(&rb
->event_lock
, flags
);
3799 list_del_init(&event
->rb_entry
);
3800 wake_up_all(&event
->waitq
);
3801 spin_unlock_irqrestore(&rb
->event_lock
, flags
);
3804 static void ring_buffer_wakeup(struct perf_event
*event
)
3806 struct ring_buffer
*rb
;
3809 rb
= rcu_dereference(event
->rb
);
3811 list_for_each_entry_rcu(event
, &rb
->event_list
, rb_entry
)
3812 wake_up_all(&event
->waitq
);
3817 static void rb_free_rcu(struct rcu_head
*rcu_head
)
3819 struct ring_buffer
*rb
;
3821 rb
= container_of(rcu_head
, struct ring_buffer
, rcu_head
);
3825 static struct ring_buffer
*ring_buffer_get(struct perf_event
*event
)
3827 struct ring_buffer
*rb
;
3830 rb
= rcu_dereference(event
->rb
);
3832 if (!atomic_inc_not_zero(&rb
->refcount
))
3840 static void ring_buffer_put(struct ring_buffer
*rb
)
3842 if (!atomic_dec_and_test(&rb
->refcount
))
3845 WARN_ON_ONCE(!list_empty(&rb
->event_list
));
3847 call_rcu(&rb
->rcu_head
, rb_free_rcu
);
3850 static void perf_mmap_open(struct vm_area_struct
*vma
)
3852 struct perf_event
*event
= vma
->vm_file
->private_data
;
3854 atomic_inc(&event
->mmap_count
);
3855 atomic_inc(&event
->rb
->mmap_count
);
3859 * A buffer can be mmap()ed multiple times; either directly through the same
3860 * event, or through other events by use of perf_event_set_output().
3862 * In order to undo the VM accounting done by perf_mmap() we need to destroy
3863 * the buffer here, where we still have a VM context. This means we need
3864 * to detach all events redirecting to us.
3866 static void perf_mmap_close(struct vm_area_struct
*vma
)
3868 struct perf_event
*event
= vma
->vm_file
->private_data
;
3870 struct ring_buffer
*rb
= event
->rb
;
3871 struct user_struct
*mmap_user
= rb
->mmap_user
;
3872 int mmap_locked
= rb
->mmap_locked
;
3873 unsigned long size
= perf_data_size(rb
);
3875 atomic_dec(&rb
->mmap_count
);
3877 if (!atomic_dec_and_mutex_lock(&event
->mmap_count
, &event
->mmap_mutex
))
3880 /* Detach current event from the buffer. */
3881 rcu_assign_pointer(event
->rb
, NULL
);
3882 ring_buffer_detach(event
, rb
);
3883 mutex_unlock(&event
->mmap_mutex
);
3885 /* If there's still other mmap()s of this buffer, we're done. */
3886 if (atomic_read(&rb
->mmap_count
)) {
3887 ring_buffer_put(rb
); /* can't be last */
3892 * No other mmap()s, detach from all other events that might redirect
3893 * into the now unreachable buffer. Somewhat complicated by the
3894 * fact that rb::event_lock otherwise nests inside mmap_mutex.
3898 list_for_each_entry_rcu(event
, &rb
->event_list
, rb_entry
) {
3899 if (!atomic_long_inc_not_zero(&event
->refcount
)) {
3901 * This event is en-route to free_event() which will
3902 * detach it and remove it from the list.
3908 mutex_lock(&event
->mmap_mutex
);
3910 * Check we didn't race with perf_event_set_output() which can
3911 * swizzle the rb from under us while we were waiting to
3912 * acquire mmap_mutex.
3914 * If we find a different rb; ignore this event, a next
3915 * iteration will no longer find it on the list. We have to
3916 * still restart the iteration to make sure we're not now
3917 * iterating the wrong list.
3919 if (event
->rb
== rb
) {
3920 rcu_assign_pointer(event
->rb
, NULL
);
3921 ring_buffer_detach(event
, rb
);
3922 ring_buffer_put(rb
); /* can't be last, we still have one */
3924 mutex_unlock(&event
->mmap_mutex
);
3928 * Restart the iteration; either we're on the wrong list or
3929 * destroyed its integrity by doing a deletion.
3936 * It could be there's still a few 0-ref events on the list; they'll
3937 * get cleaned up by free_event() -- they'll also still have their
3938 * ref on the rb and will free it whenever they are done with it.
3940 * Aside from that, this buffer is 'fully' detached and unmapped,
3941 * undo the VM accounting.
3944 atomic_long_sub((size
>> PAGE_SHIFT
) + 1, &mmap_user
->locked_vm
);
3945 vma
->vm_mm
->pinned_vm
-= mmap_locked
;
3946 free_uid(mmap_user
);
3948 ring_buffer_put(rb
); /* could be last */
3951 static const struct vm_operations_struct perf_mmap_vmops
= {
3952 .open
= perf_mmap_open
,
3953 .close
= perf_mmap_close
,
3954 .fault
= perf_mmap_fault
,
3955 .page_mkwrite
= perf_mmap_fault
,
3958 static int perf_mmap(struct file
*file
, struct vm_area_struct
*vma
)
3960 struct perf_event
*event
= file
->private_data
;
3961 unsigned long user_locked
, user_lock_limit
;
3962 struct user_struct
*user
= current_user();
3963 unsigned long locked
, lock_limit
;
3964 struct ring_buffer
*rb
;
3965 unsigned long vma_size
;
3966 unsigned long nr_pages
;
3967 long user_extra
, extra
;
3968 int ret
= 0, flags
= 0;
3971 * Don't allow mmap() of inherited per-task counters. This would
3972 * create a performance issue due to all children writing to the
3975 if (event
->cpu
== -1 && event
->attr
.inherit
)
3978 if (!(vma
->vm_flags
& VM_SHARED
))
3981 vma_size
= vma
->vm_end
- vma
->vm_start
;
3982 nr_pages
= (vma_size
/ PAGE_SIZE
) - 1;
3985 * If we have rb pages ensure they're a power-of-two number, so we
3986 * can do bitmasks instead of modulo.
3988 if (nr_pages
!= 0 && !is_power_of_2(nr_pages
))
3991 if (vma_size
!= PAGE_SIZE
* (1 + nr_pages
))
3994 if (vma
->vm_pgoff
!= 0)
3997 WARN_ON_ONCE(event
->ctx
->parent_ctx
);
3999 mutex_lock(&event
->mmap_mutex
);
4001 if (event
->rb
->nr_pages
!= nr_pages
) {
4006 if (!atomic_inc_not_zero(&event
->rb
->mmap_count
)) {
4008 * Raced against perf_mmap_close() through
4009 * perf_event_set_output(). Try again, hope for better
4012 mutex_unlock(&event
->mmap_mutex
);
4019 user_extra
= nr_pages
+ 1;
4020 user_lock_limit
= sysctl_perf_event_mlock
>> (PAGE_SHIFT
- 10);
4023 * Increase the limit linearly with more CPUs:
4025 user_lock_limit
*= num_online_cpus();
4027 user_locked
= atomic_long_read(&user
->locked_vm
) + user_extra
;
4030 if (user_locked
> user_lock_limit
)
4031 extra
= user_locked
- user_lock_limit
;
4033 lock_limit
= rlimit(RLIMIT_MEMLOCK
);
4034 lock_limit
>>= PAGE_SHIFT
;
4035 locked
= vma
->vm_mm
->pinned_vm
+ extra
;
4037 if ((locked
> lock_limit
) && perf_paranoid_tracepoint_raw() &&
4038 !capable(CAP_IPC_LOCK
)) {
4045 if (vma
->vm_flags
& VM_WRITE
)
4046 flags
|= RING_BUFFER_WRITABLE
;
4048 rb
= rb_alloc(nr_pages
,
4049 event
->attr
.watermark
? event
->attr
.wakeup_watermark
: 0,
4057 atomic_set(&rb
->mmap_count
, 1);
4058 rb
->mmap_locked
= extra
;
4059 rb
->mmap_user
= get_current_user();
4061 atomic_long_add(user_extra
, &user
->locked_vm
);
4062 vma
->vm_mm
->pinned_vm
+= extra
;
4064 ring_buffer_attach(event
, rb
);
4065 rcu_assign_pointer(event
->rb
, rb
);
4067 perf_event_init_userpage(event
);
4068 perf_event_update_userpage(event
);
4072 atomic_inc(&event
->mmap_count
);
4073 mutex_unlock(&event
->mmap_mutex
);
4076 * Since pinned accounting is per vm we cannot allow fork() to copy our
4079 vma
->vm_flags
|= VM_DONTCOPY
| VM_DONTEXPAND
| VM_DONTDUMP
;
4080 vma
->vm_ops
= &perf_mmap_vmops
;
4085 static int perf_fasync(int fd
, struct file
*filp
, int on
)
4087 struct inode
*inode
= file_inode(filp
);
4088 struct perf_event
*event
= filp
->private_data
;
4091 mutex_lock(&inode
->i_mutex
);
4092 retval
= fasync_helper(fd
, filp
, on
, &event
->fasync
);
4093 mutex_unlock(&inode
->i_mutex
);
4101 static const struct file_operations perf_fops
= {
4102 .llseek
= no_llseek
,
4103 .release
= perf_release
,
4106 .unlocked_ioctl
= perf_ioctl
,
4107 .compat_ioctl
= perf_ioctl
,
4109 .fasync
= perf_fasync
,
4115 * If there's data, ensure we set the poll() state and publish everything
4116 * to user-space before waking everybody up.
4119 void perf_event_wakeup(struct perf_event
*event
)
4121 ring_buffer_wakeup(event
);
4123 if (event
->pending_kill
) {
4124 kill_fasync(&event
->fasync
, SIGIO
, event
->pending_kill
);
4125 event
->pending_kill
= 0;
4129 static void perf_pending_event(struct irq_work
*entry
)
4131 struct perf_event
*event
= container_of(entry
,
4132 struct perf_event
, pending
);
4134 if (event
->pending_disable
) {
4135 event
->pending_disable
= 0;
4136 __perf_event_disable(event
);
4139 if (event
->pending_wakeup
) {
4140 event
->pending_wakeup
= 0;
4141 perf_event_wakeup(event
);
4146 * We assume there is only KVM supporting the callbacks.
4147 * Later on, we might change it to a list if there is
4148 * another virtualization implementation supporting the callbacks.
4150 struct perf_guest_info_callbacks
*perf_guest_cbs
;
4152 int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks
*cbs
)
4154 perf_guest_cbs
= cbs
;
4157 EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks
);
4159 int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks
*cbs
)
4161 perf_guest_cbs
= NULL
;
4164 EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks
);
4167 perf_output_sample_regs(struct perf_output_handle
*handle
,
4168 struct pt_regs
*regs
, u64 mask
)
4172 for_each_set_bit(bit
, (const unsigned long *) &mask
,
4173 sizeof(mask
) * BITS_PER_BYTE
) {
4176 val
= perf_reg_value(regs
, bit
);
4177 perf_output_put(handle
, val
);
4181 static void perf_sample_regs_user(struct perf_regs_user
*regs_user
,
4182 struct pt_regs
*regs
)
4184 if (!user_mode(regs
)) {
4186 regs
= task_pt_regs(current
);
4192 regs_user
->regs
= regs
;
4193 regs_user
->abi
= perf_reg_abi(current
);
4198 * Get remaining task size from user stack pointer.
4200 * It'd be better to take stack vma map and limit this more
4201 * precisly, but there's no way to get it safely under interrupt,
4202 * so using TASK_SIZE as limit.
4204 static u64
perf_ustack_task_size(struct pt_regs
*regs
)
4206 unsigned long addr
= perf_user_stack_pointer(regs
);
4208 if (!addr
|| addr
>= TASK_SIZE
)
4211 return TASK_SIZE
- addr
;
4215 perf_sample_ustack_size(u16 stack_size
, u16 header_size
,
4216 struct pt_regs
*regs
)
4220 /* No regs, no stack pointer, no dump. */
4225 * Check if we fit in with the requested stack size into the:
4227 * If we don't, we limit the size to the TASK_SIZE.
4229 * - remaining sample size
4230 * If we don't, we customize the stack size to
4231 * fit in to the remaining sample size.
4234 task_size
= min((u64
) USHRT_MAX
, perf_ustack_task_size(regs
));
4235 stack_size
= min(stack_size
, (u16
) task_size
);
4237 /* Current header size plus static size and dynamic size. */
4238 header_size
+= 2 * sizeof(u64
);
4240 /* Do we fit in with the current stack dump size? */
4241 if ((u16
) (header_size
+ stack_size
) < header_size
) {
4243 * If we overflow the maximum size for the sample,
4244 * we customize the stack dump size to fit in.
4246 stack_size
= USHRT_MAX
- header_size
- sizeof(u64
);
4247 stack_size
= round_up(stack_size
, sizeof(u64
));
4254 perf_output_sample_ustack(struct perf_output_handle
*handle
, u64 dump_size
,
4255 struct pt_regs
*regs
)
4257 /* Case of a kernel thread, nothing to dump */
4260 perf_output_put(handle
, size
);
4269 * - the size requested by user or the best one we can fit
4270 * in to the sample max size
4272 * - user stack dump data
4274 * - the actual dumped size
4278 perf_output_put(handle
, dump_size
);
4281 sp
= perf_user_stack_pointer(regs
);
4282 rem
= __output_copy_user(handle
, (void *) sp
, dump_size
);
4283 dyn_size
= dump_size
- rem
;
4285 perf_output_skip(handle
, rem
);
4288 perf_output_put(handle
, dyn_size
);
4292 static void __perf_event_header__init_id(struct perf_event_header
*header
,
4293 struct perf_sample_data
*data
,
4294 struct perf_event
*event
)
4296 u64 sample_type
= event
->attr
.sample_type
;
4298 data
->type
= sample_type
;
4299 header
->size
+= event
->id_header_size
;
4301 if (sample_type
& PERF_SAMPLE_TID
) {
4302 /* namespace issues */
4303 data
->tid_entry
.pid
= perf_event_pid(event
, current
);
4304 data
->tid_entry
.tid
= perf_event_tid(event
, current
);
4307 if (sample_type
& PERF_SAMPLE_TIME
)
4308 data
->time
= perf_clock();
4310 if (sample_type
& (PERF_SAMPLE_ID
| PERF_SAMPLE_IDENTIFIER
))
4311 data
->id
= primary_event_id(event
);
4313 if (sample_type
& PERF_SAMPLE_STREAM_ID
)
4314 data
->stream_id
= event
->id
;
4316 if (sample_type
& PERF_SAMPLE_CPU
) {
4317 data
->cpu_entry
.cpu
= raw_smp_processor_id();
4318 data
->cpu_entry
.reserved
= 0;
4322 void perf_event_header__init_id(struct perf_event_header
*header
,
4323 struct perf_sample_data
*data
,
4324 struct perf_event
*event
)
4326 if (event
->attr
.sample_id_all
)
4327 __perf_event_header__init_id(header
, data
, event
);
4330 static void __perf_event__output_id_sample(struct perf_output_handle
*handle
,
4331 struct perf_sample_data
*data
)
4333 u64 sample_type
= data
->type
;
4335 if (sample_type
& PERF_SAMPLE_TID
)
4336 perf_output_put(handle
, data
->tid_entry
);
4338 if (sample_type
& PERF_SAMPLE_TIME
)
4339 perf_output_put(handle
, data
->time
);
4341 if (sample_type
& PERF_SAMPLE_ID
)
4342 perf_output_put(handle
, data
->id
);
4344 if (sample_type
& PERF_SAMPLE_STREAM_ID
)
4345 perf_output_put(handle
, data
->stream_id
);
4347 if (sample_type
& PERF_SAMPLE_CPU
)
4348 perf_output_put(handle
, data
->cpu_entry
);
4350 if (sample_type
& PERF_SAMPLE_IDENTIFIER
)
4351 perf_output_put(handle
, data
->id
);
4354 void perf_event__output_id_sample(struct perf_event
*event
,
4355 struct perf_output_handle
*handle
,
4356 struct perf_sample_data
*sample
)
4358 if (event
->attr
.sample_id_all
)
4359 __perf_event__output_id_sample(handle
, sample
);
4362 static void perf_output_read_one(struct perf_output_handle
*handle
,
4363 struct perf_event
*event
,
4364 u64 enabled
, u64 running
)
4366 u64 read_format
= event
->attr
.read_format
;
4370 values
[n
++] = perf_event_count(event
);
4371 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
) {
4372 values
[n
++] = enabled
+
4373 atomic64_read(&event
->child_total_time_enabled
);
4375 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
) {
4376 values
[n
++] = running
+
4377 atomic64_read(&event
->child_total_time_running
);
4379 if (read_format
& PERF_FORMAT_ID
)
4380 values
[n
++] = primary_event_id(event
);
4382 __output_copy(handle
, values
, n
* sizeof(u64
));
4386 * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
4388 static void perf_output_read_group(struct perf_output_handle
*handle
,
4389 struct perf_event
*event
,
4390 u64 enabled
, u64 running
)
4392 struct perf_event
*leader
= event
->group_leader
, *sub
;
4393 u64 read_format
= event
->attr
.read_format
;
4397 values
[n
++] = 1 + leader
->nr_siblings
;
4399 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
4400 values
[n
++] = enabled
;
4402 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
4403 values
[n
++] = running
;
4405 if (leader
!= event
)
4406 leader
->pmu
->read(leader
);
4408 values
[n
++] = perf_event_count(leader
);
4409 if (read_format
& PERF_FORMAT_ID
)
4410 values
[n
++] = primary_event_id(leader
);
4412 __output_copy(handle
, values
, n
* sizeof(u64
));
4414 list_for_each_entry(sub
, &leader
->sibling_list
, group_entry
) {
4417 if ((sub
!= event
) &&
4418 (sub
->state
== PERF_EVENT_STATE_ACTIVE
))
4419 sub
->pmu
->read(sub
);
4421 values
[n
++] = perf_event_count(sub
);
4422 if (read_format
& PERF_FORMAT_ID
)
4423 values
[n
++] = primary_event_id(sub
);
4425 __output_copy(handle
, values
, n
* sizeof(u64
));
4429 #define PERF_FORMAT_TOTAL_TIMES (PERF_FORMAT_TOTAL_TIME_ENABLED|\
4430 PERF_FORMAT_TOTAL_TIME_RUNNING)
4432 static void perf_output_read(struct perf_output_handle
*handle
,
4433 struct perf_event
*event
)
4435 u64 enabled
= 0, running
= 0, now
;
4436 u64 read_format
= event
->attr
.read_format
;
4439 * compute total_time_enabled, total_time_running
4440 * based on snapshot values taken when the event
4441 * was last scheduled in.
4443 * we cannot simply called update_context_time()
4444 * because of locking issue as we are called in
4447 if (read_format
& PERF_FORMAT_TOTAL_TIMES
)
4448 calc_timer_values(event
, &now
, &enabled
, &running
);
4450 if (event
->attr
.read_format
& PERF_FORMAT_GROUP
)
4451 perf_output_read_group(handle
, event
, enabled
, running
);
4453 perf_output_read_one(handle
, event
, enabled
, running
);
4456 void perf_output_sample(struct perf_output_handle
*handle
,
4457 struct perf_event_header
*header
,
4458 struct perf_sample_data
*data
,
4459 struct perf_event
*event
)
4461 u64 sample_type
= data
->type
;
4463 perf_output_put(handle
, *header
);
4465 if (sample_type
& PERF_SAMPLE_IDENTIFIER
)
4466 perf_output_put(handle
, data
->id
);
4468 if (sample_type
& PERF_SAMPLE_IP
)
4469 perf_output_put(handle
, data
->ip
);
4471 if (sample_type
& PERF_SAMPLE_TID
)
4472 perf_output_put(handle
, data
->tid_entry
);
4474 if (sample_type
& PERF_SAMPLE_TIME
)
4475 perf_output_put(handle
, data
->time
);
4477 if (sample_type
& PERF_SAMPLE_ADDR
)
4478 perf_output_put(handle
, data
->addr
);
4480 if (sample_type
& PERF_SAMPLE_ID
)
4481 perf_output_put(handle
, data
->id
);
4483 if (sample_type
& PERF_SAMPLE_STREAM_ID
)
4484 perf_output_put(handle
, data
->stream_id
);
4486 if (sample_type
& PERF_SAMPLE_CPU
)
4487 perf_output_put(handle
, data
->cpu_entry
);
4489 if (sample_type
& PERF_SAMPLE_PERIOD
)
4490 perf_output_put(handle
, data
->period
);
4492 if (sample_type
& PERF_SAMPLE_READ
)
4493 perf_output_read(handle
, event
);
4495 if (sample_type
& PERF_SAMPLE_CALLCHAIN
) {
4496 if (data
->callchain
) {
4499 if (data
->callchain
)
4500 size
+= data
->callchain
->nr
;
4502 size
*= sizeof(u64
);
4504 __output_copy(handle
, data
->callchain
, size
);
4507 perf_output_put(handle
, nr
);
4511 if (sample_type
& PERF_SAMPLE_RAW
) {
4513 perf_output_put(handle
, data
->raw
->size
);
4514 __output_copy(handle
, data
->raw
->data
,
4521 .size
= sizeof(u32
),
4524 perf_output_put(handle
, raw
);
4528 if (sample_type
& PERF_SAMPLE_BRANCH_STACK
) {
4529 if (data
->br_stack
) {
4532 size
= data
->br_stack
->nr
4533 * sizeof(struct perf_branch_entry
);
4535 perf_output_put(handle
, data
->br_stack
->nr
);
4536 perf_output_copy(handle
, data
->br_stack
->entries
, size
);
4539 * we always store at least the value of nr
4542 perf_output_put(handle
, nr
);
4546 if (sample_type
& PERF_SAMPLE_REGS_USER
) {
4547 u64 abi
= data
->regs_user
.abi
;
4550 * If there are no regs to dump, notice it through
4551 * first u64 being zero (PERF_SAMPLE_REGS_ABI_NONE).
4553 perf_output_put(handle
, abi
);
4556 u64 mask
= event
->attr
.sample_regs_user
;
4557 perf_output_sample_regs(handle
,
4558 data
->regs_user
.regs
,
4563 if (sample_type
& PERF_SAMPLE_STACK_USER
) {
4564 perf_output_sample_ustack(handle
,
4565 data
->stack_user_size
,
4566 data
->regs_user
.regs
);
4569 if (sample_type
& PERF_SAMPLE_WEIGHT
)
4570 perf_output_put(handle
, data
->weight
);
4572 if (sample_type
& PERF_SAMPLE_DATA_SRC
)
4573 perf_output_put(handle
, data
->data_src
.val
);
4575 if (!event
->attr
.watermark
) {
4576 int wakeup_events
= event
->attr
.wakeup_events
;
4578 if (wakeup_events
) {
4579 struct ring_buffer
*rb
= handle
->rb
;
4580 int events
= local_inc_return(&rb
->events
);
4582 if (events
>= wakeup_events
) {
4583 local_sub(wakeup_events
, &rb
->events
);
4584 local_inc(&rb
->wakeup
);
4590 void perf_prepare_sample(struct perf_event_header
*header
,
4591 struct perf_sample_data
*data
,
4592 struct perf_event
*event
,
4593 struct pt_regs
*regs
)
4595 u64 sample_type
= event
->attr
.sample_type
;
4597 header
->type
= PERF_RECORD_SAMPLE
;
4598 header
->size
= sizeof(*header
) + event
->header_size
;
4601 header
->misc
|= perf_misc_flags(regs
);
4603 __perf_event_header__init_id(header
, data
, event
);
4605 if (sample_type
& PERF_SAMPLE_IP
)
4606 data
->ip
= perf_instruction_pointer(regs
);
4608 if (sample_type
& PERF_SAMPLE_CALLCHAIN
) {
4611 data
->callchain
= perf_callchain(event
, regs
);
4613 if (data
->callchain
)
4614 size
+= data
->callchain
->nr
;
4616 header
->size
+= size
* sizeof(u64
);
4619 if (sample_type
& PERF_SAMPLE_RAW
) {
4620 int size
= sizeof(u32
);
4623 size
+= data
->raw
->size
;
4625 size
+= sizeof(u32
);
4627 WARN_ON_ONCE(size
& (sizeof(u64
)-1));
4628 header
->size
+= size
;
4631 if (sample_type
& PERF_SAMPLE_BRANCH_STACK
) {
4632 int size
= sizeof(u64
); /* nr */
4633 if (data
->br_stack
) {
4634 size
+= data
->br_stack
->nr
4635 * sizeof(struct perf_branch_entry
);
4637 header
->size
+= size
;
4640 if (sample_type
& PERF_SAMPLE_REGS_USER
) {
4641 /* regs dump ABI info */
4642 int size
= sizeof(u64
);
4644 perf_sample_regs_user(&data
->regs_user
, regs
);
4646 if (data
->regs_user
.regs
) {
4647 u64 mask
= event
->attr
.sample_regs_user
;
4648 size
+= hweight64(mask
) * sizeof(u64
);
4651 header
->size
+= size
;
4654 if (sample_type
& PERF_SAMPLE_STACK_USER
) {
4656 * Either we need PERF_SAMPLE_STACK_USER bit to be allways
4657 * processed as the last one or have additional check added
4658 * in case new sample type is added, because we could eat
4659 * up the rest of the sample size.
4661 struct perf_regs_user
*uregs
= &data
->regs_user
;
4662 u16 stack_size
= event
->attr
.sample_stack_user
;
4663 u16 size
= sizeof(u64
);
4666 perf_sample_regs_user(uregs
, regs
);
4668 stack_size
= perf_sample_ustack_size(stack_size
, header
->size
,
4672 * If there is something to dump, add space for the dump
4673 * itself and for the field that tells the dynamic size,
4674 * which is how many have been actually dumped.
4677 size
+= sizeof(u64
) + stack_size
;
4679 data
->stack_user_size
= stack_size
;
4680 header
->size
+= size
;
4684 static void perf_event_output(struct perf_event
*event
,
4685 struct perf_sample_data
*data
,
4686 struct pt_regs
*regs
)
4688 struct perf_output_handle handle
;
4689 struct perf_event_header header
;
4691 /* protect the callchain buffers */
4694 perf_prepare_sample(&header
, data
, event
, regs
);
4696 if (perf_output_begin(&handle
, event
, header
.size
))
4699 perf_output_sample(&handle
, &header
, data
, event
);
4701 perf_output_end(&handle
);
4711 struct perf_read_event
{
4712 struct perf_event_header header
;
4719 perf_event_read_event(struct perf_event
*event
,
4720 struct task_struct
*task
)
4722 struct perf_output_handle handle
;
4723 struct perf_sample_data sample
;
4724 struct perf_read_event read_event
= {
4726 .type
= PERF_RECORD_READ
,
4728 .size
= sizeof(read_event
) + event
->read_size
,
4730 .pid
= perf_event_pid(event
, task
),
4731 .tid
= perf_event_tid(event
, task
),
4735 perf_event_header__init_id(&read_event
.header
, &sample
, event
);
4736 ret
= perf_output_begin(&handle
, event
, read_event
.header
.size
);
4740 perf_output_put(&handle
, read_event
);
4741 perf_output_read(&handle
, event
);
4742 perf_event__output_id_sample(event
, &handle
, &sample
);
4744 perf_output_end(&handle
);
4747 typedef void (perf_event_aux_output_cb
)(struct perf_event
*event
, void *data
);
4750 perf_event_aux_ctx(struct perf_event_context
*ctx
,
4751 perf_event_aux_output_cb output
,
4754 struct perf_event
*event
;
4756 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
4757 if (event
->state
< PERF_EVENT_STATE_INACTIVE
)
4759 if (!event_filter_match(event
))
4761 output(event
, data
);
4766 perf_event_aux(perf_event_aux_output_cb output
, void *data
,
4767 struct perf_event_context
*task_ctx
)
4769 struct perf_cpu_context
*cpuctx
;
4770 struct perf_event_context
*ctx
;
4775 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
4776 cpuctx
= get_cpu_ptr(pmu
->pmu_cpu_context
);
4777 if (cpuctx
->unique_pmu
!= pmu
)
4779 perf_event_aux_ctx(&cpuctx
->ctx
, output
, data
);
4782 ctxn
= pmu
->task_ctx_nr
;
4785 ctx
= rcu_dereference(current
->perf_event_ctxp
[ctxn
]);
4787 perf_event_aux_ctx(ctx
, output
, data
);
4789 put_cpu_ptr(pmu
->pmu_cpu_context
);
4794 perf_event_aux_ctx(task_ctx
, output
, data
);
4801 * task tracking -- fork/exit
4803 * enabled by: attr.comm | attr.mmap | attr.mmap2 | attr.mmap_data | attr.task
4806 struct perf_task_event
{
4807 struct task_struct
*task
;
4808 struct perf_event_context
*task_ctx
;
4811 struct perf_event_header header
;
4821 static int perf_event_task_match(struct perf_event
*event
)
4823 return event
->attr
.comm
|| event
->attr
.mmap
||
4824 event
->attr
.mmap2
|| event
->attr
.mmap_data
||
4828 static void perf_event_task_output(struct perf_event
*event
,
4831 struct perf_task_event
*task_event
= data
;
4832 struct perf_output_handle handle
;
4833 struct perf_sample_data sample
;
4834 struct task_struct
*task
= task_event
->task
;
4835 int ret
, size
= task_event
->event_id
.header
.size
;
4837 if (!perf_event_task_match(event
))
4840 perf_event_header__init_id(&task_event
->event_id
.header
, &sample
, event
);
4842 ret
= perf_output_begin(&handle
, event
,
4843 task_event
->event_id
.header
.size
);
4847 task_event
->event_id
.pid
= perf_event_pid(event
, task
);
4848 task_event
->event_id
.ppid
= perf_event_pid(event
, current
);
4850 task_event
->event_id
.tid
= perf_event_tid(event
, task
);
4851 task_event
->event_id
.ptid
= perf_event_tid(event
, current
);
4853 perf_output_put(&handle
, task_event
->event_id
);
4855 perf_event__output_id_sample(event
, &handle
, &sample
);
4857 perf_output_end(&handle
);
4859 task_event
->event_id
.header
.size
= size
;
4862 static void perf_event_task(struct task_struct
*task
,
4863 struct perf_event_context
*task_ctx
,
4866 struct perf_task_event task_event
;
4868 if (!atomic_read(&nr_comm_events
) &&
4869 !atomic_read(&nr_mmap_events
) &&
4870 !atomic_read(&nr_task_events
))
4873 task_event
= (struct perf_task_event
){
4875 .task_ctx
= task_ctx
,
4878 .type
= new ? PERF_RECORD_FORK
: PERF_RECORD_EXIT
,
4880 .size
= sizeof(task_event
.event_id
),
4886 .time
= perf_clock(),
4890 perf_event_aux(perf_event_task_output
,
4895 void perf_event_fork(struct task_struct
*task
)
4897 perf_event_task(task
, NULL
, 1);
4904 struct perf_comm_event
{
4905 struct task_struct
*task
;
4910 struct perf_event_header header
;
4917 static int perf_event_comm_match(struct perf_event
*event
)
4919 return event
->attr
.comm
;
4922 static void perf_event_comm_output(struct perf_event
*event
,
4925 struct perf_comm_event
*comm_event
= data
;
4926 struct perf_output_handle handle
;
4927 struct perf_sample_data sample
;
4928 int size
= comm_event
->event_id
.header
.size
;
4931 if (!perf_event_comm_match(event
))
4934 perf_event_header__init_id(&comm_event
->event_id
.header
, &sample
, event
);
4935 ret
= perf_output_begin(&handle
, event
,
4936 comm_event
->event_id
.header
.size
);
4941 comm_event
->event_id
.pid
= perf_event_pid(event
, comm_event
->task
);
4942 comm_event
->event_id
.tid
= perf_event_tid(event
, comm_event
->task
);
4944 perf_output_put(&handle
, comm_event
->event_id
);
4945 __output_copy(&handle
, comm_event
->comm
,
4946 comm_event
->comm_size
);
4948 perf_event__output_id_sample(event
, &handle
, &sample
);
4950 perf_output_end(&handle
);
4952 comm_event
->event_id
.header
.size
= size
;
4955 static void perf_event_comm_event(struct perf_comm_event
*comm_event
)
4957 char comm
[TASK_COMM_LEN
];
4960 memset(comm
, 0, sizeof(comm
));
4961 strlcpy(comm
, comm_event
->task
->comm
, sizeof(comm
));
4962 size
= ALIGN(strlen(comm
)+1, sizeof(u64
));
4964 comm_event
->comm
= comm
;
4965 comm_event
->comm_size
= size
;
4967 comm_event
->event_id
.header
.size
= sizeof(comm_event
->event_id
) + size
;
4969 perf_event_aux(perf_event_comm_output
,
4974 void perf_event_comm(struct task_struct
*task
)
4976 struct perf_comm_event comm_event
;
4977 struct perf_event_context
*ctx
;
4981 for_each_task_context_nr(ctxn
) {
4982 ctx
= task
->perf_event_ctxp
[ctxn
];
4986 perf_event_enable_on_exec(ctx
);
4990 if (!atomic_read(&nr_comm_events
))
4993 comm_event
= (struct perf_comm_event
){
4999 .type
= PERF_RECORD_COMM
,
5008 perf_event_comm_event(&comm_event
);
5015 struct perf_mmap_event
{
5016 struct vm_area_struct
*vma
;
5018 const char *file_name
;
5025 struct perf_event_header header
;
5035 static int perf_event_mmap_match(struct perf_event
*event
,
5038 struct perf_mmap_event
*mmap_event
= data
;
5039 struct vm_area_struct
*vma
= mmap_event
->vma
;
5040 int executable
= vma
->vm_flags
& VM_EXEC
;
5042 return (!executable
&& event
->attr
.mmap_data
) ||
5043 (executable
&& (event
->attr
.mmap
|| event
->attr
.mmap2
));
5046 static void perf_event_mmap_output(struct perf_event
*event
,
5049 struct perf_mmap_event
*mmap_event
= data
;
5050 struct perf_output_handle handle
;
5051 struct perf_sample_data sample
;
5052 int size
= mmap_event
->event_id
.header
.size
;
5055 if (!perf_event_mmap_match(event
, data
))
5058 if (event
->attr
.mmap2
) {
5059 mmap_event
->event_id
.header
.type
= PERF_RECORD_MMAP2
;
5060 mmap_event
->event_id
.header
.size
+= sizeof(mmap_event
->maj
);
5061 mmap_event
->event_id
.header
.size
+= sizeof(mmap_event
->min
);
5062 mmap_event
->event_id
.header
.size
+= sizeof(mmap_event
->ino
);
5063 mmap_event
->event_id
.header
.size
+= sizeof(mmap_event
->ino_generation
);
5066 perf_event_header__init_id(&mmap_event
->event_id
.header
, &sample
, event
);
5067 ret
= perf_output_begin(&handle
, event
,
5068 mmap_event
->event_id
.header
.size
);
5072 mmap_event
->event_id
.pid
= perf_event_pid(event
, current
);
5073 mmap_event
->event_id
.tid
= perf_event_tid(event
, current
);
5075 perf_output_put(&handle
, mmap_event
->event_id
);
5077 if (event
->attr
.mmap2
) {
5078 perf_output_put(&handle
, mmap_event
->maj
);
5079 perf_output_put(&handle
, mmap_event
->min
);
5080 perf_output_put(&handle
, mmap_event
->ino
);
5081 perf_output_put(&handle
, mmap_event
->ino_generation
);
5084 __output_copy(&handle
, mmap_event
->file_name
,
5085 mmap_event
->file_size
);
5087 perf_event__output_id_sample(event
, &handle
, &sample
);
5089 perf_output_end(&handle
);
5091 mmap_event
->event_id
.header
.size
= size
;
5094 static void perf_event_mmap_event(struct perf_mmap_event
*mmap_event
)
5096 struct vm_area_struct
*vma
= mmap_event
->vma
;
5097 struct file
*file
= vma
->vm_file
;
5098 int maj
= 0, min
= 0;
5099 u64 ino
= 0, gen
= 0;
5105 memset(tmp
, 0, sizeof(tmp
));
5108 struct inode
*inode
;
5111 * d_path works from the end of the rb backwards, so we
5112 * need to add enough zero bytes after the string to handle
5113 * the 64bit alignment we do later.
5115 buf
= kzalloc(PATH_MAX
+ sizeof(u64
), GFP_KERNEL
);
5117 name
= strncpy(tmp
, "//enomem", sizeof(tmp
));
5120 name
= d_path(&file
->f_path
, buf
, PATH_MAX
);
5122 name
= strncpy(tmp
, "//toolong", sizeof(tmp
));
5125 inode
= file_inode(vma
->vm_file
);
5126 dev
= inode
->i_sb
->s_dev
;
5128 gen
= inode
->i_generation
;
5133 if (arch_vma_name(mmap_event
->vma
)) {
5134 name
= strncpy(tmp
, arch_vma_name(mmap_event
->vma
),
5136 tmp
[sizeof(tmp
) - 1] = '\0';
5141 name
= strncpy(tmp
, "[vdso]", sizeof(tmp
));
5143 } else if (vma
->vm_start
<= vma
->vm_mm
->start_brk
&&
5144 vma
->vm_end
>= vma
->vm_mm
->brk
) {
5145 name
= strncpy(tmp
, "[heap]", sizeof(tmp
));
5147 } else if (vma
->vm_start
<= vma
->vm_mm
->start_stack
&&
5148 vma
->vm_end
>= vma
->vm_mm
->start_stack
) {
5149 name
= strncpy(tmp
, "[stack]", sizeof(tmp
));
5153 name
= strncpy(tmp
, "//anon", sizeof(tmp
));
5158 size
= ALIGN(strlen(name
)+1, sizeof(u64
));
5160 mmap_event
->file_name
= name
;
5161 mmap_event
->file_size
= size
;
5162 mmap_event
->maj
= maj
;
5163 mmap_event
->min
= min
;
5164 mmap_event
->ino
= ino
;
5165 mmap_event
->ino_generation
= gen
;
5167 if (!(vma
->vm_flags
& VM_EXEC
))
5168 mmap_event
->event_id
.header
.misc
|= PERF_RECORD_MISC_MMAP_DATA
;
5170 mmap_event
->event_id
.header
.size
= sizeof(mmap_event
->event_id
) + size
;
5172 perf_event_aux(perf_event_mmap_output
,
5179 void perf_event_mmap(struct vm_area_struct
*vma
)
5181 struct perf_mmap_event mmap_event
;
5183 if (!atomic_read(&nr_mmap_events
))
5186 mmap_event
= (struct perf_mmap_event
){
5192 .type
= PERF_RECORD_MMAP
,
5193 .misc
= PERF_RECORD_MISC_USER
,
5198 .start
= vma
->vm_start
,
5199 .len
= vma
->vm_end
- vma
->vm_start
,
5200 .pgoff
= (u64
)vma
->vm_pgoff
<< PAGE_SHIFT
,
5202 /* .maj (attr_mmap2 only) */
5203 /* .min (attr_mmap2 only) */
5204 /* .ino (attr_mmap2 only) */
5205 /* .ino_generation (attr_mmap2 only) */
5208 perf_event_mmap_event(&mmap_event
);
5212 * IRQ throttle logging
5215 static void perf_log_throttle(struct perf_event
*event
, int enable
)
5217 struct perf_output_handle handle
;
5218 struct perf_sample_data sample
;
5222 struct perf_event_header header
;
5226 } throttle_event
= {
5228 .type
= PERF_RECORD_THROTTLE
,
5230 .size
= sizeof(throttle_event
),
5232 .time
= perf_clock(),
5233 .id
= primary_event_id(event
),
5234 .stream_id
= event
->id
,
5238 throttle_event
.header
.type
= PERF_RECORD_UNTHROTTLE
;
5240 perf_event_header__init_id(&throttle_event
.header
, &sample
, event
);
5242 ret
= perf_output_begin(&handle
, event
,
5243 throttle_event
.header
.size
);
5247 perf_output_put(&handle
, throttle_event
);
5248 perf_event__output_id_sample(event
, &handle
, &sample
);
5249 perf_output_end(&handle
);
5253 * Generic event overflow handling, sampling.
5256 static int __perf_event_overflow(struct perf_event
*event
,
5257 int throttle
, struct perf_sample_data
*data
,
5258 struct pt_regs
*regs
)
5260 int events
= atomic_read(&event
->event_limit
);
5261 struct hw_perf_event
*hwc
= &event
->hw
;
5266 * Non-sampling counters might still use the PMI to fold short
5267 * hardware counters, ignore those.
5269 if (unlikely(!is_sampling_event(event
)))
5272 seq
= __this_cpu_read(perf_throttled_seq
);
5273 if (seq
!= hwc
->interrupts_seq
) {
5274 hwc
->interrupts_seq
= seq
;
5275 hwc
->interrupts
= 1;
5278 if (unlikely(throttle
5279 && hwc
->interrupts
>= max_samples_per_tick
)) {
5280 __this_cpu_inc(perf_throttled_count
);
5281 hwc
->interrupts
= MAX_INTERRUPTS
;
5282 perf_log_throttle(event
, 0);
5283 tick_nohz_full_kick();
5288 if (event
->attr
.freq
) {
5289 u64 now
= perf_clock();
5290 s64 delta
= now
- hwc
->freq_time_stamp
;
5292 hwc
->freq_time_stamp
= now
;
5294 if (delta
> 0 && delta
< 2*TICK_NSEC
)
5295 perf_adjust_period(event
, delta
, hwc
->last_period
, true);
5299 * XXX event_limit might not quite work as expected on inherited
5303 event
->pending_kill
= POLL_IN
;
5304 if (events
&& atomic_dec_and_test(&event
->event_limit
)) {
5306 event
->pending_kill
= POLL_HUP
;
5307 event
->pending_disable
= 1;
5308 irq_work_queue(&event
->pending
);
5311 if (event
->overflow_handler
)
5312 event
->overflow_handler(event
, data
, regs
);
5314 perf_event_output(event
, data
, regs
);
5316 if (event
->fasync
&& event
->pending_kill
) {
5317 event
->pending_wakeup
= 1;
5318 irq_work_queue(&event
->pending
);
5324 int perf_event_overflow(struct perf_event
*event
,
5325 struct perf_sample_data
*data
,
5326 struct pt_regs
*regs
)
5328 return __perf_event_overflow(event
, 1, data
, regs
);
5332 * Generic software event infrastructure
5335 struct swevent_htable
{
5336 struct swevent_hlist
*swevent_hlist
;
5337 struct mutex hlist_mutex
;
5340 /* Recursion avoidance in each contexts */
5341 int recursion
[PERF_NR_CONTEXTS
];
5344 static DEFINE_PER_CPU(struct swevent_htable
, swevent_htable
);
5347 * We directly increment event->count and keep a second value in
5348 * event->hw.period_left to count intervals. This period event
5349 * is kept in the range [-sample_period, 0] so that we can use the
5353 u64
perf_swevent_set_period(struct perf_event
*event
)
5355 struct hw_perf_event
*hwc
= &event
->hw
;
5356 u64 period
= hwc
->last_period
;
5360 hwc
->last_period
= hwc
->sample_period
;
5363 old
= val
= local64_read(&hwc
->period_left
);
5367 nr
= div64_u64(period
+ val
, period
);
5368 offset
= nr
* period
;
5370 if (local64_cmpxchg(&hwc
->period_left
, old
, val
) != old
)
5376 static void perf_swevent_overflow(struct perf_event
*event
, u64 overflow
,
5377 struct perf_sample_data
*data
,
5378 struct pt_regs
*regs
)
5380 struct hw_perf_event
*hwc
= &event
->hw
;
5384 overflow
= perf_swevent_set_period(event
);
5386 if (hwc
->interrupts
== MAX_INTERRUPTS
)
5389 for (; overflow
; overflow
--) {
5390 if (__perf_event_overflow(event
, throttle
,
5393 * We inhibit the overflow from happening when
5394 * hwc->interrupts == MAX_INTERRUPTS.
5402 static void perf_swevent_event(struct perf_event
*event
, u64 nr
,
5403 struct perf_sample_data
*data
,
5404 struct pt_regs
*regs
)
5406 struct hw_perf_event
*hwc
= &event
->hw
;
5408 local64_add(nr
, &event
->count
);
5413 if (!is_sampling_event(event
))
5416 if ((event
->attr
.sample_type
& PERF_SAMPLE_PERIOD
) && !event
->attr
.freq
) {
5418 return perf_swevent_overflow(event
, 1, data
, regs
);
5420 data
->period
= event
->hw
.last_period
;
5422 if (nr
== 1 && hwc
->sample_period
== 1 && !event
->attr
.freq
)
5423 return perf_swevent_overflow(event
, 1, data
, regs
);
5425 if (local64_add_negative(nr
, &hwc
->period_left
))
5428 perf_swevent_overflow(event
, 0, data
, regs
);
5431 static int perf_exclude_event(struct perf_event
*event
,
5432 struct pt_regs
*regs
)
5434 if (event
->hw
.state
& PERF_HES_STOPPED
)
5438 if (event
->attr
.exclude_user
&& user_mode(regs
))
5441 if (event
->attr
.exclude_kernel
&& !user_mode(regs
))
5448 static int perf_swevent_match(struct perf_event
*event
,
5449 enum perf_type_id type
,
5451 struct perf_sample_data
*data
,
5452 struct pt_regs
*regs
)
5454 if (event
->attr
.type
!= type
)
5457 if (event
->attr
.config
!= event_id
)
5460 if (perf_exclude_event(event
, regs
))
5466 static inline u64
swevent_hash(u64 type
, u32 event_id
)
5468 u64 val
= event_id
| (type
<< 32);
5470 return hash_64(val
, SWEVENT_HLIST_BITS
);
5473 static inline struct hlist_head
*
5474 __find_swevent_head(struct swevent_hlist
*hlist
, u64 type
, u32 event_id
)
5476 u64 hash
= swevent_hash(type
, event_id
);
5478 return &hlist
->heads
[hash
];
5481 /* For the read side: events when they trigger */
5482 static inline struct hlist_head
*
5483 find_swevent_head_rcu(struct swevent_htable
*swhash
, u64 type
, u32 event_id
)
5485 struct swevent_hlist
*hlist
;
5487 hlist
= rcu_dereference(swhash
->swevent_hlist
);
5491 return __find_swevent_head(hlist
, type
, event_id
);
5494 /* For the event head insertion and removal in the hlist */
5495 static inline struct hlist_head
*
5496 find_swevent_head(struct swevent_htable
*swhash
, struct perf_event
*event
)
5498 struct swevent_hlist
*hlist
;
5499 u32 event_id
= event
->attr
.config
;
5500 u64 type
= event
->attr
.type
;
5503 * Event scheduling is always serialized against hlist allocation
5504 * and release. Which makes the protected version suitable here.
5505 * The context lock guarantees that.
5507 hlist
= rcu_dereference_protected(swhash
->swevent_hlist
,
5508 lockdep_is_held(&event
->ctx
->lock
));
5512 return __find_swevent_head(hlist
, type
, event_id
);
5515 static void do_perf_sw_event(enum perf_type_id type
, u32 event_id
,
5517 struct perf_sample_data
*data
,
5518 struct pt_regs
*regs
)
5520 struct swevent_htable
*swhash
= &__get_cpu_var(swevent_htable
);
5521 struct perf_event
*event
;
5522 struct hlist_head
*head
;
5525 head
= find_swevent_head_rcu(swhash
, type
, event_id
);
5529 hlist_for_each_entry_rcu(event
, head
, hlist_entry
) {
5530 if (perf_swevent_match(event
, type
, event_id
, data
, regs
))
5531 perf_swevent_event(event
, nr
, data
, regs
);
5537 int perf_swevent_get_recursion_context(void)
5539 struct swevent_htable
*swhash
= &__get_cpu_var(swevent_htable
);
5541 return get_recursion_context(swhash
->recursion
);
5543 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context
);
5545 inline void perf_swevent_put_recursion_context(int rctx
)
5547 struct swevent_htable
*swhash
= &__get_cpu_var(swevent_htable
);
5549 put_recursion_context(swhash
->recursion
, rctx
);
5552 void __perf_sw_event(u32 event_id
, u64 nr
, struct pt_regs
*regs
, u64 addr
)
5554 struct perf_sample_data data
;
5557 preempt_disable_notrace();
5558 rctx
= perf_swevent_get_recursion_context();
5562 perf_sample_data_init(&data
, addr
, 0);
5564 do_perf_sw_event(PERF_TYPE_SOFTWARE
, event_id
, nr
, &data
, regs
);
5566 perf_swevent_put_recursion_context(rctx
);
5567 preempt_enable_notrace();
5570 static void perf_swevent_read(struct perf_event
*event
)
5574 static int perf_swevent_add(struct perf_event
*event
, int flags
)
5576 struct swevent_htable
*swhash
= &__get_cpu_var(swevent_htable
);
5577 struct hw_perf_event
*hwc
= &event
->hw
;
5578 struct hlist_head
*head
;
5580 if (is_sampling_event(event
)) {
5581 hwc
->last_period
= hwc
->sample_period
;
5582 perf_swevent_set_period(event
);
5585 hwc
->state
= !(flags
& PERF_EF_START
);
5587 head
= find_swevent_head(swhash
, event
);
5588 if (WARN_ON_ONCE(!head
))
5591 hlist_add_head_rcu(&event
->hlist_entry
, head
);
5596 static void perf_swevent_del(struct perf_event
*event
, int flags
)
5598 hlist_del_rcu(&event
->hlist_entry
);
5601 static void perf_swevent_start(struct perf_event
*event
, int flags
)
5603 event
->hw
.state
= 0;
5606 static void perf_swevent_stop(struct perf_event
*event
, int flags
)
5608 event
->hw
.state
= PERF_HES_STOPPED
;
5611 /* Deref the hlist from the update side */
5612 static inline struct swevent_hlist
*
5613 swevent_hlist_deref(struct swevent_htable
*swhash
)
5615 return rcu_dereference_protected(swhash
->swevent_hlist
,
5616 lockdep_is_held(&swhash
->hlist_mutex
));
5619 static void swevent_hlist_release(struct swevent_htable
*swhash
)
5621 struct swevent_hlist
*hlist
= swevent_hlist_deref(swhash
);
5626 rcu_assign_pointer(swhash
->swevent_hlist
, NULL
);
5627 kfree_rcu(hlist
, rcu_head
);
5630 static void swevent_hlist_put_cpu(struct perf_event
*event
, int cpu
)
5632 struct swevent_htable
*swhash
= &per_cpu(swevent_htable
, cpu
);
5634 mutex_lock(&swhash
->hlist_mutex
);
5636 if (!--swhash
->hlist_refcount
)
5637 swevent_hlist_release(swhash
);
5639 mutex_unlock(&swhash
->hlist_mutex
);
5642 static void swevent_hlist_put(struct perf_event
*event
)
5646 if (event
->cpu
!= -1) {
5647 swevent_hlist_put_cpu(event
, event
->cpu
);
5651 for_each_possible_cpu(cpu
)
5652 swevent_hlist_put_cpu(event
, cpu
);
5655 static int swevent_hlist_get_cpu(struct perf_event
*event
, int cpu
)
5657 struct swevent_htable
*swhash
= &per_cpu(swevent_htable
, cpu
);
5660 mutex_lock(&swhash
->hlist_mutex
);
5662 if (!swevent_hlist_deref(swhash
) && cpu_online(cpu
)) {
5663 struct swevent_hlist
*hlist
;
5665 hlist
= kzalloc(sizeof(*hlist
), GFP_KERNEL
);
5670 rcu_assign_pointer(swhash
->swevent_hlist
, hlist
);
5672 swhash
->hlist_refcount
++;
5674 mutex_unlock(&swhash
->hlist_mutex
);
5679 static int swevent_hlist_get(struct perf_event
*event
)
5682 int cpu
, failed_cpu
;
5684 if (event
->cpu
!= -1)
5685 return swevent_hlist_get_cpu(event
, event
->cpu
);
5688 for_each_possible_cpu(cpu
) {
5689 err
= swevent_hlist_get_cpu(event
, cpu
);
5699 for_each_possible_cpu(cpu
) {
5700 if (cpu
== failed_cpu
)
5702 swevent_hlist_put_cpu(event
, cpu
);
5709 struct static_key perf_swevent_enabled
[PERF_COUNT_SW_MAX
];
5711 static void sw_perf_event_destroy(struct perf_event
*event
)
5713 u64 event_id
= event
->attr
.config
;
5715 WARN_ON(event
->parent
);
5717 static_key_slow_dec(&perf_swevent_enabled
[event_id
]);
5718 swevent_hlist_put(event
);
5721 static int perf_swevent_init(struct perf_event
*event
)
5723 u64 event_id
= event
->attr
.config
;
5725 if (event
->attr
.type
!= PERF_TYPE_SOFTWARE
)
5729 * no branch sampling for software events
5731 if (has_branch_stack(event
))
5735 case PERF_COUNT_SW_CPU_CLOCK
:
5736 case PERF_COUNT_SW_TASK_CLOCK
:
5743 if (event_id
>= PERF_COUNT_SW_MAX
)
5746 if (!event
->parent
) {
5749 err
= swevent_hlist_get(event
);
5753 static_key_slow_inc(&perf_swevent_enabled
[event_id
]);
5754 event
->destroy
= sw_perf_event_destroy
;
5760 static int perf_swevent_event_idx(struct perf_event
*event
)
5765 static struct pmu perf_swevent
= {
5766 .task_ctx_nr
= perf_sw_context
,
5768 .event_init
= perf_swevent_init
,
5769 .add
= perf_swevent_add
,
5770 .del
= perf_swevent_del
,
5771 .start
= perf_swevent_start
,
5772 .stop
= perf_swevent_stop
,
5773 .read
= perf_swevent_read
,
5775 .event_idx
= perf_swevent_event_idx
,
5778 #ifdef CONFIG_EVENT_TRACING
5780 static int perf_tp_filter_match(struct perf_event
*event
,
5781 struct perf_sample_data
*data
)
5783 void *record
= data
->raw
->data
;
5785 if (likely(!event
->filter
) || filter_match_preds(event
->filter
, record
))
5790 static int perf_tp_event_match(struct perf_event
*event
,
5791 struct perf_sample_data
*data
,
5792 struct pt_regs
*regs
)
5794 if (event
->hw
.state
& PERF_HES_STOPPED
)
5797 * All tracepoints are from kernel-space.
5799 if (event
->attr
.exclude_kernel
)
5802 if (!perf_tp_filter_match(event
, data
))
5808 void perf_tp_event(u64 addr
, u64 count
, void *record
, int entry_size
,
5809 struct pt_regs
*regs
, struct hlist_head
*head
, int rctx
,
5810 struct task_struct
*task
)
5812 struct perf_sample_data data
;
5813 struct perf_event
*event
;
5815 struct perf_raw_record raw
= {
5820 perf_sample_data_init(&data
, addr
, 0);
5823 hlist_for_each_entry_rcu(event
, head
, hlist_entry
) {
5824 if (perf_tp_event_match(event
, &data
, regs
))
5825 perf_swevent_event(event
, count
, &data
, regs
);
5829 * If we got specified a target task, also iterate its context and
5830 * deliver this event there too.
5832 if (task
&& task
!= current
) {
5833 struct perf_event_context
*ctx
;
5834 struct trace_entry
*entry
= record
;
5837 ctx
= rcu_dereference(task
->perf_event_ctxp
[perf_sw_context
]);
5841 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
5842 if (event
->attr
.type
!= PERF_TYPE_TRACEPOINT
)
5844 if (event
->attr
.config
!= entry
->type
)
5846 if (perf_tp_event_match(event
, &data
, regs
))
5847 perf_swevent_event(event
, count
, &data
, regs
);
5853 perf_swevent_put_recursion_context(rctx
);
5855 EXPORT_SYMBOL_GPL(perf_tp_event
);
5857 static void tp_perf_event_destroy(struct perf_event
*event
)
5859 perf_trace_destroy(event
);
5862 static int perf_tp_event_init(struct perf_event
*event
)
5866 if (event
->attr
.type
!= PERF_TYPE_TRACEPOINT
)
5870 * no branch sampling for tracepoint events
5872 if (has_branch_stack(event
))
5875 err
= perf_trace_init(event
);
5879 event
->destroy
= tp_perf_event_destroy
;
5884 static struct pmu perf_tracepoint
= {
5885 .task_ctx_nr
= perf_sw_context
,
5887 .event_init
= perf_tp_event_init
,
5888 .add
= perf_trace_add
,
5889 .del
= perf_trace_del
,
5890 .start
= perf_swevent_start
,
5891 .stop
= perf_swevent_stop
,
5892 .read
= perf_swevent_read
,
5894 .event_idx
= perf_swevent_event_idx
,
5897 static inline void perf_tp_register(void)
5899 perf_pmu_register(&perf_tracepoint
, "tracepoint", PERF_TYPE_TRACEPOINT
);
5902 static int perf_event_set_filter(struct perf_event
*event
, void __user
*arg
)
5907 if (event
->attr
.type
!= PERF_TYPE_TRACEPOINT
)
5910 filter_str
= strndup_user(arg
, PAGE_SIZE
);
5911 if (IS_ERR(filter_str
))
5912 return PTR_ERR(filter_str
);
5914 ret
= ftrace_profile_set_filter(event
, event
->attr
.config
, filter_str
);
5920 static void perf_event_free_filter(struct perf_event
*event
)
5922 ftrace_profile_free_filter(event
);
5927 static inline void perf_tp_register(void)
5931 static int perf_event_set_filter(struct perf_event
*event
, void __user
*arg
)
5936 static void perf_event_free_filter(struct perf_event
*event
)
5940 #endif /* CONFIG_EVENT_TRACING */
5942 #ifdef CONFIG_HAVE_HW_BREAKPOINT
5943 void perf_bp_event(struct perf_event
*bp
, void *data
)
5945 struct perf_sample_data sample
;
5946 struct pt_regs
*regs
= data
;
5948 perf_sample_data_init(&sample
, bp
->attr
.bp_addr
, 0);
5950 if (!bp
->hw
.state
&& !perf_exclude_event(bp
, regs
))
5951 perf_swevent_event(bp
, 1, &sample
, regs
);
5956 * hrtimer based swevent callback
5959 static enum hrtimer_restart
perf_swevent_hrtimer(struct hrtimer
*hrtimer
)
5961 enum hrtimer_restart ret
= HRTIMER_RESTART
;
5962 struct perf_sample_data data
;
5963 struct pt_regs
*regs
;
5964 struct perf_event
*event
;
5967 event
= container_of(hrtimer
, struct perf_event
, hw
.hrtimer
);
5969 if (event
->state
!= PERF_EVENT_STATE_ACTIVE
)
5970 return HRTIMER_NORESTART
;
5972 event
->pmu
->read(event
);
5974 perf_sample_data_init(&data
, 0, event
->hw
.last_period
);
5975 regs
= get_irq_regs();
5977 if (regs
&& !perf_exclude_event(event
, regs
)) {
5978 if (!(event
->attr
.exclude_idle
&& is_idle_task(current
)))
5979 if (__perf_event_overflow(event
, 1, &data
, regs
))
5980 ret
= HRTIMER_NORESTART
;
5983 period
= max_t(u64
, 10000, event
->hw
.sample_period
);
5984 hrtimer_forward_now(hrtimer
, ns_to_ktime(period
));
5989 static void perf_swevent_start_hrtimer(struct perf_event
*event
)
5991 struct hw_perf_event
*hwc
= &event
->hw
;
5994 if (!is_sampling_event(event
))
5997 period
= local64_read(&hwc
->period_left
);
6002 local64_set(&hwc
->period_left
, 0);
6004 period
= max_t(u64
, 10000, hwc
->sample_period
);
6006 __hrtimer_start_range_ns(&hwc
->hrtimer
,
6007 ns_to_ktime(period
), 0,
6008 HRTIMER_MODE_REL_PINNED
, 0);
6011 static void perf_swevent_cancel_hrtimer(struct perf_event
*event
)
6013 struct hw_perf_event
*hwc
= &event
->hw
;
6015 if (is_sampling_event(event
)) {
6016 ktime_t remaining
= hrtimer_get_remaining(&hwc
->hrtimer
);
6017 local64_set(&hwc
->period_left
, ktime_to_ns(remaining
));
6019 hrtimer_cancel(&hwc
->hrtimer
);
6023 static void perf_swevent_init_hrtimer(struct perf_event
*event
)
6025 struct hw_perf_event
*hwc
= &event
->hw
;
6027 if (!is_sampling_event(event
))
6030 hrtimer_init(&hwc
->hrtimer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
6031 hwc
->hrtimer
.function
= perf_swevent_hrtimer
;
6034 * Since hrtimers have a fixed rate, we can do a static freq->period
6035 * mapping and avoid the whole period adjust feedback stuff.
6037 if (event
->attr
.freq
) {
6038 long freq
= event
->attr
.sample_freq
;
6040 event
->attr
.sample_period
= NSEC_PER_SEC
/ freq
;
6041 hwc
->sample_period
= event
->attr
.sample_period
;
6042 local64_set(&hwc
->period_left
, hwc
->sample_period
);
6043 hwc
->last_period
= hwc
->sample_period
;
6044 event
->attr
.freq
= 0;
6049 * Software event: cpu wall time clock
6052 static void cpu_clock_event_update(struct perf_event
*event
)
6057 now
= local_clock();
6058 prev
= local64_xchg(&event
->hw
.prev_count
, now
);
6059 local64_add(now
- prev
, &event
->count
);
6062 static void cpu_clock_event_start(struct perf_event
*event
, int flags
)
6064 local64_set(&event
->hw
.prev_count
, local_clock());
6065 perf_swevent_start_hrtimer(event
);
6068 static void cpu_clock_event_stop(struct perf_event
*event
, int flags
)
6070 perf_swevent_cancel_hrtimer(event
);
6071 cpu_clock_event_update(event
);
6074 static int cpu_clock_event_add(struct perf_event
*event
, int flags
)
6076 if (flags
& PERF_EF_START
)
6077 cpu_clock_event_start(event
, flags
);
6082 static void cpu_clock_event_del(struct perf_event
*event
, int flags
)
6084 cpu_clock_event_stop(event
, flags
);
6087 static void cpu_clock_event_read(struct perf_event
*event
)
6089 cpu_clock_event_update(event
);
6092 static int cpu_clock_event_init(struct perf_event
*event
)
6094 if (event
->attr
.type
!= PERF_TYPE_SOFTWARE
)
6097 if (event
->attr
.config
!= PERF_COUNT_SW_CPU_CLOCK
)
6101 * no branch sampling for software events
6103 if (has_branch_stack(event
))
6106 perf_swevent_init_hrtimer(event
);
6111 static struct pmu perf_cpu_clock
= {
6112 .task_ctx_nr
= perf_sw_context
,
6114 .event_init
= cpu_clock_event_init
,
6115 .add
= cpu_clock_event_add
,
6116 .del
= cpu_clock_event_del
,
6117 .start
= cpu_clock_event_start
,
6118 .stop
= cpu_clock_event_stop
,
6119 .read
= cpu_clock_event_read
,
6121 .event_idx
= perf_swevent_event_idx
,
6125 * Software event: task time clock
6128 static void task_clock_event_update(struct perf_event
*event
, u64 now
)
6133 prev
= local64_xchg(&event
->hw
.prev_count
, now
);
6135 local64_add(delta
, &event
->count
);
6138 static void task_clock_event_start(struct perf_event
*event
, int flags
)
6140 local64_set(&event
->hw
.prev_count
, event
->ctx
->time
);
6141 perf_swevent_start_hrtimer(event
);
6144 static void task_clock_event_stop(struct perf_event
*event
, int flags
)
6146 perf_swevent_cancel_hrtimer(event
);
6147 task_clock_event_update(event
, event
->ctx
->time
);
6150 static int task_clock_event_add(struct perf_event
*event
, int flags
)
6152 if (flags
& PERF_EF_START
)
6153 task_clock_event_start(event
, flags
);
6158 static void task_clock_event_del(struct perf_event
*event
, int flags
)
6160 task_clock_event_stop(event
, PERF_EF_UPDATE
);
6163 static void task_clock_event_read(struct perf_event
*event
)
6165 u64 now
= perf_clock();
6166 u64 delta
= now
- event
->ctx
->timestamp
;
6167 u64 time
= event
->ctx
->time
+ delta
;
6169 task_clock_event_update(event
, time
);
6172 static int task_clock_event_init(struct perf_event
*event
)
6174 if (event
->attr
.type
!= PERF_TYPE_SOFTWARE
)
6177 if (event
->attr
.config
!= PERF_COUNT_SW_TASK_CLOCK
)
6181 * no branch sampling for software events
6183 if (has_branch_stack(event
))
6186 perf_swevent_init_hrtimer(event
);
6191 static struct pmu perf_task_clock
= {
6192 .task_ctx_nr
= perf_sw_context
,
6194 .event_init
= task_clock_event_init
,
6195 .add
= task_clock_event_add
,
6196 .del
= task_clock_event_del
,
6197 .start
= task_clock_event_start
,
6198 .stop
= task_clock_event_stop
,
6199 .read
= task_clock_event_read
,
6201 .event_idx
= perf_swevent_event_idx
,
6204 static void perf_pmu_nop_void(struct pmu
*pmu
)
6208 static int perf_pmu_nop_int(struct pmu
*pmu
)
6213 static void perf_pmu_start_txn(struct pmu
*pmu
)
6215 perf_pmu_disable(pmu
);
6218 static int perf_pmu_commit_txn(struct pmu
*pmu
)
6220 perf_pmu_enable(pmu
);
6224 static void perf_pmu_cancel_txn(struct pmu
*pmu
)
6226 perf_pmu_enable(pmu
);
6229 static int perf_event_idx_default(struct perf_event
*event
)
6231 return event
->hw
.idx
+ 1;
6235 * Ensures all contexts with the same task_ctx_nr have the same
6236 * pmu_cpu_context too.
6238 static void *find_pmu_context(int ctxn
)
6245 list_for_each_entry(pmu
, &pmus
, entry
) {
6246 if (pmu
->task_ctx_nr
== ctxn
)
6247 return pmu
->pmu_cpu_context
;
6253 static void update_pmu_context(struct pmu
*pmu
, struct pmu
*old_pmu
)
6257 for_each_possible_cpu(cpu
) {
6258 struct perf_cpu_context
*cpuctx
;
6260 cpuctx
= per_cpu_ptr(pmu
->pmu_cpu_context
, cpu
);
6262 if (cpuctx
->unique_pmu
== old_pmu
)
6263 cpuctx
->unique_pmu
= pmu
;
6267 static void free_pmu_context(struct pmu
*pmu
)
6271 mutex_lock(&pmus_lock
);
6273 * Like a real lame refcount.
6275 list_for_each_entry(i
, &pmus
, entry
) {
6276 if (i
->pmu_cpu_context
== pmu
->pmu_cpu_context
) {
6277 update_pmu_context(i
, pmu
);
6282 free_percpu(pmu
->pmu_cpu_context
);
6284 mutex_unlock(&pmus_lock
);
6286 static struct idr pmu_idr
;
6289 type_show(struct device
*dev
, struct device_attribute
*attr
, char *page
)
6291 struct pmu
*pmu
= dev_get_drvdata(dev
);
6293 return snprintf(page
, PAGE_SIZE
-1, "%d\n", pmu
->type
);
6297 perf_event_mux_interval_ms_show(struct device
*dev
,
6298 struct device_attribute
*attr
,
6301 struct pmu
*pmu
= dev_get_drvdata(dev
);
6303 return snprintf(page
, PAGE_SIZE
-1, "%d\n", pmu
->hrtimer_interval_ms
);
6307 perf_event_mux_interval_ms_store(struct device
*dev
,
6308 struct device_attribute
*attr
,
6309 const char *buf
, size_t count
)
6311 struct pmu
*pmu
= dev_get_drvdata(dev
);
6312 int timer
, cpu
, ret
;
6314 ret
= kstrtoint(buf
, 0, &timer
);
6321 /* same value, noting to do */
6322 if (timer
== pmu
->hrtimer_interval_ms
)
6325 pmu
->hrtimer_interval_ms
= timer
;
6327 /* update all cpuctx for this PMU */
6328 for_each_possible_cpu(cpu
) {
6329 struct perf_cpu_context
*cpuctx
;
6330 cpuctx
= per_cpu_ptr(pmu
->pmu_cpu_context
, cpu
);
6331 cpuctx
->hrtimer_interval
= ns_to_ktime(NSEC_PER_MSEC
* timer
);
6333 if (hrtimer_active(&cpuctx
->hrtimer
))
6334 hrtimer_forward_now(&cpuctx
->hrtimer
, cpuctx
->hrtimer_interval
);
6340 static struct device_attribute pmu_dev_attrs
[] = {
6342 __ATTR_RW(perf_event_mux_interval_ms
),
6346 static int pmu_bus_running
;
6347 static struct bus_type pmu_bus
= {
6348 .name
= "event_source",
6349 .dev_attrs
= pmu_dev_attrs
,
6352 static void pmu_dev_release(struct device
*dev
)
6357 static int pmu_dev_alloc(struct pmu
*pmu
)
6361 pmu
->dev
= kzalloc(sizeof(struct device
), GFP_KERNEL
);
6365 pmu
->dev
->groups
= pmu
->attr_groups
;
6366 device_initialize(pmu
->dev
);
6367 ret
= dev_set_name(pmu
->dev
, "%s", pmu
->name
);
6371 dev_set_drvdata(pmu
->dev
, pmu
);
6372 pmu
->dev
->bus
= &pmu_bus
;
6373 pmu
->dev
->release
= pmu_dev_release
;
6374 ret
= device_add(pmu
->dev
);
6382 put_device(pmu
->dev
);
6386 static struct lock_class_key cpuctx_mutex
;
6387 static struct lock_class_key cpuctx_lock
;
6389 int perf_pmu_register(struct pmu
*pmu
, const char *name
, int type
)
6393 mutex_lock(&pmus_lock
);
6395 pmu
->pmu_disable_count
= alloc_percpu(int);
6396 if (!pmu
->pmu_disable_count
)
6405 type
= idr_alloc(&pmu_idr
, pmu
, PERF_TYPE_MAX
, 0, GFP_KERNEL
);
6413 if (pmu_bus_running
) {
6414 ret
= pmu_dev_alloc(pmu
);
6420 pmu
->pmu_cpu_context
= find_pmu_context(pmu
->task_ctx_nr
);
6421 if (pmu
->pmu_cpu_context
)
6422 goto got_cpu_context
;
6425 pmu
->pmu_cpu_context
= alloc_percpu(struct perf_cpu_context
);
6426 if (!pmu
->pmu_cpu_context
)
6429 for_each_possible_cpu(cpu
) {
6430 struct perf_cpu_context
*cpuctx
;
6432 cpuctx
= per_cpu_ptr(pmu
->pmu_cpu_context
, cpu
);
6433 __perf_event_init_context(&cpuctx
->ctx
);
6434 lockdep_set_class(&cpuctx
->ctx
.mutex
, &cpuctx_mutex
);
6435 lockdep_set_class(&cpuctx
->ctx
.lock
, &cpuctx_lock
);
6436 cpuctx
->ctx
.type
= cpu_context
;
6437 cpuctx
->ctx
.pmu
= pmu
;
6439 __perf_cpu_hrtimer_init(cpuctx
, cpu
);
6441 INIT_LIST_HEAD(&cpuctx
->rotation_list
);
6442 cpuctx
->unique_pmu
= pmu
;
6446 if (!pmu
->start_txn
) {
6447 if (pmu
->pmu_enable
) {
6449 * If we have pmu_enable/pmu_disable calls, install
6450 * transaction stubs that use that to try and batch
6451 * hardware accesses.
6453 pmu
->start_txn
= perf_pmu_start_txn
;
6454 pmu
->commit_txn
= perf_pmu_commit_txn
;
6455 pmu
->cancel_txn
= perf_pmu_cancel_txn
;
6457 pmu
->start_txn
= perf_pmu_nop_void
;
6458 pmu
->commit_txn
= perf_pmu_nop_int
;
6459 pmu
->cancel_txn
= perf_pmu_nop_void
;
6463 if (!pmu
->pmu_enable
) {
6464 pmu
->pmu_enable
= perf_pmu_nop_void
;
6465 pmu
->pmu_disable
= perf_pmu_nop_void
;
6468 if (!pmu
->event_idx
)
6469 pmu
->event_idx
= perf_event_idx_default
;
6471 list_add_rcu(&pmu
->entry
, &pmus
);
6474 mutex_unlock(&pmus_lock
);
6479 device_del(pmu
->dev
);
6480 put_device(pmu
->dev
);
6483 if (pmu
->type
>= PERF_TYPE_MAX
)
6484 idr_remove(&pmu_idr
, pmu
->type
);
6487 free_percpu(pmu
->pmu_disable_count
);
6491 void perf_pmu_unregister(struct pmu
*pmu
)
6493 mutex_lock(&pmus_lock
);
6494 list_del_rcu(&pmu
->entry
);
6495 mutex_unlock(&pmus_lock
);
6498 * We dereference the pmu list under both SRCU and regular RCU, so
6499 * synchronize against both of those.
6501 synchronize_srcu(&pmus_srcu
);
6504 free_percpu(pmu
->pmu_disable_count
);
6505 if (pmu
->type
>= PERF_TYPE_MAX
)
6506 idr_remove(&pmu_idr
, pmu
->type
);
6507 device_del(pmu
->dev
);
6508 put_device(pmu
->dev
);
6509 free_pmu_context(pmu
);
6512 struct pmu
*perf_init_event(struct perf_event
*event
)
6514 struct pmu
*pmu
= NULL
;
6518 idx
= srcu_read_lock(&pmus_srcu
);
6521 pmu
= idr_find(&pmu_idr
, event
->attr
.type
);
6525 ret
= pmu
->event_init(event
);
6531 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
6533 ret
= pmu
->event_init(event
);
6537 if (ret
!= -ENOENT
) {
6542 pmu
= ERR_PTR(-ENOENT
);
6544 srcu_read_unlock(&pmus_srcu
, idx
);
6549 static void account_event_cpu(struct perf_event
*event
, int cpu
)
6554 if (has_branch_stack(event
)) {
6555 if (!(event
->attach_state
& PERF_ATTACH_TASK
))
6556 atomic_inc(&per_cpu(perf_branch_stack_events
, cpu
));
6558 if (is_cgroup_event(event
))
6559 atomic_inc(&per_cpu(perf_cgroup_events
, cpu
));
6562 static void account_event(struct perf_event
*event
)
6567 if (event
->attach_state
& PERF_ATTACH_TASK
)
6568 static_key_slow_inc(&perf_sched_events
.key
);
6569 if (event
->attr
.mmap
|| event
->attr
.mmap_data
)
6570 atomic_inc(&nr_mmap_events
);
6571 if (event
->attr
.comm
)
6572 atomic_inc(&nr_comm_events
);
6573 if (event
->attr
.task
)
6574 atomic_inc(&nr_task_events
);
6575 if (event
->attr
.freq
) {
6576 if (atomic_inc_return(&nr_freq_events
) == 1)
6577 tick_nohz_full_kick_all();
6579 if (has_branch_stack(event
))
6580 static_key_slow_inc(&perf_sched_events
.key
);
6581 if (is_cgroup_event(event
))
6582 static_key_slow_inc(&perf_sched_events
.key
);
6584 account_event_cpu(event
, event
->cpu
);
6588 * Allocate and initialize a event structure
6590 static struct perf_event
*
6591 perf_event_alloc(struct perf_event_attr
*attr
, int cpu
,
6592 struct task_struct
*task
,
6593 struct perf_event
*group_leader
,
6594 struct perf_event
*parent_event
,
6595 perf_overflow_handler_t overflow_handler
,
6599 struct perf_event
*event
;
6600 struct hw_perf_event
*hwc
;
6603 if ((unsigned)cpu
>= nr_cpu_ids
) {
6604 if (!task
|| cpu
!= -1)
6605 return ERR_PTR(-EINVAL
);
6608 event
= kzalloc(sizeof(*event
), GFP_KERNEL
);
6610 return ERR_PTR(-ENOMEM
);
6613 * Single events are their own group leaders, with an
6614 * empty sibling list:
6617 group_leader
= event
;
6619 mutex_init(&event
->child_mutex
);
6620 INIT_LIST_HEAD(&event
->child_list
);
6622 INIT_LIST_HEAD(&event
->group_entry
);
6623 INIT_LIST_HEAD(&event
->event_entry
);
6624 INIT_LIST_HEAD(&event
->sibling_list
);
6625 INIT_LIST_HEAD(&event
->rb_entry
);
6627 init_waitqueue_head(&event
->waitq
);
6628 init_irq_work(&event
->pending
, perf_pending_event
);
6630 mutex_init(&event
->mmap_mutex
);
6632 atomic_long_set(&event
->refcount
, 1);
6634 event
->attr
= *attr
;
6635 event
->group_leader
= group_leader
;
6639 event
->parent
= parent_event
;
6641 event
->ns
= get_pid_ns(task_active_pid_ns(current
));
6642 event
->id
= atomic64_inc_return(&perf_event_id
);
6644 event
->state
= PERF_EVENT_STATE_INACTIVE
;
6647 event
->attach_state
= PERF_ATTACH_TASK
;
6649 if (attr
->type
== PERF_TYPE_TRACEPOINT
)
6650 event
->hw
.tp_target
= task
;
6651 #ifdef CONFIG_HAVE_HW_BREAKPOINT
6653 * hw_breakpoint is a bit difficult here..
6655 else if (attr
->type
== PERF_TYPE_BREAKPOINT
)
6656 event
->hw
.bp_target
= task
;
6660 if (!overflow_handler
&& parent_event
) {
6661 overflow_handler
= parent_event
->overflow_handler
;
6662 context
= parent_event
->overflow_handler_context
;
6665 event
->overflow_handler
= overflow_handler
;
6666 event
->overflow_handler_context
= context
;
6668 perf_event__state_init(event
);
6673 hwc
->sample_period
= attr
->sample_period
;
6674 if (attr
->freq
&& attr
->sample_freq
)
6675 hwc
->sample_period
= 1;
6676 hwc
->last_period
= hwc
->sample_period
;
6678 local64_set(&hwc
->period_left
, hwc
->sample_period
);
6681 * we currently do not support PERF_FORMAT_GROUP on inherited events
6683 if (attr
->inherit
&& (attr
->read_format
& PERF_FORMAT_GROUP
))
6686 pmu
= perf_init_event(event
);
6689 else if (IS_ERR(pmu
)) {
6694 if (!event
->parent
) {
6695 if (event
->attr
.sample_type
& PERF_SAMPLE_CALLCHAIN
) {
6696 err
= get_callchain_buffers();
6706 event
->destroy(event
);
6709 put_pid_ns(event
->ns
);
6712 return ERR_PTR(err
);
6715 static int perf_copy_attr(struct perf_event_attr __user
*uattr
,
6716 struct perf_event_attr
*attr
)
6721 if (!access_ok(VERIFY_WRITE
, uattr
, PERF_ATTR_SIZE_VER0
))
6725 * zero the full structure, so that a short copy will be nice.
6727 memset(attr
, 0, sizeof(*attr
));
6729 ret
= get_user(size
, &uattr
->size
);
6733 if (size
> PAGE_SIZE
) /* silly large */
6736 if (!size
) /* abi compat */
6737 size
= PERF_ATTR_SIZE_VER0
;
6739 if (size
< PERF_ATTR_SIZE_VER0
)
6743 * If we're handed a bigger struct than we know of,
6744 * ensure all the unknown bits are 0 - i.e. new
6745 * user-space does not rely on any kernel feature
6746 * extensions we dont know about yet.
6748 if (size
> sizeof(*attr
)) {
6749 unsigned char __user
*addr
;
6750 unsigned char __user
*end
;
6753 addr
= (void __user
*)uattr
+ sizeof(*attr
);
6754 end
= (void __user
*)uattr
+ size
;
6756 for (; addr
< end
; addr
++) {
6757 ret
= get_user(val
, addr
);
6763 size
= sizeof(*attr
);
6766 ret
= copy_from_user(attr
, uattr
, size
);
6770 if (attr
->__reserved_1
)
6773 if (attr
->sample_type
& ~(PERF_SAMPLE_MAX
-1))
6776 if (attr
->read_format
& ~(PERF_FORMAT_MAX
-1))
6779 if (attr
->sample_type
& PERF_SAMPLE_BRANCH_STACK
) {
6780 u64 mask
= attr
->branch_sample_type
;
6782 /* only using defined bits */
6783 if (mask
& ~(PERF_SAMPLE_BRANCH_MAX
-1))
6786 /* at least one branch bit must be set */
6787 if (!(mask
& ~PERF_SAMPLE_BRANCH_PLM_ALL
))
6790 /* propagate priv level, when not set for branch */
6791 if (!(mask
& PERF_SAMPLE_BRANCH_PLM_ALL
)) {
6793 /* exclude_kernel checked on syscall entry */
6794 if (!attr
->exclude_kernel
)
6795 mask
|= PERF_SAMPLE_BRANCH_KERNEL
;
6797 if (!attr
->exclude_user
)
6798 mask
|= PERF_SAMPLE_BRANCH_USER
;
6800 if (!attr
->exclude_hv
)
6801 mask
|= PERF_SAMPLE_BRANCH_HV
;
6803 * adjust user setting (for HW filter setup)
6805 attr
->branch_sample_type
= mask
;
6807 /* privileged levels capture (kernel, hv): check permissions */
6808 if ((mask
& PERF_SAMPLE_BRANCH_PERM_PLM
)
6809 && perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN
))
6813 if (attr
->sample_type
& PERF_SAMPLE_REGS_USER
) {
6814 ret
= perf_reg_validate(attr
->sample_regs_user
);
6819 if (attr
->sample_type
& PERF_SAMPLE_STACK_USER
) {
6820 if (!arch_perf_have_user_stack_dump())
6824 * We have __u32 type for the size, but so far
6825 * we can only use __u16 as maximum due to the
6826 * __u16 sample size limit.
6828 if (attr
->sample_stack_user
>= USHRT_MAX
)
6830 else if (!IS_ALIGNED(attr
->sample_stack_user
, sizeof(u64
)))
6838 put_user(sizeof(*attr
), &uattr
->size
);
6844 perf_event_set_output(struct perf_event
*event
, struct perf_event
*output_event
)
6846 struct ring_buffer
*rb
= NULL
, *old_rb
= NULL
;
6852 /* don't allow circular references */
6853 if (event
== output_event
)
6857 * Don't allow cross-cpu buffers
6859 if (output_event
->cpu
!= event
->cpu
)
6863 * If its not a per-cpu rb, it must be the same task.
6865 if (output_event
->cpu
== -1 && output_event
->ctx
!= event
->ctx
)
6869 mutex_lock(&event
->mmap_mutex
);
6870 /* Can't redirect output if we've got an active mmap() */
6871 if (atomic_read(&event
->mmap_count
))
6877 /* get the rb we want to redirect to */
6878 rb
= ring_buffer_get(output_event
);
6884 ring_buffer_detach(event
, old_rb
);
6887 ring_buffer_attach(event
, rb
);
6889 rcu_assign_pointer(event
->rb
, rb
);
6892 ring_buffer_put(old_rb
);
6894 * Since we detached before setting the new rb, so that we
6895 * could attach the new rb, we could have missed a wakeup.
6898 wake_up_all(&event
->waitq
);
6903 mutex_unlock(&event
->mmap_mutex
);
6910 * sys_perf_event_open - open a performance event, associate it to a task/cpu
6912 * @attr_uptr: event_id type attributes for monitoring/sampling
6915 * @group_fd: group leader event fd
6917 SYSCALL_DEFINE5(perf_event_open
,
6918 struct perf_event_attr __user
*, attr_uptr
,
6919 pid_t
, pid
, int, cpu
, int, group_fd
, unsigned long, flags
)
6921 struct perf_event
*group_leader
= NULL
, *output_event
= NULL
;
6922 struct perf_event
*event
, *sibling
;
6923 struct perf_event_attr attr
;
6924 struct perf_event_context
*ctx
;
6925 struct file
*event_file
= NULL
;
6926 struct fd group
= {NULL
, 0};
6927 struct task_struct
*task
= NULL
;
6933 /* for future expandability... */
6934 if (flags
& ~PERF_FLAG_ALL
)
6937 err
= perf_copy_attr(attr_uptr
, &attr
);
6941 if (!attr
.exclude_kernel
) {
6942 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN
))
6947 if (attr
.sample_freq
> sysctl_perf_event_sample_rate
)
6952 * In cgroup mode, the pid argument is used to pass the fd
6953 * opened to the cgroup directory in cgroupfs. The cpu argument
6954 * designates the cpu on which to monitor threads from that
6957 if ((flags
& PERF_FLAG_PID_CGROUP
) && (pid
== -1 || cpu
== -1))
6960 event_fd
= get_unused_fd();
6964 if (group_fd
!= -1) {
6965 err
= perf_fget_light(group_fd
, &group
);
6968 group_leader
= group
.file
->private_data
;
6969 if (flags
& PERF_FLAG_FD_OUTPUT
)
6970 output_event
= group_leader
;
6971 if (flags
& PERF_FLAG_FD_NO_GROUP
)
6972 group_leader
= NULL
;
6975 if (pid
!= -1 && !(flags
& PERF_FLAG_PID_CGROUP
)) {
6976 task
= find_lively_task_by_vpid(pid
);
6978 err
= PTR_ERR(task
);
6985 event
= perf_event_alloc(&attr
, cpu
, task
, group_leader
, NULL
,
6987 if (IS_ERR(event
)) {
6988 err
= PTR_ERR(event
);
6992 if (flags
& PERF_FLAG_PID_CGROUP
) {
6993 err
= perf_cgroup_connect(pid
, event
, &attr
, group_leader
);
6995 __free_event(event
);
7000 account_event(event
);
7003 * Special case software events and allow them to be part of
7004 * any hardware group.
7009 (is_software_event(event
) != is_software_event(group_leader
))) {
7010 if (is_software_event(event
)) {
7012 * If event and group_leader are not both a software
7013 * event, and event is, then group leader is not.
7015 * Allow the addition of software events to !software
7016 * groups, this is safe because software events never
7019 pmu
= group_leader
->pmu
;
7020 } else if (is_software_event(group_leader
) &&
7021 (group_leader
->group_flags
& PERF_GROUP_SOFTWARE
)) {
7023 * In case the group is a pure software group, and we
7024 * try to add a hardware event, move the whole group to
7025 * the hardware context.
7032 * Get the target context (task or percpu):
7034 ctx
= find_get_context(pmu
, task
, event
->cpu
);
7041 put_task_struct(task
);
7046 * Look up the group leader (we will attach this event to it):
7052 * Do not allow a recursive hierarchy (this new sibling
7053 * becoming part of another group-sibling):
7055 if (group_leader
->group_leader
!= group_leader
)
7058 * Do not allow to attach to a group in a different
7059 * task or CPU context:
7062 if (group_leader
->ctx
->type
!= ctx
->type
)
7065 if (group_leader
->ctx
!= ctx
)
7070 * Only a group leader can be exclusive or pinned
7072 if (attr
.exclusive
|| attr
.pinned
)
7077 err
= perf_event_set_output(event
, output_event
);
7082 event_file
= anon_inode_getfile("[perf_event]", &perf_fops
, event
, O_RDWR
);
7083 if (IS_ERR(event_file
)) {
7084 err
= PTR_ERR(event_file
);
7089 struct perf_event_context
*gctx
= group_leader
->ctx
;
7091 mutex_lock(&gctx
->mutex
);
7092 perf_remove_from_context(group_leader
);
7095 * Removing from the context ends up with disabled
7096 * event. What we want here is event in the initial
7097 * startup state, ready to be add into new context.
7099 perf_event__state_init(group_leader
);
7100 list_for_each_entry(sibling
, &group_leader
->sibling_list
,
7102 perf_remove_from_context(sibling
);
7103 perf_event__state_init(sibling
);
7106 mutex_unlock(&gctx
->mutex
);
7110 WARN_ON_ONCE(ctx
->parent_ctx
);
7111 mutex_lock(&ctx
->mutex
);
7115 perf_install_in_context(ctx
, group_leader
, event
->cpu
);
7117 list_for_each_entry(sibling
, &group_leader
->sibling_list
,
7119 perf_install_in_context(ctx
, sibling
, event
->cpu
);
7124 perf_install_in_context(ctx
, event
, event
->cpu
);
7126 perf_unpin_context(ctx
);
7127 mutex_unlock(&ctx
->mutex
);
7131 event
->owner
= current
;
7133 mutex_lock(¤t
->perf_event_mutex
);
7134 list_add_tail(&event
->owner_entry
, ¤t
->perf_event_list
);
7135 mutex_unlock(¤t
->perf_event_mutex
);
7138 * Precalculate sample_data sizes
7140 perf_event__header_size(event
);
7141 perf_event__id_header_size(event
);
7144 * Drop the reference on the group_event after placing the
7145 * new event on the sibling_list. This ensures destruction
7146 * of the group leader will find the pointer to itself in
7147 * perf_group_detach().
7150 fd_install(event_fd
, event_file
);
7154 perf_unpin_context(ctx
);
7161 put_task_struct(task
);
7165 put_unused_fd(event_fd
);
7170 * perf_event_create_kernel_counter
7172 * @attr: attributes of the counter to create
7173 * @cpu: cpu in which the counter is bound
7174 * @task: task to profile (NULL for percpu)
7177 perf_event_create_kernel_counter(struct perf_event_attr
*attr
, int cpu
,
7178 struct task_struct
*task
,
7179 perf_overflow_handler_t overflow_handler
,
7182 struct perf_event_context
*ctx
;
7183 struct perf_event
*event
;
7187 * Get the target context (task or percpu):
7190 event
= perf_event_alloc(attr
, cpu
, task
, NULL
, NULL
,
7191 overflow_handler
, context
);
7192 if (IS_ERR(event
)) {
7193 err
= PTR_ERR(event
);
7197 account_event(event
);
7199 ctx
= find_get_context(event
->pmu
, task
, cpu
);
7205 WARN_ON_ONCE(ctx
->parent_ctx
);
7206 mutex_lock(&ctx
->mutex
);
7207 perf_install_in_context(ctx
, event
, cpu
);
7209 perf_unpin_context(ctx
);
7210 mutex_unlock(&ctx
->mutex
);
7217 return ERR_PTR(err
);
7219 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter
);
7221 void perf_pmu_migrate_context(struct pmu
*pmu
, int src_cpu
, int dst_cpu
)
7223 struct perf_event_context
*src_ctx
;
7224 struct perf_event_context
*dst_ctx
;
7225 struct perf_event
*event
, *tmp
;
7228 src_ctx
= &per_cpu_ptr(pmu
->pmu_cpu_context
, src_cpu
)->ctx
;
7229 dst_ctx
= &per_cpu_ptr(pmu
->pmu_cpu_context
, dst_cpu
)->ctx
;
7231 mutex_lock(&src_ctx
->mutex
);
7232 list_for_each_entry_safe(event
, tmp
, &src_ctx
->event_list
,
7234 perf_remove_from_context(event
);
7235 unaccount_event_cpu(event
, src_cpu
);
7237 list_add(&event
->event_entry
, &events
);
7239 mutex_unlock(&src_ctx
->mutex
);
7243 mutex_lock(&dst_ctx
->mutex
);
7244 list_for_each_entry_safe(event
, tmp
, &events
, event_entry
) {
7245 list_del(&event
->event_entry
);
7246 if (event
->state
>= PERF_EVENT_STATE_OFF
)
7247 event
->state
= PERF_EVENT_STATE_INACTIVE
;
7248 account_event_cpu(event
, dst_cpu
);
7249 perf_install_in_context(dst_ctx
, event
, dst_cpu
);
7252 mutex_unlock(&dst_ctx
->mutex
);
7254 EXPORT_SYMBOL_GPL(perf_pmu_migrate_context
);
7256 static void sync_child_event(struct perf_event
*child_event
,
7257 struct task_struct
*child
)
7259 struct perf_event
*parent_event
= child_event
->parent
;
7262 if (child_event
->attr
.inherit_stat
)
7263 perf_event_read_event(child_event
, child
);
7265 child_val
= perf_event_count(child_event
);
7268 * Add back the child's count to the parent's count:
7270 atomic64_add(child_val
, &parent_event
->child_count
);
7271 atomic64_add(child_event
->total_time_enabled
,
7272 &parent_event
->child_total_time_enabled
);
7273 atomic64_add(child_event
->total_time_running
,
7274 &parent_event
->child_total_time_running
);
7277 * Remove this event from the parent's list
7279 WARN_ON_ONCE(parent_event
->ctx
->parent_ctx
);
7280 mutex_lock(&parent_event
->child_mutex
);
7281 list_del_init(&child_event
->child_list
);
7282 mutex_unlock(&parent_event
->child_mutex
);
7285 * Release the parent event, if this was the last
7288 put_event(parent_event
);
7292 __perf_event_exit_task(struct perf_event
*child_event
,
7293 struct perf_event_context
*child_ctx
,
7294 struct task_struct
*child
)
7296 if (child_event
->parent
) {
7297 raw_spin_lock_irq(&child_ctx
->lock
);
7298 perf_group_detach(child_event
);
7299 raw_spin_unlock_irq(&child_ctx
->lock
);
7302 perf_remove_from_context(child_event
);
7305 * It can happen that the parent exits first, and has events
7306 * that are still around due to the child reference. These
7307 * events need to be zapped.
7309 if (child_event
->parent
) {
7310 sync_child_event(child_event
, child
);
7311 free_event(child_event
);
7315 static void perf_event_exit_task_context(struct task_struct
*child
, int ctxn
)
7317 struct perf_event
*child_event
, *tmp
;
7318 struct perf_event_context
*child_ctx
;
7319 unsigned long flags
;
7321 if (likely(!child
->perf_event_ctxp
[ctxn
])) {
7322 perf_event_task(child
, NULL
, 0);
7326 local_irq_save(flags
);
7328 * We can't reschedule here because interrupts are disabled,
7329 * and either child is current or it is a task that can't be
7330 * scheduled, so we are now safe from rescheduling changing
7333 child_ctx
= rcu_dereference_raw(child
->perf_event_ctxp
[ctxn
]);
7336 * Take the context lock here so that if find_get_context is
7337 * reading child->perf_event_ctxp, we wait until it has
7338 * incremented the context's refcount before we do put_ctx below.
7340 raw_spin_lock(&child_ctx
->lock
);
7341 task_ctx_sched_out(child_ctx
);
7342 child
->perf_event_ctxp
[ctxn
] = NULL
;
7344 * If this context is a clone; unclone it so it can't get
7345 * swapped to another process while we're removing all
7346 * the events from it.
7348 unclone_ctx(child_ctx
);
7349 update_context_time(child_ctx
);
7350 raw_spin_unlock_irqrestore(&child_ctx
->lock
, flags
);
7353 * Report the task dead after unscheduling the events so that we
7354 * won't get any samples after PERF_RECORD_EXIT. We can however still
7355 * get a few PERF_RECORD_READ events.
7357 perf_event_task(child
, child_ctx
, 0);
7360 * We can recurse on the same lock type through:
7362 * __perf_event_exit_task()
7363 * sync_child_event()
7365 * mutex_lock(&ctx->mutex)
7367 * But since its the parent context it won't be the same instance.
7369 mutex_lock(&child_ctx
->mutex
);
7372 list_for_each_entry_safe(child_event
, tmp
, &child_ctx
->pinned_groups
,
7374 __perf_event_exit_task(child_event
, child_ctx
, child
);
7376 list_for_each_entry_safe(child_event
, tmp
, &child_ctx
->flexible_groups
,
7378 __perf_event_exit_task(child_event
, child_ctx
, child
);
7381 * If the last event was a group event, it will have appended all
7382 * its siblings to the list, but we obtained 'tmp' before that which
7383 * will still point to the list head terminating the iteration.
7385 if (!list_empty(&child_ctx
->pinned_groups
) ||
7386 !list_empty(&child_ctx
->flexible_groups
))
7389 mutex_unlock(&child_ctx
->mutex
);
7395 * When a child task exits, feed back event values to parent events.
7397 void perf_event_exit_task(struct task_struct
*child
)
7399 struct perf_event
*event
, *tmp
;
7402 mutex_lock(&child
->perf_event_mutex
);
7403 list_for_each_entry_safe(event
, tmp
, &child
->perf_event_list
,
7405 list_del_init(&event
->owner_entry
);
7408 * Ensure the list deletion is visible before we clear
7409 * the owner, closes a race against perf_release() where
7410 * we need to serialize on the owner->perf_event_mutex.
7413 event
->owner
= NULL
;
7415 mutex_unlock(&child
->perf_event_mutex
);
7417 for_each_task_context_nr(ctxn
)
7418 perf_event_exit_task_context(child
, ctxn
);
7421 static void perf_free_event(struct perf_event
*event
,
7422 struct perf_event_context
*ctx
)
7424 struct perf_event
*parent
= event
->parent
;
7426 if (WARN_ON_ONCE(!parent
))
7429 mutex_lock(&parent
->child_mutex
);
7430 list_del_init(&event
->child_list
);
7431 mutex_unlock(&parent
->child_mutex
);
7435 perf_group_detach(event
);
7436 list_del_event(event
, ctx
);
7441 * free an unexposed, unused context as created by inheritance by
7442 * perf_event_init_task below, used by fork() in case of fail.
7444 void perf_event_free_task(struct task_struct
*task
)
7446 struct perf_event_context
*ctx
;
7447 struct perf_event
*event
, *tmp
;
7450 for_each_task_context_nr(ctxn
) {
7451 ctx
= task
->perf_event_ctxp
[ctxn
];
7455 mutex_lock(&ctx
->mutex
);
7457 list_for_each_entry_safe(event
, tmp
, &ctx
->pinned_groups
,
7459 perf_free_event(event
, ctx
);
7461 list_for_each_entry_safe(event
, tmp
, &ctx
->flexible_groups
,
7463 perf_free_event(event
, ctx
);
7465 if (!list_empty(&ctx
->pinned_groups
) ||
7466 !list_empty(&ctx
->flexible_groups
))
7469 mutex_unlock(&ctx
->mutex
);
7475 void perf_event_delayed_put(struct task_struct
*task
)
7479 for_each_task_context_nr(ctxn
)
7480 WARN_ON_ONCE(task
->perf_event_ctxp
[ctxn
]);
7484 * inherit a event from parent task to child task:
7486 static struct perf_event
*
7487 inherit_event(struct perf_event
*parent_event
,
7488 struct task_struct
*parent
,
7489 struct perf_event_context
*parent_ctx
,
7490 struct task_struct
*child
,
7491 struct perf_event
*group_leader
,
7492 struct perf_event_context
*child_ctx
)
7494 struct perf_event
*child_event
;
7495 unsigned long flags
;
7498 * Instead of creating recursive hierarchies of events,
7499 * we link inherited events back to the original parent,
7500 * which has a filp for sure, which we use as the reference
7503 if (parent_event
->parent
)
7504 parent_event
= parent_event
->parent
;
7506 child_event
= perf_event_alloc(&parent_event
->attr
,
7509 group_leader
, parent_event
,
7511 if (IS_ERR(child_event
))
7514 if (!atomic_long_inc_not_zero(&parent_event
->refcount
)) {
7515 free_event(child_event
);
7522 * Make the child state follow the state of the parent event,
7523 * not its attr.disabled bit. We hold the parent's mutex,
7524 * so we won't race with perf_event_{en, dis}able_family.
7526 if (parent_event
->state
>= PERF_EVENT_STATE_INACTIVE
)
7527 child_event
->state
= PERF_EVENT_STATE_INACTIVE
;
7529 child_event
->state
= PERF_EVENT_STATE_OFF
;
7531 if (parent_event
->attr
.freq
) {
7532 u64 sample_period
= parent_event
->hw
.sample_period
;
7533 struct hw_perf_event
*hwc
= &child_event
->hw
;
7535 hwc
->sample_period
= sample_period
;
7536 hwc
->last_period
= sample_period
;
7538 local64_set(&hwc
->period_left
, sample_period
);
7541 child_event
->ctx
= child_ctx
;
7542 child_event
->overflow_handler
= parent_event
->overflow_handler
;
7543 child_event
->overflow_handler_context
7544 = parent_event
->overflow_handler_context
;
7547 * Precalculate sample_data sizes
7549 perf_event__header_size(child_event
);
7550 perf_event__id_header_size(child_event
);
7553 * Link it up in the child's context:
7555 raw_spin_lock_irqsave(&child_ctx
->lock
, flags
);
7556 add_event_to_ctx(child_event
, child_ctx
);
7557 raw_spin_unlock_irqrestore(&child_ctx
->lock
, flags
);
7560 * Link this into the parent event's child list
7562 WARN_ON_ONCE(parent_event
->ctx
->parent_ctx
);
7563 mutex_lock(&parent_event
->child_mutex
);
7564 list_add_tail(&child_event
->child_list
, &parent_event
->child_list
);
7565 mutex_unlock(&parent_event
->child_mutex
);
7570 static int inherit_group(struct perf_event
*parent_event
,
7571 struct task_struct
*parent
,
7572 struct perf_event_context
*parent_ctx
,
7573 struct task_struct
*child
,
7574 struct perf_event_context
*child_ctx
)
7576 struct perf_event
*leader
;
7577 struct perf_event
*sub
;
7578 struct perf_event
*child_ctr
;
7580 leader
= inherit_event(parent_event
, parent
, parent_ctx
,
7581 child
, NULL
, child_ctx
);
7583 return PTR_ERR(leader
);
7584 list_for_each_entry(sub
, &parent_event
->sibling_list
, group_entry
) {
7585 child_ctr
= inherit_event(sub
, parent
, parent_ctx
,
7586 child
, leader
, child_ctx
);
7587 if (IS_ERR(child_ctr
))
7588 return PTR_ERR(child_ctr
);
7594 inherit_task_group(struct perf_event
*event
, struct task_struct
*parent
,
7595 struct perf_event_context
*parent_ctx
,
7596 struct task_struct
*child
, int ctxn
,
7600 struct perf_event_context
*child_ctx
;
7602 if (!event
->attr
.inherit
) {
7607 child_ctx
= child
->perf_event_ctxp
[ctxn
];
7610 * This is executed from the parent task context, so
7611 * inherit events that have been marked for cloning.
7612 * First allocate and initialize a context for the
7616 child_ctx
= alloc_perf_context(parent_ctx
->pmu
, child
);
7620 child
->perf_event_ctxp
[ctxn
] = child_ctx
;
7623 ret
= inherit_group(event
, parent
, parent_ctx
,
7633 * Initialize the perf_event context in task_struct
7635 int perf_event_init_context(struct task_struct
*child
, int ctxn
)
7637 struct perf_event_context
*child_ctx
, *parent_ctx
;
7638 struct perf_event_context
*cloned_ctx
;
7639 struct perf_event
*event
;
7640 struct task_struct
*parent
= current
;
7641 int inherited_all
= 1;
7642 unsigned long flags
;
7645 if (likely(!parent
->perf_event_ctxp
[ctxn
]))
7649 * If the parent's context is a clone, pin it so it won't get
7652 parent_ctx
= perf_pin_task_context(parent
, ctxn
);
7655 * No need to check if parent_ctx != NULL here; since we saw
7656 * it non-NULL earlier, the only reason for it to become NULL
7657 * is if we exit, and since we're currently in the middle of
7658 * a fork we can't be exiting at the same time.
7662 * Lock the parent list. No need to lock the child - not PID
7663 * hashed yet and not running, so nobody can access it.
7665 mutex_lock(&parent_ctx
->mutex
);
7668 * We dont have to disable NMIs - we are only looking at
7669 * the list, not manipulating it:
7671 list_for_each_entry(event
, &parent_ctx
->pinned_groups
, group_entry
) {
7672 ret
= inherit_task_group(event
, parent
, parent_ctx
,
7673 child
, ctxn
, &inherited_all
);
7679 * We can't hold ctx->lock when iterating the ->flexible_group list due
7680 * to allocations, but we need to prevent rotation because
7681 * rotate_ctx() will change the list from interrupt context.
7683 raw_spin_lock_irqsave(&parent_ctx
->lock
, flags
);
7684 parent_ctx
->rotate_disable
= 1;
7685 raw_spin_unlock_irqrestore(&parent_ctx
->lock
, flags
);
7687 list_for_each_entry(event
, &parent_ctx
->flexible_groups
, group_entry
) {
7688 ret
= inherit_task_group(event
, parent
, parent_ctx
,
7689 child
, ctxn
, &inherited_all
);
7694 raw_spin_lock_irqsave(&parent_ctx
->lock
, flags
);
7695 parent_ctx
->rotate_disable
= 0;
7697 child_ctx
= child
->perf_event_ctxp
[ctxn
];
7699 if (child_ctx
&& inherited_all
) {
7701 * Mark the child context as a clone of the parent
7702 * context, or of whatever the parent is a clone of.
7704 * Note that if the parent is a clone, the holding of
7705 * parent_ctx->lock avoids it from being uncloned.
7707 cloned_ctx
= parent_ctx
->parent_ctx
;
7709 child_ctx
->parent_ctx
= cloned_ctx
;
7710 child_ctx
->parent_gen
= parent_ctx
->parent_gen
;
7712 child_ctx
->parent_ctx
= parent_ctx
;
7713 child_ctx
->parent_gen
= parent_ctx
->generation
;
7715 get_ctx(child_ctx
->parent_ctx
);
7718 raw_spin_unlock_irqrestore(&parent_ctx
->lock
, flags
);
7719 mutex_unlock(&parent_ctx
->mutex
);
7721 perf_unpin_context(parent_ctx
);
7722 put_ctx(parent_ctx
);
7728 * Initialize the perf_event context in task_struct
7730 int perf_event_init_task(struct task_struct
*child
)
7734 memset(child
->perf_event_ctxp
, 0, sizeof(child
->perf_event_ctxp
));
7735 mutex_init(&child
->perf_event_mutex
);
7736 INIT_LIST_HEAD(&child
->perf_event_list
);
7738 for_each_task_context_nr(ctxn
) {
7739 ret
= perf_event_init_context(child
, ctxn
);
7747 static void __init
perf_event_init_all_cpus(void)
7749 struct swevent_htable
*swhash
;
7752 for_each_possible_cpu(cpu
) {
7753 swhash
= &per_cpu(swevent_htable
, cpu
);
7754 mutex_init(&swhash
->hlist_mutex
);
7755 INIT_LIST_HEAD(&per_cpu(rotation_list
, cpu
));
7759 static void perf_event_init_cpu(int cpu
)
7761 struct swevent_htable
*swhash
= &per_cpu(swevent_htable
, cpu
);
7763 mutex_lock(&swhash
->hlist_mutex
);
7764 if (swhash
->hlist_refcount
> 0) {
7765 struct swevent_hlist
*hlist
;
7767 hlist
= kzalloc_node(sizeof(*hlist
), GFP_KERNEL
, cpu_to_node(cpu
));
7769 rcu_assign_pointer(swhash
->swevent_hlist
, hlist
);
7771 mutex_unlock(&swhash
->hlist_mutex
);
7774 #if defined CONFIG_HOTPLUG_CPU || defined CONFIG_KEXEC
7775 static void perf_pmu_rotate_stop(struct pmu
*pmu
)
7777 struct perf_cpu_context
*cpuctx
= this_cpu_ptr(pmu
->pmu_cpu_context
);
7779 WARN_ON(!irqs_disabled());
7781 list_del_init(&cpuctx
->rotation_list
);
7784 static void __perf_event_exit_context(void *__info
)
7786 struct perf_event_context
*ctx
= __info
;
7787 struct perf_event
*event
, *tmp
;
7789 perf_pmu_rotate_stop(ctx
->pmu
);
7791 list_for_each_entry_safe(event
, tmp
, &ctx
->pinned_groups
, group_entry
)
7792 __perf_remove_from_context(event
);
7793 list_for_each_entry_safe(event
, tmp
, &ctx
->flexible_groups
, group_entry
)
7794 __perf_remove_from_context(event
);
7797 static void perf_event_exit_cpu_context(int cpu
)
7799 struct perf_event_context
*ctx
;
7803 idx
= srcu_read_lock(&pmus_srcu
);
7804 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
7805 ctx
= &per_cpu_ptr(pmu
->pmu_cpu_context
, cpu
)->ctx
;
7807 mutex_lock(&ctx
->mutex
);
7808 smp_call_function_single(cpu
, __perf_event_exit_context
, ctx
, 1);
7809 mutex_unlock(&ctx
->mutex
);
7811 srcu_read_unlock(&pmus_srcu
, idx
);
7814 static void perf_event_exit_cpu(int cpu
)
7816 struct swevent_htable
*swhash
= &per_cpu(swevent_htable
, cpu
);
7818 mutex_lock(&swhash
->hlist_mutex
);
7819 swevent_hlist_release(swhash
);
7820 mutex_unlock(&swhash
->hlist_mutex
);
7822 perf_event_exit_cpu_context(cpu
);
7825 static inline void perf_event_exit_cpu(int cpu
) { }
7829 perf_reboot(struct notifier_block
*notifier
, unsigned long val
, void *v
)
7833 for_each_online_cpu(cpu
)
7834 perf_event_exit_cpu(cpu
);
7840 * Run the perf reboot notifier at the very last possible moment so that
7841 * the generic watchdog code runs as long as possible.
7843 static struct notifier_block perf_reboot_notifier
= {
7844 .notifier_call
= perf_reboot
,
7845 .priority
= INT_MIN
,
7849 perf_cpu_notify(struct notifier_block
*self
, unsigned long action
, void *hcpu
)
7851 unsigned int cpu
= (long)hcpu
;
7853 switch (action
& ~CPU_TASKS_FROZEN
) {
7855 case CPU_UP_PREPARE
:
7856 case CPU_DOWN_FAILED
:
7857 perf_event_init_cpu(cpu
);
7860 case CPU_UP_CANCELED
:
7861 case CPU_DOWN_PREPARE
:
7862 perf_event_exit_cpu(cpu
);
7871 void __init
perf_event_init(void)
7877 perf_event_init_all_cpus();
7878 init_srcu_struct(&pmus_srcu
);
7879 perf_pmu_register(&perf_swevent
, "software", PERF_TYPE_SOFTWARE
);
7880 perf_pmu_register(&perf_cpu_clock
, NULL
, -1);
7881 perf_pmu_register(&perf_task_clock
, NULL
, -1);
7883 perf_cpu_notifier(perf_cpu_notify
);
7884 register_reboot_notifier(&perf_reboot_notifier
);
7886 ret
= init_hw_breakpoint();
7887 WARN(ret
, "hw_breakpoint initialization failed with: %d", ret
);
7889 /* do not patch jump label more than once per second */
7890 jump_label_rate_limit(&perf_sched_events
, HZ
);
7893 * Build time assertion that we keep the data_head at the intended
7894 * location. IOW, validation we got the __reserved[] size right.
7896 BUILD_BUG_ON((offsetof(struct perf_event_mmap_page
, data_head
))
7900 static int __init
perf_event_sysfs_init(void)
7905 mutex_lock(&pmus_lock
);
7907 ret
= bus_register(&pmu_bus
);
7911 list_for_each_entry(pmu
, &pmus
, entry
) {
7912 if (!pmu
->name
|| pmu
->type
< 0)
7915 ret
= pmu_dev_alloc(pmu
);
7916 WARN(ret
, "Failed to register pmu: %s, reason %d\n", pmu
->name
, ret
);
7918 pmu_bus_running
= 1;
7922 mutex_unlock(&pmus_lock
);
7926 device_initcall(perf_event_sysfs_init
);
7928 #ifdef CONFIG_CGROUP_PERF
7929 static struct cgroup_subsys_state
*
7930 perf_cgroup_css_alloc(struct cgroup_subsys_state
*parent_css
)
7932 struct perf_cgroup
*jc
;
7934 jc
= kzalloc(sizeof(*jc
), GFP_KERNEL
);
7936 return ERR_PTR(-ENOMEM
);
7938 jc
->info
= alloc_percpu(struct perf_cgroup_info
);
7941 return ERR_PTR(-ENOMEM
);
7947 static void perf_cgroup_css_free(struct cgroup_subsys_state
*css
)
7949 struct perf_cgroup
*jc
= container_of(css
, struct perf_cgroup
, css
);
7951 free_percpu(jc
->info
);
7955 static int __perf_cgroup_move(void *info
)
7957 struct task_struct
*task
= info
;
7958 perf_cgroup_switch(task
, PERF_CGROUP_SWOUT
| PERF_CGROUP_SWIN
);
7962 static void perf_cgroup_attach(struct cgroup_subsys_state
*css
,
7963 struct cgroup_taskset
*tset
)
7965 struct task_struct
*task
;
7967 cgroup_taskset_for_each(task
, css
, tset
)
7968 task_function_call(task
, __perf_cgroup_move
, task
);
7971 static void perf_cgroup_exit(struct cgroup_subsys_state
*css
,
7972 struct cgroup_subsys_state
*old_css
,
7973 struct task_struct
*task
)
7976 * cgroup_exit() is called in the copy_process() failure path.
7977 * Ignore this case since the task hasn't ran yet, this avoids
7978 * trying to poke a half freed task state from generic code.
7980 if (!(task
->flags
& PF_EXITING
))
7983 task_function_call(task
, __perf_cgroup_move
, task
);
7986 struct cgroup_subsys perf_subsys
= {
7987 .name
= "perf_event",
7988 .subsys_id
= perf_subsys_id
,
7989 .css_alloc
= perf_cgroup_css_alloc
,
7990 .css_free
= perf_cgroup_css_free
,
7991 .exit
= perf_cgroup_exit
,
7992 .attach
= perf_cgroup_attach
,
7994 #endif /* CONFIG_CGROUP_PERF */