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 int perf_sample_allowed_ns __read_mostly
=
179 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 ACCESS_ONCE(perf_sample_allowed_ns
) = tmp
;
190 static int perf_rotate_context(struct perf_cpu_context
*cpuctx
);
192 int perf_proc_update_handler(struct ctl_table
*table
, int write
,
193 void __user
*buffer
, size_t *lenp
,
196 int ret
= proc_dointvec_minmax(table
, write
, buffer
, lenp
, ppos
);
201 max_samples_per_tick
= DIV_ROUND_UP(sysctl_perf_event_sample_rate
, HZ
);
202 perf_sample_period_ns
= NSEC_PER_SEC
/ sysctl_perf_event_sample_rate
;
203 update_perf_cpu_limits();
208 int sysctl_perf_cpu_time_max_percent __read_mostly
= DEFAULT_CPU_TIME_MAX_PERCENT
;
210 int perf_cpu_time_max_percent_handler(struct ctl_table
*table
, int write
,
211 void __user
*buffer
, size_t *lenp
,
214 int ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
219 update_perf_cpu_limits();
225 * perf samples are done in some very critical code paths (NMIs).
226 * If they take too much CPU time, the system can lock up and not
227 * get any real work done. This will drop the sample rate when
228 * we detect that events are taking too long.
230 #define NR_ACCUMULATED_SAMPLES 128
231 static 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
;
237 u64 allowed_ns
= ACCESS_ONCE(perf_sample_allowed_ns
);
242 /* decay the counter by 1 average sample */
243 local_samples_len
= __get_cpu_var(running_sample_length
);
244 local_samples_len
-= local_samples_len
/NR_ACCUMULATED_SAMPLES
;
245 local_samples_len
+= sample_len_ns
;
246 __get_cpu_var(running_sample_length
) = local_samples_len
;
249 * note: this will be biased artifically low until we have
250 * seen NR_ACCUMULATED_SAMPLES. Doing it this way keeps us
251 * from having to maintain a count.
253 avg_local_sample_len
= local_samples_len
/NR_ACCUMULATED_SAMPLES
;
255 if (avg_local_sample_len
<= allowed_ns
)
258 if (max_samples_per_tick
<= 1)
261 max_samples_per_tick
= DIV_ROUND_UP(max_samples_per_tick
, 2);
262 sysctl_perf_event_sample_rate
= max_samples_per_tick
* HZ
;
263 perf_sample_period_ns
= NSEC_PER_SEC
/ sysctl_perf_event_sample_rate
;
265 printk_ratelimited(KERN_WARNING
266 "perf samples too long (%lld > %lld), lowering "
267 "kernel.perf_event_max_sample_rate to %d\n",
268 avg_local_sample_len
, 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
;
905 static u32
perf_event_pid(struct perf_event
*event
, struct task_struct
*p
)
908 * only top level events have the pid namespace they were created in
911 event
= event
->parent
;
913 return task_tgid_nr_ns(p
, event
->ns
);
916 static u32
perf_event_tid(struct perf_event
*event
, struct task_struct
*p
)
919 * only top level events have the pid namespace they were created in
922 event
= event
->parent
;
924 return task_pid_nr_ns(p
, event
->ns
);
928 * If we inherit events we want to return the parent event id
931 static u64
primary_event_id(struct perf_event
*event
)
936 id
= event
->parent
->id
;
942 * Get the perf_event_context for a task and lock it.
943 * This has to cope with with the fact that until it is locked,
944 * the context could get moved to another task.
946 static struct perf_event_context
*
947 perf_lock_task_context(struct task_struct
*task
, int ctxn
, unsigned long *flags
)
949 struct perf_event_context
*ctx
;
953 * One of the few rules of preemptible RCU is that one cannot do
954 * rcu_read_unlock() while holding a scheduler (or nested) lock when
955 * part of the read side critical section was preemptible -- see
956 * rcu_read_unlock_special().
958 * Since ctx->lock nests under rq->lock we must ensure the entire read
959 * side critical section is non-preemptible.
963 ctx
= rcu_dereference(task
->perf_event_ctxp
[ctxn
]);
966 * If this context is a clone of another, it might
967 * get swapped for another underneath us by
968 * perf_event_task_sched_out, though the
969 * rcu_read_lock() protects us from any context
970 * getting freed. Lock the context and check if it
971 * got swapped before we could get the lock, and retry
972 * if so. If we locked the right context, then it
973 * can't get swapped on us any more.
975 raw_spin_lock_irqsave(&ctx
->lock
, *flags
);
976 if (ctx
!= rcu_dereference(task
->perf_event_ctxp
[ctxn
])) {
977 raw_spin_unlock_irqrestore(&ctx
->lock
, *flags
);
983 if (!atomic_inc_not_zero(&ctx
->refcount
)) {
984 raw_spin_unlock_irqrestore(&ctx
->lock
, *flags
);
994 * Get the context for a task and increment its pin_count so it
995 * can't get swapped to another task. This also increments its
996 * reference count so that the context can't get freed.
998 static struct perf_event_context
*
999 perf_pin_task_context(struct task_struct
*task
, int ctxn
)
1001 struct perf_event_context
*ctx
;
1002 unsigned long flags
;
1004 ctx
= perf_lock_task_context(task
, ctxn
, &flags
);
1007 raw_spin_unlock_irqrestore(&ctx
->lock
, flags
);
1012 static void perf_unpin_context(struct perf_event_context
*ctx
)
1014 unsigned long flags
;
1016 raw_spin_lock_irqsave(&ctx
->lock
, flags
);
1018 raw_spin_unlock_irqrestore(&ctx
->lock
, flags
);
1022 * Update the record of the current time in a context.
1024 static void update_context_time(struct perf_event_context
*ctx
)
1026 u64 now
= perf_clock();
1028 ctx
->time
+= now
- ctx
->timestamp
;
1029 ctx
->timestamp
= now
;
1032 static u64
perf_event_time(struct perf_event
*event
)
1034 struct perf_event_context
*ctx
= event
->ctx
;
1036 if (is_cgroup_event(event
))
1037 return perf_cgroup_event_time(event
);
1039 return ctx
? ctx
->time
: 0;
1043 * Update the total_time_enabled and total_time_running fields for a event.
1044 * The caller of this function needs to hold the ctx->lock.
1046 static void update_event_times(struct perf_event
*event
)
1048 struct perf_event_context
*ctx
= event
->ctx
;
1051 if (event
->state
< PERF_EVENT_STATE_INACTIVE
||
1052 event
->group_leader
->state
< PERF_EVENT_STATE_INACTIVE
)
1055 * in cgroup mode, time_enabled represents
1056 * the time the event was enabled AND active
1057 * tasks were in the monitored cgroup. This is
1058 * independent of the activity of the context as
1059 * there may be a mix of cgroup and non-cgroup events.
1061 * That is why we treat cgroup events differently
1064 if (is_cgroup_event(event
))
1065 run_end
= perf_cgroup_event_time(event
);
1066 else if (ctx
->is_active
)
1067 run_end
= ctx
->time
;
1069 run_end
= event
->tstamp_stopped
;
1071 event
->total_time_enabled
= run_end
- event
->tstamp_enabled
;
1073 if (event
->state
== PERF_EVENT_STATE_INACTIVE
)
1074 run_end
= event
->tstamp_stopped
;
1076 run_end
= perf_event_time(event
);
1078 event
->total_time_running
= run_end
- event
->tstamp_running
;
1083 * Update total_time_enabled and total_time_running for all events in a group.
1085 static void update_group_times(struct perf_event
*leader
)
1087 struct perf_event
*event
;
1089 update_event_times(leader
);
1090 list_for_each_entry(event
, &leader
->sibling_list
, group_entry
)
1091 update_event_times(event
);
1094 static struct list_head
*
1095 ctx_group_list(struct perf_event
*event
, struct perf_event_context
*ctx
)
1097 if (event
->attr
.pinned
)
1098 return &ctx
->pinned_groups
;
1100 return &ctx
->flexible_groups
;
1104 * Add a event from the lists for its context.
1105 * Must be called with ctx->mutex and ctx->lock held.
1108 list_add_event(struct perf_event
*event
, struct perf_event_context
*ctx
)
1110 WARN_ON_ONCE(event
->attach_state
& PERF_ATTACH_CONTEXT
);
1111 event
->attach_state
|= PERF_ATTACH_CONTEXT
;
1114 * If we're a stand alone event or group leader, we go to the context
1115 * list, group events are kept attached to the group so that
1116 * perf_group_detach can, at all times, locate all siblings.
1118 if (event
->group_leader
== event
) {
1119 struct list_head
*list
;
1121 if (is_software_event(event
))
1122 event
->group_flags
|= PERF_GROUP_SOFTWARE
;
1124 list
= ctx_group_list(event
, ctx
);
1125 list_add_tail(&event
->group_entry
, list
);
1128 if (is_cgroup_event(event
))
1131 if (has_branch_stack(event
))
1132 ctx
->nr_branch_stack
++;
1134 list_add_rcu(&event
->event_entry
, &ctx
->event_list
);
1135 if (!ctx
->nr_events
)
1136 perf_pmu_rotate_start(ctx
->pmu
);
1138 if (event
->attr
.inherit_stat
)
1145 * Initialize event state based on the perf_event_attr::disabled.
1147 static inline void perf_event__state_init(struct perf_event
*event
)
1149 event
->state
= event
->attr
.disabled
? PERF_EVENT_STATE_OFF
:
1150 PERF_EVENT_STATE_INACTIVE
;
1154 * Called at perf_event creation and when events are attached/detached from a
1157 static void perf_event__read_size(struct perf_event
*event
)
1159 int entry
= sizeof(u64
); /* value */
1163 if (event
->attr
.read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
1164 size
+= sizeof(u64
);
1166 if (event
->attr
.read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
1167 size
+= sizeof(u64
);
1169 if (event
->attr
.read_format
& PERF_FORMAT_ID
)
1170 entry
+= sizeof(u64
);
1172 if (event
->attr
.read_format
& PERF_FORMAT_GROUP
) {
1173 nr
+= event
->group_leader
->nr_siblings
;
1174 size
+= sizeof(u64
);
1178 event
->read_size
= size
;
1181 static void perf_event__header_size(struct perf_event
*event
)
1183 struct perf_sample_data
*data
;
1184 u64 sample_type
= event
->attr
.sample_type
;
1187 perf_event__read_size(event
);
1189 if (sample_type
& PERF_SAMPLE_IP
)
1190 size
+= sizeof(data
->ip
);
1192 if (sample_type
& PERF_SAMPLE_ADDR
)
1193 size
+= sizeof(data
->addr
);
1195 if (sample_type
& PERF_SAMPLE_PERIOD
)
1196 size
+= sizeof(data
->period
);
1198 if (sample_type
& PERF_SAMPLE_WEIGHT
)
1199 size
+= sizeof(data
->weight
);
1201 if (sample_type
& PERF_SAMPLE_READ
)
1202 size
+= event
->read_size
;
1204 if (sample_type
& PERF_SAMPLE_DATA_SRC
)
1205 size
+= sizeof(data
->data_src
.val
);
1207 if (sample_type
& PERF_SAMPLE_TRANSACTION
)
1208 size
+= sizeof(data
->txn
);
1210 event
->header_size
= size
;
1213 static void perf_event__id_header_size(struct perf_event
*event
)
1215 struct perf_sample_data
*data
;
1216 u64 sample_type
= event
->attr
.sample_type
;
1219 if (sample_type
& PERF_SAMPLE_TID
)
1220 size
+= sizeof(data
->tid_entry
);
1222 if (sample_type
& PERF_SAMPLE_TIME
)
1223 size
+= sizeof(data
->time
);
1225 if (sample_type
& PERF_SAMPLE_IDENTIFIER
)
1226 size
+= sizeof(data
->id
);
1228 if (sample_type
& PERF_SAMPLE_ID
)
1229 size
+= sizeof(data
->id
);
1231 if (sample_type
& PERF_SAMPLE_STREAM_ID
)
1232 size
+= sizeof(data
->stream_id
);
1234 if (sample_type
& PERF_SAMPLE_CPU
)
1235 size
+= sizeof(data
->cpu_entry
);
1237 event
->id_header_size
= size
;
1240 static void perf_group_attach(struct perf_event
*event
)
1242 struct perf_event
*group_leader
= event
->group_leader
, *pos
;
1245 * We can have double attach due to group movement in perf_event_open.
1247 if (event
->attach_state
& PERF_ATTACH_GROUP
)
1250 event
->attach_state
|= PERF_ATTACH_GROUP
;
1252 if (group_leader
== event
)
1255 if (group_leader
->group_flags
& PERF_GROUP_SOFTWARE
&&
1256 !is_software_event(event
))
1257 group_leader
->group_flags
&= ~PERF_GROUP_SOFTWARE
;
1259 list_add_tail(&event
->group_entry
, &group_leader
->sibling_list
);
1260 group_leader
->nr_siblings
++;
1262 perf_event__header_size(group_leader
);
1264 list_for_each_entry(pos
, &group_leader
->sibling_list
, group_entry
)
1265 perf_event__header_size(pos
);
1269 * Remove a event from the lists for its context.
1270 * Must be called with ctx->mutex and ctx->lock held.
1273 list_del_event(struct perf_event
*event
, struct perf_event_context
*ctx
)
1275 struct perf_cpu_context
*cpuctx
;
1277 * We can have double detach due to exit/hot-unplug + close.
1279 if (!(event
->attach_state
& PERF_ATTACH_CONTEXT
))
1282 event
->attach_state
&= ~PERF_ATTACH_CONTEXT
;
1284 if (is_cgroup_event(event
)) {
1286 cpuctx
= __get_cpu_context(ctx
);
1288 * if there are no more cgroup events
1289 * then cler cgrp to avoid stale pointer
1290 * in update_cgrp_time_from_cpuctx()
1292 if (!ctx
->nr_cgroups
)
1293 cpuctx
->cgrp
= NULL
;
1296 if (has_branch_stack(event
))
1297 ctx
->nr_branch_stack
--;
1300 if (event
->attr
.inherit_stat
)
1303 list_del_rcu(&event
->event_entry
);
1305 if (event
->group_leader
== event
)
1306 list_del_init(&event
->group_entry
);
1308 update_group_times(event
);
1311 * If event was in error state, then keep it
1312 * that way, otherwise bogus counts will be
1313 * returned on read(). The only way to get out
1314 * of error state is by explicit re-enabling
1317 if (event
->state
> PERF_EVENT_STATE_OFF
)
1318 event
->state
= PERF_EVENT_STATE_OFF
;
1323 static void perf_group_detach(struct perf_event
*event
)
1325 struct perf_event
*sibling
, *tmp
;
1326 struct list_head
*list
= NULL
;
1329 * We can have double detach due to exit/hot-unplug + close.
1331 if (!(event
->attach_state
& PERF_ATTACH_GROUP
))
1334 event
->attach_state
&= ~PERF_ATTACH_GROUP
;
1337 * If this is a sibling, remove it from its group.
1339 if (event
->group_leader
!= event
) {
1340 list_del_init(&event
->group_entry
);
1341 event
->group_leader
->nr_siblings
--;
1345 if (!list_empty(&event
->group_entry
))
1346 list
= &event
->group_entry
;
1349 * If this was a group event with sibling events then
1350 * upgrade the siblings to singleton events by adding them
1351 * to whatever list we are on.
1353 list_for_each_entry_safe(sibling
, tmp
, &event
->sibling_list
, group_entry
) {
1355 list_move_tail(&sibling
->group_entry
, list
);
1356 sibling
->group_leader
= sibling
;
1358 /* Inherit group flags from the previous leader */
1359 sibling
->group_flags
= event
->group_flags
;
1363 perf_event__header_size(event
->group_leader
);
1365 list_for_each_entry(tmp
, &event
->group_leader
->sibling_list
, group_entry
)
1366 perf_event__header_size(tmp
);
1370 event_filter_match(struct perf_event
*event
)
1372 return (event
->cpu
== -1 || event
->cpu
== smp_processor_id())
1373 && perf_cgroup_match(event
);
1377 event_sched_out(struct perf_event
*event
,
1378 struct perf_cpu_context
*cpuctx
,
1379 struct perf_event_context
*ctx
)
1381 u64 tstamp
= perf_event_time(event
);
1384 * An event which could not be activated because of
1385 * filter mismatch still needs to have its timings
1386 * maintained, otherwise bogus information is return
1387 * via read() for time_enabled, time_running:
1389 if (event
->state
== PERF_EVENT_STATE_INACTIVE
1390 && !event_filter_match(event
)) {
1391 delta
= tstamp
- event
->tstamp_stopped
;
1392 event
->tstamp_running
+= delta
;
1393 event
->tstamp_stopped
= tstamp
;
1396 if (event
->state
!= PERF_EVENT_STATE_ACTIVE
)
1399 event
->state
= PERF_EVENT_STATE_INACTIVE
;
1400 if (event
->pending_disable
) {
1401 event
->pending_disable
= 0;
1402 event
->state
= PERF_EVENT_STATE_OFF
;
1404 event
->tstamp_stopped
= tstamp
;
1405 event
->pmu
->del(event
, 0);
1408 if (!is_software_event(event
))
1409 cpuctx
->active_oncpu
--;
1411 if (event
->attr
.freq
&& event
->attr
.sample_freq
)
1413 if (event
->attr
.exclusive
|| !cpuctx
->active_oncpu
)
1414 cpuctx
->exclusive
= 0;
1418 group_sched_out(struct perf_event
*group_event
,
1419 struct perf_cpu_context
*cpuctx
,
1420 struct perf_event_context
*ctx
)
1422 struct perf_event
*event
;
1423 int state
= group_event
->state
;
1425 event_sched_out(group_event
, cpuctx
, ctx
);
1428 * Schedule out siblings (if any):
1430 list_for_each_entry(event
, &group_event
->sibling_list
, group_entry
)
1431 event_sched_out(event
, cpuctx
, ctx
);
1433 if (state
== PERF_EVENT_STATE_ACTIVE
&& group_event
->attr
.exclusive
)
1434 cpuctx
->exclusive
= 0;
1438 * Cross CPU call to remove a performance event
1440 * We disable the event on the hardware level first. After that we
1441 * remove it from the context list.
1443 static int __perf_remove_from_context(void *info
)
1445 struct perf_event
*event
= info
;
1446 struct perf_event_context
*ctx
= event
->ctx
;
1447 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
1449 raw_spin_lock(&ctx
->lock
);
1450 event_sched_out(event
, cpuctx
, ctx
);
1451 list_del_event(event
, ctx
);
1452 if (!ctx
->nr_events
&& cpuctx
->task_ctx
== ctx
) {
1454 cpuctx
->task_ctx
= NULL
;
1456 raw_spin_unlock(&ctx
->lock
);
1463 * Remove the event from a task's (or a CPU's) list of events.
1465 * CPU events are removed with a smp call. For task events we only
1466 * call when the task is on a CPU.
1468 * If event->ctx is a cloned context, callers must make sure that
1469 * every task struct that event->ctx->task could possibly point to
1470 * remains valid. This is OK when called from perf_release since
1471 * that only calls us on the top-level context, which can't be a clone.
1472 * When called from perf_event_exit_task, it's OK because the
1473 * context has been detached from its task.
1475 static void perf_remove_from_context(struct perf_event
*event
)
1477 struct perf_event_context
*ctx
= event
->ctx
;
1478 struct task_struct
*task
= ctx
->task
;
1480 lockdep_assert_held(&ctx
->mutex
);
1484 * Per cpu events are removed via an smp call and
1485 * the removal is always successful.
1487 cpu_function_call(event
->cpu
, __perf_remove_from_context
, event
);
1492 if (!task_function_call(task
, __perf_remove_from_context
, event
))
1495 raw_spin_lock_irq(&ctx
->lock
);
1497 * If we failed to find a running task, but find the context active now
1498 * that we've acquired the ctx->lock, retry.
1500 if (ctx
->is_active
) {
1501 raw_spin_unlock_irq(&ctx
->lock
);
1506 * Since the task isn't running, its safe to remove the event, us
1507 * holding the ctx->lock ensures the task won't get scheduled in.
1509 list_del_event(event
, ctx
);
1510 raw_spin_unlock_irq(&ctx
->lock
);
1514 * Cross CPU call to disable a performance event
1516 int __perf_event_disable(void *info
)
1518 struct perf_event
*event
= info
;
1519 struct perf_event_context
*ctx
= event
->ctx
;
1520 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
1523 * If this is a per-task event, need to check whether this
1524 * event's task is the current task on this cpu.
1526 * Can trigger due to concurrent perf_event_context_sched_out()
1527 * flipping contexts around.
1529 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
)
1532 raw_spin_lock(&ctx
->lock
);
1535 * If the event is on, turn it off.
1536 * If it is in error state, leave it in error state.
1538 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
) {
1539 update_context_time(ctx
);
1540 update_cgrp_time_from_event(event
);
1541 update_group_times(event
);
1542 if (event
== event
->group_leader
)
1543 group_sched_out(event
, cpuctx
, ctx
);
1545 event_sched_out(event
, cpuctx
, ctx
);
1546 event
->state
= PERF_EVENT_STATE_OFF
;
1549 raw_spin_unlock(&ctx
->lock
);
1557 * If event->ctx is a cloned context, callers must make sure that
1558 * every task struct that event->ctx->task could possibly point to
1559 * remains valid. This condition is satisifed when called through
1560 * perf_event_for_each_child or perf_event_for_each because they
1561 * hold the top-level event's child_mutex, so any descendant that
1562 * goes to exit will block in sync_child_event.
1563 * When called from perf_pending_event it's OK because event->ctx
1564 * is the current context on this CPU and preemption is disabled,
1565 * hence we can't get into perf_event_task_sched_out for this context.
1567 void perf_event_disable(struct perf_event
*event
)
1569 struct perf_event_context
*ctx
= event
->ctx
;
1570 struct task_struct
*task
= ctx
->task
;
1574 * Disable the event on the cpu that it's on
1576 cpu_function_call(event
->cpu
, __perf_event_disable
, event
);
1581 if (!task_function_call(task
, __perf_event_disable
, event
))
1584 raw_spin_lock_irq(&ctx
->lock
);
1586 * If the event is still active, we need to retry the cross-call.
1588 if (event
->state
== PERF_EVENT_STATE_ACTIVE
) {
1589 raw_spin_unlock_irq(&ctx
->lock
);
1591 * Reload the task pointer, it might have been changed by
1592 * a concurrent perf_event_context_sched_out().
1599 * Since we have the lock this context can't be scheduled
1600 * in, so we can change the state safely.
1602 if (event
->state
== PERF_EVENT_STATE_INACTIVE
) {
1603 update_group_times(event
);
1604 event
->state
= PERF_EVENT_STATE_OFF
;
1606 raw_spin_unlock_irq(&ctx
->lock
);
1608 EXPORT_SYMBOL_GPL(perf_event_disable
);
1610 static void perf_set_shadow_time(struct perf_event
*event
,
1611 struct perf_event_context
*ctx
,
1615 * use the correct time source for the time snapshot
1617 * We could get by without this by leveraging the
1618 * fact that to get to this function, the caller
1619 * has most likely already called update_context_time()
1620 * and update_cgrp_time_xx() and thus both timestamp
1621 * are identical (or very close). Given that tstamp is,
1622 * already adjusted for cgroup, we could say that:
1623 * tstamp - ctx->timestamp
1625 * tstamp - cgrp->timestamp.
1627 * Then, in perf_output_read(), the calculation would
1628 * work with no changes because:
1629 * - event is guaranteed scheduled in
1630 * - no scheduled out in between
1631 * - thus the timestamp would be the same
1633 * But this is a bit hairy.
1635 * So instead, we have an explicit cgroup call to remain
1636 * within the time time source all along. We believe it
1637 * is cleaner and simpler to understand.
1639 if (is_cgroup_event(event
))
1640 perf_cgroup_set_shadow_time(event
, tstamp
);
1642 event
->shadow_ctx_time
= tstamp
- ctx
->timestamp
;
1645 #define MAX_INTERRUPTS (~0ULL)
1647 static void perf_log_throttle(struct perf_event
*event
, int enable
);
1650 event_sched_in(struct perf_event
*event
,
1651 struct perf_cpu_context
*cpuctx
,
1652 struct perf_event_context
*ctx
)
1654 u64 tstamp
= perf_event_time(event
);
1656 if (event
->state
<= PERF_EVENT_STATE_OFF
)
1659 event
->state
= PERF_EVENT_STATE_ACTIVE
;
1660 event
->oncpu
= smp_processor_id();
1663 * Unthrottle events, since we scheduled we might have missed several
1664 * ticks already, also for a heavily scheduling task there is little
1665 * guarantee it'll get a tick in a timely manner.
1667 if (unlikely(event
->hw
.interrupts
== MAX_INTERRUPTS
)) {
1668 perf_log_throttle(event
, 1);
1669 event
->hw
.interrupts
= 0;
1673 * The new state must be visible before we turn it on in the hardware:
1677 if (event
->pmu
->add(event
, PERF_EF_START
)) {
1678 event
->state
= PERF_EVENT_STATE_INACTIVE
;
1683 event
->tstamp_running
+= tstamp
- event
->tstamp_stopped
;
1685 perf_set_shadow_time(event
, ctx
, tstamp
);
1687 if (!is_software_event(event
))
1688 cpuctx
->active_oncpu
++;
1690 if (event
->attr
.freq
&& event
->attr
.sample_freq
)
1693 if (event
->attr
.exclusive
)
1694 cpuctx
->exclusive
= 1;
1700 group_sched_in(struct perf_event
*group_event
,
1701 struct perf_cpu_context
*cpuctx
,
1702 struct perf_event_context
*ctx
)
1704 struct perf_event
*event
, *partial_group
= NULL
;
1705 struct pmu
*pmu
= group_event
->pmu
;
1706 u64 now
= ctx
->time
;
1707 bool simulate
= false;
1709 if (group_event
->state
== PERF_EVENT_STATE_OFF
)
1712 pmu
->start_txn(pmu
);
1714 if (event_sched_in(group_event
, cpuctx
, ctx
)) {
1715 pmu
->cancel_txn(pmu
);
1716 perf_cpu_hrtimer_restart(cpuctx
);
1721 * Schedule in siblings as one group (if any):
1723 list_for_each_entry(event
, &group_event
->sibling_list
, group_entry
) {
1724 if (event_sched_in(event
, cpuctx
, ctx
)) {
1725 partial_group
= event
;
1730 if (!pmu
->commit_txn(pmu
))
1735 * Groups can be scheduled in as one unit only, so undo any
1736 * partial group before returning:
1737 * The events up to the failed event are scheduled out normally,
1738 * tstamp_stopped will be updated.
1740 * The failed events and the remaining siblings need to have
1741 * their timings updated as if they had gone thru event_sched_in()
1742 * and event_sched_out(). This is required to get consistent timings
1743 * across the group. This also takes care of the case where the group
1744 * could never be scheduled by ensuring tstamp_stopped is set to mark
1745 * the time the event was actually stopped, such that time delta
1746 * calculation in update_event_times() is correct.
1748 list_for_each_entry(event
, &group_event
->sibling_list
, group_entry
) {
1749 if (event
== partial_group
)
1753 event
->tstamp_running
+= now
- event
->tstamp_stopped
;
1754 event
->tstamp_stopped
= now
;
1756 event_sched_out(event
, cpuctx
, ctx
);
1759 event_sched_out(group_event
, cpuctx
, ctx
);
1761 pmu
->cancel_txn(pmu
);
1763 perf_cpu_hrtimer_restart(cpuctx
);
1769 * Work out whether we can put this event group on the CPU now.
1771 static int group_can_go_on(struct perf_event
*event
,
1772 struct perf_cpu_context
*cpuctx
,
1776 * Groups consisting entirely of software events can always go on.
1778 if (event
->group_flags
& PERF_GROUP_SOFTWARE
)
1781 * If an exclusive group is already on, no other hardware
1784 if (cpuctx
->exclusive
)
1787 * If this group is exclusive and there are already
1788 * events on the CPU, it can't go on.
1790 if (event
->attr
.exclusive
&& cpuctx
->active_oncpu
)
1793 * Otherwise, try to add it if all previous groups were able
1799 static void add_event_to_ctx(struct perf_event
*event
,
1800 struct perf_event_context
*ctx
)
1802 u64 tstamp
= perf_event_time(event
);
1804 list_add_event(event
, ctx
);
1805 perf_group_attach(event
);
1806 event
->tstamp_enabled
= tstamp
;
1807 event
->tstamp_running
= tstamp
;
1808 event
->tstamp_stopped
= tstamp
;
1811 static void task_ctx_sched_out(struct perf_event_context
*ctx
);
1813 ctx_sched_in(struct perf_event_context
*ctx
,
1814 struct perf_cpu_context
*cpuctx
,
1815 enum event_type_t event_type
,
1816 struct task_struct
*task
);
1818 static void perf_event_sched_in(struct perf_cpu_context
*cpuctx
,
1819 struct perf_event_context
*ctx
,
1820 struct task_struct
*task
)
1822 cpu_ctx_sched_in(cpuctx
, EVENT_PINNED
, task
);
1824 ctx_sched_in(ctx
, cpuctx
, EVENT_PINNED
, task
);
1825 cpu_ctx_sched_in(cpuctx
, EVENT_FLEXIBLE
, task
);
1827 ctx_sched_in(ctx
, cpuctx
, EVENT_FLEXIBLE
, task
);
1831 * Cross CPU call to install and enable a performance event
1833 * Must be called with ctx->mutex held
1835 static int __perf_install_in_context(void *info
)
1837 struct perf_event
*event
= info
;
1838 struct perf_event_context
*ctx
= event
->ctx
;
1839 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
1840 struct perf_event_context
*task_ctx
= cpuctx
->task_ctx
;
1841 struct task_struct
*task
= current
;
1843 perf_ctx_lock(cpuctx
, task_ctx
);
1844 perf_pmu_disable(cpuctx
->ctx
.pmu
);
1847 * If there was an active task_ctx schedule it out.
1850 task_ctx_sched_out(task_ctx
);
1853 * If the context we're installing events in is not the
1854 * active task_ctx, flip them.
1856 if (ctx
->task
&& task_ctx
!= ctx
) {
1858 raw_spin_unlock(&task_ctx
->lock
);
1859 raw_spin_lock(&ctx
->lock
);
1864 cpuctx
->task_ctx
= task_ctx
;
1865 task
= task_ctx
->task
;
1868 cpu_ctx_sched_out(cpuctx
, EVENT_ALL
);
1870 update_context_time(ctx
);
1872 * update cgrp time only if current cgrp
1873 * matches event->cgrp. Must be done before
1874 * calling add_event_to_ctx()
1876 update_cgrp_time_from_event(event
);
1878 add_event_to_ctx(event
, ctx
);
1881 * Schedule everything back in
1883 perf_event_sched_in(cpuctx
, task_ctx
, task
);
1885 perf_pmu_enable(cpuctx
->ctx
.pmu
);
1886 perf_ctx_unlock(cpuctx
, task_ctx
);
1892 * Attach a performance event to a context
1894 * First we add the event to the list with the hardware enable bit
1895 * in event->hw_config cleared.
1897 * If the event is attached to a task which is on a CPU we use a smp
1898 * call to enable it in the task context. The task might have been
1899 * scheduled away, but we check this in the smp call again.
1902 perf_install_in_context(struct perf_event_context
*ctx
,
1903 struct perf_event
*event
,
1906 struct task_struct
*task
= ctx
->task
;
1908 lockdep_assert_held(&ctx
->mutex
);
1911 if (event
->cpu
!= -1)
1916 * Per cpu events are installed via an smp call and
1917 * the install is always successful.
1919 cpu_function_call(cpu
, __perf_install_in_context
, event
);
1924 if (!task_function_call(task
, __perf_install_in_context
, event
))
1927 raw_spin_lock_irq(&ctx
->lock
);
1929 * If we failed to find a running task, but find the context active now
1930 * that we've acquired the ctx->lock, retry.
1932 if (ctx
->is_active
) {
1933 raw_spin_unlock_irq(&ctx
->lock
);
1938 * Since the task isn't running, its safe to add the event, us holding
1939 * the ctx->lock ensures the task won't get scheduled in.
1941 add_event_to_ctx(event
, ctx
);
1942 raw_spin_unlock_irq(&ctx
->lock
);
1946 * Put a event into inactive state and update time fields.
1947 * Enabling the leader of a group effectively enables all
1948 * the group members that aren't explicitly disabled, so we
1949 * have to update their ->tstamp_enabled also.
1950 * Note: this works for group members as well as group leaders
1951 * since the non-leader members' sibling_lists will be empty.
1953 static void __perf_event_mark_enabled(struct perf_event
*event
)
1955 struct perf_event
*sub
;
1956 u64 tstamp
= perf_event_time(event
);
1958 event
->state
= PERF_EVENT_STATE_INACTIVE
;
1959 event
->tstamp_enabled
= tstamp
- event
->total_time_enabled
;
1960 list_for_each_entry(sub
, &event
->sibling_list
, group_entry
) {
1961 if (sub
->state
>= PERF_EVENT_STATE_INACTIVE
)
1962 sub
->tstamp_enabled
= tstamp
- sub
->total_time_enabled
;
1967 * Cross CPU call to enable a performance event
1969 static int __perf_event_enable(void *info
)
1971 struct perf_event
*event
= info
;
1972 struct perf_event_context
*ctx
= event
->ctx
;
1973 struct perf_event
*leader
= event
->group_leader
;
1974 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
1978 * There's a time window between 'ctx->is_active' check
1979 * in perf_event_enable function and this place having:
1981 * - ctx->lock unlocked
1983 * where the task could be killed and 'ctx' deactivated
1984 * by perf_event_exit_task.
1986 if (!ctx
->is_active
)
1989 raw_spin_lock(&ctx
->lock
);
1990 update_context_time(ctx
);
1992 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
)
1996 * set current task's cgroup time reference point
1998 perf_cgroup_set_timestamp(current
, ctx
);
2000 __perf_event_mark_enabled(event
);
2002 if (!event_filter_match(event
)) {
2003 if (is_cgroup_event(event
))
2004 perf_cgroup_defer_enabled(event
);
2009 * If the event is in a group and isn't the group leader,
2010 * then don't put it on unless the group is on.
2012 if (leader
!= event
&& leader
->state
!= PERF_EVENT_STATE_ACTIVE
)
2015 if (!group_can_go_on(event
, cpuctx
, 1)) {
2018 if (event
== leader
)
2019 err
= group_sched_in(event
, cpuctx
, ctx
);
2021 err
= event_sched_in(event
, cpuctx
, ctx
);
2026 * If this event can't go on and it's part of a
2027 * group, then the whole group has to come off.
2029 if (leader
!= event
) {
2030 group_sched_out(leader
, cpuctx
, ctx
);
2031 perf_cpu_hrtimer_restart(cpuctx
);
2033 if (leader
->attr
.pinned
) {
2034 update_group_times(leader
);
2035 leader
->state
= PERF_EVENT_STATE_ERROR
;
2040 raw_spin_unlock(&ctx
->lock
);
2048 * If event->ctx is a cloned context, callers must make sure that
2049 * every task struct that event->ctx->task could possibly point to
2050 * remains valid. This condition is satisfied when called through
2051 * perf_event_for_each_child or perf_event_for_each as described
2052 * for perf_event_disable.
2054 void perf_event_enable(struct perf_event
*event
)
2056 struct perf_event_context
*ctx
= event
->ctx
;
2057 struct task_struct
*task
= ctx
->task
;
2061 * Enable the event on the cpu that it's on
2063 cpu_function_call(event
->cpu
, __perf_event_enable
, event
);
2067 raw_spin_lock_irq(&ctx
->lock
);
2068 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
)
2072 * If the event is in error state, clear that first.
2073 * That way, if we see the event in error state below, we
2074 * know that it has gone back into error state, as distinct
2075 * from the task having been scheduled away before the
2076 * cross-call arrived.
2078 if (event
->state
== PERF_EVENT_STATE_ERROR
)
2079 event
->state
= PERF_EVENT_STATE_OFF
;
2082 if (!ctx
->is_active
) {
2083 __perf_event_mark_enabled(event
);
2087 raw_spin_unlock_irq(&ctx
->lock
);
2089 if (!task_function_call(task
, __perf_event_enable
, event
))
2092 raw_spin_lock_irq(&ctx
->lock
);
2095 * If the context is active and the event is still off,
2096 * we need to retry the cross-call.
2098 if (ctx
->is_active
&& event
->state
== PERF_EVENT_STATE_OFF
) {
2100 * task could have been flipped by a concurrent
2101 * perf_event_context_sched_out()
2108 raw_spin_unlock_irq(&ctx
->lock
);
2110 EXPORT_SYMBOL_GPL(perf_event_enable
);
2112 int perf_event_refresh(struct perf_event
*event
, int refresh
)
2115 * not supported on inherited events
2117 if (event
->attr
.inherit
|| !is_sampling_event(event
))
2120 atomic_add(refresh
, &event
->event_limit
);
2121 perf_event_enable(event
);
2125 EXPORT_SYMBOL_GPL(perf_event_refresh
);
2127 static void ctx_sched_out(struct perf_event_context
*ctx
,
2128 struct perf_cpu_context
*cpuctx
,
2129 enum event_type_t event_type
)
2131 struct perf_event
*event
;
2132 int is_active
= ctx
->is_active
;
2134 ctx
->is_active
&= ~event_type
;
2135 if (likely(!ctx
->nr_events
))
2138 update_context_time(ctx
);
2139 update_cgrp_time_from_cpuctx(cpuctx
);
2140 if (!ctx
->nr_active
)
2143 perf_pmu_disable(ctx
->pmu
);
2144 if ((is_active
& EVENT_PINNED
) && (event_type
& EVENT_PINNED
)) {
2145 list_for_each_entry(event
, &ctx
->pinned_groups
, group_entry
)
2146 group_sched_out(event
, cpuctx
, ctx
);
2149 if ((is_active
& EVENT_FLEXIBLE
) && (event_type
& EVENT_FLEXIBLE
)) {
2150 list_for_each_entry(event
, &ctx
->flexible_groups
, group_entry
)
2151 group_sched_out(event
, cpuctx
, ctx
);
2153 perf_pmu_enable(ctx
->pmu
);
2157 * Test whether two contexts are equivalent, i.e. whether they have both been
2158 * cloned from the same version of the same context.
2160 * Equivalence is measured using a generation number in the context that is
2161 * incremented on each modification to it; see unclone_ctx(), list_add_event()
2162 * and list_del_event().
2164 static int context_equiv(struct perf_event_context
*ctx1
,
2165 struct perf_event_context
*ctx2
)
2167 /* Pinning disables the swap optimization */
2168 if (ctx1
->pin_count
|| ctx2
->pin_count
)
2171 /* If ctx1 is the parent of ctx2 */
2172 if (ctx1
== ctx2
->parent_ctx
&& ctx1
->generation
== ctx2
->parent_gen
)
2175 /* If ctx2 is the parent of ctx1 */
2176 if (ctx1
->parent_ctx
== ctx2
&& ctx1
->parent_gen
== ctx2
->generation
)
2180 * If ctx1 and ctx2 have the same parent; we flatten the parent
2181 * hierarchy, see perf_event_init_context().
2183 if (ctx1
->parent_ctx
&& ctx1
->parent_ctx
== ctx2
->parent_ctx
&&
2184 ctx1
->parent_gen
== ctx2
->parent_gen
)
2191 static void __perf_event_sync_stat(struct perf_event
*event
,
2192 struct perf_event
*next_event
)
2196 if (!event
->attr
.inherit_stat
)
2200 * Update the event value, we cannot use perf_event_read()
2201 * because we're in the middle of a context switch and have IRQs
2202 * disabled, which upsets smp_call_function_single(), however
2203 * we know the event must be on the current CPU, therefore we
2204 * don't need to use it.
2206 switch (event
->state
) {
2207 case PERF_EVENT_STATE_ACTIVE
:
2208 event
->pmu
->read(event
);
2211 case PERF_EVENT_STATE_INACTIVE
:
2212 update_event_times(event
);
2220 * In order to keep per-task stats reliable we need to flip the event
2221 * values when we flip the contexts.
2223 value
= local64_read(&next_event
->count
);
2224 value
= local64_xchg(&event
->count
, value
);
2225 local64_set(&next_event
->count
, value
);
2227 swap(event
->total_time_enabled
, next_event
->total_time_enabled
);
2228 swap(event
->total_time_running
, next_event
->total_time_running
);
2231 * Since we swizzled the values, update the user visible data too.
2233 perf_event_update_userpage(event
);
2234 perf_event_update_userpage(next_event
);
2237 static void perf_event_sync_stat(struct perf_event_context
*ctx
,
2238 struct perf_event_context
*next_ctx
)
2240 struct perf_event
*event
, *next_event
;
2245 update_context_time(ctx
);
2247 event
= list_first_entry(&ctx
->event_list
,
2248 struct perf_event
, event_entry
);
2250 next_event
= list_first_entry(&next_ctx
->event_list
,
2251 struct perf_event
, event_entry
);
2253 while (&event
->event_entry
!= &ctx
->event_list
&&
2254 &next_event
->event_entry
!= &next_ctx
->event_list
) {
2256 __perf_event_sync_stat(event
, next_event
);
2258 event
= list_next_entry(event
, event_entry
);
2259 next_event
= list_next_entry(next_event
, event_entry
);
2263 static void perf_event_context_sched_out(struct task_struct
*task
, int ctxn
,
2264 struct task_struct
*next
)
2266 struct perf_event_context
*ctx
= task
->perf_event_ctxp
[ctxn
];
2267 struct perf_event_context
*next_ctx
;
2268 struct perf_event_context
*parent
, *next_parent
;
2269 struct perf_cpu_context
*cpuctx
;
2275 cpuctx
= __get_cpu_context(ctx
);
2276 if (!cpuctx
->task_ctx
)
2280 next_ctx
= next
->perf_event_ctxp
[ctxn
];
2284 parent
= rcu_dereference(ctx
->parent_ctx
);
2285 next_parent
= rcu_dereference(next_ctx
->parent_ctx
);
2287 /* If neither context have a parent context; they cannot be clones. */
2288 if (!parent
&& !next_parent
)
2291 if (next_parent
== ctx
|| next_ctx
== parent
|| next_parent
== parent
) {
2293 * Looks like the two contexts are clones, so we might be
2294 * able to optimize the context switch. We lock both
2295 * contexts and check that they are clones under the
2296 * lock (including re-checking that neither has been
2297 * uncloned in the meantime). It doesn't matter which
2298 * order we take the locks because no other cpu could
2299 * be trying to lock both of these tasks.
2301 raw_spin_lock(&ctx
->lock
);
2302 raw_spin_lock_nested(&next_ctx
->lock
, SINGLE_DEPTH_NESTING
);
2303 if (context_equiv(ctx
, next_ctx
)) {
2305 * XXX do we need a memory barrier of sorts
2306 * wrt to rcu_dereference() of perf_event_ctxp
2308 task
->perf_event_ctxp
[ctxn
] = next_ctx
;
2309 next
->perf_event_ctxp
[ctxn
] = ctx
;
2311 next_ctx
->task
= task
;
2314 perf_event_sync_stat(ctx
, next_ctx
);
2316 raw_spin_unlock(&next_ctx
->lock
);
2317 raw_spin_unlock(&ctx
->lock
);
2323 raw_spin_lock(&ctx
->lock
);
2324 ctx_sched_out(ctx
, cpuctx
, EVENT_ALL
);
2325 cpuctx
->task_ctx
= NULL
;
2326 raw_spin_unlock(&ctx
->lock
);
2330 #define for_each_task_context_nr(ctxn) \
2331 for ((ctxn) = 0; (ctxn) < perf_nr_task_contexts; (ctxn)++)
2334 * Called from scheduler to remove the events of the current task,
2335 * with interrupts disabled.
2337 * We stop each event and update the event value in event->count.
2339 * This does not protect us against NMI, but disable()
2340 * sets the disabled bit in the control field of event _before_
2341 * accessing the event control register. If a NMI hits, then it will
2342 * not restart the event.
2344 void __perf_event_task_sched_out(struct task_struct
*task
,
2345 struct task_struct
*next
)
2349 for_each_task_context_nr(ctxn
)
2350 perf_event_context_sched_out(task
, ctxn
, next
);
2353 * if cgroup events exist on this CPU, then we need
2354 * to check if we have to switch out PMU state.
2355 * cgroup event are system-wide mode only
2357 if (atomic_read(&__get_cpu_var(perf_cgroup_events
)))
2358 perf_cgroup_sched_out(task
, next
);
2361 static void task_ctx_sched_out(struct perf_event_context
*ctx
)
2363 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
2365 if (!cpuctx
->task_ctx
)
2368 if (WARN_ON_ONCE(ctx
!= cpuctx
->task_ctx
))
2371 ctx_sched_out(ctx
, cpuctx
, EVENT_ALL
);
2372 cpuctx
->task_ctx
= NULL
;
2376 * Called with IRQs disabled
2378 static void cpu_ctx_sched_out(struct perf_cpu_context
*cpuctx
,
2379 enum event_type_t event_type
)
2381 ctx_sched_out(&cpuctx
->ctx
, cpuctx
, event_type
);
2385 ctx_pinned_sched_in(struct perf_event_context
*ctx
,
2386 struct perf_cpu_context
*cpuctx
)
2388 struct perf_event
*event
;
2390 list_for_each_entry(event
, &ctx
->pinned_groups
, group_entry
) {
2391 if (event
->state
<= PERF_EVENT_STATE_OFF
)
2393 if (!event_filter_match(event
))
2396 /* may need to reset tstamp_enabled */
2397 if (is_cgroup_event(event
))
2398 perf_cgroup_mark_enabled(event
, ctx
);
2400 if (group_can_go_on(event
, cpuctx
, 1))
2401 group_sched_in(event
, cpuctx
, ctx
);
2404 * If this pinned group hasn't been scheduled,
2405 * put it in error state.
2407 if (event
->state
== PERF_EVENT_STATE_INACTIVE
) {
2408 update_group_times(event
);
2409 event
->state
= PERF_EVENT_STATE_ERROR
;
2415 ctx_flexible_sched_in(struct perf_event_context
*ctx
,
2416 struct perf_cpu_context
*cpuctx
)
2418 struct perf_event
*event
;
2421 list_for_each_entry(event
, &ctx
->flexible_groups
, group_entry
) {
2422 /* Ignore events in OFF or ERROR state */
2423 if (event
->state
<= PERF_EVENT_STATE_OFF
)
2426 * Listen to the 'cpu' scheduling filter constraint
2429 if (!event_filter_match(event
))
2432 /* may need to reset tstamp_enabled */
2433 if (is_cgroup_event(event
))
2434 perf_cgroup_mark_enabled(event
, ctx
);
2436 if (group_can_go_on(event
, cpuctx
, can_add_hw
)) {
2437 if (group_sched_in(event
, cpuctx
, ctx
))
2444 ctx_sched_in(struct perf_event_context
*ctx
,
2445 struct perf_cpu_context
*cpuctx
,
2446 enum event_type_t event_type
,
2447 struct task_struct
*task
)
2450 int is_active
= ctx
->is_active
;
2452 ctx
->is_active
|= event_type
;
2453 if (likely(!ctx
->nr_events
))
2457 ctx
->timestamp
= now
;
2458 perf_cgroup_set_timestamp(task
, ctx
);
2460 * First go through the list and put on any pinned groups
2461 * in order to give them the best chance of going on.
2463 if (!(is_active
& EVENT_PINNED
) && (event_type
& EVENT_PINNED
))
2464 ctx_pinned_sched_in(ctx
, cpuctx
);
2466 /* Then walk through the lower prio flexible groups */
2467 if (!(is_active
& EVENT_FLEXIBLE
) && (event_type
& EVENT_FLEXIBLE
))
2468 ctx_flexible_sched_in(ctx
, cpuctx
);
2471 static void cpu_ctx_sched_in(struct perf_cpu_context
*cpuctx
,
2472 enum event_type_t event_type
,
2473 struct task_struct
*task
)
2475 struct perf_event_context
*ctx
= &cpuctx
->ctx
;
2477 ctx_sched_in(ctx
, cpuctx
, event_type
, task
);
2480 static void perf_event_context_sched_in(struct perf_event_context
*ctx
,
2481 struct task_struct
*task
)
2483 struct perf_cpu_context
*cpuctx
;
2485 cpuctx
= __get_cpu_context(ctx
);
2486 if (cpuctx
->task_ctx
== ctx
)
2489 perf_ctx_lock(cpuctx
, ctx
);
2490 perf_pmu_disable(ctx
->pmu
);
2492 * We want to keep the following priority order:
2493 * cpu pinned (that don't need to move), task pinned,
2494 * cpu flexible, task flexible.
2496 cpu_ctx_sched_out(cpuctx
, EVENT_FLEXIBLE
);
2499 cpuctx
->task_ctx
= ctx
;
2501 perf_event_sched_in(cpuctx
, cpuctx
->task_ctx
, task
);
2503 perf_pmu_enable(ctx
->pmu
);
2504 perf_ctx_unlock(cpuctx
, ctx
);
2507 * Since these rotations are per-cpu, we need to ensure the
2508 * cpu-context we got scheduled on is actually rotating.
2510 perf_pmu_rotate_start(ctx
->pmu
);
2514 * When sampling the branck stack in system-wide, it may be necessary
2515 * to flush the stack on context switch. This happens when the branch
2516 * stack does not tag its entries with the pid of the current task.
2517 * Otherwise it becomes impossible to associate a branch entry with a
2518 * task. This ambiguity is more likely to appear when the branch stack
2519 * supports priv level filtering and the user sets it to monitor only
2520 * at the user level (which could be a useful measurement in system-wide
2521 * mode). In that case, the risk is high of having a branch stack with
2522 * branch from multiple tasks. Flushing may mean dropping the existing
2523 * entries or stashing them somewhere in the PMU specific code layer.
2525 * This function provides the context switch callback to the lower code
2526 * layer. It is invoked ONLY when there is at least one system-wide context
2527 * with at least one active event using taken branch sampling.
2529 static void perf_branch_stack_sched_in(struct task_struct
*prev
,
2530 struct task_struct
*task
)
2532 struct perf_cpu_context
*cpuctx
;
2534 unsigned long flags
;
2536 /* no need to flush branch stack if not changing task */
2540 local_irq_save(flags
);
2544 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
2545 cpuctx
= this_cpu_ptr(pmu
->pmu_cpu_context
);
2548 * check if the context has at least one
2549 * event using PERF_SAMPLE_BRANCH_STACK
2551 if (cpuctx
->ctx
.nr_branch_stack
> 0
2552 && pmu
->flush_branch_stack
) {
2554 pmu
= cpuctx
->ctx
.pmu
;
2556 perf_ctx_lock(cpuctx
, cpuctx
->task_ctx
);
2558 perf_pmu_disable(pmu
);
2560 pmu
->flush_branch_stack();
2562 perf_pmu_enable(pmu
);
2564 perf_ctx_unlock(cpuctx
, cpuctx
->task_ctx
);
2570 local_irq_restore(flags
);
2574 * Called from scheduler to add the events of the current task
2575 * with interrupts disabled.
2577 * We restore the event value and then enable it.
2579 * This does not protect us against NMI, but enable()
2580 * sets the enabled bit in the control field of event _before_
2581 * accessing the event control register. If a NMI hits, then it will
2582 * keep the event running.
2584 void __perf_event_task_sched_in(struct task_struct
*prev
,
2585 struct task_struct
*task
)
2587 struct perf_event_context
*ctx
;
2590 for_each_task_context_nr(ctxn
) {
2591 ctx
= task
->perf_event_ctxp
[ctxn
];
2595 perf_event_context_sched_in(ctx
, task
);
2598 * if cgroup events exist on this CPU, then we need
2599 * to check if we have to switch in PMU state.
2600 * cgroup event are system-wide mode only
2602 if (atomic_read(&__get_cpu_var(perf_cgroup_events
)))
2603 perf_cgroup_sched_in(prev
, task
);
2605 /* check for system-wide branch_stack events */
2606 if (atomic_read(&__get_cpu_var(perf_branch_stack_events
)))
2607 perf_branch_stack_sched_in(prev
, task
);
2610 static u64
perf_calculate_period(struct perf_event
*event
, u64 nsec
, u64 count
)
2612 u64 frequency
= event
->attr
.sample_freq
;
2613 u64 sec
= NSEC_PER_SEC
;
2614 u64 divisor
, dividend
;
2616 int count_fls
, nsec_fls
, frequency_fls
, sec_fls
;
2618 count_fls
= fls64(count
);
2619 nsec_fls
= fls64(nsec
);
2620 frequency_fls
= fls64(frequency
);
2624 * We got @count in @nsec, with a target of sample_freq HZ
2625 * the target period becomes:
2628 * period = -------------------
2629 * @nsec * sample_freq
2634 * Reduce accuracy by one bit such that @a and @b converge
2635 * to a similar magnitude.
2637 #define REDUCE_FLS(a, b) \
2639 if (a##_fls > b##_fls) { \
2649 * Reduce accuracy until either term fits in a u64, then proceed with
2650 * the other, so that finally we can do a u64/u64 division.
2652 while (count_fls
+ sec_fls
> 64 && nsec_fls
+ frequency_fls
> 64) {
2653 REDUCE_FLS(nsec
, frequency
);
2654 REDUCE_FLS(sec
, count
);
2657 if (count_fls
+ sec_fls
> 64) {
2658 divisor
= nsec
* frequency
;
2660 while (count_fls
+ sec_fls
> 64) {
2661 REDUCE_FLS(count
, sec
);
2665 dividend
= count
* sec
;
2667 dividend
= count
* sec
;
2669 while (nsec_fls
+ frequency_fls
> 64) {
2670 REDUCE_FLS(nsec
, frequency
);
2674 divisor
= nsec
* frequency
;
2680 return div64_u64(dividend
, divisor
);
2683 static DEFINE_PER_CPU(int, perf_throttled_count
);
2684 static DEFINE_PER_CPU(u64
, perf_throttled_seq
);
2686 static void perf_adjust_period(struct perf_event
*event
, u64 nsec
, u64 count
, bool disable
)
2688 struct hw_perf_event
*hwc
= &event
->hw
;
2689 s64 period
, sample_period
;
2692 period
= perf_calculate_period(event
, nsec
, count
);
2694 delta
= (s64
)(period
- hwc
->sample_period
);
2695 delta
= (delta
+ 7) / 8; /* low pass filter */
2697 sample_period
= hwc
->sample_period
+ delta
;
2702 hwc
->sample_period
= sample_period
;
2704 if (local64_read(&hwc
->period_left
) > 8*sample_period
) {
2706 event
->pmu
->stop(event
, PERF_EF_UPDATE
);
2708 local64_set(&hwc
->period_left
, 0);
2711 event
->pmu
->start(event
, PERF_EF_RELOAD
);
2716 * combine freq adjustment with unthrottling to avoid two passes over the
2717 * events. At the same time, make sure, having freq events does not change
2718 * the rate of unthrottling as that would introduce bias.
2720 static void perf_adjust_freq_unthr_context(struct perf_event_context
*ctx
,
2723 struct perf_event
*event
;
2724 struct hw_perf_event
*hwc
;
2725 u64 now
, period
= TICK_NSEC
;
2729 * only need to iterate over all events iff:
2730 * - context have events in frequency mode (needs freq adjust)
2731 * - there are events to unthrottle on this cpu
2733 if (!(ctx
->nr_freq
|| needs_unthr
))
2736 raw_spin_lock(&ctx
->lock
);
2737 perf_pmu_disable(ctx
->pmu
);
2739 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
2740 if (event
->state
!= PERF_EVENT_STATE_ACTIVE
)
2743 if (!event_filter_match(event
))
2748 if (hwc
->interrupts
== MAX_INTERRUPTS
) {
2749 hwc
->interrupts
= 0;
2750 perf_log_throttle(event
, 1);
2751 event
->pmu
->start(event
, 0);
2754 if (!event
->attr
.freq
|| !event
->attr
.sample_freq
)
2758 * stop the event and update event->count
2760 event
->pmu
->stop(event
, PERF_EF_UPDATE
);
2762 now
= local64_read(&event
->count
);
2763 delta
= now
- hwc
->freq_count_stamp
;
2764 hwc
->freq_count_stamp
= now
;
2768 * reload only if value has changed
2769 * we have stopped the event so tell that
2770 * to perf_adjust_period() to avoid stopping it
2774 perf_adjust_period(event
, period
, delta
, false);
2776 event
->pmu
->start(event
, delta
> 0 ? PERF_EF_RELOAD
: 0);
2779 perf_pmu_enable(ctx
->pmu
);
2780 raw_spin_unlock(&ctx
->lock
);
2784 * Round-robin a context's events:
2786 static void rotate_ctx(struct perf_event_context
*ctx
)
2789 * Rotate the first entry last of non-pinned groups. Rotation might be
2790 * disabled by the inheritance code.
2792 if (!ctx
->rotate_disable
)
2793 list_rotate_left(&ctx
->flexible_groups
);
2797 * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
2798 * because they're strictly cpu affine and rotate_start is called with IRQs
2799 * disabled, while rotate_context is called from IRQ context.
2801 static int perf_rotate_context(struct perf_cpu_context
*cpuctx
)
2803 struct perf_event_context
*ctx
= NULL
;
2804 int rotate
= 0, remove
= 1;
2806 if (cpuctx
->ctx
.nr_events
) {
2808 if (cpuctx
->ctx
.nr_events
!= cpuctx
->ctx
.nr_active
)
2812 ctx
= cpuctx
->task_ctx
;
2813 if (ctx
&& ctx
->nr_events
) {
2815 if (ctx
->nr_events
!= ctx
->nr_active
)
2822 perf_ctx_lock(cpuctx
, cpuctx
->task_ctx
);
2823 perf_pmu_disable(cpuctx
->ctx
.pmu
);
2825 cpu_ctx_sched_out(cpuctx
, EVENT_FLEXIBLE
);
2827 ctx_sched_out(ctx
, cpuctx
, EVENT_FLEXIBLE
);
2829 rotate_ctx(&cpuctx
->ctx
);
2833 perf_event_sched_in(cpuctx
, ctx
, current
);
2835 perf_pmu_enable(cpuctx
->ctx
.pmu
);
2836 perf_ctx_unlock(cpuctx
, cpuctx
->task_ctx
);
2839 list_del_init(&cpuctx
->rotation_list
);
2844 #ifdef CONFIG_NO_HZ_FULL
2845 bool perf_event_can_stop_tick(void)
2847 if (atomic_read(&nr_freq_events
) ||
2848 __this_cpu_read(perf_throttled_count
))
2855 void perf_event_task_tick(void)
2857 struct list_head
*head
= &__get_cpu_var(rotation_list
);
2858 struct perf_cpu_context
*cpuctx
, *tmp
;
2859 struct perf_event_context
*ctx
;
2862 WARN_ON(!irqs_disabled());
2864 __this_cpu_inc(perf_throttled_seq
);
2865 throttled
= __this_cpu_xchg(perf_throttled_count
, 0);
2867 list_for_each_entry_safe(cpuctx
, tmp
, head
, rotation_list
) {
2869 perf_adjust_freq_unthr_context(ctx
, throttled
);
2871 ctx
= cpuctx
->task_ctx
;
2873 perf_adjust_freq_unthr_context(ctx
, throttled
);
2877 static int event_enable_on_exec(struct perf_event
*event
,
2878 struct perf_event_context
*ctx
)
2880 if (!event
->attr
.enable_on_exec
)
2883 event
->attr
.enable_on_exec
= 0;
2884 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
)
2887 __perf_event_mark_enabled(event
);
2893 * Enable all of a task's events that have been marked enable-on-exec.
2894 * This expects task == current.
2896 static void perf_event_enable_on_exec(struct perf_event_context
*ctx
)
2898 struct perf_event
*event
;
2899 unsigned long flags
;
2903 local_irq_save(flags
);
2904 if (!ctx
|| !ctx
->nr_events
)
2908 * We must ctxsw out cgroup events to avoid conflict
2909 * when invoking perf_task_event_sched_in() later on
2910 * in this function. Otherwise we end up trying to
2911 * ctxswin cgroup events which are already scheduled
2914 perf_cgroup_sched_out(current
, NULL
);
2916 raw_spin_lock(&ctx
->lock
);
2917 task_ctx_sched_out(ctx
);
2919 list_for_each_entry(event
, &ctx
->event_list
, event_entry
) {
2920 ret
= event_enable_on_exec(event
, ctx
);
2926 * Unclone this context if we enabled any event.
2931 raw_spin_unlock(&ctx
->lock
);
2934 * Also calls ctxswin for cgroup events, if any:
2936 perf_event_context_sched_in(ctx
, ctx
->task
);
2938 local_irq_restore(flags
);
2942 * Cross CPU call to read the hardware event
2944 static void __perf_event_read(void *info
)
2946 struct perf_event
*event
= info
;
2947 struct perf_event_context
*ctx
= event
->ctx
;
2948 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
2951 * If this is a task context, we need to check whether it is
2952 * the current task context of this cpu. If not it has been
2953 * scheduled out before the smp call arrived. In that case
2954 * event->count would have been updated to a recent sample
2955 * when the event was scheduled out.
2957 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
)
2960 raw_spin_lock(&ctx
->lock
);
2961 if (ctx
->is_active
) {
2962 update_context_time(ctx
);
2963 update_cgrp_time_from_event(event
);
2965 update_event_times(event
);
2966 if (event
->state
== PERF_EVENT_STATE_ACTIVE
)
2967 event
->pmu
->read(event
);
2968 raw_spin_unlock(&ctx
->lock
);
2971 static inline u64
perf_event_count(struct perf_event
*event
)
2973 return local64_read(&event
->count
) + atomic64_read(&event
->child_count
);
2976 static u64
perf_event_read(struct perf_event
*event
)
2979 * If event is enabled and currently active on a CPU, update the
2980 * value in the event structure:
2982 if (event
->state
== PERF_EVENT_STATE_ACTIVE
) {
2983 smp_call_function_single(event
->oncpu
,
2984 __perf_event_read
, event
, 1);
2985 } else if (event
->state
== PERF_EVENT_STATE_INACTIVE
) {
2986 struct perf_event_context
*ctx
= event
->ctx
;
2987 unsigned long flags
;
2989 raw_spin_lock_irqsave(&ctx
->lock
, flags
);
2991 * may read while context is not active
2992 * (e.g., thread is blocked), in that case
2993 * we cannot update context time
2995 if (ctx
->is_active
) {
2996 update_context_time(ctx
);
2997 update_cgrp_time_from_event(event
);
2999 update_event_times(event
);
3000 raw_spin_unlock_irqrestore(&ctx
->lock
, flags
);
3003 return perf_event_count(event
);
3007 * Initialize the perf_event context in a task_struct:
3009 static void __perf_event_init_context(struct perf_event_context
*ctx
)
3011 raw_spin_lock_init(&ctx
->lock
);
3012 mutex_init(&ctx
->mutex
);
3013 INIT_LIST_HEAD(&ctx
->pinned_groups
);
3014 INIT_LIST_HEAD(&ctx
->flexible_groups
);
3015 INIT_LIST_HEAD(&ctx
->event_list
);
3016 atomic_set(&ctx
->refcount
, 1);
3019 static struct perf_event_context
*
3020 alloc_perf_context(struct pmu
*pmu
, struct task_struct
*task
)
3022 struct perf_event_context
*ctx
;
3024 ctx
= kzalloc(sizeof(struct perf_event_context
), GFP_KERNEL
);
3028 __perf_event_init_context(ctx
);
3031 get_task_struct(task
);
3038 static struct task_struct
*
3039 find_lively_task_by_vpid(pid_t vpid
)
3041 struct task_struct
*task
;
3048 task
= find_task_by_vpid(vpid
);
3050 get_task_struct(task
);
3054 return ERR_PTR(-ESRCH
);
3056 /* Reuse ptrace permission checks for now. */
3058 if (!ptrace_may_access(task
, PTRACE_MODE_READ
))
3063 put_task_struct(task
);
3064 return ERR_PTR(err
);
3069 * Returns a matching context with refcount and pincount.
3071 static struct perf_event_context
*
3072 find_get_context(struct pmu
*pmu
, struct task_struct
*task
, int cpu
)
3074 struct perf_event_context
*ctx
;
3075 struct perf_cpu_context
*cpuctx
;
3076 unsigned long flags
;
3080 /* Must be root to operate on a CPU event: */
3081 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN
))
3082 return ERR_PTR(-EACCES
);
3085 * We could be clever and allow to attach a event to an
3086 * offline CPU and activate it when the CPU comes up, but
3089 if (!cpu_online(cpu
))
3090 return ERR_PTR(-ENODEV
);
3092 cpuctx
= per_cpu_ptr(pmu
->pmu_cpu_context
, cpu
);
3101 ctxn
= pmu
->task_ctx_nr
;
3106 ctx
= perf_lock_task_context(task
, ctxn
, &flags
);
3110 raw_spin_unlock_irqrestore(&ctx
->lock
, flags
);
3112 ctx
= alloc_perf_context(pmu
, task
);
3118 mutex_lock(&task
->perf_event_mutex
);
3120 * If it has already passed perf_event_exit_task().
3121 * we must see PF_EXITING, it takes this mutex too.
3123 if (task
->flags
& PF_EXITING
)
3125 else if (task
->perf_event_ctxp
[ctxn
])
3130 rcu_assign_pointer(task
->perf_event_ctxp
[ctxn
], ctx
);
3132 mutex_unlock(&task
->perf_event_mutex
);
3134 if (unlikely(err
)) {
3146 return ERR_PTR(err
);
3149 static void perf_event_free_filter(struct perf_event
*event
);
3151 static void free_event_rcu(struct rcu_head
*head
)
3153 struct perf_event
*event
;
3155 event
= container_of(head
, struct perf_event
, rcu_head
);
3157 put_pid_ns(event
->ns
);
3158 perf_event_free_filter(event
);
3162 static void ring_buffer_put(struct ring_buffer
*rb
);
3163 static void ring_buffer_detach(struct perf_event
*event
, struct ring_buffer
*rb
);
3165 static void unaccount_event_cpu(struct perf_event
*event
, int cpu
)
3170 if (has_branch_stack(event
)) {
3171 if (!(event
->attach_state
& PERF_ATTACH_TASK
))
3172 atomic_dec(&per_cpu(perf_branch_stack_events
, cpu
));
3174 if (is_cgroup_event(event
))
3175 atomic_dec(&per_cpu(perf_cgroup_events
, cpu
));
3178 static void unaccount_event(struct perf_event
*event
)
3183 if (event
->attach_state
& PERF_ATTACH_TASK
)
3184 static_key_slow_dec_deferred(&perf_sched_events
);
3185 if (event
->attr
.mmap
|| event
->attr
.mmap_data
)
3186 atomic_dec(&nr_mmap_events
);
3187 if (event
->attr
.comm
)
3188 atomic_dec(&nr_comm_events
);
3189 if (event
->attr
.task
)
3190 atomic_dec(&nr_task_events
);
3191 if (event
->attr
.freq
)
3192 atomic_dec(&nr_freq_events
);
3193 if (is_cgroup_event(event
))
3194 static_key_slow_dec_deferred(&perf_sched_events
);
3195 if (has_branch_stack(event
))
3196 static_key_slow_dec_deferred(&perf_sched_events
);
3198 unaccount_event_cpu(event
, event
->cpu
);
3201 static void __free_event(struct perf_event
*event
)
3203 if (!event
->parent
) {
3204 if (event
->attr
.sample_type
& PERF_SAMPLE_CALLCHAIN
)
3205 put_callchain_buffers();
3209 event
->destroy(event
);
3212 put_ctx(event
->ctx
);
3214 call_rcu(&event
->rcu_head
, free_event_rcu
);
3216 static void free_event(struct perf_event
*event
)
3218 irq_work_sync(&event
->pending
);
3220 unaccount_event(event
);
3223 struct ring_buffer
*rb
;
3226 * Can happen when we close an event with re-directed output.
3228 * Since we have a 0 refcount, perf_mmap_close() will skip
3229 * over us; possibly making our ring_buffer_put() the last.
3231 mutex_lock(&event
->mmap_mutex
);
3234 rcu_assign_pointer(event
->rb
, NULL
);
3235 ring_buffer_detach(event
, rb
);
3236 ring_buffer_put(rb
); /* could be last */
3238 mutex_unlock(&event
->mmap_mutex
);
3241 if (is_cgroup_event(event
))
3242 perf_detach_cgroup(event
);
3245 __free_event(event
);
3248 int perf_event_release_kernel(struct perf_event
*event
)
3250 struct perf_event_context
*ctx
= event
->ctx
;
3252 WARN_ON_ONCE(ctx
->parent_ctx
);
3254 * There are two ways this annotation is useful:
3256 * 1) there is a lock recursion from perf_event_exit_task
3257 * see the comment there.
3259 * 2) there is a lock-inversion with mmap_sem through
3260 * perf_event_read_group(), which takes faults while
3261 * holding ctx->mutex, however this is called after
3262 * the last filedesc died, so there is no possibility
3263 * to trigger the AB-BA case.
3265 mutex_lock_nested(&ctx
->mutex
, SINGLE_DEPTH_NESTING
);
3266 raw_spin_lock_irq(&ctx
->lock
);
3267 perf_group_detach(event
);
3268 raw_spin_unlock_irq(&ctx
->lock
);
3269 perf_remove_from_context(event
);
3270 mutex_unlock(&ctx
->mutex
);
3276 EXPORT_SYMBOL_GPL(perf_event_release_kernel
);
3279 * Called when the last reference to the file is gone.
3281 static void put_event(struct perf_event
*event
)
3283 struct task_struct
*owner
;
3285 if (!atomic_long_dec_and_test(&event
->refcount
))
3289 owner
= ACCESS_ONCE(event
->owner
);
3291 * Matches the smp_wmb() in perf_event_exit_task(). If we observe
3292 * !owner it means the list deletion is complete and we can indeed
3293 * free this event, otherwise we need to serialize on
3294 * owner->perf_event_mutex.
3296 smp_read_barrier_depends();
3299 * Since delayed_put_task_struct() also drops the last
3300 * task reference we can safely take a new reference
3301 * while holding the rcu_read_lock().
3303 get_task_struct(owner
);
3308 mutex_lock(&owner
->perf_event_mutex
);
3310 * We have to re-check the event->owner field, if it is cleared
3311 * we raced with perf_event_exit_task(), acquiring the mutex
3312 * ensured they're done, and we can proceed with freeing the
3316 list_del_init(&event
->owner_entry
);
3317 mutex_unlock(&owner
->perf_event_mutex
);
3318 put_task_struct(owner
);
3321 perf_event_release_kernel(event
);
3324 static int perf_release(struct inode
*inode
, struct file
*file
)
3326 put_event(file
->private_data
);
3330 u64
perf_event_read_value(struct perf_event
*event
, u64
*enabled
, u64
*running
)
3332 struct perf_event
*child
;
3338 mutex_lock(&event
->child_mutex
);
3339 total
+= perf_event_read(event
);
3340 *enabled
+= event
->total_time_enabled
+
3341 atomic64_read(&event
->child_total_time_enabled
);
3342 *running
+= event
->total_time_running
+
3343 atomic64_read(&event
->child_total_time_running
);
3345 list_for_each_entry(child
, &event
->child_list
, child_list
) {
3346 total
+= perf_event_read(child
);
3347 *enabled
+= child
->total_time_enabled
;
3348 *running
+= child
->total_time_running
;
3350 mutex_unlock(&event
->child_mutex
);
3354 EXPORT_SYMBOL_GPL(perf_event_read_value
);
3356 static int perf_event_read_group(struct perf_event
*event
,
3357 u64 read_format
, char __user
*buf
)
3359 struct perf_event
*leader
= event
->group_leader
, *sub
;
3360 int n
= 0, size
= 0, ret
= -EFAULT
;
3361 struct perf_event_context
*ctx
= leader
->ctx
;
3363 u64 count
, enabled
, running
;
3365 mutex_lock(&ctx
->mutex
);
3366 count
= perf_event_read_value(leader
, &enabled
, &running
);
3368 values
[n
++] = 1 + leader
->nr_siblings
;
3369 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
3370 values
[n
++] = enabled
;
3371 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
3372 values
[n
++] = running
;
3373 values
[n
++] = count
;
3374 if (read_format
& PERF_FORMAT_ID
)
3375 values
[n
++] = primary_event_id(leader
);
3377 size
= n
* sizeof(u64
);
3379 if (copy_to_user(buf
, values
, size
))
3384 list_for_each_entry(sub
, &leader
->sibling_list
, group_entry
) {
3387 values
[n
++] = perf_event_read_value(sub
, &enabled
, &running
);
3388 if (read_format
& PERF_FORMAT_ID
)
3389 values
[n
++] = primary_event_id(sub
);
3391 size
= n
* sizeof(u64
);
3393 if (copy_to_user(buf
+ ret
, values
, size
)) {
3401 mutex_unlock(&ctx
->mutex
);
3406 static int perf_event_read_one(struct perf_event
*event
,
3407 u64 read_format
, char __user
*buf
)
3409 u64 enabled
, running
;
3413 values
[n
++] = perf_event_read_value(event
, &enabled
, &running
);
3414 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
3415 values
[n
++] = enabled
;
3416 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
3417 values
[n
++] = running
;
3418 if (read_format
& PERF_FORMAT_ID
)
3419 values
[n
++] = primary_event_id(event
);
3421 if (copy_to_user(buf
, values
, n
* sizeof(u64
)))
3424 return n
* sizeof(u64
);
3428 * Read the performance event - simple non blocking version for now
3431 perf_read_hw(struct perf_event
*event
, char __user
*buf
, size_t count
)
3433 u64 read_format
= event
->attr
.read_format
;
3437 * Return end-of-file for a read on a event that is in
3438 * error state (i.e. because it was pinned but it couldn't be
3439 * scheduled on to the CPU at some point).
3441 if (event
->state
== PERF_EVENT_STATE_ERROR
)
3444 if (count
< event
->read_size
)
3447 WARN_ON_ONCE(event
->ctx
->parent_ctx
);
3448 if (read_format
& PERF_FORMAT_GROUP
)
3449 ret
= perf_event_read_group(event
, read_format
, buf
);
3451 ret
= perf_event_read_one(event
, read_format
, buf
);
3457 perf_read(struct file
*file
, char __user
*buf
, size_t count
, loff_t
*ppos
)
3459 struct perf_event
*event
= file
->private_data
;
3461 return perf_read_hw(event
, buf
, count
);
3464 static unsigned int perf_poll(struct file
*file
, poll_table
*wait
)
3466 struct perf_event
*event
= file
->private_data
;
3467 struct ring_buffer
*rb
;
3468 unsigned int events
= POLL_HUP
;
3471 * Pin the event->rb by taking event->mmap_mutex; otherwise
3472 * perf_event_set_output() can swizzle our rb and make us miss wakeups.
3474 mutex_lock(&event
->mmap_mutex
);
3477 events
= atomic_xchg(&rb
->poll
, 0);
3478 mutex_unlock(&event
->mmap_mutex
);
3480 poll_wait(file
, &event
->waitq
, wait
);
3485 static void perf_event_reset(struct perf_event
*event
)
3487 (void)perf_event_read(event
);
3488 local64_set(&event
->count
, 0);
3489 perf_event_update_userpage(event
);
3493 * Holding the top-level event's child_mutex means that any
3494 * descendant process that has inherited this event will block
3495 * in sync_child_event if it goes to exit, thus satisfying the
3496 * task existence requirements of perf_event_enable/disable.
3498 static void perf_event_for_each_child(struct perf_event
*event
,
3499 void (*func
)(struct perf_event
*))
3501 struct perf_event
*child
;
3503 WARN_ON_ONCE(event
->ctx
->parent_ctx
);
3504 mutex_lock(&event
->child_mutex
);
3506 list_for_each_entry(child
, &event
->child_list
, child_list
)
3508 mutex_unlock(&event
->child_mutex
);
3511 static void perf_event_for_each(struct perf_event
*event
,
3512 void (*func
)(struct perf_event
*))
3514 struct perf_event_context
*ctx
= event
->ctx
;
3515 struct perf_event
*sibling
;
3517 WARN_ON_ONCE(ctx
->parent_ctx
);
3518 mutex_lock(&ctx
->mutex
);
3519 event
= event
->group_leader
;
3521 perf_event_for_each_child(event
, func
);
3522 list_for_each_entry(sibling
, &event
->sibling_list
, group_entry
)
3523 perf_event_for_each_child(sibling
, func
);
3524 mutex_unlock(&ctx
->mutex
);
3527 static int perf_event_period(struct perf_event
*event
, u64 __user
*arg
)
3529 struct perf_event_context
*ctx
= event
->ctx
;
3530 int ret
= 0, active
;
3533 if (!is_sampling_event(event
))
3536 if (copy_from_user(&value
, arg
, sizeof(value
)))
3542 raw_spin_lock_irq(&ctx
->lock
);
3543 if (event
->attr
.freq
) {
3544 if (value
> sysctl_perf_event_sample_rate
) {
3549 event
->attr
.sample_freq
= value
;
3551 event
->attr
.sample_period
= value
;
3552 event
->hw
.sample_period
= value
;
3555 active
= (event
->state
== PERF_EVENT_STATE_ACTIVE
);
3557 perf_pmu_disable(ctx
->pmu
);
3558 event
->pmu
->stop(event
, PERF_EF_UPDATE
);
3561 local64_set(&event
->hw
.period_left
, 0);
3564 event
->pmu
->start(event
, PERF_EF_RELOAD
);
3565 perf_pmu_enable(ctx
->pmu
);
3569 raw_spin_unlock_irq(&ctx
->lock
);
3574 static const struct file_operations perf_fops
;
3576 static inline int perf_fget_light(int fd
, struct fd
*p
)
3578 struct fd f
= fdget(fd
);
3582 if (f
.file
->f_op
!= &perf_fops
) {
3590 static int perf_event_set_output(struct perf_event
*event
,
3591 struct perf_event
*output_event
);
3592 static int perf_event_set_filter(struct perf_event
*event
, void __user
*arg
);
3594 static long perf_ioctl(struct file
*file
, unsigned int cmd
, unsigned long arg
)
3596 struct perf_event
*event
= file
->private_data
;
3597 void (*func
)(struct perf_event
*);
3601 case PERF_EVENT_IOC_ENABLE
:
3602 func
= perf_event_enable
;
3604 case PERF_EVENT_IOC_DISABLE
:
3605 func
= perf_event_disable
;
3607 case PERF_EVENT_IOC_RESET
:
3608 func
= perf_event_reset
;
3611 case PERF_EVENT_IOC_REFRESH
:
3612 return perf_event_refresh(event
, arg
);
3614 case PERF_EVENT_IOC_PERIOD
:
3615 return perf_event_period(event
, (u64 __user
*)arg
);
3617 case PERF_EVENT_IOC_ID
:
3619 u64 id
= primary_event_id(event
);
3621 if (copy_to_user((void __user
*)arg
, &id
, sizeof(id
)))
3626 case PERF_EVENT_IOC_SET_OUTPUT
:
3630 struct perf_event
*output_event
;
3632 ret
= perf_fget_light(arg
, &output
);
3635 output_event
= output
.file
->private_data
;
3636 ret
= perf_event_set_output(event
, output_event
);
3639 ret
= perf_event_set_output(event
, NULL
);
3644 case PERF_EVENT_IOC_SET_FILTER
:
3645 return perf_event_set_filter(event
, (void __user
*)arg
);
3651 if (flags
& PERF_IOC_FLAG_GROUP
)
3652 perf_event_for_each(event
, func
);
3654 perf_event_for_each_child(event
, func
);
3659 int perf_event_task_enable(void)
3661 struct perf_event
*event
;
3663 mutex_lock(¤t
->perf_event_mutex
);
3664 list_for_each_entry(event
, ¤t
->perf_event_list
, owner_entry
)
3665 perf_event_for_each_child(event
, perf_event_enable
);
3666 mutex_unlock(¤t
->perf_event_mutex
);
3671 int perf_event_task_disable(void)
3673 struct perf_event
*event
;
3675 mutex_lock(¤t
->perf_event_mutex
);
3676 list_for_each_entry(event
, ¤t
->perf_event_list
, owner_entry
)
3677 perf_event_for_each_child(event
, perf_event_disable
);
3678 mutex_unlock(¤t
->perf_event_mutex
);
3683 static int perf_event_index(struct perf_event
*event
)
3685 if (event
->hw
.state
& PERF_HES_STOPPED
)
3688 if (event
->state
!= PERF_EVENT_STATE_ACTIVE
)
3691 return event
->pmu
->event_idx(event
);
3694 static void calc_timer_values(struct perf_event
*event
,
3701 *now
= perf_clock();
3702 ctx_time
= event
->shadow_ctx_time
+ *now
;
3703 *enabled
= ctx_time
- event
->tstamp_enabled
;
3704 *running
= ctx_time
- event
->tstamp_running
;
3707 static void perf_event_init_userpage(struct perf_event
*event
)
3709 struct perf_event_mmap_page
*userpg
;
3710 struct ring_buffer
*rb
;
3713 rb
= rcu_dereference(event
->rb
);
3717 userpg
= rb
->user_page
;
3719 /* Allow new userspace to detect that bit 0 is deprecated */
3720 userpg
->cap_bit0_is_deprecated
= 1;
3721 userpg
->size
= offsetof(struct perf_event_mmap_page
, __reserved
);
3727 void __weak
arch_perf_update_userpage(struct perf_event_mmap_page
*userpg
, u64 now
)
3732 * Callers need to ensure there can be no nesting of this function, otherwise
3733 * the seqlock logic goes bad. We can not serialize this because the arch
3734 * code calls this from NMI context.
3736 void perf_event_update_userpage(struct perf_event
*event
)
3738 struct perf_event_mmap_page
*userpg
;
3739 struct ring_buffer
*rb
;
3740 u64 enabled
, running
, now
;
3743 rb
= rcu_dereference(event
->rb
);
3748 * compute total_time_enabled, total_time_running
3749 * based on snapshot values taken when the event
3750 * was last scheduled in.
3752 * we cannot simply called update_context_time()
3753 * because of locking issue as we can be called in
3756 calc_timer_values(event
, &now
, &enabled
, &running
);
3758 userpg
= rb
->user_page
;
3760 * Disable preemption so as to not let the corresponding user-space
3761 * spin too long if we get preempted.
3766 userpg
->index
= perf_event_index(event
);
3767 userpg
->offset
= perf_event_count(event
);
3769 userpg
->offset
-= local64_read(&event
->hw
.prev_count
);
3771 userpg
->time_enabled
= enabled
+
3772 atomic64_read(&event
->child_total_time_enabled
);
3774 userpg
->time_running
= running
+
3775 atomic64_read(&event
->child_total_time_running
);
3777 arch_perf_update_userpage(userpg
, now
);
3786 static int perf_mmap_fault(struct vm_area_struct
*vma
, struct vm_fault
*vmf
)
3788 struct perf_event
*event
= vma
->vm_file
->private_data
;
3789 struct ring_buffer
*rb
;
3790 int ret
= VM_FAULT_SIGBUS
;
3792 if (vmf
->flags
& FAULT_FLAG_MKWRITE
) {
3793 if (vmf
->pgoff
== 0)
3799 rb
= rcu_dereference(event
->rb
);
3803 if (vmf
->pgoff
&& (vmf
->flags
& FAULT_FLAG_WRITE
))
3806 vmf
->page
= perf_mmap_to_page(rb
, vmf
->pgoff
);
3810 get_page(vmf
->page
);
3811 vmf
->page
->mapping
= vma
->vm_file
->f_mapping
;
3812 vmf
->page
->index
= vmf
->pgoff
;
3821 static void ring_buffer_attach(struct perf_event
*event
,
3822 struct ring_buffer
*rb
)
3824 unsigned long flags
;
3826 if (!list_empty(&event
->rb_entry
))
3829 spin_lock_irqsave(&rb
->event_lock
, flags
);
3830 if (list_empty(&event
->rb_entry
))
3831 list_add(&event
->rb_entry
, &rb
->event_list
);
3832 spin_unlock_irqrestore(&rb
->event_lock
, flags
);
3835 static void ring_buffer_detach(struct perf_event
*event
, struct ring_buffer
*rb
)
3837 unsigned long flags
;
3839 if (list_empty(&event
->rb_entry
))
3842 spin_lock_irqsave(&rb
->event_lock
, flags
);
3843 list_del_init(&event
->rb_entry
);
3844 wake_up_all(&event
->waitq
);
3845 spin_unlock_irqrestore(&rb
->event_lock
, flags
);
3848 static void ring_buffer_wakeup(struct perf_event
*event
)
3850 struct ring_buffer
*rb
;
3853 rb
= rcu_dereference(event
->rb
);
3855 list_for_each_entry_rcu(event
, &rb
->event_list
, rb_entry
)
3856 wake_up_all(&event
->waitq
);
3861 static void rb_free_rcu(struct rcu_head
*rcu_head
)
3863 struct ring_buffer
*rb
;
3865 rb
= container_of(rcu_head
, struct ring_buffer
, rcu_head
);
3869 static struct ring_buffer
*ring_buffer_get(struct perf_event
*event
)
3871 struct ring_buffer
*rb
;
3874 rb
= rcu_dereference(event
->rb
);
3876 if (!atomic_inc_not_zero(&rb
->refcount
))
3884 static void ring_buffer_put(struct ring_buffer
*rb
)
3886 if (!atomic_dec_and_test(&rb
->refcount
))
3889 WARN_ON_ONCE(!list_empty(&rb
->event_list
));
3891 call_rcu(&rb
->rcu_head
, rb_free_rcu
);
3894 static void perf_mmap_open(struct vm_area_struct
*vma
)
3896 struct perf_event
*event
= vma
->vm_file
->private_data
;
3898 atomic_inc(&event
->mmap_count
);
3899 atomic_inc(&event
->rb
->mmap_count
);
3903 * A buffer can be mmap()ed multiple times; either directly through the same
3904 * event, or through other events by use of perf_event_set_output().
3906 * In order to undo the VM accounting done by perf_mmap() we need to destroy
3907 * the buffer here, where we still have a VM context. This means we need
3908 * to detach all events redirecting to us.
3910 static void perf_mmap_close(struct vm_area_struct
*vma
)
3912 struct perf_event
*event
= vma
->vm_file
->private_data
;
3914 struct ring_buffer
*rb
= event
->rb
;
3915 struct user_struct
*mmap_user
= rb
->mmap_user
;
3916 int mmap_locked
= rb
->mmap_locked
;
3917 unsigned long size
= perf_data_size(rb
);
3919 atomic_dec(&rb
->mmap_count
);
3921 if (!atomic_dec_and_mutex_lock(&event
->mmap_count
, &event
->mmap_mutex
))
3924 /* Detach current event from the buffer. */
3925 rcu_assign_pointer(event
->rb
, NULL
);
3926 ring_buffer_detach(event
, rb
);
3927 mutex_unlock(&event
->mmap_mutex
);
3929 /* If there's still other mmap()s of this buffer, we're done. */
3930 if (atomic_read(&rb
->mmap_count
)) {
3931 ring_buffer_put(rb
); /* can't be last */
3936 * No other mmap()s, detach from all other events that might redirect
3937 * into the now unreachable buffer. Somewhat complicated by the
3938 * fact that rb::event_lock otherwise nests inside mmap_mutex.
3942 list_for_each_entry_rcu(event
, &rb
->event_list
, rb_entry
) {
3943 if (!atomic_long_inc_not_zero(&event
->refcount
)) {
3945 * This event is en-route to free_event() which will
3946 * detach it and remove it from the list.
3952 mutex_lock(&event
->mmap_mutex
);
3954 * Check we didn't race with perf_event_set_output() which can
3955 * swizzle the rb from under us while we were waiting to
3956 * acquire mmap_mutex.
3958 * If we find a different rb; ignore this event, a next
3959 * iteration will no longer find it on the list. We have to
3960 * still restart the iteration to make sure we're not now
3961 * iterating the wrong list.
3963 if (event
->rb
== rb
) {
3964 rcu_assign_pointer(event
->rb
, NULL
);
3965 ring_buffer_detach(event
, rb
);
3966 ring_buffer_put(rb
); /* can't be last, we still have one */
3968 mutex_unlock(&event
->mmap_mutex
);
3972 * Restart the iteration; either we're on the wrong list or
3973 * destroyed its integrity by doing a deletion.
3980 * It could be there's still a few 0-ref events on the list; they'll
3981 * get cleaned up by free_event() -- they'll also still have their
3982 * ref on the rb and will free it whenever they are done with it.
3984 * Aside from that, this buffer is 'fully' detached and unmapped,
3985 * undo the VM accounting.
3988 atomic_long_sub((size
>> PAGE_SHIFT
) + 1, &mmap_user
->locked_vm
);
3989 vma
->vm_mm
->pinned_vm
-= mmap_locked
;
3990 free_uid(mmap_user
);
3992 ring_buffer_put(rb
); /* could be last */
3995 static const struct vm_operations_struct perf_mmap_vmops
= {
3996 .open
= perf_mmap_open
,
3997 .close
= perf_mmap_close
,
3998 .fault
= perf_mmap_fault
,
3999 .page_mkwrite
= perf_mmap_fault
,
4002 static int perf_mmap(struct file
*file
, struct vm_area_struct
*vma
)
4004 struct perf_event
*event
= file
->private_data
;
4005 unsigned long user_locked
, user_lock_limit
;
4006 struct user_struct
*user
= current_user();
4007 unsigned long locked
, lock_limit
;
4008 struct ring_buffer
*rb
;
4009 unsigned long vma_size
;
4010 unsigned long nr_pages
;
4011 long user_extra
, extra
;
4012 int ret
= 0, flags
= 0;
4015 * Don't allow mmap() of inherited per-task counters. This would
4016 * create a performance issue due to all children writing to the
4019 if (event
->cpu
== -1 && event
->attr
.inherit
)
4022 if (!(vma
->vm_flags
& VM_SHARED
))
4025 vma_size
= vma
->vm_end
- vma
->vm_start
;
4026 nr_pages
= (vma_size
/ PAGE_SIZE
) - 1;
4029 * If we have rb pages ensure they're a power-of-two number, so we
4030 * can do bitmasks instead of modulo.
4032 if (nr_pages
!= 0 && !is_power_of_2(nr_pages
))
4035 if (vma_size
!= PAGE_SIZE
* (1 + nr_pages
))
4038 if (vma
->vm_pgoff
!= 0)
4041 WARN_ON_ONCE(event
->ctx
->parent_ctx
);
4043 mutex_lock(&event
->mmap_mutex
);
4045 if (event
->rb
->nr_pages
!= nr_pages
) {
4050 if (!atomic_inc_not_zero(&event
->rb
->mmap_count
)) {
4052 * Raced against perf_mmap_close() through
4053 * perf_event_set_output(). Try again, hope for better
4056 mutex_unlock(&event
->mmap_mutex
);
4063 user_extra
= nr_pages
+ 1;
4064 user_lock_limit
= sysctl_perf_event_mlock
>> (PAGE_SHIFT
- 10);
4067 * Increase the limit linearly with more CPUs:
4069 user_lock_limit
*= num_online_cpus();
4071 user_locked
= atomic_long_read(&user
->locked_vm
) + user_extra
;
4074 if (user_locked
> user_lock_limit
)
4075 extra
= user_locked
- user_lock_limit
;
4077 lock_limit
= rlimit(RLIMIT_MEMLOCK
);
4078 lock_limit
>>= PAGE_SHIFT
;
4079 locked
= vma
->vm_mm
->pinned_vm
+ extra
;
4081 if ((locked
> lock_limit
) && perf_paranoid_tracepoint_raw() &&
4082 !capable(CAP_IPC_LOCK
)) {
4089 if (vma
->vm_flags
& VM_WRITE
)
4090 flags
|= RING_BUFFER_WRITABLE
;
4092 rb
= rb_alloc(nr_pages
,
4093 event
->attr
.watermark
? event
->attr
.wakeup_watermark
: 0,
4101 atomic_set(&rb
->mmap_count
, 1);
4102 rb
->mmap_locked
= extra
;
4103 rb
->mmap_user
= get_current_user();
4105 atomic_long_add(user_extra
, &user
->locked_vm
);
4106 vma
->vm_mm
->pinned_vm
+= extra
;
4108 ring_buffer_attach(event
, rb
);
4109 rcu_assign_pointer(event
->rb
, rb
);
4111 perf_event_init_userpage(event
);
4112 perf_event_update_userpage(event
);
4116 atomic_inc(&event
->mmap_count
);
4117 mutex_unlock(&event
->mmap_mutex
);
4120 * Since pinned accounting is per vm we cannot allow fork() to copy our
4123 vma
->vm_flags
|= VM_DONTCOPY
| VM_DONTEXPAND
| VM_DONTDUMP
;
4124 vma
->vm_ops
= &perf_mmap_vmops
;
4129 static int perf_fasync(int fd
, struct file
*filp
, int on
)
4131 struct inode
*inode
= file_inode(filp
);
4132 struct perf_event
*event
= filp
->private_data
;
4135 mutex_lock(&inode
->i_mutex
);
4136 retval
= fasync_helper(fd
, filp
, on
, &event
->fasync
);
4137 mutex_unlock(&inode
->i_mutex
);
4145 static const struct file_operations perf_fops
= {
4146 .llseek
= no_llseek
,
4147 .release
= perf_release
,
4150 .unlocked_ioctl
= perf_ioctl
,
4151 .compat_ioctl
= perf_ioctl
,
4153 .fasync
= perf_fasync
,
4159 * If there's data, ensure we set the poll() state and publish everything
4160 * to user-space before waking everybody up.
4163 void perf_event_wakeup(struct perf_event
*event
)
4165 ring_buffer_wakeup(event
);
4167 if (event
->pending_kill
) {
4168 kill_fasync(&event
->fasync
, SIGIO
, event
->pending_kill
);
4169 event
->pending_kill
= 0;
4173 static void perf_pending_event(struct irq_work
*entry
)
4175 struct perf_event
*event
= container_of(entry
,
4176 struct perf_event
, pending
);
4178 if (event
->pending_disable
) {
4179 event
->pending_disable
= 0;
4180 __perf_event_disable(event
);
4183 if (event
->pending_wakeup
) {
4184 event
->pending_wakeup
= 0;
4185 perf_event_wakeup(event
);
4190 * We assume there is only KVM supporting the callbacks.
4191 * Later on, we might change it to a list if there is
4192 * another virtualization implementation supporting the callbacks.
4194 struct perf_guest_info_callbacks
*perf_guest_cbs
;
4196 int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks
*cbs
)
4198 perf_guest_cbs
= cbs
;
4201 EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks
);
4203 int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks
*cbs
)
4205 perf_guest_cbs
= NULL
;
4208 EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks
);
4211 perf_output_sample_regs(struct perf_output_handle
*handle
,
4212 struct pt_regs
*regs
, u64 mask
)
4216 for_each_set_bit(bit
, (const unsigned long *) &mask
,
4217 sizeof(mask
) * BITS_PER_BYTE
) {
4220 val
= perf_reg_value(regs
, bit
);
4221 perf_output_put(handle
, val
);
4225 static void perf_sample_regs_user(struct perf_regs_user
*regs_user
,
4226 struct pt_regs
*regs
)
4228 if (!user_mode(regs
)) {
4230 regs
= task_pt_regs(current
);
4236 regs_user
->regs
= regs
;
4237 regs_user
->abi
= perf_reg_abi(current
);
4242 * Get remaining task size from user stack pointer.
4244 * It'd be better to take stack vma map and limit this more
4245 * precisly, but there's no way to get it safely under interrupt,
4246 * so using TASK_SIZE as limit.
4248 static u64
perf_ustack_task_size(struct pt_regs
*regs
)
4250 unsigned long addr
= perf_user_stack_pointer(regs
);
4252 if (!addr
|| addr
>= TASK_SIZE
)
4255 return TASK_SIZE
- addr
;
4259 perf_sample_ustack_size(u16 stack_size
, u16 header_size
,
4260 struct pt_regs
*regs
)
4264 /* No regs, no stack pointer, no dump. */
4269 * Check if we fit in with the requested stack size into the:
4271 * If we don't, we limit the size to the TASK_SIZE.
4273 * - remaining sample size
4274 * If we don't, we customize the stack size to
4275 * fit in to the remaining sample size.
4278 task_size
= min((u64
) USHRT_MAX
, perf_ustack_task_size(regs
));
4279 stack_size
= min(stack_size
, (u16
) task_size
);
4281 /* Current header size plus static size and dynamic size. */
4282 header_size
+= 2 * sizeof(u64
);
4284 /* Do we fit in with the current stack dump size? */
4285 if ((u16
) (header_size
+ stack_size
) < header_size
) {
4287 * If we overflow the maximum size for the sample,
4288 * we customize the stack dump size to fit in.
4290 stack_size
= USHRT_MAX
- header_size
- sizeof(u64
);
4291 stack_size
= round_up(stack_size
, sizeof(u64
));
4298 perf_output_sample_ustack(struct perf_output_handle
*handle
, u64 dump_size
,
4299 struct pt_regs
*regs
)
4301 /* Case of a kernel thread, nothing to dump */
4304 perf_output_put(handle
, size
);
4313 * - the size requested by user or the best one we can fit
4314 * in to the sample max size
4316 * - user stack dump data
4318 * - the actual dumped size
4322 perf_output_put(handle
, dump_size
);
4325 sp
= perf_user_stack_pointer(regs
);
4326 rem
= __output_copy_user(handle
, (void *) sp
, dump_size
);
4327 dyn_size
= dump_size
- rem
;
4329 perf_output_skip(handle
, rem
);
4332 perf_output_put(handle
, dyn_size
);
4336 static void __perf_event_header__init_id(struct perf_event_header
*header
,
4337 struct perf_sample_data
*data
,
4338 struct perf_event
*event
)
4340 u64 sample_type
= event
->attr
.sample_type
;
4342 data
->type
= sample_type
;
4343 header
->size
+= event
->id_header_size
;
4345 if (sample_type
& PERF_SAMPLE_TID
) {
4346 /* namespace issues */
4347 data
->tid_entry
.pid
= perf_event_pid(event
, current
);
4348 data
->tid_entry
.tid
= perf_event_tid(event
, current
);
4351 if (sample_type
& PERF_SAMPLE_TIME
)
4352 data
->time
= perf_clock();
4354 if (sample_type
& (PERF_SAMPLE_ID
| PERF_SAMPLE_IDENTIFIER
))
4355 data
->id
= primary_event_id(event
);
4357 if (sample_type
& PERF_SAMPLE_STREAM_ID
)
4358 data
->stream_id
= event
->id
;
4360 if (sample_type
& PERF_SAMPLE_CPU
) {
4361 data
->cpu_entry
.cpu
= raw_smp_processor_id();
4362 data
->cpu_entry
.reserved
= 0;
4366 void perf_event_header__init_id(struct perf_event_header
*header
,
4367 struct perf_sample_data
*data
,
4368 struct perf_event
*event
)
4370 if (event
->attr
.sample_id_all
)
4371 __perf_event_header__init_id(header
, data
, event
);
4374 static void __perf_event__output_id_sample(struct perf_output_handle
*handle
,
4375 struct perf_sample_data
*data
)
4377 u64 sample_type
= data
->type
;
4379 if (sample_type
& PERF_SAMPLE_TID
)
4380 perf_output_put(handle
, data
->tid_entry
);
4382 if (sample_type
& PERF_SAMPLE_TIME
)
4383 perf_output_put(handle
, data
->time
);
4385 if (sample_type
& PERF_SAMPLE_ID
)
4386 perf_output_put(handle
, data
->id
);
4388 if (sample_type
& PERF_SAMPLE_STREAM_ID
)
4389 perf_output_put(handle
, data
->stream_id
);
4391 if (sample_type
& PERF_SAMPLE_CPU
)
4392 perf_output_put(handle
, data
->cpu_entry
);
4394 if (sample_type
& PERF_SAMPLE_IDENTIFIER
)
4395 perf_output_put(handle
, data
->id
);
4398 void perf_event__output_id_sample(struct perf_event
*event
,
4399 struct perf_output_handle
*handle
,
4400 struct perf_sample_data
*sample
)
4402 if (event
->attr
.sample_id_all
)
4403 __perf_event__output_id_sample(handle
, sample
);
4406 static void perf_output_read_one(struct perf_output_handle
*handle
,
4407 struct perf_event
*event
,
4408 u64 enabled
, u64 running
)
4410 u64 read_format
= event
->attr
.read_format
;
4414 values
[n
++] = perf_event_count(event
);
4415 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
) {
4416 values
[n
++] = enabled
+
4417 atomic64_read(&event
->child_total_time_enabled
);
4419 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
) {
4420 values
[n
++] = running
+
4421 atomic64_read(&event
->child_total_time_running
);
4423 if (read_format
& PERF_FORMAT_ID
)
4424 values
[n
++] = primary_event_id(event
);
4426 __output_copy(handle
, values
, n
* sizeof(u64
));
4430 * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
4432 static void perf_output_read_group(struct perf_output_handle
*handle
,
4433 struct perf_event
*event
,
4434 u64 enabled
, u64 running
)
4436 struct perf_event
*leader
= event
->group_leader
, *sub
;
4437 u64 read_format
= event
->attr
.read_format
;
4441 values
[n
++] = 1 + leader
->nr_siblings
;
4443 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
4444 values
[n
++] = enabled
;
4446 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
4447 values
[n
++] = running
;
4449 if (leader
!= event
)
4450 leader
->pmu
->read(leader
);
4452 values
[n
++] = perf_event_count(leader
);
4453 if (read_format
& PERF_FORMAT_ID
)
4454 values
[n
++] = primary_event_id(leader
);
4456 __output_copy(handle
, values
, n
* sizeof(u64
));
4458 list_for_each_entry(sub
, &leader
->sibling_list
, group_entry
) {
4461 if ((sub
!= event
) &&
4462 (sub
->state
== PERF_EVENT_STATE_ACTIVE
))
4463 sub
->pmu
->read(sub
);
4465 values
[n
++] = perf_event_count(sub
);
4466 if (read_format
& PERF_FORMAT_ID
)
4467 values
[n
++] = primary_event_id(sub
);
4469 __output_copy(handle
, values
, n
* sizeof(u64
));
4473 #define PERF_FORMAT_TOTAL_TIMES (PERF_FORMAT_TOTAL_TIME_ENABLED|\
4474 PERF_FORMAT_TOTAL_TIME_RUNNING)
4476 static void perf_output_read(struct perf_output_handle
*handle
,
4477 struct perf_event
*event
)
4479 u64 enabled
= 0, running
= 0, now
;
4480 u64 read_format
= event
->attr
.read_format
;
4483 * compute total_time_enabled, total_time_running
4484 * based on snapshot values taken when the event
4485 * was last scheduled in.
4487 * we cannot simply called update_context_time()
4488 * because of locking issue as we are called in
4491 if (read_format
& PERF_FORMAT_TOTAL_TIMES
)
4492 calc_timer_values(event
, &now
, &enabled
, &running
);
4494 if (event
->attr
.read_format
& PERF_FORMAT_GROUP
)
4495 perf_output_read_group(handle
, event
, enabled
, running
);
4497 perf_output_read_one(handle
, event
, enabled
, running
);
4500 void perf_output_sample(struct perf_output_handle
*handle
,
4501 struct perf_event_header
*header
,
4502 struct perf_sample_data
*data
,
4503 struct perf_event
*event
)
4505 u64 sample_type
= data
->type
;
4507 perf_output_put(handle
, *header
);
4509 if (sample_type
& PERF_SAMPLE_IDENTIFIER
)
4510 perf_output_put(handle
, data
->id
);
4512 if (sample_type
& PERF_SAMPLE_IP
)
4513 perf_output_put(handle
, data
->ip
);
4515 if (sample_type
& PERF_SAMPLE_TID
)
4516 perf_output_put(handle
, data
->tid_entry
);
4518 if (sample_type
& PERF_SAMPLE_TIME
)
4519 perf_output_put(handle
, data
->time
);
4521 if (sample_type
& PERF_SAMPLE_ADDR
)
4522 perf_output_put(handle
, data
->addr
);
4524 if (sample_type
& PERF_SAMPLE_ID
)
4525 perf_output_put(handle
, data
->id
);
4527 if (sample_type
& PERF_SAMPLE_STREAM_ID
)
4528 perf_output_put(handle
, data
->stream_id
);
4530 if (sample_type
& PERF_SAMPLE_CPU
)
4531 perf_output_put(handle
, data
->cpu_entry
);
4533 if (sample_type
& PERF_SAMPLE_PERIOD
)
4534 perf_output_put(handle
, data
->period
);
4536 if (sample_type
& PERF_SAMPLE_READ
)
4537 perf_output_read(handle
, event
);
4539 if (sample_type
& PERF_SAMPLE_CALLCHAIN
) {
4540 if (data
->callchain
) {
4543 if (data
->callchain
)
4544 size
+= data
->callchain
->nr
;
4546 size
*= sizeof(u64
);
4548 __output_copy(handle
, data
->callchain
, size
);
4551 perf_output_put(handle
, nr
);
4555 if (sample_type
& PERF_SAMPLE_RAW
) {
4557 perf_output_put(handle
, data
->raw
->size
);
4558 __output_copy(handle
, data
->raw
->data
,
4565 .size
= sizeof(u32
),
4568 perf_output_put(handle
, raw
);
4572 if (sample_type
& PERF_SAMPLE_BRANCH_STACK
) {
4573 if (data
->br_stack
) {
4576 size
= data
->br_stack
->nr
4577 * sizeof(struct perf_branch_entry
);
4579 perf_output_put(handle
, data
->br_stack
->nr
);
4580 perf_output_copy(handle
, data
->br_stack
->entries
, size
);
4583 * we always store at least the value of nr
4586 perf_output_put(handle
, nr
);
4590 if (sample_type
& PERF_SAMPLE_REGS_USER
) {
4591 u64 abi
= data
->regs_user
.abi
;
4594 * If there are no regs to dump, notice it through
4595 * first u64 being zero (PERF_SAMPLE_REGS_ABI_NONE).
4597 perf_output_put(handle
, abi
);
4600 u64 mask
= event
->attr
.sample_regs_user
;
4601 perf_output_sample_regs(handle
,
4602 data
->regs_user
.regs
,
4607 if (sample_type
& PERF_SAMPLE_STACK_USER
) {
4608 perf_output_sample_ustack(handle
,
4609 data
->stack_user_size
,
4610 data
->regs_user
.regs
);
4613 if (sample_type
& PERF_SAMPLE_WEIGHT
)
4614 perf_output_put(handle
, data
->weight
);
4616 if (sample_type
& PERF_SAMPLE_DATA_SRC
)
4617 perf_output_put(handle
, data
->data_src
.val
);
4619 if (sample_type
& PERF_SAMPLE_TRANSACTION
)
4620 perf_output_put(handle
, data
->txn
);
4622 if (!event
->attr
.watermark
) {
4623 int wakeup_events
= event
->attr
.wakeup_events
;
4625 if (wakeup_events
) {
4626 struct ring_buffer
*rb
= handle
->rb
;
4627 int events
= local_inc_return(&rb
->events
);
4629 if (events
>= wakeup_events
) {
4630 local_sub(wakeup_events
, &rb
->events
);
4631 local_inc(&rb
->wakeup
);
4637 void perf_prepare_sample(struct perf_event_header
*header
,
4638 struct perf_sample_data
*data
,
4639 struct perf_event
*event
,
4640 struct pt_regs
*regs
)
4642 u64 sample_type
= event
->attr
.sample_type
;
4644 header
->type
= PERF_RECORD_SAMPLE
;
4645 header
->size
= sizeof(*header
) + event
->header_size
;
4648 header
->misc
|= perf_misc_flags(regs
);
4650 __perf_event_header__init_id(header
, data
, event
);
4652 if (sample_type
& PERF_SAMPLE_IP
)
4653 data
->ip
= perf_instruction_pointer(regs
);
4655 if (sample_type
& PERF_SAMPLE_CALLCHAIN
) {
4658 data
->callchain
= perf_callchain(event
, regs
);
4660 if (data
->callchain
)
4661 size
+= data
->callchain
->nr
;
4663 header
->size
+= size
* sizeof(u64
);
4666 if (sample_type
& PERF_SAMPLE_RAW
) {
4667 int size
= sizeof(u32
);
4670 size
+= data
->raw
->size
;
4672 size
+= sizeof(u32
);
4674 WARN_ON_ONCE(size
& (sizeof(u64
)-1));
4675 header
->size
+= size
;
4678 if (sample_type
& PERF_SAMPLE_BRANCH_STACK
) {
4679 int size
= sizeof(u64
); /* nr */
4680 if (data
->br_stack
) {
4681 size
+= data
->br_stack
->nr
4682 * sizeof(struct perf_branch_entry
);
4684 header
->size
+= size
;
4687 if (sample_type
& PERF_SAMPLE_REGS_USER
) {
4688 /* regs dump ABI info */
4689 int size
= sizeof(u64
);
4691 perf_sample_regs_user(&data
->regs_user
, regs
);
4693 if (data
->regs_user
.regs
) {
4694 u64 mask
= event
->attr
.sample_regs_user
;
4695 size
+= hweight64(mask
) * sizeof(u64
);
4698 header
->size
+= size
;
4701 if (sample_type
& PERF_SAMPLE_STACK_USER
) {
4703 * Either we need PERF_SAMPLE_STACK_USER bit to be allways
4704 * processed as the last one or have additional check added
4705 * in case new sample type is added, because we could eat
4706 * up the rest of the sample size.
4708 struct perf_regs_user
*uregs
= &data
->regs_user
;
4709 u16 stack_size
= event
->attr
.sample_stack_user
;
4710 u16 size
= sizeof(u64
);
4713 perf_sample_regs_user(uregs
, regs
);
4715 stack_size
= perf_sample_ustack_size(stack_size
, header
->size
,
4719 * If there is something to dump, add space for the dump
4720 * itself and for the field that tells the dynamic size,
4721 * which is how many have been actually dumped.
4724 size
+= sizeof(u64
) + stack_size
;
4726 data
->stack_user_size
= stack_size
;
4727 header
->size
+= size
;
4731 static void perf_event_output(struct perf_event
*event
,
4732 struct perf_sample_data
*data
,
4733 struct pt_regs
*regs
)
4735 struct perf_output_handle handle
;
4736 struct perf_event_header header
;
4738 /* protect the callchain buffers */
4741 perf_prepare_sample(&header
, data
, event
, regs
);
4743 if (perf_output_begin(&handle
, event
, header
.size
))
4746 perf_output_sample(&handle
, &header
, data
, event
);
4748 perf_output_end(&handle
);
4758 struct perf_read_event
{
4759 struct perf_event_header header
;
4766 perf_event_read_event(struct perf_event
*event
,
4767 struct task_struct
*task
)
4769 struct perf_output_handle handle
;
4770 struct perf_sample_data sample
;
4771 struct perf_read_event read_event
= {
4773 .type
= PERF_RECORD_READ
,
4775 .size
= sizeof(read_event
) + event
->read_size
,
4777 .pid
= perf_event_pid(event
, task
),
4778 .tid
= perf_event_tid(event
, task
),
4782 perf_event_header__init_id(&read_event
.header
, &sample
, event
);
4783 ret
= perf_output_begin(&handle
, event
, read_event
.header
.size
);
4787 perf_output_put(&handle
, read_event
);
4788 perf_output_read(&handle
, event
);
4789 perf_event__output_id_sample(event
, &handle
, &sample
);
4791 perf_output_end(&handle
);
4794 typedef void (perf_event_aux_output_cb
)(struct perf_event
*event
, void *data
);
4797 perf_event_aux_ctx(struct perf_event_context
*ctx
,
4798 perf_event_aux_output_cb output
,
4801 struct perf_event
*event
;
4803 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
4804 if (event
->state
< PERF_EVENT_STATE_INACTIVE
)
4806 if (!event_filter_match(event
))
4808 output(event
, data
);
4813 perf_event_aux(perf_event_aux_output_cb output
, void *data
,
4814 struct perf_event_context
*task_ctx
)
4816 struct perf_cpu_context
*cpuctx
;
4817 struct perf_event_context
*ctx
;
4822 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
4823 cpuctx
= get_cpu_ptr(pmu
->pmu_cpu_context
);
4824 if (cpuctx
->unique_pmu
!= pmu
)
4826 perf_event_aux_ctx(&cpuctx
->ctx
, output
, data
);
4829 ctxn
= pmu
->task_ctx_nr
;
4832 ctx
= rcu_dereference(current
->perf_event_ctxp
[ctxn
]);
4834 perf_event_aux_ctx(ctx
, output
, data
);
4836 put_cpu_ptr(pmu
->pmu_cpu_context
);
4841 perf_event_aux_ctx(task_ctx
, output
, data
);
4848 * task tracking -- fork/exit
4850 * enabled by: attr.comm | attr.mmap | attr.mmap2 | attr.mmap_data | attr.task
4853 struct perf_task_event
{
4854 struct task_struct
*task
;
4855 struct perf_event_context
*task_ctx
;
4858 struct perf_event_header header
;
4868 static int perf_event_task_match(struct perf_event
*event
)
4870 return event
->attr
.comm
|| event
->attr
.mmap
||
4871 event
->attr
.mmap2
|| event
->attr
.mmap_data
||
4875 static void perf_event_task_output(struct perf_event
*event
,
4878 struct perf_task_event
*task_event
= data
;
4879 struct perf_output_handle handle
;
4880 struct perf_sample_data sample
;
4881 struct task_struct
*task
= task_event
->task
;
4882 int ret
, size
= task_event
->event_id
.header
.size
;
4884 if (!perf_event_task_match(event
))
4887 perf_event_header__init_id(&task_event
->event_id
.header
, &sample
, event
);
4889 ret
= perf_output_begin(&handle
, event
,
4890 task_event
->event_id
.header
.size
);
4894 task_event
->event_id
.pid
= perf_event_pid(event
, task
);
4895 task_event
->event_id
.ppid
= perf_event_pid(event
, current
);
4897 task_event
->event_id
.tid
= perf_event_tid(event
, task
);
4898 task_event
->event_id
.ptid
= perf_event_tid(event
, current
);
4900 perf_output_put(&handle
, task_event
->event_id
);
4902 perf_event__output_id_sample(event
, &handle
, &sample
);
4904 perf_output_end(&handle
);
4906 task_event
->event_id
.header
.size
= size
;
4909 static void perf_event_task(struct task_struct
*task
,
4910 struct perf_event_context
*task_ctx
,
4913 struct perf_task_event task_event
;
4915 if (!atomic_read(&nr_comm_events
) &&
4916 !atomic_read(&nr_mmap_events
) &&
4917 !atomic_read(&nr_task_events
))
4920 task_event
= (struct perf_task_event
){
4922 .task_ctx
= task_ctx
,
4925 .type
= new ? PERF_RECORD_FORK
: PERF_RECORD_EXIT
,
4927 .size
= sizeof(task_event
.event_id
),
4933 .time
= perf_clock(),
4937 perf_event_aux(perf_event_task_output
,
4942 void perf_event_fork(struct task_struct
*task
)
4944 perf_event_task(task
, NULL
, 1);
4951 struct perf_comm_event
{
4952 struct task_struct
*task
;
4957 struct perf_event_header header
;
4964 static int perf_event_comm_match(struct perf_event
*event
)
4966 return event
->attr
.comm
;
4969 static void perf_event_comm_output(struct perf_event
*event
,
4972 struct perf_comm_event
*comm_event
= data
;
4973 struct perf_output_handle handle
;
4974 struct perf_sample_data sample
;
4975 int size
= comm_event
->event_id
.header
.size
;
4978 if (!perf_event_comm_match(event
))
4981 perf_event_header__init_id(&comm_event
->event_id
.header
, &sample
, event
);
4982 ret
= perf_output_begin(&handle
, event
,
4983 comm_event
->event_id
.header
.size
);
4988 comm_event
->event_id
.pid
= perf_event_pid(event
, comm_event
->task
);
4989 comm_event
->event_id
.tid
= perf_event_tid(event
, comm_event
->task
);
4991 perf_output_put(&handle
, comm_event
->event_id
);
4992 __output_copy(&handle
, comm_event
->comm
,
4993 comm_event
->comm_size
);
4995 perf_event__output_id_sample(event
, &handle
, &sample
);
4997 perf_output_end(&handle
);
4999 comm_event
->event_id
.header
.size
= size
;
5002 static void perf_event_comm_event(struct perf_comm_event
*comm_event
)
5004 char comm
[TASK_COMM_LEN
];
5007 memset(comm
, 0, sizeof(comm
));
5008 strlcpy(comm
, comm_event
->task
->comm
, sizeof(comm
));
5009 size
= ALIGN(strlen(comm
)+1, sizeof(u64
));
5011 comm_event
->comm
= comm
;
5012 comm_event
->comm_size
= size
;
5014 comm_event
->event_id
.header
.size
= sizeof(comm_event
->event_id
) + size
;
5016 perf_event_aux(perf_event_comm_output
,
5021 void perf_event_comm(struct task_struct
*task
)
5023 struct perf_comm_event comm_event
;
5024 struct perf_event_context
*ctx
;
5028 for_each_task_context_nr(ctxn
) {
5029 ctx
= task
->perf_event_ctxp
[ctxn
];
5033 perf_event_enable_on_exec(ctx
);
5037 if (!atomic_read(&nr_comm_events
))
5040 comm_event
= (struct perf_comm_event
){
5046 .type
= PERF_RECORD_COMM
,
5055 perf_event_comm_event(&comm_event
);
5062 struct perf_mmap_event
{
5063 struct vm_area_struct
*vma
;
5065 const char *file_name
;
5072 struct perf_event_header header
;
5082 static int perf_event_mmap_match(struct perf_event
*event
,
5085 struct perf_mmap_event
*mmap_event
= data
;
5086 struct vm_area_struct
*vma
= mmap_event
->vma
;
5087 int executable
= vma
->vm_flags
& VM_EXEC
;
5089 return (!executable
&& event
->attr
.mmap_data
) ||
5090 (executable
&& (event
->attr
.mmap
|| event
->attr
.mmap2
));
5093 static void perf_event_mmap_output(struct perf_event
*event
,
5096 struct perf_mmap_event
*mmap_event
= data
;
5097 struct perf_output_handle handle
;
5098 struct perf_sample_data sample
;
5099 int size
= mmap_event
->event_id
.header
.size
;
5102 if (!perf_event_mmap_match(event
, data
))
5105 if (event
->attr
.mmap2
) {
5106 mmap_event
->event_id
.header
.type
= PERF_RECORD_MMAP2
;
5107 mmap_event
->event_id
.header
.size
+= sizeof(mmap_event
->maj
);
5108 mmap_event
->event_id
.header
.size
+= sizeof(mmap_event
->min
);
5109 mmap_event
->event_id
.header
.size
+= sizeof(mmap_event
->ino
);
5110 mmap_event
->event_id
.header
.size
+= sizeof(mmap_event
->ino_generation
);
5113 perf_event_header__init_id(&mmap_event
->event_id
.header
, &sample
, event
);
5114 ret
= perf_output_begin(&handle
, event
,
5115 mmap_event
->event_id
.header
.size
);
5119 mmap_event
->event_id
.pid
= perf_event_pid(event
, current
);
5120 mmap_event
->event_id
.tid
= perf_event_tid(event
, current
);
5122 perf_output_put(&handle
, mmap_event
->event_id
);
5124 if (event
->attr
.mmap2
) {
5125 perf_output_put(&handle
, mmap_event
->maj
);
5126 perf_output_put(&handle
, mmap_event
->min
);
5127 perf_output_put(&handle
, mmap_event
->ino
);
5128 perf_output_put(&handle
, mmap_event
->ino_generation
);
5131 __output_copy(&handle
, mmap_event
->file_name
,
5132 mmap_event
->file_size
);
5134 perf_event__output_id_sample(event
, &handle
, &sample
);
5136 perf_output_end(&handle
);
5138 mmap_event
->event_id
.header
.size
= size
;
5141 static void perf_event_mmap_event(struct perf_mmap_event
*mmap_event
)
5143 struct vm_area_struct
*vma
= mmap_event
->vma
;
5144 struct file
*file
= vma
->vm_file
;
5145 int maj
= 0, min
= 0;
5146 u64 ino
= 0, gen
= 0;
5153 struct inode
*inode
;
5156 buf
= kmalloc(PATH_MAX
, GFP_KERNEL
);
5162 * d_path() works from the end of the rb backwards, so we
5163 * need to add enough zero bytes after the string to handle
5164 * the 64bit alignment we do later.
5166 name
= d_path(&file
->f_path
, buf
, PATH_MAX
- sizeof(u64
));
5171 inode
= file_inode(vma
->vm_file
);
5172 dev
= inode
->i_sb
->s_dev
;
5174 gen
= inode
->i_generation
;
5179 name
= (char *)arch_vma_name(vma
);
5183 if (vma
->vm_start
<= vma
->vm_mm
->start_brk
&&
5184 vma
->vm_end
>= vma
->vm_mm
->brk
) {
5188 if (vma
->vm_start
<= vma
->vm_mm
->start_stack
&&
5189 vma
->vm_end
>= vma
->vm_mm
->start_stack
) {
5199 strlcpy(tmp
, name
, sizeof(tmp
));
5203 * Since our buffer works in 8 byte units we need to align our string
5204 * size to a multiple of 8. However, we must guarantee the tail end is
5205 * zero'd out to avoid leaking random bits to userspace.
5207 size
= strlen(name
)+1;
5208 while (!IS_ALIGNED(size
, sizeof(u64
)))
5209 name
[size
++] = '\0';
5211 mmap_event
->file_name
= name
;
5212 mmap_event
->file_size
= size
;
5213 mmap_event
->maj
= maj
;
5214 mmap_event
->min
= min
;
5215 mmap_event
->ino
= ino
;
5216 mmap_event
->ino_generation
= gen
;
5218 if (!(vma
->vm_flags
& VM_EXEC
))
5219 mmap_event
->event_id
.header
.misc
|= PERF_RECORD_MISC_MMAP_DATA
;
5221 mmap_event
->event_id
.header
.size
= sizeof(mmap_event
->event_id
) + size
;
5223 perf_event_aux(perf_event_mmap_output
,
5230 void perf_event_mmap(struct vm_area_struct
*vma
)
5232 struct perf_mmap_event mmap_event
;
5234 if (!atomic_read(&nr_mmap_events
))
5237 mmap_event
= (struct perf_mmap_event
){
5243 .type
= PERF_RECORD_MMAP
,
5244 .misc
= PERF_RECORD_MISC_USER
,
5249 .start
= vma
->vm_start
,
5250 .len
= vma
->vm_end
- vma
->vm_start
,
5251 .pgoff
= (u64
)vma
->vm_pgoff
<< PAGE_SHIFT
,
5253 /* .maj (attr_mmap2 only) */
5254 /* .min (attr_mmap2 only) */
5255 /* .ino (attr_mmap2 only) */
5256 /* .ino_generation (attr_mmap2 only) */
5259 perf_event_mmap_event(&mmap_event
);
5263 * IRQ throttle logging
5266 static void perf_log_throttle(struct perf_event
*event
, int enable
)
5268 struct perf_output_handle handle
;
5269 struct perf_sample_data sample
;
5273 struct perf_event_header header
;
5277 } throttle_event
= {
5279 .type
= PERF_RECORD_THROTTLE
,
5281 .size
= sizeof(throttle_event
),
5283 .time
= perf_clock(),
5284 .id
= primary_event_id(event
),
5285 .stream_id
= event
->id
,
5289 throttle_event
.header
.type
= PERF_RECORD_UNTHROTTLE
;
5291 perf_event_header__init_id(&throttle_event
.header
, &sample
, event
);
5293 ret
= perf_output_begin(&handle
, event
,
5294 throttle_event
.header
.size
);
5298 perf_output_put(&handle
, throttle_event
);
5299 perf_event__output_id_sample(event
, &handle
, &sample
);
5300 perf_output_end(&handle
);
5304 * Generic event overflow handling, sampling.
5307 static int __perf_event_overflow(struct perf_event
*event
,
5308 int throttle
, struct perf_sample_data
*data
,
5309 struct pt_regs
*regs
)
5311 int events
= atomic_read(&event
->event_limit
);
5312 struct hw_perf_event
*hwc
= &event
->hw
;
5317 * Non-sampling counters might still use the PMI to fold short
5318 * hardware counters, ignore those.
5320 if (unlikely(!is_sampling_event(event
)))
5323 seq
= __this_cpu_read(perf_throttled_seq
);
5324 if (seq
!= hwc
->interrupts_seq
) {
5325 hwc
->interrupts_seq
= seq
;
5326 hwc
->interrupts
= 1;
5329 if (unlikely(throttle
5330 && hwc
->interrupts
>= max_samples_per_tick
)) {
5331 __this_cpu_inc(perf_throttled_count
);
5332 hwc
->interrupts
= MAX_INTERRUPTS
;
5333 perf_log_throttle(event
, 0);
5334 tick_nohz_full_kick();
5339 if (event
->attr
.freq
) {
5340 u64 now
= perf_clock();
5341 s64 delta
= now
- hwc
->freq_time_stamp
;
5343 hwc
->freq_time_stamp
= now
;
5345 if (delta
> 0 && delta
< 2*TICK_NSEC
)
5346 perf_adjust_period(event
, delta
, hwc
->last_period
, true);
5350 * XXX event_limit might not quite work as expected on inherited
5354 event
->pending_kill
= POLL_IN
;
5355 if (events
&& atomic_dec_and_test(&event
->event_limit
)) {
5357 event
->pending_kill
= POLL_HUP
;
5358 event
->pending_disable
= 1;
5359 irq_work_queue(&event
->pending
);
5362 if (event
->overflow_handler
)
5363 event
->overflow_handler(event
, data
, regs
);
5365 perf_event_output(event
, data
, regs
);
5367 if (event
->fasync
&& event
->pending_kill
) {
5368 event
->pending_wakeup
= 1;
5369 irq_work_queue(&event
->pending
);
5375 int perf_event_overflow(struct perf_event
*event
,
5376 struct perf_sample_data
*data
,
5377 struct pt_regs
*regs
)
5379 return __perf_event_overflow(event
, 1, data
, regs
);
5383 * Generic software event infrastructure
5386 struct swevent_htable
{
5387 struct swevent_hlist
*swevent_hlist
;
5388 struct mutex hlist_mutex
;
5391 /* Recursion avoidance in each contexts */
5392 int recursion
[PERF_NR_CONTEXTS
];
5395 static DEFINE_PER_CPU(struct swevent_htable
, swevent_htable
);
5398 * We directly increment event->count and keep a second value in
5399 * event->hw.period_left to count intervals. This period event
5400 * is kept in the range [-sample_period, 0] so that we can use the
5404 u64
perf_swevent_set_period(struct perf_event
*event
)
5406 struct hw_perf_event
*hwc
= &event
->hw
;
5407 u64 period
= hwc
->last_period
;
5411 hwc
->last_period
= hwc
->sample_period
;
5414 old
= val
= local64_read(&hwc
->period_left
);
5418 nr
= div64_u64(period
+ val
, period
);
5419 offset
= nr
* period
;
5421 if (local64_cmpxchg(&hwc
->period_left
, old
, val
) != old
)
5427 static void perf_swevent_overflow(struct perf_event
*event
, u64 overflow
,
5428 struct perf_sample_data
*data
,
5429 struct pt_regs
*regs
)
5431 struct hw_perf_event
*hwc
= &event
->hw
;
5435 overflow
= perf_swevent_set_period(event
);
5437 if (hwc
->interrupts
== MAX_INTERRUPTS
)
5440 for (; overflow
; overflow
--) {
5441 if (__perf_event_overflow(event
, throttle
,
5444 * We inhibit the overflow from happening when
5445 * hwc->interrupts == MAX_INTERRUPTS.
5453 static void perf_swevent_event(struct perf_event
*event
, u64 nr
,
5454 struct perf_sample_data
*data
,
5455 struct pt_regs
*regs
)
5457 struct hw_perf_event
*hwc
= &event
->hw
;
5459 local64_add(nr
, &event
->count
);
5464 if (!is_sampling_event(event
))
5467 if ((event
->attr
.sample_type
& PERF_SAMPLE_PERIOD
) && !event
->attr
.freq
) {
5469 return perf_swevent_overflow(event
, 1, data
, regs
);
5471 data
->period
= event
->hw
.last_period
;
5473 if (nr
== 1 && hwc
->sample_period
== 1 && !event
->attr
.freq
)
5474 return perf_swevent_overflow(event
, 1, data
, regs
);
5476 if (local64_add_negative(nr
, &hwc
->period_left
))
5479 perf_swevent_overflow(event
, 0, data
, regs
);
5482 static int perf_exclude_event(struct perf_event
*event
,
5483 struct pt_regs
*regs
)
5485 if (event
->hw
.state
& PERF_HES_STOPPED
)
5489 if (event
->attr
.exclude_user
&& user_mode(regs
))
5492 if (event
->attr
.exclude_kernel
&& !user_mode(regs
))
5499 static int perf_swevent_match(struct perf_event
*event
,
5500 enum perf_type_id type
,
5502 struct perf_sample_data
*data
,
5503 struct pt_regs
*regs
)
5505 if (event
->attr
.type
!= type
)
5508 if (event
->attr
.config
!= event_id
)
5511 if (perf_exclude_event(event
, regs
))
5517 static inline u64
swevent_hash(u64 type
, u32 event_id
)
5519 u64 val
= event_id
| (type
<< 32);
5521 return hash_64(val
, SWEVENT_HLIST_BITS
);
5524 static inline struct hlist_head
*
5525 __find_swevent_head(struct swevent_hlist
*hlist
, u64 type
, u32 event_id
)
5527 u64 hash
= swevent_hash(type
, event_id
);
5529 return &hlist
->heads
[hash
];
5532 /* For the read side: events when they trigger */
5533 static inline struct hlist_head
*
5534 find_swevent_head_rcu(struct swevent_htable
*swhash
, u64 type
, u32 event_id
)
5536 struct swevent_hlist
*hlist
;
5538 hlist
= rcu_dereference(swhash
->swevent_hlist
);
5542 return __find_swevent_head(hlist
, type
, event_id
);
5545 /* For the event head insertion and removal in the hlist */
5546 static inline struct hlist_head
*
5547 find_swevent_head(struct swevent_htable
*swhash
, struct perf_event
*event
)
5549 struct swevent_hlist
*hlist
;
5550 u32 event_id
= event
->attr
.config
;
5551 u64 type
= event
->attr
.type
;
5554 * Event scheduling is always serialized against hlist allocation
5555 * and release. Which makes the protected version suitable here.
5556 * The context lock guarantees that.
5558 hlist
= rcu_dereference_protected(swhash
->swevent_hlist
,
5559 lockdep_is_held(&event
->ctx
->lock
));
5563 return __find_swevent_head(hlist
, type
, event_id
);
5566 static void do_perf_sw_event(enum perf_type_id type
, u32 event_id
,
5568 struct perf_sample_data
*data
,
5569 struct pt_regs
*regs
)
5571 struct swevent_htable
*swhash
= &__get_cpu_var(swevent_htable
);
5572 struct perf_event
*event
;
5573 struct hlist_head
*head
;
5576 head
= find_swevent_head_rcu(swhash
, type
, event_id
);
5580 hlist_for_each_entry_rcu(event
, head
, hlist_entry
) {
5581 if (perf_swevent_match(event
, type
, event_id
, data
, regs
))
5582 perf_swevent_event(event
, nr
, data
, regs
);
5588 int perf_swevent_get_recursion_context(void)
5590 struct swevent_htable
*swhash
= &__get_cpu_var(swevent_htable
);
5592 return get_recursion_context(swhash
->recursion
);
5594 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context
);
5596 inline void perf_swevent_put_recursion_context(int rctx
)
5598 struct swevent_htable
*swhash
= &__get_cpu_var(swevent_htable
);
5600 put_recursion_context(swhash
->recursion
, rctx
);
5603 void __perf_sw_event(u32 event_id
, u64 nr
, struct pt_regs
*regs
, u64 addr
)
5605 struct perf_sample_data data
;
5608 preempt_disable_notrace();
5609 rctx
= perf_swevent_get_recursion_context();
5613 perf_sample_data_init(&data
, addr
, 0);
5615 do_perf_sw_event(PERF_TYPE_SOFTWARE
, event_id
, nr
, &data
, regs
);
5617 perf_swevent_put_recursion_context(rctx
);
5618 preempt_enable_notrace();
5621 static void perf_swevent_read(struct perf_event
*event
)
5625 static int perf_swevent_add(struct perf_event
*event
, int flags
)
5627 struct swevent_htable
*swhash
= &__get_cpu_var(swevent_htable
);
5628 struct hw_perf_event
*hwc
= &event
->hw
;
5629 struct hlist_head
*head
;
5631 if (is_sampling_event(event
)) {
5632 hwc
->last_period
= hwc
->sample_period
;
5633 perf_swevent_set_period(event
);
5636 hwc
->state
= !(flags
& PERF_EF_START
);
5638 head
= find_swevent_head(swhash
, event
);
5639 if (WARN_ON_ONCE(!head
))
5642 hlist_add_head_rcu(&event
->hlist_entry
, head
);
5647 static void perf_swevent_del(struct perf_event
*event
, int flags
)
5649 hlist_del_rcu(&event
->hlist_entry
);
5652 static void perf_swevent_start(struct perf_event
*event
, int flags
)
5654 event
->hw
.state
= 0;
5657 static void perf_swevent_stop(struct perf_event
*event
, int flags
)
5659 event
->hw
.state
= PERF_HES_STOPPED
;
5662 /* Deref the hlist from the update side */
5663 static inline struct swevent_hlist
*
5664 swevent_hlist_deref(struct swevent_htable
*swhash
)
5666 return rcu_dereference_protected(swhash
->swevent_hlist
,
5667 lockdep_is_held(&swhash
->hlist_mutex
));
5670 static void swevent_hlist_release(struct swevent_htable
*swhash
)
5672 struct swevent_hlist
*hlist
= swevent_hlist_deref(swhash
);
5677 rcu_assign_pointer(swhash
->swevent_hlist
, NULL
);
5678 kfree_rcu(hlist
, rcu_head
);
5681 static void swevent_hlist_put_cpu(struct perf_event
*event
, int cpu
)
5683 struct swevent_htable
*swhash
= &per_cpu(swevent_htable
, cpu
);
5685 mutex_lock(&swhash
->hlist_mutex
);
5687 if (!--swhash
->hlist_refcount
)
5688 swevent_hlist_release(swhash
);
5690 mutex_unlock(&swhash
->hlist_mutex
);
5693 static void swevent_hlist_put(struct perf_event
*event
)
5697 for_each_possible_cpu(cpu
)
5698 swevent_hlist_put_cpu(event
, cpu
);
5701 static int swevent_hlist_get_cpu(struct perf_event
*event
, int cpu
)
5703 struct swevent_htable
*swhash
= &per_cpu(swevent_htable
, cpu
);
5706 mutex_lock(&swhash
->hlist_mutex
);
5708 if (!swevent_hlist_deref(swhash
) && cpu_online(cpu
)) {
5709 struct swevent_hlist
*hlist
;
5711 hlist
= kzalloc(sizeof(*hlist
), GFP_KERNEL
);
5716 rcu_assign_pointer(swhash
->swevent_hlist
, hlist
);
5718 swhash
->hlist_refcount
++;
5720 mutex_unlock(&swhash
->hlist_mutex
);
5725 static int swevent_hlist_get(struct perf_event
*event
)
5728 int cpu
, failed_cpu
;
5731 for_each_possible_cpu(cpu
) {
5732 err
= swevent_hlist_get_cpu(event
, cpu
);
5742 for_each_possible_cpu(cpu
) {
5743 if (cpu
== failed_cpu
)
5745 swevent_hlist_put_cpu(event
, cpu
);
5752 struct static_key perf_swevent_enabled
[PERF_COUNT_SW_MAX
];
5754 static void sw_perf_event_destroy(struct perf_event
*event
)
5756 u64 event_id
= event
->attr
.config
;
5758 WARN_ON(event
->parent
);
5760 static_key_slow_dec(&perf_swevent_enabled
[event_id
]);
5761 swevent_hlist_put(event
);
5764 static int perf_swevent_init(struct perf_event
*event
)
5766 u64 event_id
= event
->attr
.config
;
5768 if (event
->attr
.type
!= PERF_TYPE_SOFTWARE
)
5772 * no branch sampling for software events
5774 if (has_branch_stack(event
))
5778 case PERF_COUNT_SW_CPU_CLOCK
:
5779 case PERF_COUNT_SW_TASK_CLOCK
:
5786 if (event_id
>= PERF_COUNT_SW_MAX
)
5789 if (!event
->parent
) {
5792 err
= swevent_hlist_get(event
);
5796 static_key_slow_inc(&perf_swevent_enabled
[event_id
]);
5797 event
->destroy
= sw_perf_event_destroy
;
5803 static int perf_swevent_event_idx(struct perf_event
*event
)
5808 static struct pmu perf_swevent
= {
5809 .task_ctx_nr
= perf_sw_context
,
5811 .event_init
= perf_swevent_init
,
5812 .add
= perf_swevent_add
,
5813 .del
= perf_swevent_del
,
5814 .start
= perf_swevent_start
,
5815 .stop
= perf_swevent_stop
,
5816 .read
= perf_swevent_read
,
5818 .event_idx
= perf_swevent_event_idx
,
5821 #ifdef CONFIG_EVENT_TRACING
5823 static int perf_tp_filter_match(struct perf_event
*event
,
5824 struct perf_sample_data
*data
)
5826 void *record
= data
->raw
->data
;
5828 if (likely(!event
->filter
) || filter_match_preds(event
->filter
, record
))
5833 static int perf_tp_event_match(struct perf_event
*event
,
5834 struct perf_sample_data
*data
,
5835 struct pt_regs
*regs
)
5837 if (event
->hw
.state
& PERF_HES_STOPPED
)
5840 * All tracepoints are from kernel-space.
5842 if (event
->attr
.exclude_kernel
)
5845 if (!perf_tp_filter_match(event
, data
))
5851 void perf_tp_event(u64 addr
, u64 count
, void *record
, int entry_size
,
5852 struct pt_regs
*regs
, struct hlist_head
*head
, int rctx
,
5853 struct task_struct
*task
)
5855 struct perf_sample_data data
;
5856 struct perf_event
*event
;
5858 struct perf_raw_record raw
= {
5863 perf_sample_data_init(&data
, addr
, 0);
5866 hlist_for_each_entry_rcu(event
, head
, hlist_entry
) {
5867 if (perf_tp_event_match(event
, &data
, regs
))
5868 perf_swevent_event(event
, count
, &data
, regs
);
5872 * If we got specified a target task, also iterate its context and
5873 * deliver this event there too.
5875 if (task
&& task
!= current
) {
5876 struct perf_event_context
*ctx
;
5877 struct trace_entry
*entry
= record
;
5880 ctx
= rcu_dereference(task
->perf_event_ctxp
[perf_sw_context
]);
5884 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
5885 if (event
->attr
.type
!= PERF_TYPE_TRACEPOINT
)
5887 if (event
->attr
.config
!= entry
->type
)
5889 if (perf_tp_event_match(event
, &data
, regs
))
5890 perf_swevent_event(event
, count
, &data
, regs
);
5896 perf_swevent_put_recursion_context(rctx
);
5898 EXPORT_SYMBOL_GPL(perf_tp_event
);
5900 static void tp_perf_event_destroy(struct perf_event
*event
)
5902 perf_trace_destroy(event
);
5905 static int perf_tp_event_init(struct perf_event
*event
)
5909 if (event
->attr
.type
!= PERF_TYPE_TRACEPOINT
)
5913 * no branch sampling for tracepoint events
5915 if (has_branch_stack(event
))
5918 err
= perf_trace_init(event
);
5922 event
->destroy
= tp_perf_event_destroy
;
5927 static struct pmu perf_tracepoint
= {
5928 .task_ctx_nr
= perf_sw_context
,
5930 .event_init
= perf_tp_event_init
,
5931 .add
= perf_trace_add
,
5932 .del
= perf_trace_del
,
5933 .start
= perf_swevent_start
,
5934 .stop
= perf_swevent_stop
,
5935 .read
= perf_swevent_read
,
5937 .event_idx
= perf_swevent_event_idx
,
5940 static inline void perf_tp_register(void)
5942 perf_pmu_register(&perf_tracepoint
, "tracepoint", PERF_TYPE_TRACEPOINT
);
5945 static int perf_event_set_filter(struct perf_event
*event
, void __user
*arg
)
5950 if (event
->attr
.type
!= PERF_TYPE_TRACEPOINT
)
5953 filter_str
= strndup_user(arg
, PAGE_SIZE
);
5954 if (IS_ERR(filter_str
))
5955 return PTR_ERR(filter_str
);
5957 ret
= ftrace_profile_set_filter(event
, event
->attr
.config
, filter_str
);
5963 static void perf_event_free_filter(struct perf_event
*event
)
5965 ftrace_profile_free_filter(event
);
5970 static inline void perf_tp_register(void)
5974 static int perf_event_set_filter(struct perf_event
*event
, void __user
*arg
)
5979 static void perf_event_free_filter(struct perf_event
*event
)
5983 #endif /* CONFIG_EVENT_TRACING */
5985 #ifdef CONFIG_HAVE_HW_BREAKPOINT
5986 void perf_bp_event(struct perf_event
*bp
, void *data
)
5988 struct perf_sample_data sample
;
5989 struct pt_regs
*regs
= data
;
5991 perf_sample_data_init(&sample
, bp
->attr
.bp_addr
, 0);
5993 if (!bp
->hw
.state
&& !perf_exclude_event(bp
, regs
))
5994 perf_swevent_event(bp
, 1, &sample
, regs
);
5999 * hrtimer based swevent callback
6002 static enum hrtimer_restart
perf_swevent_hrtimer(struct hrtimer
*hrtimer
)
6004 enum hrtimer_restart ret
= HRTIMER_RESTART
;
6005 struct perf_sample_data data
;
6006 struct pt_regs
*regs
;
6007 struct perf_event
*event
;
6010 event
= container_of(hrtimer
, struct perf_event
, hw
.hrtimer
);
6012 if (event
->state
!= PERF_EVENT_STATE_ACTIVE
)
6013 return HRTIMER_NORESTART
;
6015 event
->pmu
->read(event
);
6017 perf_sample_data_init(&data
, 0, event
->hw
.last_period
);
6018 regs
= get_irq_regs();
6020 if (regs
&& !perf_exclude_event(event
, regs
)) {
6021 if (!(event
->attr
.exclude_idle
&& is_idle_task(current
)))
6022 if (__perf_event_overflow(event
, 1, &data
, regs
))
6023 ret
= HRTIMER_NORESTART
;
6026 period
= max_t(u64
, 10000, event
->hw
.sample_period
);
6027 hrtimer_forward_now(hrtimer
, ns_to_ktime(period
));
6032 static void perf_swevent_start_hrtimer(struct perf_event
*event
)
6034 struct hw_perf_event
*hwc
= &event
->hw
;
6037 if (!is_sampling_event(event
))
6040 period
= local64_read(&hwc
->period_left
);
6045 local64_set(&hwc
->period_left
, 0);
6047 period
= max_t(u64
, 10000, hwc
->sample_period
);
6049 __hrtimer_start_range_ns(&hwc
->hrtimer
,
6050 ns_to_ktime(period
), 0,
6051 HRTIMER_MODE_REL_PINNED
, 0);
6054 static void perf_swevent_cancel_hrtimer(struct perf_event
*event
)
6056 struct hw_perf_event
*hwc
= &event
->hw
;
6058 if (is_sampling_event(event
)) {
6059 ktime_t remaining
= hrtimer_get_remaining(&hwc
->hrtimer
);
6060 local64_set(&hwc
->period_left
, ktime_to_ns(remaining
));
6062 hrtimer_cancel(&hwc
->hrtimer
);
6066 static void perf_swevent_init_hrtimer(struct perf_event
*event
)
6068 struct hw_perf_event
*hwc
= &event
->hw
;
6070 if (!is_sampling_event(event
))
6073 hrtimer_init(&hwc
->hrtimer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
6074 hwc
->hrtimer
.function
= perf_swevent_hrtimer
;
6077 * Since hrtimers have a fixed rate, we can do a static freq->period
6078 * mapping and avoid the whole period adjust feedback stuff.
6080 if (event
->attr
.freq
) {
6081 long freq
= event
->attr
.sample_freq
;
6083 event
->attr
.sample_period
= NSEC_PER_SEC
/ freq
;
6084 hwc
->sample_period
= event
->attr
.sample_period
;
6085 local64_set(&hwc
->period_left
, hwc
->sample_period
);
6086 hwc
->last_period
= hwc
->sample_period
;
6087 event
->attr
.freq
= 0;
6092 * Software event: cpu wall time clock
6095 static void cpu_clock_event_update(struct perf_event
*event
)
6100 now
= local_clock();
6101 prev
= local64_xchg(&event
->hw
.prev_count
, now
);
6102 local64_add(now
- prev
, &event
->count
);
6105 static void cpu_clock_event_start(struct perf_event
*event
, int flags
)
6107 local64_set(&event
->hw
.prev_count
, local_clock());
6108 perf_swevent_start_hrtimer(event
);
6111 static void cpu_clock_event_stop(struct perf_event
*event
, int flags
)
6113 perf_swevent_cancel_hrtimer(event
);
6114 cpu_clock_event_update(event
);
6117 static int cpu_clock_event_add(struct perf_event
*event
, int flags
)
6119 if (flags
& PERF_EF_START
)
6120 cpu_clock_event_start(event
, flags
);
6125 static void cpu_clock_event_del(struct perf_event
*event
, int flags
)
6127 cpu_clock_event_stop(event
, flags
);
6130 static void cpu_clock_event_read(struct perf_event
*event
)
6132 cpu_clock_event_update(event
);
6135 static int cpu_clock_event_init(struct perf_event
*event
)
6137 if (event
->attr
.type
!= PERF_TYPE_SOFTWARE
)
6140 if (event
->attr
.config
!= PERF_COUNT_SW_CPU_CLOCK
)
6144 * no branch sampling for software events
6146 if (has_branch_stack(event
))
6149 perf_swevent_init_hrtimer(event
);
6154 static struct pmu perf_cpu_clock
= {
6155 .task_ctx_nr
= perf_sw_context
,
6157 .event_init
= cpu_clock_event_init
,
6158 .add
= cpu_clock_event_add
,
6159 .del
= cpu_clock_event_del
,
6160 .start
= cpu_clock_event_start
,
6161 .stop
= cpu_clock_event_stop
,
6162 .read
= cpu_clock_event_read
,
6164 .event_idx
= perf_swevent_event_idx
,
6168 * Software event: task time clock
6171 static void task_clock_event_update(struct perf_event
*event
, u64 now
)
6176 prev
= local64_xchg(&event
->hw
.prev_count
, now
);
6178 local64_add(delta
, &event
->count
);
6181 static void task_clock_event_start(struct perf_event
*event
, int flags
)
6183 local64_set(&event
->hw
.prev_count
, event
->ctx
->time
);
6184 perf_swevent_start_hrtimer(event
);
6187 static void task_clock_event_stop(struct perf_event
*event
, int flags
)
6189 perf_swevent_cancel_hrtimer(event
);
6190 task_clock_event_update(event
, event
->ctx
->time
);
6193 static int task_clock_event_add(struct perf_event
*event
, int flags
)
6195 if (flags
& PERF_EF_START
)
6196 task_clock_event_start(event
, flags
);
6201 static void task_clock_event_del(struct perf_event
*event
, int flags
)
6203 task_clock_event_stop(event
, PERF_EF_UPDATE
);
6206 static void task_clock_event_read(struct perf_event
*event
)
6208 u64 now
= perf_clock();
6209 u64 delta
= now
- event
->ctx
->timestamp
;
6210 u64 time
= event
->ctx
->time
+ delta
;
6212 task_clock_event_update(event
, time
);
6215 static int task_clock_event_init(struct perf_event
*event
)
6217 if (event
->attr
.type
!= PERF_TYPE_SOFTWARE
)
6220 if (event
->attr
.config
!= PERF_COUNT_SW_TASK_CLOCK
)
6224 * no branch sampling for software events
6226 if (has_branch_stack(event
))
6229 perf_swevent_init_hrtimer(event
);
6234 static struct pmu perf_task_clock
= {
6235 .task_ctx_nr
= perf_sw_context
,
6237 .event_init
= task_clock_event_init
,
6238 .add
= task_clock_event_add
,
6239 .del
= task_clock_event_del
,
6240 .start
= task_clock_event_start
,
6241 .stop
= task_clock_event_stop
,
6242 .read
= task_clock_event_read
,
6244 .event_idx
= perf_swevent_event_idx
,
6247 static void perf_pmu_nop_void(struct pmu
*pmu
)
6251 static int perf_pmu_nop_int(struct pmu
*pmu
)
6256 static void perf_pmu_start_txn(struct pmu
*pmu
)
6258 perf_pmu_disable(pmu
);
6261 static int perf_pmu_commit_txn(struct pmu
*pmu
)
6263 perf_pmu_enable(pmu
);
6267 static void perf_pmu_cancel_txn(struct pmu
*pmu
)
6269 perf_pmu_enable(pmu
);
6272 static int perf_event_idx_default(struct perf_event
*event
)
6274 return event
->hw
.idx
+ 1;
6278 * Ensures all contexts with the same task_ctx_nr have the same
6279 * pmu_cpu_context too.
6281 static void *find_pmu_context(int ctxn
)
6288 list_for_each_entry(pmu
, &pmus
, entry
) {
6289 if (pmu
->task_ctx_nr
== ctxn
)
6290 return pmu
->pmu_cpu_context
;
6296 static void update_pmu_context(struct pmu
*pmu
, struct pmu
*old_pmu
)
6300 for_each_possible_cpu(cpu
) {
6301 struct perf_cpu_context
*cpuctx
;
6303 cpuctx
= per_cpu_ptr(pmu
->pmu_cpu_context
, cpu
);
6305 if (cpuctx
->unique_pmu
== old_pmu
)
6306 cpuctx
->unique_pmu
= pmu
;
6310 static void free_pmu_context(struct pmu
*pmu
)
6314 mutex_lock(&pmus_lock
);
6316 * Like a real lame refcount.
6318 list_for_each_entry(i
, &pmus
, entry
) {
6319 if (i
->pmu_cpu_context
== pmu
->pmu_cpu_context
) {
6320 update_pmu_context(i
, pmu
);
6325 free_percpu(pmu
->pmu_cpu_context
);
6327 mutex_unlock(&pmus_lock
);
6329 static struct idr pmu_idr
;
6332 type_show(struct device
*dev
, struct device_attribute
*attr
, char *page
)
6334 struct pmu
*pmu
= dev_get_drvdata(dev
);
6336 return snprintf(page
, PAGE_SIZE
-1, "%d\n", pmu
->type
);
6338 static DEVICE_ATTR_RO(type
);
6341 perf_event_mux_interval_ms_show(struct device
*dev
,
6342 struct device_attribute
*attr
,
6345 struct pmu
*pmu
= dev_get_drvdata(dev
);
6347 return snprintf(page
, PAGE_SIZE
-1, "%d\n", pmu
->hrtimer_interval_ms
);
6351 perf_event_mux_interval_ms_store(struct device
*dev
,
6352 struct device_attribute
*attr
,
6353 const char *buf
, size_t count
)
6355 struct pmu
*pmu
= dev_get_drvdata(dev
);
6356 int timer
, cpu
, ret
;
6358 ret
= kstrtoint(buf
, 0, &timer
);
6365 /* same value, noting to do */
6366 if (timer
== pmu
->hrtimer_interval_ms
)
6369 pmu
->hrtimer_interval_ms
= timer
;
6371 /* update all cpuctx for this PMU */
6372 for_each_possible_cpu(cpu
) {
6373 struct perf_cpu_context
*cpuctx
;
6374 cpuctx
= per_cpu_ptr(pmu
->pmu_cpu_context
, cpu
);
6375 cpuctx
->hrtimer_interval
= ns_to_ktime(NSEC_PER_MSEC
* timer
);
6377 if (hrtimer_active(&cpuctx
->hrtimer
))
6378 hrtimer_forward_now(&cpuctx
->hrtimer
, cpuctx
->hrtimer_interval
);
6383 static DEVICE_ATTR_RW(perf_event_mux_interval_ms
);
6385 static struct attribute
*pmu_dev_attrs
[] = {
6386 &dev_attr_type
.attr
,
6387 &dev_attr_perf_event_mux_interval_ms
.attr
,
6390 ATTRIBUTE_GROUPS(pmu_dev
);
6392 static int pmu_bus_running
;
6393 static struct bus_type pmu_bus
= {
6394 .name
= "event_source",
6395 .dev_groups
= pmu_dev_groups
,
6398 static void pmu_dev_release(struct device
*dev
)
6403 static int pmu_dev_alloc(struct pmu
*pmu
)
6407 pmu
->dev
= kzalloc(sizeof(struct device
), GFP_KERNEL
);
6411 pmu
->dev
->groups
= pmu
->attr_groups
;
6412 device_initialize(pmu
->dev
);
6413 ret
= dev_set_name(pmu
->dev
, "%s", pmu
->name
);
6417 dev_set_drvdata(pmu
->dev
, pmu
);
6418 pmu
->dev
->bus
= &pmu_bus
;
6419 pmu
->dev
->release
= pmu_dev_release
;
6420 ret
= device_add(pmu
->dev
);
6428 put_device(pmu
->dev
);
6432 static struct lock_class_key cpuctx_mutex
;
6433 static struct lock_class_key cpuctx_lock
;
6435 int perf_pmu_register(struct pmu
*pmu
, const char *name
, int type
)
6439 mutex_lock(&pmus_lock
);
6441 pmu
->pmu_disable_count
= alloc_percpu(int);
6442 if (!pmu
->pmu_disable_count
)
6451 type
= idr_alloc(&pmu_idr
, pmu
, PERF_TYPE_MAX
, 0, GFP_KERNEL
);
6459 if (pmu_bus_running
) {
6460 ret
= pmu_dev_alloc(pmu
);
6466 pmu
->pmu_cpu_context
= find_pmu_context(pmu
->task_ctx_nr
);
6467 if (pmu
->pmu_cpu_context
)
6468 goto got_cpu_context
;
6471 pmu
->pmu_cpu_context
= alloc_percpu(struct perf_cpu_context
);
6472 if (!pmu
->pmu_cpu_context
)
6475 for_each_possible_cpu(cpu
) {
6476 struct perf_cpu_context
*cpuctx
;
6478 cpuctx
= per_cpu_ptr(pmu
->pmu_cpu_context
, cpu
);
6479 __perf_event_init_context(&cpuctx
->ctx
);
6480 lockdep_set_class(&cpuctx
->ctx
.mutex
, &cpuctx_mutex
);
6481 lockdep_set_class(&cpuctx
->ctx
.lock
, &cpuctx_lock
);
6482 cpuctx
->ctx
.type
= cpu_context
;
6483 cpuctx
->ctx
.pmu
= pmu
;
6485 __perf_cpu_hrtimer_init(cpuctx
, cpu
);
6487 INIT_LIST_HEAD(&cpuctx
->rotation_list
);
6488 cpuctx
->unique_pmu
= pmu
;
6492 if (!pmu
->start_txn
) {
6493 if (pmu
->pmu_enable
) {
6495 * If we have pmu_enable/pmu_disable calls, install
6496 * transaction stubs that use that to try and batch
6497 * hardware accesses.
6499 pmu
->start_txn
= perf_pmu_start_txn
;
6500 pmu
->commit_txn
= perf_pmu_commit_txn
;
6501 pmu
->cancel_txn
= perf_pmu_cancel_txn
;
6503 pmu
->start_txn
= perf_pmu_nop_void
;
6504 pmu
->commit_txn
= perf_pmu_nop_int
;
6505 pmu
->cancel_txn
= perf_pmu_nop_void
;
6509 if (!pmu
->pmu_enable
) {
6510 pmu
->pmu_enable
= perf_pmu_nop_void
;
6511 pmu
->pmu_disable
= perf_pmu_nop_void
;
6514 if (!pmu
->event_idx
)
6515 pmu
->event_idx
= perf_event_idx_default
;
6517 list_add_rcu(&pmu
->entry
, &pmus
);
6520 mutex_unlock(&pmus_lock
);
6525 device_del(pmu
->dev
);
6526 put_device(pmu
->dev
);
6529 if (pmu
->type
>= PERF_TYPE_MAX
)
6530 idr_remove(&pmu_idr
, pmu
->type
);
6533 free_percpu(pmu
->pmu_disable_count
);
6537 void perf_pmu_unregister(struct pmu
*pmu
)
6539 mutex_lock(&pmus_lock
);
6540 list_del_rcu(&pmu
->entry
);
6541 mutex_unlock(&pmus_lock
);
6544 * We dereference the pmu list under both SRCU and regular RCU, so
6545 * synchronize against both of those.
6547 synchronize_srcu(&pmus_srcu
);
6550 free_percpu(pmu
->pmu_disable_count
);
6551 if (pmu
->type
>= PERF_TYPE_MAX
)
6552 idr_remove(&pmu_idr
, pmu
->type
);
6553 device_del(pmu
->dev
);
6554 put_device(pmu
->dev
);
6555 free_pmu_context(pmu
);
6558 struct pmu
*perf_init_event(struct perf_event
*event
)
6560 struct pmu
*pmu
= NULL
;
6564 idx
= srcu_read_lock(&pmus_srcu
);
6567 pmu
= idr_find(&pmu_idr
, event
->attr
.type
);
6571 ret
= pmu
->event_init(event
);
6577 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
6579 ret
= pmu
->event_init(event
);
6583 if (ret
!= -ENOENT
) {
6588 pmu
= ERR_PTR(-ENOENT
);
6590 srcu_read_unlock(&pmus_srcu
, idx
);
6595 static void account_event_cpu(struct perf_event
*event
, int cpu
)
6600 if (has_branch_stack(event
)) {
6601 if (!(event
->attach_state
& PERF_ATTACH_TASK
))
6602 atomic_inc(&per_cpu(perf_branch_stack_events
, cpu
));
6604 if (is_cgroup_event(event
))
6605 atomic_inc(&per_cpu(perf_cgroup_events
, cpu
));
6608 static void account_event(struct perf_event
*event
)
6613 if (event
->attach_state
& PERF_ATTACH_TASK
)
6614 static_key_slow_inc(&perf_sched_events
.key
);
6615 if (event
->attr
.mmap
|| event
->attr
.mmap_data
)
6616 atomic_inc(&nr_mmap_events
);
6617 if (event
->attr
.comm
)
6618 atomic_inc(&nr_comm_events
);
6619 if (event
->attr
.task
)
6620 atomic_inc(&nr_task_events
);
6621 if (event
->attr
.freq
) {
6622 if (atomic_inc_return(&nr_freq_events
) == 1)
6623 tick_nohz_full_kick_all();
6625 if (has_branch_stack(event
))
6626 static_key_slow_inc(&perf_sched_events
.key
);
6627 if (is_cgroup_event(event
))
6628 static_key_slow_inc(&perf_sched_events
.key
);
6630 account_event_cpu(event
, event
->cpu
);
6634 * Allocate and initialize a event structure
6636 static struct perf_event
*
6637 perf_event_alloc(struct perf_event_attr
*attr
, int cpu
,
6638 struct task_struct
*task
,
6639 struct perf_event
*group_leader
,
6640 struct perf_event
*parent_event
,
6641 perf_overflow_handler_t overflow_handler
,
6645 struct perf_event
*event
;
6646 struct hw_perf_event
*hwc
;
6649 if ((unsigned)cpu
>= nr_cpu_ids
) {
6650 if (!task
|| cpu
!= -1)
6651 return ERR_PTR(-EINVAL
);
6654 event
= kzalloc(sizeof(*event
), GFP_KERNEL
);
6656 return ERR_PTR(-ENOMEM
);
6659 * Single events are their own group leaders, with an
6660 * empty sibling list:
6663 group_leader
= event
;
6665 mutex_init(&event
->child_mutex
);
6666 INIT_LIST_HEAD(&event
->child_list
);
6668 INIT_LIST_HEAD(&event
->group_entry
);
6669 INIT_LIST_HEAD(&event
->event_entry
);
6670 INIT_LIST_HEAD(&event
->sibling_list
);
6671 INIT_LIST_HEAD(&event
->rb_entry
);
6672 INIT_LIST_HEAD(&event
->active_entry
);
6673 INIT_HLIST_NODE(&event
->hlist_entry
);
6676 init_waitqueue_head(&event
->waitq
);
6677 init_irq_work(&event
->pending
, perf_pending_event
);
6679 mutex_init(&event
->mmap_mutex
);
6681 atomic_long_set(&event
->refcount
, 1);
6683 event
->attr
= *attr
;
6684 event
->group_leader
= group_leader
;
6688 event
->parent
= parent_event
;
6690 event
->ns
= get_pid_ns(task_active_pid_ns(current
));
6691 event
->id
= atomic64_inc_return(&perf_event_id
);
6693 event
->state
= PERF_EVENT_STATE_INACTIVE
;
6696 event
->attach_state
= PERF_ATTACH_TASK
;
6698 if (attr
->type
== PERF_TYPE_TRACEPOINT
)
6699 event
->hw
.tp_target
= task
;
6700 #ifdef CONFIG_HAVE_HW_BREAKPOINT
6702 * hw_breakpoint is a bit difficult here..
6704 else if (attr
->type
== PERF_TYPE_BREAKPOINT
)
6705 event
->hw
.bp_target
= task
;
6709 if (!overflow_handler
&& parent_event
) {
6710 overflow_handler
= parent_event
->overflow_handler
;
6711 context
= parent_event
->overflow_handler_context
;
6714 event
->overflow_handler
= overflow_handler
;
6715 event
->overflow_handler_context
= context
;
6717 perf_event__state_init(event
);
6722 hwc
->sample_period
= attr
->sample_period
;
6723 if (attr
->freq
&& attr
->sample_freq
)
6724 hwc
->sample_period
= 1;
6725 hwc
->last_period
= hwc
->sample_period
;
6727 local64_set(&hwc
->period_left
, hwc
->sample_period
);
6730 * we currently do not support PERF_FORMAT_GROUP on inherited events
6732 if (attr
->inherit
&& (attr
->read_format
& PERF_FORMAT_GROUP
))
6735 pmu
= perf_init_event(event
);
6738 else if (IS_ERR(pmu
)) {
6743 if (!event
->parent
) {
6744 if (event
->attr
.sample_type
& PERF_SAMPLE_CALLCHAIN
) {
6745 err
= get_callchain_buffers();
6755 event
->destroy(event
);
6758 put_pid_ns(event
->ns
);
6761 return ERR_PTR(err
);
6764 static int perf_copy_attr(struct perf_event_attr __user
*uattr
,
6765 struct perf_event_attr
*attr
)
6770 if (!access_ok(VERIFY_WRITE
, uattr
, PERF_ATTR_SIZE_VER0
))
6774 * zero the full structure, so that a short copy will be nice.
6776 memset(attr
, 0, sizeof(*attr
));
6778 ret
= get_user(size
, &uattr
->size
);
6782 if (size
> PAGE_SIZE
) /* silly large */
6785 if (!size
) /* abi compat */
6786 size
= PERF_ATTR_SIZE_VER0
;
6788 if (size
< PERF_ATTR_SIZE_VER0
)
6792 * If we're handed a bigger struct than we know of,
6793 * ensure all the unknown bits are 0 - i.e. new
6794 * user-space does not rely on any kernel feature
6795 * extensions we dont know about yet.
6797 if (size
> sizeof(*attr
)) {
6798 unsigned char __user
*addr
;
6799 unsigned char __user
*end
;
6802 addr
= (void __user
*)uattr
+ sizeof(*attr
);
6803 end
= (void __user
*)uattr
+ size
;
6805 for (; addr
< end
; addr
++) {
6806 ret
= get_user(val
, addr
);
6812 size
= sizeof(*attr
);
6815 ret
= copy_from_user(attr
, uattr
, size
);
6819 /* disabled for now */
6823 if (attr
->__reserved_1
)
6826 if (attr
->sample_type
& ~(PERF_SAMPLE_MAX
-1))
6829 if (attr
->read_format
& ~(PERF_FORMAT_MAX
-1))
6832 if (attr
->sample_type
& PERF_SAMPLE_BRANCH_STACK
) {
6833 u64 mask
= attr
->branch_sample_type
;
6835 /* only using defined bits */
6836 if (mask
& ~(PERF_SAMPLE_BRANCH_MAX
-1))
6839 /* at least one branch bit must be set */
6840 if (!(mask
& ~PERF_SAMPLE_BRANCH_PLM_ALL
))
6843 /* propagate priv level, when not set for branch */
6844 if (!(mask
& PERF_SAMPLE_BRANCH_PLM_ALL
)) {
6846 /* exclude_kernel checked on syscall entry */
6847 if (!attr
->exclude_kernel
)
6848 mask
|= PERF_SAMPLE_BRANCH_KERNEL
;
6850 if (!attr
->exclude_user
)
6851 mask
|= PERF_SAMPLE_BRANCH_USER
;
6853 if (!attr
->exclude_hv
)
6854 mask
|= PERF_SAMPLE_BRANCH_HV
;
6856 * adjust user setting (for HW filter setup)
6858 attr
->branch_sample_type
= mask
;
6860 /* privileged levels capture (kernel, hv): check permissions */
6861 if ((mask
& PERF_SAMPLE_BRANCH_PERM_PLM
)
6862 && perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN
))
6866 if (attr
->sample_type
& PERF_SAMPLE_REGS_USER
) {
6867 ret
= perf_reg_validate(attr
->sample_regs_user
);
6872 if (attr
->sample_type
& PERF_SAMPLE_STACK_USER
) {
6873 if (!arch_perf_have_user_stack_dump())
6877 * We have __u32 type for the size, but so far
6878 * we can only use __u16 as maximum due to the
6879 * __u16 sample size limit.
6881 if (attr
->sample_stack_user
>= USHRT_MAX
)
6883 else if (!IS_ALIGNED(attr
->sample_stack_user
, sizeof(u64
)))
6891 put_user(sizeof(*attr
), &uattr
->size
);
6897 perf_event_set_output(struct perf_event
*event
, struct perf_event
*output_event
)
6899 struct ring_buffer
*rb
= NULL
, *old_rb
= NULL
;
6905 /* don't allow circular references */
6906 if (event
== output_event
)
6910 * Don't allow cross-cpu buffers
6912 if (output_event
->cpu
!= event
->cpu
)
6916 * If its not a per-cpu rb, it must be the same task.
6918 if (output_event
->cpu
== -1 && output_event
->ctx
!= event
->ctx
)
6922 mutex_lock(&event
->mmap_mutex
);
6923 /* Can't redirect output if we've got an active mmap() */
6924 if (atomic_read(&event
->mmap_count
))
6930 /* get the rb we want to redirect to */
6931 rb
= ring_buffer_get(output_event
);
6937 ring_buffer_detach(event
, old_rb
);
6940 ring_buffer_attach(event
, rb
);
6942 rcu_assign_pointer(event
->rb
, rb
);
6945 ring_buffer_put(old_rb
);
6947 * Since we detached before setting the new rb, so that we
6948 * could attach the new rb, we could have missed a wakeup.
6951 wake_up_all(&event
->waitq
);
6956 mutex_unlock(&event
->mmap_mutex
);
6963 * sys_perf_event_open - open a performance event, associate it to a task/cpu
6965 * @attr_uptr: event_id type attributes for monitoring/sampling
6968 * @group_fd: group leader event fd
6970 SYSCALL_DEFINE5(perf_event_open
,
6971 struct perf_event_attr __user
*, attr_uptr
,
6972 pid_t
, pid
, int, cpu
, int, group_fd
, unsigned long, flags
)
6974 struct perf_event
*group_leader
= NULL
, *output_event
= NULL
;
6975 struct perf_event
*event
, *sibling
;
6976 struct perf_event_attr attr
;
6977 struct perf_event_context
*ctx
;
6978 struct file
*event_file
= NULL
;
6979 struct fd group
= {NULL
, 0};
6980 struct task_struct
*task
= NULL
;
6986 /* for future expandability... */
6987 if (flags
& ~PERF_FLAG_ALL
)
6990 err
= perf_copy_attr(attr_uptr
, &attr
);
6994 if (!attr
.exclude_kernel
) {
6995 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN
))
7000 if (attr
.sample_freq
> sysctl_perf_event_sample_rate
)
7005 * In cgroup mode, the pid argument is used to pass the fd
7006 * opened to the cgroup directory in cgroupfs. The cpu argument
7007 * designates the cpu on which to monitor threads from that
7010 if ((flags
& PERF_FLAG_PID_CGROUP
) && (pid
== -1 || cpu
== -1))
7013 event_fd
= get_unused_fd();
7017 if (group_fd
!= -1) {
7018 err
= perf_fget_light(group_fd
, &group
);
7021 group_leader
= group
.file
->private_data
;
7022 if (flags
& PERF_FLAG_FD_OUTPUT
)
7023 output_event
= group_leader
;
7024 if (flags
& PERF_FLAG_FD_NO_GROUP
)
7025 group_leader
= NULL
;
7028 if (pid
!= -1 && !(flags
& PERF_FLAG_PID_CGROUP
)) {
7029 task
= find_lively_task_by_vpid(pid
);
7031 err
= PTR_ERR(task
);
7038 event
= perf_event_alloc(&attr
, cpu
, task
, group_leader
, NULL
,
7040 if (IS_ERR(event
)) {
7041 err
= PTR_ERR(event
);
7045 if (flags
& PERF_FLAG_PID_CGROUP
) {
7046 err
= perf_cgroup_connect(pid
, event
, &attr
, group_leader
);
7048 __free_event(event
);
7053 account_event(event
);
7056 * Special case software events and allow them to be part of
7057 * any hardware group.
7062 (is_software_event(event
) != is_software_event(group_leader
))) {
7063 if (is_software_event(event
)) {
7065 * If event and group_leader are not both a software
7066 * event, and event is, then group leader is not.
7068 * Allow the addition of software events to !software
7069 * groups, this is safe because software events never
7072 pmu
= group_leader
->pmu
;
7073 } else if (is_software_event(group_leader
) &&
7074 (group_leader
->group_flags
& PERF_GROUP_SOFTWARE
)) {
7076 * In case the group is a pure software group, and we
7077 * try to add a hardware event, move the whole group to
7078 * the hardware context.
7085 * Get the target context (task or percpu):
7087 ctx
= find_get_context(pmu
, task
, event
->cpu
);
7094 put_task_struct(task
);
7099 * Look up the group leader (we will attach this event to it):
7105 * Do not allow a recursive hierarchy (this new sibling
7106 * becoming part of another group-sibling):
7108 if (group_leader
->group_leader
!= group_leader
)
7111 * Do not allow to attach to a group in a different
7112 * task or CPU context:
7115 if (group_leader
->ctx
->type
!= ctx
->type
)
7118 if (group_leader
->ctx
!= ctx
)
7123 * Only a group leader can be exclusive or pinned
7125 if (attr
.exclusive
|| attr
.pinned
)
7130 err
= perf_event_set_output(event
, output_event
);
7135 event_file
= anon_inode_getfile("[perf_event]", &perf_fops
, event
, O_RDWR
);
7136 if (IS_ERR(event_file
)) {
7137 err
= PTR_ERR(event_file
);
7142 struct perf_event_context
*gctx
= group_leader
->ctx
;
7144 mutex_lock(&gctx
->mutex
);
7145 perf_remove_from_context(group_leader
);
7148 * Removing from the context ends up with disabled
7149 * event. What we want here is event in the initial
7150 * startup state, ready to be add into new context.
7152 perf_event__state_init(group_leader
);
7153 list_for_each_entry(sibling
, &group_leader
->sibling_list
,
7155 perf_remove_from_context(sibling
);
7156 perf_event__state_init(sibling
);
7159 mutex_unlock(&gctx
->mutex
);
7163 WARN_ON_ONCE(ctx
->parent_ctx
);
7164 mutex_lock(&ctx
->mutex
);
7168 perf_install_in_context(ctx
, group_leader
, event
->cpu
);
7170 list_for_each_entry(sibling
, &group_leader
->sibling_list
,
7172 perf_install_in_context(ctx
, sibling
, event
->cpu
);
7177 perf_install_in_context(ctx
, event
, event
->cpu
);
7178 perf_unpin_context(ctx
);
7179 mutex_unlock(&ctx
->mutex
);
7183 event
->owner
= current
;
7185 mutex_lock(¤t
->perf_event_mutex
);
7186 list_add_tail(&event
->owner_entry
, ¤t
->perf_event_list
);
7187 mutex_unlock(¤t
->perf_event_mutex
);
7190 * Precalculate sample_data sizes
7192 perf_event__header_size(event
);
7193 perf_event__id_header_size(event
);
7196 * Drop the reference on the group_event after placing the
7197 * new event on the sibling_list. This ensures destruction
7198 * of the group leader will find the pointer to itself in
7199 * perf_group_detach().
7202 fd_install(event_fd
, event_file
);
7206 perf_unpin_context(ctx
);
7213 put_task_struct(task
);
7217 put_unused_fd(event_fd
);
7222 * perf_event_create_kernel_counter
7224 * @attr: attributes of the counter to create
7225 * @cpu: cpu in which the counter is bound
7226 * @task: task to profile (NULL for percpu)
7229 perf_event_create_kernel_counter(struct perf_event_attr
*attr
, int cpu
,
7230 struct task_struct
*task
,
7231 perf_overflow_handler_t overflow_handler
,
7234 struct perf_event_context
*ctx
;
7235 struct perf_event
*event
;
7239 * Get the target context (task or percpu):
7242 event
= perf_event_alloc(attr
, cpu
, task
, NULL
, NULL
,
7243 overflow_handler
, context
);
7244 if (IS_ERR(event
)) {
7245 err
= PTR_ERR(event
);
7249 account_event(event
);
7251 ctx
= find_get_context(event
->pmu
, task
, cpu
);
7257 WARN_ON_ONCE(ctx
->parent_ctx
);
7258 mutex_lock(&ctx
->mutex
);
7259 perf_install_in_context(ctx
, event
, cpu
);
7260 perf_unpin_context(ctx
);
7261 mutex_unlock(&ctx
->mutex
);
7268 return ERR_PTR(err
);
7270 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter
);
7272 void perf_pmu_migrate_context(struct pmu
*pmu
, int src_cpu
, int dst_cpu
)
7274 struct perf_event_context
*src_ctx
;
7275 struct perf_event_context
*dst_ctx
;
7276 struct perf_event
*event
, *tmp
;
7279 src_ctx
= &per_cpu_ptr(pmu
->pmu_cpu_context
, src_cpu
)->ctx
;
7280 dst_ctx
= &per_cpu_ptr(pmu
->pmu_cpu_context
, dst_cpu
)->ctx
;
7282 mutex_lock(&src_ctx
->mutex
);
7283 list_for_each_entry_safe(event
, tmp
, &src_ctx
->event_list
,
7285 perf_remove_from_context(event
);
7286 unaccount_event_cpu(event
, src_cpu
);
7288 list_add(&event
->migrate_entry
, &events
);
7290 mutex_unlock(&src_ctx
->mutex
);
7294 mutex_lock(&dst_ctx
->mutex
);
7295 list_for_each_entry_safe(event
, tmp
, &events
, migrate_entry
) {
7296 list_del(&event
->migrate_entry
);
7297 if (event
->state
>= PERF_EVENT_STATE_OFF
)
7298 event
->state
= PERF_EVENT_STATE_INACTIVE
;
7299 account_event_cpu(event
, dst_cpu
);
7300 perf_install_in_context(dst_ctx
, event
, dst_cpu
);
7303 mutex_unlock(&dst_ctx
->mutex
);
7305 EXPORT_SYMBOL_GPL(perf_pmu_migrate_context
);
7307 static void sync_child_event(struct perf_event
*child_event
,
7308 struct task_struct
*child
)
7310 struct perf_event
*parent_event
= child_event
->parent
;
7313 if (child_event
->attr
.inherit_stat
)
7314 perf_event_read_event(child_event
, child
);
7316 child_val
= perf_event_count(child_event
);
7319 * Add back the child's count to the parent's count:
7321 atomic64_add(child_val
, &parent_event
->child_count
);
7322 atomic64_add(child_event
->total_time_enabled
,
7323 &parent_event
->child_total_time_enabled
);
7324 atomic64_add(child_event
->total_time_running
,
7325 &parent_event
->child_total_time_running
);
7328 * Remove this event from the parent's list
7330 WARN_ON_ONCE(parent_event
->ctx
->parent_ctx
);
7331 mutex_lock(&parent_event
->child_mutex
);
7332 list_del_init(&child_event
->child_list
);
7333 mutex_unlock(&parent_event
->child_mutex
);
7336 * Release the parent event, if this was the last
7339 put_event(parent_event
);
7343 __perf_event_exit_task(struct perf_event
*child_event
,
7344 struct perf_event_context
*child_ctx
,
7345 struct task_struct
*child
)
7347 if (child_event
->parent
) {
7348 raw_spin_lock_irq(&child_ctx
->lock
);
7349 perf_group_detach(child_event
);
7350 raw_spin_unlock_irq(&child_ctx
->lock
);
7353 perf_remove_from_context(child_event
);
7356 * It can happen that the parent exits first, and has events
7357 * that are still around due to the child reference. These
7358 * events need to be zapped.
7360 if (child_event
->parent
) {
7361 sync_child_event(child_event
, child
);
7362 free_event(child_event
);
7366 static void perf_event_exit_task_context(struct task_struct
*child
, int ctxn
)
7368 struct perf_event
*child_event
, *tmp
;
7369 struct perf_event_context
*child_ctx
;
7370 unsigned long flags
;
7372 if (likely(!child
->perf_event_ctxp
[ctxn
])) {
7373 perf_event_task(child
, NULL
, 0);
7377 local_irq_save(flags
);
7379 * We can't reschedule here because interrupts are disabled,
7380 * and either child is current or it is a task that can't be
7381 * scheduled, so we are now safe from rescheduling changing
7384 child_ctx
= rcu_dereference_raw(child
->perf_event_ctxp
[ctxn
]);
7387 * Take the context lock here so that if find_get_context is
7388 * reading child->perf_event_ctxp, we wait until it has
7389 * incremented the context's refcount before we do put_ctx below.
7391 raw_spin_lock(&child_ctx
->lock
);
7392 task_ctx_sched_out(child_ctx
);
7393 child
->perf_event_ctxp
[ctxn
] = NULL
;
7395 * If this context is a clone; unclone it so it can't get
7396 * swapped to another process while we're removing all
7397 * the events from it.
7399 unclone_ctx(child_ctx
);
7400 update_context_time(child_ctx
);
7401 raw_spin_unlock_irqrestore(&child_ctx
->lock
, flags
);
7404 * Report the task dead after unscheduling the events so that we
7405 * won't get any samples after PERF_RECORD_EXIT. We can however still
7406 * get a few PERF_RECORD_READ events.
7408 perf_event_task(child
, child_ctx
, 0);
7411 * We can recurse on the same lock type through:
7413 * __perf_event_exit_task()
7414 * sync_child_event()
7416 * mutex_lock(&ctx->mutex)
7418 * But since its the parent context it won't be the same instance.
7420 mutex_lock(&child_ctx
->mutex
);
7423 list_for_each_entry_safe(child_event
, tmp
, &child_ctx
->pinned_groups
,
7425 __perf_event_exit_task(child_event
, child_ctx
, child
);
7427 list_for_each_entry_safe(child_event
, tmp
, &child_ctx
->flexible_groups
,
7429 __perf_event_exit_task(child_event
, child_ctx
, child
);
7432 * If the last event was a group event, it will have appended all
7433 * its siblings to the list, but we obtained 'tmp' before that which
7434 * will still point to the list head terminating the iteration.
7436 if (!list_empty(&child_ctx
->pinned_groups
) ||
7437 !list_empty(&child_ctx
->flexible_groups
))
7440 mutex_unlock(&child_ctx
->mutex
);
7446 * When a child task exits, feed back event values to parent events.
7448 void perf_event_exit_task(struct task_struct
*child
)
7450 struct perf_event
*event
, *tmp
;
7453 mutex_lock(&child
->perf_event_mutex
);
7454 list_for_each_entry_safe(event
, tmp
, &child
->perf_event_list
,
7456 list_del_init(&event
->owner_entry
);
7459 * Ensure the list deletion is visible before we clear
7460 * the owner, closes a race against perf_release() where
7461 * we need to serialize on the owner->perf_event_mutex.
7464 event
->owner
= NULL
;
7466 mutex_unlock(&child
->perf_event_mutex
);
7468 for_each_task_context_nr(ctxn
)
7469 perf_event_exit_task_context(child
, ctxn
);
7472 static void perf_free_event(struct perf_event
*event
,
7473 struct perf_event_context
*ctx
)
7475 struct perf_event
*parent
= event
->parent
;
7477 if (WARN_ON_ONCE(!parent
))
7480 mutex_lock(&parent
->child_mutex
);
7481 list_del_init(&event
->child_list
);
7482 mutex_unlock(&parent
->child_mutex
);
7486 perf_group_detach(event
);
7487 list_del_event(event
, ctx
);
7492 * free an unexposed, unused context as created by inheritance by
7493 * perf_event_init_task below, used by fork() in case of fail.
7495 void perf_event_free_task(struct task_struct
*task
)
7497 struct perf_event_context
*ctx
;
7498 struct perf_event
*event
, *tmp
;
7501 for_each_task_context_nr(ctxn
) {
7502 ctx
= task
->perf_event_ctxp
[ctxn
];
7506 mutex_lock(&ctx
->mutex
);
7508 list_for_each_entry_safe(event
, tmp
, &ctx
->pinned_groups
,
7510 perf_free_event(event
, ctx
);
7512 list_for_each_entry_safe(event
, tmp
, &ctx
->flexible_groups
,
7514 perf_free_event(event
, ctx
);
7516 if (!list_empty(&ctx
->pinned_groups
) ||
7517 !list_empty(&ctx
->flexible_groups
))
7520 mutex_unlock(&ctx
->mutex
);
7526 void perf_event_delayed_put(struct task_struct
*task
)
7530 for_each_task_context_nr(ctxn
)
7531 WARN_ON_ONCE(task
->perf_event_ctxp
[ctxn
]);
7535 * inherit a event from parent task to child task:
7537 static struct perf_event
*
7538 inherit_event(struct perf_event
*parent_event
,
7539 struct task_struct
*parent
,
7540 struct perf_event_context
*parent_ctx
,
7541 struct task_struct
*child
,
7542 struct perf_event
*group_leader
,
7543 struct perf_event_context
*child_ctx
)
7545 struct perf_event
*child_event
;
7546 unsigned long flags
;
7549 * Instead of creating recursive hierarchies of events,
7550 * we link inherited events back to the original parent,
7551 * which has a filp for sure, which we use as the reference
7554 if (parent_event
->parent
)
7555 parent_event
= parent_event
->parent
;
7557 child_event
= perf_event_alloc(&parent_event
->attr
,
7560 group_leader
, parent_event
,
7562 if (IS_ERR(child_event
))
7565 if (!atomic_long_inc_not_zero(&parent_event
->refcount
)) {
7566 free_event(child_event
);
7573 * Make the child state follow the state of the parent event,
7574 * not its attr.disabled bit. We hold the parent's mutex,
7575 * so we won't race with perf_event_{en, dis}able_family.
7577 if (parent_event
->state
>= PERF_EVENT_STATE_INACTIVE
)
7578 child_event
->state
= PERF_EVENT_STATE_INACTIVE
;
7580 child_event
->state
= PERF_EVENT_STATE_OFF
;
7582 if (parent_event
->attr
.freq
) {
7583 u64 sample_period
= parent_event
->hw
.sample_period
;
7584 struct hw_perf_event
*hwc
= &child_event
->hw
;
7586 hwc
->sample_period
= sample_period
;
7587 hwc
->last_period
= sample_period
;
7589 local64_set(&hwc
->period_left
, sample_period
);
7592 child_event
->ctx
= child_ctx
;
7593 child_event
->overflow_handler
= parent_event
->overflow_handler
;
7594 child_event
->overflow_handler_context
7595 = parent_event
->overflow_handler_context
;
7598 * Precalculate sample_data sizes
7600 perf_event__header_size(child_event
);
7601 perf_event__id_header_size(child_event
);
7604 * Link it up in the child's context:
7606 raw_spin_lock_irqsave(&child_ctx
->lock
, flags
);
7607 add_event_to_ctx(child_event
, child_ctx
);
7608 raw_spin_unlock_irqrestore(&child_ctx
->lock
, flags
);
7611 * Link this into the parent event's child list
7613 WARN_ON_ONCE(parent_event
->ctx
->parent_ctx
);
7614 mutex_lock(&parent_event
->child_mutex
);
7615 list_add_tail(&child_event
->child_list
, &parent_event
->child_list
);
7616 mutex_unlock(&parent_event
->child_mutex
);
7621 static int inherit_group(struct perf_event
*parent_event
,
7622 struct task_struct
*parent
,
7623 struct perf_event_context
*parent_ctx
,
7624 struct task_struct
*child
,
7625 struct perf_event_context
*child_ctx
)
7627 struct perf_event
*leader
;
7628 struct perf_event
*sub
;
7629 struct perf_event
*child_ctr
;
7631 leader
= inherit_event(parent_event
, parent
, parent_ctx
,
7632 child
, NULL
, child_ctx
);
7634 return PTR_ERR(leader
);
7635 list_for_each_entry(sub
, &parent_event
->sibling_list
, group_entry
) {
7636 child_ctr
= inherit_event(sub
, parent
, parent_ctx
,
7637 child
, leader
, child_ctx
);
7638 if (IS_ERR(child_ctr
))
7639 return PTR_ERR(child_ctr
);
7645 inherit_task_group(struct perf_event
*event
, struct task_struct
*parent
,
7646 struct perf_event_context
*parent_ctx
,
7647 struct task_struct
*child
, int ctxn
,
7651 struct perf_event_context
*child_ctx
;
7653 if (!event
->attr
.inherit
) {
7658 child_ctx
= child
->perf_event_ctxp
[ctxn
];
7661 * This is executed from the parent task context, so
7662 * inherit events that have been marked for cloning.
7663 * First allocate and initialize a context for the
7667 child_ctx
= alloc_perf_context(parent_ctx
->pmu
, child
);
7671 child
->perf_event_ctxp
[ctxn
] = child_ctx
;
7674 ret
= inherit_group(event
, parent
, parent_ctx
,
7684 * Initialize the perf_event context in task_struct
7686 int perf_event_init_context(struct task_struct
*child
, int ctxn
)
7688 struct perf_event_context
*child_ctx
, *parent_ctx
;
7689 struct perf_event_context
*cloned_ctx
;
7690 struct perf_event
*event
;
7691 struct task_struct
*parent
= current
;
7692 int inherited_all
= 1;
7693 unsigned long flags
;
7696 if (likely(!parent
->perf_event_ctxp
[ctxn
]))
7700 * If the parent's context is a clone, pin it so it won't get
7703 parent_ctx
= perf_pin_task_context(parent
, ctxn
);
7706 * No need to check if parent_ctx != NULL here; since we saw
7707 * it non-NULL earlier, the only reason for it to become NULL
7708 * is if we exit, and since we're currently in the middle of
7709 * a fork we can't be exiting at the same time.
7713 * Lock the parent list. No need to lock the child - not PID
7714 * hashed yet and not running, so nobody can access it.
7716 mutex_lock(&parent_ctx
->mutex
);
7719 * We dont have to disable NMIs - we are only looking at
7720 * the list, not manipulating it:
7722 list_for_each_entry(event
, &parent_ctx
->pinned_groups
, group_entry
) {
7723 ret
= inherit_task_group(event
, parent
, parent_ctx
,
7724 child
, ctxn
, &inherited_all
);
7730 * We can't hold ctx->lock when iterating the ->flexible_group list due
7731 * to allocations, but we need to prevent rotation because
7732 * rotate_ctx() will change the list from interrupt context.
7734 raw_spin_lock_irqsave(&parent_ctx
->lock
, flags
);
7735 parent_ctx
->rotate_disable
= 1;
7736 raw_spin_unlock_irqrestore(&parent_ctx
->lock
, flags
);
7738 list_for_each_entry(event
, &parent_ctx
->flexible_groups
, group_entry
) {
7739 ret
= inherit_task_group(event
, parent
, parent_ctx
,
7740 child
, ctxn
, &inherited_all
);
7745 raw_spin_lock_irqsave(&parent_ctx
->lock
, flags
);
7746 parent_ctx
->rotate_disable
= 0;
7748 child_ctx
= child
->perf_event_ctxp
[ctxn
];
7750 if (child_ctx
&& inherited_all
) {
7752 * Mark the child context as a clone of the parent
7753 * context, or of whatever the parent is a clone of.
7755 * Note that if the parent is a clone, the holding of
7756 * parent_ctx->lock avoids it from being uncloned.
7758 cloned_ctx
= parent_ctx
->parent_ctx
;
7760 child_ctx
->parent_ctx
= cloned_ctx
;
7761 child_ctx
->parent_gen
= parent_ctx
->parent_gen
;
7763 child_ctx
->parent_ctx
= parent_ctx
;
7764 child_ctx
->parent_gen
= parent_ctx
->generation
;
7766 get_ctx(child_ctx
->parent_ctx
);
7769 raw_spin_unlock_irqrestore(&parent_ctx
->lock
, flags
);
7770 mutex_unlock(&parent_ctx
->mutex
);
7772 perf_unpin_context(parent_ctx
);
7773 put_ctx(parent_ctx
);
7779 * Initialize the perf_event context in task_struct
7781 int perf_event_init_task(struct task_struct
*child
)
7785 memset(child
->perf_event_ctxp
, 0, sizeof(child
->perf_event_ctxp
));
7786 mutex_init(&child
->perf_event_mutex
);
7787 INIT_LIST_HEAD(&child
->perf_event_list
);
7789 for_each_task_context_nr(ctxn
) {
7790 ret
= perf_event_init_context(child
, ctxn
);
7798 static void __init
perf_event_init_all_cpus(void)
7800 struct swevent_htable
*swhash
;
7803 for_each_possible_cpu(cpu
) {
7804 swhash
= &per_cpu(swevent_htable
, cpu
);
7805 mutex_init(&swhash
->hlist_mutex
);
7806 INIT_LIST_HEAD(&per_cpu(rotation_list
, cpu
));
7810 static void perf_event_init_cpu(int cpu
)
7812 struct swevent_htable
*swhash
= &per_cpu(swevent_htable
, cpu
);
7814 mutex_lock(&swhash
->hlist_mutex
);
7815 if (swhash
->hlist_refcount
> 0) {
7816 struct swevent_hlist
*hlist
;
7818 hlist
= kzalloc_node(sizeof(*hlist
), GFP_KERNEL
, cpu_to_node(cpu
));
7820 rcu_assign_pointer(swhash
->swevent_hlist
, hlist
);
7822 mutex_unlock(&swhash
->hlist_mutex
);
7825 #if defined CONFIG_HOTPLUG_CPU || defined CONFIG_KEXEC
7826 static void perf_pmu_rotate_stop(struct pmu
*pmu
)
7828 struct perf_cpu_context
*cpuctx
= this_cpu_ptr(pmu
->pmu_cpu_context
);
7830 WARN_ON(!irqs_disabled());
7832 list_del_init(&cpuctx
->rotation_list
);
7835 static void __perf_event_exit_context(void *__info
)
7837 struct perf_event_context
*ctx
= __info
;
7838 struct perf_event
*event
, *tmp
;
7840 perf_pmu_rotate_stop(ctx
->pmu
);
7842 list_for_each_entry_safe(event
, tmp
, &ctx
->pinned_groups
, group_entry
)
7843 __perf_remove_from_context(event
);
7844 list_for_each_entry_safe(event
, tmp
, &ctx
->flexible_groups
, group_entry
)
7845 __perf_remove_from_context(event
);
7848 static void perf_event_exit_cpu_context(int cpu
)
7850 struct perf_event_context
*ctx
;
7854 idx
= srcu_read_lock(&pmus_srcu
);
7855 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
7856 ctx
= &per_cpu_ptr(pmu
->pmu_cpu_context
, cpu
)->ctx
;
7858 mutex_lock(&ctx
->mutex
);
7859 smp_call_function_single(cpu
, __perf_event_exit_context
, ctx
, 1);
7860 mutex_unlock(&ctx
->mutex
);
7862 srcu_read_unlock(&pmus_srcu
, idx
);
7865 static void perf_event_exit_cpu(int cpu
)
7867 struct swevent_htable
*swhash
= &per_cpu(swevent_htable
, cpu
);
7869 mutex_lock(&swhash
->hlist_mutex
);
7870 swevent_hlist_release(swhash
);
7871 mutex_unlock(&swhash
->hlist_mutex
);
7873 perf_event_exit_cpu_context(cpu
);
7876 static inline void perf_event_exit_cpu(int cpu
) { }
7880 perf_reboot(struct notifier_block
*notifier
, unsigned long val
, void *v
)
7884 for_each_online_cpu(cpu
)
7885 perf_event_exit_cpu(cpu
);
7891 * Run the perf reboot notifier at the very last possible moment so that
7892 * the generic watchdog code runs as long as possible.
7894 static struct notifier_block perf_reboot_notifier
= {
7895 .notifier_call
= perf_reboot
,
7896 .priority
= INT_MIN
,
7900 perf_cpu_notify(struct notifier_block
*self
, unsigned long action
, void *hcpu
)
7902 unsigned int cpu
= (long)hcpu
;
7904 switch (action
& ~CPU_TASKS_FROZEN
) {
7906 case CPU_UP_PREPARE
:
7907 case CPU_DOWN_FAILED
:
7908 perf_event_init_cpu(cpu
);
7911 case CPU_UP_CANCELED
:
7912 case CPU_DOWN_PREPARE
:
7913 perf_event_exit_cpu(cpu
);
7922 void __init
perf_event_init(void)
7928 perf_event_init_all_cpus();
7929 init_srcu_struct(&pmus_srcu
);
7930 perf_pmu_register(&perf_swevent
, "software", PERF_TYPE_SOFTWARE
);
7931 perf_pmu_register(&perf_cpu_clock
, NULL
, -1);
7932 perf_pmu_register(&perf_task_clock
, NULL
, -1);
7934 perf_cpu_notifier(perf_cpu_notify
);
7935 register_reboot_notifier(&perf_reboot_notifier
);
7937 ret
= init_hw_breakpoint();
7938 WARN(ret
, "hw_breakpoint initialization failed with: %d", ret
);
7940 /* do not patch jump label more than once per second */
7941 jump_label_rate_limit(&perf_sched_events
, HZ
);
7944 * Build time assertion that we keep the data_head at the intended
7945 * location. IOW, validation we got the __reserved[] size right.
7947 BUILD_BUG_ON((offsetof(struct perf_event_mmap_page
, data_head
))
7951 static int __init
perf_event_sysfs_init(void)
7956 mutex_lock(&pmus_lock
);
7958 ret
= bus_register(&pmu_bus
);
7962 list_for_each_entry(pmu
, &pmus
, entry
) {
7963 if (!pmu
->name
|| pmu
->type
< 0)
7966 ret
= pmu_dev_alloc(pmu
);
7967 WARN(ret
, "Failed to register pmu: %s, reason %d\n", pmu
->name
, ret
);
7969 pmu_bus_running
= 1;
7973 mutex_unlock(&pmus_lock
);
7977 device_initcall(perf_event_sysfs_init
);
7979 #ifdef CONFIG_CGROUP_PERF
7980 static struct cgroup_subsys_state
*
7981 perf_cgroup_css_alloc(struct cgroup_subsys_state
*parent_css
)
7983 struct perf_cgroup
*jc
;
7985 jc
= kzalloc(sizeof(*jc
), GFP_KERNEL
);
7987 return ERR_PTR(-ENOMEM
);
7989 jc
->info
= alloc_percpu(struct perf_cgroup_info
);
7992 return ERR_PTR(-ENOMEM
);
7998 static void perf_cgroup_css_free(struct cgroup_subsys_state
*css
)
8000 struct perf_cgroup
*jc
= container_of(css
, struct perf_cgroup
, css
);
8002 free_percpu(jc
->info
);
8006 static int __perf_cgroup_move(void *info
)
8008 struct task_struct
*task
= info
;
8009 perf_cgroup_switch(task
, PERF_CGROUP_SWOUT
| PERF_CGROUP_SWIN
);
8013 static void perf_cgroup_attach(struct cgroup_subsys_state
*css
,
8014 struct cgroup_taskset
*tset
)
8016 struct task_struct
*task
;
8018 cgroup_taskset_for_each(task
, css
, tset
)
8019 task_function_call(task
, __perf_cgroup_move
, task
);
8022 static void perf_cgroup_exit(struct cgroup_subsys_state
*css
,
8023 struct cgroup_subsys_state
*old_css
,
8024 struct task_struct
*task
)
8027 * cgroup_exit() is called in the copy_process() failure path.
8028 * Ignore this case since the task hasn't ran yet, this avoids
8029 * trying to poke a half freed task state from generic code.
8031 if (!(task
->flags
& PF_EXITING
))
8034 task_function_call(task
, __perf_cgroup_move
, task
);
8037 struct cgroup_subsys perf_subsys
= {
8038 .name
= "perf_event",
8039 .subsys_id
= perf_subsys_id
,
8040 .css_alloc
= perf_cgroup_css_alloc
,
8041 .css_free
= perf_cgroup_css_free
,
8042 .exit
= perf_cgroup_exit
,
8043 .attach
= perf_cgroup_attach
,
8045 #endif /* CONFIG_CGROUP_PERF */