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 |\
123 PERF_FLAG_FD_CLOEXEC)
126 * branch priv levels that need permission checks
128 #define PERF_SAMPLE_BRANCH_PERM_PLM \
129 (PERF_SAMPLE_BRANCH_KERNEL |\
130 PERF_SAMPLE_BRANCH_HV)
133 EVENT_FLEXIBLE
= 0x1,
135 EVENT_ALL
= EVENT_FLEXIBLE
| EVENT_PINNED
,
139 * perf_sched_events : >0 events exist
140 * perf_cgroup_events: >0 per-cpu cgroup events exist on this cpu
142 struct static_key_deferred perf_sched_events __read_mostly
;
143 static DEFINE_PER_CPU(atomic_t
, perf_cgroup_events
);
144 static DEFINE_PER_CPU(atomic_t
, perf_branch_stack_events
);
146 static atomic_t nr_mmap_events __read_mostly
;
147 static atomic_t nr_comm_events __read_mostly
;
148 static atomic_t nr_task_events __read_mostly
;
149 static atomic_t nr_freq_events __read_mostly
;
151 static LIST_HEAD(pmus
);
152 static DEFINE_MUTEX(pmus_lock
);
153 static struct srcu_struct pmus_srcu
;
156 * perf event paranoia level:
157 * -1 - not paranoid at all
158 * 0 - disallow raw tracepoint access for unpriv
159 * 1 - disallow cpu events for unpriv
160 * 2 - disallow kernel profiling for unpriv
162 int sysctl_perf_event_paranoid __read_mostly
= 1;
164 /* Minimum for 512 kiB + 1 user control page */
165 int sysctl_perf_event_mlock __read_mostly
= 512 + (PAGE_SIZE
/ 1024); /* 'free' kiB per user */
168 * max perf event sample rate
170 #define DEFAULT_MAX_SAMPLE_RATE 100000
171 #define DEFAULT_SAMPLE_PERIOD_NS (NSEC_PER_SEC / DEFAULT_MAX_SAMPLE_RATE)
172 #define DEFAULT_CPU_TIME_MAX_PERCENT 25
174 int sysctl_perf_event_sample_rate __read_mostly
= DEFAULT_MAX_SAMPLE_RATE
;
176 static int max_samples_per_tick __read_mostly
= DIV_ROUND_UP(DEFAULT_MAX_SAMPLE_RATE
, HZ
);
177 static int perf_sample_period_ns __read_mostly
= DEFAULT_SAMPLE_PERIOD_NS
;
179 static int perf_sample_allowed_ns __read_mostly
=
180 DEFAULT_SAMPLE_PERIOD_NS
* DEFAULT_CPU_TIME_MAX_PERCENT
/ 100;
182 void update_perf_cpu_limits(void)
184 u64 tmp
= perf_sample_period_ns
;
186 tmp
*= sysctl_perf_cpu_time_max_percent
;
188 ACCESS_ONCE(perf_sample_allowed_ns
) = tmp
;
191 static int perf_rotate_context(struct perf_cpu_context
*cpuctx
);
193 int perf_proc_update_handler(struct ctl_table
*table
, int write
,
194 void __user
*buffer
, size_t *lenp
,
197 int ret
= proc_dointvec_minmax(table
, write
, buffer
, lenp
, ppos
);
202 max_samples_per_tick
= DIV_ROUND_UP(sysctl_perf_event_sample_rate
, HZ
);
203 perf_sample_period_ns
= NSEC_PER_SEC
/ sysctl_perf_event_sample_rate
;
204 update_perf_cpu_limits();
209 int sysctl_perf_cpu_time_max_percent __read_mostly
= DEFAULT_CPU_TIME_MAX_PERCENT
;
211 int perf_cpu_time_max_percent_handler(struct ctl_table
*table
, int write
,
212 void __user
*buffer
, size_t *lenp
,
215 int ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
220 update_perf_cpu_limits();
226 * perf samples are done in some very critical code paths (NMIs).
227 * If they take too much CPU time, the system can lock up and not
228 * get any real work done. This will drop the sample rate when
229 * we detect that events are taking too long.
231 #define NR_ACCUMULATED_SAMPLES 128
232 static DEFINE_PER_CPU(u64
, running_sample_length
);
234 static void perf_duration_warn(struct irq_work
*w
)
236 u64 allowed_ns
= ACCESS_ONCE(perf_sample_allowed_ns
);
237 u64 avg_local_sample_len
;
238 u64 local_samples_len
;
240 local_samples_len
= __get_cpu_var(running_sample_length
);
241 avg_local_sample_len
= local_samples_len
/NR_ACCUMULATED_SAMPLES
;
243 printk_ratelimited(KERN_WARNING
244 "perf interrupt took too long (%lld > %lld), lowering "
245 "kernel.perf_event_max_sample_rate to %d\n",
246 avg_local_sample_len
, allowed_ns
>> 1,
247 sysctl_perf_event_sample_rate
);
250 static DEFINE_IRQ_WORK(perf_duration_work
, perf_duration_warn
);
252 void perf_sample_event_took(u64 sample_len_ns
)
254 u64 allowed_ns
= ACCESS_ONCE(perf_sample_allowed_ns
);
255 u64 avg_local_sample_len
;
256 u64 local_samples_len
;
261 /* decay the counter by 1 average sample */
262 local_samples_len
= __get_cpu_var(running_sample_length
);
263 local_samples_len
-= local_samples_len
/NR_ACCUMULATED_SAMPLES
;
264 local_samples_len
+= sample_len_ns
;
265 __get_cpu_var(running_sample_length
) = local_samples_len
;
268 * note: this will be biased artifically low until we have
269 * seen NR_ACCUMULATED_SAMPLES. Doing it this way keeps us
270 * from having to maintain a count.
272 avg_local_sample_len
= local_samples_len
/NR_ACCUMULATED_SAMPLES
;
274 if (avg_local_sample_len
<= allowed_ns
)
277 if (max_samples_per_tick
<= 1)
280 max_samples_per_tick
= DIV_ROUND_UP(max_samples_per_tick
, 2);
281 sysctl_perf_event_sample_rate
= max_samples_per_tick
* HZ
;
282 perf_sample_period_ns
= NSEC_PER_SEC
/ sysctl_perf_event_sample_rate
;
284 update_perf_cpu_limits();
286 if (!irq_work_queue(&perf_duration_work
)) {
287 early_printk("perf interrupt took too long (%lld > %lld), lowering "
288 "kernel.perf_event_max_sample_rate to %d\n",
289 avg_local_sample_len
, allowed_ns
>> 1,
290 sysctl_perf_event_sample_rate
);
294 static atomic64_t perf_event_id
;
296 static void cpu_ctx_sched_out(struct perf_cpu_context
*cpuctx
,
297 enum event_type_t event_type
);
299 static void cpu_ctx_sched_in(struct perf_cpu_context
*cpuctx
,
300 enum event_type_t event_type
,
301 struct task_struct
*task
);
303 static void update_context_time(struct perf_event_context
*ctx
);
304 static u64
perf_event_time(struct perf_event
*event
);
306 void __weak
perf_event_print_debug(void) { }
308 extern __weak
const char *perf_pmu_name(void)
313 static inline u64
perf_clock(void)
315 return local_clock();
318 static inline struct perf_cpu_context
*
319 __get_cpu_context(struct perf_event_context
*ctx
)
321 return this_cpu_ptr(ctx
->pmu
->pmu_cpu_context
);
324 static void perf_ctx_lock(struct perf_cpu_context
*cpuctx
,
325 struct perf_event_context
*ctx
)
327 raw_spin_lock(&cpuctx
->ctx
.lock
);
329 raw_spin_lock(&ctx
->lock
);
332 static void perf_ctx_unlock(struct perf_cpu_context
*cpuctx
,
333 struct perf_event_context
*ctx
)
336 raw_spin_unlock(&ctx
->lock
);
337 raw_spin_unlock(&cpuctx
->ctx
.lock
);
340 #ifdef CONFIG_CGROUP_PERF
343 * perf_cgroup_info keeps track of time_enabled for a cgroup.
344 * This is a per-cpu dynamically allocated data structure.
346 struct perf_cgroup_info
{
352 struct cgroup_subsys_state css
;
353 struct perf_cgroup_info __percpu
*info
;
357 * Must ensure cgroup is pinned (css_get) before calling
358 * this function. In other words, we cannot call this function
359 * if there is no cgroup event for the current CPU context.
361 static inline struct perf_cgroup
*
362 perf_cgroup_from_task(struct task_struct
*task
)
364 return container_of(task_css(task
, perf_event_cgrp_id
),
365 struct perf_cgroup
, css
);
369 perf_cgroup_match(struct perf_event
*event
)
371 struct perf_event_context
*ctx
= event
->ctx
;
372 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
374 /* @event doesn't care about cgroup */
378 /* wants specific cgroup scope but @cpuctx isn't associated with any */
383 * Cgroup scoping is recursive. An event enabled for a cgroup is
384 * also enabled for all its descendant cgroups. If @cpuctx's
385 * cgroup is a descendant of @event's (the test covers identity
386 * case), it's a match.
388 return cgroup_is_descendant(cpuctx
->cgrp
->css
.cgroup
,
389 event
->cgrp
->css
.cgroup
);
392 static inline void perf_put_cgroup(struct perf_event
*event
)
394 css_put(&event
->cgrp
->css
);
397 static inline void perf_detach_cgroup(struct perf_event
*event
)
399 perf_put_cgroup(event
);
403 static inline int is_cgroup_event(struct perf_event
*event
)
405 return event
->cgrp
!= NULL
;
408 static inline u64
perf_cgroup_event_time(struct perf_event
*event
)
410 struct perf_cgroup_info
*t
;
412 t
= per_cpu_ptr(event
->cgrp
->info
, event
->cpu
);
416 static inline void __update_cgrp_time(struct perf_cgroup
*cgrp
)
418 struct perf_cgroup_info
*info
;
423 info
= this_cpu_ptr(cgrp
->info
);
425 info
->time
+= now
- info
->timestamp
;
426 info
->timestamp
= now
;
429 static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context
*cpuctx
)
431 struct perf_cgroup
*cgrp_out
= cpuctx
->cgrp
;
433 __update_cgrp_time(cgrp_out
);
436 static inline void update_cgrp_time_from_event(struct perf_event
*event
)
438 struct perf_cgroup
*cgrp
;
441 * ensure we access cgroup data only when needed and
442 * when we know the cgroup is pinned (css_get)
444 if (!is_cgroup_event(event
))
447 cgrp
= perf_cgroup_from_task(current
);
449 * Do not update time when cgroup is not active
451 if (cgrp
== event
->cgrp
)
452 __update_cgrp_time(event
->cgrp
);
456 perf_cgroup_set_timestamp(struct task_struct
*task
,
457 struct perf_event_context
*ctx
)
459 struct perf_cgroup
*cgrp
;
460 struct perf_cgroup_info
*info
;
463 * ctx->lock held by caller
464 * ensure we do not access cgroup data
465 * unless we have the cgroup pinned (css_get)
467 if (!task
|| !ctx
->nr_cgroups
)
470 cgrp
= perf_cgroup_from_task(task
);
471 info
= this_cpu_ptr(cgrp
->info
);
472 info
->timestamp
= ctx
->timestamp
;
475 #define PERF_CGROUP_SWOUT 0x1 /* cgroup switch out every event */
476 #define PERF_CGROUP_SWIN 0x2 /* cgroup switch in events based on task */
479 * reschedule events based on the cgroup constraint of task.
481 * mode SWOUT : schedule out everything
482 * mode SWIN : schedule in based on cgroup for next
484 void perf_cgroup_switch(struct task_struct
*task
, int mode
)
486 struct perf_cpu_context
*cpuctx
;
491 * disable interrupts to avoid geting nr_cgroup
492 * changes via __perf_event_disable(). Also
495 local_irq_save(flags
);
498 * we reschedule only in the presence of cgroup
499 * constrained events.
503 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
504 cpuctx
= this_cpu_ptr(pmu
->pmu_cpu_context
);
505 if (cpuctx
->unique_pmu
!= pmu
)
506 continue; /* ensure we process each cpuctx once */
509 * perf_cgroup_events says at least one
510 * context on this CPU has cgroup events.
512 * ctx->nr_cgroups reports the number of cgroup
513 * events for a context.
515 if (cpuctx
->ctx
.nr_cgroups
> 0) {
516 perf_ctx_lock(cpuctx
, cpuctx
->task_ctx
);
517 perf_pmu_disable(cpuctx
->ctx
.pmu
);
519 if (mode
& PERF_CGROUP_SWOUT
) {
520 cpu_ctx_sched_out(cpuctx
, EVENT_ALL
);
522 * must not be done before ctxswout due
523 * to event_filter_match() in event_sched_out()
528 if (mode
& PERF_CGROUP_SWIN
) {
529 WARN_ON_ONCE(cpuctx
->cgrp
);
531 * set cgrp before ctxsw in to allow
532 * event_filter_match() to not have to pass
535 cpuctx
->cgrp
= perf_cgroup_from_task(task
);
536 cpu_ctx_sched_in(cpuctx
, EVENT_ALL
, task
);
538 perf_pmu_enable(cpuctx
->ctx
.pmu
);
539 perf_ctx_unlock(cpuctx
, cpuctx
->task_ctx
);
545 local_irq_restore(flags
);
548 static inline void perf_cgroup_sched_out(struct task_struct
*task
,
549 struct task_struct
*next
)
551 struct perf_cgroup
*cgrp1
;
552 struct perf_cgroup
*cgrp2
= NULL
;
555 * we come here when we know perf_cgroup_events > 0
557 cgrp1
= perf_cgroup_from_task(task
);
560 * next is NULL when called from perf_event_enable_on_exec()
561 * that will systematically cause a cgroup_switch()
564 cgrp2
= perf_cgroup_from_task(next
);
567 * only schedule out current cgroup events if we know
568 * that we are switching to a different cgroup. Otherwise,
569 * do no touch the cgroup events.
572 perf_cgroup_switch(task
, PERF_CGROUP_SWOUT
);
575 static inline void perf_cgroup_sched_in(struct task_struct
*prev
,
576 struct task_struct
*task
)
578 struct perf_cgroup
*cgrp1
;
579 struct perf_cgroup
*cgrp2
= NULL
;
582 * we come here when we know perf_cgroup_events > 0
584 cgrp1
= perf_cgroup_from_task(task
);
586 /* prev can never be NULL */
587 cgrp2
= perf_cgroup_from_task(prev
);
590 * only need to schedule in cgroup events if we are changing
591 * cgroup during ctxsw. Cgroup events were not scheduled
592 * out of ctxsw out if that was not the case.
595 perf_cgroup_switch(task
, PERF_CGROUP_SWIN
);
598 static inline int perf_cgroup_connect(int fd
, struct perf_event
*event
,
599 struct perf_event_attr
*attr
,
600 struct perf_event
*group_leader
)
602 struct perf_cgroup
*cgrp
;
603 struct cgroup_subsys_state
*css
;
604 struct fd f
= fdget(fd
);
610 css
= css_tryget_from_dir(f
.file
->f_dentry
, &perf_event_cgrp_subsys
);
616 cgrp
= container_of(css
, struct perf_cgroup
, css
);
620 * all events in a group must monitor
621 * the same cgroup because a task belongs
622 * to only one perf cgroup at a time
624 if (group_leader
&& group_leader
->cgrp
!= cgrp
) {
625 perf_detach_cgroup(event
);
634 perf_cgroup_set_shadow_time(struct perf_event
*event
, u64 now
)
636 struct perf_cgroup_info
*t
;
637 t
= per_cpu_ptr(event
->cgrp
->info
, event
->cpu
);
638 event
->shadow_ctx_time
= now
- t
->timestamp
;
642 perf_cgroup_defer_enabled(struct perf_event
*event
)
645 * when the current task's perf cgroup does not match
646 * the event's, we need to remember to call the
647 * perf_mark_enable() function the first time a task with
648 * a matching perf cgroup is scheduled in.
650 if (is_cgroup_event(event
) && !perf_cgroup_match(event
))
651 event
->cgrp_defer_enabled
= 1;
655 perf_cgroup_mark_enabled(struct perf_event
*event
,
656 struct perf_event_context
*ctx
)
658 struct perf_event
*sub
;
659 u64 tstamp
= perf_event_time(event
);
661 if (!event
->cgrp_defer_enabled
)
664 event
->cgrp_defer_enabled
= 0;
666 event
->tstamp_enabled
= tstamp
- event
->total_time_enabled
;
667 list_for_each_entry(sub
, &event
->sibling_list
, group_entry
) {
668 if (sub
->state
>= PERF_EVENT_STATE_INACTIVE
) {
669 sub
->tstamp_enabled
= tstamp
- sub
->total_time_enabled
;
670 sub
->cgrp_defer_enabled
= 0;
674 #else /* !CONFIG_CGROUP_PERF */
677 perf_cgroup_match(struct perf_event
*event
)
682 static inline void perf_detach_cgroup(struct perf_event
*event
)
685 static inline int is_cgroup_event(struct perf_event
*event
)
690 static inline u64
perf_cgroup_event_cgrp_time(struct perf_event
*event
)
695 static inline void update_cgrp_time_from_event(struct perf_event
*event
)
699 static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context
*cpuctx
)
703 static inline void perf_cgroup_sched_out(struct task_struct
*task
,
704 struct task_struct
*next
)
708 static inline void perf_cgroup_sched_in(struct task_struct
*prev
,
709 struct task_struct
*task
)
713 static inline int perf_cgroup_connect(pid_t pid
, struct perf_event
*event
,
714 struct perf_event_attr
*attr
,
715 struct perf_event
*group_leader
)
721 perf_cgroup_set_timestamp(struct task_struct
*task
,
722 struct perf_event_context
*ctx
)
727 perf_cgroup_switch(struct task_struct
*task
, struct task_struct
*next
)
732 perf_cgroup_set_shadow_time(struct perf_event
*event
, u64 now
)
736 static inline u64
perf_cgroup_event_time(struct perf_event
*event
)
742 perf_cgroup_defer_enabled(struct perf_event
*event
)
747 perf_cgroup_mark_enabled(struct perf_event
*event
,
748 struct perf_event_context
*ctx
)
754 * set default to be dependent on timer tick just
757 #define PERF_CPU_HRTIMER (1000 / HZ)
759 * function must be called with interrupts disbled
761 static enum hrtimer_restart
perf_cpu_hrtimer_handler(struct hrtimer
*hr
)
763 struct perf_cpu_context
*cpuctx
;
764 enum hrtimer_restart ret
= HRTIMER_NORESTART
;
767 WARN_ON(!irqs_disabled());
769 cpuctx
= container_of(hr
, struct perf_cpu_context
, hrtimer
);
771 rotations
= perf_rotate_context(cpuctx
);
774 * arm timer if needed
777 hrtimer_forward_now(hr
, cpuctx
->hrtimer_interval
);
778 ret
= HRTIMER_RESTART
;
784 /* CPU is going down */
785 void perf_cpu_hrtimer_cancel(int cpu
)
787 struct perf_cpu_context
*cpuctx
;
791 if (WARN_ON(cpu
!= smp_processor_id()))
794 local_irq_save(flags
);
798 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
799 cpuctx
= this_cpu_ptr(pmu
->pmu_cpu_context
);
801 if (pmu
->task_ctx_nr
== perf_sw_context
)
804 hrtimer_cancel(&cpuctx
->hrtimer
);
809 local_irq_restore(flags
);
812 static void __perf_cpu_hrtimer_init(struct perf_cpu_context
*cpuctx
, int cpu
)
814 struct hrtimer
*hr
= &cpuctx
->hrtimer
;
815 struct pmu
*pmu
= cpuctx
->ctx
.pmu
;
818 /* no multiplexing needed for SW PMU */
819 if (pmu
->task_ctx_nr
== perf_sw_context
)
823 * check default is sane, if not set then force to
824 * default interval (1/tick)
826 timer
= pmu
->hrtimer_interval_ms
;
828 timer
= pmu
->hrtimer_interval_ms
= PERF_CPU_HRTIMER
;
830 cpuctx
->hrtimer_interval
= ns_to_ktime(NSEC_PER_MSEC
* timer
);
832 hrtimer_init(hr
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL_PINNED
);
833 hr
->function
= perf_cpu_hrtimer_handler
;
836 static void perf_cpu_hrtimer_restart(struct perf_cpu_context
*cpuctx
)
838 struct hrtimer
*hr
= &cpuctx
->hrtimer
;
839 struct pmu
*pmu
= cpuctx
->ctx
.pmu
;
842 if (pmu
->task_ctx_nr
== perf_sw_context
)
845 if (hrtimer_active(hr
))
848 if (!hrtimer_callback_running(hr
))
849 __hrtimer_start_range_ns(hr
, cpuctx
->hrtimer_interval
,
850 0, HRTIMER_MODE_REL_PINNED
, 0);
853 void perf_pmu_disable(struct pmu
*pmu
)
855 int *count
= this_cpu_ptr(pmu
->pmu_disable_count
);
857 pmu
->pmu_disable(pmu
);
860 void perf_pmu_enable(struct pmu
*pmu
)
862 int *count
= this_cpu_ptr(pmu
->pmu_disable_count
);
864 pmu
->pmu_enable(pmu
);
867 static DEFINE_PER_CPU(struct list_head
, rotation_list
);
870 * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
871 * because they're strictly cpu affine and rotate_start is called with IRQs
872 * disabled, while rotate_context is called from IRQ context.
874 static void perf_pmu_rotate_start(struct pmu
*pmu
)
876 struct perf_cpu_context
*cpuctx
= this_cpu_ptr(pmu
->pmu_cpu_context
);
877 struct list_head
*head
= &__get_cpu_var(rotation_list
);
879 WARN_ON(!irqs_disabled());
881 if (list_empty(&cpuctx
->rotation_list
))
882 list_add(&cpuctx
->rotation_list
, head
);
885 static void get_ctx(struct perf_event_context
*ctx
)
887 WARN_ON(!atomic_inc_not_zero(&ctx
->refcount
));
890 static void put_ctx(struct perf_event_context
*ctx
)
892 if (atomic_dec_and_test(&ctx
->refcount
)) {
894 put_ctx(ctx
->parent_ctx
);
896 put_task_struct(ctx
->task
);
897 kfree_rcu(ctx
, rcu_head
);
901 static void unclone_ctx(struct perf_event_context
*ctx
)
903 if (ctx
->parent_ctx
) {
904 put_ctx(ctx
->parent_ctx
);
905 ctx
->parent_ctx
= NULL
;
910 static u32
perf_event_pid(struct perf_event
*event
, struct task_struct
*p
)
913 * only top level events have the pid namespace they were created in
916 event
= event
->parent
;
918 return task_tgid_nr_ns(p
, event
->ns
);
921 static u32
perf_event_tid(struct perf_event
*event
, struct task_struct
*p
)
924 * only top level events have the pid namespace they were created in
927 event
= event
->parent
;
929 return task_pid_nr_ns(p
, event
->ns
);
933 * If we inherit events we want to return the parent event id
936 static u64
primary_event_id(struct perf_event
*event
)
941 id
= event
->parent
->id
;
947 * Get the perf_event_context for a task and lock it.
948 * This has to cope with with the fact that until it is locked,
949 * the context could get moved to another task.
951 static struct perf_event_context
*
952 perf_lock_task_context(struct task_struct
*task
, int ctxn
, unsigned long *flags
)
954 struct perf_event_context
*ctx
;
958 * One of the few rules of preemptible RCU is that one cannot do
959 * rcu_read_unlock() while holding a scheduler (or nested) lock when
960 * part of the read side critical section was preemptible -- see
961 * rcu_read_unlock_special().
963 * Since ctx->lock nests under rq->lock we must ensure the entire read
964 * side critical section is non-preemptible.
968 ctx
= rcu_dereference(task
->perf_event_ctxp
[ctxn
]);
971 * If this context is a clone of another, it might
972 * get swapped for another underneath us by
973 * perf_event_task_sched_out, though the
974 * rcu_read_lock() protects us from any context
975 * getting freed. Lock the context and check if it
976 * got swapped before we could get the lock, and retry
977 * if so. If we locked the right context, then it
978 * can't get swapped on us any more.
980 raw_spin_lock_irqsave(&ctx
->lock
, *flags
);
981 if (ctx
!= rcu_dereference(task
->perf_event_ctxp
[ctxn
])) {
982 raw_spin_unlock_irqrestore(&ctx
->lock
, *flags
);
988 if (!atomic_inc_not_zero(&ctx
->refcount
)) {
989 raw_spin_unlock_irqrestore(&ctx
->lock
, *flags
);
999 * Get the context for a task and increment its pin_count so it
1000 * can't get swapped to another task. This also increments its
1001 * reference count so that the context can't get freed.
1003 static struct perf_event_context
*
1004 perf_pin_task_context(struct task_struct
*task
, int ctxn
)
1006 struct perf_event_context
*ctx
;
1007 unsigned long flags
;
1009 ctx
= perf_lock_task_context(task
, ctxn
, &flags
);
1012 raw_spin_unlock_irqrestore(&ctx
->lock
, flags
);
1017 static void perf_unpin_context(struct perf_event_context
*ctx
)
1019 unsigned long flags
;
1021 raw_spin_lock_irqsave(&ctx
->lock
, flags
);
1023 raw_spin_unlock_irqrestore(&ctx
->lock
, flags
);
1027 * Update the record of the current time in a context.
1029 static void update_context_time(struct perf_event_context
*ctx
)
1031 u64 now
= perf_clock();
1033 ctx
->time
+= now
- ctx
->timestamp
;
1034 ctx
->timestamp
= now
;
1037 static u64
perf_event_time(struct perf_event
*event
)
1039 struct perf_event_context
*ctx
= event
->ctx
;
1041 if (is_cgroup_event(event
))
1042 return perf_cgroup_event_time(event
);
1044 return ctx
? ctx
->time
: 0;
1048 * Update the total_time_enabled and total_time_running fields for a event.
1049 * The caller of this function needs to hold the ctx->lock.
1051 static void update_event_times(struct perf_event
*event
)
1053 struct perf_event_context
*ctx
= event
->ctx
;
1056 if (event
->state
< PERF_EVENT_STATE_INACTIVE
||
1057 event
->group_leader
->state
< PERF_EVENT_STATE_INACTIVE
)
1060 * in cgroup mode, time_enabled represents
1061 * the time the event was enabled AND active
1062 * tasks were in the monitored cgroup. This is
1063 * independent of the activity of the context as
1064 * there may be a mix of cgroup and non-cgroup events.
1066 * That is why we treat cgroup events differently
1069 if (is_cgroup_event(event
))
1070 run_end
= perf_cgroup_event_time(event
);
1071 else if (ctx
->is_active
)
1072 run_end
= ctx
->time
;
1074 run_end
= event
->tstamp_stopped
;
1076 event
->total_time_enabled
= run_end
- event
->tstamp_enabled
;
1078 if (event
->state
== PERF_EVENT_STATE_INACTIVE
)
1079 run_end
= event
->tstamp_stopped
;
1081 run_end
= perf_event_time(event
);
1083 event
->total_time_running
= run_end
- event
->tstamp_running
;
1088 * Update total_time_enabled and total_time_running for all events in a group.
1090 static void update_group_times(struct perf_event
*leader
)
1092 struct perf_event
*event
;
1094 update_event_times(leader
);
1095 list_for_each_entry(event
, &leader
->sibling_list
, group_entry
)
1096 update_event_times(event
);
1099 static struct list_head
*
1100 ctx_group_list(struct perf_event
*event
, struct perf_event_context
*ctx
)
1102 if (event
->attr
.pinned
)
1103 return &ctx
->pinned_groups
;
1105 return &ctx
->flexible_groups
;
1109 * Add a event from the lists for its context.
1110 * Must be called with ctx->mutex and ctx->lock held.
1113 list_add_event(struct perf_event
*event
, struct perf_event_context
*ctx
)
1115 WARN_ON_ONCE(event
->attach_state
& PERF_ATTACH_CONTEXT
);
1116 event
->attach_state
|= PERF_ATTACH_CONTEXT
;
1119 * If we're a stand alone event or group leader, we go to the context
1120 * list, group events are kept attached to the group so that
1121 * perf_group_detach can, at all times, locate all siblings.
1123 if (event
->group_leader
== event
) {
1124 struct list_head
*list
;
1126 if (is_software_event(event
))
1127 event
->group_flags
|= PERF_GROUP_SOFTWARE
;
1129 list
= ctx_group_list(event
, ctx
);
1130 list_add_tail(&event
->group_entry
, list
);
1133 if (is_cgroup_event(event
))
1136 if (has_branch_stack(event
))
1137 ctx
->nr_branch_stack
++;
1139 list_add_rcu(&event
->event_entry
, &ctx
->event_list
);
1140 if (!ctx
->nr_events
)
1141 perf_pmu_rotate_start(ctx
->pmu
);
1143 if (event
->attr
.inherit_stat
)
1150 * Initialize event state based on the perf_event_attr::disabled.
1152 static inline void perf_event__state_init(struct perf_event
*event
)
1154 event
->state
= event
->attr
.disabled
? PERF_EVENT_STATE_OFF
:
1155 PERF_EVENT_STATE_INACTIVE
;
1159 * Called at perf_event creation and when events are attached/detached from a
1162 static void perf_event__read_size(struct perf_event
*event
)
1164 int entry
= sizeof(u64
); /* value */
1168 if (event
->attr
.read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
1169 size
+= sizeof(u64
);
1171 if (event
->attr
.read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
1172 size
+= sizeof(u64
);
1174 if (event
->attr
.read_format
& PERF_FORMAT_ID
)
1175 entry
+= sizeof(u64
);
1177 if (event
->attr
.read_format
& PERF_FORMAT_GROUP
) {
1178 nr
+= event
->group_leader
->nr_siblings
;
1179 size
+= sizeof(u64
);
1183 event
->read_size
= size
;
1186 static void perf_event__header_size(struct perf_event
*event
)
1188 struct perf_sample_data
*data
;
1189 u64 sample_type
= event
->attr
.sample_type
;
1192 perf_event__read_size(event
);
1194 if (sample_type
& PERF_SAMPLE_IP
)
1195 size
+= sizeof(data
->ip
);
1197 if (sample_type
& PERF_SAMPLE_ADDR
)
1198 size
+= sizeof(data
->addr
);
1200 if (sample_type
& PERF_SAMPLE_PERIOD
)
1201 size
+= sizeof(data
->period
);
1203 if (sample_type
& PERF_SAMPLE_WEIGHT
)
1204 size
+= sizeof(data
->weight
);
1206 if (sample_type
& PERF_SAMPLE_READ
)
1207 size
+= event
->read_size
;
1209 if (sample_type
& PERF_SAMPLE_DATA_SRC
)
1210 size
+= sizeof(data
->data_src
.val
);
1212 if (sample_type
& PERF_SAMPLE_TRANSACTION
)
1213 size
+= sizeof(data
->txn
);
1215 event
->header_size
= size
;
1218 static void perf_event__id_header_size(struct perf_event
*event
)
1220 struct perf_sample_data
*data
;
1221 u64 sample_type
= event
->attr
.sample_type
;
1224 if (sample_type
& PERF_SAMPLE_TID
)
1225 size
+= sizeof(data
->tid_entry
);
1227 if (sample_type
& PERF_SAMPLE_TIME
)
1228 size
+= sizeof(data
->time
);
1230 if (sample_type
& PERF_SAMPLE_IDENTIFIER
)
1231 size
+= sizeof(data
->id
);
1233 if (sample_type
& PERF_SAMPLE_ID
)
1234 size
+= sizeof(data
->id
);
1236 if (sample_type
& PERF_SAMPLE_STREAM_ID
)
1237 size
+= sizeof(data
->stream_id
);
1239 if (sample_type
& PERF_SAMPLE_CPU
)
1240 size
+= sizeof(data
->cpu_entry
);
1242 event
->id_header_size
= size
;
1245 static void perf_group_attach(struct perf_event
*event
)
1247 struct perf_event
*group_leader
= event
->group_leader
, *pos
;
1250 * We can have double attach due to group movement in perf_event_open.
1252 if (event
->attach_state
& PERF_ATTACH_GROUP
)
1255 event
->attach_state
|= PERF_ATTACH_GROUP
;
1257 if (group_leader
== event
)
1260 if (group_leader
->group_flags
& PERF_GROUP_SOFTWARE
&&
1261 !is_software_event(event
))
1262 group_leader
->group_flags
&= ~PERF_GROUP_SOFTWARE
;
1264 list_add_tail(&event
->group_entry
, &group_leader
->sibling_list
);
1265 group_leader
->nr_siblings
++;
1267 perf_event__header_size(group_leader
);
1269 list_for_each_entry(pos
, &group_leader
->sibling_list
, group_entry
)
1270 perf_event__header_size(pos
);
1274 * Remove a event from the lists for its context.
1275 * Must be called with ctx->mutex and ctx->lock held.
1278 list_del_event(struct perf_event
*event
, struct perf_event_context
*ctx
)
1280 struct perf_cpu_context
*cpuctx
;
1282 * We can have double detach due to exit/hot-unplug + close.
1284 if (!(event
->attach_state
& PERF_ATTACH_CONTEXT
))
1287 event
->attach_state
&= ~PERF_ATTACH_CONTEXT
;
1289 if (is_cgroup_event(event
)) {
1291 cpuctx
= __get_cpu_context(ctx
);
1293 * if there are no more cgroup events
1294 * then cler cgrp to avoid stale pointer
1295 * in update_cgrp_time_from_cpuctx()
1297 if (!ctx
->nr_cgroups
)
1298 cpuctx
->cgrp
= NULL
;
1301 if (has_branch_stack(event
))
1302 ctx
->nr_branch_stack
--;
1305 if (event
->attr
.inherit_stat
)
1308 list_del_rcu(&event
->event_entry
);
1310 if (event
->group_leader
== event
)
1311 list_del_init(&event
->group_entry
);
1313 update_group_times(event
);
1316 * If event was in error state, then keep it
1317 * that way, otherwise bogus counts will be
1318 * returned on read(). The only way to get out
1319 * of error state is by explicit re-enabling
1322 if (event
->state
> PERF_EVENT_STATE_OFF
)
1323 event
->state
= PERF_EVENT_STATE_OFF
;
1328 static void perf_group_detach(struct perf_event
*event
)
1330 struct perf_event
*sibling
, *tmp
;
1331 struct list_head
*list
= NULL
;
1334 * We can have double detach due to exit/hot-unplug + close.
1336 if (!(event
->attach_state
& PERF_ATTACH_GROUP
))
1339 event
->attach_state
&= ~PERF_ATTACH_GROUP
;
1342 * If this is a sibling, remove it from its group.
1344 if (event
->group_leader
!= event
) {
1345 list_del_init(&event
->group_entry
);
1346 event
->group_leader
->nr_siblings
--;
1350 if (!list_empty(&event
->group_entry
))
1351 list
= &event
->group_entry
;
1354 * If this was a group event with sibling events then
1355 * upgrade the siblings to singleton events by adding them
1356 * to whatever list we are on.
1358 list_for_each_entry_safe(sibling
, tmp
, &event
->sibling_list
, group_entry
) {
1360 list_move_tail(&sibling
->group_entry
, list
);
1361 sibling
->group_leader
= sibling
;
1363 /* Inherit group flags from the previous leader */
1364 sibling
->group_flags
= event
->group_flags
;
1368 perf_event__header_size(event
->group_leader
);
1370 list_for_each_entry(tmp
, &event
->group_leader
->sibling_list
, group_entry
)
1371 perf_event__header_size(tmp
);
1375 event_filter_match(struct perf_event
*event
)
1377 return (event
->cpu
== -1 || event
->cpu
== smp_processor_id())
1378 && perf_cgroup_match(event
);
1382 event_sched_out(struct perf_event
*event
,
1383 struct perf_cpu_context
*cpuctx
,
1384 struct perf_event_context
*ctx
)
1386 u64 tstamp
= perf_event_time(event
);
1389 * An event which could not be activated because of
1390 * filter mismatch still needs to have its timings
1391 * maintained, otherwise bogus information is return
1392 * via read() for time_enabled, time_running:
1394 if (event
->state
== PERF_EVENT_STATE_INACTIVE
1395 && !event_filter_match(event
)) {
1396 delta
= tstamp
- event
->tstamp_stopped
;
1397 event
->tstamp_running
+= delta
;
1398 event
->tstamp_stopped
= tstamp
;
1401 if (event
->state
!= PERF_EVENT_STATE_ACTIVE
)
1404 perf_pmu_disable(event
->pmu
);
1406 event
->state
= PERF_EVENT_STATE_INACTIVE
;
1407 if (event
->pending_disable
) {
1408 event
->pending_disable
= 0;
1409 event
->state
= PERF_EVENT_STATE_OFF
;
1411 event
->tstamp_stopped
= tstamp
;
1412 event
->pmu
->del(event
, 0);
1415 if (!is_software_event(event
))
1416 cpuctx
->active_oncpu
--;
1418 if (event
->attr
.freq
&& event
->attr
.sample_freq
)
1420 if (event
->attr
.exclusive
|| !cpuctx
->active_oncpu
)
1421 cpuctx
->exclusive
= 0;
1423 perf_pmu_enable(event
->pmu
);
1427 group_sched_out(struct perf_event
*group_event
,
1428 struct perf_cpu_context
*cpuctx
,
1429 struct perf_event_context
*ctx
)
1431 struct perf_event
*event
;
1432 int state
= group_event
->state
;
1434 event_sched_out(group_event
, cpuctx
, ctx
);
1437 * Schedule out siblings (if any):
1439 list_for_each_entry(event
, &group_event
->sibling_list
, group_entry
)
1440 event_sched_out(event
, cpuctx
, ctx
);
1442 if (state
== PERF_EVENT_STATE_ACTIVE
&& group_event
->attr
.exclusive
)
1443 cpuctx
->exclusive
= 0;
1447 * Cross CPU call to remove a performance event
1449 * We disable the event on the hardware level first. After that we
1450 * remove it from the context list.
1452 static int __perf_remove_from_context(void *info
)
1454 struct perf_event
*event
= info
;
1455 struct perf_event_context
*ctx
= event
->ctx
;
1456 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
1458 raw_spin_lock(&ctx
->lock
);
1459 event_sched_out(event
, cpuctx
, ctx
);
1460 list_del_event(event
, ctx
);
1461 if (!ctx
->nr_events
&& cpuctx
->task_ctx
== ctx
) {
1463 cpuctx
->task_ctx
= NULL
;
1465 raw_spin_unlock(&ctx
->lock
);
1472 * Remove the event from a task's (or a CPU's) list of events.
1474 * CPU events are removed with a smp call. For task events we only
1475 * call when the task is on a CPU.
1477 * If event->ctx is a cloned context, callers must make sure that
1478 * every task struct that event->ctx->task could possibly point to
1479 * remains valid. This is OK when called from perf_release since
1480 * that only calls us on the top-level context, which can't be a clone.
1481 * When called from perf_event_exit_task, it's OK because the
1482 * context has been detached from its task.
1484 static void perf_remove_from_context(struct perf_event
*event
)
1486 struct perf_event_context
*ctx
= event
->ctx
;
1487 struct task_struct
*task
= ctx
->task
;
1489 lockdep_assert_held(&ctx
->mutex
);
1493 * Per cpu events are removed via an smp call and
1494 * the removal is always successful.
1496 cpu_function_call(event
->cpu
, __perf_remove_from_context
, event
);
1501 if (!task_function_call(task
, __perf_remove_from_context
, event
))
1504 raw_spin_lock_irq(&ctx
->lock
);
1506 * If we failed to find a running task, but find the context active now
1507 * that we've acquired the ctx->lock, retry.
1509 if (ctx
->is_active
) {
1510 raw_spin_unlock_irq(&ctx
->lock
);
1515 * Since the task isn't running, its safe to remove the event, us
1516 * holding the ctx->lock ensures the task won't get scheduled in.
1518 list_del_event(event
, ctx
);
1519 raw_spin_unlock_irq(&ctx
->lock
);
1523 * Cross CPU call to disable a performance event
1525 int __perf_event_disable(void *info
)
1527 struct perf_event
*event
= info
;
1528 struct perf_event_context
*ctx
= event
->ctx
;
1529 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
1532 * If this is a per-task event, need to check whether this
1533 * event's task is the current task on this cpu.
1535 * Can trigger due to concurrent perf_event_context_sched_out()
1536 * flipping contexts around.
1538 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
)
1541 raw_spin_lock(&ctx
->lock
);
1544 * If the event is on, turn it off.
1545 * If it is in error state, leave it in error state.
1547 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
) {
1548 update_context_time(ctx
);
1549 update_cgrp_time_from_event(event
);
1550 update_group_times(event
);
1551 if (event
== event
->group_leader
)
1552 group_sched_out(event
, cpuctx
, ctx
);
1554 event_sched_out(event
, cpuctx
, ctx
);
1555 event
->state
= PERF_EVENT_STATE_OFF
;
1558 raw_spin_unlock(&ctx
->lock
);
1566 * If event->ctx is a cloned context, callers must make sure that
1567 * every task struct that event->ctx->task could possibly point to
1568 * remains valid. This condition is satisifed when called through
1569 * perf_event_for_each_child or perf_event_for_each because they
1570 * hold the top-level event's child_mutex, so any descendant that
1571 * goes to exit will block in sync_child_event.
1572 * When called from perf_pending_event it's OK because event->ctx
1573 * is the current context on this CPU and preemption is disabled,
1574 * hence we can't get into perf_event_task_sched_out for this context.
1576 void perf_event_disable(struct perf_event
*event
)
1578 struct perf_event_context
*ctx
= event
->ctx
;
1579 struct task_struct
*task
= ctx
->task
;
1583 * Disable the event on the cpu that it's on
1585 cpu_function_call(event
->cpu
, __perf_event_disable
, event
);
1590 if (!task_function_call(task
, __perf_event_disable
, event
))
1593 raw_spin_lock_irq(&ctx
->lock
);
1595 * If the event is still active, we need to retry the cross-call.
1597 if (event
->state
== PERF_EVENT_STATE_ACTIVE
) {
1598 raw_spin_unlock_irq(&ctx
->lock
);
1600 * Reload the task pointer, it might have been changed by
1601 * a concurrent perf_event_context_sched_out().
1608 * Since we have the lock this context can't be scheduled
1609 * in, so we can change the state safely.
1611 if (event
->state
== PERF_EVENT_STATE_INACTIVE
) {
1612 update_group_times(event
);
1613 event
->state
= PERF_EVENT_STATE_OFF
;
1615 raw_spin_unlock_irq(&ctx
->lock
);
1617 EXPORT_SYMBOL_GPL(perf_event_disable
);
1619 static void perf_set_shadow_time(struct perf_event
*event
,
1620 struct perf_event_context
*ctx
,
1624 * use the correct time source for the time snapshot
1626 * We could get by without this by leveraging the
1627 * fact that to get to this function, the caller
1628 * has most likely already called update_context_time()
1629 * and update_cgrp_time_xx() and thus both timestamp
1630 * are identical (or very close). Given that tstamp is,
1631 * already adjusted for cgroup, we could say that:
1632 * tstamp - ctx->timestamp
1634 * tstamp - cgrp->timestamp.
1636 * Then, in perf_output_read(), the calculation would
1637 * work with no changes because:
1638 * - event is guaranteed scheduled in
1639 * - no scheduled out in between
1640 * - thus the timestamp would be the same
1642 * But this is a bit hairy.
1644 * So instead, we have an explicit cgroup call to remain
1645 * within the time time source all along. We believe it
1646 * is cleaner and simpler to understand.
1648 if (is_cgroup_event(event
))
1649 perf_cgroup_set_shadow_time(event
, tstamp
);
1651 event
->shadow_ctx_time
= tstamp
- ctx
->timestamp
;
1654 #define MAX_INTERRUPTS (~0ULL)
1656 static void perf_log_throttle(struct perf_event
*event
, int enable
);
1659 event_sched_in(struct perf_event
*event
,
1660 struct perf_cpu_context
*cpuctx
,
1661 struct perf_event_context
*ctx
)
1663 u64 tstamp
= perf_event_time(event
);
1666 if (event
->state
<= PERF_EVENT_STATE_OFF
)
1669 event
->state
= PERF_EVENT_STATE_ACTIVE
;
1670 event
->oncpu
= smp_processor_id();
1673 * Unthrottle events, since we scheduled we might have missed several
1674 * ticks already, also for a heavily scheduling task there is little
1675 * guarantee it'll get a tick in a timely manner.
1677 if (unlikely(event
->hw
.interrupts
== MAX_INTERRUPTS
)) {
1678 perf_log_throttle(event
, 1);
1679 event
->hw
.interrupts
= 0;
1683 * The new state must be visible before we turn it on in the hardware:
1687 perf_pmu_disable(event
->pmu
);
1689 if (event
->pmu
->add(event
, PERF_EF_START
)) {
1690 event
->state
= PERF_EVENT_STATE_INACTIVE
;
1696 event
->tstamp_running
+= tstamp
- event
->tstamp_stopped
;
1698 perf_set_shadow_time(event
, ctx
, tstamp
);
1700 if (!is_software_event(event
))
1701 cpuctx
->active_oncpu
++;
1703 if (event
->attr
.freq
&& event
->attr
.sample_freq
)
1706 if (event
->attr
.exclusive
)
1707 cpuctx
->exclusive
= 1;
1710 perf_pmu_enable(event
->pmu
);
1716 group_sched_in(struct perf_event
*group_event
,
1717 struct perf_cpu_context
*cpuctx
,
1718 struct perf_event_context
*ctx
)
1720 struct perf_event
*event
, *partial_group
= NULL
;
1721 struct pmu
*pmu
= ctx
->pmu
;
1722 u64 now
= ctx
->time
;
1723 bool simulate
= false;
1725 if (group_event
->state
== PERF_EVENT_STATE_OFF
)
1728 pmu
->start_txn(pmu
);
1730 if (event_sched_in(group_event
, cpuctx
, ctx
)) {
1731 pmu
->cancel_txn(pmu
);
1732 perf_cpu_hrtimer_restart(cpuctx
);
1737 * Schedule in siblings as one group (if any):
1739 list_for_each_entry(event
, &group_event
->sibling_list
, group_entry
) {
1740 if (event_sched_in(event
, cpuctx
, ctx
)) {
1741 partial_group
= event
;
1746 if (!pmu
->commit_txn(pmu
))
1751 * Groups can be scheduled in as one unit only, so undo any
1752 * partial group before returning:
1753 * The events up to the failed event are scheduled out normally,
1754 * tstamp_stopped will be updated.
1756 * The failed events and the remaining siblings need to have
1757 * their timings updated as if they had gone thru event_sched_in()
1758 * and event_sched_out(). This is required to get consistent timings
1759 * across the group. This also takes care of the case where the group
1760 * could never be scheduled by ensuring tstamp_stopped is set to mark
1761 * the time the event was actually stopped, such that time delta
1762 * calculation in update_event_times() is correct.
1764 list_for_each_entry(event
, &group_event
->sibling_list
, group_entry
) {
1765 if (event
== partial_group
)
1769 event
->tstamp_running
+= now
- event
->tstamp_stopped
;
1770 event
->tstamp_stopped
= now
;
1772 event_sched_out(event
, cpuctx
, ctx
);
1775 event_sched_out(group_event
, cpuctx
, ctx
);
1777 pmu
->cancel_txn(pmu
);
1779 perf_cpu_hrtimer_restart(cpuctx
);
1785 * Work out whether we can put this event group on the CPU now.
1787 static int group_can_go_on(struct perf_event
*event
,
1788 struct perf_cpu_context
*cpuctx
,
1792 * Groups consisting entirely of software events can always go on.
1794 if (event
->group_flags
& PERF_GROUP_SOFTWARE
)
1797 * If an exclusive group is already on, no other hardware
1800 if (cpuctx
->exclusive
)
1803 * If this group is exclusive and there are already
1804 * events on the CPU, it can't go on.
1806 if (event
->attr
.exclusive
&& cpuctx
->active_oncpu
)
1809 * Otherwise, try to add it if all previous groups were able
1815 static void add_event_to_ctx(struct perf_event
*event
,
1816 struct perf_event_context
*ctx
)
1818 u64 tstamp
= perf_event_time(event
);
1820 list_add_event(event
, ctx
);
1821 perf_group_attach(event
);
1822 event
->tstamp_enabled
= tstamp
;
1823 event
->tstamp_running
= tstamp
;
1824 event
->tstamp_stopped
= tstamp
;
1827 static void task_ctx_sched_out(struct perf_event_context
*ctx
);
1829 ctx_sched_in(struct perf_event_context
*ctx
,
1830 struct perf_cpu_context
*cpuctx
,
1831 enum event_type_t event_type
,
1832 struct task_struct
*task
);
1834 static void perf_event_sched_in(struct perf_cpu_context
*cpuctx
,
1835 struct perf_event_context
*ctx
,
1836 struct task_struct
*task
)
1838 cpu_ctx_sched_in(cpuctx
, EVENT_PINNED
, task
);
1840 ctx_sched_in(ctx
, cpuctx
, EVENT_PINNED
, task
);
1841 cpu_ctx_sched_in(cpuctx
, EVENT_FLEXIBLE
, task
);
1843 ctx_sched_in(ctx
, cpuctx
, EVENT_FLEXIBLE
, task
);
1847 * Cross CPU call to install and enable a performance event
1849 * Must be called with ctx->mutex held
1851 static int __perf_install_in_context(void *info
)
1853 struct perf_event
*event
= info
;
1854 struct perf_event_context
*ctx
= event
->ctx
;
1855 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
1856 struct perf_event_context
*task_ctx
= cpuctx
->task_ctx
;
1857 struct task_struct
*task
= current
;
1859 perf_ctx_lock(cpuctx
, task_ctx
);
1860 perf_pmu_disable(cpuctx
->ctx
.pmu
);
1863 * If there was an active task_ctx schedule it out.
1866 task_ctx_sched_out(task_ctx
);
1869 * If the context we're installing events in is not the
1870 * active task_ctx, flip them.
1872 if (ctx
->task
&& task_ctx
!= ctx
) {
1874 raw_spin_unlock(&task_ctx
->lock
);
1875 raw_spin_lock(&ctx
->lock
);
1880 cpuctx
->task_ctx
= task_ctx
;
1881 task
= task_ctx
->task
;
1884 cpu_ctx_sched_out(cpuctx
, EVENT_ALL
);
1886 update_context_time(ctx
);
1888 * update cgrp time only if current cgrp
1889 * matches event->cgrp. Must be done before
1890 * calling add_event_to_ctx()
1892 update_cgrp_time_from_event(event
);
1894 add_event_to_ctx(event
, ctx
);
1897 * Schedule everything back in
1899 perf_event_sched_in(cpuctx
, task_ctx
, task
);
1901 perf_pmu_enable(cpuctx
->ctx
.pmu
);
1902 perf_ctx_unlock(cpuctx
, task_ctx
);
1908 * Attach a performance event to a context
1910 * First we add the event to the list with the hardware enable bit
1911 * in event->hw_config cleared.
1913 * If the event is attached to a task which is on a CPU we use a smp
1914 * call to enable it in the task context. The task might have been
1915 * scheduled away, but we check this in the smp call again.
1918 perf_install_in_context(struct perf_event_context
*ctx
,
1919 struct perf_event
*event
,
1922 struct task_struct
*task
= ctx
->task
;
1924 lockdep_assert_held(&ctx
->mutex
);
1927 if (event
->cpu
!= -1)
1932 * Per cpu events are installed via an smp call and
1933 * the install is always successful.
1935 cpu_function_call(cpu
, __perf_install_in_context
, event
);
1940 if (!task_function_call(task
, __perf_install_in_context
, event
))
1943 raw_spin_lock_irq(&ctx
->lock
);
1945 * If we failed to find a running task, but find the context active now
1946 * that we've acquired the ctx->lock, retry.
1948 if (ctx
->is_active
) {
1949 raw_spin_unlock_irq(&ctx
->lock
);
1954 * Since the task isn't running, its safe to add the event, us holding
1955 * the ctx->lock ensures the task won't get scheduled in.
1957 add_event_to_ctx(event
, ctx
);
1958 raw_spin_unlock_irq(&ctx
->lock
);
1962 * Put a event into inactive state and update time fields.
1963 * Enabling the leader of a group effectively enables all
1964 * the group members that aren't explicitly disabled, so we
1965 * have to update their ->tstamp_enabled also.
1966 * Note: this works for group members as well as group leaders
1967 * since the non-leader members' sibling_lists will be empty.
1969 static void __perf_event_mark_enabled(struct perf_event
*event
)
1971 struct perf_event
*sub
;
1972 u64 tstamp
= perf_event_time(event
);
1974 event
->state
= PERF_EVENT_STATE_INACTIVE
;
1975 event
->tstamp_enabled
= tstamp
- event
->total_time_enabled
;
1976 list_for_each_entry(sub
, &event
->sibling_list
, group_entry
) {
1977 if (sub
->state
>= PERF_EVENT_STATE_INACTIVE
)
1978 sub
->tstamp_enabled
= tstamp
- sub
->total_time_enabled
;
1983 * Cross CPU call to enable a performance event
1985 static int __perf_event_enable(void *info
)
1987 struct perf_event
*event
= info
;
1988 struct perf_event_context
*ctx
= event
->ctx
;
1989 struct perf_event
*leader
= event
->group_leader
;
1990 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
1994 * There's a time window between 'ctx->is_active' check
1995 * in perf_event_enable function and this place having:
1997 * - ctx->lock unlocked
1999 * where the task could be killed and 'ctx' deactivated
2000 * by perf_event_exit_task.
2002 if (!ctx
->is_active
)
2005 raw_spin_lock(&ctx
->lock
);
2006 update_context_time(ctx
);
2008 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
)
2012 * set current task's cgroup time reference point
2014 perf_cgroup_set_timestamp(current
, ctx
);
2016 __perf_event_mark_enabled(event
);
2018 if (!event_filter_match(event
)) {
2019 if (is_cgroup_event(event
))
2020 perf_cgroup_defer_enabled(event
);
2025 * If the event is in a group and isn't the group leader,
2026 * then don't put it on unless the group is on.
2028 if (leader
!= event
&& leader
->state
!= PERF_EVENT_STATE_ACTIVE
)
2031 if (!group_can_go_on(event
, cpuctx
, 1)) {
2034 if (event
== leader
)
2035 err
= group_sched_in(event
, cpuctx
, ctx
);
2037 err
= event_sched_in(event
, cpuctx
, ctx
);
2042 * If this event can't go on and it's part of a
2043 * group, then the whole group has to come off.
2045 if (leader
!= event
) {
2046 group_sched_out(leader
, cpuctx
, ctx
);
2047 perf_cpu_hrtimer_restart(cpuctx
);
2049 if (leader
->attr
.pinned
) {
2050 update_group_times(leader
);
2051 leader
->state
= PERF_EVENT_STATE_ERROR
;
2056 raw_spin_unlock(&ctx
->lock
);
2064 * If event->ctx is a cloned context, callers must make sure that
2065 * every task struct that event->ctx->task could possibly point to
2066 * remains valid. This condition is satisfied when called through
2067 * perf_event_for_each_child or perf_event_for_each as described
2068 * for perf_event_disable.
2070 void perf_event_enable(struct perf_event
*event
)
2072 struct perf_event_context
*ctx
= event
->ctx
;
2073 struct task_struct
*task
= ctx
->task
;
2077 * Enable the event on the cpu that it's on
2079 cpu_function_call(event
->cpu
, __perf_event_enable
, event
);
2083 raw_spin_lock_irq(&ctx
->lock
);
2084 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
)
2088 * If the event is in error state, clear that first.
2089 * That way, if we see the event in error state below, we
2090 * know that it has gone back into error state, as distinct
2091 * from the task having been scheduled away before the
2092 * cross-call arrived.
2094 if (event
->state
== PERF_EVENT_STATE_ERROR
)
2095 event
->state
= PERF_EVENT_STATE_OFF
;
2098 if (!ctx
->is_active
) {
2099 __perf_event_mark_enabled(event
);
2103 raw_spin_unlock_irq(&ctx
->lock
);
2105 if (!task_function_call(task
, __perf_event_enable
, event
))
2108 raw_spin_lock_irq(&ctx
->lock
);
2111 * If the context is active and the event is still off,
2112 * we need to retry the cross-call.
2114 if (ctx
->is_active
&& event
->state
== PERF_EVENT_STATE_OFF
) {
2116 * task could have been flipped by a concurrent
2117 * perf_event_context_sched_out()
2124 raw_spin_unlock_irq(&ctx
->lock
);
2126 EXPORT_SYMBOL_GPL(perf_event_enable
);
2128 int perf_event_refresh(struct perf_event
*event
, int refresh
)
2131 * not supported on inherited events
2133 if (event
->attr
.inherit
|| !is_sampling_event(event
))
2136 atomic_add(refresh
, &event
->event_limit
);
2137 perf_event_enable(event
);
2141 EXPORT_SYMBOL_GPL(perf_event_refresh
);
2143 static void ctx_sched_out(struct perf_event_context
*ctx
,
2144 struct perf_cpu_context
*cpuctx
,
2145 enum event_type_t event_type
)
2147 struct perf_event
*event
;
2148 int is_active
= ctx
->is_active
;
2150 ctx
->is_active
&= ~event_type
;
2151 if (likely(!ctx
->nr_events
))
2154 update_context_time(ctx
);
2155 update_cgrp_time_from_cpuctx(cpuctx
);
2156 if (!ctx
->nr_active
)
2159 perf_pmu_disable(ctx
->pmu
);
2160 if ((is_active
& EVENT_PINNED
) && (event_type
& EVENT_PINNED
)) {
2161 list_for_each_entry(event
, &ctx
->pinned_groups
, group_entry
)
2162 group_sched_out(event
, cpuctx
, ctx
);
2165 if ((is_active
& EVENT_FLEXIBLE
) && (event_type
& EVENT_FLEXIBLE
)) {
2166 list_for_each_entry(event
, &ctx
->flexible_groups
, group_entry
)
2167 group_sched_out(event
, cpuctx
, ctx
);
2169 perf_pmu_enable(ctx
->pmu
);
2173 * Test whether two contexts are equivalent, i.e. whether they have both been
2174 * cloned from the same version of the same context.
2176 * Equivalence is measured using a generation number in the context that is
2177 * incremented on each modification to it; see unclone_ctx(), list_add_event()
2178 * and list_del_event().
2180 static int context_equiv(struct perf_event_context
*ctx1
,
2181 struct perf_event_context
*ctx2
)
2183 /* Pinning disables the swap optimization */
2184 if (ctx1
->pin_count
|| ctx2
->pin_count
)
2187 /* If ctx1 is the parent of ctx2 */
2188 if (ctx1
== ctx2
->parent_ctx
&& ctx1
->generation
== ctx2
->parent_gen
)
2191 /* If ctx2 is the parent of ctx1 */
2192 if (ctx1
->parent_ctx
== ctx2
&& ctx1
->parent_gen
== ctx2
->generation
)
2196 * If ctx1 and ctx2 have the same parent; we flatten the parent
2197 * hierarchy, see perf_event_init_context().
2199 if (ctx1
->parent_ctx
&& ctx1
->parent_ctx
== ctx2
->parent_ctx
&&
2200 ctx1
->parent_gen
== ctx2
->parent_gen
)
2207 static void __perf_event_sync_stat(struct perf_event
*event
,
2208 struct perf_event
*next_event
)
2212 if (!event
->attr
.inherit_stat
)
2216 * Update the event value, we cannot use perf_event_read()
2217 * because we're in the middle of a context switch and have IRQs
2218 * disabled, which upsets smp_call_function_single(), however
2219 * we know the event must be on the current CPU, therefore we
2220 * don't need to use it.
2222 switch (event
->state
) {
2223 case PERF_EVENT_STATE_ACTIVE
:
2224 event
->pmu
->read(event
);
2227 case PERF_EVENT_STATE_INACTIVE
:
2228 update_event_times(event
);
2236 * In order to keep per-task stats reliable we need to flip the event
2237 * values when we flip the contexts.
2239 value
= local64_read(&next_event
->count
);
2240 value
= local64_xchg(&event
->count
, value
);
2241 local64_set(&next_event
->count
, value
);
2243 swap(event
->total_time_enabled
, next_event
->total_time_enabled
);
2244 swap(event
->total_time_running
, next_event
->total_time_running
);
2247 * Since we swizzled the values, update the user visible data too.
2249 perf_event_update_userpage(event
);
2250 perf_event_update_userpage(next_event
);
2253 static void perf_event_sync_stat(struct perf_event_context
*ctx
,
2254 struct perf_event_context
*next_ctx
)
2256 struct perf_event
*event
, *next_event
;
2261 update_context_time(ctx
);
2263 event
= list_first_entry(&ctx
->event_list
,
2264 struct perf_event
, event_entry
);
2266 next_event
= list_first_entry(&next_ctx
->event_list
,
2267 struct perf_event
, event_entry
);
2269 while (&event
->event_entry
!= &ctx
->event_list
&&
2270 &next_event
->event_entry
!= &next_ctx
->event_list
) {
2272 __perf_event_sync_stat(event
, next_event
);
2274 event
= list_next_entry(event
, event_entry
);
2275 next_event
= list_next_entry(next_event
, event_entry
);
2279 static void perf_event_context_sched_out(struct task_struct
*task
, int ctxn
,
2280 struct task_struct
*next
)
2282 struct perf_event_context
*ctx
= task
->perf_event_ctxp
[ctxn
];
2283 struct perf_event_context
*next_ctx
;
2284 struct perf_event_context
*parent
, *next_parent
;
2285 struct perf_cpu_context
*cpuctx
;
2291 cpuctx
= __get_cpu_context(ctx
);
2292 if (!cpuctx
->task_ctx
)
2296 next_ctx
= next
->perf_event_ctxp
[ctxn
];
2300 parent
= rcu_dereference(ctx
->parent_ctx
);
2301 next_parent
= rcu_dereference(next_ctx
->parent_ctx
);
2303 /* If neither context have a parent context; they cannot be clones. */
2304 if (!parent
&& !next_parent
)
2307 if (next_parent
== ctx
|| next_ctx
== parent
|| next_parent
== parent
) {
2309 * Looks like the two contexts are clones, so we might be
2310 * able to optimize the context switch. We lock both
2311 * contexts and check that they are clones under the
2312 * lock (including re-checking that neither has been
2313 * uncloned in the meantime). It doesn't matter which
2314 * order we take the locks because no other cpu could
2315 * be trying to lock both of these tasks.
2317 raw_spin_lock(&ctx
->lock
);
2318 raw_spin_lock_nested(&next_ctx
->lock
, SINGLE_DEPTH_NESTING
);
2319 if (context_equiv(ctx
, next_ctx
)) {
2321 * XXX do we need a memory barrier of sorts
2322 * wrt to rcu_dereference() of perf_event_ctxp
2324 task
->perf_event_ctxp
[ctxn
] = next_ctx
;
2325 next
->perf_event_ctxp
[ctxn
] = ctx
;
2327 next_ctx
->task
= task
;
2330 perf_event_sync_stat(ctx
, next_ctx
);
2332 raw_spin_unlock(&next_ctx
->lock
);
2333 raw_spin_unlock(&ctx
->lock
);
2339 raw_spin_lock(&ctx
->lock
);
2340 ctx_sched_out(ctx
, cpuctx
, EVENT_ALL
);
2341 cpuctx
->task_ctx
= NULL
;
2342 raw_spin_unlock(&ctx
->lock
);
2346 #define for_each_task_context_nr(ctxn) \
2347 for ((ctxn) = 0; (ctxn) < perf_nr_task_contexts; (ctxn)++)
2350 * Called from scheduler to remove the events of the current task,
2351 * with interrupts disabled.
2353 * We stop each event and update the event value in event->count.
2355 * This does not protect us against NMI, but disable()
2356 * sets the disabled bit in the control field of event _before_
2357 * accessing the event control register. If a NMI hits, then it will
2358 * not restart the event.
2360 void __perf_event_task_sched_out(struct task_struct
*task
,
2361 struct task_struct
*next
)
2365 for_each_task_context_nr(ctxn
)
2366 perf_event_context_sched_out(task
, ctxn
, next
);
2369 * if cgroup events exist on this CPU, then we need
2370 * to check if we have to switch out PMU state.
2371 * cgroup event are system-wide mode only
2373 if (atomic_read(&__get_cpu_var(perf_cgroup_events
)))
2374 perf_cgroup_sched_out(task
, next
);
2377 static void task_ctx_sched_out(struct perf_event_context
*ctx
)
2379 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
2381 if (!cpuctx
->task_ctx
)
2384 if (WARN_ON_ONCE(ctx
!= cpuctx
->task_ctx
))
2387 ctx_sched_out(ctx
, cpuctx
, EVENT_ALL
);
2388 cpuctx
->task_ctx
= NULL
;
2392 * Called with IRQs disabled
2394 static void cpu_ctx_sched_out(struct perf_cpu_context
*cpuctx
,
2395 enum event_type_t event_type
)
2397 ctx_sched_out(&cpuctx
->ctx
, cpuctx
, event_type
);
2401 ctx_pinned_sched_in(struct perf_event_context
*ctx
,
2402 struct perf_cpu_context
*cpuctx
)
2404 struct perf_event
*event
;
2406 list_for_each_entry(event
, &ctx
->pinned_groups
, group_entry
) {
2407 if (event
->state
<= PERF_EVENT_STATE_OFF
)
2409 if (!event_filter_match(event
))
2412 /* may need to reset tstamp_enabled */
2413 if (is_cgroup_event(event
))
2414 perf_cgroup_mark_enabled(event
, ctx
);
2416 if (group_can_go_on(event
, cpuctx
, 1))
2417 group_sched_in(event
, cpuctx
, ctx
);
2420 * If this pinned group hasn't been scheduled,
2421 * put it in error state.
2423 if (event
->state
== PERF_EVENT_STATE_INACTIVE
) {
2424 update_group_times(event
);
2425 event
->state
= PERF_EVENT_STATE_ERROR
;
2431 ctx_flexible_sched_in(struct perf_event_context
*ctx
,
2432 struct perf_cpu_context
*cpuctx
)
2434 struct perf_event
*event
;
2437 list_for_each_entry(event
, &ctx
->flexible_groups
, group_entry
) {
2438 /* Ignore events in OFF or ERROR state */
2439 if (event
->state
<= PERF_EVENT_STATE_OFF
)
2442 * Listen to the 'cpu' scheduling filter constraint
2445 if (!event_filter_match(event
))
2448 /* may need to reset tstamp_enabled */
2449 if (is_cgroup_event(event
))
2450 perf_cgroup_mark_enabled(event
, ctx
);
2452 if (group_can_go_on(event
, cpuctx
, can_add_hw
)) {
2453 if (group_sched_in(event
, cpuctx
, ctx
))
2460 ctx_sched_in(struct perf_event_context
*ctx
,
2461 struct perf_cpu_context
*cpuctx
,
2462 enum event_type_t event_type
,
2463 struct task_struct
*task
)
2466 int is_active
= ctx
->is_active
;
2468 ctx
->is_active
|= event_type
;
2469 if (likely(!ctx
->nr_events
))
2473 ctx
->timestamp
= now
;
2474 perf_cgroup_set_timestamp(task
, ctx
);
2476 * First go through the list and put on any pinned groups
2477 * in order to give them the best chance of going on.
2479 if (!(is_active
& EVENT_PINNED
) && (event_type
& EVENT_PINNED
))
2480 ctx_pinned_sched_in(ctx
, cpuctx
);
2482 /* Then walk through the lower prio flexible groups */
2483 if (!(is_active
& EVENT_FLEXIBLE
) && (event_type
& EVENT_FLEXIBLE
))
2484 ctx_flexible_sched_in(ctx
, cpuctx
);
2487 static void cpu_ctx_sched_in(struct perf_cpu_context
*cpuctx
,
2488 enum event_type_t event_type
,
2489 struct task_struct
*task
)
2491 struct perf_event_context
*ctx
= &cpuctx
->ctx
;
2493 ctx_sched_in(ctx
, cpuctx
, event_type
, task
);
2496 static void perf_event_context_sched_in(struct perf_event_context
*ctx
,
2497 struct task_struct
*task
)
2499 struct perf_cpu_context
*cpuctx
;
2501 cpuctx
= __get_cpu_context(ctx
);
2502 if (cpuctx
->task_ctx
== ctx
)
2505 perf_ctx_lock(cpuctx
, ctx
);
2506 perf_pmu_disable(ctx
->pmu
);
2508 * We want to keep the following priority order:
2509 * cpu pinned (that don't need to move), task pinned,
2510 * cpu flexible, task flexible.
2512 cpu_ctx_sched_out(cpuctx
, EVENT_FLEXIBLE
);
2515 cpuctx
->task_ctx
= ctx
;
2517 perf_event_sched_in(cpuctx
, cpuctx
->task_ctx
, task
);
2519 perf_pmu_enable(ctx
->pmu
);
2520 perf_ctx_unlock(cpuctx
, ctx
);
2523 * Since these rotations are per-cpu, we need to ensure the
2524 * cpu-context we got scheduled on is actually rotating.
2526 perf_pmu_rotate_start(ctx
->pmu
);
2530 * When sampling the branck stack in system-wide, it may be necessary
2531 * to flush the stack on context switch. This happens when the branch
2532 * stack does not tag its entries with the pid of the current task.
2533 * Otherwise it becomes impossible to associate a branch entry with a
2534 * task. This ambiguity is more likely to appear when the branch stack
2535 * supports priv level filtering and the user sets it to monitor only
2536 * at the user level (which could be a useful measurement in system-wide
2537 * mode). In that case, the risk is high of having a branch stack with
2538 * branch from multiple tasks. Flushing may mean dropping the existing
2539 * entries or stashing them somewhere in the PMU specific code layer.
2541 * This function provides the context switch callback to the lower code
2542 * layer. It is invoked ONLY when there is at least one system-wide context
2543 * with at least one active event using taken branch sampling.
2545 static void perf_branch_stack_sched_in(struct task_struct
*prev
,
2546 struct task_struct
*task
)
2548 struct perf_cpu_context
*cpuctx
;
2550 unsigned long flags
;
2552 /* no need to flush branch stack if not changing task */
2556 local_irq_save(flags
);
2560 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
2561 cpuctx
= this_cpu_ptr(pmu
->pmu_cpu_context
);
2564 * check if the context has at least one
2565 * event using PERF_SAMPLE_BRANCH_STACK
2567 if (cpuctx
->ctx
.nr_branch_stack
> 0
2568 && pmu
->flush_branch_stack
) {
2570 perf_ctx_lock(cpuctx
, cpuctx
->task_ctx
);
2572 perf_pmu_disable(pmu
);
2574 pmu
->flush_branch_stack();
2576 perf_pmu_enable(pmu
);
2578 perf_ctx_unlock(cpuctx
, cpuctx
->task_ctx
);
2584 local_irq_restore(flags
);
2588 * Called from scheduler to add the events of the current task
2589 * with interrupts disabled.
2591 * We restore the event value and then enable it.
2593 * This does not protect us against NMI, but enable()
2594 * sets the enabled bit in the control field of event _before_
2595 * accessing the event control register. If a NMI hits, then it will
2596 * keep the event running.
2598 void __perf_event_task_sched_in(struct task_struct
*prev
,
2599 struct task_struct
*task
)
2601 struct perf_event_context
*ctx
;
2604 for_each_task_context_nr(ctxn
) {
2605 ctx
= task
->perf_event_ctxp
[ctxn
];
2609 perf_event_context_sched_in(ctx
, task
);
2612 * if cgroup events exist on this CPU, then we need
2613 * to check if we have to switch in PMU state.
2614 * cgroup event are system-wide mode only
2616 if (atomic_read(&__get_cpu_var(perf_cgroup_events
)))
2617 perf_cgroup_sched_in(prev
, task
);
2619 /* check for system-wide branch_stack events */
2620 if (atomic_read(&__get_cpu_var(perf_branch_stack_events
)))
2621 perf_branch_stack_sched_in(prev
, task
);
2624 static u64
perf_calculate_period(struct perf_event
*event
, u64 nsec
, u64 count
)
2626 u64 frequency
= event
->attr
.sample_freq
;
2627 u64 sec
= NSEC_PER_SEC
;
2628 u64 divisor
, dividend
;
2630 int count_fls
, nsec_fls
, frequency_fls
, sec_fls
;
2632 count_fls
= fls64(count
);
2633 nsec_fls
= fls64(nsec
);
2634 frequency_fls
= fls64(frequency
);
2638 * We got @count in @nsec, with a target of sample_freq HZ
2639 * the target period becomes:
2642 * period = -------------------
2643 * @nsec * sample_freq
2648 * Reduce accuracy by one bit such that @a and @b converge
2649 * to a similar magnitude.
2651 #define REDUCE_FLS(a, b) \
2653 if (a##_fls > b##_fls) { \
2663 * Reduce accuracy until either term fits in a u64, then proceed with
2664 * the other, so that finally we can do a u64/u64 division.
2666 while (count_fls
+ sec_fls
> 64 && nsec_fls
+ frequency_fls
> 64) {
2667 REDUCE_FLS(nsec
, frequency
);
2668 REDUCE_FLS(sec
, count
);
2671 if (count_fls
+ sec_fls
> 64) {
2672 divisor
= nsec
* frequency
;
2674 while (count_fls
+ sec_fls
> 64) {
2675 REDUCE_FLS(count
, sec
);
2679 dividend
= count
* sec
;
2681 dividend
= count
* sec
;
2683 while (nsec_fls
+ frequency_fls
> 64) {
2684 REDUCE_FLS(nsec
, frequency
);
2688 divisor
= nsec
* frequency
;
2694 return div64_u64(dividend
, divisor
);
2697 static DEFINE_PER_CPU(int, perf_throttled_count
);
2698 static DEFINE_PER_CPU(u64
, perf_throttled_seq
);
2700 static void perf_adjust_period(struct perf_event
*event
, u64 nsec
, u64 count
, bool disable
)
2702 struct hw_perf_event
*hwc
= &event
->hw
;
2703 s64 period
, sample_period
;
2706 period
= perf_calculate_period(event
, nsec
, count
);
2708 delta
= (s64
)(period
- hwc
->sample_period
);
2709 delta
= (delta
+ 7) / 8; /* low pass filter */
2711 sample_period
= hwc
->sample_period
+ delta
;
2716 hwc
->sample_period
= sample_period
;
2718 if (local64_read(&hwc
->period_left
) > 8*sample_period
) {
2720 event
->pmu
->stop(event
, PERF_EF_UPDATE
);
2722 local64_set(&hwc
->period_left
, 0);
2725 event
->pmu
->start(event
, PERF_EF_RELOAD
);
2730 * combine freq adjustment with unthrottling to avoid two passes over the
2731 * events. At the same time, make sure, having freq events does not change
2732 * the rate of unthrottling as that would introduce bias.
2734 static void perf_adjust_freq_unthr_context(struct perf_event_context
*ctx
,
2737 struct perf_event
*event
;
2738 struct hw_perf_event
*hwc
;
2739 u64 now
, period
= TICK_NSEC
;
2743 * only need to iterate over all events iff:
2744 * - context have events in frequency mode (needs freq adjust)
2745 * - there are events to unthrottle on this cpu
2747 if (!(ctx
->nr_freq
|| needs_unthr
))
2750 raw_spin_lock(&ctx
->lock
);
2751 perf_pmu_disable(ctx
->pmu
);
2753 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
2754 if (event
->state
!= PERF_EVENT_STATE_ACTIVE
)
2757 if (!event_filter_match(event
))
2760 perf_pmu_disable(event
->pmu
);
2764 if (hwc
->interrupts
== MAX_INTERRUPTS
) {
2765 hwc
->interrupts
= 0;
2766 perf_log_throttle(event
, 1);
2767 event
->pmu
->start(event
, 0);
2770 if (!event
->attr
.freq
|| !event
->attr
.sample_freq
)
2774 * stop the event and update event->count
2776 event
->pmu
->stop(event
, PERF_EF_UPDATE
);
2778 now
= local64_read(&event
->count
);
2779 delta
= now
- hwc
->freq_count_stamp
;
2780 hwc
->freq_count_stamp
= now
;
2784 * reload only if value has changed
2785 * we have stopped the event so tell that
2786 * to perf_adjust_period() to avoid stopping it
2790 perf_adjust_period(event
, period
, delta
, false);
2792 event
->pmu
->start(event
, delta
> 0 ? PERF_EF_RELOAD
: 0);
2794 perf_pmu_enable(event
->pmu
);
2797 perf_pmu_enable(ctx
->pmu
);
2798 raw_spin_unlock(&ctx
->lock
);
2802 * Round-robin a context's events:
2804 static void rotate_ctx(struct perf_event_context
*ctx
)
2807 * Rotate the first entry last of non-pinned groups. Rotation might be
2808 * disabled by the inheritance code.
2810 if (!ctx
->rotate_disable
)
2811 list_rotate_left(&ctx
->flexible_groups
);
2815 * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
2816 * because they're strictly cpu affine and rotate_start is called with IRQs
2817 * disabled, while rotate_context is called from IRQ context.
2819 static int perf_rotate_context(struct perf_cpu_context
*cpuctx
)
2821 struct perf_event_context
*ctx
= NULL
;
2822 int rotate
= 0, remove
= 1;
2824 if (cpuctx
->ctx
.nr_events
) {
2826 if (cpuctx
->ctx
.nr_events
!= cpuctx
->ctx
.nr_active
)
2830 ctx
= cpuctx
->task_ctx
;
2831 if (ctx
&& ctx
->nr_events
) {
2833 if (ctx
->nr_events
!= ctx
->nr_active
)
2840 perf_ctx_lock(cpuctx
, cpuctx
->task_ctx
);
2841 perf_pmu_disable(cpuctx
->ctx
.pmu
);
2843 cpu_ctx_sched_out(cpuctx
, EVENT_FLEXIBLE
);
2845 ctx_sched_out(ctx
, cpuctx
, EVENT_FLEXIBLE
);
2847 rotate_ctx(&cpuctx
->ctx
);
2851 perf_event_sched_in(cpuctx
, ctx
, current
);
2853 perf_pmu_enable(cpuctx
->ctx
.pmu
);
2854 perf_ctx_unlock(cpuctx
, cpuctx
->task_ctx
);
2857 list_del_init(&cpuctx
->rotation_list
);
2862 #ifdef CONFIG_NO_HZ_FULL
2863 bool perf_event_can_stop_tick(void)
2865 if (atomic_read(&nr_freq_events
) ||
2866 __this_cpu_read(perf_throttled_count
))
2873 void perf_event_task_tick(void)
2875 struct list_head
*head
= &__get_cpu_var(rotation_list
);
2876 struct perf_cpu_context
*cpuctx
, *tmp
;
2877 struct perf_event_context
*ctx
;
2880 WARN_ON(!irqs_disabled());
2882 __this_cpu_inc(perf_throttled_seq
);
2883 throttled
= __this_cpu_xchg(perf_throttled_count
, 0);
2885 list_for_each_entry_safe(cpuctx
, tmp
, head
, rotation_list
) {
2887 perf_adjust_freq_unthr_context(ctx
, throttled
);
2889 ctx
= cpuctx
->task_ctx
;
2891 perf_adjust_freq_unthr_context(ctx
, throttled
);
2895 static int event_enable_on_exec(struct perf_event
*event
,
2896 struct perf_event_context
*ctx
)
2898 if (!event
->attr
.enable_on_exec
)
2901 event
->attr
.enable_on_exec
= 0;
2902 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
)
2905 __perf_event_mark_enabled(event
);
2911 * Enable all of a task's events that have been marked enable-on-exec.
2912 * This expects task == current.
2914 static void perf_event_enable_on_exec(struct perf_event_context
*ctx
)
2916 struct perf_event
*event
;
2917 unsigned long flags
;
2921 local_irq_save(flags
);
2922 if (!ctx
|| !ctx
->nr_events
)
2926 * We must ctxsw out cgroup events to avoid conflict
2927 * when invoking perf_task_event_sched_in() later on
2928 * in this function. Otherwise we end up trying to
2929 * ctxswin cgroup events which are already scheduled
2932 perf_cgroup_sched_out(current
, NULL
);
2934 raw_spin_lock(&ctx
->lock
);
2935 task_ctx_sched_out(ctx
);
2937 list_for_each_entry(event
, &ctx
->event_list
, event_entry
) {
2938 ret
= event_enable_on_exec(event
, ctx
);
2944 * Unclone this context if we enabled any event.
2949 raw_spin_unlock(&ctx
->lock
);
2952 * Also calls ctxswin for cgroup events, if any:
2954 perf_event_context_sched_in(ctx
, ctx
->task
);
2956 local_irq_restore(flags
);
2960 * Cross CPU call to read the hardware event
2962 static void __perf_event_read(void *info
)
2964 struct perf_event
*event
= info
;
2965 struct perf_event_context
*ctx
= event
->ctx
;
2966 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
2969 * If this is a task context, we need to check whether it is
2970 * the current task context of this cpu. If not it has been
2971 * scheduled out before the smp call arrived. In that case
2972 * event->count would have been updated to a recent sample
2973 * when the event was scheduled out.
2975 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
)
2978 raw_spin_lock(&ctx
->lock
);
2979 if (ctx
->is_active
) {
2980 update_context_time(ctx
);
2981 update_cgrp_time_from_event(event
);
2983 update_event_times(event
);
2984 if (event
->state
== PERF_EVENT_STATE_ACTIVE
)
2985 event
->pmu
->read(event
);
2986 raw_spin_unlock(&ctx
->lock
);
2989 static inline u64
perf_event_count(struct perf_event
*event
)
2991 return local64_read(&event
->count
) + atomic64_read(&event
->child_count
);
2994 static u64
perf_event_read(struct perf_event
*event
)
2997 * If event is enabled and currently active on a CPU, update the
2998 * value in the event structure:
3000 if (event
->state
== PERF_EVENT_STATE_ACTIVE
) {
3001 smp_call_function_single(event
->oncpu
,
3002 __perf_event_read
, event
, 1);
3003 } else if (event
->state
== PERF_EVENT_STATE_INACTIVE
) {
3004 struct perf_event_context
*ctx
= event
->ctx
;
3005 unsigned long flags
;
3007 raw_spin_lock_irqsave(&ctx
->lock
, flags
);
3009 * may read while context is not active
3010 * (e.g., thread is blocked), in that case
3011 * we cannot update context time
3013 if (ctx
->is_active
) {
3014 update_context_time(ctx
);
3015 update_cgrp_time_from_event(event
);
3017 update_event_times(event
);
3018 raw_spin_unlock_irqrestore(&ctx
->lock
, flags
);
3021 return perf_event_count(event
);
3025 * Initialize the perf_event context in a task_struct:
3027 static void __perf_event_init_context(struct perf_event_context
*ctx
)
3029 raw_spin_lock_init(&ctx
->lock
);
3030 mutex_init(&ctx
->mutex
);
3031 INIT_LIST_HEAD(&ctx
->pinned_groups
);
3032 INIT_LIST_HEAD(&ctx
->flexible_groups
);
3033 INIT_LIST_HEAD(&ctx
->event_list
);
3034 atomic_set(&ctx
->refcount
, 1);
3037 static struct perf_event_context
*
3038 alloc_perf_context(struct pmu
*pmu
, struct task_struct
*task
)
3040 struct perf_event_context
*ctx
;
3042 ctx
= kzalloc(sizeof(struct perf_event_context
), GFP_KERNEL
);
3046 __perf_event_init_context(ctx
);
3049 get_task_struct(task
);
3056 static struct task_struct
*
3057 find_lively_task_by_vpid(pid_t vpid
)
3059 struct task_struct
*task
;
3066 task
= find_task_by_vpid(vpid
);
3068 get_task_struct(task
);
3072 return ERR_PTR(-ESRCH
);
3074 /* Reuse ptrace permission checks for now. */
3076 if (!ptrace_may_access(task
, PTRACE_MODE_READ
))
3081 put_task_struct(task
);
3082 return ERR_PTR(err
);
3087 * Returns a matching context with refcount and pincount.
3089 static struct perf_event_context
*
3090 find_get_context(struct pmu
*pmu
, struct task_struct
*task
, int cpu
)
3092 struct perf_event_context
*ctx
;
3093 struct perf_cpu_context
*cpuctx
;
3094 unsigned long flags
;
3098 /* Must be root to operate on a CPU event: */
3099 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN
))
3100 return ERR_PTR(-EACCES
);
3103 * We could be clever and allow to attach a event to an
3104 * offline CPU and activate it when the CPU comes up, but
3107 if (!cpu_online(cpu
))
3108 return ERR_PTR(-ENODEV
);
3110 cpuctx
= per_cpu_ptr(pmu
->pmu_cpu_context
, cpu
);
3119 ctxn
= pmu
->task_ctx_nr
;
3124 ctx
= perf_lock_task_context(task
, ctxn
, &flags
);
3128 raw_spin_unlock_irqrestore(&ctx
->lock
, flags
);
3130 ctx
= alloc_perf_context(pmu
, task
);
3136 mutex_lock(&task
->perf_event_mutex
);
3138 * If it has already passed perf_event_exit_task().
3139 * we must see PF_EXITING, it takes this mutex too.
3141 if (task
->flags
& PF_EXITING
)
3143 else if (task
->perf_event_ctxp
[ctxn
])
3148 rcu_assign_pointer(task
->perf_event_ctxp
[ctxn
], ctx
);
3150 mutex_unlock(&task
->perf_event_mutex
);
3152 if (unlikely(err
)) {
3164 return ERR_PTR(err
);
3167 static void perf_event_free_filter(struct perf_event
*event
);
3169 static void free_event_rcu(struct rcu_head
*head
)
3171 struct perf_event
*event
;
3173 event
= container_of(head
, struct perf_event
, rcu_head
);
3175 put_pid_ns(event
->ns
);
3176 perf_event_free_filter(event
);
3180 static void ring_buffer_put(struct ring_buffer
*rb
);
3181 static void ring_buffer_detach(struct perf_event
*event
, struct ring_buffer
*rb
);
3183 static void unaccount_event_cpu(struct perf_event
*event
, int cpu
)
3188 if (has_branch_stack(event
)) {
3189 if (!(event
->attach_state
& PERF_ATTACH_TASK
))
3190 atomic_dec(&per_cpu(perf_branch_stack_events
, cpu
));
3192 if (is_cgroup_event(event
))
3193 atomic_dec(&per_cpu(perf_cgroup_events
, cpu
));
3196 static void unaccount_event(struct perf_event
*event
)
3201 if (event
->attach_state
& PERF_ATTACH_TASK
)
3202 static_key_slow_dec_deferred(&perf_sched_events
);
3203 if (event
->attr
.mmap
|| event
->attr
.mmap_data
)
3204 atomic_dec(&nr_mmap_events
);
3205 if (event
->attr
.comm
)
3206 atomic_dec(&nr_comm_events
);
3207 if (event
->attr
.task
)
3208 atomic_dec(&nr_task_events
);
3209 if (event
->attr
.freq
)
3210 atomic_dec(&nr_freq_events
);
3211 if (is_cgroup_event(event
))
3212 static_key_slow_dec_deferred(&perf_sched_events
);
3213 if (has_branch_stack(event
))
3214 static_key_slow_dec_deferred(&perf_sched_events
);
3216 unaccount_event_cpu(event
, event
->cpu
);
3219 static void __free_event(struct perf_event
*event
)
3221 if (!event
->parent
) {
3222 if (event
->attr
.sample_type
& PERF_SAMPLE_CALLCHAIN
)
3223 put_callchain_buffers();
3227 event
->destroy(event
);
3230 put_ctx(event
->ctx
);
3232 call_rcu(&event
->rcu_head
, free_event_rcu
);
3234 static void free_event(struct perf_event
*event
)
3236 irq_work_sync(&event
->pending
);
3238 unaccount_event(event
);
3241 struct ring_buffer
*rb
;
3244 * Can happen when we close an event with re-directed output.
3246 * Since we have a 0 refcount, perf_mmap_close() will skip
3247 * over us; possibly making our ring_buffer_put() the last.
3249 mutex_lock(&event
->mmap_mutex
);
3252 rcu_assign_pointer(event
->rb
, NULL
);
3253 ring_buffer_detach(event
, rb
);
3254 ring_buffer_put(rb
); /* could be last */
3256 mutex_unlock(&event
->mmap_mutex
);
3259 if (is_cgroup_event(event
))
3260 perf_detach_cgroup(event
);
3263 __free_event(event
);
3266 int perf_event_release_kernel(struct perf_event
*event
)
3268 struct perf_event_context
*ctx
= event
->ctx
;
3270 WARN_ON_ONCE(ctx
->parent_ctx
);
3272 * There are two ways this annotation is useful:
3274 * 1) there is a lock recursion from perf_event_exit_task
3275 * see the comment there.
3277 * 2) there is a lock-inversion with mmap_sem through
3278 * perf_event_read_group(), which takes faults while
3279 * holding ctx->mutex, however this is called after
3280 * the last filedesc died, so there is no possibility
3281 * to trigger the AB-BA case.
3283 mutex_lock_nested(&ctx
->mutex
, SINGLE_DEPTH_NESTING
);
3284 raw_spin_lock_irq(&ctx
->lock
);
3285 perf_group_detach(event
);
3286 raw_spin_unlock_irq(&ctx
->lock
);
3287 perf_remove_from_context(event
);
3288 mutex_unlock(&ctx
->mutex
);
3294 EXPORT_SYMBOL_GPL(perf_event_release_kernel
);
3297 * Called when the last reference to the file is gone.
3299 static void put_event(struct perf_event
*event
)
3301 struct task_struct
*owner
;
3303 if (!atomic_long_dec_and_test(&event
->refcount
))
3307 owner
= ACCESS_ONCE(event
->owner
);
3309 * Matches the smp_wmb() in perf_event_exit_task(). If we observe
3310 * !owner it means the list deletion is complete and we can indeed
3311 * free this event, otherwise we need to serialize on
3312 * owner->perf_event_mutex.
3314 smp_read_barrier_depends();
3317 * Since delayed_put_task_struct() also drops the last
3318 * task reference we can safely take a new reference
3319 * while holding the rcu_read_lock().
3321 get_task_struct(owner
);
3326 mutex_lock(&owner
->perf_event_mutex
);
3328 * We have to re-check the event->owner field, if it is cleared
3329 * we raced with perf_event_exit_task(), acquiring the mutex
3330 * ensured they're done, and we can proceed with freeing the
3334 list_del_init(&event
->owner_entry
);
3335 mutex_unlock(&owner
->perf_event_mutex
);
3336 put_task_struct(owner
);
3339 perf_event_release_kernel(event
);
3342 static int perf_release(struct inode
*inode
, struct file
*file
)
3344 put_event(file
->private_data
);
3348 u64
perf_event_read_value(struct perf_event
*event
, u64
*enabled
, u64
*running
)
3350 struct perf_event
*child
;
3356 mutex_lock(&event
->child_mutex
);
3357 total
+= perf_event_read(event
);
3358 *enabled
+= event
->total_time_enabled
+
3359 atomic64_read(&event
->child_total_time_enabled
);
3360 *running
+= event
->total_time_running
+
3361 atomic64_read(&event
->child_total_time_running
);
3363 list_for_each_entry(child
, &event
->child_list
, child_list
) {
3364 total
+= perf_event_read(child
);
3365 *enabled
+= child
->total_time_enabled
;
3366 *running
+= child
->total_time_running
;
3368 mutex_unlock(&event
->child_mutex
);
3372 EXPORT_SYMBOL_GPL(perf_event_read_value
);
3374 static int perf_event_read_group(struct perf_event
*event
,
3375 u64 read_format
, char __user
*buf
)
3377 struct perf_event
*leader
= event
->group_leader
, *sub
;
3378 int n
= 0, size
= 0, ret
= -EFAULT
;
3379 struct perf_event_context
*ctx
= leader
->ctx
;
3381 u64 count
, enabled
, running
;
3383 mutex_lock(&ctx
->mutex
);
3384 count
= perf_event_read_value(leader
, &enabled
, &running
);
3386 values
[n
++] = 1 + leader
->nr_siblings
;
3387 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
3388 values
[n
++] = enabled
;
3389 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
3390 values
[n
++] = running
;
3391 values
[n
++] = count
;
3392 if (read_format
& PERF_FORMAT_ID
)
3393 values
[n
++] = primary_event_id(leader
);
3395 size
= n
* sizeof(u64
);
3397 if (copy_to_user(buf
, values
, size
))
3402 list_for_each_entry(sub
, &leader
->sibling_list
, group_entry
) {
3405 values
[n
++] = perf_event_read_value(sub
, &enabled
, &running
);
3406 if (read_format
& PERF_FORMAT_ID
)
3407 values
[n
++] = primary_event_id(sub
);
3409 size
= n
* sizeof(u64
);
3411 if (copy_to_user(buf
+ ret
, values
, size
)) {
3419 mutex_unlock(&ctx
->mutex
);
3424 static int perf_event_read_one(struct perf_event
*event
,
3425 u64 read_format
, char __user
*buf
)
3427 u64 enabled
, running
;
3431 values
[n
++] = perf_event_read_value(event
, &enabled
, &running
);
3432 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
3433 values
[n
++] = enabled
;
3434 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
3435 values
[n
++] = running
;
3436 if (read_format
& PERF_FORMAT_ID
)
3437 values
[n
++] = primary_event_id(event
);
3439 if (copy_to_user(buf
, values
, n
* sizeof(u64
)))
3442 return n
* sizeof(u64
);
3446 * Read the performance event - simple non blocking version for now
3449 perf_read_hw(struct perf_event
*event
, char __user
*buf
, size_t count
)
3451 u64 read_format
= event
->attr
.read_format
;
3455 * Return end-of-file for a read on a event that is in
3456 * error state (i.e. because it was pinned but it couldn't be
3457 * scheduled on to the CPU at some point).
3459 if (event
->state
== PERF_EVENT_STATE_ERROR
)
3462 if (count
< event
->read_size
)
3465 WARN_ON_ONCE(event
->ctx
->parent_ctx
);
3466 if (read_format
& PERF_FORMAT_GROUP
)
3467 ret
= perf_event_read_group(event
, read_format
, buf
);
3469 ret
= perf_event_read_one(event
, read_format
, buf
);
3475 perf_read(struct file
*file
, char __user
*buf
, size_t count
, loff_t
*ppos
)
3477 struct perf_event
*event
= file
->private_data
;
3479 return perf_read_hw(event
, buf
, count
);
3482 static unsigned int perf_poll(struct file
*file
, poll_table
*wait
)
3484 struct perf_event
*event
= file
->private_data
;
3485 struct ring_buffer
*rb
;
3486 unsigned int events
= POLL_HUP
;
3489 * Pin the event->rb by taking event->mmap_mutex; otherwise
3490 * perf_event_set_output() can swizzle our rb and make us miss wakeups.
3492 mutex_lock(&event
->mmap_mutex
);
3495 events
= atomic_xchg(&rb
->poll
, 0);
3496 mutex_unlock(&event
->mmap_mutex
);
3498 poll_wait(file
, &event
->waitq
, wait
);
3503 static void perf_event_reset(struct perf_event
*event
)
3505 (void)perf_event_read(event
);
3506 local64_set(&event
->count
, 0);
3507 perf_event_update_userpage(event
);
3511 * Holding the top-level event's child_mutex means that any
3512 * descendant process that has inherited this event will block
3513 * in sync_child_event if it goes to exit, thus satisfying the
3514 * task existence requirements of perf_event_enable/disable.
3516 static void perf_event_for_each_child(struct perf_event
*event
,
3517 void (*func
)(struct perf_event
*))
3519 struct perf_event
*child
;
3521 WARN_ON_ONCE(event
->ctx
->parent_ctx
);
3522 mutex_lock(&event
->child_mutex
);
3524 list_for_each_entry(child
, &event
->child_list
, child_list
)
3526 mutex_unlock(&event
->child_mutex
);
3529 static void perf_event_for_each(struct perf_event
*event
,
3530 void (*func
)(struct perf_event
*))
3532 struct perf_event_context
*ctx
= event
->ctx
;
3533 struct perf_event
*sibling
;
3535 WARN_ON_ONCE(ctx
->parent_ctx
);
3536 mutex_lock(&ctx
->mutex
);
3537 event
= event
->group_leader
;
3539 perf_event_for_each_child(event
, func
);
3540 list_for_each_entry(sibling
, &event
->sibling_list
, group_entry
)
3541 perf_event_for_each_child(sibling
, func
);
3542 mutex_unlock(&ctx
->mutex
);
3545 static int perf_event_period(struct perf_event
*event
, u64 __user
*arg
)
3547 struct perf_event_context
*ctx
= event
->ctx
;
3548 int ret
= 0, active
;
3551 if (!is_sampling_event(event
))
3554 if (copy_from_user(&value
, arg
, sizeof(value
)))
3560 raw_spin_lock_irq(&ctx
->lock
);
3561 if (event
->attr
.freq
) {
3562 if (value
> sysctl_perf_event_sample_rate
) {
3567 event
->attr
.sample_freq
= value
;
3569 event
->attr
.sample_period
= value
;
3570 event
->hw
.sample_period
= value
;
3573 active
= (event
->state
== PERF_EVENT_STATE_ACTIVE
);
3575 perf_pmu_disable(ctx
->pmu
);
3576 event
->pmu
->stop(event
, PERF_EF_UPDATE
);
3579 local64_set(&event
->hw
.period_left
, 0);
3582 event
->pmu
->start(event
, PERF_EF_RELOAD
);
3583 perf_pmu_enable(ctx
->pmu
);
3587 raw_spin_unlock_irq(&ctx
->lock
);
3592 static const struct file_operations perf_fops
;
3594 static inline int perf_fget_light(int fd
, struct fd
*p
)
3596 struct fd f
= fdget(fd
);
3600 if (f
.file
->f_op
!= &perf_fops
) {
3608 static int perf_event_set_output(struct perf_event
*event
,
3609 struct perf_event
*output_event
);
3610 static int perf_event_set_filter(struct perf_event
*event
, void __user
*arg
);
3612 static long perf_ioctl(struct file
*file
, unsigned int cmd
, unsigned long arg
)
3614 struct perf_event
*event
= file
->private_data
;
3615 void (*func
)(struct perf_event
*);
3619 case PERF_EVENT_IOC_ENABLE
:
3620 func
= perf_event_enable
;
3622 case PERF_EVENT_IOC_DISABLE
:
3623 func
= perf_event_disable
;
3625 case PERF_EVENT_IOC_RESET
:
3626 func
= perf_event_reset
;
3629 case PERF_EVENT_IOC_REFRESH
:
3630 return perf_event_refresh(event
, arg
);
3632 case PERF_EVENT_IOC_PERIOD
:
3633 return perf_event_period(event
, (u64 __user
*)arg
);
3635 case PERF_EVENT_IOC_ID
:
3637 u64 id
= primary_event_id(event
);
3639 if (copy_to_user((void __user
*)arg
, &id
, sizeof(id
)))
3644 case PERF_EVENT_IOC_SET_OUTPUT
:
3648 struct perf_event
*output_event
;
3650 ret
= perf_fget_light(arg
, &output
);
3653 output_event
= output
.file
->private_data
;
3654 ret
= perf_event_set_output(event
, output_event
);
3657 ret
= perf_event_set_output(event
, NULL
);
3662 case PERF_EVENT_IOC_SET_FILTER
:
3663 return perf_event_set_filter(event
, (void __user
*)arg
);
3669 if (flags
& PERF_IOC_FLAG_GROUP
)
3670 perf_event_for_each(event
, func
);
3672 perf_event_for_each_child(event
, func
);
3677 int perf_event_task_enable(void)
3679 struct perf_event
*event
;
3681 mutex_lock(¤t
->perf_event_mutex
);
3682 list_for_each_entry(event
, ¤t
->perf_event_list
, owner_entry
)
3683 perf_event_for_each_child(event
, perf_event_enable
);
3684 mutex_unlock(¤t
->perf_event_mutex
);
3689 int perf_event_task_disable(void)
3691 struct perf_event
*event
;
3693 mutex_lock(¤t
->perf_event_mutex
);
3694 list_for_each_entry(event
, ¤t
->perf_event_list
, owner_entry
)
3695 perf_event_for_each_child(event
, perf_event_disable
);
3696 mutex_unlock(¤t
->perf_event_mutex
);
3701 static int perf_event_index(struct perf_event
*event
)
3703 if (event
->hw
.state
& PERF_HES_STOPPED
)
3706 if (event
->state
!= PERF_EVENT_STATE_ACTIVE
)
3709 return event
->pmu
->event_idx(event
);
3712 static void calc_timer_values(struct perf_event
*event
,
3719 *now
= perf_clock();
3720 ctx_time
= event
->shadow_ctx_time
+ *now
;
3721 *enabled
= ctx_time
- event
->tstamp_enabled
;
3722 *running
= ctx_time
- event
->tstamp_running
;
3725 static void perf_event_init_userpage(struct perf_event
*event
)
3727 struct perf_event_mmap_page
*userpg
;
3728 struct ring_buffer
*rb
;
3731 rb
= rcu_dereference(event
->rb
);
3735 userpg
= rb
->user_page
;
3737 /* Allow new userspace to detect that bit 0 is deprecated */
3738 userpg
->cap_bit0_is_deprecated
= 1;
3739 userpg
->size
= offsetof(struct perf_event_mmap_page
, __reserved
);
3745 void __weak
arch_perf_update_userpage(struct perf_event_mmap_page
*userpg
, u64 now
)
3750 * Callers need to ensure there can be no nesting of this function, otherwise
3751 * the seqlock logic goes bad. We can not serialize this because the arch
3752 * code calls this from NMI context.
3754 void perf_event_update_userpage(struct perf_event
*event
)
3756 struct perf_event_mmap_page
*userpg
;
3757 struct ring_buffer
*rb
;
3758 u64 enabled
, running
, now
;
3761 rb
= rcu_dereference(event
->rb
);
3766 * compute total_time_enabled, total_time_running
3767 * based on snapshot values taken when the event
3768 * was last scheduled in.
3770 * we cannot simply called update_context_time()
3771 * because of locking issue as we can be called in
3774 calc_timer_values(event
, &now
, &enabled
, &running
);
3776 userpg
= rb
->user_page
;
3778 * Disable preemption so as to not let the corresponding user-space
3779 * spin too long if we get preempted.
3784 userpg
->index
= perf_event_index(event
);
3785 userpg
->offset
= perf_event_count(event
);
3787 userpg
->offset
-= local64_read(&event
->hw
.prev_count
);
3789 userpg
->time_enabled
= enabled
+
3790 atomic64_read(&event
->child_total_time_enabled
);
3792 userpg
->time_running
= running
+
3793 atomic64_read(&event
->child_total_time_running
);
3795 arch_perf_update_userpage(userpg
, now
);
3804 static int perf_mmap_fault(struct vm_area_struct
*vma
, struct vm_fault
*vmf
)
3806 struct perf_event
*event
= vma
->vm_file
->private_data
;
3807 struct ring_buffer
*rb
;
3808 int ret
= VM_FAULT_SIGBUS
;
3810 if (vmf
->flags
& FAULT_FLAG_MKWRITE
) {
3811 if (vmf
->pgoff
== 0)
3817 rb
= rcu_dereference(event
->rb
);
3821 if (vmf
->pgoff
&& (vmf
->flags
& FAULT_FLAG_WRITE
))
3824 vmf
->page
= perf_mmap_to_page(rb
, vmf
->pgoff
);
3828 get_page(vmf
->page
);
3829 vmf
->page
->mapping
= vma
->vm_file
->f_mapping
;
3830 vmf
->page
->index
= vmf
->pgoff
;
3839 static void ring_buffer_attach(struct perf_event
*event
,
3840 struct ring_buffer
*rb
)
3842 unsigned long flags
;
3844 if (!list_empty(&event
->rb_entry
))
3847 spin_lock_irqsave(&rb
->event_lock
, flags
);
3848 if (list_empty(&event
->rb_entry
))
3849 list_add(&event
->rb_entry
, &rb
->event_list
);
3850 spin_unlock_irqrestore(&rb
->event_lock
, flags
);
3853 static void ring_buffer_detach(struct perf_event
*event
, struct ring_buffer
*rb
)
3855 unsigned long flags
;
3857 if (list_empty(&event
->rb_entry
))
3860 spin_lock_irqsave(&rb
->event_lock
, flags
);
3861 list_del_init(&event
->rb_entry
);
3862 wake_up_all(&event
->waitq
);
3863 spin_unlock_irqrestore(&rb
->event_lock
, flags
);
3866 static void ring_buffer_wakeup(struct perf_event
*event
)
3868 struct ring_buffer
*rb
;
3871 rb
= rcu_dereference(event
->rb
);
3873 list_for_each_entry_rcu(event
, &rb
->event_list
, rb_entry
)
3874 wake_up_all(&event
->waitq
);
3879 static void rb_free_rcu(struct rcu_head
*rcu_head
)
3881 struct ring_buffer
*rb
;
3883 rb
= container_of(rcu_head
, struct ring_buffer
, rcu_head
);
3887 static struct ring_buffer
*ring_buffer_get(struct perf_event
*event
)
3889 struct ring_buffer
*rb
;
3892 rb
= rcu_dereference(event
->rb
);
3894 if (!atomic_inc_not_zero(&rb
->refcount
))
3902 static void ring_buffer_put(struct ring_buffer
*rb
)
3904 if (!atomic_dec_and_test(&rb
->refcount
))
3907 WARN_ON_ONCE(!list_empty(&rb
->event_list
));
3909 call_rcu(&rb
->rcu_head
, rb_free_rcu
);
3912 static void perf_mmap_open(struct vm_area_struct
*vma
)
3914 struct perf_event
*event
= vma
->vm_file
->private_data
;
3916 atomic_inc(&event
->mmap_count
);
3917 atomic_inc(&event
->rb
->mmap_count
);
3921 * A buffer can be mmap()ed multiple times; either directly through the same
3922 * event, or through other events by use of perf_event_set_output().
3924 * In order to undo the VM accounting done by perf_mmap() we need to destroy
3925 * the buffer here, where we still have a VM context. This means we need
3926 * to detach all events redirecting to us.
3928 static void perf_mmap_close(struct vm_area_struct
*vma
)
3930 struct perf_event
*event
= vma
->vm_file
->private_data
;
3932 struct ring_buffer
*rb
= event
->rb
;
3933 struct user_struct
*mmap_user
= rb
->mmap_user
;
3934 int mmap_locked
= rb
->mmap_locked
;
3935 unsigned long size
= perf_data_size(rb
);
3937 atomic_dec(&rb
->mmap_count
);
3939 if (!atomic_dec_and_mutex_lock(&event
->mmap_count
, &event
->mmap_mutex
))
3942 /* Detach current event from the buffer. */
3943 rcu_assign_pointer(event
->rb
, NULL
);
3944 ring_buffer_detach(event
, rb
);
3945 mutex_unlock(&event
->mmap_mutex
);
3947 /* If there's still other mmap()s of this buffer, we're done. */
3948 if (atomic_read(&rb
->mmap_count
)) {
3949 ring_buffer_put(rb
); /* can't be last */
3954 * No other mmap()s, detach from all other events that might redirect
3955 * into the now unreachable buffer. Somewhat complicated by the
3956 * fact that rb::event_lock otherwise nests inside mmap_mutex.
3960 list_for_each_entry_rcu(event
, &rb
->event_list
, rb_entry
) {
3961 if (!atomic_long_inc_not_zero(&event
->refcount
)) {
3963 * This event is en-route to free_event() which will
3964 * detach it and remove it from the list.
3970 mutex_lock(&event
->mmap_mutex
);
3972 * Check we didn't race with perf_event_set_output() which can
3973 * swizzle the rb from under us while we were waiting to
3974 * acquire mmap_mutex.
3976 * If we find a different rb; ignore this event, a next
3977 * iteration will no longer find it on the list. We have to
3978 * still restart the iteration to make sure we're not now
3979 * iterating the wrong list.
3981 if (event
->rb
== rb
) {
3982 rcu_assign_pointer(event
->rb
, NULL
);
3983 ring_buffer_detach(event
, rb
);
3984 ring_buffer_put(rb
); /* can't be last, we still have one */
3986 mutex_unlock(&event
->mmap_mutex
);
3990 * Restart the iteration; either we're on the wrong list or
3991 * destroyed its integrity by doing a deletion.
3998 * It could be there's still a few 0-ref events on the list; they'll
3999 * get cleaned up by free_event() -- they'll also still have their
4000 * ref on the rb and will free it whenever they are done with it.
4002 * Aside from that, this buffer is 'fully' detached and unmapped,
4003 * undo the VM accounting.
4006 atomic_long_sub((size
>> PAGE_SHIFT
) + 1, &mmap_user
->locked_vm
);
4007 vma
->vm_mm
->pinned_vm
-= mmap_locked
;
4008 free_uid(mmap_user
);
4010 ring_buffer_put(rb
); /* could be last */
4013 static const struct vm_operations_struct perf_mmap_vmops
= {
4014 .open
= perf_mmap_open
,
4015 .close
= perf_mmap_close
,
4016 .fault
= perf_mmap_fault
,
4017 .page_mkwrite
= perf_mmap_fault
,
4020 static int perf_mmap(struct file
*file
, struct vm_area_struct
*vma
)
4022 struct perf_event
*event
= file
->private_data
;
4023 unsigned long user_locked
, user_lock_limit
;
4024 struct user_struct
*user
= current_user();
4025 unsigned long locked
, lock_limit
;
4026 struct ring_buffer
*rb
;
4027 unsigned long vma_size
;
4028 unsigned long nr_pages
;
4029 long user_extra
, extra
;
4030 int ret
= 0, flags
= 0;
4033 * Don't allow mmap() of inherited per-task counters. This would
4034 * create a performance issue due to all children writing to the
4037 if (event
->cpu
== -1 && event
->attr
.inherit
)
4040 if (!(vma
->vm_flags
& VM_SHARED
))
4043 vma_size
= vma
->vm_end
- vma
->vm_start
;
4044 nr_pages
= (vma_size
/ PAGE_SIZE
) - 1;
4047 * If we have rb pages ensure they're a power-of-two number, so we
4048 * can do bitmasks instead of modulo.
4050 if (nr_pages
!= 0 && !is_power_of_2(nr_pages
))
4053 if (vma_size
!= PAGE_SIZE
* (1 + nr_pages
))
4056 if (vma
->vm_pgoff
!= 0)
4059 WARN_ON_ONCE(event
->ctx
->parent_ctx
);
4061 mutex_lock(&event
->mmap_mutex
);
4063 if (event
->rb
->nr_pages
!= nr_pages
) {
4068 if (!atomic_inc_not_zero(&event
->rb
->mmap_count
)) {
4070 * Raced against perf_mmap_close() through
4071 * perf_event_set_output(). Try again, hope for better
4074 mutex_unlock(&event
->mmap_mutex
);
4081 user_extra
= nr_pages
+ 1;
4082 user_lock_limit
= sysctl_perf_event_mlock
>> (PAGE_SHIFT
- 10);
4085 * Increase the limit linearly with more CPUs:
4087 user_lock_limit
*= num_online_cpus();
4089 user_locked
= atomic_long_read(&user
->locked_vm
) + user_extra
;
4092 if (user_locked
> user_lock_limit
)
4093 extra
= user_locked
- user_lock_limit
;
4095 lock_limit
= rlimit(RLIMIT_MEMLOCK
);
4096 lock_limit
>>= PAGE_SHIFT
;
4097 locked
= vma
->vm_mm
->pinned_vm
+ extra
;
4099 if ((locked
> lock_limit
) && perf_paranoid_tracepoint_raw() &&
4100 !capable(CAP_IPC_LOCK
)) {
4107 if (vma
->vm_flags
& VM_WRITE
)
4108 flags
|= RING_BUFFER_WRITABLE
;
4110 rb
= rb_alloc(nr_pages
,
4111 event
->attr
.watermark
? event
->attr
.wakeup_watermark
: 0,
4119 atomic_set(&rb
->mmap_count
, 1);
4120 rb
->mmap_locked
= extra
;
4121 rb
->mmap_user
= get_current_user();
4123 atomic_long_add(user_extra
, &user
->locked_vm
);
4124 vma
->vm_mm
->pinned_vm
+= extra
;
4126 ring_buffer_attach(event
, rb
);
4127 rcu_assign_pointer(event
->rb
, rb
);
4129 perf_event_init_userpage(event
);
4130 perf_event_update_userpage(event
);
4134 atomic_inc(&event
->mmap_count
);
4135 mutex_unlock(&event
->mmap_mutex
);
4138 * Since pinned accounting is per vm we cannot allow fork() to copy our
4141 vma
->vm_flags
|= VM_DONTCOPY
| VM_DONTEXPAND
| VM_DONTDUMP
;
4142 vma
->vm_ops
= &perf_mmap_vmops
;
4147 static int perf_fasync(int fd
, struct file
*filp
, int on
)
4149 struct inode
*inode
= file_inode(filp
);
4150 struct perf_event
*event
= filp
->private_data
;
4153 mutex_lock(&inode
->i_mutex
);
4154 retval
= fasync_helper(fd
, filp
, on
, &event
->fasync
);
4155 mutex_unlock(&inode
->i_mutex
);
4163 static const struct file_operations perf_fops
= {
4164 .llseek
= no_llseek
,
4165 .release
= perf_release
,
4168 .unlocked_ioctl
= perf_ioctl
,
4169 .compat_ioctl
= perf_ioctl
,
4171 .fasync
= perf_fasync
,
4177 * If there's data, ensure we set the poll() state and publish everything
4178 * to user-space before waking everybody up.
4181 void perf_event_wakeup(struct perf_event
*event
)
4183 ring_buffer_wakeup(event
);
4185 if (event
->pending_kill
) {
4186 kill_fasync(&event
->fasync
, SIGIO
, event
->pending_kill
);
4187 event
->pending_kill
= 0;
4191 static void perf_pending_event(struct irq_work
*entry
)
4193 struct perf_event
*event
= container_of(entry
,
4194 struct perf_event
, pending
);
4196 if (event
->pending_disable
) {
4197 event
->pending_disable
= 0;
4198 __perf_event_disable(event
);
4201 if (event
->pending_wakeup
) {
4202 event
->pending_wakeup
= 0;
4203 perf_event_wakeup(event
);
4208 * We assume there is only KVM supporting the callbacks.
4209 * Later on, we might change it to a list if there is
4210 * another virtualization implementation supporting the callbacks.
4212 struct perf_guest_info_callbacks
*perf_guest_cbs
;
4214 int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks
*cbs
)
4216 perf_guest_cbs
= cbs
;
4219 EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks
);
4221 int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks
*cbs
)
4223 perf_guest_cbs
= NULL
;
4226 EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks
);
4229 perf_output_sample_regs(struct perf_output_handle
*handle
,
4230 struct pt_regs
*regs
, u64 mask
)
4234 for_each_set_bit(bit
, (const unsigned long *) &mask
,
4235 sizeof(mask
) * BITS_PER_BYTE
) {
4238 val
= perf_reg_value(regs
, bit
);
4239 perf_output_put(handle
, val
);
4243 static void perf_sample_regs_user(struct perf_regs_user
*regs_user
,
4244 struct pt_regs
*regs
)
4246 if (!user_mode(regs
)) {
4248 regs
= task_pt_regs(current
);
4254 regs_user
->regs
= regs
;
4255 regs_user
->abi
= perf_reg_abi(current
);
4260 * Get remaining task size from user stack pointer.
4262 * It'd be better to take stack vma map and limit this more
4263 * precisly, but there's no way to get it safely under interrupt,
4264 * so using TASK_SIZE as limit.
4266 static u64
perf_ustack_task_size(struct pt_regs
*regs
)
4268 unsigned long addr
= perf_user_stack_pointer(regs
);
4270 if (!addr
|| addr
>= TASK_SIZE
)
4273 return TASK_SIZE
- addr
;
4277 perf_sample_ustack_size(u16 stack_size
, u16 header_size
,
4278 struct pt_regs
*regs
)
4282 /* No regs, no stack pointer, no dump. */
4287 * Check if we fit in with the requested stack size into the:
4289 * If we don't, we limit the size to the TASK_SIZE.
4291 * - remaining sample size
4292 * If we don't, we customize the stack size to
4293 * fit in to the remaining sample size.
4296 task_size
= min((u64
) USHRT_MAX
, perf_ustack_task_size(regs
));
4297 stack_size
= min(stack_size
, (u16
) task_size
);
4299 /* Current header size plus static size and dynamic size. */
4300 header_size
+= 2 * sizeof(u64
);
4302 /* Do we fit in with the current stack dump size? */
4303 if ((u16
) (header_size
+ stack_size
) < header_size
) {
4305 * If we overflow the maximum size for the sample,
4306 * we customize the stack dump size to fit in.
4308 stack_size
= USHRT_MAX
- header_size
- sizeof(u64
);
4309 stack_size
= round_up(stack_size
, sizeof(u64
));
4316 perf_output_sample_ustack(struct perf_output_handle
*handle
, u64 dump_size
,
4317 struct pt_regs
*regs
)
4319 /* Case of a kernel thread, nothing to dump */
4322 perf_output_put(handle
, size
);
4331 * - the size requested by user or the best one we can fit
4332 * in to the sample max size
4334 * - user stack dump data
4336 * - the actual dumped size
4340 perf_output_put(handle
, dump_size
);
4343 sp
= perf_user_stack_pointer(regs
);
4344 rem
= __output_copy_user(handle
, (void *) sp
, dump_size
);
4345 dyn_size
= dump_size
- rem
;
4347 perf_output_skip(handle
, rem
);
4350 perf_output_put(handle
, dyn_size
);
4354 static void __perf_event_header__init_id(struct perf_event_header
*header
,
4355 struct perf_sample_data
*data
,
4356 struct perf_event
*event
)
4358 u64 sample_type
= event
->attr
.sample_type
;
4360 data
->type
= sample_type
;
4361 header
->size
+= event
->id_header_size
;
4363 if (sample_type
& PERF_SAMPLE_TID
) {
4364 /* namespace issues */
4365 data
->tid_entry
.pid
= perf_event_pid(event
, current
);
4366 data
->tid_entry
.tid
= perf_event_tid(event
, current
);
4369 if (sample_type
& PERF_SAMPLE_TIME
)
4370 data
->time
= perf_clock();
4372 if (sample_type
& (PERF_SAMPLE_ID
| PERF_SAMPLE_IDENTIFIER
))
4373 data
->id
= primary_event_id(event
);
4375 if (sample_type
& PERF_SAMPLE_STREAM_ID
)
4376 data
->stream_id
= event
->id
;
4378 if (sample_type
& PERF_SAMPLE_CPU
) {
4379 data
->cpu_entry
.cpu
= raw_smp_processor_id();
4380 data
->cpu_entry
.reserved
= 0;
4384 void perf_event_header__init_id(struct perf_event_header
*header
,
4385 struct perf_sample_data
*data
,
4386 struct perf_event
*event
)
4388 if (event
->attr
.sample_id_all
)
4389 __perf_event_header__init_id(header
, data
, event
);
4392 static void __perf_event__output_id_sample(struct perf_output_handle
*handle
,
4393 struct perf_sample_data
*data
)
4395 u64 sample_type
= data
->type
;
4397 if (sample_type
& PERF_SAMPLE_TID
)
4398 perf_output_put(handle
, data
->tid_entry
);
4400 if (sample_type
& PERF_SAMPLE_TIME
)
4401 perf_output_put(handle
, data
->time
);
4403 if (sample_type
& PERF_SAMPLE_ID
)
4404 perf_output_put(handle
, data
->id
);
4406 if (sample_type
& PERF_SAMPLE_STREAM_ID
)
4407 perf_output_put(handle
, data
->stream_id
);
4409 if (sample_type
& PERF_SAMPLE_CPU
)
4410 perf_output_put(handle
, data
->cpu_entry
);
4412 if (sample_type
& PERF_SAMPLE_IDENTIFIER
)
4413 perf_output_put(handle
, data
->id
);
4416 void perf_event__output_id_sample(struct perf_event
*event
,
4417 struct perf_output_handle
*handle
,
4418 struct perf_sample_data
*sample
)
4420 if (event
->attr
.sample_id_all
)
4421 __perf_event__output_id_sample(handle
, sample
);
4424 static void perf_output_read_one(struct perf_output_handle
*handle
,
4425 struct perf_event
*event
,
4426 u64 enabled
, u64 running
)
4428 u64 read_format
= event
->attr
.read_format
;
4432 values
[n
++] = perf_event_count(event
);
4433 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
) {
4434 values
[n
++] = enabled
+
4435 atomic64_read(&event
->child_total_time_enabled
);
4437 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
) {
4438 values
[n
++] = running
+
4439 atomic64_read(&event
->child_total_time_running
);
4441 if (read_format
& PERF_FORMAT_ID
)
4442 values
[n
++] = primary_event_id(event
);
4444 __output_copy(handle
, values
, n
* sizeof(u64
));
4448 * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
4450 static void perf_output_read_group(struct perf_output_handle
*handle
,
4451 struct perf_event
*event
,
4452 u64 enabled
, u64 running
)
4454 struct perf_event
*leader
= event
->group_leader
, *sub
;
4455 u64 read_format
= event
->attr
.read_format
;
4459 values
[n
++] = 1 + leader
->nr_siblings
;
4461 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
4462 values
[n
++] = enabled
;
4464 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
4465 values
[n
++] = running
;
4467 if (leader
!= event
)
4468 leader
->pmu
->read(leader
);
4470 values
[n
++] = perf_event_count(leader
);
4471 if (read_format
& PERF_FORMAT_ID
)
4472 values
[n
++] = primary_event_id(leader
);
4474 __output_copy(handle
, values
, n
* sizeof(u64
));
4476 list_for_each_entry(sub
, &leader
->sibling_list
, group_entry
) {
4479 if ((sub
!= event
) &&
4480 (sub
->state
== PERF_EVENT_STATE_ACTIVE
))
4481 sub
->pmu
->read(sub
);
4483 values
[n
++] = perf_event_count(sub
);
4484 if (read_format
& PERF_FORMAT_ID
)
4485 values
[n
++] = primary_event_id(sub
);
4487 __output_copy(handle
, values
, n
* sizeof(u64
));
4491 #define PERF_FORMAT_TOTAL_TIMES (PERF_FORMAT_TOTAL_TIME_ENABLED|\
4492 PERF_FORMAT_TOTAL_TIME_RUNNING)
4494 static void perf_output_read(struct perf_output_handle
*handle
,
4495 struct perf_event
*event
)
4497 u64 enabled
= 0, running
= 0, now
;
4498 u64 read_format
= event
->attr
.read_format
;
4501 * compute total_time_enabled, total_time_running
4502 * based on snapshot values taken when the event
4503 * was last scheduled in.
4505 * we cannot simply called update_context_time()
4506 * because of locking issue as we are called in
4509 if (read_format
& PERF_FORMAT_TOTAL_TIMES
)
4510 calc_timer_values(event
, &now
, &enabled
, &running
);
4512 if (event
->attr
.read_format
& PERF_FORMAT_GROUP
)
4513 perf_output_read_group(handle
, event
, enabled
, running
);
4515 perf_output_read_one(handle
, event
, enabled
, running
);
4518 void perf_output_sample(struct perf_output_handle
*handle
,
4519 struct perf_event_header
*header
,
4520 struct perf_sample_data
*data
,
4521 struct perf_event
*event
)
4523 u64 sample_type
= data
->type
;
4525 perf_output_put(handle
, *header
);
4527 if (sample_type
& PERF_SAMPLE_IDENTIFIER
)
4528 perf_output_put(handle
, data
->id
);
4530 if (sample_type
& PERF_SAMPLE_IP
)
4531 perf_output_put(handle
, data
->ip
);
4533 if (sample_type
& PERF_SAMPLE_TID
)
4534 perf_output_put(handle
, data
->tid_entry
);
4536 if (sample_type
& PERF_SAMPLE_TIME
)
4537 perf_output_put(handle
, data
->time
);
4539 if (sample_type
& PERF_SAMPLE_ADDR
)
4540 perf_output_put(handle
, data
->addr
);
4542 if (sample_type
& PERF_SAMPLE_ID
)
4543 perf_output_put(handle
, data
->id
);
4545 if (sample_type
& PERF_SAMPLE_STREAM_ID
)
4546 perf_output_put(handle
, data
->stream_id
);
4548 if (sample_type
& PERF_SAMPLE_CPU
)
4549 perf_output_put(handle
, data
->cpu_entry
);
4551 if (sample_type
& PERF_SAMPLE_PERIOD
)
4552 perf_output_put(handle
, data
->period
);
4554 if (sample_type
& PERF_SAMPLE_READ
)
4555 perf_output_read(handle
, event
);
4557 if (sample_type
& PERF_SAMPLE_CALLCHAIN
) {
4558 if (data
->callchain
) {
4561 if (data
->callchain
)
4562 size
+= data
->callchain
->nr
;
4564 size
*= sizeof(u64
);
4566 __output_copy(handle
, data
->callchain
, size
);
4569 perf_output_put(handle
, nr
);
4573 if (sample_type
& PERF_SAMPLE_RAW
) {
4575 perf_output_put(handle
, data
->raw
->size
);
4576 __output_copy(handle
, data
->raw
->data
,
4583 .size
= sizeof(u32
),
4586 perf_output_put(handle
, raw
);
4590 if (sample_type
& PERF_SAMPLE_BRANCH_STACK
) {
4591 if (data
->br_stack
) {
4594 size
= data
->br_stack
->nr
4595 * sizeof(struct perf_branch_entry
);
4597 perf_output_put(handle
, data
->br_stack
->nr
);
4598 perf_output_copy(handle
, data
->br_stack
->entries
, size
);
4601 * we always store at least the value of nr
4604 perf_output_put(handle
, nr
);
4608 if (sample_type
& PERF_SAMPLE_REGS_USER
) {
4609 u64 abi
= data
->regs_user
.abi
;
4612 * If there are no regs to dump, notice it through
4613 * first u64 being zero (PERF_SAMPLE_REGS_ABI_NONE).
4615 perf_output_put(handle
, abi
);
4618 u64 mask
= event
->attr
.sample_regs_user
;
4619 perf_output_sample_regs(handle
,
4620 data
->regs_user
.regs
,
4625 if (sample_type
& PERF_SAMPLE_STACK_USER
) {
4626 perf_output_sample_ustack(handle
,
4627 data
->stack_user_size
,
4628 data
->regs_user
.regs
);
4631 if (sample_type
& PERF_SAMPLE_WEIGHT
)
4632 perf_output_put(handle
, data
->weight
);
4634 if (sample_type
& PERF_SAMPLE_DATA_SRC
)
4635 perf_output_put(handle
, data
->data_src
.val
);
4637 if (sample_type
& PERF_SAMPLE_TRANSACTION
)
4638 perf_output_put(handle
, data
->txn
);
4640 if (!event
->attr
.watermark
) {
4641 int wakeup_events
= event
->attr
.wakeup_events
;
4643 if (wakeup_events
) {
4644 struct ring_buffer
*rb
= handle
->rb
;
4645 int events
= local_inc_return(&rb
->events
);
4647 if (events
>= wakeup_events
) {
4648 local_sub(wakeup_events
, &rb
->events
);
4649 local_inc(&rb
->wakeup
);
4655 void perf_prepare_sample(struct perf_event_header
*header
,
4656 struct perf_sample_data
*data
,
4657 struct perf_event
*event
,
4658 struct pt_regs
*regs
)
4660 u64 sample_type
= event
->attr
.sample_type
;
4662 header
->type
= PERF_RECORD_SAMPLE
;
4663 header
->size
= sizeof(*header
) + event
->header_size
;
4666 header
->misc
|= perf_misc_flags(regs
);
4668 __perf_event_header__init_id(header
, data
, event
);
4670 if (sample_type
& PERF_SAMPLE_IP
)
4671 data
->ip
= perf_instruction_pointer(regs
);
4673 if (sample_type
& PERF_SAMPLE_CALLCHAIN
) {
4676 data
->callchain
= perf_callchain(event
, regs
);
4678 if (data
->callchain
)
4679 size
+= data
->callchain
->nr
;
4681 header
->size
+= size
* sizeof(u64
);
4684 if (sample_type
& PERF_SAMPLE_RAW
) {
4685 int size
= sizeof(u32
);
4688 size
+= data
->raw
->size
;
4690 size
+= sizeof(u32
);
4692 WARN_ON_ONCE(size
& (sizeof(u64
)-1));
4693 header
->size
+= size
;
4696 if (sample_type
& PERF_SAMPLE_BRANCH_STACK
) {
4697 int size
= sizeof(u64
); /* nr */
4698 if (data
->br_stack
) {
4699 size
+= data
->br_stack
->nr
4700 * sizeof(struct perf_branch_entry
);
4702 header
->size
+= size
;
4705 if (sample_type
& PERF_SAMPLE_REGS_USER
) {
4706 /* regs dump ABI info */
4707 int size
= sizeof(u64
);
4709 perf_sample_regs_user(&data
->regs_user
, regs
);
4711 if (data
->regs_user
.regs
) {
4712 u64 mask
= event
->attr
.sample_regs_user
;
4713 size
+= hweight64(mask
) * sizeof(u64
);
4716 header
->size
+= size
;
4719 if (sample_type
& PERF_SAMPLE_STACK_USER
) {
4721 * Either we need PERF_SAMPLE_STACK_USER bit to be allways
4722 * processed as the last one or have additional check added
4723 * in case new sample type is added, because we could eat
4724 * up the rest of the sample size.
4726 struct perf_regs_user
*uregs
= &data
->regs_user
;
4727 u16 stack_size
= event
->attr
.sample_stack_user
;
4728 u16 size
= sizeof(u64
);
4731 perf_sample_regs_user(uregs
, regs
);
4733 stack_size
= perf_sample_ustack_size(stack_size
, header
->size
,
4737 * If there is something to dump, add space for the dump
4738 * itself and for the field that tells the dynamic size,
4739 * which is how many have been actually dumped.
4742 size
+= sizeof(u64
) + stack_size
;
4744 data
->stack_user_size
= stack_size
;
4745 header
->size
+= size
;
4749 static void perf_event_output(struct perf_event
*event
,
4750 struct perf_sample_data
*data
,
4751 struct pt_regs
*regs
)
4753 struct perf_output_handle handle
;
4754 struct perf_event_header header
;
4756 /* protect the callchain buffers */
4759 perf_prepare_sample(&header
, data
, event
, regs
);
4761 if (perf_output_begin(&handle
, event
, header
.size
))
4764 perf_output_sample(&handle
, &header
, data
, event
);
4766 perf_output_end(&handle
);
4776 struct perf_read_event
{
4777 struct perf_event_header header
;
4784 perf_event_read_event(struct perf_event
*event
,
4785 struct task_struct
*task
)
4787 struct perf_output_handle handle
;
4788 struct perf_sample_data sample
;
4789 struct perf_read_event read_event
= {
4791 .type
= PERF_RECORD_READ
,
4793 .size
= sizeof(read_event
) + event
->read_size
,
4795 .pid
= perf_event_pid(event
, task
),
4796 .tid
= perf_event_tid(event
, task
),
4800 perf_event_header__init_id(&read_event
.header
, &sample
, event
);
4801 ret
= perf_output_begin(&handle
, event
, read_event
.header
.size
);
4805 perf_output_put(&handle
, read_event
);
4806 perf_output_read(&handle
, event
);
4807 perf_event__output_id_sample(event
, &handle
, &sample
);
4809 perf_output_end(&handle
);
4812 typedef void (perf_event_aux_output_cb
)(struct perf_event
*event
, void *data
);
4815 perf_event_aux_ctx(struct perf_event_context
*ctx
,
4816 perf_event_aux_output_cb output
,
4819 struct perf_event
*event
;
4821 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
4822 if (event
->state
< PERF_EVENT_STATE_INACTIVE
)
4824 if (!event_filter_match(event
))
4826 output(event
, data
);
4831 perf_event_aux(perf_event_aux_output_cb output
, void *data
,
4832 struct perf_event_context
*task_ctx
)
4834 struct perf_cpu_context
*cpuctx
;
4835 struct perf_event_context
*ctx
;
4840 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
4841 cpuctx
= get_cpu_ptr(pmu
->pmu_cpu_context
);
4842 if (cpuctx
->unique_pmu
!= pmu
)
4844 perf_event_aux_ctx(&cpuctx
->ctx
, output
, data
);
4847 ctxn
= pmu
->task_ctx_nr
;
4850 ctx
= rcu_dereference(current
->perf_event_ctxp
[ctxn
]);
4852 perf_event_aux_ctx(ctx
, output
, data
);
4854 put_cpu_ptr(pmu
->pmu_cpu_context
);
4859 perf_event_aux_ctx(task_ctx
, output
, data
);
4866 * task tracking -- fork/exit
4868 * enabled by: attr.comm | attr.mmap | attr.mmap2 | attr.mmap_data | attr.task
4871 struct perf_task_event
{
4872 struct task_struct
*task
;
4873 struct perf_event_context
*task_ctx
;
4876 struct perf_event_header header
;
4886 static int perf_event_task_match(struct perf_event
*event
)
4888 return event
->attr
.comm
|| event
->attr
.mmap
||
4889 event
->attr
.mmap2
|| event
->attr
.mmap_data
||
4893 static void perf_event_task_output(struct perf_event
*event
,
4896 struct perf_task_event
*task_event
= data
;
4897 struct perf_output_handle handle
;
4898 struct perf_sample_data sample
;
4899 struct task_struct
*task
= task_event
->task
;
4900 int ret
, size
= task_event
->event_id
.header
.size
;
4902 if (!perf_event_task_match(event
))
4905 perf_event_header__init_id(&task_event
->event_id
.header
, &sample
, event
);
4907 ret
= perf_output_begin(&handle
, event
,
4908 task_event
->event_id
.header
.size
);
4912 task_event
->event_id
.pid
= perf_event_pid(event
, task
);
4913 task_event
->event_id
.ppid
= perf_event_pid(event
, current
);
4915 task_event
->event_id
.tid
= perf_event_tid(event
, task
);
4916 task_event
->event_id
.ptid
= perf_event_tid(event
, current
);
4918 perf_output_put(&handle
, task_event
->event_id
);
4920 perf_event__output_id_sample(event
, &handle
, &sample
);
4922 perf_output_end(&handle
);
4924 task_event
->event_id
.header
.size
= size
;
4927 static void perf_event_task(struct task_struct
*task
,
4928 struct perf_event_context
*task_ctx
,
4931 struct perf_task_event task_event
;
4933 if (!atomic_read(&nr_comm_events
) &&
4934 !atomic_read(&nr_mmap_events
) &&
4935 !atomic_read(&nr_task_events
))
4938 task_event
= (struct perf_task_event
){
4940 .task_ctx
= task_ctx
,
4943 .type
= new ? PERF_RECORD_FORK
: PERF_RECORD_EXIT
,
4945 .size
= sizeof(task_event
.event_id
),
4951 .time
= perf_clock(),
4955 perf_event_aux(perf_event_task_output
,
4960 void perf_event_fork(struct task_struct
*task
)
4962 perf_event_task(task
, NULL
, 1);
4969 struct perf_comm_event
{
4970 struct task_struct
*task
;
4975 struct perf_event_header header
;
4982 static int perf_event_comm_match(struct perf_event
*event
)
4984 return event
->attr
.comm
;
4987 static void perf_event_comm_output(struct perf_event
*event
,
4990 struct perf_comm_event
*comm_event
= data
;
4991 struct perf_output_handle handle
;
4992 struct perf_sample_data sample
;
4993 int size
= comm_event
->event_id
.header
.size
;
4996 if (!perf_event_comm_match(event
))
4999 perf_event_header__init_id(&comm_event
->event_id
.header
, &sample
, event
);
5000 ret
= perf_output_begin(&handle
, event
,
5001 comm_event
->event_id
.header
.size
);
5006 comm_event
->event_id
.pid
= perf_event_pid(event
, comm_event
->task
);
5007 comm_event
->event_id
.tid
= perf_event_tid(event
, comm_event
->task
);
5009 perf_output_put(&handle
, comm_event
->event_id
);
5010 __output_copy(&handle
, comm_event
->comm
,
5011 comm_event
->comm_size
);
5013 perf_event__output_id_sample(event
, &handle
, &sample
);
5015 perf_output_end(&handle
);
5017 comm_event
->event_id
.header
.size
= size
;
5020 static void perf_event_comm_event(struct perf_comm_event
*comm_event
)
5022 char comm
[TASK_COMM_LEN
];
5025 memset(comm
, 0, sizeof(comm
));
5026 strlcpy(comm
, comm_event
->task
->comm
, sizeof(comm
));
5027 size
= ALIGN(strlen(comm
)+1, sizeof(u64
));
5029 comm_event
->comm
= comm
;
5030 comm_event
->comm_size
= size
;
5032 comm_event
->event_id
.header
.size
= sizeof(comm_event
->event_id
) + size
;
5034 perf_event_aux(perf_event_comm_output
,
5039 void perf_event_comm(struct task_struct
*task
)
5041 struct perf_comm_event comm_event
;
5042 struct perf_event_context
*ctx
;
5046 for_each_task_context_nr(ctxn
) {
5047 ctx
= task
->perf_event_ctxp
[ctxn
];
5051 perf_event_enable_on_exec(ctx
);
5055 if (!atomic_read(&nr_comm_events
))
5058 comm_event
= (struct perf_comm_event
){
5064 .type
= PERF_RECORD_COMM
,
5073 perf_event_comm_event(&comm_event
);
5080 struct perf_mmap_event
{
5081 struct vm_area_struct
*vma
;
5083 const char *file_name
;
5090 struct perf_event_header header
;
5100 static int perf_event_mmap_match(struct perf_event
*event
,
5103 struct perf_mmap_event
*mmap_event
= data
;
5104 struct vm_area_struct
*vma
= mmap_event
->vma
;
5105 int executable
= vma
->vm_flags
& VM_EXEC
;
5107 return (!executable
&& event
->attr
.mmap_data
) ||
5108 (executable
&& (event
->attr
.mmap
|| event
->attr
.mmap2
));
5111 static void perf_event_mmap_output(struct perf_event
*event
,
5114 struct perf_mmap_event
*mmap_event
= data
;
5115 struct perf_output_handle handle
;
5116 struct perf_sample_data sample
;
5117 int size
= mmap_event
->event_id
.header
.size
;
5120 if (!perf_event_mmap_match(event
, data
))
5123 if (event
->attr
.mmap2
) {
5124 mmap_event
->event_id
.header
.type
= PERF_RECORD_MMAP2
;
5125 mmap_event
->event_id
.header
.size
+= sizeof(mmap_event
->maj
);
5126 mmap_event
->event_id
.header
.size
+= sizeof(mmap_event
->min
);
5127 mmap_event
->event_id
.header
.size
+= sizeof(mmap_event
->ino
);
5128 mmap_event
->event_id
.header
.size
+= sizeof(mmap_event
->ino_generation
);
5131 perf_event_header__init_id(&mmap_event
->event_id
.header
, &sample
, event
);
5132 ret
= perf_output_begin(&handle
, event
,
5133 mmap_event
->event_id
.header
.size
);
5137 mmap_event
->event_id
.pid
= perf_event_pid(event
, current
);
5138 mmap_event
->event_id
.tid
= perf_event_tid(event
, current
);
5140 perf_output_put(&handle
, mmap_event
->event_id
);
5142 if (event
->attr
.mmap2
) {
5143 perf_output_put(&handle
, mmap_event
->maj
);
5144 perf_output_put(&handle
, mmap_event
->min
);
5145 perf_output_put(&handle
, mmap_event
->ino
);
5146 perf_output_put(&handle
, mmap_event
->ino_generation
);
5149 __output_copy(&handle
, mmap_event
->file_name
,
5150 mmap_event
->file_size
);
5152 perf_event__output_id_sample(event
, &handle
, &sample
);
5154 perf_output_end(&handle
);
5156 mmap_event
->event_id
.header
.size
= size
;
5159 static void perf_event_mmap_event(struct perf_mmap_event
*mmap_event
)
5161 struct vm_area_struct
*vma
= mmap_event
->vma
;
5162 struct file
*file
= vma
->vm_file
;
5163 int maj
= 0, min
= 0;
5164 u64 ino
= 0, gen
= 0;
5171 struct inode
*inode
;
5174 buf
= kmalloc(PATH_MAX
, GFP_KERNEL
);
5180 * d_path() works from the end of the rb backwards, so we
5181 * need to add enough zero bytes after the string to handle
5182 * the 64bit alignment we do later.
5184 name
= d_path(&file
->f_path
, buf
, PATH_MAX
- sizeof(u64
));
5189 inode
= file_inode(vma
->vm_file
);
5190 dev
= inode
->i_sb
->s_dev
;
5192 gen
= inode
->i_generation
;
5197 name
= (char *)arch_vma_name(vma
);
5201 if (vma
->vm_start
<= vma
->vm_mm
->start_brk
&&
5202 vma
->vm_end
>= vma
->vm_mm
->brk
) {
5206 if (vma
->vm_start
<= vma
->vm_mm
->start_stack
&&
5207 vma
->vm_end
>= vma
->vm_mm
->start_stack
) {
5217 strlcpy(tmp
, name
, sizeof(tmp
));
5221 * Since our buffer works in 8 byte units we need to align our string
5222 * size to a multiple of 8. However, we must guarantee the tail end is
5223 * zero'd out to avoid leaking random bits to userspace.
5225 size
= strlen(name
)+1;
5226 while (!IS_ALIGNED(size
, sizeof(u64
)))
5227 name
[size
++] = '\0';
5229 mmap_event
->file_name
= name
;
5230 mmap_event
->file_size
= size
;
5231 mmap_event
->maj
= maj
;
5232 mmap_event
->min
= min
;
5233 mmap_event
->ino
= ino
;
5234 mmap_event
->ino_generation
= gen
;
5236 if (!(vma
->vm_flags
& VM_EXEC
))
5237 mmap_event
->event_id
.header
.misc
|= PERF_RECORD_MISC_MMAP_DATA
;
5239 mmap_event
->event_id
.header
.size
= sizeof(mmap_event
->event_id
) + size
;
5241 perf_event_aux(perf_event_mmap_output
,
5248 void perf_event_mmap(struct vm_area_struct
*vma
)
5250 struct perf_mmap_event mmap_event
;
5252 if (!atomic_read(&nr_mmap_events
))
5255 mmap_event
= (struct perf_mmap_event
){
5261 .type
= PERF_RECORD_MMAP
,
5262 .misc
= PERF_RECORD_MISC_USER
,
5267 .start
= vma
->vm_start
,
5268 .len
= vma
->vm_end
- vma
->vm_start
,
5269 .pgoff
= (u64
)vma
->vm_pgoff
<< PAGE_SHIFT
,
5271 /* .maj (attr_mmap2 only) */
5272 /* .min (attr_mmap2 only) */
5273 /* .ino (attr_mmap2 only) */
5274 /* .ino_generation (attr_mmap2 only) */
5277 perf_event_mmap_event(&mmap_event
);
5281 * IRQ throttle logging
5284 static void perf_log_throttle(struct perf_event
*event
, int enable
)
5286 struct perf_output_handle handle
;
5287 struct perf_sample_data sample
;
5291 struct perf_event_header header
;
5295 } throttle_event
= {
5297 .type
= PERF_RECORD_THROTTLE
,
5299 .size
= sizeof(throttle_event
),
5301 .time
= perf_clock(),
5302 .id
= primary_event_id(event
),
5303 .stream_id
= event
->id
,
5307 throttle_event
.header
.type
= PERF_RECORD_UNTHROTTLE
;
5309 perf_event_header__init_id(&throttle_event
.header
, &sample
, event
);
5311 ret
= perf_output_begin(&handle
, event
,
5312 throttle_event
.header
.size
);
5316 perf_output_put(&handle
, throttle_event
);
5317 perf_event__output_id_sample(event
, &handle
, &sample
);
5318 perf_output_end(&handle
);
5322 * Generic event overflow handling, sampling.
5325 static int __perf_event_overflow(struct perf_event
*event
,
5326 int throttle
, struct perf_sample_data
*data
,
5327 struct pt_regs
*regs
)
5329 int events
= atomic_read(&event
->event_limit
);
5330 struct hw_perf_event
*hwc
= &event
->hw
;
5335 * Non-sampling counters might still use the PMI to fold short
5336 * hardware counters, ignore those.
5338 if (unlikely(!is_sampling_event(event
)))
5341 seq
= __this_cpu_read(perf_throttled_seq
);
5342 if (seq
!= hwc
->interrupts_seq
) {
5343 hwc
->interrupts_seq
= seq
;
5344 hwc
->interrupts
= 1;
5347 if (unlikely(throttle
5348 && hwc
->interrupts
>= max_samples_per_tick
)) {
5349 __this_cpu_inc(perf_throttled_count
);
5350 hwc
->interrupts
= MAX_INTERRUPTS
;
5351 perf_log_throttle(event
, 0);
5352 tick_nohz_full_kick();
5357 if (event
->attr
.freq
) {
5358 u64 now
= perf_clock();
5359 s64 delta
= now
- hwc
->freq_time_stamp
;
5361 hwc
->freq_time_stamp
= now
;
5363 if (delta
> 0 && delta
< 2*TICK_NSEC
)
5364 perf_adjust_period(event
, delta
, hwc
->last_period
, true);
5368 * XXX event_limit might not quite work as expected on inherited
5372 event
->pending_kill
= POLL_IN
;
5373 if (events
&& atomic_dec_and_test(&event
->event_limit
)) {
5375 event
->pending_kill
= POLL_HUP
;
5376 event
->pending_disable
= 1;
5377 irq_work_queue(&event
->pending
);
5380 if (event
->overflow_handler
)
5381 event
->overflow_handler(event
, data
, regs
);
5383 perf_event_output(event
, data
, regs
);
5385 if (event
->fasync
&& event
->pending_kill
) {
5386 event
->pending_wakeup
= 1;
5387 irq_work_queue(&event
->pending
);
5393 int perf_event_overflow(struct perf_event
*event
,
5394 struct perf_sample_data
*data
,
5395 struct pt_regs
*regs
)
5397 return __perf_event_overflow(event
, 1, data
, regs
);
5401 * Generic software event infrastructure
5404 struct swevent_htable
{
5405 struct swevent_hlist
*swevent_hlist
;
5406 struct mutex hlist_mutex
;
5409 /* Recursion avoidance in each contexts */
5410 int recursion
[PERF_NR_CONTEXTS
];
5413 static DEFINE_PER_CPU(struct swevent_htable
, swevent_htable
);
5416 * We directly increment event->count and keep a second value in
5417 * event->hw.period_left to count intervals. This period event
5418 * is kept in the range [-sample_period, 0] so that we can use the
5422 u64
perf_swevent_set_period(struct perf_event
*event
)
5424 struct hw_perf_event
*hwc
= &event
->hw
;
5425 u64 period
= hwc
->last_period
;
5429 hwc
->last_period
= hwc
->sample_period
;
5432 old
= val
= local64_read(&hwc
->period_left
);
5436 nr
= div64_u64(period
+ val
, period
);
5437 offset
= nr
* period
;
5439 if (local64_cmpxchg(&hwc
->period_left
, old
, val
) != old
)
5445 static void perf_swevent_overflow(struct perf_event
*event
, u64 overflow
,
5446 struct perf_sample_data
*data
,
5447 struct pt_regs
*regs
)
5449 struct hw_perf_event
*hwc
= &event
->hw
;
5453 overflow
= perf_swevent_set_period(event
);
5455 if (hwc
->interrupts
== MAX_INTERRUPTS
)
5458 for (; overflow
; overflow
--) {
5459 if (__perf_event_overflow(event
, throttle
,
5462 * We inhibit the overflow from happening when
5463 * hwc->interrupts == MAX_INTERRUPTS.
5471 static void perf_swevent_event(struct perf_event
*event
, u64 nr
,
5472 struct perf_sample_data
*data
,
5473 struct pt_regs
*regs
)
5475 struct hw_perf_event
*hwc
= &event
->hw
;
5477 local64_add(nr
, &event
->count
);
5482 if (!is_sampling_event(event
))
5485 if ((event
->attr
.sample_type
& PERF_SAMPLE_PERIOD
) && !event
->attr
.freq
) {
5487 return perf_swevent_overflow(event
, 1, data
, regs
);
5489 data
->period
= event
->hw
.last_period
;
5491 if (nr
== 1 && hwc
->sample_period
== 1 && !event
->attr
.freq
)
5492 return perf_swevent_overflow(event
, 1, data
, regs
);
5494 if (local64_add_negative(nr
, &hwc
->period_left
))
5497 perf_swevent_overflow(event
, 0, data
, regs
);
5500 static int perf_exclude_event(struct perf_event
*event
,
5501 struct pt_regs
*regs
)
5503 if (event
->hw
.state
& PERF_HES_STOPPED
)
5507 if (event
->attr
.exclude_user
&& user_mode(regs
))
5510 if (event
->attr
.exclude_kernel
&& !user_mode(regs
))
5517 static int perf_swevent_match(struct perf_event
*event
,
5518 enum perf_type_id type
,
5520 struct perf_sample_data
*data
,
5521 struct pt_regs
*regs
)
5523 if (event
->attr
.type
!= type
)
5526 if (event
->attr
.config
!= event_id
)
5529 if (perf_exclude_event(event
, regs
))
5535 static inline u64
swevent_hash(u64 type
, u32 event_id
)
5537 u64 val
= event_id
| (type
<< 32);
5539 return hash_64(val
, SWEVENT_HLIST_BITS
);
5542 static inline struct hlist_head
*
5543 __find_swevent_head(struct swevent_hlist
*hlist
, u64 type
, u32 event_id
)
5545 u64 hash
= swevent_hash(type
, event_id
);
5547 return &hlist
->heads
[hash
];
5550 /* For the read side: events when they trigger */
5551 static inline struct hlist_head
*
5552 find_swevent_head_rcu(struct swevent_htable
*swhash
, u64 type
, u32 event_id
)
5554 struct swevent_hlist
*hlist
;
5556 hlist
= rcu_dereference(swhash
->swevent_hlist
);
5560 return __find_swevent_head(hlist
, type
, event_id
);
5563 /* For the event head insertion and removal in the hlist */
5564 static inline struct hlist_head
*
5565 find_swevent_head(struct swevent_htable
*swhash
, struct perf_event
*event
)
5567 struct swevent_hlist
*hlist
;
5568 u32 event_id
= event
->attr
.config
;
5569 u64 type
= event
->attr
.type
;
5572 * Event scheduling is always serialized against hlist allocation
5573 * and release. Which makes the protected version suitable here.
5574 * The context lock guarantees that.
5576 hlist
= rcu_dereference_protected(swhash
->swevent_hlist
,
5577 lockdep_is_held(&event
->ctx
->lock
));
5581 return __find_swevent_head(hlist
, type
, event_id
);
5584 static void do_perf_sw_event(enum perf_type_id type
, u32 event_id
,
5586 struct perf_sample_data
*data
,
5587 struct pt_regs
*regs
)
5589 struct swevent_htable
*swhash
= &__get_cpu_var(swevent_htable
);
5590 struct perf_event
*event
;
5591 struct hlist_head
*head
;
5594 head
= find_swevent_head_rcu(swhash
, type
, event_id
);
5598 hlist_for_each_entry_rcu(event
, head
, hlist_entry
) {
5599 if (perf_swevent_match(event
, type
, event_id
, data
, regs
))
5600 perf_swevent_event(event
, nr
, data
, regs
);
5606 int perf_swevent_get_recursion_context(void)
5608 struct swevent_htable
*swhash
= &__get_cpu_var(swevent_htable
);
5610 return get_recursion_context(swhash
->recursion
);
5612 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context
);
5614 inline void perf_swevent_put_recursion_context(int rctx
)
5616 struct swevent_htable
*swhash
= &__get_cpu_var(swevent_htable
);
5618 put_recursion_context(swhash
->recursion
, rctx
);
5621 void __perf_sw_event(u32 event_id
, u64 nr
, struct pt_regs
*regs
, u64 addr
)
5623 struct perf_sample_data data
;
5626 preempt_disable_notrace();
5627 rctx
= perf_swevent_get_recursion_context();
5631 perf_sample_data_init(&data
, addr
, 0);
5633 do_perf_sw_event(PERF_TYPE_SOFTWARE
, event_id
, nr
, &data
, regs
);
5635 perf_swevent_put_recursion_context(rctx
);
5636 preempt_enable_notrace();
5639 static void perf_swevent_read(struct perf_event
*event
)
5643 static int perf_swevent_add(struct perf_event
*event
, int flags
)
5645 struct swevent_htable
*swhash
= &__get_cpu_var(swevent_htable
);
5646 struct hw_perf_event
*hwc
= &event
->hw
;
5647 struct hlist_head
*head
;
5649 if (is_sampling_event(event
)) {
5650 hwc
->last_period
= hwc
->sample_period
;
5651 perf_swevent_set_period(event
);
5654 hwc
->state
= !(flags
& PERF_EF_START
);
5656 head
= find_swevent_head(swhash
, event
);
5657 if (WARN_ON_ONCE(!head
))
5660 hlist_add_head_rcu(&event
->hlist_entry
, head
);
5665 static void perf_swevent_del(struct perf_event
*event
, int flags
)
5667 hlist_del_rcu(&event
->hlist_entry
);
5670 static void perf_swevent_start(struct perf_event
*event
, int flags
)
5672 event
->hw
.state
= 0;
5675 static void perf_swevent_stop(struct perf_event
*event
, int flags
)
5677 event
->hw
.state
= PERF_HES_STOPPED
;
5680 /* Deref the hlist from the update side */
5681 static inline struct swevent_hlist
*
5682 swevent_hlist_deref(struct swevent_htable
*swhash
)
5684 return rcu_dereference_protected(swhash
->swevent_hlist
,
5685 lockdep_is_held(&swhash
->hlist_mutex
));
5688 static void swevent_hlist_release(struct swevent_htable
*swhash
)
5690 struct swevent_hlist
*hlist
= swevent_hlist_deref(swhash
);
5695 rcu_assign_pointer(swhash
->swevent_hlist
, NULL
);
5696 kfree_rcu(hlist
, rcu_head
);
5699 static void swevent_hlist_put_cpu(struct perf_event
*event
, int cpu
)
5701 struct swevent_htable
*swhash
= &per_cpu(swevent_htable
, cpu
);
5703 mutex_lock(&swhash
->hlist_mutex
);
5705 if (!--swhash
->hlist_refcount
)
5706 swevent_hlist_release(swhash
);
5708 mutex_unlock(&swhash
->hlist_mutex
);
5711 static void swevent_hlist_put(struct perf_event
*event
)
5715 for_each_possible_cpu(cpu
)
5716 swevent_hlist_put_cpu(event
, cpu
);
5719 static int swevent_hlist_get_cpu(struct perf_event
*event
, int cpu
)
5721 struct swevent_htable
*swhash
= &per_cpu(swevent_htable
, cpu
);
5724 mutex_lock(&swhash
->hlist_mutex
);
5726 if (!swevent_hlist_deref(swhash
) && cpu_online(cpu
)) {
5727 struct swevent_hlist
*hlist
;
5729 hlist
= kzalloc(sizeof(*hlist
), GFP_KERNEL
);
5734 rcu_assign_pointer(swhash
->swevent_hlist
, hlist
);
5736 swhash
->hlist_refcount
++;
5738 mutex_unlock(&swhash
->hlist_mutex
);
5743 static int swevent_hlist_get(struct perf_event
*event
)
5746 int cpu
, failed_cpu
;
5749 for_each_possible_cpu(cpu
) {
5750 err
= swevent_hlist_get_cpu(event
, cpu
);
5760 for_each_possible_cpu(cpu
) {
5761 if (cpu
== failed_cpu
)
5763 swevent_hlist_put_cpu(event
, cpu
);
5770 struct static_key perf_swevent_enabled
[PERF_COUNT_SW_MAX
];
5772 static void sw_perf_event_destroy(struct perf_event
*event
)
5774 u64 event_id
= event
->attr
.config
;
5776 WARN_ON(event
->parent
);
5778 static_key_slow_dec(&perf_swevent_enabled
[event_id
]);
5779 swevent_hlist_put(event
);
5782 static int perf_swevent_init(struct perf_event
*event
)
5784 u64 event_id
= event
->attr
.config
;
5786 if (event
->attr
.type
!= PERF_TYPE_SOFTWARE
)
5790 * no branch sampling for software events
5792 if (has_branch_stack(event
))
5796 case PERF_COUNT_SW_CPU_CLOCK
:
5797 case PERF_COUNT_SW_TASK_CLOCK
:
5804 if (event_id
>= PERF_COUNT_SW_MAX
)
5807 if (!event
->parent
) {
5810 err
= swevent_hlist_get(event
);
5814 static_key_slow_inc(&perf_swevent_enabled
[event_id
]);
5815 event
->destroy
= sw_perf_event_destroy
;
5821 static int perf_swevent_event_idx(struct perf_event
*event
)
5826 static struct pmu perf_swevent
= {
5827 .task_ctx_nr
= perf_sw_context
,
5829 .event_init
= perf_swevent_init
,
5830 .add
= perf_swevent_add
,
5831 .del
= perf_swevent_del
,
5832 .start
= perf_swevent_start
,
5833 .stop
= perf_swevent_stop
,
5834 .read
= perf_swevent_read
,
5836 .event_idx
= perf_swevent_event_idx
,
5839 #ifdef CONFIG_EVENT_TRACING
5841 static int perf_tp_filter_match(struct perf_event
*event
,
5842 struct perf_sample_data
*data
)
5844 void *record
= data
->raw
->data
;
5846 if (likely(!event
->filter
) || filter_match_preds(event
->filter
, record
))
5851 static int perf_tp_event_match(struct perf_event
*event
,
5852 struct perf_sample_data
*data
,
5853 struct pt_regs
*regs
)
5855 if (event
->hw
.state
& PERF_HES_STOPPED
)
5858 * All tracepoints are from kernel-space.
5860 if (event
->attr
.exclude_kernel
)
5863 if (!perf_tp_filter_match(event
, data
))
5869 void perf_tp_event(u64 addr
, u64 count
, void *record
, int entry_size
,
5870 struct pt_regs
*regs
, struct hlist_head
*head
, int rctx
,
5871 struct task_struct
*task
)
5873 struct perf_sample_data data
;
5874 struct perf_event
*event
;
5876 struct perf_raw_record raw
= {
5881 perf_sample_data_init(&data
, addr
, 0);
5884 hlist_for_each_entry_rcu(event
, head
, hlist_entry
) {
5885 if (perf_tp_event_match(event
, &data
, regs
))
5886 perf_swevent_event(event
, count
, &data
, regs
);
5890 * If we got specified a target task, also iterate its context and
5891 * deliver this event there too.
5893 if (task
&& task
!= current
) {
5894 struct perf_event_context
*ctx
;
5895 struct trace_entry
*entry
= record
;
5898 ctx
= rcu_dereference(task
->perf_event_ctxp
[perf_sw_context
]);
5902 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
5903 if (event
->attr
.type
!= PERF_TYPE_TRACEPOINT
)
5905 if (event
->attr
.config
!= entry
->type
)
5907 if (perf_tp_event_match(event
, &data
, regs
))
5908 perf_swevent_event(event
, count
, &data
, regs
);
5914 perf_swevent_put_recursion_context(rctx
);
5916 EXPORT_SYMBOL_GPL(perf_tp_event
);
5918 static void tp_perf_event_destroy(struct perf_event
*event
)
5920 perf_trace_destroy(event
);
5923 static int perf_tp_event_init(struct perf_event
*event
)
5927 if (event
->attr
.type
!= PERF_TYPE_TRACEPOINT
)
5931 * no branch sampling for tracepoint events
5933 if (has_branch_stack(event
))
5936 err
= perf_trace_init(event
);
5940 event
->destroy
= tp_perf_event_destroy
;
5945 static struct pmu perf_tracepoint
= {
5946 .task_ctx_nr
= perf_sw_context
,
5948 .event_init
= perf_tp_event_init
,
5949 .add
= perf_trace_add
,
5950 .del
= perf_trace_del
,
5951 .start
= perf_swevent_start
,
5952 .stop
= perf_swevent_stop
,
5953 .read
= perf_swevent_read
,
5955 .event_idx
= perf_swevent_event_idx
,
5958 static inline void perf_tp_register(void)
5960 perf_pmu_register(&perf_tracepoint
, "tracepoint", PERF_TYPE_TRACEPOINT
);
5963 static int perf_event_set_filter(struct perf_event
*event
, void __user
*arg
)
5968 if (event
->attr
.type
!= PERF_TYPE_TRACEPOINT
)
5971 filter_str
= strndup_user(arg
, PAGE_SIZE
);
5972 if (IS_ERR(filter_str
))
5973 return PTR_ERR(filter_str
);
5975 ret
= ftrace_profile_set_filter(event
, event
->attr
.config
, filter_str
);
5981 static void perf_event_free_filter(struct perf_event
*event
)
5983 ftrace_profile_free_filter(event
);
5988 static inline void perf_tp_register(void)
5992 static int perf_event_set_filter(struct perf_event
*event
, void __user
*arg
)
5997 static void perf_event_free_filter(struct perf_event
*event
)
6001 #endif /* CONFIG_EVENT_TRACING */
6003 #ifdef CONFIG_HAVE_HW_BREAKPOINT
6004 void perf_bp_event(struct perf_event
*bp
, void *data
)
6006 struct perf_sample_data sample
;
6007 struct pt_regs
*regs
= data
;
6009 perf_sample_data_init(&sample
, bp
->attr
.bp_addr
, 0);
6011 if (!bp
->hw
.state
&& !perf_exclude_event(bp
, regs
))
6012 perf_swevent_event(bp
, 1, &sample
, regs
);
6017 * hrtimer based swevent callback
6020 static enum hrtimer_restart
perf_swevent_hrtimer(struct hrtimer
*hrtimer
)
6022 enum hrtimer_restart ret
= HRTIMER_RESTART
;
6023 struct perf_sample_data data
;
6024 struct pt_regs
*regs
;
6025 struct perf_event
*event
;
6028 event
= container_of(hrtimer
, struct perf_event
, hw
.hrtimer
);
6030 if (event
->state
!= PERF_EVENT_STATE_ACTIVE
)
6031 return HRTIMER_NORESTART
;
6033 event
->pmu
->read(event
);
6035 perf_sample_data_init(&data
, 0, event
->hw
.last_period
);
6036 regs
= get_irq_regs();
6038 if (regs
&& !perf_exclude_event(event
, regs
)) {
6039 if (!(event
->attr
.exclude_idle
&& is_idle_task(current
)))
6040 if (__perf_event_overflow(event
, 1, &data
, regs
))
6041 ret
= HRTIMER_NORESTART
;
6044 period
= max_t(u64
, 10000, event
->hw
.sample_period
);
6045 hrtimer_forward_now(hrtimer
, ns_to_ktime(period
));
6050 static void perf_swevent_start_hrtimer(struct perf_event
*event
)
6052 struct hw_perf_event
*hwc
= &event
->hw
;
6055 if (!is_sampling_event(event
))
6058 period
= local64_read(&hwc
->period_left
);
6063 local64_set(&hwc
->period_left
, 0);
6065 period
= max_t(u64
, 10000, hwc
->sample_period
);
6067 __hrtimer_start_range_ns(&hwc
->hrtimer
,
6068 ns_to_ktime(period
), 0,
6069 HRTIMER_MODE_REL_PINNED
, 0);
6072 static void perf_swevent_cancel_hrtimer(struct perf_event
*event
)
6074 struct hw_perf_event
*hwc
= &event
->hw
;
6076 if (is_sampling_event(event
)) {
6077 ktime_t remaining
= hrtimer_get_remaining(&hwc
->hrtimer
);
6078 local64_set(&hwc
->period_left
, ktime_to_ns(remaining
));
6080 hrtimer_cancel(&hwc
->hrtimer
);
6084 static void perf_swevent_init_hrtimer(struct perf_event
*event
)
6086 struct hw_perf_event
*hwc
= &event
->hw
;
6088 if (!is_sampling_event(event
))
6091 hrtimer_init(&hwc
->hrtimer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
6092 hwc
->hrtimer
.function
= perf_swevent_hrtimer
;
6095 * Since hrtimers have a fixed rate, we can do a static freq->period
6096 * mapping and avoid the whole period adjust feedback stuff.
6098 if (event
->attr
.freq
) {
6099 long freq
= event
->attr
.sample_freq
;
6101 event
->attr
.sample_period
= NSEC_PER_SEC
/ freq
;
6102 hwc
->sample_period
= event
->attr
.sample_period
;
6103 local64_set(&hwc
->period_left
, hwc
->sample_period
);
6104 hwc
->last_period
= hwc
->sample_period
;
6105 event
->attr
.freq
= 0;
6110 * Software event: cpu wall time clock
6113 static void cpu_clock_event_update(struct perf_event
*event
)
6118 now
= local_clock();
6119 prev
= local64_xchg(&event
->hw
.prev_count
, now
);
6120 local64_add(now
- prev
, &event
->count
);
6123 static void cpu_clock_event_start(struct perf_event
*event
, int flags
)
6125 local64_set(&event
->hw
.prev_count
, local_clock());
6126 perf_swevent_start_hrtimer(event
);
6129 static void cpu_clock_event_stop(struct perf_event
*event
, int flags
)
6131 perf_swevent_cancel_hrtimer(event
);
6132 cpu_clock_event_update(event
);
6135 static int cpu_clock_event_add(struct perf_event
*event
, int flags
)
6137 if (flags
& PERF_EF_START
)
6138 cpu_clock_event_start(event
, flags
);
6143 static void cpu_clock_event_del(struct perf_event
*event
, int flags
)
6145 cpu_clock_event_stop(event
, flags
);
6148 static void cpu_clock_event_read(struct perf_event
*event
)
6150 cpu_clock_event_update(event
);
6153 static int cpu_clock_event_init(struct perf_event
*event
)
6155 if (event
->attr
.type
!= PERF_TYPE_SOFTWARE
)
6158 if (event
->attr
.config
!= PERF_COUNT_SW_CPU_CLOCK
)
6162 * no branch sampling for software events
6164 if (has_branch_stack(event
))
6167 perf_swevent_init_hrtimer(event
);
6172 static struct pmu perf_cpu_clock
= {
6173 .task_ctx_nr
= perf_sw_context
,
6175 .event_init
= cpu_clock_event_init
,
6176 .add
= cpu_clock_event_add
,
6177 .del
= cpu_clock_event_del
,
6178 .start
= cpu_clock_event_start
,
6179 .stop
= cpu_clock_event_stop
,
6180 .read
= cpu_clock_event_read
,
6182 .event_idx
= perf_swevent_event_idx
,
6186 * Software event: task time clock
6189 static void task_clock_event_update(struct perf_event
*event
, u64 now
)
6194 prev
= local64_xchg(&event
->hw
.prev_count
, now
);
6196 local64_add(delta
, &event
->count
);
6199 static void task_clock_event_start(struct perf_event
*event
, int flags
)
6201 local64_set(&event
->hw
.prev_count
, event
->ctx
->time
);
6202 perf_swevent_start_hrtimer(event
);
6205 static void task_clock_event_stop(struct perf_event
*event
, int flags
)
6207 perf_swevent_cancel_hrtimer(event
);
6208 task_clock_event_update(event
, event
->ctx
->time
);
6211 static int task_clock_event_add(struct perf_event
*event
, int flags
)
6213 if (flags
& PERF_EF_START
)
6214 task_clock_event_start(event
, flags
);
6219 static void task_clock_event_del(struct perf_event
*event
, int flags
)
6221 task_clock_event_stop(event
, PERF_EF_UPDATE
);
6224 static void task_clock_event_read(struct perf_event
*event
)
6226 u64 now
= perf_clock();
6227 u64 delta
= now
- event
->ctx
->timestamp
;
6228 u64 time
= event
->ctx
->time
+ delta
;
6230 task_clock_event_update(event
, time
);
6233 static int task_clock_event_init(struct perf_event
*event
)
6235 if (event
->attr
.type
!= PERF_TYPE_SOFTWARE
)
6238 if (event
->attr
.config
!= PERF_COUNT_SW_TASK_CLOCK
)
6242 * no branch sampling for software events
6244 if (has_branch_stack(event
))
6247 perf_swevent_init_hrtimer(event
);
6252 static struct pmu perf_task_clock
= {
6253 .task_ctx_nr
= perf_sw_context
,
6255 .event_init
= task_clock_event_init
,
6256 .add
= task_clock_event_add
,
6257 .del
= task_clock_event_del
,
6258 .start
= task_clock_event_start
,
6259 .stop
= task_clock_event_stop
,
6260 .read
= task_clock_event_read
,
6262 .event_idx
= perf_swevent_event_idx
,
6265 static void perf_pmu_nop_void(struct pmu
*pmu
)
6269 static int perf_pmu_nop_int(struct pmu
*pmu
)
6274 static void perf_pmu_start_txn(struct pmu
*pmu
)
6276 perf_pmu_disable(pmu
);
6279 static int perf_pmu_commit_txn(struct pmu
*pmu
)
6281 perf_pmu_enable(pmu
);
6285 static void perf_pmu_cancel_txn(struct pmu
*pmu
)
6287 perf_pmu_enable(pmu
);
6290 static int perf_event_idx_default(struct perf_event
*event
)
6292 return event
->hw
.idx
+ 1;
6296 * Ensures all contexts with the same task_ctx_nr have the same
6297 * pmu_cpu_context too.
6299 static struct perf_cpu_context __percpu
*find_pmu_context(int ctxn
)
6306 list_for_each_entry(pmu
, &pmus
, entry
) {
6307 if (pmu
->task_ctx_nr
== ctxn
)
6308 return pmu
->pmu_cpu_context
;
6314 static void update_pmu_context(struct pmu
*pmu
, struct pmu
*old_pmu
)
6318 for_each_possible_cpu(cpu
) {
6319 struct perf_cpu_context
*cpuctx
;
6321 cpuctx
= per_cpu_ptr(pmu
->pmu_cpu_context
, cpu
);
6323 if (cpuctx
->unique_pmu
== old_pmu
)
6324 cpuctx
->unique_pmu
= pmu
;
6328 static void free_pmu_context(struct pmu
*pmu
)
6332 mutex_lock(&pmus_lock
);
6334 * Like a real lame refcount.
6336 list_for_each_entry(i
, &pmus
, entry
) {
6337 if (i
->pmu_cpu_context
== pmu
->pmu_cpu_context
) {
6338 update_pmu_context(i
, pmu
);
6343 free_percpu(pmu
->pmu_cpu_context
);
6345 mutex_unlock(&pmus_lock
);
6347 static struct idr pmu_idr
;
6350 type_show(struct device
*dev
, struct device_attribute
*attr
, char *page
)
6352 struct pmu
*pmu
= dev_get_drvdata(dev
);
6354 return snprintf(page
, PAGE_SIZE
-1, "%d\n", pmu
->type
);
6356 static DEVICE_ATTR_RO(type
);
6359 perf_event_mux_interval_ms_show(struct device
*dev
,
6360 struct device_attribute
*attr
,
6363 struct pmu
*pmu
= dev_get_drvdata(dev
);
6365 return snprintf(page
, PAGE_SIZE
-1, "%d\n", pmu
->hrtimer_interval_ms
);
6369 perf_event_mux_interval_ms_store(struct device
*dev
,
6370 struct device_attribute
*attr
,
6371 const char *buf
, size_t count
)
6373 struct pmu
*pmu
= dev_get_drvdata(dev
);
6374 int timer
, cpu
, ret
;
6376 ret
= kstrtoint(buf
, 0, &timer
);
6383 /* same value, noting to do */
6384 if (timer
== pmu
->hrtimer_interval_ms
)
6387 pmu
->hrtimer_interval_ms
= timer
;
6389 /* update all cpuctx for this PMU */
6390 for_each_possible_cpu(cpu
) {
6391 struct perf_cpu_context
*cpuctx
;
6392 cpuctx
= per_cpu_ptr(pmu
->pmu_cpu_context
, cpu
);
6393 cpuctx
->hrtimer_interval
= ns_to_ktime(NSEC_PER_MSEC
* timer
);
6395 if (hrtimer_active(&cpuctx
->hrtimer
))
6396 hrtimer_forward_now(&cpuctx
->hrtimer
, cpuctx
->hrtimer_interval
);
6401 static DEVICE_ATTR_RW(perf_event_mux_interval_ms
);
6403 static struct attribute
*pmu_dev_attrs
[] = {
6404 &dev_attr_type
.attr
,
6405 &dev_attr_perf_event_mux_interval_ms
.attr
,
6408 ATTRIBUTE_GROUPS(pmu_dev
);
6410 static int pmu_bus_running
;
6411 static struct bus_type pmu_bus
= {
6412 .name
= "event_source",
6413 .dev_groups
= pmu_dev_groups
,
6416 static void pmu_dev_release(struct device
*dev
)
6421 static int pmu_dev_alloc(struct pmu
*pmu
)
6425 pmu
->dev
= kzalloc(sizeof(struct device
), GFP_KERNEL
);
6429 pmu
->dev
->groups
= pmu
->attr_groups
;
6430 device_initialize(pmu
->dev
);
6431 ret
= dev_set_name(pmu
->dev
, "%s", pmu
->name
);
6435 dev_set_drvdata(pmu
->dev
, pmu
);
6436 pmu
->dev
->bus
= &pmu_bus
;
6437 pmu
->dev
->release
= pmu_dev_release
;
6438 ret
= device_add(pmu
->dev
);
6446 put_device(pmu
->dev
);
6450 static struct lock_class_key cpuctx_mutex
;
6451 static struct lock_class_key cpuctx_lock
;
6453 int perf_pmu_register(struct pmu
*pmu
, const char *name
, int type
)
6457 mutex_lock(&pmus_lock
);
6459 pmu
->pmu_disable_count
= alloc_percpu(int);
6460 if (!pmu
->pmu_disable_count
)
6469 type
= idr_alloc(&pmu_idr
, pmu
, PERF_TYPE_MAX
, 0, GFP_KERNEL
);
6477 if (pmu_bus_running
) {
6478 ret
= pmu_dev_alloc(pmu
);
6484 pmu
->pmu_cpu_context
= find_pmu_context(pmu
->task_ctx_nr
);
6485 if (pmu
->pmu_cpu_context
)
6486 goto got_cpu_context
;
6489 pmu
->pmu_cpu_context
= alloc_percpu(struct perf_cpu_context
);
6490 if (!pmu
->pmu_cpu_context
)
6493 for_each_possible_cpu(cpu
) {
6494 struct perf_cpu_context
*cpuctx
;
6496 cpuctx
= per_cpu_ptr(pmu
->pmu_cpu_context
, cpu
);
6497 __perf_event_init_context(&cpuctx
->ctx
);
6498 lockdep_set_class(&cpuctx
->ctx
.mutex
, &cpuctx_mutex
);
6499 lockdep_set_class(&cpuctx
->ctx
.lock
, &cpuctx_lock
);
6500 cpuctx
->ctx
.type
= cpu_context
;
6501 cpuctx
->ctx
.pmu
= pmu
;
6503 __perf_cpu_hrtimer_init(cpuctx
, cpu
);
6505 INIT_LIST_HEAD(&cpuctx
->rotation_list
);
6506 cpuctx
->unique_pmu
= pmu
;
6510 if (!pmu
->start_txn
) {
6511 if (pmu
->pmu_enable
) {
6513 * If we have pmu_enable/pmu_disable calls, install
6514 * transaction stubs that use that to try and batch
6515 * hardware accesses.
6517 pmu
->start_txn
= perf_pmu_start_txn
;
6518 pmu
->commit_txn
= perf_pmu_commit_txn
;
6519 pmu
->cancel_txn
= perf_pmu_cancel_txn
;
6521 pmu
->start_txn
= perf_pmu_nop_void
;
6522 pmu
->commit_txn
= perf_pmu_nop_int
;
6523 pmu
->cancel_txn
= perf_pmu_nop_void
;
6527 if (!pmu
->pmu_enable
) {
6528 pmu
->pmu_enable
= perf_pmu_nop_void
;
6529 pmu
->pmu_disable
= perf_pmu_nop_void
;
6532 if (!pmu
->event_idx
)
6533 pmu
->event_idx
= perf_event_idx_default
;
6535 list_add_rcu(&pmu
->entry
, &pmus
);
6538 mutex_unlock(&pmus_lock
);
6543 device_del(pmu
->dev
);
6544 put_device(pmu
->dev
);
6547 if (pmu
->type
>= PERF_TYPE_MAX
)
6548 idr_remove(&pmu_idr
, pmu
->type
);
6551 free_percpu(pmu
->pmu_disable_count
);
6555 void perf_pmu_unregister(struct pmu
*pmu
)
6557 mutex_lock(&pmus_lock
);
6558 list_del_rcu(&pmu
->entry
);
6559 mutex_unlock(&pmus_lock
);
6562 * We dereference the pmu list under both SRCU and regular RCU, so
6563 * synchronize against both of those.
6565 synchronize_srcu(&pmus_srcu
);
6568 free_percpu(pmu
->pmu_disable_count
);
6569 if (pmu
->type
>= PERF_TYPE_MAX
)
6570 idr_remove(&pmu_idr
, pmu
->type
);
6571 device_del(pmu
->dev
);
6572 put_device(pmu
->dev
);
6573 free_pmu_context(pmu
);
6576 struct pmu
*perf_init_event(struct perf_event
*event
)
6578 struct pmu
*pmu
= NULL
;
6582 idx
= srcu_read_lock(&pmus_srcu
);
6585 pmu
= idr_find(&pmu_idr
, event
->attr
.type
);
6589 ret
= pmu
->event_init(event
);
6595 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
6597 ret
= pmu
->event_init(event
);
6601 if (ret
!= -ENOENT
) {
6606 pmu
= ERR_PTR(-ENOENT
);
6608 srcu_read_unlock(&pmus_srcu
, idx
);
6613 static void account_event_cpu(struct perf_event
*event
, int cpu
)
6618 if (has_branch_stack(event
)) {
6619 if (!(event
->attach_state
& PERF_ATTACH_TASK
))
6620 atomic_inc(&per_cpu(perf_branch_stack_events
, cpu
));
6622 if (is_cgroup_event(event
))
6623 atomic_inc(&per_cpu(perf_cgroup_events
, cpu
));
6626 static void account_event(struct perf_event
*event
)
6631 if (event
->attach_state
& PERF_ATTACH_TASK
)
6632 static_key_slow_inc(&perf_sched_events
.key
);
6633 if (event
->attr
.mmap
|| event
->attr
.mmap_data
)
6634 atomic_inc(&nr_mmap_events
);
6635 if (event
->attr
.comm
)
6636 atomic_inc(&nr_comm_events
);
6637 if (event
->attr
.task
)
6638 atomic_inc(&nr_task_events
);
6639 if (event
->attr
.freq
) {
6640 if (atomic_inc_return(&nr_freq_events
) == 1)
6641 tick_nohz_full_kick_all();
6643 if (has_branch_stack(event
))
6644 static_key_slow_inc(&perf_sched_events
.key
);
6645 if (is_cgroup_event(event
))
6646 static_key_slow_inc(&perf_sched_events
.key
);
6648 account_event_cpu(event
, event
->cpu
);
6652 * Allocate and initialize a event structure
6654 static struct perf_event
*
6655 perf_event_alloc(struct perf_event_attr
*attr
, int cpu
,
6656 struct task_struct
*task
,
6657 struct perf_event
*group_leader
,
6658 struct perf_event
*parent_event
,
6659 perf_overflow_handler_t overflow_handler
,
6663 struct perf_event
*event
;
6664 struct hw_perf_event
*hwc
;
6667 if ((unsigned)cpu
>= nr_cpu_ids
) {
6668 if (!task
|| cpu
!= -1)
6669 return ERR_PTR(-EINVAL
);
6672 event
= kzalloc(sizeof(*event
), GFP_KERNEL
);
6674 return ERR_PTR(-ENOMEM
);
6677 * Single events are their own group leaders, with an
6678 * empty sibling list:
6681 group_leader
= event
;
6683 mutex_init(&event
->child_mutex
);
6684 INIT_LIST_HEAD(&event
->child_list
);
6686 INIT_LIST_HEAD(&event
->group_entry
);
6687 INIT_LIST_HEAD(&event
->event_entry
);
6688 INIT_LIST_HEAD(&event
->sibling_list
);
6689 INIT_LIST_HEAD(&event
->rb_entry
);
6690 INIT_LIST_HEAD(&event
->active_entry
);
6691 INIT_HLIST_NODE(&event
->hlist_entry
);
6694 init_waitqueue_head(&event
->waitq
);
6695 init_irq_work(&event
->pending
, perf_pending_event
);
6697 mutex_init(&event
->mmap_mutex
);
6699 atomic_long_set(&event
->refcount
, 1);
6701 event
->attr
= *attr
;
6702 event
->group_leader
= group_leader
;
6706 event
->parent
= parent_event
;
6708 event
->ns
= get_pid_ns(task_active_pid_ns(current
));
6709 event
->id
= atomic64_inc_return(&perf_event_id
);
6711 event
->state
= PERF_EVENT_STATE_INACTIVE
;
6714 event
->attach_state
= PERF_ATTACH_TASK
;
6716 if (attr
->type
== PERF_TYPE_TRACEPOINT
)
6717 event
->hw
.tp_target
= task
;
6718 #ifdef CONFIG_HAVE_HW_BREAKPOINT
6720 * hw_breakpoint is a bit difficult here..
6722 else if (attr
->type
== PERF_TYPE_BREAKPOINT
)
6723 event
->hw
.bp_target
= task
;
6727 if (!overflow_handler
&& parent_event
) {
6728 overflow_handler
= parent_event
->overflow_handler
;
6729 context
= parent_event
->overflow_handler_context
;
6732 event
->overflow_handler
= overflow_handler
;
6733 event
->overflow_handler_context
= context
;
6735 perf_event__state_init(event
);
6740 hwc
->sample_period
= attr
->sample_period
;
6741 if (attr
->freq
&& attr
->sample_freq
)
6742 hwc
->sample_period
= 1;
6743 hwc
->last_period
= hwc
->sample_period
;
6745 local64_set(&hwc
->period_left
, hwc
->sample_period
);
6748 * we currently do not support PERF_FORMAT_GROUP on inherited events
6750 if (attr
->inherit
&& (attr
->read_format
& PERF_FORMAT_GROUP
))
6753 pmu
= perf_init_event(event
);
6756 else if (IS_ERR(pmu
)) {
6761 if (!event
->parent
) {
6762 if (event
->attr
.sample_type
& PERF_SAMPLE_CALLCHAIN
) {
6763 err
= get_callchain_buffers();
6773 event
->destroy(event
);
6776 put_pid_ns(event
->ns
);
6779 return ERR_PTR(err
);
6782 static int perf_copy_attr(struct perf_event_attr __user
*uattr
,
6783 struct perf_event_attr
*attr
)
6788 if (!access_ok(VERIFY_WRITE
, uattr
, PERF_ATTR_SIZE_VER0
))
6792 * zero the full structure, so that a short copy will be nice.
6794 memset(attr
, 0, sizeof(*attr
));
6796 ret
= get_user(size
, &uattr
->size
);
6800 if (size
> PAGE_SIZE
) /* silly large */
6803 if (!size
) /* abi compat */
6804 size
= PERF_ATTR_SIZE_VER0
;
6806 if (size
< PERF_ATTR_SIZE_VER0
)
6810 * If we're handed a bigger struct than we know of,
6811 * ensure all the unknown bits are 0 - i.e. new
6812 * user-space does not rely on any kernel feature
6813 * extensions we dont know about yet.
6815 if (size
> sizeof(*attr
)) {
6816 unsigned char __user
*addr
;
6817 unsigned char __user
*end
;
6820 addr
= (void __user
*)uattr
+ sizeof(*attr
);
6821 end
= (void __user
*)uattr
+ size
;
6823 for (; addr
< end
; addr
++) {
6824 ret
= get_user(val
, addr
);
6830 size
= sizeof(*attr
);
6833 ret
= copy_from_user(attr
, uattr
, size
);
6837 /* disabled for now */
6841 if (attr
->__reserved_1
)
6844 if (attr
->sample_type
& ~(PERF_SAMPLE_MAX
-1))
6847 if (attr
->read_format
& ~(PERF_FORMAT_MAX
-1))
6850 if (attr
->sample_type
& PERF_SAMPLE_BRANCH_STACK
) {
6851 u64 mask
= attr
->branch_sample_type
;
6853 /* only using defined bits */
6854 if (mask
& ~(PERF_SAMPLE_BRANCH_MAX
-1))
6857 /* at least one branch bit must be set */
6858 if (!(mask
& ~PERF_SAMPLE_BRANCH_PLM_ALL
))
6861 /* propagate priv level, when not set for branch */
6862 if (!(mask
& PERF_SAMPLE_BRANCH_PLM_ALL
)) {
6864 /* exclude_kernel checked on syscall entry */
6865 if (!attr
->exclude_kernel
)
6866 mask
|= PERF_SAMPLE_BRANCH_KERNEL
;
6868 if (!attr
->exclude_user
)
6869 mask
|= PERF_SAMPLE_BRANCH_USER
;
6871 if (!attr
->exclude_hv
)
6872 mask
|= PERF_SAMPLE_BRANCH_HV
;
6874 * adjust user setting (for HW filter setup)
6876 attr
->branch_sample_type
= mask
;
6878 /* privileged levels capture (kernel, hv): check permissions */
6879 if ((mask
& PERF_SAMPLE_BRANCH_PERM_PLM
)
6880 && perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN
))
6884 if (attr
->sample_type
& PERF_SAMPLE_REGS_USER
) {
6885 ret
= perf_reg_validate(attr
->sample_regs_user
);
6890 if (attr
->sample_type
& PERF_SAMPLE_STACK_USER
) {
6891 if (!arch_perf_have_user_stack_dump())
6895 * We have __u32 type for the size, but so far
6896 * we can only use __u16 as maximum due to the
6897 * __u16 sample size limit.
6899 if (attr
->sample_stack_user
>= USHRT_MAX
)
6901 else if (!IS_ALIGNED(attr
->sample_stack_user
, sizeof(u64
)))
6909 put_user(sizeof(*attr
), &uattr
->size
);
6915 perf_event_set_output(struct perf_event
*event
, struct perf_event
*output_event
)
6917 struct ring_buffer
*rb
= NULL
, *old_rb
= NULL
;
6923 /* don't allow circular references */
6924 if (event
== output_event
)
6928 * Don't allow cross-cpu buffers
6930 if (output_event
->cpu
!= event
->cpu
)
6934 * If its not a per-cpu rb, it must be the same task.
6936 if (output_event
->cpu
== -1 && output_event
->ctx
!= event
->ctx
)
6940 mutex_lock(&event
->mmap_mutex
);
6941 /* Can't redirect output if we've got an active mmap() */
6942 if (atomic_read(&event
->mmap_count
))
6948 /* get the rb we want to redirect to */
6949 rb
= ring_buffer_get(output_event
);
6955 ring_buffer_detach(event
, old_rb
);
6958 ring_buffer_attach(event
, rb
);
6960 rcu_assign_pointer(event
->rb
, rb
);
6963 ring_buffer_put(old_rb
);
6965 * Since we detached before setting the new rb, so that we
6966 * could attach the new rb, we could have missed a wakeup.
6969 wake_up_all(&event
->waitq
);
6974 mutex_unlock(&event
->mmap_mutex
);
6981 * sys_perf_event_open - open a performance event, associate it to a task/cpu
6983 * @attr_uptr: event_id type attributes for monitoring/sampling
6986 * @group_fd: group leader event fd
6988 SYSCALL_DEFINE5(perf_event_open
,
6989 struct perf_event_attr __user
*, attr_uptr
,
6990 pid_t
, pid
, int, cpu
, int, group_fd
, unsigned long, flags
)
6992 struct perf_event
*group_leader
= NULL
, *output_event
= NULL
;
6993 struct perf_event
*event
, *sibling
;
6994 struct perf_event_attr attr
;
6995 struct perf_event_context
*ctx
;
6996 struct file
*event_file
= NULL
;
6997 struct fd group
= {NULL
, 0};
6998 struct task_struct
*task
= NULL
;
7003 int f_flags
= O_RDWR
;
7005 /* for future expandability... */
7006 if (flags
& ~PERF_FLAG_ALL
)
7009 err
= perf_copy_attr(attr_uptr
, &attr
);
7013 if (!attr
.exclude_kernel
) {
7014 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN
))
7019 if (attr
.sample_freq
> sysctl_perf_event_sample_rate
)
7024 * In cgroup mode, the pid argument is used to pass the fd
7025 * opened to the cgroup directory in cgroupfs. The cpu argument
7026 * designates the cpu on which to monitor threads from that
7029 if ((flags
& PERF_FLAG_PID_CGROUP
) && (pid
== -1 || cpu
== -1))
7032 if (flags
& PERF_FLAG_FD_CLOEXEC
)
7033 f_flags
|= O_CLOEXEC
;
7035 event_fd
= get_unused_fd_flags(f_flags
);
7039 if (group_fd
!= -1) {
7040 err
= perf_fget_light(group_fd
, &group
);
7043 group_leader
= group
.file
->private_data
;
7044 if (flags
& PERF_FLAG_FD_OUTPUT
)
7045 output_event
= group_leader
;
7046 if (flags
& PERF_FLAG_FD_NO_GROUP
)
7047 group_leader
= NULL
;
7050 if (pid
!= -1 && !(flags
& PERF_FLAG_PID_CGROUP
)) {
7051 task
= find_lively_task_by_vpid(pid
);
7053 err
= PTR_ERR(task
);
7060 event
= perf_event_alloc(&attr
, cpu
, task
, group_leader
, NULL
,
7062 if (IS_ERR(event
)) {
7063 err
= PTR_ERR(event
);
7067 if (flags
& PERF_FLAG_PID_CGROUP
) {
7068 err
= perf_cgroup_connect(pid
, event
, &attr
, group_leader
);
7070 __free_event(event
);
7075 account_event(event
);
7078 * Special case software events and allow them to be part of
7079 * any hardware group.
7084 (is_software_event(event
) != is_software_event(group_leader
))) {
7085 if (is_software_event(event
)) {
7087 * If event and group_leader are not both a software
7088 * event, and event is, then group leader is not.
7090 * Allow the addition of software events to !software
7091 * groups, this is safe because software events never
7094 pmu
= group_leader
->pmu
;
7095 } else if (is_software_event(group_leader
) &&
7096 (group_leader
->group_flags
& PERF_GROUP_SOFTWARE
)) {
7098 * In case the group is a pure software group, and we
7099 * try to add a hardware event, move the whole group to
7100 * the hardware context.
7107 * Get the target context (task or percpu):
7109 ctx
= find_get_context(pmu
, task
, event
->cpu
);
7116 put_task_struct(task
);
7121 * Look up the group leader (we will attach this event to it):
7127 * Do not allow a recursive hierarchy (this new sibling
7128 * becoming part of another group-sibling):
7130 if (group_leader
->group_leader
!= group_leader
)
7133 * Do not allow to attach to a group in a different
7134 * task or CPU context:
7137 if (group_leader
->ctx
->type
!= ctx
->type
)
7140 if (group_leader
->ctx
!= ctx
)
7145 * Only a group leader can be exclusive or pinned
7147 if (attr
.exclusive
|| attr
.pinned
)
7152 err
= perf_event_set_output(event
, output_event
);
7157 event_file
= anon_inode_getfile("[perf_event]", &perf_fops
, event
,
7159 if (IS_ERR(event_file
)) {
7160 err
= PTR_ERR(event_file
);
7165 struct perf_event_context
*gctx
= group_leader
->ctx
;
7167 mutex_lock(&gctx
->mutex
);
7168 perf_remove_from_context(group_leader
);
7171 * Removing from the context ends up with disabled
7172 * event. What we want here is event in the initial
7173 * startup state, ready to be add into new context.
7175 perf_event__state_init(group_leader
);
7176 list_for_each_entry(sibling
, &group_leader
->sibling_list
,
7178 perf_remove_from_context(sibling
);
7179 perf_event__state_init(sibling
);
7182 mutex_unlock(&gctx
->mutex
);
7186 WARN_ON_ONCE(ctx
->parent_ctx
);
7187 mutex_lock(&ctx
->mutex
);
7191 perf_install_in_context(ctx
, group_leader
, event
->cpu
);
7193 list_for_each_entry(sibling
, &group_leader
->sibling_list
,
7195 perf_install_in_context(ctx
, sibling
, event
->cpu
);
7200 perf_install_in_context(ctx
, event
, event
->cpu
);
7201 perf_unpin_context(ctx
);
7202 mutex_unlock(&ctx
->mutex
);
7206 event
->owner
= current
;
7208 mutex_lock(¤t
->perf_event_mutex
);
7209 list_add_tail(&event
->owner_entry
, ¤t
->perf_event_list
);
7210 mutex_unlock(¤t
->perf_event_mutex
);
7213 * Precalculate sample_data sizes
7215 perf_event__header_size(event
);
7216 perf_event__id_header_size(event
);
7219 * Drop the reference on the group_event after placing the
7220 * new event on the sibling_list. This ensures destruction
7221 * of the group leader will find the pointer to itself in
7222 * perf_group_detach().
7225 fd_install(event_fd
, event_file
);
7229 perf_unpin_context(ctx
);
7236 put_task_struct(task
);
7240 put_unused_fd(event_fd
);
7245 * perf_event_create_kernel_counter
7247 * @attr: attributes of the counter to create
7248 * @cpu: cpu in which the counter is bound
7249 * @task: task to profile (NULL for percpu)
7252 perf_event_create_kernel_counter(struct perf_event_attr
*attr
, int cpu
,
7253 struct task_struct
*task
,
7254 perf_overflow_handler_t overflow_handler
,
7257 struct perf_event_context
*ctx
;
7258 struct perf_event
*event
;
7262 * Get the target context (task or percpu):
7265 event
= perf_event_alloc(attr
, cpu
, task
, NULL
, NULL
,
7266 overflow_handler
, context
);
7267 if (IS_ERR(event
)) {
7268 err
= PTR_ERR(event
);
7272 account_event(event
);
7274 ctx
= find_get_context(event
->pmu
, task
, cpu
);
7280 WARN_ON_ONCE(ctx
->parent_ctx
);
7281 mutex_lock(&ctx
->mutex
);
7282 perf_install_in_context(ctx
, event
, cpu
);
7283 perf_unpin_context(ctx
);
7284 mutex_unlock(&ctx
->mutex
);
7291 return ERR_PTR(err
);
7293 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter
);
7295 void perf_pmu_migrate_context(struct pmu
*pmu
, int src_cpu
, int dst_cpu
)
7297 struct perf_event_context
*src_ctx
;
7298 struct perf_event_context
*dst_ctx
;
7299 struct perf_event
*event
, *tmp
;
7302 src_ctx
= &per_cpu_ptr(pmu
->pmu_cpu_context
, src_cpu
)->ctx
;
7303 dst_ctx
= &per_cpu_ptr(pmu
->pmu_cpu_context
, dst_cpu
)->ctx
;
7305 mutex_lock(&src_ctx
->mutex
);
7306 list_for_each_entry_safe(event
, tmp
, &src_ctx
->event_list
,
7308 perf_remove_from_context(event
);
7309 unaccount_event_cpu(event
, src_cpu
);
7311 list_add(&event
->migrate_entry
, &events
);
7313 mutex_unlock(&src_ctx
->mutex
);
7317 mutex_lock(&dst_ctx
->mutex
);
7318 list_for_each_entry_safe(event
, tmp
, &events
, migrate_entry
) {
7319 list_del(&event
->migrate_entry
);
7320 if (event
->state
>= PERF_EVENT_STATE_OFF
)
7321 event
->state
= PERF_EVENT_STATE_INACTIVE
;
7322 account_event_cpu(event
, dst_cpu
);
7323 perf_install_in_context(dst_ctx
, event
, dst_cpu
);
7326 mutex_unlock(&dst_ctx
->mutex
);
7328 EXPORT_SYMBOL_GPL(perf_pmu_migrate_context
);
7330 static void sync_child_event(struct perf_event
*child_event
,
7331 struct task_struct
*child
)
7333 struct perf_event
*parent_event
= child_event
->parent
;
7336 if (child_event
->attr
.inherit_stat
)
7337 perf_event_read_event(child_event
, child
);
7339 child_val
= perf_event_count(child_event
);
7342 * Add back the child's count to the parent's count:
7344 atomic64_add(child_val
, &parent_event
->child_count
);
7345 atomic64_add(child_event
->total_time_enabled
,
7346 &parent_event
->child_total_time_enabled
);
7347 atomic64_add(child_event
->total_time_running
,
7348 &parent_event
->child_total_time_running
);
7351 * Remove this event from the parent's list
7353 WARN_ON_ONCE(parent_event
->ctx
->parent_ctx
);
7354 mutex_lock(&parent_event
->child_mutex
);
7355 list_del_init(&child_event
->child_list
);
7356 mutex_unlock(&parent_event
->child_mutex
);
7359 * Release the parent event, if this was the last
7362 put_event(parent_event
);
7366 __perf_event_exit_task(struct perf_event
*child_event
,
7367 struct perf_event_context
*child_ctx
,
7368 struct task_struct
*child
)
7370 if (child_event
->parent
) {
7371 raw_spin_lock_irq(&child_ctx
->lock
);
7372 perf_group_detach(child_event
);
7373 raw_spin_unlock_irq(&child_ctx
->lock
);
7376 perf_remove_from_context(child_event
);
7379 * It can happen that the parent exits first, and has events
7380 * that are still around due to the child reference. These
7381 * events need to be zapped.
7383 if (child_event
->parent
) {
7384 sync_child_event(child_event
, child
);
7385 free_event(child_event
);
7389 static void perf_event_exit_task_context(struct task_struct
*child
, int ctxn
)
7391 struct perf_event
*child_event
, *tmp
;
7392 struct perf_event_context
*child_ctx
;
7393 unsigned long flags
;
7395 if (likely(!child
->perf_event_ctxp
[ctxn
])) {
7396 perf_event_task(child
, NULL
, 0);
7400 local_irq_save(flags
);
7402 * We can't reschedule here because interrupts are disabled,
7403 * and either child is current or it is a task that can't be
7404 * scheduled, so we are now safe from rescheduling changing
7407 child_ctx
= rcu_dereference_raw(child
->perf_event_ctxp
[ctxn
]);
7410 * Take the context lock here so that if find_get_context is
7411 * reading child->perf_event_ctxp, we wait until it has
7412 * incremented the context's refcount before we do put_ctx below.
7414 raw_spin_lock(&child_ctx
->lock
);
7415 task_ctx_sched_out(child_ctx
);
7416 child
->perf_event_ctxp
[ctxn
] = NULL
;
7418 * If this context is a clone; unclone it so it can't get
7419 * swapped to another process while we're removing all
7420 * the events from it.
7422 unclone_ctx(child_ctx
);
7423 update_context_time(child_ctx
);
7424 raw_spin_unlock_irqrestore(&child_ctx
->lock
, flags
);
7427 * Report the task dead after unscheduling the events so that we
7428 * won't get any samples after PERF_RECORD_EXIT. We can however still
7429 * get a few PERF_RECORD_READ events.
7431 perf_event_task(child
, child_ctx
, 0);
7434 * We can recurse on the same lock type through:
7436 * __perf_event_exit_task()
7437 * sync_child_event()
7439 * mutex_lock(&ctx->mutex)
7441 * But since its the parent context it won't be the same instance.
7443 mutex_lock(&child_ctx
->mutex
);
7446 list_for_each_entry_safe(child_event
, tmp
, &child_ctx
->pinned_groups
,
7448 __perf_event_exit_task(child_event
, child_ctx
, child
);
7450 list_for_each_entry_safe(child_event
, tmp
, &child_ctx
->flexible_groups
,
7452 __perf_event_exit_task(child_event
, child_ctx
, child
);
7455 * If the last event was a group event, it will have appended all
7456 * its siblings to the list, but we obtained 'tmp' before that which
7457 * will still point to the list head terminating the iteration.
7459 if (!list_empty(&child_ctx
->pinned_groups
) ||
7460 !list_empty(&child_ctx
->flexible_groups
))
7463 mutex_unlock(&child_ctx
->mutex
);
7469 * When a child task exits, feed back event values to parent events.
7471 void perf_event_exit_task(struct task_struct
*child
)
7473 struct perf_event
*event
, *tmp
;
7476 mutex_lock(&child
->perf_event_mutex
);
7477 list_for_each_entry_safe(event
, tmp
, &child
->perf_event_list
,
7479 list_del_init(&event
->owner_entry
);
7482 * Ensure the list deletion is visible before we clear
7483 * the owner, closes a race against perf_release() where
7484 * we need to serialize on the owner->perf_event_mutex.
7487 event
->owner
= NULL
;
7489 mutex_unlock(&child
->perf_event_mutex
);
7491 for_each_task_context_nr(ctxn
)
7492 perf_event_exit_task_context(child
, ctxn
);
7495 static void perf_free_event(struct perf_event
*event
,
7496 struct perf_event_context
*ctx
)
7498 struct perf_event
*parent
= event
->parent
;
7500 if (WARN_ON_ONCE(!parent
))
7503 mutex_lock(&parent
->child_mutex
);
7504 list_del_init(&event
->child_list
);
7505 mutex_unlock(&parent
->child_mutex
);
7509 perf_group_detach(event
);
7510 list_del_event(event
, ctx
);
7515 * free an unexposed, unused context as created by inheritance by
7516 * perf_event_init_task below, used by fork() in case of fail.
7518 void perf_event_free_task(struct task_struct
*task
)
7520 struct perf_event_context
*ctx
;
7521 struct perf_event
*event
, *tmp
;
7524 for_each_task_context_nr(ctxn
) {
7525 ctx
= task
->perf_event_ctxp
[ctxn
];
7529 mutex_lock(&ctx
->mutex
);
7531 list_for_each_entry_safe(event
, tmp
, &ctx
->pinned_groups
,
7533 perf_free_event(event
, ctx
);
7535 list_for_each_entry_safe(event
, tmp
, &ctx
->flexible_groups
,
7537 perf_free_event(event
, ctx
);
7539 if (!list_empty(&ctx
->pinned_groups
) ||
7540 !list_empty(&ctx
->flexible_groups
))
7543 mutex_unlock(&ctx
->mutex
);
7549 void perf_event_delayed_put(struct task_struct
*task
)
7553 for_each_task_context_nr(ctxn
)
7554 WARN_ON_ONCE(task
->perf_event_ctxp
[ctxn
]);
7558 * inherit a event from parent task to child task:
7560 static struct perf_event
*
7561 inherit_event(struct perf_event
*parent_event
,
7562 struct task_struct
*parent
,
7563 struct perf_event_context
*parent_ctx
,
7564 struct task_struct
*child
,
7565 struct perf_event
*group_leader
,
7566 struct perf_event_context
*child_ctx
)
7568 struct perf_event
*child_event
;
7569 unsigned long flags
;
7572 * Instead of creating recursive hierarchies of events,
7573 * we link inherited events back to the original parent,
7574 * which has a filp for sure, which we use as the reference
7577 if (parent_event
->parent
)
7578 parent_event
= parent_event
->parent
;
7580 child_event
= perf_event_alloc(&parent_event
->attr
,
7583 group_leader
, parent_event
,
7585 if (IS_ERR(child_event
))
7588 if (!atomic_long_inc_not_zero(&parent_event
->refcount
)) {
7589 free_event(child_event
);
7596 * Make the child state follow the state of the parent event,
7597 * not its attr.disabled bit. We hold the parent's mutex,
7598 * so we won't race with perf_event_{en, dis}able_family.
7600 if (parent_event
->state
>= PERF_EVENT_STATE_INACTIVE
)
7601 child_event
->state
= PERF_EVENT_STATE_INACTIVE
;
7603 child_event
->state
= PERF_EVENT_STATE_OFF
;
7605 if (parent_event
->attr
.freq
) {
7606 u64 sample_period
= parent_event
->hw
.sample_period
;
7607 struct hw_perf_event
*hwc
= &child_event
->hw
;
7609 hwc
->sample_period
= sample_period
;
7610 hwc
->last_period
= sample_period
;
7612 local64_set(&hwc
->period_left
, sample_period
);
7615 child_event
->ctx
= child_ctx
;
7616 child_event
->overflow_handler
= parent_event
->overflow_handler
;
7617 child_event
->overflow_handler_context
7618 = parent_event
->overflow_handler_context
;
7621 * Precalculate sample_data sizes
7623 perf_event__header_size(child_event
);
7624 perf_event__id_header_size(child_event
);
7627 * Link it up in the child's context:
7629 raw_spin_lock_irqsave(&child_ctx
->lock
, flags
);
7630 add_event_to_ctx(child_event
, child_ctx
);
7631 raw_spin_unlock_irqrestore(&child_ctx
->lock
, flags
);
7634 * Link this into the parent event's child list
7636 WARN_ON_ONCE(parent_event
->ctx
->parent_ctx
);
7637 mutex_lock(&parent_event
->child_mutex
);
7638 list_add_tail(&child_event
->child_list
, &parent_event
->child_list
);
7639 mutex_unlock(&parent_event
->child_mutex
);
7644 static int inherit_group(struct perf_event
*parent_event
,
7645 struct task_struct
*parent
,
7646 struct perf_event_context
*parent_ctx
,
7647 struct task_struct
*child
,
7648 struct perf_event_context
*child_ctx
)
7650 struct perf_event
*leader
;
7651 struct perf_event
*sub
;
7652 struct perf_event
*child_ctr
;
7654 leader
= inherit_event(parent_event
, parent
, parent_ctx
,
7655 child
, NULL
, child_ctx
);
7657 return PTR_ERR(leader
);
7658 list_for_each_entry(sub
, &parent_event
->sibling_list
, group_entry
) {
7659 child_ctr
= inherit_event(sub
, parent
, parent_ctx
,
7660 child
, leader
, child_ctx
);
7661 if (IS_ERR(child_ctr
))
7662 return PTR_ERR(child_ctr
);
7668 inherit_task_group(struct perf_event
*event
, struct task_struct
*parent
,
7669 struct perf_event_context
*parent_ctx
,
7670 struct task_struct
*child
, int ctxn
,
7674 struct perf_event_context
*child_ctx
;
7676 if (!event
->attr
.inherit
) {
7681 child_ctx
= child
->perf_event_ctxp
[ctxn
];
7684 * This is executed from the parent task context, so
7685 * inherit events that have been marked for cloning.
7686 * First allocate and initialize a context for the
7690 child_ctx
= alloc_perf_context(parent_ctx
->pmu
, child
);
7694 child
->perf_event_ctxp
[ctxn
] = child_ctx
;
7697 ret
= inherit_group(event
, parent
, parent_ctx
,
7707 * Initialize the perf_event context in task_struct
7709 int perf_event_init_context(struct task_struct
*child
, int ctxn
)
7711 struct perf_event_context
*child_ctx
, *parent_ctx
;
7712 struct perf_event_context
*cloned_ctx
;
7713 struct perf_event
*event
;
7714 struct task_struct
*parent
= current
;
7715 int inherited_all
= 1;
7716 unsigned long flags
;
7719 if (likely(!parent
->perf_event_ctxp
[ctxn
]))
7723 * If the parent's context is a clone, pin it so it won't get
7726 parent_ctx
= perf_pin_task_context(parent
, ctxn
);
7729 * No need to check if parent_ctx != NULL here; since we saw
7730 * it non-NULL earlier, the only reason for it to become NULL
7731 * is if we exit, and since we're currently in the middle of
7732 * a fork we can't be exiting at the same time.
7736 * Lock the parent list. No need to lock the child - not PID
7737 * hashed yet and not running, so nobody can access it.
7739 mutex_lock(&parent_ctx
->mutex
);
7742 * We dont have to disable NMIs - we are only looking at
7743 * the list, not manipulating it:
7745 list_for_each_entry(event
, &parent_ctx
->pinned_groups
, group_entry
) {
7746 ret
= inherit_task_group(event
, parent
, parent_ctx
,
7747 child
, ctxn
, &inherited_all
);
7753 * We can't hold ctx->lock when iterating the ->flexible_group list due
7754 * to allocations, but we need to prevent rotation because
7755 * rotate_ctx() will change the list from interrupt context.
7757 raw_spin_lock_irqsave(&parent_ctx
->lock
, flags
);
7758 parent_ctx
->rotate_disable
= 1;
7759 raw_spin_unlock_irqrestore(&parent_ctx
->lock
, flags
);
7761 list_for_each_entry(event
, &parent_ctx
->flexible_groups
, group_entry
) {
7762 ret
= inherit_task_group(event
, parent
, parent_ctx
,
7763 child
, ctxn
, &inherited_all
);
7768 raw_spin_lock_irqsave(&parent_ctx
->lock
, flags
);
7769 parent_ctx
->rotate_disable
= 0;
7771 child_ctx
= child
->perf_event_ctxp
[ctxn
];
7773 if (child_ctx
&& inherited_all
) {
7775 * Mark the child context as a clone of the parent
7776 * context, or of whatever the parent is a clone of.
7778 * Note that if the parent is a clone, the holding of
7779 * parent_ctx->lock avoids it from being uncloned.
7781 cloned_ctx
= parent_ctx
->parent_ctx
;
7783 child_ctx
->parent_ctx
= cloned_ctx
;
7784 child_ctx
->parent_gen
= parent_ctx
->parent_gen
;
7786 child_ctx
->parent_ctx
= parent_ctx
;
7787 child_ctx
->parent_gen
= parent_ctx
->generation
;
7789 get_ctx(child_ctx
->parent_ctx
);
7792 raw_spin_unlock_irqrestore(&parent_ctx
->lock
, flags
);
7793 mutex_unlock(&parent_ctx
->mutex
);
7795 perf_unpin_context(parent_ctx
);
7796 put_ctx(parent_ctx
);
7802 * Initialize the perf_event context in task_struct
7804 int perf_event_init_task(struct task_struct
*child
)
7808 memset(child
->perf_event_ctxp
, 0, sizeof(child
->perf_event_ctxp
));
7809 mutex_init(&child
->perf_event_mutex
);
7810 INIT_LIST_HEAD(&child
->perf_event_list
);
7812 for_each_task_context_nr(ctxn
) {
7813 ret
= perf_event_init_context(child
, ctxn
);
7821 static void __init
perf_event_init_all_cpus(void)
7823 struct swevent_htable
*swhash
;
7826 for_each_possible_cpu(cpu
) {
7827 swhash
= &per_cpu(swevent_htable
, cpu
);
7828 mutex_init(&swhash
->hlist_mutex
);
7829 INIT_LIST_HEAD(&per_cpu(rotation_list
, cpu
));
7833 static void perf_event_init_cpu(int cpu
)
7835 struct swevent_htable
*swhash
= &per_cpu(swevent_htable
, cpu
);
7837 mutex_lock(&swhash
->hlist_mutex
);
7838 if (swhash
->hlist_refcount
> 0) {
7839 struct swevent_hlist
*hlist
;
7841 hlist
= kzalloc_node(sizeof(*hlist
), GFP_KERNEL
, cpu_to_node(cpu
));
7843 rcu_assign_pointer(swhash
->swevent_hlist
, hlist
);
7845 mutex_unlock(&swhash
->hlist_mutex
);
7848 #if defined CONFIG_HOTPLUG_CPU || defined CONFIG_KEXEC
7849 static void perf_pmu_rotate_stop(struct pmu
*pmu
)
7851 struct perf_cpu_context
*cpuctx
= this_cpu_ptr(pmu
->pmu_cpu_context
);
7853 WARN_ON(!irqs_disabled());
7855 list_del_init(&cpuctx
->rotation_list
);
7858 static void __perf_event_exit_context(void *__info
)
7860 struct perf_event_context
*ctx
= __info
;
7861 struct perf_event
*event
;
7863 perf_pmu_rotate_stop(ctx
->pmu
);
7866 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
)
7867 __perf_remove_from_context(event
);
7871 static void perf_event_exit_cpu_context(int cpu
)
7873 struct perf_event_context
*ctx
;
7877 idx
= srcu_read_lock(&pmus_srcu
);
7878 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
7879 ctx
= &per_cpu_ptr(pmu
->pmu_cpu_context
, cpu
)->ctx
;
7881 mutex_lock(&ctx
->mutex
);
7882 smp_call_function_single(cpu
, __perf_event_exit_context
, ctx
, 1);
7883 mutex_unlock(&ctx
->mutex
);
7885 srcu_read_unlock(&pmus_srcu
, idx
);
7888 static void perf_event_exit_cpu(int cpu
)
7890 struct swevent_htable
*swhash
= &per_cpu(swevent_htable
, cpu
);
7892 perf_event_exit_cpu_context(cpu
);
7894 mutex_lock(&swhash
->hlist_mutex
);
7895 swevent_hlist_release(swhash
);
7896 mutex_unlock(&swhash
->hlist_mutex
);
7899 static inline void perf_event_exit_cpu(int cpu
) { }
7903 perf_reboot(struct notifier_block
*notifier
, unsigned long val
, void *v
)
7907 for_each_online_cpu(cpu
)
7908 perf_event_exit_cpu(cpu
);
7914 * Run the perf reboot notifier at the very last possible moment so that
7915 * the generic watchdog code runs as long as possible.
7917 static struct notifier_block perf_reboot_notifier
= {
7918 .notifier_call
= perf_reboot
,
7919 .priority
= INT_MIN
,
7923 perf_cpu_notify(struct notifier_block
*self
, unsigned long action
, void *hcpu
)
7925 unsigned int cpu
= (long)hcpu
;
7927 switch (action
& ~CPU_TASKS_FROZEN
) {
7929 case CPU_UP_PREPARE
:
7930 case CPU_DOWN_FAILED
:
7931 perf_event_init_cpu(cpu
);
7934 case CPU_UP_CANCELED
:
7935 case CPU_DOWN_PREPARE
:
7936 perf_event_exit_cpu(cpu
);
7945 void __init
perf_event_init(void)
7951 perf_event_init_all_cpus();
7952 init_srcu_struct(&pmus_srcu
);
7953 perf_pmu_register(&perf_swevent
, "software", PERF_TYPE_SOFTWARE
);
7954 perf_pmu_register(&perf_cpu_clock
, NULL
, -1);
7955 perf_pmu_register(&perf_task_clock
, NULL
, -1);
7957 perf_cpu_notifier(perf_cpu_notify
);
7958 register_reboot_notifier(&perf_reboot_notifier
);
7960 ret
= init_hw_breakpoint();
7961 WARN(ret
, "hw_breakpoint initialization failed with: %d", ret
);
7963 /* do not patch jump label more than once per second */
7964 jump_label_rate_limit(&perf_sched_events
, HZ
);
7967 * Build time assertion that we keep the data_head at the intended
7968 * location. IOW, validation we got the __reserved[] size right.
7970 BUILD_BUG_ON((offsetof(struct perf_event_mmap_page
, data_head
))
7974 static int __init
perf_event_sysfs_init(void)
7979 mutex_lock(&pmus_lock
);
7981 ret
= bus_register(&pmu_bus
);
7985 list_for_each_entry(pmu
, &pmus
, entry
) {
7986 if (!pmu
->name
|| pmu
->type
< 0)
7989 ret
= pmu_dev_alloc(pmu
);
7990 WARN(ret
, "Failed to register pmu: %s, reason %d\n", pmu
->name
, ret
);
7992 pmu_bus_running
= 1;
7996 mutex_unlock(&pmus_lock
);
8000 device_initcall(perf_event_sysfs_init
);
8002 #ifdef CONFIG_CGROUP_PERF
8003 static struct cgroup_subsys_state
*
8004 perf_cgroup_css_alloc(struct cgroup_subsys_state
*parent_css
)
8006 struct perf_cgroup
*jc
;
8008 jc
= kzalloc(sizeof(*jc
), GFP_KERNEL
);
8010 return ERR_PTR(-ENOMEM
);
8012 jc
->info
= alloc_percpu(struct perf_cgroup_info
);
8015 return ERR_PTR(-ENOMEM
);
8021 static void perf_cgroup_css_free(struct cgroup_subsys_state
*css
)
8023 struct perf_cgroup
*jc
= container_of(css
, struct perf_cgroup
, css
);
8025 free_percpu(jc
->info
);
8029 static int __perf_cgroup_move(void *info
)
8031 struct task_struct
*task
= info
;
8032 perf_cgroup_switch(task
, PERF_CGROUP_SWOUT
| PERF_CGROUP_SWIN
);
8036 static void perf_cgroup_attach(struct cgroup_subsys_state
*css
,
8037 struct cgroup_taskset
*tset
)
8039 struct task_struct
*task
;
8041 cgroup_taskset_for_each(task
, tset
)
8042 task_function_call(task
, __perf_cgroup_move
, task
);
8045 static void perf_cgroup_exit(struct cgroup_subsys_state
*css
,
8046 struct cgroup_subsys_state
*old_css
,
8047 struct task_struct
*task
)
8050 * cgroup_exit() is called in the copy_process() failure path.
8051 * Ignore this case since the task hasn't ran yet, this avoids
8052 * trying to poke a half freed task state from generic code.
8054 if (!(task
->flags
& PF_EXITING
))
8057 task_function_call(task
, __perf_cgroup_move
, task
);
8060 struct cgroup_subsys perf_event_cgrp_subsys
= {
8061 .css_alloc
= perf_cgroup_css_alloc
,
8062 .css_free
= perf_cgroup_css_free
,
8063 .exit
= perf_cgroup_exit
,
8064 .attach
= perf_cgroup_attach
,
8066 #endif /* CONFIG_CGROUP_PERF */