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
,
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 irq_work_queue(&perf_duration_work
);
289 static atomic64_t perf_event_id
;
291 static void cpu_ctx_sched_out(struct perf_cpu_context
*cpuctx
,
292 enum event_type_t event_type
);
294 static void cpu_ctx_sched_in(struct perf_cpu_context
*cpuctx
,
295 enum event_type_t event_type
,
296 struct task_struct
*task
);
298 static void update_context_time(struct perf_event_context
*ctx
);
299 static u64
perf_event_time(struct perf_event
*event
);
301 void __weak
perf_event_print_debug(void) { }
303 extern __weak
const char *perf_pmu_name(void)
308 static inline u64
perf_clock(void)
310 return local_clock();
313 static inline struct perf_cpu_context
*
314 __get_cpu_context(struct perf_event_context
*ctx
)
316 return this_cpu_ptr(ctx
->pmu
->pmu_cpu_context
);
319 static void perf_ctx_lock(struct perf_cpu_context
*cpuctx
,
320 struct perf_event_context
*ctx
)
322 raw_spin_lock(&cpuctx
->ctx
.lock
);
324 raw_spin_lock(&ctx
->lock
);
327 static void perf_ctx_unlock(struct perf_cpu_context
*cpuctx
,
328 struct perf_event_context
*ctx
)
331 raw_spin_unlock(&ctx
->lock
);
332 raw_spin_unlock(&cpuctx
->ctx
.lock
);
335 #ifdef CONFIG_CGROUP_PERF
338 * perf_cgroup_info keeps track of time_enabled for a cgroup.
339 * This is a per-cpu dynamically allocated data structure.
341 struct perf_cgroup_info
{
347 struct cgroup_subsys_state css
;
348 struct perf_cgroup_info __percpu
*info
;
352 * Must ensure cgroup is pinned (css_get) before calling
353 * this function. In other words, we cannot call this function
354 * if there is no cgroup event for the current CPU context.
356 static inline struct perf_cgroup
*
357 perf_cgroup_from_task(struct task_struct
*task
)
359 return container_of(task_css(task
, perf_subsys_id
),
360 struct perf_cgroup
, css
);
364 perf_cgroup_match(struct perf_event
*event
)
366 struct perf_event_context
*ctx
= event
->ctx
;
367 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
369 /* @event doesn't care about cgroup */
373 /* wants specific cgroup scope but @cpuctx isn't associated with any */
378 * Cgroup scoping is recursive. An event enabled for a cgroup is
379 * also enabled for all its descendant cgroups. If @cpuctx's
380 * cgroup is a descendant of @event's (the test covers identity
381 * case), it's a match.
383 return cgroup_is_descendant(cpuctx
->cgrp
->css
.cgroup
,
384 event
->cgrp
->css
.cgroup
);
387 static inline bool perf_tryget_cgroup(struct perf_event
*event
)
389 return css_tryget(&event
->cgrp
->css
);
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
);
612 css
= css_from_dir(f
.file
->f_dentry
, &perf_subsys
);
618 cgrp
= container_of(css
, struct perf_cgroup
, css
);
621 /* must be done before we fput() the file */
622 if (!perf_tryget_cgroup(event
)) {
629 * all events in a group must monitor
630 * the same cgroup because a task belongs
631 * to only one perf cgroup at a time
633 if (group_leader
&& group_leader
->cgrp
!= cgrp
) {
634 perf_detach_cgroup(event
);
644 perf_cgroup_set_shadow_time(struct perf_event
*event
, u64 now
)
646 struct perf_cgroup_info
*t
;
647 t
= per_cpu_ptr(event
->cgrp
->info
, event
->cpu
);
648 event
->shadow_ctx_time
= now
- t
->timestamp
;
652 perf_cgroup_defer_enabled(struct perf_event
*event
)
655 * when the current task's perf cgroup does not match
656 * the event's, we need to remember to call the
657 * perf_mark_enable() function the first time a task with
658 * a matching perf cgroup is scheduled in.
660 if (is_cgroup_event(event
) && !perf_cgroup_match(event
))
661 event
->cgrp_defer_enabled
= 1;
665 perf_cgroup_mark_enabled(struct perf_event
*event
,
666 struct perf_event_context
*ctx
)
668 struct perf_event
*sub
;
669 u64 tstamp
= perf_event_time(event
);
671 if (!event
->cgrp_defer_enabled
)
674 event
->cgrp_defer_enabled
= 0;
676 event
->tstamp_enabled
= tstamp
- event
->total_time_enabled
;
677 list_for_each_entry(sub
, &event
->sibling_list
, group_entry
) {
678 if (sub
->state
>= PERF_EVENT_STATE_INACTIVE
) {
679 sub
->tstamp_enabled
= tstamp
- sub
->total_time_enabled
;
680 sub
->cgrp_defer_enabled
= 0;
684 #else /* !CONFIG_CGROUP_PERF */
687 perf_cgroup_match(struct perf_event
*event
)
692 static inline void perf_detach_cgroup(struct perf_event
*event
)
695 static inline int is_cgroup_event(struct perf_event
*event
)
700 static inline u64
perf_cgroup_event_cgrp_time(struct perf_event
*event
)
705 static inline void update_cgrp_time_from_event(struct perf_event
*event
)
709 static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context
*cpuctx
)
713 static inline void perf_cgroup_sched_out(struct task_struct
*task
,
714 struct task_struct
*next
)
718 static inline void perf_cgroup_sched_in(struct task_struct
*prev
,
719 struct task_struct
*task
)
723 static inline int perf_cgroup_connect(pid_t pid
, struct perf_event
*event
,
724 struct perf_event_attr
*attr
,
725 struct perf_event
*group_leader
)
731 perf_cgroup_set_timestamp(struct task_struct
*task
,
732 struct perf_event_context
*ctx
)
737 perf_cgroup_switch(struct task_struct
*task
, struct task_struct
*next
)
742 perf_cgroup_set_shadow_time(struct perf_event
*event
, u64 now
)
746 static inline u64
perf_cgroup_event_time(struct perf_event
*event
)
752 perf_cgroup_defer_enabled(struct perf_event
*event
)
757 perf_cgroup_mark_enabled(struct perf_event
*event
,
758 struct perf_event_context
*ctx
)
764 * set default to be dependent on timer tick just
767 #define PERF_CPU_HRTIMER (1000 / HZ)
769 * function must be called with interrupts disbled
771 static enum hrtimer_restart
perf_cpu_hrtimer_handler(struct hrtimer
*hr
)
773 struct perf_cpu_context
*cpuctx
;
774 enum hrtimer_restart ret
= HRTIMER_NORESTART
;
777 WARN_ON(!irqs_disabled());
779 cpuctx
= container_of(hr
, struct perf_cpu_context
, hrtimer
);
781 rotations
= perf_rotate_context(cpuctx
);
784 * arm timer if needed
787 hrtimer_forward_now(hr
, cpuctx
->hrtimer_interval
);
788 ret
= HRTIMER_RESTART
;
794 /* CPU is going down */
795 void perf_cpu_hrtimer_cancel(int cpu
)
797 struct perf_cpu_context
*cpuctx
;
801 if (WARN_ON(cpu
!= smp_processor_id()))
804 local_irq_save(flags
);
808 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
809 cpuctx
= this_cpu_ptr(pmu
->pmu_cpu_context
);
811 if (pmu
->task_ctx_nr
== perf_sw_context
)
814 hrtimer_cancel(&cpuctx
->hrtimer
);
819 local_irq_restore(flags
);
822 static void __perf_cpu_hrtimer_init(struct perf_cpu_context
*cpuctx
, int cpu
)
824 struct hrtimer
*hr
= &cpuctx
->hrtimer
;
825 struct pmu
*pmu
= cpuctx
->ctx
.pmu
;
828 /* no multiplexing needed for SW PMU */
829 if (pmu
->task_ctx_nr
== perf_sw_context
)
833 * check default is sane, if not set then force to
834 * default interval (1/tick)
836 timer
= pmu
->hrtimer_interval_ms
;
838 timer
= pmu
->hrtimer_interval_ms
= PERF_CPU_HRTIMER
;
840 cpuctx
->hrtimer_interval
= ns_to_ktime(NSEC_PER_MSEC
* timer
);
842 hrtimer_init(hr
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL_PINNED
);
843 hr
->function
= perf_cpu_hrtimer_handler
;
846 static void perf_cpu_hrtimer_restart(struct perf_cpu_context
*cpuctx
)
848 struct hrtimer
*hr
= &cpuctx
->hrtimer
;
849 struct pmu
*pmu
= cpuctx
->ctx
.pmu
;
852 if (pmu
->task_ctx_nr
== perf_sw_context
)
855 if (hrtimer_active(hr
))
858 if (!hrtimer_callback_running(hr
))
859 __hrtimer_start_range_ns(hr
, cpuctx
->hrtimer_interval
,
860 0, HRTIMER_MODE_REL_PINNED
, 0);
863 void perf_pmu_disable(struct pmu
*pmu
)
865 int *count
= this_cpu_ptr(pmu
->pmu_disable_count
);
867 pmu
->pmu_disable(pmu
);
870 void perf_pmu_enable(struct pmu
*pmu
)
872 int *count
= this_cpu_ptr(pmu
->pmu_disable_count
);
874 pmu
->pmu_enable(pmu
);
877 static DEFINE_PER_CPU(struct list_head
, rotation_list
);
880 * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
881 * because they're strictly cpu affine and rotate_start is called with IRQs
882 * disabled, while rotate_context is called from IRQ context.
884 static void perf_pmu_rotate_start(struct pmu
*pmu
)
886 struct perf_cpu_context
*cpuctx
= this_cpu_ptr(pmu
->pmu_cpu_context
);
887 struct list_head
*head
= &__get_cpu_var(rotation_list
);
889 WARN_ON(!irqs_disabled());
891 if (list_empty(&cpuctx
->rotation_list
))
892 list_add(&cpuctx
->rotation_list
, head
);
895 static void get_ctx(struct perf_event_context
*ctx
)
897 WARN_ON(!atomic_inc_not_zero(&ctx
->refcount
));
900 static void put_ctx(struct perf_event_context
*ctx
)
902 if (atomic_dec_and_test(&ctx
->refcount
)) {
904 put_ctx(ctx
->parent_ctx
);
906 put_task_struct(ctx
->task
);
907 kfree_rcu(ctx
, rcu_head
);
911 static void unclone_ctx(struct perf_event_context
*ctx
)
913 if (ctx
->parent_ctx
) {
914 put_ctx(ctx
->parent_ctx
);
915 ctx
->parent_ctx
= NULL
;
920 static u32
perf_event_pid(struct perf_event
*event
, struct task_struct
*p
)
923 * only top level events have the pid namespace they were created in
926 event
= event
->parent
;
928 return task_tgid_nr_ns(p
, event
->ns
);
931 static u32
perf_event_tid(struct perf_event
*event
, struct task_struct
*p
)
934 * only top level events have the pid namespace they were created in
937 event
= event
->parent
;
939 return task_pid_nr_ns(p
, event
->ns
);
943 * If we inherit events we want to return the parent event id
946 static u64
primary_event_id(struct perf_event
*event
)
951 id
= event
->parent
->id
;
957 * Get the perf_event_context for a task and lock it.
958 * This has to cope with with the fact that until it is locked,
959 * the context could get moved to another task.
961 static struct perf_event_context
*
962 perf_lock_task_context(struct task_struct
*task
, int ctxn
, unsigned long *flags
)
964 struct perf_event_context
*ctx
;
968 * One of the few rules of preemptible RCU is that one cannot do
969 * rcu_read_unlock() while holding a scheduler (or nested) lock when
970 * part of the read side critical section was preemptible -- see
971 * rcu_read_unlock_special().
973 * Since ctx->lock nests under rq->lock we must ensure the entire read
974 * side critical section is non-preemptible.
978 ctx
= rcu_dereference(task
->perf_event_ctxp
[ctxn
]);
981 * If this context is a clone of another, it might
982 * get swapped for another underneath us by
983 * perf_event_task_sched_out, though the
984 * rcu_read_lock() protects us from any context
985 * getting freed. Lock the context and check if it
986 * got swapped before we could get the lock, and retry
987 * if so. If we locked the right context, then it
988 * can't get swapped on us any more.
990 raw_spin_lock_irqsave(&ctx
->lock
, *flags
);
991 if (ctx
!= rcu_dereference(task
->perf_event_ctxp
[ctxn
])) {
992 raw_spin_unlock_irqrestore(&ctx
->lock
, *flags
);
998 if (!atomic_inc_not_zero(&ctx
->refcount
)) {
999 raw_spin_unlock_irqrestore(&ctx
->lock
, *flags
);
1009 * Get the context for a task and increment its pin_count so it
1010 * can't get swapped to another task. This also increments its
1011 * reference count so that the context can't get freed.
1013 static struct perf_event_context
*
1014 perf_pin_task_context(struct task_struct
*task
, int ctxn
)
1016 struct perf_event_context
*ctx
;
1017 unsigned long flags
;
1019 ctx
= perf_lock_task_context(task
, ctxn
, &flags
);
1022 raw_spin_unlock_irqrestore(&ctx
->lock
, flags
);
1027 static void perf_unpin_context(struct perf_event_context
*ctx
)
1029 unsigned long flags
;
1031 raw_spin_lock_irqsave(&ctx
->lock
, flags
);
1033 raw_spin_unlock_irqrestore(&ctx
->lock
, flags
);
1037 * Update the record of the current time in a context.
1039 static void update_context_time(struct perf_event_context
*ctx
)
1041 u64 now
= perf_clock();
1043 ctx
->time
+= now
- ctx
->timestamp
;
1044 ctx
->timestamp
= now
;
1047 static u64
perf_event_time(struct perf_event
*event
)
1049 struct perf_event_context
*ctx
= event
->ctx
;
1051 if (is_cgroup_event(event
))
1052 return perf_cgroup_event_time(event
);
1054 return ctx
? ctx
->time
: 0;
1058 * Update the total_time_enabled and total_time_running fields for a event.
1059 * The caller of this function needs to hold the ctx->lock.
1061 static void update_event_times(struct perf_event
*event
)
1063 struct perf_event_context
*ctx
= event
->ctx
;
1066 if (event
->state
< PERF_EVENT_STATE_INACTIVE
||
1067 event
->group_leader
->state
< PERF_EVENT_STATE_INACTIVE
)
1070 * in cgroup mode, time_enabled represents
1071 * the time the event was enabled AND active
1072 * tasks were in the monitored cgroup. This is
1073 * independent of the activity of the context as
1074 * there may be a mix of cgroup and non-cgroup events.
1076 * That is why we treat cgroup events differently
1079 if (is_cgroup_event(event
))
1080 run_end
= perf_cgroup_event_time(event
);
1081 else if (ctx
->is_active
)
1082 run_end
= ctx
->time
;
1084 run_end
= event
->tstamp_stopped
;
1086 event
->total_time_enabled
= run_end
- event
->tstamp_enabled
;
1088 if (event
->state
== PERF_EVENT_STATE_INACTIVE
)
1089 run_end
= event
->tstamp_stopped
;
1091 run_end
= perf_event_time(event
);
1093 event
->total_time_running
= run_end
- event
->tstamp_running
;
1098 * Update total_time_enabled and total_time_running for all events in a group.
1100 static void update_group_times(struct perf_event
*leader
)
1102 struct perf_event
*event
;
1104 update_event_times(leader
);
1105 list_for_each_entry(event
, &leader
->sibling_list
, group_entry
)
1106 update_event_times(event
);
1109 static struct list_head
*
1110 ctx_group_list(struct perf_event
*event
, struct perf_event_context
*ctx
)
1112 if (event
->attr
.pinned
)
1113 return &ctx
->pinned_groups
;
1115 return &ctx
->flexible_groups
;
1119 * Add a event from the lists for its context.
1120 * Must be called with ctx->mutex and ctx->lock held.
1123 list_add_event(struct perf_event
*event
, struct perf_event_context
*ctx
)
1125 WARN_ON_ONCE(event
->attach_state
& PERF_ATTACH_CONTEXT
);
1126 event
->attach_state
|= PERF_ATTACH_CONTEXT
;
1129 * If we're a stand alone event or group leader, we go to the context
1130 * list, group events are kept attached to the group so that
1131 * perf_group_detach can, at all times, locate all siblings.
1133 if (event
->group_leader
== event
) {
1134 struct list_head
*list
;
1136 if (is_software_event(event
))
1137 event
->group_flags
|= PERF_GROUP_SOFTWARE
;
1139 list
= ctx_group_list(event
, ctx
);
1140 list_add_tail(&event
->group_entry
, list
);
1143 if (is_cgroup_event(event
))
1146 if (has_branch_stack(event
))
1147 ctx
->nr_branch_stack
++;
1149 list_add_rcu(&event
->event_entry
, &ctx
->event_list
);
1150 if (!ctx
->nr_events
)
1151 perf_pmu_rotate_start(ctx
->pmu
);
1153 if (event
->attr
.inherit_stat
)
1160 * Initialize event state based on the perf_event_attr::disabled.
1162 static inline void perf_event__state_init(struct perf_event
*event
)
1164 event
->state
= event
->attr
.disabled
? PERF_EVENT_STATE_OFF
:
1165 PERF_EVENT_STATE_INACTIVE
;
1169 * Called at perf_event creation and when events are attached/detached from a
1172 static void perf_event__read_size(struct perf_event
*event
)
1174 int entry
= sizeof(u64
); /* value */
1178 if (event
->attr
.read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
1179 size
+= sizeof(u64
);
1181 if (event
->attr
.read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
1182 size
+= sizeof(u64
);
1184 if (event
->attr
.read_format
& PERF_FORMAT_ID
)
1185 entry
+= sizeof(u64
);
1187 if (event
->attr
.read_format
& PERF_FORMAT_GROUP
) {
1188 nr
+= event
->group_leader
->nr_siblings
;
1189 size
+= sizeof(u64
);
1193 event
->read_size
= size
;
1196 static void perf_event__header_size(struct perf_event
*event
)
1198 struct perf_sample_data
*data
;
1199 u64 sample_type
= event
->attr
.sample_type
;
1202 perf_event__read_size(event
);
1204 if (sample_type
& PERF_SAMPLE_IP
)
1205 size
+= sizeof(data
->ip
);
1207 if (sample_type
& PERF_SAMPLE_ADDR
)
1208 size
+= sizeof(data
->addr
);
1210 if (sample_type
& PERF_SAMPLE_PERIOD
)
1211 size
+= sizeof(data
->period
);
1213 if (sample_type
& PERF_SAMPLE_WEIGHT
)
1214 size
+= sizeof(data
->weight
);
1216 if (sample_type
& PERF_SAMPLE_READ
)
1217 size
+= event
->read_size
;
1219 if (sample_type
& PERF_SAMPLE_DATA_SRC
)
1220 size
+= sizeof(data
->data_src
.val
);
1222 if (sample_type
& PERF_SAMPLE_TRANSACTION
)
1223 size
+= sizeof(data
->txn
);
1225 event
->header_size
= size
;
1228 static void perf_event__id_header_size(struct perf_event
*event
)
1230 struct perf_sample_data
*data
;
1231 u64 sample_type
= event
->attr
.sample_type
;
1234 if (sample_type
& PERF_SAMPLE_TID
)
1235 size
+= sizeof(data
->tid_entry
);
1237 if (sample_type
& PERF_SAMPLE_TIME
)
1238 size
+= sizeof(data
->time
);
1240 if (sample_type
& PERF_SAMPLE_IDENTIFIER
)
1241 size
+= sizeof(data
->id
);
1243 if (sample_type
& PERF_SAMPLE_ID
)
1244 size
+= sizeof(data
->id
);
1246 if (sample_type
& PERF_SAMPLE_STREAM_ID
)
1247 size
+= sizeof(data
->stream_id
);
1249 if (sample_type
& PERF_SAMPLE_CPU
)
1250 size
+= sizeof(data
->cpu_entry
);
1252 event
->id_header_size
= size
;
1255 static void perf_group_attach(struct perf_event
*event
)
1257 struct perf_event
*group_leader
= event
->group_leader
, *pos
;
1260 * We can have double attach due to group movement in perf_event_open.
1262 if (event
->attach_state
& PERF_ATTACH_GROUP
)
1265 event
->attach_state
|= PERF_ATTACH_GROUP
;
1267 if (group_leader
== event
)
1270 if (group_leader
->group_flags
& PERF_GROUP_SOFTWARE
&&
1271 !is_software_event(event
))
1272 group_leader
->group_flags
&= ~PERF_GROUP_SOFTWARE
;
1274 list_add_tail(&event
->group_entry
, &group_leader
->sibling_list
);
1275 group_leader
->nr_siblings
++;
1277 perf_event__header_size(group_leader
);
1279 list_for_each_entry(pos
, &group_leader
->sibling_list
, group_entry
)
1280 perf_event__header_size(pos
);
1284 * Remove a event from the lists for its context.
1285 * Must be called with ctx->mutex and ctx->lock held.
1288 list_del_event(struct perf_event
*event
, struct perf_event_context
*ctx
)
1290 struct perf_cpu_context
*cpuctx
;
1292 * We can have double detach due to exit/hot-unplug + close.
1294 if (!(event
->attach_state
& PERF_ATTACH_CONTEXT
))
1297 event
->attach_state
&= ~PERF_ATTACH_CONTEXT
;
1299 if (is_cgroup_event(event
)) {
1301 cpuctx
= __get_cpu_context(ctx
);
1303 * if there are no more cgroup events
1304 * then cler cgrp to avoid stale pointer
1305 * in update_cgrp_time_from_cpuctx()
1307 if (!ctx
->nr_cgroups
)
1308 cpuctx
->cgrp
= NULL
;
1311 if (has_branch_stack(event
))
1312 ctx
->nr_branch_stack
--;
1315 if (event
->attr
.inherit_stat
)
1318 list_del_rcu(&event
->event_entry
);
1320 if (event
->group_leader
== event
)
1321 list_del_init(&event
->group_entry
);
1323 update_group_times(event
);
1326 * If event was in error state, then keep it
1327 * that way, otherwise bogus counts will be
1328 * returned on read(). The only way to get out
1329 * of error state is by explicit re-enabling
1332 if (event
->state
> PERF_EVENT_STATE_OFF
)
1333 event
->state
= PERF_EVENT_STATE_OFF
;
1338 static void perf_group_detach(struct perf_event
*event
)
1340 struct perf_event
*sibling
, *tmp
;
1341 struct list_head
*list
= NULL
;
1344 * We can have double detach due to exit/hot-unplug + close.
1346 if (!(event
->attach_state
& PERF_ATTACH_GROUP
))
1349 event
->attach_state
&= ~PERF_ATTACH_GROUP
;
1352 * If this is a sibling, remove it from its group.
1354 if (event
->group_leader
!= event
) {
1355 list_del_init(&event
->group_entry
);
1356 event
->group_leader
->nr_siblings
--;
1360 if (!list_empty(&event
->group_entry
))
1361 list
= &event
->group_entry
;
1364 * If this was a group event with sibling events then
1365 * upgrade the siblings to singleton events by adding them
1366 * to whatever list we are on.
1368 list_for_each_entry_safe(sibling
, tmp
, &event
->sibling_list
, group_entry
) {
1370 list_move_tail(&sibling
->group_entry
, list
);
1371 sibling
->group_leader
= sibling
;
1373 /* Inherit group flags from the previous leader */
1374 sibling
->group_flags
= event
->group_flags
;
1378 perf_event__header_size(event
->group_leader
);
1380 list_for_each_entry(tmp
, &event
->group_leader
->sibling_list
, group_entry
)
1381 perf_event__header_size(tmp
);
1385 event_filter_match(struct perf_event
*event
)
1387 return (event
->cpu
== -1 || event
->cpu
== smp_processor_id())
1388 && perf_cgroup_match(event
);
1392 event_sched_out(struct perf_event
*event
,
1393 struct perf_cpu_context
*cpuctx
,
1394 struct perf_event_context
*ctx
)
1396 u64 tstamp
= perf_event_time(event
);
1399 * An event which could not be activated because of
1400 * filter mismatch still needs to have its timings
1401 * maintained, otherwise bogus information is return
1402 * via read() for time_enabled, time_running:
1404 if (event
->state
== PERF_EVENT_STATE_INACTIVE
1405 && !event_filter_match(event
)) {
1406 delta
= tstamp
- event
->tstamp_stopped
;
1407 event
->tstamp_running
+= delta
;
1408 event
->tstamp_stopped
= tstamp
;
1411 if (event
->state
!= PERF_EVENT_STATE_ACTIVE
)
1414 perf_pmu_disable(event
->pmu
);
1416 event
->state
= PERF_EVENT_STATE_INACTIVE
;
1417 if (event
->pending_disable
) {
1418 event
->pending_disable
= 0;
1419 event
->state
= PERF_EVENT_STATE_OFF
;
1421 event
->tstamp_stopped
= tstamp
;
1422 event
->pmu
->del(event
, 0);
1425 if (!is_software_event(event
))
1426 cpuctx
->active_oncpu
--;
1428 if (event
->attr
.freq
&& event
->attr
.sample_freq
)
1430 if (event
->attr
.exclusive
|| !cpuctx
->active_oncpu
)
1431 cpuctx
->exclusive
= 0;
1433 perf_pmu_enable(event
->pmu
);
1437 group_sched_out(struct perf_event
*group_event
,
1438 struct perf_cpu_context
*cpuctx
,
1439 struct perf_event_context
*ctx
)
1441 struct perf_event
*event
;
1442 int state
= group_event
->state
;
1444 event_sched_out(group_event
, cpuctx
, ctx
);
1447 * Schedule out siblings (if any):
1449 list_for_each_entry(event
, &group_event
->sibling_list
, group_entry
)
1450 event_sched_out(event
, cpuctx
, ctx
);
1452 if (state
== PERF_EVENT_STATE_ACTIVE
&& group_event
->attr
.exclusive
)
1453 cpuctx
->exclusive
= 0;
1457 * Cross CPU call to remove a performance event
1459 * We disable the event on the hardware level first. After that we
1460 * remove it from the context list.
1462 static int __perf_remove_from_context(void *info
)
1464 struct perf_event
*event
= info
;
1465 struct perf_event_context
*ctx
= event
->ctx
;
1466 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
1468 raw_spin_lock(&ctx
->lock
);
1469 event_sched_out(event
, cpuctx
, ctx
);
1470 list_del_event(event
, ctx
);
1471 if (!ctx
->nr_events
&& cpuctx
->task_ctx
== ctx
) {
1473 cpuctx
->task_ctx
= NULL
;
1475 raw_spin_unlock(&ctx
->lock
);
1482 * Remove the event from a task's (or a CPU's) list of events.
1484 * CPU events are removed with a smp call. For task events we only
1485 * call when the task is on a CPU.
1487 * If event->ctx is a cloned context, callers must make sure that
1488 * every task struct that event->ctx->task could possibly point to
1489 * remains valid. This is OK when called from perf_release since
1490 * that only calls us on the top-level context, which can't be a clone.
1491 * When called from perf_event_exit_task, it's OK because the
1492 * context has been detached from its task.
1494 static void perf_remove_from_context(struct perf_event
*event
)
1496 struct perf_event_context
*ctx
= event
->ctx
;
1497 struct task_struct
*task
= ctx
->task
;
1499 lockdep_assert_held(&ctx
->mutex
);
1503 * Per cpu events are removed via an smp call and
1504 * the removal is always successful.
1506 cpu_function_call(event
->cpu
, __perf_remove_from_context
, event
);
1511 if (!task_function_call(task
, __perf_remove_from_context
, event
))
1514 raw_spin_lock_irq(&ctx
->lock
);
1516 * If we failed to find a running task, but find the context active now
1517 * that we've acquired the ctx->lock, retry.
1519 if (ctx
->is_active
) {
1520 raw_spin_unlock_irq(&ctx
->lock
);
1525 * Since the task isn't running, its safe to remove the event, us
1526 * holding the ctx->lock ensures the task won't get scheduled in.
1528 list_del_event(event
, ctx
);
1529 raw_spin_unlock_irq(&ctx
->lock
);
1533 * Cross CPU call to disable a performance event
1535 int __perf_event_disable(void *info
)
1537 struct perf_event
*event
= info
;
1538 struct perf_event_context
*ctx
= event
->ctx
;
1539 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
1542 * If this is a per-task event, need to check whether this
1543 * event's task is the current task on this cpu.
1545 * Can trigger due to concurrent perf_event_context_sched_out()
1546 * flipping contexts around.
1548 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
)
1551 raw_spin_lock(&ctx
->lock
);
1554 * If the event is on, turn it off.
1555 * If it is in error state, leave it in error state.
1557 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
) {
1558 update_context_time(ctx
);
1559 update_cgrp_time_from_event(event
);
1560 update_group_times(event
);
1561 if (event
== event
->group_leader
)
1562 group_sched_out(event
, cpuctx
, ctx
);
1564 event_sched_out(event
, cpuctx
, ctx
);
1565 event
->state
= PERF_EVENT_STATE_OFF
;
1568 raw_spin_unlock(&ctx
->lock
);
1576 * If event->ctx is a cloned context, callers must make sure that
1577 * every task struct that event->ctx->task could possibly point to
1578 * remains valid. This condition is satisifed when called through
1579 * perf_event_for_each_child or perf_event_for_each because they
1580 * hold the top-level event's child_mutex, so any descendant that
1581 * goes to exit will block in sync_child_event.
1582 * When called from perf_pending_event it's OK because event->ctx
1583 * is the current context on this CPU and preemption is disabled,
1584 * hence we can't get into perf_event_task_sched_out for this context.
1586 void perf_event_disable(struct perf_event
*event
)
1588 struct perf_event_context
*ctx
= event
->ctx
;
1589 struct task_struct
*task
= ctx
->task
;
1593 * Disable the event on the cpu that it's on
1595 cpu_function_call(event
->cpu
, __perf_event_disable
, event
);
1600 if (!task_function_call(task
, __perf_event_disable
, event
))
1603 raw_spin_lock_irq(&ctx
->lock
);
1605 * If the event is still active, we need to retry the cross-call.
1607 if (event
->state
== PERF_EVENT_STATE_ACTIVE
) {
1608 raw_spin_unlock_irq(&ctx
->lock
);
1610 * Reload the task pointer, it might have been changed by
1611 * a concurrent perf_event_context_sched_out().
1618 * Since we have the lock this context can't be scheduled
1619 * in, so we can change the state safely.
1621 if (event
->state
== PERF_EVENT_STATE_INACTIVE
) {
1622 update_group_times(event
);
1623 event
->state
= PERF_EVENT_STATE_OFF
;
1625 raw_spin_unlock_irq(&ctx
->lock
);
1627 EXPORT_SYMBOL_GPL(perf_event_disable
);
1629 static void perf_set_shadow_time(struct perf_event
*event
,
1630 struct perf_event_context
*ctx
,
1634 * use the correct time source for the time snapshot
1636 * We could get by without this by leveraging the
1637 * fact that to get to this function, the caller
1638 * has most likely already called update_context_time()
1639 * and update_cgrp_time_xx() and thus both timestamp
1640 * are identical (or very close). Given that tstamp is,
1641 * already adjusted for cgroup, we could say that:
1642 * tstamp - ctx->timestamp
1644 * tstamp - cgrp->timestamp.
1646 * Then, in perf_output_read(), the calculation would
1647 * work with no changes because:
1648 * - event is guaranteed scheduled in
1649 * - no scheduled out in between
1650 * - thus the timestamp would be the same
1652 * But this is a bit hairy.
1654 * So instead, we have an explicit cgroup call to remain
1655 * within the time time source all along. We believe it
1656 * is cleaner and simpler to understand.
1658 if (is_cgroup_event(event
))
1659 perf_cgroup_set_shadow_time(event
, tstamp
);
1661 event
->shadow_ctx_time
= tstamp
- ctx
->timestamp
;
1664 #define MAX_INTERRUPTS (~0ULL)
1666 static void perf_log_throttle(struct perf_event
*event
, int enable
);
1669 event_sched_in(struct perf_event
*event
,
1670 struct perf_cpu_context
*cpuctx
,
1671 struct perf_event_context
*ctx
)
1673 u64 tstamp
= perf_event_time(event
);
1676 if (event
->state
<= PERF_EVENT_STATE_OFF
)
1679 event
->state
= PERF_EVENT_STATE_ACTIVE
;
1680 event
->oncpu
= smp_processor_id();
1683 * Unthrottle events, since we scheduled we might have missed several
1684 * ticks already, also for a heavily scheduling task there is little
1685 * guarantee it'll get a tick in a timely manner.
1687 if (unlikely(event
->hw
.interrupts
== MAX_INTERRUPTS
)) {
1688 perf_log_throttle(event
, 1);
1689 event
->hw
.interrupts
= 0;
1693 * The new state must be visible before we turn it on in the hardware:
1697 perf_pmu_disable(event
->pmu
);
1699 if (event
->pmu
->add(event
, PERF_EF_START
)) {
1700 event
->state
= PERF_EVENT_STATE_INACTIVE
;
1706 event
->tstamp_running
+= tstamp
- event
->tstamp_stopped
;
1708 perf_set_shadow_time(event
, ctx
, tstamp
);
1710 if (!is_software_event(event
))
1711 cpuctx
->active_oncpu
++;
1713 if (event
->attr
.freq
&& event
->attr
.sample_freq
)
1716 if (event
->attr
.exclusive
)
1717 cpuctx
->exclusive
= 1;
1720 perf_pmu_enable(event
->pmu
);
1726 group_sched_in(struct perf_event
*group_event
,
1727 struct perf_cpu_context
*cpuctx
,
1728 struct perf_event_context
*ctx
)
1730 struct perf_event
*event
, *partial_group
= NULL
;
1731 struct pmu
*pmu
= group_event
->pmu
;
1732 u64 now
= ctx
->time
;
1733 bool simulate
= false;
1735 if (group_event
->state
== PERF_EVENT_STATE_OFF
)
1738 pmu
->start_txn(pmu
);
1740 if (event_sched_in(group_event
, cpuctx
, ctx
)) {
1741 pmu
->cancel_txn(pmu
);
1742 perf_cpu_hrtimer_restart(cpuctx
);
1747 * Schedule in siblings as one group (if any):
1749 list_for_each_entry(event
, &group_event
->sibling_list
, group_entry
) {
1750 if (event_sched_in(event
, cpuctx
, ctx
)) {
1751 partial_group
= event
;
1756 if (!pmu
->commit_txn(pmu
))
1761 * Groups can be scheduled in as one unit only, so undo any
1762 * partial group before returning:
1763 * The events up to the failed event are scheduled out normally,
1764 * tstamp_stopped will be updated.
1766 * The failed events and the remaining siblings need to have
1767 * their timings updated as if they had gone thru event_sched_in()
1768 * and event_sched_out(). This is required to get consistent timings
1769 * across the group. This also takes care of the case where the group
1770 * could never be scheduled by ensuring tstamp_stopped is set to mark
1771 * the time the event was actually stopped, such that time delta
1772 * calculation in update_event_times() is correct.
1774 list_for_each_entry(event
, &group_event
->sibling_list
, group_entry
) {
1775 if (event
== partial_group
)
1779 event
->tstamp_running
+= now
- event
->tstamp_stopped
;
1780 event
->tstamp_stopped
= now
;
1782 event_sched_out(event
, cpuctx
, ctx
);
1785 event_sched_out(group_event
, cpuctx
, ctx
);
1787 pmu
->cancel_txn(pmu
);
1789 perf_cpu_hrtimer_restart(cpuctx
);
1795 * Work out whether we can put this event group on the CPU now.
1797 static int group_can_go_on(struct perf_event
*event
,
1798 struct perf_cpu_context
*cpuctx
,
1802 * Groups consisting entirely of software events can always go on.
1804 if (event
->group_flags
& PERF_GROUP_SOFTWARE
)
1807 * If an exclusive group is already on, no other hardware
1810 if (cpuctx
->exclusive
)
1813 * If this group is exclusive and there are already
1814 * events on the CPU, it can't go on.
1816 if (event
->attr
.exclusive
&& cpuctx
->active_oncpu
)
1819 * Otherwise, try to add it if all previous groups were able
1825 static void add_event_to_ctx(struct perf_event
*event
,
1826 struct perf_event_context
*ctx
)
1828 u64 tstamp
= perf_event_time(event
);
1830 list_add_event(event
, ctx
);
1831 perf_group_attach(event
);
1832 event
->tstamp_enabled
= tstamp
;
1833 event
->tstamp_running
= tstamp
;
1834 event
->tstamp_stopped
= tstamp
;
1837 static void task_ctx_sched_out(struct perf_event_context
*ctx
);
1839 ctx_sched_in(struct perf_event_context
*ctx
,
1840 struct perf_cpu_context
*cpuctx
,
1841 enum event_type_t event_type
,
1842 struct task_struct
*task
);
1844 static void perf_event_sched_in(struct perf_cpu_context
*cpuctx
,
1845 struct perf_event_context
*ctx
,
1846 struct task_struct
*task
)
1848 cpu_ctx_sched_in(cpuctx
, EVENT_PINNED
, task
);
1850 ctx_sched_in(ctx
, cpuctx
, EVENT_PINNED
, task
);
1851 cpu_ctx_sched_in(cpuctx
, EVENT_FLEXIBLE
, task
);
1853 ctx_sched_in(ctx
, cpuctx
, EVENT_FLEXIBLE
, task
);
1857 * Cross CPU call to install and enable a performance event
1859 * Must be called with ctx->mutex held
1861 static int __perf_install_in_context(void *info
)
1863 struct perf_event
*event
= info
;
1864 struct perf_event_context
*ctx
= event
->ctx
;
1865 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
1866 struct perf_event_context
*task_ctx
= cpuctx
->task_ctx
;
1867 struct task_struct
*task
= current
;
1869 perf_ctx_lock(cpuctx
, task_ctx
);
1870 perf_pmu_disable(cpuctx
->ctx
.pmu
);
1873 * If there was an active task_ctx schedule it out.
1876 task_ctx_sched_out(task_ctx
);
1879 * If the context we're installing events in is not the
1880 * active task_ctx, flip them.
1882 if (ctx
->task
&& task_ctx
!= ctx
) {
1884 raw_spin_unlock(&task_ctx
->lock
);
1885 raw_spin_lock(&ctx
->lock
);
1890 cpuctx
->task_ctx
= task_ctx
;
1891 task
= task_ctx
->task
;
1894 cpu_ctx_sched_out(cpuctx
, EVENT_ALL
);
1896 update_context_time(ctx
);
1898 * update cgrp time only if current cgrp
1899 * matches event->cgrp. Must be done before
1900 * calling add_event_to_ctx()
1902 update_cgrp_time_from_event(event
);
1904 add_event_to_ctx(event
, ctx
);
1907 * Schedule everything back in
1909 perf_event_sched_in(cpuctx
, task_ctx
, task
);
1911 perf_pmu_enable(cpuctx
->ctx
.pmu
);
1912 perf_ctx_unlock(cpuctx
, task_ctx
);
1918 * Attach a performance event to a context
1920 * First we add the event to the list with the hardware enable bit
1921 * in event->hw_config cleared.
1923 * If the event is attached to a task which is on a CPU we use a smp
1924 * call to enable it in the task context. The task might have been
1925 * scheduled away, but we check this in the smp call again.
1928 perf_install_in_context(struct perf_event_context
*ctx
,
1929 struct perf_event
*event
,
1932 struct task_struct
*task
= ctx
->task
;
1934 lockdep_assert_held(&ctx
->mutex
);
1937 if (event
->cpu
!= -1)
1942 * Per cpu events are installed via an smp call and
1943 * the install is always successful.
1945 cpu_function_call(cpu
, __perf_install_in_context
, event
);
1950 if (!task_function_call(task
, __perf_install_in_context
, event
))
1953 raw_spin_lock_irq(&ctx
->lock
);
1955 * If we failed to find a running task, but find the context active now
1956 * that we've acquired the ctx->lock, retry.
1958 if (ctx
->is_active
) {
1959 raw_spin_unlock_irq(&ctx
->lock
);
1964 * Since the task isn't running, its safe to add the event, us holding
1965 * the ctx->lock ensures the task won't get scheduled in.
1967 add_event_to_ctx(event
, ctx
);
1968 raw_spin_unlock_irq(&ctx
->lock
);
1972 * Put a event into inactive state and update time fields.
1973 * Enabling the leader of a group effectively enables all
1974 * the group members that aren't explicitly disabled, so we
1975 * have to update their ->tstamp_enabled also.
1976 * Note: this works for group members as well as group leaders
1977 * since the non-leader members' sibling_lists will be empty.
1979 static void __perf_event_mark_enabled(struct perf_event
*event
)
1981 struct perf_event
*sub
;
1982 u64 tstamp
= perf_event_time(event
);
1984 event
->state
= PERF_EVENT_STATE_INACTIVE
;
1985 event
->tstamp_enabled
= tstamp
- event
->total_time_enabled
;
1986 list_for_each_entry(sub
, &event
->sibling_list
, group_entry
) {
1987 if (sub
->state
>= PERF_EVENT_STATE_INACTIVE
)
1988 sub
->tstamp_enabled
= tstamp
- sub
->total_time_enabled
;
1993 * Cross CPU call to enable a performance event
1995 static int __perf_event_enable(void *info
)
1997 struct perf_event
*event
= info
;
1998 struct perf_event_context
*ctx
= event
->ctx
;
1999 struct perf_event
*leader
= event
->group_leader
;
2000 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
2004 * There's a time window between 'ctx->is_active' check
2005 * in perf_event_enable function and this place having:
2007 * - ctx->lock unlocked
2009 * where the task could be killed and 'ctx' deactivated
2010 * by perf_event_exit_task.
2012 if (!ctx
->is_active
)
2015 raw_spin_lock(&ctx
->lock
);
2016 update_context_time(ctx
);
2018 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
)
2022 * set current task's cgroup time reference point
2024 perf_cgroup_set_timestamp(current
, ctx
);
2026 __perf_event_mark_enabled(event
);
2028 if (!event_filter_match(event
)) {
2029 if (is_cgroup_event(event
))
2030 perf_cgroup_defer_enabled(event
);
2035 * If the event is in a group and isn't the group leader,
2036 * then don't put it on unless the group is on.
2038 if (leader
!= event
&& leader
->state
!= PERF_EVENT_STATE_ACTIVE
)
2041 if (!group_can_go_on(event
, cpuctx
, 1)) {
2044 if (event
== leader
)
2045 err
= group_sched_in(event
, cpuctx
, ctx
);
2047 err
= event_sched_in(event
, cpuctx
, ctx
);
2052 * If this event can't go on and it's part of a
2053 * group, then the whole group has to come off.
2055 if (leader
!= event
) {
2056 group_sched_out(leader
, cpuctx
, ctx
);
2057 perf_cpu_hrtimer_restart(cpuctx
);
2059 if (leader
->attr
.pinned
) {
2060 update_group_times(leader
);
2061 leader
->state
= PERF_EVENT_STATE_ERROR
;
2066 raw_spin_unlock(&ctx
->lock
);
2074 * If event->ctx is a cloned context, callers must make sure that
2075 * every task struct that event->ctx->task could possibly point to
2076 * remains valid. This condition is satisfied when called through
2077 * perf_event_for_each_child or perf_event_for_each as described
2078 * for perf_event_disable.
2080 void perf_event_enable(struct perf_event
*event
)
2082 struct perf_event_context
*ctx
= event
->ctx
;
2083 struct task_struct
*task
= ctx
->task
;
2087 * Enable the event on the cpu that it's on
2089 cpu_function_call(event
->cpu
, __perf_event_enable
, event
);
2093 raw_spin_lock_irq(&ctx
->lock
);
2094 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
)
2098 * If the event is in error state, clear that first.
2099 * That way, if we see the event in error state below, we
2100 * know that it has gone back into error state, as distinct
2101 * from the task having been scheduled away before the
2102 * cross-call arrived.
2104 if (event
->state
== PERF_EVENT_STATE_ERROR
)
2105 event
->state
= PERF_EVENT_STATE_OFF
;
2108 if (!ctx
->is_active
) {
2109 __perf_event_mark_enabled(event
);
2113 raw_spin_unlock_irq(&ctx
->lock
);
2115 if (!task_function_call(task
, __perf_event_enable
, event
))
2118 raw_spin_lock_irq(&ctx
->lock
);
2121 * If the context is active and the event is still off,
2122 * we need to retry the cross-call.
2124 if (ctx
->is_active
&& event
->state
== PERF_EVENT_STATE_OFF
) {
2126 * task could have been flipped by a concurrent
2127 * perf_event_context_sched_out()
2134 raw_spin_unlock_irq(&ctx
->lock
);
2136 EXPORT_SYMBOL_GPL(perf_event_enable
);
2138 int perf_event_refresh(struct perf_event
*event
, int refresh
)
2141 * not supported on inherited events
2143 if (event
->attr
.inherit
|| !is_sampling_event(event
))
2146 atomic_add(refresh
, &event
->event_limit
);
2147 perf_event_enable(event
);
2151 EXPORT_SYMBOL_GPL(perf_event_refresh
);
2153 static void ctx_sched_out(struct perf_event_context
*ctx
,
2154 struct perf_cpu_context
*cpuctx
,
2155 enum event_type_t event_type
)
2157 struct perf_event
*event
;
2158 int is_active
= ctx
->is_active
;
2160 ctx
->is_active
&= ~event_type
;
2161 if (likely(!ctx
->nr_events
))
2164 update_context_time(ctx
);
2165 update_cgrp_time_from_cpuctx(cpuctx
);
2166 if (!ctx
->nr_active
)
2169 perf_pmu_disable(ctx
->pmu
);
2170 if ((is_active
& EVENT_PINNED
) && (event_type
& EVENT_PINNED
)) {
2171 list_for_each_entry(event
, &ctx
->pinned_groups
, group_entry
)
2172 group_sched_out(event
, cpuctx
, ctx
);
2175 if ((is_active
& EVENT_FLEXIBLE
) && (event_type
& EVENT_FLEXIBLE
)) {
2176 list_for_each_entry(event
, &ctx
->flexible_groups
, group_entry
)
2177 group_sched_out(event
, cpuctx
, ctx
);
2179 perf_pmu_enable(ctx
->pmu
);
2183 * Test whether two contexts are equivalent, i.e. whether they have both been
2184 * cloned from the same version of the same context.
2186 * Equivalence is measured using a generation number in the context that is
2187 * incremented on each modification to it; see unclone_ctx(), list_add_event()
2188 * and list_del_event().
2190 static int context_equiv(struct perf_event_context
*ctx1
,
2191 struct perf_event_context
*ctx2
)
2193 /* Pinning disables the swap optimization */
2194 if (ctx1
->pin_count
|| ctx2
->pin_count
)
2197 /* If ctx1 is the parent of ctx2 */
2198 if (ctx1
== ctx2
->parent_ctx
&& ctx1
->generation
== ctx2
->parent_gen
)
2201 /* If ctx2 is the parent of ctx1 */
2202 if (ctx1
->parent_ctx
== ctx2
&& ctx1
->parent_gen
== ctx2
->generation
)
2206 * If ctx1 and ctx2 have the same parent; we flatten the parent
2207 * hierarchy, see perf_event_init_context().
2209 if (ctx1
->parent_ctx
&& ctx1
->parent_ctx
== ctx2
->parent_ctx
&&
2210 ctx1
->parent_gen
== ctx2
->parent_gen
)
2217 static void __perf_event_sync_stat(struct perf_event
*event
,
2218 struct perf_event
*next_event
)
2222 if (!event
->attr
.inherit_stat
)
2226 * Update the event value, we cannot use perf_event_read()
2227 * because we're in the middle of a context switch and have IRQs
2228 * disabled, which upsets smp_call_function_single(), however
2229 * we know the event must be on the current CPU, therefore we
2230 * don't need to use it.
2232 switch (event
->state
) {
2233 case PERF_EVENT_STATE_ACTIVE
:
2234 event
->pmu
->read(event
);
2237 case PERF_EVENT_STATE_INACTIVE
:
2238 update_event_times(event
);
2246 * In order to keep per-task stats reliable we need to flip the event
2247 * values when we flip the contexts.
2249 value
= local64_read(&next_event
->count
);
2250 value
= local64_xchg(&event
->count
, value
);
2251 local64_set(&next_event
->count
, value
);
2253 swap(event
->total_time_enabled
, next_event
->total_time_enabled
);
2254 swap(event
->total_time_running
, next_event
->total_time_running
);
2257 * Since we swizzled the values, update the user visible data too.
2259 perf_event_update_userpage(event
);
2260 perf_event_update_userpage(next_event
);
2263 static void perf_event_sync_stat(struct perf_event_context
*ctx
,
2264 struct perf_event_context
*next_ctx
)
2266 struct perf_event
*event
, *next_event
;
2271 update_context_time(ctx
);
2273 event
= list_first_entry(&ctx
->event_list
,
2274 struct perf_event
, event_entry
);
2276 next_event
= list_first_entry(&next_ctx
->event_list
,
2277 struct perf_event
, event_entry
);
2279 while (&event
->event_entry
!= &ctx
->event_list
&&
2280 &next_event
->event_entry
!= &next_ctx
->event_list
) {
2282 __perf_event_sync_stat(event
, next_event
);
2284 event
= list_next_entry(event
, event_entry
);
2285 next_event
= list_next_entry(next_event
, event_entry
);
2289 static void perf_event_context_sched_out(struct task_struct
*task
, int ctxn
,
2290 struct task_struct
*next
)
2292 struct perf_event_context
*ctx
= task
->perf_event_ctxp
[ctxn
];
2293 struct perf_event_context
*next_ctx
;
2294 struct perf_event_context
*parent
, *next_parent
;
2295 struct perf_cpu_context
*cpuctx
;
2301 cpuctx
= __get_cpu_context(ctx
);
2302 if (!cpuctx
->task_ctx
)
2306 next_ctx
= next
->perf_event_ctxp
[ctxn
];
2310 parent
= rcu_dereference(ctx
->parent_ctx
);
2311 next_parent
= rcu_dereference(next_ctx
->parent_ctx
);
2313 /* If neither context have a parent context; they cannot be clones. */
2314 if (!parent
&& !next_parent
)
2317 if (next_parent
== ctx
|| next_ctx
== parent
|| next_parent
== parent
) {
2319 * Looks like the two contexts are clones, so we might be
2320 * able to optimize the context switch. We lock both
2321 * contexts and check that they are clones under the
2322 * lock (including re-checking that neither has been
2323 * uncloned in the meantime). It doesn't matter which
2324 * order we take the locks because no other cpu could
2325 * be trying to lock both of these tasks.
2327 raw_spin_lock(&ctx
->lock
);
2328 raw_spin_lock_nested(&next_ctx
->lock
, SINGLE_DEPTH_NESTING
);
2329 if (context_equiv(ctx
, next_ctx
)) {
2331 * XXX do we need a memory barrier of sorts
2332 * wrt to rcu_dereference() of perf_event_ctxp
2334 task
->perf_event_ctxp
[ctxn
] = next_ctx
;
2335 next
->perf_event_ctxp
[ctxn
] = ctx
;
2337 next_ctx
->task
= task
;
2340 perf_event_sync_stat(ctx
, next_ctx
);
2342 raw_spin_unlock(&next_ctx
->lock
);
2343 raw_spin_unlock(&ctx
->lock
);
2349 raw_spin_lock(&ctx
->lock
);
2350 ctx_sched_out(ctx
, cpuctx
, EVENT_ALL
);
2351 cpuctx
->task_ctx
= NULL
;
2352 raw_spin_unlock(&ctx
->lock
);
2356 #define for_each_task_context_nr(ctxn) \
2357 for ((ctxn) = 0; (ctxn) < perf_nr_task_contexts; (ctxn)++)
2360 * Called from scheduler to remove the events of the current task,
2361 * with interrupts disabled.
2363 * We stop each event and update the event value in event->count.
2365 * This does not protect us against NMI, but disable()
2366 * sets the disabled bit in the control field of event _before_
2367 * accessing the event control register. If a NMI hits, then it will
2368 * not restart the event.
2370 void __perf_event_task_sched_out(struct task_struct
*task
,
2371 struct task_struct
*next
)
2375 for_each_task_context_nr(ctxn
)
2376 perf_event_context_sched_out(task
, ctxn
, next
);
2379 * if cgroup events exist on this CPU, then we need
2380 * to check if we have to switch out PMU state.
2381 * cgroup event are system-wide mode only
2383 if (atomic_read(&__get_cpu_var(perf_cgroup_events
)))
2384 perf_cgroup_sched_out(task
, next
);
2387 static void task_ctx_sched_out(struct perf_event_context
*ctx
)
2389 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
2391 if (!cpuctx
->task_ctx
)
2394 if (WARN_ON_ONCE(ctx
!= cpuctx
->task_ctx
))
2397 ctx_sched_out(ctx
, cpuctx
, EVENT_ALL
);
2398 cpuctx
->task_ctx
= NULL
;
2402 * Called with IRQs disabled
2404 static void cpu_ctx_sched_out(struct perf_cpu_context
*cpuctx
,
2405 enum event_type_t event_type
)
2407 ctx_sched_out(&cpuctx
->ctx
, cpuctx
, event_type
);
2411 ctx_pinned_sched_in(struct perf_event_context
*ctx
,
2412 struct perf_cpu_context
*cpuctx
)
2414 struct perf_event
*event
;
2416 list_for_each_entry(event
, &ctx
->pinned_groups
, group_entry
) {
2417 if (event
->state
<= PERF_EVENT_STATE_OFF
)
2419 if (!event_filter_match(event
))
2422 /* may need to reset tstamp_enabled */
2423 if (is_cgroup_event(event
))
2424 perf_cgroup_mark_enabled(event
, ctx
);
2426 if (group_can_go_on(event
, cpuctx
, 1))
2427 group_sched_in(event
, cpuctx
, ctx
);
2430 * If this pinned group hasn't been scheduled,
2431 * put it in error state.
2433 if (event
->state
== PERF_EVENT_STATE_INACTIVE
) {
2434 update_group_times(event
);
2435 event
->state
= PERF_EVENT_STATE_ERROR
;
2441 ctx_flexible_sched_in(struct perf_event_context
*ctx
,
2442 struct perf_cpu_context
*cpuctx
)
2444 struct perf_event
*event
;
2447 list_for_each_entry(event
, &ctx
->flexible_groups
, group_entry
) {
2448 /* Ignore events in OFF or ERROR state */
2449 if (event
->state
<= PERF_EVENT_STATE_OFF
)
2452 * Listen to the 'cpu' scheduling filter constraint
2455 if (!event_filter_match(event
))
2458 /* may need to reset tstamp_enabled */
2459 if (is_cgroup_event(event
))
2460 perf_cgroup_mark_enabled(event
, ctx
);
2462 if (group_can_go_on(event
, cpuctx
, can_add_hw
)) {
2463 if (group_sched_in(event
, cpuctx
, ctx
))
2470 ctx_sched_in(struct perf_event_context
*ctx
,
2471 struct perf_cpu_context
*cpuctx
,
2472 enum event_type_t event_type
,
2473 struct task_struct
*task
)
2476 int is_active
= ctx
->is_active
;
2478 ctx
->is_active
|= event_type
;
2479 if (likely(!ctx
->nr_events
))
2483 ctx
->timestamp
= now
;
2484 perf_cgroup_set_timestamp(task
, ctx
);
2486 * First go through the list and put on any pinned groups
2487 * in order to give them the best chance of going on.
2489 if (!(is_active
& EVENT_PINNED
) && (event_type
& EVENT_PINNED
))
2490 ctx_pinned_sched_in(ctx
, cpuctx
);
2492 /* Then walk through the lower prio flexible groups */
2493 if (!(is_active
& EVENT_FLEXIBLE
) && (event_type
& EVENT_FLEXIBLE
))
2494 ctx_flexible_sched_in(ctx
, cpuctx
);
2497 static void cpu_ctx_sched_in(struct perf_cpu_context
*cpuctx
,
2498 enum event_type_t event_type
,
2499 struct task_struct
*task
)
2501 struct perf_event_context
*ctx
= &cpuctx
->ctx
;
2503 ctx_sched_in(ctx
, cpuctx
, event_type
, task
);
2506 static void perf_event_context_sched_in(struct perf_event_context
*ctx
,
2507 struct task_struct
*task
)
2509 struct perf_cpu_context
*cpuctx
;
2511 cpuctx
= __get_cpu_context(ctx
);
2512 if (cpuctx
->task_ctx
== ctx
)
2515 perf_ctx_lock(cpuctx
, ctx
);
2516 perf_pmu_disable(ctx
->pmu
);
2518 * We want to keep the following priority order:
2519 * cpu pinned (that don't need to move), task pinned,
2520 * cpu flexible, task flexible.
2522 cpu_ctx_sched_out(cpuctx
, EVENT_FLEXIBLE
);
2525 cpuctx
->task_ctx
= ctx
;
2527 perf_event_sched_in(cpuctx
, cpuctx
->task_ctx
, task
);
2529 perf_pmu_enable(ctx
->pmu
);
2530 perf_ctx_unlock(cpuctx
, ctx
);
2533 * Since these rotations are per-cpu, we need to ensure the
2534 * cpu-context we got scheduled on is actually rotating.
2536 perf_pmu_rotate_start(ctx
->pmu
);
2540 * When sampling the branck stack in system-wide, it may be necessary
2541 * to flush the stack on context switch. This happens when the branch
2542 * stack does not tag its entries with the pid of the current task.
2543 * Otherwise it becomes impossible to associate a branch entry with a
2544 * task. This ambiguity is more likely to appear when the branch stack
2545 * supports priv level filtering and the user sets it to monitor only
2546 * at the user level (which could be a useful measurement in system-wide
2547 * mode). In that case, the risk is high of having a branch stack with
2548 * branch from multiple tasks. Flushing may mean dropping the existing
2549 * entries or stashing them somewhere in the PMU specific code layer.
2551 * This function provides the context switch callback to the lower code
2552 * layer. It is invoked ONLY when there is at least one system-wide context
2553 * with at least one active event using taken branch sampling.
2555 static void perf_branch_stack_sched_in(struct task_struct
*prev
,
2556 struct task_struct
*task
)
2558 struct perf_cpu_context
*cpuctx
;
2560 unsigned long flags
;
2562 /* no need to flush branch stack if not changing task */
2566 local_irq_save(flags
);
2570 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
2571 cpuctx
= this_cpu_ptr(pmu
->pmu_cpu_context
);
2574 * check if the context has at least one
2575 * event using PERF_SAMPLE_BRANCH_STACK
2577 if (cpuctx
->ctx
.nr_branch_stack
> 0
2578 && pmu
->flush_branch_stack
) {
2580 pmu
= cpuctx
->ctx
.pmu
;
2582 perf_ctx_lock(cpuctx
, cpuctx
->task_ctx
);
2584 perf_pmu_disable(pmu
);
2586 pmu
->flush_branch_stack();
2588 perf_pmu_enable(pmu
);
2590 perf_ctx_unlock(cpuctx
, cpuctx
->task_ctx
);
2596 local_irq_restore(flags
);
2600 * Called from scheduler to add the events of the current task
2601 * with interrupts disabled.
2603 * We restore the event value and then enable it.
2605 * This does not protect us against NMI, but enable()
2606 * sets the enabled bit in the control field of event _before_
2607 * accessing the event control register. If a NMI hits, then it will
2608 * keep the event running.
2610 void __perf_event_task_sched_in(struct task_struct
*prev
,
2611 struct task_struct
*task
)
2613 struct perf_event_context
*ctx
;
2616 for_each_task_context_nr(ctxn
) {
2617 ctx
= task
->perf_event_ctxp
[ctxn
];
2621 perf_event_context_sched_in(ctx
, task
);
2624 * if cgroup events exist on this CPU, then we need
2625 * to check if we have to switch in PMU state.
2626 * cgroup event are system-wide mode only
2628 if (atomic_read(&__get_cpu_var(perf_cgroup_events
)))
2629 perf_cgroup_sched_in(prev
, task
);
2631 /* check for system-wide branch_stack events */
2632 if (atomic_read(&__get_cpu_var(perf_branch_stack_events
)))
2633 perf_branch_stack_sched_in(prev
, task
);
2636 static u64
perf_calculate_period(struct perf_event
*event
, u64 nsec
, u64 count
)
2638 u64 frequency
= event
->attr
.sample_freq
;
2639 u64 sec
= NSEC_PER_SEC
;
2640 u64 divisor
, dividend
;
2642 int count_fls
, nsec_fls
, frequency_fls
, sec_fls
;
2644 count_fls
= fls64(count
);
2645 nsec_fls
= fls64(nsec
);
2646 frequency_fls
= fls64(frequency
);
2650 * We got @count in @nsec, with a target of sample_freq HZ
2651 * the target period becomes:
2654 * period = -------------------
2655 * @nsec * sample_freq
2660 * Reduce accuracy by one bit such that @a and @b converge
2661 * to a similar magnitude.
2663 #define REDUCE_FLS(a, b) \
2665 if (a##_fls > b##_fls) { \
2675 * Reduce accuracy until either term fits in a u64, then proceed with
2676 * the other, so that finally we can do a u64/u64 division.
2678 while (count_fls
+ sec_fls
> 64 && nsec_fls
+ frequency_fls
> 64) {
2679 REDUCE_FLS(nsec
, frequency
);
2680 REDUCE_FLS(sec
, count
);
2683 if (count_fls
+ sec_fls
> 64) {
2684 divisor
= nsec
* frequency
;
2686 while (count_fls
+ sec_fls
> 64) {
2687 REDUCE_FLS(count
, sec
);
2691 dividend
= count
* sec
;
2693 dividend
= count
* sec
;
2695 while (nsec_fls
+ frequency_fls
> 64) {
2696 REDUCE_FLS(nsec
, frequency
);
2700 divisor
= nsec
* frequency
;
2706 return div64_u64(dividend
, divisor
);
2709 static DEFINE_PER_CPU(int, perf_throttled_count
);
2710 static DEFINE_PER_CPU(u64
, perf_throttled_seq
);
2712 static void perf_adjust_period(struct perf_event
*event
, u64 nsec
, u64 count
, bool disable
)
2714 struct hw_perf_event
*hwc
= &event
->hw
;
2715 s64 period
, sample_period
;
2718 period
= perf_calculate_period(event
, nsec
, count
);
2720 delta
= (s64
)(period
- hwc
->sample_period
);
2721 delta
= (delta
+ 7) / 8; /* low pass filter */
2723 sample_period
= hwc
->sample_period
+ delta
;
2728 hwc
->sample_period
= sample_period
;
2730 if (local64_read(&hwc
->period_left
) > 8*sample_period
) {
2732 event
->pmu
->stop(event
, PERF_EF_UPDATE
);
2734 local64_set(&hwc
->period_left
, 0);
2737 event
->pmu
->start(event
, PERF_EF_RELOAD
);
2742 * combine freq adjustment with unthrottling to avoid two passes over the
2743 * events. At the same time, make sure, having freq events does not change
2744 * the rate of unthrottling as that would introduce bias.
2746 static void perf_adjust_freq_unthr_context(struct perf_event_context
*ctx
,
2749 struct perf_event
*event
;
2750 struct hw_perf_event
*hwc
;
2751 u64 now
, period
= TICK_NSEC
;
2755 * only need to iterate over all events iff:
2756 * - context have events in frequency mode (needs freq adjust)
2757 * - there are events to unthrottle on this cpu
2759 if (!(ctx
->nr_freq
|| needs_unthr
))
2762 raw_spin_lock(&ctx
->lock
);
2763 perf_pmu_disable(ctx
->pmu
);
2765 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
2766 if (event
->state
!= PERF_EVENT_STATE_ACTIVE
)
2769 if (!event_filter_match(event
))
2772 perf_pmu_disable(event
->pmu
);
2776 if (hwc
->interrupts
== MAX_INTERRUPTS
) {
2777 hwc
->interrupts
= 0;
2778 perf_log_throttle(event
, 1);
2779 event
->pmu
->start(event
, 0);
2782 if (!event
->attr
.freq
|| !event
->attr
.sample_freq
)
2786 * stop the event and update event->count
2788 event
->pmu
->stop(event
, PERF_EF_UPDATE
);
2790 now
= local64_read(&event
->count
);
2791 delta
= now
- hwc
->freq_count_stamp
;
2792 hwc
->freq_count_stamp
= now
;
2796 * reload only if value has changed
2797 * we have stopped the event so tell that
2798 * to perf_adjust_period() to avoid stopping it
2802 perf_adjust_period(event
, period
, delta
, false);
2804 event
->pmu
->start(event
, delta
> 0 ? PERF_EF_RELOAD
: 0);
2806 perf_pmu_enable(event
->pmu
);
2809 perf_pmu_enable(ctx
->pmu
);
2810 raw_spin_unlock(&ctx
->lock
);
2814 * Round-robin a context's events:
2816 static void rotate_ctx(struct perf_event_context
*ctx
)
2819 * Rotate the first entry last of non-pinned groups. Rotation might be
2820 * disabled by the inheritance code.
2822 if (!ctx
->rotate_disable
)
2823 list_rotate_left(&ctx
->flexible_groups
);
2827 * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
2828 * because they're strictly cpu affine and rotate_start is called with IRQs
2829 * disabled, while rotate_context is called from IRQ context.
2831 static int perf_rotate_context(struct perf_cpu_context
*cpuctx
)
2833 struct perf_event_context
*ctx
= NULL
;
2834 int rotate
= 0, remove
= 1;
2836 if (cpuctx
->ctx
.nr_events
) {
2838 if (cpuctx
->ctx
.nr_events
!= cpuctx
->ctx
.nr_active
)
2842 ctx
= cpuctx
->task_ctx
;
2843 if (ctx
&& ctx
->nr_events
) {
2845 if (ctx
->nr_events
!= ctx
->nr_active
)
2852 perf_ctx_lock(cpuctx
, cpuctx
->task_ctx
);
2853 perf_pmu_disable(cpuctx
->ctx
.pmu
);
2855 cpu_ctx_sched_out(cpuctx
, EVENT_FLEXIBLE
);
2857 ctx_sched_out(ctx
, cpuctx
, EVENT_FLEXIBLE
);
2859 rotate_ctx(&cpuctx
->ctx
);
2863 perf_event_sched_in(cpuctx
, ctx
, current
);
2865 perf_pmu_enable(cpuctx
->ctx
.pmu
);
2866 perf_ctx_unlock(cpuctx
, cpuctx
->task_ctx
);
2869 list_del_init(&cpuctx
->rotation_list
);
2874 #ifdef CONFIG_NO_HZ_FULL
2875 bool perf_event_can_stop_tick(void)
2877 if (atomic_read(&nr_freq_events
) ||
2878 __this_cpu_read(perf_throttled_count
))
2885 void perf_event_task_tick(void)
2887 struct list_head
*head
= &__get_cpu_var(rotation_list
);
2888 struct perf_cpu_context
*cpuctx
, *tmp
;
2889 struct perf_event_context
*ctx
;
2892 WARN_ON(!irqs_disabled());
2894 __this_cpu_inc(perf_throttled_seq
);
2895 throttled
= __this_cpu_xchg(perf_throttled_count
, 0);
2897 list_for_each_entry_safe(cpuctx
, tmp
, head
, rotation_list
) {
2899 perf_adjust_freq_unthr_context(ctx
, throttled
);
2901 ctx
= cpuctx
->task_ctx
;
2903 perf_adjust_freq_unthr_context(ctx
, throttled
);
2907 static int event_enable_on_exec(struct perf_event
*event
,
2908 struct perf_event_context
*ctx
)
2910 if (!event
->attr
.enable_on_exec
)
2913 event
->attr
.enable_on_exec
= 0;
2914 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
)
2917 __perf_event_mark_enabled(event
);
2923 * Enable all of a task's events that have been marked enable-on-exec.
2924 * This expects task == current.
2926 static void perf_event_enable_on_exec(struct perf_event_context
*ctx
)
2928 struct perf_event
*event
;
2929 unsigned long flags
;
2933 local_irq_save(flags
);
2934 if (!ctx
|| !ctx
->nr_events
)
2938 * We must ctxsw out cgroup events to avoid conflict
2939 * when invoking perf_task_event_sched_in() later on
2940 * in this function. Otherwise we end up trying to
2941 * ctxswin cgroup events which are already scheduled
2944 perf_cgroup_sched_out(current
, NULL
);
2946 raw_spin_lock(&ctx
->lock
);
2947 task_ctx_sched_out(ctx
);
2949 list_for_each_entry(event
, &ctx
->event_list
, event_entry
) {
2950 ret
= event_enable_on_exec(event
, ctx
);
2956 * Unclone this context if we enabled any event.
2961 raw_spin_unlock(&ctx
->lock
);
2964 * Also calls ctxswin for cgroup events, if any:
2966 perf_event_context_sched_in(ctx
, ctx
->task
);
2968 local_irq_restore(flags
);
2972 * Cross CPU call to read the hardware event
2974 static void __perf_event_read(void *info
)
2976 struct perf_event
*event
= info
;
2977 struct perf_event_context
*ctx
= event
->ctx
;
2978 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
2981 * If this is a task context, we need to check whether it is
2982 * the current task context of this cpu. If not it has been
2983 * scheduled out before the smp call arrived. In that case
2984 * event->count would have been updated to a recent sample
2985 * when the event was scheduled out.
2987 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
)
2990 raw_spin_lock(&ctx
->lock
);
2991 if (ctx
->is_active
) {
2992 update_context_time(ctx
);
2993 update_cgrp_time_from_event(event
);
2995 update_event_times(event
);
2996 if (event
->state
== PERF_EVENT_STATE_ACTIVE
)
2997 event
->pmu
->read(event
);
2998 raw_spin_unlock(&ctx
->lock
);
3001 static inline u64
perf_event_count(struct perf_event
*event
)
3003 return local64_read(&event
->count
) + atomic64_read(&event
->child_count
);
3006 static u64
perf_event_read(struct perf_event
*event
)
3009 * If event is enabled and currently active on a CPU, update the
3010 * value in the event structure:
3012 if (event
->state
== PERF_EVENT_STATE_ACTIVE
) {
3013 smp_call_function_single(event
->oncpu
,
3014 __perf_event_read
, event
, 1);
3015 } else if (event
->state
== PERF_EVENT_STATE_INACTIVE
) {
3016 struct perf_event_context
*ctx
= event
->ctx
;
3017 unsigned long flags
;
3019 raw_spin_lock_irqsave(&ctx
->lock
, flags
);
3021 * may read while context is not active
3022 * (e.g., thread is blocked), in that case
3023 * we cannot update context time
3025 if (ctx
->is_active
) {
3026 update_context_time(ctx
);
3027 update_cgrp_time_from_event(event
);
3029 update_event_times(event
);
3030 raw_spin_unlock_irqrestore(&ctx
->lock
, flags
);
3033 return perf_event_count(event
);
3037 * Initialize the perf_event context in a task_struct:
3039 static void __perf_event_init_context(struct perf_event_context
*ctx
)
3041 raw_spin_lock_init(&ctx
->lock
);
3042 mutex_init(&ctx
->mutex
);
3043 INIT_LIST_HEAD(&ctx
->pinned_groups
);
3044 INIT_LIST_HEAD(&ctx
->flexible_groups
);
3045 INIT_LIST_HEAD(&ctx
->event_list
);
3046 atomic_set(&ctx
->refcount
, 1);
3049 static struct perf_event_context
*
3050 alloc_perf_context(struct pmu
*pmu
, struct task_struct
*task
)
3052 struct perf_event_context
*ctx
;
3054 ctx
= kzalloc(sizeof(struct perf_event_context
), GFP_KERNEL
);
3058 __perf_event_init_context(ctx
);
3061 get_task_struct(task
);
3068 static struct task_struct
*
3069 find_lively_task_by_vpid(pid_t vpid
)
3071 struct task_struct
*task
;
3078 task
= find_task_by_vpid(vpid
);
3080 get_task_struct(task
);
3084 return ERR_PTR(-ESRCH
);
3086 /* Reuse ptrace permission checks for now. */
3088 if (!ptrace_may_access(task
, PTRACE_MODE_READ
))
3093 put_task_struct(task
);
3094 return ERR_PTR(err
);
3099 * Returns a matching context with refcount and pincount.
3101 static struct perf_event_context
*
3102 find_get_context(struct pmu
*pmu
, struct task_struct
*task
, int cpu
)
3104 struct perf_event_context
*ctx
;
3105 struct perf_cpu_context
*cpuctx
;
3106 unsigned long flags
;
3110 /* Must be root to operate on a CPU event: */
3111 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN
))
3112 return ERR_PTR(-EACCES
);
3115 * We could be clever and allow to attach a event to an
3116 * offline CPU and activate it when the CPU comes up, but
3119 if (!cpu_online(cpu
))
3120 return ERR_PTR(-ENODEV
);
3122 cpuctx
= per_cpu_ptr(pmu
->pmu_cpu_context
, cpu
);
3131 ctxn
= pmu
->task_ctx_nr
;
3136 ctx
= perf_lock_task_context(task
, ctxn
, &flags
);
3140 raw_spin_unlock_irqrestore(&ctx
->lock
, flags
);
3142 ctx
= alloc_perf_context(pmu
, task
);
3148 mutex_lock(&task
->perf_event_mutex
);
3150 * If it has already passed perf_event_exit_task().
3151 * we must see PF_EXITING, it takes this mutex too.
3153 if (task
->flags
& PF_EXITING
)
3155 else if (task
->perf_event_ctxp
[ctxn
])
3160 rcu_assign_pointer(task
->perf_event_ctxp
[ctxn
], ctx
);
3162 mutex_unlock(&task
->perf_event_mutex
);
3164 if (unlikely(err
)) {
3176 return ERR_PTR(err
);
3179 static void perf_event_free_filter(struct perf_event
*event
);
3181 static void free_event_rcu(struct rcu_head
*head
)
3183 struct perf_event
*event
;
3185 event
= container_of(head
, struct perf_event
, rcu_head
);
3187 put_pid_ns(event
->ns
);
3188 perf_event_free_filter(event
);
3192 static void ring_buffer_put(struct ring_buffer
*rb
);
3193 static void ring_buffer_detach(struct perf_event
*event
, struct ring_buffer
*rb
);
3195 static void unaccount_event_cpu(struct perf_event
*event
, int cpu
)
3200 if (has_branch_stack(event
)) {
3201 if (!(event
->attach_state
& PERF_ATTACH_TASK
))
3202 atomic_dec(&per_cpu(perf_branch_stack_events
, cpu
));
3204 if (is_cgroup_event(event
))
3205 atomic_dec(&per_cpu(perf_cgroup_events
, cpu
));
3208 static void unaccount_event(struct perf_event
*event
)
3213 if (event
->attach_state
& PERF_ATTACH_TASK
)
3214 static_key_slow_dec_deferred(&perf_sched_events
);
3215 if (event
->attr
.mmap
|| event
->attr
.mmap_data
)
3216 atomic_dec(&nr_mmap_events
);
3217 if (event
->attr
.comm
)
3218 atomic_dec(&nr_comm_events
);
3219 if (event
->attr
.task
)
3220 atomic_dec(&nr_task_events
);
3221 if (event
->attr
.freq
)
3222 atomic_dec(&nr_freq_events
);
3223 if (is_cgroup_event(event
))
3224 static_key_slow_dec_deferred(&perf_sched_events
);
3225 if (has_branch_stack(event
))
3226 static_key_slow_dec_deferred(&perf_sched_events
);
3228 unaccount_event_cpu(event
, event
->cpu
);
3231 static void __free_event(struct perf_event
*event
)
3233 if (!event
->parent
) {
3234 if (event
->attr
.sample_type
& PERF_SAMPLE_CALLCHAIN
)
3235 put_callchain_buffers();
3239 event
->destroy(event
);
3242 put_ctx(event
->ctx
);
3244 call_rcu(&event
->rcu_head
, free_event_rcu
);
3246 static void free_event(struct perf_event
*event
)
3248 irq_work_sync(&event
->pending
);
3250 unaccount_event(event
);
3253 struct ring_buffer
*rb
;
3256 * Can happen when we close an event with re-directed output.
3258 * Since we have a 0 refcount, perf_mmap_close() will skip
3259 * over us; possibly making our ring_buffer_put() the last.
3261 mutex_lock(&event
->mmap_mutex
);
3264 rcu_assign_pointer(event
->rb
, NULL
);
3265 ring_buffer_detach(event
, rb
);
3266 ring_buffer_put(rb
); /* could be last */
3268 mutex_unlock(&event
->mmap_mutex
);
3271 if (is_cgroup_event(event
))
3272 perf_detach_cgroup(event
);
3275 __free_event(event
);
3278 int perf_event_release_kernel(struct perf_event
*event
)
3280 struct perf_event_context
*ctx
= event
->ctx
;
3282 WARN_ON_ONCE(ctx
->parent_ctx
);
3284 * There are two ways this annotation is useful:
3286 * 1) there is a lock recursion from perf_event_exit_task
3287 * see the comment there.
3289 * 2) there is a lock-inversion with mmap_sem through
3290 * perf_event_read_group(), which takes faults while
3291 * holding ctx->mutex, however this is called after
3292 * the last filedesc died, so there is no possibility
3293 * to trigger the AB-BA case.
3295 mutex_lock_nested(&ctx
->mutex
, SINGLE_DEPTH_NESTING
);
3296 raw_spin_lock_irq(&ctx
->lock
);
3297 perf_group_detach(event
);
3298 raw_spin_unlock_irq(&ctx
->lock
);
3299 perf_remove_from_context(event
);
3300 mutex_unlock(&ctx
->mutex
);
3306 EXPORT_SYMBOL_GPL(perf_event_release_kernel
);
3309 * Called when the last reference to the file is gone.
3311 static void put_event(struct perf_event
*event
)
3313 struct task_struct
*owner
;
3315 if (!atomic_long_dec_and_test(&event
->refcount
))
3319 owner
= ACCESS_ONCE(event
->owner
);
3321 * Matches the smp_wmb() in perf_event_exit_task(). If we observe
3322 * !owner it means the list deletion is complete and we can indeed
3323 * free this event, otherwise we need to serialize on
3324 * owner->perf_event_mutex.
3326 smp_read_barrier_depends();
3329 * Since delayed_put_task_struct() also drops the last
3330 * task reference we can safely take a new reference
3331 * while holding the rcu_read_lock().
3333 get_task_struct(owner
);
3338 mutex_lock(&owner
->perf_event_mutex
);
3340 * We have to re-check the event->owner field, if it is cleared
3341 * we raced with perf_event_exit_task(), acquiring the mutex
3342 * ensured they're done, and we can proceed with freeing the
3346 list_del_init(&event
->owner_entry
);
3347 mutex_unlock(&owner
->perf_event_mutex
);
3348 put_task_struct(owner
);
3351 perf_event_release_kernel(event
);
3354 static int perf_release(struct inode
*inode
, struct file
*file
)
3356 put_event(file
->private_data
);
3360 u64
perf_event_read_value(struct perf_event
*event
, u64
*enabled
, u64
*running
)
3362 struct perf_event
*child
;
3368 mutex_lock(&event
->child_mutex
);
3369 total
+= perf_event_read(event
);
3370 *enabled
+= event
->total_time_enabled
+
3371 atomic64_read(&event
->child_total_time_enabled
);
3372 *running
+= event
->total_time_running
+
3373 atomic64_read(&event
->child_total_time_running
);
3375 list_for_each_entry(child
, &event
->child_list
, child_list
) {
3376 total
+= perf_event_read(child
);
3377 *enabled
+= child
->total_time_enabled
;
3378 *running
+= child
->total_time_running
;
3380 mutex_unlock(&event
->child_mutex
);
3384 EXPORT_SYMBOL_GPL(perf_event_read_value
);
3386 static int perf_event_read_group(struct perf_event
*event
,
3387 u64 read_format
, char __user
*buf
)
3389 struct perf_event
*leader
= event
->group_leader
, *sub
;
3390 int n
= 0, size
= 0, ret
= -EFAULT
;
3391 struct perf_event_context
*ctx
= leader
->ctx
;
3393 u64 count
, enabled
, running
;
3395 mutex_lock(&ctx
->mutex
);
3396 count
= perf_event_read_value(leader
, &enabled
, &running
);
3398 values
[n
++] = 1 + leader
->nr_siblings
;
3399 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
3400 values
[n
++] = enabled
;
3401 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
3402 values
[n
++] = running
;
3403 values
[n
++] = count
;
3404 if (read_format
& PERF_FORMAT_ID
)
3405 values
[n
++] = primary_event_id(leader
);
3407 size
= n
* sizeof(u64
);
3409 if (copy_to_user(buf
, values
, size
))
3414 list_for_each_entry(sub
, &leader
->sibling_list
, group_entry
) {
3417 values
[n
++] = perf_event_read_value(sub
, &enabled
, &running
);
3418 if (read_format
& PERF_FORMAT_ID
)
3419 values
[n
++] = primary_event_id(sub
);
3421 size
= n
* sizeof(u64
);
3423 if (copy_to_user(buf
+ ret
, values
, size
)) {
3431 mutex_unlock(&ctx
->mutex
);
3436 static int perf_event_read_one(struct perf_event
*event
,
3437 u64 read_format
, char __user
*buf
)
3439 u64 enabled
, running
;
3443 values
[n
++] = perf_event_read_value(event
, &enabled
, &running
);
3444 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
3445 values
[n
++] = enabled
;
3446 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
3447 values
[n
++] = running
;
3448 if (read_format
& PERF_FORMAT_ID
)
3449 values
[n
++] = primary_event_id(event
);
3451 if (copy_to_user(buf
, values
, n
* sizeof(u64
)))
3454 return n
* sizeof(u64
);
3458 * Read the performance event - simple non blocking version for now
3461 perf_read_hw(struct perf_event
*event
, char __user
*buf
, size_t count
)
3463 u64 read_format
= event
->attr
.read_format
;
3467 * Return end-of-file for a read on a event that is in
3468 * error state (i.e. because it was pinned but it couldn't be
3469 * scheduled on to the CPU at some point).
3471 if (event
->state
== PERF_EVENT_STATE_ERROR
)
3474 if (count
< event
->read_size
)
3477 WARN_ON_ONCE(event
->ctx
->parent_ctx
);
3478 if (read_format
& PERF_FORMAT_GROUP
)
3479 ret
= perf_event_read_group(event
, read_format
, buf
);
3481 ret
= perf_event_read_one(event
, read_format
, buf
);
3487 perf_read(struct file
*file
, char __user
*buf
, size_t count
, loff_t
*ppos
)
3489 struct perf_event
*event
= file
->private_data
;
3491 return perf_read_hw(event
, buf
, count
);
3494 static unsigned int perf_poll(struct file
*file
, poll_table
*wait
)
3496 struct perf_event
*event
= file
->private_data
;
3497 struct ring_buffer
*rb
;
3498 unsigned int events
= POLL_HUP
;
3501 * Pin the event->rb by taking event->mmap_mutex; otherwise
3502 * perf_event_set_output() can swizzle our rb and make us miss wakeups.
3504 mutex_lock(&event
->mmap_mutex
);
3507 events
= atomic_xchg(&rb
->poll
, 0);
3508 mutex_unlock(&event
->mmap_mutex
);
3510 poll_wait(file
, &event
->waitq
, wait
);
3515 static void perf_event_reset(struct perf_event
*event
)
3517 (void)perf_event_read(event
);
3518 local64_set(&event
->count
, 0);
3519 perf_event_update_userpage(event
);
3523 * Holding the top-level event's child_mutex means that any
3524 * descendant process that has inherited this event will block
3525 * in sync_child_event if it goes to exit, thus satisfying the
3526 * task existence requirements of perf_event_enable/disable.
3528 static void perf_event_for_each_child(struct perf_event
*event
,
3529 void (*func
)(struct perf_event
*))
3531 struct perf_event
*child
;
3533 WARN_ON_ONCE(event
->ctx
->parent_ctx
);
3534 mutex_lock(&event
->child_mutex
);
3536 list_for_each_entry(child
, &event
->child_list
, child_list
)
3538 mutex_unlock(&event
->child_mutex
);
3541 static void perf_event_for_each(struct perf_event
*event
,
3542 void (*func
)(struct perf_event
*))
3544 struct perf_event_context
*ctx
= event
->ctx
;
3545 struct perf_event
*sibling
;
3547 WARN_ON_ONCE(ctx
->parent_ctx
);
3548 mutex_lock(&ctx
->mutex
);
3549 event
= event
->group_leader
;
3551 perf_event_for_each_child(event
, func
);
3552 list_for_each_entry(sibling
, &event
->sibling_list
, group_entry
)
3553 perf_event_for_each_child(sibling
, func
);
3554 mutex_unlock(&ctx
->mutex
);
3557 static int perf_event_period(struct perf_event
*event
, u64 __user
*arg
)
3559 struct perf_event_context
*ctx
= event
->ctx
;
3560 int ret
= 0, active
;
3563 if (!is_sampling_event(event
))
3566 if (copy_from_user(&value
, arg
, sizeof(value
)))
3572 raw_spin_lock_irq(&ctx
->lock
);
3573 if (event
->attr
.freq
) {
3574 if (value
> sysctl_perf_event_sample_rate
) {
3579 event
->attr
.sample_freq
= value
;
3581 event
->attr
.sample_period
= value
;
3582 event
->hw
.sample_period
= value
;
3585 active
= (event
->state
== PERF_EVENT_STATE_ACTIVE
);
3587 perf_pmu_disable(ctx
->pmu
);
3588 event
->pmu
->stop(event
, PERF_EF_UPDATE
);
3591 local64_set(&event
->hw
.period_left
, 0);
3594 event
->pmu
->start(event
, PERF_EF_RELOAD
);
3595 perf_pmu_enable(ctx
->pmu
);
3599 raw_spin_unlock_irq(&ctx
->lock
);
3604 static const struct file_operations perf_fops
;
3606 static inline int perf_fget_light(int fd
, struct fd
*p
)
3608 struct fd f
= fdget(fd
);
3612 if (f
.file
->f_op
!= &perf_fops
) {
3620 static int perf_event_set_output(struct perf_event
*event
,
3621 struct perf_event
*output_event
);
3622 static int perf_event_set_filter(struct perf_event
*event
, void __user
*arg
);
3624 static long perf_ioctl(struct file
*file
, unsigned int cmd
, unsigned long arg
)
3626 struct perf_event
*event
= file
->private_data
;
3627 void (*func
)(struct perf_event
*);
3631 case PERF_EVENT_IOC_ENABLE
:
3632 func
= perf_event_enable
;
3634 case PERF_EVENT_IOC_DISABLE
:
3635 func
= perf_event_disable
;
3637 case PERF_EVENT_IOC_RESET
:
3638 func
= perf_event_reset
;
3641 case PERF_EVENT_IOC_REFRESH
:
3642 return perf_event_refresh(event
, arg
);
3644 case PERF_EVENT_IOC_PERIOD
:
3645 return perf_event_period(event
, (u64 __user
*)arg
);
3647 case PERF_EVENT_IOC_ID
:
3649 u64 id
= primary_event_id(event
);
3651 if (copy_to_user((void __user
*)arg
, &id
, sizeof(id
)))
3656 case PERF_EVENT_IOC_SET_OUTPUT
:
3660 struct perf_event
*output_event
;
3662 ret
= perf_fget_light(arg
, &output
);
3665 output_event
= output
.file
->private_data
;
3666 ret
= perf_event_set_output(event
, output_event
);
3669 ret
= perf_event_set_output(event
, NULL
);
3674 case PERF_EVENT_IOC_SET_FILTER
:
3675 return perf_event_set_filter(event
, (void __user
*)arg
);
3681 if (flags
& PERF_IOC_FLAG_GROUP
)
3682 perf_event_for_each(event
, func
);
3684 perf_event_for_each_child(event
, func
);
3689 int perf_event_task_enable(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_enable
);
3696 mutex_unlock(¤t
->perf_event_mutex
);
3701 int perf_event_task_disable(void)
3703 struct perf_event
*event
;
3705 mutex_lock(¤t
->perf_event_mutex
);
3706 list_for_each_entry(event
, ¤t
->perf_event_list
, owner_entry
)
3707 perf_event_for_each_child(event
, perf_event_disable
);
3708 mutex_unlock(¤t
->perf_event_mutex
);
3713 static int perf_event_index(struct perf_event
*event
)
3715 if (event
->hw
.state
& PERF_HES_STOPPED
)
3718 if (event
->state
!= PERF_EVENT_STATE_ACTIVE
)
3721 return event
->pmu
->event_idx(event
);
3724 static void calc_timer_values(struct perf_event
*event
,
3731 *now
= perf_clock();
3732 ctx_time
= event
->shadow_ctx_time
+ *now
;
3733 *enabled
= ctx_time
- event
->tstamp_enabled
;
3734 *running
= ctx_time
- event
->tstamp_running
;
3737 static void perf_event_init_userpage(struct perf_event
*event
)
3739 struct perf_event_mmap_page
*userpg
;
3740 struct ring_buffer
*rb
;
3743 rb
= rcu_dereference(event
->rb
);
3747 userpg
= rb
->user_page
;
3749 /* Allow new userspace to detect that bit 0 is deprecated */
3750 userpg
->cap_bit0_is_deprecated
= 1;
3751 userpg
->size
= offsetof(struct perf_event_mmap_page
, __reserved
);
3757 void __weak
arch_perf_update_userpage(struct perf_event_mmap_page
*userpg
, u64 now
)
3762 * Callers need to ensure there can be no nesting of this function, otherwise
3763 * the seqlock logic goes bad. We can not serialize this because the arch
3764 * code calls this from NMI context.
3766 void perf_event_update_userpage(struct perf_event
*event
)
3768 struct perf_event_mmap_page
*userpg
;
3769 struct ring_buffer
*rb
;
3770 u64 enabled
, running
, now
;
3773 rb
= rcu_dereference(event
->rb
);
3778 * compute total_time_enabled, total_time_running
3779 * based on snapshot values taken when the event
3780 * was last scheduled in.
3782 * we cannot simply called update_context_time()
3783 * because of locking issue as we can be called in
3786 calc_timer_values(event
, &now
, &enabled
, &running
);
3788 userpg
= rb
->user_page
;
3790 * Disable preemption so as to not let the corresponding user-space
3791 * spin too long if we get preempted.
3796 userpg
->index
= perf_event_index(event
);
3797 userpg
->offset
= perf_event_count(event
);
3799 userpg
->offset
-= local64_read(&event
->hw
.prev_count
);
3801 userpg
->time_enabled
= enabled
+
3802 atomic64_read(&event
->child_total_time_enabled
);
3804 userpg
->time_running
= running
+
3805 atomic64_read(&event
->child_total_time_running
);
3807 arch_perf_update_userpage(userpg
, now
);
3816 static int perf_mmap_fault(struct vm_area_struct
*vma
, struct vm_fault
*vmf
)
3818 struct perf_event
*event
= vma
->vm_file
->private_data
;
3819 struct ring_buffer
*rb
;
3820 int ret
= VM_FAULT_SIGBUS
;
3822 if (vmf
->flags
& FAULT_FLAG_MKWRITE
) {
3823 if (vmf
->pgoff
== 0)
3829 rb
= rcu_dereference(event
->rb
);
3833 if (vmf
->pgoff
&& (vmf
->flags
& FAULT_FLAG_WRITE
))
3836 vmf
->page
= perf_mmap_to_page(rb
, vmf
->pgoff
);
3840 get_page(vmf
->page
);
3841 vmf
->page
->mapping
= vma
->vm_file
->f_mapping
;
3842 vmf
->page
->index
= vmf
->pgoff
;
3851 static void ring_buffer_attach(struct perf_event
*event
,
3852 struct ring_buffer
*rb
)
3854 unsigned long flags
;
3856 if (!list_empty(&event
->rb_entry
))
3859 spin_lock_irqsave(&rb
->event_lock
, flags
);
3860 if (list_empty(&event
->rb_entry
))
3861 list_add(&event
->rb_entry
, &rb
->event_list
);
3862 spin_unlock_irqrestore(&rb
->event_lock
, flags
);
3865 static void ring_buffer_detach(struct perf_event
*event
, struct ring_buffer
*rb
)
3867 unsigned long flags
;
3869 if (list_empty(&event
->rb_entry
))
3872 spin_lock_irqsave(&rb
->event_lock
, flags
);
3873 list_del_init(&event
->rb_entry
);
3874 wake_up_all(&event
->waitq
);
3875 spin_unlock_irqrestore(&rb
->event_lock
, flags
);
3878 static void ring_buffer_wakeup(struct perf_event
*event
)
3880 struct ring_buffer
*rb
;
3883 rb
= rcu_dereference(event
->rb
);
3885 list_for_each_entry_rcu(event
, &rb
->event_list
, rb_entry
)
3886 wake_up_all(&event
->waitq
);
3891 static void rb_free_rcu(struct rcu_head
*rcu_head
)
3893 struct ring_buffer
*rb
;
3895 rb
= container_of(rcu_head
, struct ring_buffer
, rcu_head
);
3899 static struct ring_buffer
*ring_buffer_get(struct perf_event
*event
)
3901 struct ring_buffer
*rb
;
3904 rb
= rcu_dereference(event
->rb
);
3906 if (!atomic_inc_not_zero(&rb
->refcount
))
3914 static void ring_buffer_put(struct ring_buffer
*rb
)
3916 if (!atomic_dec_and_test(&rb
->refcount
))
3919 WARN_ON_ONCE(!list_empty(&rb
->event_list
));
3921 call_rcu(&rb
->rcu_head
, rb_free_rcu
);
3924 static void perf_mmap_open(struct vm_area_struct
*vma
)
3926 struct perf_event
*event
= vma
->vm_file
->private_data
;
3928 atomic_inc(&event
->mmap_count
);
3929 atomic_inc(&event
->rb
->mmap_count
);
3933 * A buffer can be mmap()ed multiple times; either directly through the same
3934 * event, or through other events by use of perf_event_set_output().
3936 * In order to undo the VM accounting done by perf_mmap() we need to destroy
3937 * the buffer here, where we still have a VM context. This means we need
3938 * to detach all events redirecting to us.
3940 static void perf_mmap_close(struct vm_area_struct
*vma
)
3942 struct perf_event
*event
= vma
->vm_file
->private_data
;
3944 struct ring_buffer
*rb
= event
->rb
;
3945 struct user_struct
*mmap_user
= rb
->mmap_user
;
3946 int mmap_locked
= rb
->mmap_locked
;
3947 unsigned long size
= perf_data_size(rb
);
3949 atomic_dec(&rb
->mmap_count
);
3951 if (!atomic_dec_and_mutex_lock(&event
->mmap_count
, &event
->mmap_mutex
))
3954 /* Detach current event from the buffer. */
3955 rcu_assign_pointer(event
->rb
, NULL
);
3956 ring_buffer_detach(event
, rb
);
3957 mutex_unlock(&event
->mmap_mutex
);
3959 /* If there's still other mmap()s of this buffer, we're done. */
3960 if (atomic_read(&rb
->mmap_count
)) {
3961 ring_buffer_put(rb
); /* can't be last */
3966 * No other mmap()s, detach from all other events that might redirect
3967 * into the now unreachable buffer. Somewhat complicated by the
3968 * fact that rb::event_lock otherwise nests inside mmap_mutex.
3972 list_for_each_entry_rcu(event
, &rb
->event_list
, rb_entry
) {
3973 if (!atomic_long_inc_not_zero(&event
->refcount
)) {
3975 * This event is en-route to free_event() which will
3976 * detach it and remove it from the list.
3982 mutex_lock(&event
->mmap_mutex
);
3984 * Check we didn't race with perf_event_set_output() which can
3985 * swizzle the rb from under us while we were waiting to
3986 * acquire mmap_mutex.
3988 * If we find a different rb; ignore this event, a next
3989 * iteration will no longer find it on the list. We have to
3990 * still restart the iteration to make sure we're not now
3991 * iterating the wrong list.
3993 if (event
->rb
== rb
) {
3994 rcu_assign_pointer(event
->rb
, NULL
);
3995 ring_buffer_detach(event
, rb
);
3996 ring_buffer_put(rb
); /* can't be last, we still have one */
3998 mutex_unlock(&event
->mmap_mutex
);
4002 * Restart the iteration; either we're on the wrong list or
4003 * destroyed its integrity by doing a deletion.
4010 * It could be there's still a few 0-ref events on the list; they'll
4011 * get cleaned up by free_event() -- they'll also still have their
4012 * ref on the rb and will free it whenever they are done with it.
4014 * Aside from that, this buffer is 'fully' detached and unmapped,
4015 * undo the VM accounting.
4018 atomic_long_sub((size
>> PAGE_SHIFT
) + 1, &mmap_user
->locked_vm
);
4019 vma
->vm_mm
->pinned_vm
-= mmap_locked
;
4020 free_uid(mmap_user
);
4022 ring_buffer_put(rb
); /* could be last */
4025 static const struct vm_operations_struct perf_mmap_vmops
= {
4026 .open
= perf_mmap_open
,
4027 .close
= perf_mmap_close
,
4028 .fault
= perf_mmap_fault
,
4029 .page_mkwrite
= perf_mmap_fault
,
4032 static int perf_mmap(struct file
*file
, struct vm_area_struct
*vma
)
4034 struct perf_event
*event
= file
->private_data
;
4035 unsigned long user_locked
, user_lock_limit
;
4036 struct user_struct
*user
= current_user();
4037 unsigned long locked
, lock_limit
;
4038 struct ring_buffer
*rb
;
4039 unsigned long vma_size
;
4040 unsigned long nr_pages
;
4041 long user_extra
, extra
;
4042 int ret
= 0, flags
= 0;
4045 * Don't allow mmap() of inherited per-task counters. This would
4046 * create a performance issue due to all children writing to the
4049 if (event
->cpu
== -1 && event
->attr
.inherit
)
4052 if (!(vma
->vm_flags
& VM_SHARED
))
4055 vma_size
= vma
->vm_end
- vma
->vm_start
;
4056 nr_pages
= (vma_size
/ PAGE_SIZE
) - 1;
4059 * If we have rb pages ensure they're a power-of-two number, so we
4060 * can do bitmasks instead of modulo.
4062 if (nr_pages
!= 0 && !is_power_of_2(nr_pages
))
4065 if (vma_size
!= PAGE_SIZE
* (1 + nr_pages
))
4068 if (vma
->vm_pgoff
!= 0)
4071 WARN_ON_ONCE(event
->ctx
->parent_ctx
);
4073 mutex_lock(&event
->mmap_mutex
);
4075 if (event
->rb
->nr_pages
!= nr_pages
) {
4080 if (!atomic_inc_not_zero(&event
->rb
->mmap_count
)) {
4082 * Raced against perf_mmap_close() through
4083 * perf_event_set_output(). Try again, hope for better
4086 mutex_unlock(&event
->mmap_mutex
);
4093 user_extra
= nr_pages
+ 1;
4094 user_lock_limit
= sysctl_perf_event_mlock
>> (PAGE_SHIFT
- 10);
4097 * Increase the limit linearly with more CPUs:
4099 user_lock_limit
*= num_online_cpus();
4101 user_locked
= atomic_long_read(&user
->locked_vm
) + user_extra
;
4104 if (user_locked
> user_lock_limit
)
4105 extra
= user_locked
- user_lock_limit
;
4107 lock_limit
= rlimit(RLIMIT_MEMLOCK
);
4108 lock_limit
>>= PAGE_SHIFT
;
4109 locked
= vma
->vm_mm
->pinned_vm
+ extra
;
4111 if ((locked
> lock_limit
) && perf_paranoid_tracepoint_raw() &&
4112 !capable(CAP_IPC_LOCK
)) {
4119 if (vma
->vm_flags
& VM_WRITE
)
4120 flags
|= RING_BUFFER_WRITABLE
;
4122 rb
= rb_alloc(nr_pages
,
4123 event
->attr
.watermark
? event
->attr
.wakeup_watermark
: 0,
4131 atomic_set(&rb
->mmap_count
, 1);
4132 rb
->mmap_locked
= extra
;
4133 rb
->mmap_user
= get_current_user();
4135 atomic_long_add(user_extra
, &user
->locked_vm
);
4136 vma
->vm_mm
->pinned_vm
+= extra
;
4138 ring_buffer_attach(event
, rb
);
4139 rcu_assign_pointer(event
->rb
, rb
);
4141 perf_event_init_userpage(event
);
4142 perf_event_update_userpage(event
);
4146 atomic_inc(&event
->mmap_count
);
4147 mutex_unlock(&event
->mmap_mutex
);
4150 * Since pinned accounting is per vm we cannot allow fork() to copy our
4153 vma
->vm_flags
|= VM_DONTCOPY
| VM_DONTEXPAND
| VM_DONTDUMP
;
4154 vma
->vm_ops
= &perf_mmap_vmops
;
4159 static int perf_fasync(int fd
, struct file
*filp
, int on
)
4161 struct inode
*inode
= file_inode(filp
);
4162 struct perf_event
*event
= filp
->private_data
;
4165 mutex_lock(&inode
->i_mutex
);
4166 retval
= fasync_helper(fd
, filp
, on
, &event
->fasync
);
4167 mutex_unlock(&inode
->i_mutex
);
4175 static const struct file_operations perf_fops
= {
4176 .llseek
= no_llseek
,
4177 .release
= perf_release
,
4180 .unlocked_ioctl
= perf_ioctl
,
4181 .compat_ioctl
= perf_ioctl
,
4183 .fasync
= perf_fasync
,
4189 * If there's data, ensure we set the poll() state and publish everything
4190 * to user-space before waking everybody up.
4193 void perf_event_wakeup(struct perf_event
*event
)
4195 ring_buffer_wakeup(event
);
4197 if (event
->pending_kill
) {
4198 kill_fasync(&event
->fasync
, SIGIO
, event
->pending_kill
);
4199 event
->pending_kill
= 0;
4203 static void perf_pending_event(struct irq_work
*entry
)
4205 struct perf_event
*event
= container_of(entry
,
4206 struct perf_event
, pending
);
4208 if (event
->pending_disable
) {
4209 event
->pending_disable
= 0;
4210 __perf_event_disable(event
);
4213 if (event
->pending_wakeup
) {
4214 event
->pending_wakeup
= 0;
4215 perf_event_wakeup(event
);
4220 * We assume there is only KVM supporting the callbacks.
4221 * Later on, we might change it to a list if there is
4222 * another virtualization implementation supporting the callbacks.
4224 struct perf_guest_info_callbacks
*perf_guest_cbs
;
4226 int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks
*cbs
)
4228 perf_guest_cbs
= cbs
;
4231 EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks
);
4233 int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks
*cbs
)
4235 perf_guest_cbs
= NULL
;
4238 EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks
);
4241 perf_output_sample_regs(struct perf_output_handle
*handle
,
4242 struct pt_regs
*regs
, u64 mask
)
4246 for_each_set_bit(bit
, (const unsigned long *) &mask
,
4247 sizeof(mask
) * BITS_PER_BYTE
) {
4250 val
= perf_reg_value(regs
, bit
);
4251 perf_output_put(handle
, val
);
4255 static void perf_sample_regs_user(struct perf_regs_user
*regs_user
,
4256 struct pt_regs
*regs
)
4258 if (!user_mode(regs
)) {
4260 regs
= task_pt_regs(current
);
4266 regs_user
->regs
= regs
;
4267 regs_user
->abi
= perf_reg_abi(current
);
4272 * Get remaining task size from user stack pointer.
4274 * It'd be better to take stack vma map and limit this more
4275 * precisly, but there's no way to get it safely under interrupt,
4276 * so using TASK_SIZE as limit.
4278 static u64
perf_ustack_task_size(struct pt_regs
*regs
)
4280 unsigned long addr
= perf_user_stack_pointer(regs
);
4282 if (!addr
|| addr
>= TASK_SIZE
)
4285 return TASK_SIZE
- addr
;
4289 perf_sample_ustack_size(u16 stack_size
, u16 header_size
,
4290 struct pt_regs
*regs
)
4294 /* No regs, no stack pointer, no dump. */
4299 * Check if we fit in with the requested stack size into the:
4301 * If we don't, we limit the size to the TASK_SIZE.
4303 * - remaining sample size
4304 * If we don't, we customize the stack size to
4305 * fit in to the remaining sample size.
4308 task_size
= min((u64
) USHRT_MAX
, perf_ustack_task_size(regs
));
4309 stack_size
= min(stack_size
, (u16
) task_size
);
4311 /* Current header size plus static size and dynamic size. */
4312 header_size
+= 2 * sizeof(u64
);
4314 /* Do we fit in with the current stack dump size? */
4315 if ((u16
) (header_size
+ stack_size
) < header_size
) {
4317 * If we overflow the maximum size for the sample,
4318 * we customize the stack dump size to fit in.
4320 stack_size
= USHRT_MAX
- header_size
- sizeof(u64
);
4321 stack_size
= round_up(stack_size
, sizeof(u64
));
4328 perf_output_sample_ustack(struct perf_output_handle
*handle
, u64 dump_size
,
4329 struct pt_regs
*regs
)
4331 /* Case of a kernel thread, nothing to dump */
4334 perf_output_put(handle
, size
);
4343 * - the size requested by user or the best one we can fit
4344 * in to the sample max size
4346 * - user stack dump data
4348 * - the actual dumped size
4352 perf_output_put(handle
, dump_size
);
4355 sp
= perf_user_stack_pointer(regs
);
4356 rem
= __output_copy_user(handle
, (void *) sp
, dump_size
);
4357 dyn_size
= dump_size
- rem
;
4359 perf_output_skip(handle
, rem
);
4362 perf_output_put(handle
, dyn_size
);
4366 static void __perf_event_header__init_id(struct perf_event_header
*header
,
4367 struct perf_sample_data
*data
,
4368 struct perf_event
*event
)
4370 u64 sample_type
= event
->attr
.sample_type
;
4372 data
->type
= sample_type
;
4373 header
->size
+= event
->id_header_size
;
4375 if (sample_type
& PERF_SAMPLE_TID
) {
4376 /* namespace issues */
4377 data
->tid_entry
.pid
= perf_event_pid(event
, current
);
4378 data
->tid_entry
.tid
= perf_event_tid(event
, current
);
4381 if (sample_type
& PERF_SAMPLE_TIME
)
4382 data
->time
= perf_clock();
4384 if (sample_type
& (PERF_SAMPLE_ID
| PERF_SAMPLE_IDENTIFIER
))
4385 data
->id
= primary_event_id(event
);
4387 if (sample_type
& PERF_SAMPLE_STREAM_ID
)
4388 data
->stream_id
= event
->id
;
4390 if (sample_type
& PERF_SAMPLE_CPU
) {
4391 data
->cpu_entry
.cpu
= raw_smp_processor_id();
4392 data
->cpu_entry
.reserved
= 0;
4396 void perf_event_header__init_id(struct perf_event_header
*header
,
4397 struct perf_sample_data
*data
,
4398 struct perf_event
*event
)
4400 if (event
->attr
.sample_id_all
)
4401 __perf_event_header__init_id(header
, data
, event
);
4404 static void __perf_event__output_id_sample(struct perf_output_handle
*handle
,
4405 struct perf_sample_data
*data
)
4407 u64 sample_type
= data
->type
;
4409 if (sample_type
& PERF_SAMPLE_TID
)
4410 perf_output_put(handle
, data
->tid_entry
);
4412 if (sample_type
& PERF_SAMPLE_TIME
)
4413 perf_output_put(handle
, data
->time
);
4415 if (sample_type
& PERF_SAMPLE_ID
)
4416 perf_output_put(handle
, data
->id
);
4418 if (sample_type
& PERF_SAMPLE_STREAM_ID
)
4419 perf_output_put(handle
, data
->stream_id
);
4421 if (sample_type
& PERF_SAMPLE_CPU
)
4422 perf_output_put(handle
, data
->cpu_entry
);
4424 if (sample_type
& PERF_SAMPLE_IDENTIFIER
)
4425 perf_output_put(handle
, data
->id
);
4428 void perf_event__output_id_sample(struct perf_event
*event
,
4429 struct perf_output_handle
*handle
,
4430 struct perf_sample_data
*sample
)
4432 if (event
->attr
.sample_id_all
)
4433 __perf_event__output_id_sample(handle
, sample
);
4436 static void perf_output_read_one(struct perf_output_handle
*handle
,
4437 struct perf_event
*event
,
4438 u64 enabled
, u64 running
)
4440 u64 read_format
= event
->attr
.read_format
;
4444 values
[n
++] = perf_event_count(event
);
4445 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
) {
4446 values
[n
++] = enabled
+
4447 atomic64_read(&event
->child_total_time_enabled
);
4449 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
) {
4450 values
[n
++] = running
+
4451 atomic64_read(&event
->child_total_time_running
);
4453 if (read_format
& PERF_FORMAT_ID
)
4454 values
[n
++] = primary_event_id(event
);
4456 __output_copy(handle
, values
, n
* sizeof(u64
));
4460 * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
4462 static void perf_output_read_group(struct perf_output_handle
*handle
,
4463 struct perf_event
*event
,
4464 u64 enabled
, u64 running
)
4466 struct perf_event
*leader
= event
->group_leader
, *sub
;
4467 u64 read_format
= event
->attr
.read_format
;
4471 values
[n
++] = 1 + leader
->nr_siblings
;
4473 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
4474 values
[n
++] = enabled
;
4476 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
4477 values
[n
++] = running
;
4479 if (leader
!= event
)
4480 leader
->pmu
->read(leader
);
4482 values
[n
++] = perf_event_count(leader
);
4483 if (read_format
& PERF_FORMAT_ID
)
4484 values
[n
++] = primary_event_id(leader
);
4486 __output_copy(handle
, values
, n
* sizeof(u64
));
4488 list_for_each_entry(sub
, &leader
->sibling_list
, group_entry
) {
4491 if ((sub
!= event
) &&
4492 (sub
->state
== PERF_EVENT_STATE_ACTIVE
))
4493 sub
->pmu
->read(sub
);
4495 values
[n
++] = perf_event_count(sub
);
4496 if (read_format
& PERF_FORMAT_ID
)
4497 values
[n
++] = primary_event_id(sub
);
4499 __output_copy(handle
, values
, n
* sizeof(u64
));
4503 #define PERF_FORMAT_TOTAL_TIMES (PERF_FORMAT_TOTAL_TIME_ENABLED|\
4504 PERF_FORMAT_TOTAL_TIME_RUNNING)
4506 static void perf_output_read(struct perf_output_handle
*handle
,
4507 struct perf_event
*event
)
4509 u64 enabled
= 0, running
= 0, now
;
4510 u64 read_format
= event
->attr
.read_format
;
4513 * compute total_time_enabled, total_time_running
4514 * based on snapshot values taken when the event
4515 * was last scheduled in.
4517 * we cannot simply called update_context_time()
4518 * because of locking issue as we are called in
4521 if (read_format
& PERF_FORMAT_TOTAL_TIMES
)
4522 calc_timer_values(event
, &now
, &enabled
, &running
);
4524 if (event
->attr
.read_format
& PERF_FORMAT_GROUP
)
4525 perf_output_read_group(handle
, event
, enabled
, running
);
4527 perf_output_read_one(handle
, event
, enabled
, running
);
4530 void perf_output_sample(struct perf_output_handle
*handle
,
4531 struct perf_event_header
*header
,
4532 struct perf_sample_data
*data
,
4533 struct perf_event
*event
)
4535 u64 sample_type
= data
->type
;
4537 perf_output_put(handle
, *header
);
4539 if (sample_type
& PERF_SAMPLE_IDENTIFIER
)
4540 perf_output_put(handle
, data
->id
);
4542 if (sample_type
& PERF_SAMPLE_IP
)
4543 perf_output_put(handle
, data
->ip
);
4545 if (sample_type
& PERF_SAMPLE_TID
)
4546 perf_output_put(handle
, data
->tid_entry
);
4548 if (sample_type
& PERF_SAMPLE_TIME
)
4549 perf_output_put(handle
, data
->time
);
4551 if (sample_type
& PERF_SAMPLE_ADDR
)
4552 perf_output_put(handle
, data
->addr
);
4554 if (sample_type
& PERF_SAMPLE_ID
)
4555 perf_output_put(handle
, data
->id
);
4557 if (sample_type
& PERF_SAMPLE_STREAM_ID
)
4558 perf_output_put(handle
, data
->stream_id
);
4560 if (sample_type
& PERF_SAMPLE_CPU
)
4561 perf_output_put(handle
, data
->cpu_entry
);
4563 if (sample_type
& PERF_SAMPLE_PERIOD
)
4564 perf_output_put(handle
, data
->period
);
4566 if (sample_type
& PERF_SAMPLE_READ
)
4567 perf_output_read(handle
, event
);
4569 if (sample_type
& PERF_SAMPLE_CALLCHAIN
) {
4570 if (data
->callchain
) {
4573 if (data
->callchain
)
4574 size
+= data
->callchain
->nr
;
4576 size
*= sizeof(u64
);
4578 __output_copy(handle
, data
->callchain
, size
);
4581 perf_output_put(handle
, nr
);
4585 if (sample_type
& PERF_SAMPLE_RAW
) {
4587 perf_output_put(handle
, data
->raw
->size
);
4588 __output_copy(handle
, data
->raw
->data
,
4595 .size
= sizeof(u32
),
4598 perf_output_put(handle
, raw
);
4602 if (sample_type
& PERF_SAMPLE_BRANCH_STACK
) {
4603 if (data
->br_stack
) {
4606 size
= data
->br_stack
->nr
4607 * sizeof(struct perf_branch_entry
);
4609 perf_output_put(handle
, data
->br_stack
->nr
);
4610 perf_output_copy(handle
, data
->br_stack
->entries
, size
);
4613 * we always store at least the value of nr
4616 perf_output_put(handle
, nr
);
4620 if (sample_type
& PERF_SAMPLE_REGS_USER
) {
4621 u64 abi
= data
->regs_user
.abi
;
4624 * If there are no regs to dump, notice it through
4625 * first u64 being zero (PERF_SAMPLE_REGS_ABI_NONE).
4627 perf_output_put(handle
, abi
);
4630 u64 mask
= event
->attr
.sample_regs_user
;
4631 perf_output_sample_regs(handle
,
4632 data
->regs_user
.regs
,
4637 if (sample_type
& PERF_SAMPLE_STACK_USER
) {
4638 perf_output_sample_ustack(handle
,
4639 data
->stack_user_size
,
4640 data
->regs_user
.regs
);
4643 if (sample_type
& PERF_SAMPLE_WEIGHT
)
4644 perf_output_put(handle
, data
->weight
);
4646 if (sample_type
& PERF_SAMPLE_DATA_SRC
)
4647 perf_output_put(handle
, data
->data_src
.val
);
4649 if (sample_type
& PERF_SAMPLE_TRANSACTION
)
4650 perf_output_put(handle
, data
->txn
);
4652 if (!event
->attr
.watermark
) {
4653 int wakeup_events
= event
->attr
.wakeup_events
;
4655 if (wakeup_events
) {
4656 struct ring_buffer
*rb
= handle
->rb
;
4657 int events
= local_inc_return(&rb
->events
);
4659 if (events
>= wakeup_events
) {
4660 local_sub(wakeup_events
, &rb
->events
);
4661 local_inc(&rb
->wakeup
);
4667 void perf_prepare_sample(struct perf_event_header
*header
,
4668 struct perf_sample_data
*data
,
4669 struct perf_event
*event
,
4670 struct pt_regs
*regs
)
4672 u64 sample_type
= event
->attr
.sample_type
;
4674 header
->type
= PERF_RECORD_SAMPLE
;
4675 header
->size
= sizeof(*header
) + event
->header_size
;
4678 header
->misc
|= perf_misc_flags(regs
);
4680 __perf_event_header__init_id(header
, data
, event
);
4682 if (sample_type
& PERF_SAMPLE_IP
)
4683 data
->ip
= perf_instruction_pointer(regs
);
4685 if (sample_type
& PERF_SAMPLE_CALLCHAIN
) {
4688 data
->callchain
= perf_callchain(event
, regs
);
4690 if (data
->callchain
)
4691 size
+= data
->callchain
->nr
;
4693 header
->size
+= size
* sizeof(u64
);
4696 if (sample_type
& PERF_SAMPLE_RAW
) {
4697 int size
= sizeof(u32
);
4700 size
+= data
->raw
->size
;
4702 size
+= sizeof(u32
);
4704 WARN_ON_ONCE(size
& (sizeof(u64
)-1));
4705 header
->size
+= size
;
4708 if (sample_type
& PERF_SAMPLE_BRANCH_STACK
) {
4709 int size
= sizeof(u64
); /* nr */
4710 if (data
->br_stack
) {
4711 size
+= data
->br_stack
->nr
4712 * sizeof(struct perf_branch_entry
);
4714 header
->size
+= size
;
4717 if (sample_type
& PERF_SAMPLE_REGS_USER
) {
4718 /* regs dump ABI info */
4719 int size
= sizeof(u64
);
4721 perf_sample_regs_user(&data
->regs_user
, regs
);
4723 if (data
->regs_user
.regs
) {
4724 u64 mask
= event
->attr
.sample_regs_user
;
4725 size
+= hweight64(mask
) * sizeof(u64
);
4728 header
->size
+= size
;
4731 if (sample_type
& PERF_SAMPLE_STACK_USER
) {
4733 * Either we need PERF_SAMPLE_STACK_USER bit to be allways
4734 * processed as the last one or have additional check added
4735 * in case new sample type is added, because we could eat
4736 * up the rest of the sample size.
4738 struct perf_regs_user
*uregs
= &data
->regs_user
;
4739 u16 stack_size
= event
->attr
.sample_stack_user
;
4740 u16 size
= sizeof(u64
);
4743 perf_sample_regs_user(uregs
, regs
);
4745 stack_size
= perf_sample_ustack_size(stack_size
, header
->size
,
4749 * If there is something to dump, add space for the dump
4750 * itself and for the field that tells the dynamic size,
4751 * which is how many have been actually dumped.
4754 size
+= sizeof(u64
) + stack_size
;
4756 data
->stack_user_size
= stack_size
;
4757 header
->size
+= size
;
4761 static void perf_event_output(struct perf_event
*event
,
4762 struct perf_sample_data
*data
,
4763 struct pt_regs
*regs
)
4765 struct perf_output_handle handle
;
4766 struct perf_event_header header
;
4768 /* protect the callchain buffers */
4771 perf_prepare_sample(&header
, data
, event
, regs
);
4773 if (perf_output_begin(&handle
, event
, header
.size
))
4776 perf_output_sample(&handle
, &header
, data
, event
);
4778 perf_output_end(&handle
);
4788 struct perf_read_event
{
4789 struct perf_event_header header
;
4796 perf_event_read_event(struct perf_event
*event
,
4797 struct task_struct
*task
)
4799 struct perf_output_handle handle
;
4800 struct perf_sample_data sample
;
4801 struct perf_read_event read_event
= {
4803 .type
= PERF_RECORD_READ
,
4805 .size
= sizeof(read_event
) + event
->read_size
,
4807 .pid
= perf_event_pid(event
, task
),
4808 .tid
= perf_event_tid(event
, task
),
4812 perf_event_header__init_id(&read_event
.header
, &sample
, event
);
4813 ret
= perf_output_begin(&handle
, event
, read_event
.header
.size
);
4817 perf_output_put(&handle
, read_event
);
4818 perf_output_read(&handle
, event
);
4819 perf_event__output_id_sample(event
, &handle
, &sample
);
4821 perf_output_end(&handle
);
4824 typedef void (perf_event_aux_output_cb
)(struct perf_event
*event
, void *data
);
4827 perf_event_aux_ctx(struct perf_event_context
*ctx
,
4828 perf_event_aux_output_cb output
,
4831 struct perf_event
*event
;
4833 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
4834 if (event
->state
< PERF_EVENT_STATE_INACTIVE
)
4836 if (!event_filter_match(event
))
4838 output(event
, data
);
4843 perf_event_aux(perf_event_aux_output_cb output
, void *data
,
4844 struct perf_event_context
*task_ctx
)
4846 struct perf_cpu_context
*cpuctx
;
4847 struct perf_event_context
*ctx
;
4852 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
4853 cpuctx
= get_cpu_ptr(pmu
->pmu_cpu_context
);
4854 if (cpuctx
->unique_pmu
!= pmu
)
4856 perf_event_aux_ctx(&cpuctx
->ctx
, output
, data
);
4859 ctxn
= pmu
->task_ctx_nr
;
4862 ctx
= rcu_dereference(current
->perf_event_ctxp
[ctxn
]);
4864 perf_event_aux_ctx(ctx
, output
, data
);
4866 put_cpu_ptr(pmu
->pmu_cpu_context
);
4871 perf_event_aux_ctx(task_ctx
, output
, data
);
4878 * task tracking -- fork/exit
4880 * enabled by: attr.comm | attr.mmap | attr.mmap2 | attr.mmap_data | attr.task
4883 struct perf_task_event
{
4884 struct task_struct
*task
;
4885 struct perf_event_context
*task_ctx
;
4888 struct perf_event_header header
;
4898 static int perf_event_task_match(struct perf_event
*event
)
4900 return event
->attr
.comm
|| event
->attr
.mmap
||
4901 event
->attr
.mmap2
|| event
->attr
.mmap_data
||
4905 static void perf_event_task_output(struct perf_event
*event
,
4908 struct perf_task_event
*task_event
= data
;
4909 struct perf_output_handle handle
;
4910 struct perf_sample_data sample
;
4911 struct task_struct
*task
= task_event
->task
;
4912 int ret
, size
= task_event
->event_id
.header
.size
;
4914 if (!perf_event_task_match(event
))
4917 perf_event_header__init_id(&task_event
->event_id
.header
, &sample
, event
);
4919 ret
= perf_output_begin(&handle
, event
,
4920 task_event
->event_id
.header
.size
);
4924 task_event
->event_id
.pid
= perf_event_pid(event
, task
);
4925 task_event
->event_id
.ppid
= perf_event_pid(event
, current
);
4927 task_event
->event_id
.tid
= perf_event_tid(event
, task
);
4928 task_event
->event_id
.ptid
= perf_event_tid(event
, current
);
4930 perf_output_put(&handle
, task_event
->event_id
);
4932 perf_event__output_id_sample(event
, &handle
, &sample
);
4934 perf_output_end(&handle
);
4936 task_event
->event_id
.header
.size
= size
;
4939 static void perf_event_task(struct task_struct
*task
,
4940 struct perf_event_context
*task_ctx
,
4943 struct perf_task_event task_event
;
4945 if (!atomic_read(&nr_comm_events
) &&
4946 !atomic_read(&nr_mmap_events
) &&
4947 !atomic_read(&nr_task_events
))
4950 task_event
= (struct perf_task_event
){
4952 .task_ctx
= task_ctx
,
4955 .type
= new ? PERF_RECORD_FORK
: PERF_RECORD_EXIT
,
4957 .size
= sizeof(task_event
.event_id
),
4963 .time
= perf_clock(),
4967 perf_event_aux(perf_event_task_output
,
4972 void perf_event_fork(struct task_struct
*task
)
4974 perf_event_task(task
, NULL
, 1);
4981 struct perf_comm_event
{
4982 struct task_struct
*task
;
4987 struct perf_event_header header
;
4994 static int perf_event_comm_match(struct perf_event
*event
)
4996 return event
->attr
.comm
;
4999 static void perf_event_comm_output(struct perf_event
*event
,
5002 struct perf_comm_event
*comm_event
= data
;
5003 struct perf_output_handle handle
;
5004 struct perf_sample_data sample
;
5005 int size
= comm_event
->event_id
.header
.size
;
5008 if (!perf_event_comm_match(event
))
5011 perf_event_header__init_id(&comm_event
->event_id
.header
, &sample
, event
);
5012 ret
= perf_output_begin(&handle
, event
,
5013 comm_event
->event_id
.header
.size
);
5018 comm_event
->event_id
.pid
= perf_event_pid(event
, comm_event
->task
);
5019 comm_event
->event_id
.tid
= perf_event_tid(event
, comm_event
->task
);
5021 perf_output_put(&handle
, comm_event
->event_id
);
5022 __output_copy(&handle
, comm_event
->comm
,
5023 comm_event
->comm_size
);
5025 perf_event__output_id_sample(event
, &handle
, &sample
);
5027 perf_output_end(&handle
);
5029 comm_event
->event_id
.header
.size
= size
;
5032 static void perf_event_comm_event(struct perf_comm_event
*comm_event
)
5034 char comm
[TASK_COMM_LEN
];
5037 memset(comm
, 0, sizeof(comm
));
5038 strlcpy(comm
, comm_event
->task
->comm
, sizeof(comm
));
5039 size
= ALIGN(strlen(comm
)+1, sizeof(u64
));
5041 comm_event
->comm
= comm
;
5042 comm_event
->comm_size
= size
;
5044 comm_event
->event_id
.header
.size
= sizeof(comm_event
->event_id
) + size
;
5046 perf_event_aux(perf_event_comm_output
,
5051 void perf_event_comm(struct task_struct
*task
)
5053 struct perf_comm_event comm_event
;
5054 struct perf_event_context
*ctx
;
5058 for_each_task_context_nr(ctxn
) {
5059 ctx
= task
->perf_event_ctxp
[ctxn
];
5063 perf_event_enable_on_exec(ctx
);
5067 if (!atomic_read(&nr_comm_events
))
5070 comm_event
= (struct perf_comm_event
){
5076 .type
= PERF_RECORD_COMM
,
5085 perf_event_comm_event(&comm_event
);
5092 struct perf_mmap_event
{
5093 struct vm_area_struct
*vma
;
5095 const char *file_name
;
5102 struct perf_event_header header
;
5112 static int perf_event_mmap_match(struct perf_event
*event
,
5115 struct perf_mmap_event
*mmap_event
= data
;
5116 struct vm_area_struct
*vma
= mmap_event
->vma
;
5117 int executable
= vma
->vm_flags
& VM_EXEC
;
5119 return (!executable
&& event
->attr
.mmap_data
) ||
5120 (executable
&& (event
->attr
.mmap
|| event
->attr
.mmap2
));
5123 static void perf_event_mmap_output(struct perf_event
*event
,
5126 struct perf_mmap_event
*mmap_event
= data
;
5127 struct perf_output_handle handle
;
5128 struct perf_sample_data sample
;
5129 int size
= mmap_event
->event_id
.header
.size
;
5132 if (!perf_event_mmap_match(event
, data
))
5135 if (event
->attr
.mmap2
) {
5136 mmap_event
->event_id
.header
.type
= PERF_RECORD_MMAP2
;
5137 mmap_event
->event_id
.header
.size
+= sizeof(mmap_event
->maj
);
5138 mmap_event
->event_id
.header
.size
+= sizeof(mmap_event
->min
);
5139 mmap_event
->event_id
.header
.size
+= sizeof(mmap_event
->ino
);
5140 mmap_event
->event_id
.header
.size
+= sizeof(mmap_event
->ino_generation
);
5143 perf_event_header__init_id(&mmap_event
->event_id
.header
, &sample
, event
);
5144 ret
= perf_output_begin(&handle
, event
,
5145 mmap_event
->event_id
.header
.size
);
5149 mmap_event
->event_id
.pid
= perf_event_pid(event
, current
);
5150 mmap_event
->event_id
.tid
= perf_event_tid(event
, current
);
5152 perf_output_put(&handle
, mmap_event
->event_id
);
5154 if (event
->attr
.mmap2
) {
5155 perf_output_put(&handle
, mmap_event
->maj
);
5156 perf_output_put(&handle
, mmap_event
->min
);
5157 perf_output_put(&handle
, mmap_event
->ino
);
5158 perf_output_put(&handle
, mmap_event
->ino_generation
);
5161 __output_copy(&handle
, mmap_event
->file_name
,
5162 mmap_event
->file_size
);
5164 perf_event__output_id_sample(event
, &handle
, &sample
);
5166 perf_output_end(&handle
);
5168 mmap_event
->event_id
.header
.size
= size
;
5171 static void perf_event_mmap_event(struct perf_mmap_event
*mmap_event
)
5173 struct vm_area_struct
*vma
= mmap_event
->vma
;
5174 struct file
*file
= vma
->vm_file
;
5175 int maj
= 0, min
= 0;
5176 u64 ino
= 0, gen
= 0;
5183 struct inode
*inode
;
5186 buf
= kmalloc(PATH_MAX
, GFP_KERNEL
);
5192 * d_path() works from the end of the rb backwards, so we
5193 * need to add enough zero bytes after the string to handle
5194 * the 64bit alignment we do later.
5196 name
= d_path(&file
->f_path
, buf
, PATH_MAX
- sizeof(u64
));
5201 inode
= file_inode(vma
->vm_file
);
5202 dev
= inode
->i_sb
->s_dev
;
5204 gen
= inode
->i_generation
;
5209 name
= (char *)arch_vma_name(vma
);
5213 if (vma
->vm_start
<= vma
->vm_mm
->start_brk
&&
5214 vma
->vm_end
>= vma
->vm_mm
->brk
) {
5218 if (vma
->vm_start
<= vma
->vm_mm
->start_stack
&&
5219 vma
->vm_end
>= vma
->vm_mm
->start_stack
) {
5229 strlcpy(tmp
, name
, sizeof(tmp
));
5233 * Since our buffer works in 8 byte units we need to align our string
5234 * size to a multiple of 8. However, we must guarantee the tail end is
5235 * zero'd out to avoid leaking random bits to userspace.
5237 size
= strlen(name
)+1;
5238 while (!IS_ALIGNED(size
, sizeof(u64
)))
5239 name
[size
++] = '\0';
5241 mmap_event
->file_name
= name
;
5242 mmap_event
->file_size
= size
;
5243 mmap_event
->maj
= maj
;
5244 mmap_event
->min
= min
;
5245 mmap_event
->ino
= ino
;
5246 mmap_event
->ino_generation
= gen
;
5248 if (!(vma
->vm_flags
& VM_EXEC
))
5249 mmap_event
->event_id
.header
.misc
|= PERF_RECORD_MISC_MMAP_DATA
;
5251 mmap_event
->event_id
.header
.size
= sizeof(mmap_event
->event_id
) + size
;
5253 perf_event_aux(perf_event_mmap_output
,
5260 void perf_event_mmap(struct vm_area_struct
*vma
)
5262 struct perf_mmap_event mmap_event
;
5264 if (!atomic_read(&nr_mmap_events
))
5267 mmap_event
= (struct perf_mmap_event
){
5273 .type
= PERF_RECORD_MMAP
,
5274 .misc
= PERF_RECORD_MISC_USER
,
5279 .start
= vma
->vm_start
,
5280 .len
= vma
->vm_end
- vma
->vm_start
,
5281 .pgoff
= (u64
)vma
->vm_pgoff
<< PAGE_SHIFT
,
5283 /* .maj (attr_mmap2 only) */
5284 /* .min (attr_mmap2 only) */
5285 /* .ino (attr_mmap2 only) */
5286 /* .ino_generation (attr_mmap2 only) */
5289 perf_event_mmap_event(&mmap_event
);
5293 * IRQ throttle logging
5296 static void perf_log_throttle(struct perf_event
*event
, int enable
)
5298 struct perf_output_handle handle
;
5299 struct perf_sample_data sample
;
5303 struct perf_event_header header
;
5307 } throttle_event
= {
5309 .type
= PERF_RECORD_THROTTLE
,
5311 .size
= sizeof(throttle_event
),
5313 .time
= perf_clock(),
5314 .id
= primary_event_id(event
),
5315 .stream_id
= event
->id
,
5319 throttle_event
.header
.type
= PERF_RECORD_UNTHROTTLE
;
5321 perf_event_header__init_id(&throttle_event
.header
, &sample
, event
);
5323 ret
= perf_output_begin(&handle
, event
,
5324 throttle_event
.header
.size
);
5328 perf_output_put(&handle
, throttle_event
);
5329 perf_event__output_id_sample(event
, &handle
, &sample
);
5330 perf_output_end(&handle
);
5334 * Generic event overflow handling, sampling.
5337 static int __perf_event_overflow(struct perf_event
*event
,
5338 int throttle
, struct perf_sample_data
*data
,
5339 struct pt_regs
*regs
)
5341 int events
= atomic_read(&event
->event_limit
);
5342 struct hw_perf_event
*hwc
= &event
->hw
;
5347 * Non-sampling counters might still use the PMI to fold short
5348 * hardware counters, ignore those.
5350 if (unlikely(!is_sampling_event(event
)))
5353 seq
= __this_cpu_read(perf_throttled_seq
);
5354 if (seq
!= hwc
->interrupts_seq
) {
5355 hwc
->interrupts_seq
= seq
;
5356 hwc
->interrupts
= 1;
5359 if (unlikely(throttle
5360 && hwc
->interrupts
>= max_samples_per_tick
)) {
5361 __this_cpu_inc(perf_throttled_count
);
5362 hwc
->interrupts
= MAX_INTERRUPTS
;
5363 perf_log_throttle(event
, 0);
5364 tick_nohz_full_kick();
5369 if (event
->attr
.freq
) {
5370 u64 now
= perf_clock();
5371 s64 delta
= now
- hwc
->freq_time_stamp
;
5373 hwc
->freq_time_stamp
= now
;
5375 if (delta
> 0 && delta
< 2*TICK_NSEC
)
5376 perf_adjust_period(event
, delta
, hwc
->last_period
, true);
5380 * XXX event_limit might not quite work as expected on inherited
5384 event
->pending_kill
= POLL_IN
;
5385 if (events
&& atomic_dec_and_test(&event
->event_limit
)) {
5387 event
->pending_kill
= POLL_HUP
;
5388 event
->pending_disable
= 1;
5389 irq_work_queue(&event
->pending
);
5392 if (event
->overflow_handler
)
5393 event
->overflow_handler(event
, data
, regs
);
5395 perf_event_output(event
, data
, regs
);
5397 if (event
->fasync
&& event
->pending_kill
) {
5398 event
->pending_wakeup
= 1;
5399 irq_work_queue(&event
->pending
);
5405 int perf_event_overflow(struct perf_event
*event
,
5406 struct perf_sample_data
*data
,
5407 struct pt_regs
*regs
)
5409 return __perf_event_overflow(event
, 1, data
, regs
);
5413 * Generic software event infrastructure
5416 struct swevent_htable
{
5417 struct swevent_hlist
*swevent_hlist
;
5418 struct mutex hlist_mutex
;
5421 /* Recursion avoidance in each contexts */
5422 int recursion
[PERF_NR_CONTEXTS
];
5425 static DEFINE_PER_CPU(struct swevent_htable
, swevent_htable
);
5428 * We directly increment event->count and keep a second value in
5429 * event->hw.period_left to count intervals. This period event
5430 * is kept in the range [-sample_period, 0] so that we can use the
5434 u64
perf_swevent_set_period(struct perf_event
*event
)
5436 struct hw_perf_event
*hwc
= &event
->hw
;
5437 u64 period
= hwc
->last_period
;
5441 hwc
->last_period
= hwc
->sample_period
;
5444 old
= val
= local64_read(&hwc
->period_left
);
5448 nr
= div64_u64(period
+ val
, period
);
5449 offset
= nr
* period
;
5451 if (local64_cmpxchg(&hwc
->period_left
, old
, val
) != old
)
5457 static void perf_swevent_overflow(struct perf_event
*event
, u64 overflow
,
5458 struct perf_sample_data
*data
,
5459 struct pt_regs
*regs
)
5461 struct hw_perf_event
*hwc
= &event
->hw
;
5465 overflow
= perf_swevent_set_period(event
);
5467 if (hwc
->interrupts
== MAX_INTERRUPTS
)
5470 for (; overflow
; overflow
--) {
5471 if (__perf_event_overflow(event
, throttle
,
5474 * We inhibit the overflow from happening when
5475 * hwc->interrupts == MAX_INTERRUPTS.
5483 static void perf_swevent_event(struct perf_event
*event
, u64 nr
,
5484 struct perf_sample_data
*data
,
5485 struct pt_regs
*regs
)
5487 struct hw_perf_event
*hwc
= &event
->hw
;
5489 local64_add(nr
, &event
->count
);
5494 if (!is_sampling_event(event
))
5497 if ((event
->attr
.sample_type
& PERF_SAMPLE_PERIOD
) && !event
->attr
.freq
) {
5499 return perf_swevent_overflow(event
, 1, data
, regs
);
5501 data
->period
= event
->hw
.last_period
;
5503 if (nr
== 1 && hwc
->sample_period
== 1 && !event
->attr
.freq
)
5504 return perf_swevent_overflow(event
, 1, data
, regs
);
5506 if (local64_add_negative(nr
, &hwc
->period_left
))
5509 perf_swevent_overflow(event
, 0, data
, regs
);
5512 static int perf_exclude_event(struct perf_event
*event
,
5513 struct pt_regs
*regs
)
5515 if (event
->hw
.state
& PERF_HES_STOPPED
)
5519 if (event
->attr
.exclude_user
&& user_mode(regs
))
5522 if (event
->attr
.exclude_kernel
&& !user_mode(regs
))
5529 static int perf_swevent_match(struct perf_event
*event
,
5530 enum perf_type_id type
,
5532 struct perf_sample_data
*data
,
5533 struct pt_regs
*regs
)
5535 if (event
->attr
.type
!= type
)
5538 if (event
->attr
.config
!= event_id
)
5541 if (perf_exclude_event(event
, regs
))
5547 static inline u64
swevent_hash(u64 type
, u32 event_id
)
5549 u64 val
= event_id
| (type
<< 32);
5551 return hash_64(val
, SWEVENT_HLIST_BITS
);
5554 static inline struct hlist_head
*
5555 __find_swevent_head(struct swevent_hlist
*hlist
, u64 type
, u32 event_id
)
5557 u64 hash
= swevent_hash(type
, event_id
);
5559 return &hlist
->heads
[hash
];
5562 /* For the read side: events when they trigger */
5563 static inline struct hlist_head
*
5564 find_swevent_head_rcu(struct swevent_htable
*swhash
, u64 type
, u32 event_id
)
5566 struct swevent_hlist
*hlist
;
5568 hlist
= rcu_dereference(swhash
->swevent_hlist
);
5572 return __find_swevent_head(hlist
, type
, event_id
);
5575 /* For the event head insertion and removal in the hlist */
5576 static inline struct hlist_head
*
5577 find_swevent_head(struct swevent_htable
*swhash
, struct perf_event
*event
)
5579 struct swevent_hlist
*hlist
;
5580 u32 event_id
= event
->attr
.config
;
5581 u64 type
= event
->attr
.type
;
5584 * Event scheduling is always serialized against hlist allocation
5585 * and release. Which makes the protected version suitable here.
5586 * The context lock guarantees that.
5588 hlist
= rcu_dereference_protected(swhash
->swevent_hlist
,
5589 lockdep_is_held(&event
->ctx
->lock
));
5593 return __find_swevent_head(hlist
, type
, event_id
);
5596 static void do_perf_sw_event(enum perf_type_id type
, u32 event_id
,
5598 struct perf_sample_data
*data
,
5599 struct pt_regs
*regs
)
5601 struct swevent_htable
*swhash
= &__get_cpu_var(swevent_htable
);
5602 struct perf_event
*event
;
5603 struct hlist_head
*head
;
5606 head
= find_swevent_head_rcu(swhash
, type
, event_id
);
5610 hlist_for_each_entry_rcu(event
, head
, hlist_entry
) {
5611 if (perf_swevent_match(event
, type
, event_id
, data
, regs
))
5612 perf_swevent_event(event
, nr
, data
, regs
);
5618 int perf_swevent_get_recursion_context(void)
5620 struct swevent_htable
*swhash
= &__get_cpu_var(swevent_htable
);
5622 return get_recursion_context(swhash
->recursion
);
5624 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context
);
5626 inline void perf_swevent_put_recursion_context(int rctx
)
5628 struct swevent_htable
*swhash
= &__get_cpu_var(swevent_htable
);
5630 put_recursion_context(swhash
->recursion
, rctx
);
5633 void __perf_sw_event(u32 event_id
, u64 nr
, struct pt_regs
*regs
, u64 addr
)
5635 struct perf_sample_data data
;
5638 preempt_disable_notrace();
5639 rctx
= perf_swevent_get_recursion_context();
5643 perf_sample_data_init(&data
, addr
, 0);
5645 do_perf_sw_event(PERF_TYPE_SOFTWARE
, event_id
, nr
, &data
, regs
);
5647 perf_swevent_put_recursion_context(rctx
);
5648 preempt_enable_notrace();
5651 static void perf_swevent_read(struct perf_event
*event
)
5655 static int perf_swevent_add(struct perf_event
*event
, int flags
)
5657 struct swevent_htable
*swhash
= &__get_cpu_var(swevent_htable
);
5658 struct hw_perf_event
*hwc
= &event
->hw
;
5659 struct hlist_head
*head
;
5661 if (is_sampling_event(event
)) {
5662 hwc
->last_period
= hwc
->sample_period
;
5663 perf_swevent_set_period(event
);
5666 hwc
->state
= !(flags
& PERF_EF_START
);
5668 head
= find_swevent_head(swhash
, event
);
5669 if (WARN_ON_ONCE(!head
))
5672 hlist_add_head_rcu(&event
->hlist_entry
, head
);
5677 static void perf_swevent_del(struct perf_event
*event
, int flags
)
5679 hlist_del_rcu(&event
->hlist_entry
);
5682 static void perf_swevent_start(struct perf_event
*event
, int flags
)
5684 event
->hw
.state
= 0;
5687 static void perf_swevent_stop(struct perf_event
*event
, int flags
)
5689 event
->hw
.state
= PERF_HES_STOPPED
;
5692 /* Deref the hlist from the update side */
5693 static inline struct swevent_hlist
*
5694 swevent_hlist_deref(struct swevent_htable
*swhash
)
5696 return rcu_dereference_protected(swhash
->swevent_hlist
,
5697 lockdep_is_held(&swhash
->hlist_mutex
));
5700 static void swevent_hlist_release(struct swevent_htable
*swhash
)
5702 struct swevent_hlist
*hlist
= swevent_hlist_deref(swhash
);
5707 rcu_assign_pointer(swhash
->swevent_hlist
, NULL
);
5708 kfree_rcu(hlist
, rcu_head
);
5711 static void swevent_hlist_put_cpu(struct perf_event
*event
, int cpu
)
5713 struct swevent_htable
*swhash
= &per_cpu(swevent_htable
, cpu
);
5715 mutex_lock(&swhash
->hlist_mutex
);
5717 if (!--swhash
->hlist_refcount
)
5718 swevent_hlist_release(swhash
);
5720 mutex_unlock(&swhash
->hlist_mutex
);
5723 static void swevent_hlist_put(struct perf_event
*event
)
5727 for_each_possible_cpu(cpu
)
5728 swevent_hlist_put_cpu(event
, cpu
);
5731 static int swevent_hlist_get_cpu(struct perf_event
*event
, int cpu
)
5733 struct swevent_htable
*swhash
= &per_cpu(swevent_htable
, cpu
);
5736 mutex_lock(&swhash
->hlist_mutex
);
5738 if (!swevent_hlist_deref(swhash
) && cpu_online(cpu
)) {
5739 struct swevent_hlist
*hlist
;
5741 hlist
= kzalloc(sizeof(*hlist
), GFP_KERNEL
);
5746 rcu_assign_pointer(swhash
->swevent_hlist
, hlist
);
5748 swhash
->hlist_refcount
++;
5750 mutex_unlock(&swhash
->hlist_mutex
);
5755 static int swevent_hlist_get(struct perf_event
*event
)
5758 int cpu
, failed_cpu
;
5761 for_each_possible_cpu(cpu
) {
5762 err
= swevent_hlist_get_cpu(event
, cpu
);
5772 for_each_possible_cpu(cpu
) {
5773 if (cpu
== failed_cpu
)
5775 swevent_hlist_put_cpu(event
, cpu
);
5782 struct static_key perf_swevent_enabled
[PERF_COUNT_SW_MAX
];
5784 static void sw_perf_event_destroy(struct perf_event
*event
)
5786 u64 event_id
= event
->attr
.config
;
5788 WARN_ON(event
->parent
);
5790 static_key_slow_dec(&perf_swevent_enabled
[event_id
]);
5791 swevent_hlist_put(event
);
5794 static int perf_swevent_init(struct perf_event
*event
)
5796 u64 event_id
= event
->attr
.config
;
5798 if (event
->attr
.type
!= PERF_TYPE_SOFTWARE
)
5802 * no branch sampling for software events
5804 if (has_branch_stack(event
))
5808 case PERF_COUNT_SW_CPU_CLOCK
:
5809 case PERF_COUNT_SW_TASK_CLOCK
:
5816 if (event_id
>= PERF_COUNT_SW_MAX
)
5819 if (!event
->parent
) {
5822 err
= swevent_hlist_get(event
);
5826 static_key_slow_inc(&perf_swevent_enabled
[event_id
]);
5827 event
->destroy
= sw_perf_event_destroy
;
5833 static int perf_swevent_event_idx(struct perf_event
*event
)
5838 static struct pmu perf_swevent
= {
5839 .task_ctx_nr
= perf_sw_context
,
5841 .event_init
= perf_swevent_init
,
5842 .add
= perf_swevent_add
,
5843 .del
= perf_swevent_del
,
5844 .start
= perf_swevent_start
,
5845 .stop
= perf_swevent_stop
,
5846 .read
= perf_swevent_read
,
5848 .event_idx
= perf_swevent_event_idx
,
5851 #ifdef CONFIG_EVENT_TRACING
5853 static int perf_tp_filter_match(struct perf_event
*event
,
5854 struct perf_sample_data
*data
)
5856 void *record
= data
->raw
->data
;
5858 if (likely(!event
->filter
) || filter_match_preds(event
->filter
, record
))
5863 static int perf_tp_event_match(struct perf_event
*event
,
5864 struct perf_sample_data
*data
,
5865 struct pt_regs
*regs
)
5867 if (event
->hw
.state
& PERF_HES_STOPPED
)
5870 * All tracepoints are from kernel-space.
5872 if (event
->attr
.exclude_kernel
)
5875 if (!perf_tp_filter_match(event
, data
))
5881 void perf_tp_event(u64 addr
, u64 count
, void *record
, int entry_size
,
5882 struct pt_regs
*regs
, struct hlist_head
*head
, int rctx
,
5883 struct task_struct
*task
)
5885 struct perf_sample_data data
;
5886 struct perf_event
*event
;
5888 struct perf_raw_record raw
= {
5893 perf_sample_data_init(&data
, addr
, 0);
5896 hlist_for_each_entry_rcu(event
, head
, hlist_entry
) {
5897 if (perf_tp_event_match(event
, &data
, regs
))
5898 perf_swevent_event(event
, count
, &data
, regs
);
5902 * If we got specified a target task, also iterate its context and
5903 * deliver this event there too.
5905 if (task
&& task
!= current
) {
5906 struct perf_event_context
*ctx
;
5907 struct trace_entry
*entry
= record
;
5910 ctx
= rcu_dereference(task
->perf_event_ctxp
[perf_sw_context
]);
5914 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
5915 if (event
->attr
.type
!= PERF_TYPE_TRACEPOINT
)
5917 if (event
->attr
.config
!= entry
->type
)
5919 if (perf_tp_event_match(event
, &data
, regs
))
5920 perf_swevent_event(event
, count
, &data
, regs
);
5926 perf_swevent_put_recursion_context(rctx
);
5928 EXPORT_SYMBOL_GPL(perf_tp_event
);
5930 static void tp_perf_event_destroy(struct perf_event
*event
)
5932 perf_trace_destroy(event
);
5935 static int perf_tp_event_init(struct perf_event
*event
)
5939 if (event
->attr
.type
!= PERF_TYPE_TRACEPOINT
)
5943 * no branch sampling for tracepoint events
5945 if (has_branch_stack(event
))
5948 err
= perf_trace_init(event
);
5952 event
->destroy
= tp_perf_event_destroy
;
5957 static struct pmu perf_tracepoint
= {
5958 .task_ctx_nr
= perf_sw_context
,
5960 .event_init
= perf_tp_event_init
,
5961 .add
= perf_trace_add
,
5962 .del
= perf_trace_del
,
5963 .start
= perf_swevent_start
,
5964 .stop
= perf_swevent_stop
,
5965 .read
= perf_swevent_read
,
5967 .event_idx
= perf_swevent_event_idx
,
5970 static inline void perf_tp_register(void)
5972 perf_pmu_register(&perf_tracepoint
, "tracepoint", PERF_TYPE_TRACEPOINT
);
5975 static int perf_event_set_filter(struct perf_event
*event
, void __user
*arg
)
5980 if (event
->attr
.type
!= PERF_TYPE_TRACEPOINT
)
5983 filter_str
= strndup_user(arg
, PAGE_SIZE
);
5984 if (IS_ERR(filter_str
))
5985 return PTR_ERR(filter_str
);
5987 ret
= ftrace_profile_set_filter(event
, event
->attr
.config
, filter_str
);
5993 static void perf_event_free_filter(struct perf_event
*event
)
5995 ftrace_profile_free_filter(event
);
6000 static inline void perf_tp_register(void)
6004 static int perf_event_set_filter(struct perf_event
*event
, void __user
*arg
)
6009 static void perf_event_free_filter(struct perf_event
*event
)
6013 #endif /* CONFIG_EVENT_TRACING */
6015 #ifdef CONFIG_HAVE_HW_BREAKPOINT
6016 void perf_bp_event(struct perf_event
*bp
, void *data
)
6018 struct perf_sample_data sample
;
6019 struct pt_regs
*regs
= data
;
6021 perf_sample_data_init(&sample
, bp
->attr
.bp_addr
, 0);
6023 if (!bp
->hw
.state
&& !perf_exclude_event(bp
, regs
))
6024 perf_swevent_event(bp
, 1, &sample
, regs
);
6029 * hrtimer based swevent callback
6032 static enum hrtimer_restart
perf_swevent_hrtimer(struct hrtimer
*hrtimer
)
6034 enum hrtimer_restart ret
= HRTIMER_RESTART
;
6035 struct perf_sample_data data
;
6036 struct pt_regs
*regs
;
6037 struct perf_event
*event
;
6040 event
= container_of(hrtimer
, struct perf_event
, hw
.hrtimer
);
6042 if (event
->state
!= PERF_EVENT_STATE_ACTIVE
)
6043 return HRTIMER_NORESTART
;
6045 event
->pmu
->read(event
);
6047 perf_sample_data_init(&data
, 0, event
->hw
.last_period
);
6048 regs
= get_irq_regs();
6050 if (regs
&& !perf_exclude_event(event
, regs
)) {
6051 if (!(event
->attr
.exclude_idle
&& is_idle_task(current
)))
6052 if (__perf_event_overflow(event
, 1, &data
, regs
))
6053 ret
= HRTIMER_NORESTART
;
6056 period
= max_t(u64
, 10000, event
->hw
.sample_period
);
6057 hrtimer_forward_now(hrtimer
, ns_to_ktime(period
));
6062 static void perf_swevent_start_hrtimer(struct perf_event
*event
)
6064 struct hw_perf_event
*hwc
= &event
->hw
;
6067 if (!is_sampling_event(event
))
6070 period
= local64_read(&hwc
->period_left
);
6075 local64_set(&hwc
->period_left
, 0);
6077 period
= max_t(u64
, 10000, hwc
->sample_period
);
6079 __hrtimer_start_range_ns(&hwc
->hrtimer
,
6080 ns_to_ktime(period
), 0,
6081 HRTIMER_MODE_REL_PINNED
, 0);
6084 static void perf_swevent_cancel_hrtimer(struct perf_event
*event
)
6086 struct hw_perf_event
*hwc
= &event
->hw
;
6088 if (is_sampling_event(event
)) {
6089 ktime_t remaining
= hrtimer_get_remaining(&hwc
->hrtimer
);
6090 local64_set(&hwc
->period_left
, ktime_to_ns(remaining
));
6092 hrtimer_cancel(&hwc
->hrtimer
);
6096 static void perf_swevent_init_hrtimer(struct perf_event
*event
)
6098 struct hw_perf_event
*hwc
= &event
->hw
;
6100 if (!is_sampling_event(event
))
6103 hrtimer_init(&hwc
->hrtimer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
6104 hwc
->hrtimer
.function
= perf_swevent_hrtimer
;
6107 * Since hrtimers have a fixed rate, we can do a static freq->period
6108 * mapping and avoid the whole period adjust feedback stuff.
6110 if (event
->attr
.freq
) {
6111 long freq
= event
->attr
.sample_freq
;
6113 event
->attr
.sample_period
= NSEC_PER_SEC
/ freq
;
6114 hwc
->sample_period
= event
->attr
.sample_period
;
6115 local64_set(&hwc
->period_left
, hwc
->sample_period
);
6116 hwc
->last_period
= hwc
->sample_period
;
6117 event
->attr
.freq
= 0;
6122 * Software event: cpu wall time clock
6125 static void cpu_clock_event_update(struct perf_event
*event
)
6130 now
= local_clock();
6131 prev
= local64_xchg(&event
->hw
.prev_count
, now
);
6132 local64_add(now
- prev
, &event
->count
);
6135 static void cpu_clock_event_start(struct perf_event
*event
, int flags
)
6137 local64_set(&event
->hw
.prev_count
, local_clock());
6138 perf_swevent_start_hrtimer(event
);
6141 static void cpu_clock_event_stop(struct perf_event
*event
, int flags
)
6143 perf_swevent_cancel_hrtimer(event
);
6144 cpu_clock_event_update(event
);
6147 static int cpu_clock_event_add(struct perf_event
*event
, int flags
)
6149 if (flags
& PERF_EF_START
)
6150 cpu_clock_event_start(event
, flags
);
6155 static void cpu_clock_event_del(struct perf_event
*event
, int flags
)
6157 cpu_clock_event_stop(event
, flags
);
6160 static void cpu_clock_event_read(struct perf_event
*event
)
6162 cpu_clock_event_update(event
);
6165 static int cpu_clock_event_init(struct perf_event
*event
)
6167 if (event
->attr
.type
!= PERF_TYPE_SOFTWARE
)
6170 if (event
->attr
.config
!= PERF_COUNT_SW_CPU_CLOCK
)
6174 * no branch sampling for software events
6176 if (has_branch_stack(event
))
6179 perf_swevent_init_hrtimer(event
);
6184 static struct pmu perf_cpu_clock
= {
6185 .task_ctx_nr
= perf_sw_context
,
6187 .event_init
= cpu_clock_event_init
,
6188 .add
= cpu_clock_event_add
,
6189 .del
= cpu_clock_event_del
,
6190 .start
= cpu_clock_event_start
,
6191 .stop
= cpu_clock_event_stop
,
6192 .read
= cpu_clock_event_read
,
6194 .event_idx
= perf_swevent_event_idx
,
6198 * Software event: task time clock
6201 static void task_clock_event_update(struct perf_event
*event
, u64 now
)
6206 prev
= local64_xchg(&event
->hw
.prev_count
, now
);
6208 local64_add(delta
, &event
->count
);
6211 static void task_clock_event_start(struct perf_event
*event
, int flags
)
6213 local64_set(&event
->hw
.prev_count
, event
->ctx
->time
);
6214 perf_swevent_start_hrtimer(event
);
6217 static void task_clock_event_stop(struct perf_event
*event
, int flags
)
6219 perf_swevent_cancel_hrtimer(event
);
6220 task_clock_event_update(event
, event
->ctx
->time
);
6223 static int task_clock_event_add(struct perf_event
*event
, int flags
)
6225 if (flags
& PERF_EF_START
)
6226 task_clock_event_start(event
, flags
);
6231 static void task_clock_event_del(struct perf_event
*event
, int flags
)
6233 task_clock_event_stop(event
, PERF_EF_UPDATE
);
6236 static void task_clock_event_read(struct perf_event
*event
)
6238 u64 now
= perf_clock();
6239 u64 delta
= now
- event
->ctx
->timestamp
;
6240 u64 time
= event
->ctx
->time
+ delta
;
6242 task_clock_event_update(event
, time
);
6245 static int task_clock_event_init(struct perf_event
*event
)
6247 if (event
->attr
.type
!= PERF_TYPE_SOFTWARE
)
6250 if (event
->attr
.config
!= PERF_COUNT_SW_TASK_CLOCK
)
6254 * no branch sampling for software events
6256 if (has_branch_stack(event
))
6259 perf_swevent_init_hrtimer(event
);
6264 static struct pmu perf_task_clock
= {
6265 .task_ctx_nr
= perf_sw_context
,
6267 .event_init
= task_clock_event_init
,
6268 .add
= task_clock_event_add
,
6269 .del
= task_clock_event_del
,
6270 .start
= task_clock_event_start
,
6271 .stop
= task_clock_event_stop
,
6272 .read
= task_clock_event_read
,
6274 .event_idx
= perf_swevent_event_idx
,
6277 static void perf_pmu_nop_void(struct pmu
*pmu
)
6281 static int perf_pmu_nop_int(struct pmu
*pmu
)
6286 static void perf_pmu_start_txn(struct pmu
*pmu
)
6288 perf_pmu_disable(pmu
);
6291 static int perf_pmu_commit_txn(struct pmu
*pmu
)
6293 perf_pmu_enable(pmu
);
6297 static void perf_pmu_cancel_txn(struct pmu
*pmu
)
6299 perf_pmu_enable(pmu
);
6302 static int perf_event_idx_default(struct perf_event
*event
)
6304 return event
->hw
.idx
+ 1;
6308 * Ensures all contexts with the same task_ctx_nr have the same
6309 * pmu_cpu_context too.
6311 static void *find_pmu_context(int ctxn
)
6318 list_for_each_entry(pmu
, &pmus
, entry
) {
6319 if (pmu
->task_ctx_nr
== ctxn
)
6320 return pmu
->pmu_cpu_context
;
6326 static void update_pmu_context(struct pmu
*pmu
, struct pmu
*old_pmu
)
6330 for_each_possible_cpu(cpu
) {
6331 struct perf_cpu_context
*cpuctx
;
6333 cpuctx
= per_cpu_ptr(pmu
->pmu_cpu_context
, cpu
);
6335 if (cpuctx
->unique_pmu
== old_pmu
)
6336 cpuctx
->unique_pmu
= pmu
;
6340 static void free_pmu_context(struct pmu
*pmu
)
6344 mutex_lock(&pmus_lock
);
6346 * Like a real lame refcount.
6348 list_for_each_entry(i
, &pmus
, entry
) {
6349 if (i
->pmu_cpu_context
== pmu
->pmu_cpu_context
) {
6350 update_pmu_context(i
, pmu
);
6355 free_percpu(pmu
->pmu_cpu_context
);
6357 mutex_unlock(&pmus_lock
);
6359 static struct idr pmu_idr
;
6362 type_show(struct device
*dev
, struct device_attribute
*attr
, char *page
)
6364 struct pmu
*pmu
= dev_get_drvdata(dev
);
6366 return snprintf(page
, PAGE_SIZE
-1, "%d\n", pmu
->type
);
6368 static DEVICE_ATTR_RO(type
);
6371 perf_event_mux_interval_ms_show(struct device
*dev
,
6372 struct device_attribute
*attr
,
6375 struct pmu
*pmu
= dev_get_drvdata(dev
);
6377 return snprintf(page
, PAGE_SIZE
-1, "%d\n", pmu
->hrtimer_interval_ms
);
6381 perf_event_mux_interval_ms_store(struct device
*dev
,
6382 struct device_attribute
*attr
,
6383 const char *buf
, size_t count
)
6385 struct pmu
*pmu
= dev_get_drvdata(dev
);
6386 int timer
, cpu
, ret
;
6388 ret
= kstrtoint(buf
, 0, &timer
);
6395 /* same value, noting to do */
6396 if (timer
== pmu
->hrtimer_interval_ms
)
6399 pmu
->hrtimer_interval_ms
= timer
;
6401 /* update all cpuctx for this PMU */
6402 for_each_possible_cpu(cpu
) {
6403 struct perf_cpu_context
*cpuctx
;
6404 cpuctx
= per_cpu_ptr(pmu
->pmu_cpu_context
, cpu
);
6405 cpuctx
->hrtimer_interval
= ns_to_ktime(NSEC_PER_MSEC
* timer
);
6407 if (hrtimer_active(&cpuctx
->hrtimer
))
6408 hrtimer_forward_now(&cpuctx
->hrtimer
, cpuctx
->hrtimer_interval
);
6413 static DEVICE_ATTR_RW(perf_event_mux_interval_ms
);
6415 static struct attribute
*pmu_dev_attrs
[] = {
6416 &dev_attr_type
.attr
,
6417 &dev_attr_perf_event_mux_interval_ms
.attr
,
6420 ATTRIBUTE_GROUPS(pmu_dev
);
6422 static int pmu_bus_running
;
6423 static struct bus_type pmu_bus
= {
6424 .name
= "event_source",
6425 .dev_groups
= pmu_dev_groups
,
6428 static void pmu_dev_release(struct device
*dev
)
6433 static int pmu_dev_alloc(struct pmu
*pmu
)
6437 pmu
->dev
= kzalloc(sizeof(struct device
), GFP_KERNEL
);
6441 pmu
->dev
->groups
= pmu
->attr_groups
;
6442 device_initialize(pmu
->dev
);
6443 ret
= dev_set_name(pmu
->dev
, "%s", pmu
->name
);
6447 dev_set_drvdata(pmu
->dev
, pmu
);
6448 pmu
->dev
->bus
= &pmu_bus
;
6449 pmu
->dev
->release
= pmu_dev_release
;
6450 ret
= device_add(pmu
->dev
);
6458 put_device(pmu
->dev
);
6462 static struct lock_class_key cpuctx_mutex
;
6463 static struct lock_class_key cpuctx_lock
;
6465 int perf_pmu_register(struct pmu
*pmu
, const char *name
, int type
)
6469 mutex_lock(&pmus_lock
);
6471 pmu
->pmu_disable_count
= alloc_percpu(int);
6472 if (!pmu
->pmu_disable_count
)
6481 type
= idr_alloc(&pmu_idr
, pmu
, PERF_TYPE_MAX
, 0, GFP_KERNEL
);
6489 if (pmu_bus_running
) {
6490 ret
= pmu_dev_alloc(pmu
);
6496 pmu
->pmu_cpu_context
= find_pmu_context(pmu
->task_ctx_nr
);
6497 if (pmu
->pmu_cpu_context
)
6498 goto got_cpu_context
;
6501 pmu
->pmu_cpu_context
= alloc_percpu(struct perf_cpu_context
);
6502 if (!pmu
->pmu_cpu_context
)
6505 for_each_possible_cpu(cpu
) {
6506 struct perf_cpu_context
*cpuctx
;
6508 cpuctx
= per_cpu_ptr(pmu
->pmu_cpu_context
, cpu
);
6509 __perf_event_init_context(&cpuctx
->ctx
);
6510 lockdep_set_class(&cpuctx
->ctx
.mutex
, &cpuctx_mutex
);
6511 lockdep_set_class(&cpuctx
->ctx
.lock
, &cpuctx_lock
);
6512 cpuctx
->ctx
.type
= cpu_context
;
6513 cpuctx
->ctx
.pmu
= pmu
;
6515 __perf_cpu_hrtimer_init(cpuctx
, cpu
);
6517 INIT_LIST_HEAD(&cpuctx
->rotation_list
);
6518 cpuctx
->unique_pmu
= pmu
;
6522 if (!pmu
->start_txn
) {
6523 if (pmu
->pmu_enable
) {
6525 * If we have pmu_enable/pmu_disable calls, install
6526 * transaction stubs that use that to try and batch
6527 * hardware accesses.
6529 pmu
->start_txn
= perf_pmu_start_txn
;
6530 pmu
->commit_txn
= perf_pmu_commit_txn
;
6531 pmu
->cancel_txn
= perf_pmu_cancel_txn
;
6533 pmu
->start_txn
= perf_pmu_nop_void
;
6534 pmu
->commit_txn
= perf_pmu_nop_int
;
6535 pmu
->cancel_txn
= perf_pmu_nop_void
;
6539 if (!pmu
->pmu_enable
) {
6540 pmu
->pmu_enable
= perf_pmu_nop_void
;
6541 pmu
->pmu_disable
= perf_pmu_nop_void
;
6544 if (!pmu
->event_idx
)
6545 pmu
->event_idx
= perf_event_idx_default
;
6547 list_add_rcu(&pmu
->entry
, &pmus
);
6550 mutex_unlock(&pmus_lock
);
6555 device_del(pmu
->dev
);
6556 put_device(pmu
->dev
);
6559 if (pmu
->type
>= PERF_TYPE_MAX
)
6560 idr_remove(&pmu_idr
, pmu
->type
);
6563 free_percpu(pmu
->pmu_disable_count
);
6567 void perf_pmu_unregister(struct pmu
*pmu
)
6569 mutex_lock(&pmus_lock
);
6570 list_del_rcu(&pmu
->entry
);
6571 mutex_unlock(&pmus_lock
);
6574 * We dereference the pmu list under both SRCU and regular RCU, so
6575 * synchronize against both of those.
6577 synchronize_srcu(&pmus_srcu
);
6580 free_percpu(pmu
->pmu_disable_count
);
6581 if (pmu
->type
>= PERF_TYPE_MAX
)
6582 idr_remove(&pmu_idr
, pmu
->type
);
6583 device_del(pmu
->dev
);
6584 put_device(pmu
->dev
);
6585 free_pmu_context(pmu
);
6588 struct pmu
*perf_init_event(struct perf_event
*event
)
6590 struct pmu
*pmu
= NULL
;
6594 idx
= srcu_read_lock(&pmus_srcu
);
6597 pmu
= idr_find(&pmu_idr
, event
->attr
.type
);
6601 ret
= pmu
->event_init(event
);
6607 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
6609 ret
= pmu
->event_init(event
);
6613 if (ret
!= -ENOENT
) {
6618 pmu
= ERR_PTR(-ENOENT
);
6620 srcu_read_unlock(&pmus_srcu
, idx
);
6625 static void account_event_cpu(struct perf_event
*event
, int cpu
)
6630 if (has_branch_stack(event
)) {
6631 if (!(event
->attach_state
& PERF_ATTACH_TASK
))
6632 atomic_inc(&per_cpu(perf_branch_stack_events
, cpu
));
6634 if (is_cgroup_event(event
))
6635 atomic_inc(&per_cpu(perf_cgroup_events
, cpu
));
6638 static void account_event(struct perf_event
*event
)
6643 if (event
->attach_state
& PERF_ATTACH_TASK
)
6644 static_key_slow_inc(&perf_sched_events
.key
);
6645 if (event
->attr
.mmap
|| event
->attr
.mmap_data
)
6646 atomic_inc(&nr_mmap_events
);
6647 if (event
->attr
.comm
)
6648 atomic_inc(&nr_comm_events
);
6649 if (event
->attr
.task
)
6650 atomic_inc(&nr_task_events
);
6651 if (event
->attr
.freq
) {
6652 if (atomic_inc_return(&nr_freq_events
) == 1)
6653 tick_nohz_full_kick_all();
6655 if (has_branch_stack(event
))
6656 static_key_slow_inc(&perf_sched_events
.key
);
6657 if (is_cgroup_event(event
))
6658 static_key_slow_inc(&perf_sched_events
.key
);
6660 account_event_cpu(event
, event
->cpu
);
6664 * Allocate and initialize a event structure
6666 static struct perf_event
*
6667 perf_event_alloc(struct perf_event_attr
*attr
, int cpu
,
6668 struct task_struct
*task
,
6669 struct perf_event
*group_leader
,
6670 struct perf_event
*parent_event
,
6671 perf_overflow_handler_t overflow_handler
,
6675 struct perf_event
*event
;
6676 struct hw_perf_event
*hwc
;
6679 if ((unsigned)cpu
>= nr_cpu_ids
) {
6680 if (!task
|| cpu
!= -1)
6681 return ERR_PTR(-EINVAL
);
6684 event
= kzalloc(sizeof(*event
), GFP_KERNEL
);
6686 return ERR_PTR(-ENOMEM
);
6689 * Single events are their own group leaders, with an
6690 * empty sibling list:
6693 group_leader
= event
;
6695 mutex_init(&event
->child_mutex
);
6696 INIT_LIST_HEAD(&event
->child_list
);
6698 INIT_LIST_HEAD(&event
->group_entry
);
6699 INIT_LIST_HEAD(&event
->event_entry
);
6700 INIT_LIST_HEAD(&event
->sibling_list
);
6701 INIT_LIST_HEAD(&event
->rb_entry
);
6702 INIT_LIST_HEAD(&event
->active_entry
);
6703 INIT_HLIST_NODE(&event
->hlist_entry
);
6706 init_waitqueue_head(&event
->waitq
);
6707 init_irq_work(&event
->pending
, perf_pending_event
);
6709 mutex_init(&event
->mmap_mutex
);
6711 atomic_long_set(&event
->refcount
, 1);
6713 event
->attr
= *attr
;
6714 event
->group_leader
= group_leader
;
6718 event
->parent
= parent_event
;
6720 event
->ns
= get_pid_ns(task_active_pid_ns(current
));
6721 event
->id
= atomic64_inc_return(&perf_event_id
);
6723 event
->state
= PERF_EVENT_STATE_INACTIVE
;
6726 event
->attach_state
= PERF_ATTACH_TASK
;
6728 if (attr
->type
== PERF_TYPE_TRACEPOINT
)
6729 event
->hw
.tp_target
= task
;
6730 #ifdef CONFIG_HAVE_HW_BREAKPOINT
6732 * hw_breakpoint is a bit difficult here..
6734 else if (attr
->type
== PERF_TYPE_BREAKPOINT
)
6735 event
->hw
.bp_target
= task
;
6739 if (!overflow_handler
&& parent_event
) {
6740 overflow_handler
= parent_event
->overflow_handler
;
6741 context
= parent_event
->overflow_handler_context
;
6744 event
->overflow_handler
= overflow_handler
;
6745 event
->overflow_handler_context
= context
;
6747 perf_event__state_init(event
);
6752 hwc
->sample_period
= attr
->sample_period
;
6753 if (attr
->freq
&& attr
->sample_freq
)
6754 hwc
->sample_period
= 1;
6755 hwc
->last_period
= hwc
->sample_period
;
6757 local64_set(&hwc
->period_left
, hwc
->sample_period
);
6760 * we currently do not support PERF_FORMAT_GROUP on inherited events
6762 if (attr
->inherit
&& (attr
->read_format
& PERF_FORMAT_GROUP
))
6765 pmu
= perf_init_event(event
);
6768 else if (IS_ERR(pmu
)) {
6773 if (!event
->parent
) {
6774 if (event
->attr
.sample_type
& PERF_SAMPLE_CALLCHAIN
) {
6775 err
= get_callchain_buffers();
6785 event
->destroy(event
);
6788 put_pid_ns(event
->ns
);
6791 return ERR_PTR(err
);
6794 static int perf_copy_attr(struct perf_event_attr __user
*uattr
,
6795 struct perf_event_attr
*attr
)
6800 if (!access_ok(VERIFY_WRITE
, uattr
, PERF_ATTR_SIZE_VER0
))
6804 * zero the full structure, so that a short copy will be nice.
6806 memset(attr
, 0, sizeof(*attr
));
6808 ret
= get_user(size
, &uattr
->size
);
6812 if (size
> PAGE_SIZE
) /* silly large */
6815 if (!size
) /* abi compat */
6816 size
= PERF_ATTR_SIZE_VER0
;
6818 if (size
< PERF_ATTR_SIZE_VER0
)
6822 * If we're handed a bigger struct than we know of,
6823 * ensure all the unknown bits are 0 - i.e. new
6824 * user-space does not rely on any kernel feature
6825 * extensions we dont know about yet.
6827 if (size
> sizeof(*attr
)) {
6828 unsigned char __user
*addr
;
6829 unsigned char __user
*end
;
6832 addr
= (void __user
*)uattr
+ sizeof(*attr
);
6833 end
= (void __user
*)uattr
+ size
;
6835 for (; addr
< end
; addr
++) {
6836 ret
= get_user(val
, addr
);
6842 size
= sizeof(*attr
);
6845 ret
= copy_from_user(attr
, uattr
, size
);
6849 /* disabled for now */
6853 if (attr
->__reserved_1
)
6856 if (attr
->sample_type
& ~(PERF_SAMPLE_MAX
-1))
6859 if (attr
->read_format
& ~(PERF_FORMAT_MAX
-1))
6862 if (attr
->sample_type
& PERF_SAMPLE_BRANCH_STACK
) {
6863 u64 mask
= attr
->branch_sample_type
;
6865 /* only using defined bits */
6866 if (mask
& ~(PERF_SAMPLE_BRANCH_MAX
-1))
6869 /* at least one branch bit must be set */
6870 if (!(mask
& ~PERF_SAMPLE_BRANCH_PLM_ALL
))
6873 /* propagate priv level, when not set for branch */
6874 if (!(mask
& PERF_SAMPLE_BRANCH_PLM_ALL
)) {
6876 /* exclude_kernel checked on syscall entry */
6877 if (!attr
->exclude_kernel
)
6878 mask
|= PERF_SAMPLE_BRANCH_KERNEL
;
6880 if (!attr
->exclude_user
)
6881 mask
|= PERF_SAMPLE_BRANCH_USER
;
6883 if (!attr
->exclude_hv
)
6884 mask
|= PERF_SAMPLE_BRANCH_HV
;
6886 * adjust user setting (for HW filter setup)
6888 attr
->branch_sample_type
= mask
;
6890 /* privileged levels capture (kernel, hv): check permissions */
6891 if ((mask
& PERF_SAMPLE_BRANCH_PERM_PLM
)
6892 && perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN
))
6896 if (attr
->sample_type
& PERF_SAMPLE_REGS_USER
) {
6897 ret
= perf_reg_validate(attr
->sample_regs_user
);
6902 if (attr
->sample_type
& PERF_SAMPLE_STACK_USER
) {
6903 if (!arch_perf_have_user_stack_dump())
6907 * We have __u32 type for the size, but so far
6908 * we can only use __u16 as maximum due to the
6909 * __u16 sample size limit.
6911 if (attr
->sample_stack_user
>= USHRT_MAX
)
6913 else if (!IS_ALIGNED(attr
->sample_stack_user
, sizeof(u64
)))
6921 put_user(sizeof(*attr
), &uattr
->size
);
6927 perf_event_set_output(struct perf_event
*event
, struct perf_event
*output_event
)
6929 struct ring_buffer
*rb
= NULL
, *old_rb
= NULL
;
6935 /* don't allow circular references */
6936 if (event
== output_event
)
6940 * Don't allow cross-cpu buffers
6942 if (output_event
->cpu
!= event
->cpu
)
6946 * If its not a per-cpu rb, it must be the same task.
6948 if (output_event
->cpu
== -1 && output_event
->ctx
!= event
->ctx
)
6952 mutex_lock(&event
->mmap_mutex
);
6953 /* Can't redirect output if we've got an active mmap() */
6954 if (atomic_read(&event
->mmap_count
))
6960 /* get the rb we want to redirect to */
6961 rb
= ring_buffer_get(output_event
);
6967 ring_buffer_detach(event
, old_rb
);
6970 ring_buffer_attach(event
, rb
);
6972 rcu_assign_pointer(event
->rb
, rb
);
6975 ring_buffer_put(old_rb
);
6977 * Since we detached before setting the new rb, so that we
6978 * could attach the new rb, we could have missed a wakeup.
6981 wake_up_all(&event
->waitq
);
6986 mutex_unlock(&event
->mmap_mutex
);
6993 * sys_perf_event_open - open a performance event, associate it to a task/cpu
6995 * @attr_uptr: event_id type attributes for monitoring/sampling
6998 * @group_fd: group leader event fd
7000 SYSCALL_DEFINE5(perf_event_open
,
7001 struct perf_event_attr __user
*, attr_uptr
,
7002 pid_t
, pid
, int, cpu
, int, group_fd
, unsigned long, flags
)
7004 struct perf_event
*group_leader
= NULL
, *output_event
= NULL
;
7005 struct perf_event
*event
, *sibling
;
7006 struct perf_event_attr attr
;
7007 struct perf_event_context
*ctx
;
7008 struct file
*event_file
= NULL
;
7009 struct fd group
= {NULL
, 0};
7010 struct task_struct
*task
= NULL
;
7015 int f_flags
= O_RDWR
;
7017 /* for future expandability... */
7018 if (flags
& ~PERF_FLAG_ALL
)
7021 err
= perf_copy_attr(attr_uptr
, &attr
);
7025 if (!attr
.exclude_kernel
) {
7026 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN
))
7031 if (attr
.sample_freq
> sysctl_perf_event_sample_rate
)
7036 * In cgroup mode, the pid argument is used to pass the fd
7037 * opened to the cgroup directory in cgroupfs. The cpu argument
7038 * designates the cpu on which to monitor threads from that
7041 if ((flags
& PERF_FLAG_PID_CGROUP
) && (pid
== -1 || cpu
== -1))
7044 if (flags
& PERF_FLAG_FD_CLOEXEC
)
7045 f_flags
|= O_CLOEXEC
;
7047 event_fd
= get_unused_fd_flags(f_flags
);
7051 if (group_fd
!= -1) {
7052 err
= perf_fget_light(group_fd
, &group
);
7055 group_leader
= group
.file
->private_data
;
7056 if (flags
& PERF_FLAG_FD_OUTPUT
)
7057 output_event
= group_leader
;
7058 if (flags
& PERF_FLAG_FD_NO_GROUP
)
7059 group_leader
= NULL
;
7062 if (pid
!= -1 && !(flags
& PERF_FLAG_PID_CGROUP
)) {
7063 task
= find_lively_task_by_vpid(pid
);
7065 err
= PTR_ERR(task
);
7072 event
= perf_event_alloc(&attr
, cpu
, task
, group_leader
, NULL
,
7074 if (IS_ERR(event
)) {
7075 err
= PTR_ERR(event
);
7079 if (flags
& PERF_FLAG_PID_CGROUP
) {
7080 err
= perf_cgroup_connect(pid
, event
, &attr
, group_leader
);
7082 __free_event(event
);
7087 account_event(event
);
7090 * Special case software events and allow them to be part of
7091 * any hardware group.
7096 (is_software_event(event
) != is_software_event(group_leader
))) {
7097 if (is_software_event(event
)) {
7099 * If event and group_leader are not both a software
7100 * event, and event is, then group leader is not.
7102 * Allow the addition of software events to !software
7103 * groups, this is safe because software events never
7106 pmu
= group_leader
->pmu
;
7107 } else if (is_software_event(group_leader
) &&
7108 (group_leader
->group_flags
& PERF_GROUP_SOFTWARE
)) {
7110 * In case the group is a pure software group, and we
7111 * try to add a hardware event, move the whole group to
7112 * the hardware context.
7119 * Get the target context (task or percpu):
7121 ctx
= find_get_context(pmu
, task
, event
->cpu
);
7128 put_task_struct(task
);
7133 * Look up the group leader (we will attach this event to it):
7139 * Do not allow a recursive hierarchy (this new sibling
7140 * becoming part of another group-sibling):
7142 if (group_leader
->group_leader
!= group_leader
)
7145 * Do not allow to attach to a group in a different
7146 * task or CPU context:
7149 if (group_leader
->ctx
->type
!= ctx
->type
)
7152 if (group_leader
->ctx
!= ctx
)
7157 * Only a group leader can be exclusive or pinned
7159 if (attr
.exclusive
|| attr
.pinned
)
7164 err
= perf_event_set_output(event
, output_event
);
7169 event_file
= anon_inode_getfile("[perf_event]", &perf_fops
, event
,
7171 if (IS_ERR(event_file
)) {
7172 err
= PTR_ERR(event_file
);
7177 struct perf_event_context
*gctx
= group_leader
->ctx
;
7179 mutex_lock(&gctx
->mutex
);
7180 perf_remove_from_context(group_leader
);
7183 * Removing from the context ends up with disabled
7184 * event. What we want here is event in the initial
7185 * startup state, ready to be add into new context.
7187 perf_event__state_init(group_leader
);
7188 list_for_each_entry(sibling
, &group_leader
->sibling_list
,
7190 perf_remove_from_context(sibling
);
7191 perf_event__state_init(sibling
);
7194 mutex_unlock(&gctx
->mutex
);
7198 WARN_ON_ONCE(ctx
->parent_ctx
);
7199 mutex_lock(&ctx
->mutex
);
7203 perf_install_in_context(ctx
, group_leader
, event
->cpu
);
7205 list_for_each_entry(sibling
, &group_leader
->sibling_list
,
7207 perf_install_in_context(ctx
, sibling
, event
->cpu
);
7212 perf_install_in_context(ctx
, event
, event
->cpu
);
7213 perf_unpin_context(ctx
);
7214 mutex_unlock(&ctx
->mutex
);
7218 event
->owner
= current
;
7220 mutex_lock(¤t
->perf_event_mutex
);
7221 list_add_tail(&event
->owner_entry
, ¤t
->perf_event_list
);
7222 mutex_unlock(¤t
->perf_event_mutex
);
7225 * Precalculate sample_data sizes
7227 perf_event__header_size(event
);
7228 perf_event__id_header_size(event
);
7231 * Drop the reference on the group_event after placing the
7232 * new event on the sibling_list. This ensures destruction
7233 * of the group leader will find the pointer to itself in
7234 * perf_group_detach().
7237 fd_install(event_fd
, event_file
);
7241 perf_unpin_context(ctx
);
7248 put_task_struct(task
);
7252 put_unused_fd(event_fd
);
7257 * perf_event_create_kernel_counter
7259 * @attr: attributes of the counter to create
7260 * @cpu: cpu in which the counter is bound
7261 * @task: task to profile (NULL for percpu)
7264 perf_event_create_kernel_counter(struct perf_event_attr
*attr
, int cpu
,
7265 struct task_struct
*task
,
7266 perf_overflow_handler_t overflow_handler
,
7269 struct perf_event_context
*ctx
;
7270 struct perf_event
*event
;
7274 * Get the target context (task or percpu):
7277 event
= perf_event_alloc(attr
, cpu
, task
, NULL
, NULL
,
7278 overflow_handler
, context
);
7279 if (IS_ERR(event
)) {
7280 err
= PTR_ERR(event
);
7284 account_event(event
);
7286 ctx
= find_get_context(event
->pmu
, task
, cpu
);
7292 WARN_ON_ONCE(ctx
->parent_ctx
);
7293 mutex_lock(&ctx
->mutex
);
7294 perf_install_in_context(ctx
, event
, cpu
);
7295 perf_unpin_context(ctx
);
7296 mutex_unlock(&ctx
->mutex
);
7303 return ERR_PTR(err
);
7305 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter
);
7307 void perf_pmu_migrate_context(struct pmu
*pmu
, int src_cpu
, int dst_cpu
)
7309 struct perf_event_context
*src_ctx
;
7310 struct perf_event_context
*dst_ctx
;
7311 struct perf_event
*event
, *tmp
;
7314 src_ctx
= &per_cpu_ptr(pmu
->pmu_cpu_context
, src_cpu
)->ctx
;
7315 dst_ctx
= &per_cpu_ptr(pmu
->pmu_cpu_context
, dst_cpu
)->ctx
;
7317 mutex_lock(&src_ctx
->mutex
);
7318 list_for_each_entry_safe(event
, tmp
, &src_ctx
->event_list
,
7320 perf_remove_from_context(event
);
7321 unaccount_event_cpu(event
, src_cpu
);
7323 list_add(&event
->migrate_entry
, &events
);
7325 mutex_unlock(&src_ctx
->mutex
);
7329 mutex_lock(&dst_ctx
->mutex
);
7330 list_for_each_entry_safe(event
, tmp
, &events
, migrate_entry
) {
7331 list_del(&event
->migrate_entry
);
7332 if (event
->state
>= PERF_EVENT_STATE_OFF
)
7333 event
->state
= PERF_EVENT_STATE_INACTIVE
;
7334 account_event_cpu(event
, dst_cpu
);
7335 perf_install_in_context(dst_ctx
, event
, dst_cpu
);
7338 mutex_unlock(&dst_ctx
->mutex
);
7340 EXPORT_SYMBOL_GPL(perf_pmu_migrate_context
);
7342 static void sync_child_event(struct perf_event
*child_event
,
7343 struct task_struct
*child
)
7345 struct perf_event
*parent_event
= child_event
->parent
;
7348 if (child_event
->attr
.inherit_stat
)
7349 perf_event_read_event(child_event
, child
);
7351 child_val
= perf_event_count(child_event
);
7354 * Add back the child's count to the parent's count:
7356 atomic64_add(child_val
, &parent_event
->child_count
);
7357 atomic64_add(child_event
->total_time_enabled
,
7358 &parent_event
->child_total_time_enabled
);
7359 atomic64_add(child_event
->total_time_running
,
7360 &parent_event
->child_total_time_running
);
7363 * Remove this event from the parent's list
7365 WARN_ON_ONCE(parent_event
->ctx
->parent_ctx
);
7366 mutex_lock(&parent_event
->child_mutex
);
7367 list_del_init(&child_event
->child_list
);
7368 mutex_unlock(&parent_event
->child_mutex
);
7371 * Release the parent event, if this was the last
7374 put_event(parent_event
);
7378 __perf_event_exit_task(struct perf_event
*child_event
,
7379 struct perf_event_context
*child_ctx
,
7380 struct task_struct
*child
)
7382 if (child_event
->parent
) {
7383 raw_spin_lock_irq(&child_ctx
->lock
);
7384 perf_group_detach(child_event
);
7385 raw_spin_unlock_irq(&child_ctx
->lock
);
7388 perf_remove_from_context(child_event
);
7391 * It can happen that the parent exits first, and has events
7392 * that are still around due to the child reference. These
7393 * events need to be zapped.
7395 if (child_event
->parent
) {
7396 sync_child_event(child_event
, child
);
7397 free_event(child_event
);
7401 static void perf_event_exit_task_context(struct task_struct
*child
, int ctxn
)
7403 struct perf_event
*child_event
, *tmp
;
7404 struct perf_event_context
*child_ctx
;
7405 unsigned long flags
;
7407 if (likely(!child
->perf_event_ctxp
[ctxn
])) {
7408 perf_event_task(child
, NULL
, 0);
7412 local_irq_save(flags
);
7414 * We can't reschedule here because interrupts are disabled,
7415 * and either child is current or it is a task that can't be
7416 * scheduled, so we are now safe from rescheduling changing
7419 child_ctx
= rcu_dereference_raw(child
->perf_event_ctxp
[ctxn
]);
7422 * Take the context lock here so that if find_get_context is
7423 * reading child->perf_event_ctxp, we wait until it has
7424 * incremented the context's refcount before we do put_ctx below.
7426 raw_spin_lock(&child_ctx
->lock
);
7427 task_ctx_sched_out(child_ctx
);
7428 child
->perf_event_ctxp
[ctxn
] = NULL
;
7430 * If this context is a clone; unclone it so it can't get
7431 * swapped to another process while we're removing all
7432 * the events from it.
7434 unclone_ctx(child_ctx
);
7435 update_context_time(child_ctx
);
7436 raw_spin_unlock_irqrestore(&child_ctx
->lock
, flags
);
7439 * Report the task dead after unscheduling the events so that we
7440 * won't get any samples after PERF_RECORD_EXIT. We can however still
7441 * get a few PERF_RECORD_READ events.
7443 perf_event_task(child
, child_ctx
, 0);
7446 * We can recurse on the same lock type through:
7448 * __perf_event_exit_task()
7449 * sync_child_event()
7451 * mutex_lock(&ctx->mutex)
7453 * But since its the parent context it won't be the same instance.
7455 mutex_lock(&child_ctx
->mutex
);
7458 list_for_each_entry_safe(child_event
, tmp
, &child_ctx
->pinned_groups
,
7460 __perf_event_exit_task(child_event
, child_ctx
, child
);
7462 list_for_each_entry_safe(child_event
, tmp
, &child_ctx
->flexible_groups
,
7464 __perf_event_exit_task(child_event
, child_ctx
, child
);
7467 * If the last event was a group event, it will have appended all
7468 * its siblings to the list, but we obtained 'tmp' before that which
7469 * will still point to the list head terminating the iteration.
7471 if (!list_empty(&child_ctx
->pinned_groups
) ||
7472 !list_empty(&child_ctx
->flexible_groups
))
7475 mutex_unlock(&child_ctx
->mutex
);
7481 * When a child task exits, feed back event values to parent events.
7483 void perf_event_exit_task(struct task_struct
*child
)
7485 struct perf_event
*event
, *tmp
;
7488 mutex_lock(&child
->perf_event_mutex
);
7489 list_for_each_entry_safe(event
, tmp
, &child
->perf_event_list
,
7491 list_del_init(&event
->owner_entry
);
7494 * Ensure the list deletion is visible before we clear
7495 * the owner, closes a race against perf_release() where
7496 * we need to serialize on the owner->perf_event_mutex.
7499 event
->owner
= NULL
;
7501 mutex_unlock(&child
->perf_event_mutex
);
7503 for_each_task_context_nr(ctxn
)
7504 perf_event_exit_task_context(child
, ctxn
);
7507 static void perf_free_event(struct perf_event
*event
,
7508 struct perf_event_context
*ctx
)
7510 struct perf_event
*parent
= event
->parent
;
7512 if (WARN_ON_ONCE(!parent
))
7515 mutex_lock(&parent
->child_mutex
);
7516 list_del_init(&event
->child_list
);
7517 mutex_unlock(&parent
->child_mutex
);
7521 perf_group_detach(event
);
7522 list_del_event(event
, ctx
);
7527 * free an unexposed, unused context as created by inheritance by
7528 * perf_event_init_task below, used by fork() in case of fail.
7530 void perf_event_free_task(struct task_struct
*task
)
7532 struct perf_event_context
*ctx
;
7533 struct perf_event
*event
, *tmp
;
7536 for_each_task_context_nr(ctxn
) {
7537 ctx
= task
->perf_event_ctxp
[ctxn
];
7541 mutex_lock(&ctx
->mutex
);
7543 list_for_each_entry_safe(event
, tmp
, &ctx
->pinned_groups
,
7545 perf_free_event(event
, ctx
);
7547 list_for_each_entry_safe(event
, tmp
, &ctx
->flexible_groups
,
7549 perf_free_event(event
, ctx
);
7551 if (!list_empty(&ctx
->pinned_groups
) ||
7552 !list_empty(&ctx
->flexible_groups
))
7555 mutex_unlock(&ctx
->mutex
);
7561 void perf_event_delayed_put(struct task_struct
*task
)
7565 for_each_task_context_nr(ctxn
)
7566 WARN_ON_ONCE(task
->perf_event_ctxp
[ctxn
]);
7570 * inherit a event from parent task to child task:
7572 static struct perf_event
*
7573 inherit_event(struct perf_event
*parent_event
,
7574 struct task_struct
*parent
,
7575 struct perf_event_context
*parent_ctx
,
7576 struct task_struct
*child
,
7577 struct perf_event
*group_leader
,
7578 struct perf_event_context
*child_ctx
)
7580 struct perf_event
*child_event
;
7581 unsigned long flags
;
7584 * Instead of creating recursive hierarchies of events,
7585 * we link inherited events back to the original parent,
7586 * which has a filp for sure, which we use as the reference
7589 if (parent_event
->parent
)
7590 parent_event
= parent_event
->parent
;
7592 child_event
= perf_event_alloc(&parent_event
->attr
,
7595 group_leader
, parent_event
,
7597 if (IS_ERR(child_event
))
7600 if (!atomic_long_inc_not_zero(&parent_event
->refcount
)) {
7601 free_event(child_event
);
7608 * Make the child state follow the state of the parent event,
7609 * not its attr.disabled bit. We hold the parent's mutex,
7610 * so we won't race with perf_event_{en, dis}able_family.
7612 if (parent_event
->state
>= PERF_EVENT_STATE_INACTIVE
)
7613 child_event
->state
= PERF_EVENT_STATE_INACTIVE
;
7615 child_event
->state
= PERF_EVENT_STATE_OFF
;
7617 if (parent_event
->attr
.freq
) {
7618 u64 sample_period
= parent_event
->hw
.sample_period
;
7619 struct hw_perf_event
*hwc
= &child_event
->hw
;
7621 hwc
->sample_period
= sample_period
;
7622 hwc
->last_period
= sample_period
;
7624 local64_set(&hwc
->period_left
, sample_period
);
7627 child_event
->ctx
= child_ctx
;
7628 child_event
->overflow_handler
= parent_event
->overflow_handler
;
7629 child_event
->overflow_handler_context
7630 = parent_event
->overflow_handler_context
;
7633 * Precalculate sample_data sizes
7635 perf_event__header_size(child_event
);
7636 perf_event__id_header_size(child_event
);
7639 * Link it up in the child's context:
7641 raw_spin_lock_irqsave(&child_ctx
->lock
, flags
);
7642 add_event_to_ctx(child_event
, child_ctx
);
7643 raw_spin_unlock_irqrestore(&child_ctx
->lock
, flags
);
7646 * Link this into the parent event's child list
7648 WARN_ON_ONCE(parent_event
->ctx
->parent_ctx
);
7649 mutex_lock(&parent_event
->child_mutex
);
7650 list_add_tail(&child_event
->child_list
, &parent_event
->child_list
);
7651 mutex_unlock(&parent_event
->child_mutex
);
7656 static int inherit_group(struct perf_event
*parent_event
,
7657 struct task_struct
*parent
,
7658 struct perf_event_context
*parent_ctx
,
7659 struct task_struct
*child
,
7660 struct perf_event_context
*child_ctx
)
7662 struct perf_event
*leader
;
7663 struct perf_event
*sub
;
7664 struct perf_event
*child_ctr
;
7666 leader
= inherit_event(parent_event
, parent
, parent_ctx
,
7667 child
, NULL
, child_ctx
);
7669 return PTR_ERR(leader
);
7670 list_for_each_entry(sub
, &parent_event
->sibling_list
, group_entry
) {
7671 child_ctr
= inherit_event(sub
, parent
, parent_ctx
,
7672 child
, leader
, child_ctx
);
7673 if (IS_ERR(child_ctr
))
7674 return PTR_ERR(child_ctr
);
7680 inherit_task_group(struct perf_event
*event
, struct task_struct
*parent
,
7681 struct perf_event_context
*parent_ctx
,
7682 struct task_struct
*child
, int ctxn
,
7686 struct perf_event_context
*child_ctx
;
7688 if (!event
->attr
.inherit
) {
7693 child_ctx
= child
->perf_event_ctxp
[ctxn
];
7696 * This is executed from the parent task context, so
7697 * inherit events that have been marked for cloning.
7698 * First allocate and initialize a context for the
7702 child_ctx
= alloc_perf_context(parent_ctx
->pmu
, child
);
7706 child
->perf_event_ctxp
[ctxn
] = child_ctx
;
7709 ret
= inherit_group(event
, parent
, parent_ctx
,
7719 * Initialize the perf_event context in task_struct
7721 int perf_event_init_context(struct task_struct
*child
, int ctxn
)
7723 struct perf_event_context
*child_ctx
, *parent_ctx
;
7724 struct perf_event_context
*cloned_ctx
;
7725 struct perf_event
*event
;
7726 struct task_struct
*parent
= current
;
7727 int inherited_all
= 1;
7728 unsigned long flags
;
7731 if (likely(!parent
->perf_event_ctxp
[ctxn
]))
7735 * If the parent's context is a clone, pin it so it won't get
7738 parent_ctx
= perf_pin_task_context(parent
, ctxn
);
7741 * No need to check if parent_ctx != NULL here; since we saw
7742 * it non-NULL earlier, the only reason for it to become NULL
7743 * is if we exit, and since we're currently in the middle of
7744 * a fork we can't be exiting at the same time.
7748 * Lock the parent list. No need to lock the child - not PID
7749 * hashed yet and not running, so nobody can access it.
7751 mutex_lock(&parent_ctx
->mutex
);
7754 * We dont have to disable NMIs - we are only looking at
7755 * the list, not manipulating it:
7757 list_for_each_entry(event
, &parent_ctx
->pinned_groups
, group_entry
) {
7758 ret
= inherit_task_group(event
, parent
, parent_ctx
,
7759 child
, ctxn
, &inherited_all
);
7765 * We can't hold ctx->lock when iterating the ->flexible_group list due
7766 * to allocations, but we need to prevent rotation because
7767 * rotate_ctx() will change the list from interrupt context.
7769 raw_spin_lock_irqsave(&parent_ctx
->lock
, flags
);
7770 parent_ctx
->rotate_disable
= 1;
7771 raw_spin_unlock_irqrestore(&parent_ctx
->lock
, flags
);
7773 list_for_each_entry(event
, &parent_ctx
->flexible_groups
, group_entry
) {
7774 ret
= inherit_task_group(event
, parent
, parent_ctx
,
7775 child
, ctxn
, &inherited_all
);
7780 raw_spin_lock_irqsave(&parent_ctx
->lock
, flags
);
7781 parent_ctx
->rotate_disable
= 0;
7783 child_ctx
= child
->perf_event_ctxp
[ctxn
];
7785 if (child_ctx
&& inherited_all
) {
7787 * Mark the child context as a clone of the parent
7788 * context, or of whatever the parent is a clone of.
7790 * Note that if the parent is a clone, the holding of
7791 * parent_ctx->lock avoids it from being uncloned.
7793 cloned_ctx
= parent_ctx
->parent_ctx
;
7795 child_ctx
->parent_ctx
= cloned_ctx
;
7796 child_ctx
->parent_gen
= parent_ctx
->parent_gen
;
7798 child_ctx
->parent_ctx
= parent_ctx
;
7799 child_ctx
->parent_gen
= parent_ctx
->generation
;
7801 get_ctx(child_ctx
->parent_ctx
);
7804 raw_spin_unlock_irqrestore(&parent_ctx
->lock
, flags
);
7805 mutex_unlock(&parent_ctx
->mutex
);
7807 perf_unpin_context(parent_ctx
);
7808 put_ctx(parent_ctx
);
7814 * Initialize the perf_event context in task_struct
7816 int perf_event_init_task(struct task_struct
*child
)
7820 memset(child
->perf_event_ctxp
, 0, sizeof(child
->perf_event_ctxp
));
7821 mutex_init(&child
->perf_event_mutex
);
7822 INIT_LIST_HEAD(&child
->perf_event_list
);
7824 for_each_task_context_nr(ctxn
) {
7825 ret
= perf_event_init_context(child
, ctxn
);
7833 static void __init
perf_event_init_all_cpus(void)
7835 struct swevent_htable
*swhash
;
7838 for_each_possible_cpu(cpu
) {
7839 swhash
= &per_cpu(swevent_htable
, cpu
);
7840 mutex_init(&swhash
->hlist_mutex
);
7841 INIT_LIST_HEAD(&per_cpu(rotation_list
, cpu
));
7845 static void perf_event_init_cpu(int cpu
)
7847 struct swevent_htable
*swhash
= &per_cpu(swevent_htable
, cpu
);
7849 mutex_lock(&swhash
->hlist_mutex
);
7850 if (swhash
->hlist_refcount
> 0) {
7851 struct swevent_hlist
*hlist
;
7853 hlist
= kzalloc_node(sizeof(*hlist
), GFP_KERNEL
, cpu_to_node(cpu
));
7855 rcu_assign_pointer(swhash
->swevent_hlist
, hlist
);
7857 mutex_unlock(&swhash
->hlist_mutex
);
7860 #if defined CONFIG_HOTPLUG_CPU || defined CONFIG_KEXEC
7861 static void perf_pmu_rotate_stop(struct pmu
*pmu
)
7863 struct perf_cpu_context
*cpuctx
= this_cpu_ptr(pmu
->pmu_cpu_context
);
7865 WARN_ON(!irqs_disabled());
7867 list_del_init(&cpuctx
->rotation_list
);
7870 static void __perf_event_exit_context(void *__info
)
7872 struct perf_event_context
*ctx
= __info
;
7873 struct perf_event
*event
, *tmp
;
7875 perf_pmu_rotate_stop(ctx
->pmu
);
7877 list_for_each_entry_safe(event
, tmp
, &ctx
->pinned_groups
, group_entry
)
7878 __perf_remove_from_context(event
);
7879 list_for_each_entry_safe(event
, tmp
, &ctx
->flexible_groups
, group_entry
)
7880 __perf_remove_from_context(event
);
7883 static void perf_event_exit_cpu_context(int cpu
)
7885 struct perf_event_context
*ctx
;
7889 idx
= srcu_read_lock(&pmus_srcu
);
7890 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
7891 ctx
= &per_cpu_ptr(pmu
->pmu_cpu_context
, cpu
)->ctx
;
7893 mutex_lock(&ctx
->mutex
);
7894 smp_call_function_single(cpu
, __perf_event_exit_context
, ctx
, 1);
7895 mutex_unlock(&ctx
->mutex
);
7897 srcu_read_unlock(&pmus_srcu
, idx
);
7900 static void perf_event_exit_cpu(int cpu
)
7902 struct swevent_htable
*swhash
= &per_cpu(swevent_htable
, cpu
);
7904 mutex_lock(&swhash
->hlist_mutex
);
7905 swevent_hlist_release(swhash
);
7906 mutex_unlock(&swhash
->hlist_mutex
);
7908 perf_event_exit_cpu_context(cpu
);
7911 static inline void perf_event_exit_cpu(int cpu
) { }
7915 perf_reboot(struct notifier_block
*notifier
, unsigned long val
, void *v
)
7919 for_each_online_cpu(cpu
)
7920 perf_event_exit_cpu(cpu
);
7926 * Run the perf reboot notifier at the very last possible moment so that
7927 * the generic watchdog code runs as long as possible.
7929 static struct notifier_block perf_reboot_notifier
= {
7930 .notifier_call
= perf_reboot
,
7931 .priority
= INT_MIN
,
7935 perf_cpu_notify(struct notifier_block
*self
, unsigned long action
, void *hcpu
)
7937 unsigned int cpu
= (long)hcpu
;
7939 switch (action
& ~CPU_TASKS_FROZEN
) {
7941 case CPU_UP_PREPARE
:
7942 case CPU_DOWN_FAILED
:
7943 perf_event_init_cpu(cpu
);
7946 case CPU_UP_CANCELED
:
7947 case CPU_DOWN_PREPARE
:
7948 perf_event_exit_cpu(cpu
);
7957 void __init
perf_event_init(void)
7963 perf_event_init_all_cpus();
7964 init_srcu_struct(&pmus_srcu
);
7965 perf_pmu_register(&perf_swevent
, "software", PERF_TYPE_SOFTWARE
);
7966 perf_pmu_register(&perf_cpu_clock
, NULL
, -1);
7967 perf_pmu_register(&perf_task_clock
, NULL
, -1);
7969 perf_cpu_notifier(perf_cpu_notify
);
7970 register_reboot_notifier(&perf_reboot_notifier
);
7972 ret
= init_hw_breakpoint();
7973 WARN(ret
, "hw_breakpoint initialization failed with: %d", ret
);
7975 /* do not patch jump label more than once per second */
7976 jump_label_rate_limit(&perf_sched_events
, HZ
);
7979 * Build time assertion that we keep the data_head at the intended
7980 * location. IOW, validation we got the __reserved[] size right.
7982 BUILD_BUG_ON((offsetof(struct perf_event_mmap_page
, data_head
))
7986 static int __init
perf_event_sysfs_init(void)
7991 mutex_lock(&pmus_lock
);
7993 ret
= bus_register(&pmu_bus
);
7997 list_for_each_entry(pmu
, &pmus
, entry
) {
7998 if (!pmu
->name
|| pmu
->type
< 0)
8001 ret
= pmu_dev_alloc(pmu
);
8002 WARN(ret
, "Failed to register pmu: %s, reason %d\n", pmu
->name
, ret
);
8004 pmu_bus_running
= 1;
8008 mutex_unlock(&pmus_lock
);
8012 device_initcall(perf_event_sysfs_init
);
8014 #ifdef CONFIG_CGROUP_PERF
8015 static struct cgroup_subsys_state
*
8016 perf_cgroup_css_alloc(struct cgroup_subsys_state
*parent_css
)
8018 struct perf_cgroup
*jc
;
8020 jc
= kzalloc(sizeof(*jc
), GFP_KERNEL
);
8022 return ERR_PTR(-ENOMEM
);
8024 jc
->info
= alloc_percpu(struct perf_cgroup_info
);
8027 return ERR_PTR(-ENOMEM
);
8033 static void perf_cgroup_css_free(struct cgroup_subsys_state
*css
)
8035 struct perf_cgroup
*jc
= container_of(css
, struct perf_cgroup
, css
);
8037 free_percpu(jc
->info
);
8041 static int __perf_cgroup_move(void *info
)
8043 struct task_struct
*task
= info
;
8044 perf_cgroup_switch(task
, PERF_CGROUP_SWOUT
| PERF_CGROUP_SWIN
);
8048 static void perf_cgroup_attach(struct cgroup_subsys_state
*css
,
8049 struct cgroup_taskset
*tset
)
8051 struct task_struct
*task
;
8053 cgroup_taskset_for_each(task
, css
, tset
)
8054 task_function_call(task
, __perf_cgroup_move
, task
);
8057 static void perf_cgroup_exit(struct cgroup_subsys_state
*css
,
8058 struct cgroup_subsys_state
*old_css
,
8059 struct task_struct
*task
)
8062 * cgroup_exit() is called in the copy_process() failure path.
8063 * Ignore this case since the task hasn't ran yet, this avoids
8064 * trying to poke a half freed task state from generic code.
8066 if (!(task
->flags
& PF_EXITING
))
8069 task_function_call(task
, __perf_cgroup_move
, task
);
8072 struct cgroup_subsys perf_subsys
= {
8073 .name
= "perf_event",
8074 .subsys_id
= perf_subsys_id
,
8075 .css_alloc
= perf_cgroup_css_alloc
,
8076 .css_free
= perf_cgroup_css_free
,
8077 .exit
= perf_cgroup_exit
,
8078 .attach
= perf_cgroup_attach
,
8080 #endif /* CONFIG_CGROUP_PERF */