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/sysfs.h>
22 #include <linux/dcache.h>
23 #include <linux/percpu.h>
24 #include <linux/ptrace.h>
25 #include <linux/reboot.h>
26 #include <linux/vmstat.h>
27 #include <linux/device.h>
28 #include <linux/export.h>
29 #include <linux/vmalloc.h>
30 #include <linux/hardirq.h>
31 #include <linux/rculist.h>
32 #include <linux/uaccess.h>
33 #include <linux/syscalls.h>
34 #include <linux/anon_inodes.h>
35 #include <linux/kernel_stat.h>
36 #include <linux/perf_event.h>
37 #include <linux/ftrace_event.h>
38 #include <linux/hw_breakpoint.h>
42 #include <asm/irq_regs.h>
44 struct remote_function_call
{
45 struct task_struct
*p
;
46 int (*func
)(void *info
);
51 static void remote_function(void *data
)
53 struct remote_function_call
*tfc
= data
;
54 struct task_struct
*p
= tfc
->p
;
58 if (task_cpu(p
) != smp_processor_id() || !task_curr(p
))
62 tfc
->ret
= tfc
->func(tfc
->info
);
66 * task_function_call - call a function on the cpu on which a task runs
67 * @p: the task to evaluate
68 * @func: the function to be called
69 * @info: the function call argument
71 * Calls the function @func when the task is currently running. This might
72 * be on the current CPU, which just calls the function directly
74 * returns: @func return value, or
75 * -ESRCH - when the process isn't running
76 * -EAGAIN - when the process moved away
79 task_function_call(struct task_struct
*p
, int (*func
) (void *info
), void *info
)
81 struct remote_function_call data
= {
85 .ret
= -ESRCH
, /* No such (running) process */
89 smp_call_function_single(task_cpu(p
), remote_function
, &data
, 1);
95 * cpu_function_call - call a function on the cpu
96 * @func: the function to be called
97 * @info: the function call argument
99 * Calls the function @func on the remote cpu.
101 * returns: @func return value or -ENXIO when the cpu is offline
103 static int cpu_function_call(int cpu
, int (*func
) (void *info
), void *info
)
105 struct remote_function_call data
= {
109 .ret
= -ENXIO
, /* No such CPU */
112 smp_call_function_single(cpu
, remote_function
, &data
, 1);
117 #define PERF_FLAG_ALL (PERF_FLAG_FD_NO_GROUP |\
118 PERF_FLAG_FD_OUTPUT |\
119 PERF_FLAG_PID_CGROUP)
122 * branch priv levels that need permission checks
124 #define PERF_SAMPLE_BRANCH_PERM_PLM \
125 (PERF_SAMPLE_BRANCH_KERNEL |\
126 PERF_SAMPLE_BRANCH_HV)
129 EVENT_FLEXIBLE
= 0x1,
131 EVENT_ALL
= EVENT_FLEXIBLE
| EVENT_PINNED
,
135 * perf_sched_events : >0 events exist
136 * perf_cgroup_events: >0 per-cpu cgroup events exist on this cpu
138 struct static_key_deferred perf_sched_events __read_mostly
;
139 static DEFINE_PER_CPU(atomic_t
, perf_cgroup_events
);
140 static DEFINE_PER_CPU(atomic_t
, perf_branch_stack_events
);
142 static atomic_t nr_mmap_events __read_mostly
;
143 static atomic_t nr_comm_events __read_mostly
;
144 static atomic_t nr_task_events __read_mostly
;
146 static LIST_HEAD(pmus
);
147 static DEFINE_MUTEX(pmus_lock
);
148 static struct srcu_struct pmus_srcu
;
151 * perf event paranoia level:
152 * -1 - not paranoid at all
153 * 0 - disallow raw tracepoint access for unpriv
154 * 1 - disallow cpu events for unpriv
155 * 2 - disallow kernel profiling for unpriv
157 int sysctl_perf_event_paranoid __read_mostly
= 1;
159 /* Minimum for 512 kiB + 1 user control page */
160 int sysctl_perf_event_mlock __read_mostly
= 512 + (PAGE_SIZE
/ 1024); /* 'free' kiB per user */
163 * max perf event sample rate
165 #define DEFAULT_MAX_SAMPLE_RATE 100000
166 int sysctl_perf_event_sample_rate __read_mostly
= DEFAULT_MAX_SAMPLE_RATE
;
167 static int max_samples_per_tick __read_mostly
=
168 DIV_ROUND_UP(DEFAULT_MAX_SAMPLE_RATE
, HZ
);
170 int perf_proc_update_handler(struct ctl_table
*table
, int write
,
171 void __user
*buffer
, size_t *lenp
,
174 int ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
179 max_samples_per_tick
= DIV_ROUND_UP(sysctl_perf_event_sample_rate
, HZ
);
184 static atomic64_t perf_event_id
;
186 static void cpu_ctx_sched_out(struct perf_cpu_context
*cpuctx
,
187 enum event_type_t event_type
);
189 static void cpu_ctx_sched_in(struct perf_cpu_context
*cpuctx
,
190 enum event_type_t event_type
,
191 struct task_struct
*task
);
193 static void update_context_time(struct perf_event_context
*ctx
);
194 static u64
perf_event_time(struct perf_event
*event
);
196 static void ring_buffer_attach(struct perf_event
*event
,
197 struct ring_buffer
*rb
);
199 void __weak
perf_event_print_debug(void) { }
201 extern __weak
const char *perf_pmu_name(void)
206 static inline u64
perf_clock(void)
208 return local_clock();
211 static inline struct perf_cpu_context
*
212 __get_cpu_context(struct perf_event_context
*ctx
)
214 return this_cpu_ptr(ctx
->pmu
->pmu_cpu_context
);
217 static void perf_ctx_lock(struct perf_cpu_context
*cpuctx
,
218 struct perf_event_context
*ctx
)
220 raw_spin_lock(&cpuctx
->ctx
.lock
);
222 raw_spin_lock(&ctx
->lock
);
225 static void perf_ctx_unlock(struct perf_cpu_context
*cpuctx
,
226 struct perf_event_context
*ctx
)
229 raw_spin_unlock(&ctx
->lock
);
230 raw_spin_unlock(&cpuctx
->ctx
.lock
);
233 #ifdef CONFIG_CGROUP_PERF
236 * Must ensure cgroup is pinned (css_get) before calling
237 * this function. In other words, we cannot call this function
238 * if there is no cgroup event for the current CPU context.
240 static inline struct perf_cgroup
*
241 perf_cgroup_from_task(struct task_struct
*task
)
243 return container_of(task_subsys_state(task
, perf_subsys_id
),
244 struct perf_cgroup
, css
);
248 perf_cgroup_match(struct perf_event
*event
)
250 struct perf_event_context
*ctx
= event
->ctx
;
251 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
253 return !event
->cgrp
|| event
->cgrp
== cpuctx
->cgrp
;
256 static inline void perf_get_cgroup(struct perf_event
*event
)
258 css_get(&event
->cgrp
->css
);
261 static inline void perf_put_cgroup(struct perf_event
*event
)
263 css_put(&event
->cgrp
->css
);
266 static inline void perf_detach_cgroup(struct perf_event
*event
)
268 perf_put_cgroup(event
);
272 static inline int is_cgroup_event(struct perf_event
*event
)
274 return event
->cgrp
!= NULL
;
277 static inline u64
perf_cgroup_event_time(struct perf_event
*event
)
279 struct perf_cgroup_info
*t
;
281 t
= per_cpu_ptr(event
->cgrp
->info
, event
->cpu
);
285 static inline void __update_cgrp_time(struct perf_cgroup
*cgrp
)
287 struct perf_cgroup_info
*info
;
292 info
= this_cpu_ptr(cgrp
->info
);
294 info
->time
+= now
- info
->timestamp
;
295 info
->timestamp
= now
;
298 static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context
*cpuctx
)
300 struct perf_cgroup
*cgrp_out
= cpuctx
->cgrp
;
302 __update_cgrp_time(cgrp_out
);
305 static inline void update_cgrp_time_from_event(struct perf_event
*event
)
307 struct perf_cgroup
*cgrp
;
310 * ensure we access cgroup data only when needed and
311 * when we know the cgroup is pinned (css_get)
313 if (!is_cgroup_event(event
))
316 cgrp
= perf_cgroup_from_task(current
);
318 * Do not update time when cgroup is not active
320 if (cgrp
== event
->cgrp
)
321 __update_cgrp_time(event
->cgrp
);
325 perf_cgroup_set_timestamp(struct task_struct
*task
,
326 struct perf_event_context
*ctx
)
328 struct perf_cgroup
*cgrp
;
329 struct perf_cgroup_info
*info
;
332 * ctx->lock held by caller
333 * ensure we do not access cgroup data
334 * unless we have the cgroup pinned (css_get)
336 if (!task
|| !ctx
->nr_cgroups
)
339 cgrp
= perf_cgroup_from_task(task
);
340 info
= this_cpu_ptr(cgrp
->info
);
341 info
->timestamp
= ctx
->timestamp
;
344 #define PERF_CGROUP_SWOUT 0x1 /* cgroup switch out every event */
345 #define PERF_CGROUP_SWIN 0x2 /* cgroup switch in events based on task */
348 * reschedule events based on the cgroup constraint of task.
350 * mode SWOUT : schedule out everything
351 * mode SWIN : schedule in based on cgroup for next
353 void perf_cgroup_switch(struct task_struct
*task
, int mode
)
355 struct perf_cpu_context
*cpuctx
;
360 * disable interrupts to avoid geting nr_cgroup
361 * changes via __perf_event_disable(). Also
364 local_irq_save(flags
);
367 * we reschedule only in the presence of cgroup
368 * constrained events.
372 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
373 cpuctx
= this_cpu_ptr(pmu
->pmu_cpu_context
);
376 * perf_cgroup_events says at least one
377 * context on this CPU has cgroup events.
379 * ctx->nr_cgroups reports the number of cgroup
380 * events for a context.
382 if (cpuctx
->ctx
.nr_cgroups
> 0) {
383 perf_ctx_lock(cpuctx
, cpuctx
->task_ctx
);
384 perf_pmu_disable(cpuctx
->ctx
.pmu
);
386 if (mode
& PERF_CGROUP_SWOUT
) {
387 cpu_ctx_sched_out(cpuctx
, EVENT_ALL
);
389 * must not be done before ctxswout due
390 * to event_filter_match() in event_sched_out()
395 if (mode
& PERF_CGROUP_SWIN
) {
396 WARN_ON_ONCE(cpuctx
->cgrp
);
397 /* set cgrp before ctxsw in to
398 * allow event_filter_match() to not
399 * have to pass task around
401 cpuctx
->cgrp
= perf_cgroup_from_task(task
);
402 cpu_ctx_sched_in(cpuctx
, EVENT_ALL
, task
);
404 perf_pmu_enable(cpuctx
->ctx
.pmu
);
405 perf_ctx_unlock(cpuctx
, cpuctx
->task_ctx
);
411 local_irq_restore(flags
);
414 static inline void perf_cgroup_sched_out(struct task_struct
*task
,
415 struct task_struct
*next
)
417 struct perf_cgroup
*cgrp1
;
418 struct perf_cgroup
*cgrp2
= NULL
;
421 * we come here when we know perf_cgroup_events > 0
423 cgrp1
= perf_cgroup_from_task(task
);
426 * next is NULL when called from perf_event_enable_on_exec()
427 * that will systematically cause a cgroup_switch()
430 cgrp2
= perf_cgroup_from_task(next
);
433 * only schedule out current cgroup events if we know
434 * that we are switching to a different cgroup. Otherwise,
435 * do no touch the cgroup events.
438 perf_cgroup_switch(task
, PERF_CGROUP_SWOUT
);
441 static inline void perf_cgroup_sched_in(struct task_struct
*prev
,
442 struct task_struct
*task
)
444 struct perf_cgroup
*cgrp1
;
445 struct perf_cgroup
*cgrp2
= NULL
;
448 * we come here when we know perf_cgroup_events > 0
450 cgrp1
= perf_cgroup_from_task(task
);
452 /* prev can never be NULL */
453 cgrp2
= perf_cgroup_from_task(prev
);
456 * only need to schedule in cgroup events if we are changing
457 * cgroup during ctxsw. Cgroup events were not scheduled
458 * out of ctxsw out if that was not the case.
461 perf_cgroup_switch(task
, PERF_CGROUP_SWIN
);
464 static inline int perf_cgroup_connect(int fd
, struct perf_event
*event
,
465 struct perf_event_attr
*attr
,
466 struct perf_event
*group_leader
)
468 struct perf_cgroup
*cgrp
;
469 struct cgroup_subsys_state
*css
;
471 int ret
= 0, fput_needed
;
473 file
= fget_light(fd
, &fput_needed
);
477 css
= cgroup_css_from_dir(file
, perf_subsys_id
);
483 cgrp
= container_of(css
, struct perf_cgroup
, css
);
486 /* must be done before we fput() the file */
487 perf_get_cgroup(event
);
490 * all events in a group must monitor
491 * the same cgroup because a task belongs
492 * to only one perf cgroup at a time
494 if (group_leader
&& group_leader
->cgrp
!= cgrp
) {
495 perf_detach_cgroup(event
);
499 fput_light(file
, fput_needed
);
504 perf_cgroup_set_shadow_time(struct perf_event
*event
, u64 now
)
506 struct perf_cgroup_info
*t
;
507 t
= per_cpu_ptr(event
->cgrp
->info
, event
->cpu
);
508 event
->shadow_ctx_time
= now
- t
->timestamp
;
512 perf_cgroup_defer_enabled(struct perf_event
*event
)
515 * when the current task's perf cgroup does not match
516 * the event's, we need to remember to call the
517 * perf_mark_enable() function the first time a task with
518 * a matching perf cgroup is scheduled in.
520 if (is_cgroup_event(event
) && !perf_cgroup_match(event
))
521 event
->cgrp_defer_enabled
= 1;
525 perf_cgroup_mark_enabled(struct perf_event
*event
,
526 struct perf_event_context
*ctx
)
528 struct perf_event
*sub
;
529 u64 tstamp
= perf_event_time(event
);
531 if (!event
->cgrp_defer_enabled
)
534 event
->cgrp_defer_enabled
= 0;
536 event
->tstamp_enabled
= tstamp
- event
->total_time_enabled
;
537 list_for_each_entry(sub
, &event
->sibling_list
, group_entry
) {
538 if (sub
->state
>= PERF_EVENT_STATE_INACTIVE
) {
539 sub
->tstamp_enabled
= tstamp
- sub
->total_time_enabled
;
540 sub
->cgrp_defer_enabled
= 0;
544 #else /* !CONFIG_CGROUP_PERF */
547 perf_cgroup_match(struct perf_event
*event
)
552 static inline void perf_detach_cgroup(struct perf_event
*event
)
555 static inline int is_cgroup_event(struct perf_event
*event
)
560 static inline u64
perf_cgroup_event_cgrp_time(struct perf_event
*event
)
565 static inline void update_cgrp_time_from_event(struct perf_event
*event
)
569 static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context
*cpuctx
)
573 static inline void perf_cgroup_sched_out(struct task_struct
*task
,
574 struct task_struct
*next
)
578 static inline void perf_cgroup_sched_in(struct task_struct
*prev
,
579 struct task_struct
*task
)
583 static inline int perf_cgroup_connect(pid_t pid
, struct perf_event
*event
,
584 struct perf_event_attr
*attr
,
585 struct perf_event
*group_leader
)
591 perf_cgroup_set_timestamp(struct task_struct
*task
,
592 struct perf_event_context
*ctx
)
597 perf_cgroup_switch(struct task_struct
*task
, struct task_struct
*next
)
602 perf_cgroup_set_shadow_time(struct perf_event
*event
, u64 now
)
606 static inline u64
perf_cgroup_event_time(struct perf_event
*event
)
612 perf_cgroup_defer_enabled(struct perf_event
*event
)
617 perf_cgroup_mark_enabled(struct perf_event
*event
,
618 struct perf_event_context
*ctx
)
623 void perf_pmu_disable(struct pmu
*pmu
)
625 int *count
= this_cpu_ptr(pmu
->pmu_disable_count
);
627 pmu
->pmu_disable(pmu
);
630 void perf_pmu_enable(struct pmu
*pmu
)
632 int *count
= this_cpu_ptr(pmu
->pmu_disable_count
);
634 pmu
->pmu_enable(pmu
);
637 static DEFINE_PER_CPU(struct list_head
, rotation_list
);
640 * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
641 * because they're strictly cpu affine and rotate_start is called with IRQs
642 * disabled, while rotate_context is called from IRQ context.
644 static void perf_pmu_rotate_start(struct pmu
*pmu
)
646 struct perf_cpu_context
*cpuctx
= this_cpu_ptr(pmu
->pmu_cpu_context
);
647 struct list_head
*head
= &__get_cpu_var(rotation_list
);
649 WARN_ON(!irqs_disabled());
651 if (list_empty(&cpuctx
->rotation_list
))
652 list_add(&cpuctx
->rotation_list
, head
);
655 static void get_ctx(struct perf_event_context
*ctx
)
657 WARN_ON(!atomic_inc_not_zero(&ctx
->refcount
));
660 static void put_ctx(struct perf_event_context
*ctx
)
662 if (atomic_dec_and_test(&ctx
->refcount
)) {
664 put_ctx(ctx
->parent_ctx
);
666 put_task_struct(ctx
->task
);
667 kfree_rcu(ctx
, rcu_head
);
671 static void unclone_ctx(struct perf_event_context
*ctx
)
673 if (ctx
->parent_ctx
) {
674 put_ctx(ctx
->parent_ctx
);
675 ctx
->parent_ctx
= NULL
;
679 static u32
perf_event_pid(struct perf_event
*event
, struct task_struct
*p
)
682 * only top level events have the pid namespace they were created in
685 event
= event
->parent
;
687 return task_tgid_nr_ns(p
, event
->ns
);
690 static u32
perf_event_tid(struct perf_event
*event
, struct task_struct
*p
)
693 * only top level events have the pid namespace they were created in
696 event
= event
->parent
;
698 return task_pid_nr_ns(p
, event
->ns
);
702 * If we inherit events we want to return the parent event id
705 static u64
primary_event_id(struct perf_event
*event
)
710 id
= event
->parent
->id
;
716 * Get the perf_event_context for a task and lock it.
717 * This has to cope with with the fact that until it is locked,
718 * the context could get moved to another task.
720 static struct perf_event_context
*
721 perf_lock_task_context(struct task_struct
*task
, int ctxn
, unsigned long *flags
)
723 struct perf_event_context
*ctx
;
727 ctx
= rcu_dereference(task
->perf_event_ctxp
[ctxn
]);
730 * If this context is a clone of another, it might
731 * get swapped for another underneath us by
732 * perf_event_task_sched_out, though the
733 * rcu_read_lock() protects us from any context
734 * getting freed. Lock the context and check if it
735 * got swapped before we could get the lock, and retry
736 * if so. If we locked the right context, then it
737 * can't get swapped on us any more.
739 raw_spin_lock_irqsave(&ctx
->lock
, *flags
);
740 if (ctx
!= rcu_dereference(task
->perf_event_ctxp
[ctxn
])) {
741 raw_spin_unlock_irqrestore(&ctx
->lock
, *flags
);
745 if (!atomic_inc_not_zero(&ctx
->refcount
)) {
746 raw_spin_unlock_irqrestore(&ctx
->lock
, *flags
);
755 * Get the context for a task and increment its pin_count so it
756 * can't get swapped to another task. This also increments its
757 * reference count so that the context can't get freed.
759 static struct perf_event_context
*
760 perf_pin_task_context(struct task_struct
*task
, int ctxn
)
762 struct perf_event_context
*ctx
;
765 ctx
= perf_lock_task_context(task
, ctxn
, &flags
);
768 raw_spin_unlock_irqrestore(&ctx
->lock
, flags
);
773 static void perf_unpin_context(struct perf_event_context
*ctx
)
777 raw_spin_lock_irqsave(&ctx
->lock
, flags
);
779 raw_spin_unlock_irqrestore(&ctx
->lock
, flags
);
783 * Update the record of the current time in a context.
785 static void update_context_time(struct perf_event_context
*ctx
)
787 u64 now
= perf_clock();
789 ctx
->time
+= now
- ctx
->timestamp
;
790 ctx
->timestamp
= now
;
793 static u64
perf_event_time(struct perf_event
*event
)
795 struct perf_event_context
*ctx
= event
->ctx
;
797 if (is_cgroup_event(event
))
798 return perf_cgroup_event_time(event
);
800 return ctx
? ctx
->time
: 0;
804 * Update the total_time_enabled and total_time_running fields for a event.
805 * The caller of this function needs to hold the ctx->lock.
807 static void update_event_times(struct perf_event
*event
)
809 struct perf_event_context
*ctx
= event
->ctx
;
812 if (event
->state
< PERF_EVENT_STATE_INACTIVE
||
813 event
->group_leader
->state
< PERF_EVENT_STATE_INACTIVE
)
816 * in cgroup mode, time_enabled represents
817 * the time the event was enabled AND active
818 * tasks were in the monitored cgroup. This is
819 * independent of the activity of the context as
820 * there may be a mix of cgroup and non-cgroup events.
822 * That is why we treat cgroup events differently
825 if (is_cgroup_event(event
))
826 run_end
= perf_cgroup_event_time(event
);
827 else if (ctx
->is_active
)
830 run_end
= event
->tstamp_stopped
;
832 event
->total_time_enabled
= run_end
- event
->tstamp_enabled
;
834 if (event
->state
== PERF_EVENT_STATE_INACTIVE
)
835 run_end
= event
->tstamp_stopped
;
837 run_end
= perf_event_time(event
);
839 event
->total_time_running
= run_end
- event
->tstamp_running
;
844 * Update total_time_enabled and total_time_running for all events in a group.
846 static void update_group_times(struct perf_event
*leader
)
848 struct perf_event
*event
;
850 update_event_times(leader
);
851 list_for_each_entry(event
, &leader
->sibling_list
, group_entry
)
852 update_event_times(event
);
855 static struct list_head
*
856 ctx_group_list(struct perf_event
*event
, struct perf_event_context
*ctx
)
858 if (event
->attr
.pinned
)
859 return &ctx
->pinned_groups
;
861 return &ctx
->flexible_groups
;
865 * Add a event from the lists for its context.
866 * Must be called with ctx->mutex and ctx->lock held.
869 list_add_event(struct perf_event
*event
, struct perf_event_context
*ctx
)
871 WARN_ON_ONCE(event
->attach_state
& PERF_ATTACH_CONTEXT
);
872 event
->attach_state
|= PERF_ATTACH_CONTEXT
;
875 * If we're a stand alone event or group leader, we go to the context
876 * list, group events are kept attached to the group so that
877 * perf_group_detach can, at all times, locate all siblings.
879 if (event
->group_leader
== event
) {
880 struct list_head
*list
;
882 if (is_software_event(event
))
883 event
->group_flags
|= PERF_GROUP_SOFTWARE
;
885 list
= ctx_group_list(event
, ctx
);
886 list_add_tail(&event
->group_entry
, list
);
889 if (is_cgroup_event(event
))
892 if (has_branch_stack(event
))
893 ctx
->nr_branch_stack
++;
895 list_add_rcu(&event
->event_entry
, &ctx
->event_list
);
897 perf_pmu_rotate_start(ctx
->pmu
);
899 if (event
->attr
.inherit_stat
)
904 * Called at perf_event creation and when events are attached/detached from a
907 static void perf_event__read_size(struct perf_event
*event
)
909 int entry
= sizeof(u64
); /* value */
913 if (event
->attr
.read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
916 if (event
->attr
.read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
919 if (event
->attr
.read_format
& PERF_FORMAT_ID
)
920 entry
+= sizeof(u64
);
922 if (event
->attr
.read_format
& PERF_FORMAT_GROUP
) {
923 nr
+= event
->group_leader
->nr_siblings
;
928 event
->read_size
= size
;
931 static void perf_event__header_size(struct perf_event
*event
)
933 struct perf_sample_data
*data
;
934 u64 sample_type
= event
->attr
.sample_type
;
937 perf_event__read_size(event
);
939 if (sample_type
& PERF_SAMPLE_IP
)
940 size
+= sizeof(data
->ip
);
942 if (sample_type
& PERF_SAMPLE_ADDR
)
943 size
+= sizeof(data
->addr
);
945 if (sample_type
& PERF_SAMPLE_PERIOD
)
946 size
+= sizeof(data
->period
);
948 if (sample_type
& PERF_SAMPLE_READ
)
949 size
+= event
->read_size
;
951 event
->header_size
= size
;
954 static void perf_event__id_header_size(struct perf_event
*event
)
956 struct perf_sample_data
*data
;
957 u64 sample_type
= event
->attr
.sample_type
;
960 if (sample_type
& PERF_SAMPLE_TID
)
961 size
+= sizeof(data
->tid_entry
);
963 if (sample_type
& PERF_SAMPLE_TIME
)
964 size
+= sizeof(data
->time
);
966 if (sample_type
& PERF_SAMPLE_ID
)
967 size
+= sizeof(data
->id
);
969 if (sample_type
& PERF_SAMPLE_STREAM_ID
)
970 size
+= sizeof(data
->stream_id
);
972 if (sample_type
& PERF_SAMPLE_CPU
)
973 size
+= sizeof(data
->cpu_entry
);
975 event
->id_header_size
= size
;
978 static void perf_group_attach(struct perf_event
*event
)
980 struct perf_event
*group_leader
= event
->group_leader
, *pos
;
983 * We can have double attach due to group movement in perf_event_open.
985 if (event
->attach_state
& PERF_ATTACH_GROUP
)
988 event
->attach_state
|= PERF_ATTACH_GROUP
;
990 if (group_leader
== event
)
993 if (group_leader
->group_flags
& PERF_GROUP_SOFTWARE
&&
994 !is_software_event(event
))
995 group_leader
->group_flags
&= ~PERF_GROUP_SOFTWARE
;
997 list_add_tail(&event
->group_entry
, &group_leader
->sibling_list
);
998 group_leader
->nr_siblings
++;
1000 perf_event__header_size(group_leader
);
1002 list_for_each_entry(pos
, &group_leader
->sibling_list
, group_entry
)
1003 perf_event__header_size(pos
);
1007 * Remove a event from the lists for its context.
1008 * Must be called with ctx->mutex and ctx->lock held.
1011 list_del_event(struct perf_event
*event
, struct perf_event_context
*ctx
)
1013 struct perf_cpu_context
*cpuctx
;
1015 * We can have double detach due to exit/hot-unplug + close.
1017 if (!(event
->attach_state
& PERF_ATTACH_CONTEXT
))
1020 event
->attach_state
&= ~PERF_ATTACH_CONTEXT
;
1022 if (is_cgroup_event(event
)) {
1024 cpuctx
= __get_cpu_context(ctx
);
1026 * if there are no more cgroup events
1027 * then cler cgrp to avoid stale pointer
1028 * in update_cgrp_time_from_cpuctx()
1030 if (!ctx
->nr_cgroups
)
1031 cpuctx
->cgrp
= NULL
;
1034 if (has_branch_stack(event
))
1035 ctx
->nr_branch_stack
--;
1038 if (event
->attr
.inherit_stat
)
1041 list_del_rcu(&event
->event_entry
);
1043 if (event
->group_leader
== event
)
1044 list_del_init(&event
->group_entry
);
1046 update_group_times(event
);
1049 * If event was in error state, then keep it
1050 * that way, otherwise bogus counts will be
1051 * returned on read(). The only way to get out
1052 * of error state is by explicit re-enabling
1055 if (event
->state
> PERF_EVENT_STATE_OFF
)
1056 event
->state
= PERF_EVENT_STATE_OFF
;
1059 static void perf_group_detach(struct perf_event
*event
)
1061 struct perf_event
*sibling
, *tmp
;
1062 struct list_head
*list
= NULL
;
1065 * We can have double detach due to exit/hot-unplug + close.
1067 if (!(event
->attach_state
& PERF_ATTACH_GROUP
))
1070 event
->attach_state
&= ~PERF_ATTACH_GROUP
;
1073 * If this is a sibling, remove it from its group.
1075 if (event
->group_leader
!= event
) {
1076 list_del_init(&event
->group_entry
);
1077 event
->group_leader
->nr_siblings
--;
1081 if (!list_empty(&event
->group_entry
))
1082 list
= &event
->group_entry
;
1085 * If this was a group event with sibling events then
1086 * upgrade the siblings to singleton events by adding them
1087 * to whatever list we are on.
1089 list_for_each_entry_safe(sibling
, tmp
, &event
->sibling_list
, group_entry
) {
1091 list_move_tail(&sibling
->group_entry
, list
);
1092 sibling
->group_leader
= sibling
;
1094 /* Inherit group flags from the previous leader */
1095 sibling
->group_flags
= event
->group_flags
;
1099 perf_event__header_size(event
->group_leader
);
1101 list_for_each_entry(tmp
, &event
->group_leader
->sibling_list
, group_entry
)
1102 perf_event__header_size(tmp
);
1106 event_filter_match(struct perf_event
*event
)
1108 return (event
->cpu
== -1 || event
->cpu
== smp_processor_id())
1109 && perf_cgroup_match(event
);
1113 event_sched_out(struct perf_event
*event
,
1114 struct perf_cpu_context
*cpuctx
,
1115 struct perf_event_context
*ctx
)
1117 u64 tstamp
= perf_event_time(event
);
1120 * An event which could not be activated because of
1121 * filter mismatch still needs to have its timings
1122 * maintained, otherwise bogus information is return
1123 * via read() for time_enabled, time_running:
1125 if (event
->state
== PERF_EVENT_STATE_INACTIVE
1126 && !event_filter_match(event
)) {
1127 delta
= tstamp
- event
->tstamp_stopped
;
1128 event
->tstamp_running
+= delta
;
1129 event
->tstamp_stopped
= tstamp
;
1132 if (event
->state
!= PERF_EVENT_STATE_ACTIVE
)
1135 event
->state
= PERF_EVENT_STATE_INACTIVE
;
1136 if (event
->pending_disable
) {
1137 event
->pending_disable
= 0;
1138 event
->state
= PERF_EVENT_STATE_OFF
;
1140 event
->tstamp_stopped
= tstamp
;
1141 event
->pmu
->del(event
, 0);
1144 if (!is_software_event(event
))
1145 cpuctx
->active_oncpu
--;
1147 if (event
->attr
.freq
&& event
->attr
.sample_freq
)
1149 if (event
->attr
.exclusive
|| !cpuctx
->active_oncpu
)
1150 cpuctx
->exclusive
= 0;
1154 group_sched_out(struct perf_event
*group_event
,
1155 struct perf_cpu_context
*cpuctx
,
1156 struct perf_event_context
*ctx
)
1158 struct perf_event
*event
;
1159 int state
= group_event
->state
;
1161 event_sched_out(group_event
, cpuctx
, ctx
);
1164 * Schedule out siblings (if any):
1166 list_for_each_entry(event
, &group_event
->sibling_list
, group_entry
)
1167 event_sched_out(event
, cpuctx
, ctx
);
1169 if (state
== PERF_EVENT_STATE_ACTIVE
&& group_event
->attr
.exclusive
)
1170 cpuctx
->exclusive
= 0;
1174 * Cross CPU call to remove a performance event
1176 * We disable the event on the hardware level first. After that we
1177 * remove it from the context list.
1179 static int __perf_remove_from_context(void *info
)
1181 struct perf_event
*event
= info
;
1182 struct perf_event_context
*ctx
= event
->ctx
;
1183 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
1185 raw_spin_lock(&ctx
->lock
);
1186 event_sched_out(event
, cpuctx
, ctx
);
1187 list_del_event(event
, ctx
);
1188 if (!ctx
->nr_events
&& cpuctx
->task_ctx
== ctx
) {
1190 cpuctx
->task_ctx
= NULL
;
1192 raw_spin_unlock(&ctx
->lock
);
1199 * Remove the event from a task's (or a CPU's) list of events.
1201 * CPU events are removed with a smp call. For task events we only
1202 * call when the task is on a CPU.
1204 * If event->ctx is a cloned context, callers must make sure that
1205 * every task struct that event->ctx->task could possibly point to
1206 * remains valid. This is OK when called from perf_release since
1207 * that only calls us on the top-level context, which can't be a clone.
1208 * When called from perf_event_exit_task, it's OK because the
1209 * context has been detached from its task.
1211 static void perf_remove_from_context(struct perf_event
*event
)
1213 struct perf_event_context
*ctx
= event
->ctx
;
1214 struct task_struct
*task
= ctx
->task
;
1216 lockdep_assert_held(&ctx
->mutex
);
1220 * Per cpu events are removed via an smp call and
1221 * the removal is always successful.
1223 cpu_function_call(event
->cpu
, __perf_remove_from_context
, event
);
1228 if (!task_function_call(task
, __perf_remove_from_context
, event
))
1231 raw_spin_lock_irq(&ctx
->lock
);
1233 * If we failed to find a running task, but find the context active now
1234 * that we've acquired the ctx->lock, retry.
1236 if (ctx
->is_active
) {
1237 raw_spin_unlock_irq(&ctx
->lock
);
1242 * Since the task isn't running, its safe to remove the event, us
1243 * holding the ctx->lock ensures the task won't get scheduled in.
1245 list_del_event(event
, ctx
);
1246 raw_spin_unlock_irq(&ctx
->lock
);
1250 * Cross CPU call to disable a performance event
1252 static int __perf_event_disable(void *info
)
1254 struct perf_event
*event
= info
;
1255 struct perf_event_context
*ctx
= event
->ctx
;
1256 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
1259 * If this is a per-task event, need to check whether this
1260 * event's task is the current task on this cpu.
1262 * Can trigger due to concurrent perf_event_context_sched_out()
1263 * flipping contexts around.
1265 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
)
1268 raw_spin_lock(&ctx
->lock
);
1271 * If the event is on, turn it off.
1272 * If it is in error state, leave it in error state.
1274 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
) {
1275 update_context_time(ctx
);
1276 update_cgrp_time_from_event(event
);
1277 update_group_times(event
);
1278 if (event
== event
->group_leader
)
1279 group_sched_out(event
, cpuctx
, ctx
);
1281 event_sched_out(event
, cpuctx
, ctx
);
1282 event
->state
= PERF_EVENT_STATE_OFF
;
1285 raw_spin_unlock(&ctx
->lock
);
1293 * If event->ctx is a cloned context, callers must make sure that
1294 * every task struct that event->ctx->task could possibly point to
1295 * remains valid. This condition is satisifed when called through
1296 * perf_event_for_each_child or perf_event_for_each because they
1297 * hold the top-level event's child_mutex, so any descendant that
1298 * goes to exit will block in sync_child_event.
1299 * When called from perf_pending_event it's OK because event->ctx
1300 * is the current context on this CPU and preemption is disabled,
1301 * hence we can't get into perf_event_task_sched_out for this context.
1303 void perf_event_disable(struct perf_event
*event
)
1305 struct perf_event_context
*ctx
= event
->ctx
;
1306 struct task_struct
*task
= ctx
->task
;
1310 * Disable the event on the cpu that it's on
1312 cpu_function_call(event
->cpu
, __perf_event_disable
, event
);
1317 if (!task_function_call(task
, __perf_event_disable
, event
))
1320 raw_spin_lock_irq(&ctx
->lock
);
1322 * If the event is still active, we need to retry the cross-call.
1324 if (event
->state
== PERF_EVENT_STATE_ACTIVE
) {
1325 raw_spin_unlock_irq(&ctx
->lock
);
1327 * Reload the task pointer, it might have been changed by
1328 * a concurrent perf_event_context_sched_out().
1335 * Since we have the lock this context can't be scheduled
1336 * in, so we can change the state safely.
1338 if (event
->state
== PERF_EVENT_STATE_INACTIVE
) {
1339 update_group_times(event
);
1340 event
->state
= PERF_EVENT_STATE_OFF
;
1342 raw_spin_unlock_irq(&ctx
->lock
);
1344 EXPORT_SYMBOL_GPL(perf_event_disable
);
1346 static void perf_set_shadow_time(struct perf_event
*event
,
1347 struct perf_event_context
*ctx
,
1351 * use the correct time source for the time snapshot
1353 * We could get by without this by leveraging the
1354 * fact that to get to this function, the caller
1355 * has most likely already called update_context_time()
1356 * and update_cgrp_time_xx() and thus both timestamp
1357 * are identical (or very close). Given that tstamp is,
1358 * already adjusted for cgroup, we could say that:
1359 * tstamp - ctx->timestamp
1361 * tstamp - cgrp->timestamp.
1363 * Then, in perf_output_read(), the calculation would
1364 * work with no changes because:
1365 * - event is guaranteed scheduled in
1366 * - no scheduled out in between
1367 * - thus the timestamp would be the same
1369 * But this is a bit hairy.
1371 * So instead, we have an explicit cgroup call to remain
1372 * within the time time source all along. We believe it
1373 * is cleaner and simpler to understand.
1375 if (is_cgroup_event(event
))
1376 perf_cgroup_set_shadow_time(event
, tstamp
);
1378 event
->shadow_ctx_time
= tstamp
- ctx
->timestamp
;
1381 #define MAX_INTERRUPTS (~0ULL)
1383 static void perf_log_throttle(struct perf_event
*event
, int enable
);
1386 event_sched_in(struct perf_event
*event
,
1387 struct perf_cpu_context
*cpuctx
,
1388 struct perf_event_context
*ctx
)
1390 u64 tstamp
= perf_event_time(event
);
1392 if (event
->state
<= PERF_EVENT_STATE_OFF
)
1395 event
->state
= PERF_EVENT_STATE_ACTIVE
;
1396 event
->oncpu
= smp_processor_id();
1399 * Unthrottle events, since we scheduled we might have missed several
1400 * ticks already, also for a heavily scheduling task there is little
1401 * guarantee it'll get a tick in a timely manner.
1403 if (unlikely(event
->hw
.interrupts
== MAX_INTERRUPTS
)) {
1404 perf_log_throttle(event
, 1);
1405 event
->hw
.interrupts
= 0;
1409 * The new state must be visible before we turn it on in the hardware:
1413 if (event
->pmu
->add(event
, PERF_EF_START
)) {
1414 event
->state
= PERF_EVENT_STATE_INACTIVE
;
1419 event
->tstamp_running
+= tstamp
- event
->tstamp_stopped
;
1421 perf_set_shadow_time(event
, ctx
, tstamp
);
1423 if (!is_software_event(event
))
1424 cpuctx
->active_oncpu
++;
1426 if (event
->attr
.freq
&& event
->attr
.sample_freq
)
1429 if (event
->attr
.exclusive
)
1430 cpuctx
->exclusive
= 1;
1436 group_sched_in(struct perf_event
*group_event
,
1437 struct perf_cpu_context
*cpuctx
,
1438 struct perf_event_context
*ctx
)
1440 struct perf_event
*event
, *partial_group
= NULL
;
1441 struct pmu
*pmu
= group_event
->pmu
;
1442 u64 now
= ctx
->time
;
1443 bool simulate
= false;
1445 if (group_event
->state
== PERF_EVENT_STATE_OFF
)
1448 pmu
->start_txn(pmu
);
1450 if (event_sched_in(group_event
, cpuctx
, ctx
)) {
1451 pmu
->cancel_txn(pmu
);
1456 * Schedule in siblings as one group (if any):
1458 list_for_each_entry(event
, &group_event
->sibling_list
, group_entry
) {
1459 if (event_sched_in(event
, cpuctx
, ctx
)) {
1460 partial_group
= event
;
1465 if (!pmu
->commit_txn(pmu
))
1470 * Groups can be scheduled in as one unit only, so undo any
1471 * partial group before returning:
1472 * The events up to the failed event are scheduled out normally,
1473 * tstamp_stopped will be updated.
1475 * The failed events and the remaining siblings need to have
1476 * their timings updated as if they had gone thru event_sched_in()
1477 * and event_sched_out(). This is required to get consistent timings
1478 * across the group. This also takes care of the case where the group
1479 * could never be scheduled by ensuring tstamp_stopped is set to mark
1480 * the time the event was actually stopped, such that time delta
1481 * calculation in update_event_times() is correct.
1483 list_for_each_entry(event
, &group_event
->sibling_list
, group_entry
) {
1484 if (event
== partial_group
)
1488 event
->tstamp_running
+= now
- event
->tstamp_stopped
;
1489 event
->tstamp_stopped
= now
;
1491 event_sched_out(event
, cpuctx
, ctx
);
1494 event_sched_out(group_event
, cpuctx
, ctx
);
1496 pmu
->cancel_txn(pmu
);
1502 * Work out whether we can put this event group on the CPU now.
1504 static int group_can_go_on(struct perf_event
*event
,
1505 struct perf_cpu_context
*cpuctx
,
1509 * Groups consisting entirely of software events can always go on.
1511 if (event
->group_flags
& PERF_GROUP_SOFTWARE
)
1514 * If an exclusive group is already on, no other hardware
1517 if (cpuctx
->exclusive
)
1520 * If this group is exclusive and there are already
1521 * events on the CPU, it can't go on.
1523 if (event
->attr
.exclusive
&& cpuctx
->active_oncpu
)
1526 * Otherwise, try to add it if all previous groups were able
1532 static void add_event_to_ctx(struct perf_event
*event
,
1533 struct perf_event_context
*ctx
)
1535 u64 tstamp
= perf_event_time(event
);
1537 list_add_event(event
, ctx
);
1538 perf_group_attach(event
);
1539 event
->tstamp_enabled
= tstamp
;
1540 event
->tstamp_running
= tstamp
;
1541 event
->tstamp_stopped
= tstamp
;
1544 static void task_ctx_sched_out(struct perf_event_context
*ctx
);
1546 ctx_sched_in(struct perf_event_context
*ctx
,
1547 struct perf_cpu_context
*cpuctx
,
1548 enum event_type_t event_type
,
1549 struct task_struct
*task
);
1551 static void perf_event_sched_in(struct perf_cpu_context
*cpuctx
,
1552 struct perf_event_context
*ctx
,
1553 struct task_struct
*task
)
1555 cpu_ctx_sched_in(cpuctx
, EVENT_PINNED
, task
);
1557 ctx_sched_in(ctx
, cpuctx
, EVENT_PINNED
, task
);
1558 cpu_ctx_sched_in(cpuctx
, EVENT_FLEXIBLE
, task
);
1560 ctx_sched_in(ctx
, cpuctx
, EVENT_FLEXIBLE
, task
);
1564 * Cross CPU call to install and enable a performance event
1566 * Must be called with ctx->mutex held
1568 static int __perf_install_in_context(void *info
)
1570 struct perf_event
*event
= info
;
1571 struct perf_event_context
*ctx
= event
->ctx
;
1572 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
1573 struct perf_event_context
*task_ctx
= cpuctx
->task_ctx
;
1574 struct task_struct
*task
= current
;
1576 perf_ctx_lock(cpuctx
, task_ctx
);
1577 perf_pmu_disable(cpuctx
->ctx
.pmu
);
1580 * If there was an active task_ctx schedule it out.
1583 task_ctx_sched_out(task_ctx
);
1586 * If the context we're installing events in is not the
1587 * active task_ctx, flip them.
1589 if (ctx
->task
&& task_ctx
!= ctx
) {
1591 raw_spin_unlock(&task_ctx
->lock
);
1592 raw_spin_lock(&ctx
->lock
);
1597 cpuctx
->task_ctx
= task_ctx
;
1598 task
= task_ctx
->task
;
1601 cpu_ctx_sched_out(cpuctx
, EVENT_ALL
);
1603 update_context_time(ctx
);
1605 * update cgrp time only if current cgrp
1606 * matches event->cgrp. Must be done before
1607 * calling add_event_to_ctx()
1609 update_cgrp_time_from_event(event
);
1611 add_event_to_ctx(event
, ctx
);
1614 * Schedule everything back in
1616 perf_event_sched_in(cpuctx
, task_ctx
, task
);
1618 perf_pmu_enable(cpuctx
->ctx
.pmu
);
1619 perf_ctx_unlock(cpuctx
, task_ctx
);
1625 * Attach a performance event to a context
1627 * First we add the event to the list with the hardware enable bit
1628 * in event->hw_config cleared.
1630 * If the event is attached to a task which is on a CPU we use a smp
1631 * call to enable it in the task context. The task might have been
1632 * scheduled away, but we check this in the smp call again.
1635 perf_install_in_context(struct perf_event_context
*ctx
,
1636 struct perf_event
*event
,
1639 struct task_struct
*task
= ctx
->task
;
1641 lockdep_assert_held(&ctx
->mutex
);
1647 * Per cpu events are installed via an smp call and
1648 * the install is always successful.
1650 cpu_function_call(cpu
, __perf_install_in_context
, event
);
1655 if (!task_function_call(task
, __perf_install_in_context
, event
))
1658 raw_spin_lock_irq(&ctx
->lock
);
1660 * If we failed to find a running task, but find the context active now
1661 * that we've acquired the ctx->lock, retry.
1663 if (ctx
->is_active
) {
1664 raw_spin_unlock_irq(&ctx
->lock
);
1669 * Since the task isn't running, its safe to add the event, us holding
1670 * the ctx->lock ensures the task won't get scheduled in.
1672 add_event_to_ctx(event
, ctx
);
1673 raw_spin_unlock_irq(&ctx
->lock
);
1677 * Put a event into inactive state and update time fields.
1678 * Enabling the leader of a group effectively enables all
1679 * the group members that aren't explicitly disabled, so we
1680 * have to update their ->tstamp_enabled also.
1681 * Note: this works for group members as well as group leaders
1682 * since the non-leader members' sibling_lists will be empty.
1684 static void __perf_event_mark_enabled(struct perf_event
*event
)
1686 struct perf_event
*sub
;
1687 u64 tstamp
= perf_event_time(event
);
1689 event
->state
= PERF_EVENT_STATE_INACTIVE
;
1690 event
->tstamp_enabled
= tstamp
- event
->total_time_enabled
;
1691 list_for_each_entry(sub
, &event
->sibling_list
, group_entry
) {
1692 if (sub
->state
>= PERF_EVENT_STATE_INACTIVE
)
1693 sub
->tstamp_enabled
= tstamp
- sub
->total_time_enabled
;
1698 * Cross CPU call to enable a performance event
1700 static int __perf_event_enable(void *info
)
1702 struct perf_event
*event
= info
;
1703 struct perf_event_context
*ctx
= event
->ctx
;
1704 struct perf_event
*leader
= event
->group_leader
;
1705 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
1708 if (WARN_ON_ONCE(!ctx
->is_active
))
1711 raw_spin_lock(&ctx
->lock
);
1712 update_context_time(ctx
);
1714 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
)
1718 * set current task's cgroup time reference point
1720 perf_cgroup_set_timestamp(current
, ctx
);
1722 __perf_event_mark_enabled(event
);
1724 if (!event_filter_match(event
)) {
1725 if (is_cgroup_event(event
))
1726 perf_cgroup_defer_enabled(event
);
1731 * If the event is in a group and isn't the group leader,
1732 * then don't put it on unless the group is on.
1734 if (leader
!= event
&& leader
->state
!= PERF_EVENT_STATE_ACTIVE
)
1737 if (!group_can_go_on(event
, cpuctx
, 1)) {
1740 if (event
== leader
)
1741 err
= group_sched_in(event
, cpuctx
, ctx
);
1743 err
= event_sched_in(event
, cpuctx
, ctx
);
1748 * If this event can't go on and it's part of a
1749 * group, then the whole group has to come off.
1751 if (leader
!= event
)
1752 group_sched_out(leader
, cpuctx
, ctx
);
1753 if (leader
->attr
.pinned
) {
1754 update_group_times(leader
);
1755 leader
->state
= PERF_EVENT_STATE_ERROR
;
1760 raw_spin_unlock(&ctx
->lock
);
1768 * If event->ctx is a cloned context, callers must make sure that
1769 * every task struct that event->ctx->task could possibly point to
1770 * remains valid. This condition is satisfied when called through
1771 * perf_event_for_each_child or perf_event_for_each as described
1772 * for perf_event_disable.
1774 void perf_event_enable(struct perf_event
*event
)
1776 struct perf_event_context
*ctx
= event
->ctx
;
1777 struct task_struct
*task
= ctx
->task
;
1781 * Enable the event on the cpu that it's on
1783 cpu_function_call(event
->cpu
, __perf_event_enable
, event
);
1787 raw_spin_lock_irq(&ctx
->lock
);
1788 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
)
1792 * If the event is in error state, clear that first.
1793 * That way, if we see the event in error state below, we
1794 * know that it has gone back into error state, as distinct
1795 * from the task having been scheduled away before the
1796 * cross-call arrived.
1798 if (event
->state
== PERF_EVENT_STATE_ERROR
)
1799 event
->state
= PERF_EVENT_STATE_OFF
;
1802 if (!ctx
->is_active
) {
1803 __perf_event_mark_enabled(event
);
1807 raw_spin_unlock_irq(&ctx
->lock
);
1809 if (!task_function_call(task
, __perf_event_enable
, event
))
1812 raw_spin_lock_irq(&ctx
->lock
);
1815 * If the context is active and the event is still off,
1816 * we need to retry the cross-call.
1818 if (ctx
->is_active
&& event
->state
== PERF_EVENT_STATE_OFF
) {
1820 * task could have been flipped by a concurrent
1821 * perf_event_context_sched_out()
1828 raw_spin_unlock_irq(&ctx
->lock
);
1830 EXPORT_SYMBOL_GPL(perf_event_enable
);
1832 int perf_event_refresh(struct perf_event
*event
, int refresh
)
1835 * not supported on inherited events
1837 if (event
->attr
.inherit
|| !is_sampling_event(event
))
1840 atomic_add(refresh
, &event
->event_limit
);
1841 perf_event_enable(event
);
1845 EXPORT_SYMBOL_GPL(perf_event_refresh
);
1847 static void ctx_sched_out(struct perf_event_context
*ctx
,
1848 struct perf_cpu_context
*cpuctx
,
1849 enum event_type_t event_type
)
1851 struct perf_event
*event
;
1852 int is_active
= ctx
->is_active
;
1854 ctx
->is_active
&= ~event_type
;
1855 if (likely(!ctx
->nr_events
))
1858 update_context_time(ctx
);
1859 update_cgrp_time_from_cpuctx(cpuctx
);
1860 if (!ctx
->nr_active
)
1863 perf_pmu_disable(ctx
->pmu
);
1864 if ((is_active
& EVENT_PINNED
) && (event_type
& EVENT_PINNED
)) {
1865 list_for_each_entry(event
, &ctx
->pinned_groups
, group_entry
)
1866 group_sched_out(event
, cpuctx
, ctx
);
1869 if ((is_active
& EVENT_FLEXIBLE
) && (event_type
& EVENT_FLEXIBLE
)) {
1870 list_for_each_entry(event
, &ctx
->flexible_groups
, group_entry
)
1871 group_sched_out(event
, cpuctx
, ctx
);
1873 perf_pmu_enable(ctx
->pmu
);
1877 * Test whether two contexts are equivalent, i.e. whether they
1878 * have both been cloned from the same version of the same context
1879 * and they both have the same number of enabled events.
1880 * If the number of enabled events is the same, then the set
1881 * of enabled events should be the same, because these are both
1882 * inherited contexts, therefore we can't access individual events
1883 * in them directly with an fd; we can only enable/disable all
1884 * events via prctl, or enable/disable all events in a family
1885 * via ioctl, which will have the same effect on both contexts.
1887 static int context_equiv(struct perf_event_context
*ctx1
,
1888 struct perf_event_context
*ctx2
)
1890 return ctx1
->parent_ctx
&& ctx1
->parent_ctx
== ctx2
->parent_ctx
1891 && ctx1
->parent_gen
== ctx2
->parent_gen
1892 && !ctx1
->pin_count
&& !ctx2
->pin_count
;
1895 static void __perf_event_sync_stat(struct perf_event
*event
,
1896 struct perf_event
*next_event
)
1900 if (!event
->attr
.inherit_stat
)
1904 * Update the event value, we cannot use perf_event_read()
1905 * because we're in the middle of a context switch and have IRQs
1906 * disabled, which upsets smp_call_function_single(), however
1907 * we know the event must be on the current CPU, therefore we
1908 * don't need to use it.
1910 switch (event
->state
) {
1911 case PERF_EVENT_STATE_ACTIVE
:
1912 event
->pmu
->read(event
);
1915 case PERF_EVENT_STATE_INACTIVE
:
1916 update_event_times(event
);
1924 * In order to keep per-task stats reliable we need to flip the event
1925 * values when we flip the contexts.
1927 value
= local64_read(&next_event
->count
);
1928 value
= local64_xchg(&event
->count
, value
);
1929 local64_set(&next_event
->count
, value
);
1931 swap(event
->total_time_enabled
, next_event
->total_time_enabled
);
1932 swap(event
->total_time_running
, next_event
->total_time_running
);
1935 * Since we swizzled the values, update the user visible data too.
1937 perf_event_update_userpage(event
);
1938 perf_event_update_userpage(next_event
);
1941 #define list_next_entry(pos, member) \
1942 list_entry(pos->member.next, typeof(*pos), member)
1944 static void perf_event_sync_stat(struct perf_event_context
*ctx
,
1945 struct perf_event_context
*next_ctx
)
1947 struct perf_event
*event
, *next_event
;
1952 update_context_time(ctx
);
1954 event
= list_first_entry(&ctx
->event_list
,
1955 struct perf_event
, event_entry
);
1957 next_event
= list_first_entry(&next_ctx
->event_list
,
1958 struct perf_event
, event_entry
);
1960 while (&event
->event_entry
!= &ctx
->event_list
&&
1961 &next_event
->event_entry
!= &next_ctx
->event_list
) {
1963 __perf_event_sync_stat(event
, next_event
);
1965 event
= list_next_entry(event
, event_entry
);
1966 next_event
= list_next_entry(next_event
, event_entry
);
1970 static void perf_event_context_sched_out(struct task_struct
*task
, int ctxn
,
1971 struct task_struct
*next
)
1973 struct perf_event_context
*ctx
= task
->perf_event_ctxp
[ctxn
];
1974 struct perf_event_context
*next_ctx
;
1975 struct perf_event_context
*parent
;
1976 struct perf_cpu_context
*cpuctx
;
1982 cpuctx
= __get_cpu_context(ctx
);
1983 if (!cpuctx
->task_ctx
)
1987 parent
= rcu_dereference(ctx
->parent_ctx
);
1988 next_ctx
= next
->perf_event_ctxp
[ctxn
];
1989 if (parent
&& next_ctx
&&
1990 rcu_dereference(next_ctx
->parent_ctx
) == parent
) {
1992 * Looks like the two contexts are clones, so we might be
1993 * able to optimize the context switch. We lock both
1994 * contexts and check that they are clones under the
1995 * lock (including re-checking that neither has been
1996 * uncloned in the meantime). It doesn't matter which
1997 * order we take the locks because no other cpu could
1998 * be trying to lock both of these tasks.
2000 raw_spin_lock(&ctx
->lock
);
2001 raw_spin_lock_nested(&next_ctx
->lock
, SINGLE_DEPTH_NESTING
);
2002 if (context_equiv(ctx
, next_ctx
)) {
2004 * XXX do we need a memory barrier of sorts
2005 * wrt to rcu_dereference() of perf_event_ctxp
2007 task
->perf_event_ctxp
[ctxn
] = next_ctx
;
2008 next
->perf_event_ctxp
[ctxn
] = ctx
;
2010 next_ctx
->task
= task
;
2013 perf_event_sync_stat(ctx
, next_ctx
);
2015 raw_spin_unlock(&next_ctx
->lock
);
2016 raw_spin_unlock(&ctx
->lock
);
2021 raw_spin_lock(&ctx
->lock
);
2022 ctx_sched_out(ctx
, cpuctx
, EVENT_ALL
);
2023 cpuctx
->task_ctx
= NULL
;
2024 raw_spin_unlock(&ctx
->lock
);
2028 #define for_each_task_context_nr(ctxn) \
2029 for ((ctxn) = 0; (ctxn) < perf_nr_task_contexts; (ctxn)++)
2032 * Called from scheduler to remove the events of the current task,
2033 * with interrupts disabled.
2035 * We stop each event and update the event value in event->count.
2037 * This does not protect us against NMI, but disable()
2038 * sets the disabled bit in the control field of event _before_
2039 * accessing the event control register. If a NMI hits, then it will
2040 * not restart the event.
2042 void __perf_event_task_sched_out(struct task_struct
*task
,
2043 struct task_struct
*next
)
2047 for_each_task_context_nr(ctxn
)
2048 perf_event_context_sched_out(task
, ctxn
, next
);
2051 * if cgroup events exist on this CPU, then we need
2052 * to check if we have to switch out PMU state.
2053 * cgroup event are system-wide mode only
2055 if (atomic_read(&__get_cpu_var(perf_cgroup_events
)))
2056 perf_cgroup_sched_out(task
, next
);
2059 static void task_ctx_sched_out(struct perf_event_context
*ctx
)
2061 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
2063 if (!cpuctx
->task_ctx
)
2066 if (WARN_ON_ONCE(ctx
!= cpuctx
->task_ctx
))
2069 ctx_sched_out(ctx
, cpuctx
, EVENT_ALL
);
2070 cpuctx
->task_ctx
= NULL
;
2074 * Called with IRQs disabled
2076 static void cpu_ctx_sched_out(struct perf_cpu_context
*cpuctx
,
2077 enum event_type_t event_type
)
2079 ctx_sched_out(&cpuctx
->ctx
, cpuctx
, event_type
);
2083 ctx_pinned_sched_in(struct perf_event_context
*ctx
,
2084 struct perf_cpu_context
*cpuctx
)
2086 struct perf_event
*event
;
2088 list_for_each_entry(event
, &ctx
->pinned_groups
, group_entry
) {
2089 if (event
->state
<= PERF_EVENT_STATE_OFF
)
2091 if (!event_filter_match(event
))
2094 /* may need to reset tstamp_enabled */
2095 if (is_cgroup_event(event
))
2096 perf_cgroup_mark_enabled(event
, ctx
);
2098 if (group_can_go_on(event
, cpuctx
, 1))
2099 group_sched_in(event
, cpuctx
, ctx
);
2102 * If this pinned group hasn't been scheduled,
2103 * put it in error state.
2105 if (event
->state
== PERF_EVENT_STATE_INACTIVE
) {
2106 update_group_times(event
);
2107 event
->state
= PERF_EVENT_STATE_ERROR
;
2113 ctx_flexible_sched_in(struct perf_event_context
*ctx
,
2114 struct perf_cpu_context
*cpuctx
)
2116 struct perf_event
*event
;
2119 list_for_each_entry(event
, &ctx
->flexible_groups
, group_entry
) {
2120 /* Ignore events in OFF or ERROR state */
2121 if (event
->state
<= PERF_EVENT_STATE_OFF
)
2124 * Listen to the 'cpu' scheduling filter constraint
2127 if (!event_filter_match(event
))
2130 /* may need to reset tstamp_enabled */
2131 if (is_cgroup_event(event
))
2132 perf_cgroup_mark_enabled(event
, ctx
);
2134 if (group_can_go_on(event
, cpuctx
, can_add_hw
)) {
2135 if (group_sched_in(event
, cpuctx
, ctx
))
2142 ctx_sched_in(struct perf_event_context
*ctx
,
2143 struct perf_cpu_context
*cpuctx
,
2144 enum event_type_t event_type
,
2145 struct task_struct
*task
)
2148 int is_active
= ctx
->is_active
;
2150 ctx
->is_active
|= event_type
;
2151 if (likely(!ctx
->nr_events
))
2155 ctx
->timestamp
= now
;
2156 perf_cgroup_set_timestamp(task
, ctx
);
2158 * First go through the list and put on any pinned groups
2159 * in order to give them the best chance of going on.
2161 if (!(is_active
& EVENT_PINNED
) && (event_type
& EVENT_PINNED
))
2162 ctx_pinned_sched_in(ctx
, cpuctx
);
2164 /* Then walk through the lower prio flexible groups */
2165 if (!(is_active
& EVENT_FLEXIBLE
) && (event_type
& EVENT_FLEXIBLE
))
2166 ctx_flexible_sched_in(ctx
, cpuctx
);
2169 static void cpu_ctx_sched_in(struct perf_cpu_context
*cpuctx
,
2170 enum event_type_t event_type
,
2171 struct task_struct
*task
)
2173 struct perf_event_context
*ctx
= &cpuctx
->ctx
;
2175 ctx_sched_in(ctx
, cpuctx
, event_type
, task
);
2178 static void perf_event_context_sched_in(struct perf_event_context
*ctx
,
2179 struct task_struct
*task
)
2181 struct perf_cpu_context
*cpuctx
;
2183 cpuctx
= __get_cpu_context(ctx
);
2184 if (cpuctx
->task_ctx
== ctx
)
2187 perf_ctx_lock(cpuctx
, ctx
);
2188 perf_pmu_disable(ctx
->pmu
);
2190 * We want to keep the following priority order:
2191 * cpu pinned (that don't need to move), task pinned,
2192 * cpu flexible, task flexible.
2194 cpu_ctx_sched_out(cpuctx
, EVENT_FLEXIBLE
);
2197 cpuctx
->task_ctx
= ctx
;
2199 perf_event_sched_in(cpuctx
, cpuctx
->task_ctx
, task
);
2201 perf_pmu_enable(ctx
->pmu
);
2202 perf_ctx_unlock(cpuctx
, ctx
);
2205 * Since these rotations are per-cpu, we need to ensure the
2206 * cpu-context we got scheduled on is actually rotating.
2208 perf_pmu_rotate_start(ctx
->pmu
);
2212 * When sampling the branck stack in system-wide, it may be necessary
2213 * to flush the stack on context switch. This happens when the branch
2214 * stack does not tag its entries with the pid of the current task.
2215 * Otherwise it becomes impossible to associate a branch entry with a
2216 * task. This ambiguity is more likely to appear when the branch stack
2217 * supports priv level filtering and the user sets it to monitor only
2218 * at the user level (which could be a useful measurement in system-wide
2219 * mode). In that case, the risk is high of having a branch stack with
2220 * branch from multiple tasks. Flushing may mean dropping the existing
2221 * entries or stashing them somewhere in the PMU specific code layer.
2223 * This function provides the context switch callback to the lower code
2224 * layer. It is invoked ONLY when there is at least one system-wide context
2225 * with at least one active event using taken branch sampling.
2227 static void perf_branch_stack_sched_in(struct task_struct
*prev
,
2228 struct task_struct
*task
)
2230 struct perf_cpu_context
*cpuctx
;
2232 unsigned long flags
;
2234 /* no need to flush branch stack if not changing task */
2238 local_irq_save(flags
);
2242 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
2243 cpuctx
= this_cpu_ptr(pmu
->pmu_cpu_context
);
2246 * check if the context has at least one
2247 * event using PERF_SAMPLE_BRANCH_STACK
2249 if (cpuctx
->ctx
.nr_branch_stack
> 0
2250 && pmu
->flush_branch_stack
) {
2252 pmu
= cpuctx
->ctx
.pmu
;
2254 perf_ctx_lock(cpuctx
, cpuctx
->task_ctx
);
2256 perf_pmu_disable(pmu
);
2258 pmu
->flush_branch_stack();
2260 perf_pmu_enable(pmu
);
2262 perf_ctx_unlock(cpuctx
, cpuctx
->task_ctx
);
2268 local_irq_restore(flags
);
2272 * Called from scheduler to add the events of the current task
2273 * with interrupts disabled.
2275 * We restore the event value and then enable it.
2277 * This does not protect us against NMI, but enable()
2278 * sets the enabled bit in the control field of event _before_
2279 * accessing the event control register. If a NMI hits, then it will
2280 * keep the event running.
2282 void __perf_event_task_sched_in(struct task_struct
*prev
,
2283 struct task_struct
*task
)
2285 struct perf_event_context
*ctx
;
2288 for_each_task_context_nr(ctxn
) {
2289 ctx
= task
->perf_event_ctxp
[ctxn
];
2293 perf_event_context_sched_in(ctx
, task
);
2296 * if cgroup events exist on this CPU, then we need
2297 * to check if we have to switch in PMU state.
2298 * cgroup event are system-wide mode only
2300 if (atomic_read(&__get_cpu_var(perf_cgroup_events
)))
2301 perf_cgroup_sched_in(prev
, task
);
2303 /* check for system-wide branch_stack events */
2304 if (atomic_read(&__get_cpu_var(perf_branch_stack_events
)))
2305 perf_branch_stack_sched_in(prev
, task
);
2308 static u64
perf_calculate_period(struct perf_event
*event
, u64 nsec
, u64 count
)
2310 u64 frequency
= event
->attr
.sample_freq
;
2311 u64 sec
= NSEC_PER_SEC
;
2312 u64 divisor
, dividend
;
2314 int count_fls
, nsec_fls
, frequency_fls
, sec_fls
;
2316 count_fls
= fls64(count
);
2317 nsec_fls
= fls64(nsec
);
2318 frequency_fls
= fls64(frequency
);
2322 * We got @count in @nsec, with a target of sample_freq HZ
2323 * the target period becomes:
2326 * period = -------------------
2327 * @nsec * sample_freq
2332 * Reduce accuracy by one bit such that @a and @b converge
2333 * to a similar magnitude.
2335 #define REDUCE_FLS(a, b) \
2337 if (a##_fls > b##_fls) { \
2347 * Reduce accuracy until either term fits in a u64, then proceed with
2348 * the other, so that finally we can do a u64/u64 division.
2350 while (count_fls
+ sec_fls
> 64 && nsec_fls
+ frequency_fls
> 64) {
2351 REDUCE_FLS(nsec
, frequency
);
2352 REDUCE_FLS(sec
, count
);
2355 if (count_fls
+ sec_fls
> 64) {
2356 divisor
= nsec
* frequency
;
2358 while (count_fls
+ sec_fls
> 64) {
2359 REDUCE_FLS(count
, sec
);
2363 dividend
= count
* sec
;
2365 dividend
= count
* sec
;
2367 while (nsec_fls
+ frequency_fls
> 64) {
2368 REDUCE_FLS(nsec
, frequency
);
2372 divisor
= nsec
* frequency
;
2378 return div64_u64(dividend
, divisor
);
2381 static DEFINE_PER_CPU(int, perf_throttled_count
);
2382 static DEFINE_PER_CPU(u64
, perf_throttled_seq
);
2384 static void perf_adjust_period(struct perf_event
*event
, u64 nsec
, u64 count
, bool disable
)
2386 struct hw_perf_event
*hwc
= &event
->hw
;
2387 s64 period
, sample_period
;
2390 period
= perf_calculate_period(event
, nsec
, count
);
2392 delta
= (s64
)(period
- hwc
->sample_period
);
2393 delta
= (delta
+ 7) / 8; /* low pass filter */
2395 sample_period
= hwc
->sample_period
+ delta
;
2400 hwc
->sample_period
= sample_period
;
2402 if (local64_read(&hwc
->period_left
) > 8*sample_period
) {
2404 event
->pmu
->stop(event
, PERF_EF_UPDATE
);
2406 local64_set(&hwc
->period_left
, 0);
2409 event
->pmu
->start(event
, PERF_EF_RELOAD
);
2414 * combine freq adjustment with unthrottling to avoid two passes over the
2415 * events. At the same time, make sure, having freq events does not change
2416 * the rate of unthrottling as that would introduce bias.
2418 static void perf_adjust_freq_unthr_context(struct perf_event_context
*ctx
,
2421 struct perf_event
*event
;
2422 struct hw_perf_event
*hwc
;
2423 u64 now
, period
= TICK_NSEC
;
2427 * only need to iterate over all events iff:
2428 * - context have events in frequency mode (needs freq adjust)
2429 * - there are events to unthrottle on this cpu
2431 if (!(ctx
->nr_freq
|| needs_unthr
))
2434 raw_spin_lock(&ctx
->lock
);
2435 perf_pmu_disable(ctx
->pmu
);
2437 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
2438 if (event
->state
!= PERF_EVENT_STATE_ACTIVE
)
2441 if (!event_filter_match(event
))
2446 if (needs_unthr
&& hwc
->interrupts
== MAX_INTERRUPTS
) {
2447 hwc
->interrupts
= 0;
2448 perf_log_throttle(event
, 1);
2449 event
->pmu
->start(event
, 0);
2452 if (!event
->attr
.freq
|| !event
->attr
.sample_freq
)
2456 * stop the event and update event->count
2458 event
->pmu
->stop(event
, PERF_EF_UPDATE
);
2460 now
= local64_read(&event
->count
);
2461 delta
= now
- hwc
->freq_count_stamp
;
2462 hwc
->freq_count_stamp
= now
;
2466 * reload only if value has changed
2467 * we have stopped the event so tell that
2468 * to perf_adjust_period() to avoid stopping it
2472 perf_adjust_period(event
, period
, delta
, false);
2474 event
->pmu
->start(event
, delta
> 0 ? PERF_EF_RELOAD
: 0);
2477 perf_pmu_enable(ctx
->pmu
);
2478 raw_spin_unlock(&ctx
->lock
);
2482 * Round-robin a context's events:
2484 static void rotate_ctx(struct perf_event_context
*ctx
)
2487 * Rotate the first entry last of non-pinned groups. Rotation might be
2488 * disabled by the inheritance code.
2490 if (!ctx
->rotate_disable
)
2491 list_rotate_left(&ctx
->flexible_groups
);
2495 * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
2496 * because they're strictly cpu affine and rotate_start is called with IRQs
2497 * disabled, while rotate_context is called from IRQ context.
2499 static void perf_rotate_context(struct perf_cpu_context
*cpuctx
)
2501 struct perf_event_context
*ctx
= NULL
;
2502 int rotate
= 0, remove
= 1;
2504 if (cpuctx
->ctx
.nr_events
) {
2506 if (cpuctx
->ctx
.nr_events
!= cpuctx
->ctx
.nr_active
)
2510 ctx
= cpuctx
->task_ctx
;
2511 if (ctx
&& ctx
->nr_events
) {
2513 if (ctx
->nr_events
!= ctx
->nr_active
)
2520 perf_ctx_lock(cpuctx
, cpuctx
->task_ctx
);
2521 perf_pmu_disable(cpuctx
->ctx
.pmu
);
2523 cpu_ctx_sched_out(cpuctx
, EVENT_FLEXIBLE
);
2525 ctx_sched_out(ctx
, cpuctx
, EVENT_FLEXIBLE
);
2527 rotate_ctx(&cpuctx
->ctx
);
2531 perf_event_sched_in(cpuctx
, ctx
, current
);
2533 perf_pmu_enable(cpuctx
->ctx
.pmu
);
2534 perf_ctx_unlock(cpuctx
, cpuctx
->task_ctx
);
2537 list_del_init(&cpuctx
->rotation_list
);
2540 void perf_event_task_tick(void)
2542 struct list_head
*head
= &__get_cpu_var(rotation_list
);
2543 struct perf_cpu_context
*cpuctx
, *tmp
;
2544 struct perf_event_context
*ctx
;
2547 WARN_ON(!irqs_disabled());
2549 __this_cpu_inc(perf_throttled_seq
);
2550 throttled
= __this_cpu_xchg(perf_throttled_count
, 0);
2552 list_for_each_entry_safe(cpuctx
, tmp
, head
, rotation_list
) {
2554 perf_adjust_freq_unthr_context(ctx
, throttled
);
2556 ctx
= cpuctx
->task_ctx
;
2558 perf_adjust_freq_unthr_context(ctx
, throttled
);
2560 if (cpuctx
->jiffies_interval
== 1 ||
2561 !(jiffies
% cpuctx
->jiffies_interval
))
2562 perf_rotate_context(cpuctx
);
2566 static int event_enable_on_exec(struct perf_event
*event
,
2567 struct perf_event_context
*ctx
)
2569 if (!event
->attr
.enable_on_exec
)
2572 event
->attr
.enable_on_exec
= 0;
2573 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
)
2576 __perf_event_mark_enabled(event
);
2582 * Enable all of a task's events that have been marked enable-on-exec.
2583 * This expects task == current.
2585 static void perf_event_enable_on_exec(struct perf_event_context
*ctx
)
2587 struct perf_event
*event
;
2588 unsigned long flags
;
2592 local_irq_save(flags
);
2593 if (!ctx
|| !ctx
->nr_events
)
2597 * We must ctxsw out cgroup events to avoid conflict
2598 * when invoking perf_task_event_sched_in() later on
2599 * in this function. Otherwise we end up trying to
2600 * ctxswin cgroup events which are already scheduled
2603 perf_cgroup_sched_out(current
, NULL
);
2605 raw_spin_lock(&ctx
->lock
);
2606 task_ctx_sched_out(ctx
);
2608 list_for_each_entry(event
, &ctx
->event_list
, event_entry
) {
2609 ret
= event_enable_on_exec(event
, ctx
);
2615 * Unclone this context if we enabled any event.
2620 raw_spin_unlock(&ctx
->lock
);
2623 * Also calls ctxswin for cgroup events, if any:
2625 perf_event_context_sched_in(ctx
, ctx
->task
);
2627 local_irq_restore(flags
);
2631 * Cross CPU call to read the hardware event
2633 static void __perf_event_read(void *info
)
2635 struct perf_event
*event
= info
;
2636 struct perf_event_context
*ctx
= event
->ctx
;
2637 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
2640 * If this is a task context, we need to check whether it is
2641 * the current task context of this cpu. If not it has been
2642 * scheduled out before the smp call arrived. In that case
2643 * event->count would have been updated to a recent sample
2644 * when the event was scheduled out.
2646 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
)
2649 raw_spin_lock(&ctx
->lock
);
2650 if (ctx
->is_active
) {
2651 update_context_time(ctx
);
2652 update_cgrp_time_from_event(event
);
2654 update_event_times(event
);
2655 if (event
->state
== PERF_EVENT_STATE_ACTIVE
)
2656 event
->pmu
->read(event
);
2657 raw_spin_unlock(&ctx
->lock
);
2660 static inline u64
perf_event_count(struct perf_event
*event
)
2662 return local64_read(&event
->count
) + atomic64_read(&event
->child_count
);
2665 static u64
perf_event_read(struct perf_event
*event
)
2668 * If event is enabled and currently active on a CPU, update the
2669 * value in the event structure:
2671 if (event
->state
== PERF_EVENT_STATE_ACTIVE
) {
2672 smp_call_function_single(event
->oncpu
,
2673 __perf_event_read
, event
, 1);
2674 } else if (event
->state
== PERF_EVENT_STATE_INACTIVE
) {
2675 struct perf_event_context
*ctx
= event
->ctx
;
2676 unsigned long flags
;
2678 raw_spin_lock_irqsave(&ctx
->lock
, flags
);
2680 * may read while context is not active
2681 * (e.g., thread is blocked), in that case
2682 * we cannot update context time
2684 if (ctx
->is_active
) {
2685 update_context_time(ctx
);
2686 update_cgrp_time_from_event(event
);
2688 update_event_times(event
);
2689 raw_spin_unlock_irqrestore(&ctx
->lock
, flags
);
2692 return perf_event_count(event
);
2696 * Initialize the perf_event context in a task_struct:
2698 static void __perf_event_init_context(struct perf_event_context
*ctx
)
2700 raw_spin_lock_init(&ctx
->lock
);
2701 mutex_init(&ctx
->mutex
);
2702 INIT_LIST_HEAD(&ctx
->pinned_groups
);
2703 INIT_LIST_HEAD(&ctx
->flexible_groups
);
2704 INIT_LIST_HEAD(&ctx
->event_list
);
2705 atomic_set(&ctx
->refcount
, 1);
2708 static struct perf_event_context
*
2709 alloc_perf_context(struct pmu
*pmu
, struct task_struct
*task
)
2711 struct perf_event_context
*ctx
;
2713 ctx
= kzalloc(sizeof(struct perf_event_context
), GFP_KERNEL
);
2717 __perf_event_init_context(ctx
);
2720 get_task_struct(task
);
2727 static struct task_struct
*
2728 find_lively_task_by_vpid(pid_t vpid
)
2730 struct task_struct
*task
;
2737 task
= find_task_by_vpid(vpid
);
2739 get_task_struct(task
);
2743 return ERR_PTR(-ESRCH
);
2745 /* Reuse ptrace permission checks for now. */
2747 if (!ptrace_may_access(task
, PTRACE_MODE_READ
))
2752 put_task_struct(task
);
2753 return ERR_PTR(err
);
2758 * Returns a matching context with refcount and pincount.
2760 static struct perf_event_context
*
2761 find_get_context(struct pmu
*pmu
, struct task_struct
*task
, int cpu
)
2763 struct perf_event_context
*ctx
;
2764 struct perf_cpu_context
*cpuctx
;
2765 unsigned long flags
;
2769 /* Must be root to operate on a CPU event: */
2770 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN
))
2771 return ERR_PTR(-EACCES
);
2774 * We could be clever and allow to attach a event to an
2775 * offline CPU and activate it when the CPU comes up, but
2778 if (!cpu_online(cpu
))
2779 return ERR_PTR(-ENODEV
);
2781 cpuctx
= per_cpu_ptr(pmu
->pmu_cpu_context
, cpu
);
2790 ctxn
= pmu
->task_ctx_nr
;
2795 ctx
= perf_lock_task_context(task
, ctxn
, &flags
);
2799 raw_spin_unlock_irqrestore(&ctx
->lock
, flags
);
2801 ctx
= alloc_perf_context(pmu
, task
);
2807 mutex_lock(&task
->perf_event_mutex
);
2809 * If it has already passed perf_event_exit_task().
2810 * we must see PF_EXITING, it takes this mutex too.
2812 if (task
->flags
& PF_EXITING
)
2814 else if (task
->perf_event_ctxp
[ctxn
])
2819 rcu_assign_pointer(task
->perf_event_ctxp
[ctxn
], ctx
);
2821 mutex_unlock(&task
->perf_event_mutex
);
2823 if (unlikely(err
)) {
2835 return ERR_PTR(err
);
2838 static void perf_event_free_filter(struct perf_event
*event
);
2840 static void free_event_rcu(struct rcu_head
*head
)
2842 struct perf_event
*event
;
2844 event
= container_of(head
, struct perf_event
, rcu_head
);
2846 put_pid_ns(event
->ns
);
2847 perf_event_free_filter(event
);
2851 static void ring_buffer_put(struct ring_buffer
*rb
);
2853 static void free_event(struct perf_event
*event
)
2855 irq_work_sync(&event
->pending
);
2857 if (!event
->parent
) {
2858 if (event
->attach_state
& PERF_ATTACH_TASK
)
2859 static_key_slow_dec_deferred(&perf_sched_events
);
2860 if (event
->attr
.mmap
|| event
->attr
.mmap_data
)
2861 atomic_dec(&nr_mmap_events
);
2862 if (event
->attr
.comm
)
2863 atomic_dec(&nr_comm_events
);
2864 if (event
->attr
.task
)
2865 atomic_dec(&nr_task_events
);
2866 if (event
->attr
.sample_type
& PERF_SAMPLE_CALLCHAIN
)
2867 put_callchain_buffers();
2868 if (is_cgroup_event(event
)) {
2869 atomic_dec(&per_cpu(perf_cgroup_events
, event
->cpu
));
2870 static_key_slow_dec_deferred(&perf_sched_events
);
2873 if (has_branch_stack(event
)) {
2874 static_key_slow_dec_deferred(&perf_sched_events
);
2875 /* is system-wide event */
2876 if (!(event
->attach_state
& PERF_ATTACH_TASK
))
2877 atomic_dec(&per_cpu(perf_branch_stack_events
,
2883 ring_buffer_put(event
->rb
);
2887 if (is_cgroup_event(event
))
2888 perf_detach_cgroup(event
);
2891 event
->destroy(event
);
2894 put_ctx(event
->ctx
);
2896 call_rcu(&event
->rcu_head
, free_event_rcu
);
2899 int perf_event_release_kernel(struct perf_event
*event
)
2901 struct perf_event_context
*ctx
= event
->ctx
;
2903 WARN_ON_ONCE(ctx
->parent_ctx
);
2905 * There are two ways this annotation is useful:
2907 * 1) there is a lock recursion from perf_event_exit_task
2908 * see the comment there.
2910 * 2) there is a lock-inversion with mmap_sem through
2911 * perf_event_read_group(), which takes faults while
2912 * holding ctx->mutex, however this is called after
2913 * the last filedesc died, so there is no possibility
2914 * to trigger the AB-BA case.
2916 mutex_lock_nested(&ctx
->mutex
, SINGLE_DEPTH_NESTING
);
2917 raw_spin_lock_irq(&ctx
->lock
);
2918 perf_group_detach(event
);
2919 raw_spin_unlock_irq(&ctx
->lock
);
2920 perf_remove_from_context(event
);
2921 mutex_unlock(&ctx
->mutex
);
2927 EXPORT_SYMBOL_GPL(perf_event_release_kernel
);
2930 * Called when the last reference to the file is gone.
2932 static int perf_release(struct inode
*inode
, struct file
*file
)
2934 struct perf_event
*event
= file
->private_data
;
2935 struct task_struct
*owner
;
2937 file
->private_data
= NULL
;
2940 owner
= ACCESS_ONCE(event
->owner
);
2942 * Matches the smp_wmb() in perf_event_exit_task(). If we observe
2943 * !owner it means the list deletion is complete and we can indeed
2944 * free this event, otherwise we need to serialize on
2945 * owner->perf_event_mutex.
2947 smp_read_barrier_depends();
2950 * Since delayed_put_task_struct() also drops the last
2951 * task reference we can safely take a new reference
2952 * while holding the rcu_read_lock().
2954 get_task_struct(owner
);
2959 mutex_lock(&owner
->perf_event_mutex
);
2961 * We have to re-check the event->owner field, if it is cleared
2962 * we raced with perf_event_exit_task(), acquiring the mutex
2963 * ensured they're done, and we can proceed with freeing the
2967 list_del_init(&event
->owner_entry
);
2968 mutex_unlock(&owner
->perf_event_mutex
);
2969 put_task_struct(owner
);
2972 return perf_event_release_kernel(event
);
2975 u64
perf_event_read_value(struct perf_event
*event
, u64
*enabled
, u64
*running
)
2977 struct perf_event
*child
;
2983 mutex_lock(&event
->child_mutex
);
2984 total
+= perf_event_read(event
);
2985 *enabled
+= event
->total_time_enabled
+
2986 atomic64_read(&event
->child_total_time_enabled
);
2987 *running
+= event
->total_time_running
+
2988 atomic64_read(&event
->child_total_time_running
);
2990 list_for_each_entry(child
, &event
->child_list
, child_list
) {
2991 total
+= perf_event_read(child
);
2992 *enabled
+= child
->total_time_enabled
;
2993 *running
+= child
->total_time_running
;
2995 mutex_unlock(&event
->child_mutex
);
2999 EXPORT_SYMBOL_GPL(perf_event_read_value
);
3001 static int perf_event_read_group(struct perf_event
*event
,
3002 u64 read_format
, char __user
*buf
)
3004 struct perf_event
*leader
= event
->group_leader
, *sub
;
3005 int n
= 0, size
= 0, ret
= -EFAULT
;
3006 struct perf_event_context
*ctx
= leader
->ctx
;
3008 u64 count
, enabled
, running
;
3010 mutex_lock(&ctx
->mutex
);
3011 count
= perf_event_read_value(leader
, &enabled
, &running
);
3013 values
[n
++] = 1 + leader
->nr_siblings
;
3014 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
3015 values
[n
++] = enabled
;
3016 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
3017 values
[n
++] = running
;
3018 values
[n
++] = count
;
3019 if (read_format
& PERF_FORMAT_ID
)
3020 values
[n
++] = primary_event_id(leader
);
3022 size
= n
* sizeof(u64
);
3024 if (copy_to_user(buf
, values
, size
))
3029 list_for_each_entry(sub
, &leader
->sibling_list
, group_entry
) {
3032 values
[n
++] = perf_event_read_value(sub
, &enabled
, &running
);
3033 if (read_format
& PERF_FORMAT_ID
)
3034 values
[n
++] = primary_event_id(sub
);
3036 size
= n
* sizeof(u64
);
3038 if (copy_to_user(buf
+ ret
, values
, size
)) {
3046 mutex_unlock(&ctx
->mutex
);
3051 static int perf_event_read_one(struct perf_event
*event
,
3052 u64 read_format
, char __user
*buf
)
3054 u64 enabled
, running
;
3058 values
[n
++] = perf_event_read_value(event
, &enabled
, &running
);
3059 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
3060 values
[n
++] = enabled
;
3061 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
3062 values
[n
++] = running
;
3063 if (read_format
& PERF_FORMAT_ID
)
3064 values
[n
++] = primary_event_id(event
);
3066 if (copy_to_user(buf
, values
, n
* sizeof(u64
)))
3069 return n
* sizeof(u64
);
3073 * Read the performance event - simple non blocking version for now
3076 perf_read_hw(struct perf_event
*event
, char __user
*buf
, size_t count
)
3078 u64 read_format
= event
->attr
.read_format
;
3082 * Return end-of-file for a read on a event that is in
3083 * error state (i.e. because it was pinned but it couldn't be
3084 * scheduled on to the CPU at some point).
3086 if (event
->state
== PERF_EVENT_STATE_ERROR
)
3089 if (count
< event
->read_size
)
3092 WARN_ON_ONCE(event
->ctx
->parent_ctx
);
3093 if (read_format
& PERF_FORMAT_GROUP
)
3094 ret
= perf_event_read_group(event
, read_format
, buf
);
3096 ret
= perf_event_read_one(event
, read_format
, buf
);
3102 perf_read(struct file
*file
, char __user
*buf
, size_t count
, loff_t
*ppos
)
3104 struct perf_event
*event
= file
->private_data
;
3106 return perf_read_hw(event
, buf
, count
);
3109 static unsigned int perf_poll(struct file
*file
, poll_table
*wait
)
3111 struct perf_event
*event
= file
->private_data
;
3112 struct ring_buffer
*rb
;
3113 unsigned int events
= POLL_HUP
;
3116 * Race between perf_event_set_output() and perf_poll(): perf_poll()
3117 * grabs the rb reference but perf_event_set_output() overrides it.
3118 * Here is the timeline for two threads T1, T2:
3119 * t0: T1, rb = rcu_dereference(event->rb)
3120 * t1: T2, old_rb = event->rb
3121 * t2: T2, event->rb = new rb
3122 * t3: T2, ring_buffer_detach(old_rb)
3123 * t4: T1, ring_buffer_attach(rb1)
3124 * t5: T1, poll_wait(event->waitq)
3126 * To avoid this problem, we grab mmap_mutex in perf_poll()
3127 * thereby ensuring that the assignment of the new ring buffer
3128 * and the detachment of the old buffer appear atomic to perf_poll()
3130 mutex_lock(&event
->mmap_mutex
);
3133 rb
= rcu_dereference(event
->rb
);
3135 ring_buffer_attach(event
, rb
);
3136 events
= atomic_xchg(&rb
->poll
, 0);
3140 mutex_unlock(&event
->mmap_mutex
);
3142 poll_wait(file
, &event
->waitq
, wait
);
3147 static void perf_event_reset(struct perf_event
*event
)
3149 (void)perf_event_read(event
);
3150 local64_set(&event
->count
, 0);
3151 perf_event_update_userpage(event
);
3155 * Holding the top-level event's child_mutex means that any
3156 * descendant process that has inherited this event will block
3157 * in sync_child_event if it goes to exit, thus satisfying the
3158 * task existence requirements of perf_event_enable/disable.
3160 static void perf_event_for_each_child(struct perf_event
*event
,
3161 void (*func
)(struct perf_event
*))
3163 struct perf_event
*child
;
3165 WARN_ON_ONCE(event
->ctx
->parent_ctx
);
3166 mutex_lock(&event
->child_mutex
);
3168 list_for_each_entry(child
, &event
->child_list
, child_list
)
3170 mutex_unlock(&event
->child_mutex
);
3173 static void perf_event_for_each(struct perf_event
*event
,
3174 void (*func
)(struct perf_event
*))
3176 struct perf_event_context
*ctx
= event
->ctx
;
3177 struct perf_event
*sibling
;
3179 WARN_ON_ONCE(ctx
->parent_ctx
);
3180 mutex_lock(&ctx
->mutex
);
3181 event
= event
->group_leader
;
3183 perf_event_for_each_child(event
, func
);
3184 list_for_each_entry(sibling
, &event
->sibling_list
, group_entry
)
3185 perf_event_for_each_child(sibling
, func
);
3186 mutex_unlock(&ctx
->mutex
);
3189 static int perf_event_period(struct perf_event
*event
, u64 __user
*arg
)
3191 struct perf_event_context
*ctx
= event
->ctx
;
3195 if (!is_sampling_event(event
))
3198 if (copy_from_user(&value
, arg
, sizeof(value
)))
3204 raw_spin_lock_irq(&ctx
->lock
);
3205 if (event
->attr
.freq
) {
3206 if (value
> sysctl_perf_event_sample_rate
) {
3211 event
->attr
.sample_freq
= value
;
3213 event
->attr
.sample_period
= value
;
3214 event
->hw
.sample_period
= value
;
3217 raw_spin_unlock_irq(&ctx
->lock
);
3222 static const struct file_operations perf_fops
;
3224 static struct perf_event
*perf_fget_light(int fd
, int *fput_needed
)
3228 file
= fget_light(fd
, fput_needed
);
3230 return ERR_PTR(-EBADF
);
3232 if (file
->f_op
!= &perf_fops
) {
3233 fput_light(file
, *fput_needed
);
3235 return ERR_PTR(-EBADF
);
3238 return file
->private_data
;
3241 static int perf_event_set_output(struct perf_event
*event
,
3242 struct perf_event
*output_event
);
3243 static int perf_event_set_filter(struct perf_event
*event
, void __user
*arg
);
3245 static long perf_ioctl(struct file
*file
, unsigned int cmd
, unsigned long arg
)
3247 struct perf_event
*event
= file
->private_data
;
3248 void (*func
)(struct perf_event
*);
3252 case PERF_EVENT_IOC_ENABLE
:
3253 func
= perf_event_enable
;
3255 case PERF_EVENT_IOC_DISABLE
:
3256 func
= perf_event_disable
;
3258 case PERF_EVENT_IOC_RESET
:
3259 func
= perf_event_reset
;
3262 case PERF_EVENT_IOC_REFRESH
:
3263 return perf_event_refresh(event
, arg
);
3265 case PERF_EVENT_IOC_PERIOD
:
3266 return perf_event_period(event
, (u64 __user
*)arg
);
3268 case PERF_EVENT_IOC_SET_OUTPUT
:
3270 struct perf_event
*output_event
= NULL
;
3271 int fput_needed
= 0;
3275 output_event
= perf_fget_light(arg
, &fput_needed
);
3276 if (IS_ERR(output_event
))
3277 return PTR_ERR(output_event
);
3280 ret
= perf_event_set_output(event
, output_event
);
3282 fput_light(output_event
->filp
, fput_needed
);
3287 case PERF_EVENT_IOC_SET_FILTER
:
3288 return perf_event_set_filter(event
, (void __user
*)arg
);
3294 if (flags
& PERF_IOC_FLAG_GROUP
)
3295 perf_event_for_each(event
, func
);
3297 perf_event_for_each_child(event
, func
);
3302 int perf_event_task_enable(void)
3304 struct perf_event
*event
;
3306 mutex_lock(¤t
->perf_event_mutex
);
3307 list_for_each_entry(event
, ¤t
->perf_event_list
, owner_entry
)
3308 perf_event_for_each_child(event
, perf_event_enable
);
3309 mutex_unlock(¤t
->perf_event_mutex
);
3314 int perf_event_task_disable(void)
3316 struct perf_event
*event
;
3318 mutex_lock(¤t
->perf_event_mutex
);
3319 list_for_each_entry(event
, ¤t
->perf_event_list
, owner_entry
)
3320 perf_event_for_each_child(event
, perf_event_disable
);
3321 mutex_unlock(¤t
->perf_event_mutex
);
3326 static int perf_event_index(struct perf_event
*event
)
3328 if (event
->hw
.state
& PERF_HES_STOPPED
)
3331 if (event
->state
!= PERF_EVENT_STATE_ACTIVE
)
3334 return event
->pmu
->event_idx(event
);
3337 static void calc_timer_values(struct perf_event
*event
,
3344 *now
= perf_clock();
3345 ctx_time
= event
->shadow_ctx_time
+ *now
;
3346 *enabled
= ctx_time
- event
->tstamp_enabled
;
3347 *running
= ctx_time
- event
->tstamp_running
;
3350 void __weak
arch_perf_update_userpage(struct perf_event_mmap_page
*userpg
, u64 now
)
3355 * Callers need to ensure there can be no nesting of this function, otherwise
3356 * the seqlock logic goes bad. We can not serialize this because the arch
3357 * code calls this from NMI context.
3359 void perf_event_update_userpage(struct perf_event
*event
)
3361 struct perf_event_mmap_page
*userpg
;
3362 struct ring_buffer
*rb
;
3363 u64 enabled
, running
, now
;
3367 * compute total_time_enabled, total_time_running
3368 * based on snapshot values taken when the event
3369 * was last scheduled in.
3371 * we cannot simply called update_context_time()
3372 * because of locking issue as we can be called in
3375 calc_timer_values(event
, &now
, &enabled
, &running
);
3376 rb
= rcu_dereference(event
->rb
);
3380 userpg
= rb
->user_page
;
3383 * Disable preemption so as to not let the corresponding user-space
3384 * spin too long if we get preempted.
3389 userpg
->index
= perf_event_index(event
);
3390 userpg
->offset
= perf_event_count(event
);
3392 userpg
->offset
-= local64_read(&event
->hw
.prev_count
);
3394 userpg
->time_enabled
= enabled
+
3395 atomic64_read(&event
->child_total_time_enabled
);
3397 userpg
->time_running
= running
+
3398 atomic64_read(&event
->child_total_time_running
);
3400 arch_perf_update_userpage(userpg
, now
);
3409 static int perf_mmap_fault(struct vm_area_struct
*vma
, struct vm_fault
*vmf
)
3411 struct perf_event
*event
= vma
->vm_file
->private_data
;
3412 struct ring_buffer
*rb
;
3413 int ret
= VM_FAULT_SIGBUS
;
3415 if (vmf
->flags
& FAULT_FLAG_MKWRITE
) {
3416 if (vmf
->pgoff
== 0)
3422 rb
= rcu_dereference(event
->rb
);
3426 if (vmf
->pgoff
&& (vmf
->flags
& FAULT_FLAG_WRITE
))
3429 vmf
->page
= perf_mmap_to_page(rb
, vmf
->pgoff
);
3433 get_page(vmf
->page
);
3434 vmf
->page
->mapping
= vma
->vm_file
->f_mapping
;
3435 vmf
->page
->index
= vmf
->pgoff
;
3444 static void ring_buffer_attach(struct perf_event
*event
,
3445 struct ring_buffer
*rb
)
3447 unsigned long flags
;
3449 if (!list_empty(&event
->rb_entry
))
3452 spin_lock_irqsave(&rb
->event_lock
, flags
);
3453 if (!list_empty(&event
->rb_entry
))
3456 list_add(&event
->rb_entry
, &rb
->event_list
);
3458 spin_unlock_irqrestore(&rb
->event_lock
, flags
);
3461 static void ring_buffer_detach(struct perf_event
*event
,
3462 struct ring_buffer
*rb
)
3464 unsigned long flags
;
3466 if (list_empty(&event
->rb_entry
))
3469 spin_lock_irqsave(&rb
->event_lock
, flags
);
3470 list_del_init(&event
->rb_entry
);
3471 wake_up_all(&event
->waitq
);
3472 spin_unlock_irqrestore(&rb
->event_lock
, flags
);
3475 static void ring_buffer_wakeup(struct perf_event
*event
)
3477 struct ring_buffer
*rb
;
3480 rb
= rcu_dereference(event
->rb
);
3484 list_for_each_entry_rcu(event
, &rb
->event_list
, rb_entry
)
3485 wake_up_all(&event
->waitq
);
3491 static void rb_free_rcu(struct rcu_head
*rcu_head
)
3493 struct ring_buffer
*rb
;
3495 rb
= container_of(rcu_head
, struct ring_buffer
, rcu_head
);
3499 static struct ring_buffer
*ring_buffer_get(struct perf_event
*event
)
3501 struct ring_buffer
*rb
;
3504 rb
= rcu_dereference(event
->rb
);
3506 if (!atomic_inc_not_zero(&rb
->refcount
))
3514 static void ring_buffer_put(struct ring_buffer
*rb
)
3516 struct perf_event
*event
, *n
;
3517 unsigned long flags
;
3519 if (!atomic_dec_and_test(&rb
->refcount
))
3522 spin_lock_irqsave(&rb
->event_lock
, flags
);
3523 list_for_each_entry_safe(event
, n
, &rb
->event_list
, rb_entry
) {
3524 list_del_init(&event
->rb_entry
);
3525 wake_up_all(&event
->waitq
);
3527 spin_unlock_irqrestore(&rb
->event_lock
, flags
);
3529 call_rcu(&rb
->rcu_head
, rb_free_rcu
);
3532 static void perf_mmap_open(struct vm_area_struct
*vma
)
3534 struct perf_event
*event
= vma
->vm_file
->private_data
;
3536 atomic_inc(&event
->mmap_count
);
3539 static void perf_mmap_close(struct vm_area_struct
*vma
)
3541 struct perf_event
*event
= vma
->vm_file
->private_data
;
3543 if (atomic_dec_and_mutex_lock(&event
->mmap_count
, &event
->mmap_mutex
)) {
3544 unsigned long size
= perf_data_size(event
->rb
);
3545 struct user_struct
*user
= event
->mmap_user
;
3546 struct ring_buffer
*rb
= event
->rb
;
3548 atomic_long_sub((size
>> PAGE_SHIFT
) + 1, &user
->locked_vm
);
3549 vma
->vm_mm
->pinned_vm
-= event
->mmap_locked
;
3550 rcu_assign_pointer(event
->rb
, NULL
);
3551 ring_buffer_detach(event
, rb
);
3552 mutex_unlock(&event
->mmap_mutex
);
3554 ring_buffer_put(rb
);
3559 static const struct vm_operations_struct perf_mmap_vmops
= {
3560 .open
= perf_mmap_open
,
3561 .close
= perf_mmap_close
,
3562 .fault
= perf_mmap_fault
,
3563 .page_mkwrite
= perf_mmap_fault
,
3566 static int perf_mmap(struct file
*file
, struct vm_area_struct
*vma
)
3568 struct perf_event
*event
= file
->private_data
;
3569 unsigned long user_locked
, user_lock_limit
;
3570 struct user_struct
*user
= current_user();
3571 unsigned long locked
, lock_limit
;
3572 struct ring_buffer
*rb
;
3573 unsigned long vma_size
;
3574 unsigned long nr_pages
;
3575 long user_extra
, extra
;
3576 int ret
= 0, flags
= 0;
3579 * Don't allow mmap() of inherited per-task counters. This would
3580 * create a performance issue due to all children writing to the
3583 if (event
->cpu
== -1 && event
->attr
.inherit
)
3586 if (!(vma
->vm_flags
& VM_SHARED
))
3589 vma_size
= vma
->vm_end
- vma
->vm_start
;
3590 nr_pages
= (vma_size
/ PAGE_SIZE
) - 1;
3593 * If we have rb pages ensure they're a power-of-two number, so we
3594 * can do bitmasks instead of modulo.
3596 if (nr_pages
!= 0 && !is_power_of_2(nr_pages
))
3599 if (vma_size
!= PAGE_SIZE
* (1 + nr_pages
))
3602 if (vma
->vm_pgoff
!= 0)
3605 WARN_ON_ONCE(event
->ctx
->parent_ctx
);
3606 mutex_lock(&event
->mmap_mutex
);
3608 if (event
->rb
->nr_pages
== nr_pages
)
3609 atomic_inc(&event
->rb
->refcount
);
3615 user_extra
= nr_pages
+ 1;
3616 user_lock_limit
= sysctl_perf_event_mlock
>> (PAGE_SHIFT
- 10);
3619 * Increase the limit linearly with more CPUs:
3621 user_lock_limit
*= num_online_cpus();
3623 user_locked
= atomic_long_read(&user
->locked_vm
) + user_extra
;
3626 if (user_locked
> user_lock_limit
)
3627 extra
= user_locked
- user_lock_limit
;
3629 lock_limit
= rlimit(RLIMIT_MEMLOCK
);
3630 lock_limit
>>= PAGE_SHIFT
;
3631 locked
= vma
->vm_mm
->pinned_vm
+ extra
;
3633 if ((locked
> lock_limit
) && perf_paranoid_tracepoint_raw() &&
3634 !capable(CAP_IPC_LOCK
)) {
3641 if (vma
->vm_flags
& VM_WRITE
)
3642 flags
|= RING_BUFFER_WRITABLE
;
3644 rb
= rb_alloc(nr_pages
,
3645 event
->attr
.watermark
? event
->attr
.wakeup_watermark
: 0,
3652 rcu_assign_pointer(event
->rb
, rb
);
3654 atomic_long_add(user_extra
, &user
->locked_vm
);
3655 event
->mmap_locked
= extra
;
3656 event
->mmap_user
= get_current_user();
3657 vma
->vm_mm
->pinned_vm
+= event
->mmap_locked
;
3659 perf_event_update_userpage(event
);
3663 atomic_inc(&event
->mmap_count
);
3664 mutex_unlock(&event
->mmap_mutex
);
3666 vma
->vm_flags
|= VM_RESERVED
;
3667 vma
->vm_ops
= &perf_mmap_vmops
;
3672 static int perf_fasync(int fd
, struct file
*filp
, int on
)
3674 struct inode
*inode
= filp
->f_path
.dentry
->d_inode
;
3675 struct perf_event
*event
= filp
->private_data
;
3678 mutex_lock(&inode
->i_mutex
);
3679 retval
= fasync_helper(fd
, filp
, on
, &event
->fasync
);
3680 mutex_unlock(&inode
->i_mutex
);
3688 static const struct file_operations perf_fops
= {
3689 .llseek
= no_llseek
,
3690 .release
= perf_release
,
3693 .unlocked_ioctl
= perf_ioctl
,
3694 .compat_ioctl
= perf_ioctl
,
3696 .fasync
= perf_fasync
,
3702 * If there's data, ensure we set the poll() state and publish everything
3703 * to user-space before waking everybody up.
3706 void perf_event_wakeup(struct perf_event
*event
)
3708 ring_buffer_wakeup(event
);
3710 if (event
->pending_kill
) {
3711 kill_fasync(&event
->fasync
, SIGIO
, event
->pending_kill
);
3712 event
->pending_kill
= 0;
3716 static void perf_pending_event(struct irq_work
*entry
)
3718 struct perf_event
*event
= container_of(entry
,
3719 struct perf_event
, pending
);
3721 if (event
->pending_disable
) {
3722 event
->pending_disable
= 0;
3723 __perf_event_disable(event
);
3726 if (event
->pending_wakeup
) {
3727 event
->pending_wakeup
= 0;
3728 perf_event_wakeup(event
);
3733 * We assume there is only KVM supporting the callbacks.
3734 * Later on, we might change it to a list if there is
3735 * another virtualization implementation supporting the callbacks.
3737 struct perf_guest_info_callbacks
*perf_guest_cbs
;
3739 int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks
*cbs
)
3741 perf_guest_cbs
= cbs
;
3744 EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks
);
3746 int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks
*cbs
)
3748 perf_guest_cbs
= NULL
;
3751 EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks
);
3753 static void __perf_event_header__init_id(struct perf_event_header
*header
,
3754 struct perf_sample_data
*data
,
3755 struct perf_event
*event
)
3757 u64 sample_type
= event
->attr
.sample_type
;
3759 data
->type
= sample_type
;
3760 header
->size
+= event
->id_header_size
;
3762 if (sample_type
& PERF_SAMPLE_TID
) {
3763 /* namespace issues */
3764 data
->tid_entry
.pid
= perf_event_pid(event
, current
);
3765 data
->tid_entry
.tid
= perf_event_tid(event
, current
);
3768 if (sample_type
& PERF_SAMPLE_TIME
)
3769 data
->time
= perf_clock();
3771 if (sample_type
& PERF_SAMPLE_ID
)
3772 data
->id
= primary_event_id(event
);
3774 if (sample_type
& PERF_SAMPLE_STREAM_ID
)
3775 data
->stream_id
= event
->id
;
3777 if (sample_type
& PERF_SAMPLE_CPU
) {
3778 data
->cpu_entry
.cpu
= raw_smp_processor_id();
3779 data
->cpu_entry
.reserved
= 0;
3783 void perf_event_header__init_id(struct perf_event_header
*header
,
3784 struct perf_sample_data
*data
,
3785 struct perf_event
*event
)
3787 if (event
->attr
.sample_id_all
)
3788 __perf_event_header__init_id(header
, data
, event
);
3791 static void __perf_event__output_id_sample(struct perf_output_handle
*handle
,
3792 struct perf_sample_data
*data
)
3794 u64 sample_type
= data
->type
;
3796 if (sample_type
& PERF_SAMPLE_TID
)
3797 perf_output_put(handle
, data
->tid_entry
);
3799 if (sample_type
& PERF_SAMPLE_TIME
)
3800 perf_output_put(handle
, data
->time
);
3802 if (sample_type
& PERF_SAMPLE_ID
)
3803 perf_output_put(handle
, data
->id
);
3805 if (sample_type
& PERF_SAMPLE_STREAM_ID
)
3806 perf_output_put(handle
, data
->stream_id
);
3808 if (sample_type
& PERF_SAMPLE_CPU
)
3809 perf_output_put(handle
, data
->cpu_entry
);
3812 void perf_event__output_id_sample(struct perf_event
*event
,
3813 struct perf_output_handle
*handle
,
3814 struct perf_sample_data
*sample
)
3816 if (event
->attr
.sample_id_all
)
3817 __perf_event__output_id_sample(handle
, sample
);
3820 static void perf_output_read_one(struct perf_output_handle
*handle
,
3821 struct perf_event
*event
,
3822 u64 enabled
, u64 running
)
3824 u64 read_format
= event
->attr
.read_format
;
3828 values
[n
++] = perf_event_count(event
);
3829 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
) {
3830 values
[n
++] = enabled
+
3831 atomic64_read(&event
->child_total_time_enabled
);
3833 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
) {
3834 values
[n
++] = running
+
3835 atomic64_read(&event
->child_total_time_running
);
3837 if (read_format
& PERF_FORMAT_ID
)
3838 values
[n
++] = primary_event_id(event
);
3840 __output_copy(handle
, values
, n
* sizeof(u64
));
3844 * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
3846 static void perf_output_read_group(struct perf_output_handle
*handle
,
3847 struct perf_event
*event
,
3848 u64 enabled
, u64 running
)
3850 struct perf_event
*leader
= event
->group_leader
, *sub
;
3851 u64 read_format
= event
->attr
.read_format
;
3855 values
[n
++] = 1 + leader
->nr_siblings
;
3857 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
3858 values
[n
++] = enabled
;
3860 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
3861 values
[n
++] = running
;
3863 if (leader
!= event
)
3864 leader
->pmu
->read(leader
);
3866 values
[n
++] = perf_event_count(leader
);
3867 if (read_format
& PERF_FORMAT_ID
)
3868 values
[n
++] = primary_event_id(leader
);
3870 __output_copy(handle
, values
, n
* sizeof(u64
));
3872 list_for_each_entry(sub
, &leader
->sibling_list
, group_entry
) {
3876 sub
->pmu
->read(sub
);
3878 values
[n
++] = perf_event_count(sub
);
3879 if (read_format
& PERF_FORMAT_ID
)
3880 values
[n
++] = primary_event_id(sub
);
3882 __output_copy(handle
, values
, n
* sizeof(u64
));
3886 #define PERF_FORMAT_TOTAL_TIMES (PERF_FORMAT_TOTAL_TIME_ENABLED|\
3887 PERF_FORMAT_TOTAL_TIME_RUNNING)
3889 static void perf_output_read(struct perf_output_handle
*handle
,
3890 struct perf_event
*event
)
3892 u64 enabled
= 0, running
= 0, now
;
3893 u64 read_format
= event
->attr
.read_format
;
3896 * compute total_time_enabled, total_time_running
3897 * based on snapshot values taken when the event
3898 * was last scheduled in.
3900 * we cannot simply called update_context_time()
3901 * because of locking issue as we are called in
3904 if (read_format
& PERF_FORMAT_TOTAL_TIMES
)
3905 calc_timer_values(event
, &now
, &enabled
, &running
);
3907 if (event
->attr
.read_format
& PERF_FORMAT_GROUP
)
3908 perf_output_read_group(handle
, event
, enabled
, running
);
3910 perf_output_read_one(handle
, event
, enabled
, running
);
3913 void perf_output_sample(struct perf_output_handle
*handle
,
3914 struct perf_event_header
*header
,
3915 struct perf_sample_data
*data
,
3916 struct perf_event
*event
)
3918 u64 sample_type
= data
->type
;
3920 perf_output_put(handle
, *header
);
3922 if (sample_type
& PERF_SAMPLE_IP
)
3923 perf_output_put(handle
, data
->ip
);
3925 if (sample_type
& PERF_SAMPLE_TID
)
3926 perf_output_put(handle
, data
->tid_entry
);
3928 if (sample_type
& PERF_SAMPLE_TIME
)
3929 perf_output_put(handle
, data
->time
);
3931 if (sample_type
& PERF_SAMPLE_ADDR
)
3932 perf_output_put(handle
, data
->addr
);
3934 if (sample_type
& PERF_SAMPLE_ID
)
3935 perf_output_put(handle
, data
->id
);
3937 if (sample_type
& PERF_SAMPLE_STREAM_ID
)
3938 perf_output_put(handle
, data
->stream_id
);
3940 if (sample_type
& PERF_SAMPLE_CPU
)
3941 perf_output_put(handle
, data
->cpu_entry
);
3943 if (sample_type
& PERF_SAMPLE_PERIOD
)
3944 perf_output_put(handle
, data
->period
);
3946 if (sample_type
& PERF_SAMPLE_READ
)
3947 perf_output_read(handle
, event
);
3949 if (sample_type
& PERF_SAMPLE_CALLCHAIN
) {
3950 if (data
->callchain
) {
3953 if (data
->callchain
)
3954 size
+= data
->callchain
->nr
;
3956 size
*= sizeof(u64
);
3958 __output_copy(handle
, data
->callchain
, size
);
3961 perf_output_put(handle
, nr
);
3965 if (sample_type
& PERF_SAMPLE_RAW
) {
3967 perf_output_put(handle
, data
->raw
->size
);
3968 __output_copy(handle
, data
->raw
->data
,
3975 .size
= sizeof(u32
),
3978 perf_output_put(handle
, raw
);
3982 if (!event
->attr
.watermark
) {
3983 int wakeup_events
= event
->attr
.wakeup_events
;
3985 if (wakeup_events
) {
3986 struct ring_buffer
*rb
= handle
->rb
;
3987 int events
= local_inc_return(&rb
->events
);
3989 if (events
>= wakeup_events
) {
3990 local_sub(wakeup_events
, &rb
->events
);
3991 local_inc(&rb
->wakeup
);
3996 if (sample_type
& PERF_SAMPLE_BRANCH_STACK
) {
3997 if (data
->br_stack
) {
4000 size
= data
->br_stack
->nr
4001 * sizeof(struct perf_branch_entry
);
4003 perf_output_put(handle
, data
->br_stack
->nr
);
4004 perf_output_copy(handle
, data
->br_stack
->entries
, size
);
4007 * we always store at least the value of nr
4010 perf_output_put(handle
, nr
);
4015 void perf_prepare_sample(struct perf_event_header
*header
,
4016 struct perf_sample_data
*data
,
4017 struct perf_event
*event
,
4018 struct pt_regs
*regs
)
4020 u64 sample_type
= event
->attr
.sample_type
;
4022 header
->type
= PERF_RECORD_SAMPLE
;
4023 header
->size
= sizeof(*header
) + event
->header_size
;
4026 header
->misc
|= perf_misc_flags(regs
);
4028 __perf_event_header__init_id(header
, data
, event
);
4030 if (sample_type
& PERF_SAMPLE_IP
)
4031 data
->ip
= perf_instruction_pointer(regs
);
4033 if (sample_type
& PERF_SAMPLE_CALLCHAIN
) {
4036 data
->callchain
= perf_callchain(regs
);
4038 if (data
->callchain
)
4039 size
+= data
->callchain
->nr
;
4041 header
->size
+= size
* sizeof(u64
);
4044 if (sample_type
& PERF_SAMPLE_RAW
) {
4045 int size
= sizeof(u32
);
4048 size
+= data
->raw
->size
;
4050 size
+= sizeof(u32
);
4052 WARN_ON_ONCE(size
& (sizeof(u64
)-1));
4053 header
->size
+= size
;
4056 if (sample_type
& PERF_SAMPLE_BRANCH_STACK
) {
4057 int size
= sizeof(u64
); /* nr */
4058 if (data
->br_stack
) {
4059 size
+= data
->br_stack
->nr
4060 * sizeof(struct perf_branch_entry
);
4062 header
->size
+= size
;
4066 static void perf_event_output(struct perf_event
*event
,
4067 struct perf_sample_data
*data
,
4068 struct pt_regs
*regs
)
4070 struct perf_output_handle handle
;
4071 struct perf_event_header header
;
4073 /* protect the callchain buffers */
4076 perf_prepare_sample(&header
, data
, event
, regs
);
4078 if (perf_output_begin(&handle
, event
, header
.size
))
4081 perf_output_sample(&handle
, &header
, data
, event
);
4083 perf_output_end(&handle
);
4093 struct perf_read_event
{
4094 struct perf_event_header header
;
4101 perf_event_read_event(struct perf_event
*event
,
4102 struct task_struct
*task
)
4104 struct perf_output_handle handle
;
4105 struct perf_sample_data sample
;
4106 struct perf_read_event read_event
= {
4108 .type
= PERF_RECORD_READ
,
4110 .size
= sizeof(read_event
) + event
->read_size
,
4112 .pid
= perf_event_pid(event
, task
),
4113 .tid
= perf_event_tid(event
, task
),
4117 perf_event_header__init_id(&read_event
.header
, &sample
, event
);
4118 ret
= perf_output_begin(&handle
, event
, read_event
.header
.size
);
4122 perf_output_put(&handle
, read_event
);
4123 perf_output_read(&handle
, event
);
4124 perf_event__output_id_sample(event
, &handle
, &sample
);
4126 perf_output_end(&handle
);
4130 * task tracking -- fork/exit
4132 * enabled by: attr.comm | attr.mmap | attr.mmap_data | attr.task
4135 struct perf_task_event
{
4136 struct task_struct
*task
;
4137 struct perf_event_context
*task_ctx
;
4140 struct perf_event_header header
;
4150 static void perf_event_task_output(struct perf_event
*event
,
4151 struct perf_task_event
*task_event
)
4153 struct perf_output_handle handle
;
4154 struct perf_sample_data sample
;
4155 struct task_struct
*task
= task_event
->task
;
4156 int ret
, size
= task_event
->event_id
.header
.size
;
4158 perf_event_header__init_id(&task_event
->event_id
.header
, &sample
, event
);
4160 ret
= perf_output_begin(&handle
, event
,
4161 task_event
->event_id
.header
.size
);
4165 task_event
->event_id
.pid
= perf_event_pid(event
, task
);
4166 task_event
->event_id
.ppid
= perf_event_pid(event
, current
);
4168 task_event
->event_id
.tid
= perf_event_tid(event
, task
);
4169 task_event
->event_id
.ptid
= perf_event_tid(event
, current
);
4171 perf_output_put(&handle
, task_event
->event_id
);
4173 perf_event__output_id_sample(event
, &handle
, &sample
);
4175 perf_output_end(&handle
);
4177 task_event
->event_id
.header
.size
= size
;
4180 static int perf_event_task_match(struct perf_event
*event
)
4182 if (event
->state
< PERF_EVENT_STATE_INACTIVE
)
4185 if (!event_filter_match(event
))
4188 if (event
->attr
.comm
|| event
->attr
.mmap
||
4189 event
->attr
.mmap_data
|| event
->attr
.task
)
4195 static void perf_event_task_ctx(struct perf_event_context
*ctx
,
4196 struct perf_task_event
*task_event
)
4198 struct perf_event
*event
;
4200 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
4201 if (perf_event_task_match(event
))
4202 perf_event_task_output(event
, task_event
);
4206 static void perf_event_task_event(struct perf_task_event
*task_event
)
4208 struct perf_cpu_context
*cpuctx
;
4209 struct perf_event_context
*ctx
;
4214 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
4215 cpuctx
= get_cpu_ptr(pmu
->pmu_cpu_context
);
4216 if (cpuctx
->active_pmu
!= pmu
)
4218 perf_event_task_ctx(&cpuctx
->ctx
, task_event
);
4220 ctx
= task_event
->task_ctx
;
4222 ctxn
= pmu
->task_ctx_nr
;
4225 ctx
= rcu_dereference(current
->perf_event_ctxp
[ctxn
]);
4228 perf_event_task_ctx(ctx
, task_event
);
4230 put_cpu_ptr(pmu
->pmu_cpu_context
);
4235 static void perf_event_task(struct task_struct
*task
,
4236 struct perf_event_context
*task_ctx
,
4239 struct perf_task_event task_event
;
4241 if (!atomic_read(&nr_comm_events
) &&
4242 !atomic_read(&nr_mmap_events
) &&
4243 !atomic_read(&nr_task_events
))
4246 task_event
= (struct perf_task_event
){
4248 .task_ctx
= task_ctx
,
4251 .type
= new ? PERF_RECORD_FORK
: PERF_RECORD_EXIT
,
4253 .size
= sizeof(task_event
.event_id
),
4259 .time
= perf_clock(),
4263 perf_event_task_event(&task_event
);
4266 void perf_event_fork(struct task_struct
*task
)
4268 perf_event_task(task
, NULL
, 1);
4275 struct perf_comm_event
{
4276 struct task_struct
*task
;
4281 struct perf_event_header header
;
4288 static void perf_event_comm_output(struct perf_event
*event
,
4289 struct perf_comm_event
*comm_event
)
4291 struct perf_output_handle handle
;
4292 struct perf_sample_data sample
;
4293 int size
= comm_event
->event_id
.header
.size
;
4296 perf_event_header__init_id(&comm_event
->event_id
.header
, &sample
, event
);
4297 ret
= perf_output_begin(&handle
, event
,
4298 comm_event
->event_id
.header
.size
);
4303 comm_event
->event_id
.pid
= perf_event_pid(event
, comm_event
->task
);
4304 comm_event
->event_id
.tid
= perf_event_tid(event
, comm_event
->task
);
4306 perf_output_put(&handle
, comm_event
->event_id
);
4307 __output_copy(&handle
, comm_event
->comm
,
4308 comm_event
->comm_size
);
4310 perf_event__output_id_sample(event
, &handle
, &sample
);
4312 perf_output_end(&handle
);
4314 comm_event
->event_id
.header
.size
= size
;
4317 static int perf_event_comm_match(struct perf_event
*event
)
4319 if (event
->state
< PERF_EVENT_STATE_INACTIVE
)
4322 if (!event_filter_match(event
))
4325 if (event
->attr
.comm
)
4331 static void perf_event_comm_ctx(struct perf_event_context
*ctx
,
4332 struct perf_comm_event
*comm_event
)
4334 struct perf_event
*event
;
4336 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
4337 if (perf_event_comm_match(event
))
4338 perf_event_comm_output(event
, comm_event
);
4342 static void perf_event_comm_event(struct perf_comm_event
*comm_event
)
4344 struct perf_cpu_context
*cpuctx
;
4345 struct perf_event_context
*ctx
;
4346 char comm
[TASK_COMM_LEN
];
4351 memset(comm
, 0, sizeof(comm
));
4352 strlcpy(comm
, comm_event
->task
->comm
, sizeof(comm
));
4353 size
= ALIGN(strlen(comm
)+1, sizeof(u64
));
4355 comm_event
->comm
= comm
;
4356 comm_event
->comm_size
= size
;
4358 comm_event
->event_id
.header
.size
= sizeof(comm_event
->event_id
) + size
;
4360 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
4361 cpuctx
= get_cpu_ptr(pmu
->pmu_cpu_context
);
4362 if (cpuctx
->active_pmu
!= pmu
)
4364 perf_event_comm_ctx(&cpuctx
->ctx
, comm_event
);
4366 ctxn
= pmu
->task_ctx_nr
;
4370 ctx
= rcu_dereference(current
->perf_event_ctxp
[ctxn
]);
4372 perf_event_comm_ctx(ctx
, comm_event
);
4374 put_cpu_ptr(pmu
->pmu_cpu_context
);
4379 void perf_event_comm(struct task_struct
*task
)
4381 struct perf_comm_event comm_event
;
4382 struct perf_event_context
*ctx
;
4385 for_each_task_context_nr(ctxn
) {
4386 ctx
= task
->perf_event_ctxp
[ctxn
];
4390 perf_event_enable_on_exec(ctx
);
4393 if (!atomic_read(&nr_comm_events
))
4396 comm_event
= (struct perf_comm_event
){
4402 .type
= PERF_RECORD_COMM
,
4411 perf_event_comm_event(&comm_event
);
4418 struct perf_mmap_event
{
4419 struct vm_area_struct
*vma
;
4421 const char *file_name
;
4425 struct perf_event_header header
;
4435 static void perf_event_mmap_output(struct perf_event
*event
,
4436 struct perf_mmap_event
*mmap_event
)
4438 struct perf_output_handle handle
;
4439 struct perf_sample_data sample
;
4440 int size
= mmap_event
->event_id
.header
.size
;
4443 perf_event_header__init_id(&mmap_event
->event_id
.header
, &sample
, event
);
4444 ret
= perf_output_begin(&handle
, event
,
4445 mmap_event
->event_id
.header
.size
);
4449 mmap_event
->event_id
.pid
= perf_event_pid(event
, current
);
4450 mmap_event
->event_id
.tid
= perf_event_tid(event
, current
);
4452 perf_output_put(&handle
, mmap_event
->event_id
);
4453 __output_copy(&handle
, mmap_event
->file_name
,
4454 mmap_event
->file_size
);
4456 perf_event__output_id_sample(event
, &handle
, &sample
);
4458 perf_output_end(&handle
);
4460 mmap_event
->event_id
.header
.size
= size
;
4463 static int perf_event_mmap_match(struct perf_event
*event
,
4464 struct perf_mmap_event
*mmap_event
,
4467 if (event
->state
< PERF_EVENT_STATE_INACTIVE
)
4470 if (!event_filter_match(event
))
4473 if ((!executable
&& event
->attr
.mmap_data
) ||
4474 (executable
&& event
->attr
.mmap
))
4480 static void perf_event_mmap_ctx(struct perf_event_context
*ctx
,
4481 struct perf_mmap_event
*mmap_event
,
4484 struct perf_event
*event
;
4486 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
4487 if (perf_event_mmap_match(event
, mmap_event
, executable
))
4488 perf_event_mmap_output(event
, mmap_event
);
4492 static void perf_event_mmap_event(struct perf_mmap_event
*mmap_event
)
4494 struct perf_cpu_context
*cpuctx
;
4495 struct perf_event_context
*ctx
;
4496 struct vm_area_struct
*vma
= mmap_event
->vma
;
4497 struct file
*file
= vma
->vm_file
;
4505 memset(tmp
, 0, sizeof(tmp
));
4509 * d_path works from the end of the rb backwards, so we
4510 * need to add enough zero bytes after the string to handle
4511 * the 64bit alignment we do later.
4513 buf
= kzalloc(PATH_MAX
+ sizeof(u64
), GFP_KERNEL
);
4515 name
= strncpy(tmp
, "//enomem", sizeof(tmp
));
4518 name
= d_path(&file
->f_path
, buf
, PATH_MAX
);
4520 name
= strncpy(tmp
, "//toolong", sizeof(tmp
));
4524 if (arch_vma_name(mmap_event
->vma
)) {
4525 name
= strncpy(tmp
, arch_vma_name(mmap_event
->vma
),
4531 name
= strncpy(tmp
, "[vdso]", sizeof(tmp
));
4533 } else if (vma
->vm_start
<= vma
->vm_mm
->start_brk
&&
4534 vma
->vm_end
>= vma
->vm_mm
->brk
) {
4535 name
= strncpy(tmp
, "[heap]", sizeof(tmp
));
4537 } else if (vma
->vm_start
<= vma
->vm_mm
->start_stack
&&
4538 vma
->vm_end
>= vma
->vm_mm
->start_stack
) {
4539 name
= strncpy(tmp
, "[stack]", sizeof(tmp
));
4543 name
= strncpy(tmp
, "//anon", sizeof(tmp
));
4548 size
= ALIGN(strlen(name
)+1, sizeof(u64
));
4550 mmap_event
->file_name
= name
;
4551 mmap_event
->file_size
= size
;
4553 mmap_event
->event_id
.header
.size
= sizeof(mmap_event
->event_id
) + size
;
4556 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
4557 cpuctx
= get_cpu_ptr(pmu
->pmu_cpu_context
);
4558 if (cpuctx
->active_pmu
!= pmu
)
4560 perf_event_mmap_ctx(&cpuctx
->ctx
, mmap_event
,
4561 vma
->vm_flags
& VM_EXEC
);
4563 ctxn
= pmu
->task_ctx_nr
;
4567 ctx
= rcu_dereference(current
->perf_event_ctxp
[ctxn
]);
4569 perf_event_mmap_ctx(ctx
, mmap_event
,
4570 vma
->vm_flags
& VM_EXEC
);
4573 put_cpu_ptr(pmu
->pmu_cpu_context
);
4580 void perf_event_mmap(struct vm_area_struct
*vma
)
4582 struct perf_mmap_event mmap_event
;
4584 if (!atomic_read(&nr_mmap_events
))
4587 mmap_event
= (struct perf_mmap_event
){
4593 .type
= PERF_RECORD_MMAP
,
4594 .misc
= PERF_RECORD_MISC_USER
,
4599 .start
= vma
->vm_start
,
4600 .len
= vma
->vm_end
- vma
->vm_start
,
4601 .pgoff
= (u64
)vma
->vm_pgoff
<< PAGE_SHIFT
,
4605 perf_event_mmap_event(&mmap_event
);
4609 * IRQ throttle logging
4612 static void perf_log_throttle(struct perf_event
*event
, int enable
)
4614 struct perf_output_handle handle
;
4615 struct perf_sample_data sample
;
4619 struct perf_event_header header
;
4623 } throttle_event
= {
4625 .type
= PERF_RECORD_THROTTLE
,
4627 .size
= sizeof(throttle_event
),
4629 .time
= perf_clock(),
4630 .id
= primary_event_id(event
),
4631 .stream_id
= event
->id
,
4635 throttle_event
.header
.type
= PERF_RECORD_UNTHROTTLE
;
4637 perf_event_header__init_id(&throttle_event
.header
, &sample
, event
);
4639 ret
= perf_output_begin(&handle
, event
,
4640 throttle_event
.header
.size
);
4644 perf_output_put(&handle
, throttle_event
);
4645 perf_event__output_id_sample(event
, &handle
, &sample
);
4646 perf_output_end(&handle
);
4650 * Generic event overflow handling, sampling.
4653 static int __perf_event_overflow(struct perf_event
*event
,
4654 int throttle
, struct perf_sample_data
*data
,
4655 struct pt_regs
*regs
)
4657 int events
= atomic_read(&event
->event_limit
);
4658 struct hw_perf_event
*hwc
= &event
->hw
;
4663 * Non-sampling counters might still use the PMI to fold short
4664 * hardware counters, ignore those.
4666 if (unlikely(!is_sampling_event(event
)))
4669 seq
= __this_cpu_read(perf_throttled_seq
);
4670 if (seq
!= hwc
->interrupts_seq
) {
4671 hwc
->interrupts_seq
= seq
;
4672 hwc
->interrupts
= 1;
4675 if (unlikely(throttle
4676 && hwc
->interrupts
>= max_samples_per_tick
)) {
4677 __this_cpu_inc(perf_throttled_count
);
4678 hwc
->interrupts
= MAX_INTERRUPTS
;
4679 perf_log_throttle(event
, 0);
4684 if (event
->attr
.freq
) {
4685 u64 now
= perf_clock();
4686 s64 delta
= now
- hwc
->freq_time_stamp
;
4688 hwc
->freq_time_stamp
= now
;
4690 if (delta
> 0 && delta
< 2*TICK_NSEC
)
4691 perf_adjust_period(event
, delta
, hwc
->last_period
, true);
4695 * XXX event_limit might not quite work as expected on inherited
4699 event
->pending_kill
= POLL_IN
;
4700 if (events
&& atomic_dec_and_test(&event
->event_limit
)) {
4702 event
->pending_kill
= POLL_HUP
;
4703 event
->pending_disable
= 1;
4704 irq_work_queue(&event
->pending
);
4707 if (event
->overflow_handler
)
4708 event
->overflow_handler(event
, data
, regs
);
4710 perf_event_output(event
, data
, regs
);
4712 if (event
->fasync
&& event
->pending_kill
) {
4713 event
->pending_wakeup
= 1;
4714 irq_work_queue(&event
->pending
);
4720 int perf_event_overflow(struct perf_event
*event
,
4721 struct perf_sample_data
*data
,
4722 struct pt_regs
*regs
)
4724 return __perf_event_overflow(event
, 1, data
, regs
);
4728 * Generic software event infrastructure
4731 struct swevent_htable
{
4732 struct swevent_hlist
*swevent_hlist
;
4733 struct mutex hlist_mutex
;
4736 /* Recursion avoidance in each contexts */
4737 int recursion
[PERF_NR_CONTEXTS
];
4740 static DEFINE_PER_CPU(struct swevent_htable
, swevent_htable
);
4743 * We directly increment event->count and keep a second value in
4744 * event->hw.period_left to count intervals. This period event
4745 * is kept in the range [-sample_period, 0] so that we can use the
4749 static u64
perf_swevent_set_period(struct perf_event
*event
)
4751 struct hw_perf_event
*hwc
= &event
->hw
;
4752 u64 period
= hwc
->last_period
;
4756 hwc
->last_period
= hwc
->sample_period
;
4759 old
= val
= local64_read(&hwc
->period_left
);
4763 nr
= div64_u64(period
+ val
, period
);
4764 offset
= nr
* period
;
4766 if (local64_cmpxchg(&hwc
->period_left
, old
, val
) != old
)
4772 static void perf_swevent_overflow(struct perf_event
*event
, u64 overflow
,
4773 struct perf_sample_data
*data
,
4774 struct pt_regs
*regs
)
4776 struct hw_perf_event
*hwc
= &event
->hw
;
4780 overflow
= perf_swevent_set_period(event
);
4782 if (hwc
->interrupts
== MAX_INTERRUPTS
)
4785 for (; overflow
; overflow
--) {
4786 if (__perf_event_overflow(event
, throttle
,
4789 * We inhibit the overflow from happening when
4790 * hwc->interrupts == MAX_INTERRUPTS.
4798 static void perf_swevent_event(struct perf_event
*event
, u64 nr
,
4799 struct perf_sample_data
*data
,
4800 struct pt_regs
*regs
)
4802 struct hw_perf_event
*hwc
= &event
->hw
;
4804 local64_add(nr
, &event
->count
);
4809 if (!is_sampling_event(event
))
4812 if ((event
->attr
.sample_type
& PERF_SAMPLE_PERIOD
) && !event
->attr
.freq
) {
4814 return perf_swevent_overflow(event
, 1, data
, regs
);
4816 data
->period
= event
->hw
.last_period
;
4818 if (nr
== 1 && hwc
->sample_period
== 1 && !event
->attr
.freq
)
4819 return perf_swevent_overflow(event
, 1, data
, regs
);
4821 if (local64_add_negative(nr
, &hwc
->period_left
))
4824 perf_swevent_overflow(event
, 0, data
, regs
);
4827 static int perf_exclude_event(struct perf_event
*event
,
4828 struct pt_regs
*regs
)
4830 if (event
->hw
.state
& PERF_HES_STOPPED
)
4834 if (event
->attr
.exclude_user
&& user_mode(regs
))
4837 if (event
->attr
.exclude_kernel
&& !user_mode(regs
))
4844 static int perf_swevent_match(struct perf_event
*event
,
4845 enum perf_type_id type
,
4847 struct perf_sample_data
*data
,
4848 struct pt_regs
*regs
)
4850 if (event
->attr
.type
!= type
)
4853 if (event
->attr
.config
!= event_id
)
4856 if (perf_exclude_event(event
, regs
))
4862 static inline u64
swevent_hash(u64 type
, u32 event_id
)
4864 u64 val
= event_id
| (type
<< 32);
4866 return hash_64(val
, SWEVENT_HLIST_BITS
);
4869 static inline struct hlist_head
*
4870 __find_swevent_head(struct swevent_hlist
*hlist
, u64 type
, u32 event_id
)
4872 u64 hash
= swevent_hash(type
, event_id
);
4874 return &hlist
->heads
[hash
];
4877 /* For the read side: events when they trigger */
4878 static inline struct hlist_head
*
4879 find_swevent_head_rcu(struct swevent_htable
*swhash
, u64 type
, u32 event_id
)
4881 struct swevent_hlist
*hlist
;
4883 hlist
= rcu_dereference(swhash
->swevent_hlist
);
4887 return __find_swevent_head(hlist
, type
, event_id
);
4890 /* For the event head insertion and removal in the hlist */
4891 static inline struct hlist_head
*
4892 find_swevent_head(struct swevent_htable
*swhash
, struct perf_event
*event
)
4894 struct swevent_hlist
*hlist
;
4895 u32 event_id
= event
->attr
.config
;
4896 u64 type
= event
->attr
.type
;
4899 * Event scheduling is always serialized against hlist allocation
4900 * and release. Which makes the protected version suitable here.
4901 * The context lock guarantees that.
4903 hlist
= rcu_dereference_protected(swhash
->swevent_hlist
,
4904 lockdep_is_held(&event
->ctx
->lock
));
4908 return __find_swevent_head(hlist
, type
, event_id
);
4911 static void do_perf_sw_event(enum perf_type_id type
, u32 event_id
,
4913 struct perf_sample_data
*data
,
4914 struct pt_regs
*regs
)
4916 struct swevent_htable
*swhash
= &__get_cpu_var(swevent_htable
);
4917 struct perf_event
*event
;
4918 struct hlist_node
*node
;
4919 struct hlist_head
*head
;
4922 head
= find_swevent_head_rcu(swhash
, type
, event_id
);
4926 hlist_for_each_entry_rcu(event
, node
, head
, hlist_entry
) {
4927 if (perf_swevent_match(event
, type
, event_id
, data
, regs
))
4928 perf_swevent_event(event
, nr
, data
, regs
);
4934 int perf_swevent_get_recursion_context(void)
4936 struct swevent_htable
*swhash
= &__get_cpu_var(swevent_htable
);
4938 return get_recursion_context(swhash
->recursion
);
4940 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context
);
4942 inline void perf_swevent_put_recursion_context(int rctx
)
4944 struct swevent_htable
*swhash
= &__get_cpu_var(swevent_htable
);
4946 put_recursion_context(swhash
->recursion
, rctx
);
4949 void __perf_sw_event(u32 event_id
, u64 nr
, struct pt_regs
*regs
, u64 addr
)
4951 struct perf_sample_data data
;
4954 preempt_disable_notrace();
4955 rctx
= perf_swevent_get_recursion_context();
4959 perf_sample_data_init(&data
, addr
, 0);
4961 do_perf_sw_event(PERF_TYPE_SOFTWARE
, event_id
, nr
, &data
, regs
);
4963 perf_swevent_put_recursion_context(rctx
);
4964 preempt_enable_notrace();
4967 static void perf_swevent_read(struct perf_event
*event
)
4971 static int perf_swevent_add(struct perf_event
*event
, int flags
)
4973 struct swevent_htable
*swhash
= &__get_cpu_var(swevent_htable
);
4974 struct hw_perf_event
*hwc
= &event
->hw
;
4975 struct hlist_head
*head
;
4977 if (is_sampling_event(event
)) {
4978 hwc
->last_period
= hwc
->sample_period
;
4979 perf_swevent_set_period(event
);
4982 hwc
->state
= !(flags
& PERF_EF_START
);
4984 head
= find_swevent_head(swhash
, event
);
4985 if (WARN_ON_ONCE(!head
))
4988 hlist_add_head_rcu(&event
->hlist_entry
, head
);
4993 static void perf_swevent_del(struct perf_event
*event
, int flags
)
4995 hlist_del_rcu(&event
->hlist_entry
);
4998 static void perf_swevent_start(struct perf_event
*event
, int flags
)
5000 event
->hw
.state
= 0;
5003 static void perf_swevent_stop(struct perf_event
*event
, int flags
)
5005 event
->hw
.state
= PERF_HES_STOPPED
;
5008 /* Deref the hlist from the update side */
5009 static inline struct swevent_hlist
*
5010 swevent_hlist_deref(struct swevent_htable
*swhash
)
5012 return rcu_dereference_protected(swhash
->swevent_hlist
,
5013 lockdep_is_held(&swhash
->hlist_mutex
));
5016 static void swevent_hlist_release(struct swevent_htable
*swhash
)
5018 struct swevent_hlist
*hlist
= swevent_hlist_deref(swhash
);
5023 rcu_assign_pointer(swhash
->swevent_hlist
, NULL
);
5024 kfree_rcu(hlist
, rcu_head
);
5027 static void swevent_hlist_put_cpu(struct perf_event
*event
, int cpu
)
5029 struct swevent_htable
*swhash
= &per_cpu(swevent_htable
, cpu
);
5031 mutex_lock(&swhash
->hlist_mutex
);
5033 if (!--swhash
->hlist_refcount
)
5034 swevent_hlist_release(swhash
);
5036 mutex_unlock(&swhash
->hlist_mutex
);
5039 static void swevent_hlist_put(struct perf_event
*event
)
5043 if (event
->cpu
!= -1) {
5044 swevent_hlist_put_cpu(event
, event
->cpu
);
5048 for_each_possible_cpu(cpu
)
5049 swevent_hlist_put_cpu(event
, cpu
);
5052 static int swevent_hlist_get_cpu(struct perf_event
*event
, int cpu
)
5054 struct swevent_htable
*swhash
= &per_cpu(swevent_htable
, cpu
);
5057 mutex_lock(&swhash
->hlist_mutex
);
5059 if (!swevent_hlist_deref(swhash
) && cpu_online(cpu
)) {
5060 struct swevent_hlist
*hlist
;
5062 hlist
= kzalloc(sizeof(*hlist
), GFP_KERNEL
);
5067 rcu_assign_pointer(swhash
->swevent_hlist
, hlist
);
5069 swhash
->hlist_refcount
++;
5071 mutex_unlock(&swhash
->hlist_mutex
);
5076 static int swevent_hlist_get(struct perf_event
*event
)
5079 int cpu
, failed_cpu
;
5081 if (event
->cpu
!= -1)
5082 return swevent_hlist_get_cpu(event
, event
->cpu
);
5085 for_each_possible_cpu(cpu
) {
5086 err
= swevent_hlist_get_cpu(event
, cpu
);
5096 for_each_possible_cpu(cpu
) {
5097 if (cpu
== failed_cpu
)
5099 swevent_hlist_put_cpu(event
, cpu
);
5106 struct static_key perf_swevent_enabled
[PERF_COUNT_SW_MAX
];
5108 static void sw_perf_event_destroy(struct perf_event
*event
)
5110 u64 event_id
= event
->attr
.config
;
5112 WARN_ON(event
->parent
);
5114 static_key_slow_dec(&perf_swevent_enabled
[event_id
]);
5115 swevent_hlist_put(event
);
5118 static int perf_swevent_init(struct perf_event
*event
)
5120 int event_id
= event
->attr
.config
;
5122 if (event
->attr
.type
!= PERF_TYPE_SOFTWARE
)
5126 * no branch sampling for software events
5128 if (has_branch_stack(event
))
5132 case PERF_COUNT_SW_CPU_CLOCK
:
5133 case PERF_COUNT_SW_TASK_CLOCK
:
5140 if (event_id
>= PERF_COUNT_SW_MAX
)
5143 if (!event
->parent
) {
5146 err
= swevent_hlist_get(event
);
5150 static_key_slow_inc(&perf_swevent_enabled
[event_id
]);
5151 event
->destroy
= sw_perf_event_destroy
;
5157 static int perf_swevent_event_idx(struct perf_event
*event
)
5162 static struct pmu perf_swevent
= {
5163 .task_ctx_nr
= perf_sw_context
,
5165 .event_init
= perf_swevent_init
,
5166 .add
= perf_swevent_add
,
5167 .del
= perf_swevent_del
,
5168 .start
= perf_swevent_start
,
5169 .stop
= perf_swevent_stop
,
5170 .read
= perf_swevent_read
,
5172 .event_idx
= perf_swevent_event_idx
,
5175 #ifdef CONFIG_EVENT_TRACING
5177 static int perf_tp_filter_match(struct perf_event
*event
,
5178 struct perf_sample_data
*data
)
5180 void *record
= data
->raw
->data
;
5182 if (likely(!event
->filter
) || filter_match_preds(event
->filter
, record
))
5187 static int perf_tp_event_match(struct perf_event
*event
,
5188 struct perf_sample_data
*data
,
5189 struct pt_regs
*regs
)
5191 if (event
->hw
.state
& PERF_HES_STOPPED
)
5194 * All tracepoints are from kernel-space.
5196 if (event
->attr
.exclude_kernel
)
5199 if (!perf_tp_filter_match(event
, data
))
5205 void perf_tp_event(u64 addr
, u64 count
, void *record
, int entry_size
,
5206 struct pt_regs
*regs
, struct hlist_head
*head
, int rctx
)
5208 struct perf_sample_data data
;
5209 struct perf_event
*event
;
5210 struct hlist_node
*node
;
5212 struct perf_raw_record raw
= {
5217 perf_sample_data_init(&data
, addr
, 0);
5220 hlist_for_each_entry_rcu(event
, node
, head
, hlist_entry
) {
5221 if (perf_tp_event_match(event
, &data
, regs
))
5222 perf_swevent_event(event
, count
, &data
, regs
);
5225 perf_swevent_put_recursion_context(rctx
);
5227 EXPORT_SYMBOL_GPL(perf_tp_event
);
5229 static void tp_perf_event_destroy(struct perf_event
*event
)
5231 perf_trace_destroy(event
);
5234 static int perf_tp_event_init(struct perf_event
*event
)
5238 if (event
->attr
.type
!= PERF_TYPE_TRACEPOINT
)
5242 * no branch sampling for tracepoint events
5244 if (has_branch_stack(event
))
5247 err
= perf_trace_init(event
);
5251 event
->destroy
= tp_perf_event_destroy
;
5256 static struct pmu perf_tracepoint
= {
5257 .task_ctx_nr
= perf_sw_context
,
5259 .event_init
= perf_tp_event_init
,
5260 .add
= perf_trace_add
,
5261 .del
= perf_trace_del
,
5262 .start
= perf_swevent_start
,
5263 .stop
= perf_swevent_stop
,
5264 .read
= perf_swevent_read
,
5266 .event_idx
= perf_swevent_event_idx
,
5269 static inline void perf_tp_register(void)
5271 perf_pmu_register(&perf_tracepoint
, "tracepoint", PERF_TYPE_TRACEPOINT
);
5274 static int perf_event_set_filter(struct perf_event
*event
, void __user
*arg
)
5279 if (event
->attr
.type
!= PERF_TYPE_TRACEPOINT
)
5282 filter_str
= strndup_user(arg
, PAGE_SIZE
);
5283 if (IS_ERR(filter_str
))
5284 return PTR_ERR(filter_str
);
5286 ret
= ftrace_profile_set_filter(event
, event
->attr
.config
, filter_str
);
5292 static void perf_event_free_filter(struct perf_event
*event
)
5294 ftrace_profile_free_filter(event
);
5299 static inline void perf_tp_register(void)
5303 static int perf_event_set_filter(struct perf_event
*event
, void __user
*arg
)
5308 static void perf_event_free_filter(struct perf_event
*event
)
5312 #endif /* CONFIG_EVENT_TRACING */
5314 #ifdef CONFIG_HAVE_HW_BREAKPOINT
5315 void perf_bp_event(struct perf_event
*bp
, void *data
)
5317 struct perf_sample_data sample
;
5318 struct pt_regs
*regs
= data
;
5320 perf_sample_data_init(&sample
, bp
->attr
.bp_addr
, 0);
5322 if (!bp
->hw
.state
&& !perf_exclude_event(bp
, regs
))
5323 perf_swevent_event(bp
, 1, &sample
, regs
);
5328 * hrtimer based swevent callback
5331 static enum hrtimer_restart
perf_swevent_hrtimer(struct hrtimer
*hrtimer
)
5333 enum hrtimer_restart ret
= HRTIMER_RESTART
;
5334 struct perf_sample_data data
;
5335 struct pt_regs
*regs
;
5336 struct perf_event
*event
;
5339 event
= container_of(hrtimer
, struct perf_event
, hw
.hrtimer
);
5341 if (event
->state
!= PERF_EVENT_STATE_ACTIVE
)
5342 return HRTIMER_NORESTART
;
5344 event
->pmu
->read(event
);
5346 perf_sample_data_init(&data
, 0, event
->hw
.last_period
);
5347 regs
= get_irq_regs();
5349 if (regs
&& !perf_exclude_event(event
, regs
)) {
5350 if (!(event
->attr
.exclude_idle
&& is_idle_task(current
)))
5351 if (__perf_event_overflow(event
, 1, &data
, regs
))
5352 ret
= HRTIMER_NORESTART
;
5355 period
= max_t(u64
, 10000, event
->hw
.sample_period
);
5356 hrtimer_forward_now(hrtimer
, ns_to_ktime(period
));
5361 static void perf_swevent_start_hrtimer(struct perf_event
*event
)
5363 struct hw_perf_event
*hwc
= &event
->hw
;
5366 if (!is_sampling_event(event
))
5369 period
= local64_read(&hwc
->period_left
);
5374 local64_set(&hwc
->period_left
, 0);
5376 period
= max_t(u64
, 10000, hwc
->sample_period
);
5378 __hrtimer_start_range_ns(&hwc
->hrtimer
,
5379 ns_to_ktime(period
), 0,
5380 HRTIMER_MODE_REL_PINNED
, 0);
5383 static void perf_swevent_cancel_hrtimer(struct perf_event
*event
)
5385 struct hw_perf_event
*hwc
= &event
->hw
;
5387 if (is_sampling_event(event
)) {
5388 ktime_t remaining
= hrtimer_get_remaining(&hwc
->hrtimer
);
5389 local64_set(&hwc
->period_left
, ktime_to_ns(remaining
));
5391 hrtimer_cancel(&hwc
->hrtimer
);
5395 static void perf_swevent_init_hrtimer(struct perf_event
*event
)
5397 struct hw_perf_event
*hwc
= &event
->hw
;
5399 if (!is_sampling_event(event
))
5402 hrtimer_init(&hwc
->hrtimer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
5403 hwc
->hrtimer
.function
= perf_swevent_hrtimer
;
5406 * Since hrtimers have a fixed rate, we can do a static freq->period
5407 * mapping and avoid the whole period adjust feedback stuff.
5409 if (event
->attr
.freq
) {
5410 long freq
= event
->attr
.sample_freq
;
5412 event
->attr
.sample_period
= NSEC_PER_SEC
/ freq
;
5413 hwc
->sample_period
= event
->attr
.sample_period
;
5414 local64_set(&hwc
->period_left
, hwc
->sample_period
);
5415 event
->attr
.freq
= 0;
5420 * Software event: cpu wall time clock
5423 static void cpu_clock_event_update(struct perf_event
*event
)
5428 now
= local_clock();
5429 prev
= local64_xchg(&event
->hw
.prev_count
, now
);
5430 local64_add(now
- prev
, &event
->count
);
5433 static void cpu_clock_event_start(struct perf_event
*event
, int flags
)
5435 local64_set(&event
->hw
.prev_count
, local_clock());
5436 perf_swevent_start_hrtimer(event
);
5439 static void cpu_clock_event_stop(struct perf_event
*event
, int flags
)
5441 perf_swevent_cancel_hrtimer(event
);
5442 cpu_clock_event_update(event
);
5445 static int cpu_clock_event_add(struct perf_event
*event
, int flags
)
5447 if (flags
& PERF_EF_START
)
5448 cpu_clock_event_start(event
, flags
);
5453 static void cpu_clock_event_del(struct perf_event
*event
, int flags
)
5455 cpu_clock_event_stop(event
, flags
);
5458 static void cpu_clock_event_read(struct perf_event
*event
)
5460 cpu_clock_event_update(event
);
5463 static int cpu_clock_event_init(struct perf_event
*event
)
5465 if (event
->attr
.type
!= PERF_TYPE_SOFTWARE
)
5468 if (event
->attr
.config
!= PERF_COUNT_SW_CPU_CLOCK
)
5472 * no branch sampling for software events
5474 if (has_branch_stack(event
))
5477 perf_swevent_init_hrtimer(event
);
5482 static struct pmu perf_cpu_clock
= {
5483 .task_ctx_nr
= perf_sw_context
,
5485 .event_init
= cpu_clock_event_init
,
5486 .add
= cpu_clock_event_add
,
5487 .del
= cpu_clock_event_del
,
5488 .start
= cpu_clock_event_start
,
5489 .stop
= cpu_clock_event_stop
,
5490 .read
= cpu_clock_event_read
,
5492 .event_idx
= perf_swevent_event_idx
,
5496 * Software event: task time clock
5499 static void task_clock_event_update(struct perf_event
*event
, u64 now
)
5504 prev
= local64_xchg(&event
->hw
.prev_count
, now
);
5506 local64_add(delta
, &event
->count
);
5509 static void task_clock_event_start(struct perf_event
*event
, int flags
)
5511 local64_set(&event
->hw
.prev_count
, event
->ctx
->time
);
5512 perf_swevent_start_hrtimer(event
);
5515 static void task_clock_event_stop(struct perf_event
*event
, int flags
)
5517 perf_swevent_cancel_hrtimer(event
);
5518 task_clock_event_update(event
, event
->ctx
->time
);
5521 static int task_clock_event_add(struct perf_event
*event
, int flags
)
5523 if (flags
& PERF_EF_START
)
5524 task_clock_event_start(event
, flags
);
5529 static void task_clock_event_del(struct perf_event
*event
, int flags
)
5531 task_clock_event_stop(event
, PERF_EF_UPDATE
);
5534 static void task_clock_event_read(struct perf_event
*event
)
5536 u64 now
= perf_clock();
5537 u64 delta
= now
- event
->ctx
->timestamp
;
5538 u64 time
= event
->ctx
->time
+ delta
;
5540 task_clock_event_update(event
, time
);
5543 static int task_clock_event_init(struct perf_event
*event
)
5545 if (event
->attr
.type
!= PERF_TYPE_SOFTWARE
)
5548 if (event
->attr
.config
!= PERF_COUNT_SW_TASK_CLOCK
)
5552 * no branch sampling for software events
5554 if (has_branch_stack(event
))
5557 perf_swevent_init_hrtimer(event
);
5562 static struct pmu perf_task_clock
= {
5563 .task_ctx_nr
= perf_sw_context
,
5565 .event_init
= task_clock_event_init
,
5566 .add
= task_clock_event_add
,
5567 .del
= task_clock_event_del
,
5568 .start
= task_clock_event_start
,
5569 .stop
= task_clock_event_stop
,
5570 .read
= task_clock_event_read
,
5572 .event_idx
= perf_swevent_event_idx
,
5575 static void perf_pmu_nop_void(struct pmu
*pmu
)
5579 static int perf_pmu_nop_int(struct pmu
*pmu
)
5584 static void perf_pmu_start_txn(struct pmu
*pmu
)
5586 perf_pmu_disable(pmu
);
5589 static int perf_pmu_commit_txn(struct pmu
*pmu
)
5591 perf_pmu_enable(pmu
);
5595 static void perf_pmu_cancel_txn(struct pmu
*pmu
)
5597 perf_pmu_enable(pmu
);
5600 static int perf_event_idx_default(struct perf_event
*event
)
5602 return event
->hw
.idx
+ 1;
5606 * Ensures all contexts with the same task_ctx_nr have the same
5607 * pmu_cpu_context too.
5609 static void *find_pmu_context(int ctxn
)
5616 list_for_each_entry(pmu
, &pmus
, entry
) {
5617 if (pmu
->task_ctx_nr
== ctxn
)
5618 return pmu
->pmu_cpu_context
;
5624 static void update_pmu_context(struct pmu
*pmu
, struct pmu
*old_pmu
)
5628 for_each_possible_cpu(cpu
) {
5629 struct perf_cpu_context
*cpuctx
;
5631 cpuctx
= per_cpu_ptr(pmu
->pmu_cpu_context
, cpu
);
5633 if (cpuctx
->active_pmu
== old_pmu
)
5634 cpuctx
->active_pmu
= pmu
;
5638 static void free_pmu_context(struct pmu
*pmu
)
5642 mutex_lock(&pmus_lock
);
5644 * Like a real lame refcount.
5646 list_for_each_entry(i
, &pmus
, entry
) {
5647 if (i
->pmu_cpu_context
== pmu
->pmu_cpu_context
) {
5648 update_pmu_context(i
, pmu
);
5653 free_percpu(pmu
->pmu_cpu_context
);
5655 mutex_unlock(&pmus_lock
);
5657 static struct idr pmu_idr
;
5660 type_show(struct device
*dev
, struct device_attribute
*attr
, char *page
)
5662 struct pmu
*pmu
= dev_get_drvdata(dev
);
5664 return snprintf(page
, PAGE_SIZE
-1, "%d\n", pmu
->type
);
5667 static struct device_attribute pmu_dev_attrs
[] = {
5672 static int pmu_bus_running
;
5673 static struct bus_type pmu_bus
= {
5674 .name
= "event_source",
5675 .dev_attrs
= pmu_dev_attrs
,
5678 static void pmu_dev_release(struct device
*dev
)
5683 static int pmu_dev_alloc(struct pmu
*pmu
)
5687 pmu
->dev
= kzalloc(sizeof(struct device
), GFP_KERNEL
);
5691 pmu
->dev
->groups
= pmu
->attr_groups
;
5692 device_initialize(pmu
->dev
);
5693 ret
= dev_set_name(pmu
->dev
, "%s", pmu
->name
);
5697 dev_set_drvdata(pmu
->dev
, pmu
);
5698 pmu
->dev
->bus
= &pmu_bus
;
5699 pmu
->dev
->release
= pmu_dev_release
;
5700 ret
= device_add(pmu
->dev
);
5708 put_device(pmu
->dev
);
5712 static struct lock_class_key cpuctx_mutex
;
5713 static struct lock_class_key cpuctx_lock
;
5715 int perf_pmu_register(struct pmu
*pmu
, char *name
, int type
)
5719 mutex_lock(&pmus_lock
);
5721 pmu
->pmu_disable_count
= alloc_percpu(int);
5722 if (!pmu
->pmu_disable_count
)
5731 int err
= idr_pre_get(&pmu_idr
, GFP_KERNEL
);
5735 err
= idr_get_new_above(&pmu_idr
, pmu
, PERF_TYPE_MAX
, &type
);
5743 if (pmu_bus_running
) {
5744 ret
= pmu_dev_alloc(pmu
);
5750 pmu
->pmu_cpu_context
= find_pmu_context(pmu
->task_ctx_nr
);
5751 if (pmu
->pmu_cpu_context
)
5752 goto got_cpu_context
;
5754 pmu
->pmu_cpu_context
= alloc_percpu(struct perf_cpu_context
);
5755 if (!pmu
->pmu_cpu_context
)
5758 for_each_possible_cpu(cpu
) {
5759 struct perf_cpu_context
*cpuctx
;
5761 cpuctx
= per_cpu_ptr(pmu
->pmu_cpu_context
, cpu
);
5762 __perf_event_init_context(&cpuctx
->ctx
);
5763 lockdep_set_class(&cpuctx
->ctx
.mutex
, &cpuctx_mutex
);
5764 lockdep_set_class(&cpuctx
->ctx
.lock
, &cpuctx_lock
);
5765 cpuctx
->ctx
.type
= cpu_context
;
5766 cpuctx
->ctx
.pmu
= pmu
;
5767 cpuctx
->jiffies_interval
= 1;
5768 INIT_LIST_HEAD(&cpuctx
->rotation_list
);
5769 cpuctx
->active_pmu
= pmu
;
5773 if (!pmu
->start_txn
) {
5774 if (pmu
->pmu_enable
) {
5776 * If we have pmu_enable/pmu_disable calls, install
5777 * transaction stubs that use that to try and batch
5778 * hardware accesses.
5780 pmu
->start_txn
= perf_pmu_start_txn
;
5781 pmu
->commit_txn
= perf_pmu_commit_txn
;
5782 pmu
->cancel_txn
= perf_pmu_cancel_txn
;
5784 pmu
->start_txn
= perf_pmu_nop_void
;
5785 pmu
->commit_txn
= perf_pmu_nop_int
;
5786 pmu
->cancel_txn
= perf_pmu_nop_void
;
5790 if (!pmu
->pmu_enable
) {
5791 pmu
->pmu_enable
= perf_pmu_nop_void
;
5792 pmu
->pmu_disable
= perf_pmu_nop_void
;
5795 if (!pmu
->event_idx
)
5796 pmu
->event_idx
= perf_event_idx_default
;
5798 list_add_rcu(&pmu
->entry
, &pmus
);
5801 mutex_unlock(&pmus_lock
);
5806 device_del(pmu
->dev
);
5807 put_device(pmu
->dev
);
5810 if (pmu
->type
>= PERF_TYPE_MAX
)
5811 idr_remove(&pmu_idr
, pmu
->type
);
5814 free_percpu(pmu
->pmu_disable_count
);
5818 void perf_pmu_unregister(struct pmu
*pmu
)
5820 mutex_lock(&pmus_lock
);
5821 list_del_rcu(&pmu
->entry
);
5822 mutex_unlock(&pmus_lock
);
5825 * We dereference the pmu list under both SRCU and regular RCU, so
5826 * synchronize against both of those.
5828 synchronize_srcu(&pmus_srcu
);
5831 free_percpu(pmu
->pmu_disable_count
);
5832 if (pmu
->type
>= PERF_TYPE_MAX
)
5833 idr_remove(&pmu_idr
, pmu
->type
);
5834 device_del(pmu
->dev
);
5835 put_device(pmu
->dev
);
5836 free_pmu_context(pmu
);
5839 struct pmu
*perf_init_event(struct perf_event
*event
)
5841 struct pmu
*pmu
= NULL
;
5845 idx
= srcu_read_lock(&pmus_srcu
);
5848 pmu
= idr_find(&pmu_idr
, event
->attr
.type
);
5852 ret
= pmu
->event_init(event
);
5858 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
5860 ret
= pmu
->event_init(event
);
5864 if (ret
!= -ENOENT
) {
5869 pmu
= ERR_PTR(-ENOENT
);
5871 srcu_read_unlock(&pmus_srcu
, idx
);
5877 * Allocate and initialize a event structure
5879 static struct perf_event
*
5880 perf_event_alloc(struct perf_event_attr
*attr
, int cpu
,
5881 struct task_struct
*task
,
5882 struct perf_event
*group_leader
,
5883 struct perf_event
*parent_event
,
5884 perf_overflow_handler_t overflow_handler
,
5888 struct perf_event
*event
;
5889 struct hw_perf_event
*hwc
;
5892 if ((unsigned)cpu
>= nr_cpu_ids
) {
5893 if (!task
|| cpu
!= -1)
5894 return ERR_PTR(-EINVAL
);
5897 event
= kzalloc(sizeof(*event
), GFP_KERNEL
);
5899 return ERR_PTR(-ENOMEM
);
5902 * Single events are their own group leaders, with an
5903 * empty sibling list:
5906 group_leader
= event
;
5908 mutex_init(&event
->child_mutex
);
5909 INIT_LIST_HEAD(&event
->child_list
);
5911 INIT_LIST_HEAD(&event
->group_entry
);
5912 INIT_LIST_HEAD(&event
->event_entry
);
5913 INIT_LIST_HEAD(&event
->sibling_list
);
5914 INIT_LIST_HEAD(&event
->rb_entry
);
5916 init_waitqueue_head(&event
->waitq
);
5917 init_irq_work(&event
->pending
, perf_pending_event
);
5919 mutex_init(&event
->mmap_mutex
);
5922 event
->attr
= *attr
;
5923 event
->group_leader
= group_leader
;
5927 event
->parent
= parent_event
;
5929 event
->ns
= get_pid_ns(current
->nsproxy
->pid_ns
);
5930 event
->id
= atomic64_inc_return(&perf_event_id
);
5932 event
->state
= PERF_EVENT_STATE_INACTIVE
;
5935 event
->attach_state
= PERF_ATTACH_TASK
;
5936 #ifdef CONFIG_HAVE_HW_BREAKPOINT
5938 * hw_breakpoint is a bit difficult here..
5940 if (attr
->type
== PERF_TYPE_BREAKPOINT
)
5941 event
->hw
.bp_target
= task
;
5945 if (!overflow_handler
&& parent_event
) {
5946 overflow_handler
= parent_event
->overflow_handler
;
5947 context
= parent_event
->overflow_handler_context
;
5950 event
->overflow_handler
= overflow_handler
;
5951 event
->overflow_handler_context
= context
;
5954 event
->state
= PERF_EVENT_STATE_OFF
;
5959 hwc
->sample_period
= attr
->sample_period
;
5960 if (attr
->freq
&& attr
->sample_freq
)
5961 hwc
->sample_period
= 1;
5962 hwc
->last_period
= hwc
->sample_period
;
5964 local64_set(&hwc
->period_left
, hwc
->sample_period
);
5967 * we currently do not support PERF_FORMAT_GROUP on inherited events
5969 if (attr
->inherit
&& (attr
->read_format
& PERF_FORMAT_GROUP
))
5972 pmu
= perf_init_event(event
);
5978 else if (IS_ERR(pmu
))
5983 put_pid_ns(event
->ns
);
5985 return ERR_PTR(err
);
5988 if (!event
->parent
) {
5989 if (event
->attach_state
& PERF_ATTACH_TASK
)
5990 static_key_slow_inc(&perf_sched_events
.key
);
5991 if (event
->attr
.mmap
|| event
->attr
.mmap_data
)
5992 atomic_inc(&nr_mmap_events
);
5993 if (event
->attr
.comm
)
5994 atomic_inc(&nr_comm_events
);
5995 if (event
->attr
.task
)
5996 atomic_inc(&nr_task_events
);
5997 if (event
->attr
.sample_type
& PERF_SAMPLE_CALLCHAIN
) {
5998 err
= get_callchain_buffers();
6001 return ERR_PTR(err
);
6004 if (has_branch_stack(event
)) {
6005 static_key_slow_inc(&perf_sched_events
.key
);
6006 if (!(event
->attach_state
& PERF_ATTACH_TASK
))
6007 atomic_inc(&per_cpu(perf_branch_stack_events
,
6015 static int perf_copy_attr(struct perf_event_attr __user
*uattr
,
6016 struct perf_event_attr
*attr
)
6021 if (!access_ok(VERIFY_WRITE
, uattr
, PERF_ATTR_SIZE_VER0
))
6025 * zero the full structure, so that a short copy will be nice.
6027 memset(attr
, 0, sizeof(*attr
));
6029 ret
= get_user(size
, &uattr
->size
);
6033 if (size
> PAGE_SIZE
) /* silly large */
6036 if (!size
) /* abi compat */
6037 size
= PERF_ATTR_SIZE_VER0
;
6039 if (size
< PERF_ATTR_SIZE_VER0
)
6043 * If we're handed a bigger struct than we know of,
6044 * ensure all the unknown bits are 0 - i.e. new
6045 * user-space does not rely on any kernel feature
6046 * extensions we dont know about yet.
6048 if (size
> sizeof(*attr
)) {
6049 unsigned char __user
*addr
;
6050 unsigned char __user
*end
;
6053 addr
= (void __user
*)uattr
+ sizeof(*attr
);
6054 end
= (void __user
*)uattr
+ size
;
6056 for (; addr
< end
; addr
++) {
6057 ret
= get_user(val
, addr
);
6063 size
= sizeof(*attr
);
6066 ret
= copy_from_user(attr
, uattr
, size
);
6070 if (attr
->__reserved_1
)
6073 if (attr
->sample_type
& ~(PERF_SAMPLE_MAX
-1))
6076 if (attr
->read_format
& ~(PERF_FORMAT_MAX
-1))
6079 if (attr
->sample_type
& PERF_SAMPLE_BRANCH_STACK
) {
6080 u64 mask
= attr
->branch_sample_type
;
6082 /* only using defined bits */
6083 if (mask
& ~(PERF_SAMPLE_BRANCH_MAX
-1))
6086 /* at least one branch bit must be set */
6087 if (!(mask
& ~PERF_SAMPLE_BRANCH_PLM_ALL
))
6090 /* kernel level capture: check permissions */
6091 if ((mask
& PERF_SAMPLE_BRANCH_PERM_PLM
)
6092 && perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN
))
6095 /* propagate priv level, when not set for branch */
6096 if (!(mask
& PERF_SAMPLE_BRANCH_PLM_ALL
)) {
6098 /* exclude_kernel checked on syscall entry */
6099 if (!attr
->exclude_kernel
)
6100 mask
|= PERF_SAMPLE_BRANCH_KERNEL
;
6102 if (!attr
->exclude_user
)
6103 mask
|= PERF_SAMPLE_BRANCH_USER
;
6105 if (!attr
->exclude_hv
)
6106 mask
|= PERF_SAMPLE_BRANCH_HV
;
6108 * adjust user setting (for HW filter setup)
6110 attr
->branch_sample_type
= mask
;
6117 put_user(sizeof(*attr
), &uattr
->size
);
6123 perf_event_set_output(struct perf_event
*event
, struct perf_event
*output_event
)
6125 struct ring_buffer
*rb
= NULL
, *old_rb
= NULL
;
6131 /* don't allow circular references */
6132 if (event
== output_event
)
6136 * Don't allow cross-cpu buffers
6138 if (output_event
->cpu
!= event
->cpu
)
6142 * If its not a per-cpu rb, it must be the same task.
6144 if (output_event
->cpu
== -1 && output_event
->ctx
!= event
->ctx
)
6148 mutex_lock(&event
->mmap_mutex
);
6149 /* Can't redirect output if we've got an active mmap() */
6150 if (atomic_read(&event
->mmap_count
))
6154 /* get the rb we want to redirect to */
6155 rb
= ring_buffer_get(output_event
);
6161 rcu_assign_pointer(event
->rb
, rb
);
6163 ring_buffer_detach(event
, old_rb
);
6166 mutex_unlock(&event
->mmap_mutex
);
6169 ring_buffer_put(old_rb
);
6175 * sys_perf_event_open - open a performance event, associate it to a task/cpu
6177 * @attr_uptr: event_id type attributes for monitoring/sampling
6180 * @group_fd: group leader event fd
6182 SYSCALL_DEFINE5(perf_event_open
,
6183 struct perf_event_attr __user
*, attr_uptr
,
6184 pid_t
, pid
, int, cpu
, int, group_fd
, unsigned long, flags
)
6186 struct perf_event
*group_leader
= NULL
, *output_event
= NULL
;
6187 struct perf_event
*event
, *sibling
;
6188 struct perf_event_attr attr
;
6189 struct perf_event_context
*ctx
;
6190 struct file
*event_file
= NULL
;
6191 struct file
*group_file
= NULL
;
6192 struct task_struct
*task
= NULL
;
6196 int fput_needed
= 0;
6199 /* for future expandability... */
6200 if (flags
& ~PERF_FLAG_ALL
)
6203 err
= perf_copy_attr(attr_uptr
, &attr
);
6207 if (!attr
.exclude_kernel
) {
6208 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN
))
6213 if (attr
.sample_freq
> sysctl_perf_event_sample_rate
)
6218 * In cgroup mode, the pid argument is used to pass the fd
6219 * opened to the cgroup directory in cgroupfs. The cpu argument
6220 * designates the cpu on which to monitor threads from that
6223 if ((flags
& PERF_FLAG_PID_CGROUP
) && (pid
== -1 || cpu
== -1))
6226 event_fd
= get_unused_fd_flags(O_RDWR
);
6230 if (group_fd
!= -1) {
6231 group_leader
= perf_fget_light(group_fd
, &fput_needed
);
6232 if (IS_ERR(group_leader
)) {
6233 err
= PTR_ERR(group_leader
);
6236 group_file
= group_leader
->filp
;
6237 if (flags
& PERF_FLAG_FD_OUTPUT
)
6238 output_event
= group_leader
;
6239 if (flags
& PERF_FLAG_FD_NO_GROUP
)
6240 group_leader
= NULL
;
6243 if (pid
!= -1 && !(flags
& PERF_FLAG_PID_CGROUP
)) {
6244 task
= find_lively_task_by_vpid(pid
);
6246 err
= PTR_ERR(task
);
6251 event
= perf_event_alloc(&attr
, cpu
, task
, group_leader
, NULL
,
6253 if (IS_ERR(event
)) {
6254 err
= PTR_ERR(event
);
6258 if (flags
& PERF_FLAG_PID_CGROUP
) {
6259 err
= perf_cgroup_connect(pid
, event
, &attr
, group_leader
);
6264 * - that has cgroup constraint on event->cpu
6265 * - that may need work on context switch
6267 atomic_inc(&per_cpu(perf_cgroup_events
, event
->cpu
));
6268 static_key_slow_inc(&perf_sched_events
.key
);
6272 * Special case software events and allow them to be part of
6273 * any hardware group.
6278 (is_software_event(event
) != is_software_event(group_leader
))) {
6279 if (is_software_event(event
)) {
6281 * If event and group_leader are not both a software
6282 * event, and event is, then group leader is not.
6284 * Allow the addition of software events to !software
6285 * groups, this is safe because software events never
6288 pmu
= group_leader
->pmu
;
6289 } else if (is_software_event(group_leader
) &&
6290 (group_leader
->group_flags
& PERF_GROUP_SOFTWARE
)) {
6292 * In case the group is a pure software group, and we
6293 * try to add a hardware event, move the whole group to
6294 * the hardware context.
6301 * Get the target context (task or percpu):
6303 ctx
= find_get_context(pmu
, task
, cpu
);
6310 put_task_struct(task
);
6315 * Look up the group leader (we will attach this event to it):
6321 * Do not allow a recursive hierarchy (this new sibling
6322 * becoming part of another group-sibling):
6324 if (group_leader
->group_leader
!= group_leader
)
6327 * Do not allow to attach to a group in a different
6328 * task or CPU context:
6331 if (group_leader
->ctx
->type
!= ctx
->type
)
6334 if (group_leader
->ctx
!= ctx
)
6339 * Only a group leader can be exclusive or pinned
6341 if (attr
.exclusive
|| attr
.pinned
)
6346 err
= perf_event_set_output(event
, output_event
);
6351 event_file
= anon_inode_getfile("[perf_event]", &perf_fops
, event
, O_RDWR
);
6352 if (IS_ERR(event_file
)) {
6353 err
= PTR_ERR(event_file
);
6358 struct perf_event_context
*gctx
= group_leader
->ctx
;
6360 mutex_lock(&gctx
->mutex
);
6361 perf_remove_from_context(group_leader
);
6362 list_for_each_entry(sibling
, &group_leader
->sibling_list
,
6364 perf_remove_from_context(sibling
);
6367 mutex_unlock(&gctx
->mutex
);
6371 event
->filp
= event_file
;
6372 WARN_ON_ONCE(ctx
->parent_ctx
);
6373 mutex_lock(&ctx
->mutex
);
6376 perf_install_in_context(ctx
, group_leader
, cpu
);
6378 list_for_each_entry(sibling
, &group_leader
->sibling_list
,
6380 perf_install_in_context(ctx
, sibling
, cpu
);
6385 perf_install_in_context(ctx
, event
, cpu
);
6387 perf_unpin_context(ctx
);
6388 mutex_unlock(&ctx
->mutex
);
6390 event
->owner
= current
;
6392 mutex_lock(¤t
->perf_event_mutex
);
6393 list_add_tail(&event
->owner_entry
, ¤t
->perf_event_list
);
6394 mutex_unlock(¤t
->perf_event_mutex
);
6397 * Precalculate sample_data sizes
6399 perf_event__header_size(event
);
6400 perf_event__id_header_size(event
);
6403 * Drop the reference on the group_event after placing the
6404 * new event on the sibling_list. This ensures destruction
6405 * of the group leader will find the pointer to itself in
6406 * perf_group_detach().
6408 fput_light(group_file
, fput_needed
);
6409 fd_install(event_fd
, event_file
);
6413 perf_unpin_context(ctx
);
6419 put_task_struct(task
);
6421 fput_light(group_file
, fput_needed
);
6423 put_unused_fd(event_fd
);
6428 * perf_event_create_kernel_counter
6430 * @attr: attributes of the counter to create
6431 * @cpu: cpu in which the counter is bound
6432 * @task: task to profile (NULL for percpu)
6435 perf_event_create_kernel_counter(struct perf_event_attr
*attr
, int cpu
,
6436 struct task_struct
*task
,
6437 perf_overflow_handler_t overflow_handler
,
6440 struct perf_event_context
*ctx
;
6441 struct perf_event
*event
;
6445 * Get the target context (task or percpu):
6448 event
= perf_event_alloc(attr
, cpu
, task
, NULL
, NULL
,
6449 overflow_handler
, context
);
6450 if (IS_ERR(event
)) {
6451 err
= PTR_ERR(event
);
6455 ctx
= find_get_context(event
->pmu
, task
, cpu
);
6462 WARN_ON_ONCE(ctx
->parent_ctx
);
6463 mutex_lock(&ctx
->mutex
);
6464 perf_install_in_context(ctx
, event
, cpu
);
6466 perf_unpin_context(ctx
);
6467 mutex_unlock(&ctx
->mutex
);
6474 return ERR_PTR(err
);
6476 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter
);
6478 static void sync_child_event(struct perf_event
*child_event
,
6479 struct task_struct
*child
)
6481 struct perf_event
*parent_event
= child_event
->parent
;
6484 if (child_event
->attr
.inherit_stat
)
6485 perf_event_read_event(child_event
, child
);
6487 child_val
= perf_event_count(child_event
);
6490 * Add back the child's count to the parent's count:
6492 atomic64_add(child_val
, &parent_event
->child_count
);
6493 atomic64_add(child_event
->total_time_enabled
,
6494 &parent_event
->child_total_time_enabled
);
6495 atomic64_add(child_event
->total_time_running
,
6496 &parent_event
->child_total_time_running
);
6499 * Remove this event from the parent's list
6501 WARN_ON_ONCE(parent_event
->ctx
->parent_ctx
);
6502 mutex_lock(&parent_event
->child_mutex
);
6503 list_del_init(&child_event
->child_list
);
6504 mutex_unlock(&parent_event
->child_mutex
);
6507 * Release the parent event, if this was the last
6510 fput(parent_event
->filp
);
6514 __perf_event_exit_task(struct perf_event
*child_event
,
6515 struct perf_event_context
*child_ctx
,
6516 struct task_struct
*child
)
6518 if (child_event
->parent
) {
6519 raw_spin_lock_irq(&child_ctx
->lock
);
6520 perf_group_detach(child_event
);
6521 raw_spin_unlock_irq(&child_ctx
->lock
);
6524 perf_remove_from_context(child_event
);
6527 * It can happen that the parent exits first, and has events
6528 * that are still around due to the child reference. These
6529 * events need to be zapped.
6531 if (child_event
->parent
) {
6532 sync_child_event(child_event
, child
);
6533 free_event(child_event
);
6537 static void perf_event_exit_task_context(struct task_struct
*child
, int ctxn
)
6539 struct perf_event
*child_event
, *tmp
;
6540 struct perf_event_context
*child_ctx
;
6541 unsigned long flags
;
6543 if (likely(!child
->perf_event_ctxp
[ctxn
])) {
6544 perf_event_task(child
, NULL
, 0);
6548 local_irq_save(flags
);
6550 * We can't reschedule here because interrupts are disabled,
6551 * and either child is current or it is a task that can't be
6552 * scheduled, so we are now safe from rescheduling changing
6555 child_ctx
= rcu_dereference_raw(child
->perf_event_ctxp
[ctxn
]);
6558 * Take the context lock here so that if find_get_context is
6559 * reading child->perf_event_ctxp, we wait until it has
6560 * incremented the context's refcount before we do put_ctx below.
6562 raw_spin_lock(&child_ctx
->lock
);
6563 task_ctx_sched_out(child_ctx
);
6564 child
->perf_event_ctxp
[ctxn
] = NULL
;
6566 * If this context is a clone; unclone it so it can't get
6567 * swapped to another process while we're removing all
6568 * the events from it.
6570 unclone_ctx(child_ctx
);
6571 update_context_time(child_ctx
);
6572 raw_spin_unlock_irqrestore(&child_ctx
->lock
, flags
);
6575 * Report the task dead after unscheduling the events so that we
6576 * won't get any samples after PERF_RECORD_EXIT. We can however still
6577 * get a few PERF_RECORD_READ events.
6579 perf_event_task(child
, child_ctx
, 0);
6582 * We can recurse on the same lock type through:
6584 * __perf_event_exit_task()
6585 * sync_child_event()
6586 * fput(parent_event->filp)
6588 * mutex_lock(&ctx->mutex)
6590 * But since its the parent context it won't be the same instance.
6592 mutex_lock(&child_ctx
->mutex
);
6595 list_for_each_entry_safe(child_event
, tmp
, &child_ctx
->pinned_groups
,
6597 __perf_event_exit_task(child_event
, child_ctx
, child
);
6599 list_for_each_entry_safe(child_event
, tmp
, &child_ctx
->flexible_groups
,
6601 __perf_event_exit_task(child_event
, child_ctx
, child
);
6604 * If the last event was a group event, it will have appended all
6605 * its siblings to the list, but we obtained 'tmp' before that which
6606 * will still point to the list head terminating the iteration.
6608 if (!list_empty(&child_ctx
->pinned_groups
) ||
6609 !list_empty(&child_ctx
->flexible_groups
))
6612 mutex_unlock(&child_ctx
->mutex
);
6618 * When a child task exits, feed back event values to parent events.
6620 void perf_event_exit_task(struct task_struct
*child
)
6622 struct perf_event
*event
, *tmp
;
6625 mutex_lock(&child
->perf_event_mutex
);
6626 list_for_each_entry_safe(event
, tmp
, &child
->perf_event_list
,
6628 list_del_init(&event
->owner_entry
);
6631 * Ensure the list deletion is visible before we clear
6632 * the owner, closes a race against perf_release() where
6633 * we need to serialize on the owner->perf_event_mutex.
6636 event
->owner
= NULL
;
6638 mutex_unlock(&child
->perf_event_mutex
);
6640 for_each_task_context_nr(ctxn
)
6641 perf_event_exit_task_context(child
, ctxn
);
6644 static void perf_free_event(struct perf_event
*event
,
6645 struct perf_event_context
*ctx
)
6647 struct perf_event
*parent
= event
->parent
;
6649 if (WARN_ON_ONCE(!parent
))
6652 mutex_lock(&parent
->child_mutex
);
6653 list_del_init(&event
->child_list
);
6654 mutex_unlock(&parent
->child_mutex
);
6658 perf_group_detach(event
);
6659 list_del_event(event
, ctx
);
6664 * free an unexposed, unused context as created by inheritance by
6665 * perf_event_init_task below, used by fork() in case of fail.
6667 void perf_event_free_task(struct task_struct
*task
)
6669 struct perf_event_context
*ctx
;
6670 struct perf_event
*event
, *tmp
;
6673 for_each_task_context_nr(ctxn
) {
6674 ctx
= task
->perf_event_ctxp
[ctxn
];
6678 mutex_lock(&ctx
->mutex
);
6680 list_for_each_entry_safe(event
, tmp
, &ctx
->pinned_groups
,
6682 perf_free_event(event
, ctx
);
6684 list_for_each_entry_safe(event
, tmp
, &ctx
->flexible_groups
,
6686 perf_free_event(event
, ctx
);
6688 if (!list_empty(&ctx
->pinned_groups
) ||
6689 !list_empty(&ctx
->flexible_groups
))
6692 mutex_unlock(&ctx
->mutex
);
6698 void perf_event_delayed_put(struct task_struct
*task
)
6702 for_each_task_context_nr(ctxn
)
6703 WARN_ON_ONCE(task
->perf_event_ctxp
[ctxn
]);
6707 * inherit a event from parent task to child task:
6709 static struct perf_event
*
6710 inherit_event(struct perf_event
*parent_event
,
6711 struct task_struct
*parent
,
6712 struct perf_event_context
*parent_ctx
,
6713 struct task_struct
*child
,
6714 struct perf_event
*group_leader
,
6715 struct perf_event_context
*child_ctx
)
6717 struct perf_event
*child_event
;
6718 unsigned long flags
;
6721 * Instead of creating recursive hierarchies of events,
6722 * we link inherited events back to the original parent,
6723 * which has a filp for sure, which we use as the reference
6726 if (parent_event
->parent
)
6727 parent_event
= parent_event
->parent
;
6729 child_event
= perf_event_alloc(&parent_event
->attr
,
6732 group_leader
, parent_event
,
6734 if (IS_ERR(child_event
))
6739 * Make the child state follow the state of the parent event,
6740 * not its attr.disabled bit. We hold the parent's mutex,
6741 * so we won't race with perf_event_{en, dis}able_family.
6743 if (parent_event
->state
>= PERF_EVENT_STATE_INACTIVE
)
6744 child_event
->state
= PERF_EVENT_STATE_INACTIVE
;
6746 child_event
->state
= PERF_EVENT_STATE_OFF
;
6748 if (parent_event
->attr
.freq
) {
6749 u64 sample_period
= parent_event
->hw
.sample_period
;
6750 struct hw_perf_event
*hwc
= &child_event
->hw
;
6752 hwc
->sample_period
= sample_period
;
6753 hwc
->last_period
= sample_period
;
6755 local64_set(&hwc
->period_left
, sample_period
);
6758 child_event
->ctx
= child_ctx
;
6759 child_event
->overflow_handler
= parent_event
->overflow_handler
;
6760 child_event
->overflow_handler_context
6761 = parent_event
->overflow_handler_context
;
6764 * Precalculate sample_data sizes
6766 perf_event__header_size(child_event
);
6767 perf_event__id_header_size(child_event
);
6770 * Link it up in the child's context:
6772 raw_spin_lock_irqsave(&child_ctx
->lock
, flags
);
6773 add_event_to_ctx(child_event
, child_ctx
);
6774 raw_spin_unlock_irqrestore(&child_ctx
->lock
, flags
);
6777 * Get a reference to the parent filp - we will fput it
6778 * when the child event exits. This is safe to do because
6779 * we are in the parent and we know that the filp still
6780 * exists and has a nonzero count:
6782 atomic_long_inc(&parent_event
->filp
->f_count
);
6785 * Link this into the parent event's child list
6787 WARN_ON_ONCE(parent_event
->ctx
->parent_ctx
);
6788 mutex_lock(&parent_event
->child_mutex
);
6789 list_add_tail(&child_event
->child_list
, &parent_event
->child_list
);
6790 mutex_unlock(&parent_event
->child_mutex
);
6795 static int inherit_group(struct perf_event
*parent_event
,
6796 struct task_struct
*parent
,
6797 struct perf_event_context
*parent_ctx
,
6798 struct task_struct
*child
,
6799 struct perf_event_context
*child_ctx
)
6801 struct perf_event
*leader
;
6802 struct perf_event
*sub
;
6803 struct perf_event
*child_ctr
;
6805 leader
= inherit_event(parent_event
, parent
, parent_ctx
,
6806 child
, NULL
, child_ctx
);
6808 return PTR_ERR(leader
);
6809 list_for_each_entry(sub
, &parent_event
->sibling_list
, group_entry
) {
6810 child_ctr
= inherit_event(sub
, parent
, parent_ctx
,
6811 child
, leader
, child_ctx
);
6812 if (IS_ERR(child_ctr
))
6813 return PTR_ERR(child_ctr
);
6819 inherit_task_group(struct perf_event
*event
, struct task_struct
*parent
,
6820 struct perf_event_context
*parent_ctx
,
6821 struct task_struct
*child
, int ctxn
,
6825 struct perf_event_context
*child_ctx
;
6827 if (!event
->attr
.inherit
) {
6832 child_ctx
= child
->perf_event_ctxp
[ctxn
];
6835 * This is executed from the parent task context, so
6836 * inherit events that have been marked for cloning.
6837 * First allocate and initialize a context for the
6841 child_ctx
= alloc_perf_context(event
->pmu
, child
);
6845 child
->perf_event_ctxp
[ctxn
] = child_ctx
;
6848 ret
= inherit_group(event
, parent
, parent_ctx
,
6858 * Initialize the perf_event context in task_struct
6860 int perf_event_init_context(struct task_struct
*child
, int ctxn
)
6862 struct perf_event_context
*child_ctx
, *parent_ctx
;
6863 struct perf_event_context
*cloned_ctx
;
6864 struct perf_event
*event
;
6865 struct task_struct
*parent
= current
;
6866 int inherited_all
= 1;
6867 unsigned long flags
;
6870 if (likely(!parent
->perf_event_ctxp
[ctxn
]))
6874 * If the parent's context is a clone, pin it so it won't get
6877 parent_ctx
= perf_pin_task_context(parent
, ctxn
);
6880 * No need to check if parent_ctx != NULL here; since we saw
6881 * it non-NULL earlier, the only reason for it to become NULL
6882 * is if we exit, and since we're currently in the middle of
6883 * a fork we can't be exiting at the same time.
6887 * Lock the parent list. No need to lock the child - not PID
6888 * hashed yet and not running, so nobody can access it.
6890 mutex_lock(&parent_ctx
->mutex
);
6893 * We dont have to disable NMIs - we are only looking at
6894 * the list, not manipulating it:
6896 list_for_each_entry(event
, &parent_ctx
->pinned_groups
, group_entry
) {
6897 ret
= inherit_task_group(event
, parent
, parent_ctx
,
6898 child
, ctxn
, &inherited_all
);
6904 * We can't hold ctx->lock when iterating the ->flexible_group list due
6905 * to allocations, but we need to prevent rotation because
6906 * rotate_ctx() will change the list from interrupt context.
6908 raw_spin_lock_irqsave(&parent_ctx
->lock
, flags
);
6909 parent_ctx
->rotate_disable
= 1;
6910 raw_spin_unlock_irqrestore(&parent_ctx
->lock
, flags
);
6912 list_for_each_entry(event
, &parent_ctx
->flexible_groups
, group_entry
) {
6913 ret
= inherit_task_group(event
, parent
, parent_ctx
,
6914 child
, ctxn
, &inherited_all
);
6919 raw_spin_lock_irqsave(&parent_ctx
->lock
, flags
);
6920 parent_ctx
->rotate_disable
= 0;
6922 child_ctx
= child
->perf_event_ctxp
[ctxn
];
6924 if (child_ctx
&& inherited_all
) {
6926 * Mark the child context as a clone of the parent
6927 * context, or of whatever the parent is a clone of.
6929 * Note that if the parent is a clone, the holding of
6930 * parent_ctx->lock avoids it from being uncloned.
6932 cloned_ctx
= parent_ctx
->parent_ctx
;
6934 child_ctx
->parent_ctx
= cloned_ctx
;
6935 child_ctx
->parent_gen
= parent_ctx
->parent_gen
;
6937 child_ctx
->parent_ctx
= parent_ctx
;
6938 child_ctx
->parent_gen
= parent_ctx
->generation
;
6940 get_ctx(child_ctx
->parent_ctx
);
6943 raw_spin_unlock_irqrestore(&parent_ctx
->lock
, flags
);
6944 mutex_unlock(&parent_ctx
->mutex
);
6946 perf_unpin_context(parent_ctx
);
6947 put_ctx(parent_ctx
);
6953 * Initialize the perf_event context in task_struct
6955 int perf_event_init_task(struct task_struct
*child
)
6959 memset(child
->perf_event_ctxp
, 0, sizeof(child
->perf_event_ctxp
));
6960 mutex_init(&child
->perf_event_mutex
);
6961 INIT_LIST_HEAD(&child
->perf_event_list
);
6963 for_each_task_context_nr(ctxn
) {
6964 ret
= perf_event_init_context(child
, ctxn
);
6972 static void __init
perf_event_init_all_cpus(void)
6974 struct swevent_htable
*swhash
;
6977 for_each_possible_cpu(cpu
) {
6978 swhash
= &per_cpu(swevent_htable
, cpu
);
6979 mutex_init(&swhash
->hlist_mutex
);
6980 INIT_LIST_HEAD(&per_cpu(rotation_list
, cpu
));
6984 static void __cpuinit
perf_event_init_cpu(int cpu
)
6986 struct swevent_htable
*swhash
= &per_cpu(swevent_htable
, cpu
);
6988 mutex_lock(&swhash
->hlist_mutex
);
6989 if (swhash
->hlist_refcount
> 0) {
6990 struct swevent_hlist
*hlist
;
6992 hlist
= kzalloc_node(sizeof(*hlist
), GFP_KERNEL
, cpu_to_node(cpu
));
6994 rcu_assign_pointer(swhash
->swevent_hlist
, hlist
);
6996 mutex_unlock(&swhash
->hlist_mutex
);
6999 #if defined CONFIG_HOTPLUG_CPU || defined CONFIG_KEXEC
7000 static void perf_pmu_rotate_stop(struct pmu
*pmu
)
7002 struct perf_cpu_context
*cpuctx
= this_cpu_ptr(pmu
->pmu_cpu_context
);
7004 WARN_ON(!irqs_disabled());
7006 list_del_init(&cpuctx
->rotation_list
);
7009 static void __perf_event_exit_context(void *__info
)
7011 struct perf_event_context
*ctx
= __info
;
7012 struct perf_event
*event
, *tmp
;
7014 perf_pmu_rotate_stop(ctx
->pmu
);
7016 list_for_each_entry_safe(event
, tmp
, &ctx
->pinned_groups
, group_entry
)
7017 __perf_remove_from_context(event
);
7018 list_for_each_entry_safe(event
, tmp
, &ctx
->flexible_groups
, group_entry
)
7019 __perf_remove_from_context(event
);
7022 static void perf_event_exit_cpu_context(int cpu
)
7024 struct perf_event_context
*ctx
;
7028 idx
= srcu_read_lock(&pmus_srcu
);
7029 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
7030 ctx
= &per_cpu_ptr(pmu
->pmu_cpu_context
, cpu
)->ctx
;
7032 mutex_lock(&ctx
->mutex
);
7033 smp_call_function_single(cpu
, __perf_event_exit_context
, ctx
, 1);
7034 mutex_unlock(&ctx
->mutex
);
7036 srcu_read_unlock(&pmus_srcu
, idx
);
7039 static void perf_event_exit_cpu(int cpu
)
7041 struct swevent_htable
*swhash
= &per_cpu(swevent_htable
, cpu
);
7043 mutex_lock(&swhash
->hlist_mutex
);
7044 swevent_hlist_release(swhash
);
7045 mutex_unlock(&swhash
->hlist_mutex
);
7047 perf_event_exit_cpu_context(cpu
);
7050 static inline void perf_event_exit_cpu(int cpu
) { }
7054 perf_reboot(struct notifier_block
*notifier
, unsigned long val
, void *v
)
7058 for_each_online_cpu(cpu
)
7059 perf_event_exit_cpu(cpu
);
7065 * Run the perf reboot notifier at the very last possible moment so that
7066 * the generic watchdog code runs as long as possible.
7068 static struct notifier_block perf_reboot_notifier
= {
7069 .notifier_call
= perf_reboot
,
7070 .priority
= INT_MIN
,
7073 static int __cpuinit
7074 perf_cpu_notify(struct notifier_block
*self
, unsigned long action
, void *hcpu
)
7076 unsigned int cpu
= (long)hcpu
;
7078 switch (action
& ~CPU_TASKS_FROZEN
) {
7080 case CPU_UP_PREPARE
:
7081 case CPU_DOWN_FAILED
:
7082 perf_event_init_cpu(cpu
);
7085 case CPU_UP_CANCELED
:
7086 case CPU_DOWN_PREPARE
:
7087 perf_event_exit_cpu(cpu
);
7097 void __init
perf_event_init(void)
7103 perf_event_init_all_cpus();
7104 init_srcu_struct(&pmus_srcu
);
7105 perf_pmu_register(&perf_swevent
, "software", PERF_TYPE_SOFTWARE
);
7106 perf_pmu_register(&perf_cpu_clock
, NULL
, -1);
7107 perf_pmu_register(&perf_task_clock
, NULL
, -1);
7109 perf_cpu_notifier(perf_cpu_notify
);
7110 register_reboot_notifier(&perf_reboot_notifier
);
7112 ret
= init_hw_breakpoint();
7113 WARN(ret
, "hw_breakpoint initialization failed with: %d", ret
);
7115 /* do not patch jump label more than once per second */
7116 jump_label_rate_limit(&perf_sched_events
, HZ
);
7119 * Build time assertion that we keep the data_head at the intended
7120 * location. IOW, validation we got the __reserved[] size right.
7122 BUILD_BUG_ON((offsetof(struct perf_event_mmap_page
, data_head
))
7126 static int __init
perf_event_sysfs_init(void)
7131 mutex_lock(&pmus_lock
);
7133 ret
= bus_register(&pmu_bus
);
7137 list_for_each_entry(pmu
, &pmus
, entry
) {
7138 if (!pmu
->name
|| pmu
->type
< 0)
7141 ret
= pmu_dev_alloc(pmu
);
7142 WARN(ret
, "Failed to register pmu: %s, reason %d\n", pmu
->name
, ret
);
7144 pmu_bus_running
= 1;
7148 mutex_unlock(&pmus_lock
);
7152 device_initcall(perf_event_sysfs_init
);
7154 #ifdef CONFIG_CGROUP_PERF
7155 static struct cgroup_subsys_state
*perf_cgroup_create(struct cgroup
*cont
)
7157 struct perf_cgroup
*jc
;
7159 jc
= kzalloc(sizeof(*jc
), GFP_KERNEL
);
7161 return ERR_PTR(-ENOMEM
);
7163 jc
->info
= alloc_percpu(struct perf_cgroup_info
);
7166 return ERR_PTR(-ENOMEM
);
7172 static void perf_cgroup_destroy(struct cgroup
*cont
)
7174 struct perf_cgroup
*jc
;
7175 jc
= container_of(cgroup_subsys_state(cont
, perf_subsys_id
),
7176 struct perf_cgroup
, css
);
7177 free_percpu(jc
->info
);
7181 static int __perf_cgroup_move(void *info
)
7183 struct task_struct
*task
= info
;
7184 perf_cgroup_switch(task
, PERF_CGROUP_SWOUT
| PERF_CGROUP_SWIN
);
7188 static void perf_cgroup_attach(struct cgroup
*cgrp
, struct cgroup_taskset
*tset
)
7190 struct task_struct
*task
;
7192 cgroup_taskset_for_each(task
, cgrp
, tset
)
7193 task_function_call(task
, __perf_cgroup_move
, task
);
7196 static void perf_cgroup_exit(struct cgroup
*cgrp
, struct cgroup
*old_cgrp
,
7197 struct task_struct
*task
)
7200 * cgroup_exit() is called in the copy_process() failure path.
7201 * Ignore this case since the task hasn't ran yet, this avoids
7202 * trying to poke a half freed task state from generic code.
7204 if (!(task
->flags
& PF_EXITING
))
7207 task_function_call(task
, __perf_cgroup_move
, task
);
7210 struct cgroup_subsys perf_subsys
= {
7211 .name
= "perf_event",
7212 .subsys_id
= perf_subsys_id
,
7213 .create
= perf_cgroup_create
,
7214 .destroy
= perf_cgroup_destroy
,
7215 .exit
= perf_cgroup_exit
,
7216 .attach
= perf_cgroup_attach
,
7218 #endif /* CONFIG_CGROUP_PERF */