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 bool perf_tryget_cgroup(struct perf_event
*event
)
258 return css_tryget(&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
;
470 struct fd f
= fdget(fd
);
476 css
= cgroup_css_from_dir(f
.file
, perf_subsys_id
);
482 cgrp
= container_of(css
, struct perf_cgroup
, css
);
485 /* must be done before we fput() the file */
486 if (!perf_tryget_cgroup(event
)) {
493 * all events in a group must monitor
494 * the same cgroup because a task belongs
495 * to only one perf cgroup at a time
497 if (group_leader
&& group_leader
->cgrp
!= cgrp
) {
498 perf_detach_cgroup(event
);
507 perf_cgroup_set_shadow_time(struct perf_event
*event
, u64 now
)
509 struct perf_cgroup_info
*t
;
510 t
= per_cpu_ptr(event
->cgrp
->info
, event
->cpu
);
511 event
->shadow_ctx_time
= now
- t
->timestamp
;
515 perf_cgroup_defer_enabled(struct perf_event
*event
)
518 * when the current task's perf cgroup does not match
519 * the event's, we need to remember to call the
520 * perf_mark_enable() function the first time a task with
521 * a matching perf cgroup is scheduled in.
523 if (is_cgroup_event(event
) && !perf_cgroup_match(event
))
524 event
->cgrp_defer_enabled
= 1;
528 perf_cgroup_mark_enabled(struct perf_event
*event
,
529 struct perf_event_context
*ctx
)
531 struct perf_event
*sub
;
532 u64 tstamp
= perf_event_time(event
);
534 if (!event
->cgrp_defer_enabled
)
537 event
->cgrp_defer_enabled
= 0;
539 event
->tstamp_enabled
= tstamp
- event
->total_time_enabled
;
540 list_for_each_entry(sub
, &event
->sibling_list
, group_entry
) {
541 if (sub
->state
>= PERF_EVENT_STATE_INACTIVE
) {
542 sub
->tstamp_enabled
= tstamp
- sub
->total_time_enabled
;
543 sub
->cgrp_defer_enabled
= 0;
547 #else /* !CONFIG_CGROUP_PERF */
550 perf_cgroup_match(struct perf_event
*event
)
555 static inline void perf_detach_cgroup(struct perf_event
*event
)
558 static inline int is_cgroup_event(struct perf_event
*event
)
563 static inline u64
perf_cgroup_event_cgrp_time(struct perf_event
*event
)
568 static inline void update_cgrp_time_from_event(struct perf_event
*event
)
572 static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context
*cpuctx
)
576 static inline void perf_cgroup_sched_out(struct task_struct
*task
,
577 struct task_struct
*next
)
581 static inline void perf_cgroup_sched_in(struct task_struct
*prev
,
582 struct task_struct
*task
)
586 static inline int perf_cgroup_connect(pid_t pid
, struct perf_event
*event
,
587 struct perf_event_attr
*attr
,
588 struct perf_event
*group_leader
)
594 perf_cgroup_set_timestamp(struct task_struct
*task
,
595 struct perf_event_context
*ctx
)
600 perf_cgroup_switch(struct task_struct
*task
, struct task_struct
*next
)
605 perf_cgroup_set_shadow_time(struct perf_event
*event
, u64 now
)
609 static inline u64
perf_cgroup_event_time(struct perf_event
*event
)
615 perf_cgroup_defer_enabled(struct perf_event
*event
)
620 perf_cgroup_mark_enabled(struct perf_event
*event
,
621 struct perf_event_context
*ctx
)
626 void perf_pmu_disable(struct pmu
*pmu
)
628 int *count
= this_cpu_ptr(pmu
->pmu_disable_count
);
630 pmu
->pmu_disable(pmu
);
633 void perf_pmu_enable(struct pmu
*pmu
)
635 int *count
= this_cpu_ptr(pmu
->pmu_disable_count
);
637 pmu
->pmu_enable(pmu
);
640 static DEFINE_PER_CPU(struct list_head
, rotation_list
);
643 * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
644 * because they're strictly cpu affine and rotate_start is called with IRQs
645 * disabled, while rotate_context is called from IRQ context.
647 static void perf_pmu_rotate_start(struct pmu
*pmu
)
649 struct perf_cpu_context
*cpuctx
= this_cpu_ptr(pmu
->pmu_cpu_context
);
650 struct list_head
*head
= &__get_cpu_var(rotation_list
);
652 WARN_ON(!irqs_disabled());
654 if (list_empty(&cpuctx
->rotation_list
))
655 list_add(&cpuctx
->rotation_list
, head
);
658 static void get_ctx(struct perf_event_context
*ctx
)
660 WARN_ON(!atomic_inc_not_zero(&ctx
->refcount
));
663 static void put_ctx(struct perf_event_context
*ctx
)
665 if (atomic_dec_and_test(&ctx
->refcount
)) {
667 put_ctx(ctx
->parent_ctx
);
669 put_task_struct(ctx
->task
);
670 kfree_rcu(ctx
, rcu_head
);
674 static void unclone_ctx(struct perf_event_context
*ctx
)
676 if (ctx
->parent_ctx
) {
677 put_ctx(ctx
->parent_ctx
);
678 ctx
->parent_ctx
= NULL
;
682 static u32
perf_event_pid(struct perf_event
*event
, struct task_struct
*p
)
685 * only top level events have the pid namespace they were created in
688 event
= event
->parent
;
690 return task_tgid_nr_ns(p
, event
->ns
);
693 static u32
perf_event_tid(struct perf_event
*event
, struct task_struct
*p
)
696 * only top level events have the pid namespace they were created in
699 event
= event
->parent
;
701 return task_pid_nr_ns(p
, event
->ns
);
705 * If we inherit events we want to return the parent event id
708 static u64
primary_event_id(struct perf_event
*event
)
713 id
= event
->parent
->id
;
719 * Get the perf_event_context for a task and lock it.
720 * This has to cope with with the fact that until it is locked,
721 * the context could get moved to another task.
723 static struct perf_event_context
*
724 perf_lock_task_context(struct task_struct
*task
, int ctxn
, unsigned long *flags
)
726 struct perf_event_context
*ctx
;
730 ctx
= rcu_dereference(task
->perf_event_ctxp
[ctxn
]);
733 * If this context is a clone of another, it might
734 * get swapped for another underneath us by
735 * perf_event_task_sched_out, though the
736 * rcu_read_lock() protects us from any context
737 * getting freed. Lock the context and check if it
738 * got swapped before we could get the lock, and retry
739 * if so. If we locked the right context, then it
740 * can't get swapped on us any more.
742 raw_spin_lock_irqsave(&ctx
->lock
, *flags
);
743 if (ctx
!= rcu_dereference(task
->perf_event_ctxp
[ctxn
])) {
744 raw_spin_unlock_irqrestore(&ctx
->lock
, *flags
);
748 if (!atomic_inc_not_zero(&ctx
->refcount
)) {
749 raw_spin_unlock_irqrestore(&ctx
->lock
, *flags
);
758 * Get the context for a task and increment its pin_count so it
759 * can't get swapped to another task. This also increments its
760 * reference count so that the context can't get freed.
762 static struct perf_event_context
*
763 perf_pin_task_context(struct task_struct
*task
, int ctxn
)
765 struct perf_event_context
*ctx
;
768 ctx
= perf_lock_task_context(task
, ctxn
, &flags
);
771 raw_spin_unlock_irqrestore(&ctx
->lock
, flags
);
776 static void perf_unpin_context(struct perf_event_context
*ctx
)
780 raw_spin_lock_irqsave(&ctx
->lock
, flags
);
782 raw_spin_unlock_irqrestore(&ctx
->lock
, flags
);
786 * Update the record of the current time in a context.
788 static void update_context_time(struct perf_event_context
*ctx
)
790 u64 now
= perf_clock();
792 ctx
->time
+= now
- ctx
->timestamp
;
793 ctx
->timestamp
= now
;
796 static u64
perf_event_time(struct perf_event
*event
)
798 struct perf_event_context
*ctx
= event
->ctx
;
800 if (is_cgroup_event(event
))
801 return perf_cgroup_event_time(event
);
803 return ctx
? ctx
->time
: 0;
807 * Update the total_time_enabled and total_time_running fields for a event.
808 * The caller of this function needs to hold the ctx->lock.
810 static void update_event_times(struct perf_event
*event
)
812 struct perf_event_context
*ctx
= event
->ctx
;
815 if (event
->state
< PERF_EVENT_STATE_INACTIVE
||
816 event
->group_leader
->state
< PERF_EVENT_STATE_INACTIVE
)
819 * in cgroup mode, time_enabled represents
820 * the time the event was enabled AND active
821 * tasks were in the monitored cgroup. This is
822 * independent of the activity of the context as
823 * there may be a mix of cgroup and non-cgroup events.
825 * That is why we treat cgroup events differently
828 if (is_cgroup_event(event
))
829 run_end
= perf_cgroup_event_time(event
);
830 else if (ctx
->is_active
)
833 run_end
= event
->tstamp_stopped
;
835 event
->total_time_enabled
= run_end
- event
->tstamp_enabled
;
837 if (event
->state
== PERF_EVENT_STATE_INACTIVE
)
838 run_end
= event
->tstamp_stopped
;
840 run_end
= perf_event_time(event
);
842 event
->total_time_running
= run_end
- event
->tstamp_running
;
847 * Update total_time_enabled and total_time_running for all events in a group.
849 static void update_group_times(struct perf_event
*leader
)
851 struct perf_event
*event
;
853 update_event_times(leader
);
854 list_for_each_entry(event
, &leader
->sibling_list
, group_entry
)
855 update_event_times(event
);
858 static struct list_head
*
859 ctx_group_list(struct perf_event
*event
, struct perf_event_context
*ctx
)
861 if (event
->attr
.pinned
)
862 return &ctx
->pinned_groups
;
864 return &ctx
->flexible_groups
;
868 * Add a event from the lists for its context.
869 * Must be called with ctx->mutex and ctx->lock held.
872 list_add_event(struct perf_event
*event
, struct perf_event_context
*ctx
)
874 WARN_ON_ONCE(event
->attach_state
& PERF_ATTACH_CONTEXT
);
875 event
->attach_state
|= PERF_ATTACH_CONTEXT
;
878 * If we're a stand alone event or group leader, we go to the context
879 * list, group events are kept attached to the group so that
880 * perf_group_detach can, at all times, locate all siblings.
882 if (event
->group_leader
== event
) {
883 struct list_head
*list
;
885 if (is_software_event(event
))
886 event
->group_flags
|= PERF_GROUP_SOFTWARE
;
888 list
= ctx_group_list(event
, ctx
);
889 list_add_tail(&event
->group_entry
, list
);
892 if (is_cgroup_event(event
))
895 if (has_branch_stack(event
))
896 ctx
->nr_branch_stack
++;
898 list_add_rcu(&event
->event_entry
, &ctx
->event_list
);
900 perf_pmu_rotate_start(ctx
->pmu
);
902 if (event
->attr
.inherit_stat
)
907 * Called at perf_event creation and when events are attached/detached from a
910 static void perf_event__read_size(struct perf_event
*event
)
912 int entry
= sizeof(u64
); /* value */
916 if (event
->attr
.read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
919 if (event
->attr
.read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
922 if (event
->attr
.read_format
& PERF_FORMAT_ID
)
923 entry
+= sizeof(u64
);
925 if (event
->attr
.read_format
& PERF_FORMAT_GROUP
) {
926 nr
+= event
->group_leader
->nr_siblings
;
931 event
->read_size
= size
;
934 static void perf_event__header_size(struct perf_event
*event
)
936 struct perf_sample_data
*data
;
937 u64 sample_type
= event
->attr
.sample_type
;
940 perf_event__read_size(event
);
942 if (sample_type
& PERF_SAMPLE_IP
)
943 size
+= sizeof(data
->ip
);
945 if (sample_type
& PERF_SAMPLE_ADDR
)
946 size
+= sizeof(data
->addr
);
948 if (sample_type
& PERF_SAMPLE_PERIOD
)
949 size
+= sizeof(data
->period
);
951 if (sample_type
& PERF_SAMPLE_READ
)
952 size
+= event
->read_size
;
954 event
->header_size
= size
;
957 static void perf_event__id_header_size(struct perf_event
*event
)
959 struct perf_sample_data
*data
;
960 u64 sample_type
= event
->attr
.sample_type
;
963 if (sample_type
& PERF_SAMPLE_TID
)
964 size
+= sizeof(data
->tid_entry
);
966 if (sample_type
& PERF_SAMPLE_TIME
)
967 size
+= sizeof(data
->time
);
969 if (sample_type
& PERF_SAMPLE_ID
)
970 size
+= sizeof(data
->id
);
972 if (sample_type
& PERF_SAMPLE_STREAM_ID
)
973 size
+= sizeof(data
->stream_id
);
975 if (sample_type
& PERF_SAMPLE_CPU
)
976 size
+= sizeof(data
->cpu_entry
);
978 event
->id_header_size
= size
;
981 static void perf_group_attach(struct perf_event
*event
)
983 struct perf_event
*group_leader
= event
->group_leader
, *pos
;
986 * We can have double attach due to group movement in perf_event_open.
988 if (event
->attach_state
& PERF_ATTACH_GROUP
)
991 event
->attach_state
|= PERF_ATTACH_GROUP
;
993 if (group_leader
== event
)
996 if (group_leader
->group_flags
& PERF_GROUP_SOFTWARE
&&
997 !is_software_event(event
))
998 group_leader
->group_flags
&= ~PERF_GROUP_SOFTWARE
;
1000 list_add_tail(&event
->group_entry
, &group_leader
->sibling_list
);
1001 group_leader
->nr_siblings
++;
1003 perf_event__header_size(group_leader
);
1005 list_for_each_entry(pos
, &group_leader
->sibling_list
, group_entry
)
1006 perf_event__header_size(pos
);
1010 * Remove a event from the lists for its context.
1011 * Must be called with ctx->mutex and ctx->lock held.
1014 list_del_event(struct perf_event
*event
, struct perf_event_context
*ctx
)
1016 struct perf_cpu_context
*cpuctx
;
1018 * We can have double detach due to exit/hot-unplug + close.
1020 if (!(event
->attach_state
& PERF_ATTACH_CONTEXT
))
1023 event
->attach_state
&= ~PERF_ATTACH_CONTEXT
;
1025 if (is_cgroup_event(event
)) {
1027 cpuctx
= __get_cpu_context(ctx
);
1029 * if there are no more cgroup events
1030 * then cler cgrp to avoid stale pointer
1031 * in update_cgrp_time_from_cpuctx()
1033 if (!ctx
->nr_cgroups
)
1034 cpuctx
->cgrp
= NULL
;
1037 if (has_branch_stack(event
))
1038 ctx
->nr_branch_stack
--;
1041 if (event
->attr
.inherit_stat
)
1044 list_del_rcu(&event
->event_entry
);
1046 if (event
->group_leader
== event
)
1047 list_del_init(&event
->group_entry
);
1049 update_group_times(event
);
1052 * If event was in error state, then keep it
1053 * that way, otherwise bogus counts will be
1054 * returned on read(). The only way to get out
1055 * of error state is by explicit re-enabling
1058 if (event
->state
> PERF_EVENT_STATE_OFF
)
1059 event
->state
= PERF_EVENT_STATE_OFF
;
1062 static void perf_group_detach(struct perf_event
*event
)
1064 struct perf_event
*sibling
, *tmp
;
1065 struct list_head
*list
= NULL
;
1068 * We can have double detach due to exit/hot-unplug + close.
1070 if (!(event
->attach_state
& PERF_ATTACH_GROUP
))
1073 event
->attach_state
&= ~PERF_ATTACH_GROUP
;
1076 * If this is a sibling, remove it from its group.
1078 if (event
->group_leader
!= event
) {
1079 list_del_init(&event
->group_entry
);
1080 event
->group_leader
->nr_siblings
--;
1084 if (!list_empty(&event
->group_entry
))
1085 list
= &event
->group_entry
;
1088 * If this was a group event with sibling events then
1089 * upgrade the siblings to singleton events by adding them
1090 * to whatever list we are on.
1092 list_for_each_entry_safe(sibling
, tmp
, &event
->sibling_list
, group_entry
) {
1094 list_move_tail(&sibling
->group_entry
, list
);
1095 sibling
->group_leader
= sibling
;
1097 /* Inherit group flags from the previous leader */
1098 sibling
->group_flags
= event
->group_flags
;
1102 perf_event__header_size(event
->group_leader
);
1104 list_for_each_entry(tmp
, &event
->group_leader
->sibling_list
, group_entry
)
1105 perf_event__header_size(tmp
);
1109 event_filter_match(struct perf_event
*event
)
1111 return (event
->cpu
== -1 || event
->cpu
== smp_processor_id())
1112 && perf_cgroup_match(event
);
1116 event_sched_out(struct perf_event
*event
,
1117 struct perf_cpu_context
*cpuctx
,
1118 struct perf_event_context
*ctx
)
1120 u64 tstamp
= perf_event_time(event
);
1123 * An event which could not be activated because of
1124 * filter mismatch still needs to have its timings
1125 * maintained, otherwise bogus information is return
1126 * via read() for time_enabled, time_running:
1128 if (event
->state
== PERF_EVENT_STATE_INACTIVE
1129 && !event_filter_match(event
)) {
1130 delta
= tstamp
- event
->tstamp_stopped
;
1131 event
->tstamp_running
+= delta
;
1132 event
->tstamp_stopped
= tstamp
;
1135 if (event
->state
!= PERF_EVENT_STATE_ACTIVE
)
1138 event
->state
= PERF_EVENT_STATE_INACTIVE
;
1139 if (event
->pending_disable
) {
1140 event
->pending_disable
= 0;
1141 event
->state
= PERF_EVENT_STATE_OFF
;
1143 event
->tstamp_stopped
= tstamp
;
1144 event
->pmu
->del(event
, 0);
1147 if (!is_software_event(event
))
1148 cpuctx
->active_oncpu
--;
1150 if (event
->attr
.freq
&& event
->attr
.sample_freq
)
1152 if (event
->attr
.exclusive
|| !cpuctx
->active_oncpu
)
1153 cpuctx
->exclusive
= 0;
1157 group_sched_out(struct perf_event
*group_event
,
1158 struct perf_cpu_context
*cpuctx
,
1159 struct perf_event_context
*ctx
)
1161 struct perf_event
*event
;
1162 int state
= group_event
->state
;
1164 event_sched_out(group_event
, cpuctx
, ctx
);
1167 * Schedule out siblings (if any):
1169 list_for_each_entry(event
, &group_event
->sibling_list
, group_entry
)
1170 event_sched_out(event
, cpuctx
, ctx
);
1172 if (state
== PERF_EVENT_STATE_ACTIVE
&& group_event
->attr
.exclusive
)
1173 cpuctx
->exclusive
= 0;
1177 * Cross CPU call to remove a performance event
1179 * We disable the event on the hardware level first. After that we
1180 * remove it from the context list.
1182 static int __perf_remove_from_context(void *info
)
1184 struct perf_event
*event
= info
;
1185 struct perf_event_context
*ctx
= event
->ctx
;
1186 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
1188 raw_spin_lock(&ctx
->lock
);
1189 event_sched_out(event
, cpuctx
, ctx
);
1190 list_del_event(event
, ctx
);
1191 if (!ctx
->nr_events
&& cpuctx
->task_ctx
== ctx
) {
1193 cpuctx
->task_ctx
= NULL
;
1195 raw_spin_unlock(&ctx
->lock
);
1202 * Remove the event from a task's (or a CPU's) list of events.
1204 * CPU events are removed with a smp call. For task events we only
1205 * call when the task is on a CPU.
1207 * If event->ctx is a cloned context, callers must make sure that
1208 * every task struct that event->ctx->task could possibly point to
1209 * remains valid. This is OK when called from perf_release since
1210 * that only calls us on the top-level context, which can't be a clone.
1211 * When called from perf_event_exit_task, it's OK because the
1212 * context has been detached from its task.
1214 static void perf_remove_from_context(struct perf_event
*event
)
1216 struct perf_event_context
*ctx
= event
->ctx
;
1217 struct task_struct
*task
= ctx
->task
;
1219 lockdep_assert_held(&ctx
->mutex
);
1223 * Per cpu events are removed via an smp call and
1224 * the removal is always successful.
1226 cpu_function_call(event
->cpu
, __perf_remove_from_context
, event
);
1231 if (!task_function_call(task
, __perf_remove_from_context
, event
))
1234 raw_spin_lock_irq(&ctx
->lock
);
1236 * If we failed to find a running task, but find the context active now
1237 * that we've acquired the ctx->lock, retry.
1239 if (ctx
->is_active
) {
1240 raw_spin_unlock_irq(&ctx
->lock
);
1245 * Since the task isn't running, its safe to remove the event, us
1246 * holding the ctx->lock ensures the task won't get scheduled in.
1248 list_del_event(event
, ctx
);
1249 raw_spin_unlock_irq(&ctx
->lock
);
1253 * Cross CPU call to disable a performance event
1255 int __perf_event_disable(void *info
)
1257 struct perf_event
*event
= info
;
1258 struct perf_event_context
*ctx
= event
->ctx
;
1259 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
1262 * If this is a per-task event, need to check whether this
1263 * event's task is the current task on this cpu.
1265 * Can trigger due to concurrent perf_event_context_sched_out()
1266 * flipping contexts around.
1268 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
)
1271 raw_spin_lock(&ctx
->lock
);
1274 * If the event is on, turn it off.
1275 * If it is in error state, leave it in error state.
1277 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
) {
1278 update_context_time(ctx
);
1279 update_cgrp_time_from_event(event
);
1280 update_group_times(event
);
1281 if (event
== event
->group_leader
)
1282 group_sched_out(event
, cpuctx
, ctx
);
1284 event_sched_out(event
, cpuctx
, ctx
);
1285 event
->state
= PERF_EVENT_STATE_OFF
;
1288 raw_spin_unlock(&ctx
->lock
);
1296 * If event->ctx is a cloned context, callers must make sure that
1297 * every task struct that event->ctx->task could possibly point to
1298 * remains valid. This condition is satisifed when called through
1299 * perf_event_for_each_child or perf_event_for_each because they
1300 * hold the top-level event's child_mutex, so any descendant that
1301 * goes to exit will block in sync_child_event.
1302 * When called from perf_pending_event it's OK because event->ctx
1303 * is the current context on this CPU and preemption is disabled,
1304 * hence we can't get into perf_event_task_sched_out for this context.
1306 void perf_event_disable(struct perf_event
*event
)
1308 struct perf_event_context
*ctx
= event
->ctx
;
1309 struct task_struct
*task
= ctx
->task
;
1313 * Disable the event on the cpu that it's on
1315 cpu_function_call(event
->cpu
, __perf_event_disable
, event
);
1320 if (!task_function_call(task
, __perf_event_disable
, event
))
1323 raw_spin_lock_irq(&ctx
->lock
);
1325 * If the event is still active, we need to retry the cross-call.
1327 if (event
->state
== PERF_EVENT_STATE_ACTIVE
) {
1328 raw_spin_unlock_irq(&ctx
->lock
);
1330 * Reload the task pointer, it might have been changed by
1331 * a concurrent perf_event_context_sched_out().
1338 * Since we have the lock this context can't be scheduled
1339 * in, so we can change the state safely.
1341 if (event
->state
== PERF_EVENT_STATE_INACTIVE
) {
1342 update_group_times(event
);
1343 event
->state
= PERF_EVENT_STATE_OFF
;
1345 raw_spin_unlock_irq(&ctx
->lock
);
1347 EXPORT_SYMBOL_GPL(perf_event_disable
);
1349 static void perf_set_shadow_time(struct perf_event
*event
,
1350 struct perf_event_context
*ctx
,
1354 * use the correct time source for the time snapshot
1356 * We could get by without this by leveraging the
1357 * fact that to get to this function, the caller
1358 * has most likely already called update_context_time()
1359 * and update_cgrp_time_xx() and thus both timestamp
1360 * are identical (or very close). Given that tstamp is,
1361 * already adjusted for cgroup, we could say that:
1362 * tstamp - ctx->timestamp
1364 * tstamp - cgrp->timestamp.
1366 * Then, in perf_output_read(), the calculation would
1367 * work with no changes because:
1368 * - event is guaranteed scheduled in
1369 * - no scheduled out in between
1370 * - thus the timestamp would be the same
1372 * But this is a bit hairy.
1374 * So instead, we have an explicit cgroup call to remain
1375 * within the time time source all along. We believe it
1376 * is cleaner and simpler to understand.
1378 if (is_cgroup_event(event
))
1379 perf_cgroup_set_shadow_time(event
, tstamp
);
1381 event
->shadow_ctx_time
= tstamp
- ctx
->timestamp
;
1384 #define MAX_INTERRUPTS (~0ULL)
1386 static void perf_log_throttle(struct perf_event
*event
, int enable
);
1389 event_sched_in(struct perf_event
*event
,
1390 struct perf_cpu_context
*cpuctx
,
1391 struct perf_event_context
*ctx
)
1393 u64 tstamp
= perf_event_time(event
);
1395 if (event
->state
<= PERF_EVENT_STATE_OFF
)
1398 event
->state
= PERF_EVENT_STATE_ACTIVE
;
1399 event
->oncpu
= smp_processor_id();
1402 * Unthrottle events, since we scheduled we might have missed several
1403 * ticks already, also for a heavily scheduling task there is little
1404 * guarantee it'll get a tick in a timely manner.
1406 if (unlikely(event
->hw
.interrupts
== MAX_INTERRUPTS
)) {
1407 perf_log_throttle(event
, 1);
1408 event
->hw
.interrupts
= 0;
1412 * The new state must be visible before we turn it on in the hardware:
1416 if (event
->pmu
->add(event
, PERF_EF_START
)) {
1417 event
->state
= PERF_EVENT_STATE_INACTIVE
;
1422 event
->tstamp_running
+= tstamp
- event
->tstamp_stopped
;
1424 perf_set_shadow_time(event
, ctx
, tstamp
);
1426 if (!is_software_event(event
))
1427 cpuctx
->active_oncpu
++;
1429 if (event
->attr
.freq
&& event
->attr
.sample_freq
)
1432 if (event
->attr
.exclusive
)
1433 cpuctx
->exclusive
= 1;
1439 group_sched_in(struct perf_event
*group_event
,
1440 struct perf_cpu_context
*cpuctx
,
1441 struct perf_event_context
*ctx
)
1443 struct perf_event
*event
, *partial_group
= NULL
;
1444 struct pmu
*pmu
= group_event
->pmu
;
1445 u64 now
= ctx
->time
;
1446 bool simulate
= false;
1448 if (group_event
->state
== PERF_EVENT_STATE_OFF
)
1451 pmu
->start_txn(pmu
);
1453 if (event_sched_in(group_event
, cpuctx
, ctx
)) {
1454 pmu
->cancel_txn(pmu
);
1459 * Schedule in siblings as one group (if any):
1461 list_for_each_entry(event
, &group_event
->sibling_list
, group_entry
) {
1462 if (event_sched_in(event
, cpuctx
, ctx
)) {
1463 partial_group
= event
;
1468 if (!pmu
->commit_txn(pmu
))
1473 * Groups can be scheduled in as one unit only, so undo any
1474 * partial group before returning:
1475 * The events up to the failed event are scheduled out normally,
1476 * tstamp_stopped will be updated.
1478 * The failed events and the remaining siblings need to have
1479 * their timings updated as if they had gone thru event_sched_in()
1480 * and event_sched_out(). This is required to get consistent timings
1481 * across the group. This also takes care of the case where the group
1482 * could never be scheduled by ensuring tstamp_stopped is set to mark
1483 * the time the event was actually stopped, such that time delta
1484 * calculation in update_event_times() is correct.
1486 list_for_each_entry(event
, &group_event
->sibling_list
, group_entry
) {
1487 if (event
== partial_group
)
1491 event
->tstamp_running
+= now
- event
->tstamp_stopped
;
1492 event
->tstamp_stopped
= now
;
1494 event_sched_out(event
, cpuctx
, ctx
);
1497 event_sched_out(group_event
, cpuctx
, ctx
);
1499 pmu
->cancel_txn(pmu
);
1505 * Work out whether we can put this event group on the CPU now.
1507 static int group_can_go_on(struct perf_event
*event
,
1508 struct perf_cpu_context
*cpuctx
,
1512 * Groups consisting entirely of software events can always go on.
1514 if (event
->group_flags
& PERF_GROUP_SOFTWARE
)
1517 * If an exclusive group is already on, no other hardware
1520 if (cpuctx
->exclusive
)
1523 * If this group is exclusive and there are already
1524 * events on the CPU, it can't go on.
1526 if (event
->attr
.exclusive
&& cpuctx
->active_oncpu
)
1529 * Otherwise, try to add it if all previous groups were able
1535 static void add_event_to_ctx(struct perf_event
*event
,
1536 struct perf_event_context
*ctx
)
1538 u64 tstamp
= perf_event_time(event
);
1540 list_add_event(event
, ctx
);
1541 perf_group_attach(event
);
1542 event
->tstamp_enabled
= tstamp
;
1543 event
->tstamp_running
= tstamp
;
1544 event
->tstamp_stopped
= tstamp
;
1547 static void task_ctx_sched_out(struct perf_event_context
*ctx
);
1549 ctx_sched_in(struct perf_event_context
*ctx
,
1550 struct perf_cpu_context
*cpuctx
,
1551 enum event_type_t event_type
,
1552 struct task_struct
*task
);
1554 static void perf_event_sched_in(struct perf_cpu_context
*cpuctx
,
1555 struct perf_event_context
*ctx
,
1556 struct task_struct
*task
)
1558 cpu_ctx_sched_in(cpuctx
, EVENT_PINNED
, task
);
1560 ctx_sched_in(ctx
, cpuctx
, EVENT_PINNED
, task
);
1561 cpu_ctx_sched_in(cpuctx
, EVENT_FLEXIBLE
, task
);
1563 ctx_sched_in(ctx
, cpuctx
, EVENT_FLEXIBLE
, task
);
1567 * Cross CPU call to install and enable a performance event
1569 * Must be called with ctx->mutex held
1571 static int __perf_install_in_context(void *info
)
1573 struct perf_event
*event
= info
;
1574 struct perf_event_context
*ctx
= event
->ctx
;
1575 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
1576 struct perf_event_context
*task_ctx
= cpuctx
->task_ctx
;
1577 struct task_struct
*task
= current
;
1579 perf_ctx_lock(cpuctx
, task_ctx
);
1580 perf_pmu_disable(cpuctx
->ctx
.pmu
);
1583 * If there was an active task_ctx schedule it out.
1586 task_ctx_sched_out(task_ctx
);
1589 * If the context we're installing events in is not the
1590 * active task_ctx, flip them.
1592 if (ctx
->task
&& task_ctx
!= ctx
) {
1594 raw_spin_unlock(&task_ctx
->lock
);
1595 raw_spin_lock(&ctx
->lock
);
1600 cpuctx
->task_ctx
= task_ctx
;
1601 task
= task_ctx
->task
;
1604 cpu_ctx_sched_out(cpuctx
, EVENT_ALL
);
1606 update_context_time(ctx
);
1608 * update cgrp time only if current cgrp
1609 * matches event->cgrp. Must be done before
1610 * calling add_event_to_ctx()
1612 update_cgrp_time_from_event(event
);
1614 add_event_to_ctx(event
, ctx
);
1617 * Schedule everything back in
1619 perf_event_sched_in(cpuctx
, task_ctx
, task
);
1621 perf_pmu_enable(cpuctx
->ctx
.pmu
);
1622 perf_ctx_unlock(cpuctx
, task_ctx
);
1628 * Attach a performance event to a context
1630 * First we add the event to the list with the hardware enable bit
1631 * in event->hw_config cleared.
1633 * If the event is attached to a task which is on a CPU we use a smp
1634 * call to enable it in the task context. The task might have been
1635 * scheduled away, but we check this in the smp call again.
1638 perf_install_in_context(struct perf_event_context
*ctx
,
1639 struct perf_event
*event
,
1642 struct task_struct
*task
= ctx
->task
;
1644 lockdep_assert_held(&ctx
->mutex
);
1647 if (event
->cpu
!= -1)
1652 * Per cpu events are installed via an smp call and
1653 * the install is always successful.
1655 cpu_function_call(cpu
, __perf_install_in_context
, event
);
1660 if (!task_function_call(task
, __perf_install_in_context
, event
))
1663 raw_spin_lock_irq(&ctx
->lock
);
1665 * If we failed to find a running task, but find the context active now
1666 * that we've acquired the ctx->lock, retry.
1668 if (ctx
->is_active
) {
1669 raw_spin_unlock_irq(&ctx
->lock
);
1674 * Since the task isn't running, its safe to add the event, us holding
1675 * the ctx->lock ensures the task won't get scheduled in.
1677 add_event_to_ctx(event
, ctx
);
1678 raw_spin_unlock_irq(&ctx
->lock
);
1682 * Put a event into inactive state and update time fields.
1683 * Enabling the leader of a group effectively enables all
1684 * the group members that aren't explicitly disabled, so we
1685 * have to update their ->tstamp_enabled also.
1686 * Note: this works for group members as well as group leaders
1687 * since the non-leader members' sibling_lists will be empty.
1689 static void __perf_event_mark_enabled(struct perf_event
*event
)
1691 struct perf_event
*sub
;
1692 u64 tstamp
= perf_event_time(event
);
1694 event
->state
= PERF_EVENT_STATE_INACTIVE
;
1695 event
->tstamp_enabled
= tstamp
- event
->total_time_enabled
;
1696 list_for_each_entry(sub
, &event
->sibling_list
, group_entry
) {
1697 if (sub
->state
>= PERF_EVENT_STATE_INACTIVE
)
1698 sub
->tstamp_enabled
= tstamp
- sub
->total_time_enabled
;
1703 * Cross CPU call to enable a performance event
1705 static int __perf_event_enable(void *info
)
1707 struct perf_event
*event
= info
;
1708 struct perf_event_context
*ctx
= event
->ctx
;
1709 struct perf_event
*leader
= event
->group_leader
;
1710 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
1713 if (WARN_ON_ONCE(!ctx
->is_active
))
1716 raw_spin_lock(&ctx
->lock
);
1717 update_context_time(ctx
);
1719 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
)
1723 * set current task's cgroup time reference point
1725 perf_cgroup_set_timestamp(current
, ctx
);
1727 __perf_event_mark_enabled(event
);
1729 if (!event_filter_match(event
)) {
1730 if (is_cgroup_event(event
))
1731 perf_cgroup_defer_enabled(event
);
1736 * If the event is in a group and isn't the group leader,
1737 * then don't put it on unless the group is on.
1739 if (leader
!= event
&& leader
->state
!= PERF_EVENT_STATE_ACTIVE
)
1742 if (!group_can_go_on(event
, cpuctx
, 1)) {
1745 if (event
== leader
)
1746 err
= group_sched_in(event
, cpuctx
, ctx
);
1748 err
= event_sched_in(event
, cpuctx
, ctx
);
1753 * If this event can't go on and it's part of a
1754 * group, then the whole group has to come off.
1756 if (leader
!= event
)
1757 group_sched_out(leader
, cpuctx
, ctx
);
1758 if (leader
->attr
.pinned
) {
1759 update_group_times(leader
);
1760 leader
->state
= PERF_EVENT_STATE_ERROR
;
1765 raw_spin_unlock(&ctx
->lock
);
1773 * If event->ctx is a cloned context, callers must make sure that
1774 * every task struct that event->ctx->task could possibly point to
1775 * remains valid. This condition is satisfied when called through
1776 * perf_event_for_each_child or perf_event_for_each as described
1777 * for perf_event_disable.
1779 void perf_event_enable(struct perf_event
*event
)
1781 struct perf_event_context
*ctx
= event
->ctx
;
1782 struct task_struct
*task
= ctx
->task
;
1786 * Enable the event on the cpu that it's on
1788 cpu_function_call(event
->cpu
, __perf_event_enable
, event
);
1792 raw_spin_lock_irq(&ctx
->lock
);
1793 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
)
1797 * If the event is in error state, clear that first.
1798 * That way, if we see the event in error state below, we
1799 * know that it has gone back into error state, as distinct
1800 * from the task having been scheduled away before the
1801 * cross-call arrived.
1803 if (event
->state
== PERF_EVENT_STATE_ERROR
)
1804 event
->state
= PERF_EVENT_STATE_OFF
;
1807 if (!ctx
->is_active
) {
1808 __perf_event_mark_enabled(event
);
1812 raw_spin_unlock_irq(&ctx
->lock
);
1814 if (!task_function_call(task
, __perf_event_enable
, event
))
1817 raw_spin_lock_irq(&ctx
->lock
);
1820 * If the context is active and the event is still off,
1821 * we need to retry the cross-call.
1823 if (ctx
->is_active
&& event
->state
== PERF_EVENT_STATE_OFF
) {
1825 * task could have been flipped by a concurrent
1826 * perf_event_context_sched_out()
1833 raw_spin_unlock_irq(&ctx
->lock
);
1835 EXPORT_SYMBOL_GPL(perf_event_enable
);
1837 int perf_event_refresh(struct perf_event
*event
, int refresh
)
1840 * not supported on inherited events
1842 if (event
->attr
.inherit
|| !is_sampling_event(event
))
1845 atomic_add(refresh
, &event
->event_limit
);
1846 perf_event_enable(event
);
1850 EXPORT_SYMBOL_GPL(perf_event_refresh
);
1852 static void ctx_sched_out(struct perf_event_context
*ctx
,
1853 struct perf_cpu_context
*cpuctx
,
1854 enum event_type_t event_type
)
1856 struct perf_event
*event
;
1857 int is_active
= ctx
->is_active
;
1859 ctx
->is_active
&= ~event_type
;
1860 if (likely(!ctx
->nr_events
))
1863 update_context_time(ctx
);
1864 update_cgrp_time_from_cpuctx(cpuctx
);
1865 if (!ctx
->nr_active
)
1868 perf_pmu_disable(ctx
->pmu
);
1869 if ((is_active
& EVENT_PINNED
) && (event_type
& EVENT_PINNED
)) {
1870 list_for_each_entry(event
, &ctx
->pinned_groups
, group_entry
)
1871 group_sched_out(event
, cpuctx
, ctx
);
1874 if ((is_active
& EVENT_FLEXIBLE
) && (event_type
& EVENT_FLEXIBLE
)) {
1875 list_for_each_entry(event
, &ctx
->flexible_groups
, group_entry
)
1876 group_sched_out(event
, cpuctx
, ctx
);
1878 perf_pmu_enable(ctx
->pmu
);
1882 * Test whether two contexts are equivalent, i.e. whether they
1883 * have both been cloned from the same version of the same context
1884 * and they both have the same number of enabled events.
1885 * If the number of enabled events is the same, then the set
1886 * of enabled events should be the same, because these are both
1887 * inherited contexts, therefore we can't access individual events
1888 * in them directly with an fd; we can only enable/disable all
1889 * events via prctl, or enable/disable all events in a family
1890 * via ioctl, which will have the same effect on both contexts.
1892 static int context_equiv(struct perf_event_context
*ctx1
,
1893 struct perf_event_context
*ctx2
)
1895 return ctx1
->parent_ctx
&& ctx1
->parent_ctx
== ctx2
->parent_ctx
1896 && ctx1
->parent_gen
== ctx2
->parent_gen
1897 && !ctx1
->pin_count
&& !ctx2
->pin_count
;
1900 static void __perf_event_sync_stat(struct perf_event
*event
,
1901 struct perf_event
*next_event
)
1905 if (!event
->attr
.inherit_stat
)
1909 * Update the event value, we cannot use perf_event_read()
1910 * because we're in the middle of a context switch and have IRQs
1911 * disabled, which upsets smp_call_function_single(), however
1912 * we know the event must be on the current CPU, therefore we
1913 * don't need to use it.
1915 switch (event
->state
) {
1916 case PERF_EVENT_STATE_ACTIVE
:
1917 event
->pmu
->read(event
);
1920 case PERF_EVENT_STATE_INACTIVE
:
1921 update_event_times(event
);
1929 * In order to keep per-task stats reliable we need to flip the event
1930 * values when we flip the contexts.
1932 value
= local64_read(&next_event
->count
);
1933 value
= local64_xchg(&event
->count
, value
);
1934 local64_set(&next_event
->count
, value
);
1936 swap(event
->total_time_enabled
, next_event
->total_time_enabled
);
1937 swap(event
->total_time_running
, next_event
->total_time_running
);
1940 * Since we swizzled the values, update the user visible data too.
1942 perf_event_update_userpage(event
);
1943 perf_event_update_userpage(next_event
);
1946 #define list_next_entry(pos, member) \
1947 list_entry(pos->member.next, typeof(*pos), member)
1949 static void perf_event_sync_stat(struct perf_event_context
*ctx
,
1950 struct perf_event_context
*next_ctx
)
1952 struct perf_event
*event
, *next_event
;
1957 update_context_time(ctx
);
1959 event
= list_first_entry(&ctx
->event_list
,
1960 struct perf_event
, event_entry
);
1962 next_event
= list_first_entry(&next_ctx
->event_list
,
1963 struct perf_event
, event_entry
);
1965 while (&event
->event_entry
!= &ctx
->event_list
&&
1966 &next_event
->event_entry
!= &next_ctx
->event_list
) {
1968 __perf_event_sync_stat(event
, next_event
);
1970 event
= list_next_entry(event
, event_entry
);
1971 next_event
= list_next_entry(next_event
, event_entry
);
1975 static void perf_event_context_sched_out(struct task_struct
*task
, int ctxn
,
1976 struct task_struct
*next
)
1978 struct perf_event_context
*ctx
= task
->perf_event_ctxp
[ctxn
];
1979 struct perf_event_context
*next_ctx
;
1980 struct perf_event_context
*parent
;
1981 struct perf_cpu_context
*cpuctx
;
1987 cpuctx
= __get_cpu_context(ctx
);
1988 if (!cpuctx
->task_ctx
)
1992 parent
= rcu_dereference(ctx
->parent_ctx
);
1993 next_ctx
= next
->perf_event_ctxp
[ctxn
];
1994 if (parent
&& next_ctx
&&
1995 rcu_dereference(next_ctx
->parent_ctx
) == parent
) {
1997 * Looks like the two contexts are clones, so we might be
1998 * able to optimize the context switch. We lock both
1999 * contexts and check that they are clones under the
2000 * lock (including re-checking that neither has been
2001 * uncloned in the meantime). It doesn't matter which
2002 * order we take the locks because no other cpu could
2003 * be trying to lock both of these tasks.
2005 raw_spin_lock(&ctx
->lock
);
2006 raw_spin_lock_nested(&next_ctx
->lock
, SINGLE_DEPTH_NESTING
);
2007 if (context_equiv(ctx
, next_ctx
)) {
2009 * XXX do we need a memory barrier of sorts
2010 * wrt to rcu_dereference() of perf_event_ctxp
2012 task
->perf_event_ctxp
[ctxn
] = next_ctx
;
2013 next
->perf_event_ctxp
[ctxn
] = ctx
;
2015 next_ctx
->task
= task
;
2018 perf_event_sync_stat(ctx
, next_ctx
);
2020 raw_spin_unlock(&next_ctx
->lock
);
2021 raw_spin_unlock(&ctx
->lock
);
2026 raw_spin_lock(&ctx
->lock
);
2027 ctx_sched_out(ctx
, cpuctx
, EVENT_ALL
);
2028 cpuctx
->task_ctx
= NULL
;
2029 raw_spin_unlock(&ctx
->lock
);
2033 #define for_each_task_context_nr(ctxn) \
2034 for ((ctxn) = 0; (ctxn) < perf_nr_task_contexts; (ctxn)++)
2037 * Called from scheduler to remove the events of the current task,
2038 * with interrupts disabled.
2040 * We stop each event and update the event value in event->count.
2042 * This does not protect us against NMI, but disable()
2043 * sets the disabled bit in the control field of event _before_
2044 * accessing the event control register. If a NMI hits, then it will
2045 * not restart the event.
2047 void __perf_event_task_sched_out(struct task_struct
*task
,
2048 struct task_struct
*next
)
2052 for_each_task_context_nr(ctxn
)
2053 perf_event_context_sched_out(task
, ctxn
, next
);
2056 * if cgroup events exist on this CPU, then we need
2057 * to check if we have to switch out PMU state.
2058 * cgroup event are system-wide mode only
2060 if (atomic_read(&__get_cpu_var(perf_cgroup_events
)))
2061 perf_cgroup_sched_out(task
, next
);
2064 static void task_ctx_sched_out(struct perf_event_context
*ctx
)
2066 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
2068 if (!cpuctx
->task_ctx
)
2071 if (WARN_ON_ONCE(ctx
!= cpuctx
->task_ctx
))
2074 ctx_sched_out(ctx
, cpuctx
, EVENT_ALL
);
2075 cpuctx
->task_ctx
= NULL
;
2079 * Called with IRQs disabled
2081 static void cpu_ctx_sched_out(struct perf_cpu_context
*cpuctx
,
2082 enum event_type_t event_type
)
2084 ctx_sched_out(&cpuctx
->ctx
, cpuctx
, event_type
);
2088 ctx_pinned_sched_in(struct perf_event_context
*ctx
,
2089 struct perf_cpu_context
*cpuctx
)
2091 struct perf_event
*event
;
2093 list_for_each_entry(event
, &ctx
->pinned_groups
, group_entry
) {
2094 if (event
->state
<= PERF_EVENT_STATE_OFF
)
2096 if (!event_filter_match(event
))
2099 /* may need to reset tstamp_enabled */
2100 if (is_cgroup_event(event
))
2101 perf_cgroup_mark_enabled(event
, ctx
);
2103 if (group_can_go_on(event
, cpuctx
, 1))
2104 group_sched_in(event
, cpuctx
, ctx
);
2107 * If this pinned group hasn't been scheduled,
2108 * put it in error state.
2110 if (event
->state
== PERF_EVENT_STATE_INACTIVE
) {
2111 update_group_times(event
);
2112 event
->state
= PERF_EVENT_STATE_ERROR
;
2118 ctx_flexible_sched_in(struct perf_event_context
*ctx
,
2119 struct perf_cpu_context
*cpuctx
)
2121 struct perf_event
*event
;
2124 list_for_each_entry(event
, &ctx
->flexible_groups
, group_entry
) {
2125 /* Ignore events in OFF or ERROR state */
2126 if (event
->state
<= PERF_EVENT_STATE_OFF
)
2129 * Listen to the 'cpu' scheduling filter constraint
2132 if (!event_filter_match(event
))
2135 /* may need to reset tstamp_enabled */
2136 if (is_cgroup_event(event
))
2137 perf_cgroup_mark_enabled(event
, ctx
);
2139 if (group_can_go_on(event
, cpuctx
, can_add_hw
)) {
2140 if (group_sched_in(event
, cpuctx
, ctx
))
2147 ctx_sched_in(struct perf_event_context
*ctx
,
2148 struct perf_cpu_context
*cpuctx
,
2149 enum event_type_t event_type
,
2150 struct task_struct
*task
)
2153 int is_active
= ctx
->is_active
;
2155 ctx
->is_active
|= event_type
;
2156 if (likely(!ctx
->nr_events
))
2160 ctx
->timestamp
= now
;
2161 perf_cgroup_set_timestamp(task
, ctx
);
2163 * First go through the list and put on any pinned groups
2164 * in order to give them the best chance of going on.
2166 if (!(is_active
& EVENT_PINNED
) && (event_type
& EVENT_PINNED
))
2167 ctx_pinned_sched_in(ctx
, cpuctx
);
2169 /* Then walk through the lower prio flexible groups */
2170 if (!(is_active
& EVENT_FLEXIBLE
) && (event_type
& EVENT_FLEXIBLE
))
2171 ctx_flexible_sched_in(ctx
, cpuctx
);
2174 static void cpu_ctx_sched_in(struct perf_cpu_context
*cpuctx
,
2175 enum event_type_t event_type
,
2176 struct task_struct
*task
)
2178 struct perf_event_context
*ctx
= &cpuctx
->ctx
;
2180 ctx_sched_in(ctx
, cpuctx
, event_type
, task
);
2183 static void perf_event_context_sched_in(struct perf_event_context
*ctx
,
2184 struct task_struct
*task
)
2186 struct perf_cpu_context
*cpuctx
;
2188 cpuctx
= __get_cpu_context(ctx
);
2189 if (cpuctx
->task_ctx
== ctx
)
2192 perf_ctx_lock(cpuctx
, ctx
);
2193 perf_pmu_disable(ctx
->pmu
);
2195 * We want to keep the following priority order:
2196 * cpu pinned (that don't need to move), task pinned,
2197 * cpu flexible, task flexible.
2199 cpu_ctx_sched_out(cpuctx
, EVENT_FLEXIBLE
);
2202 cpuctx
->task_ctx
= ctx
;
2204 perf_event_sched_in(cpuctx
, cpuctx
->task_ctx
, task
);
2206 perf_pmu_enable(ctx
->pmu
);
2207 perf_ctx_unlock(cpuctx
, ctx
);
2210 * Since these rotations are per-cpu, we need to ensure the
2211 * cpu-context we got scheduled on is actually rotating.
2213 perf_pmu_rotate_start(ctx
->pmu
);
2217 * When sampling the branck stack in system-wide, it may be necessary
2218 * to flush the stack on context switch. This happens when the branch
2219 * stack does not tag its entries with the pid of the current task.
2220 * Otherwise it becomes impossible to associate a branch entry with a
2221 * task. This ambiguity is more likely to appear when the branch stack
2222 * supports priv level filtering and the user sets it to monitor only
2223 * at the user level (which could be a useful measurement in system-wide
2224 * mode). In that case, the risk is high of having a branch stack with
2225 * branch from multiple tasks. Flushing may mean dropping the existing
2226 * entries or stashing them somewhere in the PMU specific code layer.
2228 * This function provides the context switch callback to the lower code
2229 * layer. It is invoked ONLY when there is at least one system-wide context
2230 * with at least one active event using taken branch sampling.
2232 static void perf_branch_stack_sched_in(struct task_struct
*prev
,
2233 struct task_struct
*task
)
2235 struct perf_cpu_context
*cpuctx
;
2237 unsigned long flags
;
2239 /* no need to flush branch stack if not changing task */
2243 local_irq_save(flags
);
2247 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
2248 cpuctx
= this_cpu_ptr(pmu
->pmu_cpu_context
);
2251 * check if the context has at least one
2252 * event using PERF_SAMPLE_BRANCH_STACK
2254 if (cpuctx
->ctx
.nr_branch_stack
> 0
2255 && pmu
->flush_branch_stack
) {
2257 pmu
= cpuctx
->ctx
.pmu
;
2259 perf_ctx_lock(cpuctx
, cpuctx
->task_ctx
);
2261 perf_pmu_disable(pmu
);
2263 pmu
->flush_branch_stack();
2265 perf_pmu_enable(pmu
);
2267 perf_ctx_unlock(cpuctx
, cpuctx
->task_ctx
);
2273 local_irq_restore(flags
);
2277 * Called from scheduler to add the events of the current task
2278 * with interrupts disabled.
2280 * We restore the event value and then enable it.
2282 * This does not protect us against NMI, but enable()
2283 * sets the enabled bit in the control field of event _before_
2284 * accessing the event control register. If a NMI hits, then it will
2285 * keep the event running.
2287 void __perf_event_task_sched_in(struct task_struct
*prev
,
2288 struct task_struct
*task
)
2290 struct perf_event_context
*ctx
;
2293 for_each_task_context_nr(ctxn
) {
2294 ctx
= task
->perf_event_ctxp
[ctxn
];
2298 perf_event_context_sched_in(ctx
, task
);
2301 * if cgroup events exist on this CPU, then we need
2302 * to check if we have to switch in PMU state.
2303 * cgroup event are system-wide mode only
2305 if (atomic_read(&__get_cpu_var(perf_cgroup_events
)))
2306 perf_cgroup_sched_in(prev
, task
);
2308 /* check for system-wide branch_stack events */
2309 if (atomic_read(&__get_cpu_var(perf_branch_stack_events
)))
2310 perf_branch_stack_sched_in(prev
, task
);
2313 static u64
perf_calculate_period(struct perf_event
*event
, u64 nsec
, u64 count
)
2315 u64 frequency
= event
->attr
.sample_freq
;
2316 u64 sec
= NSEC_PER_SEC
;
2317 u64 divisor
, dividend
;
2319 int count_fls
, nsec_fls
, frequency_fls
, sec_fls
;
2321 count_fls
= fls64(count
);
2322 nsec_fls
= fls64(nsec
);
2323 frequency_fls
= fls64(frequency
);
2327 * We got @count in @nsec, with a target of sample_freq HZ
2328 * the target period becomes:
2331 * period = -------------------
2332 * @nsec * sample_freq
2337 * Reduce accuracy by one bit such that @a and @b converge
2338 * to a similar magnitude.
2340 #define REDUCE_FLS(a, b) \
2342 if (a##_fls > b##_fls) { \
2352 * Reduce accuracy until either term fits in a u64, then proceed with
2353 * the other, so that finally we can do a u64/u64 division.
2355 while (count_fls
+ sec_fls
> 64 && nsec_fls
+ frequency_fls
> 64) {
2356 REDUCE_FLS(nsec
, frequency
);
2357 REDUCE_FLS(sec
, count
);
2360 if (count_fls
+ sec_fls
> 64) {
2361 divisor
= nsec
* frequency
;
2363 while (count_fls
+ sec_fls
> 64) {
2364 REDUCE_FLS(count
, sec
);
2368 dividend
= count
* sec
;
2370 dividend
= count
* sec
;
2372 while (nsec_fls
+ frequency_fls
> 64) {
2373 REDUCE_FLS(nsec
, frequency
);
2377 divisor
= nsec
* frequency
;
2383 return div64_u64(dividend
, divisor
);
2386 static DEFINE_PER_CPU(int, perf_throttled_count
);
2387 static DEFINE_PER_CPU(u64
, perf_throttled_seq
);
2389 static void perf_adjust_period(struct perf_event
*event
, u64 nsec
, u64 count
, bool disable
)
2391 struct hw_perf_event
*hwc
= &event
->hw
;
2392 s64 period
, sample_period
;
2395 period
= perf_calculate_period(event
, nsec
, count
);
2397 delta
= (s64
)(period
- hwc
->sample_period
);
2398 delta
= (delta
+ 7) / 8; /* low pass filter */
2400 sample_period
= hwc
->sample_period
+ delta
;
2405 hwc
->sample_period
= sample_period
;
2407 if (local64_read(&hwc
->period_left
) > 8*sample_period
) {
2409 event
->pmu
->stop(event
, PERF_EF_UPDATE
);
2411 local64_set(&hwc
->period_left
, 0);
2414 event
->pmu
->start(event
, PERF_EF_RELOAD
);
2419 * combine freq adjustment with unthrottling to avoid two passes over the
2420 * events. At the same time, make sure, having freq events does not change
2421 * the rate of unthrottling as that would introduce bias.
2423 static void perf_adjust_freq_unthr_context(struct perf_event_context
*ctx
,
2426 struct perf_event
*event
;
2427 struct hw_perf_event
*hwc
;
2428 u64 now
, period
= TICK_NSEC
;
2432 * only need to iterate over all events iff:
2433 * - context have events in frequency mode (needs freq adjust)
2434 * - there are events to unthrottle on this cpu
2436 if (!(ctx
->nr_freq
|| needs_unthr
))
2439 raw_spin_lock(&ctx
->lock
);
2440 perf_pmu_disable(ctx
->pmu
);
2442 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
2443 if (event
->state
!= PERF_EVENT_STATE_ACTIVE
)
2446 if (!event_filter_match(event
))
2451 if (needs_unthr
&& hwc
->interrupts
== MAX_INTERRUPTS
) {
2452 hwc
->interrupts
= 0;
2453 perf_log_throttle(event
, 1);
2454 event
->pmu
->start(event
, 0);
2457 if (!event
->attr
.freq
|| !event
->attr
.sample_freq
)
2461 * stop the event and update event->count
2463 event
->pmu
->stop(event
, PERF_EF_UPDATE
);
2465 now
= local64_read(&event
->count
);
2466 delta
= now
- hwc
->freq_count_stamp
;
2467 hwc
->freq_count_stamp
= now
;
2471 * reload only if value has changed
2472 * we have stopped the event so tell that
2473 * to perf_adjust_period() to avoid stopping it
2477 perf_adjust_period(event
, period
, delta
, false);
2479 event
->pmu
->start(event
, delta
> 0 ? PERF_EF_RELOAD
: 0);
2482 perf_pmu_enable(ctx
->pmu
);
2483 raw_spin_unlock(&ctx
->lock
);
2487 * Round-robin a context's events:
2489 static void rotate_ctx(struct perf_event_context
*ctx
)
2492 * Rotate the first entry last of non-pinned groups. Rotation might be
2493 * disabled by the inheritance code.
2495 if (!ctx
->rotate_disable
)
2496 list_rotate_left(&ctx
->flexible_groups
);
2500 * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
2501 * because they're strictly cpu affine and rotate_start is called with IRQs
2502 * disabled, while rotate_context is called from IRQ context.
2504 static void perf_rotate_context(struct perf_cpu_context
*cpuctx
)
2506 struct perf_event_context
*ctx
= NULL
;
2507 int rotate
= 0, remove
= 1;
2509 if (cpuctx
->ctx
.nr_events
) {
2511 if (cpuctx
->ctx
.nr_events
!= cpuctx
->ctx
.nr_active
)
2515 ctx
= cpuctx
->task_ctx
;
2516 if (ctx
&& ctx
->nr_events
) {
2518 if (ctx
->nr_events
!= ctx
->nr_active
)
2525 perf_ctx_lock(cpuctx
, cpuctx
->task_ctx
);
2526 perf_pmu_disable(cpuctx
->ctx
.pmu
);
2528 cpu_ctx_sched_out(cpuctx
, EVENT_FLEXIBLE
);
2530 ctx_sched_out(ctx
, cpuctx
, EVENT_FLEXIBLE
);
2532 rotate_ctx(&cpuctx
->ctx
);
2536 perf_event_sched_in(cpuctx
, ctx
, current
);
2538 perf_pmu_enable(cpuctx
->ctx
.pmu
);
2539 perf_ctx_unlock(cpuctx
, cpuctx
->task_ctx
);
2542 list_del_init(&cpuctx
->rotation_list
);
2545 void perf_event_task_tick(void)
2547 struct list_head
*head
= &__get_cpu_var(rotation_list
);
2548 struct perf_cpu_context
*cpuctx
, *tmp
;
2549 struct perf_event_context
*ctx
;
2552 WARN_ON(!irqs_disabled());
2554 __this_cpu_inc(perf_throttled_seq
);
2555 throttled
= __this_cpu_xchg(perf_throttled_count
, 0);
2557 list_for_each_entry_safe(cpuctx
, tmp
, head
, rotation_list
) {
2559 perf_adjust_freq_unthr_context(ctx
, throttled
);
2561 ctx
= cpuctx
->task_ctx
;
2563 perf_adjust_freq_unthr_context(ctx
, throttled
);
2565 if (cpuctx
->jiffies_interval
== 1 ||
2566 !(jiffies
% cpuctx
->jiffies_interval
))
2567 perf_rotate_context(cpuctx
);
2571 static int event_enable_on_exec(struct perf_event
*event
,
2572 struct perf_event_context
*ctx
)
2574 if (!event
->attr
.enable_on_exec
)
2577 event
->attr
.enable_on_exec
= 0;
2578 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
)
2581 __perf_event_mark_enabled(event
);
2587 * Enable all of a task's events that have been marked enable-on-exec.
2588 * This expects task == current.
2590 static void perf_event_enable_on_exec(struct perf_event_context
*ctx
)
2592 struct perf_event
*event
;
2593 unsigned long flags
;
2597 local_irq_save(flags
);
2598 if (!ctx
|| !ctx
->nr_events
)
2602 * We must ctxsw out cgroup events to avoid conflict
2603 * when invoking perf_task_event_sched_in() later on
2604 * in this function. Otherwise we end up trying to
2605 * ctxswin cgroup events which are already scheduled
2608 perf_cgroup_sched_out(current
, NULL
);
2610 raw_spin_lock(&ctx
->lock
);
2611 task_ctx_sched_out(ctx
);
2613 list_for_each_entry(event
, &ctx
->event_list
, event_entry
) {
2614 ret
= event_enable_on_exec(event
, ctx
);
2620 * Unclone this context if we enabled any event.
2625 raw_spin_unlock(&ctx
->lock
);
2628 * Also calls ctxswin for cgroup events, if any:
2630 perf_event_context_sched_in(ctx
, ctx
->task
);
2632 local_irq_restore(flags
);
2636 * Cross CPU call to read the hardware event
2638 static void __perf_event_read(void *info
)
2640 struct perf_event
*event
= info
;
2641 struct perf_event_context
*ctx
= event
->ctx
;
2642 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
2645 * If this is a task context, we need to check whether it is
2646 * the current task context of this cpu. If not it has been
2647 * scheduled out before the smp call arrived. In that case
2648 * event->count would have been updated to a recent sample
2649 * when the event was scheduled out.
2651 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
)
2654 raw_spin_lock(&ctx
->lock
);
2655 if (ctx
->is_active
) {
2656 update_context_time(ctx
);
2657 update_cgrp_time_from_event(event
);
2659 update_event_times(event
);
2660 if (event
->state
== PERF_EVENT_STATE_ACTIVE
)
2661 event
->pmu
->read(event
);
2662 raw_spin_unlock(&ctx
->lock
);
2665 static inline u64
perf_event_count(struct perf_event
*event
)
2667 return local64_read(&event
->count
) + atomic64_read(&event
->child_count
);
2670 static u64
perf_event_read(struct perf_event
*event
)
2673 * If event is enabled and currently active on a CPU, update the
2674 * value in the event structure:
2676 if (event
->state
== PERF_EVENT_STATE_ACTIVE
) {
2677 smp_call_function_single(event
->oncpu
,
2678 __perf_event_read
, event
, 1);
2679 } else if (event
->state
== PERF_EVENT_STATE_INACTIVE
) {
2680 struct perf_event_context
*ctx
= event
->ctx
;
2681 unsigned long flags
;
2683 raw_spin_lock_irqsave(&ctx
->lock
, flags
);
2685 * may read while context is not active
2686 * (e.g., thread is blocked), in that case
2687 * we cannot update context time
2689 if (ctx
->is_active
) {
2690 update_context_time(ctx
);
2691 update_cgrp_time_from_event(event
);
2693 update_event_times(event
);
2694 raw_spin_unlock_irqrestore(&ctx
->lock
, flags
);
2697 return perf_event_count(event
);
2701 * Initialize the perf_event context in a task_struct:
2703 static void __perf_event_init_context(struct perf_event_context
*ctx
)
2705 raw_spin_lock_init(&ctx
->lock
);
2706 mutex_init(&ctx
->mutex
);
2707 INIT_LIST_HEAD(&ctx
->pinned_groups
);
2708 INIT_LIST_HEAD(&ctx
->flexible_groups
);
2709 INIT_LIST_HEAD(&ctx
->event_list
);
2710 atomic_set(&ctx
->refcount
, 1);
2713 static struct perf_event_context
*
2714 alloc_perf_context(struct pmu
*pmu
, struct task_struct
*task
)
2716 struct perf_event_context
*ctx
;
2718 ctx
= kzalloc(sizeof(struct perf_event_context
), GFP_KERNEL
);
2722 __perf_event_init_context(ctx
);
2725 get_task_struct(task
);
2732 static struct task_struct
*
2733 find_lively_task_by_vpid(pid_t vpid
)
2735 struct task_struct
*task
;
2742 task
= find_task_by_vpid(vpid
);
2744 get_task_struct(task
);
2748 return ERR_PTR(-ESRCH
);
2750 /* Reuse ptrace permission checks for now. */
2752 if (!ptrace_may_access(task
, PTRACE_MODE_READ
))
2757 put_task_struct(task
);
2758 return ERR_PTR(err
);
2763 * Returns a matching context with refcount and pincount.
2765 static struct perf_event_context
*
2766 find_get_context(struct pmu
*pmu
, struct task_struct
*task
, int cpu
)
2768 struct perf_event_context
*ctx
;
2769 struct perf_cpu_context
*cpuctx
;
2770 unsigned long flags
;
2774 /* Must be root to operate on a CPU event: */
2775 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN
))
2776 return ERR_PTR(-EACCES
);
2779 * We could be clever and allow to attach a event to an
2780 * offline CPU and activate it when the CPU comes up, but
2783 if (!cpu_online(cpu
))
2784 return ERR_PTR(-ENODEV
);
2786 cpuctx
= per_cpu_ptr(pmu
->pmu_cpu_context
, cpu
);
2795 ctxn
= pmu
->task_ctx_nr
;
2800 ctx
= perf_lock_task_context(task
, ctxn
, &flags
);
2804 raw_spin_unlock_irqrestore(&ctx
->lock
, flags
);
2806 ctx
= alloc_perf_context(pmu
, task
);
2812 mutex_lock(&task
->perf_event_mutex
);
2814 * If it has already passed perf_event_exit_task().
2815 * we must see PF_EXITING, it takes this mutex too.
2817 if (task
->flags
& PF_EXITING
)
2819 else if (task
->perf_event_ctxp
[ctxn
])
2824 rcu_assign_pointer(task
->perf_event_ctxp
[ctxn
], ctx
);
2826 mutex_unlock(&task
->perf_event_mutex
);
2828 if (unlikely(err
)) {
2840 return ERR_PTR(err
);
2843 static void perf_event_free_filter(struct perf_event
*event
);
2845 static void free_event_rcu(struct rcu_head
*head
)
2847 struct perf_event
*event
;
2849 event
= container_of(head
, struct perf_event
, rcu_head
);
2851 put_pid_ns(event
->ns
);
2852 perf_event_free_filter(event
);
2856 static void ring_buffer_put(struct ring_buffer
*rb
);
2858 static void free_event(struct perf_event
*event
)
2860 irq_work_sync(&event
->pending
);
2862 if (!event
->parent
) {
2863 if (event
->attach_state
& PERF_ATTACH_TASK
)
2864 static_key_slow_dec_deferred(&perf_sched_events
);
2865 if (event
->attr
.mmap
|| event
->attr
.mmap_data
)
2866 atomic_dec(&nr_mmap_events
);
2867 if (event
->attr
.comm
)
2868 atomic_dec(&nr_comm_events
);
2869 if (event
->attr
.task
)
2870 atomic_dec(&nr_task_events
);
2871 if (event
->attr
.sample_type
& PERF_SAMPLE_CALLCHAIN
)
2872 put_callchain_buffers();
2873 if (is_cgroup_event(event
)) {
2874 atomic_dec(&per_cpu(perf_cgroup_events
, event
->cpu
));
2875 static_key_slow_dec_deferred(&perf_sched_events
);
2878 if (has_branch_stack(event
)) {
2879 static_key_slow_dec_deferred(&perf_sched_events
);
2880 /* is system-wide event */
2881 if (!(event
->attach_state
& PERF_ATTACH_TASK
))
2882 atomic_dec(&per_cpu(perf_branch_stack_events
,
2888 ring_buffer_put(event
->rb
);
2892 if (is_cgroup_event(event
))
2893 perf_detach_cgroup(event
);
2896 event
->destroy(event
);
2899 put_ctx(event
->ctx
);
2901 call_rcu(&event
->rcu_head
, free_event_rcu
);
2904 int perf_event_release_kernel(struct perf_event
*event
)
2906 struct perf_event_context
*ctx
= event
->ctx
;
2908 WARN_ON_ONCE(ctx
->parent_ctx
);
2910 * There are two ways this annotation is useful:
2912 * 1) there is a lock recursion from perf_event_exit_task
2913 * see the comment there.
2915 * 2) there is a lock-inversion with mmap_sem through
2916 * perf_event_read_group(), which takes faults while
2917 * holding ctx->mutex, however this is called after
2918 * the last filedesc died, so there is no possibility
2919 * to trigger the AB-BA case.
2921 mutex_lock_nested(&ctx
->mutex
, SINGLE_DEPTH_NESTING
);
2922 raw_spin_lock_irq(&ctx
->lock
);
2923 perf_group_detach(event
);
2924 raw_spin_unlock_irq(&ctx
->lock
);
2925 perf_remove_from_context(event
);
2926 mutex_unlock(&ctx
->mutex
);
2932 EXPORT_SYMBOL_GPL(perf_event_release_kernel
);
2935 * Called when the last reference to the file is gone.
2937 static void put_event(struct perf_event
*event
)
2939 struct task_struct
*owner
;
2941 if (!atomic_long_dec_and_test(&event
->refcount
))
2945 owner
= ACCESS_ONCE(event
->owner
);
2947 * Matches the smp_wmb() in perf_event_exit_task(). If we observe
2948 * !owner it means the list deletion is complete and we can indeed
2949 * free this event, otherwise we need to serialize on
2950 * owner->perf_event_mutex.
2952 smp_read_barrier_depends();
2955 * Since delayed_put_task_struct() also drops the last
2956 * task reference we can safely take a new reference
2957 * while holding the rcu_read_lock().
2959 get_task_struct(owner
);
2964 mutex_lock(&owner
->perf_event_mutex
);
2966 * We have to re-check the event->owner field, if it is cleared
2967 * we raced with perf_event_exit_task(), acquiring the mutex
2968 * ensured they're done, and we can proceed with freeing the
2972 list_del_init(&event
->owner_entry
);
2973 mutex_unlock(&owner
->perf_event_mutex
);
2974 put_task_struct(owner
);
2977 perf_event_release_kernel(event
);
2980 static int perf_release(struct inode
*inode
, struct file
*file
)
2982 put_event(file
->private_data
);
2986 u64
perf_event_read_value(struct perf_event
*event
, u64
*enabled
, u64
*running
)
2988 struct perf_event
*child
;
2994 mutex_lock(&event
->child_mutex
);
2995 total
+= perf_event_read(event
);
2996 *enabled
+= event
->total_time_enabled
+
2997 atomic64_read(&event
->child_total_time_enabled
);
2998 *running
+= event
->total_time_running
+
2999 atomic64_read(&event
->child_total_time_running
);
3001 list_for_each_entry(child
, &event
->child_list
, child_list
) {
3002 total
+= perf_event_read(child
);
3003 *enabled
+= child
->total_time_enabled
;
3004 *running
+= child
->total_time_running
;
3006 mutex_unlock(&event
->child_mutex
);
3010 EXPORT_SYMBOL_GPL(perf_event_read_value
);
3012 static int perf_event_read_group(struct perf_event
*event
,
3013 u64 read_format
, char __user
*buf
)
3015 struct perf_event
*leader
= event
->group_leader
, *sub
;
3016 int n
= 0, size
= 0, ret
= -EFAULT
;
3017 struct perf_event_context
*ctx
= leader
->ctx
;
3019 u64 count
, enabled
, running
;
3021 mutex_lock(&ctx
->mutex
);
3022 count
= perf_event_read_value(leader
, &enabled
, &running
);
3024 values
[n
++] = 1 + leader
->nr_siblings
;
3025 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
3026 values
[n
++] = enabled
;
3027 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
3028 values
[n
++] = running
;
3029 values
[n
++] = count
;
3030 if (read_format
& PERF_FORMAT_ID
)
3031 values
[n
++] = primary_event_id(leader
);
3033 size
= n
* sizeof(u64
);
3035 if (copy_to_user(buf
, values
, size
))
3040 list_for_each_entry(sub
, &leader
->sibling_list
, group_entry
) {
3043 values
[n
++] = perf_event_read_value(sub
, &enabled
, &running
);
3044 if (read_format
& PERF_FORMAT_ID
)
3045 values
[n
++] = primary_event_id(sub
);
3047 size
= n
* sizeof(u64
);
3049 if (copy_to_user(buf
+ ret
, values
, size
)) {
3057 mutex_unlock(&ctx
->mutex
);
3062 static int perf_event_read_one(struct perf_event
*event
,
3063 u64 read_format
, char __user
*buf
)
3065 u64 enabled
, running
;
3069 values
[n
++] = perf_event_read_value(event
, &enabled
, &running
);
3070 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
3071 values
[n
++] = enabled
;
3072 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
3073 values
[n
++] = running
;
3074 if (read_format
& PERF_FORMAT_ID
)
3075 values
[n
++] = primary_event_id(event
);
3077 if (copy_to_user(buf
, values
, n
* sizeof(u64
)))
3080 return n
* sizeof(u64
);
3084 * Read the performance event - simple non blocking version for now
3087 perf_read_hw(struct perf_event
*event
, char __user
*buf
, size_t count
)
3089 u64 read_format
= event
->attr
.read_format
;
3093 * Return end-of-file for a read on a event that is in
3094 * error state (i.e. because it was pinned but it couldn't be
3095 * scheduled on to the CPU at some point).
3097 if (event
->state
== PERF_EVENT_STATE_ERROR
)
3100 if (count
< event
->read_size
)
3103 WARN_ON_ONCE(event
->ctx
->parent_ctx
);
3104 if (read_format
& PERF_FORMAT_GROUP
)
3105 ret
= perf_event_read_group(event
, read_format
, buf
);
3107 ret
= perf_event_read_one(event
, read_format
, buf
);
3113 perf_read(struct file
*file
, char __user
*buf
, size_t count
, loff_t
*ppos
)
3115 struct perf_event
*event
= file
->private_data
;
3117 return perf_read_hw(event
, buf
, count
);
3120 static unsigned int perf_poll(struct file
*file
, poll_table
*wait
)
3122 struct perf_event
*event
= file
->private_data
;
3123 struct ring_buffer
*rb
;
3124 unsigned int events
= POLL_HUP
;
3127 * Race between perf_event_set_output() and perf_poll(): perf_poll()
3128 * grabs the rb reference but perf_event_set_output() overrides it.
3129 * Here is the timeline for two threads T1, T2:
3130 * t0: T1, rb = rcu_dereference(event->rb)
3131 * t1: T2, old_rb = event->rb
3132 * t2: T2, event->rb = new rb
3133 * t3: T2, ring_buffer_detach(old_rb)
3134 * t4: T1, ring_buffer_attach(rb1)
3135 * t5: T1, poll_wait(event->waitq)
3137 * To avoid this problem, we grab mmap_mutex in perf_poll()
3138 * thereby ensuring that the assignment of the new ring buffer
3139 * and the detachment of the old buffer appear atomic to perf_poll()
3141 mutex_lock(&event
->mmap_mutex
);
3144 rb
= rcu_dereference(event
->rb
);
3146 ring_buffer_attach(event
, rb
);
3147 events
= atomic_xchg(&rb
->poll
, 0);
3151 mutex_unlock(&event
->mmap_mutex
);
3153 poll_wait(file
, &event
->waitq
, wait
);
3158 static void perf_event_reset(struct perf_event
*event
)
3160 (void)perf_event_read(event
);
3161 local64_set(&event
->count
, 0);
3162 perf_event_update_userpage(event
);
3166 * Holding the top-level event's child_mutex means that any
3167 * descendant process that has inherited this event will block
3168 * in sync_child_event if it goes to exit, thus satisfying the
3169 * task existence requirements of perf_event_enable/disable.
3171 static void perf_event_for_each_child(struct perf_event
*event
,
3172 void (*func
)(struct perf_event
*))
3174 struct perf_event
*child
;
3176 WARN_ON_ONCE(event
->ctx
->parent_ctx
);
3177 mutex_lock(&event
->child_mutex
);
3179 list_for_each_entry(child
, &event
->child_list
, child_list
)
3181 mutex_unlock(&event
->child_mutex
);
3184 static void perf_event_for_each(struct perf_event
*event
,
3185 void (*func
)(struct perf_event
*))
3187 struct perf_event_context
*ctx
= event
->ctx
;
3188 struct perf_event
*sibling
;
3190 WARN_ON_ONCE(ctx
->parent_ctx
);
3191 mutex_lock(&ctx
->mutex
);
3192 event
= event
->group_leader
;
3194 perf_event_for_each_child(event
, func
);
3195 list_for_each_entry(sibling
, &event
->sibling_list
, group_entry
)
3196 perf_event_for_each_child(sibling
, func
);
3197 mutex_unlock(&ctx
->mutex
);
3200 static int perf_event_period(struct perf_event
*event
, u64 __user
*arg
)
3202 struct perf_event_context
*ctx
= event
->ctx
;
3206 if (!is_sampling_event(event
))
3209 if (copy_from_user(&value
, arg
, sizeof(value
)))
3215 raw_spin_lock_irq(&ctx
->lock
);
3216 if (event
->attr
.freq
) {
3217 if (value
> sysctl_perf_event_sample_rate
) {
3222 event
->attr
.sample_freq
= value
;
3224 event
->attr
.sample_period
= value
;
3225 event
->hw
.sample_period
= value
;
3228 raw_spin_unlock_irq(&ctx
->lock
);
3233 static const struct file_operations perf_fops
;
3235 static inline int perf_fget_light(int fd
, struct fd
*p
)
3237 struct fd f
= fdget(fd
);
3241 if (f
.file
->f_op
!= &perf_fops
) {
3249 static int perf_event_set_output(struct perf_event
*event
,
3250 struct perf_event
*output_event
);
3251 static int perf_event_set_filter(struct perf_event
*event
, void __user
*arg
);
3253 static long perf_ioctl(struct file
*file
, unsigned int cmd
, unsigned long arg
)
3255 struct perf_event
*event
= file
->private_data
;
3256 void (*func
)(struct perf_event
*);
3260 case PERF_EVENT_IOC_ENABLE
:
3261 func
= perf_event_enable
;
3263 case PERF_EVENT_IOC_DISABLE
:
3264 func
= perf_event_disable
;
3266 case PERF_EVENT_IOC_RESET
:
3267 func
= perf_event_reset
;
3270 case PERF_EVENT_IOC_REFRESH
:
3271 return perf_event_refresh(event
, arg
);
3273 case PERF_EVENT_IOC_PERIOD
:
3274 return perf_event_period(event
, (u64 __user
*)arg
);
3276 case PERF_EVENT_IOC_SET_OUTPUT
:
3280 struct perf_event
*output_event
;
3282 ret
= perf_fget_light(arg
, &output
);
3285 output_event
= output
.file
->private_data
;
3286 ret
= perf_event_set_output(event
, output_event
);
3289 ret
= perf_event_set_output(event
, NULL
);
3294 case PERF_EVENT_IOC_SET_FILTER
:
3295 return perf_event_set_filter(event
, (void __user
*)arg
);
3301 if (flags
& PERF_IOC_FLAG_GROUP
)
3302 perf_event_for_each(event
, func
);
3304 perf_event_for_each_child(event
, func
);
3309 int perf_event_task_enable(void)
3311 struct perf_event
*event
;
3313 mutex_lock(¤t
->perf_event_mutex
);
3314 list_for_each_entry(event
, ¤t
->perf_event_list
, owner_entry
)
3315 perf_event_for_each_child(event
, perf_event_enable
);
3316 mutex_unlock(¤t
->perf_event_mutex
);
3321 int perf_event_task_disable(void)
3323 struct perf_event
*event
;
3325 mutex_lock(¤t
->perf_event_mutex
);
3326 list_for_each_entry(event
, ¤t
->perf_event_list
, owner_entry
)
3327 perf_event_for_each_child(event
, perf_event_disable
);
3328 mutex_unlock(¤t
->perf_event_mutex
);
3333 static int perf_event_index(struct perf_event
*event
)
3335 if (event
->hw
.state
& PERF_HES_STOPPED
)
3338 if (event
->state
!= PERF_EVENT_STATE_ACTIVE
)
3341 return event
->pmu
->event_idx(event
);
3344 static void calc_timer_values(struct perf_event
*event
,
3351 *now
= perf_clock();
3352 ctx_time
= event
->shadow_ctx_time
+ *now
;
3353 *enabled
= ctx_time
- event
->tstamp_enabled
;
3354 *running
= ctx_time
- event
->tstamp_running
;
3357 void __weak
arch_perf_update_userpage(struct perf_event_mmap_page
*userpg
, u64 now
)
3362 * Callers need to ensure there can be no nesting of this function, otherwise
3363 * the seqlock logic goes bad. We can not serialize this because the arch
3364 * code calls this from NMI context.
3366 void perf_event_update_userpage(struct perf_event
*event
)
3368 struct perf_event_mmap_page
*userpg
;
3369 struct ring_buffer
*rb
;
3370 u64 enabled
, running
, now
;
3374 * compute total_time_enabled, total_time_running
3375 * based on snapshot values taken when the event
3376 * was last scheduled in.
3378 * we cannot simply called update_context_time()
3379 * because of locking issue as we can be called in
3382 calc_timer_values(event
, &now
, &enabled
, &running
);
3383 rb
= rcu_dereference(event
->rb
);
3387 userpg
= rb
->user_page
;
3390 * Disable preemption so as to not let the corresponding user-space
3391 * spin too long if we get preempted.
3396 userpg
->index
= perf_event_index(event
);
3397 userpg
->offset
= perf_event_count(event
);
3399 userpg
->offset
-= local64_read(&event
->hw
.prev_count
);
3401 userpg
->time_enabled
= enabled
+
3402 atomic64_read(&event
->child_total_time_enabled
);
3404 userpg
->time_running
= running
+
3405 atomic64_read(&event
->child_total_time_running
);
3407 arch_perf_update_userpage(userpg
, now
);
3416 static int perf_mmap_fault(struct vm_area_struct
*vma
, struct vm_fault
*vmf
)
3418 struct perf_event
*event
= vma
->vm_file
->private_data
;
3419 struct ring_buffer
*rb
;
3420 int ret
= VM_FAULT_SIGBUS
;
3422 if (vmf
->flags
& FAULT_FLAG_MKWRITE
) {
3423 if (vmf
->pgoff
== 0)
3429 rb
= rcu_dereference(event
->rb
);
3433 if (vmf
->pgoff
&& (vmf
->flags
& FAULT_FLAG_WRITE
))
3436 vmf
->page
= perf_mmap_to_page(rb
, vmf
->pgoff
);
3440 get_page(vmf
->page
);
3441 vmf
->page
->mapping
= vma
->vm_file
->f_mapping
;
3442 vmf
->page
->index
= vmf
->pgoff
;
3451 static void ring_buffer_attach(struct perf_event
*event
,
3452 struct ring_buffer
*rb
)
3454 unsigned long flags
;
3456 if (!list_empty(&event
->rb_entry
))
3459 spin_lock_irqsave(&rb
->event_lock
, flags
);
3460 if (!list_empty(&event
->rb_entry
))
3463 list_add(&event
->rb_entry
, &rb
->event_list
);
3465 spin_unlock_irqrestore(&rb
->event_lock
, flags
);
3468 static void ring_buffer_detach(struct perf_event
*event
,
3469 struct ring_buffer
*rb
)
3471 unsigned long flags
;
3473 if (list_empty(&event
->rb_entry
))
3476 spin_lock_irqsave(&rb
->event_lock
, flags
);
3477 list_del_init(&event
->rb_entry
);
3478 wake_up_all(&event
->waitq
);
3479 spin_unlock_irqrestore(&rb
->event_lock
, flags
);
3482 static void ring_buffer_wakeup(struct perf_event
*event
)
3484 struct ring_buffer
*rb
;
3487 rb
= rcu_dereference(event
->rb
);
3491 list_for_each_entry_rcu(event
, &rb
->event_list
, rb_entry
)
3492 wake_up_all(&event
->waitq
);
3498 static void rb_free_rcu(struct rcu_head
*rcu_head
)
3500 struct ring_buffer
*rb
;
3502 rb
= container_of(rcu_head
, struct ring_buffer
, rcu_head
);
3506 static struct ring_buffer
*ring_buffer_get(struct perf_event
*event
)
3508 struct ring_buffer
*rb
;
3511 rb
= rcu_dereference(event
->rb
);
3513 if (!atomic_inc_not_zero(&rb
->refcount
))
3521 static void ring_buffer_put(struct ring_buffer
*rb
)
3523 struct perf_event
*event
, *n
;
3524 unsigned long flags
;
3526 if (!atomic_dec_and_test(&rb
->refcount
))
3529 spin_lock_irqsave(&rb
->event_lock
, flags
);
3530 list_for_each_entry_safe(event
, n
, &rb
->event_list
, rb_entry
) {
3531 list_del_init(&event
->rb_entry
);
3532 wake_up_all(&event
->waitq
);
3534 spin_unlock_irqrestore(&rb
->event_lock
, flags
);
3536 call_rcu(&rb
->rcu_head
, rb_free_rcu
);
3539 static void perf_mmap_open(struct vm_area_struct
*vma
)
3541 struct perf_event
*event
= vma
->vm_file
->private_data
;
3543 atomic_inc(&event
->mmap_count
);
3546 static void perf_mmap_close(struct vm_area_struct
*vma
)
3548 struct perf_event
*event
= vma
->vm_file
->private_data
;
3550 if (atomic_dec_and_mutex_lock(&event
->mmap_count
, &event
->mmap_mutex
)) {
3551 unsigned long size
= perf_data_size(event
->rb
);
3552 struct user_struct
*user
= event
->mmap_user
;
3553 struct ring_buffer
*rb
= event
->rb
;
3555 atomic_long_sub((size
>> PAGE_SHIFT
) + 1, &user
->locked_vm
);
3556 vma
->vm_mm
->pinned_vm
-= event
->mmap_locked
;
3557 rcu_assign_pointer(event
->rb
, NULL
);
3558 ring_buffer_detach(event
, rb
);
3559 mutex_unlock(&event
->mmap_mutex
);
3561 ring_buffer_put(rb
);
3566 static const struct vm_operations_struct perf_mmap_vmops
= {
3567 .open
= perf_mmap_open
,
3568 .close
= perf_mmap_close
,
3569 .fault
= perf_mmap_fault
,
3570 .page_mkwrite
= perf_mmap_fault
,
3573 static int perf_mmap(struct file
*file
, struct vm_area_struct
*vma
)
3575 struct perf_event
*event
= file
->private_data
;
3576 unsigned long user_locked
, user_lock_limit
;
3577 struct user_struct
*user
= current_user();
3578 unsigned long locked
, lock_limit
;
3579 struct ring_buffer
*rb
;
3580 unsigned long vma_size
;
3581 unsigned long nr_pages
;
3582 long user_extra
, extra
;
3583 int ret
= 0, flags
= 0;
3586 * Don't allow mmap() of inherited per-task counters. This would
3587 * create a performance issue due to all children writing to the
3590 if (event
->cpu
== -1 && event
->attr
.inherit
)
3593 if (!(vma
->vm_flags
& VM_SHARED
))
3596 vma_size
= vma
->vm_end
- vma
->vm_start
;
3597 nr_pages
= (vma_size
/ PAGE_SIZE
) - 1;
3600 * If we have rb pages ensure they're a power-of-two number, so we
3601 * can do bitmasks instead of modulo.
3603 if (nr_pages
!= 0 && !is_power_of_2(nr_pages
))
3606 if (vma_size
!= PAGE_SIZE
* (1 + nr_pages
))
3609 if (vma
->vm_pgoff
!= 0)
3612 WARN_ON_ONCE(event
->ctx
->parent_ctx
);
3613 mutex_lock(&event
->mmap_mutex
);
3615 if (event
->rb
->nr_pages
== nr_pages
)
3616 atomic_inc(&event
->rb
->refcount
);
3622 user_extra
= nr_pages
+ 1;
3623 user_lock_limit
= sysctl_perf_event_mlock
>> (PAGE_SHIFT
- 10);
3626 * Increase the limit linearly with more CPUs:
3628 user_lock_limit
*= num_online_cpus();
3630 user_locked
= atomic_long_read(&user
->locked_vm
) + user_extra
;
3633 if (user_locked
> user_lock_limit
)
3634 extra
= user_locked
- user_lock_limit
;
3636 lock_limit
= rlimit(RLIMIT_MEMLOCK
);
3637 lock_limit
>>= PAGE_SHIFT
;
3638 locked
= vma
->vm_mm
->pinned_vm
+ extra
;
3640 if ((locked
> lock_limit
) && perf_paranoid_tracepoint_raw() &&
3641 !capable(CAP_IPC_LOCK
)) {
3648 if (vma
->vm_flags
& VM_WRITE
)
3649 flags
|= RING_BUFFER_WRITABLE
;
3651 rb
= rb_alloc(nr_pages
,
3652 event
->attr
.watermark
? event
->attr
.wakeup_watermark
: 0,
3659 rcu_assign_pointer(event
->rb
, rb
);
3661 atomic_long_add(user_extra
, &user
->locked_vm
);
3662 event
->mmap_locked
= extra
;
3663 event
->mmap_user
= get_current_user();
3664 vma
->vm_mm
->pinned_vm
+= event
->mmap_locked
;
3666 perf_event_update_userpage(event
);
3670 atomic_inc(&event
->mmap_count
);
3671 mutex_unlock(&event
->mmap_mutex
);
3673 vma
->vm_flags
|= VM_RESERVED
;
3674 vma
->vm_ops
= &perf_mmap_vmops
;
3679 static int perf_fasync(int fd
, struct file
*filp
, int on
)
3681 struct inode
*inode
= filp
->f_path
.dentry
->d_inode
;
3682 struct perf_event
*event
= filp
->private_data
;
3685 mutex_lock(&inode
->i_mutex
);
3686 retval
= fasync_helper(fd
, filp
, on
, &event
->fasync
);
3687 mutex_unlock(&inode
->i_mutex
);
3695 static const struct file_operations perf_fops
= {
3696 .llseek
= no_llseek
,
3697 .release
= perf_release
,
3700 .unlocked_ioctl
= perf_ioctl
,
3701 .compat_ioctl
= perf_ioctl
,
3703 .fasync
= perf_fasync
,
3709 * If there's data, ensure we set the poll() state and publish everything
3710 * to user-space before waking everybody up.
3713 void perf_event_wakeup(struct perf_event
*event
)
3715 ring_buffer_wakeup(event
);
3717 if (event
->pending_kill
) {
3718 kill_fasync(&event
->fasync
, SIGIO
, event
->pending_kill
);
3719 event
->pending_kill
= 0;
3723 static void perf_pending_event(struct irq_work
*entry
)
3725 struct perf_event
*event
= container_of(entry
,
3726 struct perf_event
, pending
);
3728 if (event
->pending_disable
) {
3729 event
->pending_disable
= 0;
3730 __perf_event_disable(event
);
3733 if (event
->pending_wakeup
) {
3734 event
->pending_wakeup
= 0;
3735 perf_event_wakeup(event
);
3740 * We assume there is only KVM supporting the callbacks.
3741 * Later on, we might change it to a list if there is
3742 * another virtualization implementation supporting the callbacks.
3744 struct perf_guest_info_callbacks
*perf_guest_cbs
;
3746 int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks
*cbs
)
3748 perf_guest_cbs
= cbs
;
3751 EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks
);
3753 int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks
*cbs
)
3755 perf_guest_cbs
= NULL
;
3758 EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks
);
3760 static void __perf_event_header__init_id(struct perf_event_header
*header
,
3761 struct perf_sample_data
*data
,
3762 struct perf_event
*event
)
3764 u64 sample_type
= event
->attr
.sample_type
;
3766 data
->type
= sample_type
;
3767 header
->size
+= event
->id_header_size
;
3769 if (sample_type
& PERF_SAMPLE_TID
) {
3770 /* namespace issues */
3771 data
->tid_entry
.pid
= perf_event_pid(event
, current
);
3772 data
->tid_entry
.tid
= perf_event_tid(event
, current
);
3775 if (sample_type
& PERF_SAMPLE_TIME
)
3776 data
->time
= perf_clock();
3778 if (sample_type
& PERF_SAMPLE_ID
)
3779 data
->id
= primary_event_id(event
);
3781 if (sample_type
& PERF_SAMPLE_STREAM_ID
)
3782 data
->stream_id
= event
->id
;
3784 if (sample_type
& PERF_SAMPLE_CPU
) {
3785 data
->cpu_entry
.cpu
= raw_smp_processor_id();
3786 data
->cpu_entry
.reserved
= 0;
3790 void perf_event_header__init_id(struct perf_event_header
*header
,
3791 struct perf_sample_data
*data
,
3792 struct perf_event
*event
)
3794 if (event
->attr
.sample_id_all
)
3795 __perf_event_header__init_id(header
, data
, event
);
3798 static void __perf_event__output_id_sample(struct perf_output_handle
*handle
,
3799 struct perf_sample_data
*data
)
3801 u64 sample_type
= data
->type
;
3803 if (sample_type
& PERF_SAMPLE_TID
)
3804 perf_output_put(handle
, data
->tid_entry
);
3806 if (sample_type
& PERF_SAMPLE_TIME
)
3807 perf_output_put(handle
, data
->time
);
3809 if (sample_type
& PERF_SAMPLE_ID
)
3810 perf_output_put(handle
, data
->id
);
3812 if (sample_type
& PERF_SAMPLE_STREAM_ID
)
3813 perf_output_put(handle
, data
->stream_id
);
3815 if (sample_type
& PERF_SAMPLE_CPU
)
3816 perf_output_put(handle
, data
->cpu_entry
);
3819 void perf_event__output_id_sample(struct perf_event
*event
,
3820 struct perf_output_handle
*handle
,
3821 struct perf_sample_data
*sample
)
3823 if (event
->attr
.sample_id_all
)
3824 __perf_event__output_id_sample(handle
, sample
);
3827 static void perf_output_read_one(struct perf_output_handle
*handle
,
3828 struct perf_event
*event
,
3829 u64 enabled
, u64 running
)
3831 u64 read_format
= event
->attr
.read_format
;
3835 values
[n
++] = perf_event_count(event
);
3836 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
) {
3837 values
[n
++] = enabled
+
3838 atomic64_read(&event
->child_total_time_enabled
);
3840 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
) {
3841 values
[n
++] = running
+
3842 atomic64_read(&event
->child_total_time_running
);
3844 if (read_format
& PERF_FORMAT_ID
)
3845 values
[n
++] = primary_event_id(event
);
3847 __output_copy(handle
, values
, n
* sizeof(u64
));
3851 * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
3853 static void perf_output_read_group(struct perf_output_handle
*handle
,
3854 struct perf_event
*event
,
3855 u64 enabled
, u64 running
)
3857 struct perf_event
*leader
= event
->group_leader
, *sub
;
3858 u64 read_format
= event
->attr
.read_format
;
3862 values
[n
++] = 1 + leader
->nr_siblings
;
3864 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
3865 values
[n
++] = enabled
;
3867 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
3868 values
[n
++] = running
;
3870 if (leader
!= event
)
3871 leader
->pmu
->read(leader
);
3873 values
[n
++] = perf_event_count(leader
);
3874 if (read_format
& PERF_FORMAT_ID
)
3875 values
[n
++] = primary_event_id(leader
);
3877 __output_copy(handle
, values
, n
* sizeof(u64
));
3879 list_for_each_entry(sub
, &leader
->sibling_list
, group_entry
) {
3883 sub
->pmu
->read(sub
);
3885 values
[n
++] = perf_event_count(sub
);
3886 if (read_format
& PERF_FORMAT_ID
)
3887 values
[n
++] = primary_event_id(sub
);
3889 __output_copy(handle
, values
, n
* sizeof(u64
));
3893 #define PERF_FORMAT_TOTAL_TIMES (PERF_FORMAT_TOTAL_TIME_ENABLED|\
3894 PERF_FORMAT_TOTAL_TIME_RUNNING)
3896 static void perf_output_read(struct perf_output_handle
*handle
,
3897 struct perf_event
*event
)
3899 u64 enabled
= 0, running
= 0, now
;
3900 u64 read_format
= event
->attr
.read_format
;
3903 * compute total_time_enabled, total_time_running
3904 * based on snapshot values taken when the event
3905 * was last scheduled in.
3907 * we cannot simply called update_context_time()
3908 * because of locking issue as we are called in
3911 if (read_format
& PERF_FORMAT_TOTAL_TIMES
)
3912 calc_timer_values(event
, &now
, &enabled
, &running
);
3914 if (event
->attr
.read_format
& PERF_FORMAT_GROUP
)
3915 perf_output_read_group(handle
, event
, enabled
, running
);
3917 perf_output_read_one(handle
, event
, enabled
, running
);
3920 void perf_output_sample(struct perf_output_handle
*handle
,
3921 struct perf_event_header
*header
,
3922 struct perf_sample_data
*data
,
3923 struct perf_event
*event
)
3925 u64 sample_type
= data
->type
;
3927 perf_output_put(handle
, *header
);
3929 if (sample_type
& PERF_SAMPLE_IP
)
3930 perf_output_put(handle
, data
->ip
);
3932 if (sample_type
& PERF_SAMPLE_TID
)
3933 perf_output_put(handle
, data
->tid_entry
);
3935 if (sample_type
& PERF_SAMPLE_TIME
)
3936 perf_output_put(handle
, data
->time
);
3938 if (sample_type
& PERF_SAMPLE_ADDR
)
3939 perf_output_put(handle
, data
->addr
);
3941 if (sample_type
& PERF_SAMPLE_ID
)
3942 perf_output_put(handle
, data
->id
);
3944 if (sample_type
& PERF_SAMPLE_STREAM_ID
)
3945 perf_output_put(handle
, data
->stream_id
);
3947 if (sample_type
& PERF_SAMPLE_CPU
)
3948 perf_output_put(handle
, data
->cpu_entry
);
3950 if (sample_type
& PERF_SAMPLE_PERIOD
)
3951 perf_output_put(handle
, data
->period
);
3953 if (sample_type
& PERF_SAMPLE_READ
)
3954 perf_output_read(handle
, event
);
3956 if (sample_type
& PERF_SAMPLE_CALLCHAIN
) {
3957 if (data
->callchain
) {
3960 if (data
->callchain
)
3961 size
+= data
->callchain
->nr
;
3963 size
*= sizeof(u64
);
3965 __output_copy(handle
, data
->callchain
, size
);
3968 perf_output_put(handle
, nr
);
3972 if (sample_type
& PERF_SAMPLE_RAW
) {
3974 perf_output_put(handle
, data
->raw
->size
);
3975 __output_copy(handle
, data
->raw
->data
,
3982 .size
= sizeof(u32
),
3985 perf_output_put(handle
, raw
);
3989 if (!event
->attr
.watermark
) {
3990 int wakeup_events
= event
->attr
.wakeup_events
;
3992 if (wakeup_events
) {
3993 struct ring_buffer
*rb
= handle
->rb
;
3994 int events
= local_inc_return(&rb
->events
);
3996 if (events
>= wakeup_events
) {
3997 local_sub(wakeup_events
, &rb
->events
);
3998 local_inc(&rb
->wakeup
);
4003 if (sample_type
& PERF_SAMPLE_BRANCH_STACK
) {
4004 if (data
->br_stack
) {
4007 size
= data
->br_stack
->nr
4008 * sizeof(struct perf_branch_entry
);
4010 perf_output_put(handle
, data
->br_stack
->nr
);
4011 perf_output_copy(handle
, data
->br_stack
->entries
, size
);
4014 * we always store at least the value of nr
4017 perf_output_put(handle
, nr
);
4022 void perf_prepare_sample(struct perf_event_header
*header
,
4023 struct perf_sample_data
*data
,
4024 struct perf_event
*event
,
4025 struct pt_regs
*regs
)
4027 u64 sample_type
= event
->attr
.sample_type
;
4029 header
->type
= PERF_RECORD_SAMPLE
;
4030 header
->size
= sizeof(*header
) + event
->header_size
;
4033 header
->misc
|= perf_misc_flags(regs
);
4035 __perf_event_header__init_id(header
, data
, event
);
4037 if (sample_type
& PERF_SAMPLE_IP
)
4038 data
->ip
= perf_instruction_pointer(regs
);
4040 if (sample_type
& PERF_SAMPLE_CALLCHAIN
) {
4043 data
->callchain
= perf_callchain(event
, regs
);
4045 if (data
->callchain
)
4046 size
+= data
->callchain
->nr
;
4048 header
->size
+= size
* sizeof(u64
);
4051 if (sample_type
& PERF_SAMPLE_RAW
) {
4052 int size
= sizeof(u32
);
4055 size
+= data
->raw
->size
;
4057 size
+= sizeof(u32
);
4059 WARN_ON_ONCE(size
& (sizeof(u64
)-1));
4060 header
->size
+= size
;
4063 if (sample_type
& PERF_SAMPLE_BRANCH_STACK
) {
4064 int size
= sizeof(u64
); /* nr */
4065 if (data
->br_stack
) {
4066 size
+= data
->br_stack
->nr
4067 * sizeof(struct perf_branch_entry
);
4069 header
->size
+= size
;
4073 static void perf_event_output(struct perf_event
*event
,
4074 struct perf_sample_data
*data
,
4075 struct pt_regs
*regs
)
4077 struct perf_output_handle handle
;
4078 struct perf_event_header header
;
4080 /* protect the callchain buffers */
4083 perf_prepare_sample(&header
, data
, event
, regs
);
4085 if (perf_output_begin(&handle
, event
, header
.size
))
4088 perf_output_sample(&handle
, &header
, data
, event
);
4090 perf_output_end(&handle
);
4100 struct perf_read_event
{
4101 struct perf_event_header header
;
4108 perf_event_read_event(struct perf_event
*event
,
4109 struct task_struct
*task
)
4111 struct perf_output_handle handle
;
4112 struct perf_sample_data sample
;
4113 struct perf_read_event read_event
= {
4115 .type
= PERF_RECORD_READ
,
4117 .size
= sizeof(read_event
) + event
->read_size
,
4119 .pid
= perf_event_pid(event
, task
),
4120 .tid
= perf_event_tid(event
, task
),
4124 perf_event_header__init_id(&read_event
.header
, &sample
, event
);
4125 ret
= perf_output_begin(&handle
, event
, read_event
.header
.size
);
4129 perf_output_put(&handle
, read_event
);
4130 perf_output_read(&handle
, event
);
4131 perf_event__output_id_sample(event
, &handle
, &sample
);
4133 perf_output_end(&handle
);
4137 * task tracking -- fork/exit
4139 * enabled by: attr.comm | attr.mmap | attr.mmap_data | attr.task
4142 struct perf_task_event
{
4143 struct task_struct
*task
;
4144 struct perf_event_context
*task_ctx
;
4147 struct perf_event_header header
;
4157 static void perf_event_task_output(struct perf_event
*event
,
4158 struct perf_task_event
*task_event
)
4160 struct perf_output_handle handle
;
4161 struct perf_sample_data sample
;
4162 struct task_struct
*task
= task_event
->task
;
4163 int ret
, size
= task_event
->event_id
.header
.size
;
4165 perf_event_header__init_id(&task_event
->event_id
.header
, &sample
, event
);
4167 ret
= perf_output_begin(&handle
, event
,
4168 task_event
->event_id
.header
.size
);
4172 task_event
->event_id
.pid
= perf_event_pid(event
, task
);
4173 task_event
->event_id
.ppid
= perf_event_pid(event
, current
);
4175 task_event
->event_id
.tid
= perf_event_tid(event
, task
);
4176 task_event
->event_id
.ptid
= perf_event_tid(event
, current
);
4178 perf_output_put(&handle
, task_event
->event_id
);
4180 perf_event__output_id_sample(event
, &handle
, &sample
);
4182 perf_output_end(&handle
);
4184 task_event
->event_id
.header
.size
= size
;
4187 static int perf_event_task_match(struct perf_event
*event
)
4189 if (event
->state
< PERF_EVENT_STATE_INACTIVE
)
4192 if (!event_filter_match(event
))
4195 if (event
->attr
.comm
|| event
->attr
.mmap
||
4196 event
->attr
.mmap_data
|| event
->attr
.task
)
4202 static void perf_event_task_ctx(struct perf_event_context
*ctx
,
4203 struct perf_task_event
*task_event
)
4205 struct perf_event
*event
;
4207 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
4208 if (perf_event_task_match(event
))
4209 perf_event_task_output(event
, task_event
);
4213 static void perf_event_task_event(struct perf_task_event
*task_event
)
4215 struct perf_cpu_context
*cpuctx
;
4216 struct perf_event_context
*ctx
;
4221 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
4222 cpuctx
= get_cpu_ptr(pmu
->pmu_cpu_context
);
4223 if (cpuctx
->active_pmu
!= pmu
)
4225 perf_event_task_ctx(&cpuctx
->ctx
, task_event
);
4227 ctx
= task_event
->task_ctx
;
4229 ctxn
= pmu
->task_ctx_nr
;
4232 ctx
= rcu_dereference(current
->perf_event_ctxp
[ctxn
]);
4235 perf_event_task_ctx(ctx
, task_event
);
4237 put_cpu_ptr(pmu
->pmu_cpu_context
);
4242 static void perf_event_task(struct task_struct
*task
,
4243 struct perf_event_context
*task_ctx
,
4246 struct perf_task_event task_event
;
4248 if (!atomic_read(&nr_comm_events
) &&
4249 !atomic_read(&nr_mmap_events
) &&
4250 !atomic_read(&nr_task_events
))
4253 task_event
= (struct perf_task_event
){
4255 .task_ctx
= task_ctx
,
4258 .type
= new ? PERF_RECORD_FORK
: PERF_RECORD_EXIT
,
4260 .size
= sizeof(task_event
.event_id
),
4266 .time
= perf_clock(),
4270 perf_event_task_event(&task_event
);
4273 void perf_event_fork(struct task_struct
*task
)
4275 perf_event_task(task
, NULL
, 1);
4282 struct perf_comm_event
{
4283 struct task_struct
*task
;
4288 struct perf_event_header header
;
4295 static void perf_event_comm_output(struct perf_event
*event
,
4296 struct perf_comm_event
*comm_event
)
4298 struct perf_output_handle handle
;
4299 struct perf_sample_data sample
;
4300 int size
= comm_event
->event_id
.header
.size
;
4303 perf_event_header__init_id(&comm_event
->event_id
.header
, &sample
, event
);
4304 ret
= perf_output_begin(&handle
, event
,
4305 comm_event
->event_id
.header
.size
);
4310 comm_event
->event_id
.pid
= perf_event_pid(event
, comm_event
->task
);
4311 comm_event
->event_id
.tid
= perf_event_tid(event
, comm_event
->task
);
4313 perf_output_put(&handle
, comm_event
->event_id
);
4314 __output_copy(&handle
, comm_event
->comm
,
4315 comm_event
->comm_size
);
4317 perf_event__output_id_sample(event
, &handle
, &sample
);
4319 perf_output_end(&handle
);
4321 comm_event
->event_id
.header
.size
= size
;
4324 static int perf_event_comm_match(struct perf_event
*event
)
4326 if (event
->state
< PERF_EVENT_STATE_INACTIVE
)
4329 if (!event_filter_match(event
))
4332 if (event
->attr
.comm
)
4338 static void perf_event_comm_ctx(struct perf_event_context
*ctx
,
4339 struct perf_comm_event
*comm_event
)
4341 struct perf_event
*event
;
4343 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
4344 if (perf_event_comm_match(event
))
4345 perf_event_comm_output(event
, comm_event
);
4349 static void perf_event_comm_event(struct perf_comm_event
*comm_event
)
4351 struct perf_cpu_context
*cpuctx
;
4352 struct perf_event_context
*ctx
;
4353 char comm
[TASK_COMM_LEN
];
4358 memset(comm
, 0, sizeof(comm
));
4359 strlcpy(comm
, comm_event
->task
->comm
, sizeof(comm
));
4360 size
= ALIGN(strlen(comm
)+1, sizeof(u64
));
4362 comm_event
->comm
= comm
;
4363 comm_event
->comm_size
= size
;
4365 comm_event
->event_id
.header
.size
= sizeof(comm_event
->event_id
) + size
;
4367 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
4368 cpuctx
= get_cpu_ptr(pmu
->pmu_cpu_context
);
4369 if (cpuctx
->active_pmu
!= pmu
)
4371 perf_event_comm_ctx(&cpuctx
->ctx
, comm_event
);
4373 ctxn
= pmu
->task_ctx_nr
;
4377 ctx
= rcu_dereference(current
->perf_event_ctxp
[ctxn
]);
4379 perf_event_comm_ctx(ctx
, comm_event
);
4381 put_cpu_ptr(pmu
->pmu_cpu_context
);
4386 void perf_event_comm(struct task_struct
*task
)
4388 struct perf_comm_event comm_event
;
4389 struct perf_event_context
*ctx
;
4392 for_each_task_context_nr(ctxn
) {
4393 ctx
= task
->perf_event_ctxp
[ctxn
];
4397 perf_event_enable_on_exec(ctx
);
4400 if (!atomic_read(&nr_comm_events
))
4403 comm_event
= (struct perf_comm_event
){
4409 .type
= PERF_RECORD_COMM
,
4418 perf_event_comm_event(&comm_event
);
4425 struct perf_mmap_event
{
4426 struct vm_area_struct
*vma
;
4428 const char *file_name
;
4432 struct perf_event_header header
;
4442 static void perf_event_mmap_output(struct perf_event
*event
,
4443 struct perf_mmap_event
*mmap_event
)
4445 struct perf_output_handle handle
;
4446 struct perf_sample_data sample
;
4447 int size
= mmap_event
->event_id
.header
.size
;
4450 perf_event_header__init_id(&mmap_event
->event_id
.header
, &sample
, event
);
4451 ret
= perf_output_begin(&handle
, event
,
4452 mmap_event
->event_id
.header
.size
);
4456 mmap_event
->event_id
.pid
= perf_event_pid(event
, current
);
4457 mmap_event
->event_id
.tid
= perf_event_tid(event
, current
);
4459 perf_output_put(&handle
, mmap_event
->event_id
);
4460 __output_copy(&handle
, mmap_event
->file_name
,
4461 mmap_event
->file_size
);
4463 perf_event__output_id_sample(event
, &handle
, &sample
);
4465 perf_output_end(&handle
);
4467 mmap_event
->event_id
.header
.size
= size
;
4470 static int perf_event_mmap_match(struct perf_event
*event
,
4471 struct perf_mmap_event
*mmap_event
,
4474 if (event
->state
< PERF_EVENT_STATE_INACTIVE
)
4477 if (!event_filter_match(event
))
4480 if ((!executable
&& event
->attr
.mmap_data
) ||
4481 (executable
&& event
->attr
.mmap
))
4487 static void perf_event_mmap_ctx(struct perf_event_context
*ctx
,
4488 struct perf_mmap_event
*mmap_event
,
4491 struct perf_event
*event
;
4493 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
4494 if (perf_event_mmap_match(event
, mmap_event
, executable
))
4495 perf_event_mmap_output(event
, mmap_event
);
4499 static void perf_event_mmap_event(struct perf_mmap_event
*mmap_event
)
4501 struct perf_cpu_context
*cpuctx
;
4502 struct perf_event_context
*ctx
;
4503 struct vm_area_struct
*vma
= mmap_event
->vma
;
4504 struct file
*file
= vma
->vm_file
;
4512 memset(tmp
, 0, sizeof(tmp
));
4516 * d_path works from the end of the rb backwards, so we
4517 * need to add enough zero bytes after the string to handle
4518 * the 64bit alignment we do later.
4520 buf
= kzalloc(PATH_MAX
+ sizeof(u64
), GFP_KERNEL
);
4522 name
= strncpy(tmp
, "//enomem", sizeof(tmp
));
4525 name
= d_path(&file
->f_path
, buf
, PATH_MAX
);
4527 name
= strncpy(tmp
, "//toolong", sizeof(tmp
));
4531 if (arch_vma_name(mmap_event
->vma
)) {
4532 name
= strncpy(tmp
, arch_vma_name(mmap_event
->vma
),
4538 name
= strncpy(tmp
, "[vdso]", sizeof(tmp
));
4540 } else if (vma
->vm_start
<= vma
->vm_mm
->start_brk
&&
4541 vma
->vm_end
>= vma
->vm_mm
->brk
) {
4542 name
= strncpy(tmp
, "[heap]", sizeof(tmp
));
4544 } else if (vma
->vm_start
<= vma
->vm_mm
->start_stack
&&
4545 vma
->vm_end
>= vma
->vm_mm
->start_stack
) {
4546 name
= strncpy(tmp
, "[stack]", sizeof(tmp
));
4550 name
= strncpy(tmp
, "//anon", sizeof(tmp
));
4555 size
= ALIGN(strlen(name
)+1, sizeof(u64
));
4557 mmap_event
->file_name
= name
;
4558 mmap_event
->file_size
= size
;
4560 mmap_event
->event_id
.header
.size
= sizeof(mmap_event
->event_id
) + size
;
4563 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
4564 cpuctx
= get_cpu_ptr(pmu
->pmu_cpu_context
);
4565 if (cpuctx
->active_pmu
!= pmu
)
4567 perf_event_mmap_ctx(&cpuctx
->ctx
, mmap_event
,
4568 vma
->vm_flags
& VM_EXEC
);
4570 ctxn
= pmu
->task_ctx_nr
;
4574 ctx
= rcu_dereference(current
->perf_event_ctxp
[ctxn
]);
4576 perf_event_mmap_ctx(ctx
, mmap_event
,
4577 vma
->vm_flags
& VM_EXEC
);
4580 put_cpu_ptr(pmu
->pmu_cpu_context
);
4587 void perf_event_mmap(struct vm_area_struct
*vma
)
4589 struct perf_mmap_event mmap_event
;
4591 if (!atomic_read(&nr_mmap_events
))
4594 mmap_event
= (struct perf_mmap_event
){
4600 .type
= PERF_RECORD_MMAP
,
4601 .misc
= PERF_RECORD_MISC_USER
,
4606 .start
= vma
->vm_start
,
4607 .len
= vma
->vm_end
- vma
->vm_start
,
4608 .pgoff
= (u64
)vma
->vm_pgoff
<< PAGE_SHIFT
,
4612 perf_event_mmap_event(&mmap_event
);
4616 * IRQ throttle logging
4619 static void perf_log_throttle(struct perf_event
*event
, int enable
)
4621 struct perf_output_handle handle
;
4622 struct perf_sample_data sample
;
4626 struct perf_event_header header
;
4630 } throttle_event
= {
4632 .type
= PERF_RECORD_THROTTLE
,
4634 .size
= sizeof(throttle_event
),
4636 .time
= perf_clock(),
4637 .id
= primary_event_id(event
),
4638 .stream_id
= event
->id
,
4642 throttle_event
.header
.type
= PERF_RECORD_UNTHROTTLE
;
4644 perf_event_header__init_id(&throttle_event
.header
, &sample
, event
);
4646 ret
= perf_output_begin(&handle
, event
,
4647 throttle_event
.header
.size
);
4651 perf_output_put(&handle
, throttle_event
);
4652 perf_event__output_id_sample(event
, &handle
, &sample
);
4653 perf_output_end(&handle
);
4657 * Generic event overflow handling, sampling.
4660 static int __perf_event_overflow(struct perf_event
*event
,
4661 int throttle
, struct perf_sample_data
*data
,
4662 struct pt_regs
*regs
)
4664 int events
= atomic_read(&event
->event_limit
);
4665 struct hw_perf_event
*hwc
= &event
->hw
;
4670 * Non-sampling counters might still use the PMI to fold short
4671 * hardware counters, ignore those.
4673 if (unlikely(!is_sampling_event(event
)))
4676 seq
= __this_cpu_read(perf_throttled_seq
);
4677 if (seq
!= hwc
->interrupts_seq
) {
4678 hwc
->interrupts_seq
= seq
;
4679 hwc
->interrupts
= 1;
4682 if (unlikely(throttle
4683 && hwc
->interrupts
>= max_samples_per_tick
)) {
4684 __this_cpu_inc(perf_throttled_count
);
4685 hwc
->interrupts
= MAX_INTERRUPTS
;
4686 perf_log_throttle(event
, 0);
4691 if (event
->attr
.freq
) {
4692 u64 now
= perf_clock();
4693 s64 delta
= now
- hwc
->freq_time_stamp
;
4695 hwc
->freq_time_stamp
= now
;
4697 if (delta
> 0 && delta
< 2*TICK_NSEC
)
4698 perf_adjust_period(event
, delta
, hwc
->last_period
, true);
4702 * XXX event_limit might not quite work as expected on inherited
4706 event
->pending_kill
= POLL_IN
;
4707 if (events
&& atomic_dec_and_test(&event
->event_limit
)) {
4709 event
->pending_kill
= POLL_HUP
;
4710 event
->pending_disable
= 1;
4711 irq_work_queue(&event
->pending
);
4714 if (event
->overflow_handler
)
4715 event
->overflow_handler(event
, data
, regs
);
4717 perf_event_output(event
, data
, regs
);
4719 if (event
->fasync
&& event
->pending_kill
) {
4720 event
->pending_wakeup
= 1;
4721 irq_work_queue(&event
->pending
);
4727 int perf_event_overflow(struct perf_event
*event
,
4728 struct perf_sample_data
*data
,
4729 struct pt_regs
*regs
)
4731 return __perf_event_overflow(event
, 1, data
, regs
);
4735 * Generic software event infrastructure
4738 struct swevent_htable
{
4739 struct swevent_hlist
*swevent_hlist
;
4740 struct mutex hlist_mutex
;
4743 /* Recursion avoidance in each contexts */
4744 int recursion
[PERF_NR_CONTEXTS
];
4747 static DEFINE_PER_CPU(struct swevent_htable
, swevent_htable
);
4750 * We directly increment event->count and keep a second value in
4751 * event->hw.period_left to count intervals. This period event
4752 * is kept in the range [-sample_period, 0] so that we can use the
4756 static u64
perf_swevent_set_period(struct perf_event
*event
)
4758 struct hw_perf_event
*hwc
= &event
->hw
;
4759 u64 period
= hwc
->last_period
;
4763 hwc
->last_period
= hwc
->sample_period
;
4766 old
= val
= local64_read(&hwc
->period_left
);
4770 nr
= div64_u64(period
+ val
, period
);
4771 offset
= nr
* period
;
4773 if (local64_cmpxchg(&hwc
->period_left
, old
, val
) != old
)
4779 static void perf_swevent_overflow(struct perf_event
*event
, u64 overflow
,
4780 struct perf_sample_data
*data
,
4781 struct pt_regs
*regs
)
4783 struct hw_perf_event
*hwc
= &event
->hw
;
4787 overflow
= perf_swevent_set_period(event
);
4789 if (hwc
->interrupts
== MAX_INTERRUPTS
)
4792 for (; overflow
; overflow
--) {
4793 if (__perf_event_overflow(event
, throttle
,
4796 * We inhibit the overflow from happening when
4797 * hwc->interrupts == MAX_INTERRUPTS.
4805 static void perf_swevent_event(struct perf_event
*event
, u64 nr
,
4806 struct perf_sample_data
*data
,
4807 struct pt_regs
*regs
)
4809 struct hw_perf_event
*hwc
= &event
->hw
;
4811 local64_add(nr
, &event
->count
);
4816 if (!is_sampling_event(event
))
4819 if ((event
->attr
.sample_type
& PERF_SAMPLE_PERIOD
) && !event
->attr
.freq
) {
4821 return perf_swevent_overflow(event
, 1, data
, regs
);
4823 data
->period
= event
->hw
.last_period
;
4825 if (nr
== 1 && hwc
->sample_period
== 1 && !event
->attr
.freq
)
4826 return perf_swevent_overflow(event
, 1, data
, regs
);
4828 if (local64_add_negative(nr
, &hwc
->period_left
))
4831 perf_swevent_overflow(event
, 0, data
, regs
);
4834 static int perf_exclude_event(struct perf_event
*event
,
4835 struct pt_regs
*regs
)
4837 if (event
->hw
.state
& PERF_HES_STOPPED
)
4841 if (event
->attr
.exclude_user
&& user_mode(regs
))
4844 if (event
->attr
.exclude_kernel
&& !user_mode(regs
))
4851 static int perf_swevent_match(struct perf_event
*event
,
4852 enum perf_type_id type
,
4854 struct perf_sample_data
*data
,
4855 struct pt_regs
*regs
)
4857 if (event
->attr
.type
!= type
)
4860 if (event
->attr
.config
!= event_id
)
4863 if (perf_exclude_event(event
, regs
))
4869 static inline u64
swevent_hash(u64 type
, u32 event_id
)
4871 u64 val
= event_id
| (type
<< 32);
4873 return hash_64(val
, SWEVENT_HLIST_BITS
);
4876 static inline struct hlist_head
*
4877 __find_swevent_head(struct swevent_hlist
*hlist
, u64 type
, u32 event_id
)
4879 u64 hash
= swevent_hash(type
, event_id
);
4881 return &hlist
->heads
[hash
];
4884 /* For the read side: events when they trigger */
4885 static inline struct hlist_head
*
4886 find_swevent_head_rcu(struct swevent_htable
*swhash
, u64 type
, u32 event_id
)
4888 struct swevent_hlist
*hlist
;
4890 hlist
= rcu_dereference(swhash
->swevent_hlist
);
4894 return __find_swevent_head(hlist
, type
, event_id
);
4897 /* For the event head insertion and removal in the hlist */
4898 static inline struct hlist_head
*
4899 find_swevent_head(struct swevent_htable
*swhash
, struct perf_event
*event
)
4901 struct swevent_hlist
*hlist
;
4902 u32 event_id
= event
->attr
.config
;
4903 u64 type
= event
->attr
.type
;
4906 * Event scheduling is always serialized against hlist allocation
4907 * and release. Which makes the protected version suitable here.
4908 * The context lock guarantees that.
4910 hlist
= rcu_dereference_protected(swhash
->swevent_hlist
,
4911 lockdep_is_held(&event
->ctx
->lock
));
4915 return __find_swevent_head(hlist
, type
, event_id
);
4918 static void do_perf_sw_event(enum perf_type_id type
, u32 event_id
,
4920 struct perf_sample_data
*data
,
4921 struct pt_regs
*regs
)
4923 struct swevent_htable
*swhash
= &__get_cpu_var(swevent_htable
);
4924 struct perf_event
*event
;
4925 struct hlist_node
*node
;
4926 struct hlist_head
*head
;
4929 head
= find_swevent_head_rcu(swhash
, type
, event_id
);
4933 hlist_for_each_entry_rcu(event
, node
, head
, hlist_entry
) {
4934 if (perf_swevent_match(event
, type
, event_id
, data
, regs
))
4935 perf_swevent_event(event
, nr
, data
, regs
);
4941 int perf_swevent_get_recursion_context(void)
4943 struct swevent_htable
*swhash
= &__get_cpu_var(swevent_htable
);
4945 return get_recursion_context(swhash
->recursion
);
4947 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context
);
4949 inline void perf_swevent_put_recursion_context(int rctx
)
4951 struct swevent_htable
*swhash
= &__get_cpu_var(swevent_htable
);
4953 put_recursion_context(swhash
->recursion
, rctx
);
4956 void __perf_sw_event(u32 event_id
, u64 nr
, struct pt_regs
*regs
, u64 addr
)
4958 struct perf_sample_data data
;
4961 preempt_disable_notrace();
4962 rctx
= perf_swevent_get_recursion_context();
4966 perf_sample_data_init(&data
, addr
, 0);
4968 do_perf_sw_event(PERF_TYPE_SOFTWARE
, event_id
, nr
, &data
, regs
);
4970 perf_swevent_put_recursion_context(rctx
);
4971 preempt_enable_notrace();
4974 static void perf_swevent_read(struct perf_event
*event
)
4978 static int perf_swevent_add(struct perf_event
*event
, int flags
)
4980 struct swevent_htable
*swhash
= &__get_cpu_var(swevent_htable
);
4981 struct hw_perf_event
*hwc
= &event
->hw
;
4982 struct hlist_head
*head
;
4984 if (is_sampling_event(event
)) {
4985 hwc
->last_period
= hwc
->sample_period
;
4986 perf_swevent_set_period(event
);
4989 hwc
->state
= !(flags
& PERF_EF_START
);
4991 head
= find_swevent_head(swhash
, event
);
4992 if (WARN_ON_ONCE(!head
))
4995 hlist_add_head_rcu(&event
->hlist_entry
, head
);
5000 static void perf_swevent_del(struct perf_event
*event
, int flags
)
5002 hlist_del_rcu(&event
->hlist_entry
);
5005 static void perf_swevent_start(struct perf_event
*event
, int flags
)
5007 event
->hw
.state
= 0;
5010 static void perf_swevent_stop(struct perf_event
*event
, int flags
)
5012 event
->hw
.state
= PERF_HES_STOPPED
;
5015 /* Deref the hlist from the update side */
5016 static inline struct swevent_hlist
*
5017 swevent_hlist_deref(struct swevent_htable
*swhash
)
5019 return rcu_dereference_protected(swhash
->swevent_hlist
,
5020 lockdep_is_held(&swhash
->hlist_mutex
));
5023 static void swevent_hlist_release(struct swevent_htable
*swhash
)
5025 struct swevent_hlist
*hlist
= swevent_hlist_deref(swhash
);
5030 rcu_assign_pointer(swhash
->swevent_hlist
, NULL
);
5031 kfree_rcu(hlist
, rcu_head
);
5034 static void swevent_hlist_put_cpu(struct perf_event
*event
, int cpu
)
5036 struct swevent_htable
*swhash
= &per_cpu(swevent_htable
, cpu
);
5038 mutex_lock(&swhash
->hlist_mutex
);
5040 if (!--swhash
->hlist_refcount
)
5041 swevent_hlist_release(swhash
);
5043 mutex_unlock(&swhash
->hlist_mutex
);
5046 static void swevent_hlist_put(struct perf_event
*event
)
5050 if (event
->cpu
!= -1) {
5051 swevent_hlist_put_cpu(event
, event
->cpu
);
5055 for_each_possible_cpu(cpu
)
5056 swevent_hlist_put_cpu(event
, cpu
);
5059 static int swevent_hlist_get_cpu(struct perf_event
*event
, int cpu
)
5061 struct swevent_htable
*swhash
= &per_cpu(swevent_htable
, cpu
);
5064 mutex_lock(&swhash
->hlist_mutex
);
5066 if (!swevent_hlist_deref(swhash
) && cpu_online(cpu
)) {
5067 struct swevent_hlist
*hlist
;
5069 hlist
= kzalloc(sizeof(*hlist
), GFP_KERNEL
);
5074 rcu_assign_pointer(swhash
->swevent_hlist
, hlist
);
5076 swhash
->hlist_refcount
++;
5078 mutex_unlock(&swhash
->hlist_mutex
);
5083 static int swevent_hlist_get(struct perf_event
*event
)
5086 int cpu
, failed_cpu
;
5088 if (event
->cpu
!= -1)
5089 return swevent_hlist_get_cpu(event
, event
->cpu
);
5092 for_each_possible_cpu(cpu
) {
5093 err
= swevent_hlist_get_cpu(event
, cpu
);
5103 for_each_possible_cpu(cpu
) {
5104 if (cpu
== failed_cpu
)
5106 swevent_hlist_put_cpu(event
, cpu
);
5113 struct static_key perf_swevent_enabled
[PERF_COUNT_SW_MAX
];
5115 static void sw_perf_event_destroy(struct perf_event
*event
)
5117 u64 event_id
= event
->attr
.config
;
5119 WARN_ON(event
->parent
);
5121 static_key_slow_dec(&perf_swevent_enabled
[event_id
]);
5122 swevent_hlist_put(event
);
5125 static int perf_swevent_init(struct perf_event
*event
)
5127 int event_id
= event
->attr
.config
;
5129 if (event
->attr
.type
!= PERF_TYPE_SOFTWARE
)
5133 * no branch sampling for software events
5135 if (has_branch_stack(event
))
5139 case PERF_COUNT_SW_CPU_CLOCK
:
5140 case PERF_COUNT_SW_TASK_CLOCK
:
5147 if (event_id
>= PERF_COUNT_SW_MAX
)
5150 if (!event
->parent
) {
5153 err
= swevent_hlist_get(event
);
5157 static_key_slow_inc(&perf_swevent_enabled
[event_id
]);
5158 event
->destroy
= sw_perf_event_destroy
;
5164 static int perf_swevent_event_idx(struct perf_event
*event
)
5169 static struct pmu perf_swevent
= {
5170 .task_ctx_nr
= perf_sw_context
,
5172 .event_init
= perf_swevent_init
,
5173 .add
= perf_swevent_add
,
5174 .del
= perf_swevent_del
,
5175 .start
= perf_swevent_start
,
5176 .stop
= perf_swevent_stop
,
5177 .read
= perf_swevent_read
,
5179 .event_idx
= perf_swevent_event_idx
,
5182 #ifdef CONFIG_EVENT_TRACING
5184 static int perf_tp_filter_match(struct perf_event
*event
,
5185 struct perf_sample_data
*data
)
5187 void *record
= data
->raw
->data
;
5189 if (likely(!event
->filter
) || filter_match_preds(event
->filter
, record
))
5194 static int perf_tp_event_match(struct perf_event
*event
,
5195 struct perf_sample_data
*data
,
5196 struct pt_regs
*regs
)
5198 if (event
->hw
.state
& PERF_HES_STOPPED
)
5201 * All tracepoints are from kernel-space.
5203 if (event
->attr
.exclude_kernel
)
5206 if (!perf_tp_filter_match(event
, data
))
5212 void perf_tp_event(u64 addr
, u64 count
, void *record
, int entry_size
,
5213 struct pt_regs
*regs
, struct hlist_head
*head
, int rctx
,
5214 struct task_struct
*task
)
5216 struct perf_sample_data data
;
5217 struct perf_event
*event
;
5218 struct hlist_node
*node
;
5220 struct perf_raw_record raw
= {
5225 perf_sample_data_init(&data
, addr
, 0);
5228 hlist_for_each_entry_rcu(event
, node
, head
, hlist_entry
) {
5229 if (perf_tp_event_match(event
, &data
, regs
))
5230 perf_swevent_event(event
, count
, &data
, regs
);
5234 * If we got specified a target task, also iterate its context and
5235 * deliver this event there too.
5237 if (task
&& task
!= current
) {
5238 struct perf_event_context
*ctx
;
5239 struct trace_entry
*entry
= record
;
5242 ctx
= rcu_dereference(task
->perf_event_ctxp
[perf_sw_context
]);
5246 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
5247 if (event
->attr
.type
!= PERF_TYPE_TRACEPOINT
)
5249 if (event
->attr
.config
!= entry
->type
)
5251 if (perf_tp_event_match(event
, &data
, regs
))
5252 perf_swevent_event(event
, count
, &data
, regs
);
5258 perf_swevent_put_recursion_context(rctx
);
5260 EXPORT_SYMBOL_GPL(perf_tp_event
);
5262 static void tp_perf_event_destroy(struct perf_event
*event
)
5264 perf_trace_destroy(event
);
5267 static int perf_tp_event_init(struct perf_event
*event
)
5271 if (event
->attr
.type
!= PERF_TYPE_TRACEPOINT
)
5275 * no branch sampling for tracepoint events
5277 if (has_branch_stack(event
))
5280 err
= perf_trace_init(event
);
5284 event
->destroy
= tp_perf_event_destroy
;
5289 static struct pmu perf_tracepoint
= {
5290 .task_ctx_nr
= perf_sw_context
,
5292 .event_init
= perf_tp_event_init
,
5293 .add
= perf_trace_add
,
5294 .del
= perf_trace_del
,
5295 .start
= perf_swevent_start
,
5296 .stop
= perf_swevent_stop
,
5297 .read
= perf_swevent_read
,
5299 .event_idx
= perf_swevent_event_idx
,
5302 static inline void perf_tp_register(void)
5304 perf_pmu_register(&perf_tracepoint
, "tracepoint", PERF_TYPE_TRACEPOINT
);
5307 static int perf_event_set_filter(struct perf_event
*event
, void __user
*arg
)
5312 if (event
->attr
.type
!= PERF_TYPE_TRACEPOINT
)
5315 filter_str
= strndup_user(arg
, PAGE_SIZE
);
5316 if (IS_ERR(filter_str
))
5317 return PTR_ERR(filter_str
);
5319 ret
= ftrace_profile_set_filter(event
, event
->attr
.config
, filter_str
);
5325 static void perf_event_free_filter(struct perf_event
*event
)
5327 ftrace_profile_free_filter(event
);
5332 static inline void perf_tp_register(void)
5336 static int perf_event_set_filter(struct perf_event
*event
, void __user
*arg
)
5341 static void perf_event_free_filter(struct perf_event
*event
)
5345 #endif /* CONFIG_EVENT_TRACING */
5347 #ifdef CONFIG_HAVE_HW_BREAKPOINT
5348 void perf_bp_event(struct perf_event
*bp
, void *data
)
5350 struct perf_sample_data sample
;
5351 struct pt_regs
*regs
= data
;
5353 perf_sample_data_init(&sample
, bp
->attr
.bp_addr
, 0);
5355 if (!bp
->hw
.state
&& !perf_exclude_event(bp
, regs
))
5356 perf_swevent_event(bp
, 1, &sample
, regs
);
5361 * hrtimer based swevent callback
5364 static enum hrtimer_restart
perf_swevent_hrtimer(struct hrtimer
*hrtimer
)
5366 enum hrtimer_restart ret
= HRTIMER_RESTART
;
5367 struct perf_sample_data data
;
5368 struct pt_regs
*regs
;
5369 struct perf_event
*event
;
5372 event
= container_of(hrtimer
, struct perf_event
, hw
.hrtimer
);
5374 if (event
->state
!= PERF_EVENT_STATE_ACTIVE
)
5375 return HRTIMER_NORESTART
;
5377 event
->pmu
->read(event
);
5379 perf_sample_data_init(&data
, 0, event
->hw
.last_period
);
5380 regs
= get_irq_regs();
5382 if (regs
&& !perf_exclude_event(event
, regs
)) {
5383 if (!(event
->attr
.exclude_idle
&& is_idle_task(current
)))
5384 if (__perf_event_overflow(event
, 1, &data
, regs
))
5385 ret
= HRTIMER_NORESTART
;
5388 period
= max_t(u64
, 10000, event
->hw
.sample_period
);
5389 hrtimer_forward_now(hrtimer
, ns_to_ktime(period
));
5394 static void perf_swevent_start_hrtimer(struct perf_event
*event
)
5396 struct hw_perf_event
*hwc
= &event
->hw
;
5399 if (!is_sampling_event(event
))
5402 period
= local64_read(&hwc
->period_left
);
5407 local64_set(&hwc
->period_left
, 0);
5409 period
= max_t(u64
, 10000, hwc
->sample_period
);
5411 __hrtimer_start_range_ns(&hwc
->hrtimer
,
5412 ns_to_ktime(period
), 0,
5413 HRTIMER_MODE_REL_PINNED
, 0);
5416 static void perf_swevent_cancel_hrtimer(struct perf_event
*event
)
5418 struct hw_perf_event
*hwc
= &event
->hw
;
5420 if (is_sampling_event(event
)) {
5421 ktime_t remaining
= hrtimer_get_remaining(&hwc
->hrtimer
);
5422 local64_set(&hwc
->period_left
, ktime_to_ns(remaining
));
5424 hrtimer_cancel(&hwc
->hrtimer
);
5428 static void perf_swevent_init_hrtimer(struct perf_event
*event
)
5430 struct hw_perf_event
*hwc
= &event
->hw
;
5432 if (!is_sampling_event(event
))
5435 hrtimer_init(&hwc
->hrtimer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
5436 hwc
->hrtimer
.function
= perf_swevent_hrtimer
;
5439 * Since hrtimers have a fixed rate, we can do a static freq->period
5440 * mapping and avoid the whole period adjust feedback stuff.
5442 if (event
->attr
.freq
) {
5443 long freq
= event
->attr
.sample_freq
;
5445 event
->attr
.sample_period
= NSEC_PER_SEC
/ freq
;
5446 hwc
->sample_period
= event
->attr
.sample_period
;
5447 local64_set(&hwc
->period_left
, hwc
->sample_period
);
5448 event
->attr
.freq
= 0;
5453 * Software event: cpu wall time clock
5456 static void cpu_clock_event_update(struct perf_event
*event
)
5461 now
= local_clock();
5462 prev
= local64_xchg(&event
->hw
.prev_count
, now
);
5463 local64_add(now
- prev
, &event
->count
);
5466 static void cpu_clock_event_start(struct perf_event
*event
, int flags
)
5468 local64_set(&event
->hw
.prev_count
, local_clock());
5469 perf_swevent_start_hrtimer(event
);
5472 static void cpu_clock_event_stop(struct perf_event
*event
, int flags
)
5474 perf_swevent_cancel_hrtimer(event
);
5475 cpu_clock_event_update(event
);
5478 static int cpu_clock_event_add(struct perf_event
*event
, int flags
)
5480 if (flags
& PERF_EF_START
)
5481 cpu_clock_event_start(event
, flags
);
5486 static void cpu_clock_event_del(struct perf_event
*event
, int flags
)
5488 cpu_clock_event_stop(event
, flags
);
5491 static void cpu_clock_event_read(struct perf_event
*event
)
5493 cpu_clock_event_update(event
);
5496 static int cpu_clock_event_init(struct perf_event
*event
)
5498 if (event
->attr
.type
!= PERF_TYPE_SOFTWARE
)
5501 if (event
->attr
.config
!= PERF_COUNT_SW_CPU_CLOCK
)
5505 * no branch sampling for software events
5507 if (has_branch_stack(event
))
5510 perf_swevent_init_hrtimer(event
);
5515 static struct pmu perf_cpu_clock
= {
5516 .task_ctx_nr
= perf_sw_context
,
5518 .event_init
= cpu_clock_event_init
,
5519 .add
= cpu_clock_event_add
,
5520 .del
= cpu_clock_event_del
,
5521 .start
= cpu_clock_event_start
,
5522 .stop
= cpu_clock_event_stop
,
5523 .read
= cpu_clock_event_read
,
5525 .event_idx
= perf_swevent_event_idx
,
5529 * Software event: task time clock
5532 static void task_clock_event_update(struct perf_event
*event
, u64 now
)
5537 prev
= local64_xchg(&event
->hw
.prev_count
, now
);
5539 local64_add(delta
, &event
->count
);
5542 static void task_clock_event_start(struct perf_event
*event
, int flags
)
5544 local64_set(&event
->hw
.prev_count
, event
->ctx
->time
);
5545 perf_swevent_start_hrtimer(event
);
5548 static void task_clock_event_stop(struct perf_event
*event
, int flags
)
5550 perf_swevent_cancel_hrtimer(event
);
5551 task_clock_event_update(event
, event
->ctx
->time
);
5554 static int task_clock_event_add(struct perf_event
*event
, int flags
)
5556 if (flags
& PERF_EF_START
)
5557 task_clock_event_start(event
, flags
);
5562 static void task_clock_event_del(struct perf_event
*event
, int flags
)
5564 task_clock_event_stop(event
, PERF_EF_UPDATE
);
5567 static void task_clock_event_read(struct perf_event
*event
)
5569 u64 now
= perf_clock();
5570 u64 delta
= now
- event
->ctx
->timestamp
;
5571 u64 time
= event
->ctx
->time
+ delta
;
5573 task_clock_event_update(event
, time
);
5576 static int task_clock_event_init(struct perf_event
*event
)
5578 if (event
->attr
.type
!= PERF_TYPE_SOFTWARE
)
5581 if (event
->attr
.config
!= PERF_COUNT_SW_TASK_CLOCK
)
5585 * no branch sampling for software events
5587 if (has_branch_stack(event
))
5590 perf_swevent_init_hrtimer(event
);
5595 static struct pmu perf_task_clock
= {
5596 .task_ctx_nr
= perf_sw_context
,
5598 .event_init
= task_clock_event_init
,
5599 .add
= task_clock_event_add
,
5600 .del
= task_clock_event_del
,
5601 .start
= task_clock_event_start
,
5602 .stop
= task_clock_event_stop
,
5603 .read
= task_clock_event_read
,
5605 .event_idx
= perf_swevent_event_idx
,
5608 static void perf_pmu_nop_void(struct pmu
*pmu
)
5612 static int perf_pmu_nop_int(struct pmu
*pmu
)
5617 static void perf_pmu_start_txn(struct pmu
*pmu
)
5619 perf_pmu_disable(pmu
);
5622 static int perf_pmu_commit_txn(struct pmu
*pmu
)
5624 perf_pmu_enable(pmu
);
5628 static void perf_pmu_cancel_txn(struct pmu
*pmu
)
5630 perf_pmu_enable(pmu
);
5633 static int perf_event_idx_default(struct perf_event
*event
)
5635 return event
->hw
.idx
+ 1;
5639 * Ensures all contexts with the same task_ctx_nr have the same
5640 * pmu_cpu_context too.
5642 static void *find_pmu_context(int ctxn
)
5649 list_for_each_entry(pmu
, &pmus
, entry
) {
5650 if (pmu
->task_ctx_nr
== ctxn
)
5651 return pmu
->pmu_cpu_context
;
5657 static void update_pmu_context(struct pmu
*pmu
, struct pmu
*old_pmu
)
5661 for_each_possible_cpu(cpu
) {
5662 struct perf_cpu_context
*cpuctx
;
5664 cpuctx
= per_cpu_ptr(pmu
->pmu_cpu_context
, cpu
);
5666 if (cpuctx
->active_pmu
== old_pmu
)
5667 cpuctx
->active_pmu
= pmu
;
5671 static void free_pmu_context(struct pmu
*pmu
)
5675 mutex_lock(&pmus_lock
);
5677 * Like a real lame refcount.
5679 list_for_each_entry(i
, &pmus
, entry
) {
5680 if (i
->pmu_cpu_context
== pmu
->pmu_cpu_context
) {
5681 update_pmu_context(i
, pmu
);
5686 free_percpu(pmu
->pmu_cpu_context
);
5688 mutex_unlock(&pmus_lock
);
5690 static struct idr pmu_idr
;
5693 type_show(struct device
*dev
, struct device_attribute
*attr
, char *page
)
5695 struct pmu
*pmu
= dev_get_drvdata(dev
);
5697 return snprintf(page
, PAGE_SIZE
-1, "%d\n", pmu
->type
);
5700 static struct device_attribute pmu_dev_attrs
[] = {
5705 static int pmu_bus_running
;
5706 static struct bus_type pmu_bus
= {
5707 .name
= "event_source",
5708 .dev_attrs
= pmu_dev_attrs
,
5711 static void pmu_dev_release(struct device
*dev
)
5716 static int pmu_dev_alloc(struct pmu
*pmu
)
5720 pmu
->dev
= kzalloc(sizeof(struct device
), GFP_KERNEL
);
5724 pmu
->dev
->groups
= pmu
->attr_groups
;
5725 device_initialize(pmu
->dev
);
5726 ret
= dev_set_name(pmu
->dev
, "%s", pmu
->name
);
5730 dev_set_drvdata(pmu
->dev
, pmu
);
5731 pmu
->dev
->bus
= &pmu_bus
;
5732 pmu
->dev
->release
= pmu_dev_release
;
5733 ret
= device_add(pmu
->dev
);
5741 put_device(pmu
->dev
);
5745 static struct lock_class_key cpuctx_mutex
;
5746 static struct lock_class_key cpuctx_lock
;
5748 int perf_pmu_register(struct pmu
*pmu
, char *name
, int type
)
5752 mutex_lock(&pmus_lock
);
5754 pmu
->pmu_disable_count
= alloc_percpu(int);
5755 if (!pmu
->pmu_disable_count
)
5764 int err
= idr_pre_get(&pmu_idr
, GFP_KERNEL
);
5768 err
= idr_get_new_above(&pmu_idr
, pmu
, PERF_TYPE_MAX
, &type
);
5776 if (pmu_bus_running
) {
5777 ret
= pmu_dev_alloc(pmu
);
5783 pmu
->pmu_cpu_context
= find_pmu_context(pmu
->task_ctx_nr
);
5784 if (pmu
->pmu_cpu_context
)
5785 goto got_cpu_context
;
5787 pmu
->pmu_cpu_context
= alloc_percpu(struct perf_cpu_context
);
5788 if (!pmu
->pmu_cpu_context
)
5791 for_each_possible_cpu(cpu
) {
5792 struct perf_cpu_context
*cpuctx
;
5794 cpuctx
= per_cpu_ptr(pmu
->pmu_cpu_context
, cpu
);
5795 __perf_event_init_context(&cpuctx
->ctx
);
5796 lockdep_set_class(&cpuctx
->ctx
.mutex
, &cpuctx_mutex
);
5797 lockdep_set_class(&cpuctx
->ctx
.lock
, &cpuctx_lock
);
5798 cpuctx
->ctx
.type
= cpu_context
;
5799 cpuctx
->ctx
.pmu
= pmu
;
5800 cpuctx
->jiffies_interval
= 1;
5801 INIT_LIST_HEAD(&cpuctx
->rotation_list
);
5802 cpuctx
->active_pmu
= pmu
;
5806 if (!pmu
->start_txn
) {
5807 if (pmu
->pmu_enable
) {
5809 * If we have pmu_enable/pmu_disable calls, install
5810 * transaction stubs that use that to try and batch
5811 * hardware accesses.
5813 pmu
->start_txn
= perf_pmu_start_txn
;
5814 pmu
->commit_txn
= perf_pmu_commit_txn
;
5815 pmu
->cancel_txn
= perf_pmu_cancel_txn
;
5817 pmu
->start_txn
= perf_pmu_nop_void
;
5818 pmu
->commit_txn
= perf_pmu_nop_int
;
5819 pmu
->cancel_txn
= perf_pmu_nop_void
;
5823 if (!pmu
->pmu_enable
) {
5824 pmu
->pmu_enable
= perf_pmu_nop_void
;
5825 pmu
->pmu_disable
= perf_pmu_nop_void
;
5828 if (!pmu
->event_idx
)
5829 pmu
->event_idx
= perf_event_idx_default
;
5831 list_add_rcu(&pmu
->entry
, &pmus
);
5834 mutex_unlock(&pmus_lock
);
5839 device_del(pmu
->dev
);
5840 put_device(pmu
->dev
);
5843 if (pmu
->type
>= PERF_TYPE_MAX
)
5844 idr_remove(&pmu_idr
, pmu
->type
);
5847 free_percpu(pmu
->pmu_disable_count
);
5851 void perf_pmu_unregister(struct pmu
*pmu
)
5853 mutex_lock(&pmus_lock
);
5854 list_del_rcu(&pmu
->entry
);
5855 mutex_unlock(&pmus_lock
);
5858 * We dereference the pmu list under both SRCU and regular RCU, so
5859 * synchronize against both of those.
5861 synchronize_srcu(&pmus_srcu
);
5864 free_percpu(pmu
->pmu_disable_count
);
5865 if (pmu
->type
>= PERF_TYPE_MAX
)
5866 idr_remove(&pmu_idr
, pmu
->type
);
5867 device_del(pmu
->dev
);
5868 put_device(pmu
->dev
);
5869 free_pmu_context(pmu
);
5872 struct pmu
*perf_init_event(struct perf_event
*event
)
5874 struct pmu
*pmu
= NULL
;
5878 idx
= srcu_read_lock(&pmus_srcu
);
5881 pmu
= idr_find(&pmu_idr
, event
->attr
.type
);
5885 ret
= pmu
->event_init(event
);
5891 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
5893 ret
= pmu
->event_init(event
);
5897 if (ret
!= -ENOENT
) {
5902 pmu
= ERR_PTR(-ENOENT
);
5904 srcu_read_unlock(&pmus_srcu
, idx
);
5910 * Allocate and initialize a event structure
5912 static struct perf_event
*
5913 perf_event_alloc(struct perf_event_attr
*attr
, int cpu
,
5914 struct task_struct
*task
,
5915 struct perf_event
*group_leader
,
5916 struct perf_event
*parent_event
,
5917 perf_overflow_handler_t overflow_handler
,
5921 struct perf_event
*event
;
5922 struct hw_perf_event
*hwc
;
5925 if ((unsigned)cpu
>= nr_cpu_ids
) {
5926 if (!task
|| cpu
!= -1)
5927 return ERR_PTR(-EINVAL
);
5930 event
= kzalloc(sizeof(*event
), GFP_KERNEL
);
5932 return ERR_PTR(-ENOMEM
);
5935 * Single events are their own group leaders, with an
5936 * empty sibling list:
5939 group_leader
= event
;
5941 mutex_init(&event
->child_mutex
);
5942 INIT_LIST_HEAD(&event
->child_list
);
5944 INIT_LIST_HEAD(&event
->group_entry
);
5945 INIT_LIST_HEAD(&event
->event_entry
);
5946 INIT_LIST_HEAD(&event
->sibling_list
);
5947 INIT_LIST_HEAD(&event
->rb_entry
);
5949 init_waitqueue_head(&event
->waitq
);
5950 init_irq_work(&event
->pending
, perf_pending_event
);
5952 mutex_init(&event
->mmap_mutex
);
5954 atomic_long_set(&event
->refcount
, 1);
5956 event
->attr
= *attr
;
5957 event
->group_leader
= group_leader
;
5961 event
->parent
= parent_event
;
5963 event
->ns
= get_pid_ns(current
->nsproxy
->pid_ns
);
5964 event
->id
= atomic64_inc_return(&perf_event_id
);
5966 event
->state
= PERF_EVENT_STATE_INACTIVE
;
5969 event
->attach_state
= PERF_ATTACH_TASK
;
5970 #ifdef CONFIG_HAVE_HW_BREAKPOINT
5972 * hw_breakpoint is a bit difficult here..
5974 if (attr
->type
== PERF_TYPE_BREAKPOINT
)
5975 event
->hw
.bp_target
= task
;
5979 if (!overflow_handler
&& parent_event
) {
5980 overflow_handler
= parent_event
->overflow_handler
;
5981 context
= parent_event
->overflow_handler_context
;
5984 event
->overflow_handler
= overflow_handler
;
5985 event
->overflow_handler_context
= context
;
5988 event
->state
= PERF_EVENT_STATE_OFF
;
5993 hwc
->sample_period
= attr
->sample_period
;
5994 if (attr
->freq
&& attr
->sample_freq
)
5995 hwc
->sample_period
= 1;
5996 hwc
->last_period
= hwc
->sample_period
;
5998 local64_set(&hwc
->period_left
, hwc
->sample_period
);
6001 * we currently do not support PERF_FORMAT_GROUP on inherited events
6003 if (attr
->inherit
&& (attr
->read_format
& PERF_FORMAT_GROUP
))
6006 pmu
= perf_init_event(event
);
6012 else if (IS_ERR(pmu
))
6017 put_pid_ns(event
->ns
);
6019 return ERR_PTR(err
);
6022 if (!event
->parent
) {
6023 if (event
->attach_state
& PERF_ATTACH_TASK
)
6024 static_key_slow_inc(&perf_sched_events
.key
);
6025 if (event
->attr
.mmap
|| event
->attr
.mmap_data
)
6026 atomic_inc(&nr_mmap_events
);
6027 if (event
->attr
.comm
)
6028 atomic_inc(&nr_comm_events
);
6029 if (event
->attr
.task
)
6030 atomic_inc(&nr_task_events
);
6031 if (event
->attr
.sample_type
& PERF_SAMPLE_CALLCHAIN
) {
6032 err
= get_callchain_buffers();
6035 return ERR_PTR(err
);
6038 if (has_branch_stack(event
)) {
6039 static_key_slow_inc(&perf_sched_events
.key
);
6040 if (!(event
->attach_state
& PERF_ATTACH_TASK
))
6041 atomic_inc(&per_cpu(perf_branch_stack_events
,
6049 static int perf_copy_attr(struct perf_event_attr __user
*uattr
,
6050 struct perf_event_attr
*attr
)
6055 if (!access_ok(VERIFY_WRITE
, uattr
, PERF_ATTR_SIZE_VER0
))
6059 * zero the full structure, so that a short copy will be nice.
6061 memset(attr
, 0, sizeof(*attr
));
6063 ret
= get_user(size
, &uattr
->size
);
6067 if (size
> PAGE_SIZE
) /* silly large */
6070 if (!size
) /* abi compat */
6071 size
= PERF_ATTR_SIZE_VER0
;
6073 if (size
< PERF_ATTR_SIZE_VER0
)
6077 * If we're handed a bigger struct than we know of,
6078 * ensure all the unknown bits are 0 - i.e. new
6079 * user-space does not rely on any kernel feature
6080 * extensions we dont know about yet.
6082 if (size
> sizeof(*attr
)) {
6083 unsigned char __user
*addr
;
6084 unsigned char __user
*end
;
6087 addr
= (void __user
*)uattr
+ sizeof(*attr
);
6088 end
= (void __user
*)uattr
+ size
;
6090 for (; addr
< end
; addr
++) {
6091 ret
= get_user(val
, addr
);
6097 size
= sizeof(*attr
);
6100 ret
= copy_from_user(attr
, uattr
, size
);
6104 if (attr
->__reserved_1
)
6107 if (attr
->sample_type
& ~(PERF_SAMPLE_MAX
-1))
6110 if (attr
->read_format
& ~(PERF_FORMAT_MAX
-1))
6113 if (attr
->sample_type
& PERF_SAMPLE_BRANCH_STACK
) {
6114 u64 mask
= attr
->branch_sample_type
;
6116 /* only using defined bits */
6117 if (mask
& ~(PERF_SAMPLE_BRANCH_MAX
-1))
6120 /* at least one branch bit must be set */
6121 if (!(mask
& ~PERF_SAMPLE_BRANCH_PLM_ALL
))
6124 /* kernel level capture: check permissions */
6125 if ((mask
& PERF_SAMPLE_BRANCH_PERM_PLM
)
6126 && perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN
))
6129 /* propagate priv level, when not set for branch */
6130 if (!(mask
& PERF_SAMPLE_BRANCH_PLM_ALL
)) {
6132 /* exclude_kernel checked on syscall entry */
6133 if (!attr
->exclude_kernel
)
6134 mask
|= PERF_SAMPLE_BRANCH_KERNEL
;
6136 if (!attr
->exclude_user
)
6137 mask
|= PERF_SAMPLE_BRANCH_USER
;
6139 if (!attr
->exclude_hv
)
6140 mask
|= PERF_SAMPLE_BRANCH_HV
;
6142 * adjust user setting (for HW filter setup)
6144 attr
->branch_sample_type
= mask
;
6151 put_user(sizeof(*attr
), &uattr
->size
);
6157 perf_event_set_output(struct perf_event
*event
, struct perf_event
*output_event
)
6159 struct ring_buffer
*rb
= NULL
, *old_rb
= NULL
;
6165 /* don't allow circular references */
6166 if (event
== output_event
)
6170 * Don't allow cross-cpu buffers
6172 if (output_event
->cpu
!= event
->cpu
)
6176 * If its not a per-cpu rb, it must be the same task.
6178 if (output_event
->cpu
== -1 && output_event
->ctx
!= event
->ctx
)
6182 mutex_lock(&event
->mmap_mutex
);
6183 /* Can't redirect output if we've got an active mmap() */
6184 if (atomic_read(&event
->mmap_count
))
6188 /* get the rb we want to redirect to */
6189 rb
= ring_buffer_get(output_event
);
6195 rcu_assign_pointer(event
->rb
, rb
);
6197 ring_buffer_detach(event
, old_rb
);
6200 mutex_unlock(&event
->mmap_mutex
);
6203 ring_buffer_put(old_rb
);
6209 * sys_perf_event_open - open a performance event, associate it to a task/cpu
6211 * @attr_uptr: event_id type attributes for monitoring/sampling
6214 * @group_fd: group leader event fd
6216 SYSCALL_DEFINE5(perf_event_open
,
6217 struct perf_event_attr __user
*, attr_uptr
,
6218 pid_t
, pid
, int, cpu
, int, group_fd
, unsigned long, flags
)
6220 struct perf_event
*group_leader
= NULL
, *output_event
= NULL
;
6221 struct perf_event
*event
, *sibling
;
6222 struct perf_event_attr attr
;
6223 struct perf_event_context
*ctx
;
6224 struct file
*event_file
= NULL
;
6225 struct fd group
= {NULL
, 0};
6226 struct task_struct
*task
= NULL
;
6232 /* for future expandability... */
6233 if (flags
& ~PERF_FLAG_ALL
)
6236 err
= perf_copy_attr(attr_uptr
, &attr
);
6240 if (!attr
.exclude_kernel
) {
6241 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN
))
6246 if (attr
.sample_freq
> sysctl_perf_event_sample_rate
)
6251 * In cgroup mode, the pid argument is used to pass the fd
6252 * opened to the cgroup directory in cgroupfs. The cpu argument
6253 * designates the cpu on which to monitor threads from that
6256 if ((flags
& PERF_FLAG_PID_CGROUP
) && (pid
== -1 || cpu
== -1))
6259 event_fd
= get_unused_fd();
6263 if (group_fd
!= -1) {
6264 err
= perf_fget_light(group_fd
, &group
);
6267 group_leader
= group
.file
->private_data
;
6268 if (flags
& PERF_FLAG_FD_OUTPUT
)
6269 output_event
= group_leader
;
6270 if (flags
& PERF_FLAG_FD_NO_GROUP
)
6271 group_leader
= NULL
;
6274 if (pid
!= -1 && !(flags
& PERF_FLAG_PID_CGROUP
)) {
6275 task
= find_lively_task_by_vpid(pid
);
6277 err
= PTR_ERR(task
);
6284 event
= perf_event_alloc(&attr
, cpu
, task
, group_leader
, NULL
,
6286 if (IS_ERR(event
)) {
6287 err
= PTR_ERR(event
);
6291 if (flags
& PERF_FLAG_PID_CGROUP
) {
6292 err
= perf_cgroup_connect(pid
, event
, &attr
, group_leader
);
6297 * - that has cgroup constraint on event->cpu
6298 * - that may need work on context switch
6300 atomic_inc(&per_cpu(perf_cgroup_events
, event
->cpu
));
6301 static_key_slow_inc(&perf_sched_events
.key
);
6305 * Special case software events and allow them to be part of
6306 * any hardware group.
6311 (is_software_event(event
) != is_software_event(group_leader
))) {
6312 if (is_software_event(event
)) {
6314 * If event and group_leader are not both a software
6315 * event, and event is, then group leader is not.
6317 * Allow the addition of software events to !software
6318 * groups, this is safe because software events never
6321 pmu
= group_leader
->pmu
;
6322 } else if (is_software_event(group_leader
) &&
6323 (group_leader
->group_flags
& PERF_GROUP_SOFTWARE
)) {
6325 * In case the group is a pure software group, and we
6326 * try to add a hardware event, move the whole group to
6327 * the hardware context.
6334 * Get the target context (task or percpu):
6336 ctx
= find_get_context(pmu
, task
, event
->cpu
);
6343 put_task_struct(task
);
6348 * Look up the group leader (we will attach this event to it):
6354 * Do not allow a recursive hierarchy (this new sibling
6355 * becoming part of another group-sibling):
6357 if (group_leader
->group_leader
!= group_leader
)
6360 * Do not allow to attach to a group in a different
6361 * task or CPU context:
6364 if (group_leader
->ctx
->type
!= ctx
->type
)
6367 if (group_leader
->ctx
!= ctx
)
6372 * Only a group leader can be exclusive or pinned
6374 if (attr
.exclusive
|| attr
.pinned
)
6379 err
= perf_event_set_output(event
, output_event
);
6384 event_file
= anon_inode_getfile("[perf_event]", &perf_fops
, event
, O_RDWR
);
6385 if (IS_ERR(event_file
)) {
6386 err
= PTR_ERR(event_file
);
6391 struct perf_event_context
*gctx
= group_leader
->ctx
;
6393 mutex_lock(&gctx
->mutex
);
6394 perf_remove_from_context(group_leader
);
6395 list_for_each_entry(sibling
, &group_leader
->sibling_list
,
6397 perf_remove_from_context(sibling
);
6400 mutex_unlock(&gctx
->mutex
);
6404 WARN_ON_ONCE(ctx
->parent_ctx
);
6405 mutex_lock(&ctx
->mutex
);
6409 perf_install_in_context(ctx
, group_leader
, event
->cpu
);
6411 list_for_each_entry(sibling
, &group_leader
->sibling_list
,
6413 perf_install_in_context(ctx
, sibling
, event
->cpu
);
6418 perf_install_in_context(ctx
, event
, event
->cpu
);
6420 perf_unpin_context(ctx
);
6421 mutex_unlock(&ctx
->mutex
);
6425 event
->owner
= current
;
6427 mutex_lock(¤t
->perf_event_mutex
);
6428 list_add_tail(&event
->owner_entry
, ¤t
->perf_event_list
);
6429 mutex_unlock(¤t
->perf_event_mutex
);
6432 * Precalculate sample_data sizes
6434 perf_event__header_size(event
);
6435 perf_event__id_header_size(event
);
6438 * Drop the reference on the group_event after placing the
6439 * new event on the sibling_list. This ensures destruction
6440 * of the group leader will find the pointer to itself in
6441 * perf_group_detach().
6444 fd_install(event_fd
, event_file
);
6448 perf_unpin_context(ctx
);
6455 put_task_struct(task
);
6459 put_unused_fd(event_fd
);
6464 * perf_event_create_kernel_counter
6466 * @attr: attributes of the counter to create
6467 * @cpu: cpu in which the counter is bound
6468 * @task: task to profile (NULL for percpu)
6471 perf_event_create_kernel_counter(struct perf_event_attr
*attr
, int cpu
,
6472 struct task_struct
*task
,
6473 perf_overflow_handler_t overflow_handler
,
6476 struct perf_event_context
*ctx
;
6477 struct perf_event
*event
;
6481 * Get the target context (task or percpu):
6484 event
= perf_event_alloc(attr
, cpu
, task
, NULL
, NULL
,
6485 overflow_handler
, context
);
6486 if (IS_ERR(event
)) {
6487 err
= PTR_ERR(event
);
6491 ctx
= find_get_context(event
->pmu
, task
, cpu
);
6497 WARN_ON_ONCE(ctx
->parent_ctx
);
6498 mutex_lock(&ctx
->mutex
);
6499 perf_install_in_context(ctx
, event
, cpu
);
6501 perf_unpin_context(ctx
);
6502 mutex_unlock(&ctx
->mutex
);
6509 return ERR_PTR(err
);
6511 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter
);
6513 void perf_pmu_migrate_context(struct pmu
*pmu
, int src_cpu
, int dst_cpu
)
6515 struct perf_event_context
*src_ctx
;
6516 struct perf_event_context
*dst_ctx
;
6517 struct perf_event
*event
, *tmp
;
6520 src_ctx
= &per_cpu_ptr(pmu
->pmu_cpu_context
, src_cpu
)->ctx
;
6521 dst_ctx
= &per_cpu_ptr(pmu
->pmu_cpu_context
, dst_cpu
)->ctx
;
6523 mutex_lock(&src_ctx
->mutex
);
6524 list_for_each_entry_safe(event
, tmp
, &src_ctx
->event_list
,
6526 perf_remove_from_context(event
);
6528 list_add(&event
->event_entry
, &events
);
6530 mutex_unlock(&src_ctx
->mutex
);
6534 mutex_lock(&dst_ctx
->mutex
);
6535 list_for_each_entry_safe(event
, tmp
, &events
, event_entry
) {
6536 list_del(&event
->event_entry
);
6537 if (event
->state
>= PERF_EVENT_STATE_OFF
)
6538 event
->state
= PERF_EVENT_STATE_INACTIVE
;
6539 perf_install_in_context(dst_ctx
, event
, dst_cpu
);
6542 mutex_unlock(&dst_ctx
->mutex
);
6544 EXPORT_SYMBOL_GPL(perf_pmu_migrate_context
);
6546 static void sync_child_event(struct perf_event
*child_event
,
6547 struct task_struct
*child
)
6549 struct perf_event
*parent_event
= child_event
->parent
;
6552 if (child_event
->attr
.inherit_stat
)
6553 perf_event_read_event(child_event
, child
);
6555 child_val
= perf_event_count(child_event
);
6558 * Add back the child's count to the parent's count:
6560 atomic64_add(child_val
, &parent_event
->child_count
);
6561 atomic64_add(child_event
->total_time_enabled
,
6562 &parent_event
->child_total_time_enabled
);
6563 atomic64_add(child_event
->total_time_running
,
6564 &parent_event
->child_total_time_running
);
6567 * Remove this event from the parent's list
6569 WARN_ON_ONCE(parent_event
->ctx
->parent_ctx
);
6570 mutex_lock(&parent_event
->child_mutex
);
6571 list_del_init(&child_event
->child_list
);
6572 mutex_unlock(&parent_event
->child_mutex
);
6575 * Release the parent event, if this was the last
6578 put_event(parent_event
);
6582 __perf_event_exit_task(struct perf_event
*child_event
,
6583 struct perf_event_context
*child_ctx
,
6584 struct task_struct
*child
)
6586 if (child_event
->parent
) {
6587 raw_spin_lock_irq(&child_ctx
->lock
);
6588 perf_group_detach(child_event
);
6589 raw_spin_unlock_irq(&child_ctx
->lock
);
6592 perf_remove_from_context(child_event
);
6595 * It can happen that the parent exits first, and has events
6596 * that are still around due to the child reference. These
6597 * events need to be zapped.
6599 if (child_event
->parent
) {
6600 sync_child_event(child_event
, child
);
6601 free_event(child_event
);
6605 static void perf_event_exit_task_context(struct task_struct
*child
, int ctxn
)
6607 struct perf_event
*child_event
, *tmp
;
6608 struct perf_event_context
*child_ctx
;
6609 unsigned long flags
;
6611 if (likely(!child
->perf_event_ctxp
[ctxn
])) {
6612 perf_event_task(child
, NULL
, 0);
6616 local_irq_save(flags
);
6618 * We can't reschedule here because interrupts are disabled,
6619 * and either child is current or it is a task that can't be
6620 * scheduled, so we are now safe from rescheduling changing
6623 child_ctx
= rcu_dereference_raw(child
->perf_event_ctxp
[ctxn
]);
6626 * Take the context lock here so that if find_get_context is
6627 * reading child->perf_event_ctxp, we wait until it has
6628 * incremented the context's refcount before we do put_ctx below.
6630 raw_spin_lock(&child_ctx
->lock
);
6631 task_ctx_sched_out(child_ctx
);
6632 child
->perf_event_ctxp
[ctxn
] = NULL
;
6634 * If this context is a clone; unclone it so it can't get
6635 * swapped to another process while we're removing all
6636 * the events from it.
6638 unclone_ctx(child_ctx
);
6639 update_context_time(child_ctx
);
6640 raw_spin_unlock_irqrestore(&child_ctx
->lock
, flags
);
6643 * Report the task dead after unscheduling the events so that we
6644 * won't get any samples after PERF_RECORD_EXIT. We can however still
6645 * get a few PERF_RECORD_READ events.
6647 perf_event_task(child
, child_ctx
, 0);
6650 * We can recurse on the same lock type through:
6652 * __perf_event_exit_task()
6653 * sync_child_event()
6655 * mutex_lock(&ctx->mutex)
6657 * But since its the parent context it won't be the same instance.
6659 mutex_lock(&child_ctx
->mutex
);
6662 list_for_each_entry_safe(child_event
, tmp
, &child_ctx
->pinned_groups
,
6664 __perf_event_exit_task(child_event
, child_ctx
, child
);
6666 list_for_each_entry_safe(child_event
, tmp
, &child_ctx
->flexible_groups
,
6668 __perf_event_exit_task(child_event
, child_ctx
, child
);
6671 * If the last event was a group event, it will have appended all
6672 * its siblings to the list, but we obtained 'tmp' before that which
6673 * will still point to the list head terminating the iteration.
6675 if (!list_empty(&child_ctx
->pinned_groups
) ||
6676 !list_empty(&child_ctx
->flexible_groups
))
6679 mutex_unlock(&child_ctx
->mutex
);
6685 * When a child task exits, feed back event values to parent events.
6687 void perf_event_exit_task(struct task_struct
*child
)
6689 struct perf_event
*event
, *tmp
;
6692 mutex_lock(&child
->perf_event_mutex
);
6693 list_for_each_entry_safe(event
, tmp
, &child
->perf_event_list
,
6695 list_del_init(&event
->owner_entry
);
6698 * Ensure the list deletion is visible before we clear
6699 * the owner, closes a race against perf_release() where
6700 * we need to serialize on the owner->perf_event_mutex.
6703 event
->owner
= NULL
;
6705 mutex_unlock(&child
->perf_event_mutex
);
6707 for_each_task_context_nr(ctxn
)
6708 perf_event_exit_task_context(child
, ctxn
);
6711 static void perf_free_event(struct perf_event
*event
,
6712 struct perf_event_context
*ctx
)
6714 struct perf_event
*parent
= event
->parent
;
6716 if (WARN_ON_ONCE(!parent
))
6719 mutex_lock(&parent
->child_mutex
);
6720 list_del_init(&event
->child_list
);
6721 mutex_unlock(&parent
->child_mutex
);
6725 perf_group_detach(event
);
6726 list_del_event(event
, ctx
);
6731 * free an unexposed, unused context as created by inheritance by
6732 * perf_event_init_task below, used by fork() in case of fail.
6734 void perf_event_free_task(struct task_struct
*task
)
6736 struct perf_event_context
*ctx
;
6737 struct perf_event
*event
, *tmp
;
6740 for_each_task_context_nr(ctxn
) {
6741 ctx
= task
->perf_event_ctxp
[ctxn
];
6745 mutex_lock(&ctx
->mutex
);
6747 list_for_each_entry_safe(event
, tmp
, &ctx
->pinned_groups
,
6749 perf_free_event(event
, ctx
);
6751 list_for_each_entry_safe(event
, tmp
, &ctx
->flexible_groups
,
6753 perf_free_event(event
, ctx
);
6755 if (!list_empty(&ctx
->pinned_groups
) ||
6756 !list_empty(&ctx
->flexible_groups
))
6759 mutex_unlock(&ctx
->mutex
);
6765 void perf_event_delayed_put(struct task_struct
*task
)
6769 for_each_task_context_nr(ctxn
)
6770 WARN_ON_ONCE(task
->perf_event_ctxp
[ctxn
]);
6774 * inherit a event from parent task to child task:
6776 static struct perf_event
*
6777 inherit_event(struct perf_event
*parent_event
,
6778 struct task_struct
*parent
,
6779 struct perf_event_context
*parent_ctx
,
6780 struct task_struct
*child
,
6781 struct perf_event
*group_leader
,
6782 struct perf_event_context
*child_ctx
)
6784 struct perf_event
*child_event
;
6785 unsigned long flags
;
6788 * Instead of creating recursive hierarchies of events,
6789 * we link inherited events back to the original parent,
6790 * which has a filp for sure, which we use as the reference
6793 if (parent_event
->parent
)
6794 parent_event
= parent_event
->parent
;
6796 child_event
= perf_event_alloc(&parent_event
->attr
,
6799 group_leader
, parent_event
,
6801 if (IS_ERR(child_event
))
6804 if (!atomic_long_inc_not_zero(&parent_event
->refcount
)) {
6805 free_event(child_event
);
6812 * Make the child state follow the state of the parent event,
6813 * not its attr.disabled bit. We hold the parent's mutex,
6814 * so we won't race with perf_event_{en, dis}able_family.
6816 if (parent_event
->state
>= PERF_EVENT_STATE_INACTIVE
)
6817 child_event
->state
= PERF_EVENT_STATE_INACTIVE
;
6819 child_event
->state
= PERF_EVENT_STATE_OFF
;
6821 if (parent_event
->attr
.freq
) {
6822 u64 sample_period
= parent_event
->hw
.sample_period
;
6823 struct hw_perf_event
*hwc
= &child_event
->hw
;
6825 hwc
->sample_period
= sample_period
;
6826 hwc
->last_period
= sample_period
;
6828 local64_set(&hwc
->period_left
, sample_period
);
6831 child_event
->ctx
= child_ctx
;
6832 child_event
->overflow_handler
= parent_event
->overflow_handler
;
6833 child_event
->overflow_handler_context
6834 = parent_event
->overflow_handler_context
;
6837 * Precalculate sample_data sizes
6839 perf_event__header_size(child_event
);
6840 perf_event__id_header_size(child_event
);
6843 * Link it up in the child's context:
6845 raw_spin_lock_irqsave(&child_ctx
->lock
, flags
);
6846 add_event_to_ctx(child_event
, child_ctx
);
6847 raw_spin_unlock_irqrestore(&child_ctx
->lock
, flags
);
6850 * Link this into the parent event's child list
6852 WARN_ON_ONCE(parent_event
->ctx
->parent_ctx
);
6853 mutex_lock(&parent_event
->child_mutex
);
6854 list_add_tail(&child_event
->child_list
, &parent_event
->child_list
);
6855 mutex_unlock(&parent_event
->child_mutex
);
6860 static int inherit_group(struct perf_event
*parent_event
,
6861 struct task_struct
*parent
,
6862 struct perf_event_context
*parent_ctx
,
6863 struct task_struct
*child
,
6864 struct perf_event_context
*child_ctx
)
6866 struct perf_event
*leader
;
6867 struct perf_event
*sub
;
6868 struct perf_event
*child_ctr
;
6870 leader
= inherit_event(parent_event
, parent
, parent_ctx
,
6871 child
, NULL
, child_ctx
);
6873 return PTR_ERR(leader
);
6874 list_for_each_entry(sub
, &parent_event
->sibling_list
, group_entry
) {
6875 child_ctr
= inherit_event(sub
, parent
, parent_ctx
,
6876 child
, leader
, child_ctx
);
6877 if (IS_ERR(child_ctr
))
6878 return PTR_ERR(child_ctr
);
6884 inherit_task_group(struct perf_event
*event
, struct task_struct
*parent
,
6885 struct perf_event_context
*parent_ctx
,
6886 struct task_struct
*child
, int ctxn
,
6890 struct perf_event_context
*child_ctx
;
6892 if (!event
->attr
.inherit
) {
6897 child_ctx
= child
->perf_event_ctxp
[ctxn
];
6900 * This is executed from the parent task context, so
6901 * inherit events that have been marked for cloning.
6902 * First allocate and initialize a context for the
6906 child_ctx
= alloc_perf_context(event
->pmu
, child
);
6910 child
->perf_event_ctxp
[ctxn
] = child_ctx
;
6913 ret
= inherit_group(event
, parent
, parent_ctx
,
6923 * Initialize the perf_event context in task_struct
6925 int perf_event_init_context(struct task_struct
*child
, int ctxn
)
6927 struct perf_event_context
*child_ctx
, *parent_ctx
;
6928 struct perf_event_context
*cloned_ctx
;
6929 struct perf_event
*event
;
6930 struct task_struct
*parent
= current
;
6931 int inherited_all
= 1;
6932 unsigned long flags
;
6935 if (likely(!parent
->perf_event_ctxp
[ctxn
]))
6939 * If the parent's context is a clone, pin it so it won't get
6942 parent_ctx
= perf_pin_task_context(parent
, ctxn
);
6945 * No need to check if parent_ctx != NULL here; since we saw
6946 * it non-NULL earlier, the only reason for it to become NULL
6947 * is if we exit, and since we're currently in the middle of
6948 * a fork we can't be exiting at the same time.
6952 * Lock the parent list. No need to lock the child - not PID
6953 * hashed yet and not running, so nobody can access it.
6955 mutex_lock(&parent_ctx
->mutex
);
6958 * We dont have to disable NMIs - we are only looking at
6959 * the list, not manipulating it:
6961 list_for_each_entry(event
, &parent_ctx
->pinned_groups
, group_entry
) {
6962 ret
= inherit_task_group(event
, parent
, parent_ctx
,
6963 child
, ctxn
, &inherited_all
);
6969 * We can't hold ctx->lock when iterating the ->flexible_group list due
6970 * to allocations, but we need to prevent rotation because
6971 * rotate_ctx() will change the list from interrupt context.
6973 raw_spin_lock_irqsave(&parent_ctx
->lock
, flags
);
6974 parent_ctx
->rotate_disable
= 1;
6975 raw_spin_unlock_irqrestore(&parent_ctx
->lock
, flags
);
6977 list_for_each_entry(event
, &parent_ctx
->flexible_groups
, group_entry
) {
6978 ret
= inherit_task_group(event
, parent
, parent_ctx
,
6979 child
, ctxn
, &inherited_all
);
6984 raw_spin_lock_irqsave(&parent_ctx
->lock
, flags
);
6985 parent_ctx
->rotate_disable
= 0;
6987 child_ctx
= child
->perf_event_ctxp
[ctxn
];
6989 if (child_ctx
&& inherited_all
) {
6991 * Mark the child context as a clone of the parent
6992 * context, or of whatever the parent is a clone of.
6994 * Note that if the parent is a clone, the holding of
6995 * parent_ctx->lock avoids it from being uncloned.
6997 cloned_ctx
= parent_ctx
->parent_ctx
;
6999 child_ctx
->parent_ctx
= cloned_ctx
;
7000 child_ctx
->parent_gen
= parent_ctx
->parent_gen
;
7002 child_ctx
->parent_ctx
= parent_ctx
;
7003 child_ctx
->parent_gen
= parent_ctx
->generation
;
7005 get_ctx(child_ctx
->parent_ctx
);
7008 raw_spin_unlock_irqrestore(&parent_ctx
->lock
, flags
);
7009 mutex_unlock(&parent_ctx
->mutex
);
7011 perf_unpin_context(parent_ctx
);
7012 put_ctx(parent_ctx
);
7018 * Initialize the perf_event context in task_struct
7020 int perf_event_init_task(struct task_struct
*child
)
7024 memset(child
->perf_event_ctxp
, 0, sizeof(child
->perf_event_ctxp
));
7025 mutex_init(&child
->perf_event_mutex
);
7026 INIT_LIST_HEAD(&child
->perf_event_list
);
7028 for_each_task_context_nr(ctxn
) {
7029 ret
= perf_event_init_context(child
, ctxn
);
7037 static void __init
perf_event_init_all_cpus(void)
7039 struct swevent_htable
*swhash
;
7042 for_each_possible_cpu(cpu
) {
7043 swhash
= &per_cpu(swevent_htable
, cpu
);
7044 mutex_init(&swhash
->hlist_mutex
);
7045 INIT_LIST_HEAD(&per_cpu(rotation_list
, cpu
));
7049 static void __cpuinit
perf_event_init_cpu(int cpu
)
7051 struct swevent_htable
*swhash
= &per_cpu(swevent_htable
, cpu
);
7053 mutex_lock(&swhash
->hlist_mutex
);
7054 if (swhash
->hlist_refcount
> 0) {
7055 struct swevent_hlist
*hlist
;
7057 hlist
= kzalloc_node(sizeof(*hlist
), GFP_KERNEL
, cpu_to_node(cpu
));
7059 rcu_assign_pointer(swhash
->swevent_hlist
, hlist
);
7061 mutex_unlock(&swhash
->hlist_mutex
);
7064 #if defined CONFIG_HOTPLUG_CPU || defined CONFIG_KEXEC
7065 static void perf_pmu_rotate_stop(struct pmu
*pmu
)
7067 struct perf_cpu_context
*cpuctx
= this_cpu_ptr(pmu
->pmu_cpu_context
);
7069 WARN_ON(!irqs_disabled());
7071 list_del_init(&cpuctx
->rotation_list
);
7074 static void __perf_event_exit_context(void *__info
)
7076 struct perf_event_context
*ctx
= __info
;
7077 struct perf_event
*event
, *tmp
;
7079 perf_pmu_rotate_stop(ctx
->pmu
);
7081 list_for_each_entry_safe(event
, tmp
, &ctx
->pinned_groups
, group_entry
)
7082 __perf_remove_from_context(event
);
7083 list_for_each_entry_safe(event
, tmp
, &ctx
->flexible_groups
, group_entry
)
7084 __perf_remove_from_context(event
);
7087 static void perf_event_exit_cpu_context(int cpu
)
7089 struct perf_event_context
*ctx
;
7093 idx
= srcu_read_lock(&pmus_srcu
);
7094 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
7095 ctx
= &per_cpu_ptr(pmu
->pmu_cpu_context
, cpu
)->ctx
;
7097 mutex_lock(&ctx
->mutex
);
7098 smp_call_function_single(cpu
, __perf_event_exit_context
, ctx
, 1);
7099 mutex_unlock(&ctx
->mutex
);
7101 srcu_read_unlock(&pmus_srcu
, idx
);
7104 static void perf_event_exit_cpu(int cpu
)
7106 struct swevent_htable
*swhash
= &per_cpu(swevent_htable
, cpu
);
7108 mutex_lock(&swhash
->hlist_mutex
);
7109 swevent_hlist_release(swhash
);
7110 mutex_unlock(&swhash
->hlist_mutex
);
7112 perf_event_exit_cpu_context(cpu
);
7115 static inline void perf_event_exit_cpu(int cpu
) { }
7119 perf_reboot(struct notifier_block
*notifier
, unsigned long val
, void *v
)
7123 for_each_online_cpu(cpu
)
7124 perf_event_exit_cpu(cpu
);
7130 * Run the perf reboot notifier at the very last possible moment so that
7131 * the generic watchdog code runs as long as possible.
7133 static struct notifier_block perf_reboot_notifier
= {
7134 .notifier_call
= perf_reboot
,
7135 .priority
= INT_MIN
,
7138 static int __cpuinit
7139 perf_cpu_notify(struct notifier_block
*self
, unsigned long action
, void *hcpu
)
7141 unsigned int cpu
= (long)hcpu
;
7143 switch (action
& ~CPU_TASKS_FROZEN
) {
7145 case CPU_UP_PREPARE
:
7146 case CPU_DOWN_FAILED
:
7147 perf_event_init_cpu(cpu
);
7150 case CPU_UP_CANCELED
:
7151 case CPU_DOWN_PREPARE
:
7152 perf_event_exit_cpu(cpu
);
7162 void __init
perf_event_init(void)
7168 perf_event_init_all_cpus();
7169 init_srcu_struct(&pmus_srcu
);
7170 perf_pmu_register(&perf_swevent
, "software", PERF_TYPE_SOFTWARE
);
7171 perf_pmu_register(&perf_cpu_clock
, NULL
, -1);
7172 perf_pmu_register(&perf_task_clock
, NULL
, -1);
7174 perf_cpu_notifier(perf_cpu_notify
);
7175 register_reboot_notifier(&perf_reboot_notifier
);
7177 ret
= init_hw_breakpoint();
7178 WARN(ret
, "hw_breakpoint initialization failed with: %d", ret
);
7180 /* do not patch jump label more than once per second */
7181 jump_label_rate_limit(&perf_sched_events
, HZ
);
7184 * Build time assertion that we keep the data_head at the intended
7185 * location. IOW, validation we got the __reserved[] size right.
7187 BUILD_BUG_ON((offsetof(struct perf_event_mmap_page
, data_head
))
7191 static int __init
perf_event_sysfs_init(void)
7196 mutex_lock(&pmus_lock
);
7198 ret
= bus_register(&pmu_bus
);
7202 list_for_each_entry(pmu
, &pmus
, entry
) {
7203 if (!pmu
->name
|| pmu
->type
< 0)
7206 ret
= pmu_dev_alloc(pmu
);
7207 WARN(ret
, "Failed to register pmu: %s, reason %d\n", pmu
->name
, ret
);
7209 pmu_bus_running
= 1;
7213 mutex_unlock(&pmus_lock
);
7217 device_initcall(perf_event_sysfs_init
);
7219 #ifdef CONFIG_CGROUP_PERF
7220 static struct cgroup_subsys_state
*perf_cgroup_create(struct cgroup
*cont
)
7222 struct perf_cgroup
*jc
;
7224 jc
= kzalloc(sizeof(*jc
), GFP_KERNEL
);
7226 return ERR_PTR(-ENOMEM
);
7228 jc
->info
= alloc_percpu(struct perf_cgroup_info
);
7231 return ERR_PTR(-ENOMEM
);
7237 static void perf_cgroup_destroy(struct cgroup
*cont
)
7239 struct perf_cgroup
*jc
;
7240 jc
= container_of(cgroup_subsys_state(cont
, perf_subsys_id
),
7241 struct perf_cgroup
, css
);
7242 free_percpu(jc
->info
);
7246 static int __perf_cgroup_move(void *info
)
7248 struct task_struct
*task
= info
;
7249 perf_cgroup_switch(task
, PERF_CGROUP_SWOUT
| PERF_CGROUP_SWIN
);
7253 static void perf_cgroup_attach(struct cgroup
*cgrp
, struct cgroup_taskset
*tset
)
7255 struct task_struct
*task
;
7257 cgroup_taskset_for_each(task
, cgrp
, tset
)
7258 task_function_call(task
, __perf_cgroup_move
, task
);
7261 static void perf_cgroup_exit(struct cgroup
*cgrp
, struct cgroup
*old_cgrp
,
7262 struct task_struct
*task
)
7265 * cgroup_exit() is called in the copy_process() failure path.
7266 * Ignore this case since the task hasn't ran yet, this avoids
7267 * trying to poke a half freed task state from generic code.
7269 if (!(task
->flags
& PF_EXITING
))
7272 task_function_call(task
, __perf_cgroup_move
, task
);
7275 struct cgroup_subsys perf_subsys
= {
7276 .name
= "perf_event",
7277 .subsys_id
= perf_subsys_id
,
7278 .create
= perf_cgroup_create
,
7279 .destroy
= perf_cgroup_destroy
,
7280 .exit
= perf_cgroup_exit
,
7281 .attach
= perf_cgroup_attach
,
7283 #endif /* CONFIG_CGROUP_PERF */