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>
39 #include <linux/mm_types.h>
43 #include <asm/irq_regs.h>
45 struct remote_function_call
{
46 struct task_struct
*p
;
47 int (*func
)(void *info
);
52 static void remote_function(void *data
)
54 struct remote_function_call
*tfc
= data
;
55 struct task_struct
*p
= tfc
->p
;
59 if (task_cpu(p
) != smp_processor_id() || !task_curr(p
))
63 tfc
->ret
= tfc
->func(tfc
->info
);
67 * task_function_call - call a function on the cpu on which a task runs
68 * @p: the task to evaluate
69 * @func: the function to be called
70 * @info: the function call argument
72 * Calls the function @func when the task is currently running. This might
73 * be on the current CPU, which just calls the function directly
75 * returns: @func return value, or
76 * -ESRCH - when the process isn't running
77 * -EAGAIN - when the process moved away
80 task_function_call(struct task_struct
*p
, int (*func
) (void *info
), void *info
)
82 struct remote_function_call data
= {
86 .ret
= -ESRCH
, /* No such (running) process */
90 smp_call_function_single(task_cpu(p
), remote_function
, &data
, 1);
96 * cpu_function_call - call a function on the cpu
97 * @func: the function to be called
98 * @info: the function call argument
100 * Calls the function @func on the remote cpu.
102 * returns: @func return value or -ENXIO when the cpu is offline
104 static int cpu_function_call(int cpu
, int (*func
) (void *info
), void *info
)
106 struct remote_function_call data
= {
110 .ret
= -ENXIO
, /* No such CPU */
113 smp_call_function_single(cpu
, remote_function
, &data
, 1);
118 #define PERF_FLAG_ALL (PERF_FLAG_FD_NO_GROUP |\
119 PERF_FLAG_FD_OUTPUT |\
120 PERF_FLAG_PID_CGROUP)
123 * branch priv levels that need permission checks
125 #define PERF_SAMPLE_BRANCH_PERM_PLM \
126 (PERF_SAMPLE_BRANCH_KERNEL |\
127 PERF_SAMPLE_BRANCH_HV)
130 EVENT_FLEXIBLE
= 0x1,
132 EVENT_ALL
= EVENT_FLEXIBLE
| EVENT_PINNED
,
136 * perf_sched_events : >0 events exist
137 * perf_cgroup_events: >0 per-cpu cgroup events exist on this cpu
139 struct static_key_deferred perf_sched_events __read_mostly
;
140 static DEFINE_PER_CPU(atomic_t
, perf_cgroup_events
);
141 static DEFINE_PER_CPU(atomic_t
, perf_branch_stack_events
);
143 static atomic_t nr_mmap_events __read_mostly
;
144 static atomic_t nr_comm_events __read_mostly
;
145 static atomic_t nr_task_events __read_mostly
;
147 static LIST_HEAD(pmus
);
148 static DEFINE_MUTEX(pmus_lock
);
149 static struct srcu_struct pmus_srcu
;
152 * perf event paranoia level:
153 * -1 - not paranoid at all
154 * 0 - disallow raw tracepoint access for unpriv
155 * 1 - disallow cpu events for unpriv
156 * 2 - disallow kernel profiling for unpriv
158 int sysctl_perf_event_paranoid __read_mostly
= 1;
160 /* Minimum for 512 kiB + 1 user control page */
161 int sysctl_perf_event_mlock __read_mostly
= 512 + (PAGE_SIZE
/ 1024); /* 'free' kiB per user */
164 * max perf event sample rate
166 #define DEFAULT_MAX_SAMPLE_RATE 100000
167 int sysctl_perf_event_sample_rate __read_mostly
= DEFAULT_MAX_SAMPLE_RATE
;
168 static int max_samples_per_tick __read_mostly
=
169 DIV_ROUND_UP(DEFAULT_MAX_SAMPLE_RATE
, HZ
);
171 int perf_proc_update_handler(struct ctl_table
*table
, int write
,
172 void __user
*buffer
, size_t *lenp
,
175 int ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
180 max_samples_per_tick
= DIV_ROUND_UP(sysctl_perf_event_sample_rate
, HZ
);
185 static atomic64_t perf_event_id
;
187 static void cpu_ctx_sched_out(struct perf_cpu_context
*cpuctx
,
188 enum event_type_t event_type
);
190 static void cpu_ctx_sched_in(struct perf_cpu_context
*cpuctx
,
191 enum event_type_t event_type
,
192 struct task_struct
*task
);
194 static void update_context_time(struct perf_event_context
*ctx
);
195 static u64
perf_event_time(struct perf_event
*event
);
197 static void ring_buffer_attach(struct perf_event
*event
,
198 struct ring_buffer
*rb
);
200 void __weak
perf_event_print_debug(void) { }
202 extern __weak
const char *perf_pmu_name(void)
207 static inline u64
perf_clock(void)
209 return local_clock();
212 static inline struct perf_cpu_context
*
213 __get_cpu_context(struct perf_event_context
*ctx
)
215 return this_cpu_ptr(ctx
->pmu
->pmu_cpu_context
);
218 static void perf_ctx_lock(struct perf_cpu_context
*cpuctx
,
219 struct perf_event_context
*ctx
)
221 raw_spin_lock(&cpuctx
->ctx
.lock
);
223 raw_spin_lock(&ctx
->lock
);
226 static void perf_ctx_unlock(struct perf_cpu_context
*cpuctx
,
227 struct perf_event_context
*ctx
)
230 raw_spin_unlock(&ctx
->lock
);
231 raw_spin_unlock(&cpuctx
->ctx
.lock
);
234 #ifdef CONFIG_CGROUP_PERF
237 * Must ensure cgroup is pinned (css_get) before calling
238 * this function. In other words, we cannot call this function
239 * if there is no cgroup event for the current CPU context.
241 static inline struct perf_cgroup
*
242 perf_cgroup_from_task(struct task_struct
*task
)
244 return container_of(task_subsys_state(task
, perf_subsys_id
),
245 struct perf_cgroup
, css
);
249 perf_cgroup_match(struct perf_event
*event
)
251 struct perf_event_context
*ctx
= event
->ctx
;
252 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
254 return !event
->cgrp
|| event
->cgrp
== cpuctx
->cgrp
;
257 static inline bool perf_tryget_cgroup(struct perf_event
*event
)
259 return css_tryget(&event
->cgrp
->css
);
262 static inline void perf_put_cgroup(struct perf_event
*event
)
264 css_put(&event
->cgrp
->css
);
267 static inline void perf_detach_cgroup(struct perf_event
*event
)
269 perf_put_cgroup(event
);
273 static inline int is_cgroup_event(struct perf_event
*event
)
275 return event
->cgrp
!= NULL
;
278 static inline u64
perf_cgroup_event_time(struct perf_event
*event
)
280 struct perf_cgroup_info
*t
;
282 t
= per_cpu_ptr(event
->cgrp
->info
, event
->cpu
);
286 static inline void __update_cgrp_time(struct perf_cgroup
*cgrp
)
288 struct perf_cgroup_info
*info
;
293 info
= this_cpu_ptr(cgrp
->info
);
295 info
->time
+= now
- info
->timestamp
;
296 info
->timestamp
= now
;
299 static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context
*cpuctx
)
301 struct perf_cgroup
*cgrp_out
= cpuctx
->cgrp
;
303 __update_cgrp_time(cgrp_out
);
306 static inline void update_cgrp_time_from_event(struct perf_event
*event
)
308 struct perf_cgroup
*cgrp
;
311 * ensure we access cgroup data only when needed and
312 * when we know the cgroup is pinned (css_get)
314 if (!is_cgroup_event(event
))
317 cgrp
= perf_cgroup_from_task(current
);
319 * Do not update time when cgroup is not active
321 if (cgrp
== event
->cgrp
)
322 __update_cgrp_time(event
->cgrp
);
326 perf_cgroup_set_timestamp(struct task_struct
*task
,
327 struct perf_event_context
*ctx
)
329 struct perf_cgroup
*cgrp
;
330 struct perf_cgroup_info
*info
;
333 * ctx->lock held by caller
334 * ensure we do not access cgroup data
335 * unless we have the cgroup pinned (css_get)
337 if (!task
|| !ctx
->nr_cgroups
)
340 cgrp
= perf_cgroup_from_task(task
);
341 info
= this_cpu_ptr(cgrp
->info
);
342 info
->timestamp
= ctx
->timestamp
;
345 #define PERF_CGROUP_SWOUT 0x1 /* cgroup switch out every event */
346 #define PERF_CGROUP_SWIN 0x2 /* cgroup switch in events based on task */
349 * reschedule events based on the cgroup constraint of task.
351 * mode SWOUT : schedule out everything
352 * mode SWIN : schedule in based on cgroup for next
354 void perf_cgroup_switch(struct task_struct
*task
, int mode
)
356 struct perf_cpu_context
*cpuctx
;
361 * disable interrupts to avoid geting nr_cgroup
362 * changes via __perf_event_disable(). Also
365 local_irq_save(flags
);
368 * we reschedule only in the presence of cgroup
369 * constrained events.
373 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
374 cpuctx
= this_cpu_ptr(pmu
->pmu_cpu_context
);
377 * perf_cgroup_events says at least one
378 * context on this CPU has cgroup events.
380 * ctx->nr_cgroups reports the number of cgroup
381 * events for a context.
383 if (cpuctx
->ctx
.nr_cgroups
> 0) {
384 perf_ctx_lock(cpuctx
, cpuctx
->task_ctx
);
385 perf_pmu_disable(cpuctx
->ctx
.pmu
);
387 if (mode
& PERF_CGROUP_SWOUT
) {
388 cpu_ctx_sched_out(cpuctx
, EVENT_ALL
);
390 * must not be done before ctxswout due
391 * to event_filter_match() in event_sched_out()
396 if (mode
& PERF_CGROUP_SWIN
) {
397 WARN_ON_ONCE(cpuctx
->cgrp
);
398 /* set cgrp before ctxsw in to
399 * allow event_filter_match() to not
400 * have to pass task around
402 cpuctx
->cgrp
= perf_cgroup_from_task(task
);
403 cpu_ctx_sched_in(cpuctx
, EVENT_ALL
, task
);
405 perf_pmu_enable(cpuctx
->ctx
.pmu
);
406 perf_ctx_unlock(cpuctx
, cpuctx
->task_ctx
);
412 local_irq_restore(flags
);
415 static inline void perf_cgroup_sched_out(struct task_struct
*task
,
416 struct task_struct
*next
)
418 struct perf_cgroup
*cgrp1
;
419 struct perf_cgroup
*cgrp2
= NULL
;
422 * we come here when we know perf_cgroup_events > 0
424 cgrp1
= perf_cgroup_from_task(task
);
427 * next is NULL when called from perf_event_enable_on_exec()
428 * that will systematically cause a cgroup_switch()
431 cgrp2
= perf_cgroup_from_task(next
);
434 * only schedule out current cgroup events if we know
435 * that we are switching to a different cgroup. Otherwise,
436 * do no touch the cgroup events.
439 perf_cgroup_switch(task
, PERF_CGROUP_SWOUT
);
442 static inline void perf_cgroup_sched_in(struct task_struct
*prev
,
443 struct task_struct
*task
)
445 struct perf_cgroup
*cgrp1
;
446 struct perf_cgroup
*cgrp2
= NULL
;
449 * we come here when we know perf_cgroup_events > 0
451 cgrp1
= perf_cgroup_from_task(task
);
453 /* prev can never be NULL */
454 cgrp2
= perf_cgroup_from_task(prev
);
457 * only need to schedule in cgroup events if we are changing
458 * cgroup during ctxsw. Cgroup events were not scheduled
459 * out of ctxsw out if that was not the case.
462 perf_cgroup_switch(task
, PERF_CGROUP_SWIN
);
465 static inline int perf_cgroup_connect(int fd
, struct perf_event
*event
,
466 struct perf_event_attr
*attr
,
467 struct perf_event
*group_leader
)
469 struct perf_cgroup
*cgrp
;
470 struct cgroup_subsys_state
*css
;
471 struct fd f
= fdget(fd
);
477 css
= cgroup_css_from_dir(f
.file
, perf_subsys_id
);
483 cgrp
= container_of(css
, struct perf_cgroup
, css
);
486 /* must be done before we fput() the file */
487 if (!perf_tryget_cgroup(event
)) {
494 * all events in a group must monitor
495 * the same cgroup because a task belongs
496 * to only one perf cgroup at a time
498 if (group_leader
&& group_leader
->cgrp
!= cgrp
) {
499 perf_detach_cgroup(event
);
508 perf_cgroup_set_shadow_time(struct perf_event
*event
, u64 now
)
510 struct perf_cgroup_info
*t
;
511 t
= per_cpu_ptr(event
->cgrp
->info
, event
->cpu
);
512 event
->shadow_ctx_time
= now
- t
->timestamp
;
516 perf_cgroup_defer_enabled(struct perf_event
*event
)
519 * when the current task's perf cgroup does not match
520 * the event's, we need to remember to call the
521 * perf_mark_enable() function the first time a task with
522 * a matching perf cgroup is scheduled in.
524 if (is_cgroup_event(event
) && !perf_cgroup_match(event
))
525 event
->cgrp_defer_enabled
= 1;
529 perf_cgroup_mark_enabled(struct perf_event
*event
,
530 struct perf_event_context
*ctx
)
532 struct perf_event
*sub
;
533 u64 tstamp
= perf_event_time(event
);
535 if (!event
->cgrp_defer_enabled
)
538 event
->cgrp_defer_enabled
= 0;
540 event
->tstamp_enabled
= tstamp
- event
->total_time_enabled
;
541 list_for_each_entry(sub
, &event
->sibling_list
, group_entry
) {
542 if (sub
->state
>= PERF_EVENT_STATE_INACTIVE
) {
543 sub
->tstamp_enabled
= tstamp
- sub
->total_time_enabled
;
544 sub
->cgrp_defer_enabled
= 0;
548 #else /* !CONFIG_CGROUP_PERF */
551 perf_cgroup_match(struct perf_event
*event
)
556 static inline void perf_detach_cgroup(struct perf_event
*event
)
559 static inline int is_cgroup_event(struct perf_event
*event
)
564 static inline u64
perf_cgroup_event_cgrp_time(struct perf_event
*event
)
569 static inline void update_cgrp_time_from_event(struct perf_event
*event
)
573 static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context
*cpuctx
)
577 static inline void perf_cgroup_sched_out(struct task_struct
*task
,
578 struct task_struct
*next
)
582 static inline void perf_cgroup_sched_in(struct task_struct
*prev
,
583 struct task_struct
*task
)
587 static inline int perf_cgroup_connect(pid_t pid
, struct perf_event
*event
,
588 struct perf_event_attr
*attr
,
589 struct perf_event
*group_leader
)
595 perf_cgroup_set_timestamp(struct task_struct
*task
,
596 struct perf_event_context
*ctx
)
601 perf_cgroup_switch(struct task_struct
*task
, struct task_struct
*next
)
606 perf_cgroup_set_shadow_time(struct perf_event
*event
, u64 now
)
610 static inline u64
perf_cgroup_event_time(struct perf_event
*event
)
616 perf_cgroup_defer_enabled(struct perf_event
*event
)
621 perf_cgroup_mark_enabled(struct perf_event
*event
,
622 struct perf_event_context
*ctx
)
627 void perf_pmu_disable(struct pmu
*pmu
)
629 int *count
= this_cpu_ptr(pmu
->pmu_disable_count
);
631 pmu
->pmu_disable(pmu
);
634 void perf_pmu_enable(struct pmu
*pmu
)
636 int *count
= this_cpu_ptr(pmu
->pmu_disable_count
);
638 pmu
->pmu_enable(pmu
);
641 static DEFINE_PER_CPU(struct list_head
, rotation_list
);
644 * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
645 * because they're strictly cpu affine and rotate_start is called with IRQs
646 * disabled, while rotate_context is called from IRQ context.
648 static void perf_pmu_rotate_start(struct pmu
*pmu
)
650 struct perf_cpu_context
*cpuctx
= this_cpu_ptr(pmu
->pmu_cpu_context
);
651 struct list_head
*head
= &__get_cpu_var(rotation_list
);
653 WARN_ON(!irqs_disabled());
655 if (list_empty(&cpuctx
->rotation_list
))
656 list_add(&cpuctx
->rotation_list
, head
);
659 static void get_ctx(struct perf_event_context
*ctx
)
661 WARN_ON(!atomic_inc_not_zero(&ctx
->refcount
));
664 static void put_ctx(struct perf_event_context
*ctx
)
666 if (atomic_dec_and_test(&ctx
->refcount
)) {
668 put_ctx(ctx
->parent_ctx
);
670 put_task_struct(ctx
->task
);
671 kfree_rcu(ctx
, rcu_head
);
675 static void unclone_ctx(struct perf_event_context
*ctx
)
677 if (ctx
->parent_ctx
) {
678 put_ctx(ctx
->parent_ctx
);
679 ctx
->parent_ctx
= NULL
;
683 static u32
perf_event_pid(struct perf_event
*event
, struct task_struct
*p
)
686 * only top level events have the pid namespace they were created in
689 event
= event
->parent
;
691 return task_tgid_nr_ns(p
, event
->ns
);
694 static u32
perf_event_tid(struct perf_event
*event
, struct task_struct
*p
)
697 * only top level events have the pid namespace they were created in
700 event
= event
->parent
;
702 return task_pid_nr_ns(p
, event
->ns
);
706 * If we inherit events we want to return the parent event id
709 static u64
primary_event_id(struct perf_event
*event
)
714 id
= event
->parent
->id
;
720 * Get the perf_event_context for a task and lock it.
721 * This has to cope with with the fact that until it is locked,
722 * the context could get moved to another task.
724 static struct perf_event_context
*
725 perf_lock_task_context(struct task_struct
*task
, int ctxn
, unsigned long *flags
)
727 struct perf_event_context
*ctx
;
731 ctx
= rcu_dereference(task
->perf_event_ctxp
[ctxn
]);
734 * If this context is a clone of another, it might
735 * get swapped for another underneath us by
736 * perf_event_task_sched_out, though the
737 * rcu_read_lock() protects us from any context
738 * getting freed. Lock the context and check if it
739 * got swapped before we could get the lock, and retry
740 * if so. If we locked the right context, then it
741 * can't get swapped on us any more.
743 raw_spin_lock_irqsave(&ctx
->lock
, *flags
);
744 if (ctx
!= rcu_dereference(task
->perf_event_ctxp
[ctxn
])) {
745 raw_spin_unlock_irqrestore(&ctx
->lock
, *flags
);
749 if (!atomic_inc_not_zero(&ctx
->refcount
)) {
750 raw_spin_unlock_irqrestore(&ctx
->lock
, *flags
);
759 * Get the context for a task and increment its pin_count so it
760 * can't get swapped to another task. This also increments its
761 * reference count so that the context can't get freed.
763 static struct perf_event_context
*
764 perf_pin_task_context(struct task_struct
*task
, int ctxn
)
766 struct perf_event_context
*ctx
;
769 ctx
= perf_lock_task_context(task
, ctxn
, &flags
);
772 raw_spin_unlock_irqrestore(&ctx
->lock
, flags
);
777 static void perf_unpin_context(struct perf_event_context
*ctx
)
781 raw_spin_lock_irqsave(&ctx
->lock
, flags
);
783 raw_spin_unlock_irqrestore(&ctx
->lock
, flags
);
787 * Update the record of the current time in a context.
789 static void update_context_time(struct perf_event_context
*ctx
)
791 u64 now
= perf_clock();
793 ctx
->time
+= now
- ctx
->timestamp
;
794 ctx
->timestamp
= now
;
797 static u64
perf_event_time(struct perf_event
*event
)
799 struct perf_event_context
*ctx
= event
->ctx
;
801 if (is_cgroup_event(event
))
802 return perf_cgroup_event_time(event
);
804 return ctx
? ctx
->time
: 0;
808 * Update the total_time_enabled and total_time_running fields for a event.
809 * The caller of this function needs to hold the ctx->lock.
811 static void update_event_times(struct perf_event
*event
)
813 struct perf_event_context
*ctx
= event
->ctx
;
816 if (event
->state
< PERF_EVENT_STATE_INACTIVE
||
817 event
->group_leader
->state
< PERF_EVENT_STATE_INACTIVE
)
820 * in cgroup mode, time_enabled represents
821 * the time the event was enabled AND active
822 * tasks were in the monitored cgroup. This is
823 * independent of the activity of the context as
824 * there may be a mix of cgroup and non-cgroup events.
826 * That is why we treat cgroup events differently
829 if (is_cgroup_event(event
))
830 run_end
= perf_cgroup_event_time(event
);
831 else if (ctx
->is_active
)
834 run_end
= event
->tstamp_stopped
;
836 event
->total_time_enabled
= run_end
- event
->tstamp_enabled
;
838 if (event
->state
== PERF_EVENT_STATE_INACTIVE
)
839 run_end
= event
->tstamp_stopped
;
841 run_end
= perf_event_time(event
);
843 event
->total_time_running
= run_end
- event
->tstamp_running
;
848 * Update total_time_enabled and total_time_running for all events in a group.
850 static void update_group_times(struct perf_event
*leader
)
852 struct perf_event
*event
;
854 update_event_times(leader
);
855 list_for_each_entry(event
, &leader
->sibling_list
, group_entry
)
856 update_event_times(event
);
859 static struct list_head
*
860 ctx_group_list(struct perf_event
*event
, struct perf_event_context
*ctx
)
862 if (event
->attr
.pinned
)
863 return &ctx
->pinned_groups
;
865 return &ctx
->flexible_groups
;
869 * Add a event from the lists for its context.
870 * Must be called with ctx->mutex and ctx->lock held.
873 list_add_event(struct perf_event
*event
, struct perf_event_context
*ctx
)
875 WARN_ON_ONCE(event
->attach_state
& PERF_ATTACH_CONTEXT
);
876 event
->attach_state
|= PERF_ATTACH_CONTEXT
;
879 * If we're a stand alone event or group leader, we go to the context
880 * list, group events are kept attached to the group so that
881 * perf_group_detach can, at all times, locate all siblings.
883 if (event
->group_leader
== event
) {
884 struct list_head
*list
;
886 if (is_software_event(event
))
887 event
->group_flags
|= PERF_GROUP_SOFTWARE
;
889 list
= ctx_group_list(event
, ctx
);
890 list_add_tail(&event
->group_entry
, list
);
893 if (is_cgroup_event(event
))
896 if (has_branch_stack(event
))
897 ctx
->nr_branch_stack
++;
899 list_add_rcu(&event
->event_entry
, &ctx
->event_list
);
901 perf_pmu_rotate_start(ctx
->pmu
);
903 if (event
->attr
.inherit_stat
)
908 * Called at perf_event creation and when events are attached/detached from a
911 static void perf_event__read_size(struct perf_event
*event
)
913 int entry
= sizeof(u64
); /* value */
917 if (event
->attr
.read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
920 if (event
->attr
.read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
923 if (event
->attr
.read_format
& PERF_FORMAT_ID
)
924 entry
+= sizeof(u64
);
926 if (event
->attr
.read_format
& PERF_FORMAT_GROUP
) {
927 nr
+= event
->group_leader
->nr_siblings
;
932 event
->read_size
= size
;
935 static void perf_event__header_size(struct perf_event
*event
)
937 struct perf_sample_data
*data
;
938 u64 sample_type
= event
->attr
.sample_type
;
941 perf_event__read_size(event
);
943 if (sample_type
& PERF_SAMPLE_IP
)
944 size
+= sizeof(data
->ip
);
946 if (sample_type
& PERF_SAMPLE_ADDR
)
947 size
+= sizeof(data
->addr
);
949 if (sample_type
& PERF_SAMPLE_PERIOD
)
950 size
+= sizeof(data
->period
);
952 if (sample_type
& PERF_SAMPLE_READ
)
953 size
+= event
->read_size
;
955 event
->header_size
= size
;
958 static void perf_event__id_header_size(struct perf_event
*event
)
960 struct perf_sample_data
*data
;
961 u64 sample_type
= event
->attr
.sample_type
;
964 if (sample_type
& PERF_SAMPLE_TID
)
965 size
+= sizeof(data
->tid_entry
);
967 if (sample_type
& PERF_SAMPLE_TIME
)
968 size
+= sizeof(data
->time
);
970 if (sample_type
& PERF_SAMPLE_ID
)
971 size
+= sizeof(data
->id
);
973 if (sample_type
& PERF_SAMPLE_STREAM_ID
)
974 size
+= sizeof(data
->stream_id
);
976 if (sample_type
& PERF_SAMPLE_CPU
)
977 size
+= sizeof(data
->cpu_entry
);
979 event
->id_header_size
= size
;
982 static void perf_group_attach(struct perf_event
*event
)
984 struct perf_event
*group_leader
= event
->group_leader
, *pos
;
987 * We can have double attach due to group movement in perf_event_open.
989 if (event
->attach_state
& PERF_ATTACH_GROUP
)
992 event
->attach_state
|= PERF_ATTACH_GROUP
;
994 if (group_leader
== event
)
997 if (group_leader
->group_flags
& PERF_GROUP_SOFTWARE
&&
998 !is_software_event(event
))
999 group_leader
->group_flags
&= ~PERF_GROUP_SOFTWARE
;
1001 list_add_tail(&event
->group_entry
, &group_leader
->sibling_list
);
1002 group_leader
->nr_siblings
++;
1004 perf_event__header_size(group_leader
);
1006 list_for_each_entry(pos
, &group_leader
->sibling_list
, group_entry
)
1007 perf_event__header_size(pos
);
1011 * Remove a event from the lists for its context.
1012 * Must be called with ctx->mutex and ctx->lock held.
1015 list_del_event(struct perf_event
*event
, struct perf_event_context
*ctx
)
1017 struct perf_cpu_context
*cpuctx
;
1019 * We can have double detach due to exit/hot-unplug + close.
1021 if (!(event
->attach_state
& PERF_ATTACH_CONTEXT
))
1024 event
->attach_state
&= ~PERF_ATTACH_CONTEXT
;
1026 if (is_cgroup_event(event
)) {
1028 cpuctx
= __get_cpu_context(ctx
);
1030 * if there are no more cgroup events
1031 * then cler cgrp to avoid stale pointer
1032 * in update_cgrp_time_from_cpuctx()
1034 if (!ctx
->nr_cgroups
)
1035 cpuctx
->cgrp
= NULL
;
1038 if (has_branch_stack(event
))
1039 ctx
->nr_branch_stack
--;
1042 if (event
->attr
.inherit_stat
)
1045 list_del_rcu(&event
->event_entry
);
1047 if (event
->group_leader
== event
)
1048 list_del_init(&event
->group_entry
);
1050 update_group_times(event
);
1053 * If event was in error state, then keep it
1054 * that way, otherwise bogus counts will be
1055 * returned on read(). The only way to get out
1056 * of error state is by explicit re-enabling
1059 if (event
->state
> PERF_EVENT_STATE_OFF
)
1060 event
->state
= PERF_EVENT_STATE_OFF
;
1063 static void perf_group_detach(struct perf_event
*event
)
1065 struct perf_event
*sibling
, *tmp
;
1066 struct list_head
*list
= NULL
;
1069 * We can have double detach due to exit/hot-unplug + close.
1071 if (!(event
->attach_state
& PERF_ATTACH_GROUP
))
1074 event
->attach_state
&= ~PERF_ATTACH_GROUP
;
1077 * If this is a sibling, remove it from its group.
1079 if (event
->group_leader
!= event
) {
1080 list_del_init(&event
->group_entry
);
1081 event
->group_leader
->nr_siblings
--;
1085 if (!list_empty(&event
->group_entry
))
1086 list
= &event
->group_entry
;
1089 * If this was a group event with sibling events then
1090 * upgrade the siblings to singleton events by adding them
1091 * to whatever list we are on.
1093 list_for_each_entry_safe(sibling
, tmp
, &event
->sibling_list
, group_entry
) {
1095 list_move_tail(&sibling
->group_entry
, list
);
1096 sibling
->group_leader
= sibling
;
1098 /* Inherit group flags from the previous leader */
1099 sibling
->group_flags
= event
->group_flags
;
1103 perf_event__header_size(event
->group_leader
);
1105 list_for_each_entry(tmp
, &event
->group_leader
->sibling_list
, group_entry
)
1106 perf_event__header_size(tmp
);
1110 event_filter_match(struct perf_event
*event
)
1112 return (event
->cpu
== -1 || event
->cpu
== smp_processor_id())
1113 && perf_cgroup_match(event
);
1117 event_sched_out(struct perf_event
*event
,
1118 struct perf_cpu_context
*cpuctx
,
1119 struct perf_event_context
*ctx
)
1121 u64 tstamp
= perf_event_time(event
);
1124 * An event which could not be activated because of
1125 * filter mismatch still needs to have its timings
1126 * maintained, otherwise bogus information is return
1127 * via read() for time_enabled, time_running:
1129 if (event
->state
== PERF_EVENT_STATE_INACTIVE
1130 && !event_filter_match(event
)) {
1131 delta
= tstamp
- event
->tstamp_stopped
;
1132 event
->tstamp_running
+= delta
;
1133 event
->tstamp_stopped
= tstamp
;
1136 if (event
->state
!= PERF_EVENT_STATE_ACTIVE
)
1139 event
->state
= PERF_EVENT_STATE_INACTIVE
;
1140 if (event
->pending_disable
) {
1141 event
->pending_disable
= 0;
1142 event
->state
= PERF_EVENT_STATE_OFF
;
1144 event
->tstamp_stopped
= tstamp
;
1145 event
->pmu
->del(event
, 0);
1148 if (!is_software_event(event
))
1149 cpuctx
->active_oncpu
--;
1151 if (event
->attr
.freq
&& event
->attr
.sample_freq
)
1153 if (event
->attr
.exclusive
|| !cpuctx
->active_oncpu
)
1154 cpuctx
->exclusive
= 0;
1158 group_sched_out(struct perf_event
*group_event
,
1159 struct perf_cpu_context
*cpuctx
,
1160 struct perf_event_context
*ctx
)
1162 struct perf_event
*event
;
1163 int state
= group_event
->state
;
1165 event_sched_out(group_event
, cpuctx
, ctx
);
1168 * Schedule out siblings (if any):
1170 list_for_each_entry(event
, &group_event
->sibling_list
, group_entry
)
1171 event_sched_out(event
, cpuctx
, ctx
);
1173 if (state
== PERF_EVENT_STATE_ACTIVE
&& group_event
->attr
.exclusive
)
1174 cpuctx
->exclusive
= 0;
1178 * Cross CPU call to remove a performance event
1180 * We disable the event on the hardware level first. After that we
1181 * remove it from the context list.
1183 static int __perf_remove_from_context(void *info
)
1185 struct perf_event
*event
= info
;
1186 struct perf_event_context
*ctx
= event
->ctx
;
1187 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
1189 raw_spin_lock(&ctx
->lock
);
1190 event_sched_out(event
, cpuctx
, ctx
);
1191 list_del_event(event
, ctx
);
1192 if (!ctx
->nr_events
&& cpuctx
->task_ctx
== ctx
) {
1194 cpuctx
->task_ctx
= NULL
;
1196 raw_spin_unlock(&ctx
->lock
);
1203 * Remove the event from a task's (or a CPU's) list of events.
1205 * CPU events are removed with a smp call. For task events we only
1206 * call when the task is on a CPU.
1208 * If event->ctx is a cloned context, callers must make sure that
1209 * every task struct that event->ctx->task could possibly point to
1210 * remains valid. This is OK when called from perf_release since
1211 * that only calls us on the top-level context, which can't be a clone.
1212 * When called from perf_event_exit_task, it's OK because the
1213 * context has been detached from its task.
1215 static void perf_remove_from_context(struct perf_event
*event
)
1217 struct perf_event_context
*ctx
= event
->ctx
;
1218 struct task_struct
*task
= ctx
->task
;
1220 lockdep_assert_held(&ctx
->mutex
);
1224 * Per cpu events are removed via an smp call and
1225 * the removal is always successful.
1227 cpu_function_call(event
->cpu
, __perf_remove_from_context
, event
);
1232 if (!task_function_call(task
, __perf_remove_from_context
, event
))
1235 raw_spin_lock_irq(&ctx
->lock
);
1237 * If we failed to find a running task, but find the context active now
1238 * that we've acquired the ctx->lock, retry.
1240 if (ctx
->is_active
) {
1241 raw_spin_unlock_irq(&ctx
->lock
);
1246 * Since the task isn't running, its safe to remove the event, us
1247 * holding the ctx->lock ensures the task won't get scheduled in.
1249 list_del_event(event
, ctx
);
1250 raw_spin_unlock_irq(&ctx
->lock
);
1254 * Cross CPU call to disable a performance event
1256 int __perf_event_disable(void *info
)
1258 struct perf_event
*event
= info
;
1259 struct perf_event_context
*ctx
= event
->ctx
;
1260 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
1263 * If this is a per-task event, need to check whether this
1264 * event's task is the current task on this cpu.
1266 * Can trigger due to concurrent perf_event_context_sched_out()
1267 * flipping contexts around.
1269 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
)
1272 raw_spin_lock(&ctx
->lock
);
1275 * If the event is on, turn it off.
1276 * If it is in error state, leave it in error state.
1278 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
) {
1279 update_context_time(ctx
);
1280 update_cgrp_time_from_event(event
);
1281 update_group_times(event
);
1282 if (event
== event
->group_leader
)
1283 group_sched_out(event
, cpuctx
, ctx
);
1285 event_sched_out(event
, cpuctx
, ctx
);
1286 event
->state
= PERF_EVENT_STATE_OFF
;
1289 raw_spin_unlock(&ctx
->lock
);
1297 * If event->ctx is a cloned context, callers must make sure that
1298 * every task struct that event->ctx->task could possibly point to
1299 * remains valid. This condition is satisifed when called through
1300 * perf_event_for_each_child or perf_event_for_each because they
1301 * hold the top-level event's child_mutex, so any descendant that
1302 * goes to exit will block in sync_child_event.
1303 * When called from perf_pending_event it's OK because event->ctx
1304 * is the current context on this CPU and preemption is disabled,
1305 * hence we can't get into perf_event_task_sched_out for this context.
1307 void perf_event_disable(struct perf_event
*event
)
1309 struct perf_event_context
*ctx
= event
->ctx
;
1310 struct task_struct
*task
= ctx
->task
;
1314 * Disable the event on the cpu that it's on
1316 cpu_function_call(event
->cpu
, __perf_event_disable
, event
);
1321 if (!task_function_call(task
, __perf_event_disable
, event
))
1324 raw_spin_lock_irq(&ctx
->lock
);
1326 * If the event is still active, we need to retry the cross-call.
1328 if (event
->state
== PERF_EVENT_STATE_ACTIVE
) {
1329 raw_spin_unlock_irq(&ctx
->lock
);
1331 * Reload the task pointer, it might have been changed by
1332 * a concurrent perf_event_context_sched_out().
1339 * Since we have the lock this context can't be scheduled
1340 * in, so we can change the state safely.
1342 if (event
->state
== PERF_EVENT_STATE_INACTIVE
) {
1343 update_group_times(event
);
1344 event
->state
= PERF_EVENT_STATE_OFF
;
1346 raw_spin_unlock_irq(&ctx
->lock
);
1348 EXPORT_SYMBOL_GPL(perf_event_disable
);
1350 static void perf_set_shadow_time(struct perf_event
*event
,
1351 struct perf_event_context
*ctx
,
1355 * use the correct time source for the time snapshot
1357 * We could get by without this by leveraging the
1358 * fact that to get to this function, the caller
1359 * has most likely already called update_context_time()
1360 * and update_cgrp_time_xx() and thus both timestamp
1361 * are identical (or very close). Given that tstamp is,
1362 * already adjusted for cgroup, we could say that:
1363 * tstamp - ctx->timestamp
1365 * tstamp - cgrp->timestamp.
1367 * Then, in perf_output_read(), the calculation would
1368 * work with no changes because:
1369 * - event is guaranteed scheduled in
1370 * - no scheduled out in between
1371 * - thus the timestamp would be the same
1373 * But this is a bit hairy.
1375 * So instead, we have an explicit cgroup call to remain
1376 * within the time time source all along. We believe it
1377 * is cleaner and simpler to understand.
1379 if (is_cgroup_event(event
))
1380 perf_cgroup_set_shadow_time(event
, tstamp
);
1382 event
->shadow_ctx_time
= tstamp
- ctx
->timestamp
;
1385 #define MAX_INTERRUPTS (~0ULL)
1387 static void perf_log_throttle(struct perf_event
*event
, int enable
);
1390 event_sched_in(struct perf_event
*event
,
1391 struct perf_cpu_context
*cpuctx
,
1392 struct perf_event_context
*ctx
)
1394 u64 tstamp
= perf_event_time(event
);
1396 if (event
->state
<= PERF_EVENT_STATE_OFF
)
1399 event
->state
= PERF_EVENT_STATE_ACTIVE
;
1400 event
->oncpu
= smp_processor_id();
1403 * Unthrottle events, since we scheduled we might have missed several
1404 * ticks already, also for a heavily scheduling task there is little
1405 * guarantee it'll get a tick in a timely manner.
1407 if (unlikely(event
->hw
.interrupts
== MAX_INTERRUPTS
)) {
1408 perf_log_throttle(event
, 1);
1409 event
->hw
.interrupts
= 0;
1413 * The new state must be visible before we turn it on in the hardware:
1417 if (event
->pmu
->add(event
, PERF_EF_START
)) {
1418 event
->state
= PERF_EVENT_STATE_INACTIVE
;
1423 event
->tstamp_running
+= tstamp
- event
->tstamp_stopped
;
1425 perf_set_shadow_time(event
, ctx
, tstamp
);
1427 if (!is_software_event(event
))
1428 cpuctx
->active_oncpu
++;
1430 if (event
->attr
.freq
&& event
->attr
.sample_freq
)
1433 if (event
->attr
.exclusive
)
1434 cpuctx
->exclusive
= 1;
1440 group_sched_in(struct perf_event
*group_event
,
1441 struct perf_cpu_context
*cpuctx
,
1442 struct perf_event_context
*ctx
)
1444 struct perf_event
*event
, *partial_group
= NULL
;
1445 struct pmu
*pmu
= group_event
->pmu
;
1446 u64 now
= ctx
->time
;
1447 bool simulate
= false;
1449 if (group_event
->state
== PERF_EVENT_STATE_OFF
)
1452 pmu
->start_txn(pmu
);
1454 if (event_sched_in(group_event
, cpuctx
, ctx
)) {
1455 pmu
->cancel_txn(pmu
);
1460 * Schedule in siblings as one group (if any):
1462 list_for_each_entry(event
, &group_event
->sibling_list
, group_entry
) {
1463 if (event_sched_in(event
, cpuctx
, ctx
)) {
1464 partial_group
= event
;
1469 if (!pmu
->commit_txn(pmu
))
1474 * Groups can be scheduled in as one unit only, so undo any
1475 * partial group before returning:
1476 * The events up to the failed event are scheduled out normally,
1477 * tstamp_stopped will be updated.
1479 * The failed events and the remaining siblings need to have
1480 * their timings updated as if they had gone thru event_sched_in()
1481 * and event_sched_out(). This is required to get consistent timings
1482 * across the group. This also takes care of the case where the group
1483 * could never be scheduled by ensuring tstamp_stopped is set to mark
1484 * the time the event was actually stopped, such that time delta
1485 * calculation in update_event_times() is correct.
1487 list_for_each_entry(event
, &group_event
->sibling_list
, group_entry
) {
1488 if (event
== partial_group
)
1492 event
->tstamp_running
+= now
- event
->tstamp_stopped
;
1493 event
->tstamp_stopped
= now
;
1495 event_sched_out(event
, cpuctx
, ctx
);
1498 event_sched_out(group_event
, cpuctx
, ctx
);
1500 pmu
->cancel_txn(pmu
);
1506 * Work out whether we can put this event group on the CPU now.
1508 static int group_can_go_on(struct perf_event
*event
,
1509 struct perf_cpu_context
*cpuctx
,
1513 * Groups consisting entirely of software events can always go on.
1515 if (event
->group_flags
& PERF_GROUP_SOFTWARE
)
1518 * If an exclusive group is already on, no other hardware
1521 if (cpuctx
->exclusive
)
1524 * If this group is exclusive and there are already
1525 * events on the CPU, it can't go on.
1527 if (event
->attr
.exclusive
&& cpuctx
->active_oncpu
)
1530 * Otherwise, try to add it if all previous groups were able
1536 static void add_event_to_ctx(struct perf_event
*event
,
1537 struct perf_event_context
*ctx
)
1539 u64 tstamp
= perf_event_time(event
);
1541 list_add_event(event
, ctx
);
1542 perf_group_attach(event
);
1543 event
->tstamp_enabled
= tstamp
;
1544 event
->tstamp_running
= tstamp
;
1545 event
->tstamp_stopped
= tstamp
;
1548 static void task_ctx_sched_out(struct perf_event_context
*ctx
);
1550 ctx_sched_in(struct perf_event_context
*ctx
,
1551 struct perf_cpu_context
*cpuctx
,
1552 enum event_type_t event_type
,
1553 struct task_struct
*task
);
1555 static void perf_event_sched_in(struct perf_cpu_context
*cpuctx
,
1556 struct perf_event_context
*ctx
,
1557 struct task_struct
*task
)
1559 cpu_ctx_sched_in(cpuctx
, EVENT_PINNED
, task
);
1561 ctx_sched_in(ctx
, cpuctx
, EVENT_PINNED
, task
);
1562 cpu_ctx_sched_in(cpuctx
, EVENT_FLEXIBLE
, task
);
1564 ctx_sched_in(ctx
, cpuctx
, EVENT_FLEXIBLE
, task
);
1568 * Cross CPU call to install and enable a performance event
1570 * Must be called with ctx->mutex held
1572 static int __perf_install_in_context(void *info
)
1574 struct perf_event
*event
= info
;
1575 struct perf_event_context
*ctx
= event
->ctx
;
1576 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
1577 struct perf_event_context
*task_ctx
= cpuctx
->task_ctx
;
1578 struct task_struct
*task
= current
;
1580 perf_ctx_lock(cpuctx
, task_ctx
);
1581 perf_pmu_disable(cpuctx
->ctx
.pmu
);
1584 * If there was an active task_ctx schedule it out.
1587 task_ctx_sched_out(task_ctx
);
1590 * If the context we're installing events in is not the
1591 * active task_ctx, flip them.
1593 if (ctx
->task
&& task_ctx
!= ctx
) {
1595 raw_spin_unlock(&task_ctx
->lock
);
1596 raw_spin_lock(&ctx
->lock
);
1601 cpuctx
->task_ctx
= task_ctx
;
1602 task
= task_ctx
->task
;
1605 cpu_ctx_sched_out(cpuctx
, EVENT_ALL
);
1607 update_context_time(ctx
);
1609 * update cgrp time only if current cgrp
1610 * matches event->cgrp. Must be done before
1611 * calling add_event_to_ctx()
1613 update_cgrp_time_from_event(event
);
1615 add_event_to_ctx(event
, ctx
);
1618 * Schedule everything back in
1620 perf_event_sched_in(cpuctx
, task_ctx
, task
);
1622 perf_pmu_enable(cpuctx
->ctx
.pmu
);
1623 perf_ctx_unlock(cpuctx
, task_ctx
);
1629 * Attach a performance event to a context
1631 * First we add the event to the list with the hardware enable bit
1632 * in event->hw_config cleared.
1634 * If the event is attached to a task which is on a CPU we use a smp
1635 * call to enable it in the task context. The task might have been
1636 * scheduled away, but we check this in the smp call again.
1639 perf_install_in_context(struct perf_event_context
*ctx
,
1640 struct perf_event
*event
,
1643 struct task_struct
*task
= ctx
->task
;
1645 lockdep_assert_held(&ctx
->mutex
);
1648 if (event
->cpu
!= -1)
1653 * Per cpu events are installed via an smp call and
1654 * the install is always successful.
1656 cpu_function_call(cpu
, __perf_install_in_context
, event
);
1661 if (!task_function_call(task
, __perf_install_in_context
, event
))
1664 raw_spin_lock_irq(&ctx
->lock
);
1666 * If we failed to find a running task, but find the context active now
1667 * that we've acquired the ctx->lock, retry.
1669 if (ctx
->is_active
) {
1670 raw_spin_unlock_irq(&ctx
->lock
);
1675 * Since the task isn't running, its safe to add the event, us holding
1676 * the ctx->lock ensures the task won't get scheduled in.
1678 add_event_to_ctx(event
, ctx
);
1679 raw_spin_unlock_irq(&ctx
->lock
);
1683 * Put a event into inactive state and update time fields.
1684 * Enabling the leader of a group effectively enables all
1685 * the group members that aren't explicitly disabled, so we
1686 * have to update their ->tstamp_enabled also.
1687 * Note: this works for group members as well as group leaders
1688 * since the non-leader members' sibling_lists will be empty.
1690 static void __perf_event_mark_enabled(struct perf_event
*event
)
1692 struct perf_event
*sub
;
1693 u64 tstamp
= perf_event_time(event
);
1695 event
->state
= PERF_EVENT_STATE_INACTIVE
;
1696 event
->tstamp_enabled
= tstamp
- event
->total_time_enabled
;
1697 list_for_each_entry(sub
, &event
->sibling_list
, group_entry
) {
1698 if (sub
->state
>= PERF_EVENT_STATE_INACTIVE
)
1699 sub
->tstamp_enabled
= tstamp
- sub
->total_time_enabled
;
1704 * Cross CPU call to enable a performance event
1706 static int __perf_event_enable(void *info
)
1708 struct perf_event
*event
= info
;
1709 struct perf_event_context
*ctx
= event
->ctx
;
1710 struct perf_event
*leader
= event
->group_leader
;
1711 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
1714 if (WARN_ON_ONCE(!ctx
->is_active
))
1717 raw_spin_lock(&ctx
->lock
);
1718 update_context_time(ctx
);
1720 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
)
1724 * set current task's cgroup time reference point
1726 perf_cgroup_set_timestamp(current
, ctx
);
1728 __perf_event_mark_enabled(event
);
1730 if (!event_filter_match(event
)) {
1731 if (is_cgroup_event(event
))
1732 perf_cgroup_defer_enabled(event
);
1737 * If the event is in a group and isn't the group leader,
1738 * then don't put it on unless the group is on.
1740 if (leader
!= event
&& leader
->state
!= PERF_EVENT_STATE_ACTIVE
)
1743 if (!group_can_go_on(event
, cpuctx
, 1)) {
1746 if (event
== leader
)
1747 err
= group_sched_in(event
, cpuctx
, ctx
);
1749 err
= event_sched_in(event
, cpuctx
, ctx
);
1754 * If this event can't go on and it's part of a
1755 * group, then the whole group has to come off.
1757 if (leader
!= event
)
1758 group_sched_out(leader
, cpuctx
, ctx
);
1759 if (leader
->attr
.pinned
) {
1760 update_group_times(leader
);
1761 leader
->state
= PERF_EVENT_STATE_ERROR
;
1766 raw_spin_unlock(&ctx
->lock
);
1774 * If event->ctx is a cloned context, callers must make sure that
1775 * every task struct that event->ctx->task could possibly point to
1776 * remains valid. This condition is satisfied when called through
1777 * perf_event_for_each_child or perf_event_for_each as described
1778 * for perf_event_disable.
1780 void perf_event_enable(struct perf_event
*event
)
1782 struct perf_event_context
*ctx
= event
->ctx
;
1783 struct task_struct
*task
= ctx
->task
;
1787 * Enable the event on the cpu that it's on
1789 cpu_function_call(event
->cpu
, __perf_event_enable
, event
);
1793 raw_spin_lock_irq(&ctx
->lock
);
1794 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
)
1798 * If the event is in error state, clear that first.
1799 * That way, if we see the event in error state below, we
1800 * know that it has gone back into error state, as distinct
1801 * from the task having been scheduled away before the
1802 * cross-call arrived.
1804 if (event
->state
== PERF_EVENT_STATE_ERROR
)
1805 event
->state
= PERF_EVENT_STATE_OFF
;
1808 if (!ctx
->is_active
) {
1809 __perf_event_mark_enabled(event
);
1813 raw_spin_unlock_irq(&ctx
->lock
);
1815 if (!task_function_call(task
, __perf_event_enable
, event
))
1818 raw_spin_lock_irq(&ctx
->lock
);
1821 * If the context is active and the event is still off,
1822 * we need to retry the cross-call.
1824 if (ctx
->is_active
&& event
->state
== PERF_EVENT_STATE_OFF
) {
1826 * task could have been flipped by a concurrent
1827 * perf_event_context_sched_out()
1834 raw_spin_unlock_irq(&ctx
->lock
);
1836 EXPORT_SYMBOL_GPL(perf_event_enable
);
1838 int perf_event_refresh(struct perf_event
*event
, int refresh
)
1841 * not supported on inherited events
1843 if (event
->attr
.inherit
|| !is_sampling_event(event
))
1846 atomic_add(refresh
, &event
->event_limit
);
1847 perf_event_enable(event
);
1851 EXPORT_SYMBOL_GPL(perf_event_refresh
);
1853 static void ctx_sched_out(struct perf_event_context
*ctx
,
1854 struct perf_cpu_context
*cpuctx
,
1855 enum event_type_t event_type
)
1857 struct perf_event
*event
;
1858 int is_active
= ctx
->is_active
;
1860 ctx
->is_active
&= ~event_type
;
1861 if (likely(!ctx
->nr_events
))
1864 update_context_time(ctx
);
1865 update_cgrp_time_from_cpuctx(cpuctx
);
1866 if (!ctx
->nr_active
)
1869 perf_pmu_disable(ctx
->pmu
);
1870 if ((is_active
& EVENT_PINNED
) && (event_type
& EVENT_PINNED
)) {
1871 list_for_each_entry(event
, &ctx
->pinned_groups
, group_entry
)
1872 group_sched_out(event
, cpuctx
, ctx
);
1875 if ((is_active
& EVENT_FLEXIBLE
) && (event_type
& EVENT_FLEXIBLE
)) {
1876 list_for_each_entry(event
, &ctx
->flexible_groups
, group_entry
)
1877 group_sched_out(event
, cpuctx
, ctx
);
1879 perf_pmu_enable(ctx
->pmu
);
1883 * Test whether two contexts are equivalent, i.e. whether they
1884 * have both been cloned from the same version of the same context
1885 * and they both have the same number of enabled events.
1886 * If the number of enabled events is the same, then the set
1887 * of enabled events should be the same, because these are both
1888 * inherited contexts, therefore we can't access individual events
1889 * in them directly with an fd; we can only enable/disable all
1890 * events via prctl, or enable/disable all events in a family
1891 * via ioctl, which will have the same effect on both contexts.
1893 static int context_equiv(struct perf_event_context
*ctx1
,
1894 struct perf_event_context
*ctx2
)
1896 return ctx1
->parent_ctx
&& ctx1
->parent_ctx
== ctx2
->parent_ctx
1897 && ctx1
->parent_gen
== ctx2
->parent_gen
1898 && !ctx1
->pin_count
&& !ctx2
->pin_count
;
1901 static void __perf_event_sync_stat(struct perf_event
*event
,
1902 struct perf_event
*next_event
)
1906 if (!event
->attr
.inherit_stat
)
1910 * Update the event value, we cannot use perf_event_read()
1911 * because we're in the middle of a context switch and have IRQs
1912 * disabled, which upsets smp_call_function_single(), however
1913 * we know the event must be on the current CPU, therefore we
1914 * don't need to use it.
1916 switch (event
->state
) {
1917 case PERF_EVENT_STATE_ACTIVE
:
1918 event
->pmu
->read(event
);
1921 case PERF_EVENT_STATE_INACTIVE
:
1922 update_event_times(event
);
1930 * In order to keep per-task stats reliable we need to flip the event
1931 * values when we flip the contexts.
1933 value
= local64_read(&next_event
->count
);
1934 value
= local64_xchg(&event
->count
, value
);
1935 local64_set(&next_event
->count
, value
);
1937 swap(event
->total_time_enabled
, next_event
->total_time_enabled
);
1938 swap(event
->total_time_running
, next_event
->total_time_running
);
1941 * Since we swizzled the values, update the user visible data too.
1943 perf_event_update_userpage(event
);
1944 perf_event_update_userpage(next_event
);
1947 #define list_next_entry(pos, member) \
1948 list_entry(pos->member.next, typeof(*pos), member)
1950 static void perf_event_sync_stat(struct perf_event_context
*ctx
,
1951 struct perf_event_context
*next_ctx
)
1953 struct perf_event
*event
, *next_event
;
1958 update_context_time(ctx
);
1960 event
= list_first_entry(&ctx
->event_list
,
1961 struct perf_event
, event_entry
);
1963 next_event
= list_first_entry(&next_ctx
->event_list
,
1964 struct perf_event
, event_entry
);
1966 while (&event
->event_entry
!= &ctx
->event_list
&&
1967 &next_event
->event_entry
!= &next_ctx
->event_list
) {
1969 __perf_event_sync_stat(event
, next_event
);
1971 event
= list_next_entry(event
, event_entry
);
1972 next_event
= list_next_entry(next_event
, event_entry
);
1976 static void perf_event_context_sched_out(struct task_struct
*task
, int ctxn
,
1977 struct task_struct
*next
)
1979 struct perf_event_context
*ctx
= task
->perf_event_ctxp
[ctxn
];
1980 struct perf_event_context
*next_ctx
;
1981 struct perf_event_context
*parent
;
1982 struct perf_cpu_context
*cpuctx
;
1988 cpuctx
= __get_cpu_context(ctx
);
1989 if (!cpuctx
->task_ctx
)
1993 parent
= rcu_dereference(ctx
->parent_ctx
);
1994 next_ctx
= next
->perf_event_ctxp
[ctxn
];
1995 if (parent
&& next_ctx
&&
1996 rcu_dereference(next_ctx
->parent_ctx
) == parent
) {
1998 * Looks like the two contexts are clones, so we might be
1999 * able to optimize the context switch. We lock both
2000 * contexts and check that they are clones under the
2001 * lock (including re-checking that neither has been
2002 * uncloned in the meantime). It doesn't matter which
2003 * order we take the locks because no other cpu could
2004 * be trying to lock both of these tasks.
2006 raw_spin_lock(&ctx
->lock
);
2007 raw_spin_lock_nested(&next_ctx
->lock
, SINGLE_DEPTH_NESTING
);
2008 if (context_equiv(ctx
, next_ctx
)) {
2010 * XXX do we need a memory barrier of sorts
2011 * wrt to rcu_dereference() of perf_event_ctxp
2013 task
->perf_event_ctxp
[ctxn
] = next_ctx
;
2014 next
->perf_event_ctxp
[ctxn
] = ctx
;
2016 next_ctx
->task
= task
;
2019 perf_event_sync_stat(ctx
, next_ctx
);
2021 raw_spin_unlock(&next_ctx
->lock
);
2022 raw_spin_unlock(&ctx
->lock
);
2027 raw_spin_lock(&ctx
->lock
);
2028 ctx_sched_out(ctx
, cpuctx
, EVENT_ALL
);
2029 cpuctx
->task_ctx
= NULL
;
2030 raw_spin_unlock(&ctx
->lock
);
2034 #define for_each_task_context_nr(ctxn) \
2035 for ((ctxn) = 0; (ctxn) < perf_nr_task_contexts; (ctxn)++)
2038 * Called from scheduler to remove the events of the current task,
2039 * with interrupts disabled.
2041 * We stop each event and update the event value in event->count.
2043 * This does not protect us against NMI, but disable()
2044 * sets the disabled bit in the control field of event _before_
2045 * accessing the event control register. If a NMI hits, then it will
2046 * not restart the event.
2048 void __perf_event_task_sched_out(struct task_struct
*task
,
2049 struct task_struct
*next
)
2053 for_each_task_context_nr(ctxn
)
2054 perf_event_context_sched_out(task
, ctxn
, next
);
2057 * if cgroup events exist on this CPU, then we need
2058 * to check if we have to switch out PMU state.
2059 * cgroup event are system-wide mode only
2061 if (atomic_read(&__get_cpu_var(perf_cgroup_events
)))
2062 perf_cgroup_sched_out(task
, next
);
2065 static void task_ctx_sched_out(struct perf_event_context
*ctx
)
2067 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
2069 if (!cpuctx
->task_ctx
)
2072 if (WARN_ON_ONCE(ctx
!= cpuctx
->task_ctx
))
2075 ctx_sched_out(ctx
, cpuctx
, EVENT_ALL
);
2076 cpuctx
->task_ctx
= NULL
;
2080 * Called with IRQs disabled
2082 static void cpu_ctx_sched_out(struct perf_cpu_context
*cpuctx
,
2083 enum event_type_t event_type
)
2085 ctx_sched_out(&cpuctx
->ctx
, cpuctx
, event_type
);
2089 ctx_pinned_sched_in(struct perf_event_context
*ctx
,
2090 struct perf_cpu_context
*cpuctx
)
2092 struct perf_event
*event
;
2094 list_for_each_entry(event
, &ctx
->pinned_groups
, group_entry
) {
2095 if (event
->state
<= PERF_EVENT_STATE_OFF
)
2097 if (!event_filter_match(event
))
2100 /* may need to reset tstamp_enabled */
2101 if (is_cgroup_event(event
))
2102 perf_cgroup_mark_enabled(event
, ctx
);
2104 if (group_can_go_on(event
, cpuctx
, 1))
2105 group_sched_in(event
, cpuctx
, ctx
);
2108 * If this pinned group hasn't been scheduled,
2109 * put it in error state.
2111 if (event
->state
== PERF_EVENT_STATE_INACTIVE
) {
2112 update_group_times(event
);
2113 event
->state
= PERF_EVENT_STATE_ERROR
;
2119 ctx_flexible_sched_in(struct perf_event_context
*ctx
,
2120 struct perf_cpu_context
*cpuctx
)
2122 struct perf_event
*event
;
2125 list_for_each_entry(event
, &ctx
->flexible_groups
, group_entry
) {
2126 /* Ignore events in OFF or ERROR state */
2127 if (event
->state
<= PERF_EVENT_STATE_OFF
)
2130 * Listen to the 'cpu' scheduling filter constraint
2133 if (!event_filter_match(event
))
2136 /* may need to reset tstamp_enabled */
2137 if (is_cgroup_event(event
))
2138 perf_cgroup_mark_enabled(event
, ctx
);
2140 if (group_can_go_on(event
, cpuctx
, can_add_hw
)) {
2141 if (group_sched_in(event
, cpuctx
, ctx
))
2148 ctx_sched_in(struct perf_event_context
*ctx
,
2149 struct perf_cpu_context
*cpuctx
,
2150 enum event_type_t event_type
,
2151 struct task_struct
*task
)
2154 int is_active
= ctx
->is_active
;
2156 ctx
->is_active
|= event_type
;
2157 if (likely(!ctx
->nr_events
))
2161 ctx
->timestamp
= now
;
2162 perf_cgroup_set_timestamp(task
, ctx
);
2164 * First go through the list and put on any pinned groups
2165 * in order to give them the best chance of going on.
2167 if (!(is_active
& EVENT_PINNED
) && (event_type
& EVENT_PINNED
))
2168 ctx_pinned_sched_in(ctx
, cpuctx
);
2170 /* Then walk through the lower prio flexible groups */
2171 if (!(is_active
& EVENT_FLEXIBLE
) && (event_type
& EVENT_FLEXIBLE
))
2172 ctx_flexible_sched_in(ctx
, cpuctx
);
2175 static void cpu_ctx_sched_in(struct perf_cpu_context
*cpuctx
,
2176 enum event_type_t event_type
,
2177 struct task_struct
*task
)
2179 struct perf_event_context
*ctx
= &cpuctx
->ctx
;
2181 ctx_sched_in(ctx
, cpuctx
, event_type
, task
);
2184 static void perf_event_context_sched_in(struct perf_event_context
*ctx
,
2185 struct task_struct
*task
)
2187 struct perf_cpu_context
*cpuctx
;
2189 cpuctx
= __get_cpu_context(ctx
);
2190 if (cpuctx
->task_ctx
== ctx
)
2193 perf_ctx_lock(cpuctx
, ctx
);
2194 perf_pmu_disable(ctx
->pmu
);
2196 * We want to keep the following priority order:
2197 * cpu pinned (that don't need to move), task pinned,
2198 * cpu flexible, task flexible.
2200 cpu_ctx_sched_out(cpuctx
, EVENT_FLEXIBLE
);
2203 cpuctx
->task_ctx
= ctx
;
2205 perf_event_sched_in(cpuctx
, cpuctx
->task_ctx
, task
);
2207 perf_pmu_enable(ctx
->pmu
);
2208 perf_ctx_unlock(cpuctx
, ctx
);
2211 * Since these rotations are per-cpu, we need to ensure the
2212 * cpu-context we got scheduled on is actually rotating.
2214 perf_pmu_rotate_start(ctx
->pmu
);
2218 * When sampling the branck stack in system-wide, it may be necessary
2219 * to flush the stack on context switch. This happens when the branch
2220 * stack does not tag its entries with the pid of the current task.
2221 * Otherwise it becomes impossible to associate a branch entry with a
2222 * task. This ambiguity is more likely to appear when the branch stack
2223 * supports priv level filtering and the user sets it to monitor only
2224 * at the user level (which could be a useful measurement in system-wide
2225 * mode). In that case, the risk is high of having a branch stack with
2226 * branch from multiple tasks. Flushing may mean dropping the existing
2227 * entries or stashing them somewhere in the PMU specific code layer.
2229 * This function provides the context switch callback to the lower code
2230 * layer. It is invoked ONLY when there is at least one system-wide context
2231 * with at least one active event using taken branch sampling.
2233 static void perf_branch_stack_sched_in(struct task_struct
*prev
,
2234 struct task_struct
*task
)
2236 struct perf_cpu_context
*cpuctx
;
2238 unsigned long flags
;
2240 /* no need to flush branch stack if not changing task */
2244 local_irq_save(flags
);
2248 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
2249 cpuctx
= this_cpu_ptr(pmu
->pmu_cpu_context
);
2252 * check if the context has at least one
2253 * event using PERF_SAMPLE_BRANCH_STACK
2255 if (cpuctx
->ctx
.nr_branch_stack
> 0
2256 && pmu
->flush_branch_stack
) {
2258 pmu
= cpuctx
->ctx
.pmu
;
2260 perf_ctx_lock(cpuctx
, cpuctx
->task_ctx
);
2262 perf_pmu_disable(pmu
);
2264 pmu
->flush_branch_stack();
2266 perf_pmu_enable(pmu
);
2268 perf_ctx_unlock(cpuctx
, cpuctx
->task_ctx
);
2274 local_irq_restore(flags
);
2278 * Called from scheduler to add the events of the current task
2279 * with interrupts disabled.
2281 * We restore the event value and then enable it.
2283 * This does not protect us against NMI, but enable()
2284 * sets the enabled bit in the control field of event _before_
2285 * accessing the event control register. If a NMI hits, then it will
2286 * keep the event running.
2288 void __perf_event_task_sched_in(struct task_struct
*prev
,
2289 struct task_struct
*task
)
2291 struct perf_event_context
*ctx
;
2294 for_each_task_context_nr(ctxn
) {
2295 ctx
= task
->perf_event_ctxp
[ctxn
];
2299 perf_event_context_sched_in(ctx
, task
);
2302 * if cgroup events exist on this CPU, then we need
2303 * to check if we have to switch in PMU state.
2304 * cgroup event are system-wide mode only
2306 if (atomic_read(&__get_cpu_var(perf_cgroup_events
)))
2307 perf_cgroup_sched_in(prev
, task
);
2309 /* check for system-wide branch_stack events */
2310 if (atomic_read(&__get_cpu_var(perf_branch_stack_events
)))
2311 perf_branch_stack_sched_in(prev
, task
);
2314 static u64
perf_calculate_period(struct perf_event
*event
, u64 nsec
, u64 count
)
2316 u64 frequency
= event
->attr
.sample_freq
;
2317 u64 sec
= NSEC_PER_SEC
;
2318 u64 divisor
, dividend
;
2320 int count_fls
, nsec_fls
, frequency_fls
, sec_fls
;
2322 count_fls
= fls64(count
);
2323 nsec_fls
= fls64(nsec
);
2324 frequency_fls
= fls64(frequency
);
2328 * We got @count in @nsec, with a target of sample_freq HZ
2329 * the target period becomes:
2332 * period = -------------------
2333 * @nsec * sample_freq
2338 * Reduce accuracy by one bit such that @a and @b converge
2339 * to a similar magnitude.
2341 #define REDUCE_FLS(a, b) \
2343 if (a##_fls > b##_fls) { \
2353 * Reduce accuracy until either term fits in a u64, then proceed with
2354 * the other, so that finally we can do a u64/u64 division.
2356 while (count_fls
+ sec_fls
> 64 && nsec_fls
+ frequency_fls
> 64) {
2357 REDUCE_FLS(nsec
, frequency
);
2358 REDUCE_FLS(sec
, count
);
2361 if (count_fls
+ sec_fls
> 64) {
2362 divisor
= nsec
* frequency
;
2364 while (count_fls
+ sec_fls
> 64) {
2365 REDUCE_FLS(count
, sec
);
2369 dividend
= count
* sec
;
2371 dividend
= count
* sec
;
2373 while (nsec_fls
+ frequency_fls
> 64) {
2374 REDUCE_FLS(nsec
, frequency
);
2378 divisor
= nsec
* frequency
;
2384 return div64_u64(dividend
, divisor
);
2387 static DEFINE_PER_CPU(int, perf_throttled_count
);
2388 static DEFINE_PER_CPU(u64
, perf_throttled_seq
);
2390 static void perf_adjust_period(struct perf_event
*event
, u64 nsec
, u64 count
, bool disable
)
2392 struct hw_perf_event
*hwc
= &event
->hw
;
2393 s64 period
, sample_period
;
2396 period
= perf_calculate_period(event
, nsec
, count
);
2398 delta
= (s64
)(period
- hwc
->sample_period
);
2399 delta
= (delta
+ 7) / 8; /* low pass filter */
2401 sample_period
= hwc
->sample_period
+ delta
;
2406 hwc
->sample_period
= sample_period
;
2408 if (local64_read(&hwc
->period_left
) > 8*sample_period
) {
2410 event
->pmu
->stop(event
, PERF_EF_UPDATE
);
2412 local64_set(&hwc
->period_left
, 0);
2415 event
->pmu
->start(event
, PERF_EF_RELOAD
);
2420 * combine freq adjustment with unthrottling to avoid two passes over the
2421 * events. At the same time, make sure, having freq events does not change
2422 * the rate of unthrottling as that would introduce bias.
2424 static void perf_adjust_freq_unthr_context(struct perf_event_context
*ctx
,
2427 struct perf_event
*event
;
2428 struct hw_perf_event
*hwc
;
2429 u64 now
, period
= TICK_NSEC
;
2433 * only need to iterate over all events iff:
2434 * - context have events in frequency mode (needs freq adjust)
2435 * - there are events to unthrottle on this cpu
2437 if (!(ctx
->nr_freq
|| needs_unthr
))
2440 raw_spin_lock(&ctx
->lock
);
2441 perf_pmu_disable(ctx
->pmu
);
2443 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
2444 if (event
->state
!= PERF_EVENT_STATE_ACTIVE
)
2447 if (!event_filter_match(event
))
2452 if (needs_unthr
&& hwc
->interrupts
== MAX_INTERRUPTS
) {
2453 hwc
->interrupts
= 0;
2454 perf_log_throttle(event
, 1);
2455 event
->pmu
->start(event
, 0);
2458 if (!event
->attr
.freq
|| !event
->attr
.sample_freq
)
2462 * stop the event and update event->count
2464 event
->pmu
->stop(event
, PERF_EF_UPDATE
);
2466 now
= local64_read(&event
->count
);
2467 delta
= now
- hwc
->freq_count_stamp
;
2468 hwc
->freq_count_stamp
= now
;
2472 * reload only if value has changed
2473 * we have stopped the event so tell that
2474 * to perf_adjust_period() to avoid stopping it
2478 perf_adjust_period(event
, period
, delta
, false);
2480 event
->pmu
->start(event
, delta
> 0 ? PERF_EF_RELOAD
: 0);
2483 perf_pmu_enable(ctx
->pmu
);
2484 raw_spin_unlock(&ctx
->lock
);
2488 * Round-robin a context's events:
2490 static void rotate_ctx(struct perf_event_context
*ctx
)
2493 * Rotate the first entry last of non-pinned groups. Rotation might be
2494 * disabled by the inheritance code.
2496 if (!ctx
->rotate_disable
)
2497 list_rotate_left(&ctx
->flexible_groups
);
2501 * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
2502 * because they're strictly cpu affine and rotate_start is called with IRQs
2503 * disabled, while rotate_context is called from IRQ context.
2505 static void perf_rotate_context(struct perf_cpu_context
*cpuctx
)
2507 struct perf_event_context
*ctx
= NULL
;
2508 int rotate
= 0, remove
= 1;
2510 if (cpuctx
->ctx
.nr_events
) {
2512 if (cpuctx
->ctx
.nr_events
!= cpuctx
->ctx
.nr_active
)
2516 ctx
= cpuctx
->task_ctx
;
2517 if (ctx
&& ctx
->nr_events
) {
2519 if (ctx
->nr_events
!= ctx
->nr_active
)
2526 perf_ctx_lock(cpuctx
, cpuctx
->task_ctx
);
2527 perf_pmu_disable(cpuctx
->ctx
.pmu
);
2529 cpu_ctx_sched_out(cpuctx
, EVENT_FLEXIBLE
);
2531 ctx_sched_out(ctx
, cpuctx
, EVENT_FLEXIBLE
);
2533 rotate_ctx(&cpuctx
->ctx
);
2537 perf_event_sched_in(cpuctx
, ctx
, current
);
2539 perf_pmu_enable(cpuctx
->ctx
.pmu
);
2540 perf_ctx_unlock(cpuctx
, cpuctx
->task_ctx
);
2543 list_del_init(&cpuctx
->rotation_list
);
2546 void perf_event_task_tick(void)
2548 struct list_head
*head
= &__get_cpu_var(rotation_list
);
2549 struct perf_cpu_context
*cpuctx
, *tmp
;
2550 struct perf_event_context
*ctx
;
2553 WARN_ON(!irqs_disabled());
2555 __this_cpu_inc(perf_throttled_seq
);
2556 throttled
= __this_cpu_xchg(perf_throttled_count
, 0);
2558 list_for_each_entry_safe(cpuctx
, tmp
, head
, rotation_list
) {
2560 perf_adjust_freq_unthr_context(ctx
, throttled
);
2562 ctx
= cpuctx
->task_ctx
;
2564 perf_adjust_freq_unthr_context(ctx
, throttled
);
2566 if (cpuctx
->jiffies_interval
== 1 ||
2567 !(jiffies
% cpuctx
->jiffies_interval
))
2568 perf_rotate_context(cpuctx
);
2572 static int event_enable_on_exec(struct perf_event
*event
,
2573 struct perf_event_context
*ctx
)
2575 if (!event
->attr
.enable_on_exec
)
2578 event
->attr
.enable_on_exec
= 0;
2579 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
)
2582 __perf_event_mark_enabled(event
);
2588 * Enable all of a task's events that have been marked enable-on-exec.
2589 * This expects task == current.
2591 static void perf_event_enable_on_exec(struct perf_event_context
*ctx
)
2593 struct perf_event
*event
;
2594 unsigned long flags
;
2598 local_irq_save(flags
);
2599 if (!ctx
|| !ctx
->nr_events
)
2603 * We must ctxsw out cgroup events to avoid conflict
2604 * when invoking perf_task_event_sched_in() later on
2605 * in this function. Otherwise we end up trying to
2606 * ctxswin cgroup events which are already scheduled
2609 perf_cgroup_sched_out(current
, NULL
);
2611 raw_spin_lock(&ctx
->lock
);
2612 task_ctx_sched_out(ctx
);
2614 list_for_each_entry(event
, &ctx
->event_list
, event_entry
) {
2615 ret
= event_enable_on_exec(event
, ctx
);
2621 * Unclone this context if we enabled any event.
2626 raw_spin_unlock(&ctx
->lock
);
2629 * Also calls ctxswin for cgroup events, if any:
2631 perf_event_context_sched_in(ctx
, ctx
->task
);
2633 local_irq_restore(flags
);
2637 * Cross CPU call to read the hardware event
2639 static void __perf_event_read(void *info
)
2641 struct perf_event
*event
= info
;
2642 struct perf_event_context
*ctx
= event
->ctx
;
2643 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
2646 * If this is a task context, we need to check whether it is
2647 * the current task context of this cpu. If not it has been
2648 * scheduled out before the smp call arrived. In that case
2649 * event->count would have been updated to a recent sample
2650 * when the event was scheduled out.
2652 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
)
2655 raw_spin_lock(&ctx
->lock
);
2656 if (ctx
->is_active
) {
2657 update_context_time(ctx
);
2658 update_cgrp_time_from_event(event
);
2660 update_event_times(event
);
2661 if (event
->state
== PERF_EVENT_STATE_ACTIVE
)
2662 event
->pmu
->read(event
);
2663 raw_spin_unlock(&ctx
->lock
);
2666 static inline u64
perf_event_count(struct perf_event
*event
)
2668 return local64_read(&event
->count
) + atomic64_read(&event
->child_count
);
2671 static u64
perf_event_read(struct perf_event
*event
)
2674 * If event is enabled and currently active on a CPU, update the
2675 * value in the event structure:
2677 if (event
->state
== PERF_EVENT_STATE_ACTIVE
) {
2678 smp_call_function_single(event
->oncpu
,
2679 __perf_event_read
, event
, 1);
2680 } else if (event
->state
== PERF_EVENT_STATE_INACTIVE
) {
2681 struct perf_event_context
*ctx
= event
->ctx
;
2682 unsigned long flags
;
2684 raw_spin_lock_irqsave(&ctx
->lock
, flags
);
2686 * may read while context is not active
2687 * (e.g., thread is blocked), in that case
2688 * we cannot update context time
2690 if (ctx
->is_active
) {
2691 update_context_time(ctx
);
2692 update_cgrp_time_from_event(event
);
2694 update_event_times(event
);
2695 raw_spin_unlock_irqrestore(&ctx
->lock
, flags
);
2698 return perf_event_count(event
);
2702 * Initialize the perf_event context in a task_struct:
2704 static void __perf_event_init_context(struct perf_event_context
*ctx
)
2706 raw_spin_lock_init(&ctx
->lock
);
2707 mutex_init(&ctx
->mutex
);
2708 INIT_LIST_HEAD(&ctx
->pinned_groups
);
2709 INIT_LIST_HEAD(&ctx
->flexible_groups
);
2710 INIT_LIST_HEAD(&ctx
->event_list
);
2711 atomic_set(&ctx
->refcount
, 1);
2714 static struct perf_event_context
*
2715 alloc_perf_context(struct pmu
*pmu
, struct task_struct
*task
)
2717 struct perf_event_context
*ctx
;
2719 ctx
= kzalloc(sizeof(struct perf_event_context
), GFP_KERNEL
);
2723 __perf_event_init_context(ctx
);
2726 get_task_struct(task
);
2733 static struct task_struct
*
2734 find_lively_task_by_vpid(pid_t vpid
)
2736 struct task_struct
*task
;
2743 task
= find_task_by_vpid(vpid
);
2745 get_task_struct(task
);
2749 return ERR_PTR(-ESRCH
);
2751 /* Reuse ptrace permission checks for now. */
2753 if (!ptrace_may_access(task
, PTRACE_MODE_READ
))
2758 put_task_struct(task
);
2759 return ERR_PTR(err
);
2764 * Returns a matching context with refcount and pincount.
2766 static struct perf_event_context
*
2767 find_get_context(struct pmu
*pmu
, struct task_struct
*task
, int cpu
)
2769 struct perf_event_context
*ctx
;
2770 struct perf_cpu_context
*cpuctx
;
2771 unsigned long flags
;
2775 /* Must be root to operate on a CPU event: */
2776 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN
))
2777 return ERR_PTR(-EACCES
);
2780 * We could be clever and allow to attach a event to an
2781 * offline CPU and activate it when the CPU comes up, but
2784 if (!cpu_online(cpu
))
2785 return ERR_PTR(-ENODEV
);
2787 cpuctx
= per_cpu_ptr(pmu
->pmu_cpu_context
, cpu
);
2796 ctxn
= pmu
->task_ctx_nr
;
2801 ctx
= perf_lock_task_context(task
, ctxn
, &flags
);
2805 raw_spin_unlock_irqrestore(&ctx
->lock
, flags
);
2807 ctx
= alloc_perf_context(pmu
, task
);
2813 mutex_lock(&task
->perf_event_mutex
);
2815 * If it has already passed perf_event_exit_task().
2816 * we must see PF_EXITING, it takes this mutex too.
2818 if (task
->flags
& PF_EXITING
)
2820 else if (task
->perf_event_ctxp
[ctxn
])
2825 rcu_assign_pointer(task
->perf_event_ctxp
[ctxn
], ctx
);
2827 mutex_unlock(&task
->perf_event_mutex
);
2829 if (unlikely(err
)) {
2841 return ERR_PTR(err
);
2844 static void perf_event_free_filter(struct perf_event
*event
);
2846 static void free_event_rcu(struct rcu_head
*head
)
2848 struct perf_event
*event
;
2850 event
= container_of(head
, struct perf_event
, rcu_head
);
2852 put_pid_ns(event
->ns
);
2853 perf_event_free_filter(event
);
2857 static void ring_buffer_put(struct ring_buffer
*rb
);
2859 static void free_event(struct perf_event
*event
)
2861 irq_work_sync(&event
->pending
);
2863 if (!event
->parent
) {
2864 if (event
->attach_state
& PERF_ATTACH_TASK
)
2865 static_key_slow_dec_deferred(&perf_sched_events
);
2866 if (event
->attr
.mmap
|| event
->attr
.mmap_data
)
2867 atomic_dec(&nr_mmap_events
);
2868 if (event
->attr
.comm
)
2869 atomic_dec(&nr_comm_events
);
2870 if (event
->attr
.task
)
2871 atomic_dec(&nr_task_events
);
2872 if (event
->attr
.sample_type
& PERF_SAMPLE_CALLCHAIN
)
2873 put_callchain_buffers();
2874 if (is_cgroup_event(event
)) {
2875 atomic_dec(&per_cpu(perf_cgroup_events
, event
->cpu
));
2876 static_key_slow_dec_deferred(&perf_sched_events
);
2879 if (has_branch_stack(event
)) {
2880 static_key_slow_dec_deferred(&perf_sched_events
);
2881 /* is system-wide event */
2882 if (!(event
->attach_state
& PERF_ATTACH_TASK
))
2883 atomic_dec(&per_cpu(perf_branch_stack_events
,
2889 ring_buffer_put(event
->rb
);
2893 if (is_cgroup_event(event
))
2894 perf_detach_cgroup(event
);
2897 event
->destroy(event
);
2900 put_ctx(event
->ctx
);
2902 call_rcu(&event
->rcu_head
, free_event_rcu
);
2905 int perf_event_release_kernel(struct perf_event
*event
)
2907 struct perf_event_context
*ctx
= event
->ctx
;
2909 WARN_ON_ONCE(ctx
->parent_ctx
);
2911 * There are two ways this annotation is useful:
2913 * 1) there is a lock recursion from perf_event_exit_task
2914 * see the comment there.
2916 * 2) there is a lock-inversion with mmap_sem through
2917 * perf_event_read_group(), which takes faults while
2918 * holding ctx->mutex, however this is called after
2919 * the last filedesc died, so there is no possibility
2920 * to trigger the AB-BA case.
2922 mutex_lock_nested(&ctx
->mutex
, SINGLE_DEPTH_NESTING
);
2923 raw_spin_lock_irq(&ctx
->lock
);
2924 perf_group_detach(event
);
2925 raw_spin_unlock_irq(&ctx
->lock
);
2926 perf_remove_from_context(event
);
2927 mutex_unlock(&ctx
->mutex
);
2933 EXPORT_SYMBOL_GPL(perf_event_release_kernel
);
2936 * Called when the last reference to the file is gone.
2938 static void put_event(struct perf_event
*event
)
2940 struct task_struct
*owner
;
2942 if (!atomic_long_dec_and_test(&event
->refcount
))
2946 owner
= ACCESS_ONCE(event
->owner
);
2948 * Matches the smp_wmb() in perf_event_exit_task(). If we observe
2949 * !owner it means the list deletion is complete and we can indeed
2950 * free this event, otherwise we need to serialize on
2951 * owner->perf_event_mutex.
2953 smp_read_barrier_depends();
2956 * Since delayed_put_task_struct() also drops the last
2957 * task reference we can safely take a new reference
2958 * while holding the rcu_read_lock().
2960 get_task_struct(owner
);
2965 mutex_lock(&owner
->perf_event_mutex
);
2967 * We have to re-check the event->owner field, if it is cleared
2968 * we raced with perf_event_exit_task(), acquiring the mutex
2969 * ensured they're done, and we can proceed with freeing the
2973 list_del_init(&event
->owner_entry
);
2974 mutex_unlock(&owner
->perf_event_mutex
);
2975 put_task_struct(owner
);
2978 perf_event_release_kernel(event
);
2981 static int perf_release(struct inode
*inode
, struct file
*file
)
2983 put_event(file
->private_data
);
2987 u64
perf_event_read_value(struct perf_event
*event
, u64
*enabled
, u64
*running
)
2989 struct perf_event
*child
;
2995 mutex_lock(&event
->child_mutex
);
2996 total
+= perf_event_read(event
);
2997 *enabled
+= event
->total_time_enabled
+
2998 atomic64_read(&event
->child_total_time_enabled
);
2999 *running
+= event
->total_time_running
+
3000 atomic64_read(&event
->child_total_time_running
);
3002 list_for_each_entry(child
, &event
->child_list
, child_list
) {
3003 total
+= perf_event_read(child
);
3004 *enabled
+= child
->total_time_enabled
;
3005 *running
+= child
->total_time_running
;
3007 mutex_unlock(&event
->child_mutex
);
3011 EXPORT_SYMBOL_GPL(perf_event_read_value
);
3013 static int perf_event_read_group(struct perf_event
*event
,
3014 u64 read_format
, char __user
*buf
)
3016 struct perf_event
*leader
= event
->group_leader
, *sub
;
3017 int n
= 0, size
= 0, ret
= -EFAULT
;
3018 struct perf_event_context
*ctx
= leader
->ctx
;
3020 u64 count
, enabled
, running
;
3022 mutex_lock(&ctx
->mutex
);
3023 count
= perf_event_read_value(leader
, &enabled
, &running
);
3025 values
[n
++] = 1 + leader
->nr_siblings
;
3026 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
3027 values
[n
++] = enabled
;
3028 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
3029 values
[n
++] = running
;
3030 values
[n
++] = count
;
3031 if (read_format
& PERF_FORMAT_ID
)
3032 values
[n
++] = primary_event_id(leader
);
3034 size
= n
* sizeof(u64
);
3036 if (copy_to_user(buf
, values
, size
))
3041 list_for_each_entry(sub
, &leader
->sibling_list
, group_entry
) {
3044 values
[n
++] = perf_event_read_value(sub
, &enabled
, &running
);
3045 if (read_format
& PERF_FORMAT_ID
)
3046 values
[n
++] = primary_event_id(sub
);
3048 size
= n
* sizeof(u64
);
3050 if (copy_to_user(buf
+ ret
, values
, size
)) {
3058 mutex_unlock(&ctx
->mutex
);
3063 static int perf_event_read_one(struct perf_event
*event
,
3064 u64 read_format
, char __user
*buf
)
3066 u64 enabled
, running
;
3070 values
[n
++] = perf_event_read_value(event
, &enabled
, &running
);
3071 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
3072 values
[n
++] = enabled
;
3073 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
3074 values
[n
++] = running
;
3075 if (read_format
& PERF_FORMAT_ID
)
3076 values
[n
++] = primary_event_id(event
);
3078 if (copy_to_user(buf
, values
, n
* sizeof(u64
)))
3081 return n
* sizeof(u64
);
3085 * Read the performance event - simple non blocking version for now
3088 perf_read_hw(struct perf_event
*event
, char __user
*buf
, size_t count
)
3090 u64 read_format
= event
->attr
.read_format
;
3094 * Return end-of-file for a read on a event that is in
3095 * error state (i.e. because it was pinned but it couldn't be
3096 * scheduled on to the CPU at some point).
3098 if (event
->state
== PERF_EVENT_STATE_ERROR
)
3101 if (count
< event
->read_size
)
3104 WARN_ON_ONCE(event
->ctx
->parent_ctx
);
3105 if (read_format
& PERF_FORMAT_GROUP
)
3106 ret
= perf_event_read_group(event
, read_format
, buf
);
3108 ret
= perf_event_read_one(event
, read_format
, buf
);
3114 perf_read(struct file
*file
, char __user
*buf
, size_t count
, loff_t
*ppos
)
3116 struct perf_event
*event
= file
->private_data
;
3118 return perf_read_hw(event
, buf
, count
);
3121 static unsigned int perf_poll(struct file
*file
, poll_table
*wait
)
3123 struct perf_event
*event
= file
->private_data
;
3124 struct ring_buffer
*rb
;
3125 unsigned int events
= POLL_HUP
;
3128 * Race between perf_event_set_output() and perf_poll(): perf_poll()
3129 * grabs the rb reference but perf_event_set_output() overrides it.
3130 * Here is the timeline for two threads T1, T2:
3131 * t0: T1, rb = rcu_dereference(event->rb)
3132 * t1: T2, old_rb = event->rb
3133 * t2: T2, event->rb = new rb
3134 * t3: T2, ring_buffer_detach(old_rb)
3135 * t4: T1, ring_buffer_attach(rb1)
3136 * t5: T1, poll_wait(event->waitq)
3138 * To avoid this problem, we grab mmap_mutex in perf_poll()
3139 * thereby ensuring that the assignment of the new ring buffer
3140 * and the detachment of the old buffer appear atomic to perf_poll()
3142 mutex_lock(&event
->mmap_mutex
);
3145 rb
= rcu_dereference(event
->rb
);
3147 ring_buffer_attach(event
, rb
);
3148 events
= atomic_xchg(&rb
->poll
, 0);
3152 mutex_unlock(&event
->mmap_mutex
);
3154 poll_wait(file
, &event
->waitq
, wait
);
3159 static void perf_event_reset(struct perf_event
*event
)
3161 (void)perf_event_read(event
);
3162 local64_set(&event
->count
, 0);
3163 perf_event_update_userpage(event
);
3167 * Holding the top-level event's child_mutex means that any
3168 * descendant process that has inherited this event will block
3169 * in sync_child_event if it goes to exit, thus satisfying the
3170 * task existence requirements of perf_event_enable/disable.
3172 static void perf_event_for_each_child(struct perf_event
*event
,
3173 void (*func
)(struct perf_event
*))
3175 struct perf_event
*child
;
3177 WARN_ON_ONCE(event
->ctx
->parent_ctx
);
3178 mutex_lock(&event
->child_mutex
);
3180 list_for_each_entry(child
, &event
->child_list
, child_list
)
3182 mutex_unlock(&event
->child_mutex
);
3185 static void perf_event_for_each(struct perf_event
*event
,
3186 void (*func
)(struct perf_event
*))
3188 struct perf_event_context
*ctx
= event
->ctx
;
3189 struct perf_event
*sibling
;
3191 WARN_ON_ONCE(ctx
->parent_ctx
);
3192 mutex_lock(&ctx
->mutex
);
3193 event
= event
->group_leader
;
3195 perf_event_for_each_child(event
, func
);
3196 list_for_each_entry(sibling
, &event
->sibling_list
, group_entry
)
3197 perf_event_for_each_child(sibling
, func
);
3198 mutex_unlock(&ctx
->mutex
);
3201 static int perf_event_period(struct perf_event
*event
, u64 __user
*arg
)
3203 struct perf_event_context
*ctx
= event
->ctx
;
3207 if (!is_sampling_event(event
))
3210 if (copy_from_user(&value
, arg
, sizeof(value
)))
3216 raw_spin_lock_irq(&ctx
->lock
);
3217 if (event
->attr
.freq
) {
3218 if (value
> sysctl_perf_event_sample_rate
) {
3223 event
->attr
.sample_freq
= value
;
3225 event
->attr
.sample_period
= value
;
3226 event
->hw
.sample_period
= value
;
3229 raw_spin_unlock_irq(&ctx
->lock
);
3234 static const struct file_operations perf_fops
;
3236 static inline int perf_fget_light(int fd
, struct fd
*p
)
3238 struct fd f
= fdget(fd
);
3242 if (f
.file
->f_op
!= &perf_fops
) {
3250 static int perf_event_set_output(struct perf_event
*event
,
3251 struct perf_event
*output_event
);
3252 static int perf_event_set_filter(struct perf_event
*event
, void __user
*arg
);
3254 static long perf_ioctl(struct file
*file
, unsigned int cmd
, unsigned long arg
)
3256 struct perf_event
*event
= file
->private_data
;
3257 void (*func
)(struct perf_event
*);
3261 case PERF_EVENT_IOC_ENABLE
:
3262 func
= perf_event_enable
;
3264 case PERF_EVENT_IOC_DISABLE
:
3265 func
= perf_event_disable
;
3267 case PERF_EVENT_IOC_RESET
:
3268 func
= perf_event_reset
;
3271 case PERF_EVENT_IOC_REFRESH
:
3272 return perf_event_refresh(event
, arg
);
3274 case PERF_EVENT_IOC_PERIOD
:
3275 return perf_event_period(event
, (u64 __user
*)arg
);
3277 case PERF_EVENT_IOC_SET_OUTPUT
:
3281 struct perf_event
*output_event
;
3283 ret
= perf_fget_light(arg
, &output
);
3286 output_event
= output
.file
->private_data
;
3287 ret
= perf_event_set_output(event
, output_event
);
3290 ret
= perf_event_set_output(event
, NULL
);
3295 case PERF_EVENT_IOC_SET_FILTER
:
3296 return perf_event_set_filter(event
, (void __user
*)arg
);
3302 if (flags
& PERF_IOC_FLAG_GROUP
)
3303 perf_event_for_each(event
, func
);
3305 perf_event_for_each_child(event
, func
);
3310 int perf_event_task_enable(void)
3312 struct perf_event
*event
;
3314 mutex_lock(¤t
->perf_event_mutex
);
3315 list_for_each_entry(event
, ¤t
->perf_event_list
, owner_entry
)
3316 perf_event_for_each_child(event
, perf_event_enable
);
3317 mutex_unlock(¤t
->perf_event_mutex
);
3322 int perf_event_task_disable(void)
3324 struct perf_event
*event
;
3326 mutex_lock(¤t
->perf_event_mutex
);
3327 list_for_each_entry(event
, ¤t
->perf_event_list
, owner_entry
)
3328 perf_event_for_each_child(event
, perf_event_disable
);
3329 mutex_unlock(¤t
->perf_event_mutex
);
3334 static int perf_event_index(struct perf_event
*event
)
3336 if (event
->hw
.state
& PERF_HES_STOPPED
)
3339 if (event
->state
!= PERF_EVENT_STATE_ACTIVE
)
3342 return event
->pmu
->event_idx(event
);
3345 static void calc_timer_values(struct perf_event
*event
,
3352 *now
= perf_clock();
3353 ctx_time
= event
->shadow_ctx_time
+ *now
;
3354 *enabled
= ctx_time
- event
->tstamp_enabled
;
3355 *running
= ctx_time
- event
->tstamp_running
;
3358 void __weak
arch_perf_update_userpage(struct perf_event_mmap_page
*userpg
, u64 now
)
3363 * Callers need to ensure there can be no nesting of this function, otherwise
3364 * the seqlock logic goes bad. We can not serialize this because the arch
3365 * code calls this from NMI context.
3367 void perf_event_update_userpage(struct perf_event
*event
)
3369 struct perf_event_mmap_page
*userpg
;
3370 struct ring_buffer
*rb
;
3371 u64 enabled
, running
, now
;
3375 * compute total_time_enabled, total_time_running
3376 * based on snapshot values taken when the event
3377 * was last scheduled in.
3379 * we cannot simply called update_context_time()
3380 * because of locking issue as we can be called in
3383 calc_timer_values(event
, &now
, &enabled
, &running
);
3384 rb
= rcu_dereference(event
->rb
);
3388 userpg
= rb
->user_page
;
3391 * Disable preemption so as to not let the corresponding user-space
3392 * spin too long if we get preempted.
3397 userpg
->index
= perf_event_index(event
);
3398 userpg
->offset
= perf_event_count(event
);
3400 userpg
->offset
-= local64_read(&event
->hw
.prev_count
);
3402 userpg
->time_enabled
= enabled
+
3403 atomic64_read(&event
->child_total_time_enabled
);
3405 userpg
->time_running
= running
+
3406 atomic64_read(&event
->child_total_time_running
);
3408 arch_perf_update_userpage(userpg
, now
);
3417 static int perf_mmap_fault(struct vm_area_struct
*vma
, struct vm_fault
*vmf
)
3419 struct perf_event
*event
= vma
->vm_file
->private_data
;
3420 struct ring_buffer
*rb
;
3421 int ret
= VM_FAULT_SIGBUS
;
3423 if (vmf
->flags
& FAULT_FLAG_MKWRITE
) {
3424 if (vmf
->pgoff
== 0)
3430 rb
= rcu_dereference(event
->rb
);
3434 if (vmf
->pgoff
&& (vmf
->flags
& FAULT_FLAG_WRITE
))
3437 vmf
->page
= perf_mmap_to_page(rb
, vmf
->pgoff
);
3441 get_page(vmf
->page
);
3442 vmf
->page
->mapping
= vma
->vm_file
->f_mapping
;
3443 vmf
->page
->index
= vmf
->pgoff
;
3452 static void ring_buffer_attach(struct perf_event
*event
,
3453 struct ring_buffer
*rb
)
3455 unsigned long flags
;
3457 if (!list_empty(&event
->rb_entry
))
3460 spin_lock_irqsave(&rb
->event_lock
, flags
);
3461 if (!list_empty(&event
->rb_entry
))
3464 list_add(&event
->rb_entry
, &rb
->event_list
);
3466 spin_unlock_irqrestore(&rb
->event_lock
, flags
);
3469 static void ring_buffer_detach(struct perf_event
*event
,
3470 struct ring_buffer
*rb
)
3472 unsigned long flags
;
3474 if (list_empty(&event
->rb_entry
))
3477 spin_lock_irqsave(&rb
->event_lock
, flags
);
3478 list_del_init(&event
->rb_entry
);
3479 wake_up_all(&event
->waitq
);
3480 spin_unlock_irqrestore(&rb
->event_lock
, flags
);
3483 static void ring_buffer_wakeup(struct perf_event
*event
)
3485 struct ring_buffer
*rb
;
3488 rb
= rcu_dereference(event
->rb
);
3492 list_for_each_entry_rcu(event
, &rb
->event_list
, rb_entry
)
3493 wake_up_all(&event
->waitq
);
3499 static void rb_free_rcu(struct rcu_head
*rcu_head
)
3501 struct ring_buffer
*rb
;
3503 rb
= container_of(rcu_head
, struct ring_buffer
, rcu_head
);
3507 static struct ring_buffer
*ring_buffer_get(struct perf_event
*event
)
3509 struct ring_buffer
*rb
;
3512 rb
= rcu_dereference(event
->rb
);
3514 if (!atomic_inc_not_zero(&rb
->refcount
))
3522 static void ring_buffer_put(struct ring_buffer
*rb
)
3524 struct perf_event
*event
, *n
;
3525 unsigned long flags
;
3527 if (!atomic_dec_and_test(&rb
->refcount
))
3530 spin_lock_irqsave(&rb
->event_lock
, flags
);
3531 list_for_each_entry_safe(event
, n
, &rb
->event_list
, rb_entry
) {
3532 list_del_init(&event
->rb_entry
);
3533 wake_up_all(&event
->waitq
);
3535 spin_unlock_irqrestore(&rb
->event_lock
, flags
);
3537 call_rcu(&rb
->rcu_head
, rb_free_rcu
);
3540 static void perf_mmap_open(struct vm_area_struct
*vma
)
3542 struct perf_event
*event
= vma
->vm_file
->private_data
;
3544 atomic_inc(&event
->mmap_count
);
3547 static void perf_mmap_close(struct vm_area_struct
*vma
)
3549 struct perf_event
*event
= vma
->vm_file
->private_data
;
3551 if (atomic_dec_and_mutex_lock(&event
->mmap_count
, &event
->mmap_mutex
)) {
3552 unsigned long size
= perf_data_size(event
->rb
);
3553 struct user_struct
*user
= event
->mmap_user
;
3554 struct ring_buffer
*rb
= event
->rb
;
3556 atomic_long_sub((size
>> PAGE_SHIFT
) + 1, &user
->locked_vm
);
3557 vma
->vm_mm
->pinned_vm
-= event
->mmap_locked
;
3558 rcu_assign_pointer(event
->rb
, NULL
);
3559 ring_buffer_detach(event
, rb
);
3560 mutex_unlock(&event
->mmap_mutex
);
3562 ring_buffer_put(rb
);
3567 static const struct vm_operations_struct perf_mmap_vmops
= {
3568 .open
= perf_mmap_open
,
3569 .close
= perf_mmap_close
,
3570 .fault
= perf_mmap_fault
,
3571 .page_mkwrite
= perf_mmap_fault
,
3574 static int perf_mmap(struct file
*file
, struct vm_area_struct
*vma
)
3576 struct perf_event
*event
= file
->private_data
;
3577 unsigned long user_locked
, user_lock_limit
;
3578 struct user_struct
*user
= current_user();
3579 unsigned long locked
, lock_limit
;
3580 struct ring_buffer
*rb
;
3581 unsigned long vma_size
;
3582 unsigned long nr_pages
;
3583 long user_extra
, extra
;
3584 int ret
= 0, flags
= 0;
3587 * Don't allow mmap() of inherited per-task counters. This would
3588 * create a performance issue due to all children writing to the
3591 if (event
->cpu
== -1 && event
->attr
.inherit
)
3594 if (!(vma
->vm_flags
& VM_SHARED
))
3597 vma_size
= vma
->vm_end
- vma
->vm_start
;
3598 nr_pages
= (vma_size
/ PAGE_SIZE
) - 1;
3601 * If we have rb pages ensure they're a power-of-two number, so we
3602 * can do bitmasks instead of modulo.
3604 if (nr_pages
!= 0 && !is_power_of_2(nr_pages
))
3607 if (vma_size
!= PAGE_SIZE
* (1 + nr_pages
))
3610 if (vma
->vm_pgoff
!= 0)
3613 WARN_ON_ONCE(event
->ctx
->parent_ctx
);
3614 mutex_lock(&event
->mmap_mutex
);
3616 if (event
->rb
->nr_pages
== nr_pages
)
3617 atomic_inc(&event
->rb
->refcount
);
3623 user_extra
= nr_pages
+ 1;
3624 user_lock_limit
= sysctl_perf_event_mlock
>> (PAGE_SHIFT
- 10);
3627 * Increase the limit linearly with more CPUs:
3629 user_lock_limit
*= num_online_cpus();
3631 user_locked
= atomic_long_read(&user
->locked_vm
) + user_extra
;
3634 if (user_locked
> user_lock_limit
)
3635 extra
= user_locked
- user_lock_limit
;
3637 lock_limit
= rlimit(RLIMIT_MEMLOCK
);
3638 lock_limit
>>= PAGE_SHIFT
;
3639 locked
= vma
->vm_mm
->pinned_vm
+ extra
;
3641 if ((locked
> lock_limit
) && perf_paranoid_tracepoint_raw() &&
3642 !capable(CAP_IPC_LOCK
)) {
3649 if (vma
->vm_flags
& VM_WRITE
)
3650 flags
|= RING_BUFFER_WRITABLE
;
3652 rb
= rb_alloc(nr_pages
,
3653 event
->attr
.watermark
? event
->attr
.wakeup_watermark
: 0,
3660 rcu_assign_pointer(event
->rb
, rb
);
3662 atomic_long_add(user_extra
, &user
->locked_vm
);
3663 event
->mmap_locked
= extra
;
3664 event
->mmap_user
= get_current_user();
3665 vma
->vm_mm
->pinned_vm
+= event
->mmap_locked
;
3667 perf_event_update_userpage(event
);
3671 atomic_inc(&event
->mmap_count
);
3672 mutex_unlock(&event
->mmap_mutex
);
3674 vma
->vm_flags
|= VM_DONTEXPAND
| VM_DONTDUMP
;
3675 vma
->vm_ops
= &perf_mmap_vmops
;
3680 static int perf_fasync(int fd
, struct file
*filp
, int on
)
3682 struct inode
*inode
= filp
->f_path
.dentry
->d_inode
;
3683 struct perf_event
*event
= filp
->private_data
;
3686 mutex_lock(&inode
->i_mutex
);
3687 retval
= fasync_helper(fd
, filp
, on
, &event
->fasync
);
3688 mutex_unlock(&inode
->i_mutex
);
3696 static const struct file_operations perf_fops
= {
3697 .llseek
= no_llseek
,
3698 .release
= perf_release
,
3701 .unlocked_ioctl
= perf_ioctl
,
3702 .compat_ioctl
= perf_ioctl
,
3704 .fasync
= perf_fasync
,
3710 * If there's data, ensure we set the poll() state and publish everything
3711 * to user-space before waking everybody up.
3714 void perf_event_wakeup(struct perf_event
*event
)
3716 ring_buffer_wakeup(event
);
3718 if (event
->pending_kill
) {
3719 kill_fasync(&event
->fasync
, SIGIO
, event
->pending_kill
);
3720 event
->pending_kill
= 0;
3724 static void perf_pending_event(struct irq_work
*entry
)
3726 struct perf_event
*event
= container_of(entry
,
3727 struct perf_event
, pending
);
3729 if (event
->pending_disable
) {
3730 event
->pending_disable
= 0;
3731 __perf_event_disable(event
);
3734 if (event
->pending_wakeup
) {
3735 event
->pending_wakeup
= 0;
3736 perf_event_wakeup(event
);
3741 * We assume there is only KVM supporting the callbacks.
3742 * Later on, we might change it to a list if there is
3743 * another virtualization implementation supporting the callbacks.
3745 struct perf_guest_info_callbacks
*perf_guest_cbs
;
3747 int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks
*cbs
)
3749 perf_guest_cbs
= cbs
;
3752 EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks
);
3754 int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks
*cbs
)
3756 perf_guest_cbs
= NULL
;
3759 EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks
);
3762 perf_output_sample_regs(struct perf_output_handle
*handle
,
3763 struct pt_regs
*regs
, u64 mask
)
3767 for_each_set_bit(bit
, (const unsigned long *) &mask
,
3768 sizeof(mask
) * BITS_PER_BYTE
) {
3771 val
= perf_reg_value(regs
, bit
);
3772 perf_output_put(handle
, val
);
3776 static void perf_sample_regs_user(struct perf_regs_user
*regs_user
,
3777 struct pt_regs
*regs
)
3779 if (!user_mode(regs
)) {
3781 regs
= task_pt_regs(current
);
3787 regs_user
->regs
= regs
;
3788 regs_user
->abi
= perf_reg_abi(current
);
3793 * Get remaining task size from user stack pointer.
3795 * It'd be better to take stack vma map and limit this more
3796 * precisly, but there's no way to get it safely under interrupt,
3797 * so using TASK_SIZE as limit.
3799 static u64
perf_ustack_task_size(struct pt_regs
*regs
)
3801 unsigned long addr
= perf_user_stack_pointer(regs
);
3803 if (!addr
|| addr
>= TASK_SIZE
)
3806 return TASK_SIZE
- addr
;
3810 perf_sample_ustack_size(u16 stack_size
, u16 header_size
,
3811 struct pt_regs
*regs
)
3815 /* No regs, no stack pointer, no dump. */
3820 * Check if we fit in with the requested stack size into the:
3822 * If we don't, we limit the size to the TASK_SIZE.
3824 * - remaining sample size
3825 * If we don't, we customize the stack size to
3826 * fit in to the remaining sample size.
3829 task_size
= min((u64
) USHRT_MAX
, perf_ustack_task_size(regs
));
3830 stack_size
= min(stack_size
, (u16
) task_size
);
3832 /* Current header size plus static size and dynamic size. */
3833 header_size
+= 2 * sizeof(u64
);
3835 /* Do we fit in with the current stack dump size? */
3836 if ((u16
) (header_size
+ stack_size
) < header_size
) {
3838 * If we overflow the maximum size for the sample,
3839 * we customize the stack dump size to fit in.
3841 stack_size
= USHRT_MAX
- header_size
- sizeof(u64
);
3842 stack_size
= round_up(stack_size
, sizeof(u64
));
3849 perf_output_sample_ustack(struct perf_output_handle
*handle
, u64 dump_size
,
3850 struct pt_regs
*regs
)
3852 /* Case of a kernel thread, nothing to dump */
3855 perf_output_put(handle
, size
);
3864 * - the size requested by user or the best one we can fit
3865 * in to the sample max size
3867 * - user stack dump data
3869 * - the actual dumped size
3873 perf_output_put(handle
, dump_size
);
3876 sp
= perf_user_stack_pointer(regs
);
3877 rem
= __output_copy_user(handle
, (void *) sp
, dump_size
);
3878 dyn_size
= dump_size
- rem
;
3880 perf_output_skip(handle
, rem
);
3883 perf_output_put(handle
, dyn_size
);
3887 static void __perf_event_header__init_id(struct perf_event_header
*header
,
3888 struct perf_sample_data
*data
,
3889 struct perf_event
*event
)
3891 u64 sample_type
= event
->attr
.sample_type
;
3893 data
->type
= sample_type
;
3894 header
->size
+= event
->id_header_size
;
3896 if (sample_type
& PERF_SAMPLE_TID
) {
3897 /* namespace issues */
3898 data
->tid_entry
.pid
= perf_event_pid(event
, current
);
3899 data
->tid_entry
.tid
= perf_event_tid(event
, current
);
3902 if (sample_type
& PERF_SAMPLE_TIME
)
3903 data
->time
= perf_clock();
3905 if (sample_type
& PERF_SAMPLE_ID
)
3906 data
->id
= primary_event_id(event
);
3908 if (sample_type
& PERF_SAMPLE_STREAM_ID
)
3909 data
->stream_id
= event
->id
;
3911 if (sample_type
& PERF_SAMPLE_CPU
) {
3912 data
->cpu_entry
.cpu
= raw_smp_processor_id();
3913 data
->cpu_entry
.reserved
= 0;
3917 void perf_event_header__init_id(struct perf_event_header
*header
,
3918 struct perf_sample_data
*data
,
3919 struct perf_event
*event
)
3921 if (event
->attr
.sample_id_all
)
3922 __perf_event_header__init_id(header
, data
, event
);
3925 static void __perf_event__output_id_sample(struct perf_output_handle
*handle
,
3926 struct perf_sample_data
*data
)
3928 u64 sample_type
= data
->type
;
3930 if (sample_type
& PERF_SAMPLE_TID
)
3931 perf_output_put(handle
, data
->tid_entry
);
3933 if (sample_type
& PERF_SAMPLE_TIME
)
3934 perf_output_put(handle
, data
->time
);
3936 if (sample_type
& PERF_SAMPLE_ID
)
3937 perf_output_put(handle
, data
->id
);
3939 if (sample_type
& PERF_SAMPLE_STREAM_ID
)
3940 perf_output_put(handle
, data
->stream_id
);
3942 if (sample_type
& PERF_SAMPLE_CPU
)
3943 perf_output_put(handle
, data
->cpu_entry
);
3946 void perf_event__output_id_sample(struct perf_event
*event
,
3947 struct perf_output_handle
*handle
,
3948 struct perf_sample_data
*sample
)
3950 if (event
->attr
.sample_id_all
)
3951 __perf_event__output_id_sample(handle
, sample
);
3954 static void perf_output_read_one(struct perf_output_handle
*handle
,
3955 struct perf_event
*event
,
3956 u64 enabled
, u64 running
)
3958 u64 read_format
= event
->attr
.read_format
;
3962 values
[n
++] = perf_event_count(event
);
3963 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
) {
3964 values
[n
++] = enabled
+
3965 atomic64_read(&event
->child_total_time_enabled
);
3967 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
) {
3968 values
[n
++] = running
+
3969 atomic64_read(&event
->child_total_time_running
);
3971 if (read_format
& PERF_FORMAT_ID
)
3972 values
[n
++] = primary_event_id(event
);
3974 __output_copy(handle
, values
, n
* sizeof(u64
));
3978 * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
3980 static void perf_output_read_group(struct perf_output_handle
*handle
,
3981 struct perf_event
*event
,
3982 u64 enabled
, u64 running
)
3984 struct perf_event
*leader
= event
->group_leader
, *sub
;
3985 u64 read_format
= event
->attr
.read_format
;
3989 values
[n
++] = 1 + leader
->nr_siblings
;
3991 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
3992 values
[n
++] = enabled
;
3994 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
3995 values
[n
++] = running
;
3997 if (leader
!= event
)
3998 leader
->pmu
->read(leader
);
4000 values
[n
++] = perf_event_count(leader
);
4001 if (read_format
& PERF_FORMAT_ID
)
4002 values
[n
++] = primary_event_id(leader
);
4004 __output_copy(handle
, values
, n
* sizeof(u64
));
4006 list_for_each_entry(sub
, &leader
->sibling_list
, group_entry
) {
4010 sub
->pmu
->read(sub
);
4012 values
[n
++] = perf_event_count(sub
);
4013 if (read_format
& PERF_FORMAT_ID
)
4014 values
[n
++] = primary_event_id(sub
);
4016 __output_copy(handle
, values
, n
* sizeof(u64
));
4020 #define PERF_FORMAT_TOTAL_TIMES (PERF_FORMAT_TOTAL_TIME_ENABLED|\
4021 PERF_FORMAT_TOTAL_TIME_RUNNING)
4023 static void perf_output_read(struct perf_output_handle
*handle
,
4024 struct perf_event
*event
)
4026 u64 enabled
= 0, running
= 0, now
;
4027 u64 read_format
= event
->attr
.read_format
;
4030 * compute total_time_enabled, total_time_running
4031 * based on snapshot values taken when the event
4032 * was last scheduled in.
4034 * we cannot simply called update_context_time()
4035 * because of locking issue as we are called in
4038 if (read_format
& PERF_FORMAT_TOTAL_TIMES
)
4039 calc_timer_values(event
, &now
, &enabled
, &running
);
4041 if (event
->attr
.read_format
& PERF_FORMAT_GROUP
)
4042 perf_output_read_group(handle
, event
, enabled
, running
);
4044 perf_output_read_one(handle
, event
, enabled
, running
);
4047 void perf_output_sample(struct perf_output_handle
*handle
,
4048 struct perf_event_header
*header
,
4049 struct perf_sample_data
*data
,
4050 struct perf_event
*event
)
4052 u64 sample_type
= data
->type
;
4054 perf_output_put(handle
, *header
);
4056 if (sample_type
& PERF_SAMPLE_IP
)
4057 perf_output_put(handle
, data
->ip
);
4059 if (sample_type
& PERF_SAMPLE_TID
)
4060 perf_output_put(handle
, data
->tid_entry
);
4062 if (sample_type
& PERF_SAMPLE_TIME
)
4063 perf_output_put(handle
, data
->time
);
4065 if (sample_type
& PERF_SAMPLE_ADDR
)
4066 perf_output_put(handle
, data
->addr
);
4068 if (sample_type
& PERF_SAMPLE_ID
)
4069 perf_output_put(handle
, data
->id
);
4071 if (sample_type
& PERF_SAMPLE_STREAM_ID
)
4072 perf_output_put(handle
, data
->stream_id
);
4074 if (sample_type
& PERF_SAMPLE_CPU
)
4075 perf_output_put(handle
, data
->cpu_entry
);
4077 if (sample_type
& PERF_SAMPLE_PERIOD
)
4078 perf_output_put(handle
, data
->period
);
4080 if (sample_type
& PERF_SAMPLE_READ
)
4081 perf_output_read(handle
, event
);
4083 if (sample_type
& PERF_SAMPLE_CALLCHAIN
) {
4084 if (data
->callchain
) {
4087 if (data
->callchain
)
4088 size
+= data
->callchain
->nr
;
4090 size
*= sizeof(u64
);
4092 __output_copy(handle
, data
->callchain
, size
);
4095 perf_output_put(handle
, nr
);
4099 if (sample_type
& PERF_SAMPLE_RAW
) {
4101 perf_output_put(handle
, data
->raw
->size
);
4102 __output_copy(handle
, data
->raw
->data
,
4109 .size
= sizeof(u32
),
4112 perf_output_put(handle
, raw
);
4116 if (!event
->attr
.watermark
) {
4117 int wakeup_events
= event
->attr
.wakeup_events
;
4119 if (wakeup_events
) {
4120 struct ring_buffer
*rb
= handle
->rb
;
4121 int events
= local_inc_return(&rb
->events
);
4123 if (events
>= wakeup_events
) {
4124 local_sub(wakeup_events
, &rb
->events
);
4125 local_inc(&rb
->wakeup
);
4130 if (sample_type
& PERF_SAMPLE_BRANCH_STACK
) {
4131 if (data
->br_stack
) {
4134 size
= data
->br_stack
->nr
4135 * sizeof(struct perf_branch_entry
);
4137 perf_output_put(handle
, data
->br_stack
->nr
);
4138 perf_output_copy(handle
, data
->br_stack
->entries
, size
);
4141 * we always store at least the value of nr
4144 perf_output_put(handle
, nr
);
4148 if (sample_type
& PERF_SAMPLE_REGS_USER
) {
4149 u64 abi
= data
->regs_user
.abi
;
4152 * If there are no regs to dump, notice it through
4153 * first u64 being zero (PERF_SAMPLE_REGS_ABI_NONE).
4155 perf_output_put(handle
, abi
);
4158 u64 mask
= event
->attr
.sample_regs_user
;
4159 perf_output_sample_regs(handle
,
4160 data
->regs_user
.regs
,
4165 if (sample_type
& PERF_SAMPLE_STACK_USER
)
4166 perf_output_sample_ustack(handle
,
4167 data
->stack_user_size
,
4168 data
->regs_user
.regs
);
4171 void perf_prepare_sample(struct perf_event_header
*header
,
4172 struct perf_sample_data
*data
,
4173 struct perf_event
*event
,
4174 struct pt_regs
*regs
)
4176 u64 sample_type
= event
->attr
.sample_type
;
4178 header
->type
= PERF_RECORD_SAMPLE
;
4179 header
->size
= sizeof(*header
) + event
->header_size
;
4182 header
->misc
|= perf_misc_flags(regs
);
4184 __perf_event_header__init_id(header
, data
, event
);
4186 if (sample_type
& PERF_SAMPLE_IP
)
4187 data
->ip
= perf_instruction_pointer(regs
);
4189 if (sample_type
& PERF_SAMPLE_CALLCHAIN
) {
4192 data
->callchain
= perf_callchain(event
, regs
);
4194 if (data
->callchain
)
4195 size
+= data
->callchain
->nr
;
4197 header
->size
+= size
* sizeof(u64
);
4200 if (sample_type
& PERF_SAMPLE_RAW
) {
4201 int size
= sizeof(u32
);
4204 size
+= data
->raw
->size
;
4206 size
+= sizeof(u32
);
4208 WARN_ON_ONCE(size
& (sizeof(u64
)-1));
4209 header
->size
+= size
;
4212 if (sample_type
& PERF_SAMPLE_BRANCH_STACK
) {
4213 int size
= sizeof(u64
); /* nr */
4214 if (data
->br_stack
) {
4215 size
+= data
->br_stack
->nr
4216 * sizeof(struct perf_branch_entry
);
4218 header
->size
+= size
;
4221 if (sample_type
& PERF_SAMPLE_REGS_USER
) {
4222 /* regs dump ABI info */
4223 int size
= sizeof(u64
);
4225 perf_sample_regs_user(&data
->regs_user
, regs
);
4227 if (data
->regs_user
.regs
) {
4228 u64 mask
= event
->attr
.sample_regs_user
;
4229 size
+= hweight64(mask
) * sizeof(u64
);
4232 header
->size
+= size
;
4235 if (sample_type
& PERF_SAMPLE_STACK_USER
) {
4237 * Either we need PERF_SAMPLE_STACK_USER bit to be allways
4238 * processed as the last one or have additional check added
4239 * in case new sample type is added, because we could eat
4240 * up the rest of the sample size.
4242 struct perf_regs_user
*uregs
= &data
->regs_user
;
4243 u16 stack_size
= event
->attr
.sample_stack_user
;
4244 u16 size
= sizeof(u64
);
4247 perf_sample_regs_user(uregs
, regs
);
4249 stack_size
= perf_sample_ustack_size(stack_size
, header
->size
,
4253 * If there is something to dump, add space for the dump
4254 * itself and for the field that tells the dynamic size,
4255 * which is how many have been actually dumped.
4258 size
+= sizeof(u64
) + stack_size
;
4260 data
->stack_user_size
= stack_size
;
4261 header
->size
+= size
;
4265 static void perf_event_output(struct perf_event
*event
,
4266 struct perf_sample_data
*data
,
4267 struct pt_regs
*regs
)
4269 struct perf_output_handle handle
;
4270 struct perf_event_header header
;
4272 /* protect the callchain buffers */
4275 perf_prepare_sample(&header
, data
, event
, regs
);
4277 if (perf_output_begin(&handle
, event
, header
.size
))
4280 perf_output_sample(&handle
, &header
, data
, event
);
4282 perf_output_end(&handle
);
4292 struct perf_read_event
{
4293 struct perf_event_header header
;
4300 perf_event_read_event(struct perf_event
*event
,
4301 struct task_struct
*task
)
4303 struct perf_output_handle handle
;
4304 struct perf_sample_data sample
;
4305 struct perf_read_event read_event
= {
4307 .type
= PERF_RECORD_READ
,
4309 .size
= sizeof(read_event
) + event
->read_size
,
4311 .pid
= perf_event_pid(event
, task
),
4312 .tid
= perf_event_tid(event
, task
),
4316 perf_event_header__init_id(&read_event
.header
, &sample
, event
);
4317 ret
= perf_output_begin(&handle
, event
, read_event
.header
.size
);
4321 perf_output_put(&handle
, read_event
);
4322 perf_output_read(&handle
, event
);
4323 perf_event__output_id_sample(event
, &handle
, &sample
);
4325 perf_output_end(&handle
);
4329 * task tracking -- fork/exit
4331 * enabled by: attr.comm | attr.mmap | attr.mmap_data | attr.task
4334 struct perf_task_event
{
4335 struct task_struct
*task
;
4336 struct perf_event_context
*task_ctx
;
4339 struct perf_event_header header
;
4349 static void perf_event_task_output(struct perf_event
*event
,
4350 struct perf_task_event
*task_event
)
4352 struct perf_output_handle handle
;
4353 struct perf_sample_data sample
;
4354 struct task_struct
*task
= task_event
->task
;
4355 int ret
, size
= task_event
->event_id
.header
.size
;
4357 perf_event_header__init_id(&task_event
->event_id
.header
, &sample
, event
);
4359 ret
= perf_output_begin(&handle
, event
,
4360 task_event
->event_id
.header
.size
);
4364 task_event
->event_id
.pid
= perf_event_pid(event
, task
);
4365 task_event
->event_id
.ppid
= perf_event_pid(event
, current
);
4367 task_event
->event_id
.tid
= perf_event_tid(event
, task
);
4368 task_event
->event_id
.ptid
= perf_event_tid(event
, current
);
4370 perf_output_put(&handle
, task_event
->event_id
);
4372 perf_event__output_id_sample(event
, &handle
, &sample
);
4374 perf_output_end(&handle
);
4376 task_event
->event_id
.header
.size
= size
;
4379 static int perf_event_task_match(struct perf_event
*event
)
4381 if (event
->state
< PERF_EVENT_STATE_INACTIVE
)
4384 if (!event_filter_match(event
))
4387 if (event
->attr
.comm
|| event
->attr
.mmap
||
4388 event
->attr
.mmap_data
|| event
->attr
.task
)
4394 static void perf_event_task_ctx(struct perf_event_context
*ctx
,
4395 struct perf_task_event
*task_event
)
4397 struct perf_event
*event
;
4399 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
4400 if (perf_event_task_match(event
))
4401 perf_event_task_output(event
, task_event
);
4405 static void perf_event_task_event(struct perf_task_event
*task_event
)
4407 struct perf_cpu_context
*cpuctx
;
4408 struct perf_event_context
*ctx
;
4413 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
4414 cpuctx
= get_cpu_ptr(pmu
->pmu_cpu_context
);
4415 if (cpuctx
->active_pmu
!= pmu
)
4417 perf_event_task_ctx(&cpuctx
->ctx
, task_event
);
4419 ctx
= task_event
->task_ctx
;
4421 ctxn
= pmu
->task_ctx_nr
;
4424 ctx
= rcu_dereference(current
->perf_event_ctxp
[ctxn
]);
4427 perf_event_task_ctx(ctx
, task_event
);
4429 put_cpu_ptr(pmu
->pmu_cpu_context
);
4434 static void perf_event_task(struct task_struct
*task
,
4435 struct perf_event_context
*task_ctx
,
4438 struct perf_task_event task_event
;
4440 if (!atomic_read(&nr_comm_events
) &&
4441 !atomic_read(&nr_mmap_events
) &&
4442 !atomic_read(&nr_task_events
))
4445 task_event
= (struct perf_task_event
){
4447 .task_ctx
= task_ctx
,
4450 .type
= new ? PERF_RECORD_FORK
: PERF_RECORD_EXIT
,
4452 .size
= sizeof(task_event
.event_id
),
4458 .time
= perf_clock(),
4462 perf_event_task_event(&task_event
);
4465 void perf_event_fork(struct task_struct
*task
)
4467 perf_event_task(task
, NULL
, 1);
4474 struct perf_comm_event
{
4475 struct task_struct
*task
;
4480 struct perf_event_header header
;
4487 static void perf_event_comm_output(struct perf_event
*event
,
4488 struct perf_comm_event
*comm_event
)
4490 struct perf_output_handle handle
;
4491 struct perf_sample_data sample
;
4492 int size
= comm_event
->event_id
.header
.size
;
4495 perf_event_header__init_id(&comm_event
->event_id
.header
, &sample
, event
);
4496 ret
= perf_output_begin(&handle
, event
,
4497 comm_event
->event_id
.header
.size
);
4502 comm_event
->event_id
.pid
= perf_event_pid(event
, comm_event
->task
);
4503 comm_event
->event_id
.tid
= perf_event_tid(event
, comm_event
->task
);
4505 perf_output_put(&handle
, comm_event
->event_id
);
4506 __output_copy(&handle
, comm_event
->comm
,
4507 comm_event
->comm_size
);
4509 perf_event__output_id_sample(event
, &handle
, &sample
);
4511 perf_output_end(&handle
);
4513 comm_event
->event_id
.header
.size
= size
;
4516 static int perf_event_comm_match(struct perf_event
*event
)
4518 if (event
->state
< PERF_EVENT_STATE_INACTIVE
)
4521 if (!event_filter_match(event
))
4524 if (event
->attr
.comm
)
4530 static void perf_event_comm_ctx(struct perf_event_context
*ctx
,
4531 struct perf_comm_event
*comm_event
)
4533 struct perf_event
*event
;
4535 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
4536 if (perf_event_comm_match(event
))
4537 perf_event_comm_output(event
, comm_event
);
4541 static void perf_event_comm_event(struct perf_comm_event
*comm_event
)
4543 struct perf_cpu_context
*cpuctx
;
4544 struct perf_event_context
*ctx
;
4545 char comm
[TASK_COMM_LEN
];
4550 memset(comm
, 0, sizeof(comm
));
4551 strlcpy(comm
, comm_event
->task
->comm
, sizeof(comm
));
4552 size
= ALIGN(strlen(comm
)+1, sizeof(u64
));
4554 comm_event
->comm
= comm
;
4555 comm_event
->comm_size
= size
;
4557 comm_event
->event_id
.header
.size
= sizeof(comm_event
->event_id
) + size
;
4559 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
4560 cpuctx
= get_cpu_ptr(pmu
->pmu_cpu_context
);
4561 if (cpuctx
->active_pmu
!= pmu
)
4563 perf_event_comm_ctx(&cpuctx
->ctx
, comm_event
);
4565 ctxn
= pmu
->task_ctx_nr
;
4569 ctx
= rcu_dereference(current
->perf_event_ctxp
[ctxn
]);
4571 perf_event_comm_ctx(ctx
, comm_event
);
4573 put_cpu_ptr(pmu
->pmu_cpu_context
);
4578 void perf_event_comm(struct task_struct
*task
)
4580 struct perf_comm_event comm_event
;
4581 struct perf_event_context
*ctx
;
4584 for_each_task_context_nr(ctxn
) {
4585 ctx
= task
->perf_event_ctxp
[ctxn
];
4589 perf_event_enable_on_exec(ctx
);
4592 if (!atomic_read(&nr_comm_events
))
4595 comm_event
= (struct perf_comm_event
){
4601 .type
= PERF_RECORD_COMM
,
4610 perf_event_comm_event(&comm_event
);
4617 struct perf_mmap_event
{
4618 struct vm_area_struct
*vma
;
4620 const char *file_name
;
4624 struct perf_event_header header
;
4634 static void perf_event_mmap_output(struct perf_event
*event
,
4635 struct perf_mmap_event
*mmap_event
)
4637 struct perf_output_handle handle
;
4638 struct perf_sample_data sample
;
4639 int size
= mmap_event
->event_id
.header
.size
;
4642 perf_event_header__init_id(&mmap_event
->event_id
.header
, &sample
, event
);
4643 ret
= perf_output_begin(&handle
, event
,
4644 mmap_event
->event_id
.header
.size
);
4648 mmap_event
->event_id
.pid
= perf_event_pid(event
, current
);
4649 mmap_event
->event_id
.tid
= perf_event_tid(event
, current
);
4651 perf_output_put(&handle
, mmap_event
->event_id
);
4652 __output_copy(&handle
, mmap_event
->file_name
,
4653 mmap_event
->file_size
);
4655 perf_event__output_id_sample(event
, &handle
, &sample
);
4657 perf_output_end(&handle
);
4659 mmap_event
->event_id
.header
.size
= size
;
4662 static int perf_event_mmap_match(struct perf_event
*event
,
4663 struct perf_mmap_event
*mmap_event
,
4666 if (event
->state
< PERF_EVENT_STATE_INACTIVE
)
4669 if (!event_filter_match(event
))
4672 if ((!executable
&& event
->attr
.mmap_data
) ||
4673 (executable
&& event
->attr
.mmap
))
4679 static void perf_event_mmap_ctx(struct perf_event_context
*ctx
,
4680 struct perf_mmap_event
*mmap_event
,
4683 struct perf_event
*event
;
4685 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
4686 if (perf_event_mmap_match(event
, mmap_event
, executable
))
4687 perf_event_mmap_output(event
, mmap_event
);
4691 static void perf_event_mmap_event(struct perf_mmap_event
*mmap_event
)
4693 struct perf_cpu_context
*cpuctx
;
4694 struct perf_event_context
*ctx
;
4695 struct vm_area_struct
*vma
= mmap_event
->vma
;
4696 struct file
*file
= vma
->vm_file
;
4704 memset(tmp
, 0, sizeof(tmp
));
4708 * d_path works from the end of the rb backwards, so we
4709 * need to add enough zero bytes after the string to handle
4710 * the 64bit alignment we do later.
4712 buf
= kzalloc(PATH_MAX
+ sizeof(u64
), GFP_KERNEL
);
4714 name
= strncpy(tmp
, "//enomem", sizeof(tmp
));
4717 name
= d_path(&file
->f_path
, buf
, PATH_MAX
);
4719 name
= strncpy(tmp
, "//toolong", sizeof(tmp
));
4723 if (arch_vma_name(mmap_event
->vma
)) {
4724 name
= strncpy(tmp
, arch_vma_name(mmap_event
->vma
),
4730 name
= strncpy(tmp
, "[vdso]", sizeof(tmp
));
4732 } else if (vma
->vm_start
<= vma
->vm_mm
->start_brk
&&
4733 vma
->vm_end
>= vma
->vm_mm
->brk
) {
4734 name
= strncpy(tmp
, "[heap]", sizeof(tmp
));
4736 } else if (vma
->vm_start
<= vma
->vm_mm
->start_stack
&&
4737 vma
->vm_end
>= vma
->vm_mm
->start_stack
) {
4738 name
= strncpy(tmp
, "[stack]", sizeof(tmp
));
4742 name
= strncpy(tmp
, "//anon", sizeof(tmp
));
4747 size
= ALIGN(strlen(name
)+1, sizeof(u64
));
4749 mmap_event
->file_name
= name
;
4750 mmap_event
->file_size
= size
;
4752 mmap_event
->event_id
.header
.size
= sizeof(mmap_event
->event_id
) + size
;
4755 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
4756 cpuctx
= get_cpu_ptr(pmu
->pmu_cpu_context
);
4757 if (cpuctx
->active_pmu
!= pmu
)
4759 perf_event_mmap_ctx(&cpuctx
->ctx
, mmap_event
,
4760 vma
->vm_flags
& VM_EXEC
);
4762 ctxn
= pmu
->task_ctx_nr
;
4766 ctx
= rcu_dereference(current
->perf_event_ctxp
[ctxn
]);
4768 perf_event_mmap_ctx(ctx
, mmap_event
,
4769 vma
->vm_flags
& VM_EXEC
);
4772 put_cpu_ptr(pmu
->pmu_cpu_context
);
4779 void perf_event_mmap(struct vm_area_struct
*vma
)
4781 struct perf_mmap_event mmap_event
;
4783 if (!atomic_read(&nr_mmap_events
))
4786 mmap_event
= (struct perf_mmap_event
){
4792 .type
= PERF_RECORD_MMAP
,
4793 .misc
= PERF_RECORD_MISC_USER
,
4798 .start
= vma
->vm_start
,
4799 .len
= vma
->vm_end
- vma
->vm_start
,
4800 .pgoff
= (u64
)vma
->vm_pgoff
<< PAGE_SHIFT
,
4804 perf_event_mmap_event(&mmap_event
);
4808 * IRQ throttle logging
4811 static void perf_log_throttle(struct perf_event
*event
, int enable
)
4813 struct perf_output_handle handle
;
4814 struct perf_sample_data sample
;
4818 struct perf_event_header header
;
4822 } throttle_event
= {
4824 .type
= PERF_RECORD_THROTTLE
,
4826 .size
= sizeof(throttle_event
),
4828 .time
= perf_clock(),
4829 .id
= primary_event_id(event
),
4830 .stream_id
= event
->id
,
4834 throttle_event
.header
.type
= PERF_RECORD_UNTHROTTLE
;
4836 perf_event_header__init_id(&throttle_event
.header
, &sample
, event
);
4838 ret
= perf_output_begin(&handle
, event
,
4839 throttle_event
.header
.size
);
4843 perf_output_put(&handle
, throttle_event
);
4844 perf_event__output_id_sample(event
, &handle
, &sample
);
4845 perf_output_end(&handle
);
4849 * Generic event overflow handling, sampling.
4852 static int __perf_event_overflow(struct perf_event
*event
,
4853 int throttle
, struct perf_sample_data
*data
,
4854 struct pt_regs
*regs
)
4856 int events
= atomic_read(&event
->event_limit
);
4857 struct hw_perf_event
*hwc
= &event
->hw
;
4862 * Non-sampling counters might still use the PMI to fold short
4863 * hardware counters, ignore those.
4865 if (unlikely(!is_sampling_event(event
)))
4868 seq
= __this_cpu_read(perf_throttled_seq
);
4869 if (seq
!= hwc
->interrupts_seq
) {
4870 hwc
->interrupts_seq
= seq
;
4871 hwc
->interrupts
= 1;
4874 if (unlikely(throttle
4875 && hwc
->interrupts
>= max_samples_per_tick
)) {
4876 __this_cpu_inc(perf_throttled_count
);
4877 hwc
->interrupts
= MAX_INTERRUPTS
;
4878 perf_log_throttle(event
, 0);
4883 if (event
->attr
.freq
) {
4884 u64 now
= perf_clock();
4885 s64 delta
= now
- hwc
->freq_time_stamp
;
4887 hwc
->freq_time_stamp
= now
;
4889 if (delta
> 0 && delta
< 2*TICK_NSEC
)
4890 perf_adjust_period(event
, delta
, hwc
->last_period
, true);
4894 * XXX event_limit might not quite work as expected on inherited
4898 event
->pending_kill
= POLL_IN
;
4899 if (events
&& atomic_dec_and_test(&event
->event_limit
)) {
4901 event
->pending_kill
= POLL_HUP
;
4902 event
->pending_disable
= 1;
4903 irq_work_queue(&event
->pending
);
4906 if (event
->overflow_handler
)
4907 event
->overflow_handler(event
, data
, regs
);
4909 perf_event_output(event
, data
, regs
);
4911 if (event
->fasync
&& event
->pending_kill
) {
4912 event
->pending_wakeup
= 1;
4913 irq_work_queue(&event
->pending
);
4919 int perf_event_overflow(struct perf_event
*event
,
4920 struct perf_sample_data
*data
,
4921 struct pt_regs
*regs
)
4923 return __perf_event_overflow(event
, 1, data
, regs
);
4927 * Generic software event infrastructure
4930 struct swevent_htable
{
4931 struct swevent_hlist
*swevent_hlist
;
4932 struct mutex hlist_mutex
;
4935 /* Recursion avoidance in each contexts */
4936 int recursion
[PERF_NR_CONTEXTS
];
4939 static DEFINE_PER_CPU(struct swevent_htable
, swevent_htable
);
4942 * We directly increment event->count and keep a second value in
4943 * event->hw.period_left to count intervals. This period event
4944 * is kept in the range [-sample_period, 0] so that we can use the
4948 static u64
perf_swevent_set_period(struct perf_event
*event
)
4950 struct hw_perf_event
*hwc
= &event
->hw
;
4951 u64 period
= hwc
->last_period
;
4955 hwc
->last_period
= hwc
->sample_period
;
4958 old
= val
= local64_read(&hwc
->period_left
);
4962 nr
= div64_u64(period
+ val
, period
);
4963 offset
= nr
* period
;
4965 if (local64_cmpxchg(&hwc
->period_left
, old
, val
) != old
)
4971 static void perf_swevent_overflow(struct perf_event
*event
, u64 overflow
,
4972 struct perf_sample_data
*data
,
4973 struct pt_regs
*regs
)
4975 struct hw_perf_event
*hwc
= &event
->hw
;
4979 overflow
= perf_swevent_set_period(event
);
4981 if (hwc
->interrupts
== MAX_INTERRUPTS
)
4984 for (; overflow
; overflow
--) {
4985 if (__perf_event_overflow(event
, throttle
,
4988 * We inhibit the overflow from happening when
4989 * hwc->interrupts == MAX_INTERRUPTS.
4997 static void perf_swevent_event(struct perf_event
*event
, u64 nr
,
4998 struct perf_sample_data
*data
,
4999 struct pt_regs
*regs
)
5001 struct hw_perf_event
*hwc
= &event
->hw
;
5003 local64_add(nr
, &event
->count
);
5008 if (!is_sampling_event(event
))
5011 if ((event
->attr
.sample_type
& PERF_SAMPLE_PERIOD
) && !event
->attr
.freq
) {
5013 return perf_swevent_overflow(event
, 1, data
, regs
);
5015 data
->period
= event
->hw
.last_period
;
5017 if (nr
== 1 && hwc
->sample_period
== 1 && !event
->attr
.freq
)
5018 return perf_swevent_overflow(event
, 1, data
, regs
);
5020 if (local64_add_negative(nr
, &hwc
->period_left
))
5023 perf_swevent_overflow(event
, 0, data
, regs
);
5026 static int perf_exclude_event(struct perf_event
*event
,
5027 struct pt_regs
*regs
)
5029 if (event
->hw
.state
& PERF_HES_STOPPED
)
5033 if (event
->attr
.exclude_user
&& user_mode(regs
))
5036 if (event
->attr
.exclude_kernel
&& !user_mode(regs
))
5043 static int perf_swevent_match(struct perf_event
*event
,
5044 enum perf_type_id type
,
5046 struct perf_sample_data
*data
,
5047 struct pt_regs
*regs
)
5049 if (event
->attr
.type
!= type
)
5052 if (event
->attr
.config
!= event_id
)
5055 if (perf_exclude_event(event
, regs
))
5061 static inline u64
swevent_hash(u64 type
, u32 event_id
)
5063 u64 val
= event_id
| (type
<< 32);
5065 return hash_64(val
, SWEVENT_HLIST_BITS
);
5068 static inline struct hlist_head
*
5069 __find_swevent_head(struct swevent_hlist
*hlist
, u64 type
, u32 event_id
)
5071 u64 hash
= swevent_hash(type
, event_id
);
5073 return &hlist
->heads
[hash
];
5076 /* For the read side: events when they trigger */
5077 static inline struct hlist_head
*
5078 find_swevent_head_rcu(struct swevent_htable
*swhash
, u64 type
, u32 event_id
)
5080 struct swevent_hlist
*hlist
;
5082 hlist
= rcu_dereference(swhash
->swevent_hlist
);
5086 return __find_swevent_head(hlist
, type
, event_id
);
5089 /* For the event head insertion and removal in the hlist */
5090 static inline struct hlist_head
*
5091 find_swevent_head(struct swevent_htable
*swhash
, struct perf_event
*event
)
5093 struct swevent_hlist
*hlist
;
5094 u32 event_id
= event
->attr
.config
;
5095 u64 type
= event
->attr
.type
;
5098 * Event scheduling is always serialized against hlist allocation
5099 * and release. Which makes the protected version suitable here.
5100 * The context lock guarantees that.
5102 hlist
= rcu_dereference_protected(swhash
->swevent_hlist
,
5103 lockdep_is_held(&event
->ctx
->lock
));
5107 return __find_swevent_head(hlist
, type
, event_id
);
5110 static void do_perf_sw_event(enum perf_type_id type
, u32 event_id
,
5112 struct perf_sample_data
*data
,
5113 struct pt_regs
*regs
)
5115 struct swevent_htable
*swhash
= &__get_cpu_var(swevent_htable
);
5116 struct perf_event
*event
;
5117 struct hlist_node
*node
;
5118 struct hlist_head
*head
;
5121 head
= find_swevent_head_rcu(swhash
, type
, event_id
);
5125 hlist_for_each_entry_rcu(event
, node
, head
, hlist_entry
) {
5126 if (perf_swevent_match(event
, type
, event_id
, data
, regs
))
5127 perf_swevent_event(event
, nr
, data
, regs
);
5133 int perf_swevent_get_recursion_context(void)
5135 struct swevent_htable
*swhash
= &__get_cpu_var(swevent_htable
);
5137 return get_recursion_context(swhash
->recursion
);
5139 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context
);
5141 inline void perf_swevent_put_recursion_context(int rctx
)
5143 struct swevent_htable
*swhash
= &__get_cpu_var(swevent_htable
);
5145 put_recursion_context(swhash
->recursion
, rctx
);
5148 void __perf_sw_event(u32 event_id
, u64 nr
, struct pt_regs
*regs
, u64 addr
)
5150 struct perf_sample_data data
;
5153 preempt_disable_notrace();
5154 rctx
= perf_swevent_get_recursion_context();
5158 perf_sample_data_init(&data
, addr
, 0);
5160 do_perf_sw_event(PERF_TYPE_SOFTWARE
, event_id
, nr
, &data
, regs
);
5162 perf_swevent_put_recursion_context(rctx
);
5163 preempt_enable_notrace();
5166 static void perf_swevent_read(struct perf_event
*event
)
5170 static int perf_swevent_add(struct perf_event
*event
, int flags
)
5172 struct swevent_htable
*swhash
= &__get_cpu_var(swevent_htable
);
5173 struct hw_perf_event
*hwc
= &event
->hw
;
5174 struct hlist_head
*head
;
5176 if (is_sampling_event(event
)) {
5177 hwc
->last_period
= hwc
->sample_period
;
5178 perf_swevent_set_period(event
);
5181 hwc
->state
= !(flags
& PERF_EF_START
);
5183 head
= find_swevent_head(swhash
, event
);
5184 if (WARN_ON_ONCE(!head
))
5187 hlist_add_head_rcu(&event
->hlist_entry
, head
);
5192 static void perf_swevent_del(struct perf_event
*event
, int flags
)
5194 hlist_del_rcu(&event
->hlist_entry
);
5197 static void perf_swevent_start(struct perf_event
*event
, int flags
)
5199 event
->hw
.state
= 0;
5202 static void perf_swevent_stop(struct perf_event
*event
, int flags
)
5204 event
->hw
.state
= PERF_HES_STOPPED
;
5207 /* Deref the hlist from the update side */
5208 static inline struct swevent_hlist
*
5209 swevent_hlist_deref(struct swevent_htable
*swhash
)
5211 return rcu_dereference_protected(swhash
->swevent_hlist
,
5212 lockdep_is_held(&swhash
->hlist_mutex
));
5215 static void swevent_hlist_release(struct swevent_htable
*swhash
)
5217 struct swevent_hlist
*hlist
= swevent_hlist_deref(swhash
);
5222 rcu_assign_pointer(swhash
->swevent_hlist
, NULL
);
5223 kfree_rcu(hlist
, rcu_head
);
5226 static void swevent_hlist_put_cpu(struct perf_event
*event
, int cpu
)
5228 struct swevent_htable
*swhash
= &per_cpu(swevent_htable
, cpu
);
5230 mutex_lock(&swhash
->hlist_mutex
);
5232 if (!--swhash
->hlist_refcount
)
5233 swevent_hlist_release(swhash
);
5235 mutex_unlock(&swhash
->hlist_mutex
);
5238 static void swevent_hlist_put(struct perf_event
*event
)
5242 if (event
->cpu
!= -1) {
5243 swevent_hlist_put_cpu(event
, event
->cpu
);
5247 for_each_possible_cpu(cpu
)
5248 swevent_hlist_put_cpu(event
, cpu
);
5251 static int swevent_hlist_get_cpu(struct perf_event
*event
, int cpu
)
5253 struct swevent_htable
*swhash
= &per_cpu(swevent_htable
, cpu
);
5256 mutex_lock(&swhash
->hlist_mutex
);
5258 if (!swevent_hlist_deref(swhash
) && cpu_online(cpu
)) {
5259 struct swevent_hlist
*hlist
;
5261 hlist
= kzalloc(sizeof(*hlist
), GFP_KERNEL
);
5266 rcu_assign_pointer(swhash
->swevent_hlist
, hlist
);
5268 swhash
->hlist_refcount
++;
5270 mutex_unlock(&swhash
->hlist_mutex
);
5275 static int swevent_hlist_get(struct perf_event
*event
)
5278 int cpu
, failed_cpu
;
5280 if (event
->cpu
!= -1)
5281 return swevent_hlist_get_cpu(event
, event
->cpu
);
5284 for_each_possible_cpu(cpu
) {
5285 err
= swevent_hlist_get_cpu(event
, cpu
);
5295 for_each_possible_cpu(cpu
) {
5296 if (cpu
== failed_cpu
)
5298 swevent_hlist_put_cpu(event
, cpu
);
5305 struct static_key perf_swevent_enabled
[PERF_COUNT_SW_MAX
];
5307 static void sw_perf_event_destroy(struct perf_event
*event
)
5309 u64 event_id
= event
->attr
.config
;
5311 WARN_ON(event
->parent
);
5313 static_key_slow_dec(&perf_swevent_enabled
[event_id
]);
5314 swevent_hlist_put(event
);
5317 static int perf_swevent_init(struct perf_event
*event
)
5319 int event_id
= event
->attr
.config
;
5321 if (event
->attr
.type
!= PERF_TYPE_SOFTWARE
)
5325 * no branch sampling for software events
5327 if (has_branch_stack(event
))
5331 case PERF_COUNT_SW_CPU_CLOCK
:
5332 case PERF_COUNT_SW_TASK_CLOCK
:
5339 if (event_id
>= PERF_COUNT_SW_MAX
)
5342 if (!event
->parent
) {
5345 err
= swevent_hlist_get(event
);
5349 static_key_slow_inc(&perf_swevent_enabled
[event_id
]);
5350 event
->destroy
= sw_perf_event_destroy
;
5356 static int perf_swevent_event_idx(struct perf_event
*event
)
5361 static struct pmu perf_swevent
= {
5362 .task_ctx_nr
= perf_sw_context
,
5364 .event_init
= perf_swevent_init
,
5365 .add
= perf_swevent_add
,
5366 .del
= perf_swevent_del
,
5367 .start
= perf_swevent_start
,
5368 .stop
= perf_swevent_stop
,
5369 .read
= perf_swevent_read
,
5371 .event_idx
= perf_swevent_event_idx
,
5374 #ifdef CONFIG_EVENT_TRACING
5376 static int perf_tp_filter_match(struct perf_event
*event
,
5377 struct perf_sample_data
*data
)
5379 void *record
= data
->raw
->data
;
5381 if (likely(!event
->filter
) || filter_match_preds(event
->filter
, record
))
5386 static int perf_tp_event_match(struct perf_event
*event
,
5387 struct perf_sample_data
*data
,
5388 struct pt_regs
*regs
)
5390 if (event
->hw
.state
& PERF_HES_STOPPED
)
5393 * All tracepoints are from kernel-space.
5395 if (event
->attr
.exclude_kernel
)
5398 if (!perf_tp_filter_match(event
, data
))
5404 void perf_tp_event(u64 addr
, u64 count
, void *record
, int entry_size
,
5405 struct pt_regs
*regs
, struct hlist_head
*head
, int rctx
,
5406 struct task_struct
*task
)
5408 struct perf_sample_data data
;
5409 struct perf_event
*event
;
5410 struct hlist_node
*node
;
5412 struct perf_raw_record raw
= {
5417 perf_sample_data_init(&data
, addr
, 0);
5420 hlist_for_each_entry_rcu(event
, node
, head
, hlist_entry
) {
5421 if (perf_tp_event_match(event
, &data
, regs
))
5422 perf_swevent_event(event
, count
, &data
, regs
);
5426 * If we got specified a target task, also iterate its context and
5427 * deliver this event there too.
5429 if (task
&& task
!= current
) {
5430 struct perf_event_context
*ctx
;
5431 struct trace_entry
*entry
= record
;
5434 ctx
= rcu_dereference(task
->perf_event_ctxp
[perf_sw_context
]);
5438 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
5439 if (event
->attr
.type
!= PERF_TYPE_TRACEPOINT
)
5441 if (event
->attr
.config
!= entry
->type
)
5443 if (perf_tp_event_match(event
, &data
, regs
))
5444 perf_swevent_event(event
, count
, &data
, regs
);
5450 perf_swevent_put_recursion_context(rctx
);
5452 EXPORT_SYMBOL_GPL(perf_tp_event
);
5454 static void tp_perf_event_destroy(struct perf_event
*event
)
5456 perf_trace_destroy(event
);
5459 static int perf_tp_event_init(struct perf_event
*event
)
5463 if (event
->attr
.type
!= PERF_TYPE_TRACEPOINT
)
5467 * no branch sampling for tracepoint events
5469 if (has_branch_stack(event
))
5472 err
= perf_trace_init(event
);
5476 event
->destroy
= tp_perf_event_destroy
;
5481 static struct pmu perf_tracepoint
= {
5482 .task_ctx_nr
= perf_sw_context
,
5484 .event_init
= perf_tp_event_init
,
5485 .add
= perf_trace_add
,
5486 .del
= perf_trace_del
,
5487 .start
= perf_swevent_start
,
5488 .stop
= perf_swevent_stop
,
5489 .read
= perf_swevent_read
,
5491 .event_idx
= perf_swevent_event_idx
,
5494 static inline void perf_tp_register(void)
5496 perf_pmu_register(&perf_tracepoint
, "tracepoint", PERF_TYPE_TRACEPOINT
);
5499 static int perf_event_set_filter(struct perf_event
*event
, void __user
*arg
)
5504 if (event
->attr
.type
!= PERF_TYPE_TRACEPOINT
)
5507 filter_str
= strndup_user(arg
, PAGE_SIZE
);
5508 if (IS_ERR(filter_str
))
5509 return PTR_ERR(filter_str
);
5511 ret
= ftrace_profile_set_filter(event
, event
->attr
.config
, filter_str
);
5517 static void perf_event_free_filter(struct perf_event
*event
)
5519 ftrace_profile_free_filter(event
);
5524 static inline void perf_tp_register(void)
5528 static int perf_event_set_filter(struct perf_event
*event
, void __user
*arg
)
5533 static void perf_event_free_filter(struct perf_event
*event
)
5537 #endif /* CONFIG_EVENT_TRACING */
5539 #ifdef CONFIG_HAVE_HW_BREAKPOINT
5540 void perf_bp_event(struct perf_event
*bp
, void *data
)
5542 struct perf_sample_data sample
;
5543 struct pt_regs
*regs
= data
;
5545 perf_sample_data_init(&sample
, bp
->attr
.bp_addr
, 0);
5547 if (!bp
->hw
.state
&& !perf_exclude_event(bp
, regs
))
5548 perf_swevent_event(bp
, 1, &sample
, regs
);
5553 * hrtimer based swevent callback
5556 static enum hrtimer_restart
perf_swevent_hrtimer(struct hrtimer
*hrtimer
)
5558 enum hrtimer_restart ret
= HRTIMER_RESTART
;
5559 struct perf_sample_data data
;
5560 struct pt_regs
*regs
;
5561 struct perf_event
*event
;
5564 event
= container_of(hrtimer
, struct perf_event
, hw
.hrtimer
);
5566 if (event
->state
!= PERF_EVENT_STATE_ACTIVE
)
5567 return HRTIMER_NORESTART
;
5569 event
->pmu
->read(event
);
5571 perf_sample_data_init(&data
, 0, event
->hw
.last_period
);
5572 regs
= get_irq_regs();
5574 if (regs
&& !perf_exclude_event(event
, regs
)) {
5575 if (!(event
->attr
.exclude_idle
&& is_idle_task(current
)))
5576 if (__perf_event_overflow(event
, 1, &data
, regs
))
5577 ret
= HRTIMER_NORESTART
;
5580 period
= max_t(u64
, 10000, event
->hw
.sample_period
);
5581 hrtimer_forward_now(hrtimer
, ns_to_ktime(period
));
5586 static void perf_swevent_start_hrtimer(struct perf_event
*event
)
5588 struct hw_perf_event
*hwc
= &event
->hw
;
5591 if (!is_sampling_event(event
))
5594 period
= local64_read(&hwc
->period_left
);
5599 local64_set(&hwc
->period_left
, 0);
5601 period
= max_t(u64
, 10000, hwc
->sample_period
);
5603 __hrtimer_start_range_ns(&hwc
->hrtimer
,
5604 ns_to_ktime(period
), 0,
5605 HRTIMER_MODE_REL_PINNED
, 0);
5608 static void perf_swevent_cancel_hrtimer(struct perf_event
*event
)
5610 struct hw_perf_event
*hwc
= &event
->hw
;
5612 if (is_sampling_event(event
)) {
5613 ktime_t remaining
= hrtimer_get_remaining(&hwc
->hrtimer
);
5614 local64_set(&hwc
->period_left
, ktime_to_ns(remaining
));
5616 hrtimer_cancel(&hwc
->hrtimer
);
5620 static void perf_swevent_init_hrtimer(struct perf_event
*event
)
5622 struct hw_perf_event
*hwc
= &event
->hw
;
5624 if (!is_sampling_event(event
))
5627 hrtimer_init(&hwc
->hrtimer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
5628 hwc
->hrtimer
.function
= perf_swevent_hrtimer
;
5631 * Since hrtimers have a fixed rate, we can do a static freq->period
5632 * mapping and avoid the whole period adjust feedback stuff.
5634 if (event
->attr
.freq
) {
5635 long freq
= event
->attr
.sample_freq
;
5637 event
->attr
.sample_period
= NSEC_PER_SEC
/ freq
;
5638 hwc
->sample_period
= event
->attr
.sample_period
;
5639 local64_set(&hwc
->period_left
, hwc
->sample_period
);
5640 event
->attr
.freq
= 0;
5645 * Software event: cpu wall time clock
5648 static void cpu_clock_event_update(struct perf_event
*event
)
5653 now
= local_clock();
5654 prev
= local64_xchg(&event
->hw
.prev_count
, now
);
5655 local64_add(now
- prev
, &event
->count
);
5658 static void cpu_clock_event_start(struct perf_event
*event
, int flags
)
5660 local64_set(&event
->hw
.prev_count
, local_clock());
5661 perf_swevent_start_hrtimer(event
);
5664 static void cpu_clock_event_stop(struct perf_event
*event
, int flags
)
5666 perf_swevent_cancel_hrtimer(event
);
5667 cpu_clock_event_update(event
);
5670 static int cpu_clock_event_add(struct perf_event
*event
, int flags
)
5672 if (flags
& PERF_EF_START
)
5673 cpu_clock_event_start(event
, flags
);
5678 static void cpu_clock_event_del(struct perf_event
*event
, int flags
)
5680 cpu_clock_event_stop(event
, flags
);
5683 static void cpu_clock_event_read(struct perf_event
*event
)
5685 cpu_clock_event_update(event
);
5688 static int cpu_clock_event_init(struct perf_event
*event
)
5690 if (event
->attr
.type
!= PERF_TYPE_SOFTWARE
)
5693 if (event
->attr
.config
!= PERF_COUNT_SW_CPU_CLOCK
)
5697 * no branch sampling for software events
5699 if (has_branch_stack(event
))
5702 perf_swevent_init_hrtimer(event
);
5707 static struct pmu perf_cpu_clock
= {
5708 .task_ctx_nr
= perf_sw_context
,
5710 .event_init
= cpu_clock_event_init
,
5711 .add
= cpu_clock_event_add
,
5712 .del
= cpu_clock_event_del
,
5713 .start
= cpu_clock_event_start
,
5714 .stop
= cpu_clock_event_stop
,
5715 .read
= cpu_clock_event_read
,
5717 .event_idx
= perf_swevent_event_idx
,
5721 * Software event: task time clock
5724 static void task_clock_event_update(struct perf_event
*event
, u64 now
)
5729 prev
= local64_xchg(&event
->hw
.prev_count
, now
);
5731 local64_add(delta
, &event
->count
);
5734 static void task_clock_event_start(struct perf_event
*event
, int flags
)
5736 local64_set(&event
->hw
.prev_count
, event
->ctx
->time
);
5737 perf_swevent_start_hrtimer(event
);
5740 static void task_clock_event_stop(struct perf_event
*event
, int flags
)
5742 perf_swevent_cancel_hrtimer(event
);
5743 task_clock_event_update(event
, event
->ctx
->time
);
5746 static int task_clock_event_add(struct perf_event
*event
, int flags
)
5748 if (flags
& PERF_EF_START
)
5749 task_clock_event_start(event
, flags
);
5754 static void task_clock_event_del(struct perf_event
*event
, int flags
)
5756 task_clock_event_stop(event
, PERF_EF_UPDATE
);
5759 static void task_clock_event_read(struct perf_event
*event
)
5761 u64 now
= perf_clock();
5762 u64 delta
= now
- event
->ctx
->timestamp
;
5763 u64 time
= event
->ctx
->time
+ delta
;
5765 task_clock_event_update(event
, time
);
5768 static int task_clock_event_init(struct perf_event
*event
)
5770 if (event
->attr
.type
!= PERF_TYPE_SOFTWARE
)
5773 if (event
->attr
.config
!= PERF_COUNT_SW_TASK_CLOCK
)
5777 * no branch sampling for software events
5779 if (has_branch_stack(event
))
5782 perf_swevent_init_hrtimer(event
);
5787 static struct pmu perf_task_clock
= {
5788 .task_ctx_nr
= perf_sw_context
,
5790 .event_init
= task_clock_event_init
,
5791 .add
= task_clock_event_add
,
5792 .del
= task_clock_event_del
,
5793 .start
= task_clock_event_start
,
5794 .stop
= task_clock_event_stop
,
5795 .read
= task_clock_event_read
,
5797 .event_idx
= perf_swevent_event_idx
,
5800 static void perf_pmu_nop_void(struct pmu
*pmu
)
5804 static int perf_pmu_nop_int(struct pmu
*pmu
)
5809 static void perf_pmu_start_txn(struct pmu
*pmu
)
5811 perf_pmu_disable(pmu
);
5814 static int perf_pmu_commit_txn(struct pmu
*pmu
)
5816 perf_pmu_enable(pmu
);
5820 static void perf_pmu_cancel_txn(struct pmu
*pmu
)
5822 perf_pmu_enable(pmu
);
5825 static int perf_event_idx_default(struct perf_event
*event
)
5827 return event
->hw
.idx
+ 1;
5831 * Ensures all contexts with the same task_ctx_nr have the same
5832 * pmu_cpu_context too.
5834 static void *find_pmu_context(int ctxn
)
5841 list_for_each_entry(pmu
, &pmus
, entry
) {
5842 if (pmu
->task_ctx_nr
== ctxn
)
5843 return pmu
->pmu_cpu_context
;
5849 static void update_pmu_context(struct pmu
*pmu
, struct pmu
*old_pmu
)
5853 for_each_possible_cpu(cpu
) {
5854 struct perf_cpu_context
*cpuctx
;
5856 cpuctx
= per_cpu_ptr(pmu
->pmu_cpu_context
, cpu
);
5858 if (cpuctx
->active_pmu
== old_pmu
)
5859 cpuctx
->active_pmu
= pmu
;
5863 static void free_pmu_context(struct pmu
*pmu
)
5867 mutex_lock(&pmus_lock
);
5869 * Like a real lame refcount.
5871 list_for_each_entry(i
, &pmus
, entry
) {
5872 if (i
->pmu_cpu_context
== pmu
->pmu_cpu_context
) {
5873 update_pmu_context(i
, pmu
);
5878 free_percpu(pmu
->pmu_cpu_context
);
5880 mutex_unlock(&pmus_lock
);
5882 static struct idr pmu_idr
;
5885 type_show(struct device
*dev
, struct device_attribute
*attr
, char *page
)
5887 struct pmu
*pmu
= dev_get_drvdata(dev
);
5889 return snprintf(page
, PAGE_SIZE
-1, "%d\n", pmu
->type
);
5892 static struct device_attribute pmu_dev_attrs
[] = {
5897 static int pmu_bus_running
;
5898 static struct bus_type pmu_bus
= {
5899 .name
= "event_source",
5900 .dev_attrs
= pmu_dev_attrs
,
5903 static void pmu_dev_release(struct device
*dev
)
5908 static int pmu_dev_alloc(struct pmu
*pmu
)
5912 pmu
->dev
= kzalloc(sizeof(struct device
), GFP_KERNEL
);
5916 pmu
->dev
->groups
= pmu
->attr_groups
;
5917 device_initialize(pmu
->dev
);
5918 ret
= dev_set_name(pmu
->dev
, "%s", pmu
->name
);
5922 dev_set_drvdata(pmu
->dev
, pmu
);
5923 pmu
->dev
->bus
= &pmu_bus
;
5924 pmu
->dev
->release
= pmu_dev_release
;
5925 ret
= device_add(pmu
->dev
);
5933 put_device(pmu
->dev
);
5937 static struct lock_class_key cpuctx_mutex
;
5938 static struct lock_class_key cpuctx_lock
;
5940 int perf_pmu_register(struct pmu
*pmu
, char *name
, int type
)
5944 mutex_lock(&pmus_lock
);
5946 pmu
->pmu_disable_count
= alloc_percpu(int);
5947 if (!pmu
->pmu_disable_count
)
5956 int err
= idr_pre_get(&pmu_idr
, GFP_KERNEL
);
5960 err
= idr_get_new_above(&pmu_idr
, pmu
, PERF_TYPE_MAX
, &type
);
5968 if (pmu_bus_running
) {
5969 ret
= pmu_dev_alloc(pmu
);
5975 pmu
->pmu_cpu_context
= find_pmu_context(pmu
->task_ctx_nr
);
5976 if (pmu
->pmu_cpu_context
)
5977 goto got_cpu_context
;
5979 pmu
->pmu_cpu_context
= alloc_percpu(struct perf_cpu_context
);
5980 if (!pmu
->pmu_cpu_context
)
5983 for_each_possible_cpu(cpu
) {
5984 struct perf_cpu_context
*cpuctx
;
5986 cpuctx
= per_cpu_ptr(pmu
->pmu_cpu_context
, cpu
);
5987 __perf_event_init_context(&cpuctx
->ctx
);
5988 lockdep_set_class(&cpuctx
->ctx
.mutex
, &cpuctx_mutex
);
5989 lockdep_set_class(&cpuctx
->ctx
.lock
, &cpuctx_lock
);
5990 cpuctx
->ctx
.type
= cpu_context
;
5991 cpuctx
->ctx
.pmu
= pmu
;
5992 cpuctx
->jiffies_interval
= 1;
5993 INIT_LIST_HEAD(&cpuctx
->rotation_list
);
5994 cpuctx
->active_pmu
= pmu
;
5998 if (!pmu
->start_txn
) {
5999 if (pmu
->pmu_enable
) {
6001 * If we have pmu_enable/pmu_disable calls, install
6002 * transaction stubs that use that to try and batch
6003 * hardware accesses.
6005 pmu
->start_txn
= perf_pmu_start_txn
;
6006 pmu
->commit_txn
= perf_pmu_commit_txn
;
6007 pmu
->cancel_txn
= perf_pmu_cancel_txn
;
6009 pmu
->start_txn
= perf_pmu_nop_void
;
6010 pmu
->commit_txn
= perf_pmu_nop_int
;
6011 pmu
->cancel_txn
= perf_pmu_nop_void
;
6015 if (!pmu
->pmu_enable
) {
6016 pmu
->pmu_enable
= perf_pmu_nop_void
;
6017 pmu
->pmu_disable
= perf_pmu_nop_void
;
6020 if (!pmu
->event_idx
)
6021 pmu
->event_idx
= perf_event_idx_default
;
6023 list_add_rcu(&pmu
->entry
, &pmus
);
6026 mutex_unlock(&pmus_lock
);
6031 device_del(pmu
->dev
);
6032 put_device(pmu
->dev
);
6035 if (pmu
->type
>= PERF_TYPE_MAX
)
6036 idr_remove(&pmu_idr
, pmu
->type
);
6039 free_percpu(pmu
->pmu_disable_count
);
6043 void perf_pmu_unregister(struct pmu
*pmu
)
6045 mutex_lock(&pmus_lock
);
6046 list_del_rcu(&pmu
->entry
);
6047 mutex_unlock(&pmus_lock
);
6050 * We dereference the pmu list under both SRCU and regular RCU, so
6051 * synchronize against both of those.
6053 synchronize_srcu(&pmus_srcu
);
6056 free_percpu(pmu
->pmu_disable_count
);
6057 if (pmu
->type
>= PERF_TYPE_MAX
)
6058 idr_remove(&pmu_idr
, pmu
->type
);
6059 device_del(pmu
->dev
);
6060 put_device(pmu
->dev
);
6061 free_pmu_context(pmu
);
6064 struct pmu
*perf_init_event(struct perf_event
*event
)
6066 struct pmu
*pmu
= NULL
;
6070 idx
= srcu_read_lock(&pmus_srcu
);
6073 pmu
= idr_find(&pmu_idr
, event
->attr
.type
);
6077 ret
= pmu
->event_init(event
);
6083 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
6085 ret
= pmu
->event_init(event
);
6089 if (ret
!= -ENOENT
) {
6094 pmu
= ERR_PTR(-ENOENT
);
6096 srcu_read_unlock(&pmus_srcu
, idx
);
6102 * Allocate and initialize a event structure
6104 static struct perf_event
*
6105 perf_event_alloc(struct perf_event_attr
*attr
, int cpu
,
6106 struct task_struct
*task
,
6107 struct perf_event
*group_leader
,
6108 struct perf_event
*parent_event
,
6109 perf_overflow_handler_t overflow_handler
,
6113 struct perf_event
*event
;
6114 struct hw_perf_event
*hwc
;
6117 if ((unsigned)cpu
>= nr_cpu_ids
) {
6118 if (!task
|| cpu
!= -1)
6119 return ERR_PTR(-EINVAL
);
6122 event
= kzalloc(sizeof(*event
), GFP_KERNEL
);
6124 return ERR_PTR(-ENOMEM
);
6127 * Single events are their own group leaders, with an
6128 * empty sibling list:
6131 group_leader
= event
;
6133 mutex_init(&event
->child_mutex
);
6134 INIT_LIST_HEAD(&event
->child_list
);
6136 INIT_LIST_HEAD(&event
->group_entry
);
6137 INIT_LIST_HEAD(&event
->event_entry
);
6138 INIT_LIST_HEAD(&event
->sibling_list
);
6139 INIT_LIST_HEAD(&event
->rb_entry
);
6141 init_waitqueue_head(&event
->waitq
);
6142 init_irq_work(&event
->pending
, perf_pending_event
);
6144 mutex_init(&event
->mmap_mutex
);
6146 atomic_long_set(&event
->refcount
, 1);
6148 event
->attr
= *attr
;
6149 event
->group_leader
= group_leader
;
6153 event
->parent
= parent_event
;
6155 event
->ns
= get_pid_ns(current
->nsproxy
->pid_ns
);
6156 event
->id
= atomic64_inc_return(&perf_event_id
);
6158 event
->state
= PERF_EVENT_STATE_INACTIVE
;
6161 event
->attach_state
= PERF_ATTACH_TASK
;
6162 #ifdef CONFIG_HAVE_HW_BREAKPOINT
6164 * hw_breakpoint is a bit difficult here..
6166 if (attr
->type
== PERF_TYPE_BREAKPOINT
)
6167 event
->hw
.bp_target
= task
;
6171 if (!overflow_handler
&& parent_event
) {
6172 overflow_handler
= parent_event
->overflow_handler
;
6173 context
= parent_event
->overflow_handler_context
;
6176 event
->overflow_handler
= overflow_handler
;
6177 event
->overflow_handler_context
= context
;
6180 event
->state
= PERF_EVENT_STATE_OFF
;
6185 hwc
->sample_period
= attr
->sample_period
;
6186 if (attr
->freq
&& attr
->sample_freq
)
6187 hwc
->sample_period
= 1;
6188 hwc
->last_period
= hwc
->sample_period
;
6190 local64_set(&hwc
->period_left
, hwc
->sample_period
);
6193 * we currently do not support PERF_FORMAT_GROUP on inherited events
6195 if (attr
->inherit
&& (attr
->read_format
& PERF_FORMAT_GROUP
))
6198 pmu
= perf_init_event(event
);
6204 else if (IS_ERR(pmu
))
6209 put_pid_ns(event
->ns
);
6211 return ERR_PTR(err
);
6214 if (!event
->parent
) {
6215 if (event
->attach_state
& PERF_ATTACH_TASK
)
6216 static_key_slow_inc(&perf_sched_events
.key
);
6217 if (event
->attr
.mmap
|| event
->attr
.mmap_data
)
6218 atomic_inc(&nr_mmap_events
);
6219 if (event
->attr
.comm
)
6220 atomic_inc(&nr_comm_events
);
6221 if (event
->attr
.task
)
6222 atomic_inc(&nr_task_events
);
6223 if (event
->attr
.sample_type
& PERF_SAMPLE_CALLCHAIN
) {
6224 err
= get_callchain_buffers();
6227 return ERR_PTR(err
);
6230 if (has_branch_stack(event
)) {
6231 static_key_slow_inc(&perf_sched_events
.key
);
6232 if (!(event
->attach_state
& PERF_ATTACH_TASK
))
6233 atomic_inc(&per_cpu(perf_branch_stack_events
,
6241 static int perf_copy_attr(struct perf_event_attr __user
*uattr
,
6242 struct perf_event_attr
*attr
)
6247 if (!access_ok(VERIFY_WRITE
, uattr
, PERF_ATTR_SIZE_VER0
))
6251 * zero the full structure, so that a short copy will be nice.
6253 memset(attr
, 0, sizeof(*attr
));
6255 ret
= get_user(size
, &uattr
->size
);
6259 if (size
> PAGE_SIZE
) /* silly large */
6262 if (!size
) /* abi compat */
6263 size
= PERF_ATTR_SIZE_VER0
;
6265 if (size
< PERF_ATTR_SIZE_VER0
)
6269 * If we're handed a bigger struct than we know of,
6270 * ensure all the unknown bits are 0 - i.e. new
6271 * user-space does not rely on any kernel feature
6272 * extensions we dont know about yet.
6274 if (size
> sizeof(*attr
)) {
6275 unsigned char __user
*addr
;
6276 unsigned char __user
*end
;
6279 addr
= (void __user
*)uattr
+ sizeof(*attr
);
6280 end
= (void __user
*)uattr
+ size
;
6282 for (; addr
< end
; addr
++) {
6283 ret
= get_user(val
, addr
);
6289 size
= sizeof(*attr
);
6292 ret
= copy_from_user(attr
, uattr
, size
);
6296 if (attr
->__reserved_1
)
6299 if (attr
->sample_type
& ~(PERF_SAMPLE_MAX
-1))
6302 if (attr
->read_format
& ~(PERF_FORMAT_MAX
-1))
6305 if (attr
->sample_type
& PERF_SAMPLE_BRANCH_STACK
) {
6306 u64 mask
= attr
->branch_sample_type
;
6308 /* only using defined bits */
6309 if (mask
& ~(PERF_SAMPLE_BRANCH_MAX
-1))
6312 /* at least one branch bit must be set */
6313 if (!(mask
& ~PERF_SAMPLE_BRANCH_PLM_ALL
))
6316 /* kernel level capture: check permissions */
6317 if ((mask
& PERF_SAMPLE_BRANCH_PERM_PLM
)
6318 && perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN
))
6321 /* propagate priv level, when not set for branch */
6322 if (!(mask
& PERF_SAMPLE_BRANCH_PLM_ALL
)) {
6324 /* exclude_kernel checked on syscall entry */
6325 if (!attr
->exclude_kernel
)
6326 mask
|= PERF_SAMPLE_BRANCH_KERNEL
;
6328 if (!attr
->exclude_user
)
6329 mask
|= PERF_SAMPLE_BRANCH_USER
;
6331 if (!attr
->exclude_hv
)
6332 mask
|= PERF_SAMPLE_BRANCH_HV
;
6334 * adjust user setting (for HW filter setup)
6336 attr
->branch_sample_type
= mask
;
6340 if (attr
->sample_type
& PERF_SAMPLE_REGS_USER
) {
6341 ret
= perf_reg_validate(attr
->sample_regs_user
);
6346 if (attr
->sample_type
& PERF_SAMPLE_STACK_USER
) {
6347 if (!arch_perf_have_user_stack_dump())
6351 * We have __u32 type for the size, but so far
6352 * we can only use __u16 as maximum due to the
6353 * __u16 sample size limit.
6355 if (attr
->sample_stack_user
>= USHRT_MAX
)
6357 else if (!IS_ALIGNED(attr
->sample_stack_user
, sizeof(u64
)))
6365 put_user(sizeof(*attr
), &uattr
->size
);
6371 perf_event_set_output(struct perf_event
*event
, struct perf_event
*output_event
)
6373 struct ring_buffer
*rb
= NULL
, *old_rb
= NULL
;
6379 /* don't allow circular references */
6380 if (event
== output_event
)
6384 * Don't allow cross-cpu buffers
6386 if (output_event
->cpu
!= event
->cpu
)
6390 * If its not a per-cpu rb, it must be the same task.
6392 if (output_event
->cpu
== -1 && output_event
->ctx
!= event
->ctx
)
6396 mutex_lock(&event
->mmap_mutex
);
6397 /* Can't redirect output if we've got an active mmap() */
6398 if (atomic_read(&event
->mmap_count
))
6402 /* get the rb we want to redirect to */
6403 rb
= ring_buffer_get(output_event
);
6409 rcu_assign_pointer(event
->rb
, rb
);
6411 ring_buffer_detach(event
, old_rb
);
6414 mutex_unlock(&event
->mmap_mutex
);
6417 ring_buffer_put(old_rb
);
6423 * sys_perf_event_open - open a performance event, associate it to a task/cpu
6425 * @attr_uptr: event_id type attributes for monitoring/sampling
6428 * @group_fd: group leader event fd
6430 SYSCALL_DEFINE5(perf_event_open
,
6431 struct perf_event_attr __user
*, attr_uptr
,
6432 pid_t
, pid
, int, cpu
, int, group_fd
, unsigned long, flags
)
6434 struct perf_event
*group_leader
= NULL
, *output_event
= NULL
;
6435 struct perf_event
*event
, *sibling
;
6436 struct perf_event_attr attr
;
6437 struct perf_event_context
*ctx
;
6438 struct file
*event_file
= NULL
;
6439 struct fd group
= {NULL
, 0};
6440 struct task_struct
*task
= NULL
;
6446 /* for future expandability... */
6447 if (flags
& ~PERF_FLAG_ALL
)
6450 err
= perf_copy_attr(attr_uptr
, &attr
);
6454 if (!attr
.exclude_kernel
) {
6455 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN
))
6460 if (attr
.sample_freq
> sysctl_perf_event_sample_rate
)
6465 * In cgroup mode, the pid argument is used to pass the fd
6466 * opened to the cgroup directory in cgroupfs. The cpu argument
6467 * designates the cpu on which to monitor threads from that
6470 if ((flags
& PERF_FLAG_PID_CGROUP
) && (pid
== -1 || cpu
== -1))
6473 event_fd
= get_unused_fd();
6477 if (group_fd
!= -1) {
6478 err
= perf_fget_light(group_fd
, &group
);
6481 group_leader
= group
.file
->private_data
;
6482 if (flags
& PERF_FLAG_FD_OUTPUT
)
6483 output_event
= group_leader
;
6484 if (flags
& PERF_FLAG_FD_NO_GROUP
)
6485 group_leader
= NULL
;
6488 if (pid
!= -1 && !(flags
& PERF_FLAG_PID_CGROUP
)) {
6489 task
= find_lively_task_by_vpid(pid
);
6491 err
= PTR_ERR(task
);
6498 event
= perf_event_alloc(&attr
, cpu
, task
, group_leader
, NULL
,
6500 if (IS_ERR(event
)) {
6501 err
= PTR_ERR(event
);
6505 if (flags
& PERF_FLAG_PID_CGROUP
) {
6506 err
= perf_cgroup_connect(pid
, event
, &attr
, group_leader
);
6511 * - that has cgroup constraint on event->cpu
6512 * - that may need work on context switch
6514 atomic_inc(&per_cpu(perf_cgroup_events
, event
->cpu
));
6515 static_key_slow_inc(&perf_sched_events
.key
);
6519 * Special case software events and allow them to be part of
6520 * any hardware group.
6525 (is_software_event(event
) != is_software_event(group_leader
))) {
6526 if (is_software_event(event
)) {
6528 * If event and group_leader are not both a software
6529 * event, and event is, then group leader is not.
6531 * Allow the addition of software events to !software
6532 * groups, this is safe because software events never
6535 pmu
= group_leader
->pmu
;
6536 } else if (is_software_event(group_leader
) &&
6537 (group_leader
->group_flags
& PERF_GROUP_SOFTWARE
)) {
6539 * In case the group is a pure software group, and we
6540 * try to add a hardware event, move the whole group to
6541 * the hardware context.
6548 * Get the target context (task or percpu):
6550 ctx
= find_get_context(pmu
, task
, event
->cpu
);
6557 put_task_struct(task
);
6562 * Look up the group leader (we will attach this event to it):
6568 * Do not allow a recursive hierarchy (this new sibling
6569 * becoming part of another group-sibling):
6571 if (group_leader
->group_leader
!= group_leader
)
6574 * Do not allow to attach to a group in a different
6575 * task or CPU context:
6578 if (group_leader
->ctx
->type
!= ctx
->type
)
6581 if (group_leader
->ctx
!= ctx
)
6586 * Only a group leader can be exclusive or pinned
6588 if (attr
.exclusive
|| attr
.pinned
)
6593 err
= perf_event_set_output(event
, output_event
);
6598 event_file
= anon_inode_getfile("[perf_event]", &perf_fops
, event
, O_RDWR
);
6599 if (IS_ERR(event_file
)) {
6600 err
= PTR_ERR(event_file
);
6605 struct perf_event_context
*gctx
= group_leader
->ctx
;
6607 mutex_lock(&gctx
->mutex
);
6608 perf_remove_from_context(group_leader
);
6609 list_for_each_entry(sibling
, &group_leader
->sibling_list
,
6611 perf_remove_from_context(sibling
);
6614 mutex_unlock(&gctx
->mutex
);
6618 WARN_ON_ONCE(ctx
->parent_ctx
);
6619 mutex_lock(&ctx
->mutex
);
6623 perf_install_in_context(ctx
, group_leader
, event
->cpu
);
6625 list_for_each_entry(sibling
, &group_leader
->sibling_list
,
6627 perf_install_in_context(ctx
, sibling
, event
->cpu
);
6632 perf_install_in_context(ctx
, event
, event
->cpu
);
6634 perf_unpin_context(ctx
);
6635 mutex_unlock(&ctx
->mutex
);
6639 event
->owner
= current
;
6641 mutex_lock(¤t
->perf_event_mutex
);
6642 list_add_tail(&event
->owner_entry
, ¤t
->perf_event_list
);
6643 mutex_unlock(¤t
->perf_event_mutex
);
6646 * Precalculate sample_data sizes
6648 perf_event__header_size(event
);
6649 perf_event__id_header_size(event
);
6652 * Drop the reference on the group_event after placing the
6653 * new event on the sibling_list. This ensures destruction
6654 * of the group leader will find the pointer to itself in
6655 * perf_group_detach().
6658 fd_install(event_fd
, event_file
);
6662 perf_unpin_context(ctx
);
6669 put_task_struct(task
);
6673 put_unused_fd(event_fd
);
6678 * perf_event_create_kernel_counter
6680 * @attr: attributes of the counter to create
6681 * @cpu: cpu in which the counter is bound
6682 * @task: task to profile (NULL for percpu)
6685 perf_event_create_kernel_counter(struct perf_event_attr
*attr
, int cpu
,
6686 struct task_struct
*task
,
6687 perf_overflow_handler_t overflow_handler
,
6690 struct perf_event_context
*ctx
;
6691 struct perf_event
*event
;
6695 * Get the target context (task or percpu):
6698 event
= perf_event_alloc(attr
, cpu
, task
, NULL
, NULL
,
6699 overflow_handler
, context
);
6700 if (IS_ERR(event
)) {
6701 err
= PTR_ERR(event
);
6705 ctx
= find_get_context(event
->pmu
, task
, cpu
);
6711 WARN_ON_ONCE(ctx
->parent_ctx
);
6712 mutex_lock(&ctx
->mutex
);
6713 perf_install_in_context(ctx
, event
, cpu
);
6715 perf_unpin_context(ctx
);
6716 mutex_unlock(&ctx
->mutex
);
6723 return ERR_PTR(err
);
6725 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter
);
6727 void perf_pmu_migrate_context(struct pmu
*pmu
, int src_cpu
, int dst_cpu
)
6729 struct perf_event_context
*src_ctx
;
6730 struct perf_event_context
*dst_ctx
;
6731 struct perf_event
*event
, *tmp
;
6734 src_ctx
= &per_cpu_ptr(pmu
->pmu_cpu_context
, src_cpu
)->ctx
;
6735 dst_ctx
= &per_cpu_ptr(pmu
->pmu_cpu_context
, dst_cpu
)->ctx
;
6737 mutex_lock(&src_ctx
->mutex
);
6738 list_for_each_entry_safe(event
, tmp
, &src_ctx
->event_list
,
6740 perf_remove_from_context(event
);
6742 list_add(&event
->event_entry
, &events
);
6744 mutex_unlock(&src_ctx
->mutex
);
6748 mutex_lock(&dst_ctx
->mutex
);
6749 list_for_each_entry_safe(event
, tmp
, &events
, event_entry
) {
6750 list_del(&event
->event_entry
);
6751 if (event
->state
>= PERF_EVENT_STATE_OFF
)
6752 event
->state
= PERF_EVENT_STATE_INACTIVE
;
6753 perf_install_in_context(dst_ctx
, event
, dst_cpu
);
6756 mutex_unlock(&dst_ctx
->mutex
);
6758 EXPORT_SYMBOL_GPL(perf_pmu_migrate_context
);
6760 static void sync_child_event(struct perf_event
*child_event
,
6761 struct task_struct
*child
)
6763 struct perf_event
*parent_event
= child_event
->parent
;
6766 if (child_event
->attr
.inherit_stat
)
6767 perf_event_read_event(child_event
, child
);
6769 child_val
= perf_event_count(child_event
);
6772 * Add back the child's count to the parent's count:
6774 atomic64_add(child_val
, &parent_event
->child_count
);
6775 atomic64_add(child_event
->total_time_enabled
,
6776 &parent_event
->child_total_time_enabled
);
6777 atomic64_add(child_event
->total_time_running
,
6778 &parent_event
->child_total_time_running
);
6781 * Remove this event from the parent's list
6783 WARN_ON_ONCE(parent_event
->ctx
->parent_ctx
);
6784 mutex_lock(&parent_event
->child_mutex
);
6785 list_del_init(&child_event
->child_list
);
6786 mutex_unlock(&parent_event
->child_mutex
);
6789 * Release the parent event, if this was the last
6792 put_event(parent_event
);
6796 __perf_event_exit_task(struct perf_event
*child_event
,
6797 struct perf_event_context
*child_ctx
,
6798 struct task_struct
*child
)
6800 if (child_event
->parent
) {
6801 raw_spin_lock_irq(&child_ctx
->lock
);
6802 perf_group_detach(child_event
);
6803 raw_spin_unlock_irq(&child_ctx
->lock
);
6806 perf_remove_from_context(child_event
);
6809 * It can happen that the parent exits first, and has events
6810 * that are still around due to the child reference. These
6811 * events need to be zapped.
6813 if (child_event
->parent
) {
6814 sync_child_event(child_event
, child
);
6815 free_event(child_event
);
6819 static void perf_event_exit_task_context(struct task_struct
*child
, int ctxn
)
6821 struct perf_event
*child_event
, *tmp
;
6822 struct perf_event_context
*child_ctx
;
6823 unsigned long flags
;
6825 if (likely(!child
->perf_event_ctxp
[ctxn
])) {
6826 perf_event_task(child
, NULL
, 0);
6830 local_irq_save(flags
);
6832 * We can't reschedule here because interrupts are disabled,
6833 * and either child is current or it is a task that can't be
6834 * scheduled, so we are now safe from rescheduling changing
6837 child_ctx
= rcu_dereference_raw(child
->perf_event_ctxp
[ctxn
]);
6840 * Take the context lock here so that if find_get_context is
6841 * reading child->perf_event_ctxp, we wait until it has
6842 * incremented the context's refcount before we do put_ctx below.
6844 raw_spin_lock(&child_ctx
->lock
);
6845 task_ctx_sched_out(child_ctx
);
6846 child
->perf_event_ctxp
[ctxn
] = NULL
;
6848 * If this context is a clone; unclone it so it can't get
6849 * swapped to another process while we're removing all
6850 * the events from it.
6852 unclone_ctx(child_ctx
);
6853 update_context_time(child_ctx
);
6854 raw_spin_unlock_irqrestore(&child_ctx
->lock
, flags
);
6857 * Report the task dead after unscheduling the events so that we
6858 * won't get any samples after PERF_RECORD_EXIT. We can however still
6859 * get a few PERF_RECORD_READ events.
6861 perf_event_task(child
, child_ctx
, 0);
6864 * We can recurse on the same lock type through:
6866 * __perf_event_exit_task()
6867 * sync_child_event()
6869 * mutex_lock(&ctx->mutex)
6871 * But since its the parent context it won't be the same instance.
6873 mutex_lock(&child_ctx
->mutex
);
6876 list_for_each_entry_safe(child_event
, tmp
, &child_ctx
->pinned_groups
,
6878 __perf_event_exit_task(child_event
, child_ctx
, child
);
6880 list_for_each_entry_safe(child_event
, tmp
, &child_ctx
->flexible_groups
,
6882 __perf_event_exit_task(child_event
, child_ctx
, child
);
6885 * If the last event was a group event, it will have appended all
6886 * its siblings to the list, but we obtained 'tmp' before that which
6887 * will still point to the list head terminating the iteration.
6889 if (!list_empty(&child_ctx
->pinned_groups
) ||
6890 !list_empty(&child_ctx
->flexible_groups
))
6893 mutex_unlock(&child_ctx
->mutex
);
6899 * When a child task exits, feed back event values to parent events.
6901 void perf_event_exit_task(struct task_struct
*child
)
6903 struct perf_event
*event
, *tmp
;
6906 mutex_lock(&child
->perf_event_mutex
);
6907 list_for_each_entry_safe(event
, tmp
, &child
->perf_event_list
,
6909 list_del_init(&event
->owner_entry
);
6912 * Ensure the list deletion is visible before we clear
6913 * the owner, closes a race against perf_release() where
6914 * we need to serialize on the owner->perf_event_mutex.
6917 event
->owner
= NULL
;
6919 mutex_unlock(&child
->perf_event_mutex
);
6921 for_each_task_context_nr(ctxn
)
6922 perf_event_exit_task_context(child
, ctxn
);
6925 static void perf_free_event(struct perf_event
*event
,
6926 struct perf_event_context
*ctx
)
6928 struct perf_event
*parent
= event
->parent
;
6930 if (WARN_ON_ONCE(!parent
))
6933 mutex_lock(&parent
->child_mutex
);
6934 list_del_init(&event
->child_list
);
6935 mutex_unlock(&parent
->child_mutex
);
6939 perf_group_detach(event
);
6940 list_del_event(event
, ctx
);
6945 * free an unexposed, unused context as created by inheritance by
6946 * perf_event_init_task below, used by fork() in case of fail.
6948 void perf_event_free_task(struct task_struct
*task
)
6950 struct perf_event_context
*ctx
;
6951 struct perf_event
*event
, *tmp
;
6954 for_each_task_context_nr(ctxn
) {
6955 ctx
= task
->perf_event_ctxp
[ctxn
];
6959 mutex_lock(&ctx
->mutex
);
6961 list_for_each_entry_safe(event
, tmp
, &ctx
->pinned_groups
,
6963 perf_free_event(event
, ctx
);
6965 list_for_each_entry_safe(event
, tmp
, &ctx
->flexible_groups
,
6967 perf_free_event(event
, ctx
);
6969 if (!list_empty(&ctx
->pinned_groups
) ||
6970 !list_empty(&ctx
->flexible_groups
))
6973 mutex_unlock(&ctx
->mutex
);
6979 void perf_event_delayed_put(struct task_struct
*task
)
6983 for_each_task_context_nr(ctxn
)
6984 WARN_ON_ONCE(task
->perf_event_ctxp
[ctxn
]);
6988 * inherit a event from parent task to child task:
6990 static struct perf_event
*
6991 inherit_event(struct perf_event
*parent_event
,
6992 struct task_struct
*parent
,
6993 struct perf_event_context
*parent_ctx
,
6994 struct task_struct
*child
,
6995 struct perf_event
*group_leader
,
6996 struct perf_event_context
*child_ctx
)
6998 struct perf_event
*child_event
;
6999 unsigned long flags
;
7002 * Instead of creating recursive hierarchies of events,
7003 * we link inherited events back to the original parent,
7004 * which has a filp for sure, which we use as the reference
7007 if (parent_event
->parent
)
7008 parent_event
= parent_event
->parent
;
7010 child_event
= perf_event_alloc(&parent_event
->attr
,
7013 group_leader
, parent_event
,
7015 if (IS_ERR(child_event
))
7018 if (!atomic_long_inc_not_zero(&parent_event
->refcount
)) {
7019 free_event(child_event
);
7026 * Make the child state follow the state of the parent event,
7027 * not its attr.disabled bit. We hold the parent's mutex,
7028 * so we won't race with perf_event_{en, dis}able_family.
7030 if (parent_event
->state
>= PERF_EVENT_STATE_INACTIVE
)
7031 child_event
->state
= PERF_EVENT_STATE_INACTIVE
;
7033 child_event
->state
= PERF_EVENT_STATE_OFF
;
7035 if (parent_event
->attr
.freq
) {
7036 u64 sample_period
= parent_event
->hw
.sample_period
;
7037 struct hw_perf_event
*hwc
= &child_event
->hw
;
7039 hwc
->sample_period
= sample_period
;
7040 hwc
->last_period
= sample_period
;
7042 local64_set(&hwc
->period_left
, sample_period
);
7045 child_event
->ctx
= child_ctx
;
7046 child_event
->overflow_handler
= parent_event
->overflow_handler
;
7047 child_event
->overflow_handler_context
7048 = parent_event
->overflow_handler_context
;
7051 * Precalculate sample_data sizes
7053 perf_event__header_size(child_event
);
7054 perf_event__id_header_size(child_event
);
7057 * Link it up in the child's context:
7059 raw_spin_lock_irqsave(&child_ctx
->lock
, flags
);
7060 add_event_to_ctx(child_event
, child_ctx
);
7061 raw_spin_unlock_irqrestore(&child_ctx
->lock
, flags
);
7064 * Link this into the parent event's child list
7066 WARN_ON_ONCE(parent_event
->ctx
->parent_ctx
);
7067 mutex_lock(&parent_event
->child_mutex
);
7068 list_add_tail(&child_event
->child_list
, &parent_event
->child_list
);
7069 mutex_unlock(&parent_event
->child_mutex
);
7074 static int inherit_group(struct perf_event
*parent_event
,
7075 struct task_struct
*parent
,
7076 struct perf_event_context
*parent_ctx
,
7077 struct task_struct
*child
,
7078 struct perf_event_context
*child_ctx
)
7080 struct perf_event
*leader
;
7081 struct perf_event
*sub
;
7082 struct perf_event
*child_ctr
;
7084 leader
= inherit_event(parent_event
, parent
, parent_ctx
,
7085 child
, NULL
, child_ctx
);
7087 return PTR_ERR(leader
);
7088 list_for_each_entry(sub
, &parent_event
->sibling_list
, group_entry
) {
7089 child_ctr
= inherit_event(sub
, parent
, parent_ctx
,
7090 child
, leader
, child_ctx
);
7091 if (IS_ERR(child_ctr
))
7092 return PTR_ERR(child_ctr
);
7098 inherit_task_group(struct perf_event
*event
, struct task_struct
*parent
,
7099 struct perf_event_context
*parent_ctx
,
7100 struct task_struct
*child
, int ctxn
,
7104 struct perf_event_context
*child_ctx
;
7106 if (!event
->attr
.inherit
) {
7111 child_ctx
= child
->perf_event_ctxp
[ctxn
];
7114 * This is executed from the parent task context, so
7115 * inherit events that have been marked for cloning.
7116 * First allocate and initialize a context for the
7120 child_ctx
= alloc_perf_context(event
->pmu
, child
);
7124 child
->perf_event_ctxp
[ctxn
] = child_ctx
;
7127 ret
= inherit_group(event
, parent
, parent_ctx
,
7137 * Initialize the perf_event context in task_struct
7139 int perf_event_init_context(struct task_struct
*child
, int ctxn
)
7141 struct perf_event_context
*child_ctx
, *parent_ctx
;
7142 struct perf_event_context
*cloned_ctx
;
7143 struct perf_event
*event
;
7144 struct task_struct
*parent
= current
;
7145 int inherited_all
= 1;
7146 unsigned long flags
;
7149 if (likely(!parent
->perf_event_ctxp
[ctxn
]))
7153 * If the parent's context is a clone, pin it so it won't get
7156 parent_ctx
= perf_pin_task_context(parent
, ctxn
);
7159 * No need to check if parent_ctx != NULL here; since we saw
7160 * it non-NULL earlier, the only reason for it to become NULL
7161 * is if we exit, and since we're currently in the middle of
7162 * a fork we can't be exiting at the same time.
7166 * Lock the parent list. No need to lock the child - not PID
7167 * hashed yet and not running, so nobody can access it.
7169 mutex_lock(&parent_ctx
->mutex
);
7172 * We dont have to disable NMIs - we are only looking at
7173 * the list, not manipulating it:
7175 list_for_each_entry(event
, &parent_ctx
->pinned_groups
, group_entry
) {
7176 ret
= inherit_task_group(event
, parent
, parent_ctx
,
7177 child
, ctxn
, &inherited_all
);
7183 * We can't hold ctx->lock when iterating the ->flexible_group list due
7184 * to allocations, but we need to prevent rotation because
7185 * rotate_ctx() will change the list from interrupt context.
7187 raw_spin_lock_irqsave(&parent_ctx
->lock
, flags
);
7188 parent_ctx
->rotate_disable
= 1;
7189 raw_spin_unlock_irqrestore(&parent_ctx
->lock
, flags
);
7191 list_for_each_entry(event
, &parent_ctx
->flexible_groups
, group_entry
) {
7192 ret
= inherit_task_group(event
, parent
, parent_ctx
,
7193 child
, ctxn
, &inherited_all
);
7198 raw_spin_lock_irqsave(&parent_ctx
->lock
, flags
);
7199 parent_ctx
->rotate_disable
= 0;
7201 child_ctx
= child
->perf_event_ctxp
[ctxn
];
7203 if (child_ctx
&& inherited_all
) {
7205 * Mark the child context as a clone of the parent
7206 * context, or of whatever the parent is a clone of.
7208 * Note that if the parent is a clone, the holding of
7209 * parent_ctx->lock avoids it from being uncloned.
7211 cloned_ctx
= parent_ctx
->parent_ctx
;
7213 child_ctx
->parent_ctx
= cloned_ctx
;
7214 child_ctx
->parent_gen
= parent_ctx
->parent_gen
;
7216 child_ctx
->parent_ctx
= parent_ctx
;
7217 child_ctx
->parent_gen
= parent_ctx
->generation
;
7219 get_ctx(child_ctx
->parent_ctx
);
7222 raw_spin_unlock_irqrestore(&parent_ctx
->lock
, flags
);
7223 mutex_unlock(&parent_ctx
->mutex
);
7225 perf_unpin_context(parent_ctx
);
7226 put_ctx(parent_ctx
);
7232 * Initialize the perf_event context in task_struct
7234 int perf_event_init_task(struct task_struct
*child
)
7238 memset(child
->perf_event_ctxp
, 0, sizeof(child
->perf_event_ctxp
));
7239 mutex_init(&child
->perf_event_mutex
);
7240 INIT_LIST_HEAD(&child
->perf_event_list
);
7242 for_each_task_context_nr(ctxn
) {
7243 ret
= perf_event_init_context(child
, ctxn
);
7251 static void __init
perf_event_init_all_cpus(void)
7253 struct swevent_htable
*swhash
;
7256 for_each_possible_cpu(cpu
) {
7257 swhash
= &per_cpu(swevent_htable
, cpu
);
7258 mutex_init(&swhash
->hlist_mutex
);
7259 INIT_LIST_HEAD(&per_cpu(rotation_list
, cpu
));
7263 static void __cpuinit
perf_event_init_cpu(int cpu
)
7265 struct swevent_htable
*swhash
= &per_cpu(swevent_htable
, cpu
);
7267 mutex_lock(&swhash
->hlist_mutex
);
7268 if (swhash
->hlist_refcount
> 0) {
7269 struct swevent_hlist
*hlist
;
7271 hlist
= kzalloc_node(sizeof(*hlist
), GFP_KERNEL
, cpu_to_node(cpu
));
7273 rcu_assign_pointer(swhash
->swevent_hlist
, hlist
);
7275 mutex_unlock(&swhash
->hlist_mutex
);
7278 #if defined CONFIG_HOTPLUG_CPU || defined CONFIG_KEXEC
7279 static void perf_pmu_rotate_stop(struct pmu
*pmu
)
7281 struct perf_cpu_context
*cpuctx
= this_cpu_ptr(pmu
->pmu_cpu_context
);
7283 WARN_ON(!irqs_disabled());
7285 list_del_init(&cpuctx
->rotation_list
);
7288 static void __perf_event_exit_context(void *__info
)
7290 struct perf_event_context
*ctx
= __info
;
7291 struct perf_event
*event
, *tmp
;
7293 perf_pmu_rotate_stop(ctx
->pmu
);
7295 list_for_each_entry_safe(event
, tmp
, &ctx
->pinned_groups
, group_entry
)
7296 __perf_remove_from_context(event
);
7297 list_for_each_entry_safe(event
, tmp
, &ctx
->flexible_groups
, group_entry
)
7298 __perf_remove_from_context(event
);
7301 static void perf_event_exit_cpu_context(int cpu
)
7303 struct perf_event_context
*ctx
;
7307 idx
= srcu_read_lock(&pmus_srcu
);
7308 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
7309 ctx
= &per_cpu_ptr(pmu
->pmu_cpu_context
, cpu
)->ctx
;
7311 mutex_lock(&ctx
->mutex
);
7312 smp_call_function_single(cpu
, __perf_event_exit_context
, ctx
, 1);
7313 mutex_unlock(&ctx
->mutex
);
7315 srcu_read_unlock(&pmus_srcu
, idx
);
7318 static void perf_event_exit_cpu(int cpu
)
7320 struct swevent_htable
*swhash
= &per_cpu(swevent_htable
, cpu
);
7322 mutex_lock(&swhash
->hlist_mutex
);
7323 swevent_hlist_release(swhash
);
7324 mutex_unlock(&swhash
->hlist_mutex
);
7326 perf_event_exit_cpu_context(cpu
);
7329 static inline void perf_event_exit_cpu(int cpu
) { }
7333 perf_reboot(struct notifier_block
*notifier
, unsigned long val
, void *v
)
7337 for_each_online_cpu(cpu
)
7338 perf_event_exit_cpu(cpu
);
7344 * Run the perf reboot notifier at the very last possible moment so that
7345 * the generic watchdog code runs as long as possible.
7347 static struct notifier_block perf_reboot_notifier
= {
7348 .notifier_call
= perf_reboot
,
7349 .priority
= INT_MIN
,
7352 static int __cpuinit
7353 perf_cpu_notify(struct notifier_block
*self
, unsigned long action
, void *hcpu
)
7355 unsigned int cpu
= (long)hcpu
;
7357 switch (action
& ~CPU_TASKS_FROZEN
) {
7359 case CPU_UP_PREPARE
:
7360 case CPU_DOWN_FAILED
:
7361 perf_event_init_cpu(cpu
);
7364 case CPU_UP_CANCELED
:
7365 case CPU_DOWN_PREPARE
:
7366 perf_event_exit_cpu(cpu
);
7376 void __init
perf_event_init(void)
7382 perf_event_init_all_cpus();
7383 init_srcu_struct(&pmus_srcu
);
7384 perf_pmu_register(&perf_swevent
, "software", PERF_TYPE_SOFTWARE
);
7385 perf_pmu_register(&perf_cpu_clock
, NULL
, -1);
7386 perf_pmu_register(&perf_task_clock
, NULL
, -1);
7388 perf_cpu_notifier(perf_cpu_notify
);
7389 register_reboot_notifier(&perf_reboot_notifier
);
7391 ret
= init_hw_breakpoint();
7392 WARN(ret
, "hw_breakpoint initialization failed with: %d", ret
);
7394 /* do not patch jump label more than once per second */
7395 jump_label_rate_limit(&perf_sched_events
, HZ
);
7398 * Build time assertion that we keep the data_head at the intended
7399 * location. IOW, validation we got the __reserved[] size right.
7401 BUILD_BUG_ON((offsetof(struct perf_event_mmap_page
, data_head
))
7405 static int __init
perf_event_sysfs_init(void)
7410 mutex_lock(&pmus_lock
);
7412 ret
= bus_register(&pmu_bus
);
7416 list_for_each_entry(pmu
, &pmus
, entry
) {
7417 if (!pmu
->name
|| pmu
->type
< 0)
7420 ret
= pmu_dev_alloc(pmu
);
7421 WARN(ret
, "Failed to register pmu: %s, reason %d\n", pmu
->name
, ret
);
7423 pmu_bus_running
= 1;
7427 mutex_unlock(&pmus_lock
);
7431 device_initcall(perf_event_sysfs_init
);
7433 #ifdef CONFIG_CGROUP_PERF
7434 static struct cgroup_subsys_state
*perf_cgroup_create(struct cgroup
*cont
)
7436 struct perf_cgroup
*jc
;
7438 jc
= kzalloc(sizeof(*jc
), GFP_KERNEL
);
7440 return ERR_PTR(-ENOMEM
);
7442 jc
->info
= alloc_percpu(struct perf_cgroup_info
);
7445 return ERR_PTR(-ENOMEM
);
7451 static void perf_cgroup_destroy(struct cgroup
*cont
)
7453 struct perf_cgroup
*jc
;
7454 jc
= container_of(cgroup_subsys_state(cont
, perf_subsys_id
),
7455 struct perf_cgroup
, css
);
7456 free_percpu(jc
->info
);
7460 static int __perf_cgroup_move(void *info
)
7462 struct task_struct
*task
= info
;
7463 perf_cgroup_switch(task
, PERF_CGROUP_SWOUT
| PERF_CGROUP_SWIN
);
7467 static void perf_cgroup_attach(struct cgroup
*cgrp
, struct cgroup_taskset
*tset
)
7469 struct task_struct
*task
;
7471 cgroup_taskset_for_each(task
, cgrp
, tset
)
7472 task_function_call(task
, __perf_cgroup_move
, task
);
7475 static void perf_cgroup_exit(struct cgroup
*cgrp
, struct cgroup
*old_cgrp
,
7476 struct task_struct
*task
)
7479 * cgroup_exit() is called in the copy_process() failure path.
7480 * Ignore this case since the task hasn't ran yet, this avoids
7481 * trying to poke a half freed task state from generic code.
7483 if (!(task
->flags
& PF_EXITING
))
7486 task_function_call(task
, __perf_cgroup_move
, task
);
7489 struct cgroup_subsys perf_subsys
= {
7490 .name
= "perf_event",
7491 .subsys_id
= perf_subsys_id
,
7492 .create
= perf_cgroup_create
,
7493 .destroy
= perf_cgroup_destroy
,
7494 .exit
= perf_cgroup_exit
,
7495 .attach
= perf_cgroup_attach
,
7498 * perf_event cgroup doesn't handle nesting correctly.
7499 * ctx->nr_cgroups adjustments should be propagated through the
7500 * cgroup hierarchy. Fix it and remove the following.
7502 .broken_hierarchy
= true,
7504 #endif /* CONFIG_CGROUP_PERF */