2 * Performance events core code:
4 * Copyright (C) 2008 Thomas Gleixner <tglx@linutronix.de>
5 * Copyright (C) 2008-2011 Red Hat, Inc., Ingo Molnar
6 * Copyright (C) 2008-2011 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
7 * Copyright © 2009 Paul Mackerras, IBM Corp. <paulus@au1.ibm.com>
9 * For licensing details see kernel-base/COPYING
14 #include <linux/cpu.h>
15 #include <linux/smp.h>
16 #include <linux/idr.h>
17 #include <linux/file.h>
18 #include <linux/poll.h>
19 #include <linux/slab.h>
20 #include <linux/hash.h>
21 #include <linux/sysfs.h>
22 #include <linux/dcache.h>
23 #include <linux/percpu.h>
24 #include <linux/ptrace.h>
25 #include <linux/reboot.h>
26 #include <linux/vmstat.h>
27 #include <linux/device.h>
28 #include <linux/export.h>
29 #include <linux/vmalloc.h>
30 #include <linux/hardirq.h>
31 #include <linux/rculist.h>
32 #include <linux/uaccess.h>
33 #include <linux/syscalls.h>
34 #include <linux/anon_inodes.h>
35 #include <linux/kernel_stat.h>
36 #include <linux/perf_event.h>
37 #include <linux/ftrace_event.h>
38 #include <linux/hw_breakpoint.h>
42 #include <asm/irq_regs.h>
44 struct remote_function_call
{
45 struct task_struct
*p
;
46 int (*func
)(void *info
);
51 static void remote_function(void *data
)
53 struct remote_function_call
*tfc
= data
;
54 struct task_struct
*p
= tfc
->p
;
58 if (task_cpu(p
) != smp_processor_id() || !task_curr(p
))
62 tfc
->ret
= tfc
->func(tfc
->info
);
66 * task_function_call - call a function on the cpu on which a task runs
67 * @p: the task to evaluate
68 * @func: the function to be called
69 * @info: the function call argument
71 * Calls the function @func when the task is currently running. This might
72 * be on the current CPU, which just calls the function directly
74 * returns: @func return value, or
75 * -ESRCH - when the process isn't running
76 * -EAGAIN - when the process moved away
79 task_function_call(struct task_struct
*p
, int (*func
) (void *info
), void *info
)
81 struct remote_function_call data
= {
85 .ret
= -ESRCH
, /* No such (running) process */
89 smp_call_function_single(task_cpu(p
), remote_function
, &data
, 1);
95 * cpu_function_call - call a function on the cpu
96 * @func: the function to be called
97 * @info: the function call argument
99 * Calls the function @func on the remote cpu.
101 * returns: @func return value or -ENXIO when the cpu is offline
103 static int cpu_function_call(int cpu
, int (*func
) (void *info
), void *info
)
105 struct remote_function_call data
= {
109 .ret
= -ENXIO
, /* No such CPU */
112 smp_call_function_single(cpu
, remote_function
, &data
, 1);
117 #define PERF_FLAG_ALL (PERF_FLAG_FD_NO_GROUP |\
118 PERF_FLAG_FD_OUTPUT |\
119 PERF_FLAG_PID_CGROUP)
122 EVENT_FLEXIBLE
= 0x1,
124 EVENT_ALL
= EVENT_FLEXIBLE
| EVENT_PINNED
,
128 * perf_sched_events : >0 events exist
129 * perf_cgroup_events: >0 per-cpu cgroup events exist on this cpu
131 struct jump_label_key perf_sched_events __read_mostly
;
132 static DEFINE_PER_CPU(atomic_t
, perf_cgroup_events
);
134 static atomic_t nr_mmap_events __read_mostly
;
135 static atomic_t nr_comm_events __read_mostly
;
136 static atomic_t nr_task_events __read_mostly
;
138 static LIST_HEAD(pmus
);
139 static DEFINE_MUTEX(pmus_lock
);
140 static struct srcu_struct pmus_srcu
;
143 * perf event paranoia level:
144 * -1 - not paranoid at all
145 * 0 - disallow raw tracepoint access for unpriv
146 * 1 - disallow cpu events for unpriv
147 * 2 - disallow kernel profiling for unpriv
149 int sysctl_perf_event_paranoid __read_mostly
= 1;
151 /* Minimum for 512 kiB + 1 user control page */
152 int sysctl_perf_event_mlock __read_mostly
= 512 + (PAGE_SIZE
/ 1024); /* 'free' kiB per user */
155 * max perf event sample rate
157 #define DEFAULT_MAX_SAMPLE_RATE 100000
158 int sysctl_perf_event_sample_rate __read_mostly
= DEFAULT_MAX_SAMPLE_RATE
;
159 static int max_samples_per_tick __read_mostly
=
160 DIV_ROUND_UP(DEFAULT_MAX_SAMPLE_RATE
, HZ
);
162 int perf_proc_update_handler(struct ctl_table
*table
, int write
,
163 void __user
*buffer
, size_t *lenp
,
166 int ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
171 max_samples_per_tick
= DIV_ROUND_UP(sysctl_perf_event_sample_rate
, HZ
);
176 static atomic64_t perf_event_id
;
178 static void cpu_ctx_sched_out(struct perf_cpu_context
*cpuctx
,
179 enum event_type_t event_type
);
181 static void cpu_ctx_sched_in(struct perf_cpu_context
*cpuctx
,
182 enum event_type_t event_type
,
183 struct task_struct
*task
);
185 static void update_context_time(struct perf_event_context
*ctx
);
186 static u64
perf_event_time(struct perf_event
*event
);
188 static void ring_buffer_attach(struct perf_event
*event
,
189 struct ring_buffer
*rb
);
191 void __weak
perf_event_print_debug(void) { }
193 extern __weak
const char *perf_pmu_name(void)
198 static inline u64
perf_clock(void)
200 return local_clock();
203 static inline struct perf_cpu_context
*
204 __get_cpu_context(struct perf_event_context
*ctx
)
206 return this_cpu_ptr(ctx
->pmu
->pmu_cpu_context
);
209 static void perf_ctx_lock(struct perf_cpu_context
*cpuctx
,
210 struct perf_event_context
*ctx
)
212 raw_spin_lock(&cpuctx
->ctx
.lock
);
214 raw_spin_lock(&ctx
->lock
);
217 static void perf_ctx_unlock(struct perf_cpu_context
*cpuctx
,
218 struct perf_event_context
*ctx
)
221 raw_spin_unlock(&ctx
->lock
);
222 raw_spin_unlock(&cpuctx
->ctx
.lock
);
225 #ifdef CONFIG_CGROUP_PERF
228 * Must ensure cgroup is pinned (css_get) before calling
229 * this function. In other words, we cannot call this function
230 * if there is no cgroup event for the current CPU context.
232 static inline struct perf_cgroup
*
233 perf_cgroup_from_task(struct task_struct
*task
)
235 return container_of(task_subsys_state(task
, perf_subsys_id
),
236 struct perf_cgroup
, css
);
240 perf_cgroup_match(struct perf_event
*event
)
242 struct perf_event_context
*ctx
= event
->ctx
;
243 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
245 return !event
->cgrp
|| event
->cgrp
== cpuctx
->cgrp
;
248 static inline void perf_get_cgroup(struct perf_event
*event
)
250 css_get(&event
->cgrp
->css
);
253 static inline void perf_put_cgroup(struct perf_event
*event
)
255 css_put(&event
->cgrp
->css
);
258 static inline void perf_detach_cgroup(struct perf_event
*event
)
260 perf_put_cgroup(event
);
264 static inline int is_cgroup_event(struct perf_event
*event
)
266 return event
->cgrp
!= NULL
;
269 static inline u64
perf_cgroup_event_time(struct perf_event
*event
)
271 struct perf_cgroup_info
*t
;
273 t
= per_cpu_ptr(event
->cgrp
->info
, event
->cpu
);
277 static inline void __update_cgrp_time(struct perf_cgroup
*cgrp
)
279 struct perf_cgroup_info
*info
;
284 info
= this_cpu_ptr(cgrp
->info
);
286 info
->time
+= now
- info
->timestamp
;
287 info
->timestamp
= now
;
290 static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context
*cpuctx
)
292 struct perf_cgroup
*cgrp_out
= cpuctx
->cgrp
;
294 __update_cgrp_time(cgrp_out
);
297 static inline void update_cgrp_time_from_event(struct perf_event
*event
)
299 struct perf_cgroup
*cgrp
;
302 * ensure we access cgroup data only when needed and
303 * when we know the cgroup is pinned (css_get)
305 if (!is_cgroup_event(event
))
308 cgrp
= perf_cgroup_from_task(current
);
310 * Do not update time when cgroup is not active
312 if (cgrp
== event
->cgrp
)
313 __update_cgrp_time(event
->cgrp
);
317 perf_cgroup_set_timestamp(struct task_struct
*task
,
318 struct perf_event_context
*ctx
)
320 struct perf_cgroup
*cgrp
;
321 struct perf_cgroup_info
*info
;
324 * ctx->lock held by caller
325 * ensure we do not access cgroup data
326 * unless we have the cgroup pinned (css_get)
328 if (!task
|| !ctx
->nr_cgroups
)
331 cgrp
= perf_cgroup_from_task(task
);
332 info
= this_cpu_ptr(cgrp
->info
);
333 info
->timestamp
= ctx
->timestamp
;
336 #define PERF_CGROUP_SWOUT 0x1 /* cgroup switch out every event */
337 #define PERF_CGROUP_SWIN 0x2 /* cgroup switch in events based on task */
340 * reschedule events based on the cgroup constraint of task.
342 * mode SWOUT : schedule out everything
343 * mode SWIN : schedule in based on cgroup for next
345 void perf_cgroup_switch(struct task_struct
*task
, int mode
)
347 struct perf_cpu_context
*cpuctx
;
352 * disable interrupts to avoid geting nr_cgroup
353 * changes via __perf_event_disable(). Also
356 local_irq_save(flags
);
359 * we reschedule only in the presence of cgroup
360 * constrained events.
364 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
365 cpuctx
= this_cpu_ptr(pmu
->pmu_cpu_context
);
368 * perf_cgroup_events says at least one
369 * context on this CPU has cgroup events.
371 * ctx->nr_cgroups reports the number of cgroup
372 * events for a context.
374 if (cpuctx
->ctx
.nr_cgroups
> 0) {
375 perf_ctx_lock(cpuctx
, cpuctx
->task_ctx
);
376 perf_pmu_disable(cpuctx
->ctx
.pmu
);
378 if (mode
& PERF_CGROUP_SWOUT
) {
379 cpu_ctx_sched_out(cpuctx
, EVENT_ALL
);
381 * must not be done before ctxswout due
382 * to event_filter_match() in event_sched_out()
387 if (mode
& PERF_CGROUP_SWIN
) {
388 WARN_ON_ONCE(cpuctx
->cgrp
);
389 /* set cgrp before ctxsw in to
390 * allow event_filter_match() to not
391 * have to pass task around
393 cpuctx
->cgrp
= perf_cgroup_from_task(task
);
394 cpu_ctx_sched_in(cpuctx
, EVENT_ALL
, task
);
396 perf_pmu_enable(cpuctx
->ctx
.pmu
);
397 perf_ctx_unlock(cpuctx
, cpuctx
->task_ctx
);
403 local_irq_restore(flags
);
406 static inline void perf_cgroup_sched_out(struct task_struct
*task
,
407 struct task_struct
*next
)
409 struct perf_cgroup
*cgrp1
;
410 struct perf_cgroup
*cgrp2
= NULL
;
413 * we come here when we know perf_cgroup_events > 0
415 cgrp1
= perf_cgroup_from_task(task
);
418 * next is NULL when called from perf_event_enable_on_exec()
419 * that will systematically cause a cgroup_switch()
422 cgrp2
= perf_cgroup_from_task(next
);
425 * only schedule out current cgroup events if we know
426 * that we are switching to a different cgroup. Otherwise,
427 * do no touch the cgroup events.
430 perf_cgroup_switch(task
, PERF_CGROUP_SWOUT
);
433 static inline void perf_cgroup_sched_in(struct task_struct
*prev
,
434 struct task_struct
*task
)
436 struct perf_cgroup
*cgrp1
;
437 struct perf_cgroup
*cgrp2
= NULL
;
440 * we come here when we know perf_cgroup_events > 0
442 cgrp1
= perf_cgroup_from_task(task
);
444 /* prev can never be NULL */
445 cgrp2
= perf_cgroup_from_task(prev
);
448 * only need to schedule in cgroup events if we are changing
449 * cgroup during ctxsw. Cgroup events were not scheduled
450 * out of ctxsw out if that was not the case.
453 perf_cgroup_switch(task
, PERF_CGROUP_SWIN
);
456 static inline int perf_cgroup_connect(int fd
, struct perf_event
*event
,
457 struct perf_event_attr
*attr
,
458 struct perf_event
*group_leader
)
460 struct perf_cgroup
*cgrp
;
461 struct cgroup_subsys_state
*css
;
463 int ret
= 0, fput_needed
;
465 file
= fget_light(fd
, &fput_needed
);
469 css
= cgroup_css_from_dir(file
, perf_subsys_id
);
475 cgrp
= container_of(css
, struct perf_cgroup
, css
);
478 /* must be done before we fput() the file */
479 perf_get_cgroup(event
);
482 * all events in a group must monitor
483 * the same cgroup because a task belongs
484 * to only one perf cgroup at a time
486 if (group_leader
&& group_leader
->cgrp
!= cgrp
) {
487 perf_detach_cgroup(event
);
491 fput_light(file
, fput_needed
);
496 perf_cgroup_set_shadow_time(struct perf_event
*event
, u64 now
)
498 struct perf_cgroup_info
*t
;
499 t
= per_cpu_ptr(event
->cgrp
->info
, event
->cpu
);
500 event
->shadow_ctx_time
= now
- t
->timestamp
;
504 perf_cgroup_defer_enabled(struct perf_event
*event
)
507 * when the current task's perf cgroup does not match
508 * the event's, we need to remember to call the
509 * perf_mark_enable() function the first time a task with
510 * a matching perf cgroup is scheduled in.
512 if (is_cgroup_event(event
) && !perf_cgroup_match(event
))
513 event
->cgrp_defer_enabled
= 1;
517 perf_cgroup_mark_enabled(struct perf_event
*event
,
518 struct perf_event_context
*ctx
)
520 struct perf_event
*sub
;
521 u64 tstamp
= perf_event_time(event
);
523 if (!event
->cgrp_defer_enabled
)
526 event
->cgrp_defer_enabled
= 0;
528 event
->tstamp_enabled
= tstamp
- event
->total_time_enabled
;
529 list_for_each_entry(sub
, &event
->sibling_list
, group_entry
) {
530 if (sub
->state
>= PERF_EVENT_STATE_INACTIVE
) {
531 sub
->tstamp_enabled
= tstamp
- sub
->total_time_enabled
;
532 sub
->cgrp_defer_enabled
= 0;
536 #else /* !CONFIG_CGROUP_PERF */
539 perf_cgroup_match(struct perf_event
*event
)
544 static inline void perf_detach_cgroup(struct perf_event
*event
)
547 static inline int is_cgroup_event(struct perf_event
*event
)
552 static inline u64
perf_cgroup_event_cgrp_time(struct perf_event
*event
)
557 static inline void update_cgrp_time_from_event(struct perf_event
*event
)
561 static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context
*cpuctx
)
565 static inline void perf_cgroup_sched_out(struct task_struct
*task
,
566 struct task_struct
*next
)
570 static inline void perf_cgroup_sched_in(struct task_struct
*prev
,
571 struct task_struct
*task
)
575 static inline int perf_cgroup_connect(pid_t pid
, struct perf_event
*event
,
576 struct perf_event_attr
*attr
,
577 struct perf_event
*group_leader
)
583 perf_cgroup_set_timestamp(struct task_struct
*task
,
584 struct perf_event_context
*ctx
)
589 perf_cgroup_switch(struct task_struct
*task
, struct task_struct
*next
)
594 perf_cgroup_set_shadow_time(struct perf_event
*event
, u64 now
)
598 static inline u64
perf_cgroup_event_time(struct perf_event
*event
)
604 perf_cgroup_defer_enabled(struct perf_event
*event
)
609 perf_cgroup_mark_enabled(struct perf_event
*event
,
610 struct perf_event_context
*ctx
)
615 void perf_pmu_disable(struct pmu
*pmu
)
617 int *count
= this_cpu_ptr(pmu
->pmu_disable_count
);
619 pmu
->pmu_disable(pmu
);
622 void perf_pmu_enable(struct pmu
*pmu
)
624 int *count
= this_cpu_ptr(pmu
->pmu_disable_count
);
626 pmu
->pmu_enable(pmu
);
629 static DEFINE_PER_CPU(struct list_head
, rotation_list
);
632 * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
633 * because they're strictly cpu affine and rotate_start is called with IRQs
634 * disabled, while rotate_context is called from IRQ context.
636 static void perf_pmu_rotate_start(struct pmu
*pmu
)
638 struct perf_cpu_context
*cpuctx
= this_cpu_ptr(pmu
->pmu_cpu_context
);
639 struct list_head
*head
= &__get_cpu_var(rotation_list
);
641 WARN_ON(!irqs_disabled());
643 if (list_empty(&cpuctx
->rotation_list
))
644 list_add(&cpuctx
->rotation_list
, head
);
647 static void get_ctx(struct perf_event_context
*ctx
)
649 WARN_ON(!atomic_inc_not_zero(&ctx
->refcount
));
652 static void put_ctx(struct perf_event_context
*ctx
)
654 if (atomic_dec_and_test(&ctx
->refcount
)) {
656 put_ctx(ctx
->parent_ctx
);
658 put_task_struct(ctx
->task
);
659 kfree_rcu(ctx
, rcu_head
);
663 static void unclone_ctx(struct perf_event_context
*ctx
)
665 if (ctx
->parent_ctx
) {
666 put_ctx(ctx
->parent_ctx
);
667 ctx
->parent_ctx
= NULL
;
671 static u32
perf_event_pid(struct perf_event
*event
, struct task_struct
*p
)
674 * only top level events have the pid namespace they were created in
677 event
= event
->parent
;
679 return task_tgid_nr_ns(p
, event
->ns
);
682 static u32
perf_event_tid(struct perf_event
*event
, struct task_struct
*p
)
685 * only top level events have the pid namespace they were created in
688 event
= event
->parent
;
690 return task_pid_nr_ns(p
, event
->ns
);
694 * If we inherit events we want to return the parent event id
697 static u64
primary_event_id(struct perf_event
*event
)
702 id
= event
->parent
->id
;
708 * Get the perf_event_context for a task and lock it.
709 * This has to cope with with the fact that until it is locked,
710 * the context could get moved to another task.
712 static struct perf_event_context
*
713 perf_lock_task_context(struct task_struct
*task
, int ctxn
, unsigned long *flags
)
715 struct perf_event_context
*ctx
;
719 ctx
= rcu_dereference(task
->perf_event_ctxp
[ctxn
]);
722 * If this context is a clone of another, it might
723 * get swapped for another underneath us by
724 * perf_event_task_sched_out, though the
725 * rcu_read_lock() protects us from any context
726 * getting freed. Lock the context and check if it
727 * got swapped before we could get the lock, and retry
728 * if so. If we locked the right context, then it
729 * can't get swapped on us any more.
731 raw_spin_lock_irqsave(&ctx
->lock
, *flags
);
732 if (ctx
!= rcu_dereference(task
->perf_event_ctxp
[ctxn
])) {
733 raw_spin_unlock_irqrestore(&ctx
->lock
, *flags
);
737 if (!atomic_inc_not_zero(&ctx
->refcount
)) {
738 raw_spin_unlock_irqrestore(&ctx
->lock
, *flags
);
747 * Get the context for a task and increment its pin_count so it
748 * can't get swapped to another task. This also increments its
749 * reference count so that the context can't get freed.
751 static struct perf_event_context
*
752 perf_pin_task_context(struct task_struct
*task
, int ctxn
)
754 struct perf_event_context
*ctx
;
757 ctx
= perf_lock_task_context(task
, ctxn
, &flags
);
760 raw_spin_unlock_irqrestore(&ctx
->lock
, flags
);
765 static void perf_unpin_context(struct perf_event_context
*ctx
)
769 raw_spin_lock_irqsave(&ctx
->lock
, flags
);
771 raw_spin_unlock_irqrestore(&ctx
->lock
, flags
);
775 * Update the record of the current time in a context.
777 static void update_context_time(struct perf_event_context
*ctx
)
779 u64 now
= perf_clock();
781 ctx
->time
+= now
- ctx
->timestamp
;
782 ctx
->timestamp
= now
;
785 static u64
perf_event_time(struct perf_event
*event
)
787 struct perf_event_context
*ctx
= event
->ctx
;
789 if (is_cgroup_event(event
))
790 return perf_cgroup_event_time(event
);
792 return ctx
? ctx
->time
: 0;
796 * Update the total_time_enabled and total_time_running fields for a event.
797 * The caller of this function needs to hold the ctx->lock.
799 static void update_event_times(struct perf_event
*event
)
801 struct perf_event_context
*ctx
= event
->ctx
;
804 if (event
->state
< PERF_EVENT_STATE_INACTIVE
||
805 event
->group_leader
->state
< PERF_EVENT_STATE_INACTIVE
)
808 * in cgroup mode, time_enabled represents
809 * the time the event was enabled AND active
810 * tasks were in the monitored cgroup. This is
811 * independent of the activity of the context as
812 * there may be a mix of cgroup and non-cgroup events.
814 * That is why we treat cgroup events differently
817 if (is_cgroup_event(event
))
818 run_end
= perf_event_time(event
);
819 else if (ctx
->is_active
)
822 run_end
= event
->tstamp_stopped
;
824 event
->total_time_enabled
= run_end
- event
->tstamp_enabled
;
826 if (event
->state
== PERF_EVENT_STATE_INACTIVE
)
827 run_end
= event
->tstamp_stopped
;
829 run_end
= perf_event_time(event
);
831 event
->total_time_running
= run_end
- event
->tstamp_running
;
836 * Update total_time_enabled and total_time_running for all events in a group.
838 static void update_group_times(struct perf_event
*leader
)
840 struct perf_event
*event
;
842 update_event_times(leader
);
843 list_for_each_entry(event
, &leader
->sibling_list
, group_entry
)
844 update_event_times(event
);
847 static struct list_head
*
848 ctx_group_list(struct perf_event
*event
, struct perf_event_context
*ctx
)
850 if (event
->attr
.pinned
)
851 return &ctx
->pinned_groups
;
853 return &ctx
->flexible_groups
;
857 * Add a event from the lists for its context.
858 * Must be called with ctx->mutex and ctx->lock held.
861 list_add_event(struct perf_event
*event
, struct perf_event_context
*ctx
)
863 WARN_ON_ONCE(event
->attach_state
& PERF_ATTACH_CONTEXT
);
864 event
->attach_state
|= PERF_ATTACH_CONTEXT
;
867 * If we're a stand alone event or group leader, we go to the context
868 * list, group events are kept attached to the group so that
869 * perf_group_detach can, at all times, locate all siblings.
871 if (event
->group_leader
== event
) {
872 struct list_head
*list
;
874 if (is_software_event(event
))
875 event
->group_flags
|= PERF_GROUP_SOFTWARE
;
877 list
= ctx_group_list(event
, ctx
);
878 list_add_tail(&event
->group_entry
, list
);
881 if (is_cgroup_event(event
))
884 list_add_rcu(&event
->event_entry
, &ctx
->event_list
);
886 perf_pmu_rotate_start(ctx
->pmu
);
888 if (event
->attr
.inherit_stat
)
893 * Called at perf_event creation and when events are attached/detached from a
896 static void perf_event__read_size(struct perf_event
*event
)
898 int entry
= sizeof(u64
); /* value */
902 if (event
->attr
.read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
905 if (event
->attr
.read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
908 if (event
->attr
.read_format
& PERF_FORMAT_ID
)
909 entry
+= sizeof(u64
);
911 if (event
->attr
.read_format
& PERF_FORMAT_GROUP
) {
912 nr
+= event
->group_leader
->nr_siblings
;
917 event
->read_size
= size
;
920 static void perf_event__header_size(struct perf_event
*event
)
922 struct perf_sample_data
*data
;
923 u64 sample_type
= event
->attr
.sample_type
;
926 perf_event__read_size(event
);
928 if (sample_type
& PERF_SAMPLE_IP
)
929 size
+= sizeof(data
->ip
);
931 if (sample_type
& PERF_SAMPLE_ADDR
)
932 size
+= sizeof(data
->addr
);
934 if (sample_type
& PERF_SAMPLE_PERIOD
)
935 size
+= sizeof(data
->period
);
937 if (sample_type
& PERF_SAMPLE_READ
)
938 size
+= event
->read_size
;
940 event
->header_size
= size
;
943 static void perf_event__id_header_size(struct perf_event
*event
)
945 struct perf_sample_data
*data
;
946 u64 sample_type
= event
->attr
.sample_type
;
949 if (sample_type
& PERF_SAMPLE_TID
)
950 size
+= sizeof(data
->tid_entry
);
952 if (sample_type
& PERF_SAMPLE_TIME
)
953 size
+= sizeof(data
->time
);
955 if (sample_type
& PERF_SAMPLE_ID
)
956 size
+= sizeof(data
->id
);
958 if (sample_type
& PERF_SAMPLE_STREAM_ID
)
959 size
+= sizeof(data
->stream_id
);
961 if (sample_type
& PERF_SAMPLE_CPU
)
962 size
+= sizeof(data
->cpu_entry
);
964 event
->id_header_size
= size
;
967 static void perf_group_attach(struct perf_event
*event
)
969 struct perf_event
*group_leader
= event
->group_leader
, *pos
;
972 * We can have double attach due to group movement in perf_event_open.
974 if (event
->attach_state
& PERF_ATTACH_GROUP
)
977 event
->attach_state
|= PERF_ATTACH_GROUP
;
979 if (group_leader
== event
)
982 if (group_leader
->group_flags
& PERF_GROUP_SOFTWARE
&&
983 !is_software_event(event
))
984 group_leader
->group_flags
&= ~PERF_GROUP_SOFTWARE
;
986 list_add_tail(&event
->group_entry
, &group_leader
->sibling_list
);
987 group_leader
->nr_siblings
++;
989 perf_event__header_size(group_leader
);
991 list_for_each_entry(pos
, &group_leader
->sibling_list
, group_entry
)
992 perf_event__header_size(pos
);
996 * Remove a event from the lists for its context.
997 * Must be called with ctx->mutex and ctx->lock held.
1000 list_del_event(struct perf_event
*event
, struct perf_event_context
*ctx
)
1002 struct perf_cpu_context
*cpuctx
;
1004 * We can have double detach due to exit/hot-unplug + close.
1006 if (!(event
->attach_state
& PERF_ATTACH_CONTEXT
))
1009 event
->attach_state
&= ~PERF_ATTACH_CONTEXT
;
1011 if (is_cgroup_event(event
)) {
1013 cpuctx
= __get_cpu_context(ctx
);
1015 * if there are no more cgroup events
1016 * then cler cgrp to avoid stale pointer
1017 * in update_cgrp_time_from_cpuctx()
1019 if (!ctx
->nr_cgroups
)
1020 cpuctx
->cgrp
= NULL
;
1024 if (event
->attr
.inherit_stat
)
1027 list_del_rcu(&event
->event_entry
);
1029 if (event
->group_leader
== event
)
1030 list_del_init(&event
->group_entry
);
1032 update_group_times(event
);
1035 * If event was in error state, then keep it
1036 * that way, otherwise bogus counts will be
1037 * returned on read(). The only way to get out
1038 * of error state is by explicit re-enabling
1041 if (event
->state
> PERF_EVENT_STATE_OFF
)
1042 event
->state
= PERF_EVENT_STATE_OFF
;
1045 static void perf_group_detach(struct perf_event
*event
)
1047 struct perf_event
*sibling
, *tmp
;
1048 struct list_head
*list
= NULL
;
1051 * We can have double detach due to exit/hot-unplug + close.
1053 if (!(event
->attach_state
& PERF_ATTACH_GROUP
))
1056 event
->attach_state
&= ~PERF_ATTACH_GROUP
;
1059 * If this is a sibling, remove it from its group.
1061 if (event
->group_leader
!= event
) {
1062 list_del_init(&event
->group_entry
);
1063 event
->group_leader
->nr_siblings
--;
1067 if (!list_empty(&event
->group_entry
))
1068 list
= &event
->group_entry
;
1071 * If this was a group event with sibling events then
1072 * upgrade the siblings to singleton events by adding them
1073 * to whatever list we are on.
1075 list_for_each_entry_safe(sibling
, tmp
, &event
->sibling_list
, group_entry
) {
1077 list_move_tail(&sibling
->group_entry
, list
);
1078 sibling
->group_leader
= sibling
;
1080 /* Inherit group flags from the previous leader */
1081 sibling
->group_flags
= event
->group_flags
;
1085 perf_event__header_size(event
->group_leader
);
1087 list_for_each_entry(tmp
, &event
->group_leader
->sibling_list
, group_entry
)
1088 perf_event__header_size(tmp
);
1092 event_filter_match(struct perf_event
*event
)
1094 return (event
->cpu
== -1 || event
->cpu
== smp_processor_id())
1095 && perf_cgroup_match(event
);
1099 event_sched_out(struct perf_event
*event
,
1100 struct perf_cpu_context
*cpuctx
,
1101 struct perf_event_context
*ctx
)
1103 u64 tstamp
= perf_event_time(event
);
1106 * An event which could not be activated because of
1107 * filter mismatch still needs to have its timings
1108 * maintained, otherwise bogus information is return
1109 * via read() for time_enabled, time_running:
1111 if (event
->state
== PERF_EVENT_STATE_INACTIVE
1112 && !event_filter_match(event
)) {
1113 delta
= tstamp
- event
->tstamp_stopped
;
1114 event
->tstamp_running
+= delta
;
1115 event
->tstamp_stopped
= tstamp
;
1118 if (event
->state
!= PERF_EVENT_STATE_ACTIVE
)
1121 event
->state
= PERF_EVENT_STATE_INACTIVE
;
1122 if (event
->pending_disable
) {
1123 event
->pending_disable
= 0;
1124 event
->state
= PERF_EVENT_STATE_OFF
;
1126 event
->tstamp_stopped
= tstamp
;
1127 event
->pmu
->del(event
, 0);
1130 if (!is_software_event(event
))
1131 cpuctx
->active_oncpu
--;
1133 if (event
->attr
.exclusive
|| !cpuctx
->active_oncpu
)
1134 cpuctx
->exclusive
= 0;
1138 group_sched_out(struct perf_event
*group_event
,
1139 struct perf_cpu_context
*cpuctx
,
1140 struct perf_event_context
*ctx
)
1142 struct perf_event
*event
;
1143 int state
= group_event
->state
;
1145 event_sched_out(group_event
, cpuctx
, ctx
);
1148 * Schedule out siblings (if any):
1150 list_for_each_entry(event
, &group_event
->sibling_list
, group_entry
)
1151 event_sched_out(event
, cpuctx
, ctx
);
1153 if (state
== PERF_EVENT_STATE_ACTIVE
&& group_event
->attr
.exclusive
)
1154 cpuctx
->exclusive
= 0;
1158 * Cross CPU call to remove a performance event
1160 * We disable the event on the hardware level first. After that we
1161 * remove it from the context list.
1163 static int __perf_remove_from_context(void *info
)
1165 struct perf_event
*event
= info
;
1166 struct perf_event_context
*ctx
= event
->ctx
;
1167 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
1169 raw_spin_lock(&ctx
->lock
);
1170 event_sched_out(event
, cpuctx
, ctx
);
1171 list_del_event(event
, ctx
);
1172 if (!ctx
->nr_events
&& cpuctx
->task_ctx
== ctx
) {
1174 cpuctx
->task_ctx
= NULL
;
1176 raw_spin_unlock(&ctx
->lock
);
1183 * Remove the event from a task's (or a CPU's) list of events.
1185 * CPU events are removed with a smp call. For task events we only
1186 * call when the task is on a CPU.
1188 * If event->ctx is a cloned context, callers must make sure that
1189 * every task struct that event->ctx->task could possibly point to
1190 * remains valid. This is OK when called from perf_release since
1191 * that only calls us on the top-level context, which can't be a clone.
1192 * When called from perf_event_exit_task, it's OK because the
1193 * context has been detached from its task.
1195 static void perf_remove_from_context(struct perf_event
*event
)
1197 struct perf_event_context
*ctx
= event
->ctx
;
1198 struct task_struct
*task
= ctx
->task
;
1200 lockdep_assert_held(&ctx
->mutex
);
1204 * Per cpu events are removed via an smp call and
1205 * the removal is always successful.
1207 cpu_function_call(event
->cpu
, __perf_remove_from_context
, event
);
1212 if (!task_function_call(task
, __perf_remove_from_context
, event
))
1215 raw_spin_lock_irq(&ctx
->lock
);
1217 * If we failed to find a running task, but find the context active now
1218 * that we've acquired the ctx->lock, retry.
1220 if (ctx
->is_active
) {
1221 raw_spin_unlock_irq(&ctx
->lock
);
1226 * Since the task isn't running, its safe to remove the event, us
1227 * holding the ctx->lock ensures the task won't get scheduled in.
1229 list_del_event(event
, ctx
);
1230 raw_spin_unlock_irq(&ctx
->lock
);
1234 * Cross CPU call to disable a performance event
1236 static int __perf_event_disable(void *info
)
1238 struct perf_event
*event
= info
;
1239 struct perf_event_context
*ctx
= event
->ctx
;
1240 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
1243 * If this is a per-task event, need to check whether this
1244 * event's task is the current task on this cpu.
1246 * Can trigger due to concurrent perf_event_context_sched_out()
1247 * flipping contexts around.
1249 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
)
1252 raw_spin_lock(&ctx
->lock
);
1255 * If the event is on, turn it off.
1256 * If it is in error state, leave it in error state.
1258 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
) {
1259 update_context_time(ctx
);
1260 update_cgrp_time_from_event(event
);
1261 update_group_times(event
);
1262 if (event
== event
->group_leader
)
1263 group_sched_out(event
, cpuctx
, ctx
);
1265 event_sched_out(event
, cpuctx
, ctx
);
1266 event
->state
= PERF_EVENT_STATE_OFF
;
1269 raw_spin_unlock(&ctx
->lock
);
1277 * If event->ctx is a cloned context, callers must make sure that
1278 * every task struct that event->ctx->task could possibly point to
1279 * remains valid. This condition is satisifed when called through
1280 * perf_event_for_each_child or perf_event_for_each because they
1281 * hold the top-level event's child_mutex, so any descendant that
1282 * goes to exit will block in sync_child_event.
1283 * When called from perf_pending_event it's OK because event->ctx
1284 * is the current context on this CPU and preemption is disabled,
1285 * hence we can't get into perf_event_task_sched_out for this context.
1287 void perf_event_disable(struct perf_event
*event
)
1289 struct perf_event_context
*ctx
= event
->ctx
;
1290 struct task_struct
*task
= ctx
->task
;
1294 * Disable the event on the cpu that it's on
1296 cpu_function_call(event
->cpu
, __perf_event_disable
, event
);
1301 if (!task_function_call(task
, __perf_event_disable
, event
))
1304 raw_spin_lock_irq(&ctx
->lock
);
1306 * If the event is still active, we need to retry the cross-call.
1308 if (event
->state
== PERF_EVENT_STATE_ACTIVE
) {
1309 raw_spin_unlock_irq(&ctx
->lock
);
1311 * Reload the task pointer, it might have been changed by
1312 * a concurrent perf_event_context_sched_out().
1319 * Since we have the lock this context can't be scheduled
1320 * in, so we can change the state safely.
1322 if (event
->state
== PERF_EVENT_STATE_INACTIVE
) {
1323 update_group_times(event
);
1324 event
->state
= PERF_EVENT_STATE_OFF
;
1326 raw_spin_unlock_irq(&ctx
->lock
);
1329 static void perf_set_shadow_time(struct perf_event
*event
,
1330 struct perf_event_context
*ctx
,
1334 * use the correct time source for the time snapshot
1336 * We could get by without this by leveraging the
1337 * fact that to get to this function, the caller
1338 * has most likely already called update_context_time()
1339 * and update_cgrp_time_xx() and thus both timestamp
1340 * are identical (or very close). Given that tstamp is,
1341 * already adjusted for cgroup, we could say that:
1342 * tstamp - ctx->timestamp
1344 * tstamp - cgrp->timestamp.
1346 * Then, in perf_output_read(), the calculation would
1347 * work with no changes because:
1348 * - event is guaranteed scheduled in
1349 * - no scheduled out in between
1350 * - thus the timestamp would be the same
1352 * But this is a bit hairy.
1354 * So instead, we have an explicit cgroup call to remain
1355 * within the time time source all along. We believe it
1356 * is cleaner and simpler to understand.
1358 if (is_cgroup_event(event
))
1359 perf_cgroup_set_shadow_time(event
, tstamp
);
1361 event
->shadow_ctx_time
= tstamp
- ctx
->timestamp
;
1364 #define MAX_INTERRUPTS (~0ULL)
1366 static void perf_log_throttle(struct perf_event
*event
, int enable
);
1369 event_sched_in(struct perf_event
*event
,
1370 struct perf_cpu_context
*cpuctx
,
1371 struct perf_event_context
*ctx
)
1373 u64 tstamp
= perf_event_time(event
);
1375 if (event
->state
<= PERF_EVENT_STATE_OFF
)
1378 event
->state
= PERF_EVENT_STATE_ACTIVE
;
1379 event
->oncpu
= smp_processor_id();
1382 * Unthrottle events, since we scheduled we might have missed several
1383 * ticks already, also for a heavily scheduling task there is little
1384 * guarantee it'll get a tick in a timely manner.
1386 if (unlikely(event
->hw
.interrupts
== MAX_INTERRUPTS
)) {
1387 perf_log_throttle(event
, 1);
1388 event
->hw
.interrupts
= 0;
1392 * The new state must be visible before we turn it on in the hardware:
1396 if (event
->pmu
->add(event
, PERF_EF_START
)) {
1397 event
->state
= PERF_EVENT_STATE_INACTIVE
;
1402 event
->tstamp_running
+= tstamp
- event
->tstamp_stopped
;
1404 perf_set_shadow_time(event
, ctx
, tstamp
);
1406 if (!is_software_event(event
))
1407 cpuctx
->active_oncpu
++;
1410 if (event
->attr
.exclusive
)
1411 cpuctx
->exclusive
= 1;
1417 group_sched_in(struct perf_event
*group_event
,
1418 struct perf_cpu_context
*cpuctx
,
1419 struct perf_event_context
*ctx
)
1421 struct perf_event
*event
, *partial_group
= NULL
;
1422 struct pmu
*pmu
= group_event
->pmu
;
1423 u64 now
= ctx
->time
;
1424 bool simulate
= false;
1426 if (group_event
->state
== PERF_EVENT_STATE_OFF
)
1429 pmu
->start_txn(pmu
);
1431 if (event_sched_in(group_event
, cpuctx
, ctx
)) {
1432 pmu
->cancel_txn(pmu
);
1437 * Schedule in siblings as one group (if any):
1439 list_for_each_entry(event
, &group_event
->sibling_list
, group_entry
) {
1440 if (event_sched_in(event
, cpuctx
, ctx
)) {
1441 partial_group
= event
;
1446 if (!pmu
->commit_txn(pmu
))
1451 * Groups can be scheduled in as one unit only, so undo any
1452 * partial group before returning:
1453 * The events up to the failed event are scheduled out normally,
1454 * tstamp_stopped will be updated.
1456 * The failed events and the remaining siblings need to have
1457 * their timings updated as if they had gone thru event_sched_in()
1458 * and event_sched_out(). This is required to get consistent timings
1459 * across the group. This also takes care of the case where the group
1460 * could never be scheduled by ensuring tstamp_stopped is set to mark
1461 * the time the event was actually stopped, such that time delta
1462 * calculation in update_event_times() is correct.
1464 list_for_each_entry(event
, &group_event
->sibling_list
, group_entry
) {
1465 if (event
== partial_group
)
1469 event
->tstamp_running
+= now
- event
->tstamp_stopped
;
1470 event
->tstamp_stopped
= now
;
1472 event_sched_out(event
, cpuctx
, ctx
);
1475 event_sched_out(group_event
, cpuctx
, ctx
);
1477 pmu
->cancel_txn(pmu
);
1483 * Work out whether we can put this event group on the CPU now.
1485 static int group_can_go_on(struct perf_event
*event
,
1486 struct perf_cpu_context
*cpuctx
,
1490 * Groups consisting entirely of software events can always go on.
1492 if (event
->group_flags
& PERF_GROUP_SOFTWARE
)
1495 * If an exclusive group is already on, no other hardware
1498 if (cpuctx
->exclusive
)
1501 * If this group is exclusive and there are already
1502 * events on the CPU, it can't go on.
1504 if (event
->attr
.exclusive
&& cpuctx
->active_oncpu
)
1507 * Otherwise, try to add it if all previous groups were able
1513 static void add_event_to_ctx(struct perf_event
*event
,
1514 struct perf_event_context
*ctx
)
1516 u64 tstamp
= perf_event_time(event
);
1518 list_add_event(event
, ctx
);
1519 perf_group_attach(event
);
1520 event
->tstamp_enabled
= tstamp
;
1521 event
->tstamp_running
= tstamp
;
1522 event
->tstamp_stopped
= tstamp
;
1525 static void task_ctx_sched_out(struct perf_event_context
*ctx
);
1527 ctx_sched_in(struct perf_event_context
*ctx
,
1528 struct perf_cpu_context
*cpuctx
,
1529 enum event_type_t event_type
,
1530 struct task_struct
*task
);
1532 static void perf_event_sched_in(struct perf_cpu_context
*cpuctx
,
1533 struct perf_event_context
*ctx
,
1534 struct task_struct
*task
)
1536 cpu_ctx_sched_in(cpuctx
, EVENT_PINNED
, task
);
1538 ctx_sched_in(ctx
, cpuctx
, EVENT_PINNED
, task
);
1539 cpu_ctx_sched_in(cpuctx
, EVENT_FLEXIBLE
, task
);
1541 ctx_sched_in(ctx
, cpuctx
, EVENT_FLEXIBLE
, task
);
1545 * Cross CPU call to install and enable a performance event
1547 * Must be called with ctx->mutex held
1549 static int __perf_install_in_context(void *info
)
1551 struct perf_event
*event
= info
;
1552 struct perf_event_context
*ctx
= event
->ctx
;
1553 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
1554 struct perf_event_context
*task_ctx
= cpuctx
->task_ctx
;
1555 struct task_struct
*task
= current
;
1557 perf_ctx_lock(cpuctx
, task_ctx
);
1558 perf_pmu_disable(cpuctx
->ctx
.pmu
);
1561 * If there was an active task_ctx schedule it out.
1564 task_ctx_sched_out(task_ctx
);
1567 * If the context we're installing events in is not the
1568 * active task_ctx, flip them.
1570 if (ctx
->task
&& task_ctx
!= ctx
) {
1572 raw_spin_unlock(&task_ctx
->lock
);
1573 raw_spin_lock(&ctx
->lock
);
1578 cpuctx
->task_ctx
= task_ctx
;
1579 task
= task_ctx
->task
;
1582 cpu_ctx_sched_out(cpuctx
, EVENT_ALL
);
1584 update_context_time(ctx
);
1586 * update cgrp time only if current cgrp
1587 * matches event->cgrp. Must be done before
1588 * calling add_event_to_ctx()
1590 update_cgrp_time_from_event(event
);
1592 add_event_to_ctx(event
, ctx
);
1595 * Schedule everything back in
1597 perf_event_sched_in(cpuctx
, task_ctx
, task
);
1599 perf_pmu_enable(cpuctx
->ctx
.pmu
);
1600 perf_ctx_unlock(cpuctx
, task_ctx
);
1606 * Attach a performance event to a context
1608 * First we add the event to the list with the hardware enable bit
1609 * in event->hw_config cleared.
1611 * If the event is attached to a task which is on a CPU we use a smp
1612 * call to enable it in the task context. The task might have been
1613 * scheduled away, but we check this in the smp call again.
1616 perf_install_in_context(struct perf_event_context
*ctx
,
1617 struct perf_event
*event
,
1620 struct task_struct
*task
= ctx
->task
;
1622 lockdep_assert_held(&ctx
->mutex
);
1628 * Per cpu events are installed via an smp call and
1629 * the install is always successful.
1631 cpu_function_call(cpu
, __perf_install_in_context
, event
);
1636 if (!task_function_call(task
, __perf_install_in_context
, event
))
1639 raw_spin_lock_irq(&ctx
->lock
);
1641 * If we failed to find a running task, but find the context active now
1642 * that we've acquired the ctx->lock, retry.
1644 if (ctx
->is_active
) {
1645 raw_spin_unlock_irq(&ctx
->lock
);
1650 * Since the task isn't running, its safe to add the event, us holding
1651 * the ctx->lock ensures the task won't get scheduled in.
1653 add_event_to_ctx(event
, ctx
);
1654 raw_spin_unlock_irq(&ctx
->lock
);
1658 * Put a event into inactive state and update time fields.
1659 * Enabling the leader of a group effectively enables all
1660 * the group members that aren't explicitly disabled, so we
1661 * have to update their ->tstamp_enabled also.
1662 * Note: this works for group members as well as group leaders
1663 * since the non-leader members' sibling_lists will be empty.
1665 static void __perf_event_mark_enabled(struct perf_event
*event
,
1666 struct perf_event_context
*ctx
)
1668 struct perf_event
*sub
;
1669 u64 tstamp
= perf_event_time(event
);
1671 event
->state
= PERF_EVENT_STATE_INACTIVE
;
1672 event
->tstamp_enabled
= tstamp
- event
->total_time_enabled
;
1673 list_for_each_entry(sub
, &event
->sibling_list
, group_entry
) {
1674 if (sub
->state
>= PERF_EVENT_STATE_INACTIVE
)
1675 sub
->tstamp_enabled
= tstamp
- sub
->total_time_enabled
;
1680 * Cross CPU call to enable a performance event
1682 static int __perf_event_enable(void *info
)
1684 struct perf_event
*event
= info
;
1685 struct perf_event_context
*ctx
= event
->ctx
;
1686 struct perf_event
*leader
= event
->group_leader
;
1687 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
1690 if (WARN_ON_ONCE(!ctx
->is_active
))
1693 raw_spin_lock(&ctx
->lock
);
1694 update_context_time(ctx
);
1696 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
)
1700 * set current task's cgroup time reference point
1702 perf_cgroup_set_timestamp(current
, ctx
);
1704 __perf_event_mark_enabled(event
, ctx
);
1706 if (!event_filter_match(event
)) {
1707 if (is_cgroup_event(event
))
1708 perf_cgroup_defer_enabled(event
);
1713 * If the event is in a group and isn't the group leader,
1714 * then don't put it on unless the group is on.
1716 if (leader
!= event
&& leader
->state
!= PERF_EVENT_STATE_ACTIVE
)
1719 if (!group_can_go_on(event
, cpuctx
, 1)) {
1722 if (event
== leader
)
1723 err
= group_sched_in(event
, cpuctx
, ctx
);
1725 err
= event_sched_in(event
, cpuctx
, ctx
);
1730 * If this event can't go on and it's part of a
1731 * group, then the whole group has to come off.
1733 if (leader
!= event
)
1734 group_sched_out(leader
, cpuctx
, ctx
);
1735 if (leader
->attr
.pinned
) {
1736 update_group_times(leader
);
1737 leader
->state
= PERF_EVENT_STATE_ERROR
;
1742 raw_spin_unlock(&ctx
->lock
);
1750 * If event->ctx is a cloned context, callers must make sure that
1751 * every task struct that event->ctx->task could possibly point to
1752 * remains valid. This condition is satisfied when called through
1753 * perf_event_for_each_child or perf_event_for_each as described
1754 * for perf_event_disable.
1756 void perf_event_enable(struct perf_event
*event
)
1758 struct perf_event_context
*ctx
= event
->ctx
;
1759 struct task_struct
*task
= ctx
->task
;
1763 * Enable the event on the cpu that it's on
1765 cpu_function_call(event
->cpu
, __perf_event_enable
, event
);
1769 raw_spin_lock_irq(&ctx
->lock
);
1770 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
)
1774 * If the event is in error state, clear that first.
1775 * That way, if we see the event in error state below, we
1776 * know that it has gone back into error state, as distinct
1777 * from the task having been scheduled away before the
1778 * cross-call arrived.
1780 if (event
->state
== PERF_EVENT_STATE_ERROR
)
1781 event
->state
= PERF_EVENT_STATE_OFF
;
1784 if (!ctx
->is_active
) {
1785 __perf_event_mark_enabled(event
, ctx
);
1789 raw_spin_unlock_irq(&ctx
->lock
);
1791 if (!task_function_call(task
, __perf_event_enable
, event
))
1794 raw_spin_lock_irq(&ctx
->lock
);
1797 * If the context is active and the event is still off,
1798 * we need to retry the cross-call.
1800 if (ctx
->is_active
&& event
->state
== PERF_EVENT_STATE_OFF
) {
1802 * task could have been flipped by a concurrent
1803 * perf_event_context_sched_out()
1810 raw_spin_unlock_irq(&ctx
->lock
);
1813 int perf_event_refresh(struct perf_event
*event
, int refresh
)
1816 * not supported on inherited events
1818 if (event
->attr
.inherit
|| !is_sampling_event(event
))
1821 atomic_add(refresh
, &event
->event_limit
);
1822 perf_event_enable(event
);
1826 EXPORT_SYMBOL_GPL(perf_event_refresh
);
1828 static void ctx_sched_out(struct perf_event_context
*ctx
,
1829 struct perf_cpu_context
*cpuctx
,
1830 enum event_type_t event_type
)
1832 struct perf_event
*event
;
1833 int is_active
= ctx
->is_active
;
1835 ctx
->is_active
&= ~event_type
;
1836 if (likely(!ctx
->nr_events
))
1839 update_context_time(ctx
);
1840 update_cgrp_time_from_cpuctx(cpuctx
);
1841 if (!ctx
->nr_active
)
1844 perf_pmu_disable(ctx
->pmu
);
1845 if ((is_active
& EVENT_PINNED
) && (event_type
& EVENT_PINNED
)) {
1846 list_for_each_entry(event
, &ctx
->pinned_groups
, group_entry
)
1847 group_sched_out(event
, cpuctx
, ctx
);
1850 if ((is_active
& EVENT_FLEXIBLE
) && (event_type
& EVENT_FLEXIBLE
)) {
1851 list_for_each_entry(event
, &ctx
->flexible_groups
, group_entry
)
1852 group_sched_out(event
, cpuctx
, ctx
);
1854 perf_pmu_enable(ctx
->pmu
);
1858 * Test whether two contexts are equivalent, i.e. whether they
1859 * have both been cloned from the same version of the same context
1860 * and they both have the same number of enabled events.
1861 * If the number of enabled events is the same, then the set
1862 * of enabled events should be the same, because these are both
1863 * inherited contexts, therefore we can't access individual events
1864 * in them directly with an fd; we can only enable/disable all
1865 * events via prctl, or enable/disable all events in a family
1866 * via ioctl, which will have the same effect on both contexts.
1868 static int context_equiv(struct perf_event_context
*ctx1
,
1869 struct perf_event_context
*ctx2
)
1871 return ctx1
->parent_ctx
&& ctx1
->parent_ctx
== ctx2
->parent_ctx
1872 && ctx1
->parent_gen
== ctx2
->parent_gen
1873 && !ctx1
->pin_count
&& !ctx2
->pin_count
;
1876 static void __perf_event_sync_stat(struct perf_event
*event
,
1877 struct perf_event
*next_event
)
1881 if (!event
->attr
.inherit_stat
)
1885 * Update the event value, we cannot use perf_event_read()
1886 * because we're in the middle of a context switch and have IRQs
1887 * disabled, which upsets smp_call_function_single(), however
1888 * we know the event must be on the current CPU, therefore we
1889 * don't need to use it.
1891 switch (event
->state
) {
1892 case PERF_EVENT_STATE_ACTIVE
:
1893 event
->pmu
->read(event
);
1896 case PERF_EVENT_STATE_INACTIVE
:
1897 update_event_times(event
);
1905 * In order to keep per-task stats reliable we need to flip the event
1906 * values when we flip the contexts.
1908 value
= local64_read(&next_event
->count
);
1909 value
= local64_xchg(&event
->count
, value
);
1910 local64_set(&next_event
->count
, value
);
1912 swap(event
->total_time_enabled
, next_event
->total_time_enabled
);
1913 swap(event
->total_time_running
, next_event
->total_time_running
);
1916 * Since we swizzled the values, update the user visible data too.
1918 perf_event_update_userpage(event
);
1919 perf_event_update_userpage(next_event
);
1922 #define list_next_entry(pos, member) \
1923 list_entry(pos->member.next, typeof(*pos), member)
1925 static void perf_event_sync_stat(struct perf_event_context
*ctx
,
1926 struct perf_event_context
*next_ctx
)
1928 struct perf_event
*event
, *next_event
;
1933 update_context_time(ctx
);
1935 event
= list_first_entry(&ctx
->event_list
,
1936 struct perf_event
, event_entry
);
1938 next_event
= list_first_entry(&next_ctx
->event_list
,
1939 struct perf_event
, event_entry
);
1941 while (&event
->event_entry
!= &ctx
->event_list
&&
1942 &next_event
->event_entry
!= &next_ctx
->event_list
) {
1944 __perf_event_sync_stat(event
, next_event
);
1946 event
= list_next_entry(event
, event_entry
);
1947 next_event
= list_next_entry(next_event
, event_entry
);
1951 static void perf_event_context_sched_out(struct task_struct
*task
, int ctxn
,
1952 struct task_struct
*next
)
1954 struct perf_event_context
*ctx
= task
->perf_event_ctxp
[ctxn
];
1955 struct perf_event_context
*next_ctx
;
1956 struct perf_event_context
*parent
;
1957 struct perf_cpu_context
*cpuctx
;
1963 cpuctx
= __get_cpu_context(ctx
);
1964 if (!cpuctx
->task_ctx
)
1968 parent
= rcu_dereference(ctx
->parent_ctx
);
1969 next_ctx
= next
->perf_event_ctxp
[ctxn
];
1970 if (parent
&& next_ctx
&&
1971 rcu_dereference(next_ctx
->parent_ctx
) == parent
) {
1973 * Looks like the two contexts are clones, so we might be
1974 * able to optimize the context switch. We lock both
1975 * contexts and check that they are clones under the
1976 * lock (including re-checking that neither has been
1977 * uncloned in the meantime). It doesn't matter which
1978 * order we take the locks because no other cpu could
1979 * be trying to lock both of these tasks.
1981 raw_spin_lock(&ctx
->lock
);
1982 raw_spin_lock_nested(&next_ctx
->lock
, SINGLE_DEPTH_NESTING
);
1983 if (context_equiv(ctx
, next_ctx
)) {
1985 * XXX do we need a memory barrier of sorts
1986 * wrt to rcu_dereference() of perf_event_ctxp
1988 task
->perf_event_ctxp
[ctxn
] = next_ctx
;
1989 next
->perf_event_ctxp
[ctxn
] = ctx
;
1991 next_ctx
->task
= task
;
1994 perf_event_sync_stat(ctx
, next_ctx
);
1996 raw_spin_unlock(&next_ctx
->lock
);
1997 raw_spin_unlock(&ctx
->lock
);
2002 raw_spin_lock(&ctx
->lock
);
2003 ctx_sched_out(ctx
, cpuctx
, EVENT_ALL
);
2004 cpuctx
->task_ctx
= NULL
;
2005 raw_spin_unlock(&ctx
->lock
);
2009 #define for_each_task_context_nr(ctxn) \
2010 for ((ctxn) = 0; (ctxn) < perf_nr_task_contexts; (ctxn)++)
2013 * Called from scheduler to remove the events of the current task,
2014 * with interrupts disabled.
2016 * We stop each event and update the event value in event->count.
2018 * This does not protect us against NMI, but disable()
2019 * sets the disabled bit in the control field of event _before_
2020 * accessing the event control register. If a NMI hits, then it will
2021 * not restart the event.
2023 void __perf_event_task_sched_out(struct task_struct
*task
,
2024 struct task_struct
*next
)
2028 for_each_task_context_nr(ctxn
)
2029 perf_event_context_sched_out(task
, ctxn
, next
);
2032 * if cgroup events exist on this CPU, then we need
2033 * to check if we have to switch out PMU state.
2034 * cgroup event are system-wide mode only
2036 if (atomic_read(&__get_cpu_var(perf_cgroup_events
)))
2037 perf_cgroup_sched_out(task
, next
);
2040 static void task_ctx_sched_out(struct perf_event_context
*ctx
)
2042 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
2044 if (!cpuctx
->task_ctx
)
2047 if (WARN_ON_ONCE(ctx
!= cpuctx
->task_ctx
))
2050 ctx_sched_out(ctx
, cpuctx
, EVENT_ALL
);
2051 cpuctx
->task_ctx
= NULL
;
2055 * Called with IRQs disabled
2057 static void cpu_ctx_sched_out(struct perf_cpu_context
*cpuctx
,
2058 enum event_type_t event_type
)
2060 ctx_sched_out(&cpuctx
->ctx
, cpuctx
, event_type
);
2064 ctx_pinned_sched_in(struct perf_event_context
*ctx
,
2065 struct perf_cpu_context
*cpuctx
)
2067 struct perf_event
*event
;
2069 list_for_each_entry(event
, &ctx
->pinned_groups
, group_entry
) {
2070 if (event
->state
<= PERF_EVENT_STATE_OFF
)
2072 if (!event_filter_match(event
))
2075 /* may need to reset tstamp_enabled */
2076 if (is_cgroup_event(event
))
2077 perf_cgroup_mark_enabled(event
, ctx
);
2079 if (group_can_go_on(event
, cpuctx
, 1))
2080 group_sched_in(event
, cpuctx
, ctx
);
2083 * If this pinned group hasn't been scheduled,
2084 * put it in error state.
2086 if (event
->state
== PERF_EVENT_STATE_INACTIVE
) {
2087 update_group_times(event
);
2088 event
->state
= PERF_EVENT_STATE_ERROR
;
2094 ctx_flexible_sched_in(struct perf_event_context
*ctx
,
2095 struct perf_cpu_context
*cpuctx
)
2097 struct perf_event
*event
;
2100 list_for_each_entry(event
, &ctx
->flexible_groups
, group_entry
) {
2101 /* Ignore events in OFF or ERROR state */
2102 if (event
->state
<= PERF_EVENT_STATE_OFF
)
2105 * Listen to the 'cpu' scheduling filter constraint
2108 if (!event_filter_match(event
))
2111 /* may need to reset tstamp_enabled */
2112 if (is_cgroup_event(event
))
2113 perf_cgroup_mark_enabled(event
, ctx
);
2115 if (group_can_go_on(event
, cpuctx
, can_add_hw
)) {
2116 if (group_sched_in(event
, cpuctx
, ctx
))
2123 ctx_sched_in(struct perf_event_context
*ctx
,
2124 struct perf_cpu_context
*cpuctx
,
2125 enum event_type_t event_type
,
2126 struct task_struct
*task
)
2129 int is_active
= ctx
->is_active
;
2131 ctx
->is_active
|= event_type
;
2132 if (likely(!ctx
->nr_events
))
2136 ctx
->timestamp
= now
;
2137 perf_cgroup_set_timestamp(task
, ctx
);
2139 * First go through the list and put on any pinned groups
2140 * in order to give them the best chance of going on.
2142 if (!(is_active
& EVENT_PINNED
) && (event_type
& EVENT_PINNED
))
2143 ctx_pinned_sched_in(ctx
, cpuctx
);
2145 /* Then walk through the lower prio flexible groups */
2146 if (!(is_active
& EVENT_FLEXIBLE
) && (event_type
& EVENT_FLEXIBLE
))
2147 ctx_flexible_sched_in(ctx
, cpuctx
);
2150 static void cpu_ctx_sched_in(struct perf_cpu_context
*cpuctx
,
2151 enum event_type_t event_type
,
2152 struct task_struct
*task
)
2154 struct perf_event_context
*ctx
= &cpuctx
->ctx
;
2156 ctx_sched_in(ctx
, cpuctx
, event_type
, task
);
2159 static void perf_event_context_sched_in(struct perf_event_context
*ctx
,
2160 struct task_struct
*task
)
2162 struct perf_cpu_context
*cpuctx
;
2164 cpuctx
= __get_cpu_context(ctx
);
2165 if (cpuctx
->task_ctx
== ctx
)
2168 perf_ctx_lock(cpuctx
, ctx
);
2169 perf_pmu_disable(ctx
->pmu
);
2171 * We want to keep the following priority order:
2172 * cpu pinned (that don't need to move), task pinned,
2173 * cpu flexible, task flexible.
2175 cpu_ctx_sched_out(cpuctx
, EVENT_FLEXIBLE
);
2177 perf_event_sched_in(cpuctx
, ctx
, task
);
2180 cpuctx
->task_ctx
= ctx
;
2182 perf_pmu_enable(ctx
->pmu
);
2183 perf_ctx_unlock(cpuctx
, ctx
);
2186 * Since these rotations are per-cpu, we need to ensure the
2187 * cpu-context we got scheduled on is actually rotating.
2189 perf_pmu_rotate_start(ctx
->pmu
);
2193 * Called from scheduler to add the events of the current task
2194 * with interrupts disabled.
2196 * We restore the event value and then enable it.
2198 * This does not protect us against NMI, but enable()
2199 * sets the enabled bit in the control field of event _before_
2200 * accessing the event control register. If a NMI hits, then it will
2201 * keep the event running.
2203 void __perf_event_task_sched_in(struct task_struct
*prev
,
2204 struct task_struct
*task
)
2206 struct perf_event_context
*ctx
;
2209 for_each_task_context_nr(ctxn
) {
2210 ctx
= task
->perf_event_ctxp
[ctxn
];
2214 perf_event_context_sched_in(ctx
, task
);
2217 * if cgroup events exist on this CPU, then we need
2218 * to check if we have to switch in PMU state.
2219 * cgroup event are system-wide mode only
2221 if (atomic_read(&__get_cpu_var(perf_cgroup_events
)))
2222 perf_cgroup_sched_in(prev
, task
);
2225 static u64
perf_calculate_period(struct perf_event
*event
, u64 nsec
, u64 count
)
2227 u64 frequency
= event
->attr
.sample_freq
;
2228 u64 sec
= NSEC_PER_SEC
;
2229 u64 divisor
, dividend
;
2231 int count_fls
, nsec_fls
, frequency_fls
, sec_fls
;
2233 count_fls
= fls64(count
);
2234 nsec_fls
= fls64(nsec
);
2235 frequency_fls
= fls64(frequency
);
2239 * We got @count in @nsec, with a target of sample_freq HZ
2240 * the target period becomes:
2243 * period = -------------------
2244 * @nsec * sample_freq
2249 * Reduce accuracy by one bit such that @a and @b converge
2250 * to a similar magnitude.
2252 #define REDUCE_FLS(a, b) \
2254 if (a##_fls > b##_fls) { \
2264 * Reduce accuracy until either term fits in a u64, then proceed with
2265 * the other, so that finally we can do a u64/u64 division.
2267 while (count_fls
+ sec_fls
> 64 && nsec_fls
+ frequency_fls
> 64) {
2268 REDUCE_FLS(nsec
, frequency
);
2269 REDUCE_FLS(sec
, count
);
2272 if (count_fls
+ sec_fls
> 64) {
2273 divisor
= nsec
* frequency
;
2275 while (count_fls
+ sec_fls
> 64) {
2276 REDUCE_FLS(count
, sec
);
2280 dividend
= count
* sec
;
2282 dividend
= count
* sec
;
2284 while (nsec_fls
+ frequency_fls
> 64) {
2285 REDUCE_FLS(nsec
, frequency
);
2289 divisor
= nsec
* frequency
;
2295 return div64_u64(dividend
, divisor
);
2298 static void perf_adjust_period(struct perf_event
*event
, u64 nsec
, u64 count
)
2300 struct hw_perf_event
*hwc
= &event
->hw
;
2301 s64 period
, sample_period
;
2304 period
= perf_calculate_period(event
, nsec
, count
);
2306 delta
= (s64
)(period
- hwc
->sample_period
);
2307 delta
= (delta
+ 7) / 8; /* low pass filter */
2309 sample_period
= hwc
->sample_period
+ delta
;
2314 hwc
->sample_period
= sample_period
;
2316 if (local64_read(&hwc
->period_left
) > 8*sample_period
) {
2317 event
->pmu
->stop(event
, PERF_EF_UPDATE
);
2318 local64_set(&hwc
->period_left
, 0);
2319 event
->pmu
->start(event
, PERF_EF_RELOAD
);
2323 static void perf_ctx_adjust_freq(struct perf_event_context
*ctx
, u64 period
)
2325 struct perf_event
*event
;
2326 struct hw_perf_event
*hwc
;
2327 u64 interrupts
, now
;
2330 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
2331 if (event
->state
!= PERF_EVENT_STATE_ACTIVE
)
2334 if (!event_filter_match(event
))
2339 interrupts
= hwc
->interrupts
;
2340 hwc
->interrupts
= 0;
2343 * unthrottle events on the tick
2345 if (interrupts
== MAX_INTERRUPTS
) {
2346 perf_log_throttle(event
, 1);
2347 event
->pmu
->start(event
, 0);
2350 if (!event
->attr
.freq
|| !event
->attr
.sample_freq
)
2353 event
->pmu
->read(event
);
2354 now
= local64_read(&event
->count
);
2355 delta
= now
- hwc
->freq_count_stamp
;
2356 hwc
->freq_count_stamp
= now
;
2359 perf_adjust_period(event
, period
, delta
);
2364 * Round-robin a context's events:
2366 static void rotate_ctx(struct perf_event_context
*ctx
)
2369 * Rotate the first entry last of non-pinned groups. Rotation might be
2370 * disabled by the inheritance code.
2372 if (!ctx
->rotate_disable
)
2373 list_rotate_left(&ctx
->flexible_groups
);
2377 * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
2378 * because they're strictly cpu affine and rotate_start is called with IRQs
2379 * disabled, while rotate_context is called from IRQ context.
2381 static void perf_rotate_context(struct perf_cpu_context
*cpuctx
)
2383 u64 interval
= (u64
)cpuctx
->jiffies_interval
* TICK_NSEC
;
2384 struct perf_event_context
*ctx
= NULL
;
2385 int rotate
= 0, remove
= 1;
2387 if (cpuctx
->ctx
.nr_events
) {
2389 if (cpuctx
->ctx
.nr_events
!= cpuctx
->ctx
.nr_active
)
2393 ctx
= cpuctx
->task_ctx
;
2394 if (ctx
&& ctx
->nr_events
) {
2396 if (ctx
->nr_events
!= ctx
->nr_active
)
2400 perf_ctx_lock(cpuctx
, cpuctx
->task_ctx
);
2401 perf_pmu_disable(cpuctx
->ctx
.pmu
);
2402 perf_ctx_adjust_freq(&cpuctx
->ctx
, interval
);
2404 perf_ctx_adjust_freq(ctx
, interval
);
2409 cpu_ctx_sched_out(cpuctx
, EVENT_FLEXIBLE
);
2411 ctx_sched_out(ctx
, cpuctx
, EVENT_FLEXIBLE
);
2413 rotate_ctx(&cpuctx
->ctx
);
2417 perf_event_sched_in(cpuctx
, ctx
, current
);
2421 list_del_init(&cpuctx
->rotation_list
);
2423 perf_pmu_enable(cpuctx
->ctx
.pmu
);
2424 perf_ctx_unlock(cpuctx
, cpuctx
->task_ctx
);
2427 void perf_event_task_tick(void)
2429 struct list_head
*head
= &__get_cpu_var(rotation_list
);
2430 struct perf_cpu_context
*cpuctx
, *tmp
;
2432 WARN_ON(!irqs_disabled());
2434 list_for_each_entry_safe(cpuctx
, tmp
, head
, rotation_list
) {
2435 if (cpuctx
->jiffies_interval
== 1 ||
2436 !(jiffies
% cpuctx
->jiffies_interval
))
2437 perf_rotate_context(cpuctx
);
2441 static int event_enable_on_exec(struct perf_event
*event
,
2442 struct perf_event_context
*ctx
)
2444 if (!event
->attr
.enable_on_exec
)
2447 event
->attr
.enable_on_exec
= 0;
2448 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
)
2451 __perf_event_mark_enabled(event
, ctx
);
2457 * Enable all of a task's events that have been marked enable-on-exec.
2458 * This expects task == current.
2460 static void perf_event_enable_on_exec(struct perf_event_context
*ctx
)
2462 struct perf_event
*event
;
2463 unsigned long flags
;
2467 local_irq_save(flags
);
2468 if (!ctx
|| !ctx
->nr_events
)
2472 * We must ctxsw out cgroup events to avoid conflict
2473 * when invoking perf_task_event_sched_in() later on
2474 * in this function. Otherwise we end up trying to
2475 * ctxswin cgroup events which are already scheduled
2478 perf_cgroup_sched_out(current
, NULL
);
2480 raw_spin_lock(&ctx
->lock
);
2481 task_ctx_sched_out(ctx
);
2483 list_for_each_entry(event
, &ctx
->pinned_groups
, group_entry
) {
2484 ret
= event_enable_on_exec(event
, ctx
);
2489 list_for_each_entry(event
, &ctx
->flexible_groups
, group_entry
) {
2490 ret
= event_enable_on_exec(event
, ctx
);
2496 * Unclone this context if we enabled any event.
2501 raw_spin_unlock(&ctx
->lock
);
2504 * Also calls ctxswin for cgroup events, if any:
2506 perf_event_context_sched_in(ctx
, ctx
->task
);
2508 local_irq_restore(flags
);
2512 * Cross CPU call to read the hardware event
2514 static void __perf_event_read(void *info
)
2516 struct perf_event
*event
= info
;
2517 struct perf_event_context
*ctx
= event
->ctx
;
2518 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
2521 * If this is a task context, we need to check whether it is
2522 * the current task context of this cpu. If not it has been
2523 * scheduled out before the smp call arrived. In that case
2524 * event->count would have been updated to a recent sample
2525 * when the event was scheduled out.
2527 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
)
2530 raw_spin_lock(&ctx
->lock
);
2531 if (ctx
->is_active
) {
2532 update_context_time(ctx
);
2533 update_cgrp_time_from_event(event
);
2535 update_event_times(event
);
2536 if (event
->state
== PERF_EVENT_STATE_ACTIVE
)
2537 event
->pmu
->read(event
);
2538 raw_spin_unlock(&ctx
->lock
);
2541 static inline u64
perf_event_count(struct perf_event
*event
)
2543 return local64_read(&event
->count
) + atomic64_read(&event
->child_count
);
2546 static u64
perf_event_read(struct perf_event
*event
)
2549 * If event is enabled and currently active on a CPU, update the
2550 * value in the event structure:
2552 if (event
->state
== PERF_EVENT_STATE_ACTIVE
) {
2553 smp_call_function_single(event
->oncpu
,
2554 __perf_event_read
, event
, 1);
2555 } else if (event
->state
== PERF_EVENT_STATE_INACTIVE
) {
2556 struct perf_event_context
*ctx
= event
->ctx
;
2557 unsigned long flags
;
2559 raw_spin_lock_irqsave(&ctx
->lock
, flags
);
2561 * may read while context is not active
2562 * (e.g., thread is blocked), in that case
2563 * we cannot update context time
2565 if (ctx
->is_active
) {
2566 update_context_time(ctx
);
2567 update_cgrp_time_from_event(event
);
2569 update_event_times(event
);
2570 raw_spin_unlock_irqrestore(&ctx
->lock
, flags
);
2573 return perf_event_count(event
);
2580 struct callchain_cpus_entries
{
2581 struct rcu_head rcu_head
;
2582 struct perf_callchain_entry
*cpu_entries
[0];
2585 static DEFINE_PER_CPU(int, callchain_recursion
[PERF_NR_CONTEXTS
]);
2586 static atomic_t nr_callchain_events
;
2587 static DEFINE_MUTEX(callchain_mutex
);
2588 struct callchain_cpus_entries
*callchain_cpus_entries
;
2591 __weak
void perf_callchain_kernel(struct perf_callchain_entry
*entry
,
2592 struct pt_regs
*regs
)
2596 __weak
void perf_callchain_user(struct perf_callchain_entry
*entry
,
2597 struct pt_regs
*regs
)
2601 static void release_callchain_buffers_rcu(struct rcu_head
*head
)
2603 struct callchain_cpus_entries
*entries
;
2606 entries
= container_of(head
, struct callchain_cpus_entries
, rcu_head
);
2608 for_each_possible_cpu(cpu
)
2609 kfree(entries
->cpu_entries
[cpu
]);
2614 static void release_callchain_buffers(void)
2616 struct callchain_cpus_entries
*entries
;
2618 entries
= callchain_cpus_entries
;
2619 rcu_assign_pointer(callchain_cpus_entries
, NULL
);
2620 call_rcu(&entries
->rcu_head
, release_callchain_buffers_rcu
);
2623 static int alloc_callchain_buffers(void)
2627 struct callchain_cpus_entries
*entries
;
2630 * We can't use the percpu allocation API for data that can be
2631 * accessed from NMI. Use a temporary manual per cpu allocation
2632 * until that gets sorted out.
2634 size
= offsetof(struct callchain_cpus_entries
, cpu_entries
[nr_cpu_ids
]);
2636 entries
= kzalloc(size
, GFP_KERNEL
);
2640 size
= sizeof(struct perf_callchain_entry
) * PERF_NR_CONTEXTS
;
2642 for_each_possible_cpu(cpu
) {
2643 entries
->cpu_entries
[cpu
] = kmalloc_node(size
, GFP_KERNEL
,
2645 if (!entries
->cpu_entries
[cpu
])
2649 rcu_assign_pointer(callchain_cpus_entries
, entries
);
2654 for_each_possible_cpu(cpu
)
2655 kfree(entries
->cpu_entries
[cpu
]);
2661 static int get_callchain_buffers(void)
2666 mutex_lock(&callchain_mutex
);
2668 count
= atomic_inc_return(&nr_callchain_events
);
2669 if (WARN_ON_ONCE(count
< 1)) {
2675 /* If the allocation failed, give up */
2676 if (!callchain_cpus_entries
)
2681 err
= alloc_callchain_buffers();
2683 release_callchain_buffers();
2685 mutex_unlock(&callchain_mutex
);
2690 static void put_callchain_buffers(void)
2692 if (atomic_dec_and_mutex_lock(&nr_callchain_events
, &callchain_mutex
)) {
2693 release_callchain_buffers();
2694 mutex_unlock(&callchain_mutex
);
2698 static int get_recursion_context(int *recursion
)
2706 else if (in_softirq())
2711 if (recursion
[rctx
])
2720 static inline void put_recursion_context(int *recursion
, int rctx
)
2726 static struct perf_callchain_entry
*get_callchain_entry(int *rctx
)
2729 struct callchain_cpus_entries
*entries
;
2731 *rctx
= get_recursion_context(__get_cpu_var(callchain_recursion
));
2735 entries
= rcu_dereference(callchain_cpus_entries
);
2739 cpu
= smp_processor_id();
2741 return &entries
->cpu_entries
[cpu
][*rctx
];
2745 put_callchain_entry(int rctx
)
2747 put_recursion_context(__get_cpu_var(callchain_recursion
), rctx
);
2750 static struct perf_callchain_entry
*perf_callchain(struct pt_regs
*regs
)
2753 struct perf_callchain_entry
*entry
;
2756 entry
= get_callchain_entry(&rctx
);
2765 if (!user_mode(regs
)) {
2766 perf_callchain_store(entry
, PERF_CONTEXT_KERNEL
);
2767 perf_callchain_kernel(entry
, regs
);
2769 regs
= task_pt_regs(current
);
2775 perf_callchain_store(entry
, PERF_CONTEXT_USER
);
2776 perf_callchain_user(entry
, regs
);
2780 put_callchain_entry(rctx
);
2786 * Initialize the perf_event context in a task_struct:
2788 static void __perf_event_init_context(struct perf_event_context
*ctx
)
2790 raw_spin_lock_init(&ctx
->lock
);
2791 mutex_init(&ctx
->mutex
);
2792 INIT_LIST_HEAD(&ctx
->pinned_groups
);
2793 INIT_LIST_HEAD(&ctx
->flexible_groups
);
2794 INIT_LIST_HEAD(&ctx
->event_list
);
2795 atomic_set(&ctx
->refcount
, 1);
2798 static struct perf_event_context
*
2799 alloc_perf_context(struct pmu
*pmu
, struct task_struct
*task
)
2801 struct perf_event_context
*ctx
;
2803 ctx
= kzalloc(sizeof(struct perf_event_context
), GFP_KERNEL
);
2807 __perf_event_init_context(ctx
);
2810 get_task_struct(task
);
2817 static struct task_struct
*
2818 find_lively_task_by_vpid(pid_t vpid
)
2820 struct task_struct
*task
;
2827 task
= find_task_by_vpid(vpid
);
2829 get_task_struct(task
);
2833 return ERR_PTR(-ESRCH
);
2835 /* Reuse ptrace permission checks for now. */
2837 if (!ptrace_may_access(task
, PTRACE_MODE_READ
))
2842 put_task_struct(task
);
2843 return ERR_PTR(err
);
2848 * Returns a matching context with refcount and pincount.
2850 static struct perf_event_context
*
2851 find_get_context(struct pmu
*pmu
, struct task_struct
*task
, int cpu
)
2853 struct perf_event_context
*ctx
;
2854 struct perf_cpu_context
*cpuctx
;
2855 unsigned long flags
;
2859 /* Must be root to operate on a CPU event: */
2860 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN
))
2861 return ERR_PTR(-EACCES
);
2864 * We could be clever and allow to attach a event to an
2865 * offline CPU and activate it when the CPU comes up, but
2868 if (!cpu_online(cpu
))
2869 return ERR_PTR(-ENODEV
);
2871 cpuctx
= per_cpu_ptr(pmu
->pmu_cpu_context
, cpu
);
2880 ctxn
= pmu
->task_ctx_nr
;
2885 ctx
= perf_lock_task_context(task
, ctxn
, &flags
);
2889 raw_spin_unlock_irqrestore(&ctx
->lock
, flags
);
2891 ctx
= alloc_perf_context(pmu
, task
);
2897 mutex_lock(&task
->perf_event_mutex
);
2899 * If it has already passed perf_event_exit_task().
2900 * we must see PF_EXITING, it takes this mutex too.
2902 if (task
->flags
& PF_EXITING
)
2904 else if (task
->perf_event_ctxp
[ctxn
])
2909 rcu_assign_pointer(task
->perf_event_ctxp
[ctxn
], ctx
);
2911 mutex_unlock(&task
->perf_event_mutex
);
2913 if (unlikely(err
)) {
2925 return ERR_PTR(err
);
2928 static void perf_event_free_filter(struct perf_event
*event
);
2930 static void free_event_rcu(struct rcu_head
*head
)
2932 struct perf_event
*event
;
2934 event
= container_of(head
, struct perf_event
, rcu_head
);
2936 put_pid_ns(event
->ns
);
2937 perf_event_free_filter(event
);
2941 static void ring_buffer_put(struct ring_buffer
*rb
);
2943 static void free_event(struct perf_event
*event
)
2945 irq_work_sync(&event
->pending
);
2947 if (!event
->parent
) {
2948 if (event
->attach_state
& PERF_ATTACH_TASK
)
2949 jump_label_dec(&perf_sched_events
);
2950 if (event
->attr
.mmap
|| event
->attr
.mmap_data
)
2951 atomic_dec(&nr_mmap_events
);
2952 if (event
->attr
.comm
)
2953 atomic_dec(&nr_comm_events
);
2954 if (event
->attr
.task
)
2955 atomic_dec(&nr_task_events
);
2956 if (event
->attr
.sample_type
& PERF_SAMPLE_CALLCHAIN
)
2957 put_callchain_buffers();
2958 if (is_cgroup_event(event
)) {
2959 atomic_dec(&per_cpu(perf_cgroup_events
, event
->cpu
));
2960 jump_label_dec(&perf_sched_events
);
2965 ring_buffer_put(event
->rb
);
2969 if (is_cgroup_event(event
))
2970 perf_detach_cgroup(event
);
2973 event
->destroy(event
);
2976 put_ctx(event
->ctx
);
2978 call_rcu(&event
->rcu_head
, free_event_rcu
);
2981 int perf_event_release_kernel(struct perf_event
*event
)
2983 struct perf_event_context
*ctx
= event
->ctx
;
2985 WARN_ON_ONCE(ctx
->parent_ctx
);
2987 * There are two ways this annotation is useful:
2989 * 1) there is a lock recursion from perf_event_exit_task
2990 * see the comment there.
2992 * 2) there is a lock-inversion with mmap_sem through
2993 * perf_event_read_group(), which takes faults while
2994 * holding ctx->mutex, however this is called after
2995 * the last filedesc died, so there is no possibility
2996 * to trigger the AB-BA case.
2998 mutex_lock_nested(&ctx
->mutex
, SINGLE_DEPTH_NESTING
);
2999 raw_spin_lock_irq(&ctx
->lock
);
3000 perf_group_detach(event
);
3001 raw_spin_unlock_irq(&ctx
->lock
);
3002 perf_remove_from_context(event
);
3003 mutex_unlock(&ctx
->mutex
);
3009 EXPORT_SYMBOL_GPL(perf_event_release_kernel
);
3012 * Called when the last reference to the file is gone.
3014 static int perf_release(struct inode
*inode
, struct file
*file
)
3016 struct perf_event
*event
= file
->private_data
;
3017 struct task_struct
*owner
;
3019 file
->private_data
= NULL
;
3022 owner
= ACCESS_ONCE(event
->owner
);
3024 * Matches the smp_wmb() in perf_event_exit_task(). If we observe
3025 * !owner it means the list deletion is complete and we can indeed
3026 * free this event, otherwise we need to serialize on
3027 * owner->perf_event_mutex.
3029 smp_read_barrier_depends();
3032 * Since delayed_put_task_struct() also drops the last
3033 * task reference we can safely take a new reference
3034 * while holding the rcu_read_lock().
3036 get_task_struct(owner
);
3041 mutex_lock(&owner
->perf_event_mutex
);
3043 * We have to re-check the event->owner field, if it is cleared
3044 * we raced with perf_event_exit_task(), acquiring the mutex
3045 * ensured they're done, and we can proceed with freeing the
3049 list_del_init(&event
->owner_entry
);
3050 mutex_unlock(&owner
->perf_event_mutex
);
3051 put_task_struct(owner
);
3054 return perf_event_release_kernel(event
);
3057 u64
perf_event_read_value(struct perf_event
*event
, u64
*enabled
, u64
*running
)
3059 struct perf_event
*child
;
3065 mutex_lock(&event
->child_mutex
);
3066 total
+= perf_event_read(event
);
3067 *enabled
+= event
->total_time_enabled
+
3068 atomic64_read(&event
->child_total_time_enabled
);
3069 *running
+= event
->total_time_running
+
3070 atomic64_read(&event
->child_total_time_running
);
3072 list_for_each_entry(child
, &event
->child_list
, child_list
) {
3073 total
+= perf_event_read(child
);
3074 *enabled
+= child
->total_time_enabled
;
3075 *running
+= child
->total_time_running
;
3077 mutex_unlock(&event
->child_mutex
);
3081 EXPORT_SYMBOL_GPL(perf_event_read_value
);
3083 static int perf_event_read_group(struct perf_event
*event
,
3084 u64 read_format
, char __user
*buf
)
3086 struct perf_event
*leader
= event
->group_leader
, *sub
;
3087 int n
= 0, size
= 0, ret
= -EFAULT
;
3088 struct perf_event_context
*ctx
= leader
->ctx
;
3090 u64 count
, enabled
, running
;
3092 mutex_lock(&ctx
->mutex
);
3093 count
= perf_event_read_value(leader
, &enabled
, &running
);
3095 values
[n
++] = 1 + leader
->nr_siblings
;
3096 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
3097 values
[n
++] = enabled
;
3098 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
3099 values
[n
++] = running
;
3100 values
[n
++] = count
;
3101 if (read_format
& PERF_FORMAT_ID
)
3102 values
[n
++] = primary_event_id(leader
);
3104 size
= n
* sizeof(u64
);
3106 if (copy_to_user(buf
, values
, size
))
3111 list_for_each_entry(sub
, &leader
->sibling_list
, group_entry
) {
3114 values
[n
++] = perf_event_read_value(sub
, &enabled
, &running
);
3115 if (read_format
& PERF_FORMAT_ID
)
3116 values
[n
++] = primary_event_id(sub
);
3118 size
= n
* sizeof(u64
);
3120 if (copy_to_user(buf
+ ret
, values
, size
)) {
3128 mutex_unlock(&ctx
->mutex
);
3133 static int perf_event_read_one(struct perf_event
*event
,
3134 u64 read_format
, char __user
*buf
)
3136 u64 enabled
, running
;
3140 values
[n
++] = perf_event_read_value(event
, &enabled
, &running
);
3141 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
3142 values
[n
++] = enabled
;
3143 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
3144 values
[n
++] = running
;
3145 if (read_format
& PERF_FORMAT_ID
)
3146 values
[n
++] = primary_event_id(event
);
3148 if (copy_to_user(buf
, values
, n
* sizeof(u64
)))
3151 return n
* sizeof(u64
);
3155 * Read the performance event - simple non blocking version for now
3158 perf_read_hw(struct perf_event
*event
, char __user
*buf
, size_t count
)
3160 u64 read_format
= event
->attr
.read_format
;
3164 * Return end-of-file for a read on a event that is in
3165 * error state (i.e. because it was pinned but it couldn't be
3166 * scheduled on to the CPU at some point).
3168 if (event
->state
== PERF_EVENT_STATE_ERROR
)
3171 if (count
< event
->read_size
)
3174 WARN_ON_ONCE(event
->ctx
->parent_ctx
);
3175 if (read_format
& PERF_FORMAT_GROUP
)
3176 ret
= perf_event_read_group(event
, read_format
, buf
);
3178 ret
= perf_event_read_one(event
, read_format
, buf
);
3184 perf_read(struct file
*file
, char __user
*buf
, size_t count
, loff_t
*ppos
)
3186 struct perf_event
*event
= file
->private_data
;
3188 return perf_read_hw(event
, buf
, count
);
3191 static unsigned int perf_poll(struct file
*file
, poll_table
*wait
)
3193 struct perf_event
*event
= file
->private_data
;
3194 struct ring_buffer
*rb
;
3195 unsigned int events
= POLL_HUP
;
3198 * Race between perf_event_set_output() and perf_poll(): perf_poll()
3199 * grabs the rb reference but perf_event_set_output() overrides it.
3200 * Here is the timeline for two threads T1, T2:
3201 * t0: T1, rb = rcu_dereference(event->rb)
3202 * t1: T2, old_rb = event->rb
3203 * t2: T2, event->rb = new rb
3204 * t3: T2, ring_buffer_detach(old_rb)
3205 * t4: T1, ring_buffer_attach(rb1)
3206 * t5: T1, poll_wait(event->waitq)
3208 * To avoid this problem, we grab mmap_mutex in perf_poll()
3209 * thereby ensuring that the assignment of the new ring buffer
3210 * and the detachment of the old buffer appear atomic to perf_poll()
3212 mutex_lock(&event
->mmap_mutex
);
3215 rb
= rcu_dereference(event
->rb
);
3217 ring_buffer_attach(event
, rb
);
3218 events
= atomic_xchg(&rb
->poll
, 0);
3222 mutex_unlock(&event
->mmap_mutex
);
3224 poll_wait(file
, &event
->waitq
, wait
);
3229 static void perf_event_reset(struct perf_event
*event
)
3231 (void)perf_event_read(event
);
3232 local64_set(&event
->count
, 0);
3233 perf_event_update_userpage(event
);
3237 * Holding the top-level event's child_mutex means that any
3238 * descendant process that has inherited this event will block
3239 * in sync_child_event if it goes to exit, thus satisfying the
3240 * task existence requirements of perf_event_enable/disable.
3242 static void perf_event_for_each_child(struct perf_event
*event
,
3243 void (*func
)(struct perf_event
*))
3245 struct perf_event
*child
;
3247 WARN_ON_ONCE(event
->ctx
->parent_ctx
);
3248 mutex_lock(&event
->child_mutex
);
3250 list_for_each_entry(child
, &event
->child_list
, child_list
)
3252 mutex_unlock(&event
->child_mutex
);
3255 static void perf_event_for_each(struct perf_event
*event
,
3256 void (*func
)(struct perf_event
*))
3258 struct perf_event_context
*ctx
= event
->ctx
;
3259 struct perf_event
*sibling
;
3261 WARN_ON_ONCE(ctx
->parent_ctx
);
3262 mutex_lock(&ctx
->mutex
);
3263 event
= event
->group_leader
;
3265 perf_event_for_each_child(event
, func
);
3267 list_for_each_entry(sibling
, &event
->sibling_list
, group_entry
)
3268 perf_event_for_each_child(event
, func
);
3269 mutex_unlock(&ctx
->mutex
);
3272 static int perf_event_period(struct perf_event
*event
, u64 __user
*arg
)
3274 struct perf_event_context
*ctx
= event
->ctx
;
3278 if (!is_sampling_event(event
))
3281 if (copy_from_user(&value
, arg
, sizeof(value
)))
3287 raw_spin_lock_irq(&ctx
->lock
);
3288 if (event
->attr
.freq
) {
3289 if (value
> sysctl_perf_event_sample_rate
) {
3294 event
->attr
.sample_freq
= value
;
3296 event
->attr
.sample_period
= value
;
3297 event
->hw
.sample_period
= value
;
3300 raw_spin_unlock_irq(&ctx
->lock
);
3305 static const struct file_operations perf_fops
;
3307 static struct perf_event
*perf_fget_light(int fd
, int *fput_needed
)
3311 file
= fget_light(fd
, fput_needed
);
3313 return ERR_PTR(-EBADF
);
3315 if (file
->f_op
!= &perf_fops
) {
3316 fput_light(file
, *fput_needed
);
3318 return ERR_PTR(-EBADF
);
3321 return file
->private_data
;
3324 static int perf_event_set_output(struct perf_event
*event
,
3325 struct perf_event
*output_event
);
3326 static int perf_event_set_filter(struct perf_event
*event
, void __user
*arg
);
3328 static long perf_ioctl(struct file
*file
, unsigned int cmd
, unsigned long arg
)
3330 struct perf_event
*event
= file
->private_data
;
3331 void (*func
)(struct perf_event
*);
3335 case PERF_EVENT_IOC_ENABLE
:
3336 func
= perf_event_enable
;
3338 case PERF_EVENT_IOC_DISABLE
:
3339 func
= perf_event_disable
;
3341 case PERF_EVENT_IOC_RESET
:
3342 func
= perf_event_reset
;
3345 case PERF_EVENT_IOC_REFRESH
:
3346 return perf_event_refresh(event
, arg
);
3348 case PERF_EVENT_IOC_PERIOD
:
3349 return perf_event_period(event
, (u64 __user
*)arg
);
3351 case PERF_EVENT_IOC_SET_OUTPUT
:
3353 struct perf_event
*output_event
= NULL
;
3354 int fput_needed
= 0;
3358 output_event
= perf_fget_light(arg
, &fput_needed
);
3359 if (IS_ERR(output_event
))
3360 return PTR_ERR(output_event
);
3363 ret
= perf_event_set_output(event
, output_event
);
3365 fput_light(output_event
->filp
, fput_needed
);
3370 case PERF_EVENT_IOC_SET_FILTER
:
3371 return perf_event_set_filter(event
, (void __user
*)arg
);
3377 if (flags
& PERF_IOC_FLAG_GROUP
)
3378 perf_event_for_each(event
, func
);
3380 perf_event_for_each_child(event
, func
);
3385 int perf_event_task_enable(void)
3387 struct perf_event
*event
;
3389 mutex_lock(¤t
->perf_event_mutex
);
3390 list_for_each_entry(event
, ¤t
->perf_event_list
, owner_entry
)
3391 perf_event_for_each_child(event
, perf_event_enable
);
3392 mutex_unlock(¤t
->perf_event_mutex
);
3397 int perf_event_task_disable(void)
3399 struct perf_event
*event
;
3401 mutex_lock(¤t
->perf_event_mutex
);
3402 list_for_each_entry(event
, ¤t
->perf_event_list
, owner_entry
)
3403 perf_event_for_each_child(event
, perf_event_disable
);
3404 mutex_unlock(¤t
->perf_event_mutex
);
3409 #ifndef PERF_EVENT_INDEX_OFFSET
3410 # define PERF_EVENT_INDEX_OFFSET 0
3413 static int perf_event_index(struct perf_event
*event
)
3415 if (event
->hw
.state
& PERF_HES_STOPPED
)
3418 if (event
->state
!= PERF_EVENT_STATE_ACTIVE
)
3421 return event
->hw
.idx
+ 1 - PERF_EVENT_INDEX_OFFSET
;
3424 static void calc_timer_values(struct perf_event
*event
,
3431 ctx_time
= event
->shadow_ctx_time
+ now
;
3432 *enabled
= ctx_time
- event
->tstamp_enabled
;
3433 *running
= ctx_time
- event
->tstamp_running
;
3437 * Callers need to ensure there can be no nesting of this function, otherwise
3438 * the seqlock logic goes bad. We can not serialize this because the arch
3439 * code calls this from NMI context.
3441 void perf_event_update_userpage(struct perf_event
*event
)
3443 struct perf_event_mmap_page
*userpg
;
3444 struct ring_buffer
*rb
;
3445 u64 enabled
, running
;
3449 * compute total_time_enabled, total_time_running
3450 * based on snapshot values taken when the event
3451 * was last scheduled in.
3453 * we cannot simply called update_context_time()
3454 * because of locking issue as we can be called in
3457 calc_timer_values(event
, &enabled
, &running
);
3458 rb
= rcu_dereference(event
->rb
);
3462 userpg
= rb
->user_page
;
3465 * Disable preemption so as to not let the corresponding user-space
3466 * spin too long if we get preempted.
3471 userpg
->index
= perf_event_index(event
);
3472 userpg
->offset
= perf_event_count(event
);
3473 if (event
->state
== PERF_EVENT_STATE_ACTIVE
)
3474 userpg
->offset
-= local64_read(&event
->hw
.prev_count
);
3476 userpg
->time_enabled
= enabled
+
3477 atomic64_read(&event
->child_total_time_enabled
);
3479 userpg
->time_running
= running
+
3480 atomic64_read(&event
->child_total_time_running
);
3489 static int perf_mmap_fault(struct vm_area_struct
*vma
, struct vm_fault
*vmf
)
3491 struct perf_event
*event
= vma
->vm_file
->private_data
;
3492 struct ring_buffer
*rb
;
3493 int ret
= VM_FAULT_SIGBUS
;
3495 if (vmf
->flags
& FAULT_FLAG_MKWRITE
) {
3496 if (vmf
->pgoff
== 0)
3502 rb
= rcu_dereference(event
->rb
);
3506 if (vmf
->pgoff
&& (vmf
->flags
& FAULT_FLAG_WRITE
))
3509 vmf
->page
= perf_mmap_to_page(rb
, vmf
->pgoff
);
3513 get_page(vmf
->page
);
3514 vmf
->page
->mapping
= vma
->vm_file
->f_mapping
;
3515 vmf
->page
->index
= vmf
->pgoff
;
3524 static void ring_buffer_attach(struct perf_event
*event
,
3525 struct ring_buffer
*rb
)
3527 unsigned long flags
;
3529 if (!list_empty(&event
->rb_entry
))
3532 spin_lock_irqsave(&rb
->event_lock
, flags
);
3533 if (!list_empty(&event
->rb_entry
))
3536 list_add(&event
->rb_entry
, &rb
->event_list
);
3538 spin_unlock_irqrestore(&rb
->event_lock
, flags
);
3541 static void ring_buffer_detach(struct perf_event
*event
,
3542 struct ring_buffer
*rb
)
3544 unsigned long flags
;
3546 if (list_empty(&event
->rb_entry
))
3549 spin_lock_irqsave(&rb
->event_lock
, flags
);
3550 list_del_init(&event
->rb_entry
);
3551 wake_up_all(&event
->waitq
);
3552 spin_unlock_irqrestore(&rb
->event_lock
, flags
);
3555 static void ring_buffer_wakeup(struct perf_event
*event
)
3557 struct ring_buffer
*rb
;
3560 rb
= rcu_dereference(event
->rb
);
3561 list_for_each_entry_rcu(event
, &rb
->event_list
, rb_entry
) {
3562 wake_up_all(&event
->waitq
);
3567 static void rb_free_rcu(struct rcu_head
*rcu_head
)
3569 struct ring_buffer
*rb
;
3571 rb
= container_of(rcu_head
, struct ring_buffer
, rcu_head
);
3575 static struct ring_buffer
*ring_buffer_get(struct perf_event
*event
)
3577 struct ring_buffer
*rb
;
3580 rb
= rcu_dereference(event
->rb
);
3582 if (!atomic_inc_not_zero(&rb
->refcount
))
3590 static void ring_buffer_put(struct ring_buffer
*rb
)
3592 struct perf_event
*event
, *n
;
3593 unsigned long flags
;
3595 if (!atomic_dec_and_test(&rb
->refcount
))
3598 spin_lock_irqsave(&rb
->event_lock
, flags
);
3599 list_for_each_entry_safe(event
, n
, &rb
->event_list
, rb_entry
) {
3600 list_del_init(&event
->rb_entry
);
3601 wake_up_all(&event
->waitq
);
3603 spin_unlock_irqrestore(&rb
->event_lock
, flags
);
3605 call_rcu(&rb
->rcu_head
, rb_free_rcu
);
3608 static void perf_mmap_open(struct vm_area_struct
*vma
)
3610 struct perf_event
*event
= vma
->vm_file
->private_data
;
3612 atomic_inc(&event
->mmap_count
);
3615 static void perf_mmap_close(struct vm_area_struct
*vma
)
3617 struct perf_event
*event
= vma
->vm_file
->private_data
;
3619 if (atomic_dec_and_mutex_lock(&event
->mmap_count
, &event
->mmap_mutex
)) {
3620 unsigned long size
= perf_data_size(event
->rb
);
3621 struct user_struct
*user
= event
->mmap_user
;
3622 struct ring_buffer
*rb
= event
->rb
;
3624 atomic_long_sub((size
>> PAGE_SHIFT
) + 1, &user
->locked_vm
);
3625 vma
->vm_mm
->pinned_vm
-= event
->mmap_locked
;
3626 rcu_assign_pointer(event
->rb
, NULL
);
3627 ring_buffer_detach(event
, rb
);
3628 mutex_unlock(&event
->mmap_mutex
);
3630 ring_buffer_put(rb
);
3635 static const struct vm_operations_struct perf_mmap_vmops
= {
3636 .open
= perf_mmap_open
,
3637 .close
= perf_mmap_close
,
3638 .fault
= perf_mmap_fault
,
3639 .page_mkwrite
= perf_mmap_fault
,
3642 static int perf_mmap(struct file
*file
, struct vm_area_struct
*vma
)
3644 struct perf_event
*event
= file
->private_data
;
3645 unsigned long user_locked
, user_lock_limit
;
3646 struct user_struct
*user
= current_user();
3647 unsigned long locked
, lock_limit
;
3648 struct ring_buffer
*rb
;
3649 unsigned long vma_size
;
3650 unsigned long nr_pages
;
3651 long user_extra
, extra
;
3652 int ret
= 0, flags
= 0;
3655 * Don't allow mmap() of inherited per-task counters. This would
3656 * create a performance issue due to all children writing to the
3659 if (event
->cpu
== -1 && event
->attr
.inherit
)
3662 if (!(vma
->vm_flags
& VM_SHARED
))
3665 vma_size
= vma
->vm_end
- vma
->vm_start
;
3666 nr_pages
= (vma_size
/ PAGE_SIZE
) - 1;
3669 * If we have rb pages ensure they're a power-of-two number, so we
3670 * can do bitmasks instead of modulo.
3672 if (nr_pages
!= 0 && !is_power_of_2(nr_pages
))
3675 if (vma_size
!= PAGE_SIZE
* (1 + nr_pages
))
3678 if (vma
->vm_pgoff
!= 0)
3681 WARN_ON_ONCE(event
->ctx
->parent_ctx
);
3682 mutex_lock(&event
->mmap_mutex
);
3684 if (event
->rb
->nr_pages
== nr_pages
)
3685 atomic_inc(&event
->rb
->refcount
);
3691 user_extra
= nr_pages
+ 1;
3692 user_lock_limit
= sysctl_perf_event_mlock
>> (PAGE_SHIFT
- 10);
3695 * Increase the limit linearly with more CPUs:
3697 user_lock_limit
*= num_online_cpus();
3699 user_locked
= atomic_long_read(&user
->locked_vm
) + user_extra
;
3702 if (user_locked
> user_lock_limit
)
3703 extra
= user_locked
- user_lock_limit
;
3705 lock_limit
= rlimit(RLIMIT_MEMLOCK
);
3706 lock_limit
>>= PAGE_SHIFT
;
3707 locked
= vma
->vm_mm
->pinned_vm
+ extra
;
3709 if ((locked
> lock_limit
) && perf_paranoid_tracepoint_raw() &&
3710 !capable(CAP_IPC_LOCK
)) {
3717 if (vma
->vm_flags
& VM_WRITE
)
3718 flags
|= RING_BUFFER_WRITABLE
;
3720 rb
= rb_alloc(nr_pages
,
3721 event
->attr
.watermark
? event
->attr
.wakeup_watermark
: 0,
3728 rcu_assign_pointer(event
->rb
, rb
);
3730 atomic_long_add(user_extra
, &user
->locked_vm
);
3731 event
->mmap_locked
= extra
;
3732 event
->mmap_user
= get_current_user();
3733 vma
->vm_mm
->pinned_vm
+= event
->mmap_locked
;
3737 atomic_inc(&event
->mmap_count
);
3738 mutex_unlock(&event
->mmap_mutex
);
3740 vma
->vm_flags
|= VM_RESERVED
;
3741 vma
->vm_ops
= &perf_mmap_vmops
;
3746 static int perf_fasync(int fd
, struct file
*filp
, int on
)
3748 struct inode
*inode
= filp
->f_path
.dentry
->d_inode
;
3749 struct perf_event
*event
= filp
->private_data
;
3752 mutex_lock(&inode
->i_mutex
);
3753 retval
= fasync_helper(fd
, filp
, on
, &event
->fasync
);
3754 mutex_unlock(&inode
->i_mutex
);
3762 static const struct file_operations perf_fops
= {
3763 .llseek
= no_llseek
,
3764 .release
= perf_release
,
3767 .unlocked_ioctl
= perf_ioctl
,
3768 .compat_ioctl
= perf_ioctl
,
3770 .fasync
= perf_fasync
,
3776 * If there's data, ensure we set the poll() state and publish everything
3777 * to user-space before waking everybody up.
3780 void perf_event_wakeup(struct perf_event
*event
)
3782 ring_buffer_wakeup(event
);
3784 if (event
->pending_kill
) {
3785 kill_fasync(&event
->fasync
, SIGIO
, event
->pending_kill
);
3786 event
->pending_kill
= 0;
3790 static void perf_pending_event(struct irq_work
*entry
)
3792 struct perf_event
*event
= container_of(entry
,
3793 struct perf_event
, pending
);
3795 if (event
->pending_disable
) {
3796 event
->pending_disable
= 0;
3797 __perf_event_disable(event
);
3800 if (event
->pending_wakeup
) {
3801 event
->pending_wakeup
= 0;
3802 perf_event_wakeup(event
);
3807 * We assume there is only KVM supporting the callbacks.
3808 * Later on, we might change it to a list if there is
3809 * another virtualization implementation supporting the callbacks.
3811 struct perf_guest_info_callbacks
*perf_guest_cbs
;
3813 int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks
*cbs
)
3815 perf_guest_cbs
= cbs
;
3818 EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks
);
3820 int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks
*cbs
)
3822 perf_guest_cbs
= NULL
;
3825 EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks
);
3827 static void __perf_event_header__init_id(struct perf_event_header
*header
,
3828 struct perf_sample_data
*data
,
3829 struct perf_event
*event
)
3831 u64 sample_type
= event
->attr
.sample_type
;
3833 data
->type
= sample_type
;
3834 header
->size
+= event
->id_header_size
;
3836 if (sample_type
& PERF_SAMPLE_TID
) {
3837 /* namespace issues */
3838 data
->tid_entry
.pid
= perf_event_pid(event
, current
);
3839 data
->tid_entry
.tid
= perf_event_tid(event
, current
);
3842 if (sample_type
& PERF_SAMPLE_TIME
)
3843 data
->time
= perf_clock();
3845 if (sample_type
& PERF_SAMPLE_ID
)
3846 data
->id
= primary_event_id(event
);
3848 if (sample_type
& PERF_SAMPLE_STREAM_ID
)
3849 data
->stream_id
= event
->id
;
3851 if (sample_type
& PERF_SAMPLE_CPU
) {
3852 data
->cpu_entry
.cpu
= raw_smp_processor_id();
3853 data
->cpu_entry
.reserved
= 0;
3857 void perf_event_header__init_id(struct perf_event_header
*header
,
3858 struct perf_sample_data
*data
,
3859 struct perf_event
*event
)
3861 if (event
->attr
.sample_id_all
)
3862 __perf_event_header__init_id(header
, data
, event
);
3865 static void __perf_event__output_id_sample(struct perf_output_handle
*handle
,
3866 struct perf_sample_data
*data
)
3868 u64 sample_type
= data
->type
;
3870 if (sample_type
& PERF_SAMPLE_TID
)
3871 perf_output_put(handle
, data
->tid_entry
);
3873 if (sample_type
& PERF_SAMPLE_TIME
)
3874 perf_output_put(handle
, data
->time
);
3876 if (sample_type
& PERF_SAMPLE_ID
)
3877 perf_output_put(handle
, data
->id
);
3879 if (sample_type
& PERF_SAMPLE_STREAM_ID
)
3880 perf_output_put(handle
, data
->stream_id
);
3882 if (sample_type
& PERF_SAMPLE_CPU
)
3883 perf_output_put(handle
, data
->cpu_entry
);
3886 void perf_event__output_id_sample(struct perf_event
*event
,
3887 struct perf_output_handle
*handle
,
3888 struct perf_sample_data
*sample
)
3890 if (event
->attr
.sample_id_all
)
3891 __perf_event__output_id_sample(handle
, sample
);
3894 static void perf_output_read_one(struct perf_output_handle
*handle
,
3895 struct perf_event
*event
,
3896 u64 enabled
, u64 running
)
3898 u64 read_format
= event
->attr
.read_format
;
3902 values
[n
++] = perf_event_count(event
);
3903 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
) {
3904 values
[n
++] = enabled
+
3905 atomic64_read(&event
->child_total_time_enabled
);
3907 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
) {
3908 values
[n
++] = running
+
3909 atomic64_read(&event
->child_total_time_running
);
3911 if (read_format
& PERF_FORMAT_ID
)
3912 values
[n
++] = primary_event_id(event
);
3914 __output_copy(handle
, values
, n
* sizeof(u64
));
3918 * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
3920 static void perf_output_read_group(struct perf_output_handle
*handle
,
3921 struct perf_event
*event
,
3922 u64 enabled
, u64 running
)
3924 struct perf_event
*leader
= event
->group_leader
, *sub
;
3925 u64 read_format
= event
->attr
.read_format
;
3929 values
[n
++] = 1 + leader
->nr_siblings
;
3931 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
3932 values
[n
++] = enabled
;
3934 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
3935 values
[n
++] = running
;
3937 if (leader
!= event
)
3938 leader
->pmu
->read(leader
);
3940 values
[n
++] = perf_event_count(leader
);
3941 if (read_format
& PERF_FORMAT_ID
)
3942 values
[n
++] = primary_event_id(leader
);
3944 __output_copy(handle
, values
, n
* sizeof(u64
));
3946 list_for_each_entry(sub
, &leader
->sibling_list
, group_entry
) {
3950 sub
->pmu
->read(sub
);
3952 values
[n
++] = perf_event_count(sub
);
3953 if (read_format
& PERF_FORMAT_ID
)
3954 values
[n
++] = primary_event_id(sub
);
3956 __output_copy(handle
, values
, n
* sizeof(u64
));
3960 #define PERF_FORMAT_TOTAL_TIMES (PERF_FORMAT_TOTAL_TIME_ENABLED|\
3961 PERF_FORMAT_TOTAL_TIME_RUNNING)
3963 static void perf_output_read(struct perf_output_handle
*handle
,
3964 struct perf_event
*event
)
3966 u64 enabled
= 0, running
= 0;
3967 u64 read_format
= event
->attr
.read_format
;
3970 * compute total_time_enabled, total_time_running
3971 * based on snapshot values taken when the event
3972 * was last scheduled in.
3974 * we cannot simply called update_context_time()
3975 * because of locking issue as we are called in
3978 if (read_format
& PERF_FORMAT_TOTAL_TIMES
)
3979 calc_timer_values(event
, &enabled
, &running
);
3981 if (event
->attr
.read_format
& PERF_FORMAT_GROUP
)
3982 perf_output_read_group(handle
, event
, enabled
, running
);
3984 perf_output_read_one(handle
, event
, enabled
, running
);
3987 void perf_output_sample(struct perf_output_handle
*handle
,
3988 struct perf_event_header
*header
,
3989 struct perf_sample_data
*data
,
3990 struct perf_event
*event
)
3992 u64 sample_type
= data
->type
;
3994 perf_output_put(handle
, *header
);
3996 if (sample_type
& PERF_SAMPLE_IP
)
3997 perf_output_put(handle
, data
->ip
);
3999 if (sample_type
& PERF_SAMPLE_TID
)
4000 perf_output_put(handle
, data
->tid_entry
);
4002 if (sample_type
& PERF_SAMPLE_TIME
)
4003 perf_output_put(handle
, data
->time
);
4005 if (sample_type
& PERF_SAMPLE_ADDR
)
4006 perf_output_put(handle
, data
->addr
);
4008 if (sample_type
& PERF_SAMPLE_ID
)
4009 perf_output_put(handle
, data
->id
);
4011 if (sample_type
& PERF_SAMPLE_STREAM_ID
)
4012 perf_output_put(handle
, data
->stream_id
);
4014 if (sample_type
& PERF_SAMPLE_CPU
)
4015 perf_output_put(handle
, data
->cpu_entry
);
4017 if (sample_type
& PERF_SAMPLE_PERIOD
)
4018 perf_output_put(handle
, data
->period
);
4020 if (sample_type
& PERF_SAMPLE_READ
)
4021 perf_output_read(handle
, event
);
4023 if (sample_type
& PERF_SAMPLE_CALLCHAIN
) {
4024 if (data
->callchain
) {
4027 if (data
->callchain
)
4028 size
+= data
->callchain
->nr
;
4030 size
*= sizeof(u64
);
4032 __output_copy(handle
, data
->callchain
, size
);
4035 perf_output_put(handle
, nr
);
4039 if (sample_type
& PERF_SAMPLE_RAW
) {
4041 perf_output_put(handle
, data
->raw
->size
);
4042 __output_copy(handle
, data
->raw
->data
,
4049 .size
= sizeof(u32
),
4052 perf_output_put(handle
, raw
);
4056 if (!event
->attr
.watermark
) {
4057 int wakeup_events
= event
->attr
.wakeup_events
;
4059 if (wakeup_events
) {
4060 struct ring_buffer
*rb
= handle
->rb
;
4061 int events
= local_inc_return(&rb
->events
);
4063 if (events
>= wakeup_events
) {
4064 local_sub(wakeup_events
, &rb
->events
);
4065 local_inc(&rb
->wakeup
);
4071 void perf_prepare_sample(struct perf_event_header
*header
,
4072 struct perf_sample_data
*data
,
4073 struct perf_event
*event
,
4074 struct pt_regs
*regs
)
4076 u64 sample_type
= event
->attr
.sample_type
;
4078 header
->type
= PERF_RECORD_SAMPLE
;
4079 header
->size
= sizeof(*header
) + event
->header_size
;
4082 header
->misc
|= perf_misc_flags(regs
);
4084 __perf_event_header__init_id(header
, data
, event
);
4086 if (sample_type
& PERF_SAMPLE_IP
)
4087 data
->ip
= perf_instruction_pointer(regs
);
4089 if (sample_type
& PERF_SAMPLE_CALLCHAIN
) {
4092 data
->callchain
= perf_callchain(regs
);
4094 if (data
->callchain
)
4095 size
+= data
->callchain
->nr
;
4097 header
->size
+= size
* sizeof(u64
);
4100 if (sample_type
& PERF_SAMPLE_RAW
) {
4101 int size
= sizeof(u32
);
4104 size
+= data
->raw
->size
;
4106 size
+= sizeof(u32
);
4108 WARN_ON_ONCE(size
& (sizeof(u64
)-1));
4109 header
->size
+= size
;
4113 static void perf_event_output(struct perf_event
*event
,
4114 struct perf_sample_data
*data
,
4115 struct pt_regs
*regs
)
4117 struct perf_output_handle handle
;
4118 struct perf_event_header header
;
4120 /* protect the callchain buffers */
4123 perf_prepare_sample(&header
, data
, event
, regs
);
4125 if (perf_output_begin(&handle
, event
, header
.size
))
4128 perf_output_sample(&handle
, &header
, data
, event
);
4130 perf_output_end(&handle
);
4140 struct perf_read_event
{
4141 struct perf_event_header header
;
4148 perf_event_read_event(struct perf_event
*event
,
4149 struct task_struct
*task
)
4151 struct perf_output_handle handle
;
4152 struct perf_sample_data sample
;
4153 struct perf_read_event read_event
= {
4155 .type
= PERF_RECORD_READ
,
4157 .size
= sizeof(read_event
) + event
->read_size
,
4159 .pid
= perf_event_pid(event
, task
),
4160 .tid
= perf_event_tid(event
, task
),
4164 perf_event_header__init_id(&read_event
.header
, &sample
, event
);
4165 ret
= perf_output_begin(&handle
, event
, read_event
.header
.size
);
4169 perf_output_put(&handle
, read_event
);
4170 perf_output_read(&handle
, event
);
4171 perf_event__output_id_sample(event
, &handle
, &sample
);
4173 perf_output_end(&handle
);
4177 * task tracking -- fork/exit
4179 * enabled by: attr.comm | attr.mmap | attr.mmap_data | attr.task
4182 struct perf_task_event
{
4183 struct task_struct
*task
;
4184 struct perf_event_context
*task_ctx
;
4187 struct perf_event_header header
;
4197 static void perf_event_task_output(struct perf_event
*event
,
4198 struct perf_task_event
*task_event
)
4200 struct perf_output_handle handle
;
4201 struct perf_sample_data sample
;
4202 struct task_struct
*task
= task_event
->task
;
4203 int ret
, size
= task_event
->event_id
.header
.size
;
4205 perf_event_header__init_id(&task_event
->event_id
.header
, &sample
, event
);
4207 ret
= perf_output_begin(&handle
, event
,
4208 task_event
->event_id
.header
.size
);
4212 task_event
->event_id
.pid
= perf_event_pid(event
, task
);
4213 task_event
->event_id
.ppid
= perf_event_pid(event
, current
);
4215 task_event
->event_id
.tid
= perf_event_tid(event
, task
);
4216 task_event
->event_id
.ptid
= perf_event_tid(event
, current
);
4218 perf_output_put(&handle
, task_event
->event_id
);
4220 perf_event__output_id_sample(event
, &handle
, &sample
);
4222 perf_output_end(&handle
);
4224 task_event
->event_id
.header
.size
= size
;
4227 static int perf_event_task_match(struct perf_event
*event
)
4229 if (event
->state
< PERF_EVENT_STATE_INACTIVE
)
4232 if (!event_filter_match(event
))
4235 if (event
->attr
.comm
|| event
->attr
.mmap
||
4236 event
->attr
.mmap_data
|| event
->attr
.task
)
4242 static void perf_event_task_ctx(struct perf_event_context
*ctx
,
4243 struct perf_task_event
*task_event
)
4245 struct perf_event
*event
;
4247 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
4248 if (perf_event_task_match(event
))
4249 perf_event_task_output(event
, task_event
);
4253 static void perf_event_task_event(struct perf_task_event
*task_event
)
4255 struct perf_cpu_context
*cpuctx
;
4256 struct perf_event_context
*ctx
;
4261 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
4262 cpuctx
= get_cpu_ptr(pmu
->pmu_cpu_context
);
4263 if (cpuctx
->active_pmu
!= pmu
)
4265 perf_event_task_ctx(&cpuctx
->ctx
, task_event
);
4267 ctx
= task_event
->task_ctx
;
4269 ctxn
= pmu
->task_ctx_nr
;
4272 ctx
= rcu_dereference(current
->perf_event_ctxp
[ctxn
]);
4275 perf_event_task_ctx(ctx
, task_event
);
4277 put_cpu_ptr(pmu
->pmu_cpu_context
);
4282 static void perf_event_task(struct task_struct
*task
,
4283 struct perf_event_context
*task_ctx
,
4286 struct perf_task_event task_event
;
4288 if (!atomic_read(&nr_comm_events
) &&
4289 !atomic_read(&nr_mmap_events
) &&
4290 !atomic_read(&nr_task_events
))
4293 task_event
= (struct perf_task_event
){
4295 .task_ctx
= task_ctx
,
4298 .type
= new ? PERF_RECORD_FORK
: PERF_RECORD_EXIT
,
4300 .size
= sizeof(task_event
.event_id
),
4306 .time
= perf_clock(),
4310 perf_event_task_event(&task_event
);
4313 void perf_event_fork(struct task_struct
*task
)
4315 perf_event_task(task
, NULL
, 1);
4322 struct perf_comm_event
{
4323 struct task_struct
*task
;
4328 struct perf_event_header header
;
4335 static void perf_event_comm_output(struct perf_event
*event
,
4336 struct perf_comm_event
*comm_event
)
4338 struct perf_output_handle handle
;
4339 struct perf_sample_data sample
;
4340 int size
= comm_event
->event_id
.header
.size
;
4343 perf_event_header__init_id(&comm_event
->event_id
.header
, &sample
, event
);
4344 ret
= perf_output_begin(&handle
, event
,
4345 comm_event
->event_id
.header
.size
);
4350 comm_event
->event_id
.pid
= perf_event_pid(event
, comm_event
->task
);
4351 comm_event
->event_id
.tid
= perf_event_tid(event
, comm_event
->task
);
4353 perf_output_put(&handle
, comm_event
->event_id
);
4354 __output_copy(&handle
, comm_event
->comm
,
4355 comm_event
->comm_size
);
4357 perf_event__output_id_sample(event
, &handle
, &sample
);
4359 perf_output_end(&handle
);
4361 comm_event
->event_id
.header
.size
= size
;
4364 static int perf_event_comm_match(struct perf_event
*event
)
4366 if (event
->state
< PERF_EVENT_STATE_INACTIVE
)
4369 if (!event_filter_match(event
))
4372 if (event
->attr
.comm
)
4378 static void perf_event_comm_ctx(struct perf_event_context
*ctx
,
4379 struct perf_comm_event
*comm_event
)
4381 struct perf_event
*event
;
4383 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
4384 if (perf_event_comm_match(event
))
4385 perf_event_comm_output(event
, comm_event
);
4389 static void perf_event_comm_event(struct perf_comm_event
*comm_event
)
4391 struct perf_cpu_context
*cpuctx
;
4392 struct perf_event_context
*ctx
;
4393 char comm
[TASK_COMM_LEN
];
4398 memset(comm
, 0, sizeof(comm
));
4399 strlcpy(comm
, comm_event
->task
->comm
, sizeof(comm
));
4400 size
= ALIGN(strlen(comm
)+1, sizeof(u64
));
4402 comm_event
->comm
= comm
;
4403 comm_event
->comm_size
= size
;
4405 comm_event
->event_id
.header
.size
= sizeof(comm_event
->event_id
) + size
;
4407 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
4408 cpuctx
= get_cpu_ptr(pmu
->pmu_cpu_context
);
4409 if (cpuctx
->active_pmu
!= pmu
)
4411 perf_event_comm_ctx(&cpuctx
->ctx
, comm_event
);
4413 ctxn
= pmu
->task_ctx_nr
;
4417 ctx
= rcu_dereference(current
->perf_event_ctxp
[ctxn
]);
4419 perf_event_comm_ctx(ctx
, comm_event
);
4421 put_cpu_ptr(pmu
->pmu_cpu_context
);
4426 void perf_event_comm(struct task_struct
*task
)
4428 struct perf_comm_event comm_event
;
4429 struct perf_event_context
*ctx
;
4432 for_each_task_context_nr(ctxn
) {
4433 ctx
= task
->perf_event_ctxp
[ctxn
];
4437 perf_event_enable_on_exec(ctx
);
4440 if (!atomic_read(&nr_comm_events
))
4443 comm_event
= (struct perf_comm_event
){
4449 .type
= PERF_RECORD_COMM
,
4458 perf_event_comm_event(&comm_event
);
4465 struct perf_mmap_event
{
4466 struct vm_area_struct
*vma
;
4468 const char *file_name
;
4472 struct perf_event_header header
;
4482 static void perf_event_mmap_output(struct perf_event
*event
,
4483 struct perf_mmap_event
*mmap_event
)
4485 struct perf_output_handle handle
;
4486 struct perf_sample_data sample
;
4487 int size
= mmap_event
->event_id
.header
.size
;
4490 perf_event_header__init_id(&mmap_event
->event_id
.header
, &sample
, event
);
4491 ret
= perf_output_begin(&handle
, event
,
4492 mmap_event
->event_id
.header
.size
);
4496 mmap_event
->event_id
.pid
= perf_event_pid(event
, current
);
4497 mmap_event
->event_id
.tid
= perf_event_tid(event
, current
);
4499 perf_output_put(&handle
, mmap_event
->event_id
);
4500 __output_copy(&handle
, mmap_event
->file_name
,
4501 mmap_event
->file_size
);
4503 perf_event__output_id_sample(event
, &handle
, &sample
);
4505 perf_output_end(&handle
);
4507 mmap_event
->event_id
.header
.size
= size
;
4510 static int perf_event_mmap_match(struct perf_event
*event
,
4511 struct perf_mmap_event
*mmap_event
,
4514 if (event
->state
< PERF_EVENT_STATE_INACTIVE
)
4517 if (!event_filter_match(event
))
4520 if ((!executable
&& event
->attr
.mmap_data
) ||
4521 (executable
&& event
->attr
.mmap
))
4527 static void perf_event_mmap_ctx(struct perf_event_context
*ctx
,
4528 struct perf_mmap_event
*mmap_event
,
4531 struct perf_event
*event
;
4533 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
4534 if (perf_event_mmap_match(event
, mmap_event
, executable
))
4535 perf_event_mmap_output(event
, mmap_event
);
4539 static void perf_event_mmap_event(struct perf_mmap_event
*mmap_event
)
4541 struct perf_cpu_context
*cpuctx
;
4542 struct perf_event_context
*ctx
;
4543 struct vm_area_struct
*vma
= mmap_event
->vma
;
4544 struct file
*file
= vma
->vm_file
;
4552 memset(tmp
, 0, sizeof(tmp
));
4556 * d_path works from the end of the rb backwards, so we
4557 * need to add enough zero bytes after the string to handle
4558 * the 64bit alignment we do later.
4560 buf
= kzalloc(PATH_MAX
+ sizeof(u64
), GFP_KERNEL
);
4562 name
= strncpy(tmp
, "//enomem", sizeof(tmp
));
4565 name
= d_path(&file
->f_path
, buf
, PATH_MAX
);
4567 name
= strncpy(tmp
, "//toolong", sizeof(tmp
));
4571 if (arch_vma_name(mmap_event
->vma
)) {
4572 name
= strncpy(tmp
, arch_vma_name(mmap_event
->vma
),
4578 name
= strncpy(tmp
, "[vdso]", sizeof(tmp
));
4580 } else if (vma
->vm_start
<= vma
->vm_mm
->start_brk
&&
4581 vma
->vm_end
>= vma
->vm_mm
->brk
) {
4582 name
= strncpy(tmp
, "[heap]", sizeof(tmp
));
4584 } else if (vma
->vm_start
<= vma
->vm_mm
->start_stack
&&
4585 vma
->vm_end
>= vma
->vm_mm
->start_stack
) {
4586 name
= strncpy(tmp
, "[stack]", sizeof(tmp
));
4590 name
= strncpy(tmp
, "//anon", sizeof(tmp
));
4595 size
= ALIGN(strlen(name
)+1, sizeof(u64
));
4597 mmap_event
->file_name
= name
;
4598 mmap_event
->file_size
= size
;
4600 mmap_event
->event_id
.header
.size
= sizeof(mmap_event
->event_id
) + size
;
4603 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
4604 cpuctx
= get_cpu_ptr(pmu
->pmu_cpu_context
);
4605 if (cpuctx
->active_pmu
!= pmu
)
4607 perf_event_mmap_ctx(&cpuctx
->ctx
, mmap_event
,
4608 vma
->vm_flags
& VM_EXEC
);
4610 ctxn
= pmu
->task_ctx_nr
;
4614 ctx
= rcu_dereference(current
->perf_event_ctxp
[ctxn
]);
4616 perf_event_mmap_ctx(ctx
, mmap_event
,
4617 vma
->vm_flags
& VM_EXEC
);
4620 put_cpu_ptr(pmu
->pmu_cpu_context
);
4627 void perf_event_mmap(struct vm_area_struct
*vma
)
4629 struct perf_mmap_event mmap_event
;
4631 if (!atomic_read(&nr_mmap_events
))
4634 mmap_event
= (struct perf_mmap_event
){
4640 .type
= PERF_RECORD_MMAP
,
4641 .misc
= PERF_RECORD_MISC_USER
,
4646 .start
= vma
->vm_start
,
4647 .len
= vma
->vm_end
- vma
->vm_start
,
4648 .pgoff
= (u64
)vma
->vm_pgoff
<< PAGE_SHIFT
,
4652 perf_event_mmap_event(&mmap_event
);
4656 * IRQ throttle logging
4659 static void perf_log_throttle(struct perf_event
*event
, int enable
)
4661 struct perf_output_handle handle
;
4662 struct perf_sample_data sample
;
4666 struct perf_event_header header
;
4670 } throttle_event
= {
4672 .type
= PERF_RECORD_THROTTLE
,
4674 .size
= sizeof(throttle_event
),
4676 .time
= perf_clock(),
4677 .id
= primary_event_id(event
),
4678 .stream_id
= event
->id
,
4682 throttle_event
.header
.type
= PERF_RECORD_UNTHROTTLE
;
4684 perf_event_header__init_id(&throttle_event
.header
, &sample
, event
);
4686 ret
= perf_output_begin(&handle
, event
,
4687 throttle_event
.header
.size
);
4691 perf_output_put(&handle
, throttle_event
);
4692 perf_event__output_id_sample(event
, &handle
, &sample
);
4693 perf_output_end(&handle
);
4697 * Generic event overflow handling, sampling.
4700 static int __perf_event_overflow(struct perf_event
*event
,
4701 int throttle
, struct perf_sample_data
*data
,
4702 struct pt_regs
*regs
)
4704 int events
= atomic_read(&event
->event_limit
);
4705 struct hw_perf_event
*hwc
= &event
->hw
;
4709 * Non-sampling counters might still use the PMI to fold short
4710 * hardware counters, ignore those.
4712 if (unlikely(!is_sampling_event(event
)))
4715 if (unlikely(hwc
->interrupts
>= max_samples_per_tick
)) {
4717 hwc
->interrupts
= MAX_INTERRUPTS
;
4718 perf_log_throttle(event
, 0);
4724 if (event
->attr
.freq
) {
4725 u64 now
= perf_clock();
4726 s64 delta
= now
- hwc
->freq_time_stamp
;
4728 hwc
->freq_time_stamp
= now
;
4730 if (delta
> 0 && delta
< 2*TICK_NSEC
)
4731 perf_adjust_period(event
, delta
, hwc
->last_period
);
4735 * XXX event_limit might not quite work as expected on inherited
4739 event
->pending_kill
= POLL_IN
;
4740 if (events
&& atomic_dec_and_test(&event
->event_limit
)) {
4742 event
->pending_kill
= POLL_HUP
;
4743 event
->pending_disable
= 1;
4744 irq_work_queue(&event
->pending
);
4747 if (event
->overflow_handler
)
4748 event
->overflow_handler(event
, data
, regs
);
4750 perf_event_output(event
, data
, regs
);
4752 if (event
->fasync
&& event
->pending_kill
) {
4753 event
->pending_wakeup
= 1;
4754 irq_work_queue(&event
->pending
);
4760 int perf_event_overflow(struct perf_event
*event
,
4761 struct perf_sample_data
*data
,
4762 struct pt_regs
*regs
)
4764 return __perf_event_overflow(event
, 1, data
, regs
);
4768 * Generic software event infrastructure
4771 struct swevent_htable
{
4772 struct swevent_hlist
*swevent_hlist
;
4773 struct mutex hlist_mutex
;
4776 /* Recursion avoidance in each contexts */
4777 int recursion
[PERF_NR_CONTEXTS
];
4780 static DEFINE_PER_CPU(struct swevent_htable
, swevent_htable
);
4783 * We directly increment event->count and keep a second value in
4784 * event->hw.period_left to count intervals. This period event
4785 * is kept in the range [-sample_period, 0] so that we can use the
4789 static u64
perf_swevent_set_period(struct perf_event
*event
)
4791 struct hw_perf_event
*hwc
= &event
->hw
;
4792 u64 period
= hwc
->last_period
;
4796 hwc
->last_period
= hwc
->sample_period
;
4799 old
= val
= local64_read(&hwc
->period_left
);
4803 nr
= div64_u64(period
+ val
, period
);
4804 offset
= nr
* period
;
4806 if (local64_cmpxchg(&hwc
->period_left
, old
, val
) != old
)
4812 static void perf_swevent_overflow(struct perf_event
*event
, u64 overflow
,
4813 struct perf_sample_data
*data
,
4814 struct pt_regs
*regs
)
4816 struct hw_perf_event
*hwc
= &event
->hw
;
4819 data
->period
= event
->hw
.last_period
;
4821 overflow
= perf_swevent_set_period(event
);
4823 if (hwc
->interrupts
== MAX_INTERRUPTS
)
4826 for (; overflow
; overflow
--) {
4827 if (__perf_event_overflow(event
, throttle
,
4830 * We inhibit the overflow from happening when
4831 * hwc->interrupts == MAX_INTERRUPTS.
4839 static void perf_swevent_event(struct perf_event
*event
, u64 nr
,
4840 struct perf_sample_data
*data
,
4841 struct pt_regs
*regs
)
4843 struct hw_perf_event
*hwc
= &event
->hw
;
4845 local64_add(nr
, &event
->count
);
4850 if (!is_sampling_event(event
))
4853 if (nr
== 1 && hwc
->sample_period
== 1 && !event
->attr
.freq
)
4854 return perf_swevent_overflow(event
, 1, data
, regs
);
4856 if (local64_add_negative(nr
, &hwc
->period_left
))
4859 perf_swevent_overflow(event
, 0, data
, regs
);
4862 static int perf_exclude_event(struct perf_event
*event
,
4863 struct pt_regs
*regs
)
4865 if (event
->hw
.state
& PERF_HES_STOPPED
)
4869 if (event
->attr
.exclude_user
&& user_mode(regs
))
4872 if (event
->attr
.exclude_kernel
&& !user_mode(regs
))
4879 static int perf_swevent_match(struct perf_event
*event
,
4880 enum perf_type_id type
,
4882 struct perf_sample_data
*data
,
4883 struct pt_regs
*regs
)
4885 if (event
->attr
.type
!= type
)
4888 if (event
->attr
.config
!= event_id
)
4891 if (perf_exclude_event(event
, regs
))
4897 static inline u64
swevent_hash(u64 type
, u32 event_id
)
4899 u64 val
= event_id
| (type
<< 32);
4901 return hash_64(val
, SWEVENT_HLIST_BITS
);
4904 static inline struct hlist_head
*
4905 __find_swevent_head(struct swevent_hlist
*hlist
, u64 type
, u32 event_id
)
4907 u64 hash
= swevent_hash(type
, event_id
);
4909 return &hlist
->heads
[hash
];
4912 /* For the read side: events when they trigger */
4913 static inline struct hlist_head
*
4914 find_swevent_head_rcu(struct swevent_htable
*swhash
, u64 type
, u32 event_id
)
4916 struct swevent_hlist
*hlist
;
4918 hlist
= rcu_dereference(swhash
->swevent_hlist
);
4922 return __find_swevent_head(hlist
, type
, event_id
);
4925 /* For the event head insertion and removal in the hlist */
4926 static inline struct hlist_head
*
4927 find_swevent_head(struct swevent_htable
*swhash
, struct perf_event
*event
)
4929 struct swevent_hlist
*hlist
;
4930 u32 event_id
= event
->attr
.config
;
4931 u64 type
= event
->attr
.type
;
4934 * Event scheduling is always serialized against hlist allocation
4935 * and release. Which makes the protected version suitable here.
4936 * The context lock guarantees that.
4938 hlist
= rcu_dereference_protected(swhash
->swevent_hlist
,
4939 lockdep_is_held(&event
->ctx
->lock
));
4943 return __find_swevent_head(hlist
, type
, event_id
);
4946 static void do_perf_sw_event(enum perf_type_id type
, u32 event_id
,
4948 struct perf_sample_data
*data
,
4949 struct pt_regs
*regs
)
4951 struct swevent_htable
*swhash
= &__get_cpu_var(swevent_htable
);
4952 struct perf_event
*event
;
4953 struct hlist_node
*node
;
4954 struct hlist_head
*head
;
4957 head
= find_swevent_head_rcu(swhash
, type
, event_id
);
4961 hlist_for_each_entry_rcu(event
, node
, head
, hlist_entry
) {
4962 if (perf_swevent_match(event
, type
, event_id
, data
, regs
))
4963 perf_swevent_event(event
, nr
, data
, regs
);
4969 int perf_swevent_get_recursion_context(void)
4971 struct swevent_htable
*swhash
= &__get_cpu_var(swevent_htable
);
4973 return get_recursion_context(swhash
->recursion
);
4975 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context
);
4977 inline void perf_swevent_put_recursion_context(int rctx
)
4979 struct swevent_htable
*swhash
= &__get_cpu_var(swevent_htable
);
4981 put_recursion_context(swhash
->recursion
, rctx
);
4984 void __perf_sw_event(u32 event_id
, u64 nr
, struct pt_regs
*regs
, u64 addr
)
4986 struct perf_sample_data data
;
4989 preempt_disable_notrace();
4990 rctx
= perf_swevent_get_recursion_context();
4994 perf_sample_data_init(&data
, addr
);
4996 do_perf_sw_event(PERF_TYPE_SOFTWARE
, event_id
, nr
, &data
, regs
);
4998 perf_swevent_put_recursion_context(rctx
);
4999 preempt_enable_notrace();
5002 static void perf_swevent_read(struct perf_event
*event
)
5006 static int perf_swevent_add(struct perf_event
*event
, int flags
)
5008 struct swevent_htable
*swhash
= &__get_cpu_var(swevent_htable
);
5009 struct hw_perf_event
*hwc
= &event
->hw
;
5010 struct hlist_head
*head
;
5012 if (is_sampling_event(event
)) {
5013 hwc
->last_period
= hwc
->sample_period
;
5014 perf_swevent_set_period(event
);
5017 hwc
->state
= !(flags
& PERF_EF_START
);
5019 head
= find_swevent_head(swhash
, event
);
5020 if (WARN_ON_ONCE(!head
))
5023 hlist_add_head_rcu(&event
->hlist_entry
, head
);
5028 static void perf_swevent_del(struct perf_event
*event
, int flags
)
5030 hlist_del_rcu(&event
->hlist_entry
);
5033 static void perf_swevent_start(struct perf_event
*event
, int flags
)
5035 event
->hw
.state
= 0;
5038 static void perf_swevent_stop(struct perf_event
*event
, int flags
)
5040 event
->hw
.state
= PERF_HES_STOPPED
;
5043 /* Deref the hlist from the update side */
5044 static inline struct swevent_hlist
*
5045 swevent_hlist_deref(struct swevent_htable
*swhash
)
5047 return rcu_dereference_protected(swhash
->swevent_hlist
,
5048 lockdep_is_held(&swhash
->hlist_mutex
));
5051 static void swevent_hlist_release(struct swevent_htable
*swhash
)
5053 struct swevent_hlist
*hlist
= swevent_hlist_deref(swhash
);
5058 rcu_assign_pointer(swhash
->swevent_hlist
, NULL
);
5059 kfree_rcu(hlist
, rcu_head
);
5062 static void swevent_hlist_put_cpu(struct perf_event
*event
, int cpu
)
5064 struct swevent_htable
*swhash
= &per_cpu(swevent_htable
, cpu
);
5066 mutex_lock(&swhash
->hlist_mutex
);
5068 if (!--swhash
->hlist_refcount
)
5069 swevent_hlist_release(swhash
);
5071 mutex_unlock(&swhash
->hlist_mutex
);
5074 static void swevent_hlist_put(struct perf_event
*event
)
5078 if (event
->cpu
!= -1) {
5079 swevent_hlist_put_cpu(event
, event
->cpu
);
5083 for_each_possible_cpu(cpu
)
5084 swevent_hlist_put_cpu(event
, cpu
);
5087 static int swevent_hlist_get_cpu(struct perf_event
*event
, int cpu
)
5089 struct swevent_htable
*swhash
= &per_cpu(swevent_htable
, cpu
);
5092 mutex_lock(&swhash
->hlist_mutex
);
5094 if (!swevent_hlist_deref(swhash
) && cpu_online(cpu
)) {
5095 struct swevent_hlist
*hlist
;
5097 hlist
= kzalloc(sizeof(*hlist
), GFP_KERNEL
);
5102 rcu_assign_pointer(swhash
->swevent_hlist
, hlist
);
5104 swhash
->hlist_refcount
++;
5106 mutex_unlock(&swhash
->hlist_mutex
);
5111 static int swevent_hlist_get(struct perf_event
*event
)
5114 int cpu
, failed_cpu
;
5116 if (event
->cpu
!= -1)
5117 return swevent_hlist_get_cpu(event
, event
->cpu
);
5120 for_each_possible_cpu(cpu
) {
5121 err
= swevent_hlist_get_cpu(event
, cpu
);
5131 for_each_possible_cpu(cpu
) {
5132 if (cpu
== failed_cpu
)
5134 swevent_hlist_put_cpu(event
, cpu
);
5141 struct jump_label_key perf_swevent_enabled
[PERF_COUNT_SW_MAX
];
5143 static void sw_perf_event_destroy(struct perf_event
*event
)
5145 u64 event_id
= event
->attr
.config
;
5147 WARN_ON(event
->parent
);
5149 jump_label_dec(&perf_swevent_enabled
[event_id
]);
5150 swevent_hlist_put(event
);
5153 static int perf_swevent_init(struct perf_event
*event
)
5155 int event_id
= event
->attr
.config
;
5157 if (event
->attr
.type
!= PERF_TYPE_SOFTWARE
)
5161 case PERF_COUNT_SW_CPU_CLOCK
:
5162 case PERF_COUNT_SW_TASK_CLOCK
:
5169 if (event_id
>= PERF_COUNT_SW_MAX
)
5172 if (!event
->parent
) {
5175 err
= swevent_hlist_get(event
);
5179 jump_label_inc(&perf_swevent_enabled
[event_id
]);
5180 event
->destroy
= sw_perf_event_destroy
;
5186 static struct pmu perf_swevent
= {
5187 .task_ctx_nr
= perf_sw_context
,
5189 .event_init
= perf_swevent_init
,
5190 .add
= perf_swevent_add
,
5191 .del
= perf_swevent_del
,
5192 .start
= perf_swevent_start
,
5193 .stop
= perf_swevent_stop
,
5194 .read
= perf_swevent_read
,
5197 #ifdef CONFIG_EVENT_TRACING
5199 static int perf_tp_filter_match(struct perf_event
*event
,
5200 struct perf_sample_data
*data
)
5202 void *record
= data
->raw
->data
;
5204 if (likely(!event
->filter
) || filter_match_preds(event
->filter
, record
))
5209 static int perf_tp_event_match(struct perf_event
*event
,
5210 struct perf_sample_data
*data
,
5211 struct pt_regs
*regs
)
5213 if (event
->hw
.state
& PERF_HES_STOPPED
)
5216 * All tracepoints are from kernel-space.
5218 if (event
->attr
.exclude_kernel
)
5221 if (!perf_tp_filter_match(event
, data
))
5227 void perf_tp_event(u64 addr
, u64 count
, void *record
, int entry_size
,
5228 struct pt_regs
*regs
, struct hlist_head
*head
, int rctx
)
5230 struct perf_sample_data data
;
5231 struct perf_event
*event
;
5232 struct hlist_node
*node
;
5234 struct perf_raw_record raw
= {
5239 perf_sample_data_init(&data
, addr
);
5242 hlist_for_each_entry_rcu(event
, node
, head
, hlist_entry
) {
5243 if (perf_tp_event_match(event
, &data
, regs
))
5244 perf_swevent_event(event
, count
, &data
, regs
);
5247 perf_swevent_put_recursion_context(rctx
);
5249 EXPORT_SYMBOL_GPL(perf_tp_event
);
5251 static void tp_perf_event_destroy(struct perf_event
*event
)
5253 perf_trace_destroy(event
);
5256 static int perf_tp_event_init(struct perf_event
*event
)
5260 if (event
->attr
.type
!= PERF_TYPE_TRACEPOINT
)
5263 err
= perf_trace_init(event
);
5267 event
->destroy
= tp_perf_event_destroy
;
5272 static struct pmu perf_tracepoint
= {
5273 .task_ctx_nr
= perf_sw_context
,
5275 .event_init
= perf_tp_event_init
,
5276 .add
= perf_trace_add
,
5277 .del
= perf_trace_del
,
5278 .start
= perf_swevent_start
,
5279 .stop
= perf_swevent_stop
,
5280 .read
= perf_swevent_read
,
5283 static inline void perf_tp_register(void)
5285 perf_pmu_register(&perf_tracepoint
, "tracepoint", PERF_TYPE_TRACEPOINT
);
5288 static int perf_event_set_filter(struct perf_event
*event
, void __user
*arg
)
5293 if (event
->attr
.type
!= PERF_TYPE_TRACEPOINT
)
5296 filter_str
= strndup_user(arg
, PAGE_SIZE
);
5297 if (IS_ERR(filter_str
))
5298 return PTR_ERR(filter_str
);
5300 ret
= ftrace_profile_set_filter(event
, event
->attr
.config
, filter_str
);
5306 static void perf_event_free_filter(struct perf_event
*event
)
5308 ftrace_profile_free_filter(event
);
5313 static inline void perf_tp_register(void)
5317 static int perf_event_set_filter(struct perf_event
*event
, void __user
*arg
)
5322 static void perf_event_free_filter(struct perf_event
*event
)
5326 #endif /* CONFIG_EVENT_TRACING */
5328 #ifdef CONFIG_HAVE_HW_BREAKPOINT
5329 void perf_bp_event(struct perf_event
*bp
, void *data
)
5331 struct perf_sample_data sample
;
5332 struct pt_regs
*regs
= data
;
5334 perf_sample_data_init(&sample
, bp
->attr
.bp_addr
);
5336 if (!bp
->hw
.state
&& !perf_exclude_event(bp
, regs
))
5337 perf_swevent_event(bp
, 1, &sample
, regs
);
5342 * hrtimer based swevent callback
5345 static enum hrtimer_restart
perf_swevent_hrtimer(struct hrtimer
*hrtimer
)
5347 enum hrtimer_restart ret
= HRTIMER_RESTART
;
5348 struct perf_sample_data data
;
5349 struct pt_regs
*regs
;
5350 struct perf_event
*event
;
5353 event
= container_of(hrtimer
, struct perf_event
, hw
.hrtimer
);
5355 if (event
->state
!= PERF_EVENT_STATE_ACTIVE
)
5356 return HRTIMER_NORESTART
;
5358 event
->pmu
->read(event
);
5360 perf_sample_data_init(&data
, 0);
5361 data
.period
= event
->hw
.last_period
;
5362 regs
= get_irq_regs();
5364 if (regs
&& !perf_exclude_event(event
, regs
)) {
5365 if (!(event
->attr
.exclude_idle
&& current
->pid
== 0))
5366 if (perf_event_overflow(event
, &data
, regs
))
5367 ret
= HRTIMER_NORESTART
;
5370 period
= max_t(u64
, 10000, event
->hw
.sample_period
);
5371 hrtimer_forward_now(hrtimer
, ns_to_ktime(period
));
5376 static void perf_swevent_start_hrtimer(struct perf_event
*event
)
5378 struct hw_perf_event
*hwc
= &event
->hw
;
5381 if (!is_sampling_event(event
))
5384 period
= local64_read(&hwc
->period_left
);
5389 local64_set(&hwc
->period_left
, 0);
5391 period
= max_t(u64
, 10000, hwc
->sample_period
);
5393 __hrtimer_start_range_ns(&hwc
->hrtimer
,
5394 ns_to_ktime(period
), 0,
5395 HRTIMER_MODE_REL_PINNED
, 0);
5398 static void perf_swevent_cancel_hrtimer(struct perf_event
*event
)
5400 struct hw_perf_event
*hwc
= &event
->hw
;
5402 if (is_sampling_event(event
)) {
5403 ktime_t remaining
= hrtimer_get_remaining(&hwc
->hrtimer
);
5404 local64_set(&hwc
->period_left
, ktime_to_ns(remaining
));
5406 hrtimer_cancel(&hwc
->hrtimer
);
5410 static void perf_swevent_init_hrtimer(struct perf_event
*event
)
5412 struct hw_perf_event
*hwc
= &event
->hw
;
5414 if (!is_sampling_event(event
))
5417 hrtimer_init(&hwc
->hrtimer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
5418 hwc
->hrtimer
.function
= perf_swevent_hrtimer
;
5421 * Since hrtimers have a fixed rate, we can do a static freq->period
5422 * mapping and avoid the whole period adjust feedback stuff.
5424 if (event
->attr
.freq
) {
5425 long freq
= event
->attr
.sample_freq
;
5427 event
->attr
.sample_period
= NSEC_PER_SEC
/ freq
;
5428 hwc
->sample_period
= event
->attr
.sample_period
;
5429 local64_set(&hwc
->period_left
, hwc
->sample_period
);
5430 event
->attr
.freq
= 0;
5435 * Software event: cpu wall time clock
5438 static void cpu_clock_event_update(struct perf_event
*event
)
5443 now
= local_clock();
5444 prev
= local64_xchg(&event
->hw
.prev_count
, now
);
5445 local64_add(now
- prev
, &event
->count
);
5448 static void cpu_clock_event_start(struct perf_event
*event
, int flags
)
5450 local64_set(&event
->hw
.prev_count
, local_clock());
5451 perf_swevent_start_hrtimer(event
);
5454 static void cpu_clock_event_stop(struct perf_event
*event
, int flags
)
5456 perf_swevent_cancel_hrtimer(event
);
5457 cpu_clock_event_update(event
);
5460 static int cpu_clock_event_add(struct perf_event
*event
, int flags
)
5462 if (flags
& PERF_EF_START
)
5463 cpu_clock_event_start(event
, flags
);
5468 static void cpu_clock_event_del(struct perf_event
*event
, int flags
)
5470 cpu_clock_event_stop(event
, flags
);
5473 static void cpu_clock_event_read(struct perf_event
*event
)
5475 cpu_clock_event_update(event
);
5478 static int cpu_clock_event_init(struct perf_event
*event
)
5480 if (event
->attr
.type
!= PERF_TYPE_SOFTWARE
)
5483 if (event
->attr
.config
!= PERF_COUNT_SW_CPU_CLOCK
)
5486 perf_swevent_init_hrtimer(event
);
5491 static struct pmu perf_cpu_clock
= {
5492 .task_ctx_nr
= perf_sw_context
,
5494 .event_init
= cpu_clock_event_init
,
5495 .add
= cpu_clock_event_add
,
5496 .del
= cpu_clock_event_del
,
5497 .start
= cpu_clock_event_start
,
5498 .stop
= cpu_clock_event_stop
,
5499 .read
= cpu_clock_event_read
,
5503 * Software event: task time clock
5506 static void task_clock_event_update(struct perf_event
*event
, u64 now
)
5511 prev
= local64_xchg(&event
->hw
.prev_count
, now
);
5513 local64_add(delta
, &event
->count
);
5516 static void task_clock_event_start(struct perf_event
*event
, int flags
)
5518 local64_set(&event
->hw
.prev_count
, event
->ctx
->time
);
5519 perf_swevent_start_hrtimer(event
);
5522 static void task_clock_event_stop(struct perf_event
*event
, int flags
)
5524 perf_swevent_cancel_hrtimer(event
);
5525 task_clock_event_update(event
, event
->ctx
->time
);
5528 static int task_clock_event_add(struct perf_event
*event
, int flags
)
5530 if (flags
& PERF_EF_START
)
5531 task_clock_event_start(event
, flags
);
5536 static void task_clock_event_del(struct perf_event
*event
, int flags
)
5538 task_clock_event_stop(event
, PERF_EF_UPDATE
);
5541 static void task_clock_event_read(struct perf_event
*event
)
5543 u64 now
= perf_clock();
5544 u64 delta
= now
- event
->ctx
->timestamp
;
5545 u64 time
= event
->ctx
->time
+ delta
;
5547 task_clock_event_update(event
, time
);
5550 static int task_clock_event_init(struct perf_event
*event
)
5552 if (event
->attr
.type
!= PERF_TYPE_SOFTWARE
)
5555 if (event
->attr
.config
!= PERF_COUNT_SW_TASK_CLOCK
)
5558 perf_swevent_init_hrtimer(event
);
5563 static struct pmu perf_task_clock
= {
5564 .task_ctx_nr
= perf_sw_context
,
5566 .event_init
= task_clock_event_init
,
5567 .add
= task_clock_event_add
,
5568 .del
= task_clock_event_del
,
5569 .start
= task_clock_event_start
,
5570 .stop
= task_clock_event_stop
,
5571 .read
= task_clock_event_read
,
5574 static void perf_pmu_nop_void(struct pmu
*pmu
)
5578 static int perf_pmu_nop_int(struct pmu
*pmu
)
5583 static void perf_pmu_start_txn(struct pmu
*pmu
)
5585 perf_pmu_disable(pmu
);
5588 static int perf_pmu_commit_txn(struct pmu
*pmu
)
5590 perf_pmu_enable(pmu
);
5594 static void perf_pmu_cancel_txn(struct pmu
*pmu
)
5596 perf_pmu_enable(pmu
);
5600 * Ensures all contexts with the same task_ctx_nr have the same
5601 * pmu_cpu_context too.
5603 static void *find_pmu_context(int ctxn
)
5610 list_for_each_entry(pmu
, &pmus
, entry
) {
5611 if (pmu
->task_ctx_nr
== ctxn
)
5612 return pmu
->pmu_cpu_context
;
5618 static void update_pmu_context(struct pmu
*pmu
, struct pmu
*old_pmu
)
5622 for_each_possible_cpu(cpu
) {
5623 struct perf_cpu_context
*cpuctx
;
5625 cpuctx
= per_cpu_ptr(pmu
->pmu_cpu_context
, cpu
);
5627 if (cpuctx
->active_pmu
== old_pmu
)
5628 cpuctx
->active_pmu
= pmu
;
5632 static void free_pmu_context(struct pmu
*pmu
)
5636 mutex_lock(&pmus_lock
);
5638 * Like a real lame refcount.
5640 list_for_each_entry(i
, &pmus
, entry
) {
5641 if (i
->pmu_cpu_context
== pmu
->pmu_cpu_context
) {
5642 update_pmu_context(i
, pmu
);
5647 free_percpu(pmu
->pmu_cpu_context
);
5649 mutex_unlock(&pmus_lock
);
5651 static struct idr pmu_idr
;
5654 type_show(struct device
*dev
, struct device_attribute
*attr
, char *page
)
5656 struct pmu
*pmu
= dev_get_drvdata(dev
);
5658 return snprintf(page
, PAGE_SIZE
-1, "%d\n", pmu
->type
);
5661 static struct device_attribute pmu_dev_attrs
[] = {
5666 static int pmu_bus_running
;
5667 static struct bus_type pmu_bus
= {
5668 .name
= "event_source",
5669 .dev_attrs
= pmu_dev_attrs
,
5672 static void pmu_dev_release(struct device
*dev
)
5677 static int pmu_dev_alloc(struct pmu
*pmu
)
5681 pmu
->dev
= kzalloc(sizeof(struct device
), GFP_KERNEL
);
5685 device_initialize(pmu
->dev
);
5686 ret
= dev_set_name(pmu
->dev
, "%s", pmu
->name
);
5690 dev_set_drvdata(pmu
->dev
, pmu
);
5691 pmu
->dev
->bus
= &pmu_bus
;
5692 pmu
->dev
->release
= pmu_dev_release
;
5693 ret
= device_add(pmu
->dev
);
5701 put_device(pmu
->dev
);
5705 static struct lock_class_key cpuctx_mutex
;
5706 static struct lock_class_key cpuctx_lock
;
5708 int perf_pmu_register(struct pmu
*pmu
, char *name
, int type
)
5712 mutex_lock(&pmus_lock
);
5714 pmu
->pmu_disable_count
= alloc_percpu(int);
5715 if (!pmu
->pmu_disable_count
)
5724 int err
= idr_pre_get(&pmu_idr
, GFP_KERNEL
);
5728 err
= idr_get_new_above(&pmu_idr
, pmu
, PERF_TYPE_MAX
, &type
);
5736 if (pmu_bus_running
) {
5737 ret
= pmu_dev_alloc(pmu
);
5743 pmu
->pmu_cpu_context
= find_pmu_context(pmu
->task_ctx_nr
);
5744 if (pmu
->pmu_cpu_context
)
5745 goto got_cpu_context
;
5747 pmu
->pmu_cpu_context
= alloc_percpu(struct perf_cpu_context
);
5748 if (!pmu
->pmu_cpu_context
)
5751 for_each_possible_cpu(cpu
) {
5752 struct perf_cpu_context
*cpuctx
;
5754 cpuctx
= per_cpu_ptr(pmu
->pmu_cpu_context
, cpu
);
5755 __perf_event_init_context(&cpuctx
->ctx
);
5756 lockdep_set_class(&cpuctx
->ctx
.mutex
, &cpuctx_mutex
);
5757 lockdep_set_class(&cpuctx
->ctx
.lock
, &cpuctx_lock
);
5758 cpuctx
->ctx
.type
= cpu_context
;
5759 cpuctx
->ctx
.pmu
= pmu
;
5760 cpuctx
->jiffies_interval
= 1;
5761 INIT_LIST_HEAD(&cpuctx
->rotation_list
);
5762 cpuctx
->active_pmu
= pmu
;
5766 if (!pmu
->start_txn
) {
5767 if (pmu
->pmu_enable
) {
5769 * If we have pmu_enable/pmu_disable calls, install
5770 * transaction stubs that use that to try and batch
5771 * hardware accesses.
5773 pmu
->start_txn
= perf_pmu_start_txn
;
5774 pmu
->commit_txn
= perf_pmu_commit_txn
;
5775 pmu
->cancel_txn
= perf_pmu_cancel_txn
;
5777 pmu
->start_txn
= perf_pmu_nop_void
;
5778 pmu
->commit_txn
= perf_pmu_nop_int
;
5779 pmu
->cancel_txn
= perf_pmu_nop_void
;
5783 if (!pmu
->pmu_enable
) {
5784 pmu
->pmu_enable
= perf_pmu_nop_void
;
5785 pmu
->pmu_disable
= perf_pmu_nop_void
;
5788 list_add_rcu(&pmu
->entry
, &pmus
);
5791 mutex_unlock(&pmus_lock
);
5796 device_del(pmu
->dev
);
5797 put_device(pmu
->dev
);
5800 if (pmu
->type
>= PERF_TYPE_MAX
)
5801 idr_remove(&pmu_idr
, pmu
->type
);
5804 free_percpu(pmu
->pmu_disable_count
);
5808 void perf_pmu_unregister(struct pmu
*pmu
)
5810 mutex_lock(&pmus_lock
);
5811 list_del_rcu(&pmu
->entry
);
5812 mutex_unlock(&pmus_lock
);
5815 * We dereference the pmu list under both SRCU and regular RCU, so
5816 * synchronize against both of those.
5818 synchronize_srcu(&pmus_srcu
);
5821 free_percpu(pmu
->pmu_disable_count
);
5822 if (pmu
->type
>= PERF_TYPE_MAX
)
5823 idr_remove(&pmu_idr
, pmu
->type
);
5824 device_del(pmu
->dev
);
5825 put_device(pmu
->dev
);
5826 free_pmu_context(pmu
);
5829 struct pmu
*perf_init_event(struct perf_event
*event
)
5831 struct pmu
*pmu
= NULL
;
5835 idx
= srcu_read_lock(&pmus_srcu
);
5838 pmu
= idr_find(&pmu_idr
, event
->attr
.type
);
5842 ret
= pmu
->event_init(event
);
5848 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
5850 ret
= pmu
->event_init(event
);
5854 if (ret
!= -ENOENT
) {
5859 pmu
= ERR_PTR(-ENOENT
);
5861 srcu_read_unlock(&pmus_srcu
, idx
);
5867 * Allocate and initialize a event structure
5869 static struct perf_event
*
5870 perf_event_alloc(struct perf_event_attr
*attr
, int cpu
,
5871 struct task_struct
*task
,
5872 struct perf_event
*group_leader
,
5873 struct perf_event
*parent_event
,
5874 perf_overflow_handler_t overflow_handler
,
5878 struct perf_event
*event
;
5879 struct hw_perf_event
*hwc
;
5882 if ((unsigned)cpu
>= nr_cpu_ids
) {
5883 if (!task
|| cpu
!= -1)
5884 return ERR_PTR(-EINVAL
);
5887 event
= kzalloc(sizeof(*event
), GFP_KERNEL
);
5889 return ERR_PTR(-ENOMEM
);
5892 * Single events are their own group leaders, with an
5893 * empty sibling list:
5896 group_leader
= event
;
5898 mutex_init(&event
->child_mutex
);
5899 INIT_LIST_HEAD(&event
->child_list
);
5901 INIT_LIST_HEAD(&event
->group_entry
);
5902 INIT_LIST_HEAD(&event
->event_entry
);
5903 INIT_LIST_HEAD(&event
->sibling_list
);
5904 INIT_LIST_HEAD(&event
->rb_entry
);
5906 init_waitqueue_head(&event
->waitq
);
5907 init_irq_work(&event
->pending
, perf_pending_event
);
5909 mutex_init(&event
->mmap_mutex
);
5912 event
->attr
= *attr
;
5913 event
->group_leader
= group_leader
;
5917 event
->parent
= parent_event
;
5919 event
->ns
= get_pid_ns(current
->nsproxy
->pid_ns
);
5920 event
->id
= atomic64_inc_return(&perf_event_id
);
5922 event
->state
= PERF_EVENT_STATE_INACTIVE
;
5925 event
->attach_state
= PERF_ATTACH_TASK
;
5926 #ifdef CONFIG_HAVE_HW_BREAKPOINT
5928 * hw_breakpoint is a bit difficult here..
5930 if (attr
->type
== PERF_TYPE_BREAKPOINT
)
5931 event
->hw
.bp_target
= task
;
5935 if (!overflow_handler
&& parent_event
) {
5936 overflow_handler
= parent_event
->overflow_handler
;
5937 context
= parent_event
->overflow_handler_context
;
5940 event
->overflow_handler
= overflow_handler
;
5941 event
->overflow_handler_context
= context
;
5944 event
->state
= PERF_EVENT_STATE_OFF
;
5949 hwc
->sample_period
= attr
->sample_period
;
5950 if (attr
->freq
&& attr
->sample_freq
)
5951 hwc
->sample_period
= 1;
5952 hwc
->last_period
= hwc
->sample_period
;
5954 local64_set(&hwc
->period_left
, hwc
->sample_period
);
5957 * we currently do not support PERF_FORMAT_GROUP on inherited events
5959 if (attr
->inherit
&& (attr
->read_format
& PERF_FORMAT_GROUP
))
5962 pmu
= perf_init_event(event
);
5968 else if (IS_ERR(pmu
))
5973 put_pid_ns(event
->ns
);
5975 return ERR_PTR(err
);
5978 if (!event
->parent
) {
5979 if (event
->attach_state
& PERF_ATTACH_TASK
)
5980 jump_label_inc(&perf_sched_events
);
5981 if (event
->attr
.mmap
|| event
->attr
.mmap_data
)
5982 atomic_inc(&nr_mmap_events
);
5983 if (event
->attr
.comm
)
5984 atomic_inc(&nr_comm_events
);
5985 if (event
->attr
.task
)
5986 atomic_inc(&nr_task_events
);
5987 if (event
->attr
.sample_type
& PERF_SAMPLE_CALLCHAIN
) {
5988 err
= get_callchain_buffers();
5991 return ERR_PTR(err
);
5999 static int perf_copy_attr(struct perf_event_attr __user
*uattr
,
6000 struct perf_event_attr
*attr
)
6005 if (!access_ok(VERIFY_WRITE
, uattr
, PERF_ATTR_SIZE_VER0
))
6009 * zero the full structure, so that a short copy will be nice.
6011 memset(attr
, 0, sizeof(*attr
));
6013 ret
= get_user(size
, &uattr
->size
);
6017 if (size
> PAGE_SIZE
) /* silly large */
6020 if (!size
) /* abi compat */
6021 size
= PERF_ATTR_SIZE_VER0
;
6023 if (size
< PERF_ATTR_SIZE_VER0
)
6027 * If we're handed a bigger struct than we know of,
6028 * ensure all the unknown bits are 0 - i.e. new
6029 * user-space does not rely on any kernel feature
6030 * extensions we dont know about yet.
6032 if (size
> sizeof(*attr
)) {
6033 unsigned char __user
*addr
;
6034 unsigned char __user
*end
;
6037 addr
= (void __user
*)uattr
+ sizeof(*attr
);
6038 end
= (void __user
*)uattr
+ size
;
6040 for (; addr
< end
; addr
++) {
6041 ret
= get_user(val
, addr
);
6047 size
= sizeof(*attr
);
6050 ret
= copy_from_user(attr
, uattr
, size
);
6054 if (attr
->__reserved_1
)
6057 if (attr
->sample_type
& ~(PERF_SAMPLE_MAX
-1))
6060 if (attr
->read_format
& ~(PERF_FORMAT_MAX
-1))
6067 put_user(sizeof(*attr
), &uattr
->size
);
6073 perf_event_set_output(struct perf_event
*event
, struct perf_event
*output_event
)
6075 struct ring_buffer
*rb
= NULL
, *old_rb
= NULL
;
6081 /* don't allow circular references */
6082 if (event
== output_event
)
6086 * Don't allow cross-cpu buffers
6088 if (output_event
->cpu
!= event
->cpu
)
6092 * If its not a per-cpu rb, it must be the same task.
6094 if (output_event
->cpu
== -1 && output_event
->ctx
!= event
->ctx
)
6098 mutex_lock(&event
->mmap_mutex
);
6099 /* Can't redirect output if we've got an active mmap() */
6100 if (atomic_read(&event
->mmap_count
))
6104 /* get the rb we want to redirect to */
6105 rb
= ring_buffer_get(output_event
);
6111 rcu_assign_pointer(event
->rb
, rb
);
6113 ring_buffer_detach(event
, old_rb
);
6116 mutex_unlock(&event
->mmap_mutex
);
6119 ring_buffer_put(old_rb
);
6125 * sys_perf_event_open - open a performance event, associate it to a task/cpu
6127 * @attr_uptr: event_id type attributes for monitoring/sampling
6130 * @group_fd: group leader event fd
6132 SYSCALL_DEFINE5(perf_event_open
,
6133 struct perf_event_attr __user
*, attr_uptr
,
6134 pid_t
, pid
, int, cpu
, int, group_fd
, unsigned long, flags
)
6136 struct perf_event
*group_leader
= NULL
, *output_event
= NULL
;
6137 struct perf_event
*event
, *sibling
;
6138 struct perf_event_attr attr
;
6139 struct perf_event_context
*ctx
;
6140 struct file
*event_file
= NULL
;
6141 struct file
*group_file
= NULL
;
6142 struct task_struct
*task
= NULL
;
6146 int fput_needed
= 0;
6149 /* for future expandability... */
6150 if (flags
& ~PERF_FLAG_ALL
)
6153 err
= perf_copy_attr(attr_uptr
, &attr
);
6157 if (!attr
.exclude_kernel
) {
6158 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN
))
6163 if (attr
.sample_freq
> sysctl_perf_event_sample_rate
)
6168 * In cgroup mode, the pid argument is used to pass the fd
6169 * opened to the cgroup directory in cgroupfs. The cpu argument
6170 * designates the cpu on which to monitor threads from that
6173 if ((flags
& PERF_FLAG_PID_CGROUP
) && (pid
== -1 || cpu
== -1))
6176 event_fd
= get_unused_fd_flags(O_RDWR
);
6180 if (group_fd
!= -1) {
6181 group_leader
= perf_fget_light(group_fd
, &fput_needed
);
6182 if (IS_ERR(group_leader
)) {
6183 err
= PTR_ERR(group_leader
);
6186 group_file
= group_leader
->filp
;
6187 if (flags
& PERF_FLAG_FD_OUTPUT
)
6188 output_event
= group_leader
;
6189 if (flags
& PERF_FLAG_FD_NO_GROUP
)
6190 group_leader
= NULL
;
6193 if (pid
!= -1 && !(flags
& PERF_FLAG_PID_CGROUP
)) {
6194 task
= find_lively_task_by_vpid(pid
);
6196 err
= PTR_ERR(task
);
6201 event
= perf_event_alloc(&attr
, cpu
, task
, group_leader
, NULL
,
6203 if (IS_ERR(event
)) {
6204 err
= PTR_ERR(event
);
6208 if (flags
& PERF_FLAG_PID_CGROUP
) {
6209 err
= perf_cgroup_connect(pid
, event
, &attr
, group_leader
);
6214 * - that has cgroup constraint on event->cpu
6215 * - that may need work on context switch
6217 atomic_inc(&per_cpu(perf_cgroup_events
, event
->cpu
));
6218 jump_label_inc(&perf_sched_events
);
6222 * Special case software events and allow them to be part of
6223 * any hardware group.
6228 (is_software_event(event
) != is_software_event(group_leader
))) {
6229 if (is_software_event(event
)) {
6231 * If event and group_leader are not both a software
6232 * event, and event is, then group leader is not.
6234 * Allow the addition of software events to !software
6235 * groups, this is safe because software events never
6238 pmu
= group_leader
->pmu
;
6239 } else if (is_software_event(group_leader
) &&
6240 (group_leader
->group_flags
& PERF_GROUP_SOFTWARE
)) {
6242 * In case the group is a pure software group, and we
6243 * try to add a hardware event, move the whole group to
6244 * the hardware context.
6251 * Get the target context (task or percpu):
6253 ctx
= find_get_context(pmu
, task
, cpu
);
6260 put_task_struct(task
);
6265 * Look up the group leader (we will attach this event to it):
6271 * Do not allow a recursive hierarchy (this new sibling
6272 * becoming part of another group-sibling):
6274 if (group_leader
->group_leader
!= group_leader
)
6277 * Do not allow to attach to a group in a different
6278 * task or CPU context:
6281 if (group_leader
->ctx
->type
!= ctx
->type
)
6284 if (group_leader
->ctx
!= ctx
)
6289 * Only a group leader can be exclusive or pinned
6291 if (attr
.exclusive
|| attr
.pinned
)
6296 err
= perf_event_set_output(event
, output_event
);
6301 event_file
= anon_inode_getfile("[perf_event]", &perf_fops
, event
, O_RDWR
);
6302 if (IS_ERR(event_file
)) {
6303 err
= PTR_ERR(event_file
);
6308 struct perf_event_context
*gctx
= group_leader
->ctx
;
6310 mutex_lock(&gctx
->mutex
);
6311 perf_remove_from_context(group_leader
);
6312 list_for_each_entry(sibling
, &group_leader
->sibling_list
,
6314 perf_remove_from_context(sibling
);
6317 mutex_unlock(&gctx
->mutex
);
6321 event
->filp
= event_file
;
6322 WARN_ON_ONCE(ctx
->parent_ctx
);
6323 mutex_lock(&ctx
->mutex
);
6326 perf_install_in_context(ctx
, group_leader
, cpu
);
6328 list_for_each_entry(sibling
, &group_leader
->sibling_list
,
6330 perf_install_in_context(ctx
, sibling
, cpu
);
6335 perf_install_in_context(ctx
, event
, cpu
);
6337 perf_unpin_context(ctx
);
6338 mutex_unlock(&ctx
->mutex
);
6340 event
->owner
= current
;
6342 mutex_lock(¤t
->perf_event_mutex
);
6343 list_add_tail(&event
->owner_entry
, ¤t
->perf_event_list
);
6344 mutex_unlock(¤t
->perf_event_mutex
);
6347 * Precalculate sample_data sizes
6349 perf_event__header_size(event
);
6350 perf_event__id_header_size(event
);
6353 * Drop the reference on the group_event after placing the
6354 * new event on the sibling_list. This ensures destruction
6355 * of the group leader will find the pointer to itself in
6356 * perf_group_detach().
6358 fput_light(group_file
, fput_needed
);
6359 fd_install(event_fd
, event_file
);
6363 perf_unpin_context(ctx
);
6369 put_task_struct(task
);
6371 fput_light(group_file
, fput_needed
);
6373 put_unused_fd(event_fd
);
6378 * perf_event_create_kernel_counter
6380 * @attr: attributes of the counter to create
6381 * @cpu: cpu in which the counter is bound
6382 * @task: task to profile (NULL for percpu)
6385 perf_event_create_kernel_counter(struct perf_event_attr
*attr
, int cpu
,
6386 struct task_struct
*task
,
6387 perf_overflow_handler_t overflow_handler
,
6390 struct perf_event_context
*ctx
;
6391 struct perf_event
*event
;
6395 * Get the target context (task or percpu):
6398 event
= perf_event_alloc(attr
, cpu
, task
, NULL
, NULL
,
6399 overflow_handler
, context
);
6400 if (IS_ERR(event
)) {
6401 err
= PTR_ERR(event
);
6405 ctx
= find_get_context(event
->pmu
, task
, cpu
);
6412 WARN_ON_ONCE(ctx
->parent_ctx
);
6413 mutex_lock(&ctx
->mutex
);
6414 perf_install_in_context(ctx
, event
, cpu
);
6416 perf_unpin_context(ctx
);
6417 mutex_unlock(&ctx
->mutex
);
6424 return ERR_PTR(err
);
6426 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter
);
6428 static void sync_child_event(struct perf_event
*child_event
,
6429 struct task_struct
*child
)
6431 struct perf_event
*parent_event
= child_event
->parent
;
6434 if (child_event
->attr
.inherit_stat
)
6435 perf_event_read_event(child_event
, child
);
6437 child_val
= perf_event_count(child_event
);
6440 * Add back the child's count to the parent's count:
6442 atomic64_add(child_val
, &parent_event
->child_count
);
6443 atomic64_add(child_event
->total_time_enabled
,
6444 &parent_event
->child_total_time_enabled
);
6445 atomic64_add(child_event
->total_time_running
,
6446 &parent_event
->child_total_time_running
);
6449 * Remove this event from the parent's list
6451 WARN_ON_ONCE(parent_event
->ctx
->parent_ctx
);
6452 mutex_lock(&parent_event
->child_mutex
);
6453 list_del_init(&child_event
->child_list
);
6454 mutex_unlock(&parent_event
->child_mutex
);
6457 * Release the parent event, if this was the last
6460 fput(parent_event
->filp
);
6464 __perf_event_exit_task(struct perf_event
*child_event
,
6465 struct perf_event_context
*child_ctx
,
6466 struct task_struct
*child
)
6468 if (child_event
->parent
) {
6469 raw_spin_lock_irq(&child_ctx
->lock
);
6470 perf_group_detach(child_event
);
6471 raw_spin_unlock_irq(&child_ctx
->lock
);
6474 perf_remove_from_context(child_event
);
6477 * It can happen that the parent exits first, and has events
6478 * that are still around due to the child reference. These
6479 * events need to be zapped.
6481 if (child_event
->parent
) {
6482 sync_child_event(child_event
, child
);
6483 free_event(child_event
);
6487 static void perf_event_exit_task_context(struct task_struct
*child
, int ctxn
)
6489 struct perf_event
*child_event
, *tmp
;
6490 struct perf_event_context
*child_ctx
;
6491 unsigned long flags
;
6493 if (likely(!child
->perf_event_ctxp
[ctxn
])) {
6494 perf_event_task(child
, NULL
, 0);
6498 local_irq_save(flags
);
6500 * We can't reschedule here because interrupts are disabled,
6501 * and either child is current or it is a task that can't be
6502 * scheduled, so we are now safe from rescheduling changing
6505 child_ctx
= rcu_dereference_raw(child
->perf_event_ctxp
[ctxn
]);
6508 * Take the context lock here so that if find_get_context is
6509 * reading child->perf_event_ctxp, we wait until it has
6510 * incremented the context's refcount before we do put_ctx below.
6512 raw_spin_lock(&child_ctx
->lock
);
6513 task_ctx_sched_out(child_ctx
);
6514 child
->perf_event_ctxp
[ctxn
] = NULL
;
6516 * If this context is a clone; unclone it so it can't get
6517 * swapped to another process while we're removing all
6518 * the events from it.
6520 unclone_ctx(child_ctx
);
6521 update_context_time(child_ctx
);
6522 raw_spin_unlock_irqrestore(&child_ctx
->lock
, flags
);
6525 * Report the task dead after unscheduling the events so that we
6526 * won't get any samples after PERF_RECORD_EXIT. We can however still
6527 * get a few PERF_RECORD_READ events.
6529 perf_event_task(child
, child_ctx
, 0);
6532 * We can recurse on the same lock type through:
6534 * __perf_event_exit_task()
6535 * sync_child_event()
6536 * fput(parent_event->filp)
6538 * mutex_lock(&ctx->mutex)
6540 * But since its the parent context it won't be the same instance.
6542 mutex_lock(&child_ctx
->mutex
);
6545 list_for_each_entry_safe(child_event
, tmp
, &child_ctx
->pinned_groups
,
6547 __perf_event_exit_task(child_event
, child_ctx
, child
);
6549 list_for_each_entry_safe(child_event
, tmp
, &child_ctx
->flexible_groups
,
6551 __perf_event_exit_task(child_event
, child_ctx
, child
);
6554 * If the last event was a group event, it will have appended all
6555 * its siblings to the list, but we obtained 'tmp' before that which
6556 * will still point to the list head terminating the iteration.
6558 if (!list_empty(&child_ctx
->pinned_groups
) ||
6559 !list_empty(&child_ctx
->flexible_groups
))
6562 mutex_unlock(&child_ctx
->mutex
);
6568 * When a child task exits, feed back event values to parent events.
6570 void perf_event_exit_task(struct task_struct
*child
)
6572 struct perf_event
*event
, *tmp
;
6575 mutex_lock(&child
->perf_event_mutex
);
6576 list_for_each_entry_safe(event
, tmp
, &child
->perf_event_list
,
6578 list_del_init(&event
->owner_entry
);
6581 * Ensure the list deletion is visible before we clear
6582 * the owner, closes a race against perf_release() where
6583 * we need to serialize on the owner->perf_event_mutex.
6586 event
->owner
= NULL
;
6588 mutex_unlock(&child
->perf_event_mutex
);
6590 for_each_task_context_nr(ctxn
)
6591 perf_event_exit_task_context(child
, ctxn
);
6594 static void perf_free_event(struct perf_event
*event
,
6595 struct perf_event_context
*ctx
)
6597 struct perf_event
*parent
= event
->parent
;
6599 if (WARN_ON_ONCE(!parent
))
6602 mutex_lock(&parent
->child_mutex
);
6603 list_del_init(&event
->child_list
);
6604 mutex_unlock(&parent
->child_mutex
);
6608 perf_group_detach(event
);
6609 list_del_event(event
, ctx
);
6614 * free an unexposed, unused context as created by inheritance by
6615 * perf_event_init_task below, used by fork() in case of fail.
6617 void perf_event_free_task(struct task_struct
*task
)
6619 struct perf_event_context
*ctx
;
6620 struct perf_event
*event
, *tmp
;
6623 for_each_task_context_nr(ctxn
) {
6624 ctx
= task
->perf_event_ctxp
[ctxn
];
6628 mutex_lock(&ctx
->mutex
);
6630 list_for_each_entry_safe(event
, tmp
, &ctx
->pinned_groups
,
6632 perf_free_event(event
, ctx
);
6634 list_for_each_entry_safe(event
, tmp
, &ctx
->flexible_groups
,
6636 perf_free_event(event
, ctx
);
6638 if (!list_empty(&ctx
->pinned_groups
) ||
6639 !list_empty(&ctx
->flexible_groups
))
6642 mutex_unlock(&ctx
->mutex
);
6648 void perf_event_delayed_put(struct task_struct
*task
)
6652 for_each_task_context_nr(ctxn
)
6653 WARN_ON_ONCE(task
->perf_event_ctxp
[ctxn
]);
6657 * inherit a event from parent task to child task:
6659 static struct perf_event
*
6660 inherit_event(struct perf_event
*parent_event
,
6661 struct task_struct
*parent
,
6662 struct perf_event_context
*parent_ctx
,
6663 struct task_struct
*child
,
6664 struct perf_event
*group_leader
,
6665 struct perf_event_context
*child_ctx
)
6667 struct perf_event
*child_event
;
6668 unsigned long flags
;
6671 * Instead of creating recursive hierarchies of events,
6672 * we link inherited events back to the original parent,
6673 * which has a filp for sure, which we use as the reference
6676 if (parent_event
->parent
)
6677 parent_event
= parent_event
->parent
;
6679 child_event
= perf_event_alloc(&parent_event
->attr
,
6682 group_leader
, parent_event
,
6684 if (IS_ERR(child_event
))
6689 * Make the child state follow the state of the parent event,
6690 * not its attr.disabled bit. We hold the parent's mutex,
6691 * so we won't race with perf_event_{en, dis}able_family.
6693 if (parent_event
->state
>= PERF_EVENT_STATE_INACTIVE
)
6694 child_event
->state
= PERF_EVENT_STATE_INACTIVE
;
6696 child_event
->state
= PERF_EVENT_STATE_OFF
;
6698 if (parent_event
->attr
.freq
) {
6699 u64 sample_period
= parent_event
->hw
.sample_period
;
6700 struct hw_perf_event
*hwc
= &child_event
->hw
;
6702 hwc
->sample_period
= sample_period
;
6703 hwc
->last_period
= sample_period
;
6705 local64_set(&hwc
->period_left
, sample_period
);
6708 child_event
->ctx
= child_ctx
;
6709 child_event
->overflow_handler
= parent_event
->overflow_handler
;
6710 child_event
->overflow_handler_context
6711 = parent_event
->overflow_handler_context
;
6714 * Precalculate sample_data sizes
6716 perf_event__header_size(child_event
);
6717 perf_event__id_header_size(child_event
);
6720 * Link it up in the child's context:
6722 raw_spin_lock_irqsave(&child_ctx
->lock
, flags
);
6723 add_event_to_ctx(child_event
, child_ctx
);
6724 raw_spin_unlock_irqrestore(&child_ctx
->lock
, flags
);
6727 * Get a reference to the parent filp - we will fput it
6728 * when the child event exits. This is safe to do because
6729 * we are in the parent and we know that the filp still
6730 * exists and has a nonzero count:
6732 atomic_long_inc(&parent_event
->filp
->f_count
);
6735 * Link this into the parent event's child list
6737 WARN_ON_ONCE(parent_event
->ctx
->parent_ctx
);
6738 mutex_lock(&parent_event
->child_mutex
);
6739 list_add_tail(&child_event
->child_list
, &parent_event
->child_list
);
6740 mutex_unlock(&parent_event
->child_mutex
);
6745 static int inherit_group(struct perf_event
*parent_event
,
6746 struct task_struct
*parent
,
6747 struct perf_event_context
*parent_ctx
,
6748 struct task_struct
*child
,
6749 struct perf_event_context
*child_ctx
)
6751 struct perf_event
*leader
;
6752 struct perf_event
*sub
;
6753 struct perf_event
*child_ctr
;
6755 leader
= inherit_event(parent_event
, parent
, parent_ctx
,
6756 child
, NULL
, child_ctx
);
6758 return PTR_ERR(leader
);
6759 list_for_each_entry(sub
, &parent_event
->sibling_list
, group_entry
) {
6760 child_ctr
= inherit_event(sub
, parent
, parent_ctx
,
6761 child
, leader
, child_ctx
);
6762 if (IS_ERR(child_ctr
))
6763 return PTR_ERR(child_ctr
);
6769 inherit_task_group(struct perf_event
*event
, struct task_struct
*parent
,
6770 struct perf_event_context
*parent_ctx
,
6771 struct task_struct
*child
, int ctxn
,
6775 struct perf_event_context
*child_ctx
;
6777 if (!event
->attr
.inherit
) {
6782 child_ctx
= child
->perf_event_ctxp
[ctxn
];
6785 * This is executed from the parent task context, so
6786 * inherit events that have been marked for cloning.
6787 * First allocate and initialize a context for the
6791 child_ctx
= alloc_perf_context(event
->pmu
, child
);
6795 child
->perf_event_ctxp
[ctxn
] = child_ctx
;
6798 ret
= inherit_group(event
, parent
, parent_ctx
,
6808 * Initialize the perf_event context in task_struct
6810 int perf_event_init_context(struct task_struct
*child
, int ctxn
)
6812 struct perf_event_context
*child_ctx
, *parent_ctx
;
6813 struct perf_event_context
*cloned_ctx
;
6814 struct perf_event
*event
;
6815 struct task_struct
*parent
= current
;
6816 int inherited_all
= 1;
6817 unsigned long flags
;
6820 if (likely(!parent
->perf_event_ctxp
[ctxn
]))
6824 * If the parent's context is a clone, pin it so it won't get
6827 parent_ctx
= perf_pin_task_context(parent
, ctxn
);
6830 * No need to check if parent_ctx != NULL here; since we saw
6831 * it non-NULL earlier, the only reason for it to become NULL
6832 * is if we exit, and since we're currently in the middle of
6833 * a fork we can't be exiting at the same time.
6837 * Lock the parent list. No need to lock the child - not PID
6838 * hashed yet and not running, so nobody can access it.
6840 mutex_lock(&parent_ctx
->mutex
);
6843 * We dont have to disable NMIs - we are only looking at
6844 * the list, not manipulating it:
6846 list_for_each_entry(event
, &parent_ctx
->pinned_groups
, group_entry
) {
6847 ret
= inherit_task_group(event
, parent
, parent_ctx
,
6848 child
, ctxn
, &inherited_all
);
6854 * We can't hold ctx->lock when iterating the ->flexible_group list due
6855 * to allocations, but we need to prevent rotation because
6856 * rotate_ctx() will change the list from interrupt context.
6858 raw_spin_lock_irqsave(&parent_ctx
->lock
, flags
);
6859 parent_ctx
->rotate_disable
= 1;
6860 raw_spin_unlock_irqrestore(&parent_ctx
->lock
, flags
);
6862 list_for_each_entry(event
, &parent_ctx
->flexible_groups
, group_entry
) {
6863 ret
= inherit_task_group(event
, parent
, parent_ctx
,
6864 child
, ctxn
, &inherited_all
);
6869 raw_spin_lock_irqsave(&parent_ctx
->lock
, flags
);
6870 parent_ctx
->rotate_disable
= 0;
6872 child_ctx
= child
->perf_event_ctxp
[ctxn
];
6874 if (child_ctx
&& inherited_all
) {
6876 * Mark the child context as a clone of the parent
6877 * context, or of whatever the parent is a clone of.
6879 * Note that if the parent is a clone, the holding of
6880 * parent_ctx->lock avoids it from being uncloned.
6882 cloned_ctx
= parent_ctx
->parent_ctx
;
6884 child_ctx
->parent_ctx
= cloned_ctx
;
6885 child_ctx
->parent_gen
= parent_ctx
->parent_gen
;
6887 child_ctx
->parent_ctx
= parent_ctx
;
6888 child_ctx
->parent_gen
= parent_ctx
->generation
;
6890 get_ctx(child_ctx
->parent_ctx
);
6893 raw_spin_unlock_irqrestore(&parent_ctx
->lock
, flags
);
6894 mutex_unlock(&parent_ctx
->mutex
);
6896 perf_unpin_context(parent_ctx
);
6897 put_ctx(parent_ctx
);
6903 * Initialize the perf_event context in task_struct
6905 int perf_event_init_task(struct task_struct
*child
)
6909 memset(child
->perf_event_ctxp
, 0, sizeof(child
->perf_event_ctxp
));
6910 mutex_init(&child
->perf_event_mutex
);
6911 INIT_LIST_HEAD(&child
->perf_event_list
);
6913 for_each_task_context_nr(ctxn
) {
6914 ret
= perf_event_init_context(child
, ctxn
);
6922 static void __init
perf_event_init_all_cpus(void)
6924 struct swevent_htable
*swhash
;
6927 for_each_possible_cpu(cpu
) {
6928 swhash
= &per_cpu(swevent_htable
, cpu
);
6929 mutex_init(&swhash
->hlist_mutex
);
6930 INIT_LIST_HEAD(&per_cpu(rotation_list
, cpu
));
6934 static void __cpuinit
perf_event_init_cpu(int cpu
)
6936 struct swevent_htable
*swhash
= &per_cpu(swevent_htable
, cpu
);
6938 mutex_lock(&swhash
->hlist_mutex
);
6939 if (swhash
->hlist_refcount
> 0) {
6940 struct swevent_hlist
*hlist
;
6942 hlist
= kzalloc_node(sizeof(*hlist
), GFP_KERNEL
, cpu_to_node(cpu
));
6944 rcu_assign_pointer(swhash
->swevent_hlist
, hlist
);
6946 mutex_unlock(&swhash
->hlist_mutex
);
6949 #if defined CONFIG_HOTPLUG_CPU || defined CONFIG_KEXEC
6950 static void perf_pmu_rotate_stop(struct pmu
*pmu
)
6952 struct perf_cpu_context
*cpuctx
= this_cpu_ptr(pmu
->pmu_cpu_context
);
6954 WARN_ON(!irqs_disabled());
6956 list_del_init(&cpuctx
->rotation_list
);
6959 static void __perf_event_exit_context(void *__info
)
6961 struct perf_event_context
*ctx
= __info
;
6962 struct perf_event
*event
, *tmp
;
6964 perf_pmu_rotate_stop(ctx
->pmu
);
6966 list_for_each_entry_safe(event
, tmp
, &ctx
->pinned_groups
, group_entry
)
6967 __perf_remove_from_context(event
);
6968 list_for_each_entry_safe(event
, tmp
, &ctx
->flexible_groups
, group_entry
)
6969 __perf_remove_from_context(event
);
6972 static void perf_event_exit_cpu_context(int cpu
)
6974 struct perf_event_context
*ctx
;
6978 idx
= srcu_read_lock(&pmus_srcu
);
6979 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
6980 ctx
= &per_cpu_ptr(pmu
->pmu_cpu_context
, cpu
)->ctx
;
6982 mutex_lock(&ctx
->mutex
);
6983 smp_call_function_single(cpu
, __perf_event_exit_context
, ctx
, 1);
6984 mutex_unlock(&ctx
->mutex
);
6986 srcu_read_unlock(&pmus_srcu
, idx
);
6989 static void perf_event_exit_cpu(int cpu
)
6991 struct swevent_htable
*swhash
= &per_cpu(swevent_htable
, cpu
);
6993 mutex_lock(&swhash
->hlist_mutex
);
6994 swevent_hlist_release(swhash
);
6995 mutex_unlock(&swhash
->hlist_mutex
);
6997 perf_event_exit_cpu_context(cpu
);
7000 static inline void perf_event_exit_cpu(int cpu
) { }
7004 perf_reboot(struct notifier_block
*notifier
, unsigned long val
, void *v
)
7008 for_each_online_cpu(cpu
)
7009 perf_event_exit_cpu(cpu
);
7015 * Run the perf reboot notifier at the very last possible moment so that
7016 * the generic watchdog code runs as long as possible.
7018 static struct notifier_block perf_reboot_notifier
= {
7019 .notifier_call
= perf_reboot
,
7020 .priority
= INT_MIN
,
7023 static int __cpuinit
7024 perf_cpu_notify(struct notifier_block
*self
, unsigned long action
, void *hcpu
)
7026 unsigned int cpu
= (long)hcpu
;
7028 switch (action
& ~CPU_TASKS_FROZEN
) {
7030 case CPU_UP_PREPARE
:
7031 case CPU_DOWN_FAILED
:
7032 perf_event_init_cpu(cpu
);
7035 case CPU_UP_CANCELED
:
7036 case CPU_DOWN_PREPARE
:
7037 perf_event_exit_cpu(cpu
);
7047 void __init
perf_event_init(void)
7053 perf_event_init_all_cpus();
7054 init_srcu_struct(&pmus_srcu
);
7055 perf_pmu_register(&perf_swevent
, "software", PERF_TYPE_SOFTWARE
);
7056 perf_pmu_register(&perf_cpu_clock
, NULL
, -1);
7057 perf_pmu_register(&perf_task_clock
, NULL
, -1);
7059 perf_cpu_notifier(perf_cpu_notify
);
7060 register_reboot_notifier(&perf_reboot_notifier
);
7062 ret
= init_hw_breakpoint();
7063 WARN(ret
, "hw_breakpoint initialization failed with: %d", ret
);
7066 static int __init
perf_event_sysfs_init(void)
7071 mutex_lock(&pmus_lock
);
7073 ret
= bus_register(&pmu_bus
);
7077 list_for_each_entry(pmu
, &pmus
, entry
) {
7078 if (!pmu
->name
|| pmu
->type
< 0)
7081 ret
= pmu_dev_alloc(pmu
);
7082 WARN(ret
, "Failed to register pmu: %s, reason %d\n", pmu
->name
, ret
);
7084 pmu_bus_running
= 1;
7088 mutex_unlock(&pmus_lock
);
7092 device_initcall(perf_event_sysfs_init
);
7094 #ifdef CONFIG_CGROUP_PERF
7095 static struct cgroup_subsys_state
*perf_cgroup_create(
7096 struct cgroup_subsys
*ss
, struct cgroup
*cont
)
7098 struct perf_cgroup
*jc
;
7100 jc
= kzalloc(sizeof(*jc
), GFP_KERNEL
);
7102 return ERR_PTR(-ENOMEM
);
7104 jc
->info
= alloc_percpu(struct perf_cgroup_info
);
7107 return ERR_PTR(-ENOMEM
);
7113 static void perf_cgroup_destroy(struct cgroup_subsys
*ss
,
7114 struct cgroup
*cont
)
7116 struct perf_cgroup
*jc
;
7117 jc
= container_of(cgroup_subsys_state(cont
, perf_subsys_id
),
7118 struct perf_cgroup
, css
);
7119 free_percpu(jc
->info
);
7123 static int __perf_cgroup_move(void *info
)
7125 struct task_struct
*task
= info
;
7126 perf_cgroup_switch(task
, PERF_CGROUP_SWOUT
| PERF_CGROUP_SWIN
);
7131 perf_cgroup_attach_task(struct cgroup
*cgrp
, struct task_struct
*task
)
7133 task_function_call(task
, __perf_cgroup_move
, task
);
7136 static void perf_cgroup_exit(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
7137 struct cgroup
*old_cgrp
, struct task_struct
*task
)
7140 * cgroup_exit() is called in the copy_process() failure path.
7141 * Ignore this case since the task hasn't ran yet, this avoids
7142 * trying to poke a half freed task state from generic code.
7144 if (!(task
->flags
& PF_EXITING
))
7147 perf_cgroup_attach_task(cgrp
, task
);
7150 struct cgroup_subsys perf_subsys
= {
7151 .name
= "perf_event",
7152 .subsys_id
= perf_subsys_id
,
7153 .create
= perf_cgroup_create
,
7154 .destroy
= perf_cgroup_destroy
,
7155 .exit
= perf_cgroup_exit
,
7156 .attach_task
= perf_cgroup_attach_task
,
7158 #endif /* CONFIG_CGROUP_PERF */