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 static_key_deferred 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_cgroup_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
.freq
&& event
->attr
.sample_freq
)
1135 if (event
->attr
.exclusive
|| !cpuctx
->active_oncpu
)
1136 cpuctx
->exclusive
= 0;
1140 group_sched_out(struct perf_event
*group_event
,
1141 struct perf_cpu_context
*cpuctx
,
1142 struct perf_event_context
*ctx
)
1144 struct perf_event
*event
;
1145 int state
= group_event
->state
;
1147 event_sched_out(group_event
, cpuctx
, ctx
);
1150 * Schedule out siblings (if any):
1152 list_for_each_entry(event
, &group_event
->sibling_list
, group_entry
)
1153 event_sched_out(event
, cpuctx
, ctx
);
1155 if (state
== PERF_EVENT_STATE_ACTIVE
&& group_event
->attr
.exclusive
)
1156 cpuctx
->exclusive
= 0;
1160 * Cross CPU call to remove a performance event
1162 * We disable the event on the hardware level first. After that we
1163 * remove it from the context list.
1165 static int __perf_remove_from_context(void *info
)
1167 struct perf_event
*event
= info
;
1168 struct perf_event_context
*ctx
= event
->ctx
;
1169 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
1171 raw_spin_lock(&ctx
->lock
);
1172 event_sched_out(event
, cpuctx
, ctx
);
1173 list_del_event(event
, ctx
);
1174 if (!ctx
->nr_events
&& cpuctx
->task_ctx
== ctx
) {
1176 cpuctx
->task_ctx
= NULL
;
1178 raw_spin_unlock(&ctx
->lock
);
1185 * Remove the event from a task's (or a CPU's) list of events.
1187 * CPU events are removed with a smp call. For task events we only
1188 * call when the task is on a CPU.
1190 * If event->ctx is a cloned context, callers must make sure that
1191 * every task struct that event->ctx->task could possibly point to
1192 * remains valid. This is OK when called from perf_release since
1193 * that only calls us on the top-level context, which can't be a clone.
1194 * When called from perf_event_exit_task, it's OK because the
1195 * context has been detached from its task.
1197 static void perf_remove_from_context(struct perf_event
*event
)
1199 struct perf_event_context
*ctx
= event
->ctx
;
1200 struct task_struct
*task
= ctx
->task
;
1202 lockdep_assert_held(&ctx
->mutex
);
1206 * Per cpu events are removed via an smp call and
1207 * the removal is always successful.
1209 cpu_function_call(event
->cpu
, __perf_remove_from_context
, event
);
1214 if (!task_function_call(task
, __perf_remove_from_context
, event
))
1217 raw_spin_lock_irq(&ctx
->lock
);
1219 * If we failed to find a running task, but find the context active now
1220 * that we've acquired the ctx->lock, retry.
1222 if (ctx
->is_active
) {
1223 raw_spin_unlock_irq(&ctx
->lock
);
1228 * Since the task isn't running, its safe to remove the event, us
1229 * holding the ctx->lock ensures the task won't get scheduled in.
1231 list_del_event(event
, ctx
);
1232 raw_spin_unlock_irq(&ctx
->lock
);
1236 * Cross CPU call to disable a performance event
1238 static int __perf_event_disable(void *info
)
1240 struct perf_event
*event
= info
;
1241 struct perf_event_context
*ctx
= event
->ctx
;
1242 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
1245 * If this is a per-task event, need to check whether this
1246 * event's task is the current task on this cpu.
1248 * Can trigger due to concurrent perf_event_context_sched_out()
1249 * flipping contexts around.
1251 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
)
1254 raw_spin_lock(&ctx
->lock
);
1257 * If the event is on, turn it off.
1258 * If it is in error state, leave it in error state.
1260 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
) {
1261 update_context_time(ctx
);
1262 update_cgrp_time_from_event(event
);
1263 update_group_times(event
);
1264 if (event
== event
->group_leader
)
1265 group_sched_out(event
, cpuctx
, ctx
);
1267 event_sched_out(event
, cpuctx
, ctx
);
1268 event
->state
= PERF_EVENT_STATE_OFF
;
1271 raw_spin_unlock(&ctx
->lock
);
1279 * If event->ctx is a cloned context, callers must make sure that
1280 * every task struct that event->ctx->task could possibly point to
1281 * remains valid. This condition is satisifed when called through
1282 * perf_event_for_each_child or perf_event_for_each because they
1283 * hold the top-level event's child_mutex, so any descendant that
1284 * goes to exit will block in sync_child_event.
1285 * When called from perf_pending_event it's OK because event->ctx
1286 * is the current context on this CPU and preemption is disabled,
1287 * hence we can't get into perf_event_task_sched_out for this context.
1289 void perf_event_disable(struct perf_event
*event
)
1291 struct perf_event_context
*ctx
= event
->ctx
;
1292 struct task_struct
*task
= ctx
->task
;
1296 * Disable the event on the cpu that it's on
1298 cpu_function_call(event
->cpu
, __perf_event_disable
, event
);
1303 if (!task_function_call(task
, __perf_event_disable
, event
))
1306 raw_spin_lock_irq(&ctx
->lock
);
1308 * If the event is still active, we need to retry the cross-call.
1310 if (event
->state
== PERF_EVENT_STATE_ACTIVE
) {
1311 raw_spin_unlock_irq(&ctx
->lock
);
1313 * Reload the task pointer, it might have been changed by
1314 * a concurrent perf_event_context_sched_out().
1321 * Since we have the lock this context can't be scheduled
1322 * in, so we can change the state safely.
1324 if (event
->state
== PERF_EVENT_STATE_INACTIVE
) {
1325 update_group_times(event
);
1326 event
->state
= PERF_EVENT_STATE_OFF
;
1328 raw_spin_unlock_irq(&ctx
->lock
);
1330 EXPORT_SYMBOL_GPL(perf_event_disable
);
1332 static void perf_set_shadow_time(struct perf_event
*event
,
1333 struct perf_event_context
*ctx
,
1337 * use the correct time source for the time snapshot
1339 * We could get by without this by leveraging the
1340 * fact that to get to this function, the caller
1341 * has most likely already called update_context_time()
1342 * and update_cgrp_time_xx() and thus both timestamp
1343 * are identical (or very close). Given that tstamp is,
1344 * already adjusted for cgroup, we could say that:
1345 * tstamp - ctx->timestamp
1347 * tstamp - cgrp->timestamp.
1349 * Then, in perf_output_read(), the calculation would
1350 * work with no changes because:
1351 * - event is guaranteed scheduled in
1352 * - no scheduled out in between
1353 * - thus the timestamp would be the same
1355 * But this is a bit hairy.
1357 * So instead, we have an explicit cgroup call to remain
1358 * within the time time source all along. We believe it
1359 * is cleaner and simpler to understand.
1361 if (is_cgroup_event(event
))
1362 perf_cgroup_set_shadow_time(event
, tstamp
);
1364 event
->shadow_ctx_time
= tstamp
- ctx
->timestamp
;
1367 #define MAX_INTERRUPTS (~0ULL)
1369 static void perf_log_throttle(struct perf_event
*event
, int enable
);
1372 event_sched_in(struct perf_event
*event
,
1373 struct perf_cpu_context
*cpuctx
,
1374 struct perf_event_context
*ctx
)
1376 u64 tstamp
= perf_event_time(event
);
1378 if (event
->state
<= PERF_EVENT_STATE_OFF
)
1381 event
->state
= PERF_EVENT_STATE_ACTIVE
;
1382 event
->oncpu
= smp_processor_id();
1385 * Unthrottle events, since we scheduled we might have missed several
1386 * ticks already, also for a heavily scheduling task there is little
1387 * guarantee it'll get a tick in a timely manner.
1389 if (unlikely(event
->hw
.interrupts
== MAX_INTERRUPTS
)) {
1390 perf_log_throttle(event
, 1);
1391 event
->hw
.interrupts
= 0;
1395 * The new state must be visible before we turn it on in the hardware:
1399 if (event
->pmu
->add(event
, PERF_EF_START
)) {
1400 event
->state
= PERF_EVENT_STATE_INACTIVE
;
1405 event
->tstamp_running
+= tstamp
- event
->tstamp_stopped
;
1407 perf_set_shadow_time(event
, ctx
, tstamp
);
1409 if (!is_software_event(event
))
1410 cpuctx
->active_oncpu
++;
1412 if (event
->attr
.freq
&& event
->attr
.sample_freq
)
1415 if (event
->attr
.exclusive
)
1416 cpuctx
->exclusive
= 1;
1422 group_sched_in(struct perf_event
*group_event
,
1423 struct perf_cpu_context
*cpuctx
,
1424 struct perf_event_context
*ctx
)
1426 struct perf_event
*event
, *partial_group
= NULL
;
1427 struct pmu
*pmu
= group_event
->pmu
;
1428 u64 now
= ctx
->time
;
1429 bool simulate
= false;
1431 if (group_event
->state
== PERF_EVENT_STATE_OFF
)
1434 pmu
->start_txn(pmu
);
1436 if (event_sched_in(group_event
, cpuctx
, ctx
)) {
1437 pmu
->cancel_txn(pmu
);
1442 * Schedule in siblings as one group (if any):
1444 list_for_each_entry(event
, &group_event
->sibling_list
, group_entry
) {
1445 if (event_sched_in(event
, cpuctx
, ctx
)) {
1446 partial_group
= event
;
1451 if (!pmu
->commit_txn(pmu
))
1456 * Groups can be scheduled in as one unit only, so undo any
1457 * partial group before returning:
1458 * The events up to the failed event are scheduled out normally,
1459 * tstamp_stopped will be updated.
1461 * The failed events and the remaining siblings need to have
1462 * their timings updated as if they had gone thru event_sched_in()
1463 * and event_sched_out(). This is required to get consistent timings
1464 * across the group. This also takes care of the case where the group
1465 * could never be scheduled by ensuring tstamp_stopped is set to mark
1466 * the time the event was actually stopped, such that time delta
1467 * calculation in update_event_times() is correct.
1469 list_for_each_entry(event
, &group_event
->sibling_list
, group_entry
) {
1470 if (event
== partial_group
)
1474 event
->tstamp_running
+= now
- event
->tstamp_stopped
;
1475 event
->tstamp_stopped
= now
;
1477 event_sched_out(event
, cpuctx
, ctx
);
1480 event_sched_out(group_event
, cpuctx
, ctx
);
1482 pmu
->cancel_txn(pmu
);
1488 * Work out whether we can put this event group on the CPU now.
1490 static int group_can_go_on(struct perf_event
*event
,
1491 struct perf_cpu_context
*cpuctx
,
1495 * Groups consisting entirely of software events can always go on.
1497 if (event
->group_flags
& PERF_GROUP_SOFTWARE
)
1500 * If an exclusive group is already on, no other hardware
1503 if (cpuctx
->exclusive
)
1506 * If this group is exclusive and there are already
1507 * events on the CPU, it can't go on.
1509 if (event
->attr
.exclusive
&& cpuctx
->active_oncpu
)
1512 * Otherwise, try to add it if all previous groups were able
1518 static void add_event_to_ctx(struct perf_event
*event
,
1519 struct perf_event_context
*ctx
)
1521 u64 tstamp
= perf_event_time(event
);
1523 list_add_event(event
, ctx
);
1524 perf_group_attach(event
);
1525 event
->tstamp_enabled
= tstamp
;
1526 event
->tstamp_running
= tstamp
;
1527 event
->tstamp_stopped
= tstamp
;
1530 static void task_ctx_sched_out(struct perf_event_context
*ctx
);
1532 ctx_sched_in(struct perf_event_context
*ctx
,
1533 struct perf_cpu_context
*cpuctx
,
1534 enum event_type_t event_type
,
1535 struct task_struct
*task
);
1537 static void perf_event_sched_in(struct perf_cpu_context
*cpuctx
,
1538 struct perf_event_context
*ctx
,
1539 struct task_struct
*task
)
1541 cpu_ctx_sched_in(cpuctx
, EVENT_PINNED
, task
);
1543 ctx_sched_in(ctx
, cpuctx
, EVENT_PINNED
, task
);
1544 cpu_ctx_sched_in(cpuctx
, EVENT_FLEXIBLE
, task
);
1546 ctx_sched_in(ctx
, cpuctx
, EVENT_FLEXIBLE
, task
);
1550 * Cross CPU call to install and enable a performance event
1552 * Must be called with ctx->mutex held
1554 static int __perf_install_in_context(void *info
)
1556 struct perf_event
*event
= info
;
1557 struct perf_event_context
*ctx
= event
->ctx
;
1558 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
1559 struct perf_event_context
*task_ctx
= cpuctx
->task_ctx
;
1560 struct task_struct
*task
= current
;
1562 perf_ctx_lock(cpuctx
, task_ctx
);
1563 perf_pmu_disable(cpuctx
->ctx
.pmu
);
1566 * If there was an active task_ctx schedule it out.
1569 task_ctx_sched_out(task_ctx
);
1572 * If the context we're installing events in is not the
1573 * active task_ctx, flip them.
1575 if (ctx
->task
&& task_ctx
!= ctx
) {
1577 raw_spin_unlock(&task_ctx
->lock
);
1578 raw_spin_lock(&ctx
->lock
);
1583 cpuctx
->task_ctx
= task_ctx
;
1584 task
= task_ctx
->task
;
1587 cpu_ctx_sched_out(cpuctx
, EVENT_ALL
);
1589 update_context_time(ctx
);
1591 * update cgrp time only if current cgrp
1592 * matches event->cgrp. Must be done before
1593 * calling add_event_to_ctx()
1595 update_cgrp_time_from_event(event
);
1597 add_event_to_ctx(event
, ctx
);
1600 * Schedule everything back in
1602 perf_event_sched_in(cpuctx
, task_ctx
, task
);
1604 perf_pmu_enable(cpuctx
->ctx
.pmu
);
1605 perf_ctx_unlock(cpuctx
, task_ctx
);
1611 * Attach a performance event to a context
1613 * First we add the event to the list with the hardware enable bit
1614 * in event->hw_config cleared.
1616 * If the event is attached to a task which is on a CPU we use a smp
1617 * call to enable it in the task context. The task might have been
1618 * scheduled away, but we check this in the smp call again.
1621 perf_install_in_context(struct perf_event_context
*ctx
,
1622 struct perf_event
*event
,
1625 struct task_struct
*task
= ctx
->task
;
1627 lockdep_assert_held(&ctx
->mutex
);
1633 * Per cpu events are installed via an smp call and
1634 * the install is always successful.
1636 cpu_function_call(cpu
, __perf_install_in_context
, event
);
1641 if (!task_function_call(task
, __perf_install_in_context
, event
))
1644 raw_spin_lock_irq(&ctx
->lock
);
1646 * If we failed to find a running task, but find the context active now
1647 * that we've acquired the ctx->lock, retry.
1649 if (ctx
->is_active
) {
1650 raw_spin_unlock_irq(&ctx
->lock
);
1655 * Since the task isn't running, its safe to add the event, us holding
1656 * the ctx->lock ensures the task won't get scheduled in.
1658 add_event_to_ctx(event
, ctx
);
1659 raw_spin_unlock_irq(&ctx
->lock
);
1663 * Put a event into inactive state and update time fields.
1664 * Enabling the leader of a group effectively enables all
1665 * the group members that aren't explicitly disabled, so we
1666 * have to update their ->tstamp_enabled also.
1667 * Note: this works for group members as well as group leaders
1668 * since the non-leader members' sibling_lists will be empty.
1670 static void __perf_event_mark_enabled(struct perf_event
*event
)
1672 struct perf_event
*sub
;
1673 u64 tstamp
= perf_event_time(event
);
1675 event
->state
= PERF_EVENT_STATE_INACTIVE
;
1676 event
->tstamp_enabled
= tstamp
- event
->total_time_enabled
;
1677 list_for_each_entry(sub
, &event
->sibling_list
, group_entry
) {
1678 if (sub
->state
>= PERF_EVENT_STATE_INACTIVE
)
1679 sub
->tstamp_enabled
= tstamp
- sub
->total_time_enabled
;
1684 * Cross CPU call to enable a performance event
1686 static int __perf_event_enable(void *info
)
1688 struct perf_event
*event
= info
;
1689 struct perf_event_context
*ctx
= event
->ctx
;
1690 struct perf_event
*leader
= event
->group_leader
;
1691 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
1694 if (WARN_ON_ONCE(!ctx
->is_active
))
1697 raw_spin_lock(&ctx
->lock
);
1698 update_context_time(ctx
);
1700 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
)
1704 * set current task's cgroup time reference point
1706 perf_cgroup_set_timestamp(current
, ctx
);
1708 __perf_event_mark_enabled(event
);
1710 if (!event_filter_match(event
)) {
1711 if (is_cgroup_event(event
))
1712 perf_cgroup_defer_enabled(event
);
1717 * If the event is in a group and isn't the group leader,
1718 * then don't put it on unless the group is on.
1720 if (leader
!= event
&& leader
->state
!= PERF_EVENT_STATE_ACTIVE
)
1723 if (!group_can_go_on(event
, cpuctx
, 1)) {
1726 if (event
== leader
)
1727 err
= group_sched_in(event
, cpuctx
, ctx
);
1729 err
= event_sched_in(event
, cpuctx
, ctx
);
1734 * If this event can't go on and it's part of a
1735 * group, then the whole group has to come off.
1737 if (leader
!= event
)
1738 group_sched_out(leader
, cpuctx
, ctx
);
1739 if (leader
->attr
.pinned
) {
1740 update_group_times(leader
);
1741 leader
->state
= PERF_EVENT_STATE_ERROR
;
1746 raw_spin_unlock(&ctx
->lock
);
1754 * If event->ctx is a cloned context, callers must make sure that
1755 * every task struct that event->ctx->task could possibly point to
1756 * remains valid. This condition is satisfied when called through
1757 * perf_event_for_each_child or perf_event_for_each as described
1758 * for perf_event_disable.
1760 void perf_event_enable(struct perf_event
*event
)
1762 struct perf_event_context
*ctx
= event
->ctx
;
1763 struct task_struct
*task
= ctx
->task
;
1767 * Enable the event on the cpu that it's on
1769 cpu_function_call(event
->cpu
, __perf_event_enable
, event
);
1773 raw_spin_lock_irq(&ctx
->lock
);
1774 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
)
1778 * If the event is in error state, clear that first.
1779 * That way, if we see the event in error state below, we
1780 * know that it has gone back into error state, as distinct
1781 * from the task having been scheduled away before the
1782 * cross-call arrived.
1784 if (event
->state
== PERF_EVENT_STATE_ERROR
)
1785 event
->state
= PERF_EVENT_STATE_OFF
;
1788 if (!ctx
->is_active
) {
1789 __perf_event_mark_enabled(event
);
1793 raw_spin_unlock_irq(&ctx
->lock
);
1795 if (!task_function_call(task
, __perf_event_enable
, event
))
1798 raw_spin_lock_irq(&ctx
->lock
);
1801 * If the context is active and the event is still off,
1802 * we need to retry the cross-call.
1804 if (ctx
->is_active
&& event
->state
== PERF_EVENT_STATE_OFF
) {
1806 * task could have been flipped by a concurrent
1807 * perf_event_context_sched_out()
1814 raw_spin_unlock_irq(&ctx
->lock
);
1816 EXPORT_SYMBOL_GPL(perf_event_enable
);
1818 int perf_event_refresh(struct perf_event
*event
, int refresh
)
1821 * not supported on inherited events
1823 if (event
->attr
.inherit
|| !is_sampling_event(event
))
1826 atomic_add(refresh
, &event
->event_limit
);
1827 perf_event_enable(event
);
1831 EXPORT_SYMBOL_GPL(perf_event_refresh
);
1833 static void ctx_sched_out(struct perf_event_context
*ctx
,
1834 struct perf_cpu_context
*cpuctx
,
1835 enum event_type_t event_type
)
1837 struct perf_event
*event
;
1838 int is_active
= ctx
->is_active
;
1840 ctx
->is_active
&= ~event_type
;
1841 if (likely(!ctx
->nr_events
))
1844 update_context_time(ctx
);
1845 update_cgrp_time_from_cpuctx(cpuctx
);
1846 if (!ctx
->nr_active
)
1849 perf_pmu_disable(ctx
->pmu
);
1850 if ((is_active
& EVENT_PINNED
) && (event_type
& EVENT_PINNED
)) {
1851 list_for_each_entry(event
, &ctx
->pinned_groups
, group_entry
)
1852 group_sched_out(event
, cpuctx
, ctx
);
1855 if ((is_active
& EVENT_FLEXIBLE
) && (event_type
& EVENT_FLEXIBLE
)) {
1856 list_for_each_entry(event
, &ctx
->flexible_groups
, group_entry
)
1857 group_sched_out(event
, cpuctx
, ctx
);
1859 perf_pmu_enable(ctx
->pmu
);
1863 * Test whether two contexts are equivalent, i.e. whether they
1864 * have both been cloned from the same version of the same context
1865 * and they both have the same number of enabled events.
1866 * If the number of enabled events is the same, then the set
1867 * of enabled events should be the same, because these are both
1868 * inherited contexts, therefore we can't access individual events
1869 * in them directly with an fd; we can only enable/disable all
1870 * events via prctl, or enable/disable all events in a family
1871 * via ioctl, which will have the same effect on both contexts.
1873 static int context_equiv(struct perf_event_context
*ctx1
,
1874 struct perf_event_context
*ctx2
)
1876 return ctx1
->parent_ctx
&& ctx1
->parent_ctx
== ctx2
->parent_ctx
1877 && ctx1
->parent_gen
== ctx2
->parent_gen
1878 && !ctx1
->pin_count
&& !ctx2
->pin_count
;
1881 static void __perf_event_sync_stat(struct perf_event
*event
,
1882 struct perf_event
*next_event
)
1886 if (!event
->attr
.inherit_stat
)
1890 * Update the event value, we cannot use perf_event_read()
1891 * because we're in the middle of a context switch and have IRQs
1892 * disabled, which upsets smp_call_function_single(), however
1893 * we know the event must be on the current CPU, therefore we
1894 * don't need to use it.
1896 switch (event
->state
) {
1897 case PERF_EVENT_STATE_ACTIVE
:
1898 event
->pmu
->read(event
);
1901 case PERF_EVENT_STATE_INACTIVE
:
1902 update_event_times(event
);
1910 * In order to keep per-task stats reliable we need to flip the event
1911 * values when we flip the contexts.
1913 value
= local64_read(&next_event
->count
);
1914 value
= local64_xchg(&event
->count
, value
);
1915 local64_set(&next_event
->count
, value
);
1917 swap(event
->total_time_enabled
, next_event
->total_time_enabled
);
1918 swap(event
->total_time_running
, next_event
->total_time_running
);
1921 * Since we swizzled the values, update the user visible data too.
1923 perf_event_update_userpage(event
);
1924 perf_event_update_userpage(next_event
);
1927 #define list_next_entry(pos, member) \
1928 list_entry(pos->member.next, typeof(*pos), member)
1930 static void perf_event_sync_stat(struct perf_event_context
*ctx
,
1931 struct perf_event_context
*next_ctx
)
1933 struct perf_event
*event
, *next_event
;
1938 update_context_time(ctx
);
1940 event
= list_first_entry(&ctx
->event_list
,
1941 struct perf_event
, event_entry
);
1943 next_event
= list_first_entry(&next_ctx
->event_list
,
1944 struct perf_event
, event_entry
);
1946 while (&event
->event_entry
!= &ctx
->event_list
&&
1947 &next_event
->event_entry
!= &next_ctx
->event_list
) {
1949 __perf_event_sync_stat(event
, next_event
);
1951 event
= list_next_entry(event
, event_entry
);
1952 next_event
= list_next_entry(next_event
, event_entry
);
1956 static void perf_event_context_sched_out(struct task_struct
*task
, int ctxn
,
1957 struct task_struct
*next
)
1959 struct perf_event_context
*ctx
= task
->perf_event_ctxp
[ctxn
];
1960 struct perf_event_context
*next_ctx
;
1961 struct perf_event_context
*parent
;
1962 struct perf_cpu_context
*cpuctx
;
1968 cpuctx
= __get_cpu_context(ctx
);
1969 if (!cpuctx
->task_ctx
)
1973 parent
= rcu_dereference(ctx
->parent_ctx
);
1974 next_ctx
= next
->perf_event_ctxp
[ctxn
];
1975 if (parent
&& next_ctx
&&
1976 rcu_dereference(next_ctx
->parent_ctx
) == parent
) {
1978 * Looks like the two contexts are clones, so we might be
1979 * able to optimize the context switch. We lock both
1980 * contexts and check that they are clones under the
1981 * lock (including re-checking that neither has been
1982 * uncloned in the meantime). It doesn't matter which
1983 * order we take the locks because no other cpu could
1984 * be trying to lock both of these tasks.
1986 raw_spin_lock(&ctx
->lock
);
1987 raw_spin_lock_nested(&next_ctx
->lock
, SINGLE_DEPTH_NESTING
);
1988 if (context_equiv(ctx
, next_ctx
)) {
1990 * XXX do we need a memory barrier of sorts
1991 * wrt to rcu_dereference() of perf_event_ctxp
1993 task
->perf_event_ctxp
[ctxn
] = next_ctx
;
1994 next
->perf_event_ctxp
[ctxn
] = ctx
;
1996 next_ctx
->task
= task
;
1999 perf_event_sync_stat(ctx
, next_ctx
);
2001 raw_spin_unlock(&next_ctx
->lock
);
2002 raw_spin_unlock(&ctx
->lock
);
2007 raw_spin_lock(&ctx
->lock
);
2008 ctx_sched_out(ctx
, cpuctx
, EVENT_ALL
);
2009 cpuctx
->task_ctx
= NULL
;
2010 raw_spin_unlock(&ctx
->lock
);
2014 #define for_each_task_context_nr(ctxn) \
2015 for ((ctxn) = 0; (ctxn) < perf_nr_task_contexts; (ctxn)++)
2018 * Called from scheduler to remove the events of the current task,
2019 * with interrupts disabled.
2021 * We stop each event and update the event value in event->count.
2023 * This does not protect us against NMI, but disable()
2024 * sets the disabled bit in the control field of event _before_
2025 * accessing the event control register. If a NMI hits, then it will
2026 * not restart the event.
2028 void __perf_event_task_sched_out(struct task_struct
*task
,
2029 struct task_struct
*next
)
2033 for_each_task_context_nr(ctxn
)
2034 perf_event_context_sched_out(task
, ctxn
, next
);
2037 * if cgroup events exist on this CPU, then we need
2038 * to check if we have to switch out PMU state.
2039 * cgroup event are system-wide mode only
2041 if (atomic_read(&__get_cpu_var(perf_cgroup_events
)))
2042 perf_cgroup_sched_out(task
, next
);
2045 static void task_ctx_sched_out(struct perf_event_context
*ctx
)
2047 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
2049 if (!cpuctx
->task_ctx
)
2052 if (WARN_ON_ONCE(ctx
!= cpuctx
->task_ctx
))
2055 ctx_sched_out(ctx
, cpuctx
, EVENT_ALL
);
2056 cpuctx
->task_ctx
= NULL
;
2060 * Called with IRQs disabled
2062 static void cpu_ctx_sched_out(struct perf_cpu_context
*cpuctx
,
2063 enum event_type_t event_type
)
2065 ctx_sched_out(&cpuctx
->ctx
, cpuctx
, event_type
);
2069 ctx_pinned_sched_in(struct perf_event_context
*ctx
,
2070 struct perf_cpu_context
*cpuctx
)
2072 struct perf_event
*event
;
2074 list_for_each_entry(event
, &ctx
->pinned_groups
, group_entry
) {
2075 if (event
->state
<= PERF_EVENT_STATE_OFF
)
2077 if (!event_filter_match(event
))
2080 /* may need to reset tstamp_enabled */
2081 if (is_cgroup_event(event
))
2082 perf_cgroup_mark_enabled(event
, ctx
);
2084 if (group_can_go_on(event
, cpuctx
, 1))
2085 group_sched_in(event
, cpuctx
, ctx
);
2088 * If this pinned group hasn't been scheduled,
2089 * put it in error state.
2091 if (event
->state
== PERF_EVENT_STATE_INACTIVE
) {
2092 update_group_times(event
);
2093 event
->state
= PERF_EVENT_STATE_ERROR
;
2099 ctx_flexible_sched_in(struct perf_event_context
*ctx
,
2100 struct perf_cpu_context
*cpuctx
)
2102 struct perf_event
*event
;
2105 list_for_each_entry(event
, &ctx
->flexible_groups
, group_entry
) {
2106 /* Ignore events in OFF or ERROR state */
2107 if (event
->state
<= PERF_EVENT_STATE_OFF
)
2110 * Listen to the 'cpu' scheduling filter constraint
2113 if (!event_filter_match(event
))
2116 /* may need to reset tstamp_enabled */
2117 if (is_cgroup_event(event
))
2118 perf_cgroup_mark_enabled(event
, ctx
);
2120 if (group_can_go_on(event
, cpuctx
, can_add_hw
)) {
2121 if (group_sched_in(event
, cpuctx
, ctx
))
2128 ctx_sched_in(struct perf_event_context
*ctx
,
2129 struct perf_cpu_context
*cpuctx
,
2130 enum event_type_t event_type
,
2131 struct task_struct
*task
)
2134 int is_active
= ctx
->is_active
;
2136 ctx
->is_active
|= event_type
;
2137 if (likely(!ctx
->nr_events
))
2141 ctx
->timestamp
= now
;
2142 perf_cgroup_set_timestamp(task
, ctx
);
2144 * First go through the list and put on any pinned groups
2145 * in order to give them the best chance of going on.
2147 if (!(is_active
& EVENT_PINNED
) && (event_type
& EVENT_PINNED
))
2148 ctx_pinned_sched_in(ctx
, cpuctx
);
2150 /* Then walk through the lower prio flexible groups */
2151 if (!(is_active
& EVENT_FLEXIBLE
) && (event_type
& EVENT_FLEXIBLE
))
2152 ctx_flexible_sched_in(ctx
, cpuctx
);
2155 static void cpu_ctx_sched_in(struct perf_cpu_context
*cpuctx
,
2156 enum event_type_t event_type
,
2157 struct task_struct
*task
)
2159 struct perf_event_context
*ctx
= &cpuctx
->ctx
;
2161 ctx_sched_in(ctx
, cpuctx
, event_type
, task
);
2164 static void perf_event_context_sched_in(struct perf_event_context
*ctx
,
2165 struct task_struct
*task
)
2167 struct perf_cpu_context
*cpuctx
;
2169 cpuctx
= __get_cpu_context(ctx
);
2170 if (cpuctx
->task_ctx
== ctx
)
2173 perf_ctx_lock(cpuctx
, ctx
);
2174 perf_pmu_disable(ctx
->pmu
);
2176 * We want to keep the following priority order:
2177 * cpu pinned (that don't need to move), task pinned,
2178 * cpu flexible, task flexible.
2180 cpu_ctx_sched_out(cpuctx
, EVENT_FLEXIBLE
);
2183 cpuctx
->task_ctx
= ctx
;
2185 perf_event_sched_in(cpuctx
, cpuctx
->task_ctx
, task
);
2187 perf_pmu_enable(ctx
->pmu
);
2188 perf_ctx_unlock(cpuctx
, ctx
);
2191 * Since these rotations are per-cpu, we need to ensure the
2192 * cpu-context we got scheduled on is actually rotating.
2194 perf_pmu_rotate_start(ctx
->pmu
);
2198 * Called from scheduler to add the events of the current task
2199 * with interrupts disabled.
2201 * We restore the event value and then enable it.
2203 * This does not protect us against NMI, but enable()
2204 * sets the enabled bit in the control field of event _before_
2205 * accessing the event control register. If a NMI hits, then it will
2206 * keep the event running.
2208 void __perf_event_task_sched_in(struct task_struct
*prev
,
2209 struct task_struct
*task
)
2211 struct perf_event_context
*ctx
;
2214 for_each_task_context_nr(ctxn
) {
2215 ctx
= task
->perf_event_ctxp
[ctxn
];
2219 perf_event_context_sched_in(ctx
, task
);
2222 * if cgroup events exist on this CPU, then we need
2223 * to check if we have to switch in PMU state.
2224 * cgroup event are system-wide mode only
2226 if (atomic_read(&__get_cpu_var(perf_cgroup_events
)))
2227 perf_cgroup_sched_in(prev
, task
);
2230 static u64
perf_calculate_period(struct perf_event
*event
, u64 nsec
, u64 count
)
2232 u64 frequency
= event
->attr
.sample_freq
;
2233 u64 sec
= NSEC_PER_SEC
;
2234 u64 divisor
, dividend
;
2236 int count_fls
, nsec_fls
, frequency_fls
, sec_fls
;
2238 count_fls
= fls64(count
);
2239 nsec_fls
= fls64(nsec
);
2240 frequency_fls
= fls64(frequency
);
2244 * We got @count in @nsec, with a target of sample_freq HZ
2245 * the target period becomes:
2248 * period = -------------------
2249 * @nsec * sample_freq
2254 * Reduce accuracy by one bit such that @a and @b converge
2255 * to a similar magnitude.
2257 #define REDUCE_FLS(a, b) \
2259 if (a##_fls > b##_fls) { \
2269 * Reduce accuracy until either term fits in a u64, then proceed with
2270 * the other, so that finally we can do a u64/u64 division.
2272 while (count_fls
+ sec_fls
> 64 && nsec_fls
+ frequency_fls
> 64) {
2273 REDUCE_FLS(nsec
, frequency
);
2274 REDUCE_FLS(sec
, count
);
2277 if (count_fls
+ sec_fls
> 64) {
2278 divisor
= nsec
* frequency
;
2280 while (count_fls
+ sec_fls
> 64) {
2281 REDUCE_FLS(count
, sec
);
2285 dividend
= count
* sec
;
2287 dividend
= count
* sec
;
2289 while (nsec_fls
+ frequency_fls
> 64) {
2290 REDUCE_FLS(nsec
, frequency
);
2294 divisor
= nsec
* frequency
;
2300 return div64_u64(dividend
, divisor
);
2303 static DEFINE_PER_CPU(int, perf_throttled_count
);
2304 static DEFINE_PER_CPU(u64
, perf_throttled_seq
);
2306 static void perf_adjust_period(struct perf_event
*event
, u64 nsec
, u64 count
)
2308 struct hw_perf_event
*hwc
= &event
->hw
;
2309 s64 period
, sample_period
;
2312 period
= perf_calculate_period(event
, nsec
, count
);
2314 delta
= (s64
)(period
- hwc
->sample_period
);
2315 delta
= (delta
+ 7) / 8; /* low pass filter */
2317 sample_period
= hwc
->sample_period
+ delta
;
2322 hwc
->sample_period
= sample_period
;
2324 if (local64_read(&hwc
->period_left
) > 8*sample_period
) {
2325 event
->pmu
->stop(event
, PERF_EF_UPDATE
);
2326 local64_set(&hwc
->period_left
, 0);
2327 event
->pmu
->start(event
, PERF_EF_RELOAD
);
2332 * combine freq adjustment with unthrottling to avoid two passes over the
2333 * events. At the same time, make sure, having freq events does not change
2334 * the rate of unthrottling as that would introduce bias.
2336 static void perf_adjust_freq_unthr_context(struct perf_event_context
*ctx
,
2339 struct perf_event
*event
;
2340 struct hw_perf_event
*hwc
;
2341 u64 now
, period
= TICK_NSEC
;
2345 * only need to iterate over all events iff:
2346 * - context have events in frequency mode (needs freq adjust)
2347 * - there are events to unthrottle on this cpu
2349 if (!(ctx
->nr_freq
|| needs_unthr
))
2352 raw_spin_lock(&ctx
->lock
);
2354 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
2355 if (event
->state
!= PERF_EVENT_STATE_ACTIVE
)
2358 if (!event_filter_match(event
))
2363 if (needs_unthr
&& hwc
->interrupts
== MAX_INTERRUPTS
) {
2364 hwc
->interrupts
= 0;
2365 perf_log_throttle(event
, 1);
2366 event
->pmu
->start(event
, 0);
2369 if (!event
->attr
.freq
|| !event
->attr
.sample_freq
)
2373 * stop the event and update event->count
2375 event
->pmu
->stop(event
, PERF_EF_UPDATE
);
2377 now
= local64_read(&event
->count
);
2378 delta
= now
- hwc
->freq_count_stamp
;
2379 hwc
->freq_count_stamp
= now
;
2383 * reload only if value has changed
2386 perf_adjust_period(event
, period
, delta
);
2388 event
->pmu
->start(event
, delta
> 0 ? PERF_EF_RELOAD
: 0);
2391 raw_spin_unlock(&ctx
->lock
);
2395 * Round-robin a context's events:
2397 static void rotate_ctx(struct perf_event_context
*ctx
)
2400 * Rotate the first entry last of non-pinned groups. Rotation might be
2401 * disabled by the inheritance code.
2403 if (!ctx
->rotate_disable
)
2404 list_rotate_left(&ctx
->flexible_groups
);
2408 * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
2409 * because they're strictly cpu affine and rotate_start is called with IRQs
2410 * disabled, while rotate_context is called from IRQ context.
2412 static void perf_rotate_context(struct perf_cpu_context
*cpuctx
)
2414 struct perf_event_context
*ctx
= NULL
;
2415 int rotate
= 0, remove
= 1;
2417 if (cpuctx
->ctx
.nr_events
) {
2419 if (cpuctx
->ctx
.nr_events
!= cpuctx
->ctx
.nr_active
)
2423 ctx
= cpuctx
->task_ctx
;
2424 if (ctx
&& ctx
->nr_events
) {
2426 if (ctx
->nr_events
!= ctx
->nr_active
)
2433 perf_ctx_lock(cpuctx
, cpuctx
->task_ctx
);
2434 perf_pmu_disable(cpuctx
->ctx
.pmu
);
2436 cpu_ctx_sched_out(cpuctx
, EVENT_FLEXIBLE
);
2438 ctx_sched_out(ctx
, cpuctx
, EVENT_FLEXIBLE
);
2440 rotate_ctx(&cpuctx
->ctx
);
2444 perf_event_sched_in(cpuctx
, ctx
, current
);
2446 perf_pmu_enable(cpuctx
->ctx
.pmu
);
2447 perf_ctx_unlock(cpuctx
, cpuctx
->task_ctx
);
2450 list_del_init(&cpuctx
->rotation_list
);
2453 void perf_event_task_tick(void)
2455 struct list_head
*head
= &__get_cpu_var(rotation_list
);
2456 struct perf_cpu_context
*cpuctx
, *tmp
;
2457 struct perf_event_context
*ctx
;
2460 WARN_ON(!irqs_disabled());
2462 __this_cpu_inc(perf_throttled_seq
);
2463 throttled
= __this_cpu_xchg(perf_throttled_count
, 0);
2465 list_for_each_entry_safe(cpuctx
, tmp
, head
, rotation_list
) {
2467 perf_adjust_freq_unthr_context(ctx
, throttled
);
2469 ctx
= cpuctx
->task_ctx
;
2471 perf_adjust_freq_unthr_context(ctx
, throttled
);
2473 if (cpuctx
->jiffies_interval
== 1 ||
2474 !(jiffies
% cpuctx
->jiffies_interval
))
2475 perf_rotate_context(cpuctx
);
2479 static int event_enable_on_exec(struct perf_event
*event
,
2480 struct perf_event_context
*ctx
)
2482 if (!event
->attr
.enable_on_exec
)
2485 event
->attr
.enable_on_exec
= 0;
2486 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
)
2489 __perf_event_mark_enabled(event
);
2495 * Enable all of a task's events that have been marked enable-on-exec.
2496 * This expects task == current.
2498 static void perf_event_enable_on_exec(struct perf_event_context
*ctx
)
2500 struct perf_event
*event
;
2501 unsigned long flags
;
2505 local_irq_save(flags
);
2506 if (!ctx
|| !ctx
->nr_events
)
2510 * We must ctxsw out cgroup events to avoid conflict
2511 * when invoking perf_task_event_sched_in() later on
2512 * in this function. Otherwise we end up trying to
2513 * ctxswin cgroup events which are already scheduled
2516 perf_cgroup_sched_out(current
, NULL
);
2518 raw_spin_lock(&ctx
->lock
);
2519 task_ctx_sched_out(ctx
);
2521 list_for_each_entry(event
, &ctx
->event_list
, event_entry
) {
2522 ret
= event_enable_on_exec(event
, ctx
);
2528 * Unclone this context if we enabled any event.
2533 raw_spin_unlock(&ctx
->lock
);
2536 * Also calls ctxswin for cgroup events, if any:
2538 perf_event_context_sched_in(ctx
, ctx
->task
);
2540 local_irq_restore(flags
);
2544 * Cross CPU call to read the hardware event
2546 static void __perf_event_read(void *info
)
2548 struct perf_event
*event
= info
;
2549 struct perf_event_context
*ctx
= event
->ctx
;
2550 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
2553 * If this is a task context, we need to check whether it is
2554 * the current task context of this cpu. If not it has been
2555 * scheduled out before the smp call arrived. In that case
2556 * event->count would have been updated to a recent sample
2557 * when the event was scheduled out.
2559 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
)
2562 raw_spin_lock(&ctx
->lock
);
2563 if (ctx
->is_active
) {
2564 update_context_time(ctx
);
2565 update_cgrp_time_from_event(event
);
2567 update_event_times(event
);
2568 if (event
->state
== PERF_EVENT_STATE_ACTIVE
)
2569 event
->pmu
->read(event
);
2570 raw_spin_unlock(&ctx
->lock
);
2573 static inline u64
perf_event_count(struct perf_event
*event
)
2575 return local64_read(&event
->count
) + atomic64_read(&event
->child_count
);
2578 static u64
perf_event_read(struct perf_event
*event
)
2581 * If event is enabled and currently active on a CPU, update the
2582 * value in the event structure:
2584 if (event
->state
== PERF_EVENT_STATE_ACTIVE
) {
2585 smp_call_function_single(event
->oncpu
,
2586 __perf_event_read
, event
, 1);
2587 } else if (event
->state
== PERF_EVENT_STATE_INACTIVE
) {
2588 struct perf_event_context
*ctx
= event
->ctx
;
2589 unsigned long flags
;
2591 raw_spin_lock_irqsave(&ctx
->lock
, flags
);
2593 * may read while context is not active
2594 * (e.g., thread is blocked), in that case
2595 * we cannot update context time
2597 if (ctx
->is_active
) {
2598 update_context_time(ctx
);
2599 update_cgrp_time_from_event(event
);
2601 update_event_times(event
);
2602 raw_spin_unlock_irqrestore(&ctx
->lock
, flags
);
2605 return perf_event_count(event
);
2609 * Initialize the perf_event context in a task_struct:
2611 static void __perf_event_init_context(struct perf_event_context
*ctx
)
2613 raw_spin_lock_init(&ctx
->lock
);
2614 mutex_init(&ctx
->mutex
);
2615 INIT_LIST_HEAD(&ctx
->pinned_groups
);
2616 INIT_LIST_HEAD(&ctx
->flexible_groups
);
2617 INIT_LIST_HEAD(&ctx
->event_list
);
2618 atomic_set(&ctx
->refcount
, 1);
2621 static struct perf_event_context
*
2622 alloc_perf_context(struct pmu
*pmu
, struct task_struct
*task
)
2624 struct perf_event_context
*ctx
;
2626 ctx
= kzalloc(sizeof(struct perf_event_context
), GFP_KERNEL
);
2630 __perf_event_init_context(ctx
);
2633 get_task_struct(task
);
2640 static struct task_struct
*
2641 find_lively_task_by_vpid(pid_t vpid
)
2643 struct task_struct
*task
;
2650 task
= find_task_by_vpid(vpid
);
2652 get_task_struct(task
);
2656 return ERR_PTR(-ESRCH
);
2658 /* Reuse ptrace permission checks for now. */
2660 if (!ptrace_may_access(task
, PTRACE_MODE_READ
))
2665 put_task_struct(task
);
2666 return ERR_PTR(err
);
2671 * Returns a matching context with refcount and pincount.
2673 static struct perf_event_context
*
2674 find_get_context(struct pmu
*pmu
, struct task_struct
*task
, int cpu
)
2676 struct perf_event_context
*ctx
;
2677 struct perf_cpu_context
*cpuctx
;
2678 unsigned long flags
;
2682 /* Must be root to operate on a CPU event: */
2683 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN
))
2684 return ERR_PTR(-EACCES
);
2687 * We could be clever and allow to attach a event to an
2688 * offline CPU and activate it when the CPU comes up, but
2691 if (!cpu_online(cpu
))
2692 return ERR_PTR(-ENODEV
);
2694 cpuctx
= per_cpu_ptr(pmu
->pmu_cpu_context
, cpu
);
2703 ctxn
= pmu
->task_ctx_nr
;
2708 ctx
= perf_lock_task_context(task
, ctxn
, &flags
);
2712 raw_spin_unlock_irqrestore(&ctx
->lock
, flags
);
2714 ctx
= alloc_perf_context(pmu
, task
);
2720 mutex_lock(&task
->perf_event_mutex
);
2722 * If it has already passed perf_event_exit_task().
2723 * we must see PF_EXITING, it takes this mutex too.
2725 if (task
->flags
& PF_EXITING
)
2727 else if (task
->perf_event_ctxp
[ctxn
])
2732 rcu_assign_pointer(task
->perf_event_ctxp
[ctxn
], ctx
);
2734 mutex_unlock(&task
->perf_event_mutex
);
2736 if (unlikely(err
)) {
2748 return ERR_PTR(err
);
2751 static void perf_event_free_filter(struct perf_event
*event
);
2753 static void free_event_rcu(struct rcu_head
*head
)
2755 struct perf_event
*event
;
2757 event
= container_of(head
, struct perf_event
, rcu_head
);
2759 put_pid_ns(event
->ns
);
2760 perf_event_free_filter(event
);
2764 static void ring_buffer_put(struct ring_buffer
*rb
);
2766 static void free_event(struct perf_event
*event
)
2768 irq_work_sync(&event
->pending
);
2770 if (!event
->parent
) {
2771 if (event
->attach_state
& PERF_ATTACH_TASK
)
2772 static_key_slow_dec_deferred(&perf_sched_events
);
2773 if (event
->attr
.mmap
|| event
->attr
.mmap_data
)
2774 atomic_dec(&nr_mmap_events
);
2775 if (event
->attr
.comm
)
2776 atomic_dec(&nr_comm_events
);
2777 if (event
->attr
.task
)
2778 atomic_dec(&nr_task_events
);
2779 if (event
->attr
.sample_type
& PERF_SAMPLE_CALLCHAIN
)
2780 put_callchain_buffers();
2781 if (is_cgroup_event(event
)) {
2782 atomic_dec(&per_cpu(perf_cgroup_events
, event
->cpu
));
2783 static_key_slow_dec_deferred(&perf_sched_events
);
2788 ring_buffer_put(event
->rb
);
2792 if (is_cgroup_event(event
))
2793 perf_detach_cgroup(event
);
2796 event
->destroy(event
);
2799 put_ctx(event
->ctx
);
2801 call_rcu(&event
->rcu_head
, free_event_rcu
);
2804 int perf_event_release_kernel(struct perf_event
*event
)
2806 struct perf_event_context
*ctx
= event
->ctx
;
2808 WARN_ON_ONCE(ctx
->parent_ctx
);
2810 * There are two ways this annotation is useful:
2812 * 1) there is a lock recursion from perf_event_exit_task
2813 * see the comment there.
2815 * 2) there is a lock-inversion with mmap_sem through
2816 * perf_event_read_group(), which takes faults while
2817 * holding ctx->mutex, however this is called after
2818 * the last filedesc died, so there is no possibility
2819 * to trigger the AB-BA case.
2821 mutex_lock_nested(&ctx
->mutex
, SINGLE_DEPTH_NESTING
);
2822 raw_spin_lock_irq(&ctx
->lock
);
2823 perf_group_detach(event
);
2824 raw_spin_unlock_irq(&ctx
->lock
);
2825 perf_remove_from_context(event
);
2826 mutex_unlock(&ctx
->mutex
);
2832 EXPORT_SYMBOL_GPL(perf_event_release_kernel
);
2835 * Called when the last reference to the file is gone.
2837 static int perf_release(struct inode
*inode
, struct file
*file
)
2839 struct perf_event
*event
= file
->private_data
;
2840 struct task_struct
*owner
;
2842 file
->private_data
= NULL
;
2845 owner
= ACCESS_ONCE(event
->owner
);
2847 * Matches the smp_wmb() in perf_event_exit_task(). If we observe
2848 * !owner it means the list deletion is complete and we can indeed
2849 * free this event, otherwise we need to serialize on
2850 * owner->perf_event_mutex.
2852 smp_read_barrier_depends();
2855 * Since delayed_put_task_struct() also drops the last
2856 * task reference we can safely take a new reference
2857 * while holding the rcu_read_lock().
2859 get_task_struct(owner
);
2864 mutex_lock(&owner
->perf_event_mutex
);
2866 * We have to re-check the event->owner field, if it is cleared
2867 * we raced with perf_event_exit_task(), acquiring the mutex
2868 * ensured they're done, and we can proceed with freeing the
2872 list_del_init(&event
->owner_entry
);
2873 mutex_unlock(&owner
->perf_event_mutex
);
2874 put_task_struct(owner
);
2877 return perf_event_release_kernel(event
);
2880 u64
perf_event_read_value(struct perf_event
*event
, u64
*enabled
, u64
*running
)
2882 struct perf_event
*child
;
2888 mutex_lock(&event
->child_mutex
);
2889 total
+= perf_event_read(event
);
2890 *enabled
+= event
->total_time_enabled
+
2891 atomic64_read(&event
->child_total_time_enabled
);
2892 *running
+= event
->total_time_running
+
2893 atomic64_read(&event
->child_total_time_running
);
2895 list_for_each_entry(child
, &event
->child_list
, child_list
) {
2896 total
+= perf_event_read(child
);
2897 *enabled
+= child
->total_time_enabled
;
2898 *running
+= child
->total_time_running
;
2900 mutex_unlock(&event
->child_mutex
);
2904 EXPORT_SYMBOL_GPL(perf_event_read_value
);
2906 static int perf_event_read_group(struct perf_event
*event
,
2907 u64 read_format
, char __user
*buf
)
2909 struct perf_event
*leader
= event
->group_leader
, *sub
;
2910 int n
= 0, size
= 0, ret
= -EFAULT
;
2911 struct perf_event_context
*ctx
= leader
->ctx
;
2913 u64 count
, enabled
, running
;
2915 mutex_lock(&ctx
->mutex
);
2916 count
= perf_event_read_value(leader
, &enabled
, &running
);
2918 values
[n
++] = 1 + leader
->nr_siblings
;
2919 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
2920 values
[n
++] = enabled
;
2921 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
2922 values
[n
++] = running
;
2923 values
[n
++] = count
;
2924 if (read_format
& PERF_FORMAT_ID
)
2925 values
[n
++] = primary_event_id(leader
);
2927 size
= n
* sizeof(u64
);
2929 if (copy_to_user(buf
, values
, size
))
2934 list_for_each_entry(sub
, &leader
->sibling_list
, group_entry
) {
2937 values
[n
++] = perf_event_read_value(sub
, &enabled
, &running
);
2938 if (read_format
& PERF_FORMAT_ID
)
2939 values
[n
++] = primary_event_id(sub
);
2941 size
= n
* sizeof(u64
);
2943 if (copy_to_user(buf
+ ret
, values
, size
)) {
2951 mutex_unlock(&ctx
->mutex
);
2956 static int perf_event_read_one(struct perf_event
*event
,
2957 u64 read_format
, char __user
*buf
)
2959 u64 enabled
, running
;
2963 values
[n
++] = perf_event_read_value(event
, &enabled
, &running
);
2964 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
2965 values
[n
++] = enabled
;
2966 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
2967 values
[n
++] = running
;
2968 if (read_format
& PERF_FORMAT_ID
)
2969 values
[n
++] = primary_event_id(event
);
2971 if (copy_to_user(buf
, values
, n
* sizeof(u64
)))
2974 return n
* sizeof(u64
);
2978 * Read the performance event - simple non blocking version for now
2981 perf_read_hw(struct perf_event
*event
, char __user
*buf
, size_t count
)
2983 u64 read_format
= event
->attr
.read_format
;
2987 * Return end-of-file for a read on a event that is in
2988 * error state (i.e. because it was pinned but it couldn't be
2989 * scheduled on to the CPU at some point).
2991 if (event
->state
== PERF_EVENT_STATE_ERROR
)
2994 if (count
< event
->read_size
)
2997 WARN_ON_ONCE(event
->ctx
->parent_ctx
);
2998 if (read_format
& PERF_FORMAT_GROUP
)
2999 ret
= perf_event_read_group(event
, read_format
, buf
);
3001 ret
= perf_event_read_one(event
, read_format
, buf
);
3007 perf_read(struct file
*file
, char __user
*buf
, size_t count
, loff_t
*ppos
)
3009 struct perf_event
*event
= file
->private_data
;
3011 return perf_read_hw(event
, buf
, count
);
3014 static unsigned int perf_poll(struct file
*file
, poll_table
*wait
)
3016 struct perf_event
*event
= file
->private_data
;
3017 struct ring_buffer
*rb
;
3018 unsigned int events
= POLL_HUP
;
3021 * Race between perf_event_set_output() and perf_poll(): perf_poll()
3022 * grabs the rb reference but perf_event_set_output() overrides it.
3023 * Here is the timeline for two threads T1, T2:
3024 * t0: T1, rb = rcu_dereference(event->rb)
3025 * t1: T2, old_rb = event->rb
3026 * t2: T2, event->rb = new rb
3027 * t3: T2, ring_buffer_detach(old_rb)
3028 * t4: T1, ring_buffer_attach(rb1)
3029 * t5: T1, poll_wait(event->waitq)
3031 * To avoid this problem, we grab mmap_mutex in perf_poll()
3032 * thereby ensuring that the assignment of the new ring buffer
3033 * and the detachment of the old buffer appear atomic to perf_poll()
3035 mutex_lock(&event
->mmap_mutex
);
3038 rb
= rcu_dereference(event
->rb
);
3040 ring_buffer_attach(event
, rb
);
3041 events
= atomic_xchg(&rb
->poll
, 0);
3045 mutex_unlock(&event
->mmap_mutex
);
3047 poll_wait(file
, &event
->waitq
, wait
);
3052 static void perf_event_reset(struct perf_event
*event
)
3054 (void)perf_event_read(event
);
3055 local64_set(&event
->count
, 0);
3056 perf_event_update_userpage(event
);
3060 * Holding the top-level event's child_mutex means that any
3061 * descendant process that has inherited this event will block
3062 * in sync_child_event if it goes to exit, thus satisfying the
3063 * task existence requirements of perf_event_enable/disable.
3065 static void perf_event_for_each_child(struct perf_event
*event
,
3066 void (*func
)(struct perf_event
*))
3068 struct perf_event
*child
;
3070 WARN_ON_ONCE(event
->ctx
->parent_ctx
);
3071 mutex_lock(&event
->child_mutex
);
3073 list_for_each_entry(child
, &event
->child_list
, child_list
)
3075 mutex_unlock(&event
->child_mutex
);
3078 static void perf_event_for_each(struct perf_event
*event
,
3079 void (*func
)(struct perf_event
*))
3081 struct perf_event_context
*ctx
= event
->ctx
;
3082 struct perf_event
*sibling
;
3084 WARN_ON_ONCE(ctx
->parent_ctx
);
3085 mutex_lock(&ctx
->mutex
);
3086 event
= event
->group_leader
;
3088 perf_event_for_each_child(event
, func
);
3090 list_for_each_entry(sibling
, &event
->sibling_list
, group_entry
)
3091 perf_event_for_each_child(event
, func
);
3092 mutex_unlock(&ctx
->mutex
);
3095 static int perf_event_period(struct perf_event
*event
, u64 __user
*arg
)
3097 struct perf_event_context
*ctx
= event
->ctx
;
3101 if (!is_sampling_event(event
))
3104 if (copy_from_user(&value
, arg
, sizeof(value
)))
3110 raw_spin_lock_irq(&ctx
->lock
);
3111 if (event
->attr
.freq
) {
3112 if (value
> sysctl_perf_event_sample_rate
) {
3117 event
->attr
.sample_freq
= value
;
3119 event
->attr
.sample_period
= value
;
3120 event
->hw
.sample_period
= value
;
3123 raw_spin_unlock_irq(&ctx
->lock
);
3128 static const struct file_operations perf_fops
;
3130 static struct perf_event
*perf_fget_light(int fd
, int *fput_needed
)
3134 file
= fget_light(fd
, fput_needed
);
3136 return ERR_PTR(-EBADF
);
3138 if (file
->f_op
!= &perf_fops
) {
3139 fput_light(file
, *fput_needed
);
3141 return ERR_PTR(-EBADF
);
3144 return file
->private_data
;
3147 static int perf_event_set_output(struct perf_event
*event
,
3148 struct perf_event
*output_event
);
3149 static int perf_event_set_filter(struct perf_event
*event
, void __user
*arg
);
3151 static long perf_ioctl(struct file
*file
, unsigned int cmd
, unsigned long arg
)
3153 struct perf_event
*event
= file
->private_data
;
3154 void (*func
)(struct perf_event
*);
3158 case PERF_EVENT_IOC_ENABLE
:
3159 func
= perf_event_enable
;
3161 case PERF_EVENT_IOC_DISABLE
:
3162 func
= perf_event_disable
;
3164 case PERF_EVENT_IOC_RESET
:
3165 func
= perf_event_reset
;
3168 case PERF_EVENT_IOC_REFRESH
:
3169 return perf_event_refresh(event
, arg
);
3171 case PERF_EVENT_IOC_PERIOD
:
3172 return perf_event_period(event
, (u64 __user
*)arg
);
3174 case PERF_EVENT_IOC_SET_OUTPUT
:
3176 struct perf_event
*output_event
= NULL
;
3177 int fput_needed
= 0;
3181 output_event
= perf_fget_light(arg
, &fput_needed
);
3182 if (IS_ERR(output_event
))
3183 return PTR_ERR(output_event
);
3186 ret
= perf_event_set_output(event
, output_event
);
3188 fput_light(output_event
->filp
, fput_needed
);
3193 case PERF_EVENT_IOC_SET_FILTER
:
3194 return perf_event_set_filter(event
, (void __user
*)arg
);
3200 if (flags
& PERF_IOC_FLAG_GROUP
)
3201 perf_event_for_each(event
, func
);
3203 perf_event_for_each_child(event
, func
);
3208 int perf_event_task_enable(void)
3210 struct perf_event
*event
;
3212 mutex_lock(¤t
->perf_event_mutex
);
3213 list_for_each_entry(event
, ¤t
->perf_event_list
, owner_entry
)
3214 perf_event_for_each_child(event
, perf_event_enable
);
3215 mutex_unlock(¤t
->perf_event_mutex
);
3220 int perf_event_task_disable(void)
3222 struct perf_event
*event
;
3224 mutex_lock(¤t
->perf_event_mutex
);
3225 list_for_each_entry(event
, ¤t
->perf_event_list
, owner_entry
)
3226 perf_event_for_each_child(event
, perf_event_disable
);
3227 mutex_unlock(¤t
->perf_event_mutex
);
3232 static int perf_event_index(struct perf_event
*event
)
3234 if (event
->hw
.state
& PERF_HES_STOPPED
)
3237 if (event
->state
!= PERF_EVENT_STATE_ACTIVE
)
3240 return event
->pmu
->event_idx(event
);
3243 static void calc_timer_values(struct perf_event
*event
,
3250 *now
= perf_clock();
3251 ctx_time
= event
->shadow_ctx_time
+ *now
;
3252 *enabled
= ctx_time
- event
->tstamp_enabled
;
3253 *running
= ctx_time
- event
->tstamp_running
;
3256 void __weak
perf_update_user_clock(struct perf_event_mmap_page
*userpg
, u64 now
)
3261 * Callers need to ensure there can be no nesting of this function, otherwise
3262 * the seqlock logic goes bad. We can not serialize this because the arch
3263 * code calls this from NMI context.
3265 void perf_event_update_userpage(struct perf_event
*event
)
3267 struct perf_event_mmap_page
*userpg
;
3268 struct ring_buffer
*rb
;
3269 u64 enabled
, running
, now
;
3273 * compute total_time_enabled, total_time_running
3274 * based on snapshot values taken when the event
3275 * was last scheduled in.
3277 * we cannot simply called update_context_time()
3278 * because of locking issue as we can be called in
3281 calc_timer_values(event
, &now
, &enabled
, &running
);
3282 rb
= rcu_dereference(event
->rb
);
3286 userpg
= rb
->user_page
;
3289 * Disable preemption so as to not let the corresponding user-space
3290 * spin too long if we get preempted.
3295 userpg
->index
= perf_event_index(event
);
3296 userpg
->offset
= perf_event_count(event
);
3298 userpg
->offset
-= local64_read(&event
->hw
.prev_count
);
3300 userpg
->time_enabled
= enabled
+
3301 atomic64_read(&event
->child_total_time_enabled
);
3303 userpg
->time_running
= running
+
3304 atomic64_read(&event
->child_total_time_running
);
3306 perf_update_user_clock(userpg
, now
);
3315 static int perf_mmap_fault(struct vm_area_struct
*vma
, struct vm_fault
*vmf
)
3317 struct perf_event
*event
= vma
->vm_file
->private_data
;
3318 struct ring_buffer
*rb
;
3319 int ret
= VM_FAULT_SIGBUS
;
3321 if (vmf
->flags
& FAULT_FLAG_MKWRITE
) {
3322 if (vmf
->pgoff
== 0)
3328 rb
= rcu_dereference(event
->rb
);
3332 if (vmf
->pgoff
&& (vmf
->flags
& FAULT_FLAG_WRITE
))
3335 vmf
->page
= perf_mmap_to_page(rb
, vmf
->pgoff
);
3339 get_page(vmf
->page
);
3340 vmf
->page
->mapping
= vma
->vm_file
->f_mapping
;
3341 vmf
->page
->index
= vmf
->pgoff
;
3350 static void ring_buffer_attach(struct perf_event
*event
,
3351 struct ring_buffer
*rb
)
3353 unsigned long flags
;
3355 if (!list_empty(&event
->rb_entry
))
3358 spin_lock_irqsave(&rb
->event_lock
, flags
);
3359 if (!list_empty(&event
->rb_entry
))
3362 list_add(&event
->rb_entry
, &rb
->event_list
);
3364 spin_unlock_irqrestore(&rb
->event_lock
, flags
);
3367 static void ring_buffer_detach(struct perf_event
*event
,
3368 struct ring_buffer
*rb
)
3370 unsigned long flags
;
3372 if (list_empty(&event
->rb_entry
))
3375 spin_lock_irqsave(&rb
->event_lock
, flags
);
3376 list_del_init(&event
->rb_entry
);
3377 wake_up_all(&event
->waitq
);
3378 spin_unlock_irqrestore(&rb
->event_lock
, flags
);
3381 static void ring_buffer_wakeup(struct perf_event
*event
)
3383 struct ring_buffer
*rb
;
3386 rb
= rcu_dereference(event
->rb
);
3390 list_for_each_entry_rcu(event
, &rb
->event_list
, rb_entry
)
3391 wake_up_all(&event
->waitq
);
3397 static void rb_free_rcu(struct rcu_head
*rcu_head
)
3399 struct ring_buffer
*rb
;
3401 rb
= container_of(rcu_head
, struct ring_buffer
, rcu_head
);
3405 static struct ring_buffer
*ring_buffer_get(struct perf_event
*event
)
3407 struct ring_buffer
*rb
;
3410 rb
= rcu_dereference(event
->rb
);
3412 if (!atomic_inc_not_zero(&rb
->refcount
))
3420 static void ring_buffer_put(struct ring_buffer
*rb
)
3422 struct perf_event
*event
, *n
;
3423 unsigned long flags
;
3425 if (!atomic_dec_and_test(&rb
->refcount
))
3428 spin_lock_irqsave(&rb
->event_lock
, flags
);
3429 list_for_each_entry_safe(event
, n
, &rb
->event_list
, rb_entry
) {
3430 list_del_init(&event
->rb_entry
);
3431 wake_up_all(&event
->waitq
);
3433 spin_unlock_irqrestore(&rb
->event_lock
, flags
);
3435 call_rcu(&rb
->rcu_head
, rb_free_rcu
);
3438 static void perf_mmap_open(struct vm_area_struct
*vma
)
3440 struct perf_event
*event
= vma
->vm_file
->private_data
;
3442 atomic_inc(&event
->mmap_count
);
3445 static void perf_mmap_close(struct vm_area_struct
*vma
)
3447 struct perf_event
*event
= vma
->vm_file
->private_data
;
3449 if (atomic_dec_and_mutex_lock(&event
->mmap_count
, &event
->mmap_mutex
)) {
3450 unsigned long size
= perf_data_size(event
->rb
);
3451 struct user_struct
*user
= event
->mmap_user
;
3452 struct ring_buffer
*rb
= event
->rb
;
3454 atomic_long_sub((size
>> PAGE_SHIFT
) + 1, &user
->locked_vm
);
3455 vma
->vm_mm
->pinned_vm
-= event
->mmap_locked
;
3456 rcu_assign_pointer(event
->rb
, NULL
);
3457 ring_buffer_detach(event
, rb
);
3458 mutex_unlock(&event
->mmap_mutex
);
3460 ring_buffer_put(rb
);
3465 static const struct vm_operations_struct perf_mmap_vmops
= {
3466 .open
= perf_mmap_open
,
3467 .close
= perf_mmap_close
,
3468 .fault
= perf_mmap_fault
,
3469 .page_mkwrite
= perf_mmap_fault
,
3472 static int perf_mmap(struct file
*file
, struct vm_area_struct
*vma
)
3474 struct perf_event
*event
= file
->private_data
;
3475 unsigned long user_locked
, user_lock_limit
;
3476 struct user_struct
*user
= current_user();
3477 unsigned long locked
, lock_limit
;
3478 struct ring_buffer
*rb
;
3479 unsigned long vma_size
;
3480 unsigned long nr_pages
;
3481 long user_extra
, extra
;
3482 int ret
= 0, flags
= 0;
3485 * Don't allow mmap() of inherited per-task counters. This would
3486 * create a performance issue due to all children writing to the
3489 if (event
->cpu
== -1 && event
->attr
.inherit
)
3492 if (!(vma
->vm_flags
& VM_SHARED
))
3495 vma_size
= vma
->vm_end
- vma
->vm_start
;
3496 nr_pages
= (vma_size
/ PAGE_SIZE
) - 1;
3499 * If we have rb pages ensure they're a power-of-two number, so we
3500 * can do bitmasks instead of modulo.
3502 if (nr_pages
!= 0 && !is_power_of_2(nr_pages
))
3505 if (vma_size
!= PAGE_SIZE
* (1 + nr_pages
))
3508 if (vma
->vm_pgoff
!= 0)
3511 WARN_ON_ONCE(event
->ctx
->parent_ctx
);
3512 mutex_lock(&event
->mmap_mutex
);
3514 if (event
->rb
->nr_pages
== nr_pages
)
3515 atomic_inc(&event
->rb
->refcount
);
3521 user_extra
= nr_pages
+ 1;
3522 user_lock_limit
= sysctl_perf_event_mlock
>> (PAGE_SHIFT
- 10);
3525 * Increase the limit linearly with more CPUs:
3527 user_lock_limit
*= num_online_cpus();
3529 user_locked
= atomic_long_read(&user
->locked_vm
) + user_extra
;
3532 if (user_locked
> user_lock_limit
)
3533 extra
= user_locked
- user_lock_limit
;
3535 lock_limit
= rlimit(RLIMIT_MEMLOCK
);
3536 lock_limit
>>= PAGE_SHIFT
;
3537 locked
= vma
->vm_mm
->pinned_vm
+ extra
;
3539 if ((locked
> lock_limit
) && perf_paranoid_tracepoint_raw() &&
3540 !capable(CAP_IPC_LOCK
)) {
3547 if (vma
->vm_flags
& VM_WRITE
)
3548 flags
|= RING_BUFFER_WRITABLE
;
3550 rb
= rb_alloc(nr_pages
,
3551 event
->attr
.watermark
? event
->attr
.wakeup_watermark
: 0,
3558 rcu_assign_pointer(event
->rb
, rb
);
3560 atomic_long_add(user_extra
, &user
->locked_vm
);
3561 event
->mmap_locked
= extra
;
3562 event
->mmap_user
= get_current_user();
3563 vma
->vm_mm
->pinned_vm
+= event
->mmap_locked
;
3565 perf_event_update_userpage(event
);
3569 atomic_inc(&event
->mmap_count
);
3570 mutex_unlock(&event
->mmap_mutex
);
3572 vma
->vm_flags
|= VM_RESERVED
;
3573 vma
->vm_ops
= &perf_mmap_vmops
;
3578 static int perf_fasync(int fd
, struct file
*filp
, int on
)
3580 struct inode
*inode
= filp
->f_path
.dentry
->d_inode
;
3581 struct perf_event
*event
= filp
->private_data
;
3584 mutex_lock(&inode
->i_mutex
);
3585 retval
= fasync_helper(fd
, filp
, on
, &event
->fasync
);
3586 mutex_unlock(&inode
->i_mutex
);
3594 static const struct file_operations perf_fops
= {
3595 .llseek
= no_llseek
,
3596 .release
= perf_release
,
3599 .unlocked_ioctl
= perf_ioctl
,
3600 .compat_ioctl
= perf_ioctl
,
3602 .fasync
= perf_fasync
,
3608 * If there's data, ensure we set the poll() state and publish everything
3609 * to user-space before waking everybody up.
3612 void perf_event_wakeup(struct perf_event
*event
)
3614 ring_buffer_wakeup(event
);
3616 if (event
->pending_kill
) {
3617 kill_fasync(&event
->fasync
, SIGIO
, event
->pending_kill
);
3618 event
->pending_kill
= 0;
3622 static void perf_pending_event(struct irq_work
*entry
)
3624 struct perf_event
*event
= container_of(entry
,
3625 struct perf_event
, pending
);
3627 if (event
->pending_disable
) {
3628 event
->pending_disable
= 0;
3629 __perf_event_disable(event
);
3632 if (event
->pending_wakeup
) {
3633 event
->pending_wakeup
= 0;
3634 perf_event_wakeup(event
);
3639 * We assume there is only KVM supporting the callbacks.
3640 * Later on, we might change it to a list if there is
3641 * another virtualization implementation supporting the callbacks.
3643 struct perf_guest_info_callbacks
*perf_guest_cbs
;
3645 int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks
*cbs
)
3647 perf_guest_cbs
= cbs
;
3650 EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks
);
3652 int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks
*cbs
)
3654 perf_guest_cbs
= NULL
;
3657 EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks
);
3659 static void __perf_event_header__init_id(struct perf_event_header
*header
,
3660 struct perf_sample_data
*data
,
3661 struct perf_event
*event
)
3663 u64 sample_type
= event
->attr
.sample_type
;
3665 data
->type
= sample_type
;
3666 header
->size
+= event
->id_header_size
;
3668 if (sample_type
& PERF_SAMPLE_TID
) {
3669 /* namespace issues */
3670 data
->tid_entry
.pid
= perf_event_pid(event
, current
);
3671 data
->tid_entry
.tid
= perf_event_tid(event
, current
);
3674 if (sample_type
& PERF_SAMPLE_TIME
)
3675 data
->time
= perf_clock();
3677 if (sample_type
& PERF_SAMPLE_ID
)
3678 data
->id
= primary_event_id(event
);
3680 if (sample_type
& PERF_SAMPLE_STREAM_ID
)
3681 data
->stream_id
= event
->id
;
3683 if (sample_type
& PERF_SAMPLE_CPU
) {
3684 data
->cpu_entry
.cpu
= raw_smp_processor_id();
3685 data
->cpu_entry
.reserved
= 0;
3689 void perf_event_header__init_id(struct perf_event_header
*header
,
3690 struct perf_sample_data
*data
,
3691 struct perf_event
*event
)
3693 if (event
->attr
.sample_id_all
)
3694 __perf_event_header__init_id(header
, data
, event
);
3697 static void __perf_event__output_id_sample(struct perf_output_handle
*handle
,
3698 struct perf_sample_data
*data
)
3700 u64 sample_type
= data
->type
;
3702 if (sample_type
& PERF_SAMPLE_TID
)
3703 perf_output_put(handle
, data
->tid_entry
);
3705 if (sample_type
& PERF_SAMPLE_TIME
)
3706 perf_output_put(handle
, data
->time
);
3708 if (sample_type
& PERF_SAMPLE_ID
)
3709 perf_output_put(handle
, data
->id
);
3711 if (sample_type
& PERF_SAMPLE_STREAM_ID
)
3712 perf_output_put(handle
, data
->stream_id
);
3714 if (sample_type
& PERF_SAMPLE_CPU
)
3715 perf_output_put(handle
, data
->cpu_entry
);
3718 void perf_event__output_id_sample(struct perf_event
*event
,
3719 struct perf_output_handle
*handle
,
3720 struct perf_sample_data
*sample
)
3722 if (event
->attr
.sample_id_all
)
3723 __perf_event__output_id_sample(handle
, sample
);
3726 static void perf_output_read_one(struct perf_output_handle
*handle
,
3727 struct perf_event
*event
,
3728 u64 enabled
, u64 running
)
3730 u64 read_format
= event
->attr
.read_format
;
3734 values
[n
++] = perf_event_count(event
);
3735 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
) {
3736 values
[n
++] = enabled
+
3737 atomic64_read(&event
->child_total_time_enabled
);
3739 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
) {
3740 values
[n
++] = running
+
3741 atomic64_read(&event
->child_total_time_running
);
3743 if (read_format
& PERF_FORMAT_ID
)
3744 values
[n
++] = primary_event_id(event
);
3746 __output_copy(handle
, values
, n
* sizeof(u64
));
3750 * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
3752 static void perf_output_read_group(struct perf_output_handle
*handle
,
3753 struct perf_event
*event
,
3754 u64 enabled
, u64 running
)
3756 struct perf_event
*leader
= event
->group_leader
, *sub
;
3757 u64 read_format
= event
->attr
.read_format
;
3761 values
[n
++] = 1 + leader
->nr_siblings
;
3763 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
3764 values
[n
++] = enabled
;
3766 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
3767 values
[n
++] = running
;
3769 if (leader
!= event
)
3770 leader
->pmu
->read(leader
);
3772 values
[n
++] = perf_event_count(leader
);
3773 if (read_format
& PERF_FORMAT_ID
)
3774 values
[n
++] = primary_event_id(leader
);
3776 __output_copy(handle
, values
, n
* sizeof(u64
));
3778 list_for_each_entry(sub
, &leader
->sibling_list
, group_entry
) {
3782 sub
->pmu
->read(sub
);
3784 values
[n
++] = perf_event_count(sub
);
3785 if (read_format
& PERF_FORMAT_ID
)
3786 values
[n
++] = primary_event_id(sub
);
3788 __output_copy(handle
, values
, n
* sizeof(u64
));
3792 #define PERF_FORMAT_TOTAL_TIMES (PERF_FORMAT_TOTAL_TIME_ENABLED|\
3793 PERF_FORMAT_TOTAL_TIME_RUNNING)
3795 static void perf_output_read(struct perf_output_handle
*handle
,
3796 struct perf_event
*event
)
3798 u64 enabled
= 0, running
= 0, now
;
3799 u64 read_format
= event
->attr
.read_format
;
3802 * compute total_time_enabled, total_time_running
3803 * based on snapshot values taken when the event
3804 * was last scheduled in.
3806 * we cannot simply called update_context_time()
3807 * because of locking issue as we are called in
3810 if (read_format
& PERF_FORMAT_TOTAL_TIMES
)
3811 calc_timer_values(event
, &now
, &enabled
, &running
);
3813 if (event
->attr
.read_format
& PERF_FORMAT_GROUP
)
3814 perf_output_read_group(handle
, event
, enabled
, running
);
3816 perf_output_read_one(handle
, event
, enabled
, running
);
3819 void perf_output_sample(struct perf_output_handle
*handle
,
3820 struct perf_event_header
*header
,
3821 struct perf_sample_data
*data
,
3822 struct perf_event
*event
)
3824 u64 sample_type
= data
->type
;
3826 perf_output_put(handle
, *header
);
3828 if (sample_type
& PERF_SAMPLE_IP
)
3829 perf_output_put(handle
, data
->ip
);
3831 if (sample_type
& PERF_SAMPLE_TID
)
3832 perf_output_put(handle
, data
->tid_entry
);
3834 if (sample_type
& PERF_SAMPLE_TIME
)
3835 perf_output_put(handle
, data
->time
);
3837 if (sample_type
& PERF_SAMPLE_ADDR
)
3838 perf_output_put(handle
, data
->addr
);
3840 if (sample_type
& PERF_SAMPLE_ID
)
3841 perf_output_put(handle
, data
->id
);
3843 if (sample_type
& PERF_SAMPLE_STREAM_ID
)
3844 perf_output_put(handle
, data
->stream_id
);
3846 if (sample_type
& PERF_SAMPLE_CPU
)
3847 perf_output_put(handle
, data
->cpu_entry
);
3849 if (sample_type
& PERF_SAMPLE_PERIOD
)
3850 perf_output_put(handle
, data
->period
);
3852 if (sample_type
& PERF_SAMPLE_READ
)
3853 perf_output_read(handle
, event
);
3855 if (sample_type
& PERF_SAMPLE_CALLCHAIN
) {
3856 if (data
->callchain
) {
3859 if (data
->callchain
)
3860 size
+= data
->callchain
->nr
;
3862 size
*= sizeof(u64
);
3864 __output_copy(handle
, data
->callchain
, size
);
3867 perf_output_put(handle
, nr
);
3871 if (sample_type
& PERF_SAMPLE_RAW
) {
3873 perf_output_put(handle
, data
->raw
->size
);
3874 __output_copy(handle
, data
->raw
->data
,
3881 .size
= sizeof(u32
),
3884 perf_output_put(handle
, raw
);
3888 if (!event
->attr
.watermark
) {
3889 int wakeup_events
= event
->attr
.wakeup_events
;
3891 if (wakeup_events
) {
3892 struct ring_buffer
*rb
= handle
->rb
;
3893 int events
= local_inc_return(&rb
->events
);
3895 if (events
>= wakeup_events
) {
3896 local_sub(wakeup_events
, &rb
->events
);
3897 local_inc(&rb
->wakeup
);
3903 void perf_prepare_sample(struct perf_event_header
*header
,
3904 struct perf_sample_data
*data
,
3905 struct perf_event
*event
,
3906 struct pt_regs
*regs
)
3908 u64 sample_type
= event
->attr
.sample_type
;
3910 header
->type
= PERF_RECORD_SAMPLE
;
3911 header
->size
= sizeof(*header
) + event
->header_size
;
3914 header
->misc
|= perf_misc_flags(regs
);
3916 __perf_event_header__init_id(header
, data
, event
);
3918 if (sample_type
& PERF_SAMPLE_IP
)
3919 data
->ip
= perf_instruction_pointer(regs
);
3921 if (sample_type
& PERF_SAMPLE_CALLCHAIN
) {
3924 data
->callchain
= perf_callchain(regs
);
3926 if (data
->callchain
)
3927 size
+= data
->callchain
->nr
;
3929 header
->size
+= size
* sizeof(u64
);
3932 if (sample_type
& PERF_SAMPLE_RAW
) {
3933 int size
= sizeof(u32
);
3936 size
+= data
->raw
->size
;
3938 size
+= sizeof(u32
);
3940 WARN_ON_ONCE(size
& (sizeof(u64
)-1));
3941 header
->size
+= size
;
3945 static void perf_event_output(struct perf_event
*event
,
3946 struct perf_sample_data
*data
,
3947 struct pt_regs
*regs
)
3949 struct perf_output_handle handle
;
3950 struct perf_event_header header
;
3952 /* protect the callchain buffers */
3955 perf_prepare_sample(&header
, data
, event
, regs
);
3957 if (perf_output_begin(&handle
, event
, header
.size
))
3960 perf_output_sample(&handle
, &header
, data
, event
);
3962 perf_output_end(&handle
);
3972 struct perf_read_event
{
3973 struct perf_event_header header
;
3980 perf_event_read_event(struct perf_event
*event
,
3981 struct task_struct
*task
)
3983 struct perf_output_handle handle
;
3984 struct perf_sample_data sample
;
3985 struct perf_read_event read_event
= {
3987 .type
= PERF_RECORD_READ
,
3989 .size
= sizeof(read_event
) + event
->read_size
,
3991 .pid
= perf_event_pid(event
, task
),
3992 .tid
= perf_event_tid(event
, task
),
3996 perf_event_header__init_id(&read_event
.header
, &sample
, event
);
3997 ret
= perf_output_begin(&handle
, event
, read_event
.header
.size
);
4001 perf_output_put(&handle
, read_event
);
4002 perf_output_read(&handle
, event
);
4003 perf_event__output_id_sample(event
, &handle
, &sample
);
4005 perf_output_end(&handle
);
4009 * task tracking -- fork/exit
4011 * enabled by: attr.comm | attr.mmap | attr.mmap_data | attr.task
4014 struct perf_task_event
{
4015 struct task_struct
*task
;
4016 struct perf_event_context
*task_ctx
;
4019 struct perf_event_header header
;
4029 static void perf_event_task_output(struct perf_event
*event
,
4030 struct perf_task_event
*task_event
)
4032 struct perf_output_handle handle
;
4033 struct perf_sample_data sample
;
4034 struct task_struct
*task
= task_event
->task
;
4035 int ret
, size
= task_event
->event_id
.header
.size
;
4037 perf_event_header__init_id(&task_event
->event_id
.header
, &sample
, event
);
4039 ret
= perf_output_begin(&handle
, event
,
4040 task_event
->event_id
.header
.size
);
4044 task_event
->event_id
.pid
= perf_event_pid(event
, task
);
4045 task_event
->event_id
.ppid
= perf_event_pid(event
, current
);
4047 task_event
->event_id
.tid
= perf_event_tid(event
, task
);
4048 task_event
->event_id
.ptid
= perf_event_tid(event
, current
);
4050 perf_output_put(&handle
, task_event
->event_id
);
4052 perf_event__output_id_sample(event
, &handle
, &sample
);
4054 perf_output_end(&handle
);
4056 task_event
->event_id
.header
.size
= size
;
4059 static int perf_event_task_match(struct perf_event
*event
)
4061 if (event
->state
< PERF_EVENT_STATE_INACTIVE
)
4064 if (!event_filter_match(event
))
4067 if (event
->attr
.comm
|| event
->attr
.mmap
||
4068 event
->attr
.mmap_data
|| event
->attr
.task
)
4074 static void perf_event_task_ctx(struct perf_event_context
*ctx
,
4075 struct perf_task_event
*task_event
)
4077 struct perf_event
*event
;
4079 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
4080 if (perf_event_task_match(event
))
4081 perf_event_task_output(event
, task_event
);
4085 static void perf_event_task_event(struct perf_task_event
*task_event
)
4087 struct perf_cpu_context
*cpuctx
;
4088 struct perf_event_context
*ctx
;
4093 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
4094 cpuctx
= get_cpu_ptr(pmu
->pmu_cpu_context
);
4095 if (cpuctx
->active_pmu
!= pmu
)
4097 perf_event_task_ctx(&cpuctx
->ctx
, task_event
);
4099 ctx
= task_event
->task_ctx
;
4101 ctxn
= pmu
->task_ctx_nr
;
4104 ctx
= rcu_dereference(current
->perf_event_ctxp
[ctxn
]);
4107 perf_event_task_ctx(ctx
, task_event
);
4109 put_cpu_ptr(pmu
->pmu_cpu_context
);
4114 static void perf_event_task(struct task_struct
*task
,
4115 struct perf_event_context
*task_ctx
,
4118 struct perf_task_event task_event
;
4120 if (!atomic_read(&nr_comm_events
) &&
4121 !atomic_read(&nr_mmap_events
) &&
4122 !atomic_read(&nr_task_events
))
4125 task_event
= (struct perf_task_event
){
4127 .task_ctx
= task_ctx
,
4130 .type
= new ? PERF_RECORD_FORK
: PERF_RECORD_EXIT
,
4132 .size
= sizeof(task_event
.event_id
),
4138 .time
= perf_clock(),
4142 perf_event_task_event(&task_event
);
4145 void perf_event_fork(struct task_struct
*task
)
4147 perf_event_task(task
, NULL
, 1);
4154 struct perf_comm_event
{
4155 struct task_struct
*task
;
4160 struct perf_event_header header
;
4167 static void perf_event_comm_output(struct perf_event
*event
,
4168 struct perf_comm_event
*comm_event
)
4170 struct perf_output_handle handle
;
4171 struct perf_sample_data sample
;
4172 int size
= comm_event
->event_id
.header
.size
;
4175 perf_event_header__init_id(&comm_event
->event_id
.header
, &sample
, event
);
4176 ret
= perf_output_begin(&handle
, event
,
4177 comm_event
->event_id
.header
.size
);
4182 comm_event
->event_id
.pid
= perf_event_pid(event
, comm_event
->task
);
4183 comm_event
->event_id
.tid
= perf_event_tid(event
, comm_event
->task
);
4185 perf_output_put(&handle
, comm_event
->event_id
);
4186 __output_copy(&handle
, comm_event
->comm
,
4187 comm_event
->comm_size
);
4189 perf_event__output_id_sample(event
, &handle
, &sample
);
4191 perf_output_end(&handle
);
4193 comm_event
->event_id
.header
.size
= size
;
4196 static int perf_event_comm_match(struct perf_event
*event
)
4198 if (event
->state
< PERF_EVENT_STATE_INACTIVE
)
4201 if (!event_filter_match(event
))
4204 if (event
->attr
.comm
)
4210 static void perf_event_comm_ctx(struct perf_event_context
*ctx
,
4211 struct perf_comm_event
*comm_event
)
4213 struct perf_event
*event
;
4215 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
4216 if (perf_event_comm_match(event
))
4217 perf_event_comm_output(event
, comm_event
);
4221 static void perf_event_comm_event(struct perf_comm_event
*comm_event
)
4223 struct perf_cpu_context
*cpuctx
;
4224 struct perf_event_context
*ctx
;
4225 char comm
[TASK_COMM_LEN
];
4230 memset(comm
, 0, sizeof(comm
));
4231 strlcpy(comm
, comm_event
->task
->comm
, sizeof(comm
));
4232 size
= ALIGN(strlen(comm
)+1, sizeof(u64
));
4234 comm_event
->comm
= comm
;
4235 comm_event
->comm_size
= size
;
4237 comm_event
->event_id
.header
.size
= sizeof(comm_event
->event_id
) + size
;
4239 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
4240 cpuctx
= get_cpu_ptr(pmu
->pmu_cpu_context
);
4241 if (cpuctx
->active_pmu
!= pmu
)
4243 perf_event_comm_ctx(&cpuctx
->ctx
, comm_event
);
4245 ctxn
= pmu
->task_ctx_nr
;
4249 ctx
= rcu_dereference(current
->perf_event_ctxp
[ctxn
]);
4251 perf_event_comm_ctx(ctx
, comm_event
);
4253 put_cpu_ptr(pmu
->pmu_cpu_context
);
4258 void perf_event_comm(struct task_struct
*task
)
4260 struct perf_comm_event comm_event
;
4261 struct perf_event_context
*ctx
;
4264 for_each_task_context_nr(ctxn
) {
4265 ctx
= task
->perf_event_ctxp
[ctxn
];
4269 perf_event_enable_on_exec(ctx
);
4272 if (!atomic_read(&nr_comm_events
))
4275 comm_event
= (struct perf_comm_event
){
4281 .type
= PERF_RECORD_COMM
,
4290 perf_event_comm_event(&comm_event
);
4297 struct perf_mmap_event
{
4298 struct vm_area_struct
*vma
;
4300 const char *file_name
;
4304 struct perf_event_header header
;
4314 static void perf_event_mmap_output(struct perf_event
*event
,
4315 struct perf_mmap_event
*mmap_event
)
4317 struct perf_output_handle handle
;
4318 struct perf_sample_data sample
;
4319 int size
= mmap_event
->event_id
.header
.size
;
4322 perf_event_header__init_id(&mmap_event
->event_id
.header
, &sample
, event
);
4323 ret
= perf_output_begin(&handle
, event
,
4324 mmap_event
->event_id
.header
.size
);
4328 mmap_event
->event_id
.pid
= perf_event_pid(event
, current
);
4329 mmap_event
->event_id
.tid
= perf_event_tid(event
, current
);
4331 perf_output_put(&handle
, mmap_event
->event_id
);
4332 __output_copy(&handle
, mmap_event
->file_name
,
4333 mmap_event
->file_size
);
4335 perf_event__output_id_sample(event
, &handle
, &sample
);
4337 perf_output_end(&handle
);
4339 mmap_event
->event_id
.header
.size
= size
;
4342 static int perf_event_mmap_match(struct perf_event
*event
,
4343 struct perf_mmap_event
*mmap_event
,
4346 if (event
->state
< PERF_EVENT_STATE_INACTIVE
)
4349 if (!event_filter_match(event
))
4352 if ((!executable
&& event
->attr
.mmap_data
) ||
4353 (executable
&& event
->attr
.mmap
))
4359 static void perf_event_mmap_ctx(struct perf_event_context
*ctx
,
4360 struct perf_mmap_event
*mmap_event
,
4363 struct perf_event
*event
;
4365 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
4366 if (perf_event_mmap_match(event
, mmap_event
, executable
))
4367 perf_event_mmap_output(event
, mmap_event
);
4371 static void perf_event_mmap_event(struct perf_mmap_event
*mmap_event
)
4373 struct perf_cpu_context
*cpuctx
;
4374 struct perf_event_context
*ctx
;
4375 struct vm_area_struct
*vma
= mmap_event
->vma
;
4376 struct file
*file
= vma
->vm_file
;
4384 memset(tmp
, 0, sizeof(tmp
));
4388 * d_path works from the end of the rb backwards, so we
4389 * need to add enough zero bytes after the string to handle
4390 * the 64bit alignment we do later.
4392 buf
= kzalloc(PATH_MAX
+ sizeof(u64
), GFP_KERNEL
);
4394 name
= strncpy(tmp
, "//enomem", sizeof(tmp
));
4397 name
= d_path(&file
->f_path
, buf
, PATH_MAX
);
4399 name
= strncpy(tmp
, "//toolong", sizeof(tmp
));
4403 if (arch_vma_name(mmap_event
->vma
)) {
4404 name
= strncpy(tmp
, arch_vma_name(mmap_event
->vma
),
4410 name
= strncpy(tmp
, "[vdso]", sizeof(tmp
));
4412 } else if (vma
->vm_start
<= vma
->vm_mm
->start_brk
&&
4413 vma
->vm_end
>= vma
->vm_mm
->brk
) {
4414 name
= strncpy(tmp
, "[heap]", sizeof(tmp
));
4416 } else if (vma
->vm_start
<= vma
->vm_mm
->start_stack
&&
4417 vma
->vm_end
>= vma
->vm_mm
->start_stack
) {
4418 name
= strncpy(tmp
, "[stack]", sizeof(tmp
));
4422 name
= strncpy(tmp
, "//anon", sizeof(tmp
));
4427 size
= ALIGN(strlen(name
)+1, sizeof(u64
));
4429 mmap_event
->file_name
= name
;
4430 mmap_event
->file_size
= size
;
4432 mmap_event
->event_id
.header
.size
= sizeof(mmap_event
->event_id
) + size
;
4435 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
4436 cpuctx
= get_cpu_ptr(pmu
->pmu_cpu_context
);
4437 if (cpuctx
->active_pmu
!= pmu
)
4439 perf_event_mmap_ctx(&cpuctx
->ctx
, mmap_event
,
4440 vma
->vm_flags
& VM_EXEC
);
4442 ctxn
= pmu
->task_ctx_nr
;
4446 ctx
= rcu_dereference(current
->perf_event_ctxp
[ctxn
]);
4448 perf_event_mmap_ctx(ctx
, mmap_event
,
4449 vma
->vm_flags
& VM_EXEC
);
4452 put_cpu_ptr(pmu
->pmu_cpu_context
);
4459 void perf_event_mmap(struct vm_area_struct
*vma
)
4461 struct perf_mmap_event mmap_event
;
4463 if (!atomic_read(&nr_mmap_events
))
4466 mmap_event
= (struct perf_mmap_event
){
4472 .type
= PERF_RECORD_MMAP
,
4473 .misc
= PERF_RECORD_MISC_USER
,
4478 .start
= vma
->vm_start
,
4479 .len
= vma
->vm_end
- vma
->vm_start
,
4480 .pgoff
= (u64
)vma
->vm_pgoff
<< PAGE_SHIFT
,
4484 perf_event_mmap_event(&mmap_event
);
4488 * IRQ throttle logging
4491 static void perf_log_throttle(struct perf_event
*event
, int enable
)
4493 struct perf_output_handle handle
;
4494 struct perf_sample_data sample
;
4498 struct perf_event_header header
;
4502 } throttle_event
= {
4504 .type
= PERF_RECORD_THROTTLE
,
4506 .size
= sizeof(throttle_event
),
4508 .time
= perf_clock(),
4509 .id
= primary_event_id(event
),
4510 .stream_id
= event
->id
,
4514 throttle_event
.header
.type
= PERF_RECORD_UNTHROTTLE
;
4516 perf_event_header__init_id(&throttle_event
.header
, &sample
, event
);
4518 ret
= perf_output_begin(&handle
, event
,
4519 throttle_event
.header
.size
);
4523 perf_output_put(&handle
, throttle_event
);
4524 perf_event__output_id_sample(event
, &handle
, &sample
);
4525 perf_output_end(&handle
);
4529 * Generic event overflow handling, sampling.
4532 static int __perf_event_overflow(struct perf_event
*event
,
4533 int throttle
, struct perf_sample_data
*data
,
4534 struct pt_regs
*regs
)
4536 int events
= atomic_read(&event
->event_limit
);
4537 struct hw_perf_event
*hwc
= &event
->hw
;
4542 * Non-sampling counters might still use the PMI to fold short
4543 * hardware counters, ignore those.
4545 if (unlikely(!is_sampling_event(event
)))
4548 seq
= __this_cpu_read(perf_throttled_seq
);
4549 if (seq
!= hwc
->interrupts_seq
) {
4550 hwc
->interrupts_seq
= seq
;
4551 hwc
->interrupts
= 1;
4554 if (unlikely(throttle
4555 && hwc
->interrupts
>= max_samples_per_tick
)) {
4556 __this_cpu_inc(perf_throttled_count
);
4557 hwc
->interrupts
= MAX_INTERRUPTS
;
4558 perf_log_throttle(event
, 0);
4563 if (event
->attr
.freq
) {
4564 u64 now
= perf_clock();
4565 s64 delta
= now
- hwc
->freq_time_stamp
;
4567 hwc
->freq_time_stamp
= now
;
4569 if (delta
> 0 && delta
< 2*TICK_NSEC
)
4570 perf_adjust_period(event
, delta
, hwc
->last_period
);
4574 * XXX event_limit might not quite work as expected on inherited
4578 event
->pending_kill
= POLL_IN
;
4579 if (events
&& atomic_dec_and_test(&event
->event_limit
)) {
4581 event
->pending_kill
= POLL_HUP
;
4582 event
->pending_disable
= 1;
4583 irq_work_queue(&event
->pending
);
4586 if (event
->overflow_handler
)
4587 event
->overflow_handler(event
, data
, regs
);
4589 perf_event_output(event
, data
, regs
);
4591 if (event
->fasync
&& event
->pending_kill
) {
4592 event
->pending_wakeup
= 1;
4593 irq_work_queue(&event
->pending
);
4599 int perf_event_overflow(struct perf_event
*event
,
4600 struct perf_sample_data
*data
,
4601 struct pt_regs
*regs
)
4603 return __perf_event_overflow(event
, 1, data
, regs
);
4607 * Generic software event infrastructure
4610 struct swevent_htable
{
4611 struct swevent_hlist
*swevent_hlist
;
4612 struct mutex hlist_mutex
;
4615 /* Recursion avoidance in each contexts */
4616 int recursion
[PERF_NR_CONTEXTS
];
4619 static DEFINE_PER_CPU(struct swevent_htable
, swevent_htable
);
4622 * We directly increment event->count and keep a second value in
4623 * event->hw.period_left to count intervals. This period event
4624 * is kept in the range [-sample_period, 0] so that we can use the
4628 static u64
perf_swevent_set_period(struct perf_event
*event
)
4630 struct hw_perf_event
*hwc
= &event
->hw
;
4631 u64 period
= hwc
->last_period
;
4635 hwc
->last_period
= hwc
->sample_period
;
4638 old
= val
= local64_read(&hwc
->period_left
);
4642 nr
= div64_u64(period
+ val
, period
);
4643 offset
= nr
* period
;
4645 if (local64_cmpxchg(&hwc
->period_left
, old
, val
) != old
)
4651 static void perf_swevent_overflow(struct perf_event
*event
, u64 overflow
,
4652 struct perf_sample_data
*data
,
4653 struct pt_regs
*regs
)
4655 struct hw_perf_event
*hwc
= &event
->hw
;
4659 overflow
= perf_swevent_set_period(event
);
4661 if (hwc
->interrupts
== MAX_INTERRUPTS
)
4664 for (; overflow
; overflow
--) {
4665 if (__perf_event_overflow(event
, throttle
,
4668 * We inhibit the overflow from happening when
4669 * hwc->interrupts == MAX_INTERRUPTS.
4677 static void perf_swevent_event(struct perf_event
*event
, u64 nr
,
4678 struct perf_sample_data
*data
,
4679 struct pt_regs
*regs
)
4681 struct hw_perf_event
*hwc
= &event
->hw
;
4683 local64_add(nr
, &event
->count
);
4688 if (!is_sampling_event(event
))
4691 if ((event
->attr
.sample_type
& PERF_SAMPLE_PERIOD
) && !event
->attr
.freq
) {
4693 return perf_swevent_overflow(event
, 1, data
, regs
);
4695 data
->period
= event
->hw
.last_period
;
4697 if (nr
== 1 && hwc
->sample_period
== 1 && !event
->attr
.freq
)
4698 return perf_swevent_overflow(event
, 1, data
, regs
);
4700 if (local64_add_negative(nr
, &hwc
->period_left
))
4703 perf_swevent_overflow(event
, 0, data
, regs
);
4706 static int perf_exclude_event(struct perf_event
*event
,
4707 struct pt_regs
*regs
)
4709 if (event
->hw
.state
& PERF_HES_STOPPED
)
4713 if (event
->attr
.exclude_user
&& user_mode(regs
))
4716 if (event
->attr
.exclude_kernel
&& !user_mode(regs
))
4723 static int perf_swevent_match(struct perf_event
*event
,
4724 enum perf_type_id type
,
4726 struct perf_sample_data
*data
,
4727 struct pt_regs
*regs
)
4729 if (event
->attr
.type
!= type
)
4732 if (event
->attr
.config
!= event_id
)
4735 if (perf_exclude_event(event
, regs
))
4741 static inline u64
swevent_hash(u64 type
, u32 event_id
)
4743 u64 val
= event_id
| (type
<< 32);
4745 return hash_64(val
, SWEVENT_HLIST_BITS
);
4748 static inline struct hlist_head
*
4749 __find_swevent_head(struct swevent_hlist
*hlist
, u64 type
, u32 event_id
)
4751 u64 hash
= swevent_hash(type
, event_id
);
4753 return &hlist
->heads
[hash
];
4756 /* For the read side: events when they trigger */
4757 static inline struct hlist_head
*
4758 find_swevent_head_rcu(struct swevent_htable
*swhash
, u64 type
, u32 event_id
)
4760 struct swevent_hlist
*hlist
;
4762 hlist
= rcu_dereference(swhash
->swevent_hlist
);
4766 return __find_swevent_head(hlist
, type
, event_id
);
4769 /* For the event head insertion and removal in the hlist */
4770 static inline struct hlist_head
*
4771 find_swevent_head(struct swevent_htable
*swhash
, struct perf_event
*event
)
4773 struct swevent_hlist
*hlist
;
4774 u32 event_id
= event
->attr
.config
;
4775 u64 type
= event
->attr
.type
;
4778 * Event scheduling is always serialized against hlist allocation
4779 * and release. Which makes the protected version suitable here.
4780 * The context lock guarantees that.
4782 hlist
= rcu_dereference_protected(swhash
->swevent_hlist
,
4783 lockdep_is_held(&event
->ctx
->lock
));
4787 return __find_swevent_head(hlist
, type
, event_id
);
4790 static void do_perf_sw_event(enum perf_type_id type
, u32 event_id
,
4792 struct perf_sample_data
*data
,
4793 struct pt_regs
*regs
)
4795 struct swevent_htable
*swhash
= &__get_cpu_var(swevent_htable
);
4796 struct perf_event
*event
;
4797 struct hlist_node
*node
;
4798 struct hlist_head
*head
;
4801 head
= find_swevent_head_rcu(swhash
, type
, event_id
);
4805 hlist_for_each_entry_rcu(event
, node
, head
, hlist_entry
) {
4806 if (perf_swevent_match(event
, type
, event_id
, data
, regs
))
4807 perf_swevent_event(event
, nr
, data
, regs
);
4813 int perf_swevent_get_recursion_context(void)
4815 struct swevent_htable
*swhash
= &__get_cpu_var(swevent_htable
);
4817 return get_recursion_context(swhash
->recursion
);
4819 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context
);
4821 inline void perf_swevent_put_recursion_context(int rctx
)
4823 struct swevent_htable
*swhash
= &__get_cpu_var(swevent_htable
);
4825 put_recursion_context(swhash
->recursion
, rctx
);
4828 void __perf_sw_event(u32 event_id
, u64 nr
, struct pt_regs
*regs
, u64 addr
)
4830 struct perf_sample_data data
;
4833 preempt_disable_notrace();
4834 rctx
= perf_swevent_get_recursion_context();
4838 perf_sample_data_init(&data
, addr
);
4840 do_perf_sw_event(PERF_TYPE_SOFTWARE
, event_id
, nr
, &data
, regs
);
4842 perf_swevent_put_recursion_context(rctx
);
4843 preempt_enable_notrace();
4846 static void perf_swevent_read(struct perf_event
*event
)
4850 static int perf_swevent_add(struct perf_event
*event
, int flags
)
4852 struct swevent_htable
*swhash
= &__get_cpu_var(swevent_htable
);
4853 struct hw_perf_event
*hwc
= &event
->hw
;
4854 struct hlist_head
*head
;
4856 if (is_sampling_event(event
)) {
4857 hwc
->last_period
= hwc
->sample_period
;
4858 perf_swevent_set_period(event
);
4861 hwc
->state
= !(flags
& PERF_EF_START
);
4863 head
= find_swevent_head(swhash
, event
);
4864 if (WARN_ON_ONCE(!head
))
4867 hlist_add_head_rcu(&event
->hlist_entry
, head
);
4872 static void perf_swevent_del(struct perf_event
*event
, int flags
)
4874 hlist_del_rcu(&event
->hlist_entry
);
4877 static void perf_swevent_start(struct perf_event
*event
, int flags
)
4879 event
->hw
.state
= 0;
4882 static void perf_swevent_stop(struct perf_event
*event
, int flags
)
4884 event
->hw
.state
= PERF_HES_STOPPED
;
4887 /* Deref the hlist from the update side */
4888 static inline struct swevent_hlist
*
4889 swevent_hlist_deref(struct swevent_htable
*swhash
)
4891 return rcu_dereference_protected(swhash
->swevent_hlist
,
4892 lockdep_is_held(&swhash
->hlist_mutex
));
4895 static void swevent_hlist_release(struct swevent_htable
*swhash
)
4897 struct swevent_hlist
*hlist
= swevent_hlist_deref(swhash
);
4902 rcu_assign_pointer(swhash
->swevent_hlist
, NULL
);
4903 kfree_rcu(hlist
, rcu_head
);
4906 static void swevent_hlist_put_cpu(struct perf_event
*event
, int cpu
)
4908 struct swevent_htable
*swhash
= &per_cpu(swevent_htable
, cpu
);
4910 mutex_lock(&swhash
->hlist_mutex
);
4912 if (!--swhash
->hlist_refcount
)
4913 swevent_hlist_release(swhash
);
4915 mutex_unlock(&swhash
->hlist_mutex
);
4918 static void swevent_hlist_put(struct perf_event
*event
)
4922 if (event
->cpu
!= -1) {
4923 swevent_hlist_put_cpu(event
, event
->cpu
);
4927 for_each_possible_cpu(cpu
)
4928 swevent_hlist_put_cpu(event
, cpu
);
4931 static int swevent_hlist_get_cpu(struct perf_event
*event
, int cpu
)
4933 struct swevent_htable
*swhash
= &per_cpu(swevent_htable
, cpu
);
4936 mutex_lock(&swhash
->hlist_mutex
);
4938 if (!swevent_hlist_deref(swhash
) && cpu_online(cpu
)) {
4939 struct swevent_hlist
*hlist
;
4941 hlist
= kzalloc(sizeof(*hlist
), GFP_KERNEL
);
4946 rcu_assign_pointer(swhash
->swevent_hlist
, hlist
);
4948 swhash
->hlist_refcount
++;
4950 mutex_unlock(&swhash
->hlist_mutex
);
4955 static int swevent_hlist_get(struct perf_event
*event
)
4958 int cpu
, failed_cpu
;
4960 if (event
->cpu
!= -1)
4961 return swevent_hlist_get_cpu(event
, event
->cpu
);
4964 for_each_possible_cpu(cpu
) {
4965 err
= swevent_hlist_get_cpu(event
, cpu
);
4975 for_each_possible_cpu(cpu
) {
4976 if (cpu
== failed_cpu
)
4978 swevent_hlist_put_cpu(event
, cpu
);
4985 struct static_key perf_swevent_enabled
[PERF_COUNT_SW_MAX
];
4987 static void sw_perf_event_destroy(struct perf_event
*event
)
4989 u64 event_id
= event
->attr
.config
;
4991 WARN_ON(event
->parent
);
4993 static_key_slow_dec(&perf_swevent_enabled
[event_id
]);
4994 swevent_hlist_put(event
);
4997 static int perf_swevent_init(struct perf_event
*event
)
4999 int event_id
= event
->attr
.config
;
5001 if (event
->attr
.type
!= PERF_TYPE_SOFTWARE
)
5005 case PERF_COUNT_SW_CPU_CLOCK
:
5006 case PERF_COUNT_SW_TASK_CLOCK
:
5013 if (event_id
>= PERF_COUNT_SW_MAX
)
5016 if (!event
->parent
) {
5019 err
= swevent_hlist_get(event
);
5023 static_key_slow_inc(&perf_swevent_enabled
[event_id
]);
5024 event
->destroy
= sw_perf_event_destroy
;
5030 static int perf_swevent_event_idx(struct perf_event
*event
)
5035 static struct pmu perf_swevent
= {
5036 .task_ctx_nr
= perf_sw_context
,
5038 .event_init
= perf_swevent_init
,
5039 .add
= perf_swevent_add
,
5040 .del
= perf_swevent_del
,
5041 .start
= perf_swevent_start
,
5042 .stop
= perf_swevent_stop
,
5043 .read
= perf_swevent_read
,
5045 .event_idx
= perf_swevent_event_idx
,
5048 #ifdef CONFIG_EVENT_TRACING
5050 static int perf_tp_filter_match(struct perf_event
*event
,
5051 struct perf_sample_data
*data
)
5053 void *record
= data
->raw
->data
;
5055 if (likely(!event
->filter
) || filter_match_preds(event
->filter
, record
))
5060 static int perf_tp_event_match(struct perf_event
*event
,
5061 struct perf_sample_data
*data
,
5062 struct pt_regs
*regs
)
5064 if (event
->hw
.state
& PERF_HES_STOPPED
)
5067 * All tracepoints are from kernel-space.
5069 if (event
->attr
.exclude_kernel
)
5072 if (!perf_tp_filter_match(event
, data
))
5078 void perf_tp_event(u64 addr
, u64 count
, void *record
, int entry_size
,
5079 struct pt_regs
*regs
, struct hlist_head
*head
, int rctx
)
5081 struct perf_sample_data data
;
5082 struct perf_event
*event
;
5083 struct hlist_node
*node
;
5085 struct perf_raw_record raw
= {
5090 perf_sample_data_init(&data
, addr
);
5093 hlist_for_each_entry_rcu(event
, node
, head
, hlist_entry
) {
5094 if (perf_tp_event_match(event
, &data
, regs
))
5095 perf_swevent_event(event
, count
, &data
, regs
);
5098 perf_swevent_put_recursion_context(rctx
);
5100 EXPORT_SYMBOL_GPL(perf_tp_event
);
5102 static void tp_perf_event_destroy(struct perf_event
*event
)
5104 perf_trace_destroy(event
);
5107 static int perf_tp_event_init(struct perf_event
*event
)
5111 if (event
->attr
.type
!= PERF_TYPE_TRACEPOINT
)
5114 err
= perf_trace_init(event
);
5118 event
->destroy
= tp_perf_event_destroy
;
5123 static struct pmu perf_tracepoint
= {
5124 .task_ctx_nr
= perf_sw_context
,
5126 .event_init
= perf_tp_event_init
,
5127 .add
= perf_trace_add
,
5128 .del
= perf_trace_del
,
5129 .start
= perf_swevent_start
,
5130 .stop
= perf_swevent_stop
,
5131 .read
= perf_swevent_read
,
5133 .event_idx
= perf_swevent_event_idx
,
5136 static inline void perf_tp_register(void)
5138 perf_pmu_register(&perf_tracepoint
, "tracepoint", PERF_TYPE_TRACEPOINT
);
5141 static int perf_event_set_filter(struct perf_event
*event
, void __user
*arg
)
5146 if (event
->attr
.type
!= PERF_TYPE_TRACEPOINT
)
5149 filter_str
= strndup_user(arg
, PAGE_SIZE
);
5150 if (IS_ERR(filter_str
))
5151 return PTR_ERR(filter_str
);
5153 ret
= ftrace_profile_set_filter(event
, event
->attr
.config
, filter_str
);
5159 static void perf_event_free_filter(struct perf_event
*event
)
5161 ftrace_profile_free_filter(event
);
5166 static inline void perf_tp_register(void)
5170 static int perf_event_set_filter(struct perf_event
*event
, void __user
*arg
)
5175 static void perf_event_free_filter(struct perf_event
*event
)
5179 #endif /* CONFIG_EVENT_TRACING */
5181 #ifdef CONFIG_HAVE_HW_BREAKPOINT
5182 void perf_bp_event(struct perf_event
*bp
, void *data
)
5184 struct perf_sample_data sample
;
5185 struct pt_regs
*regs
= data
;
5187 perf_sample_data_init(&sample
, bp
->attr
.bp_addr
);
5189 if (!bp
->hw
.state
&& !perf_exclude_event(bp
, regs
))
5190 perf_swevent_event(bp
, 1, &sample
, regs
);
5195 * hrtimer based swevent callback
5198 static enum hrtimer_restart
perf_swevent_hrtimer(struct hrtimer
*hrtimer
)
5200 enum hrtimer_restart ret
= HRTIMER_RESTART
;
5201 struct perf_sample_data data
;
5202 struct pt_regs
*regs
;
5203 struct perf_event
*event
;
5206 event
= container_of(hrtimer
, struct perf_event
, hw
.hrtimer
);
5208 if (event
->state
!= PERF_EVENT_STATE_ACTIVE
)
5209 return HRTIMER_NORESTART
;
5211 event
->pmu
->read(event
);
5213 perf_sample_data_init(&data
, 0);
5214 data
.period
= event
->hw
.last_period
;
5215 regs
= get_irq_regs();
5217 if (regs
&& !perf_exclude_event(event
, regs
)) {
5218 if (!(event
->attr
.exclude_idle
&& is_idle_task(current
)))
5219 if (perf_event_overflow(event
, &data
, regs
))
5220 ret
= HRTIMER_NORESTART
;
5223 period
= max_t(u64
, 10000, event
->hw
.sample_period
);
5224 hrtimer_forward_now(hrtimer
, ns_to_ktime(period
));
5229 static void perf_swevent_start_hrtimer(struct perf_event
*event
)
5231 struct hw_perf_event
*hwc
= &event
->hw
;
5234 if (!is_sampling_event(event
))
5237 period
= local64_read(&hwc
->period_left
);
5242 local64_set(&hwc
->period_left
, 0);
5244 period
= max_t(u64
, 10000, hwc
->sample_period
);
5246 __hrtimer_start_range_ns(&hwc
->hrtimer
,
5247 ns_to_ktime(period
), 0,
5248 HRTIMER_MODE_REL_PINNED
, 0);
5251 static void perf_swevent_cancel_hrtimer(struct perf_event
*event
)
5253 struct hw_perf_event
*hwc
= &event
->hw
;
5255 if (is_sampling_event(event
)) {
5256 ktime_t remaining
= hrtimer_get_remaining(&hwc
->hrtimer
);
5257 local64_set(&hwc
->period_left
, ktime_to_ns(remaining
));
5259 hrtimer_cancel(&hwc
->hrtimer
);
5263 static void perf_swevent_init_hrtimer(struct perf_event
*event
)
5265 struct hw_perf_event
*hwc
= &event
->hw
;
5267 if (!is_sampling_event(event
))
5270 hrtimer_init(&hwc
->hrtimer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
5271 hwc
->hrtimer
.function
= perf_swevent_hrtimer
;
5274 * Since hrtimers have a fixed rate, we can do a static freq->period
5275 * mapping and avoid the whole period adjust feedback stuff.
5277 if (event
->attr
.freq
) {
5278 long freq
= event
->attr
.sample_freq
;
5280 event
->attr
.sample_period
= NSEC_PER_SEC
/ freq
;
5281 hwc
->sample_period
= event
->attr
.sample_period
;
5282 local64_set(&hwc
->period_left
, hwc
->sample_period
);
5283 event
->attr
.freq
= 0;
5288 * Software event: cpu wall time clock
5291 static void cpu_clock_event_update(struct perf_event
*event
)
5296 now
= local_clock();
5297 prev
= local64_xchg(&event
->hw
.prev_count
, now
);
5298 local64_add(now
- prev
, &event
->count
);
5301 static void cpu_clock_event_start(struct perf_event
*event
, int flags
)
5303 local64_set(&event
->hw
.prev_count
, local_clock());
5304 perf_swevent_start_hrtimer(event
);
5307 static void cpu_clock_event_stop(struct perf_event
*event
, int flags
)
5309 perf_swevent_cancel_hrtimer(event
);
5310 cpu_clock_event_update(event
);
5313 static int cpu_clock_event_add(struct perf_event
*event
, int flags
)
5315 if (flags
& PERF_EF_START
)
5316 cpu_clock_event_start(event
, flags
);
5321 static void cpu_clock_event_del(struct perf_event
*event
, int flags
)
5323 cpu_clock_event_stop(event
, flags
);
5326 static void cpu_clock_event_read(struct perf_event
*event
)
5328 cpu_clock_event_update(event
);
5331 static int cpu_clock_event_init(struct perf_event
*event
)
5333 if (event
->attr
.type
!= PERF_TYPE_SOFTWARE
)
5336 if (event
->attr
.config
!= PERF_COUNT_SW_CPU_CLOCK
)
5339 perf_swevent_init_hrtimer(event
);
5344 static struct pmu perf_cpu_clock
= {
5345 .task_ctx_nr
= perf_sw_context
,
5347 .event_init
= cpu_clock_event_init
,
5348 .add
= cpu_clock_event_add
,
5349 .del
= cpu_clock_event_del
,
5350 .start
= cpu_clock_event_start
,
5351 .stop
= cpu_clock_event_stop
,
5352 .read
= cpu_clock_event_read
,
5354 .event_idx
= perf_swevent_event_idx
,
5358 * Software event: task time clock
5361 static void task_clock_event_update(struct perf_event
*event
, u64 now
)
5366 prev
= local64_xchg(&event
->hw
.prev_count
, now
);
5368 local64_add(delta
, &event
->count
);
5371 static void task_clock_event_start(struct perf_event
*event
, int flags
)
5373 local64_set(&event
->hw
.prev_count
, event
->ctx
->time
);
5374 perf_swevent_start_hrtimer(event
);
5377 static void task_clock_event_stop(struct perf_event
*event
, int flags
)
5379 perf_swevent_cancel_hrtimer(event
);
5380 task_clock_event_update(event
, event
->ctx
->time
);
5383 static int task_clock_event_add(struct perf_event
*event
, int flags
)
5385 if (flags
& PERF_EF_START
)
5386 task_clock_event_start(event
, flags
);
5391 static void task_clock_event_del(struct perf_event
*event
, int flags
)
5393 task_clock_event_stop(event
, PERF_EF_UPDATE
);
5396 static void task_clock_event_read(struct perf_event
*event
)
5398 u64 now
= perf_clock();
5399 u64 delta
= now
- event
->ctx
->timestamp
;
5400 u64 time
= event
->ctx
->time
+ delta
;
5402 task_clock_event_update(event
, time
);
5405 static int task_clock_event_init(struct perf_event
*event
)
5407 if (event
->attr
.type
!= PERF_TYPE_SOFTWARE
)
5410 if (event
->attr
.config
!= PERF_COUNT_SW_TASK_CLOCK
)
5413 perf_swevent_init_hrtimer(event
);
5418 static struct pmu perf_task_clock
= {
5419 .task_ctx_nr
= perf_sw_context
,
5421 .event_init
= task_clock_event_init
,
5422 .add
= task_clock_event_add
,
5423 .del
= task_clock_event_del
,
5424 .start
= task_clock_event_start
,
5425 .stop
= task_clock_event_stop
,
5426 .read
= task_clock_event_read
,
5428 .event_idx
= perf_swevent_event_idx
,
5431 static void perf_pmu_nop_void(struct pmu
*pmu
)
5435 static int perf_pmu_nop_int(struct pmu
*pmu
)
5440 static void perf_pmu_start_txn(struct pmu
*pmu
)
5442 perf_pmu_disable(pmu
);
5445 static int perf_pmu_commit_txn(struct pmu
*pmu
)
5447 perf_pmu_enable(pmu
);
5451 static void perf_pmu_cancel_txn(struct pmu
*pmu
)
5453 perf_pmu_enable(pmu
);
5456 static int perf_event_idx_default(struct perf_event
*event
)
5458 return event
->hw
.idx
+ 1;
5462 * Ensures all contexts with the same task_ctx_nr have the same
5463 * pmu_cpu_context too.
5465 static void *find_pmu_context(int ctxn
)
5472 list_for_each_entry(pmu
, &pmus
, entry
) {
5473 if (pmu
->task_ctx_nr
== ctxn
)
5474 return pmu
->pmu_cpu_context
;
5480 static void update_pmu_context(struct pmu
*pmu
, struct pmu
*old_pmu
)
5484 for_each_possible_cpu(cpu
) {
5485 struct perf_cpu_context
*cpuctx
;
5487 cpuctx
= per_cpu_ptr(pmu
->pmu_cpu_context
, cpu
);
5489 if (cpuctx
->active_pmu
== old_pmu
)
5490 cpuctx
->active_pmu
= pmu
;
5494 static void free_pmu_context(struct pmu
*pmu
)
5498 mutex_lock(&pmus_lock
);
5500 * Like a real lame refcount.
5502 list_for_each_entry(i
, &pmus
, entry
) {
5503 if (i
->pmu_cpu_context
== pmu
->pmu_cpu_context
) {
5504 update_pmu_context(i
, pmu
);
5509 free_percpu(pmu
->pmu_cpu_context
);
5511 mutex_unlock(&pmus_lock
);
5513 static struct idr pmu_idr
;
5516 type_show(struct device
*dev
, struct device_attribute
*attr
, char *page
)
5518 struct pmu
*pmu
= dev_get_drvdata(dev
);
5520 return snprintf(page
, PAGE_SIZE
-1, "%d\n", pmu
->type
);
5523 static struct device_attribute pmu_dev_attrs
[] = {
5528 static int pmu_bus_running
;
5529 static struct bus_type pmu_bus
= {
5530 .name
= "event_source",
5531 .dev_attrs
= pmu_dev_attrs
,
5534 static void pmu_dev_release(struct device
*dev
)
5539 static int pmu_dev_alloc(struct pmu
*pmu
)
5543 pmu
->dev
= kzalloc(sizeof(struct device
), GFP_KERNEL
);
5547 pmu
->dev
->groups
= pmu
->attr_groups
;
5548 device_initialize(pmu
->dev
);
5549 ret
= dev_set_name(pmu
->dev
, "%s", pmu
->name
);
5553 dev_set_drvdata(pmu
->dev
, pmu
);
5554 pmu
->dev
->bus
= &pmu_bus
;
5555 pmu
->dev
->release
= pmu_dev_release
;
5556 ret
= device_add(pmu
->dev
);
5564 put_device(pmu
->dev
);
5568 static struct lock_class_key cpuctx_mutex
;
5569 static struct lock_class_key cpuctx_lock
;
5571 int perf_pmu_register(struct pmu
*pmu
, char *name
, int type
)
5575 mutex_lock(&pmus_lock
);
5577 pmu
->pmu_disable_count
= alloc_percpu(int);
5578 if (!pmu
->pmu_disable_count
)
5587 int err
= idr_pre_get(&pmu_idr
, GFP_KERNEL
);
5591 err
= idr_get_new_above(&pmu_idr
, pmu
, PERF_TYPE_MAX
, &type
);
5599 if (pmu_bus_running
) {
5600 ret
= pmu_dev_alloc(pmu
);
5606 pmu
->pmu_cpu_context
= find_pmu_context(pmu
->task_ctx_nr
);
5607 if (pmu
->pmu_cpu_context
)
5608 goto got_cpu_context
;
5610 pmu
->pmu_cpu_context
= alloc_percpu(struct perf_cpu_context
);
5611 if (!pmu
->pmu_cpu_context
)
5614 for_each_possible_cpu(cpu
) {
5615 struct perf_cpu_context
*cpuctx
;
5617 cpuctx
= per_cpu_ptr(pmu
->pmu_cpu_context
, cpu
);
5618 __perf_event_init_context(&cpuctx
->ctx
);
5619 lockdep_set_class(&cpuctx
->ctx
.mutex
, &cpuctx_mutex
);
5620 lockdep_set_class(&cpuctx
->ctx
.lock
, &cpuctx_lock
);
5621 cpuctx
->ctx
.type
= cpu_context
;
5622 cpuctx
->ctx
.pmu
= pmu
;
5623 cpuctx
->jiffies_interval
= 1;
5624 INIT_LIST_HEAD(&cpuctx
->rotation_list
);
5625 cpuctx
->active_pmu
= pmu
;
5629 if (!pmu
->start_txn
) {
5630 if (pmu
->pmu_enable
) {
5632 * If we have pmu_enable/pmu_disable calls, install
5633 * transaction stubs that use that to try and batch
5634 * hardware accesses.
5636 pmu
->start_txn
= perf_pmu_start_txn
;
5637 pmu
->commit_txn
= perf_pmu_commit_txn
;
5638 pmu
->cancel_txn
= perf_pmu_cancel_txn
;
5640 pmu
->start_txn
= perf_pmu_nop_void
;
5641 pmu
->commit_txn
= perf_pmu_nop_int
;
5642 pmu
->cancel_txn
= perf_pmu_nop_void
;
5646 if (!pmu
->pmu_enable
) {
5647 pmu
->pmu_enable
= perf_pmu_nop_void
;
5648 pmu
->pmu_disable
= perf_pmu_nop_void
;
5651 if (!pmu
->event_idx
)
5652 pmu
->event_idx
= perf_event_idx_default
;
5654 list_add_rcu(&pmu
->entry
, &pmus
);
5657 mutex_unlock(&pmus_lock
);
5662 device_del(pmu
->dev
);
5663 put_device(pmu
->dev
);
5666 if (pmu
->type
>= PERF_TYPE_MAX
)
5667 idr_remove(&pmu_idr
, pmu
->type
);
5670 free_percpu(pmu
->pmu_disable_count
);
5674 void perf_pmu_unregister(struct pmu
*pmu
)
5676 mutex_lock(&pmus_lock
);
5677 list_del_rcu(&pmu
->entry
);
5678 mutex_unlock(&pmus_lock
);
5681 * We dereference the pmu list under both SRCU and regular RCU, so
5682 * synchronize against both of those.
5684 synchronize_srcu(&pmus_srcu
);
5687 free_percpu(pmu
->pmu_disable_count
);
5688 if (pmu
->type
>= PERF_TYPE_MAX
)
5689 idr_remove(&pmu_idr
, pmu
->type
);
5690 device_del(pmu
->dev
);
5691 put_device(pmu
->dev
);
5692 free_pmu_context(pmu
);
5695 struct pmu
*perf_init_event(struct perf_event
*event
)
5697 struct pmu
*pmu
= NULL
;
5701 idx
= srcu_read_lock(&pmus_srcu
);
5704 pmu
= idr_find(&pmu_idr
, event
->attr
.type
);
5708 ret
= pmu
->event_init(event
);
5714 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
5716 ret
= pmu
->event_init(event
);
5720 if (ret
!= -ENOENT
) {
5725 pmu
= ERR_PTR(-ENOENT
);
5727 srcu_read_unlock(&pmus_srcu
, idx
);
5733 * Allocate and initialize a event structure
5735 static struct perf_event
*
5736 perf_event_alloc(struct perf_event_attr
*attr
, int cpu
,
5737 struct task_struct
*task
,
5738 struct perf_event
*group_leader
,
5739 struct perf_event
*parent_event
,
5740 perf_overflow_handler_t overflow_handler
,
5744 struct perf_event
*event
;
5745 struct hw_perf_event
*hwc
;
5748 if ((unsigned)cpu
>= nr_cpu_ids
) {
5749 if (!task
|| cpu
!= -1)
5750 return ERR_PTR(-EINVAL
);
5753 event
= kzalloc(sizeof(*event
), GFP_KERNEL
);
5755 return ERR_PTR(-ENOMEM
);
5758 * Single events are their own group leaders, with an
5759 * empty sibling list:
5762 group_leader
= event
;
5764 mutex_init(&event
->child_mutex
);
5765 INIT_LIST_HEAD(&event
->child_list
);
5767 INIT_LIST_HEAD(&event
->group_entry
);
5768 INIT_LIST_HEAD(&event
->event_entry
);
5769 INIT_LIST_HEAD(&event
->sibling_list
);
5770 INIT_LIST_HEAD(&event
->rb_entry
);
5772 init_waitqueue_head(&event
->waitq
);
5773 init_irq_work(&event
->pending
, perf_pending_event
);
5775 mutex_init(&event
->mmap_mutex
);
5778 event
->attr
= *attr
;
5779 event
->group_leader
= group_leader
;
5783 event
->parent
= parent_event
;
5785 event
->ns
= get_pid_ns(current
->nsproxy
->pid_ns
);
5786 event
->id
= atomic64_inc_return(&perf_event_id
);
5788 event
->state
= PERF_EVENT_STATE_INACTIVE
;
5791 event
->attach_state
= PERF_ATTACH_TASK
;
5792 #ifdef CONFIG_HAVE_HW_BREAKPOINT
5794 * hw_breakpoint is a bit difficult here..
5796 if (attr
->type
== PERF_TYPE_BREAKPOINT
)
5797 event
->hw
.bp_target
= task
;
5801 if (!overflow_handler
&& parent_event
) {
5802 overflow_handler
= parent_event
->overflow_handler
;
5803 context
= parent_event
->overflow_handler_context
;
5806 event
->overflow_handler
= overflow_handler
;
5807 event
->overflow_handler_context
= context
;
5810 event
->state
= PERF_EVENT_STATE_OFF
;
5815 hwc
->sample_period
= attr
->sample_period
;
5816 if (attr
->freq
&& attr
->sample_freq
)
5817 hwc
->sample_period
= 1;
5818 hwc
->last_period
= hwc
->sample_period
;
5820 local64_set(&hwc
->period_left
, hwc
->sample_period
);
5823 * we currently do not support PERF_FORMAT_GROUP on inherited events
5825 if (attr
->inherit
&& (attr
->read_format
& PERF_FORMAT_GROUP
))
5828 pmu
= perf_init_event(event
);
5834 else if (IS_ERR(pmu
))
5839 put_pid_ns(event
->ns
);
5841 return ERR_PTR(err
);
5844 if (!event
->parent
) {
5845 if (event
->attach_state
& PERF_ATTACH_TASK
)
5846 static_key_slow_inc(&perf_sched_events
.key
);
5847 if (event
->attr
.mmap
|| event
->attr
.mmap_data
)
5848 atomic_inc(&nr_mmap_events
);
5849 if (event
->attr
.comm
)
5850 atomic_inc(&nr_comm_events
);
5851 if (event
->attr
.task
)
5852 atomic_inc(&nr_task_events
);
5853 if (event
->attr
.sample_type
& PERF_SAMPLE_CALLCHAIN
) {
5854 err
= get_callchain_buffers();
5857 return ERR_PTR(err
);
5865 static int perf_copy_attr(struct perf_event_attr __user
*uattr
,
5866 struct perf_event_attr
*attr
)
5871 if (!access_ok(VERIFY_WRITE
, uattr
, PERF_ATTR_SIZE_VER0
))
5875 * zero the full structure, so that a short copy will be nice.
5877 memset(attr
, 0, sizeof(*attr
));
5879 ret
= get_user(size
, &uattr
->size
);
5883 if (size
> PAGE_SIZE
) /* silly large */
5886 if (!size
) /* abi compat */
5887 size
= PERF_ATTR_SIZE_VER0
;
5889 if (size
< PERF_ATTR_SIZE_VER0
)
5893 * If we're handed a bigger struct than we know of,
5894 * ensure all the unknown bits are 0 - i.e. new
5895 * user-space does not rely on any kernel feature
5896 * extensions we dont know about yet.
5898 if (size
> sizeof(*attr
)) {
5899 unsigned char __user
*addr
;
5900 unsigned char __user
*end
;
5903 addr
= (void __user
*)uattr
+ sizeof(*attr
);
5904 end
= (void __user
*)uattr
+ size
;
5906 for (; addr
< end
; addr
++) {
5907 ret
= get_user(val
, addr
);
5913 size
= sizeof(*attr
);
5916 ret
= copy_from_user(attr
, uattr
, size
);
5920 if (attr
->__reserved_1
)
5923 if (attr
->sample_type
& ~(PERF_SAMPLE_MAX
-1))
5926 if (attr
->read_format
& ~(PERF_FORMAT_MAX
-1))
5933 put_user(sizeof(*attr
), &uattr
->size
);
5939 perf_event_set_output(struct perf_event
*event
, struct perf_event
*output_event
)
5941 struct ring_buffer
*rb
= NULL
, *old_rb
= NULL
;
5947 /* don't allow circular references */
5948 if (event
== output_event
)
5952 * Don't allow cross-cpu buffers
5954 if (output_event
->cpu
!= event
->cpu
)
5958 * If its not a per-cpu rb, it must be the same task.
5960 if (output_event
->cpu
== -1 && output_event
->ctx
!= event
->ctx
)
5964 mutex_lock(&event
->mmap_mutex
);
5965 /* Can't redirect output if we've got an active mmap() */
5966 if (atomic_read(&event
->mmap_count
))
5970 /* get the rb we want to redirect to */
5971 rb
= ring_buffer_get(output_event
);
5977 rcu_assign_pointer(event
->rb
, rb
);
5979 ring_buffer_detach(event
, old_rb
);
5982 mutex_unlock(&event
->mmap_mutex
);
5985 ring_buffer_put(old_rb
);
5991 * sys_perf_event_open - open a performance event, associate it to a task/cpu
5993 * @attr_uptr: event_id type attributes for monitoring/sampling
5996 * @group_fd: group leader event fd
5998 SYSCALL_DEFINE5(perf_event_open
,
5999 struct perf_event_attr __user
*, attr_uptr
,
6000 pid_t
, pid
, int, cpu
, int, group_fd
, unsigned long, flags
)
6002 struct perf_event
*group_leader
= NULL
, *output_event
= NULL
;
6003 struct perf_event
*event
, *sibling
;
6004 struct perf_event_attr attr
;
6005 struct perf_event_context
*ctx
;
6006 struct file
*event_file
= NULL
;
6007 struct file
*group_file
= NULL
;
6008 struct task_struct
*task
= NULL
;
6012 int fput_needed
= 0;
6015 /* for future expandability... */
6016 if (flags
& ~PERF_FLAG_ALL
)
6019 err
= perf_copy_attr(attr_uptr
, &attr
);
6023 if (!attr
.exclude_kernel
) {
6024 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN
))
6029 if (attr
.sample_freq
> sysctl_perf_event_sample_rate
)
6034 * In cgroup mode, the pid argument is used to pass the fd
6035 * opened to the cgroup directory in cgroupfs. The cpu argument
6036 * designates the cpu on which to monitor threads from that
6039 if ((flags
& PERF_FLAG_PID_CGROUP
) && (pid
== -1 || cpu
== -1))
6042 event_fd
= get_unused_fd_flags(O_RDWR
);
6046 if (group_fd
!= -1) {
6047 group_leader
= perf_fget_light(group_fd
, &fput_needed
);
6048 if (IS_ERR(group_leader
)) {
6049 err
= PTR_ERR(group_leader
);
6052 group_file
= group_leader
->filp
;
6053 if (flags
& PERF_FLAG_FD_OUTPUT
)
6054 output_event
= group_leader
;
6055 if (flags
& PERF_FLAG_FD_NO_GROUP
)
6056 group_leader
= NULL
;
6059 if (pid
!= -1 && !(flags
& PERF_FLAG_PID_CGROUP
)) {
6060 task
= find_lively_task_by_vpid(pid
);
6062 err
= PTR_ERR(task
);
6067 event
= perf_event_alloc(&attr
, cpu
, task
, group_leader
, NULL
,
6069 if (IS_ERR(event
)) {
6070 err
= PTR_ERR(event
);
6074 if (flags
& PERF_FLAG_PID_CGROUP
) {
6075 err
= perf_cgroup_connect(pid
, event
, &attr
, group_leader
);
6080 * - that has cgroup constraint on event->cpu
6081 * - that may need work on context switch
6083 atomic_inc(&per_cpu(perf_cgroup_events
, event
->cpu
));
6084 static_key_slow_inc(&perf_sched_events
.key
);
6088 * Special case software events and allow them to be part of
6089 * any hardware group.
6094 (is_software_event(event
) != is_software_event(group_leader
))) {
6095 if (is_software_event(event
)) {
6097 * If event and group_leader are not both a software
6098 * event, and event is, then group leader is not.
6100 * Allow the addition of software events to !software
6101 * groups, this is safe because software events never
6104 pmu
= group_leader
->pmu
;
6105 } else if (is_software_event(group_leader
) &&
6106 (group_leader
->group_flags
& PERF_GROUP_SOFTWARE
)) {
6108 * In case the group is a pure software group, and we
6109 * try to add a hardware event, move the whole group to
6110 * the hardware context.
6117 * Get the target context (task or percpu):
6119 ctx
= find_get_context(pmu
, task
, cpu
);
6126 put_task_struct(task
);
6131 * Look up the group leader (we will attach this event to it):
6137 * Do not allow a recursive hierarchy (this new sibling
6138 * becoming part of another group-sibling):
6140 if (group_leader
->group_leader
!= group_leader
)
6143 * Do not allow to attach to a group in a different
6144 * task or CPU context:
6147 if (group_leader
->ctx
->type
!= ctx
->type
)
6150 if (group_leader
->ctx
!= ctx
)
6155 * Only a group leader can be exclusive or pinned
6157 if (attr
.exclusive
|| attr
.pinned
)
6162 err
= perf_event_set_output(event
, output_event
);
6167 event_file
= anon_inode_getfile("[perf_event]", &perf_fops
, event
, O_RDWR
);
6168 if (IS_ERR(event_file
)) {
6169 err
= PTR_ERR(event_file
);
6174 struct perf_event_context
*gctx
= group_leader
->ctx
;
6176 mutex_lock(&gctx
->mutex
);
6177 perf_remove_from_context(group_leader
);
6178 list_for_each_entry(sibling
, &group_leader
->sibling_list
,
6180 perf_remove_from_context(sibling
);
6183 mutex_unlock(&gctx
->mutex
);
6187 event
->filp
= event_file
;
6188 WARN_ON_ONCE(ctx
->parent_ctx
);
6189 mutex_lock(&ctx
->mutex
);
6192 perf_install_in_context(ctx
, group_leader
, cpu
);
6194 list_for_each_entry(sibling
, &group_leader
->sibling_list
,
6196 perf_install_in_context(ctx
, sibling
, cpu
);
6201 perf_install_in_context(ctx
, event
, cpu
);
6203 perf_unpin_context(ctx
);
6204 mutex_unlock(&ctx
->mutex
);
6206 event
->owner
= current
;
6208 mutex_lock(¤t
->perf_event_mutex
);
6209 list_add_tail(&event
->owner_entry
, ¤t
->perf_event_list
);
6210 mutex_unlock(¤t
->perf_event_mutex
);
6213 * Precalculate sample_data sizes
6215 perf_event__header_size(event
);
6216 perf_event__id_header_size(event
);
6219 * Drop the reference on the group_event after placing the
6220 * new event on the sibling_list. This ensures destruction
6221 * of the group leader will find the pointer to itself in
6222 * perf_group_detach().
6224 fput_light(group_file
, fput_needed
);
6225 fd_install(event_fd
, event_file
);
6229 perf_unpin_context(ctx
);
6235 put_task_struct(task
);
6237 fput_light(group_file
, fput_needed
);
6239 put_unused_fd(event_fd
);
6244 * perf_event_create_kernel_counter
6246 * @attr: attributes of the counter to create
6247 * @cpu: cpu in which the counter is bound
6248 * @task: task to profile (NULL for percpu)
6251 perf_event_create_kernel_counter(struct perf_event_attr
*attr
, int cpu
,
6252 struct task_struct
*task
,
6253 perf_overflow_handler_t overflow_handler
,
6256 struct perf_event_context
*ctx
;
6257 struct perf_event
*event
;
6261 * Get the target context (task or percpu):
6264 event
= perf_event_alloc(attr
, cpu
, task
, NULL
, NULL
,
6265 overflow_handler
, context
);
6266 if (IS_ERR(event
)) {
6267 err
= PTR_ERR(event
);
6271 ctx
= find_get_context(event
->pmu
, task
, cpu
);
6278 WARN_ON_ONCE(ctx
->parent_ctx
);
6279 mutex_lock(&ctx
->mutex
);
6280 perf_install_in_context(ctx
, event
, cpu
);
6282 perf_unpin_context(ctx
);
6283 mutex_unlock(&ctx
->mutex
);
6290 return ERR_PTR(err
);
6292 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter
);
6294 static void sync_child_event(struct perf_event
*child_event
,
6295 struct task_struct
*child
)
6297 struct perf_event
*parent_event
= child_event
->parent
;
6300 if (child_event
->attr
.inherit_stat
)
6301 perf_event_read_event(child_event
, child
);
6303 child_val
= perf_event_count(child_event
);
6306 * Add back the child's count to the parent's count:
6308 atomic64_add(child_val
, &parent_event
->child_count
);
6309 atomic64_add(child_event
->total_time_enabled
,
6310 &parent_event
->child_total_time_enabled
);
6311 atomic64_add(child_event
->total_time_running
,
6312 &parent_event
->child_total_time_running
);
6315 * Remove this event from the parent's list
6317 WARN_ON_ONCE(parent_event
->ctx
->parent_ctx
);
6318 mutex_lock(&parent_event
->child_mutex
);
6319 list_del_init(&child_event
->child_list
);
6320 mutex_unlock(&parent_event
->child_mutex
);
6323 * Release the parent event, if this was the last
6326 fput(parent_event
->filp
);
6330 __perf_event_exit_task(struct perf_event
*child_event
,
6331 struct perf_event_context
*child_ctx
,
6332 struct task_struct
*child
)
6334 if (child_event
->parent
) {
6335 raw_spin_lock_irq(&child_ctx
->lock
);
6336 perf_group_detach(child_event
);
6337 raw_spin_unlock_irq(&child_ctx
->lock
);
6340 perf_remove_from_context(child_event
);
6343 * It can happen that the parent exits first, and has events
6344 * that are still around due to the child reference. These
6345 * events need to be zapped.
6347 if (child_event
->parent
) {
6348 sync_child_event(child_event
, child
);
6349 free_event(child_event
);
6353 static void perf_event_exit_task_context(struct task_struct
*child
, int ctxn
)
6355 struct perf_event
*child_event
, *tmp
;
6356 struct perf_event_context
*child_ctx
;
6357 unsigned long flags
;
6359 if (likely(!child
->perf_event_ctxp
[ctxn
])) {
6360 perf_event_task(child
, NULL
, 0);
6364 local_irq_save(flags
);
6366 * We can't reschedule here because interrupts are disabled,
6367 * and either child is current or it is a task that can't be
6368 * scheduled, so we are now safe from rescheduling changing
6371 child_ctx
= rcu_dereference_raw(child
->perf_event_ctxp
[ctxn
]);
6374 * Take the context lock here so that if find_get_context is
6375 * reading child->perf_event_ctxp, we wait until it has
6376 * incremented the context's refcount before we do put_ctx below.
6378 raw_spin_lock(&child_ctx
->lock
);
6379 task_ctx_sched_out(child_ctx
);
6380 child
->perf_event_ctxp
[ctxn
] = NULL
;
6382 * If this context is a clone; unclone it so it can't get
6383 * swapped to another process while we're removing all
6384 * the events from it.
6386 unclone_ctx(child_ctx
);
6387 update_context_time(child_ctx
);
6388 raw_spin_unlock_irqrestore(&child_ctx
->lock
, flags
);
6391 * Report the task dead after unscheduling the events so that we
6392 * won't get any samples after PERF_RECORD_EXIT. We can however still
6393 * get a few PERF_RECORD_READ events.
6395 perf_event_task(child
, child_ctx
, 0);
6398 * We can recurse on the same lock type through:
6400 * __perf_event_exit_task()
6401 * sync_child_event()
6402 * fput(parent_event->filp)
6404 * mutex_lock(&ctx->mutex)
6406 * But since its the parent context it won't be the same instance.
6408 mutex_lock(&child_ctx
->mutex
);
6411 list_for_each_entry_safe(child_event
, tmp
, &child_ctx
->pinned_groups
,
6413 __perf_event_exit_task(child_event
, child_ctx
, child
);
6415 list_for_each_entry_safe(child_event
, tmp
, &child_ctx
->flexible_groups
,
6417 __perf_event_exit_task(child_event
, child_ctx
, child
);
6420 * If the last event was a group event, it will have appended all
6421 * its siblings to the list, but we obtained 'tmp' before that which
6422 * will still point to the list head terminating the iteration.
6424 if (!list_empty(&child_ctx
->pinned_groups
) ||
6425 !list_empty(&child_ctx
->flexible_groups
))
6428 mutex_unlock(&child_ctx
->mutex
);
6434 * When a child task exits, feed back event values to parent events.
6436 void perf_event_exit_task(struct task_struct
*child
)
6438 struct perf_event
*event
, *tmp
;
6441 mutex_lock(&child
->perf_event_mutex
);
6442 list_for_each_entry_safe(event
, tmp
, &child
->perf_event_list
,
6444 list_del_init(&event
->owner_entry
);
6447 * Ensure the list deletion is visible before we clear
6448 * the owner, closes a race against perf_release() where
6449 * we need to serialize on the owner->perf_event_mutex.
6452 event
->owner
= NULL
;
6454 mutex_unlock(&child
->perf_event_mutex
);
6456 for_each_task_context_nr(ctxn
)
6457 perf_event_exit_task_context(child
, ctxn
);
6460 static void perf_free_event(struct perf_event
*event
,
6461 struct perf_event_context
*ctx
)
6463 struct perf_event
*parent
= event
->parent
;
6465 if (WARN_ON_ONCE(!parent
))
6468 mutex_lock(&parent
->child_mutex
);
6469 list_del_init(&event
->child_list
);
6470 mutex_unlock(&parent
->child_mutex
);
6474 perf_group_detach(event
);
6475 list_del_event(event
, ctx
);
6480 * free an unexposed, unused context as created by inheritance by
6481 * perf_event_init_task below, used by fork() in case of fail.
6483 void perf_event_free_task(struct task_struct
*task
)
6485 struct perf_event_context
*ctx
;
6486 struct perf_event
*event
, *tmp
;
6489 for_each_task_context_nr(ctxn
) {
6490 ctx
= task
->perf_event_ctxp
[ctxn
];
6494 mutex_lock(&ctx
->mutex
);
6496 list_for_each_entry_safe(event
, tmp
, &ctx
->pinned_groups
,
6498 perf_free_event(event
, ctx
);
6500 list_for_each_entry_safe(event
, tmp
, &ctx
->flexible_groups
,
6502 perf_free_event(event
, ctx
);
6504 if (!list_empty(&ctx
->pinned_groups
) ||
6505 !list_empty(&ctx
->flexible_groups
))
6508 mutex_unlock(&ctx
->mutex
);
6514 void perf_event_delayed_put(struct task_struct
*task
)
6518 for_each_task_context_nr(ctxn
)
6519 WARN_ON_ONCE(task
->perf_event_ctxp
[ctxn
]);
6523 * inherit a event from parent task to child task:
6525 static struct perf_event
*
6526 inherit_event(struct perf_event
*parent_event
,
6527 struct task_struct
*parent
,
6528 struct perf_event_context
*parent_ctx
,
6529 struct task_struct
*child
,
6530 struct perf_event
*group_leader
,
6531 struct perf_event_context
*child_ctx
)
6533 struct perf_event
*child_event
;
6534 unsigned long flags
;
6537 * Instead of creating recursive hierarchies of events,
6538 * we link inherited events back to the original parent,
6539 * which has a filp for sure, which we use as the reference
6542 if (parent_event
->parent
)
6543 parent_event
= parent_event
->parent
;
6545 child_event
= perf_event_alloc(&parent_event
->attr
,
6548 group_leader
, parent_event
,
6550 if (IS_ERR(child_event
))
6555 * Make the child state follow the state of the parent event,
6556 * not its attr.disabled bit. We hold the parent's mutex,
6557 * so we won't race with perf_event_{en, dis}able_family.
6559 if (parent_event
->state
>= PERF_EVENT_STATE_INACTIVE
)
6560 child_event
->state
= PERF_EVENT_STATE_INACTIVE
;
6562 child_event
->state
= PERF_EVENT_STATE_OFF
;
6564 if (parent_event
->attr
.freq
) {
6565 u64 sample_period
= parent_event
->hw
.sample_period
;
6566 struct hw_perf_event
*hwc
= &child_event
->hw
;
6568 hwc
->sample_period
= sample_period
;
6569 hwc
->last_period
= sample_period
;
6571 local64_set(&hwc
->period_left
, sample_period
);
6574 child_event
->ctx
= child_ctx
;
6575 child_event
->overflow_handler
= parent_event
->overflow_handler
;
6576 child_event
->overflow_handler_context
6577 = parent_event
->overflow_handler_context
;
6580 * Precalculate sample_data sizes
6582 perf_event__header_size(child_event
);
6583 perf_event__id_header_size(child_event
);
6586 * Link it up in the child's context:
6588 raw_spin_lock_irqsave(&child_ctx
->lock
, flags
);
6589 add_event_to_ctx(child_event
, child_ctx
);
6590 raw_spin_unlock_irqrestore(&child_ctx
->lock
, flags
);
6593 * Get a reference to the parent filp - we will fput it
6594 * when the child event exits. This is safe to do because
6595 * we are in the parent and we know that the filp still
6596 * exists and has a nonzero count:
6598 atomic_long_inc(&parent_event
->filp
->f_count
);
6601 * Link this into the parent event's child list
6603 WARN_ON_ONCE(parent_event
->ctx
->parent_ctx
);
6604 mutex_lock(&parent_event
->child_mutex
);
6605 list_add_tail(&child_event
->child_list
, &parent_event
->child_list
);
6606 mutex_unlock(&parent_event
->child_mutex
);
6611 static int inherit_group(struct perf_event
*parent_event
,
6612 struct task_struct
*parent
,
6613 struct perf_event_context
*parent_ctx
,
6614 struct task_struct
*child
,
6615 struct perf_event_context
*child_ctx
)
6617 struct perf_event
*leader
;
6618 struct perf_event
*sub
;
6619 struct perf_event
*child_ctr
;
6621 leader
= inherit_event(parent_event
, parent
, parent_ctx
,
6622 child
, NULL
, child_ctx
);
6624 return PTR_ERR(leader
);
6625 list_for_each_entry(sub
, &parent_event
->sibling_list
, group_entry
) {
6626 child_ctr
= inherit_event(sub
, parent
, parent_ctx
,
6627 child
, leader
, child_ctx
);
6628 if (IS_ERR(child_ctr
))
6629 return PTR_ERR(child_ctr
);
6635 inherit_task_group(struct perf_event
*event
, struct task_struct
*parent
,
6636 struct perf_event_context
*parent_ctx
,
6637 struct task_struct
*child
, int ctxn
,
6641 struct perf_event_context
*child_ctx
;
6643 if (!event
->attr
.inherit
) {
6648 child_ctx
= child
->perf_event_ctxp
[ctxn
];
6651 * This is executed from the parent task context, so
6652 * inherit events that have been marked for cloning.
6653 * First allocate and initialize a context for the
6657 child_ctx
= alloc_perf_context(event
->pmu
, child
);
6661 child
->perf_event_ctxp
[ctxn
] = child_ctx
;
6664 ret
= inherit_group(event
, parent
, parent_ctx
,
6674 * Initialize the perf_event context in task_struct
6676 int perf_event_init_context(struct task_struct
*child
, int ctxn
)
6678 struct perf_event_context
*child_ctx
, *parent_ctx
;
6679 struct perf_event_context
*cloned_ctx
;
6680 struct perf_event
*event
;
6681 struct task_struct
*parent
= current
;
6682 int inherited_all
= 1;
6683 unsigned long flags
;
6686 if (likely(!parent
->perf_event_ctxp
[ctxn
]))
6690 * If the parent's context is a clone, pin it so it won't get
6693 parent_ctx
= perf_pin_task_context(parent
, ctxn
);
6696 * No need to check if parent_ctx != NULL here; since we saw
6697 * it non-NULL earlier, the only reason for it to become NULL
6698 * is if we exit, and since we're currently in the middle of
6699 * a fork we can't be exiting at the same time.
6703 * Lock the parent list. No need to lock the child - not PID
6704 * hashed yet and not running, so nobody can access it.
6706 mutex_lock(&parent_ctx
->mutex
);
6709 * We dont have to disable NMIs - we are only looking at
6710 * the list, not manipulating it:
6712 list_for_each_entry(event
, &parent_ctx
->pinned_groups
, group_entry
) {
6713 ret
= inherit_task_group(event
, parent
, parent_ctx
,
6714 child
, ctxn
, &inherited_all
);
6720 * We can't hold ctx->lock when iterating the ->flexible_group list due
6721 * to allocations, but we need to prevent rotation because
6722 * rotate_ctx() will change the list from interrupt context.
6724 raw_spin_lock_irqsave(&parent_ctx
->lock
, flags
);
6725 parent_ctx
->rotate_disable
= 1;
6726 raw_spin_unlock_irqrestore(&parent_ctx
->lock
, flags
);
6728 list_for_each_entry(event
, &parent_ctx
->flexible_groups
, group_entry
) {
6729 ret
= inherit_task_group(event
, parent
, parent_ctx
,
6730 child
, ctxn
, &inherited_all
);
6735 raw_spin_lock_irqsave(&parent_ctx
->lock
, flags
);
6736 parent_ctx
->rotate_disable
= 0;
6738 child_ctx
= child
->perf_event_ctxp
[ctxn
];
6740 if (child_ctx
&& inherited_all
) {
6742 * Mark the child context as a clone of the parent
6743 * context, or of whatever the parent is a clone of.
6745 * Note that if the parent is a clone, the holding of
6746 * parent_ctx->lock avoids it from being uncloned.
6748 cloned_ctx
= parent_ctx
->parent_ctx
;
6750 child_ctx
->parent_ctx
= cloned_ctx
;
6751 child_ctx
->parent_gen
= parent_ctx
->parent_gen
;
6753 child_ctx
->parent_ctx
= parent_ctx
;
6754 child_ctx
->parent_gen
= parent_ctx
->generation
;
6756 get_ctx(child_ctx
->parent_ctx
);
6759 raw_spin_unlock_irqrestore(&parent_ctx
->lock
, flags
);
6760 mutex_unlock(&parent_ctx
->mutex
);
6762 perf_unpin_context(parent_ctx
);
6763 put_ctx(parent_ctx
);
6769 * Initialize the perf_event context in task_struct
6771 int perf_event_init_task(struct task_struct
*child
)
6775 memset(child
->perf_event_ctxp
, 0, sizeof(child
->perf_event_ctxp
));
6776 mutex_init(&child
->perf_event_mutex
);
6777 INIT_LIST_HEAD(&child
->perf_event_list
);
6779 for_each_task_context_nr(ctxn
) {
6780 ret
= perf_event_init_context(child
, ctxn
);
6788 static void __init
perf_event_init_all_cpus(void)
6790 struct swevent_htable
*swhash
;
6793 for_each_possible_cpu(cpu
) {
6794 swhash
= &per_cpu(swevent_htable
, cpu
);
6795 mutex_init(&swhash
->hlist_mutex
);
6796 INIT_LIST_HEAD(&per_cpu(rotation_list
, cpu
));
6800 static void __cpuinit
perf_event_init_cpu(int cpu
)
6802 struct swevent_htable
*swhash
= &per_cpu(swevent_htable
, cpu
);
6804 mutex_lock(&swhash
->hlist_mutex
);
6805 if (swhash
->hlist_refcount
> 0) {
6806 struct swevent_hlist
*hlist
;
6808 hlist
= kzalloc_node(sizeof(*hlist
), GFP_KERNEL
, cpu_to_node(cpu
));
6810 rcu_assign_pointer(swhash
->swevent_hlist
, hlist
);
6812 mutex_unlock(&swhash
->hlist_mutex
);
6815 #if defined CONFIG_HOTPLUG_CPU || defined CONFIG_KEXEC
6816 static void perf_pmu_rotate_stop(struct pmu
*pmu
)
6818 struct perf_cpu_context
*cpuctx
= this_cpu_ptr(pmu
->pmu_cpu_context
);
6820 WARN_ON(!irqs_disabled());
6822 list_del_init(&cpuctx
->rotation_list
);
6825 static void __perf_event_exit_context(void *__info
)
6827 struct perf_event_context
*ctx
= __info
;
6828 struct perf_event
*event
, *tmp
;
6830 perf_pmu_rotate_stop(ctx
->pmu
);
6832 list_for_each_entry_safe(event
, tmp
, &ctx
->pinned_groups
, group_entry
)
6833 __perf_remove_from_context(event
);
6834 list_for_each_entry_safe(event
, tmp
, &ctx
->flexible_groups
, group_entry
)
6835 __perf_remove_from_context(event
);
6838 static void perf_event_exit_cpu_context(int cpu
)
6840 struct perf_event_context
*ctx
;
6844 idx
= srcu_read_lock(&pmus_srcu
);
6845 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
6846 ctx
= &per_cpu_ptr(pmu
->pmu_cpu_context
, cpu
)->ctx
;
6848 mutex_lock(&ctx
->mutex
);
6849 smp_call_function_single(cpu
, __perf_event_exit_context
, ctx
, 1);
6850 mutex_unlock(&ctx
->mutex
);
6852 srcu_read_unlock(&pmus_srcu
, idx
);
6855 static void perf_event_exit_cpu(int cpu
)
6857 struct swevent_htable
*swhash
= &per_cpu(swevent_htable
, cpu
);
6859 mutex_lock(&swhash
->hlist_mutex
);
6860 swevent_hlist_release(swhash
);
6861 mutex_unlock(&swhash
->hlist_mutex
);
6863 perf_event_exit_cpu_context(cpu
);
6866 static inline void perf_event_exit_cpu(int cpu
) { }
6870 perf_reboot(struct notifier_block
*notifier
, unsigned long val
, void *v
)
6874 for_each_online_cpu(cpu
)
6875 perf_event_exit_cpu(cpu
);
6881 * Run the perf reboot notifier at the very last possible moment so that
6882 * the generic watchdog code runs as long as possible.
6884 static struct notifier_block perf_reboot_notifier
= {
6885 .notifier_call
= perf_reboot
,
6886 .priority
= INT_MIN
,
6889 static int __cpuinit
6890 perf_cpu_notify(struct notifier_block
*self
, unsigned long action
, void *hcpu
)
6892 unsigned int cpu
= (long)hcpu
;
6894 switch (action
& ~CPU_TASKS_FROZEN
) {
6896 case CPU_UP_PREPARE
:
6897 case CPU_DOWN_FAILED
:
6898 perf_event_init_cpu(cpu
);
6901 case CPU_UP_CANCELED
:
6902 case CPU_DOWN_PREPARE
:
6903 perf_event_exit_cpu(cpu
);
6913 void __init
perf_event_init(void)
6919 perf_event_init_all_cpus();
6920 init_srcu_struct(&pmus_srcu
);
6921 perf_pmu_register(&perf_swevent
, "software", PERF_TYPE_SOFTWARE
);
6922 perf_pmu_register(&perf_cpu_clock
, NULL
, -1);
6923 perf_pmu_register(&perf_task_clock
, NULL
, -1);
6925 perf_cpu_notifier(perf_cpu_notify
);
6926 register_reboot_notifier(&perf_reboot_notifier
);
6928 ret
= init_hw_breakpoint();
6929 WARN(ret
, "hw_breakpoint initialization failed with: %d", ret
);
6931 /* do not patch jump label more than once per second */
6932 jump_label_rate_limit(&perf_sched_events
, HZ
);
6935 static int __init
perf_event_sysfs_init(void)
6940 mutex_lock(&pmus_lock
);
6942 ret
= bus_register(&pmu_bus
);
6946 list_for_each_entry(pmu
, &pmus
, entry
) {
6947 if (!pmu
->name
|| pmu
->type
< 0)
6950 ret
= pmu_dev_alloc(pmu
);
6951 WARN(ret
, "Failed to register pmu: %s, reason %d\n", pmu
->name
, ret
);
6953 pmu_bus_running
= 1;
6957 mutex_unlock(&pmus_lock
);
6961 device_initcall(perf_event_sysfs_init
);
6963 #ifdef CONFIG_CGROUP_PERF
6964 static struct cgroup_subsys_state
*perf_cgroup_create(
6965 struct cgroup_subsys
*ss
, struct cgroup
*cont
)
6967 struct perf_cgroup
*jc
;
6969 jc
= kzalloc(sizeof(*jc
), GFP_KERNEL
);
6971 return ERR_PTR(-ENOMEM
);
6973 jc
->info
= alloc_percpu(struct perf_cgroup_info
);
6976 return ERR_PTR(-ENOMEM
);
6982 static void perf_cgroup_destroy(struct cgroup_subsys
*ss
,
6983 struct cgroup
*cont
)
6985 struct perf_cgroup
*jc
;
6986 jc
= container_of(cgroup_subsys_state(cont
, perf_subsys_id
),
6987 struct perf_cgroup
, css
);
6988 free_percpu(jc
->info
);
6992 static int __perf_cgroup_move(void *info
)
6994 struct task_struct
*task
= info
;
6995 perf_cgroup_switch(task
, PERF_CGROUP_SWOUT
| PERF_CGROUP_SWIN
);
6999 static void perf_cgroup_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
7000 struct cgroup_taskset
*tset
)
7002 struct task_struct
*task
;
7004 cgroup_taskset_for_each(task
, cgrp
, tset
)
7005 task_function_call(task
, __perf_cgroup_move
, task
);
7008 static void perf_cgroup_exit(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
7009 struct cgroup
*old_cgrp
, struct task_struct
*task
)
7012 * cgroup_exit() is called in the copy_process() failure path.
7013 * Ignore this case since the task hasn't ran yet, this avoids
7014 * trying to poke a half freed task state from generic code.
7016 if (!(task
->flags
& PF_EXITING
))
7019 task_function_call(task
, __perf_cgroup_move
, task
);
7022 struct cgroup_subsys perf_subsys
= {
7023 .name
= "perf_event",
7024 .subsys_id
= perf_subsys_id
,
7025 .create
= perf_cgroup_create
,
7026 .destroy
= perf_cgroup_destroy
,
7027 .exit
= perf_cgroup_exit
,
7028 .attach
= perf_cgroup_attach
,
7030 #endif /* CONFIG_CGROUP_PERF */