2 * Performance events core code:
4 * Copyright (C) 2008 Thomas Gleixner <tglx@linutronix.de>
5 * Copyright (C) 2008-2009 Red Hat, Inc., Ingo Molnar
6 * Copyright (C) 2008-2009 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/vmalloc.h>
29 #include <linux/hardirq.h>
30 #include <linux/rculist.h>
31 #include <linux/uaccess.h>
32 #include <linux/syscalls.h>
33 #include <linux/anon_inodes.h>
34 #include <linux/kernel_stat.h>
35 #include <linux/perf_event.h>
36 #include <linux/ftrace_event.h>
37 #include <linux/hw_breakpoint.h>
39 #include <asm/irq_regs.h>
41 struct remote_function_call
{
42 struct task_struct
*p
;
43 int (*func
)(void *info
);
48 static void remote_function(void *data
)
50 struct remote_function_call
*tfc
= data
;
51 struct task_struct
*p
= tfc
->p
;
55 if (task_cpu(p
) != smp_processor_id() || !task_curr(p
))
59 tfc
->ret
= tfc
->func(tfc
->info
);
63 * task_function_call - call a function on the cpu on which a task runs
64 * @p: the task to evaluate
65 * @func: the function to be called
66 * @info: the function call argument
68 * Calls the function @func when the task is currently running. This might
69 * be on the current CPU, which just calls the function directly
71 * returns: @func return value, or
72 * -ESRCH - when the process isn't running
73 * -EAGAIN - when the process moved away
76 task_function_call(struct task_struct
*p
, int (*func
) (void *info
), void *info
)
78 struct remote_function_call data
= {
82 .ret
= -ESRCH
, /* No such (running) process */
86 smp_call_function_single(task_cpu(p
), remote_function
, &data
, 1);
92 * cpu_function_call - call a function on the cpu
93 * @func: the function to be called
94 * @info: the function call argument
96 * Calls the function @func on the remote cpu.
98 * returns: @func return value or -ENXIO when the cpu is offline
100 static int cpu_function_call(int cpu
, int (*func
) (void *info
), void *info
)
102 struct remote_function_call data
= {
106 .ret
= -ENXIO
, /* No such CPU */
109 smp_call_function_single(cpu
, remote_function
, &data
, 1);
114 #define PERF_FLAG_ALL (PERF_FLAG_FD_NO_GROUP |\
115 PERF_FLAG_FD_OUTPUT |\
116 PERF_FLAG_PID_CGROUP)
119 EVENT_FLEXIBLE
= 0x1,
121 EVENT_ALL
= EVENT_FLEXIBLE
| EVENT_PINNED
,
125 * perf_sched_events : >0 events exist
126 * perf_cgroup_events: >0 per-cpu cgroup events exist on this cpu
128 struct jump_label_key perf_sched_events __read_mostly
;
129 static DEFINE_PER_CPU(atomic_t
, perf_cgroup_events
);
131 static atomic_t nr_mmap_events __read_mostly
;
132 static atomic_t nr_comm_events __read_mostly
;
133 static atomic_t nr_task_events __read_mostly
;
135 static LIST_HEAD(pmus
);
136 static DEFINE_MUTEX(pmus_lock
);
137 static struct srcu_struct pmus_srcu
;
140 * perf event paranoia level:
141 * -1 - not paranoid at all
142 * 0 - disallow raw tracepoint access for unpriv
143 * 1 - disallow cpu events for unpriv
144 * 2 - disallow kernel profiling for unpriv
146 int sysctl_perf_event_paranoid __read_mostly
= 1;
148 /* Minimum for 512 kiB + 1 user control page */
149 int sysctl_perf_event_mlock __read_mostly
= 512 + (PAGE_SIZE
/ 1024); /* 'free' kiB per user */
152 * max perf event sample rate
154 #define DEFAULT_MAX_SAMPLE_RATE 100000
155 int sysctl_perf_event_sample_rate __read_mostly
= DEFAULT_MAX_SAMPLE_RATE
;
156 static int max_samples_per_tick __read_mostly
=
157 DIV_ROUND_UP(DEFAULT_MAX_SAMPLE_RATE
, HZ
);
159 int perf_proc_update_handler(struct ctl_table
*table
, int write
,
160 void __user
*buffer
, size_t *lenp
,
163 int ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
168 max_samples_per_tick
= DIV_ROUND_UP(sysctl_perf_event_sample_rate
, HZ
);
173 static atomic64_t perf_event_id
;
175 static void cpu_ctx_sched_out(struct perf_cpu_context
*cpuctx
,
176 enum event_type_t event_type
);
178 static void cpu_ctx_sched_in(struct perf_cpu_context
*cpuctx
,
179 enum event_type_t event_type
,
180 struct task_struct
*task
);
182 static void update_context_time(struct perf_event_context
*ctx
);
183 static u64
perf_event_time(struct perf_event
*event
);
185 void __weak
perf_event_print_debug(void) { }
187 extern __weak
const char *perf_pmu_name(void)
192 static inline u64
perf_clock(void)
194 return local_clock();
197 static inline struct perf_cpu_context
*
198 __get_cpu_context(struct perf_event_context
*ctx
)
200 return this_cpu_ptr(ctx
->pmu
->pmu_cpu_context
);
203 #ifdef CONFIG_CGROUP_PERF
206 * Must ensure cgroup is pinned (css_get) before calling
207 * this function. In other words, we cannot call this function
208 * if there is no cgroup event for the current CPU context.
210 static inline struct perf_cgroup
*
211 perf_cgroup_from_task(struct task_struct
*task
)
213 return container_of(task_subsys_state(task
, perf_subsys_id
),
214 struct perf_cgroup
, css
);
218 perf_cgroup_match(struct perf_event
*event
)
220 struct perf_event_context
*ctx
= event
->ctx
;
221 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
223 return !event
->cgrp
|| event
->cgrp
== cpuctx
->cgrp
;
226 static inline void perf_get_cgroup(struct perf_event
*event
)
228 css_get(&event
->cgrp
->css
);
231 static inline void perf_put_cgroup(struct perf_event
*event
)
233 css_put(&event
->cgrp
->css
);
236 static inline void perf_detach_cgroup(struct perf_event
*event
)
238 perf_put_cgroup(event
);
242 static inline int is_cgroup_event(struct perf_event
*event
)
244 return event
->cgrp
!= NULL
;
247 static inline u64
perf_cgroup_event_time(struct perf_event
*event
)
249 struct perf_cgroup_info
*t
;
251 t
= per_cpu_ptr(event
->cgrp
->info
, event
->cpu
);
255 static inline void __update_cgrp_time(struct perf_cgroup
*cgrp
)
257 struct perf_cgroup_info
*info
;
262 info
= this_cpu_ptr(cgrp
->info
);
264 info
->time
+= now
- info
->timestamp
;
265 info
->timestamp
= now
;
268 static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context
*cpuctx
)
270 struct perf_cgroup
*cgrp_out
= cpuctx
->cgrp
;
272 __update_cgrp_time(cgrp_out
);
275 static inline void update_cgrp_time_from_event(struct perf_event
*event
)
277 struct perf_cgroup
*cgrp
;
280 * ensure we access cgroup data only when needed and
281 * when we know the cgroup is pinned (css_get)
283 if (!is_cgroup_event(event
))
286 cgrp
= perf_cgroup_from_task(current
);
288 * Do not update time when cgroup is not active
290 if (cgrp
== event
->cgrp
)
291 __update_cgrp_time(event
->cgrp
);
295 perf_cgroup_set_timestamp(struct task_struct
*task
,
296 struct perf_event_context
*ctx
)
298 struct perf_cgroup
*cgrp
;
299 struct perf_cgroup_info
*info
;
302 * ctx->lock held by caller
303 * ensure we do not access cgroup data
304 * unless we have the cgroup pinned (css_get)
306 if (!task
|| !ctx
->nr_cgroups
)
309 cgrp
= perf_cgroup_from_task(task
);
310 info
= this_cpu_ptr(cgrp
->info
);
311 info
->timestamp
= ctx
->timestamp
;
314 #define PERF_CGROUP_SWOUT 0x1 /* cgroup switch out every event */
315 #define PERF_CGROUP_SWIN 0x2 /* cgroup switch in events based on task */
318 * reschedule events based on the cgroup constraint of task.
320 * mode SWOUT : schedule out everything
321 * mode SWIN : schedule in based on cgroup for next
323 void perf_cgroup_switch(struct task_struct
*task
, int mode
)
325 struct perf_cpu_context
*cpuctx
;
330 * disable interrupts to avoid geting nr_cgroup
331 * changes via __perf_event_disable(). Also
334 local_irq_save(flags
);
337 * we reschedule only in the presence of cgroup
338 * constrained events.
342 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
344 cpuctx
= this_cpu_ptr(pmu
->pmu_cpu_context
);
346 perf_pmu_disable(cpuctx
->ctx
.pmu
);
349 * perf_cgroup_events says at least one
350 * context on this CPU has cgroup events.
352 * ctx->nr_cgroups reports the number of cgroup
353 * events for a context.
355 if (cpuctx
->ctx
.nr_cgroups
> 0) {
357 if (mode
& PERF_CGROUP_SWOUT
) {
358 cpu_ctx_sched_out(cpuctx
, EVENT_ALL
);
360 * must not be done before ctxswout due
361 * to event_filter_match() in event_sched_out()
366 if (mode
& PERF_CGROUP_SWIN
) {
367 WARN_ON_ONCE(cpuctx
->cgrp
);
368 /* set cgrp before ctxsw in to
369 * allow event_filter_match() to not
370 * have to pass task around
372 cpuctx
->cgrp
= perf_cgroup_from_task(task
);
373 cpu_ctx_sched_in(cpuctx
, EVENT_ALL
, task
);
377 perf_pmu_enable(cpuctx
->ctx
.pmu
);
382 local_irq_restore(flags
);
385 static inline void perf_cgroup_sched_out(struct task_struct
*task
)
387 perf_cgroup_switch(task
, PERF_CGROUP_SWOUT
);
390 static inline void perf_cgroup_sched_in(struct task_struct
*task
)
392 perf_cgroup_switch(task
, PERF_CGROUP_SWIN
);
395 static inline int perf_cgroup_connect(int fd
, struct perf_event
*event
,
396 struct perf_event_attr
*attr
,
397 struct perf_event
*group_leader
)
399 struct perf_cgroup
*cgrp
;
400 struct cgroup_subsys_state
*css
;
402 int ret
= 0, fput_needed
;
404 file
= fget_light(fd
, &fput_needed
);
408 css
= cgroup_css_from_dir(file
, perf_subsys_id
);
414 cgrp
= container_of(css
, struct perf_cgroup
, css
);
417 /* must be done before we fput() the file */
418 perf_get_cgroup(event
);
421 * all events in a group must monitor
422 * the same cgroup because a task belongs
423 * to only one perf cgroup at a time
425 if (group_leader
&& group_leader
->cgrp
!= cgrp
) {
426 perf_detach_cgroup(event
);
430 fput_light(file
, fput_needed
);
435 perf_cgroup_set_shadow_time(struct perf_event
*event
, u64 now
)
437 struct perf_cgroup_info
*t
;
438 t
= per_cpu_ptr(event
->cgrp
->info
, event
->cpu
);
439 event
->shadow_ctx_time
= now
- t
->timestamp
;
443 perf_cgroup_defer_enabled(struct perf_event
*event
)
446 * when the current task's perf cgroup does not match
447 * the event's, we need to remember to call the
448 * perf_mark_enable() function the first time a task with
449 * a matching perf cgroup is scheduled in.
451 if (is_cgroup_event(event
) && !perf_cgroup_match(event
))
452 event
->cgrp_defer_enabled
= 1;
456 perf_cgroup_mark_enabled(struct perf_event
*event
,
457 struct perf_event_context
*ctx
)
459 struct perf_event
*sub
;
460 u64 tstamp
= perf_event_time(event
);
462 if (!event
->cgrp_defer_enabled
)
465 event
->cgrp_defer_enabled
= 0;
467 event
->tstamp_enabled
= tstamp
- event
->total_time_enabled
;
468 list_for_each_entry(sub
, &event
->sibling_list
, group_entry
) {
469 if (sub
->state
>= PERF_EVENT_STATE_INACTIVE
) {
470 sub
->tstamp_enabled
= tstamp
- sub
->total_time_enabled
;
471 sub
->cgrp_defer_enabled
= 0;
475 #else /* !CONFIG_CGROUP_PERF */
478 perf_cgroup_match(struct perf_event
*event
)
483 static inline void perf_detach_cgroup(struct perf_event
*event
)
486 static inline int is_cgroup_event(struct perf_event
*event
)
491 static inline u64
perf_cgroup_event_cgrp_time(struct perf_event
*event
)
496 static inline void update_cgrp_time_from_event(struct perf_event
*event
)
500 static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context
*cpuctx
)
504 static inline void perf_cgroup_sched_out(struct task_struct
*task
)
508 static inline void perf_cgroup_sched_in(struct task_struct
*task
)
512 static inline int perf_cgroup_connect(pid_t pid
, struct perf_event
*event
,
513 struct perf_event_attr
*attr
,
514 struct perf_event
*group_leader
)
520 perf_cgroup_set_timestamp(struct task_struct
*task
,
521 struct perf_event_context
*ctx
)
526 perf_cgroup_switch(struct task_struct
*task
, struct task_struct
*next
)
531 perf_cgroup_set_shadow_time(struct perf_event
*event
, u64 now
)
535 static inline u64
perf_cgroup_event_time(struct perf_event
*event
)
541 perf_cgroup_defer_enabled(struct perf_event
*event
)
546 perf_cgroup_mark_enabled(struct perf_event
*event
,
547 struct perf_event_context
*ctx
)
552 void perf_pmu_disable(struct pmu
*pmu
)
554 int *count
= this_cpu_ptr(pmu
->pmu_disable_count
);
556 pmu
->pmu_disable(pmu
);
559 void perf_pmu_enable(struct pmu
*pmu
)
561 int *count
= this_cpu_ptr(pmu
->pmu_disable_count
);
563 pmu
->pmu_enable(pmu
);
566 static DEFINE_PER_CPU(struct list_head
, rotation_list
);
569 * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
570 * because they're strictly cpu affine and rotate_start is called with IRQs
571 * disabled, while rotate_context is called from IRQ context.
573 static void perf_pmu_rotate_start(struct pmu
*pmu
)
575 struct perf_cpu_context
*cpuctx
= this_cpu_ptr(pmu
->pmu_cpu_context
);
576 struct list_head
*head
= &__get_cpu_var(rotation_list
);
578 WARN_ON(!irqs_disabled());
580 if (list_empty(&cpuctx
->rotation_list
))
581 list_add(&cpuctx
->rotation_list
, head
);
584 static void get_ctx(struct perf_event_context
*ctx
)
586 WARN_ON(!atomic_inc_not_zero(&ctx
->refcount
));
589 static void free_ctx(struct rcu_head
*head
)
591 struct perf_event_context
*ctx
;
593 ctx
= container_of(head
, struct perf_event_context
, rcu_head
);
597 static void put_ctx(struct perf_event_context
*ctx
)
599 if (atomic_dec_and_test(&ctx
->refcount
)) {
601 put_ctx(ctx
->parent_ctx
);
603 put_task_struct(ctx
->task
);
604 call_rcu(&ctx
->rcu_head
, free_ctx
);
608 static void unclone_ctx(struct perf_event_context
*ctx
)
610 if (ctx
->parent_ctx
) {
611 put_ctx(ctx
->parent_ctx
);
612 ctx
->parent_ctx
= NULL
;
616 static u32
perf_event_pid(struct perf_event
*event
, struct task_struct
*p
)
619 * only top level events have the pid namespace they were created in
622 event
= event
->parent
;
624 return task_tgid_nr_ns(p
, event
->ns
);
627 static u32
perf_event_tid(struct perf_event
*event
, struct task_struct
*p
)
630 * only top level events have the pid namespace they were created in
633 event
= event
->parent
;
635 return task_pid_nr_ns(p
, event
->ns
);
639 * If we inherit events we want to return the parent event id
642 static u64
primary_event_id(struct perf_event
*event
)
647 id
= event
->parent
->id
;
653 * Get the perf_event_context for a task and lock it.
654 * This has to cope with with the fact that until it is locked,
655 * the context could get moved to another task.
657 static struct perf_event_context
*
658 perf_lock_task_context(struct task_struct
*task
, int ctxn
, unsigned long *flags
)
660 struct perf_event_context
*ctx
;
664 ctx
= rcu_dereference(task
->perf_event_ctxp
[ctxn
]);
667 * If this context is a clone of another, it might
668 * get swapped for another underneath us by
669 * perf_event_task_sched_out, though the
670 * rcu_read_lock() protects us from any context
671 * getting freed. Lock the context and check if it
672 * got swapped before we could get the lock, and retry
673 * if so. If we locked the right context, then it
674 * can't get swapped on us any more.
676 raw_spin_lock_irqsave(&ctx
->lock
, *flags
);
677 if (ctx
!= rcu_dereference(task
->perf_event_ctxp
[ctxn
])) {
678 raw_spin_unlock_irqrestore(&ctx
->lock
, *flags
);
682 if (!atomic_inc_not_zero(&ctx
->refcount
)) {
683 raw_spin_unlock_irqrestore(&ctx
->lock
, *flags
);
692 * Get the context for a task and increment its pin_count so it
693 * can't get swapped to another task. This also increments its
694 * reference count so that the context can't get freed.
696 static struct perf_event_context
*
697 perf_pin_task_context(struct task_struct
*task
, int ctxn
)
699 struct perf_event_context
*ctx
;
702 ctx
= perf_lock_task_context(task
, ctxn
, &flags
);
705 raw_spin_unlock_irqrestore(&ctx
->lock
, flags
);
710 static void perf_unpin_context(struct perf_event_context
*ctx
)
714 raw_spin_lock_irqsave(&ctx
->lock
, flags
);
716 raw_spin_unlock_irqrestore(&ctx
->lock
, flags
);
720 * Update the record of the current time in a context.
722 static void update_context_time(struct perf_event_context
*ctx
)
724 u64 now
= perf_clock();
726 ctx
->time
+= now
- ctx
->timestamp
;
727 ctx
->timestamp
= now
;
730 static u64
perf_event_time(struct perf_event
*event
)
732 struct perf_event_context
*ctx
= event
->ctx
;
734 if (is_cgroup_event(event
))
735 return perf_cgroup_event_time(event
);
737 return ctx
? ctx
->time
: 0;
741 * Update the total_time_enabled and total_time_running fields for a event.
743 static void update_event_times(struct perf_event
*event
)
745 struct perf_event_context
*ctx
= event
->ctx
;
748 if (event
->state
< PERF_EVENT_STATE_INACTIVE
||
749 event
->group_leader
->state
< PERF_EVENT_STATE_INACTIVE
)
752 * in cgroup mode, time_enabled represents
753 * the time the event was enabled AND active
754 * tasks were in the monitored cgroup. This is
755 * independent of the activity of the context as
756 * there may be a mix of cgroup and non-cgroup events.
758 * That is why we treat cgroup events differently
761 if (is_cgroup_event(event
))
762 run_end
= perf_event_time(event
);
763 else if (ctx
->is_active
)
766 run_end
= event
->tstamp_stopped
;
768 event
->total_time_enabled
= run_end
- event
->tstamp_enabled
;
770 if (event
->state
== PERF_EVENT_STATE_INACTIVE
)
771 run_end
= event
->tstamp_stopped
;
773 run_end
= perf_event_time(event
);
775 event
->total_time_running
= run_end
- event
->tstamp_running
;
780 * Update total_time_enabled and total_time_running for all events in a group.
782 static void update_group_times(struct perf_event
*leader
)
784 struct perf_event
*event
;
786 update_event_times(leader
);
787 list_for_each_entry(event
, &leader
->sibling_list
, group_entry
)
788 update_event_times(event
);
791 static struct list_head
*
792 ctx_group_list(struct perf_event
*event
, struct perf_event_context
*ctx
)
794 if (event
->attr
.pinned
)
795 return &ctx
->pinned_groups
;
797 return &ctx
->flexible_groups
;
801 * Add a event from the lists for its context.
802 * Must be called with ctx->mutex and ctx->lock held.
805 list_add_event(struct perf_event
*event
, struct perf_event_context
*ctx
)
807 WARN_ON_ONCE(event
->attach_state
& PERF_ATTACH_CONTEXT
);
808 event
->attach_state
|= PERF_ATTACH_CONTEXT
;
811 * If we're a stand alone event or group leader, we go to the context
812 * list, group events are kept attached to the group so that
813 * perf_group_detach can, at all times, locate all siblings.
815 if (event
->group_leader
== event
) {
816 struct list_head
*list
;
818 if (is_software_event(event
))
819 event
->group_flags
|= PERF_GROUP_SOFTWARE
;
821 list
= ctx_group_list(event
, ctx
);
822 list_add_tail(&event
->group_entry
, list
);
825 if (is_cgroup_event(event
))
828 list_add_rcu(&event
->event_entry
, &ctx
->event_list
);
830 perf_pmu_rotate_start(ctx
->pmu
);
832 if (event
->attr
.inherit_stat
)
837 * Called at perf_event creation and when events are attached/detached from a
840 static void perf_event__read_size(struct perf_event
*event
)
842 int entry
= sizeof(u64
); /* value */
846 if (event
->attr
.read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
849 if (event
->attr
.read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
852 if (event
->attr
.read_format
& PERF_FORMAT_ID
)
853 entry
+= sizeof(u64
);
855 if (event
->attr
.read_format
& PERF_FORMAT_GROUP
) {
856 nr
+= event
->group_leader
->nr_siblings
;
861 event
->read_size
= size
;
864 static void perf_event__header_size(struct perf_event
*event
)
866 struct perf_sample_data
*data
;
867 u64 sample_type
= event
->attr
.sample_type
;
870 perf_event__read_size(event
);
872 if (sample_type
& PERF_SAMPLE_IP
)
873 size
+= sizeof(data
->ip
);
875 if (sample_type
& PERF_SAMPLE_ADDR
)
876 size
+= sizeof(data
->addr
);
878 if (sample_type
& PERF_SAMPLE_PERIOD
)
879 size
+= sizeof(data
->period
);
881 if (sample_type
& PERF_SAMPLE_READ
)
882 size
+= event
->read_size
;
884 event
->header_size
= size
;
887 static void perf_event__id_header_size(struct perf_event
*event
)
889 struct perf_sample_data
*data
;
890 u64 sample_type
= event
->attr
.sample_type
;
893 if (sample_type
& PERF_SAMPLE_TID
)
894 size
+= sizeof(data
->tid_entry
);
896 if (sample_type
& PERF_SAMPLE_TIME
)
897 size
+= sizeof(data
->time
);
899 if (sample_type
& PERF_SAMPLE_ID
)
900 size
+= sizeof(data
->id
);
902 if (sample_type
& PERF_SAMPLE_STREAM_ID
)
903 size
+= sizeof(data
->stream_id
);
905 if (sample_type
& PERF_SAMPLE_CPU
)
906 size
+= sizeof(data
->cpu_entry
);
908 event
->id_header_size
= size
;
911 static void perf_group_attach(struct perf_event
*event
)
913 struct perf_event
*group_leader
= event
->group_leader
, *pos
;
916 * We can have double attach due to group movement in perf_event_open.
918 if (event
->attach_state
& PERF_ATTACH_GROUP
)
921 event
->attach_state
|= PERF_ATTACH_GROUP
;
923 if (group_leader
== event
)
926 if (group_leader
->group_flags
& PERF_GROUP_SOFTWARE
&&
927 !is_software_event(event
))
928 group_leader
->group_flags
&= ~PERF_GROUP_SOFTWARE
;
930 list_add_tail(&event
->group_entry
, &group_leader
->sibling_list
);
931 group_leader
->nr_siblings
++;
933 perf_event__header_size(group_leader
);
935 list_for_each_entry(pos
, &group_leader
->sibling_list
, group_entry
)
936 perf_event__header_size(pos
);
940 * Remove a event from the lists for its context.
941 * Must be called with ctx->mutex and ctx->lock held.
944 list_del_event(struct perf_event
*event
, struct perf_event_context
*ctx
)
946 struct perf_cpu_context
*cpuctx
;
948 * We can have double detach due to exit/hot-unplug + close.
950 if (!(event
->attach_state
& PERF_ATTACH_CONTEXT
))
953 event
->attach_state
&= ~PERF_ATTACH_CONTEXT
;
955 if (is_cgroup_event(event
)) {
957 cpuctx
= __get_cpu_context(ctx
);
959 * if there are no more cgroup events
960 * then cler cgrp to avoid stale pointer
961 * in update_cgrp_time_from_cpuctx()
963 if (!ctx
->nr_cgroups
)
968 if (event
->attr
.inherit_stat
)
971 list_del_rcu(&event
->event_entry
);
973 if (event
->group_leader
== event
)
974 list_del_init(&event
->group_entry
);
976 update_group_times(event
);
979 * If event was in error state, then keep it
980 * that way, otherwise bogus counts will be
981 * returned on read(). The only way to get out
982 * of error state is by explicit re-enabling
985 if (event
->state
> PERF_EVENT_STATE_OFF
)
986 event
->state
= PERF_EVENT_STATE_OFF
;
989 static void perf_group_detach(struct perf_event
*event
)
991 struct perf_event
*sibling
, *tmp
;
992 struct list_head
*list
= NULL
;
995 * We can have double detach due to exit/hot-unplug + close.
997 if (!(event
->attach_state
& PERF_ATTACH_GROUP
))
1000 event
->attach_state
&= ~PERF_ATTACH_GROUP
;
1003 * If this is a sibling, remove it from its group.
1005 if (event
->group_leader
!= event
) {
1006 list_del_init(&event
->group_entry
);
1007 event
->group_leader
->nr_siblings
--;
1011 if (!list_empty(&event
->group_entry
))
1012 list
= &event
->group_entry
;
1015 * If this was a group event with sibling events then
1016 * upgrade the siblings to singleton events by adding them
1017 * to whatever list we are on.
1019 list_for_each_entry_safe(sibling
, tmp
, &event
->sibling_list
, group_entry
) {
1021 list_move_tail(&sibling
->group_entry
, list
);
1022 sibling
->group_leader
= sibling
;
1024 /* Inherit group flags from the previous leader */
1025 sibling
->group_flags
= event
->group_flags
;
1029 perf_event__header_size(event
->group_leader
);
1031 list_for_each_entry(tmp
, &event
->group_leader
->sibling_list
, group_entry
)
1032 perf_event__header_size(tmp
);
1036 event_filter_match(struct perf_event
*event
)
1038 return (event
->cpu
== -1 || event
->cpu
== smp_processor_id())
1039 && perf_cgroup_match(event
);
1043 event_sched_out(struct perf_event
*event
,
1044 struct perf_cpu_context
*cpuctx
,
1045 struct perf_event_context
*ctx
)
1047 u64 tstamp
= perf_event_time(event
);
1050 * An event which could not be activated because of
1051 * filter mismatch still needs to have its timings
1052 * maintained, otherwise bogus information is return
1053 * via read() for time_enabled, time_running:
1055 if (event
->state
== PERF_EVENT_STATE_INACTIVE
1056 && !event_filter_match(event
)) {
1057 delta
= tstamp
- event
->tstamp_stopped
;
1058 event
->tstamp_running
+= delta
;
1059 event
->tstamp_stopped
= tstamp
;
1062 if (event
->state
!= PERF_EVENT_STATE_ACTIVE
)
1065 event
->state
= PERF_EVENT_STATE_INACTIVE
;
1066 if (event
->pending_disable
) {
1067 event
->pending_disable
= 0;
1068 event
->state
= PERF_EVENT_STATE_OFF
;
1070 event
->tstamp_stopped
= tstamp
;
1071 event
->pmu
->del(event
, 0);
1074 if (!is_software_event(event
))
1075 cpuctx
->active_oncpu
--;
1077 if (event
->attr
.exclusive
|| !cpuctx
->active_oncpu
)
1078 cpuctx
->exclusive
= 0;
1082 group_sched_out(struct perf_event
*group_event
,
1083 struct perf_cpu_context
*cpuctx
,
1084 struct perf_event_context
*ctx
)
1086 struct perf_event
*event
;
1087 int state
= group_event
->state
;
1089 event_sched_out(group_event
, cpuctx
, ctx
);
1092 * Schedule out siblings (if any):
1094 list_for_each_entry(event
, &group_event
->sibling_list
, group_entry
)
1095 event_sched_out(event
, cpuctx
, ctx
);
1097 if (state
== PERF_EVENT_STATE_ACTIVE
&& group_event
->attr
.exclusive
)
1098 cpuctx
->exclusive
= 0;
1102 * Cross CPU call to remove a performance event
1104 * We disable the event on the hardware level first. After that we
1105 * remove it from the context list.
1107 static int __perf_remove_from_context(void *info
)
1109 struct perf_event
*event
= info
;
1110 struct perf_event_context
*ctx
= event
->ctx
;
1111 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
1113 raw_spin_lock(&ctx
->lock
);
1114 event_sched_out(event
, cpuctx
, ctx
);
1115 list_del_event(event
, ctx
);
1116 raw_spin_unlock(&ctx
->lock
);
1123 * Remove the event from a task's (or a CPU's) list of events.
1125 * CPU events are removed with a smp call. For task events we only
1126 * call when the task is on a CPU.
1128 * If event->ctx is a cloned context, callers must make sure that
1129 * every task struct that event->ctx->task could possibly point to
1130 * remains valid. This is OK when called from perf_release since
1131 * that only calls us on the top-level context, which can't be a clone.
1132 * When called from perf_event_exit_task, it's OK because the
1133 * context has been detached from its task.
1135 static void perf_remove_from_context(struct perf_event
*event
)
1137 struct perf_event_context
*ctx
= event
->ctx
;
1138 struct task_struct
*task
= ctx
->task
;
1140 lockdep_assert_held(&ctx
->mutex
);
1144 * Per cpu events are removed via an smp call and
1145 * the removal is always successful.
1147 cpu_function_call(event
->cpu
, __perf_remove_from_context
, event
);
1152 if (!task_function_call(task
, __perf_remove_from_context
, event
))
1155 raw_spin_lock_irq(&ctx
->lock
);
1157 * If we failed to find a running task, but find the context active now
1158 * that we've acquired the ctx->lock, retry.
1160 if (ctx
->is_active
) {
1161 raw_spin_unlock_irq(&ctx
->lock
);
1166 * Since the task isn't running, its safe to remove the event, us
1167 * holding the ctx->lock ensures the task won't get scheduled in.
1169 list_del_event(event
, ctx
);
1170 raw_spin_unlock_irq(&ctx
->lock
);
1174 * Cross CPU call to disable a performance event
1176 static int __perf_event_disable(void *info
)
1178 struct perf_event
*event
= info
;
1179 struct perf_event_context
*ctx
= event
->ctx
;
1180 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
1183 * If this is a per-task event, need to check whether this
1184 * event's task is the current task on this cpu.
1186 * Can trigger due to concurrent perf_event_context_sched_out()
1187 * flipping contexts around.
1189 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
)
1192 raw_spin_lock(&ctx
->lock
);
1195 * If the event is on, turn it off.
1196 * If it is in error state, leave it in error state.
1198 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
) {
1199 update_context_time(ctx
);
1200 update_cgrp_time_from_event(event
);
1201 update_group_times(event
);
1202 if (event
== event
->group_leader
)
1203 group_sched_out(event
, cpuctx
, ctx
);
1205 event_sched_out(event
, cpuctx
, ctx
);
1206 event
->state
= PERF_EVENT_STATE_OFF
;
1209 raw_spin_unlock(&ctx
->lock
);
1217 * If event->ctx is a cloned context, callers must make sure that
1218 * every task struct that event->ctx->task could possibly point to
1219 * remains valid. This condition is satisifed when called through
1220 * perf_event_for_each_child or perf_event_for_each because they
1221 * hold the top-level event's child_mutex, so any descendant that
1222 * goes to exit will block in sync_child_event.
1223 * When called from perf_pending_event it's OK because event->ctx
1224 * is the current context on this CPU and preemption is disabled,
1225 * hence we can't get into perf_event_task_sched_out for this context.
1227 void perf_event_disable(struct perf_event
*event
)
1229 struct perf_event_context
*ctx
= event
->ctx
;
1230 struct task_struct
*task
= ctx
->task
;
1234 * Disable the event on the cpu that it's on
1236 cpu_function_call(event
->cpu
, __perf_event_disable
, event
);
1241 if (!task_function_call(task
, __perf_event_disable
, event
))
1244 raw_spin_lock_irq(&ctx
->lock
);
1246 * If the event is still active, we need to retry the cross-call.
1248 if (event
->state
== PERF_EVENT_STATE_ACTIVE
) {
1249 raw_spin_unlock_irq(&ctx
->lock
);
1251 * Reload the task pointer, it might have been changed by
1252 * a concurrent perf_event_context_sched_out().
1259 * Since we have the lock this context can't be scheduled
1260 * in, so we can change the state safely.
1262 if (event
->state
== PERF_EVENT_STATE_INACTIVE
) {
1263 update_group_times(event
);
1264 event
->state
= PERF_EVENT_STATE_OFF
;
1266 raw_spin_unlock_irq(&ctx
->lock
);
1269 static void perf_set_shadow_time(struct perf_event
*event
,
1270 struct perf_event_context
*ctx
,
1274 * use the correct time source for the time snapshot
1276 * We could get by without this by leveraging the
1277 * fact that to get to this function, the caller
1278 * has most likely already called update_context_time()
1279 * and update_cgrp_time_xx() and thus both timestamp
1280 * are identical (or very close). Given that tstamp is,
1281 * already adjusted for cgroup, we could say that:
1282 * tstamp - ctx->timestamp
1284 * tstamp - cgrp->timestamp.
1286 * Then, in perf_output_read(), the calculation would
1287 * work with no changes because:
1288 * - event is guaranteed scheduled in
1289 * - no scheduled out in between
1290 * - thus the timestamp would be the same
1292 * But this is a bit hairy.
1294 * So instead, we have an explicit cgroup call to remain
1295 * within the time time source all along. We believe it
1296 * is cleaner and simpler to understand.
1298 if (is_cgroup_event(event
))
1299 perf_cgroup_set_shadow_time(event
, tstamp
);
1301 event
->shadow_ctx_time
= tstamp
- ctx
->timestamp
;
1304 #define MAX_INTERRUPTS (~0ULL)
1306 static void perf_log_throttle(struct perf_event
*event
, int enable
);
1309 event_sched_in(struct perf_event
*event
,
1310 struct perf_cpu_context
*cpuctx
,
1311 struct perf_event_context
*ctx
)
1313 u64 tstamp
= perf_event_time(event
);
1315 if (event
->state
<= PERF_EVENT_STATE_OFF
)
1318 event
->state
= PERF_EVENT_STATE_ACTIVE
;
1319 event
->oncpu
= smp_processor_id();
1322 * Unthrottle events, since we scheduled we might have missed several
1323 * ticks already, also for a heavily scheduling task there is little
1324 * guarantee it'll get a tick in a timely manner.
1326 if (unlikely(event
->hw
.interrupts
== MAX_INTERRUPTS
)) {
1327 perf_log_throttle(event
, 1);
1328 event
->hw
.interrupts
= 0;
1332 * The new state must be visible before we turn it on in the hardware:
1336 if (event
->pmu
->add(event
, PERF_EF_START
)) {
1337 event
->state
= PERF_EVENT_STATE_INACTIVE
;
1342 event
->tstamp_running
+= tstamp
- event
->tstamp_stopped
;
1344 perf_set_shadow_time(event
, ctx
, tstamp
);
1346 if (!is_software_event(event
))
1347 cpuctx
->active_oncpu
++;
1350 if (event
->attr
.exclusive
)
1351 cpuctx
->exclusive
= 1;
1357 group_sched_in(struct perf_event
*group_event
,
1358 struct perf_cpu_context
*cpuctx
,
1359 struct perf_event_context
*ctx
)
1361 struct perf_event
*event
, *partial_group
= NULL
;
1362 struct pmu
*pmu
= group_event
->pmu
;
1363 u64 now
= ctx
->time
;
1364 bool simulate
= false;
1366 if (group_event
->state
== PERF_EVENT_STATE_OFF
)
1369 pmu
->start_txn(pmu
);
1371 if (event_sched_in(group_event
, cpuctx
, ctx
)) {
1372 pmu
->cancel_txn(pmu
);
1377 * Schedule in siblings as one group (if any):
1379 list_for_each_entry(event
, &group_event
->sibling_list
, group_entry
) {
1380 if (event_sched_in(event
, cpuctx
, ctx
)) {
1381 partial_group
= event
;
1386 if (!pmu
->commit_txn(pmu
))
1391 * Groups can be scheduled in as one unit only, so undo any
1392 * partial group before returning:
1393 * The events up to the failed event are scheduled out normally,
1394 * tstamp_stopped will be updated.
1396 * The failed events and the remaining siblings need to have
1397 * their timings updated as if they had gone thru event_sched_in()
1398 * and event_sched_out(). This is required to get consistent timings
1399 * across the group. This also takes care of the case where the group
1400 * could never be scheduled by ensuring tstamp_stopped is set to mark
1401 * the time the event was actually stopped, such that time delta
1402 * calculation in update_event_times() is correct.
1404 list_for_each_entry(event
, &group_event
->sibling_list
, group_entry
) {
1405 if (event
== partial_group
)
1409 event
->tstamp_running
+= now
- event
->tstamp_stopped
;
1410 event
->tstamp_stopped
= now
;
1412 event_sched_out(event
, cpuctx
, ctx
);
1415 event_sched_out(group_event
, cpuctx
, ctx
);
1417 pmu
->cancel_txn(pmu
);
1423 * Work out whether we can put this event group on the CPU now.
1425 static int group_can_go_on(struct perf_event
*event
,
1426 struct perf_cpu_context
*cpuctx
,
1430 * Groups consisting entirely of software events can always go on.
1432 if (event
->group_flags
& PERF_GROUP_SOFTWARE
)
1435 * If an exclusive group is already on, no other hardware
1438 if (cpuctx
->exclusive
)
1441 * If this group is exclusive and there are already
1442 * events on the CPU, it can't go on.
1444 if (event
->attr
.exclusive
&& cpuctx
->active_oncpu
)
1447 * Otherwise, try to add it if all previous groups were able
1453 static void add_event_to_ctx(struct perf_event
*event
,
1454 struct perf_event_context
*ctx
)
1456 u64 tstamp
= perf_event_time(event
);
1458 list_add_event(event
, ctx
);
1459 perf_group_attach(event
);
1460 event
->tstamp_enabled
= tstamp
;
1461 event
->tstamp_running
= tstamp
;
1462 event
->tstamp_stopped
= tstamp
;
1465 static void perf_event_context_sched_in(struct perf_event_context
*ctx
,
1466 struct task_struct
*tsk
);
1469 * Cross CPU call to install and enable a performance event
1471 * Must be called with ctx->mutex held
1473 static int __perf_install_in_context(void *info
)
1475 struct perf_event
*event
= info
;
1476 struct perf_event_context
*ctx
= event
->ctx
;
1477 struct perf_event
*leader
= event
->group_leader
;
1478 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
1482 * In case we're installing a new context to an already running task,
1483 * could also happen before perf_event_task_sched_in() on architectures
1484 * which do context switches with IRQs enabled.
1486 if (ctx
->task
&& !cpuctx
->task_ctx
)
1487 perf_event_context_sched_in(ctx
, ctx
->task
);
1489 raw_spin_lock(&ctx
->lock
);
1491 update_context_time(ctx
);
1493 * update cgrp time only if current cgrp
1494 * matches event->cgrp. Must be done before
1495 * calling add_event_to_ctx()
1497 update_cgrp_time_from_event(event
);
1499 add_event_to_ctx(event
, ctx
);
1501 if (!event_filter_match(event
))
1505 * Don't put the event on if it is disabled or if
1506 * it is in a group and the group isn't on.
1508 if (event
->state
!= PERF_EVENT_STATE_INACTIVE
||
1509 (leader
!= event
&& leader
->state
!= PERF_EVENT_STATE_ACTIVE
))
1513 * An exclusive event can't go on if there are already active
1514 * hardware events, and no hardware event can go on if there
1515 * is already an exclusive event on.
1517 if (!group_can_go_on(event
, cpuctx
, 1))
1520 err
= event_sched_in(event
, cpuctx
, ctx
);
1524 * This event couldn't go on. If it is in a group
1525 * then we have to pull the whole group off.
1526 * If the event group is pinned then put it in error state.
1528 if (leader
!= event
)
1529 group_sched_out(leader
, cpuctx
, ctx
);
1530 if (leader
->attr
.pinned
) {
1531 update_group_times(leader
);
1532 leader
->state
= PERF_EVENT_STATE_ERROR
;
1537 raw_spin_unlock(&ctx
->lock
);
1543 * Attach a performance event to a context
1545 * First we add the event to the list with the hardware enable bit
1546 * in event->hw_config cleared.
1548 * If the event is attached to a task which is on a CPU we use a smp
1549 * call to enable it in the task context. The task might have been
1550 * scheduled away, but we check this in the smp call again.
1553 perf_install_in_context(struct perf_event_context
*ctx
,
1554 struct perf_event
*event
,
1557 struct task_struct
*task
= ctx
->task
;
1559 lockdep_assert_held(&ctx
->mutex
);
1565 * Per cpu events are installed via an smp call and
1566 * the install is always successful.
1568 cpu_function_call(cpu
, __perf_install_in_context
, event
);
1573 if (!task_function_call(task
, __perf_install_in_context
, event
))
1576 raw_spin_lock_irq(&ctx
->lock
);
1578 * If we failed to find a running task, but find the context active now
1579 * that we've acquired the ctx->lock, retry.
1581 if (ctx
->is_active
) {
1582 raw_spin_unlock_irq(&ctx
->lock
);
1587 * Since the task isn't running, its safe to add the event, us holding
1588 * the ctx->lock ensures the task won't get scheduled in.
1590 add_event_to_ctx(event
, ctx
);
1591 raw_spin_unlock_irq(&ctx
->lock
);
1595 * Put a event into inactive state and update time fields.
1596 * Enabling the leader of a group effectively enables all
1597 * the group members that aren't explicitly disabled, so we
1598 * have to update their ->tstamp_enabled also.
1599 * Note: this works for group members as well as group leaders
1600 * since the non-leader members' sibling_lists will be empty.
1602 static void __perf_event_mark_enabled(struct perf_event
*event
,
1603 struct perf_event_context
*ctx
)
1605 struct perf_event
*sub
;
1606 u64 tstamp
= perf_event_time(event
);
1608 event
->state
= PERF_EVENT_STATE_INACTIVE
;
1609 event
->tstamp_enabled
= tstamp
- event
->total_time_enabled
;
1610 list_for_each_entry(sub
, &event
->sibling_list
, group_entry
) {
1611 if (sub
->state
>= PERF_EVENT_STATE_INACTIVE
)
1612 sub
->tstamp_enabled
= tstamp
- sub
->total_time_enabled
;
1617 * Cross CPU call to enable a performance event
1619 static int __perf_event_enable(void *info
)
1621 struct perf_event
*event
= info
;
1622 struct perf_event_context
*ctx
= event
->ctx
;
1623 struct perf_event
*leader
= event
->group_leader
;
1624 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
1627 if (WARN_ON_ONCE(!ctx
->is_active
))
1630 raw_spin_lock(&ctx
->lock
);
1631 update_context_time(ctx
);
1633 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
)
1637 * set current task's cgroup time reference point
1639 perf_cgroup_set_timestamp(current
, ctx
);
1641 __perf_event_mark_enabled(event
, ctx
);
1643 if (!event_filter_match(event
)) {
1644 if (is_cgroup_event(event
))
1645 perf_cgroup_defer_enabled(event
);
1650 * If the event is in a group and isn't the group leader,
1651 * then don't put it on unless the group is on.
1653 if (leader
!= event
&& leader
->state
!= PERF_EVENT_STATE_ACTIVE
)
1656 if (!group_can_go_on(event
, cpuctx
, 1)) {
1659 if (event
== leader
)
1660 err
= group_sched_in(event
, cpuctx
, ctx
);
1662 err
= event_sched_in(event
, cpuctx
, ctx
);
1667 * If this event can't go on and it's part of a
1668 * group, then the whole group has to come off.
1670 if (leader
!= event
)
1671 group_sched_out(leader
, cpuctx
, ctx
);
1672 if (leader
->attr
.pinned
) {
1673 update_group_times(leader
);
1674 leader
->state
= PERF_EVENT_STATE_ERROR
;
1679 raw_spin_unlock(&ctx
->lock
);
1687 * If event->ctx is a cloned context, callers must make sure that
1688 * every task struct that event->ctx->task could possibly point to
1689 * remains valid. This condition is satisfied when called through
1690 * perf_event_for_each_child or perf_event_for_each as described
1691 * for perf_event_disable.
1693 void perf_event_enable(struct perf_event
*event
)
1695 struct perf_event_context
*ctx
= event
->ctx
;
1696 struct task_struct
*task
= ctx
->task
;
1700 * Enable the event on the cpu that it's on
1702 cpu_function_call(event
->cpu
, __perf_event_enable
, event
);
1706 raw_spin_lock_irq(&ctx
->lock
);
1707 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
)
1711 * If the event is in error state, clear that first.
1712 * That way, if we see the event in error state below, we
1713 * know that it has gone back into error state, as distinct
1714 * from the task having been scheduled away before the
1715 * cross-call arrived.
1717 if (event
->state
== PERF_EVENT_STATE_ERROR
)
1718 event
->state
= PERF_EVENT_STATE_OFF
;
1721 if (!ctx
->is_active
) {
1722 __perf_event_mark_enabled(event
, ctx
);
1726 raw_spin_unlock_irq(&ctx
->lock
);
1728 if (!task_function_call(task
, __perf_event_enable
, event
))
1731 raw_spin_lock_irq(&ctx
->lock
);
1734 * If the context is active and the event is still off,
1735 * we need to retry the cross-call.
1737 if (ctx
->is_active
&& event
->state
== PERF_EVENT_STATE_OFF
) {
1739 * task could have been flipped by a concurrent
1740 * perf_event_context_sched_out()
1747 raw_spin_unlock_irq(&ctx
->lock
);
1750 static int perf_event_refresh(struct perf_event
*event
, int refresh
)
1753 * not supported on inherited events
1755 if (event
->attr
.inherit
|| !is_sampling_event(event
))
1758 atomic_add(refresh
, &event
->event_limit
);
1759 perf_event_enable(event
);
1764 static void ctx_sched_out(struct perf_event_context
*ctx
,
1765 struct perf_cpu_context
*cpuctx
,
1766 enum event_type_t event_type
)
1768 struct perf_event
*event
;
1770 raw_spin_lock(&ctx
->lock
);
1771 perf_pmu_disable(ctx
->pmu
);
1773 if (likely(!ctx
->nr_events
))
1775 update_context_time(ctx
);
1776 update_cgrp_time_from_cpuctx(cpuctx
);
1778 if (!ctx
->nr_active
)
1781 if (event_type
& EVENT_PINNED
) {
1782 list_for_each_entry(event
, &ctx
->pinned_groups
, group_entry
)
1783 group_sched_out(event
, cpuctx
, ctx
);
1786 if (event_type
& EVENT_FLEXIBLE
) {
1787 list_for_each_entry(event
, &ctx
->flexible_groups
, group_entry
)
1788 group_sched_out(event
, cpuctx
, ctx
);
1791 perf_pmu_enable(ctx
->pmu
);
1792 raw_spin_unlock(&ctx
->lock
);
1796 * Test whether two contexts are equivalent, i.e. whether they
1797 * have both been cloned from the same version of the same context
1798 * and they both have the same number of enabled events.
1799 * If the number of enabled events is the same, then the set
1800 * of enabled events should be the same, because these are both
1801 * inherited contexts, therefore we can't access individual events
1802 * in them directly with an fd; we can only enable/disable all
1803 * events via prctl, or enable/disable all events in a family
1804 * via ioctl, which will have the same effect on both contexts.
1806 static int context_equiv(struct perf_event_context
*ctx1
,
1807 struct perf_event_context
*ctx2
)
1809 return ctx1
->parent_ctx
&& ctx1
->parent_ctx
== ctx2
->parent_ctx
1810 && ctx1
->parent_gen
== ctx2
->parent_gen
1811 && !ctx1
->pin_count
&& !ctx2
->pin_count
;
1814 static void __perf_event_sync_stat(struct perf_event
*event
,
1815 struct perf_event
*next_event
)
1819 if (!event
->attr
.inherit_stat
)
1823 * Update the event value, we cannot use perf_event_read()
1824 * because we're in the middle of a context switch and have IRQs
1825 * disabled, which upsets smp_call_function_single(), however
1826 * we know the event must be on the current CPU, therefore we
1827 * don't need to use it.
1829 switch (event
->state
) {
1830 case PERF_EVENT_STATE_ACTIVE
:
1831 event
->pmu
->read(event
);
1834 case PERF_EVENT_STATE_INACTIVE
:
1835 update_event_times(event
);
1843 * In order to keep per-task stats reliable we need to flip the event
1844 * values when we flip the contexts.
1846 value
= local64_read(&next_event
->count
);
1847 value
= local64_xchg(&event
->count
, value
);
1848 local64_set(&next_event
->count
, value
);
1850 swap(event
->total_time_enabled
, next_event
->total_time_enabled
);
1851 swap(event
->total_time_running
, next_event
->total_time_running
);
1854 * Since we swizzled the values, update the user visible data too.
1856 perf_event_update_userpage(event
);
1857 perf_event_update_userpage(next_event
);
1860 #define list_next_entry(pos, member) \
1861 list_entry(pos->member.next, typeof(*pos), member)
1863 static void perf_event_sync_stat(struct perf_event_context
*ctx
,
1864 struct perf_event_context
*next_ctx
)
1866 struct perf_event
*event
, *next_event
;
1871 update_context_time(ctx
);
1873 event
= list_first_entry(&ctx
->event_list
,
1874 struct perf_event
, event_entry
);
1876 next_event
= list_first_entry(&next_ctx
->event_list
,
1877 struct perf_event
, event_entry
);
1879 while (&event
->event_entry
!= &ctx
->event_list
&&
1880 &next_event
->event_entry
!= &next_ctx
->event_list
) {
1882 __perf_event_sync_stat(event
, next_event
);
1884 event
= list_next_entry(event
, event_entry
);
1885 next_event
= list_next_entry(next_event
, event_entry
);
1889 static void perf_event_context_sched_out(struct task_struct
*task
, int ctxn
,
1890 struct task_struct
*next
)
1892 struct perf_event_context
*ctx
= task
->perf_event_ctxp
[ctxn
];
1893 struct perf_event_context
*next_ctx
;
1894 struct perf_event_context
*parent
;
1895 struct perf_cpu_context
*cpuctx
;
1901 cpuctx
= __get_cpu_context(ctx
);
1902 if (!cpuctx
->task_ctx
)
1906 parent
= rcu_dereference(ctx
->parent_ctx
);
1907 next_ctx
= next
->perf_event_ctxp
[ctxn
];
1908 if (parent
&& next_ctx
&&
1909 rcu_dereference(next_ctx
->parent_ctx
) == parent
) {
1911 * Looks like the two contexts are clones, so we might be
1912 * able to optimize the context switch. We lock both
1913 * contexts and check that they are clones under the
1914 * lock (including re-checking that neither has been
1915 * uncloned in the meantime). It doesn't matter which
1916 * order we take the locks because no other cpu could
1917 * be trying to lock both of these tasks.
1919 raw_spin_lock(&ctx
->lock
);
1920 raw_spin_lock_nested(&next_ctx
->lock
, SINGLE_DEPTH_NESTING
);
1921 if (context_equiv(ctx
, next_ctx
)) {
1923 * XXX do we need a memory barrier of sorts
1924 * wrt to rcu_dereference() of perf_event_ctxp
1926 task
->perf_event_ctxp
[ctxn
] = next_ctx
;
1927 next
->perf_event_ctxp
[ctxn
] = ctx
;
1929 next_ctx
->task
= task
;
1932 perf_event_sync_stat(ctx
, next_ctx
);
1934 raw_spin_unlock(&next_ctx
->lock
);
1935 raw_spin_unlock(&ctx
->lock
);
1940 ctx_sched_out(ctx
, cpuctx
, EVENT_ALL
);
1941 cpuctx
->task_ctx
= NULL
;
1945 #define for_each_task_context_nr(ctxn) \
1946 for ((ctxn) = 0; (ctxn) < perf_nr_task_contexts; (ctxn)++)
1949 * Called from scheduler to remove the events of the current task,
1950 * with interrupts disabled.
1952 * We stop each event and update the event value in event->count.
1954 * This does not protect us against NMI, but disable()
1955 * sets the disabled bit in the control field of event _before_
1956 * accessing the event control register. If a NMI hits, then it will
1957 * not restart the event.
1959 void __perf_event_task_sched_out(struct task_struct
*task
,
1960 struct task_struct
*next
)
1964 for_each_task_context_nr(ctxn
)
1965 perf_event_context_sched_out(task
, ctxn
, next
);
1968 * if cgroup events exist on this CPU, then we need
1969 * to check if we have to switch out PMU state.
1970 * cgroup event are system-wide mode only
1972 if (atomic_read(&__get_cpu_var(perf_cgroup_events
)))
1973 perf_cgroup_sched_out(task
);
1976 static void task_ctx_sched_out(struct perf_event_context
*ctx
,
1977 enum event_type_t event_type
)
1979 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
1981 if (!cpuctx
->task_ctx
)
1984 if (WARN_ON_ONCE(ctx
!= cpuctx
->task_ctx
))
1987 ctx_sched_out(ctx
, cpuctx
, event_type
);
1988 cpuctx
->task_ctx
= NULL
;
1992 * Called with IRQs disabled
1994 static void cpu_ctx_sched_out(struct perf_cpu_context
*cpuctx
,
1995 enum event_type_t event_type
)
1997 ctx_sched_out(&cpuctx
->ctx
, cpuctx
, event_type
);
2001 ctx_pinned_sched_in(struct perf_event_context
*ctx
,
2002 struct perf_cpu_context
*cpuctx
)
2004 struct perf_event
*event
;
2006 list_for_each_entry(event
, &ctx
->pinned_groups
, group_entry
) {
2007 if (event
->state
<= PERF_EVENT_STATE_OFF
)
2009 if (!event_filter_match(event
))
2012 /* may need to reset tstamp_enabled */
2013 if (is_cgroup_event(event
))
2014 perf_cgroup_mark_enabled(event
, ctx
);
2016 if (group_can_go_on(event
, cpuctx
, 1))
2017 group_sched_in(event
, cpuctx
, ctx
);
2020 * If this pinned group hasn't been scheduled,
2021 * put it in error state.
2023 if (event
->state
== PERF_EVENT_STATE_INACTIVE
) {
2024 update_group_times(event
);
2025 event
->state
= PERF_EVENT_STATE_ERROR
;
2031 ctx_flexible_sched_in(struct perf_event_context
*ctx
,
2032 struct perf_cpu_context
*cpuctx
)
2034 struct perf_event
*event
;
2037 list_for_each_entry(event
, &ctx
->flexible_groups
, group_entry
) {
2038 /* Ignore events in OFF or ERROR state */
2039 if (event
->state
<= PERF_EVENT_STATE_OFF
)
2042 * Listen to the 'cpu' scheduling filter constraint
2045 if (!event_filter_match(event
))
2048 /* may need to reset tstamp_enabled */
2049 if (is_cgroup_event(event
))
2050 perf_cgroup_mark_enabled(event
, ctx
);
2052 if (group_can_go_on(event
, cpuctx
, can_add_hw
)) {
2053 if (group_sched_in(event
, cpuctx
, ctx
))
2060 ctx_sched_in(struct perf_event_context
*ctx
,
2061 struct perf_cpu_context
*cpuctx
,
2062 enum event_type_t event_type
,
2063 struct task_struct
*task
)
2067 raw_spin_lock(&ctx
->lock
);
2069 if (likely(!ctx
->nr_events
))
2073 ctx
->timestamp
= now
;
2074 perf_cgroup_set_timestamp(task
, ctx
);
2076 * First go through the list and put on any pinned groups
2077 * in order to give them the best chance of going on.
2079 if (event_type
& EVENT_PINNED
)
2080 ctx_pinned_sched_in(ctx
, cpuctx
);
2082 /* Then walk through the lower prio flexible groups */
2083 if (event_type
& EVENT_FLEXIBLE
)
2084 ctx_flexible_sched_in(ctx
, cpuctx
);
2087 raw_spin_unlock(&ctx
->lock
);
2090 static void cpu_ctx_sched_in(struct perf_cpu_context
*cpuctx
,
2091 enum event_type_t event_type
,
2092 struct task_struct
*task
)
2094 struct perf_event_context
*ctx
= &cpuctx
->ctx
;
2096 ctx_sched_in(ctx
, cpuctx
, event_type
, task
);
2099 static void task_ctx_sched_in(struct perf_event_context
*ctx
,
2100 enum event_type_t event_type
)
2102 struct perf_cpu_context
*cpuctx
;
2104 cpuctx
= __get_cpu_context(ctx
);
2105 if (cpuctx
->task_ctx
== ctx
)
2108 ctx_sched_in(ctx
, cpuctx
, event_type
, NULL
);
2109 cpuctx
->task_ctx
= ctx
;
2112 static void perf_event_context_sched_in(struct perf_event_context
*ctx
,
2113 struct task_struct
*task
)
2115 struct perf_cpu_context
*cpuctx
;
2117 cpuctx
= __get_cpu_context(ctx
);
2118 if (cpuctx
->task_ctx
== ctx
)
2121 perf_pmu_disable(ctx
->pmu
);
2123 * We want to keep the following priority order:
2124 * cpu pinned (that don't need to move), task pinned,
2125 * cpu flexible, task flexible.
2127 cpu_ctx_sched_out(cpuctx
, EVENT_FLEXIBLE
);
2129 ctx_sched_in(ctx
, cpuctx
, EVENT_PINNED
, task
);
2130 cpu_ctx_sched_in(cpuctx
, EVENT_FLEXIBLE
, task
);
2131 ctx_sched_in(ctx
, cpuctx
, EVENT_FLEXIBLE
, task
);
2133 cpuctx
->task_ctx
= ctx
;
2136 * Since these rotations are per-cpu, we need to ensure the
2137 * cpu-context we got scheduled on is actually rotating.
2139 perf_pmu_rotate_start(ctx
->pmu
);
2140 perf_pmu_enable(ctx
->pmu
);
2144 * Called from scheduler to add the events of the current task
2145 * with interrupts disabled.
2147 * We restore the event value and then enable it.
2149 * This does not protect us against NMI, but enable()
2150 * sets the enabled bit in the control field of event _before_
2151 * accessing the event control register. If a NMI hits, then it will
2152 * keep the event running.
2154 void __perf_event_task_sched_in(struct task_struct
*task
)
2156 struct perf_event_context
*ctx
;
2159 for_each_task_context_nr(ctxn
) {
2160 ctx
= task
->perf_event_ctxp
[ctxn
];
2164 perf_event_context_sched_in(ctx
, task
);
2167 * if cgroup events exist on this CPU, then we need
2168 * to check if we have to switch in PMU state.
2169 * cgroup event are system-wide mode only
2171 if (atomic_read(&__get_cpu_var(perf_cgroup_events
)))
2172 perf_cgroup_sched_in(task
);
2175 static u64
perf_calculate_period(struct perf_event
*event
, u64 nsec
, u64 count
)
2177 u64 frequency
= event
->attr
.sample_freq
;
2178 u64 sec
= NSEC_PER_SEC
;
2179 u64 divisor
, dividend
;
2181 int count_fls
, nsec_fls
, frequency_fls
, sec_fls
;
2183 count_fls
= fls64(count
);
2184 nsec_fls
= fls64(nsec
);
2185 frequency_fls
= fls64(frequency
);
2189 * We got @count in @nsec, with a target of sample_freq HZ
2190 * the target period becomes:
2193 * period = -------------------
2194 * @nsec * sample_freq
2199 * Reduce accuracy by one bit such that @a and @b converge
2200 * to a similar magnitude.
2202 #define REDUCE_FLS(a, b) \
2204 if (a##_fls > b##_fls) { \
2214 * Reduce accuracy until either term fits in a u64, then proceed with
2215 * the other, so that finally we can do a u64/u64 division.
2217 while (count_fls
+ sec_fls
> 64 && nsec_fls
+ frequency_fls
> 64) {
2218 REDUCE_FLS(nsec
, frequency
);
2219 REDUCE_FLS(sec
, count
);
2222 if (count_fls
+ sec_fls
> 64) {
2223 divisor
= nsec
* frequency
;
2225 while (count_fls
+ sec_fls
> 64) {
2226 REDUCE_FLS(count
, sec
);
2230 dividend
= count
* sec
;
2232 dividend
= count
* sec
;
2234 while (nsec_fls
+ frequency_fls
> 64) {
2235 REDUCE_FLS(nsec
, frequency
);
2239 divisor
= nsec
* frequency
;
2245 return div64_u64(dividend
, divisor
);
2248 static void perf_adjust_period(struct perf_event
*event
, u64 nsec
, u64 count
)
2250 struct hw_perf_event
*hwc
= &event
->hw
;
2251 s64 period
, sample_period
;
2254 period
= perf_calculate_period(event
, nsec
, count
);
2256 delta
= (s64
)(period
- hwc
->sample_period
);
2257 delta
= (delta
+ 7) / 8; /* low pass filter */
2259 sample_period
= hwc
->sample_period
+ delta
;
2264 hwc
->sample_period
= sample_period
;
2266 if (local64_read(&hwc
->period_left
) > 8*sample_period
) {
2267 event
->pmu
->stop(event
, PERF_EF_UPDATE
);
2268 local64_set(&hwc
->period_left
, 0);
2269 event
->pmu
->start(event
, PERF_EF_RELOAD
);
2273 static void perf_ctx_adjust_freq(struct perf_event_context
*ctx
, u64 period
)
2275 struct perf_event
*event
;
2276 struct hw_perf_event
*hwc
;
2277 u64 interrupts
, now
;
2280 raw_spin_lock(&ctx
->lock
);
2281 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
2282 if (event
->state
!= PERF_EVENT_STATE_ACTIVE
)
2285 if (!event_filter_match(event
))
2290 interrupts
= hwc
->interrupts
;
2291 hwc
->interrupts
= 0;
2294 * unthrottle events on the tick
2296 if (interrupts
== MAX_INTERRUPTS
) {
2297 perf_log_throttle(event
, 1);
2298 event
->pmu
->start(event
, 0);
2301 if (!event
->attr
.freq
|| !event
->attr
.sample_freq
)
2304 event
->pmu
->read(event
);
2305 now
= local64_read(&event
->count
);
2306 delta
= now
- hwc
->freq_count_stamp
;
2307 hwc
->freq_count_stamp
= now
;
2310 perf_adjust_period(event
, period
, delta
);
2312 raw_spin_unlock(&ctx
->lock
);
2316 * Round-robin a context's events:
2318 static void rotate_ctx(struct perf_event_context
*ctx
)
2320 raw_spin_lock(&ctx
->lock
);
2323 * Rotate the first entry last of non-pinned groups. Rotation might be
2324 * disabled by the inheritance code.
2326 if (!ctx
->rotate_disable
)
2327 list_rotate_left(&ctx
->flexible_groups
);
2329 raw_spin_unlock(&ctx
->lock
);
2333 * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
2334 * because they're strictly cpu affine and rotate_start is called with IRQs
2335 * disabled, while rotate_context is called from IRQ context.
2337 static void perf_rotate_context(struct perf_cpu_context
*cpuctx
)
2339 u64 interval
= (u64
)cpuctx
->jiffies_interval
* TICK_NSEC
;
2340 struct perf_event_context
*ctx
= NULL
;
2341 int rotate
= 0, remove
= 1;
2343 if (cpuctx
->ctx
.nr_events
) {
2345 if (cpuctx
->ctx
.nr_events
!= cpuctx
->ctx
.nr_active
)
2349 ctx
= cpuctx
->task_ctx
;
2350 if (ctx
&& ctx
->nr_events
) {
2352 if (ctx
->nr_events
!= ctx
->nr_active
)
2356 perf_pmu_disable(cpuctx
->ctx
.pmu
);
2357 perf_ctx_adjust_freq(&cpuctx
->ctx
, interval
);
2359 perf_ctx_adjust_freq(ctx
, interval
);
2364 cpu_ctx_sched_out(cpuctx
, EVENT_FLEXIBLE
);
2366 task_ctx_sched_out(ctx
, EVENT_FLEXIBLE
);
2368 rotate_ctx(&cpuctx
->ctx
);
2372 cpu_ctx_sched_in(cpuctx
, EVENT_FLEXIBLE
, current
);
2374 task_ctx_sched_in(ctx
, EVENT_FLEXIBLE
);
2378 list_del_init(&cpuctx
->rotation_list
);
2380 perf_pmu_enable(cpuctx
->ctx
.pmu
);
2383 void perf_event_task_tick(void)
2385 struct list_head
*head
= &__get_cpu_var(rotation_list
);
2386 struct perf_cpu_context
*cpuctx
, *tmp
;
2388 WARN_ON(!irqs_disabled());
2390 list_for_each_entry_safe(cpuctx
, tmp
, head
, rotation_list
) {
2391 if (cpuctx
->jiffies_interval
== 1 ||
2392 !(jiffies
% cpuctx
->jiffies_interval
))
2393 perf_rotate_context(cpuctx
);
2397 static int event_enable_on_exec(struct perf_event
*event
,
2398 struct perf_event_context
*ctx
)
2400 if (!event
->attr
.enable_on_exec
)
2403 event
->attr
.enable_on_exec
= 0;
2404 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
)
2407 __perf_event_mark_enabled(event
, ctx
);
2413 * Enable all of a task's events that have been marked enable-on-exec.
2414 * This expects task == current.
2416 static void perf_event_enable_on_exec(struct perf_event_context
*ctx
)
2418 struct perf_event
*event
;
2419 unsigned long flags
;
2423 local_irq_save(flags
);
2424 if (!ctx
|| !ctx
->nr_events
)
2428 * We must ctxsw out cgroup events to avoid conflict
2429 * when invoking perf_task_event_sched_in() later on
2430 * in this function. Otherwise we end up trying to
2431 * ctxswin cgroup events which are already scheduled
2434 perf_cgroup_sched_out(current
);
2435 task_ctx_sched_out(ctx
, EVENT_ALL
);
2437 raw_spin_lock(&ctx
->lock
);
2439 list_for_each_entry(event
, &ctx
->pinned_groups
, group_entry
) {
2440 ret
= event_enable_on_exec(event
, ctx
);
2445 list_for_each_entry(event
, &ctx
->flexible_groups
, group_entry
) {
2446 ret
= event_enable_on_exec(event
, ctx
);
2452 * Unclone this context if we enabled any event.
2457 raw_spin_unlock(&ctx
->lock
);
2460 * Also calls ctxswin for cgroup events, if any:
2462 perf_event_context_sched_in(ctx
, ctx
->task
);
2464 local_irq_restore(flags
);
2468 * Cross CPU call to read the hardware event
2470 static void __perf_event_read(void *info
)
2472 struct perf_event
*event
= info
;
2473 struct perf_event_context
*ctx
= event
->ctx
;
2474 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
2477 * If this is a task context, we need to check whether it is
2478 * the current task context of this cpu. If not it has been
2479 * scheduled out before the smp call arrived. In that case
2480 * event->count would have been updated to a recent sample
2481 * when the event was scheduled out.
2483 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
)
2486 raw_spin_lock(&ctx
->lock
);
2487 if (ctx
->is_active
) {
2488 update_context_time(ctx
);
2489 update_cgrp_time_from_event(event
);
2491 update_event_times(event
);
2492 if (event
->state
== PERF_EVENT_STATE_ACTIVE
)
2493 event
->pmu
->read(event
);
2494 raw_spin_unlock(&ctx
->lock
);
2497 static inline u64
perf_event_count(struct perf_event
*event
)
2499 return local64_read(&event
->count
) + atomic64_read(&event
->child_count
);
2502 static u64
perf_event_read(struct perf_event
*event
)
2505 * If event is enabled and currently active on a CPU, update the
2506 * value in the event structure:
2508 if (event
->state
== PERF_EVENT_STATE_ACTIVE
) {
2509 smp_call_function_single(event
->oncpu
,
2510 __perf_event_read
, event
, 1);
2511 } else if (event
->state
== PERF_EVENT_STATE_INACTIVE
) {
2512 struct perf_event_context
*ctx
= event
->ctx
;
2513 unsigned long flags
;
2515 raw_spin_lock_irqsave(&ctx
->lock
, flags
);
2517 * may read while context is not active
2518 * (e.g., thread is blocked), in that case
2519 * we cannot update context time
2521 if (ctx
->is_active
) {
2522 update_context_time(ctx
);
2523 update_cgrp_time_from_event(event
);
2525 update_event_times(event
);
2526 raw_spin_unlock_irqrestore(&ctx
->lock
, flags
);
2529 return perf_event_count(event
);
2536 struct callchain_cpus_entries
{
2537 struct rcu_head rcu_head
;
2538 struct perf_callchain_entry
*cpu_entries
[0];
2541 static DEFINE_PER_CPU(int, callchain_recursion
[PERF_NR_CONTEXTS
]);
2542 static atomic_t nr_callchain_events
;
2543 static DEFINE_MUTEX(callchain_mutex
);
2544 struct callchain_cpus_entries
*callchain_cpus_entries
;
2547 __weak
void perf_callchain_kernel(struct perf_callchain_entry
*entry
,
2548 struct pt_regs
*regs
)
2552 __weak
void perf_callchain_user(struct perf_callchain_entry
*entry
,
2553 struct pt_regs
*regs
)
2557 static void release_callchain_buffers_rcu(struct rcu_head
*head
)
2559 struct callchain_cpus_entries
*entries
;
2562 entries
= container_of(head
, struct callchain_cpus_entries
, rcu_head
);
2564 for_each_possible_cpu(cpu
)
2565 kfree(entries
->cpu_entries
[cpu
]);
2570 static void release_callchain_buffers(void)
2572 struct callchain_cpus_entries
*entries
;
2574 entries
= callchain_cpus_entries
;
2575 rcu_assign_pointer(callchain_cpus_entries
, NULL
);
2576 call_rcu(&entries
->rcu_head
, release_callchain_buffers_rcu
);
2579 static int alloc_callchain_buffers(void)
2583 struct callchain_cpus_entries
*entries
;
2586 * We can't use the percpu allocation API for data that can be
2587 * accessed from NMI. Use a temporary manual per cpu allocation
2588 * until that gets sorted out.
2590 size
= offsetof(struct callchain_cpus_entries
, cpu_entries
[nr_cpu_ids
]);
2592 entries
= kzalloc(size
, GFP_KERNEL
);
2596 size
= sizeof(struct perf_callchain_entry
) * PERF_NR_CONTEXTS
;
2598 for_each_possible_cpu(cpu
) {
2599 entries
->cpu_entries
[cpu
] = kmalloc_node(size
, GFP_KERNEL
,
2601 if (!entries
->cpu_entries
[cpu
])
2605 rcu_assign_pointer(callchain_cpus_entries
, entries
);
2610 for_each_possible_cpu(cpu
)
2611 kfree(entries
->cpu_entries
[cpu
]);
2617 static int get_callchain_buffers(void)
2622 mutex_lock(&callchain_mutex
);
2624 count
= atomic_inc_return(&nr_callchain_events
);
2625 if (WARN_ON_ONCE(count
< 1)) {
2631 /* If the allocation failed, give up */
2632 if (!callchain_cpus_entries
)
2637 err
= alloc_callchain_buffers();
2639 release_callchain_buffers();
2641 mutex_unlock(&callchain_mutex
);
2646 static void put_callchain_buffers(void)
2648 if (atomic_dec_and_mutex_lock(&nr_callchain_events
, &callchain_mutex
)) {
2649 release_callchain_buffers();
2650 mutex_unlock(&callchain_mutex
);
2654 static int get_recursion_context(int *recursion
)
2662 else if (in_softirq())
2667 if (recursion
[rctx
])
2676 static inline void put_recursion_context(int *recursion
, int rctx
)
2682 static struct perf_callchain_entry
*get_callchain_entry(int *rctx
)
2685 struct callchain_cpus_entries
*entries
;
2687 *rctx
= get_recursion_context(__get_cpu_var(callchain_recursion
));
2691 entries
= rcu_dereference(callchain_cpus_entries
);
2695 cpu
= smp_processor_id();
2697 return &entries
->cpu_entries
[cpu
][*rctx
];
2701 put_callchain_entry(int rctx
)
2703 put_recursion_context(__get_cpu_var(callchain_recursion
), rctx
);
2706 static struct perf_callchain_entry
*perf_callchain(struct pt_regs
*regs
)
2709 struct perf_callchain_entry
*entry
;
2712 entry
= get_callchain_entry(&rctx
);
2721 if (!user_mode(regs
)) {
2722 perf_callchain_store(entry
, PERF_CONTEXT_KERNEL
);
2723 perf_callchain_kernel(entry
, regs
);
2725 regs
= task_pt_regs(current
);
2731 perf_callchain_store(entry
, PERF_CONTEXT_USER
);
2732 perf_callchain_user(entry
, regs
);
2736 put_callchain_entry(rctx
);
2742 * Initialize the perf_event context in a task_struct:
2744 static void __perf_event_init_context(struct perf_event_context
*ctx
)
2746 raw_spin_lock_init(&ctx
->lock
);
2747 mutex_init(&ctx
->mutex
);
2748 INIT_LIST_HEAD(&ctx
->pinned_groups
);
2749 INIT_LIST_HEAD(&ctx
->flexible_groups
);
2750 INIT_LIST_HEAD(&ctx
->event_list
);
2751 atomic_set(&ctx
->refcount
, 1);
2754 static struct perf_event_context
*
2755 alloc_perf_context(struct pmu
*pmu
, struct task_struct
*task
)
2757 struct perf_event_context
*ctx
;
2759 ctx
= kzalloc(sizeof(struct perf_event_context
), GFP_KERNEL
);
2763 __perf_event_init_context(ctx
);
2766 get_task_struct(task
);
2773 static struct task_struct
*
2774 find_lively_task_by_vpid(pid_t vpid
)
2776 struct task_struct
*task
;
2783 task
= find_task_by_vpid(vpid
);
2785 get_task_struct(task
);
2789 return ERR_PTR(-ESRCH
);
2791 /* Reuse ptrace permission checks for now. */
2793 if (!ptrace_may_access(task
, PTRACE_MODE_READ
))
2798 put_task_struct(task
);
2799 return ERR_PTR(err
);
2804 * Returns a matching context with refcount and pincount.
2806 static struct perf_event_context
*
2807 find_get_context(struct pmu
*pmu
, struct task_struct
*task
, int cpu
)
2809 struct perf_event_context
*ctx
;
2810 struct perf_cpu_context
*cpuctx
;
2811 unsigned long flags
;
2815 /* Must be root to operate on a CPU event: */
2816 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN
))
2817 return ERR_PTR(-EACCES
);
2820 * We could be clever and allow to attach a event to an
2821 * offline CPU and activate it when the CPU comes up, but
2824 if (!cpu_online(cpu
))
2825 return ERR_PTR(-ENODEV
);
2827 cpuctx
= per_cpu_ptr(pmu
->pmu_cpu_context
, cpu
);
2836 ctxn
= pmu
->task_ctx_nr
;
2841 ctx
= perf_lock_task_context(task
, ctxn
, &flags
);
2845 raw_spin_unlock_irqrestore(&ctx
->lock
, flags
);
2849 ctx
= alloc_perf_context(pmu
, task
);
2857 mutex_lock(&task
->perf_event_mutex
);
2859 * If it has already passed perf_event_exit_task().
2860 * we must see PF_EXITING, it takes this mutex too.
2862 if (task
->flags
& PF_EXITING
)
2864 else if (task
->perf_event_ctxp
[ctxn
])
2868 rcu_assign_pointer(task
->perf_event_ctxp
[ctxn
], ctx
);
2870 mutex_unlock(&task
->perf_event_mutex
);
2872 if (unlikely(err
)) {
2873 put_task_struct(task
);
2885 return ERR_PTR(err
);
2888 static void perf_event_free_filter(struct perf_event
*event
);
2890 static void free_event_rcu(struct rcu_head
*head
)
2892 struct perf_event
*event
;
2894 event
= container_of(head
, struct perf_event
, rcu_head
);
2896 put_pid_ns(event
->ns
);
2897 perf_event_free_filter(event
);
2901 static void perf_buffer_put(struct perf_buffer
*buffer
);
2903 static void free_event(struct perf_event
*event
)
2905 irq_work_sync(&event
->pending
);
2907 if (!event
->parent
) {
2908 if (event
->attach_state
& PERF_ATTACH_TASK
)
2909 jump_label_dec(&perf_sched_events
);
2910 if (event
->attr
.mmap
|| event
->attr
.mmap_data
)
2911 atomic_dec(&nr_mmap_events
);
2912 if (event
->attr
.comm
)
2913 atomic_dec(&nr_comm_events
);
2914 if (event
->attr
.task
)
2915 atomic_dec(&nr_task_events
);
2916 if (event
->attr
.sample_type
& PERF_SAMPLE_CALLCHAIN
)
2917 put_callchain_buffers();
2918 if (is_cgroup_event(event
)) {
2919 atomic_dec(&per_cpu(perf_cgroup_events
, event
->cpu
));
2920 jump_label_dec(&perf_sched_events
);
2924 if (event
->buffer
) {
2925 perf_buffer_put(event
->buffer
);
2926 event
->buffer
= NULL
;
2929 if (is_cgroup_event(event
))
2930 perf_detach_cgroup(event
);
2933 event
->destroy(event
);
2936 put_ctx(event
->ctx
);
2938 call_rcu(&event
->rcu_head
, free_event_rcu
);
2941 int perf_event_release_kernel(struct perf_event
*event
)
2943 struct perf_event_context
*ctx
= event
->ctx
;
2946 * Remove from the PMU, can't get re-enabled since we got
2947 * here because the last ref went.
2949 perf_event_disable(event
);
2951 WARN_ON_ONCE(ctx
->parent_ctx
);
2953 * There are two ways this annotation is useful:
2955 * 1) there is a lock recursion from perf_event_exit_task
2956 * see the comment there.
2958 * 2) there is a lock-inversion with mmap_sem through
2959 * perf_event_read_group(), which takes faults while
2960 * holding ctx->mutex, however this is called after
2961 * the last filedesc died, so there is no possibility
2962 * to trigger the AB-BA case.
2964 mutex_lock_nested(&ctx
->mutex
, SINGLE_DEPTH_NESTING
);
2965 raw_spin_lock_irq(&ctx
->lock
);
2966 perf_group_detach(event
);
2967 list_del_event(event
, ctx
);
2968 raw_spin_unlock_irq(&ctx
->lock
);
2969 mutex_unlock(&ctx
->mutex
);
2975 EXPORT_SYMBOL_GPL(perf_event_release_kernel
);
2978 * Called when the last reference to the file is gone.
2980 static int perf_release(struct inode
*inode
, struct file
*file
)
2982 struct perf_event
*event
= file
->private_data
;
2983 struct task_struct
*owner
;
2985 file
->private_data
= NULL
;
2988 owner
= ACCESS_ONCE(event
->owner
);
2990 * Matches the smp_wmb() in perf_event_exit_task(). If we observe
2991 * !owner it means the list deletion is complete and we can indeed
2992 * free this event, otherwise we need to serialize on
2993 * owner->perf_event_mutex.
2995 smp_read_barrier_depends();
2998 * Since delayed_put_task_struct() also drops the last
2999 * task reference we can safely take a new reference
3000 * while holding the rcu_read_lock().
3002 get_task_struct(owner
);
3007 mutex_lock(&owner
->perf_event_mutex
);
3009 * We have to re-check the event->owner field, if it is cleared
3010 * we raced with perf_event_exit_task(), acquiring the mutex
3011 * ensured they're done, and we can proceed with freeing the
3015 list_del_init(&event
->owner_entry
);
3016 mutex_unlock(&owner
->perf_event_mutex
);
3017 put_task_struct(owner
);
3020 return perf_event_release_kernel(event
);
3023 u64
perf_event_read_value(struct perf_event
*event
, u64
*enabled
, u64
*running
)
3025 struct perf_event
*child
;
3031 mutex_lock(&event
->child_mutex
);
3032 total
+= perf_event_read(event
);
3033 *enabled
+= event
->total_time_enabled
+
3034 atomic64_read(&event
->child_total_time_enabled
);
3035 *running
+= event
->total_time_running
+
3036 atomic64_read(&event
->child_total_time_running
);
3038 list_for_each_entry(child
, &event
->child_list
, child_list
) {
3039 total
+= perf_event_read(child
);
3040 *enabled
+= child
->total_time_enabled
;
3041 *running
+= child
->total_time_running
;
3043 mutex_unlock(&event
->child_mutex
);
3047 EXPORT_SYMBOL_GPL(perf_event_read_value
);
3049 static int perf_event_read_group(struct perf_event
*event
,
3050 u64 read_format
, char __user
*buf
)
3052 struct perf_event
*leader
= event
->group_leader
, *sub
;
3053 int n
= 0, size
= 0, ret
= -EFAULT
;
3054 struct perf_event_context
*ctx
= leader
->ctx
;
3056 u64 count
, enabled
, running
;
3058 mutex_lock(&ctx
->mutex
);
3059 count
= perf_event_read_value(leader
, &enabled
, &running
);
3061 values
[n
++] = 1 + leader
->nr_siblings
;
3062 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
3063 values
[n
++] = enabled
;
3064 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
3065 values
[n
++] = running
;
3066 values
[n
++] = count
;
3067 if (read_format
& PERF_FORMAT_ID
)
3068 values
[n
++] = primary_event_id(leader
);
3070 size
= n
* sizeof(u64
);
3072 if (copy_to_user(buf
, values
, size
))
3077 list_for_each_entry(sub
, &leader
->sibling_list
, group_entry
) {
3080 values
[n
++] = perf_event_read_value(sub
, &enabled
, &running
);
3081 if (read_format
& PERF_FORMAT_ID
)
3082 values
[n
++] = primary_event_id(sub
);
3084 size
= n
* sizeof(u64
);
3086 if (copy_to_user(buf
+ ret
, values
, size
)) {
3094 mutex_unlock(&ctx
->mutex
);
3099 static int perf_event_read_one(struct perf_event
*event
,
3100 u64 read_format
, char __user
*buf
)
3102 u64 enabled
, running
;
3106 values
[n
++] = perf_event_read_value(event
, &enabled
, &running
);
3107 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
3108 values
[n
++] = enabled
;
3109 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
3110 values
[n
++] = running
;
3111 if (read_format
& PERF_FORMAT_ID
)
3112 values
[n
++] = primary_event_id(event
);
3114 if (copy_to_user(buf
, values
, n
* sizeof(u64
)))
3117 return n
* sizeof(u64
);
3121 * Read the performance event - simple non blocking version for now
3124 perf_read_hw(struct perf_event
*event
, char __user
*buf
, size_t count
)
3126 u64 read_format
= event
->attr
.read_format
;
3130 * Return end-of-file for a read on a event that is in
3131 * error state (i.e. because it was pinned but it couldn't be
3132 * scheduled on to the CPU at some point).
3134 if (event
->state
== PERF_EVENT_STATE_ERROR
)
3137 if (count
< event
->read_size
)
3140 WARN_ON_ONCE(event
->ctx
->parent_ctx
);
3141 if (read_format
& PERF_FORMAT_GROUP
)
3142 ret
= perf_event_read_group(event
, read_format
, buf
);
3144 ret
= perf_event_read_one(event
, read_format
, buf
);
3150 perf_read(struct file
*file
, char __user
*buf
, size_t count
, loff_t
*ppos
)
3152 struct perf_event
*event
= file
->private_data
;
3154 return perf_read_hw(event
, buf
, count
);
3157 static unsigned int perf_poll(struct file
*file
, poll_table
*wait
)
3159 struct perf_event
*event
= file
->private_data
;
3160 struct perf_buffer
*buffer
;
3161 unsigned int events
= POLL_HUP
;
3164 buffer
= rcu_dereference(event
->buffer
);
3166 events
= atomic_xchg(&buffer
->poll
, 0);
3169 poll_wait(file
, &event
->waitq
, wait
);
3174 static void perf_event_reset(struct perf_event
*event
)
3176 (void)perf_event_read(event
);
3177 local64_set(&event
->count
, 0);
3178 perf_event_update_userpage(event
);
3182 * Holding the top-level event's child_mutex means that any
3183 * descendant process that has inherited this event will block
3184 * in sync_child_event if it goes to exit, thus satisfying the
3185 * task existence requirements of perf_event_enable/disable.
3187 static void perf_event_for_each_child(struct perf_event
*event
,
3188 void (*func
)(struct perf_event
*))
3190 struct perf_event
*child
;
3192 WARN_ON_ONCE(event
->ctx
->parent_ctx
);
3193 mutex_lock(&event
->child_mutex
);
3195 list_for_each_entry(child
, &event
->child_list
, child_list
)
3197 mutex_unlock(&event
->child_mutex
);
3200 static void perf_event_for_each(struct perf_event
*event
,
3201 void (*func
)(struct perf_event
*))
3203 struct perf_event_context
*ctx
= event
->ctx
;
3204 struct perf_event
*sibling
;
3206 WARN_ON_ONCE(ctx
->parent_ctx
);
3207 mutex_lock(&ctx
->mutex
);
3208 event
= event
->group_leader
;
3210 perf_event_for_each_child(event
, func
);
3212 list_for_each_entry(sibling
, &event
->sibling_list
, group_entry
)
3213 perf_event_for_each_child(event
, func
);
3214 mutex_unlock(&ctx
->mutex
);
3217 static int perf_event_period(struct perf_event
*event
, u64 __user
*arg
)
3219 struct perf_event_context
*ctx
= event
->ctx
;
3223 if (!is_sampling_event(event
))
3226 if (copy_from_user(&value
, arg
, sizeof(value
)))
3232 raw_spin_lock_irq(&ctx
->lock
);
3233 if (event
->attr
.freq
) {
3234 if (value
> sysctl_perf_event_sample_rate
) {
3239 event
->attr
.sample_freq
= value
;
3241 event
->attr
.sample_period
= value
;
3242 event
->hw
.sample_period
= value
;
3245 raw_spin_unlock_irq(&ctx
->lock
);
3250 static const struct file_operations perf_fops
;
3252 static struct perf_event
*perf_fget_light(int fd
, int *fput_needed
)
3256 file
= fget_light(fd
, fput_needed
);
3258 return ERR_PTR(-EBADF
);
3260 if (file
->f_op
!= &perf_fops
) {
3261 fput_light(file
, *fput_needed
);
3263 return ERR_PTR(-EBADF
);
3266 return file
->private_data
;
3269 static int perf_event_set_output(struct perf_event
*event
,
3270 struct perf_event
*output_event
);
3271 static int perf_event_set_filter(struct perf_event
*event
, void __user
*arg
);
3273 static long perf_ioctl(struct file
*file
, unsigned int cmd
, unsigned long arg
)
3275 struct perf_event
*event
= file
->private_data
;
3276 void (*func
)(struct perf_event
*);
3280 case PERF_EVENT_IOC_ENABLE
:
3281 func
= perf_event_enable
;
3283 case PERF_EVENT_IOC_DISABLE
:
3284 func
= perf_event_disable
;
3286 case PERF_EVENT_IOC_RESET
:
3287 func
= perf_event_reset
;
3290 case PERF_EVENT_IOC_REFRESH
:
3291 return perf_event_refresh(event
, arg
);
3293 case PERF_EVENT_IOC_PERIOD
:
3294 return perf_event_period(event
, (u64 __user
*)arg
);
3296 case PERF_EVENT_IOC_SET_OUTPUT
:
3298 struct perf_event
*output_event
= NULL
;
3299 int fput_needed
= 0;
3303 output_event
= perf_fget_light(arg
, &fput_needed
);
3304 if (IS_ERR(output_event
))
3305 return PTR_ERR(output_event
);
3308 ret
= perf_event_set_output(event
, output_event
);
3310 fput_light(output_event
->filp
, fput_needed
);
3315 case PERF_EVENT_IOC_SET_FILTER
:
3316 return perf_event_set_filter(event
, (void __user
*)arg
);
3322 if (flags
& PERF_IOC_FLAG_GROUP
)
3323 perf_event_for_each(event
, func
);
3325 perf_event_for_each_child(event
, func
);
3330 int perf_event_task_enable(void)
3332 struct perf_event
*event
;
3334 mutex_lock(¤t
->perf_event_mutex
);
3335 list_for_each_entry(event
, ¤t
->perf_event_list
, owner_entry
)
3336 perf_event_for_each_child(event
, perf_event_enable
);
3337 mutex_unlock(¤t
->perf_event_mutex
);
3342 int perf_event_task_disable(void)
3344 struct perf_event
*event
;
3346 mutex_lock(¤t
->perf_event_mutex
);
3347 list_for_each_entry(event
, ¤t
->perf_event_list
, owner_entry
)
3348 perf_event_for_each_child(event
, perf_event_disable
);
3349 mutex_unlock(¤t
->perf_event_mutex
);
3354 #ifndef PERF_EVENT_INDEX_OFFSET
3355 # define PERF_EVENT_INDEX_OFFSET 0
3358 static int perf_event_index(struct perf_event
*event
)
3360 if (event
->hw
.state
& PERF_HES_STOPPED
)
3363 if (event
->state
!= PERF_EVENT_STATE_ACTIVE
)
3366 return event
->hw
.idx
+ 1 - PERF_EVENT_INDEX_OFFSET
;
3370 * Callers need to ensure there can be no nesting of this function, otherwise
3371 * the seqlock logic goes bad. We can not serialize this because the arch
3372 * code calls this from NMI context.
3374 void perf_event_update_userpage(struct perf_event
*event
)
3376 struct perf_event_mmap_page
*userpg
;
3377 struct perf_buffer
*buffer
;
3380 buffer
= rcu_dereference(event
->buffer
);
3384 userpg
= buffer
->user_page
;
3387 * Disable preemption so as to not let the corresponding user-space
3388 * spin too long if we get preempted.
3393 userpg
->index
= perf_event_index(event
);
3394 userpg
->offset
= perf_event_count(event
);
3395 if (event
->state
== PERF_EVENT_STATE_ACTIVE
)
3396 userpg
->offset
-= local64_read(&event
->hw
.prev_count
);
3398 userpg
->time_enabled
= event
->total_time_enabled
+
3399 atomic64_read(&event
->child_total_time_enabled
);
3401 userpg
->time_running
= event
->total_time_running
+
3402 atomic64_read(&event
->child_total_time_running
);
3411 static unsigned long perf_data_size(struct perf_buffer
*buffer
);
3414 perf_buffer_init(struct perf_buffer
*buffer
, long watermark
, int flags
)
3416 long max_size
= perf_data_size(buffer
);
3419 buffer
->watermark
= min(max_size
, watermark
);
3421 if (!buffer
->watermark
)
3422 buffer
->watermark
= max_size
/ 2;
3424 if (flags
& PERF_BUFFER_WRITABLE
)
3425 buffer
->writable
= 1;
3427 atomic_set(&buffer
->refcount
, 1);
3430 #ifndef CONFIG_PERF_USE_VMALLOC
3433 * Back perf_mmap() with regular GFP_KERNEL-0 pages.
3436 static struct page
*
3437 perf_mmap_to_page(struct perf_buffer
*buffer
, unsigned long pgoff
)
3439 if (pgoff
> buffer
->nr_pages
)
3443 return virt_to_page(buffer
->user_page
);
3445 return virt_to_page(buffer
->data_pages
[pgoff
- 1]);
3448 static void *perf_mmap_alloc_page(int cpu
)
3453 node
= (cpu
== -1) ? cpu
: cpu_to_node(cpu
);
3454 page
= alloc_pages_node(node
, GFP_KERNEL
| __GFP_ZERO
, 0);
3458 return page_address(page
);
3461 static struct perf_buffer
*
3462 perf_buffer_alloc(int nr_pages
, long watermark
, int cpu
, int flags
)
3464 struct perf_buffer
*buffer
;
3468 size
= sizeof(struct perf_buffer
);
3469 size
+= nr_pages
* sizeof(void *);
3471 buffer
= kzalloc(size
, GFP_KERNEL
);
3475 buffer
->user_page
= perf_mmap_alloc_page(cpu
);
3476 if (!buffer
->user_page
)
3477 goto fail_user_page
;
3479 for (i
= 0; i
< nr_pages
; i
++) {
3480 buffer
->data_pages
[i
] = perf_mmap_alloc_page(cpu
);
3481 if (!buffer
->data_pages
[i
])
3482 goto fail_data_pages
;
3485 buffer
->nr_pages
= nr_pages
;
3487 perf_buffer_init(buffer
, watermark
, flags
);
3492 for (i
--; i
>= 0; i
--)
3493 free_page((unsigned long)buffer
->data_pages
[i
]);
3495 free_page((unsigned long)buffer
->user_page
);
3504 static void perf_mmap_free_page(unsigned long addr
)
3506 struct page
*page
= virt_to_page((void *)addr
);
3508 page
->mapping
= NULL
;
3512 static void perf_buffer_free(struct perf_buffer
*buffer
)
3516 perf_mmap_free_page((unsigned long)buffer
->user_page
);
3517 for (i
= 0; i
< buffer
->nr_pages
; i
++)
3518 perf_mmap_free_page((unsigned long)buffer
->data_pages
[i
]);
3522 static inline int page_order(struct perf_buffer
*buffer
)
3530 * Back perf_mmap() with vmalloc memory.
3532 * Required for architectures that have d-cache aliasing issues.
3535 static inline int page_order(struct perf_buffer
*buffer
)
3537 return buffer
->page_order
;
3540 static struct page
*
3541 perf_mmap_to_page(struct perf_buffer
*buffer
, unsigned long pgoff
)
3543 if (pgoff
> (1UL << page_order(buffer
)))
3546 return vmalloc_to_page((void *)buffer
->user_page
+ pgoff
* PAGE_SIZE
);
3549 static void perf_mmap_unmark_page(void *addr
)
3551 struct page
*page
= vmalloc_to_page(addr
);
3553 page
->mapping
= NULL
;
3556 static void perf_buffer_free_work(struct work_struct
*work
)
3558 struct perf_buffer
*buffer
;
3562 buffer
= container_of(work
, struct perf_buffer
, work
);
3563 nr
= 1 << page_order(buffer
);
3565 base
= buffer
->user_page
;
3566 for (i
= 0; i
< nr
+ 1; i
++)
3567 perf_mmap_unmark_page(base
+ (i
* PAGE_SIZE
));
3573 static void perf_buffer_free(struct perf_buffer
*buffer
)
3575 schedule_work(&buffer
->work
);
3578 static struct perf_buffer
*
3579 perf_buffer_alloc(int nr_pages
, long watermark
, int cpu
, int flags
)
3581 struct perf_buffer
*buffer
;
3585 size
= sizeof(struct perf_buffer
);
3586 size
+= sizeof(void *);
3588 buffer
= kzalloc(size
, GFP_KERNEL
);
3592 INIT_WORK(&buffer
->work
, perf_buffer_free_work
);
3594 all_buf
= vmalloc_user((nr_pages
+ 1) * PAGE_SIZE
);
3598 buffer
->user_page
= all_buf
;
3599 buffer
->data_pages
[0] = all_buf
+ PAGE_SIZE
;
3600 buffer
->page_order
= ilog2(nr_pages
);
3601 buffer
->nr_pages
= 1;
3603 perf_buffer_init(buffer
, watermark
, flags
);
3616 static unsigned long perf_data_size(struct perf_buffer
*buffer
)
3618 return buffer
->nr_pages
<< (PAGE_SHIFT
+ page_order(buffer
));
3621 static int perf_mmap_fault(struct vm_area_struct
*vma
, struct vm_fault
*vmf
)
3623 struct perf_event
*event
= vma
->vm_file
->private_data
;
3624 struct perf_buffer
*buffer
;
3625 int ret
= VM_FAULT_SIGBUS
;
3627 if (vmf
->flags
& FAULT_FLAG_MKWRITE
) {
3628 if (vmf
->pgoff
== 0)
3634 buffer
= rcu_dereference(event
->buffer
);
3638 if (vmf
->pgoff
&& (vmf
->flags
& FAULT_FLAG_WRITE
))
3641 vmf
->page
= perf_mmap_to_page(buffer
, vmf
->pgoff
);
3645 get_page(vmf
->page
);
3646 vmf
->page
->mapping
= vma
->vm_file
->f_mapping
;
3647 vmf
->page
->index
= vmf
->pgoff
;
3656 static void perf_buffer_free_rcu(struct rcu_head
*rcu_head
)
3658 struct perf_buffer
*buffer
;
3660 buffer
= container_of(rcu_head
, struct perf_buffer
, rcu_head
);
3661 perf_buffer_free(buffer
);
3664 static struct perf_buffer
*perf_buffer_get(struct perf_event
*event
)
3666 struct perf_buffer
*buffer
;
3669 buffer
= rcu_dereference(event
->buffer
);
3671 if (!atomic_inc_not_zero(&buffer
->refcount
))
3679 static void perf_buffer_put(struct perf_buffer
*buffer
)
3681 if (!atomic_dec_and_test(&buffer
->refcount
))
3684 call_rcu(&buffer
->rcu_head
, perf_buffer_free_rcu
);
3687 static void perf_mmap_open(struct vm_area_struct
*vma
)
3689 struct perf_event
*event
= vma
->vm_file
->private_data
;
3691 atomic_inc(&event
->mmap_count
);
3694 static void perf_mmap_close(struct vm_area_struct
*vma
)
3696 struct perf_event
*event
= vma
->vm_file
->private_data
;
3698 if (atomic_dec_and_mutex_lock(&event
->mmap_count
, &event
->mmap_mutex
)) {
3699 unsigned long size
= perf_data_size(event
->buffer
);
3700 struct user_struct
*user
= event
->mmap_user
;
3701 struct perf_buffer
*buffer
= event
->buffer
;
3703 atomic_long_sub((size
>> PAGE_SHIFT
) + 1, &user
->locked_vm
);
3704 vma
->vm_mm
->locked_vm
-= event
->mmap_locked
;
3705 rcu_assign_pointer(event
->buffer
, NULL
);
3706 mutex_unlock(&event
->mmap_mutex
);
3708 perf_buffer_put(buffer
);
3713 static const struct vm_operations_struct perf_mmap_vmops
= {
3714 .open
= perf_mmap_open
,
3715 .close
= perf_mmap_close
,
3716 .fault
= perf_mmap_fault
,
3717 .page_mkwrite
= perf_mmap_fault
,
3720 static int perf_mmap(struct file
*file
, struct vm_area_struct
*vma
)
3722 struct perf_event
*event
= file
->private_data
;
3723 unsigned long user_locked
, user_lock_limit
;
3724 struct user_struct
*user
= current_user();
3725 unsigned long locked
, lock_limit
;
3726 struct perf_buffer
*buffer
;
3727 unsigned long vma_size
;
3728 unsigned long nr_pages
;
3729 long user_extra
, extra
;
3730 int ret
= 0, flags
= 0;
3733 * Don't allow mmap() of inherited per-task counters. This would
3734 * create a performance issue due to all children writing to the
3737 if (event
->cpu
== -1 && event
->attr
.inherit
)
3740 if (!(vma
->vm_flags
& VM_SHARED
))
3743 vma_size
= vma
->vm_end
- vma
->vm_start
;
3744 nr_pages
= (vma_size
/ PAGE_SIZE
) - 1;
3747 * If we have buffer pages ensure they're a power-of-two number, so we
3748 * can do bitmasks instead of modulo.
3750 if (nr_pages
!= 0 && !is_power_of_2(nr_pages
))
3753 if (vma_size
!= PAGE_SIZE
* (1 + nr_pages
))
3756 if (vma
->vm_pgoff
!= 0)
3759 WARN_ON_ONCE(event
->ctx
->parent_ctx
);
3760 mutex_lock(&event
->mmap_mutex
);
3761 if (event
->buffer
) {
3762 if (event
->buffer
->nr_pages
== nr_pages
)
3763 atomic_inc(&event
->buffer
->refcount
);
3769 user_extra
= nr_pages
+ 1;
3770 user_lock_limit
= sysctl_perf_event_mlock
>> (PAGE_SHIFT
- 10);
3773 * Increase the limit linearly with more CPUs:
3775 user_lock_limit
*= num_online_cpus();
3777 user_locked
= atomic_long_read(&user
->locked_vm
) + user_extra
;
3780 if (user_locked
> user_lock_limit
)
3781 extra
= user_locked
- user_lock_limit
;
3783 lock_limit
= rlimit(RLIMIT_MEMLOCK
);
3784 lock_limit
>>= PAGE_SHIFT
;
3785 locked
= vma
->vm_mm
->locked_vm
+ extra
;
3787 if ((locked
> lock_limit
) && perf_paranoid_tracepoint_raw() &&
3788 !capable(CAP_IPC_LOCK
)) {
3793 WARN_ON(event
->buffer
);
3795 if (vma
->vm_flags
& VM_WRITE
)
3796 flags
|= PERF_BUFFER_WRITABLE
;
3798 buffer
= perf_buffer_alloc(nr_pages
, event
->attr
.wakeup_watermark
,
3804 rcu_assign_pointer(event
->buffer
, buffer
);
3806 atomic_long_add(user_extra
, &user
->locked_vm
);
3807 event
->mmap_locked
= extra
;
3808 event
->mmap_user
= get_current_user();
3809 vma
->vm_mm
->locked_vm
+= event
->mmap_locked
;
3813 atomic_inc(&event
->mmap_count
);
3814 mutex_unlock(&event
->mmap_mutex
);
3816 vma
->vm_flags
|= VM_RESERVED
;
3817 vma
->vm_ops
= &perf_mmap_vmops
;
3822 static int perf_fasync(int fd
, struct file
*filp
, int on
)
3824 struct inode
*inode
= filp
->f_path
.dentry
->d_inode
;
3825 struct perf_event
*event
= filp
->private_data
;
3828 mutex_lock(&inode
->i_mutex
);
3829 retval
= fasync_helper(fd
, filp
, on
, &event
->fasync
);
3830 mutex_unlock(&inode
->i_mutex
);
3838 static const struct file_operations perf_fops
= {
3839 .llseek
= no_llseek
,
3840 .release
= perf_release
,
3843 .unlocked_ioctl
= perf_ioctl
,
3844 .compat_ioctl
= perf_ioctl
,
3846 .fasync
= perf_fasync
,
3852 * If there's data, ensure we set the poll() state and publish everything
3853 * to user-space before waking everybody up.
3856 void perf_event_wakeup(struct perf_event
*event
)
3858 wake_up_all(&event
->waitq
);
3860 if (event
->pending_kill
) {
3861 kill_fasync(&event
->fasync
, SIGIO
, event
->pending_kill
);
3862 event
->pending_kill
= 0;
3866 static void perf_pending_event(struct irq_work
*entry
)
3868 struct perf_event
*event
= container_of(entry
,
3869 struct perf_event
, pending
);
3871 if (event
->pending_disable
) {
3872 event
->pending_disable
= 0;
3873 __perf_event_disable(event
);
3876 if (event
->pending_wakeup
) {
3877 event
->pending_wakeup
= 0;
3878 perf_event_wakeup(event
);
3883 * We assume there is only KVM supporting the callbacks.
3884 * Later on, we might change it to a list if there is
3885 * another virtualization implementation supporting the callbacks.
3887 struct perf_guest_info_callbacks
*perf_guest_cbs
;
3889 int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks
*cbs
)
3891 perf_guest_cbs
= cbs
;
3894 EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks
);
3896 int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks
*cbs
)
3898 perf_guest_cbs
= NULL
;
3901 EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks
);
3906 static bool perf_output_space(struct perf_buffer
*buffer
, unsigned long tail
,
3907 unsigned long offset
, unsigned long head
)
3911 if (!buffer
->writable
)
3914 mask
= perf_data_size(buffer
) - 1;
3916 offset
= (offset
- tail
) & mask
;
3917 head
= (head
- tail
) & mask
;
3919 if ((int)(head
- offset
) < 0)
3925 static void perf_output_wakeup(struct perf_output_handle
*handle
)
3927 atomic_set(&handle
->buffer
->poll
, POLL_IN
);
3930 handle
->event
->pending_wakeup
= 1;
3931 irq_work_queue(&handle
->event
->pending
);
3933 perf_event_wakeup(handle
->event
);
3937 * We need to ensure a later event_id doesn't publish a head when a former
3938 * event isn't done writing. However since we need to deal with NMIs we
3939 * cannot fully serialize things.
3941 * We only publish the head (and generate a wakeup) when the outer-most
3944 static void perf_output_get_handle(struct perf_output_handle
*handle
)
3946 struct perf_buffer
*buffer
= handle
->buffer
;
3949 local_inc(&buffer
->nest
);
3950 handle
->wakeup
= local_read(&buffer
->wakeup
);
3953 static void perf_output_put_handle(struct perf_output_handle
*handle
)
3955 struct perf_buffer
*buffer
= handle
->buffer
;
3959 head
= local_read(&buffer
->head
);
3962 * IRQ/NMI can happen here, which means we can miss a head update.
3965 if (!local_dec_and_test(&buffer
->nest
))
3969 * Publish the known good head. Rely on the full barrier implied
3970 * by atomic_dec_and_test() order the buffer->head read and this
3973 buffer
->user_page
->data_head
= head
;
3976 * Now check if we missed an update, rely on the (compiler)
3977 * barrier in atomic_dec_and_test() to re-read buffer->head.
3979 if (unlikely(head
!= local_read(&buffer
->head
))) {
3980 local_inc(&buffer
->nest
);
3984 if (handle
->wakeup
!= local_read(&buffer
->wakeup
))
3985 perf_output_wakeup(handle
);
3991 __always_inline
void perf_output_copy(struct perf_output_handle
*handle
,
3992 const void *buf
, unsigned int len
)
3995 unsigned long size
= min_t(unsigned long, handle
->size
, len
);
3997 memcpy(handle
->addr
, buf
, size
);
4000 handle
->addr
+= size
;
4002 handle
->size
-= size
;
4003 if (!handle
->size
) {
4004 struct perf_buffer
*buffer
= handle
->buffer
;
4007 handle
->page
&= buffer
->nr_pages
- 1;
4008 handle
->addr
= buffer
->data_pages
[handle
->page
];
4009 handle
->size
= PAGE_SIZE
<< page_order(buffer
);
4014 static void __perf_event_header__init_id(struct perf_event_header
*header
,
4015 struct perf_sample_data
*data
,
4016 struct perf_event
*event
)
4018 u64 sample_type
= event
->attr
.sample_type
;
4020 data
->type
= sample_type
;
4021 header
->size
+= event
->id_header_size
;
4023 if (sample_type
& PERF_SAMPLE_TID
) {
4024 /* namespace issues */
4025 data
->tid_entry
.pid
= perf_event_pid(event
, current
);
4026 data
->tid_entry
.tid
= perf_event_tid(event
, current
);
4029 if (sample_type
& PERF_SAMPLE_TIME
)
4030 data
->time
= perf_clock();
4032 if (sample_type
& PERF_SAMPLE_ID
)
4033 data
->id
= primary_event_id(event
);
4035 if (sample_type
& PERF_SAMPLE_STREAM_ID
)
4036 data
->stream_id
= event
->id
;
4038 if (sample_type
& PERF_SAMPLE_CPU
) {
4039 data
->cpu_entry
.cpu
= raw_smp_processor_id();
4040 data
->cpu_entry
.reserved
= 0;
4044 static void perf_event_header__init_id(struct perf_event_header
*header
,
4045 struct perf_sample_data
*data
,
4046 struct perf_event
*event
)
4048 if (event
->attr
.sample_id_all
)
4049 __perf_event_header__init_id(header
, data
, event
);
4052 static void __perf_event__output_id_sample(struct perf_output_handle
*handle
,
4053 struct perf_sample_data
*data
)
4055 u64 sample_type
= data
->type
;
4057 if (sample_type
& PERF_SAMPLE_TID
)
4058 perf_output_put(handle
, data
->tid_entry
);
4060 if (sample_type
& PERF_SAMPLE_TIME
)
4061 perf_output_put(handle
, data
->time
);
4063 if (sample_type
& PERF_SAMPLE_ID
)
4064 perf_output_put(handle
, data
->id
);
4066 if (sample_type
& PERF_SAMPLE_STREAM_ID
)
4067 perf_output_put(handle
, data
->stream_id
);
4069 if (sample_type
& PERF_SAMPLE_CPU
)
4070 perf_output_put(handle
, data
->cpu_entry
);
4073 static void perf_event__output_id_sample(struct perf_event
*event
,
4074 struct perf_output_handle
*handle
,
4075 struct perf_sample_data
*sample
)
4077 if (event
->attr
.sample_id_all
)
4078 __perf_event__output_id_sample(handle
, sample
);
4081 int perf_output_begin(struct perf_output_handle
*handle
,
4082 struct perf_event
*event
, unsigned int size
,
4083 int nmi
, int sample
)
4085 struct perf_buffer
*buffer
;
4086 unsigned long tail
, offset
, head
;
4088 struct perf_sample_data sample_data
;
4090 struct perf_event_header header
;
4097 * For inherited events we send all the output towards the parent.
4100 event
= event
->parent
;
4102 buffer
= rcu_dereference(event
->buffer
);
4106 handle
->buffer
= buffer
;
4107 handle
->event
= event
;
4109 handle
->sample
= sample
;
4111 if (!buffer
->nr_pages
)
4114 have_lost
= local_read(&buffer
->lost
);
4116 lost_event
.header
.size
= sizeof(lost_event
);
4117 perf_event_header__init_id(&lost_event
.header
, &sample_data
,
4119 size
+= lost_event
.header
.size
;
4122 perf_output_get_handle(handle
);
4126 * Userspace could choose to issue a mb() before updating the
4127 * tail pointer. So that all reads will be completed before the
4130 tail
= ACCESS_ONCE(buffer
->user_page
->data_tail
);
4132 offset
= head
= local_read(&buffer
->head
);
4134 if (unlikely(!perf_output_space(buffer
, tail
, offset
, head
)))
4136 } while (local_cmpxchg(&buffer
->head
, offset
, head
) != offset
);
4138 if (head
- local_read(&buffer
->wakeup
) > buffer
->watermark
)
4139 local_add(buffer
->watermark
, &buffer
->wakeup
);
4141 handle
->page
= offset
>> (PAGE_SHIFT
+ page_order(buffer
));
4142 handle
->page
&= buffer
->nr_pages
- 1;
4143 handle
->size
= offset
& ((PAGE_SIZE
<< page_order(buffer
)) - 1);
4144 handle
->addr
= buffer
->data_pages
[handle
->page
];
4145 handle
->addr
+= handle
->size
;
4146 handle
->size
= (PAGE_SIZE
<< page_order(buffer
)) - handle
->size
;
4149 lost_event
.header
.type
= PERF_RECORD_LOST
;
4150 lost_event
.header
.misc
= 0;
4151 lost_event
.id
= event
->id
;
4152 lost_event
.lost
= local_xchg(&buffer
->lost
, 0);
4154 perf_output_put(handle
, lost_event
);
4155 perf_event__output_id_sample(event
, handle
, &sample_data
);
4161 local_inc(&buffer
->lost
);
4162 perf_output_put_handle(handle
);
4169 void perf_output_end(struct perf_output_handle
*handle
)
4171 struct perf_event
*event
= handle
->event
;
4172 struct perf_buffer
*buffer
= handle
->buffer
;
4174 int wakeup_events
= event
->attr
.wakeup_events
;
4176 if (handle
->sample
&& wakeup_events
) {
4177 int events
= local_inc_return(&buffer
->events
);
4178 if (events
>= wakeup_events
) {
4179 local_sub(wakeup_events
, &buffer
->events
);
4180 local_inc(&buffer
->wakeup
);
4184 perf_output_put_handle(handle
);
4188 static void perf_output_read_one(struct perf_output_handle
*handle
,
4189 struct perf_event
*event
,
4190 u64 enabled
, u64 running
)
4192 u64 read_format
= event
->attr
.read_format
;
4196 values
[n
++] = perf_event_count(event
);
4197 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
) {
4198 values
[n
++] = enabled
+
4199 atomic64_read(&event
->child_total_time_enabled
);
4201 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
) {
4202 values
[n
++] = running
+
4203 atomic64_read(&event
->child_total_time_running
);
4205 if (read_format
& PERF_FORMAT_ID
)
4206 values
[n
++] = primary_event_id(event
);
4208 perf_output_copy(handle
, values
, n
* sizeof(u64
));
4212 * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
4214 static void perf_output_read_group(struct perf_output_handle
*handle
,
4215 struct perf_event
*event
,
4216 u64 enabled
, u64 running
)
4218 struct perf_event
*leader
= event
->group_leader
, *sub
;
4219 u64 read_format
= event
->attr
.read_format
;
4223 values
[n
++] = 1 + leader
->nr_siblings
;
4225 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
4226 values
[n
++] = enabled
;
4228 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
4229 values
[n
++] = running
;
4231 if (leader
!= event
)
4232 leader
->pmu
->read(leader
);
4234 values
[n
++] = perf_event_count(leader
);
4235 if (read_format
& PERF_FORMAT_ID
)
4236 values
[n
++] = primary_event_id(leader
);
4238 perf_output_copy(handle
, values
, n
* sizeof(u64
));
4240 list_for_each_entry(sub
, &leader
->sibling_list
, group_entry
) {
4244 sub
->pmu
->read(sub
);
4246 values
[n
++] = perf_event_count(sub
);
4247 if (read_format
& PERF_FORMAT_ID
)
4248 values
[n
++] = primary_event_id(sub
);
4250 perf_output_copy(handle
, values
, n
* sizeof(u64
));
4254 #define PERF_FORMAT_TOTAL_TIMES (PERF_FORMAT_TOTAL_TIME_ENABLED|\
4255 PERF_FORMAT_TOTAL_TIME_RUNNING)
4257 static void perf_output_read(struct perf_output_handle
*handle
,
4258 struct perf_event
*event
)
4260 u64 enabled
= 0, running
= 0, now
, ctx_time
;
4261 u64 read_format
= event
->attr
.read_format
;
4264 * compute total_time_enabled, total_time_running
4265 * based on snapshot values taken when the event
4266 * was last scheduled in.
4268 * we cannot simply called update_context_time()
4269 * because of locking issue as we are called in
4272 if (read_format
& PERF_FORMAT_TOTAL_TIMES
) {
4274 ctx_time
= event
->shadow_ctx_time
+ now
;
4275 enabled
= ctx_time
- event
->tstamp_enabled
;
4276 running
= ctx_time
- event
->tstamp_running
;
4279 if (event
->attr
.read_format
& PERF_FORMAT_GROUP
)
4280 perf_output_read_group(handle
, event
, enabled
, running
);
4282 perf_output_read_one(handle
, event
, enabled
, running
);
4285 void perf_output_sample(struct perf_output_handle
*handle
,
4286 struct perf_event_header
*header
,
4287 struct perf_sample_data
*data
,
4288 struct perf_event
*event
)
4290 u64 sample_type
= data
->type
;
4292 perf_output_put(handle
, *header
);
4294 if (sample_type
& PERF_SAMPLE_IP
)
4295 perf_output_put(handle
, data
->ip
);
4297 if (sample_type
& PERF_SAMPLE_TID
)
4298 perf_output_put(handle
, data
->tid_entry
);
4300 if (sample_type
& PERF_SAMPLE_TIME
)
4301 perf_output_put(handle
, data
->time
);
4303 if (sample_type
& PERF_SAMPLE_ADDR
)
4304 perf_output_put(handle
, data
->addr
);
4306 if (sample_type
& PERF_SAMPLE_ID
)
4307 perf_output_put(handle
, data
->id
);
4309 if (sample_type
& PERF_SAMPLE_STREAM_ID
)
4310 perf_output_put(handle
, data
->stream_id
);
4312 if (sample_type
& PERF_SAMPLE_CPU
)
4313 perf_output_put(handle
, data
->cpu_entry
);
4315 if (sample_type
& PERF_SAMPLE_PERIOD
)
4316 perf_output_put(handle
, data
->period
);
4318 if (sample_type
& PERF_SAMPLE_READ
)
4319 perf_output_read(handle
, event
);
4321 if (sample_type
& PERF_SAMPLE_CALLCHAIN
) {
4322 if (data
->callchain
) {
4325 if (data
->callchain
)
4326 size
+= data
->callchain
->nr
;
4328 size
*= sizeof(u64
);
4330 perf_output_copy(handle
, data
->callchain
, size
);
4333 perf_output_put(handle
, nr
);
4337 if (sample_type
& PERF_SAMPLE_RAW
) {
4339 perf_output_put(handle
, data
->raw
->size
);
4340 perf_output_copy(handle
, data
->raw
->data
,
4347 .size
= sizeof(u32
),
4350 perf_output_put(handle
, raw
);
4355 void perf_prepare_sample(struct perf_event_header
*header
,
4356 struct perf_sample_data
*data
,
4357 struct perf_event
*event
,
4358 struct pt_regs
*regs
)
4360 u64 sample_type
= event
->attr
.sample_type
;
4362 header
->type
= PERF_RECORD_SAMPLE
;
4363 header
->size
= sizeof(*header
) + event
->header_size
;
4366 header
->misc
|= perf_misc_flags(regs
);
4368 __perf_event_header__init_id(header
, data
, event
);
4370 if (sample_type
& PERF_SAMPLE_IP
)
4371 data
->ip
= perf_instruction_pointer(regs
);
4373 if (sample_type
& PERF_SAMPLE_CALLCHAIN
) {
4376 data
->callchain
= perf_callchain(regs
);
4378 if (data
->callchain
)
4379 size
+= data
->callchain
->nr
;
4381 header
->size
+= size
* sizeof(u64
);
4384 if (sample_type
& PERF_SAMPLE_RAW
) {
4385 int size
= sizeof(u32
);
4388 size
+= data
->raw
->size
;
4390 size
+= sizeof(u32
);
4392 WARN_ON_ONCE(size
& (sizeof(u64
)-1));
4393 header
->size
+= size
;
4397 static void perf_event_output(struct perf_event
*event
, int nmi
,
4398 struct perf_sample_data
*data
,
4399 struct pt_regs
*regs
)
4401 struct perf_output_handle handle
;
4402 struct perf_event_header header
;
4404 /* protect the callchain buffers */
4407 perf_prepare_sample(&header
, data
, event
, regs
);
4409 if (perf_output_begin(&handle
, event
, header
.size
, nmi
, 1))
4412 perf_output_sample(&handle
, &header
, data
, event
);
4414 perf_output_end(&handle
);
4424 struct perf_read_event
{
4425 struct perf_event_header header
;
4432 perf_event_read_event(struct perf_event
*event
,
4433 struct task_struct
*task
)
4435 struct perf_output_handle handle
;
4436 struct perf_sample_data sample
;
4437 struct perf_read_event read_event
= {
4439 .type
= PERF_RECORD_READ
,
4441 .size
= sizeof(read_event
) + event
->read_size
,
4443 .pid
= perf_event_pid(event
, task
),
4444 .tid
= perf_event_tid(event
, task
),
4448 perf_event_header__init_id(&read_event
.header
, &sample
, event
);
4449 ret
= perf_output_begin(&handle
, event
, read_event
.header
.size
, 0, 0);
4453 perf_output_put(&handle
, read_event
);
4454 perf_output_read(&handle
, event
);
4455 perf_event__output_id_sample(event
, &handle
, &sample
);
4457 perf_output_end(&handle
);
4461 * task tracking -- fork/exit
4463 * enabled by: attr.comm | attr.mmap | attr.mmap_data | attr.task
4466 struct perf_task_event
{
4467 struct task_struct
*task
;
4468 struct perf_event_context
*task_ctx
;
4471 struct perf_event_header header
;
4481 static void perf_event_task_output(struct perf_event
*event
,
4482 struct perf_task_event
*task_event
)
4484 struct perf_output_handle handle
;
4485 struct perf_sample_data sample
;
4486 struct task_struct
*task
= task_event
->task
;
4487 int ret
, size
= task_event
->event_id
.header
.size
;
4489 perf_event_header__init_id(&task_event
->event_id
.header
, &sample
, event
);
4491 ret
= perf_output_begin(&handle
, event
,
4492 task_event
->event_id
.header
.size
, 0, 0);
4496 task_event
->event_id
.pid
= perf_event_pid(event
, task
);
4497 task_event
->event_id
.ppid
= perf_event_pid(event
, current
);
4499 task_event
->event_id
.tid
= perf_event_tid(event
, task
);
4500 task_event
->event_id
.ptid
= perf_event_tid(event
, current
);
4502 perf_output_put(&handle
, task_event
->event_id
);
4504 perf_event__output_id_sample(event
, &handle
, &sample
);
4506 perf_output_end(&handle
);
4508 task_event
->event_id
.header
.size
= size
;
4511 static int perf_event_task_match(struct perf_event
*event
)
4513 if (event
->state
< PERF_EVENT_STATE_INACTIVE
)
4516 if (!event_filter_match(event
))
4519 if (event
->attr
.comm
|| event
->attr
.mmap
||
4520 event
->attr
.mmap_data
|| event
->attr
.task
)
4526 static void perf_event_task_ctx(struct perf_event_context
*ctx
,
4527 struct perf_task_event
*task_event
)
4529 struct perf_event
*event
;
4531 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
4532 if (perf_event_task_match(event
))
4533 perf_event_task_output(event
, task_event
);
4537 static void perf_event_task_event(struct perf_task_event
*task_event
)
4539 struct perf_cpu_context
*cpuctx
;
4540 struct perf_event_context
*ctx
;
4545 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
4546 cpuctx
= get_cpu_ptr(pmu
->pmu_cpu_context
);
4547 if (cpuctx
->active_pmu
!= pmu
)
4549 perf_event_task_ctx(&cpuctx
->ctx
, task_event
);
4551 ctx
= task_event
->task_ctx
;
4553 ctxn
= pmu
->task_ctx_nr
;
4556 ctx
= rcu_dereference(current
->perf_event_ctxp
[ctxn
]);
4559 perf_event_task_ctx(ctx
, task_event
);
4561 put_cpu_ptr(pmu
->pmu_cpu_context
);
4566 static void perf_event_task(struct task_struct
*task
,
4567 struct perf_event_context
*task_ctx
,
4570 struct perf_task_event task_event
;
4572 if (!atomic_read(&nr_comm_events
) &&
4573 !atomic_read(&nr_mmap_events
) &&
4574 !atomic_read(&nr_task_events
))
4577 task_event
= (struct perf_task_event
){
4579 .task_ctx
= task_ctx
,
4582 .type
= new ? PERF_RECORD_FORK
: PERF_RECORD_EXIT
,
4584 .size
= sizeof(task_event
.event_id
),
4590 .time
= perf_clock(),
4594 perf_event_task_event(&task_event
);
4597 void perf_event_fork(struct task_struct
*task
)
4599 perf_event_task(task
, NULL
, 1);
4606 struct perf_comm_event
{
4607 struct task_struct
*task
;
4612 struct perf_event_header header
;
4619 static void perf_event_comm_output(struct perf_event
*event
,
4620 struct perf_comm_event
*comm_event
)
4622 struct perf_output_handle handle
;
4623 struct perf_sample_data sample
;
4624 int size
= comm_event
->event_id
.header
.size
;
4627 perf_event_header__init_id(&comm_event
->event_id
.header
, &sample
, event
);
4628 ret
= perf_output_begin(&handle
, event
,
4629 comm_event
->event_id
.header
.size
, 0, 0);
4634 comm_event
->event_id
.pid
= perf_event_pid(event
, comm_event
->task
);
4635 comm_event
->event_id
.tid
= perf_event_tid(event
, comm_event
->task
);
4637 perf_output_put(&handle
, comm_event
->event_id
);
4638 perf_output_copy(&handle
, comm_event
->comm
,
4639 comm_event
->comm_size
);
4641 perf_event__output_id_sample(event
, &handle
, &sample
);
4643 perf_output_end(&handle
);
4645 comm_event
->event_id
.header
.size
= size
;
4648 static int perf_event_comm_match(struct perf_event
*event
)
4650 if (event
->state
< PERF_EVENT_STATE_INACTIVE
)
4653 if (!event_filter_match(event
))
4656 if (event
->attr
.comm
)
4662 static void perf_event_comm_ctx(struct perf_event_context
*ctx
,
4663 struct perf_comm_event
*comm_event
)
4665 struct perf_event
*event
;
4667 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
4668 if (perf_event_comm_match(event
))
4669 perf_event_comm_output(event
, comm_event
);
4673 static void perf_event_comm_event(struct perf_comm_event
*comm_event
)
4675 struct perf_cpu_context
*cpuctx
;
4676 struct perf_event_context
*ctx
;
4677 char comm
[TASK_COMM_LEN
];
4682 memset(comm
, 0, sizeof(comm
));
4683 strlcpy(comm
, comm_event
->task
->comm
, sizeof(comm
));
4684 size
= ALIGN(strlen(comm
)+1, sizeof(u64
));
4686 comm_event
->comm
= comm
;
4687 comm_event
->comm_size
= size
;
4689 comm_event
->event_id
.header
.size
= sizeof(comm_event
->event_id
) + size
;
4691 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
4692 cpuctx
= get_cpu_ptr(pmu
->pmu_cpu_context
);
4693 if (cpuctx
->active_pmu
!= pmu
)
4695 perf_event_comm_ctx(&cpuctx
->ctx
, comm_event
);
4697 ctxn
= pmu
->task_ctx_nr
;
4701 ctx
= rcu_dereference(current
->perf_event_ctxp
[ctxn
]);
4703 perf_event_comm_ctx(ctx
, comm_event
);
4705 put_cpu_ptr(pmu
->pmu_cpu_context
);
4710 void perf_event_comm(struct task_struct
*task
)
4712 struct perf_comm_event comm_event
;
4713 struct perf_event_context
*ctx
;
4716 for_each_task_context_nr(ctxn
) {
4717 ctx
= task
->perf_event_ctxp
[ctxn
];
4721 perf_event_enable_on_exec(ctx
);
4724 if (!atomic_read(&nr_comm_events
))
4727 comm_event
= (struct perf_comm_event
){
4733 .type
= PERF_RECORD_COMM
,
4742 perf_event_comm_event(&comm_event
);
4749 struct perf_mmap_event
{
4750 struct vm_area_struct
*vma
;
4752 const char *file_name
;
4756 struct perf_event_header header
;
4766 static void perf_event_mmap_output(struct perf_event
*event
,
4767 struct perf_mmap_event
*mmap_event
)
4769 struct perf_output_handle handle
;
4770 struct perf_sample_data sample
;
4771 int size
= mmap_event
->event_id
.header
.size
;
4774 perf_event_header__init_id(&mmap_event
->event_id
.header
, &sample
, event
);
4775 ret
= perf_output_begin(&handle
, event
,
4776 mmap_event
->event_id
.header
.size
, 0, 0);
4780 mmap_event
->event_id
.pid
= perf_event_pid(event
, current
);
4781 mmap_event
->event_id
.tid
= perf_event_tid(event
, current
);
4783 perf_output_put(&handle
, mmap_event
->event_id
);
4784 perf_output_copy(&handle
, mmap_event
->file_name
,
4785 mmap_event
->file_size
);
4787 perf_event__output_id_sample(event
, &handle
, &sample
);
4789 perf_output_end(&handle
);
4791 mmap_event
->event_id
.header
.size
= size
;
4794 static int perf_event_mmap_match(struct perf_event
*event
,
4795 struct perf_mmap_event
*mmap_event
,
4798 if (event
->state
< PERF_EVENT_STATE_INACTIVE
)
4801 if (!event_filter_match(event
))
4804 if ((!executable
&& event
->attr
.mmap_data
) ||
4805 (executable
&& event
->attr
.mmap
))
4811 static void perf_event_mmap_ctx(struct perf_event_context
*ctx
,
4812 struct perf_mmap_event
*mmap_event
,
4815 struct perf_event
*event
;
4817 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
4818 if (perf_event_mmap_match(event
, mmap_event
, executable
))
4819 perf_event_mmap_output(event
, mmap_event
);
4823 static void perf_event_mmap_event(struct perf_mmap_event
*mmap_event
)
4825 struct perf_cpu_context
*cpuctx
;
4826 struct perf_event_context
*ctx
;
4827 struct vm_area_struct
*vma
= mmap_event
->vma
;
4828 struct file
*file
= vma
->vm_file
;
4836 memset(tmp
, 0, sizeof(tmp
));
4840 * d_path works from the end of the buffer backwards, so we
4841 * need to add enough zero bytes after the string to handle
4842 * the 64bit alignment we do later.
4844 buf
= kzalloc(PATH_MAX
+ sizeof(u64
), GFP_KERNEL
);
4846 name
= strncpy(tmp
, "//enomem", sizeof(tmp
));
4849 name
= d_path(&file
->f_path
, buf
, PATH_MAX
);
4851 name
= strncpy(tmp
, "//toolong", sizeof(tmp
));
4855 if (arch_vma_name(mmap_event
->vma
)) {
4856 name
= strncpy(tmp
, arch_vma_name(mmap_event
->vma
),
4862 name
= strncpy(tmp
, "[vdso]", sizeof(tmp
));
4864 } else if (vma
->vm_start
<= vma
->vm_mm
->start_brk
&&
4865 vma
->vm_end
>= vma
->vm_mm
->brk
) {
4866 name
= strncpy(tmp
, "[heap]", sizeof(tmp
));
4868 } else if (vma
->vm_start
<= vma
->vm_mm
->start_stack
&&
4869 vma
->vm_end
>= vma
->vm_mm
->start_stack
) {
4870 name
= strncpy(tmp
, "[stack]", sizeof(tmp
));
4874 name
= strncpy(tmp
, "//anon", sizeof(tmp
));
4879 size
= ALIGN(strlen(name
)+1, sizeof(u64
));
4881 mmap_event
->file_name
= name
;
4882 mmap_event
->file_size
= size
;
4884 mmap_event
->event_id
.header
.size
= sizeof(mmap_event
->event_id
) + size
;
4887 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
4888 cpuctx
= get_cpu_ptr(pmu
->pmu_cpu_context
);
4889 if (cpuctx
->active_pmu
!= pmu
)
4891 perf_event_mmap_ctx(&cpuctx
->ctx
, mmap_event
,
4892 vma
->vm_flags
& VM_EXEC
);
4894 ctxn
= pmu
->task_ctx_nr
;
4898 ctx
= rcu_dereference(current
->perf_event_ctxp
[ctxn
]);
4900 perf_event_mmap_ctx(ctx
, mmap_event
,
4901 vma
->vm_flags
& VM_EXEC
);
4904 put_cpu_ptr(pmu
->pmu_cpu_context
);
4911 void perf_event_mmap(struct vm_area_struct
*vma
)
4913 struct perf_mmap_event mmap_event
;
4915 if (!atomic_read(&nr_mmap_events
))
4918 mmap_event
= (struct perf_mmap_event
){
4924 .type
= PERF_RECORD_MMAP
,
4925 .misc
= PERF_RECORD_MISC_USER
,
4930 .start
= vma
->vm_start
,
4931 .len
= vma
->vm_end
- vma
->vm_start
,
4932 .pgoff
= (u64
)vma
->vm_pgoff
<< PAGE_SHIFT
,
4936 perf_event_mmap_event(&mmap_event
);
4940 * IRQ throttle logging
4943 static void perf_log_throttle(struct perf_event
*event
, int enable
)
4945 struct perf_output_handle handle
;
4946 struct perf_sample_data sample
;
4950 struct perf_event_header header
;
4954 } throttle_event
= {
4956 .type
= PERF_RECORD_THROTTLE
,
4958 .size
= sizeof(throttle_event
),
4960 .time
= perf_clock(),
4961 .id
= primary_event_id(event
),
4962 .stream_id
= event
->id
,
4966 throttle_event
.header
.type
= PERF_RECORD_UNTHROTTLE
;
4968 perf_event_header__init_id(&throttle_event
.header
, &sample
, event
);
4970 ret
= perf_output_begin(&handle
, event
,
4971 throttle_event
.header
.size
, 1, 0);
4975 perf_output_put(&handle
, throttle_event
);
4976 perf_event__output_id_sample(event
, &handle
, &sample
);
4977 perf_output_end(&handle
);
4981 * Generic event overflow handling, sampling.
4984 static int __perf_event_overflow(struct perf_event
*event
, int nmi
,
4985 int throttle
, struct perf_sample_data
*data
,
4986 struct pt_regs
*regs
)
4988 int events
= atomic_read(&event
->event_limit
);
4989 struct hw_perf_event
*hwc
= &event
->hw
;
4993 * Non-sampling counters might still use the PMI to fold short
4994 * hardware counters, ignore those.
4996 if (unlikely(!is_sampling_event(event
)))
4999 if (unlikely(hwc
->interrupts
>= max_samples_per_tick
)) {
5001 hwc
->interrupts
= MAX_INTERRUPTS
;
5002 perf_log_throttle(event
, 0);
5008 if (event
->attr
.freq
) {
5009 u64 now
= perf_clock();
5010 s64 delta
= now
- hwc
->freq_time_stamp
;
5012 hwc
->freq_time_stamp
= now
;
5014 if (delta
> 0 && delta
< 2*TICK_NSEC
)
5015 perf_adjust_period(event
, delta
, hwc
->last_period
);
5019 * XXX event_limit might not quite work as expected on inherited
5023 event
->pending_kill
= POLL_IN
;
5024 if (events
&& atomic_dec_and_test(&event
->event_limit
)) {
5026 event
->pending_kill
= POLL_HUP
;
5028 event
->pending_disable
= 1;
5029 irq_work_queue(&event
->pending
);
5031 perf_event_disable(event
);
5034 if (event
->overflow_handler
)
5035 event
->overflow_handler(event
, nmi
, data
, regs
);
5037 perf_event_output(event
, nmi
, data
, regs
);
5042 int perf_event_overflow(struct perf_event
*event
, int nmi
,
5043 struct perf_sample_data
*data
,
5044 struct pt_regs
*regs
)
5046 return __perf_event_overflow(event
, nmi
, 1, data
, regs
);
5050 * Generic software event infrastructure
5053 struct swevent_htable
{
5054 struct swevent_hlist
*swevent_hlist
;
5055 struct mutex hlist_mutex
;
5058 /* Recursion avoidance in each contexts */
5059 int recursion
[PERF_NR_CONTEXTS
];
5062 static DEFINE_PER_CPU(struct swevent_htable
, swevent_htable
);
5065 * We directly increment event->count and keep a second value in
5066 * event->hw.period_left to count intervals. This period event
5067 * is kept in the range [-sample_period, 0] so that we can use the
5071 static u64
perf_swevent_set_period(struct perf_event
*event
)
5073 struct hw_perf_event
*hwc
= &event
->hw
;
5074 u64 period
= hwc
->last_period
;
5078 hwc
->last_period
= hwc
->sample_period
;
5081 old
= val
= local64_read(&hwc
->period_left
);
5085 nr
= div64_u64(period
+ val
, period
);
5086 offset
= nr
* period
;
5088 if (local64_cmpxchg(&hwc
->period_left
, old
, val
) != old
)
5094 static void perf_swevent_overflow(struct perf_event
*event
, u64 overflow
,
5095 int nmi
, struct perf_sample_data
*data
,
5096 struct pt_regs
*regs
)
5098 struct hw_perf_event
*hwc
= &event
->hw
;
5101 data
->period
= event
->hw
.last_period
;
5103 overflow
= perf_swevent_set_period(event
);
5105 if (hwc
->interrupts
== MAX_INTERRUPTS
)
5108 for (; overflow
; overflow
--) {
5109 if (__perf_event_overflow(event
, nmi
, throttle
,
5112 * We inhibit the overflow from happening when
5113 * hwc->interrupts == MAX_INTERRUPTS.
5121 static void perf_swevent_event(struct perf_event
*event
, u64 nr
,
5122 int nmi
, struct perf_sample_data
*data
,
5123 struct pt_regs
*regs
)
5125 struct hw_perf_event
*hwc
= &event
->hw
;
5127 local64_add(nr
, &event
->count
);
5132 if (!is_sampling_event(event
))
5135 if (nr
== 1 && hwc
->sample_period
== 1 && !event
->attr
.freq
)
5136 return perf_swevent_overflow(event
, 1, nmi
, data
, regs
);
5138 if (local64_add_negative(nr
, &hwc
->period_left
))
5141 perf_swevent_overflow(event
, 0, nmi
, data
, regs
);
5144 static int perf_exclude_event(struct perf_event
*event
,
5145 struct pt_regs
*regs
)
5147 if (event
->hw
.state
& PERF_HES_STOPPED
)
5151 if (event
->attr
.exclude_user
&& user_mode(regs
))
5154 if (event
->attr
.exclude_kernel
&& !user_mode(regs
))
5161 static int perf_swevent_match(struct perf_event
*event
,
5162 enum perf_type_id type
,
5164 struct perf_sample_data
*data
,
5165 struct pt_regs
*regs
)
5167 if (event
->attr
.type
!= type
)
5170 if (event
->attr
.config
!= event_id
)
5173 if (perf_exclude_event(event
, regs
))
5179 static inline u64
swevent_hash(u64 type
, u32 event_id
)
5181 u64 val
= event_id
| (type
<< 32);
5183 return hash_64(val
, SWEVENT_HLIST_BITS
);
5186 static inline struct hlist_head
*
5187 __find_swevent_head(struct swevent_hlist
*hlist
, u64 type
, u32 event_id
)
5189 u64 hash
= swevent_hash(type
, event_id
);
5191 return &hlist
->heads
[hash
];
5194 /* For the read side: events when they trigger */
5195 static inline struct hlist_head
*
5196 find_swevent_head_rcu(struct swevent_htable
*swhash
, u64 type
, u32 event_id
)
5198 struct swevent_hlist
*hlist
;
5200 hlist
= rcu_dereference(swhash
->swevent_hlist
);
5204 return __find_swevent_head(hlist
, type
, event_id
);
5207 /* For the event head insertion and removal in the hlist */
5208 static inline struct hlist_head
*
5209 find_swevent_head(struct swevent_htable
*swhash
, struct perf_event
*event
)
5211 struct swevent_hlist
*hlist
;
5212 u32 event_id
= event
->attr
.config
;
5213 u64 type
= event
->attr
.type
;
5216 * Event scheduling is always serialized against hlist allocation
5217 * and release. Which makes the protected version suitable here.
5218 * The context lock guarantees that.
5220 hlist
= rcu_dereference_protected(swhash
->swevent_hlist
,
5221 lockdep_is_held(&event
->ctx
->lock
));
5225 return __find_swevent_head(hlist
, type
, event_id
);
5228 static void do_perf_sw_event(enum perf_type_id type
, u32 event_id
,
5230 struct perf_sample_data
*data
,
5231 struct pt_regs
*regs
)
5233 struct swevent_htable
*swhash
= &__get_cpu_var(swevent_htable
);
5234 struct perf_event
*event
;
5235 struct hlist_node
*node
;
5236 struct hlist_head
*head
;
5239 head
= find_swevent_head_rcu(swhash
, type
, event_id
);
5243 hlist_for_each_entry_rcu(event
, node
, head
, hlist_entry
) {
5244 if (perf_swevent_match(event
, type
, event_id
, data
, regs
))
5245 perf_swevent_event(event
, nr
, nmi
, data
, regs
);
5251 int perf_swevent_get_recursion_context(void)
5253 struct swevent_htable
*swhash
= &__get_cpu_var(swevent_htable
);
5255 return get_recursion_context(swhash
->recursion
);
5257 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context
);
5259 inline void perf_swevent_put_recursion_context(int rctx
)
5261 struct swevent_htable
*swhash
= &__get_cpu_var(swevent_htable
);
5263 put_recursion_context(swhash
->recursion
, rctx
);
5266 void __perf_sw_event(u32 event_id
, u64 nr
, int nmi
,
5267 struct pt_regs
*regs
, u64 addr
)
5269 struct perf_sample_data data
;
5272 preempt_disable_notrace();
5273 rctx
= perf_swevent_get_recursion_context();
5277 perf_sample_data_init(&data
, addr
);
5279 do_perf_sw_event(PERF_TYPE_SOFTWARE
, event_id
, nr
, nmi
, &data
, regs
);
5281 perf_swevent_put_recursion_context(rctx
);
5282 preempt_enable_notrace();
5285 static void perf_swevent_read(struct perf_event
*event
)
5289 static int perf_swevent_add(struct perf_event
*event
, int flags
)
5291 struct swevent_htable
*swhash
= &__get_cpu_var(swevent_htable
);
5292 struct hw_perf_event
*hwc
= &event
->hw
;
5293 struct hlist_head
*head
;
5295 if (is_sampling_event(event
)) {
5296 hwc
->last_period
= hwc
->sample_period
;
5297 perf_swevent_set_period(event
);
5300 hwc
->state
= !(flags
& PERF_EF_START
);
5302 head
= find_swevent_head(swhash
, event
);
5303 if (WARN_ON_ONCE(!head
))
5306 hlist_add_head_rcu(&event
->hlist_entry
, head
);
5311 static void perf_swevent_del(struct perf_event
*event
, int flags
)
5313 hlist_del_rcu(&event
->hlist_entry
);
5316 static void perf_swevent_start(struct perf_event
*event
, int flags
)
5318 event
->hw
.state
= 0;
5321 static void perf_swevent_stop(struct perf_event
*event
, int flags
)
5323 event
->hw
.state
= PERF_HES_STOPPED
;
5326 /* Deref the hlist from the update side */
5327 static inline struct swevent_hlist
*
5328 swevent_hlist_deref(struct swevent_htable
*swhash
)
5330 return rcu_dereference_protected(swhash
->swevent_hlist
,
5331 lockdep_is_held(&swhash
->hlist_mutex
));
5334 static void swevent_hlist_release_rcu(struct rcu_head
*rcu_head
)
5336 struct swevent_hlist
*hlist
;
5338 hlist
= container_of(rcu_head
, struct swevent_hlist
, rcu_head
);
5342 static void swevent_hlist_release(struct swevent_htable
*swhash
)
5344 struct swevent_hlist
*hlist
= swevent_hlist_deref(swhash
);
5349 rcu_assign_pointer(swhash
->swevent_hlist
, NULL
);
5350 call_rcu(&hlist
->rcu_head
, swevent_hlist_release_rcu
);
5353 static void swevent_hlist_put_cpu(struct perf_event
*event
, int cpu
)
5355 struct swevent_htable
*swhash
= &per_cpu(swevent_htable
, cpu
);
5357 mutex_lock(&swhash
->hlist_mutex
);
5359 if (!--swhash
->hlist_refcount
)
5360 swevent_hlist_release(swhash
);
5362 mutex_unlock(&swhash
->hlist_mutex
);
5365 static void swevent_hlist_put(struct perf_event
*event
)
5369 if (event
->cpu
!= -1) {
5370 swevent_hlist_put_cpu(event
, event
->cpu
);
5374 for_each_possible_cpu(cpu
)
5375 swevent_hlist_put_cpu(event
, cpu
);
5378 static int swevent_hlist_get_cpu(struct perf_event
*event
, int cpu
)
5380 struct swevent_htable
*swhash
= &per_cpu(swevent_htable
, cpu
);
5383 mutex_lock(&swhash
->hlist_mutex
);
5385 if (!swevent_hlist_deref(swhash
) && cpu_online(cpu
)) {
5386 struct swevent_hlist
*hlist
;
5388 hlist
= kzalloc(sizeof(*hlist
), GFP_KERNEL
);
5393 rcu_assign_pointer(swhash
->swevent_hlist
, hlist
);
5395 swhash
->hlist_refcount
++;
5397 mutex_unlock(&swhash
->hlist_mutex
);
5402 static int swevent_hlist_get(struct perf_event
*event
)
5405 int cpu
, failed_cpu
;
5407 if (event
->cpu
!= -1)
5408 return swevent_hlist_get_cpu(event
, event
->cpu
);
5411 for_each_possible_cpu(cpu
) {
5412 err
= swevent_hlist_get_cpu(event
, cpu
);
5422 for_each_possible_cpu(cpu
) {
5423 if (cpu
== failed_cpu
)
5425 swevent_hlist_put_cpu(event
, cpu
);
5432 struct jump_label_key perf_swevent_enabled
[PERF_COUNT_SW_MAX
];
5434 static void sw_perf_event_destroy(struct perf_event
*event
)
5436 u64 event_id
= event
->attr
.config
;
5438 WARN_ON(event
->parent
);
5440 jump_label_dec(&perf_swevent_enabled
[event_id
]);
5441 swevent_hlist_put(event
);
5444 static int perf_swevent_init(struct perf_event
*event
)
5446 int event_id
= event
->attr
.config
;
5448 if (event
->attr
.type
!= PERF_TYPE_SOFTWARE
)
5452 case PERF_COUNT_SW_CPU_CLOCK
:
5453 case PERF_COUNT_SW_TASK_CLOCK
:
5460 if (event_id
>= PERF_COUNT_SW_MAX
)
5463 if (!event
->parent
) {
5466 err
= swevent_hlist_get(event
);
5470 jump_label_inc(&perf_swevent_enabled
[event_id
]);
5471 event
->destroy
= sw_perf_event_destroy
;
5477 static struct pmu perf_swevent
= {
5478 .task_ctx_nr
= perf_sw_context
,
5480 .event_init
= perf_swevent_init
,
5481 .add
= perf_swevent_add
,
5482 .del
= perf_swevent_del
,
5483 .start
= perf_swevent_start
,
5484 .stop
= perf_swevent_stop
,
5485 .read
= perf_swevent_read
,
5488 #ifdef CONFIG_EVENT_TRACING
5490 static int perf_tp_filter_match(struct perf_event
*event
,
5491 struct perf_sample_data
*data
)
5493 void *record
= data
->raw
->data
;
5495 if (likely(!event
->filter
) || filter_match_preds(event
->filter
, record
))
5500 static int perf_tp_event_match(struct perf_event
*event
,
5501 struct perf_sample_data
*data
,
5502 struct pt_regs
*regs
)
5504 if (event
->hw
.state
& PERF_HES_STOPPED
)
5507 * All tracepoints are from kernel-space.
5509 if (event
->attr
.exclude_kernel
)
5512 if (!perf_tp_filter_match(event
, data
))
5518 void perf_tp_event(u64 addr
, u64 count
, void *record
, int entry_size
,
5519 struct pt_regs
*regs
, struct hlist_head
*head
, int rctx
)
5521 struct perf_sample_data data
;
5522 struct perf_event
*event
;
5523 struct hlist_node
*node
;
5525 struct perf_raw_record raw
= {
5530 perf_sample_data_init(&data
, addr
);
5533 hlist_for_each_entry_rcu(event
, node
, head
, hlist_entry
) {
5534 if (perf_tp_event_match(event
, &data
, regs
))
5535 perf_swevent_event(event
, count
, 1, &data
, regs
);
5538 perf_swevent_put_recursion_context(rctx
);
5540 EXPORT_SYMBOL_GPL(perf_tp_event
);
5542 static void tp_perf_event_destroy(struct perf_event
*event
)
5544 perf_trace_destroy(event
);
5547 static int perf_tp_event_init(struct perf_event
*event
)
5551 if (event
->attr
.type
!= PERF_TYPE_TRACEPOINT
)
5554 err
= perf_trace_init(event
);
5558 event
->destroy
= tp_perf_event_destroy
;
5563 static struct pmu perf_tracepoint
= {
5564 .task_ctx_nr
= perf_sw_context
,
5566 .event_init
= perf_tp_event_init
,
5567 .add
= perf_trace_add
,
5568 .del
= perf_trace_del
,
5569 .start
= perf_swevent_start
,
5570 .stop
= perf_swevent_stop
,
5571 .read
= perf_swevent_read
,
5574 static inline void perf_tp_register(void)
5576 perf_pmu_register(&perf_tracepoint
, "tracepoint", PERF_TYPE_TRACEPOINT
);
5579 static int perf_event_set_filter(struct perf_event
*event
, void __user
*arg
)
5584 if (event
->attr
.type
!= PERF_TYPE_TRACEPOINT
)
5587 filter_str
= strndup_user(arg
, PAGE_SIZE
);
5588 if (IS_ERR(filter_str
))
5589 return PTR_ERR(filter_str
);
5591 ret
= ftrace_profile_set_filter(event
, event
->attr
.config
, filter_str
);
5597 static void perf_event_free_filter(struct perf_event
*event
)
5599 ftrace_profile_free_filter(event
);
5604 static inline void perf_tp_register(void)
5608 static int perf_event_set_filter(struct perf_event
*event
, void __user
*arg
)
5613 static void perf_event_free_filter(struct perf_event
*event
)
5617 #endif /* CONFIG_EVENT_TRACING */
5619 #ifdef CONFIG_HAVE_HW_BREAKPOINT
5620 void perf_bp_event(struct perf_event
*bp
, void *data
)
5622 struct perf_sample_data sample
;
5623 struct pt_regs
*regs
= data
;
5625 perf_sample_data_init(&sample
, bp
->attr
.bp_addr
);
5627 if (!bp
->hw
.state
&& !perf_exclude_event(bp
, regs
))
5628 perf_swevent_event(bp
, 1, 1, &sample
, regs
);
5633 * hrtimer based swevent callback
5636 static enum hrtimer_restart
perf_swevent_hrtimer(struct hrtimer
*hrtimer
)
5638 enum hrtimer_restart ret
= HRTIMER_RESTART
;
5639 struct perf_sample_data data
;
5640 struct pt_regs
*regs
;
5641 struct perf_event
*event
;
5644 event
= container_of(hrtimer
, struct perf_event
, hw
.hrtimer
);
5646 if (event
->state
!= PERF_EVENT_STATE_ACTIVE
)
5647 return HRTIMER_NORESTART
;
5649 event
->pmu
->read(event
);
5651 perf_sample_data_init(&data
, 0);
5652 data
.period
= event
->hw
.last_period
;
5653 regs
= get_irq_regs();
5655 if (regs
&& !perf_exclude_event(event
, regs
)) {
5656 if (!(event
->attr
.exclude_idle
&& current
->pid
== 0))
5657 if (perf_event_overflow(event
, 0, &data
, regs
))
5658 ret
= HRTIMER_NORESTART
;
5661 period
= max_t(u64
, 10000, event
->hw
.sample_period
);
5662 hrtimer_forward_now(hrtimer
, ns_to_ktime(period
));
5667 static void perf_swevent_start_hrtimer(struct perf_event
*event
)
5669 struct hw_perf_event
*hwc
= &event
->hw
;
5672 if (!is_sampling_event(event
))
5675 period
= local64_read(&hwc
->period_left
);
5680 local64_set(&hwc
->period_left
, 0);
5682 period
= max_t(u64
, 10000, hwc
->sample_period
);
5684 __hrtimer_start_range_ns(&hwc
->hrtimer
,
5685 ns_to_ktime(period
), 0,
5686 HRTIMER_MODE_REL_PINNED
, 0);
5689 static void perf_swevent_cancel_hrtimer(struct perf_event
*event
)
5691 struct hw_perf_event
*hwc
= &event
->hw
;
5693 if (is_sampling_event(event
)) {
5694 ktime_t remaining
= hrtimer_get_remaining(&hwc
->hrtimer
);
5695 local64_set(&hwc
->period_left
, ktime_to_ns(remaining
));
5697 hrtimer_cancel(&hwc
->hrtimer
);
5701 static void perf_swevent_init_hrtimer(struct perf_event
*event
)
5703 struct hw_perf_event
*hwc
= &event
->hw
;
5705 if (!is_sampling_event(event
))
5708 hrtimer_init(&hwc
->hrtimer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
5709 hwc
->hrtimer
.function
= perf_swevent_hrtimer
;
5712 * Since hrtimers have a fixed rate, we can do a static freq->period
5713 * mapping and avoid the whole period adjust feedback stuff.
5715 if (event
->attr
.freq
) {
5716 long freq
= event
->attr
.sample_freq
;
5718 event
->attr
.sample_period
= NSEC_PER_SEC
/ freq
;
5719 hwc
->sample_period
= event
->attr
.sample_period
;
5720 local64_set(&hwc
->period_left
, hwc
->sample_period
);
5721 event
->attr
.freq
= 0;
5726 * Software event: cpu wall time clock
5729 static void cpu_clock_event_update(struct perf_event
*event
)
5734 now
= local_clock();
5735 prev
= local64_xchg(&event
->hw
.prev_count
, now
);
5736 local64_add(now
- prev
, &event
->count
);
5739 static void cpu_clock_event_start(struct perf_event
*event
, int flags
)
5741 local64_set(&event
->hw
.prev_count
, local_clock());
5742 perf_swevent_start_hrtimer(event
);
5745 static void cpu_clock_event_stop(struct perf_event
*event
, int flags
)
5747 perf_swevent_cancel_hrtimer(event
);
5748 cpu_clock_event_update(event
);
5751 static int cpu_clock_event_add(struct perf_event
*event
, int flags
)
5753 if (flags
& PERF_EF_START
)
5754 cpu_clock_event_start(event
, flags
);
5759 static void cpu_clock_event_del(struct perf_event
*event
, int flags
)
5761 cpu_clock_event_stop(event
, flags
);
5764 static void cpu_clock_event_read(struct perf_event
*event
)
5766 cpu_clock_event_update(event
);
5769 static int cpu_clock_event_init(struct perf_event
*event
)
5771 if (event
->attr
.type
!= PERF_TYPE_SOFTWARE
)
5774 if (event
->attr
.config
!= PERF_COUNT_SW_CPU_CLOCK
)
5777 perf_swevent_init_hrtimer(event
);
5782 static struct pmu perf_cpu_clock
= {
5783 .task_ctx_nr
= perf_sw_context
,
5785 .event_init
= cpu_clock_event_init
,
5786 .add
= cpu_clock_event_add
,
5787 .del
= cpu_clock_event_del
,
5788 .start
= cpu_clock_event_start
,
5789 .stop
= cpu_clock_event_stop
,
5790 .read
= cpu_clock_event_read
,
5794 * Software event: task time clock
5797 static void task_clock_event_update(struct perf_event
*event
, u64 now
)
5802 prev
= local64_xchg(&event
->hw
.prev_count
, now
);
5804 local64_add(delta
, &event
->count
);
5807 static void task_clock_event_start(struct perf_event
*event
, int flags
)
5809 local64_set(&event
->hw
.prev_count
, event
->ctx
->time
);
5810 perf_swevent_start_hrtimer(event
);
5813 static void task_clock_event_stop(struct perf_event
*event
, int flags
)
5815 perf_swevent_cancel_hrtimer(event
);
5816 task_clock_event_update(event
, event
->ctx
->time
);
5819 static int task_clock_event_add(struct perf_event
*event
, int flags
)
5821 if (flags
& PERF_EF_START
)
5822 task_clock_event_start(event
, flags
);
5827 static void task_clock_event_del(struct perf_event
*event
, int flags
)
5829 task_clock_event_stop(event
, PERF_EF_UPDATE
);
5832 static void task_clock_event_read(struct perf_event
*event
)
5834 u64 now
= perf_clock();
5835 u64 delta
= now
- event
->ctx
->timestamp
;
5836 u64 time
= event
->ctx
->time
+ delta
;
5838 task_clock_event_update(event
, time
);
5841 static int task_clock_event_init(struct perf_event
*event
)
5843 if (event
->attr
.type
!= PERF_TYPE_SOFTWARE
)
5846 if (event
->attr
.config
!= PERF_COUNT_SW_TASK_CLOCK
)
5849 perf_swevent_init_hrtimer(event
);
5854 static struct pmu perf_task_clock
= {
5855 .task_ctx_nr
= perf_sw_context
,
5857 .event_init
= task_clock_event_init
,
5858 .add
= task_clock_event_add
,
5859 .del
= task_clock_event_del
,
5860 .start
= task_clock_event_start
,
5861 .stop
= task_clock_event_stop
,
5862 .read
= task_clock_event_read
,
5865 static void perf_pmu_nop_void(struct pmu
*pmu
)
5869 static int perf_pmu_nop_int(struct pmu
*pmu
)
5874 static void perf_pmu_start_txn(struct pmu
*pmu
)
5876 perf_pmu_disable(pmu
);
5879 static int perf_pmu_commit_txn(struct pmu
*pmu
)
5881 perf_pmu_enable(pmu
);
5885 static void perf_pmu_cancel_txn(struct pmu
*pmu
)
5887 perf_pmu_enable(pmu
);
5891 * Ensures all contexts with the same task_ctx_nr have the same
5892 * pmu_cpu_context too.
5894 static void *find_pmu_context(int ctxn
)
5901 list_for_each_entry(pmu
, &pmus
, entry
) {
5902 if (pmu
->task_ctx_nr
== ctxn
)
5903 return pmu
->pmu_cpu_context
;
5909 static void update_pmu_context(struct pmu
*pmu
, struct pmu
*old_pmu
)
5913 for_each_possible_cpu(cpu
) {
5914 struct perf_cpu_context
*cpuctx
;
5916 cpuctx
= per_cpu_ptr(pmu
->pmu_cpu_context
, cpu
);
5918 if (cpuctx
->active_pmu
== old_pmu
)
5919 cpuctx
->active_pmu
= pmu
;
5923 static void free_pmu_context(struct pmu
*pmu
)
5927 mutex_lock(&pmus_lock
);
5929 * Like a real lame refcount.
5931 list_for_each_entry(i
, &pmus
, entry
) {
5932 if (i
->pmu_cpu_context
== pmu
->pmu_cpu_context
) {
5933 update_pmu_context(i
, pmu
);
5938 free_percpu(pmu
->pmu_cpu_context
);
5940 mutex_unlock(&pmus_lock
);
5942 static struct idr pmu_idr
;
5945 type_show(struct device
*dev
, struct device_attribute
*attr
, char *page
)
5947 struct pmu
*pmu
= dev_get_drvdata(dev
);
5949 return snprintf(page
, PAGE_SIZE
-1, "%d\n", pmu
->type
);
5952 static struct device_attribute pmu_dev_attrs
[] = {
5957 static int pmu_bus_running
;
5958 static struct bus_type pmu_bus
= {
5959 .name
= "event_source",
5960 .dev_attrs
= pmu_dev_attrs
,
5963 static void pmu_dev_release(struct device
*dev
)
5968 static int pmu_dev_alloc(struct pmu
*pmu
)
5972 pmu
->dev
= kzalloc(sizeof(struct device
), GFP_KERNEL
);
5976 device_initialize(pmu
->dev
);
5977 ret
= dev_set_name(pmu
->dev
, "%s", pmu
->name
);
5981 dev_set_drvdata(pmu
->dev
, pmu
);
5982 pmu
->dev
->bus
= &pmu_bus
;
5983 pmu
->dev
->release
= pmu_dev_release
;
5984 ret
= device_add(pmu
->dev
);
5992 put_device(pmu
->dev
);
5996 static struct lock_class_key cpuctx_mutex
;
5998 int perf_pmu_register(struct pmu
*pmu
, char *name
, int type
)
6002 mutex_lock(&pmus_lock
);
6004 pmu
->pmu_disable_count
= alloc_percpu(int);
6005 if (!pmu
->pmu_disable_count
)
6014 int err
= idr_pre_get(&pmu_idr
, GFP_KERNEL
);
6018 err
= idr_get_new_above(&pmu_idr
, pmu
, PERF_TYPE_MAX
, &type
);
6026 if (pmu_bus_running
) {
6027 ret
= pmu_dev_alloc(pmu
);
6033 pmu
->pmu_cpu_context
= find_pmu_context(pmu
->task_ctx_nr
);
6034 if (pmu
->pmu_cpu_context
)
6035 goto got_cpu_context
;
6037 pmu
->pmu_cpu_context
= alloc_percpu(struct perf_cpu_context
);
6038 if (!pmu
->pmu_cpu_context
)
6041 for_each_possible_cpu(cpu
) {
6042 struct perf_cpu_context
*cpuctx
;
6044 cpuctx
= per_cpu_ptr(pmu
->pmu_cpu_context
, cpu
);
6045 __perf_event_init_context(&cpuctx
->ctx
);
6046 lockdep_set_class(&cpuctx
->ctx
.mutex
, &cpuctx_mutex
);
6047 cpuctx
->ctx
.type
= cpu_context
;
6048 cpuctx
->ctx
.pmu
= pmu
;
6049 cpuctx
->jiffies_interval
= 1;
6050 INIT_LIST_HEAD(&cpuctx
->rotation_list
);
6051 cpuctx
->active_pmu
= pmu
;
6055 if (!pmu
->start_txn
) {
6056 if (pmu
->pmu_enable
) {
6058 * If we have pmu_enable/pmu_disable calls, install
6059 * transaction stubs that use that to try and batch
6060 * hardware accesses.
6062 pmu
->start_txn
= perf_pmu_start_txn
;
6063 pmu
->commit_txn
= perf_pmu_commit_txn
;
6064 pmu
->cancel_txn
= perf_pmu_cancel_txn
;
6066 pmu
->start_txn
= perf_pmu_nop_void
;
6067 pmu
->commit_txn
= perf_pmu_nop_int
;
6068 pmu
->cancel_txn
= perf_pmu_nop_void
;
6072 if (!pmu
->pmu_enable
) {
6073 pmu
->pmu_enable
= perf_pmu_nop_void
;
6074 pmu
->pmu_disable
= perf_pmu_nop_void
;
6077 list_add_rcu(&pmu
->entry
, &pmus
);
6080 mutex_unlock(&pmus_lock
);
6085 device_del(pmu
->dev
);
6086 put_device(pmu
->dev
);
6089 if (pmu
->type
>= PERF_TYPE_MAX
)
6090 idr_remove(&pmu_idr
, pmu
->type
);
6093 free_percpu(pmu
->pmu_disable_count
);
6097 void perf_pmu_unregister(struct pmu
*pmu
)
6099 mutex_lock(&pmus_lock
);
6100 list_del_rcu(&pmu
->entry
);
6101 mutex_unlock(&pmus_lock
);
6104 * We dereference the pmu list under both SRCU and regular RCU, so
6105 * synchronize against both of those.
6107 synchronize_srcu(&pmus_srcu
);
6110 free_percpu(pmu
->pmu_disable_count
);
6111 if (pmu
->type
>= PERF_TYPE_MAX
)
6112 idr_remove(&pmu_idr
, pmu
->type
);
6113 device_del(pmu
->dev
);
6114 put_device(pmu
->dev
);
6115 free_pmu_context(pmu
);
6118 struct pmu
*perf_init_event(struct perf_event
*event
)
6120 struct pmu
*pmu
= NULL
;
6124 idx
= srcu_read_lock(&pmus_srcu
);
6127 pmu
= idr_find(&pmu_idr
, event
->attr
.type
);
6130 ret
= pmu
->event_init(event
);
6136 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
6137 ret
= pmu
->event_init(event
);
6141 if (ret
!= -ENOENT
) {
6146 pmu
= ERR_PTR(-ENOENT
);
6148 srcu_read_unlock(&pmus_srcu
, idx
);
6154 * Allocate and initialize a event structure
6156 static struct perf_event
*
6157 perf_event_alloc(struct perf_event_attr
*attr
, int cpu
,
6158 struct task_struct
*task
,
6159 struct perf_event
*group_leader
,
6160 struct perf_event
*parent_event
,
6161 perf_overflow_handler_t overflow_handler
)
6164 struct perf_event
*event
;
6165 struct hw_perf_event
*hwc
;
6168 if ((unsigned)cpu
>= nr_cpu_ids
) {
6169 if (!task
|| cpu
!= -1)
6170 return ERR_PTR(-EINVAL
);
6173 event
= kzalloc(sizeof(*event
), GFP_KERNEL
);
6175 return ERR_PTR(-ENOMEM
);
6178 * Single events are their own group leaders, with an
6179 * empty sibling list:
6182 group_leader
= event
;
6184 mutex_init(&event
->child_mutex
);
6185 INIT_LIST_HEAD(&event
->child_list
);
6187 INIT_LIST_HEAD(&event
->group_entry
);
6188 INIT_LIST_HEAD(&event
->event_entry
);
6189 INIT_LIST_HEAD(&event
->sibling_list
);
6190 init_waitqueue_head(&event
->waitq
);
6191 init_irq_work(&event
->pending
, perf_pending_event
);
6193 mutex_init(&event
->mmap_mutex
);
6196 event
->attr
= *attr
;
6197 event
->group_leader
= group_leader
;
6201 event
->parent
= parent_event
;
6203 event
->ns
= get_pid_ns(current
->nsproxy
->pid_ns
);
6204 event
->id
= atomic64_inc_return(&perf_event_id
);
6206 event
->state
= PERF_EVENT_STATE_INACTIVE
;
6209 event
->attach_state
= PERF_ATTACH_TASK
;
6210 #ifdef CONFIG_HAVE_HW_BREAKPOINT
6212 * hw_breakpoint is a bit difficult here..
6214 if (attr
->type
== PERF_TYPE_BREAKPOINT
)
6215 event
->hw
.bp_target
= task
;
6219 if (!overflow_handler
&& parent_event
)
6220 overflow_handler
= parent_event
->overflow_handler
;
6222 event
->overflow_handler
= overflow_handler
;
6225 event
->state
= PERF_EVENT_STATE_OFF
;
6230 hwc
->sample_period
= attr
->sample_period
;
6231 if (attr
->freq
&& attr
->sample_freq
)
6232 hwc
->sample_period
= 1;
6233 hwc
->last_period
= hwc
->sample_period
;
6235 local64_set(&hwc
->period_left
, hwc
->sample_period
);
6238 * we currently do not support PERF_FORMAT_GROUP on inherited events
6240 if (attr
->inherit
&& (attr
->read_format
& PERF_FORMAT_GROUP
))
6243 pmu
= perf_init_event(event
);
6249 else if (IS_ERR(pmu
))
6254 put_pid_ns(event
->ns
);
6256 return ERR_PTR(err
);
6261 if (!event
->parent
) {
6262 if (event
->attach_state
& PERF_ATTACH_TASK
)
6263 jump_label_inc(&perf_sched_events
);
6264 if (event
->attr
.mmap
|| event
->attr
.mmap_data
)
6265 atomic_inc(&nr_mmap_events
);
6266 if (event
->attr
.comm
)
6267 atomic_inc(&nr_comm_events
);
6268 if (event
->attr
.task
)
6269 atomic_inc(&nr_task_events
);
6270 if (event
->attr
.sample_type
& PERF_SAMPLE_CALLCHAIN
) {
6271 err
= get_callchain_buffers();
6274 return ERR_PTR(err
);
6282 static int perf_copy_attr(struct perf_event_attr __user
*uattr
,
6283 struct perf_event_attr
*attr
)
6288 if (!access_ok(VERIFY_WRITE
, uattr
, PERF_ATTR_SIZE_VER0
))
6292 * zero the full structure, so that a short copy will be nice.
6294 memset(attr
, 0, sizeof(*attr
));
6296 ret
= get_user(size
, &uattr
->size
);
6300 if (size
> PAGE_SIZE
) /* silly large */
6303 if (!size
) /* abi compat */
6304 size
= PERF_ATTR_SIZE_VER0
;
6306 if (size
< PERF_ATTR_SIZE_VER0
)
6310 * If we're handed a bigger struct than we know of,
6311 * ensure all the unknown bits are 0 - i.e. new
6312 * user-space does not rely on any kernel feature
6313 * extensions we dont know about yet.
6315 if (size
> sizeof(*attr
)) {
6316 unsigned char __user
*addr
;
6317 unsigned char __user
*end
;
6320 addr
= (void __user
*)uattr
+ sizeof(*attr
);
6321 end
= (void __user
*)uattr
+ size
;
6323 for (; addr
< end
; addr
++) {
6324 ret
= get_user(val
, addr
);
6330 size
= sizeof(*attr
);
6333 ret
= copy_from_user(attr
, uattr
, size
);
6338 * If the type exists, the corresponding creation will verify
6341 if (attr
->type
>= PERF_TYPE_MAX
)
6344 if (attr
->__reserved_1
)
6347 if (attr
->sample_type
& ~(PERF_SAMPLE_MAX
-1))
6350 if (attr
->read_format
& ~(PERF_FORMAT_MAX
-1))
6357 put_user(sizeof(*attr
), &uattr
->size
);
6363 perf_event_set_output(struct perf_event
*event
, struct perf_event
*output_event
)
6365 struct perf_buffer
*buffer
= NULL
, *old_buffer
= NULL
;
6371 /* don't allow circular references */
6372 if (event
== output_event
)
6376 * Don't allow cross-cpu buffers
6378 if (output_event
->cpu
!= event
->cpu
)
6382 * If its not a per-cpu buffer, it must be the same task.
6384 if (output_event
->cpu
== -1 && output_event
->ctx
!= event
->ctx
)
6388 mutex_lock(&event
->mmap_mutex
);
6389 /* Can't redirect output if we've got an active mmap() */
6390 if (atomic_read(&event
->mmap_count
))
6394 /* get the buffer we want to redirect to */
6395 buffer
= perf_buffer_get(output_event
);
6400 old_buffer
= event
->buffer
;
6401 rcu_assign_pointer(event
->buffer
, buffer
);
6404 mutex_unlock(&event
->mmap_mutex
);
6407 perf_buffer_put(old_buffer
);
6413 * sys_perf_event_open - open a performance event, associate it to a task/cpu
6415 * @attr_uptr: event_id type attributes for monitoring/sampling
6418 * @group_fd: group leader event fd
6420 SYSCALL_DEFINE5(perf_event_open
,
6421 struct perf_event_attr __user
*, attr_uptr
,
6422 pid_t
, pid
, int, cpu
, int, group_fd
, unsigned long, flags
)
6424 struct perf_event
*group_leader
= NULL
, *output_event
= NULL
;
6425 struct perf_event
*event
, *sibling
;
6426 struct perf_event_attr attr
;
6427 struct perf_event_context
*ctx
;
6428 struct file
*event_file
= NULL
;
6429 struct file
*group_file
= NULL
;
6430 struct task_struct
*task
= NULL
;
6434 int fput_needed
= 0;
6437 /* for future expandability... */
6438 if (flags
& ~PERF_FLAG_ALL
)
6441 err
= perf_copy_attr(attr_uptr
, &attr
);
6445 if (!attr
.exclude_kernel
) {
6446 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN
))
6451 if (attr
.sample_freq
> sysctl_perf_event_sample_rate
)
6456 * In cgroup mode, the pid argument is used to pass the fd
6457 * opened to the cgroup directory in cgroupfs. The cpu argument
6458 * designates the cpu on which to monitor threads from that
6461 if ((flags
& PERF_FLAG_PID_CGROUP
) && (pid
== -1 || cpu
== -1))
6464 event_fd
= get_unused_fd_flags(O_RDWR
);
6468 if (group_fd
!= -1) {
6469 group_leader
= perf_fget_light(group_fd
, &fput_needed
);
6470 if (IS_ERR(group_leader
)) {
6471 err
= PTR_ERR(group_leader
);
6474 group_file
= group_leader
->filp
;
6475 if (flags
& PERF_FLAG_FD_OUTPUT
)
6476 output_event
= group_leader
;
6477 if (flags
& PERF_FLAG_FD_NO_GROUP
)
6478 group_leader
= NULL
;
6481 if (pid
!= -1 && !(flags
& PERF_FLAG_PID_CGROUP
)) {
6482 task
= find_lively_task_by_vpid(pid
);
6484 err
= PTR_ERR(task
);
6489 event
= perf_event_alloc(&attr
, cpu
, task
, group_leader
, NULL
, NULL
);
6490 if (IS_ERR(event
)) {
6491 err
= PTR_ERR(event
);
6495 if (flags
& PERF_FLAG_PID_CGROUP
) {
6496 err
= perf_cgroup_connect(pid
, event
, &attr
, group_leader
);
6501 * - that has cgroup constraint on event->cpu
6502 * - that may need work on context switch
6504 atomic_inc(&per_cpu(perf_cgroup_events
, event
->cpu
));
6505 jump_label_inc(&perf_sched_events
);
6509 * Special case software events and allow them to be part of
6510 * any hardware group.
6515 (is_software_event(event
) != is_software_event(group_leader
))) {
6516 if (is_software_event(event
)) {
6518 * If event and group_leader are not both a software
6519 * event, and event is, then group leader is not.
6521 * Allow the addition of software events to !software
6522 * groups, this is safe because software events never
6525 pmu
= group_leader
->pmu
;
6526 } else if (is_software_event(group_leader
) &&
6527 (group_leader
->group_flags
& PERF_GROUP_SOFTWARE
)) {
6529 * In case the group is a pure software group, and we
6530 * try to add a hardware event, move the whole group to
6531 * the hardware context.
6538 * Get the target context (task or percpu):
6540 ctx
= find_get_context(pmu
, task
, cpu
);
6547 put_task_struct(task
);
6552 * Look up the group leader (we will attach this event to it):
6558 * Do not allow a recursive hierarchy (this new sibling
6559 * becoming part of another group-sibling):
6561 if (group_leader
->group_leader
!= group_leader
)
6564 * Do not allow to attach to a group in a different
6565 * task or CPU context:
6568 if (group_leader
->ctx
->type
!= ctx
->type
)
6571 if (group_leader
->ctx
!= ctx
)
6576 * Only a group leader can be exclusive or pinned
6578 if (attr
.exclusive
|| attr
.pinned
)
6583 err
= perf_event_set_output(event
, output_event
);
6588 event_file
= anon_inode_getfile("[perf_event]", &perf_fops
, event
, O_RDWR
);
6589 if (IS_ERR(event_file
)) {
6590 err
= PTR_ERR(event_file
);
6595 struct perf_event_context
*gctx
= group_leader
->ctx
;
6597 mutex_lock(&gctx
->mutex
);
6598 perf_remove_from_context(group_leader
);
6599 list_for_each_entry(sibling
, &group_leader
->sibling_list
,
6601 perf_remove_from_context(sibling
);
6604 mutex_unlock(&gctx
->mutex
);
6608 event
->filp
= event_file
;
6609 WARN_ON_ONCE(ctx
->parent_ctx
);
6610 mutex_lock(&ctx
->mutex
);
6613 perf_install_in_context(ctx
, group_leader
, cpu
);
6615 list_for_each_entry(sibling
, &group_leader
->sibling_list
,
6617 perf_install_in_context(ctx
, sibling
, cpu
);
6622 perf_install_in_context(ctx
, event
, cpu
);
6624 perf_unpin_context(ctx
);
6625 mutex_unlock(&ctx
->mutex
);
6627 event
->owner
= current
;
6629 mutex_lock(¤t
->perf_event_mutex
);
6630 list_add_tail(&event
->owner_entry
, ¤t
->perf_event_list
);
6631 mutex_unlock(¤t
->perf_event_mutex
);
6634 * Precalculate sample_data sizes
6636 perf_event__header_size(event
);
6637 perf_event__id_header_size(event
);
6640 * Drop the reference on the group_event after placing the
6641 * new event on the sibling_list. This ensures destruction
6642 * of the group leader will find the pointer to itself in
6643 * perf_group_detach().
6645 fput_light(group_file
, fput_needed
);
6646 fd_install(event_fd
, event_file
);
6650 perf_unpin_context(ctx
);
6656 put_task_struct(task
);
6658 fput_light(group_file
, fput_needed
);
6660 put_unused_fd(event_fd
);
6665 * perf_event_create_kernel_counter
6667 * @attr: attributes of the counter to create
6668 * @cpu: cpu in which the counter is bound
6669 * @task: task to profile (NULL for percpu)
6672 perf_event_create_kernel_counter(struct perf_event_attr
*attr
, int cpu
,
6673 struct task_struct
*task
,
6674 perf_overflow_handler_t overflow_handler
)
6676 struct perf_event_context
*ctx
;
6677 struct perf_event
*event
;
6681 * Get the target context (task or percpu):
6684 event
= perf_event_alloc(attr
, cpu
, task
, NULL
, NULL
, overflow_handler
);
6685 if (IS_ERR(event
)) {
6686 err
= PTR_ERR(event
);
6690 ctx
= find_get_context(event
->pmu
, task
, cpu
);
6697 WARN_ON_ONCE(ctx
->parent_ctx
);
6698 mutex_lock(&ctx
->mutex
);
6699 perf_install_in_context(ctx
, event
, cpu
);
6701 perf_unpin_context(ctx
);
6702 mutex_unlock(&ctx
->mutex
);
6709 return ERR_PTR(err
);
6711 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter
);
6713 static void sync_child_event(struct perf_event
*child_event
,
6714 struct task_struct
*child
)
6716 struct perf_event
*parent_event
= child_event
->parent
;
6719 if (child_event
->attr
.inherit_stat
)
6720 perf_event_read_event(child_event
, child
);
6722 child_val
= perf_event_count(child_event
);
6725 * Add back the child's count to the parent's count:
6727 atomic64_add(child_val
, &parent_event
->child_count
);
6728 atomic64_add(child_event
->total_time_enabled
,
6729 &parent_event
->child_total_time_enabled
);
6730 atomic64_add(child_event
->total_time_running
,
6731 &parent_event
->child_total_time_running
);
6734 * Remove this event from the parent's list
6736 WARN_ON_ONCE(parent_event
->ctx
->parent_ctx
);
6737 mutex_lock(&parent_event
->child_mutex
);
6738 list_del_init(&child_event
->child_list
);
6739 mutex_unlock(&parent_event
->child_mutex
);
6742 * Release the parent event, if this was the last
6745 fput(parent_event
->filp
);
6749 __perf_event_exit_task(struct perf_event
*child_event
,
6750 struct perf_event_context
*child_ctx
,
6751 struct task_struct
*child
)
6753 if (child_event
->parent
) {
6754 raw_spin_lock_irq(&child_ctx
->lock
);
6755 perf_group_detach(child_event
);
6756 raw_spin_unlock_irq(&child_ctx
->lock
);
6759 perf_remove_from_context(child_event
);
6762 * It can happen that the parent exits first, and has events
6763 * that are still around due to the child reference. These
6764 * events need to be zapped.
6766 if (child_event
->parent
) {
6767 sync_child_event(child_event
, child
);
6768 free_event(child_event
);
6772 static void perf_event_exit_task_context(struct task_struct
*child
, int ctxn
)
6774 struct perf_event
*child_event
, *tmp
;
6775 struct perf_event_context
*child_ctx
;
6776 unsigned long flags
;
6778 if (likely(!child
->perf_event_ctxp
[ctxn
])) {
6779 perf_event_task(child
, NULL
, 0);
6783 local_irq_save(flags
);
6785 * We can't reschedule here because interrupts are disabled,
6786 * and either child is current or it is a task that can't be
6787 * scheduled, so we are now safe from rescheduling changing
6790 child_ctx
= rcu_dereference_raw(child
->perf_event_ctxp
[ctxn
]);
6791 task_ctx_sched_out(child_ctx
, EVENT_ALL
);
6794 * Take the context lock here so that if find_get_context is
6795 * reading child->perf_event_ctxp, we wait until it has
6796 * incremented the context's refcount before we do put_ctx below.
6798 raw_spin_lock(&child_ctx
->lock
);
6799 child
->perf_event_ctxp
[ctxn
] = NULL
;
6801 * If this context is a clone; unclone it so it can't get
6802 * swapped to another process while we're removing all
6803 * the events from it.
6805 unclone_ctx(child_ctx
);
6806 update_context_time(child_ctx
);
6807 raw_spin_unlock_irqrestore(&child_ctx
->lock
, flags
);
6810 * Report the task dead after unscheduling the events so that we
6811 * won't get any samples after PERF_RECORD_EXIT. We can however still
6812 * get a few PERF_RECORD_READ events.
6814 perf_event_task(child
, child_ctx
, 0);
6817 * We can recurse on the same lock type through:
6819 * __perf_event_exit_task()
6820 * sync_child_event()
6821 * fput(parent_event->filp)
6823 * mutex_lock(&ctx->mutex)
6825 * But since its the parent context it won't be the same instance.
6827 mutex_lock(&child_ctx
->mutex
);
6830 list_for_each_entry_safe(child_event
, tmp
, &child_ctx
->pinned_groups
,
6832 __perf_event_exit_task(child_event
, child_ctx
, child
);
6834 list_for_each_entry_safe(child_event
, tmp
, &child_ctx
->flexible_groups
,
6836 __perf_event_exit_task(child_event
, child_ctx
, child
);
6839 * If the last event was a group event, it will have appended all
6840 * its siblings to the list, but we obtained 'tmp' before that which
6841 * will still point to the list head terminating the iteration.
6843 if (!list_empty(&child_ctx
->pinned_groups
) ||
6844 !list_empty(&child_ctx
->flexible_groups
))
6847 mutex_unlock(&child_ctx
->mutex
);
6853 * When a child task exits, feed back event values to parent events.
6855 void perf_event_exit_task(struct task_struct
*child
)
6857 struct perf_event
*event
, *tmp
;
6860 mutex_lock(&child
->perf_event_mutex
);
6861 list_for_each_entry_safe(event
, tmp
, &child
->perf_event_list
,
6863 list_del_init(&event
->owner_entry
);
6866 * Ensure the list deletion is visible before we clear
6867 * the owner, closes a race against perf_release() where
6868 * we need to serialize on the owner->perf_event_mutex.
6871 event
->owner
= NULL
;
6873 mutex_unlock(&child
->perf_event_mutex
);
6875 for_each_task_context_nr(ctxn
)
6876 perf_event_exit_task_context(child
, ctxn
);
6879 static void perf_free_event(struct perf_event
*event
,
6880 struct perf_event_context
*ctx
)
6882 struct perf_event
*parent
= event
->parent
;
6884 if (WARN_ON_ONCE(!parent
))
6887 mutex_lock(&parent
->child_mutex
);
6888 list_del_init(&event
->child_list
);
6889 mutex_unlock(&parent
->child_mutex
);
6893 perf_group_detach(event
);
6894 list_del_event(event
, ctx
);
6899 * free an unexposed, unused context as created by inheritance by
6900 * perf_event_init_task below, used by fork() in case of fail.
6902 void perf_event_free_task(struct task_struct
*task
)
6904 struct perf_event_context
*ctx
;
6905 struct perf_event
*event
, *tmp
;
6908 for_each_task_context_nr(ctxn
) {
6909 ctx
= task
->perf_event_ctxp
[ctxn
];
6913 mutex_lock(&ctx
->mutex
);
6915 list_for_each_entry_safe(event
, tmp
, &ctx
->pinned_groups
,
6917 perf_free_event(event
, ctx
);
6919 list_for_each_entry_safe(event
, tmp
, &ctx
->flexible_groups
,
6921 perf_free_event(event
, ctx
);
6923 if (!list_empty(&ctx
->pinned_groups
) ||
6924 !list_empty(&ctx
->flexible_groups
))
6927 mutex_unlock(&ctx
->mutex
);
6933 void perf_event_delayed_put(struct task_struct
*task
)
6937 for_each_task_context_nr(ctxn
)
6938 WARN_ON_ONCE(task
->perf_event_ctxp
[ctxn
]);
6942 * inherit a event from parent task to child task:
6944 static struct perf_event
*
6945 inherit_event(struct perf_event
*parent_event
,
6946 struct task_struct
*parent
,
6947 struct perf_event_context
*parent_ctx
,
6948 struct task_struct
*child
,
6949 struct perf_event
*group_leader
,
6950 struct perf_event_context
*child_ctx
)
6952 struct perf_event
*child_event
;
6953 unsigned long flags
;
6956 * Instead of creating recursive hierarchies of events,
6957 * we link inherited events back to the original parent,
6958 * which has a filp for sure, which we use as the reference
6961 if (parent_event
->parent
)
6962 parent_event
= parent_event
->parent
;
6964 child_event
= perf_event_alloc(&parent_event
->attr
,
6967 group_leader
, parent_event
,
6969 if (IS_ERR(child_event
))
6974 * Make the child state follow the state of the parent event,
6975 * not its attr.disabled bit. We hold the parent's mutex,
6976 * so we won't race with perf_event_{en, dis}able_family.
6978 if (parent_event
->state
>= PERF_EVENT_STATE_INACTIVE
)
6979 child_event
->state
= PERF_EVENT_STATE_INACTIVE
;
6981 child_event
->state
= PERF_EVENT_STATE_OFF
;
6983 if (parent_event
->attr
.freq
) {
6984 u64 sample_period
= parent_event
->hw
.sample_period
;
6985 struct hw_perf_event
*hwc
= &child_event
->hw
;
6987 hwc
->sample_period
= sample_period
;
6988 hwc
->last_period
= sample_period
;
6990 local64_set(&hwc
->period_left
, sample_period
);
6993 child_event
->ctx
= child_ctx
;
6994 child_event
->overflow_handler
= parent_event
->overflow_handler
;
6997 * Precalculate sample_data sizes
6999 perf_event__header_size(child_event
);
7000 perf_event__id_header_size(child_event
);
7003 * Link it up in the child's context:
7005 raw_spin_lock_irqsave(&child_ctx
->lock
, flags
);
7006 add_event_to_ctx(child_event
, child_ctx
);
7007 raw_spin_unlock_irqrestore(&child_ctx
->lock
, flags
);
7010 * Get a reference to the parent filp - we will fput it
7011 * when the child event exits. This is safe to do because
7012 * we are in the parent and we know that the filp still
7013 * exists and has a nonzero count:
7015 atomic_long_inc(&parent_event
->filp
->f_count
);
7018 * Link this into the parent event's child list
7020 WARN_ON_ONCE(parent_event
->ctx
->parent_ctx
);
7021 mutex_lock(&parent_event
->child_mutex
);
7022 list_add_tail(&child_event
->child_list
, &parent_event
->child_list
);
7023 mutex_unlock(&parent_event
->child_mutex
);
7028 static int inherit_group(struct perf_event
*parent_event
,
7029 struct task_struct
*parent
,
7030 struct perf_event_context
*parent_ctx
,
7031 struct task_struct
*child
,
7032 struct perf_event_context
*child_ctx
)
7034 struct perf_event
*leader
;
7035 struct perf_event
*sub
;
7036 struct perf_event
*child_ctr
;
7038 leader
= inherit_event(parent_event
, parent
, parent_ctx
,
7039 child
, NULL
, child_ctx
);
7041 return PTR_ERR(leader
);
7042 list_for_each_entry(sub
, &parent_event
->sibling_list
, group_entry
) {
7043 child_ctr
= inherit_event(sub
, parent
, parent_ctx
,
7044 child
, leader
, child_ctx
);
7045 if (IS_ERR(child_ctr
))
7046 return PTR_ERR(child_ctr
);
7052 inherit_task_group(struct perf_event
*event
, struct task_struct
*parent
,
7053 struct perf_event_context
*parent_ctx
,
7054 struct task_struct
*child
, int ctxn
,
7058 struct perf_event_context
*child_ctx
;
7060 if (!event
->attr
.inherit
) {
7065 child_ctx
= child
->perf_event_ctxp
[ctxn
];
7068 * This is executed from the parent task context, so
7069 * inherit events that have been marked for cloning.
7070 * First allocate and initialize a context for the
7074 child_ctx
= alloc_perf_context(event
->pmu
, child
);
7078 child
->perf_event_ctxp
[ctxn
] = child_ctx
;
7081 ret
= inherit_group(event
, parent
, parent_ctx
,
7091 * Initialize the perf_event context in task_struct
7093 int perf_event_init_context(struct task_struct
*child
, int ctxn
)
7095 struct perf_event_context
*child_ctx
, *parent_ctx
;
7096 struct perf_event_context
*cloned_ctx
;
7097 struct perf_event
*event
;
7098 struct task_struct
*parent
= current
;
7099 int inherited_all
= 1;
7100 unsigned long flags
;
7103 if (likely(!parent
->perf_event_ctxp
[ctxn
]))
7107 * If the parent's context is a clone, pin it so it won't get
7110 parent_ctx
= perf_pin_task_context(parent
, ctxn
);
7113 * No need to check if parent_ctx != NULL here; since we saw
7114 * it non-NULL earlier, the only reason for it to become NULL
7115 * is if we exit, and since we're currently in the middle of
7116 * a fork we can't be exiting at the same time.
7120 * Lock the parent list. No need to lock the child - not PID
7121 * hashed yet and not running, so nobody can access it.
7123 mutex_lock(&parent_ctx
->mutex
);
7126 * We dont have to disable NMIs - we are only looking at
7127 * the list, not manipulating it:
7129 list_for_each_entry(event
, &parent_ctx
->pinned_groups
, group_entry
) {
7130 ret
= inherit_task_group(event
, parent
, parent_ctx
,
7131 child
, ctxn
, &inherited_all
);
7137 * We can't hold ctx->lock when iterating the ->flexible_group list due
7138 * to allocations, but we need to prevent rotation because
7139 * rotate_ctx() will change the list from interrupt context.
7141 raw_spin_lock_irqsave(&parent_ctx
->lock
, flags
);
7142 parent_ctx
->rotate_disable
= 1;
7143 raw_spin_unlock_irqrestore(&parent_ctx
->lock
, flags
);
7145 list_for_each_entry(event
, &parent_ctx
->flexible_groups
, group_entry
) {
7146 ret
= inherit_task_group(event
, parent
, parent_ctx
,
7147 child
, ctxn
, &inherited_all
);
7152 raw_spin_lock_irqsave(&parent_ctx
->lock
, flags
);
7153 parent_ctx
->rotate_disable
= 0;
7155 child_ctx
= child
->perf_event_ctxp
[ctxn
];
7157 if (child_ctx
&& inherited_all
) {
7159 * Mark the child context as a clone of the parent
7160 * context, or of whatever the parent is a clone of.
7162 * Note that if the parent is a clone, the holding of
7163 * parent_ctx->lock avoids it from being uncloned.
7165 cloned_ctx
= parent_ctx
->parent_ctx
;
7167 child_ctx
->parent_ctx
= cloned_ctx
;
7168 child_ctx
->parent_gen
= parent_ctx
->parent_gen
;
7170 child_ctx
->parent_ctx
= parent_ctx
;
7171 child_ctx
->parent_gen
= parent_ctx
->generation
;
7173 get_ctx(child_ctx
->parent_ctx
);
7176 raw_spin_unlock_irqrestore(&parent_ctx
->lock
, flags
);
7177 mutex_unlock(&parent_ctx
->mutex
);
7179 perf_unpin_context(parent_ctx
);
7180 put_ctx(parent_ctx
);
7186 * Initialize the perf_event context in task_struct
7188 int perf_event_init_task(struct task_struct
*child
)
7192 memset(child
->perf_event_ctxp
, 0, sizeof(child
->perf_event_ctxp
));
7193 mutex_init(&child
->perf_event_mutex
);
7194 INIT_LIST_HEAD(&child
->perf_event_list
);
7196 for_each_task_context_nr(ctxn
) {
7197 ret
= perf_event_init_context(child
, ctxn
);
7205 static void __init
perf_event_init_all_cpus(void)
7207 struct swevent_htable
*swhash
;
7210 for_each_possible_cpu(cpu
) {
7211 swhash
= &per_cpu(swevent_htable
, cpu
);
7212 mutex_init(&swhash
->hlist_mutex
);
7213 INIT_LIST_HEAD(&per_cpu(rotation_list
, cpu
));
7217 static void __cpuinit
perf_event_init_cpu(int cpu
)
7219 struct swevent_htable
*swhash
= &per_cpu(swevent_htable
, cpu
);
7221 mutex_lock(&swhash
->hlist_mutex
);
7222 if (swhash
->hlist_refcount
> 0) {
7223 struct swevent_hlist
*hlist
;
7225 hlist
= kzalloc_node(sizeof(*hlist
), GFP_KERNEL
, cpu_to_node(cpu
));
7227 rcu_assign_pointer(swhash
->swevent_hlist
, hlist
);
7229 mutex_unlock(&swhash
->hlist_mutex
);
7232 #if defined CONFIG_HOTPLUG_CPU || defined CONFIG_KEXEC
7233 static void perf_pmu_rotate_stop(struct pmu
*pmu
)
7235 struct perf_cpu_context
*cpuctx
= this_cpu_ptr(pmu
->pmu_cpu_context
);
7237 WARN_ON(!irqs_disabled());
7239 list_del_init(&cpuctx
->rotation_list
);
7242 static void __perf_event_exit_context(void *__info
)
7244 struct perf_event_context
*ctx
= __info
;
7245 struct perf_event
*event
, *tmp
;
7247 perf_pmu_rotate_stop(ctx
->pmu
);
7249 list_for_each_entry_safe(event
, tmp
, &ctx
->pinned_groups
, group_entry
)
7250 __perf_remove_from_context(event
);
7251 list_for_each_entry_safe(event
, tmp
, &ctx
->flexible_groups
, group_entry
)
7252 __perf_remove_from_context(event
);
7255 static void perf_event_exit_cpu_context(int cpu
)
7257 struct perf_event_context
*ctx
;
7261 idx
= srcu_read_lock(&pmus_srcu
);
7262 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
7263 ctx
= &per_cpu_ptr(pmu
->pmu_cpu_context
, cpu
)->ctx
;
7265 mutex_lock(&ctx
->mutex
);
7266 smp_call_function_single(cpu
, __perf_event_exit_context
, ctx
, 1);
7267 mutex_unlock(&ctx
->mutex
);
7269 srcu_read_unlock(&pmus_srcu
, idx
);
7272 static void perf_event_exit_cpu(int cpu
)
7274 struct swevent_htable
*swhash
= &per_cpu(swevent_htable
, cpu
);
7276 mutex_lock(&swhash
->hlist_mutex
);
7277 swevent_hlist_release(swhash
);
7278 mutex_unlock(&swhash
->hlist_mutex
);
7280 perf_event_exit_cpu_context(cpu
);
7283 static inline void perf_event_exit_cpu(int cpu
) { }
7287 perf_reboot(struct notifier_block
*notifier
, unsigned long val
, void *v
)
7291 for_each_online_cpu(cpu
)
7292 perf_event_exit_cpu(cpu
);
7298 * Run the perf reboot notifier at the very last possible moment so that
7299 * the generic watchdog code runs as long as possible.
7301 static struct notifier_block perf_reboot_notifier
= {
7302 .notifier_call
= perf_reboot
,
7303 .priority
= INT_MIN
,
7306 static int __cpuinit
7307 perf_cpu_notify(struct notifier_block
*self
, unsigned long action
, void *hcpu
)
7309 unsigned int cpu
= (long)hcpu
;
7311 switch (action
& ~CPU_TASKS_FROZEN
) {
7313 case CPU_UP_PREPARE
:
7314 case CPU_DOWN_FAILED
:
7315 perf_event_init_cpu(cpu
);
7318 case CPU_UP_CANCELED
:
7319 case CPU_DOWN_PREPARE
:
7320 perf_event_exit_cpu(cpu
);
7330 void __init
perf_event_init(void)
7336 perf_event_init_all_cpus();
7337 init_srcu_struct(&pmus_srcu
);
7338 perf_pmu_register(&perf_swevent
, "software", PERF_TYPE_SOFTWARE
);
7339 perf_pmu_register(&perf_cpu_clock
, NULL
, -1);
7340 perf_pmu_register(&perf_task_clock
, NULL
, -1);
7342 perf_cpu_notifier(perf_cpu_notify
);
7343 register_reboot_notifier(&perf_reboot_notifier
);
7345 ret
= init_hw_breakpoint();
7346 WARN(ret
, "hw_breakpoint initialization failed with: %d", ret
);
7349 static int __init
perf_event_sysfs_init(void)
7354 mutex_lock(&pmus_lock
);
7356 ret
= bus_register(&pmu_bus
);
7360 list_for_each_entry(pmu
, &pmus
, entry
) {
7361 if (!pmu
->name
|| pmu
->type
< 0)
7364 ret
= pmu_dev_alloc(pmu
);
7365 WARN(ret
, "Failed to register pmu: %s, reason %d\n", pmu
->name
, ret
);
7367 pmu_bus_running
= 1;
7371 mutex_unlock(&pmus_lock
);
7375 device_initcall(perf_event_sysfs_init
);
7377 #ifdef CONFIG_CGROUP_PERF
7378 static struct cgroup_subsys_state
*perf_cgroup_create(
7379 struct cgroup_subsys
*ss
, struct cgroup
*cont
)
7381 struct perf_cgroup
*jc
;
7383 jc
= kzalloc(sizeof(*jc
), GFP_KERNEL
);
7385 return ERR_PTR(-ENOMEM
);
7387 jc
->info
= alloc_percpu(struct perf_cgroup_info
);
7390 return ERR_PTR(-ENOMEM
);
7396 static void perf_cgroup_destroy(struct cgroup_subsys
*ss
,
7397 struct cgroup
*cont
)
7399 struct perf_cgroup
*jc
;
7400 jc
= container_of(cgroup_subsys_state(cont
, perf_subsys_id
),
7401 struct perf_cgroup
, css
);
7402 free_percpu(jc
->info
);
7406 static int __perf_cgroup_move(void *info
)
7408 struct task_struct
*task
= info
;
7409 perf_cgroup_switch(task
, PERF_CGROUP_SWOUT
| PERF_CGROUP_SWIN
);
7413 static void perf_cgroup_move(struct task_struct
*task
)
7415 task_function_call(task
, __perf_cgroup_move
, task
);
7418 static void perf_cgroup_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
7419 struct cgroup
*old_cgrp
, struct task_struct
*task
,
7422 perf_cgroup_move(task
);
7424 struct task_struct
*c
;
7426 list_for_each_entry_rcu(c
, &task
->thread_group
, thread_group
) {
7427 perf_cgroup_move(c
);
7433 static void perf_cgroup_exit(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
7434 struct cgroup
*old_cgrp
, struct task_struct
*task
)
7437 * cgroup_exit() is called in the copy_process() failure path.
7438 * Ignore this case since the task hasn't ran yet, this avoids
7439 * trying to poke a half freed task state from generic code.
7441 if (!(task
->flags
& PF_EXITING
))
7444 perf_cgroup_move(task
);
7447 struct cgroup_subsys perf_subsys
= {
7448 .name
= "perf_event",
7449 .subsys_id
= perf_subsys_id
,
7450 .create
= perf_cgroup_create
,
7451 .destroy
= perf_cgroup_destroy
,
7452 .exit
= perf_cgroup_exit
,
7453 .attach
= perf_cgroup_attach
,
7455 #endif /* CONFIG_CGROUP_PERF */