2 * Performance events core code:
4 * Copyright (C) 2008 Thomas Gleixner <tglx@linutronix.de>
5 * Copyright (C) 2008-2011 Red Hat, Inc., Ingo Molnar
6 * Copyright (C) 2008-2011 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
7 * Copyright © 2009 Paul Mackerras, IBM Corp. <paulus@au1.ibm.com>
9 * For licensing details see kernel-base/COPYING
14 #include <linux/cpu.h>
15 #include <linux/smp.h>
16 #include <linux/idr.h>
17 #include <linux/file.h>
18 #include <linux/poll.h>
19 #include <linux/slab.h>
20 #include <linux/hash.h>
21 #include <linux/sysfs.h>
22 #include <linux/dcache.h>
23 #include <linux/percpu.h>
24 #include <linux/ptrace.h>
25 #include <linux/reboot.h>
26 #include <linux/vmstat.h>
27 #include <linux/device.h>
28 #include <linux/export.h>
29 #include <linux/vmalloc.h>
30 #include <linux/hardirq.h>
31 #include <linux/rculist.h>
32 #include <linux/uaccess.h>
33 #include <linux/syscalls.h>
34 #include <linux/anon_inodes.h>
35 #include <linux/kernel_stat.h>
36 #include <linux/perf_event.h>
37 #include <linux/ftrace_event.h>
38 #include <linux/hw_breakpoint.h>
42 #include <asm/irq_regs.h>
44 struct remote_function_call
{
45 struct task_struct
*p
;
46 int (*func
)(void *info
);
51 static void remote_function(void *data
)
53 struct remote_function_call
*tfc
= data
;
54 struct task_struct
*p
= tfc
->p
;
58 if (task_cpu(p
) != smp_processor_id() || !task_curr(p
))
62 tfc
->ret
= tfc
->func(tfc
->info
);
66 * task_function_call - call a function on the cpu on which a task runs
67 * @p: the task to evaluate
68 * @func: the function to be called
69 * @info: the function call argument
71 * Calls the function @func when the task is currently running. This might
72 * be on the current CPU, which just calls the function directly
74 * returns: @func return value, or
75 * -ESRCH - when the process isn't running
76 * -EAGAIN - when the process moved away
79 task_function_call(struct task_struct
*p
, int (*func
) (void *info
), void *info
)
81 struct remote_function_call data
= {
85 .ret
= -ESRCH
, /* No such (running) process */
89 smp_call_function_single(task_cpu(p
), remote_function
, &data
, 1);
95 * cpu_function_call - call a function on the cpu
96 * @func: the function to be called
97 * @info: the function call argument
99 * Calls the function @func on the remote cpu.
101 * returns: @func return value or -ENXIO when the cpu is offline
103 static int cpu_function_call(int cpu
, int (*func
) (void *info
), void *info
)
105 struct remote_function_call data
= {
109 .ret
= -ENXIO
, /* No such CPU */
112 smp_call_function_single(cpu
, remote_function
, &data
, 1);
117 #define PERF_FLAG_ALL (PERF_FLAG_FD_NO_GROUP |\
118 PERF_FLAG_FD_OUTPUT |\
119 PERF_FLAG_PID_CGROUP)
122 EVENT_FLEXIBLE
= 0x1,
124 EVENT_ALL
= EVENT_FLEXIBLE
| EVENT_PINNED
,
128 * perf_sched_events : >0 events exist
129 * perf_cgroup_events: >0 per-cpu cgroup events exist on this cpu
131 struct jump_label_key perf_sched_events __read_mostly
;
132 static DEFINE_PER_CPU(atomic_t
, perf_cgroup_events
);
134 static atomic_t nr_mmap_events __read_mostly
;
135 static atomic_t nr_comm_events __read_mostly
;
136 static atomic_t nr_task_events __read_mostly
;
138 static LIST_HEAD(pmus
);
139 static DEFINE_MUTEX(pmus_lock
);
140 static struct srcu_struct pmus_srcu
;
143 * perf event paranoia level:
144 * -1 - not paranoid at all
145 * 0 - disallow raw tracepoint access for unpriv
146 * 1 - disallow cpu events for unpriv
147 * 2 - disallow kernel profiling for unpriv
149 int sysctl_perf_event_paranoid __read_mostly
= 1;
151 /* Minimum for 512 kiB + 1 user control page */
152 int sysctl_perf_event_mlock __read_mostly
= 512 + (PAGE_SIZE
/ 1024); /* 'free' kiB per user */
155 * max perf event sample rate
157 #define DEFAULT_MAX_SAMPLE_RATE 100000
158 int sysctl_perf_event_sample_rate __read_mostly
= DEFAULT_MAX_SAMPLE_RATE
;
159 static int max_samples_per_tick __read_mostly
=
160 DIV_ROUND_UP(DEFAULT_MAX_SAMPLE_RATE
, HZ
);
162 int perf_proc_update_handler(struct ctl_table
*table
, int write
,
163 void __user
*buffer
, size_t *lenp
,
166 int ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
171 max_samples_per_tick
= DIV_ROUND_UP(sysctl_perf_event_sample_rate
, HZ
);
176 static atomic64_t perf_event_id
;
178 static void cpu_ctx_sched_out(struct perf_cpu_context
*cpuctx
,
179 enum event_type_t event_type
);
181 static void cpu_ctx_sched_in(struct perf_cpu_context
*cpuctx
,
182 enum event_type_t event_type
,
183 struct task_struct
*task
);
185 static void update_context_time(struct perf_event_context
*ctx
);
186 static u64
perf_event_time(struct perf_event
*event
);
188 void __weak
perf_event_print_debug(void) { }
190 extern __weak
const char *perf_pmu_name(void)
195 static inline u64
perf_clock(void)
197 return local_clock();
200 static inline struct perf_cpu_context
*
201 __get_cpu_context(struct perf_event_context
*ctx
)
203 return this_cpu_ptr(ctx
->pmu
->pmu_cpu_context
);
206 static void perf_ctx_lock(struct perf_cpu_context
*cpuctx
,
207 struct perf_event_context
*ctx
)
209 raw_spin_lock(&cpuctx
->ctx
.lock
);
211 raw_spin_lock(&ctx
->lock
);
214 static void perf_ctx_unlock(struct perf_cpu_context
*cpuctx
,
215 struct perf_event_context
*ctx
)
218 raw_spin_unlock(&ctx
->lock
);
219 raw_spin_unlock(&cpuctx
->ctx
.lock
);
222 #ifdef CONFIG_CGROUP_PERF
225 * Must ensure cgroup is pinned (css_get) before calling
226 * this function. In other words, we cannot call this function
227 * if there is no cgroup event for the current CPU context.
229 static inline struct perf_cgroup
*
230 perf_cgroup_from_task(struct task_struct
*task
)
232 return container_of(task_subsys_state(task
, perf_subsys_id
),
233 struct perf_cgroup
, css
);
237 perf_cgroup_match(struct perf_event
*event
)
239 struct perf_event_context
*ctx
= event
->ctx
;
240 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
242 return !event
->cgrp
|| event
->cgrp
== cpuctx
->cgrp
;
245 static inline void perf_get_cgroup(struct perf_event
*event
)
247 css_get(&event
->cgrp
->css
);
250 static inline void perf_put_cgroup(struct perf_event
*event
)
252 css_put(&event
->cgrp
->css
);
255 static inline void perf_detach_cgroup(struct perf_event
*event
)
257 perf_put_cgroup(event
);
261 static inline int is_cgroup_event(struct perf_event
*event
)
263 return event
->cgrp
!= NULL
;
266 static inline u64
perf_cgroup_event_time(struct perf_event
*event
)
268 struct perf_cgroup_info
*t
;
270 t
= per_cpu_ptr(event
->cgrp
->info
, event
->cpu
);
274 static inline void __update_cgrp_time(struct perf_cgroup
*cgrp
)
276 struct perf_cgroup_info
*info
;
281 info
= this_cpu_ptr(cgrp
->info
);
283 info
->time
+= now
- info
->timestamp
;
284 info
->timestamp
= now
;
287 static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context
*cpuctx
)
289 struct perf_cgroup
*cgrp_out
= cpuctx
->cgrp
;
291 __update_cgrp_time(cgrp_out
);
294 static inline void update_cgrp_time_from_event(struct perf_event
*event
)
296 struct perf_cgroup
*cgrp
;
299 * ensure we access cgroup data only when needed and
300 * when we know the cgroup is pinned (css_get)
302 if (!is_cgroup_event(event
))
305 cgrp
= perf_cgroup_from_task(current
);
307 * Do not update time when cgroup is not active
309 if (cgrp
== event
->cgrp
)
310 __update_cgrp_time(event
->cgrp
);
314 perf_cgroup_set_timestamp(struct task_struct
*task
,
315 struct perf_event_context
*ctx
)
317 struct perf_cgroup
*cgrp
;
318 struct perf_cgroup_info
*info
;
321 * ctx->lock held by caller
322 * ensure we do not access cgroup data
323 * unless we have the cgroup pinned (css_get)
325 if (!task
|| !ctx
->nr_cgroups
)
328 cgrp
= perf_cgroup_from_task(task
);
329 info
= this_cpu_ptr(cgrp
->info
);
330 info
->timestamp
= ctx
->timestamp
;
333 #define PERF_CGROUP_SWOUT 0x1 /* cgroup switch out every event */
334 #define PERF_CGROUP_SWIN 0x2 /* cgroup switch in events based on task */
337 * reschedule events based on the cgroup constraint of task.
339 * mode SWOUT : schedule out everything
340 * mode SWIN : schedule in based on cgroup for next
342 void perf_cgroup_switch(struct task_struct
*task
, int mode
)
344 struct perf_cpu_context
*cpuctx
;
349 * disable interrupts to avoid geting nr_cgroup
350 * changes via __perf_event_disable(). Also
353 local_irq_save(flags
);
356 * we reschedule only in the presence of cgroup
357 * constrained events.
361 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
362 cpuctx
= this_cpu_ptr(pmu
->pmu_cpu_context
);
365 * perf_cgroup_events says at least one
366 * context on this CPU has cgroup events.
368 * ctx->nr_cgroups reports the number of cgroup
369 * events for a context.
371 if (cpuctx
->ctx
.nr_cgroups
> 0) {
372 perf_ctx_lock(cpuctx
, cpuctx
->task_ctx
);
373 perf_pmu_disable(cpuctx
->ctx
.pmu
);
375 if (mode
& PERF_CGROUP_SWOUT
) {
376 cpu_ctx_sched_out(cpuctx
, EVENT_ALL
);
378 * must not be done before ctxswout due
379 * to event_filter_match() in event_sched_out()
384 if (mode
& PERF_CGROUP_SWIN
) {
385 WARN_ON_ONCE(cpuctx
->cgrp
);
386 /* set cgrp before ctxsw in to
387 * allow event_filter_match() to not
388 * have to pass task around
390 cpuctx
->cgrp
= perf_cgroup_from_task(task
);
391 cpu_ctx_sched_in(cpuctx
, EVENT_ALL
, task
);
393 perf_pmu_enable(cpuctx
->ctx
.pmu
);
394 perf_ctx_unlock(cpuctx
, cpuctx
->task_ctx
);
400 local_irq_restore(flags
);
403 static inline void perf_cgroup_sched_out(struct task_struct
*task
,
404 struct task_struct
*next
)
406 struct perf_cgroup
*cgrp1
;
407 struct perf_cgroup
*cgrp2
= NULL
;
410 * we come here when we know perf_cgroup_events > 0
412 cgrp1
= perf_cgroup_from_task(task
);
415 * next is NULL when called from perf_event_enable_on_exec()
416 * that will systematically cause a cgroup_switch()
419 cgrp2
= perf_cgroup_from_task(next
);
422 * only schedule out current cgroup events if we know
423 * that we are switching to a different cgroup. Otherwise,
424 * do no touch the cgroup events.
427 perf_cgroup_switch(task
, PERF_CGROUP_SWOUT
);
430 static inline void perf_cgroup_sched_in(struct task_struct
*prev
,
431 struct task_struct
*task
)
433 struct perf_cgroup
*cgrp1
;
434 struct perf_cgroup
*cgrp2
= NULL
;
437 * we come here when we know perf_cgroup_events > 0
439 cgrp1
= perf_cgroup_from_task(task
);
441 /* prev can never be NULL */
442 cgrp2
= perf_cgroup_from_task(prev
);
445 * only need to schedule in cgroup events if we are changing
446 * cgroup during ctxsw. Cgroup events were not scheduled
447 * out of ctxsw out if that was not the case.
450 perf_cgroup_switch(task
, PERF_CGROUP_SWIN
);
453 static inline int perf_cgroup_connect(int fd
, struct perf_event
*event
,
454 struct perf_event_attr
*attr
,
455 struct perf_event
*group_leader
)
457 struct perf_cgroup
*cgrp
;
458 struct cgroup_subsys_state
*css
;
460 int ret
= 0, fput_needed
;
462 file
= fget_light(fd
, &fput_needed
);
466 css
= cgroup_css_from_dir(file
, perf_subsys_id
);
472 cgrp
= container_of(css
, struct perf_cgroup
, css
);
475 /* must be done before we fput() the file */
476 perf_get_cgroup(event
);
479 * all events in a group must monitor
480 * the same cgroup because a task belongs
481 * to only one perf cgroup at a time
483 if (group_leader
&& group_leader
->cgrp
!= cgrp
) {
484 perf_detach_cgroup(event
);
488 fput_light(file
, fput_needed
);
493 perf_cgroup_set_shadow_time(struct perf_event
*event
, u64 now
)
495 struct perf_cgroup_info
*t
;
496 t
= per_cpu_ptr(event
->cgrp
->info
, event
->cpu
);
497 event
->shadow_ctx_time
= now
- t
->timestamp
;
501 perf_cgroup_defer_enabled(struct perf_event
*event
)
504 * when the current task's perf cgroup does not match
505 * the event's, we need to remember to call the
506 * perf_mark_enable() function the first time a task with
507 * a matching perf cgroup is scheduled in.
509 if (is_cgroup_event(event
) && !perf_cgroup_match(event
))
510 event
->cgrp_defer_enabled
= 1;
514 perf_cgroup_mark_enabled(struct perf_event
*event
,
515 struct perf_event_context
*ctx
)
517 struct perf_event
*sub
;
518 u64 tstamp
= perf_event_time(event
);
520 if (!event
->cgrp_defer_enabled
)
523 event
->cgrp_defer_enabled
= 0;
525 event
->tstamp_enabled
= tstamp
- event
->total_time_enabled
;
526 list_for_each_entry(sub
, &event
->sibling_list
, group_entry
) {
527 if (sub
->state
>= PERF_EVENT_STATE_INACTIVE
) {
528 sub
->tstamp_enabled
= tstamp
- sub
->total_time_enabled
;
529 sub
->cgrp_defer_enabled
= 0;
533 #else /* !CONFIG_CGROUP_PERF */
536 perf_cgroup_match(struct perf_event
*event
)
541 static inline void perf_detach_cgroup(struct perf_event
*event
)
544 static inline int is_cgroup_event(struct perf_event
*event
)
549 static inline u64
perf_cgroup_event_cgrp_time(struct perf_event
*event
)
554 static inline void update_cgrp_time_from_event(struct perf_event
*event
)
558 static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context
*cpuctx
)
562 static inline void perf_cgroup_sched_out(struct task_struct
*task
,
563 struct task_struct
*next
)
567 static inline void perf_cgroup_sched_in(struct task_struct
*prev
,
568 struct task_struct
*task
)
572 static inline int perf_cgroup_connect(pid_t pid
, struct perf_event
*event
,
573 struct perf_event_attr
*attr
,
574 struct perf_event
*group_leader
)
580 perf_cgroup_set_timestamp(struct task_struct
*task
,
581 struct perf_event_context
*ctx
)
586 perf_cgroup_switch(struct task_struct
*task
, struct task_struct
*next
)
591 perf_cgroup_set_shadow_time(struct perf_event
*event
, u64 now
)
595 static inline u64
perf_cgroup_event_time(struct perf_event
*event
)
601 perf_cgroup_defer_enabled(struct perf_event
*event
)
606 perf_cgroup_mark_enabled(struct perf_event
*event
,
607 struct perf_event_context
*ctx
)
612 void perf_pmu_disable(struct pmu
*pmu
)
614 int *count
= this_cpu_ptr(pmu
->pmu_disable_count
);
616 pmu
->pmu_disable(pmu
);
619 void perf_pmu_enable(struct pmu
*pmu
)
621 int *count
= this_cpu_ptr(pmu
->pmu_disable_count
);
623 pmu
->pmu_enable(pmu
);
626 static DEFINE_PER_CPU(struct list_head
, rotation_list
);
629 * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
630 * because they're strictly cpu affine and rotate_start is called with IRQs
631 * disabled, while rotate_context is called from IRQ context.
633 static void perf_pmu_rotate_start(struct pmu
*pmu
)
635 struct perf_cpu_context
*cpuctx
= this_cpu_ptr(pmu
->pmu_cpu_context
);
636 struct list_head
*head
= &__get_cpu_var(rotation_list
);
638 WARN_ON(!irqs_disabled());
640 if (list_empty(&cpuctx
->rotation_list
))
641 list_add(&cpuctx
->rotation_list
, head
);
644 static void get_ctx(struct perf_event_context
*ctx
)
646 WARN_ON(!atomic_inc_not_zero(&ctx
->refcount
));
649 static void put_ctx(struct perf_event_context
*ctx
)
651 if (atomic_dec_and_test(&ctx
->refcount
)) {
653 put_ctx(ctx
->parent_ctx
);
655 put_task_struct(ctx
->task
);
656 kfree_rcu(ctx
, rcu_head
);
660 static void unclone_ctx(struct perf_event_context
*ctx
)
662 if (ctx
->parent_ctx
) {
663 put_ctx(ctx
->parent_ctx
);
664 ctx
->parent_ctx
= NULL
;
668 static u32
perf_event_pid(struct perf_event
*event
, struct task_struct
*p
)
671 * only top level events have the pid namespace they were created in
674 event
= event
->parent
;
676 return task_tgid_nr_ns(p
, event
->ns
);
679 static u32
perf_event_tid(struct perf_event
*event
, struct task_struct
*p
)
682 * only top level events have the pid namespace they were created in
685 event
= event
->parent
;
687 return task_pid_nr_ns(p
, event
->ns
);
691 * If we inherit events we want to return the parent event id
694 static u64
primary_event_id(struct perf_event
*event
)
699 id
= event
->parent
->id
;
705 * Get the perf_event_context for a task and lock it.
706 * This has to cope with with the fact that until it is locked,
707 * the context could get moved to another task.
709 static struct perf_event_context
*
710 perf_lock_task_context(struct task_struct
*task
, int ctxn
, unsigned long *flags
)
712 struct perf_event_context
*ctx
;
716 ctx
= rcu_dereference(task
->perf_event_ctxp
[ctxn
]);
719 * If this context is a clone of another, it might
720 * get swapped for another underneath us by
721 * perf_event_task_sched_out, though the
722 * rcu_read_lock() protects us from any context
723 * getting freed. Lock the context and check if it
724 * got swapped before we could get the lock, and retry
725 * if so. If we locked the right context, then it
726 * can't get swapped on us any more.
728 raw_spin_lock_irqsave(&ctx
->lock
, *flags
);
729 if (ctx
!= rcu_dereference(task
->perf_event_ctxp
[ctxn
])) {
730 raw_spin_unlock_irqrestore(&ctx
->lock
, *flags
);
734 if (!atomic_inc_not_zero(&ctx
->refcount
)) {
735 raw_spin_unlock_irqrestore(&ctx
->lock
, *flags
);
744 * Get the context for a task and increment its pin_count so it
745 * can't get swapped to another task. This also increments its
746 * reference count so that the context can't get freed.
748 static struct perf_event_context
*
749 perf_pin_task_context(struct task_struct
*task
, int ctxn
)
751 struct perf_event_context
*ctx
;
754 ctx
= perf_lock_task_context(task
, ctxn
, &flags
);
757 raw_spin_unlock_irqrestore(&ctx
->lock
, flags
);
762 static void perf_unpin_context(struct perf_event_context
*ctx
)
766 raw_spin_lock_irqsave(&ctx
->lock
, flags
);
768 raw_spin_unlock_irqrestore(&ctx
->lock
, flags
);
772 * Update the record of the current time in a context.
774 static void update_context_time(struct perf_event_context
*ctx
)
776 u64 now
= perf_clock();
778 ctx
->time
+= now
- ctx
->timestamp
;
779 ctx
->timestamp
= now
;
782 static u64
perf_event_time(struct perf_event
*event
)
784 struct perf_event_context
*ctx
= event
->ctx
;
786 if (is_cgroup_event(event
))
787 return perf_cgroup_event_time(event
);
789 return ctx
? ctx
->time
: 0;
793 * Update the total_time_enabled and total_time_running fields for a event.
794 * The caller of this function needs to hold the ctx->lock.
796 static void update_event_times(struct perf_event
*event
)
798 struct perf_event_context
*ctx
= event
->ctx
;
801 if (event
->state
< PERF_EVENT_STATE_INACTIVE
||
802 event
->group_leader
->state
< PERF_EVENT_STATE_INACTIVE
)
805 * in cgroup mode, time_enabled represents
806 * the time the event was enabled AND active
807 * tasks were in the monitored cgroup. This is
808 * independent of the activity of the context as
809 * there may be a mix of cgroup and non-cgroup events.
811 * That is why we treat cgroup events differently
814 if (is_cgroup_event(event
))
815 run_end
= perf_event_time(event
);
816 else if (ctx
->is_active
)
819 run_end
= event
->tstamp_stopped
;
821 event
->total_time_enabled
= run_end
- event
->tstamp_enabled
;
823 if (event
->state
== PERF_EVENT_STATE_INACTIVE
)
824 run_end
= event
->tstamp_stopped
;
826 run_end
= perf_event_time(event
);
828 event
->total_time_running
= run_end
- event
->tstamp_running
;
833 * Update total_time_enabled and total_time_running for all events in a group.
835 static void update_group_times(struct perf_event
*leader
)
837 struct perf_event
*event
;
839 update_event_times(leader
);
840 list_for_each_entry(event
, &leader
->sibling_list
, group_entry
)
841 update_event_times(event
);
844 static struct list_head
*
845 ctx_group_list(struct perf_event
*event
, struct perf_event_context
*ctx
)
847 if (event
->attr
.pinned
)
848 return &ctx
->pinned_groups
;
850 return &ctx
->flexible_groups
;
854 * Add a event from the lists for its context.
855 * Must be called with ctx->mutex and ctx->lock held.
858 list_add_event(struct perf_event
*event
, struct perf_event_context
*ctx
)
860 WARN_ON_ONCE(event
->attach_state
& PERF_ATTACH_CONTEXT
);
861 event
->attach_state
|= PERF_ATTACH_CONTEXT
;
864 * If we're a stand alone event or group leader, we go to the context
865 * list, group events are kept attached to the group so that
866 * perf_group_detach can, at all times, locate all siblings.
868 if (event
->group_leader
== event
) {
869 struct list_head
*list
;
871 if (is_software_event(event
))
872 event
->group_flags
|= PERF_GROUP_SOFTWARE
;
874 list
= ctx_group_list(event
, ctx
);
875 list_add_tail(&event
->group_entry
, list
);
878 if (is_cgroup_event(event
))
881 list_add_rcu(&event
->event_entry
, &ctx
->event_list
);
883 perf_pmu_rotate_start(ctx
->pmu
);
885 if (event
->attr
.inherit_stat
)
890 * Called at perf_event creation and when events are attached/detached from a
893 static void perf_event__read_size(struct perf_event
*event
)
895 int entry
= sizeof(u64
); /* value */
899 if (event
->attr
.read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
902 if (event
->attr
.read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
905 if (event
->attr
.read_format
& PERF_FORMAT_ID
)
906 entry
+= sizeof(u64
);
908 if (event
->attr
.read_format
& PERF_FORMAT_GROUP
) {
909 nr
+= event
->group_leader
->nr_siblings
;
914 event
->read_size
= size
;
917 static void perf_event__header_size(struct perf_event
*event
)
919 struct perf_sample_data
*data
;
920 u64 sample_type
= event
->attr
.sample_type
;
923 perf_event__read_size(event
);
925 if (sample_type
& PERF_SAMPLE_IP
)
926 size
+= sizeof(data
->ip
);
928 if (sample_type
& PERF_SAMPLE_ADDR
)
929 size
+= sizeof(data
->addr
);
931 if (sample_type
& PERF_SAMPLE_PERIOD
)
932 size
+= sizeof(data
->period
);
934 if (sample_type
& PERF_SAMPLE_READ
)
935 size
+= event
->read_size
;
937 event
->header_size
= size
;
940 static void perf_event__id_header_size(struct perf_event
*event
)
942 struct perf_sample_data
*data
;
943 u64 sample_type
= event
->attr
.sample_type
;
946 if (sample_type
& PERF_SAMPLE_TID
)
947 size
+= sizeof(data
->tid_entry
);
949 if (sample_type
& PERF_SAMPLE_TIME
)
950 size
+= sizeof(data
->time
);
952 if (sample_type
& PERF_SAMPLE_ID
)
953 size
+= sizeof(data
->id
);
955 if (sample_type
& PERF_SAMPLE_STREAM_ID
)
956 size
+= sizeof(data
->stream_id
);
958 if (sample_type
& PERF_SAMPLE_CPU
)
959 size
+= sizeof(data
->cpu_entry
);
961 event
->id_header_size
= size
;
964 static void perf_group_attach(struct perf_event
*event
)
966 struct perf_event
*group_leader
= event
->group_leader
, *pos
;
969 * We can have double attach due to group movement in perf_event_open.
971 if (event
->attach_state
& PERF_ATTACH_GROUP
)
974 event
->attach_state
|= PERF_ATTACH_GROUP
;
976 if (group_leader
== event
)
979 if (group_leader
->group_flags
& PERF_GROUP_SOFTWARE
&&
980 !is_software_event(event
))
981 group_leader
->group_flags
&= ~PERF_GROUP_SOFTWARE
;
983 list_add_tail(&event
->group_entry
, &group_leader
->sibling_list
);
984 group_leader
->nr_siblings
++;
986 perf_event__header_size(group_leader
);
988 list_for_each_entry(pos
, &group_leader
->sibling_list
, group_entry
)
989 perf_event__header_size(pos
);
993 * Remove a event from the lists for its context.
994 * Must be called with ctx->mutex and ctx->lock held.
997 list_del_event(struct perf_event
*event
, struct perf_event_context
*ctx
)
999 struct perf_cpu_context
*cpuctx
;
1001 * We can have double detach due to exit/hot-unplug + close.
1003 if (!(event
->attach_state
& PERF_ATTACH_CONTEXT
))
1006 event
->attach_state
&= ~PERF_ATTACH_CONTEXT
;
1008 if (is_cgroup_event(event
)) {
1010 cpuctx
= __get_cpu_context(ctx
);
1012 * if there are no more cgroup events
1013 * then cler cgrp to avoid stale pointer
1014 * in update_cgrp_time_from_cpuctx()
1016 if (!ctx
->nr_cgroups
)
1017 cpuctx
->cgrp
= NULL
;
1021 if (event
->attr
.inherit_stat
)
1024 list_del_rcu(&event
->event_entry
);
1026 if (event
->group_leader
== event
)
1027 list_del_init(&event
->group_entry
);
1029 update_group_times(event
);
1032 * If event was in error state, then keep it
1033 * that way, otherwise bogus counts will be
1034 * returned on read(). The only way to get out
1035 * of error state is by explicit re-enabling
1038 if (event
->state
> PERF_EVENT_STATE_OFF
)
1039 event
->state
= PERF_EVENT_STATE_OFF
;
1042 static void perf_group_detach(struct perf_event
*event
)
1044 struct perf_event
*sibling
, *tmp
;
1045 struct list_head
*list
= NULL
;
1048 * We can have double detach due to exit/hot-unplug + close.
1050 if (!(event
->attach_state
& PERF_ATTACH_GROUP
))
1053 event
->attach_state
&= ~PERF_ATTACH_GROUP
;
1056 * If this is a sibling, remove it from its group.
1058 if (event
->group_leader
!= event
) {
1059 list_del_init(&event
->group_entry
);
1060 event
->group_leader
->nr_siblings
--;
1064 if (!list_empty(&event
->group_entry
))
1065 list
= &event
->group_entry
;
1068 * If this was a group event with sibling events then
1069 * upgrade the siblings to singleton events by adding them
1070 * to whatever list we are on.
1072 list_for_each_entry_safe(sibling
, tmp
, &event
->sibling_list
, group_entry
) {
1074 list_move_tail(&sibling
->group_entry
, list
);
1075 sibling
->group_leader
= sibling
;
1077 /* Inherit group flags from the previous leader */
1078 sibling
->group_flags
= event
->group_flags
;
1082 perf_event__header_size(event
->group_leader
);
1084 list_for_each_entry(tmp
, &event
->group_leader
->sibling_list
, group_entry
)
1085 perf_event__header_size(tmp
);
1089 event_filter_match(struct perf_event
*event
)
1091 return (event
->cpu
== -1 || event
->cpu
== smp_processor_id())
1092 && perf_cgroup_match(event
);
1096 event_sched_out(struct perf_event
*event
,
1097 struct perf_cpu_context
*cpuctx
,
1098 struct perf_event_context
*ctx
)
1100 u64 tstamp
= perf_event_time(event
);
1103 * An event which could not be activated because of
1104 * filter mismatch still needs to have its timings
1105 * maintained, otherwise bogus information is return
1106 * via read() for time_enabled, time_running:
1108 if (event
->state
== PERF_EVENT_STATE_INACTIVE
1109 && !event_filter_match(event
)) {
1110 delta
= tstamp
- event
->tstamp_stopped
;
1111 event
->tstamp_running
+= delta
;
1112 event
->tstamp_stopped
= tstamp
;
1115 if (event
->state
!= PERF_EVENT_STATE_ACTIVE
)
1118 event
->state
= PERF_EVENT_STATE_INACTIVE
;
1119 if (event
->pending_disable
) {
1120 event
->pending_disable
= 0;
1121 event
->state
= PERF_EVENT_STATE_OFF
;
1123 event
->tstamp_stopped
= tstamp
;
1124 event
->pmu
->del(event
, 0);
1127 if (!is_software_event(event
))
1128 cpuctx
->active_oncpu
--;
1130 if (event
->attr
.exclusive
|| !cpuctx
->active_oncpu
)
1131 cpuctx
->exclusive
= 0;
1135 group_sched_out(struct perf_event
*group_event
,
1136 struct perf_cpu_context
*cpuctx
,
1137 struct perf_event_context
*ctx
)
1139 struct perf_event
*event
;
1140 int state
= group_event
->state
;
1142 event_sched_out(group_event
, cpuctx
, ctx
);
1145 * Schedule out siblings (if any):
1147 list_for_each_entry(event
, &group_event
->sibling_list
, group_entry
)
1148 event_sched_out(event
, cpuctx
, ctx
);
1150 if (state
== PERF_EVENT_STATE_ACTIVE
&& group_event
->attr
.exclusive
)
1151 cpuctx
->exclusive
= 0;
1155 * Cross CPU call to remove a performance event
1157 * We disable the event on the hardware level first. After that we
1158 * remove it from the context list.
1160 static int __perf_remove_from_context(void *info
)
1162 struct perf_event
*event
= info
;
1163 struct perf_event_context
*ctx
= event
->ctx
;
1164 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
1166 raw_spin_lock(&ctx
->lock
);
1167 event_sched_out(event
, cpuctx
, ctx
);
1168 list_del_event(event
, ctx
);
1169 if (!ctx
->nr_events
&& cpuctx
->task_ctx
== ctx
) {
1171 cpuctx
->task_ctx
= NULL
;
1173 raw_spin_unlock(&ctx
->lock
);
1180 * Remove the event from a task's (or a CPU's) list of events.
1182 * CPU events are removed with a smp call. For task events we only
1183 * call when the task is on a CPU.
1185 * If event->ctx is a cloned context, callers must make sure that
1186 * every task struct that event->ctx->task could possibly point to
1187 * remains valid. This is OK when called from perf_release since
1188 * that only calls us on the top-level context, which can't be a clone.
1189 * When called from perf_event_exit_task, it's OK because the
1190 * context has been detached from its task.
1192 static void perf_remove_from_context(struct perf_event
*event
)
1194 struct perf_event_context
*ctx
= event
->ctx
;
1195 struct task_struct
*task
= ctx
->task
;
1197 lockdep_assert_held(&ctx
->mutex
);
1201 * Per cpu events are removed via an smp call and
1202 * the removal is always successful.
1204 cpu_function_call(event
->cpu
, __perf_remove_from_context
, event
);
1209 if (!task_function_call(task
, __perf_remove_from_context
, event
))
1212 raw_spin_lock_irq(&ctx
->lock
);
1214 * If we failed to find a running task, but find the context active now
1215 * that we've acquired the ctx->lock, retry.
1217 if (ctx
->is_active
) {
1218 raw_spin_unlock_irq(&ctx
->lock
);
1223 * Since the task isn't running, its safe to remove the event, us
1224 * holding the ctx->lock ensures the task won't get scheduled in.
1226 list_del_event(event
, ctx
);
1227 raw_spin_unlock_irq(&ctx
->lock
);
1231 * Cross CPU call to disable a performance event
1233 static int __perf_event_disable(void *info
)
1235 struct perf_event
*event
= info
;
1236 struct perf_event_context
*ctx
= event
->ctx
;
1237 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
1240 * If this is a per-task event, need to check whether this
1241 * event's task is the current task on this cpu.
1243 * Can trigger due to concurrent perf_event_context_sched_out()
1244 * flipping contexts around.
1246 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
)
1249 raw_spin_lock(&ctx
->lock
);
1252 * If the event is on, turn it off.
1253 * If it is in error state, leave it in error state.
1255 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
) {
1256 update_context_time(ctx
);
1257 update_cgrp_time_from_event(event
);
1258 update_group_times(event
);
1259 if (event
== event
->group_leader
)
1260 group_sched_out(event
, cpuctx
, ctx
);
1262 event_sched_out(event
, cpuctx
, ctx
);
1263 event
->state
= PERF_EVENT_STATE_OFF
;
1266 raw_spin_unlock(&ctx
->lock
);
1274 * If event->ctx is a cloned context, callers must make sure that
1275 * every task struct that event->ctx->task could possibly point to
1276 * remains valid. This condition is satisifed when called through
1277 * perf_event_for_each_child or perf_event_for_each because they
1278 * hold the top-level event's child_mutex, so any descendant that
1279 * goes to exit will block in sync_child_event.
1280 * When called from perf_pending_event it's OK because event->ctx
1281 * is the current context on this CPU and preemption is disabled,
1282 * hence we can't get into perf_event_task_sched_out for this context.
1284 void perf_event_disable(struct perf_event
*event
)
1286 struct perf_event_context
*ctx
= event
->ctx
;
1287 struct task_struct
*task
= ctx
->task
;
1291 * Disable the event on the cpu that it's on
1293 cpu_function_call(event
->cpu
, __perf_event_disable
, event
);
1298 if (!task_function_call(task
, __perf_event_disable
, event
))
1301 raw_spin_lock_irq(&ctx
->lock
);
1303 * If the event is still active, we need to retry the cross-call.
1305 if (event
->state
== PERF_EVENT_STATE_ACTIVE
) {
1306 raw_spin_unlock_irq(&ctx
->lock
);
1308 * Reload the task pointer, it might have been changed by
1309 * a concurrent perf_event_context_sched_out().
1316 * Since we have the lock this context can't be scheduled
1317 * in, so we can change the state safely.
1319 if (event
->state
== PERF_EVENT_STATE_INACTIVE
) {
1320 update_group_times(event
);
1321 event
->state
= PERF_EVENT_STATE_OFF
;
1323 raw_spin_unlock_irq(&ctx
->lock
);
1326 static void perf_set_shadow_time(struct perf_event
*event
,
1327 struct perf_event_context
*ctx
,
1331 * use the correct time source for the time snapshot
1333 * We could get by without this by leveraging the
1334 * fact that to get to this function, the caller
1335 * has most likely already called update_context_time()
1336 * and update_cgrp_time_xx() and thus both timestamp
1337 * are identical (or very close). Given that tstamp is,
1338 * already adjusted for cgroup, we could say that:
1339 * tstamp - ctx->timestamp
1341 * tstamp - cgrp->timestamp.
1343 * Then, in perf_output_read(), the calculation would
1344 * work with no changes because:
1345 * - event is guaranteed scheduled in
1346 * - no scheduled out in between
1347 * - thus the timestamp would be the same
1349 * But this is a bit hairy.
1351 * So instead, we have an explicit cgroup call to remain
1352 * within the time time source all along. We believe it
1353 * is cleaner and simpler to understand.
1355 if (is_cgroup_event(event
))
1356 perf_cgroup_set_shadow_time(event
, tstamp
);
1358 event
->shadow_ctx_time
= tstamp
- ctx
->timestamp
;
1361 #define MAX_INTERRUPTS (~0ULL)
1363 static void perf_log_throttle(struct perf_event
*event
, int enable
);
1366 event_sched_in(struct perf_event
*event
,
1367 struct perf_cpu_context
*cpuctx
,
1368 struct perf_event_context
*ctx
)
1370 u64 tstamp
= perf_event_time(event
);
1372 if (event
->state
<= PERF_EVENT_STATE_OFF
)
1375 event
->state
= PERF_EVENT_STATE_ACTIVE
;
1376 event
->oncpu
= smp_processor_id();
1379 * Unthrottle events, since we scheduled we might have missed several
1380 * ticks already, also for a heavily scheduling task there is little
1381 * guarantee it'll get a tick in a timely manner.
1383 if (unlikely(event
->hw
.interrupts
== MAX_INTERRUPTS
)) {
1384 perf_log_throttle(event
, 1);
1385 event
->hw
.interrupts
= 0;
1389 * The new state must be visible before we turn it on in the hardware:
1393 if (event
->pmu
->add(event
, PERF_EF_START
)) {
1394 event
->state
= PERF_EVENT_STATE_INACTIVE
;
1399 event
->tstamp_running
+= tstamp
- event
->tstamp_stopped
;
1401 perf_set_shadow_time(event
, ctx
, tstamp
);
1403 if (!is_software_event(event
))
1404 cpuctx
->active_oncpu
++;
1407 if (event
->attr
.exclusive
)
1408 cpuctx
->exclusive
= 1;
1414 group_sched_in(struct perf_event
*group_event
,
1415 struct perf_cpu_context
*cpuctx
,
1416 struct perf_event_context
*ctx
)
1418 struct perf_event
*event
, *partial_group
= NULL
;
1419 struct pmu
*pmu
= group_event
->pmu
;
1420 u64 now
= ctx
->time
;
1421 bool simulate
= false;
1423 if (group_event
->state
== PERF_EVENT_STATE_OFF
)
1426 pmu
->start_txn(pmu
);
1428 if (event_sched_in(group_event
, cpuctx
, ctx
)) {
1429 pmu
->cancel_txn(pmu
);
1434 * Schedule in siblings as one group (if any):
1436 list_for_each_entry(event
, &group_event
->sibling_list
, group_entry
) {
1437 if (event_sched_in(event
, cpuctx
, ctx
)) {
1438 partial_group
= event
;
1443 if (!pmu
->commit_txn(pmu
))
1448 * Groups can be scheduled in as one unit only, so undo any
1449 * partial group before returning:
1450 * The events up to the failed event are scheduled out normally,
1451 * tstamp_stopped will be updated.
1453 * The failed events and the remaining siblings need to have
1454 * their timings updated as if they had gone thru event_sched_in()
1455 * and event_sched_out(). This is required to get consistent timings
1456 * across the group. This also takes care of the case where the group
1457 * could never be scheduled by ensuring tstamp_stopped is set to mark
1458 * the time the event was actually stopped, such that time delta
1459 * calculation in update_event_times() is correct.
1461 list_for_each_entry(event
, &group_event
->sibling_list
, group_entry
) {
1462 if (event
== partial_group
)
1466 event
->tstamp_running
+= now
- event
->tstamp_stopped
;
1467 event
->tstamp_stopped
= now
;
1469 event_sched_out(event
, cpuctx
, ctx
);
1472 event_sched_out(group_event
, cpuctx
, ctx
);
1474 pmu
->cancel_txn(pmu
);
1480 * Work out whether we can put this event group on the CPU now.
1482 static int group_can_go_on(struct perf_event
*event
,
1483 struct perf_cpu_context
*cpuctx
,
1487 * Groups consisting entirely of software events can always go on.
1489 if (event
->group_flags
& PERF_GROUP_SOFTWARE
)
1492 * If an exclusive group is already on, no other hardware
1495 if (cpuctx
->exclusive
)
1498 * If this group is exclusive and there are already
1499 * events on the CPU, it can't go on.
1501 if (event
->attr
.exclusive
&& cpuctx
->active_oncpu
)
1504 * Otherwise, try to add it if all previous groups were able
1510 static void add_event_to_ctx(struct perf_event
*event
,
1511 struct perf_event_context
*ctx
)
1513 u64 tstamp
= perf_event_time(event
);
1515 list_add_event(event
, ctx
);
1516 perf_group_attach(event
);
1517 event
->tstamp_enabled
= tstamp
;
1518 event
->tstamp_running
= tstamp
;
1519 event
->tstamp_stopped
= tstamp
;
1522 static void task_ctx_sched_out(struct perf_event_context
*ctx
);
1524 ctx_sched_in(struct perf_event_context
*ctx
,
1525 struct perf_cpu_context
*cpuctx
,
1526 enum event_type_t event_type
,
1527 struct task_struct
*task
);
1529 static void perf_event_sched_in(struct perf_cpu_context
*cpuctx
,
1530 struct perf_event_context
*ctx
,
1531 struct task_struct
*task
)
1533 cpu_ctx_sched_in(cpuctx
, EVENT_PINNED
, task
);
1535 ctx_sched_in(ctx
, cpuctx
, EVENT_PINNED
, task
);
1536 cpu_ctx_sched_in(cpuctx
, EVENT_FLEXIBLE
, task
);
1538 ctx_sched_in(ctx
, cpuctx
, EVENT_FLEXIBLE
, task
);
1542 * Cross CPU call to install and enable a performance event
1544 * Must be called with ctx->mutex held
1546 static int __perf_install_in_context(void *info
)
1548 struct perf_event
*event
= info
;
1549 struct perf_event_context
*ctx
= event
->ctx
;
1550 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
1551 struct perf_event_context
*task_ctx
= cpuctx
->task_ctx
;
1552 struct task_struct
*task
= current
;
1554 perf_ctx_lock(cpuctx
, task_ctx
);
1555 perf_pmu_disable(cpuctx
->ctx
.pmu
);
1558 * If there was an active task_ctx schedule it out.
1561 task_ctx_sched_out(task_ctx
);
1564 * If the context we're installing events in is not the
1565 * active task_ctx, flip them.
1567 if (ctx
->task
&& task_ctx
!= ctx
) {
1569 raw_spin_unlock(&task_ctx
->lock
);
1570 raw_spin_lock(&ctx
->lock
);
1575 cpuctx
->task_ctx
= task_ctx
;
1576 task
= task_ctx
->task
;
1579 cpu_ctx_sched_out(cpuctx
, EVENT_ALL
);
1581 update_context_time(ctx
);
1583 * update cgrp time only if current cgrp
1584 * matches event->cgrp. Must be done before
1585 * calling add_event_to_ctx()
1587 update_cgrp_time_from_event(event
);
1589 add_event_to_ctx(event
, ctx
);
1592 * Schedule everything back in
1594 perf_event_sched_in(cpuctx
, task_ctx
, task
);
1596 perf_pmu_enable(cpuctx
->ctx
.pmu
);
1597 perf_ctx_unlock(cpuctx
, task_ctx
);
1603 * Attach a performance event to a context
1605 * First we add the event to the list with the hardware enable bit
1606 * in event->hw_config cleared.
1608 * If the event is attached to a task which is on a CPU we use a smp
1609 * call to enable it in the task context. The task might have been
1610 * scheduled away, but we check this in the smp call again.
1613 perf_install_in_context(struct perf_event_context
*ctx
,
1614 struct perf_event
*event
,
1617 struct task_struct
*task
= ctx
->task
;
1619 lockdep_assert_held(&ctx
->mutex
);
1625 * Per cpu events are installed via an smp call and
1626 * the install is always successful.
1628 cpu_function_call(cpu
, __perf_install_in_context
, event
);
1633 if (!task_function_call(task
, __perf_install_in_context
, event
))
1636 raw_spin_lock_irq(&ctx
->lock
);
1638 * If we failed to find a running task, but find the context active now
1639 * that we've acquired the ctx->lock, retry.
1641 if (ctx
->is_active
) {
1642 raw_spin_unlock_irq(&ctx
->lock
);
1647 * Since the task isn't running, its safe to add the event, us holding
1648 * the ctx->lock ensures the task won't get scheduled in.
1650 add_event_to_ctx(event
, ctx
);
1651 raw_spin_unlock_irq(&ctx
->lock
);
1655 * Put a event into inactive state and update time fields.
1656 * Enabling the leader of a group effectively enables all
1657 * the group members that aren't explicitly disabled, so we
1658 * have to update their ->tstamp_enabled also.
1659 * Note: this works for group members as well as group leaders
1660 * since the non-leader members' sibling_lists will be empty.
1662 static void __perf_event_mark_enabled(struct perf_event
*event
,
1663 struct perf_event_context
*ctx
)
1665 struct perf_event
*sub
;
1666 u64 tstamp
= perf_event_time(event
);
1668 event
->state
= PERF_EVENT_STATE_INACTIVE
;
1669 event
->tstamp_enabled
= tstamp
- event
->total_time_enabled
;
1670 list_for_each_entry(sub
, &event
->sibling_list
, group_entry
) {
1671 if (sub
->state
>= PERF_EVENT_STATE_INACTIVE
)
1672 sub
->tstamp_enabled
= tstamp
- sub
->total_time_enabled
;
1677 * Cross CPU call to enable a performance event
1679 static int __perf_event_enable(void *info
)
1681 struct perf_event
*event
= info
;
1682 struct perf_event_context
*ctx
= event
->ctx
;
1683 struct perf_event
*leader
= event
->group_leader
;
1684 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
1687 if (WARN_ON_ONCE(!ctx
->is_active
))
1690 raw_spin_lock(&ctx
->lock
);
1691 update_context_time(ctx
);
1693 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
)
1697 * set current task's cgroup time reference point
1699 perf_cgroup_set_timestamp(current
, ctx
);
1701 __perf_event_mark_enabled(event
, ctx
);
1703 if (!event_filter_match(event
)) {
1704 if (is_cgroup_event(event
))
1705 perf_cgroup_defer_enabled(event
);
1710 * If the event is in a group and isn't the group leader,
1711 * then don't put it on unless the group is on.
1713 if (leader
!= event
&& leader
->state
!= PERF_EVENT_STATE_ACTIVE
)
1716 if (!group_can_go_on(event
, cpuctx
, 1)) {
1719 if (event
== leader
)
1720 err
= group_sched_in(event
, cpuctx
, ctx
);
1722 err
= event_sched_in(event
, cpuctx
, ctx
);
1727 * If this event can't go on and it's part of a
1728 * group, then the whole group has to come off.
1730 if (leader
!= event
)
1731 group_sched_out(leader
, cpuctx
, ctx
);
1732 if (leader
->attr
.pinned
) {
1733 update_group_times(leader
);
1734 leader
->state
= PERF_EVENT_STATE_ERROR
;
1739 raw_spin_unlock(&ctx
->lock
);
1747 * If event->ctx is a cloned context, callers must make sure that
1748 * every task struct that event->ctx->task could possibly point to
1749 * remains valid. This condition is satisfied when called through
1750 * perf_event_for_each_child or perf_event_for_each as described
1751 * for perf_event_disable.
1753 void perf_event_enable(struct perf_event
*event
)
1755 struct perf_event_context
*ctx
= event
->ctx
;
1756 struct task_struct
*task
= ctx
->task
;
1760 * Enable the event on the cpu that it's on
1762 cpu_function_call(event
->cpu
, __perf_event_enable
, event
);
1766 raw_spin_lock_irq(&ctx
->lock
);
1767 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
)
1771 * If the event is in error state, clear that first.
1772 * That way, if we see the event in error state below, we
1773 * know that it has gone back into error state, as distinct
1774 * from the task having been scheduled away before the
1775 * cross-call arrived.
1777 if (event
->state
== PERF_EVENT_STATE_ERROR
)
1778 event
->state
= PERF_EVENT_STATE_OFF
;
1781 if (!ctx
->is_active
) {
1782 __perf_event_mark_enabled(event
, ctx
);
1786 raw_spin_unlock_irq(&ctx
->lock
);
1788 if (!task_function_call(task
, __perf_event_enable
, event
))
1791 raw_spin_lock_irq(&ctx
->lock
);
1794 * If the context is active and the event is still off,
1795 * we need to retry the cross-call.
1797 if (ctx
->is_active
&& event
->state
== PERF_EVENT_STATE_OFF
) {
1799 * task could have been flipped by a concurrent
1800 * perf_event_context_sched_out()
1807 raw_spin_unlock_irq(&ctx
->lock
);
1810 int perf_event_refresh(struct perf_event
*event
, int refresh
)
1813 * not supported on inherited events
1815 if (event
->attr
.inherit
|| !is_sampling_event(event
))
1818 atomic_add(refresh
, &event
->event_limit
);
1819 perf_event_enable(event
);
1823 EXPORT_SYMBOL_GPL(perf_event_refresh
);
1825 static void ctx_sched_out(struct perf_event_context
*ctx
,
1826 struct perf_cpu_context
*cpuctx
,
1827 enum event_type_t event_type
)
1829 struct perf_event
*event
;
1830 int is_active
= ctx
->is_active
;
1832 ctx
->is_active
&= ~event_type
;
1833 if (likely(!ctx
->nr_events
))
1836 update_context_time(ctx
);
1837 update_cgrp_time_from_cpuctx(cpuctx
);
1838 if (!ctx
->nr_active
)
1841 perf_pmu_disable(ctx
->pmu
);
1842 if ((is_active
& EVENT_PINNED
) && (event_type
& EVENT_PINNED
)) {
1843 list_for_each_entry(event
, &ctx
->pinned_groups
, group_entry
)
1844 group_sched_out(event
, cpuctx
, ctx
);
1847 if ((is_active
& EVENT_FLEXIBLE
) && (event_type
& EVENT_FLEXIBLE
)) {
1848 list_for_each_entry(event
, &ctx
->flexible_groups
, group_entry
)
1849 group_sched_out(event
, cpuctx
, ctx
);
1851 perf_pmu_enable(ctx
->pmu
);
1855 * Test whether two contexts are equivalent, i.e. whether they
1856 * have both been cloned from the same version of the same context
1857 * and they both have the same number of enabled events.
1858 * If the number of enabled events is the same, then the set
1859 * of enabled events should be the same, because these are both
1860 * inherited contexts, therefore we can't access individual events
1861 * in them directly with an fd; we can only enable/disable all
1862 * events via prctl, or enable/disable all events in a family
1863 * via ioctl, which will have the same effect on both contexts.
1865 static int context_equiv(struct perf_event_context
*ctx1
,
1866 struct perf_event_context
*ctx2
)
1868 return ctx1
->parent_ctx
&& ctx1
->parent_ctx
== ctx2
->parent_ctx
1869 && ctx1
->parent_gen
== ctx2
->parent_gen
1870 && !ctx1
->pin_count
&& !ctx2
->pin_count
;
1873 static void __perf_event_sync_stat(struct perf_event
*event
,
1874 struct perf_event
*next_event
)
1878 if (!event
->attr
.inherit_stat
)
1882 * Update the event value, we cannot use perf_event_read()
1883 * because we're in the middle of a context switch and have IRQs
1884 * disabled, which upsets smp_call_function_single(), however
1885 * we know the event must be on the current CPU, therefore we
1886 * don't need to use it.
1888 switch (event
->state
) {
1889 case PERF_EVENT_STATE_ACTIVE
:
1890 event
->pmu
->read(event
);
1893 case PERF_EVENT_STATE_INACTIVE
:
1894 update_event_times(event
);
1902 * In order to keep per-task stats reliable we need to flip the event
1903 * values when we flip the contexts.
1905 value
= local64_read(&next_event
->count
);
1906 value
= local64_xchg(&event
->count
, value
);
1907 local64_set(&next_event
->count
, value
);
1909 swap(event
->total_time_enabled
, next_event
->total_time_enabled
);
1910 swap(event
->total_time_running
, next_event
->total_time_running
);
1913 * Since we swizzled the values, update the user visible data too.
1915 perf_event_update_userpage(event
);
1916 perf_event_update_userpage(next_event
);
1919 #define list_next_entry(pos, member) \
1920 list_entry(pos->member.next, typeof(*pos), member)
1922 static void perf_event_sync_stat(struct perf_event_context
*ctx
,
1923 struct perf_event_context
*next_ctx
)
1925 struct perf_event
*event
, *next_event
;
1930 update_context_time(ctx
);
1932 event
= list_first_entry(&ctx
->event_list
,
1933 struct perf_event
, event_entry
);
1935 next_event
= list_first_entry(&next_ctx
->event_list
,
1936 struct perf_event
, event_entry
);
1938 while (&event
->event_entry
!= &ctx
->event_list
&&
1939 &next_event
->event_entry
!= &next_ctx
->event_list
) {
1941 __perf_event_sync_stat(event
, next_event
);
1943 event
= list_next_entry(event
, event_entry
);
1944 next_event
= list_next_entry(next_event
, event_entry
);
1948 static void perf_event_context_sched_out(struct task_struct
*task
, int ctxn
,
1949 struct task_struct
*next
)
1951 struct perf_event_context
*ctx
= task
->perf_event_ctxp
[ctxn
];
1952 struct perf_event_context
*next_ctx
;
1953 struct perf_event_context
*parent
;
1954 struct perf_cpu_context
*cpuctx
;
1960 cpuctx
= __get_cpu_context(ctx
);
1961 if (!cpuctx
->task_ctx
)
1965 parent
= rcu_dereference(ctx
->parent_ctx
);
1966 next_ctx
= next
->perf_event_ctxp
[ctxn
];
1967 if (parent
&& next_ctx
&&
1968 rcu_dereference(next_ctx
->parent_ctx
) == parent
) {
1970 * Looks like the two contexts are clones, so we might be
1971 * able to optimize the context switch. We lock both
1972 * contexts and check that they are clones under the
1973 * lock (including re-checking that neither has been
1974 * uncloned in the meantime). It doesn't matter which
1975 * order we take the locks because no other cpu could
1976 * be trying to lock both of these tasks.
1978 raw_spin_lock(&ctx
->lock
);
1979 raw_spin_lock_nested(&next_ctx
->lock
, SINGLE_DEPTH_NESTING
);
1980 if (context_equiv(ctx
, next_ctx
)) {
1982 * XXX do we need a memory barrier of sorts
1983 * wrt to rcu_dereference() of perf_event_ctxp
1985 task
->perf_event_ctxp
[ctxn
] = next_ctx
;
1986 next
->perf_event_ctxp
[ctxn
] = ctx
;
1988 next_ctx
->task
= task
;
1991 perf_event_sync_stat(ctx
, next_ctx
);
1993 raw_spin_unlock(&next_ctx
->lock
);
1994 raw_spin_unlock(&ctx
->lock
);
1999 raw_spin_lock(&ctx
->lock
);
2000 ctx_sched_out(ctx
, cpuctx
, EVENT_ALL
);
2001 cpuctx
->task_ctx
= NULL
;
2002 raw_spin_unlock(&ctx
->lock
);
2006 #define for_each_task_context_nr(ctxn) \
2007 for ((ctxn) = 0; (ctxn) < perf_nr_task_contexts; (ctxn)++)
2010 * Called from scheduler to remove the events of the current task,
2011 * with interrupts disabled.
2013 * We stop each event and update the event value in event->count.
2015 * This does not protect us against NMI, but disable()
2016 * sets the disabled bit in the control field of event _before_
2017 * accessing the event control register. If a NMI hits, then it will
2018 * not restart the event.
2020 void __perf_event_task_sched_out(struct task_struct
*task
,
2021 struct task_struct
*next
)
2025 for_each_task_context_nr(ctxn
)
2026 perf_event_context_sched_out(task
, ctxn
, next
);
2029 * if cgroup events exist on this CPU, then we need
2030 * to check if we have to switch out PMU state.
2031 * cgroup event are system-wide mode only
2033 if (atomic_read(&__get_cpu_var(perf_cgroup_events
)))
2034 perf_cgroup_sched_out(task
, next
);
2037 static void task_ctx_sched_out(struct perf_event_context
*ctx
)
2039 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
2041 if (!cpuctx
->task_ctx
)
2044 if (WARN_ON_ONCE(ctx
!= cpuctx
->task_ctx
))
2047 ctx_sched_out(ctx
, cpuctx
, EVENT_ALL
);
2048 cpuctx
->task_ctx
= NULL
;
2052 * Called with IRQs disabled
2054 static void cpu_ctx_sched_out(struct perf_cpu_context
*cpuctx
,
2055 enum event_type_t event_type
)
2057 ctx_sched_out(&cpuctx
->ctx
, cpuctx
, event_type
);
2061 ctx_pinned_sched_in(struct perf_event_context
*ctx
,
2062 struct perf_cpu_context
*cpuctx
)
2064 struct perf_event
*event
;
2066 list_for_each_entry(event
, &ctx
->pinned_groups
, group_entry
) {
2067 if (event
->state
<= PERF_EVENT_STATE_OFF
)
2069 if (!event_filter_match(event
))
2072 /* may need to reset tstamp_enabled */
2073 if (is_cgroup_event(event
))
2074 perf_cgroup_mark_enabled(event
, ctx
);
2076 if (group_can_go_on(event
, cpuctx
, 1))
2077 group_sched_in(event
, cpuctx
, ctx
);
2080 * If this pinned group hasn't been scheduled,
2081 * put it in error state.
2083 if (event
->state
== PERF_EVENT_STATE_INACTIVE
) {
2084 update_group_times(event
);
2085 event
->state
= PERF_EVENT_STATE_ERROR
;
2091 ctx_flexible_sched_in(struct perf_event_context
*ctx
,
2092 struct perf_cpu_context
*cpuctx
)
2094 struct perf_event
*event
;
2097 list_for_each_entry(event
, &ctx
->flexible_groups
, group_entry
) {
2098 /* Ignore events in OFF or ERROR state */
2099 if (event
->state
<= PERF_EVENT_STATE_OFF
)
2102 * Listen to the 'cpu' scheduling filter constraint
2105 if (!event_filter_match(event
))
2108 /* may need to reset tstamp_enabled */
2109 if (is_cgroup_event(event
))
2110 perf_cgroup_mark_enabled(event
, ctx
);
2112 if (group_can_go_on(event
, cpuctx
, can_add_hw
)) {
2113 if (group_sched_in(event
, cpuctx
, ctx
))
2120 ctx_sched_in(struct perf_event_context
*ctx
,
2121 struct perf_cpu_context
*cpuctx
,
2122 enum event_type_t event_type
,
2123 struct task_struct
*task
)
2126 int is_active
= ctx
->is_active
;
2128 ctx
->is_active
|= event_type
;
2129 if (likely(!ctx
->nr_events
))
2133 ctx
->timestamp
= now
;
2134 perf_cgroup_set_timestamp(task
, ctx
);
2136 * First go through the list and put on any pinned groups
2137 * in order to give them the best chance of going on.
2139 if (!(is_active
& EVENT_PINNED
) && (event_type
& EVENT_PINNED
))
2140 ctx_pinned_sched_in(ctx
, cpuctx
);
2142 /* Then walk through the lower prio flexible groups */
2143 if (!(is_active
& EVENT_FLEXIBLE
) && (event_type
& EVENT_FLEXIBLE
))
2144 ctx_flexible_sched_in(ctx
, cpuctx
);
2147 static void cpu_ctx_sched_in(struct perf_cpu_context
*cpuctx
,
2148 enum event_type_t event_type
,
2149 struct task_struct
*task
)
2151 struct perf_event_context
*ctx
= &cpuctx
->ctx
;
2153 ctx_sched_in(ctx
, cpuctx
, event_type
, task
);
2156 static void perf_event_context_sched_in(struct perf_event_context
*ctx
,
2157 struct task_struct
*task
)
2159 struct perf_cpu_context
*cpuctx
;
2161 cpuctx
= __get_cpu_context(ctx
);
2162 if (cpuctx
->task_ctx
== ctx
)
2165 perf_ctx_lock(cpuctx
, ctx
);
2166 perf_pmu_disable(ctx
->pmu
);
2168 * We want to keep the following priority order:
2169 * cpu pinned (that don't need to move), task pinned,
2170 * cpu flexible, task flexible.
2172 cpu_ctx_sched_out(cpuctx
, EVENT_FLEXIBLE
);
2174 perf_event_sched_in(cpuctx
, ctx
, task
);
2176 cpuctx
->task_ctx
= ctx
;
2178 perf_pmu_enable(ctx
->pmu
);
2179 perf_ctx_unlock(cpuctx
, ctx
);
2182 * Since these rotations are per-cpu, we need to ensure the
2183 * cpu-context we got scheduled on is actually rotating.
2185 perf_pmu_rotate_start(ctx
->pmu
);
2189 * Called from scheduler to add the events of the current task
2190 * with interrupts disabled.
2192 * We restore the event value and then enable it.
2194 * This does not protect us against NMI, but enable()
2195 * sets the enabled bit in the control field of event _before_
2196 * accessing the event control register. If a NMI hits, then it will
2197 * keep the event running.
2199 void __perf_event_task_sched_in(struct task_struct
*prev
,
2200 struct task_struct
*task
)
2202 struct perf_event_context
*ctx
;
2205 for_each_task_context_nr(ctxn
) {
2206 ctx
= task
->perf_event_ctxp
[ctxn
];
2210 perf_event_context_sched_in(ctx
, task
);
2213 * if cgroup events exist on this CPU, then we need
2214 * to check if we have to switch in PMU state.
2215 * cgroup event are system-wide mode only
2217 if (atomic_read(&__get_cpu_var(perf_cgroup_events
)))
2218 perf_cgroup_sched_in(prev
, task
);
2221 static u64
perf_calculate_period(struct perf_event
*event
, u64 nsec
, u64 count
)
2223 u64 frequency
= event
->attr
.sample_freq
;
2224 u64 sec
= NSEC_PER_SEC
;
2225 u64 divisor
, dividend
;
2227 int count_fls
, nsec_fls
, frequency_fls
, sec_fls
;
2229 count_fls
= fls64(count
);
2230 nsec_fls
= fls64(nsec
);
2231 frequency_fls
= fls64(frequency
);
2235 * We got @count in @nsec, with a target of sample_freq HZ
2236 * the target period becomes:
2239 * period = -------------------
2240 * @nsec * sample_freq
2245 * Reduce accuracy by one bit such that @a and @b converge
2246 * to a similar magnitude.
2248 #define REDUCE_FLS(a, b) \
2250 if (a##_fls > b##_fls) { \
2260 * Reduce accuracy until either term fits in a u64, then proceed with
2261 * the other, so that finally we can do a u64/u64 division.
2263 while (count_fls
+ sec_fls
> 64 && nsec_fls
+ frequency_fls
> 64) {
2264 REDUCE_FLS(nsec
, frequency
);
2265 REDUCE_FLS(sec
, count
);
2268 if (count_fls
+ sec_fls
> 64) {
2269 divisor
= nsec
* frequency
;
2271 while (count_fls
+ sec_fls
> 64) {
2272 REDUCE_FLS(count
, sec
);
2276 dividend
= count
* sec
;
2278 dividend
= count
* sec
;
2280 while (nsec_fls
+ frequency_fls
> 64) {
2281 REDUCE_FLS(nsec
, frequency
);
2285 divisor
= nsec
* frequency
;
2291 return div64_u64(dividend
, divisor
);
2294 static void perf_adjust_period(struct perf_event
*event
, u64 nsec
, u64 count
)
2296 struct hw_perf_event
*hwc
= &event
->hw
;
2297 s64 period
, sample_period
;
2300 period
= perf_calculate_period(event
, nsec
, count
);
2302 delta
= (s64
)(period
- hwc
->sample_period
);
2303 delta
= (delta
+ 7) / 8; /* low pass filter */
2305 sample_period
= hwc
->sample_period
+ delta
;
2310 hwc
->sample_period
= sample_period
;
2312 if (local64_read(&hwc
->period_left
) > 8*sample_period
) {
2313 event
->pmu
->stop(event
, PERF_EF_UPDATE
);
2314 local64_set(&hwc
->period_left
, 0);
2315 event
->pmu
->start(event
, PERF_EF_RELOAD
);
2319 static void perf_ctx_adjust_freq(struct perf_event_context
*ctx
, u64 period
)
2321 struct perf_event
*event
;
2322 struct hw_perf_event
*hwc
;
2323 u64 interrupts
, now
;
2326 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
2327 if (event
->state
!= PERF_EVENT_STATE_ACTIVE
)
2330 if (!event_filter_match(event
))
2335 interrupts
= hwc
->interrupts
;
2336 hwc
->interrupts
= 0;
2339 * unthrottle events on the tick
2341 if (interrupts
== MAX_INTERRUPTS
) {
2342 perf_log_throttle(event
, 1);
2343 event
->pmu
->start(event
, 0);
2346 if (!event
->attr
.freq
|| !event
->attr
.sample_freq
)
2349 event
->pmu
->read(event
);
2350 now
= local64_read(&event
->count
);
2351 delta
= now
- hwc
->freq_count_stamp
;
2352 hwc
->freq_count_stamp
= now
;
2355 perf_adjust_period(event
, period
, delta
);
2360 * Round-robin a context's events:
2362 static void rotate_ctx(struct perf_event_context
*ctx
)
2365 * Rotate the first entry last of non-pinned groups. Rotation might be
2366 * disabled by the inheritance code.
2368 if (!ctx
->rotate_disable
)
2369 list_rotate_left(&ctx
->flexible_groups
);
2373 * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
2374 * because they're strictly cpu affine and rotate_start is called with IRQs
2375 * disabled, while rotate_context is called from IRQ context.
2377 static void perf_rotate_context(struct perf_cpu_context
*cpuctx
)
2379 u64 interval
= (u64
)cpuctx
->jiffies_interval
* TICK_NSEC
;
2380 struct perf_event_context
*ctx
= NULL
;
2381 int rotate
= 0, remove
= 1;
2383 if (cpuctx
->ctx
.nr_events
) {
2385 if (cpuctx
->ctx
.nr_events
!= cpuctx
->ctx
.nr_active
)
2389 ctx
= cpuctx
->task_ctx
;
2390 if (ctx
&& ctx
->nr_events
) {
2392 if (ctx
->nr_events
!= ctx
->nr_active
)
2396 perf_ctx_lock(cpuctx
, cpuctx
->task_ctx
);
2397 perf_pmu_disable(cpuctx
->ctx
.pmu
);
2398 perf_ctx_adjust_freq(&cpuctx
->ctx
, interval
);
2400 perf_ctx_adjust_freq(ctx
, interval
);
2405 cpu_ctx_sched_out(cpuctx
, EVENT_FLEXIBLE
);
2407 ctx_sched_out(ctx
, cpuctx
, EVENT_FLEXIBLE
);
2409 rotate_ctx(&cpuctx
->ctx
);
2413 perf_event_sched_in(cpuctx
, ctx
, current
);
2417 list_del_init(&cpuctx
->rotation_list
);
2419 perf_pmu_enable(cpuctx
->ctx
.pmu
);
2420 perf_ctx_unlock(cpuctx
, cpuctx
->task_ctx
);
2423 void perf_event_task_tick(void)
2425 struct list_head
*head
= &__get_cpu_var(rotation_list
);
2426 struct perf_cpu_context
*cpuctx
, *tmp
;
2428 WARN_ON(!irqs_disabled());
2430 list_for_each_entry_safe(cpuctx
, tmp
, head
, rotation_list
) {
2431 if (cpuctx
->jiffies_interval
== 1 ||
2432 !(jiffies
% cpuctx
->jiffies_interval
))
2433 perf_rotate_context(cpuctx
);
2437 static int event_enable_on_exec(struct perf_event
*event
,
2438 struct perf_event_context
*ctx
)
2440 if (!event
->attr
.enable_on_exec
)
2443 event
->attr
.enable_on_exec
= 0;
2444 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
)
2447 __perf_event_mark_enabled(event
, ctx
);
2453 * Enable all of a task's events that have been marked enable-on-exec.
2454 * This expects task == current.
2456 static void perf_event_enable_on_exec(struct perf_event_context
*ctx
)
2458 struct perf_event
*event
;
2459 unsigned long flags
;
2463 local_irq_save(flags
);
2464 if (!ctx
|| !ctx
->nr_events
)
2468 * We must ctxsw out cgroup events to avoid conflict
2469 * when invoking perf_task_event_sched_in() later on
2470 * in this function. Otherwise we end up trying to
2471 * ctxswin cgroup events which are already scheduled
2474 perf_cgroup_sched_out(current
, NULL
);
2476 raw_spin_lock(&ctx
->lock
);
2477 task_ctx_sched_out(ctx
);
2479 list_for_each_entry(event
, &ctx
->pinned_groups
, group_entry
) {
2480 ret
= event_enable_on_exec(event
, ctx
);
2485 list_for_each_entry(event
, &ctx
->flexible_groups
, group_entry
) {
2486 ret
= event_enable_on_exec(event
, ctx
);
2492 * Unclone this context if we enabled any event.
2497 raw_spin_unlock(&ctx
->lock
);
2500 * Also calls ctxswin for cgroup events, if any:
2502 perf_event_context_sched_in(ctx
, ctx
->task
);
2504 local_irq_restore(flags
);
2508 * Cross CPU call to read the hardware event
2510 static void __perf_event_read(void *info
)
2512 struct perf_event
*event
= info
;
2513 struct perf_event_context
*ctx
= event
->ctx
;
2514 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
2517 * If this is a task context, we need to check whether it is
2518 * the current task context of this cpu. If not it has been
2519 * scheduled out before the smp call arrived. In that case
2520 * event->count would have been updated to a recent sample
2521 * when the event was scheduled out.
2523 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
)
2526 raw_spin_lock(&ctx
->lock
);
2527 if (ctx
->is_active
) {
2528 update_context_time(ctx
);
2529 update_cgrp_time_from_event(event
);
2531 update_event_times(event
);
2532 if (event
->state
== PERF_EVENT_STATE_ACTIVE
)
2533 event
->pmu
->read(event
);
2534 raw_spin_unlock(&ctx
->lock
);
2537 static inline u64
perf_event_count(struct perf_event
*event
)
2539 return local64_read(&event
->count
) + atomic64_read(&event
->child_count
);
2542 static u64
perf_event_read(struct perf_event
*event
)
2545 * If event is enabled and currently active on a CPU, update the
2546 * value in the event structure:
2548 if (event
->state
== PERF_EVENT_STATE_ACTIVE
) {
2549 smp_call_function_single(event
->oncpu
,
2550 __perf_event_read
, event
, 1);
2551 } else if (event
->state
== PERF_EVENT_STATE_INACTIVE
) {
2552 struct perf_event_context
*ctx
= event
->ctx
;
2553 unsigned long flags
;
2555 raw_spin_lock_irqsave(&ctx
->lock
, flags
);
2557 * may read while context is not active
2558 * (e.g., thread is blocked), in that case
2559 * we cannot update context time
2561 if (ctx
->is_active
) {
2562 update_context_time(ctx
);
2563 update_cgrp_time_from_event(event
);
2565 update_event_times(event
);
2566 raw_spin_unlock_irqrestore(&ctx
->lock
, flags
);
2569 return perf_event_count(event
);
2573 * Initialize the perf_event context in a task_struct:
2575 static void __perf_event_init_context(struct perf_event_context
*ctx
)
2577 raw_spin_lock_init(&ctx
->lock
);
2578 mutex_init(&ctx
->mutex
);
2579 INIT_LIST_HEAD(&ctx
->pinned_groups
);
2580 INIT_LIST_HEAD(&ctx
->flexible_groups
);
2581 INIT_LIST_HEAD(&ctx
->event_list
);
2582 atomic_set(&ctx
->refcount
, 1);
2585 static struct perf_event_context
*
2586 alloc_perf_context(struct pmu
*pmu
, struct task_struct
*task
)
2588 struct perf_event_context
*ctx
;
2590 ctx
= kzalloc(sizeof(struct perf_event_context
), GFP_KERNEL
);
2594 __perf_event_init_context(ctx
);
2597 get_task_struct(task
);
2604 static struct task_struct
*
2605 find_lively_task_by_vpid(pid_t vpid
)
2607 struct task_struct
*task
;
2614 task
= find_task_by_vpid(vpid
);
2616 get_task_struct(task
);
2620 return ERR_PTR(-ESRCH
);
2622 /* Reuse ptrace permission checks for now. */
2624 if (!ptrace_may_access(task
, PTRACE_MODE_READ
))
2629 put_task_struct(task
);
2630 return ERR_PTR(err
);
2635 * Returns a matching context with refcount and pincount.
2637 static struct perf_event_context
*
2638 find_get_context(struct pmu
*pmu
, struct task_struct
*task
, int cpu
)
2640 struct perf_event_context
*ctx
;
2641 struct perf_cpu_context
*cpuctx
;
2642 unsigned long flags
;
2646 /* Must be root to operate on a CPU event: */
2647 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN
))
2648 return ERR_PTR(-EACCES
);
2651 * We could be clever and allow to attach a event to an
2652 * offline CPU and activate it when the CPU comes up, but
2655 if (!cpu_online(cpu
))
2656 return ERR_PTR(-ENODEV
);
2658 cpuctx
= per_cpu_ptr(pmu
->pmu_cpu_context
, cpu
);
2667 ctxn
= pmu
->task_ctx_nr
;
2672 ctx
= perf_lock_task_context(task
, ctxn
, &flags
);
2676 raw_spin_unlock_irqrestore(&ctx
->lock
, flags
);
2678 ctx
= alloc_perf_context(pmu
, task
);
2684 mutex_lock(&task
->perf_event_mutex
);
2686 * If it has already passed perf_event_exit_task().
2687 * we must see PF_EXITING, it takes this mutex too.
2689 if (task
->flags
& PF_EXITING
)
2691 else if (task
->perf_event_ctxp
[ctxn
])
2696 rcu_assign_pointer(task
->perf_event_ctxp
[ctxn
], ctx
);
2698 mutex_unlock(&task
->perf_event_mutex
);
2700 if (unlikely(err
)) {
2712 return ERR_PTR(err
);
2715 static void perf_event_free_filter(struct perf_event
*event
);
2717 static void free_event_rcu(struct rcu_head
*head
)
2719 struct perf_event
*event
;
2721 event
= container_of(head
, struct perf_event
, rcu_head
);
2723 put_pid_ns(event
->ns
);
2724 perf_event_free_filter(event
);
2728 static void ring_buffer_put(struct ring_buffer
*rb
);
2730 static void free_event(struct perf_event
*event
)
2732 irq_work_sync(&event
->pending
);
2734 if (!event
->parent
) {
2735 if (event
->attach_state
& PERF_ATTACH_TASK
)
2736 jump_label_dec(&perf_sched_events
);
2737 if (event
->attr
.mmap
|| event
->attr
.mmap_data
)
2738 atomic_dec(&nr_mmap_events
);
2739 if (event
->attr
.comm
)
2740 atomic_dec(&nr_comm_events
);
2741 if (event
->attr
.task
)
2742 atomic_dec(&nr_task_events
);
2743 if (event
->attr
.sample_type
& PERF_SAMPLE_CALLCHAIN
)
2744 put_callchain_buffers();
2745 if (is_cgroup_event(event
)) {
2746 atomic_dec(&per_cpu(perf_cgroup_events
, event
->cpu
));
2747 jump_label_dec(&perf_sched_events
);
2752 ring_buffer_put(event
->rb
);
2756 if (is_cgroup_event(event
))
2757 perf_detach_cgroup(event
);
2760 event
->destroy(event
);
2763 put_ctx(event
->ctx
);
2765 call_rcu(&event
->rcu_head
, free_event_rcu
);
2768 int perf_event_release_kernel(struct perf_event
*event
)
2770 struct perf_event_context
*ctx
= event
->ctx
;
2772 WARN_ON_ONCE(ctx
->parent_ctx
);
2774 * There are two ways this annotation is useful:
2776 * 1) there is a lock recursion from perf_event_exit_task
2777 * see the comment there.
2779 * 2) there is a lock-inversion with mmap_sem through
2780 * perf_event_read_group(), which takes faults while
2781 * holding ctx->mutex, however this is called after
2782 * the last filedesc died, so there is no possibility
2783 * to trigger the AB-BA case.
2785 mutex_lock_nested(&ctx
->mutex
, SINGLE_DEPTH_NESTING
);
2786 raw_spin_lock_irq(&ctx
->lock
);
2787 perf_group_detach(event
);
2788 raw_spin_unlock_irq(&ctx
->lock
);
2789 perf_remove_from_context(event
);
2790 mutex_unlock(&ctx
->mutex
);
2796 EXPORT_SYMBOL_GPL(perf_event_release_kernel
);
2799 * Called when the last reference to the file is gone.
2801 static int perf_release(struct inode
*inode
, struct file
*file
)
2803 struct perf_event
*event
= file
->private_data
;
2804 struct task_struct
*owner
;
2806 file
->private_data
= NULL
;
2809 owner
= ACCESS_ONCE(event
->owner
);
2811 * Matches the smp_wmb() in perf_event_exit_task(). If we observe
2812 * !owner it means the list deletion is complete and we can indeed
2813 * free this event, otherwise we need to serialize on
2814 * owner->perf_event_mutex.
2816 smp_read_barrier_depends();
2819 * Since delayed_put_task_struct() also drops the last
2820 * task reference we can safely take a new reference
2821 * while holding the rcu_read_lock().
2823 get_task_struct(owner
);
2828 mutex_lock(&owner
->perf_event_mutex
);
2830 * We have to re-check the event->owner field, if it is cleared
2831 * we raced with perf_event_exit_task(), acquiring the mutex
2832 * ensured they're done, and we can proceed with freeing the
2836 list_del_init(&event
->owner_entry
);
2837 mutex_unlock(&owner
->perf_event_mutex
);
2838 put_task_struct(owner
);
2841 return perf_event_release_kernel(event
);
2844 u64
perf_event_read_value(struct perf_event
*event
, u64
*enabled
, u64
*running
)
2846 struct perf_event
*child
;
2852 mutex_lock(&event
->child_mutex
);
2853 total
+= perf_event_read(event
);
2854 *enabled
+= event
->total_time_enabled
+
2855 atomic64_read(&event
->child_total_time_enabled
);
2856 *running
+= event
->total_time_running
+
2857 atomic64_read(&event
->child_total_time_running
);
2859 list_for_each_entry(child
, &event
->child_list
, child_list
) {
2860 total
+= perf_event_read(child
);
2861 *enabled
+= child
->total_time_enabled
;
2862 *running
+= child
->total_time_running
;
2864 mutex_unlock(&event
->child_mutex
);
2868 EXPORT_SYMBOL_GPL(perf_event_read_value
);
2870 static int perf_event_read_group(struct perf_event
*event
,
2871 u64 read_format
, char __user
*buf
)
2873 struct perf_event
*leader
= event
->group_leader
, *sub
;
2874 int n
= 0, size
= 0, ret
= -EFAULT
;
2875 struct perf_event_context
*ctx
= leader
->ctx
;
2877 u64 count
, enabled
, running
;
2879 mutex_lock(&ctx
->mutex
);
2880 count
= perf_event_read_value(leader
, &enabled
, &running
);
2882 values
[n
++] = 1 + leader
->nr_siblings
;
2883 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
2884 values
[n
++] = enabled
;
2885 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
2886 values
[n
++] = running
;
2887 values
[n
++] = count
;
2888 if (read_format
& PERF_FORMAT_ID
)
2889 values
[n
++] = primary_event_id(leader
);
2891 size
= n
* sizeof(u64
);
2893 if (copy_to_user(buf
, values
, size
))
2898 list_for_each_entry(sub
, &leader
->sibling_list
, group_entry
) {
2901 values
[n
++] = perf_event_read_value(sub
, &enabled
, &running
);
2902 if (read_format
& PERF_FORMAT_ID
)
2903 values
[n
++] = primary_event_id(sub
);
2905 size
= n
* sizeof(u64
);
2907 if (copy_to_user(buf
+ ret
, values
, size
)) {
2915 mutex_unlock(&ctx
->mutex
);
2920 static int perf_event_read_one(struct perf_event
*event
,
2921 u64 read_format
, char __user
*buf
)
2923 u64 enabled
, running
;
2927 values
[n
++] = perf_event_read_value(event
, &enabled
, &running
);
2928 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
2929 values
[n
++] = enabled
;
2930 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
2931 values
[n
++] = running
;
2932 if (read_format
& PERF_FORMAT_ID
)
2933 values
[n
++] = primary_event_id(event
);
2935 if (copy_to_user(buf
, values
, n
* sizeof(u64
)))
2938 return n
* sizeof(u64
);
2942 * Read the performance event - simple non blocking version for now
2945 perf_read_hw(struct perf_event
*event
, char __user
*buf
, size_t count
)
2947 u64 read_format
= event
->attr
.read_format
;
2951 * Return end-of-file for a read on a event that is in
2952 * error state (i.e. because it was pinned but it couldn't be
2953 * scheduled on to the CPU at some point).
2955 if (event
->state
== PERF_EVENT_STATE_ERROR
)
2958 if (count
< event
->read_size
)
2961 WARN_ON_ONCE(event
->ctx
->parent_ctx
);
2962 if (read_format
& PERF_FORMAT_GROUP
)
2963 ret
= perf_event_read_group(event
, read_format
, buf
);
2965 ret
= perf_event_read_one(event
, read_format
, buf
);
2971 perf_read(struct file
*file
, char __user
*buf
, size_t count
, loff_t
*ppos
)
2973 struct perf_event
*event
= file
->private_data
;
2975 return perf_read_hw(event
, buf
, count
);
2978 static unsigned int perf_poll(struct file
*file
, poll_table
*wait
)
2980 struct perf_event
*event
= file
->private_data
;
2981 struct ring_buffer
*rb
;
2982 unsigned int events
= POLL_HUP
;
2985 rb
= rcu_dereference(event
->rb
);
2987 events
= atomic_xchg(&rb
->poll
, 0);
2990 poll_wait(file
, &event
->waitq
, wait
);
2995 static void perf_event_reset(struct perf_event
*event
)
2997 (void)perf_event_read(event
);
2998 local64_set(&event
->count
, 0);
2999 perf_event_update_userpage(event
);
3003 * Holding the top-level event's child_mutex means that any
3004 * descendant process that has inherited this event will block
3005 * in sync_child_event if it goes to exit, thus satisfying the
3006 * task existence requirements of perf_event_enable/disable.
3008 static void perf_event_for_each_child(struct perf_event
*event
,
3009 void (*func
)(struct perf_event
*))
3011 struct perf_event
*child
;
3013 WARN_ON_ONCE(event
->ctx
->parent_ctx
);
3014 mutex_lock(&event
->child_mutex
);
3016 list_for_each_entry(child
, &event
->child_list
, child_list
)
3018 mutex_unlock(&event
->child_mutex
);
3021 static void perf_event_for_each(struct perf_event
*event
,
3022 void (*func
)(struct perf_event
*))
3024 struct perf_event_context
*ctx
= event
->ctx
;
3025 struct perf_event
*sibling
;
3027 WARN_ON_ONCE(ctx
->parent_ctx
);
3028 mutex_lock(&ctx
->mutex
);
3029 event
= event
->group_leader
;
3031 perf_event_for_each_child(event
, func
);
3033 list_for_each_entry(sibling
, &event
->sibling_list
, group_entry
)
3034 perf_event_for_each_child(event
, func
);
3035 mutex_unlock(&ctx
->mutex
);
3038 static int perf_event_period(struct perf_event
*event
, u64 __user
*arg
)
3040 struct perf_event_context
*ctx
= event
->ctx
;
3044 if (!is_sampling_event(event
))
3047 if (copy_from_user(&value
, arg
, sizeof(value
)))
3053 raw_spin_lock_irq(&ctx
->lock
);
3054 if (event
->attr
.freq
) {
3055 if (value
> sysctl_perf_event_sample_rate
) {
3060 event
->attr
.sample_freq
= value
;
3062 event
->attr
.sample_period
= value
;
3063 event
->hw
.sample_period
= value
;
3066 raw_spin_unlock_irq(&ctx
->lock
);
3071 static const struct file_operations perf_fops
;
3073 static struct perf_event
*perf_fget_light(int fd
, int *fput_needed
)
3077 file
= fget_light(fd
, fput_needed
);
3079 return ERR_PTR(-EBADF
);
3081 if (file
->f_op
!= &perf_fops
) {
3082 fput_light(file
, *fput_needed
);
3084 return ERR_PTR(-EBADF
);
3087 return file
->private_data
;
3090 static int perf_event_set_output(struct perf_event
*event
,
3091 struct perf_event
*output_event
);
3092 static int perf_event_set_filter(struct perf_event
*event
, void __user
*arg
);
3094 static long perf_ioctl(struct file
*file
, unsigned int cmd
, unsigned long arg
)
3096 struct perf_event
*event
= file
->private_data
;
3097 void (*func
)(struct perf_event
*);
3101 case PERF_EVENT_IOC_ENABLE
:
3102 func
= perf_event_enable
;
3104 case PERF_EVENT_IOC_DISABLE
:
3105 func
= perf_event_disable
;
3107 case PERF_EVENT_IOC_RESET
:
3108 func
= perf_event_reset
;
3111 case PERF_EVENT_IOC_REFRESH
:
3112 return perf_event_refresh(event
, arg
);
3114 case PERF_EVENT_IOC_PERIOD
:
3115 return perf_event_period(event
, (u64 __user
*)arg
);
3117 case PERF_EVENT_IOC_SET_OUTPUT
:
3119 struct perf_event
*output_event
= NULL
;
3120 int fput_needed
= 0;
3124 output_event
= perf_fget_light(arg
, &fput_needed
);
3125 if (IS_ERR(output_event
))
3126 return PTR_ERR(output_event
);
3129 ret
= perf_event_set_output(event
, output_event
);
3131 fput_light(output_event
->filp
, fput_needed
);
3136 case PERF_EVENT_IOC_SET_FILTER
:
3137 return perf_event_set_filter(event
, (void __user
*)arg
);
3143 if (flags
& PERF_IOC_FLAG_GROUP
)
3144 perf_event_for_each(event
, func
);
3146 perf_event_for_each_child(event
, func
);
3151 int perf_event_task_enable(void)
3153 struct perf_event
*event
;
3155 mutex_lock(¤t
->perf_event_mutex
);
3156 list_for_each_entry(event
, ¤t
->perf_event_list
, owner_entry
)
3157 perf_event_for_each_child(event
, perf_event_enable
);
3158 mutex_unlock(¤t
->perf_event_mutex
);
3163 int perf_event_task_disable(void)
3165 struct perf_event
*event
;
3167 mutex_lock(¤t
->perf_event_mutex
);
3168 list_for_each_entry(event
, ¤t
->perf_event_list
, owner_entry
)
3169 perf_event_for_each_child(event
, perf_event_disable
);
3170 mutex_unlock(¤t
->perf_event_mutex
);
3175 #ifndef PERF_EVENT_INDEX_OFFSET
3176 # define PERF_EVENT_INDEX_OFFSET 0
3179 static int perf_event_index(struct perf_event
*event
)
3181 if (event
->hw
.state
& PERF_HES_STOPPED
)
3184 if (event
->state
!= PERF_EVENT_STATE_ACTIVE
)
3187 return event
->hw
.idx
+ 1 - PERF_EVENT_INDEX_OFFSET
;
3190 static void calc_timer_values(struct perf_event
*event
,
3197 ctx_time
= event
->shadow_ctx_time
+ now
;
3198 *enabled
= ctx_time
- event
->tstamp_enabled
;
3199 *running
= ctx_time
- event
->tstamp_running
;
3203 * Callers need to ensure there can be no nesting of this function, otherwise
3204 * the seqlock logic goes bad. We can not serialize this because the arch
3205 * code calls this from NMI context.
3207 void perf_event_update_userpage(struct perf_event
*event
)
3209 struct perf_event_mmap_page
*userpg
;
3210 struct ring_buffer
*rb
;
3211 u64 enabled
, running
;
3215 * compute total_time_enabled, total_time_running
3216 * based on snapshot values taken when the event
3217 * was last scheduled in.
3219 * we cannot simply called update_context_time()
3220 * because of locking issue as we can be called in
3223 calc_timer_values(event
, &enabled
, &running
);
3224 rb
= rcu_dereference(event
->rb
);
3228 userpg
= rb
->user_page
;
3231 * Disable preemption so as to not let the corresponding user-space
3232 * spin too long if we get preempted.
3237 userpg
->index
= perf_event_index(event
);
3238 userpg
->offset
= perf_event_count(event
);
3239 if (event
->state
== PERF_EVENT_STATE_ACTIVE
)
3240 userpg
->offset
-= local64_read(&event
->hw
.prev_count
);
3242 userpg
->time_enabled
= enabled
+
3243 atomic64_read(&event
->child_total_time_enabled
);
3245 userpg
->time_running
= running
+
3246 atomic64_read(&event
->child_total_time_running
);
3255 static int perf_mmap_fault(struct vm_area_struct
*vma
, struct vm_fault
*vmf
)
3257 struct perf_event
*event
= vma
->vm_file
->private_data
;
3258 struct ring_buffer
*rb
;
3259 int ret
= VM_FAULT_SIGBUS
;
3261 if (vmf
->flags
& FAULT_FLAG_MKWRITE
) {
3262 if (vmf
->pgoff
== 0)
3268 rb
= rcu_dereference(event
->rb
);
3272 if (vmf
->pgoff
&& (vmf
->flags
& FAULT_FLAG_WRITE
))
3275 vmf
->page
= perf_mmap_to_page(rb
, vmf
->pgoff
);
3279 get_page(vmf
->page
);
3280 vmf
->page
->mapping
= vma
->vm_file
->f_mapping
;
3281 vmf
->page
->index
= vmf
->pgoff
;
3290 static void rb_free_rcu(struct rcu_head
*rcu_head
)
3292 struct ring_buffer
*rb
;
3294 rb
= container_of(rcu_head
, struct ring_buffer
, rcu_head
);
3298 static struct ring_buffer
*ring_buffer_get(struct perf_event
*event
)
3300 struct ring_buffer
*rb
;
3303 rb
= rcu_dereference(event
->rb
);
3305 if (!atomic_inc_not_zero(&rb
->refcount
))
3313 static void ring_buffer_put(struct ring_buffer
*rb
)
3315 if (!atomic_dec_and_test(&rb
->refcount
))
3318 call_rcu(&rb
->rcu_head
, rb_free_rcu
);
3321 static void perf_mmap_open(struct vm_area_struct
*vma
)
3323 struct perf_event
*event
= vma
->vm_file
->private_data
;
3325 atomic_inc(&event
->mmap_count
);
3328 static void perf_mmap_close(struct vm_area_struct
*vma
)
3330 struct perf_event
*event
= vma
->vm_file
->private_data
;
3332 if (atomic_dec_and_mutex_lock(&event
->mmap_count
, &event
->mmap_mutex
)) {
3333 unsigned long size
= perf_data_size(event
->rb
);
3334 struct user_struct
*user
= event
->mmap_user
;
3335 struct ring_buffer
*rb
= event
->rb
;
3337 atomic_long_sub((size
>> PAGE_SHIFT
) + 1, &user
->locked_vm
);
3338 vma
->vm_mm
->pinned_vm
-= event
->mmap_locked
;
3339 rcu_assign_pointer(event
->rb
, NULL
);
3340 mutex_unlock(&event
->mmap_mutex
);
3342 ring_buffer_put(rb
);
3347 static const struct vm_operations_struct perf_mmap_vmops
= {
3348 .open
= perf_mmap_open
,
3349 .close
= perf_mmap_close
,
3350 .fault
= perf_mmap_fault
,
3351 .page_mkwrite
= perf_mmap_fault
,
3354 static int perf_mmap(struct file
*file
, struct vm_area_struct
*vma
)
3356 struct perf_event
*event
= file
->private_data
;
3357 unsigned long user_locked
, user_lock_limit
;
3358 struct user_struct
*user
= current_user();
3359 unsigned long locked
, lock_limit
;
3360 struct ring_buffer
*rb
;
3361 unsigned long vma_size
;
3362 unsigned long nr_pages
;
3363 long user_extra
, extra
;
3364 int ret
= 0, flags
= 0;
3367 * Don't allow mmap() of inherited per-task counters. This would
3368 * create a performance issue due to all children writing to the
3371 if (event
->cpu
== -1 && event
->attr
.inherit
)
3374 if (!(vma
->vm_flags
& VM_SHARED
))
3377 vma_size
= vma
->vm_end
- vma
->vm_start
;
3378 nr_pages
= (vma_size
/ PAGE_SIZE
) - 1;
3381 * If we have rb pages ensure they're a power-of-two number, so we
3382 * can do bitmasks instead of modulo.
3384 if (nr_pages
!= 0 && !is_power_of_2(nr_pages
))
3387 if (vma_size
!= PAGE_SIZE
* (1 + nr_pages
))
3390 if (vma
->vm_pgoff
!= 0)
3393 WARN_ON_ONCE(event
->ctx
->parent_ctx
);
3394 mutex_lock(&event
->mmap_mutex
);
3396 if (event
->rb
->nr_pages
== nr_pages
)
3397 atomic_inc(&event
->rb
->refcount
);
3403 user_extra
= nr_pages
+ 1;
3404 user_lock_limit
= sysctl_perf_event_mlock
>> (PAGE_SHIFT
- 10);
3407 * Increase the limit linearly with more CPUs:
3409 user_lock_limit
*= num_online_cpus();
3411 user_locked
= atomic_long_read(&user
->locked_vm
) + user_extra
;
3414 if (user_locked
> user_lock_limit
)
3415 extra
= user_locked
- user_lock_limit
;
3417 lock_limit
= rlimit(RLIMIT_MEMLOCK
);
3418 lock_limit
>>= PAGE_SHIFT
;
3419 locked
= vma
->vm_mm
->pinned_vm
+ extra
;
3421 if ((locked
> lock_limit
) && perf_paranoid_tracepoint_raw() &&
3422 !capable(CAP_IPC_LOCK
)) {
3429 if (vma
->vm_flags
& VM_WRITE
)
3430 flags
|= RING_BUFFER_WRITABLE
;
3432 rb
= rb_alloc(nr_pages
,
3433 event
->attr
.watermark
? event
->attr
.wakeup_watermark
: 0,
3440 rcu_assign_pointer(event
->rb
, rb
);
3442 atomic_long_add(user_extra
, &user
->locked_vm
);
3443 event
->mmap_locked
= extra
;
3444 event
->mmap_user
= get_current_user();
3445 vma
->vm_mm
->pinned_vm
+= event
->mmap_locked
;
3449 atomic_inc(&event
->mmap_count
);
3450 mutex_unlock(&event
->mmap_mutex
);
3452 vma
->vm_flags
|= VM_RESERVED
;
3453 vma
->vm_ops
= &perf_mmap_vmops
;
3458 static int perf_fasync(int fd
, struct file
*filp
, int on
)
3460 struct inode
*inode
= filp
->f_path
.dentry
->d_inode
;
3461 struct perf_event
*event
= filp
->private_data
;
3464 mutex_lock(&inode
->i_mutex
);
3465 retval
= fasync_helper(fd
, filp
, on
, &event
->fasync
);
3466 mutex_unlock(&inode
->i_mutex
);
3474 static const struct file_operations perf_fops
= {
3475 .llseek
= no_llseek
,
3476 .release
= perf_release
,
3479 .unlocked_ioctl
= perf_ioctl
,
3480 .compat_ioctl
= perf_ioctl
,
3482 .fasync
= perf_fasync
,
3488 * If there's data, ensure we set the poll() state and publish everything
3489 * to user-space before waking everybody up.
3492 void perf_event_wakeup(struct perf_event
*event
)
3494 wake_up_all(&event
->waitq
);
3496 if (event
->pending_kill
) {
3497 kill_fasync(&event
->fasync
, SIGIO
, event
->pending_kill
);
3498 event
->pending_kill
= 0;
3502 static void perf_pending_event(struct irq_work
*entry
)
3504 struct perf_event
*event
= container_of(entry
,
3505 struct perf_event
, pending
);
3507 if (event
->pending_disable
) {
3508 event
->pending_disable
= 0;
3509 __perf_event_disable(event
);
3512 if (event
->pending_wakeup
) {
3513 event
->pending_wakeup
= 0;
3514 perf_event_wakeup(event
);
3519 * We assume there is only KVM supporting the callbacks.
3520 * Later on, we might change it to a list if there is
3521 * another virtualization implementation supporting the callbacks.
3523 struct perf_guest_info_callbacks
*perf_guest_cbs
;
3525 int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks
*cbs
)
3527 perf_guest_cbs
= cbs
;
3530 EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks
);
3532 int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks
*cbs
)
3534 perf_guest_cbs
= NULL
;
3537 EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks
);
3539 static void __perf_event_header__init_id(struct perf_event_header
*header
,
3540 struct perf_sample_data
*data
,
3541 struct perf_event
*event
)
3543 u64 sample_type
= event
->attr
.sample_type
;
3545 data
->type
= sample_type
;
3546 header
->size
+= event
->id_header_size
;
3548 if (sample_type
& PERF_SAMPLE_TID
) {
3549 /* namespace issues */
3550 data
->tid_entry
.pid
= perf_event_pid(event
, current
);
3551 data
->tid_entry
.tid
= perf_event_tid(event
, current
);
3554 if (sample_type
& PERF_SAMPLE_TIME
)
3555 data
->time
= perf_clock();
3557 if (sample_type
& PERF_SAMPLE_ID
)
3558 data
->id
= primary_event_id(event
);
3560 if (sample_type
& PERF_SAMPLE_STREAM_ID
)
3561 data
->stream_id
= event
->id
;
3563 if (sample_type
& PERF_SAMPLE_CPU
) {
3564 data
->cpu_entry
.cpu
= raw_smp_processor_id();
3565 data
->cpu_entry
.reserved
= 0;
3569 void perf_event_header__init_id(struct perf_event_header
*header
,
3570 struct perf_sample_data
*data
,
3571 struct perf_event
*event
)
3573 if (event
->attr
.sample_id_all
)
3574 __perf_event_header__init_id(header
, data
, event
);
3577 static void __perf_event__output_id_sample(struct perf_output_handle
*handle
,
3578 struct perf_sample_data
*data
)
3580 u64 sample_type
= data
->type
;
3582 if (sample_type
& PERF_SAMPLE_TID
)
3583 perf_output_put(handle
, data
->tid_entry
);
3585 if (sample_type
& PERF_SAMPLE_TIME
)
3586 perf_output_put(handle
, data
->time
);
3588 if (sample_type
& PERF_SAMPLE_ID
)
3589 perf_output_put(handle
, data
->id
);
3591 if (sample_type
& PERF_SAMPLE_STREAM_ID
)
3592 perf_output_put(handle
, data
->stream_id
);
3594 if (sample_type
& PERF_SAMPLE_CPU
)
3595 perf_output_put(handle
, data
->cpu_entry
);
3598 void perf_event__output_id_sample(struct perf_event
*event
,
3599 struct perf_output_handle
*handle
,
3600 struct perf_sample_data
*sample
)
3602 if (event
->attr
.sample_id_all
)
3603 __perf_event__output_id_sample(handle
, sample
);
3606 static void perf_output_read_one(struct perf_output_handle
*handle
,
3607 struct perf_event
*event
,
3608 u64 enabled
, u64 running
)
3610 u64 read_format
= event
->attr
.read_format
;
3614 values
[n
++] = perf_event_count(event
);
3615 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
) {
3616 values
[n
++] = enabled
+
3617 atomic64_read(&event
->child_total_time_enabled
);
3619 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
) {
3620 values
[n
++] = running
+
3621 atomic64_read(&event
->child_total_time_running
);
3623 if (read_format
& PERF_FORMAT_ID
)
3624 values
[n
++] = primary_event_id(event
);
3626 __output_copy(handle
, values
, n
* sizeof(u64
));
3630 * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
3632 static void perf_output_read_group(struct perf_output_handle
*handle
,
3633 struct perf_event
*event
,
3634 u64 enabled
, u64 running
)
3636 struct perf_event
*leader
= event
->group_leader
, *sub
;
3637 u64 read_format
= event
->attr
.read_format
;
3641 values
[n
++] = 1 + leader
->nr_siblings
;
3643 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
3644 values
[n
++] = enabled
;
3646 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
3647 values
[n
++] = running
;
3649 if (leader
!= event
)
3650 leader
->pmu
->read(leader
);
3652 values
[n
++] = perf_event_count(leader
);
3653 if (read_format
& PERF_FORMAT_ID
)
3654 values
[n
++] = primary_event_id(leader
);
3656 __output_copy(handle
, values
, n
* sizeof(u64
));
3658 list_for_each_entry(sub
, &leader
->sibling_list
, group_entry
) {
3662 sub
->pmu
->read(sub
);
3664 values
[n
++] = perf_event_count(sub
);
3665 if (read_format
& PERF_FORMAT_ID
)
3666 values
[n
++] = primary_event_id(sub
);
3668 __output_copy(handle
, values
, n
* sizeof(u64
));
3672 #define PERF_FORMAT_TOTAL_TIMES (PERF_FORMAT_TOTAL_TIME_ENABLED|\
3673 PERF_FORMAT_TOTAL_TIME_RUNNING)
3675 static void perf_output_read(struct perf_output_handle
*handle
,
3676 struct perf_event
*event
)
3678 u64 enabled
= 0, running
= 0;
3679 u64 read_format
= event
->attr
.read_format
;
3682 * compute total_time_enabled, total_time_running
3683 * based on snapshot values taken when the event
3684 * was last scheduled in.
3686 * we cannot simply called update_context_time()
3687 * because of locking issue as we are called in
3690 if (read_format
& PERF_FORMAT_TOTAL_TIMES
)
3691 calc_timer_values(event
, &enabled
, &running
);
3693 if (event
->attr
.read_format
& PERF_FORMAT_GROUP
)
3694 perf_output_read_group(handle
, event
, enabled
, running
);
3696 perf_output_read_one(handle
, event
, enabled
, running
);
3699 void perf_output_sample(struct perf_output_handle
*handle
,
3700 struct perf_event_header
*header
,
3701 struct perf_sample_data
*data
,
3702 struct perf_event
*event
)
3704 u64 sample_type
= data
->type
;
3706 perf_output_put(handle
, *header
);
3708 if (sample_type
& PERF_SAMPLE_IP
)
3709 perf_output_put(handle
, data
->ip
);
3711 if (sample_type
& PERF_SAMPLE_TID
)
3712 perf_output_put(handle
, data
->tid_entry
);
3714 if (sample_type
& PERF_SAMPLE_TIME
)
3715 perf_output_put(handle
, data
->time
);
3717 if (sample_type
& PERF_SAMPLE_ADDR
)
3718 perf_output_put(handle
, data
->addr
);
3720 if (sample_type
& PERF_SAMPLE_ID
)
3721 perf_output_put(handle
, data
->id
);
3723 if (sample_type
& PERF_SAMPLE_STREAM_ID
)
3724 perf_output_put(handle
, data
->stream_id
);
3726 if (sample_type
& PERF_SAMPLE_CPU
)
3727 perf_output_put(handle
, data
->cpu_entry
);
3729 if (sample_type
& PERF_SAMPLE_PERIOD
)
3730 perf_output_put(handle
, data
->period
);
3732 if (sample_type
& PERF_SAMPLE_READ
)
3733 perf_output_read(handle
, event
);
3735 if (sample_type
& PERF_SAMPLE_CALLCHAIN
) {
3736 if (data
->callchain
) {
3739 if (data
->callchain
)
3740 size
+= data
->callchain
->nr
;
3742 size
*= sizeof(u64
);
3744 __output_copy(handle
, data
->callchain
, size
);
3747 perf_output_put(handle
, nr
);
3751 if (sample_type
& PERF_SAMPLE_RAW
) {
3753 perf_output_put(handle
, data
->raw
->size
);
3754 __output_copy(handle
, data
->raw
->data
,
3761 .size
= sizeof(u32
),
3764 perf_output_put(handle
, raw
);
3768 if (!event
->attr
.watermark
) {
3769 int wakeup_events
= event
->attr
.wakeup_events
;
3771 if (wakeup_events
) {
3772 struct ring_buffer
*rb
= handle
->rb
;
3773 int events
= local_inc_return(&rb
->events
);
3775 if (events
>= wakeup_events
) {
3776 local_sub(wakeup_events
, &rb
->events
);
3777 local_inc(&rb
->wakeup
);
3783 void perf_prepare_sample(struct perf_event_header
*header
,
3784 struct perf_sample_data
*data
,
3785 struct perf_event
*event
,
3786 struct pt_regs
*regs
)
3788 u64 sample_type
= event
->attr
.sample_type
;
3790 header
->type
= PERF_RECORD_SAMPLE
;
3791 header
->size
= sizeof(*header
) + event
->header_size
;
3794 header
->misc
|= perf_misc_flags(regs
);
3796 __perf_event_header__init_id(header
, data
, event
);
3798 if (sample_type
& PERF_SAMPLE_IP
)
3799 data
->ip
= perf_instruction_pointer(regs
);
3801 if (sample_type
& PERF_SAMPLE_CALLCHAIN
) {
3804 data
->callchain
= perf_callchain(regs
);
3806 if (data
->callchain
)
3807 size
+= data
->callchain
->nr
;
3809 header
->size
+= size
* sizeof(u64
);
3812 if (sample_type
& PERF_SAMPLE_RAW
) {
3813 int size
= sizeof(u32
);
3816 size
+= data
->raw
->size
;
3818 size
+= sizeof(u32
);
3820 WARN_ON_ONCE(size
& (sizeof(u64
)-1));
3821 header
->size
+= size
;
3825 static void perf_event_output(struct perf_event
*event
,
3826 struct perf_sample_data
*data
,
3827 struct pt_regs
*regs
)
3829 struct perf_output_handle handle
;
3830 struct perf_event_header header
;
3832 /* protect the callchain buffers */
3835 perf_prepare_sample(&header
, data
, event
, regs
);
3837 if (perf_output_begin(&handle
, event
, header
.size
))
3840 perf_output_sample(&handle
, &header
, data
, event
);
3842 perf_output_end(&handle
);
3852 struct perf_read_event
{
3853 struct perf_event_header header
;
3860 perf_event_read_event(struct perf_event
*event
,
3861 struct task_struct
*task
)
3863 struct perf_output_handle handle
;
3864 struct perf_sample_data sample
;
3865 struct perf_read_event read_event
= {
3867 .type
= PERF_RECORD_READ
,
3869 .size
= sizeof(read_event
) + event
->read_size
,
3871 .pid
= perf_event_pid(event
, task
),
3872 .tid
= perf_event_tid(event
, task
),
3876 perf_event_header__init_id(&read_event
.header
, &sample
, event
);
3877 ret
= perf_output_begin(&handle
, event
, read_event
.header
.size
);
3881 perf_output_put(&handle
, read_event
);
3882 perf_output_read(&handle
, event
);
3883 perf_event__output_id_sample(event
, &handle
, &sample
);
3885 perf_output_end(&handle
);
3889 * task tracking -- fork/exit
3891 * enabled by: attr.comm | attr.mmap | attr.mmap_data | attr.task
3894 struct perf_task_event
{
3895 struct task_struct
*task
;
3896 struct perf_event_context
*task_ctx
;
3899 struct perf_event_header header
;
3909 static void perf_event_task_output(struct perf_event
*event
,
3910 struct perf_task_event
*task_event
)
3912 struct perf_output_handle handle
;
3913 struct perf_sample_data sample
;
3914 struct task_struct
*task
= task_event
->task
;
3915 int ret
, size
= task_event
->event_id
.header
.size
;
3917 perf_event_header__init_id(&task_event
->event_id
.header
, &sample
, event
);
3919 ret
= perf_output_begin(&handle
, event
,
3920 task_event
->event_id
.header
.size
);
3924 task_event
->event_id
.pid
= perf_event_pid(event
, task
);
3925 task_event
->event_id
.ppid
= perf_event_pid(event
, current
);
3927 task_event
->event_id
.tid
= perf_event_tid(event
, task
);
3928 task_event
->event_id
.ptid
= perf_event_tid(event
, current
);
3930 perf_output_put(&handle
, task_event
->event_id
);
3932 perf_event__output_id_sample(event
, &handle
, &sample
);
3934 perf_output_end(&handle
);
3936 task_event
->event_id
.header
.size
= size
;
3939 static int perf_event_task_match(struct perf_event
*event
)
3941 if (event
->state
< PERF_EVENT_STATE_INACTIVE
)
3944 if (!event_filter_match(event
))
3947 if (event
->attr
.comm
|| event
->attr
.mmap
||
3948 event
->attr
.mmap_data
|| event
->attr
.task
)
3954 static void perf_event_task_ctx(struct perf_event_context
*ctx
,
3955 struct perf_task_event
*task_event
)
3957 struct perf_event
*event
;
3959 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
3960 if (perf_event_task_match(event
))
3961 perf_event_task_output(event
, task_event
);
3965 static void perf_event_task_event(struct perf_task_event
*task_event
)
3967 struct perf_cpu_context
*cpuctx
;
3968 struct perf_event_context
*ctx
;
3973 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
3974 cpuctx
= get_cpu_ptr(pmu
->pmu_cpu_context
);
3975 if (cpuctx
->active_pmu
!= pmu
)
3977 perf_event_task_ctx(&cpuctx
->ctx
, task_event
);
3979 ctx
= task_event
->task_ctx
;
3981 ctxn
= pmu
->task_ctx_nr
;
3984 ctx
= rcu_dereference(current
->perf_event_ctxp
[ctxn
]);
3987 perf_event_task_ctx(ctx
, task_event
);
3989 put_cpu_ptr(pmu
->pmu_cpu_context
);
3994 static void perf_event_task(struct task_struct
*task
,
3995 struct perf_event_context
*task_ctx
,
3998 struct perf_task_event task_event
;
4000 if (!atomic_read(&nr_comm_events
) &&
4001 !atomic_read(&nr_mmap_events
) &&
4002 !atomic_read(&nr_task_events
))
4005 task_event
= (struct perf_task_event
){
4007 .task_ctx
= task_ctx
,
4010 .type
= new ? PERF_RECORD_FORK
: PERF_RECORD_EXIT
,
4012 .size
= sizeof(task_event
.event_id
),
4018 .time
= perf_clock(),
4022 perf_event_task_event(&task_event
);
4025 void perf_event_fork(struct task_struct
*task
)
4027 perf_event_task(task
, NULL
, 1);
4034 struct perf_comm_event
{
4035 struct task_struct
*task
;
4040 struct perf_event_header header
;
4047 static void perf_event_comm_output(struct perf_event
*event
,
4048 struct perf_comm_event
*comm_event
)
4050 struct perf_output_handle handle
;
4051 struct perf_sample_data sample
;
4052 int size
= comm_event
->event_id
.header
.size
;
4055 perf_event_header__init_id(&comm_event
->event_id
.header
, &sample
, event
);
4056 ret
= perf_output_begin(&handle
, event
,
4057 comm_event
->event_id
.header
.size
);
4062 comm_event
->event_id
.pid
= perf_event_pid(event
, comm_event
->task
);
4063 comm_event
->event_id
.tid
= perf_event_tid(event
, comm_event
->task
);
4065 perf_output_put(&handle
, comm_event
->event_id
);
4066 __output_copy(&handle
, comm_event
->comm
,
4067 comm_event
->comm_size
);
4069 perf_event__output_id_sample(event
, &handle
, &sample
);
4071 perf_output_end(&handle
);
4073 comm_event
->event_id
.header
.size
= size
;
4076 static int perf_event_comm_match(struct perf_event
*event
)
4078 if (event
->state
< PERF_EVENT_STATE_INACTIVE
)
4081 if (!event_filter_match(event
))
4084 if (event
->attr
.comm
)
4090 static void perf_event_comm_ctx(struct perf_event_context
*ctx
,
4091 struct perf_comm_event
*comm_event
)
4093 struct perf_event
*event
;
4095 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
4096 if (perf_event_comm_match(event
))
4097 perf_event_comm_output(event
, comm_event
);
4101 static void perf_event_comm_event(struct perf_comm_event
*comm_event
)
4103 struct perf_cpu_context
*cpuctx
;
4104 struct perf_event_context
*ctx
;
4105 char comm
[TASK_COMM_LEN
];
4110 memset(comm
, 0, sizeof(comm
));
4111 strlcpy(comm
, comm_event
->task
->comm
, sizeof(comm
));
4112 size
= ALIGN(strlen(comm
)+1, sizeof(u64
));
4114 comm_event
->comm
= comm
;
4115 comm_event
->comm_size
= size
;
4117 comm_event
->event_id
.header
.size
= sizeof(comm_event
->event_id
) + size
;
4119 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
4120 cpuctx
= get_cpu_ptr(pmu
->pmu_cpu_context
);
4121 if (cpuctx
->active_pmu
!= pmu
)
4123 perf_event_comm_ctx(&cpuctx
->ctx
, comm_event
);
4125 ctxn
= pmu
->task_ctx_nr
;
4129 ctx
= rcu_dereference(current
->perf_event_ctxp
[ctxn
]);
4131 perf_event_comm_ctx(ctx
, comm_event
);
4133 put_cpu_ptr(pmu
->pmu_cpu_context
);
4138 void perf_event_comm(struct task_struct
*task
)
4140 struct perf_comm_event comm_event
;
4141 struct perf_event_context
*ctx
;
4144 for_each_task_context_nr(ctxn
) {
4145 ctx
= task
->perf_event_ctxp
[ctxn
];
4149 perf_event_enable_on_exec(ctx
);
4152 if (!atomic_read(&nr_comm_events
))
4155 comm_event
= (struct perf_comm_event
){
4161 .type
= PERF_RECORD_COMM
,
4170 perf_event_comm_event(&comm_event
);
4177 struct perf_mmap_event
{
4178 struct vm_area_struct
*vma
;
4180 const char *file_name
;
4184 struct perf_event_header header
;
4194 static void perf_event_mmap_output(struct perf_event
*event
,
4195 struct perf_mmap_event
*mmap_event
)
4197 struct perf_output_handle handle
;
4198 struct perf_sample_data sample
;
4199 int size
= mmap_event
->event_id
.header
.size
;
4202 perf_event_header__init_id(&mmap_event
->event_id
.header
, &sample
, event
);
4203 ret
= perf_output_begin(&handle
, event
,
4204 mmap_event
->event_id
.header
.size
);
4208 mmap_event
->event_id
.pid
= perf_event_pid(event
, current
);
4209 mmap_event
->event_id
.tid
= perf_event_tid(event
, current
);
4211 perf_output_put(&handle
, mmap_event
->event_id
);
4212 __output_copy(&handle
, mmap_event
->file_name
,
4213 mmap_event
->file_size
);
4215 perf_event__output_id_sample(event
, &handle
, &sample
);
4217 perf_output_end(&handle
);
4219 mmap_event
->event_id
.header
.size
= size
;
4222 static int perf_event_mmap_match(struct perf_event
*event
,
4223 struct perf_mmap_event
*mmap_event
,
4226 if (event
->state
< PERF_EVENT_STATE_INACTIVE
)
4229 if (!event_filter_match(event
))
4232 if ((!executable
&& event
->attr
.mmap_data
) ||
4233 (executable
&& event
->attr
.mmap
))
4239 static void perf_event_mmap_ctx(struct perf_event_context
*ctx
,
4240 struct perf_mmap_event
*mmap_event
,
4243 struct perf_event
*event
;
4245 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
4246 if (perf_event_mmap_match(event
, mmap_event
, executable
))
4247 perf_event_mmap_output(event
, mmap_event
);
4251 static void perf_event_mmap_event(struct perf_mmap_event
*mmap_event
)
4253 struct perf_cpu_context
*cpuctx
;
4254 struct perf_event_context
*ctx
;
4255 struct vm_area_struct
*vma
= mmap_event
->vma
;
4256 struct file
*file
= vma
->vm_file
;
4264 memset(tmp
, 0, sizeof(tmp
));
4268 * d_path works from the end of the rb backwards, so we
4269 * need to add enough zero bytes after the string to handle
4270 * the 64bit alignment we do later.
4272 buf
= kzalloc(PATH_MAX
+ sizeof(u64
), GFP_KERNEL
);
4274 name
= strncpy(tmp
, "//enomem", sizeof(tmp
));
4277 name
= d_path(&file
->f_path
, buf
, PATH_MAX
);
4279 name
= strncpy(tmp
, "//toolong", sizeof(tmp
));
4283 if (arch_vma_name(mmap_event
->vma
)) {
4284 name
= strncpy(tmp
, arch_vma_name(mmap_event
->vma
),
4290 name
= strncpy(tmp
, "[vdso]", sizeof(tmp
));
4292 } else if (vma
->vm_start
<= vma
->vm_mm
->start_brk
&&
4293 vma
->vm_end
>= vma
->vm_mm
->brk
) {
4294 name
= strncpy(tmp
, "[heap]", sizeof(tmp
));
4296 } else if (vma
->vm_start
<= vma
->vm_mm
->start_stack
&&
4297 vma
->vm_end
>= vma
->vm_mm
->start_stack
) {
4298 name
= strncpy(tmp
, "[stack]", sizeof(tmp
));
4302 name
= strncpy(tmp
, "//anon", sizeof(tmp
));
4307 size
= ALIGN(strlen(name
)+1, sizeof(u64
));
4309 mmap_event
->file_name
= name
;
4310 mmap_event
->file_size
= size
;
4312 mmap_event
->event_id
.header
.size
= sizeof(mmap_event
->event_id
) + size
;
4315 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
4316 cpuctx
= get_cpu_ptr(pmu
->pmu_cpu_context
);
4317 if (cpuctx
->active_pmu
!= pmu
)
4319 perf_event_mmap_ctx(&cpuctx
->ctx
, mmap_event
,
4320 vma
->vm_flags
& VM_EXEC
);
4322 ctxn
= pmu
->task_ctx_nr
;
4326 ctx
= rcu_dereference(current
->perf_event_ctxp
[ctxn
]);
4328 perf_event_mmap_ctx(ctx
, mmap_event
,
4329 vma
->vm_flags
& VM_EXEC
);
4332 put_cpu_ptr(pmu
->pmu_cpu_context
);
4339 void perf_event_mmap(struct vm_area_struct
*vma
)
4341 struct perf_mmap_event mmap_event
;
4343 if (!atomic_read(&nr_mmap_events
))
4346 mmap_event
= (struct perf_mmap_event
){
4352 .type
= PERF_RECORD_MMAP
,
4353 .misc
= PERF_RECORD_MISC_USER
,
4358 .start
= vma
->vm_start
,
4359 .len
= vma
->vm_end
- vma
->vm_start
,
4360 .pgoff
= (u64
)vma
->vm_pgoff
<< PAGE_SHIFT
,
4364 perf_event_mmap_event(&mmap_event
);
4368 * IRQ throttle logging
4371 static void perf_log_throttle(struct perf_event
*event
, int enable
)
4373 struct perf_output_handle handle
;
4374 struct perf_sample_data sample
;
4378 struct perf_event_header header
;
4382 } throttle_event
= {
4384 .type
= PERF_RECORD_THROTTLE
,
4386 .size
= sizeof(throttle_event
),
4388 .time
= perf_clock(),
4389 .id
= primary_event_id(event
),
4390 .stream_id
= event
->id
,
4394 throttle_event
.header
.type
= PERF_RECORD_UNTHROTTLE
;
4396 perf_event_header__init_id(&throttle_event
.header
, &sample
, event
);
4398 ret
= perf_output_begin(&handle
, event
,
4399 throttle_event
.header
.size
);
4403 perf_output_put(&handle
, throttle_event
);
4404 perf_event__output_id_sample(event
, &handle
, &sample
);
4405 perf_output_end(&handle
);
4409 * Generic event overflow handling, sampling.
4412 static int __perf_event_overflow(struct perf_event
*event
,
4413 int throttle
, struct perf_sample_data
*data
,
4414 struct pt_regs
*regs
)
4416 int events
= atomic_read(&event
->event_limit
);
4417 struct hw_perf_event
*hwc
= &event
->hw
;
4421 * Non-sampling counters might still use the PMI to fold short
4422 * hardware counters, ignore those.
4424 if (unlikely(!is_sampling_event(event
)))
4427 if (unlikely(hwc
->interrupts
>= max_samples_per_tick
)) {
4429 hwc
->interrupts
= MAX_INTERRUPTS
;
4430 perf_log_throttle(event
, 0);
4436 if (event
->attr
.freq
) {
4437 u64 now
= perf_clock();
4438 s64 delta
= now
- hwc
->freq_time_stamp
;
4440 hwc
->freq_time_stamp
= now
;
4442 if (delta
> 0 && delta
< 2*TICK_NSEC
)
4443 perf_adjust_period(event
, delta
, hwc
->last_period
);
4447 * XXX event_limit might not quite work as expected on inherited
4451 event
->pending_kill
= POLL_IN
;
4452 if (events
&& atomic_dec_and_test(&event
->event_limit
)) {
4454 event
->pending_kill
= POLL_HUP
;
4455 event
->pending_disable
= 1;
4456 irq_work_queue(&event
->pending
);
4459 if (event
->overflow_handler
)
4460 event
->overflow_handler(event
, data
, regs
);
4462 perf_event_output(event
, data
, regs
);
4464 if (event
->fasync
&& event
->pending_kill
) {
4465 event
->pending_wakeup
= 1;
4466 irq_work_queue(&event
->pending
);
4472 int perf_event_overflow(struct perf_event
*event
,
4473 struct perf_sample_data
*data
,
4474 struct pt_regs
*regs
)
4476 return __perf_event_overflow(event
, 1, data
, regs
);
4480 * Generic software event infrastructure
4483 struct swevent_htable
{
4484 struct swevent_hlist
*swevent_hlist
;
4485 struct mutex hlist_mutex
;
4488 /* Recursion avoidance in each contexts */
4489 int recursion
[PERF_NR_CONTEXTS
];
4492 static DEFINE_PER_CPU(struct swevent_htable
, swevent_htable
);
4495 * We directly increment event->count and keep a second value in
4496 * event->hw.period_left to count intervals. This period event
4497 * is kept in the range [-sample_period, 0] so that we can use the
4501 static u64
perf_swevent_set_period(struct perf_event
*event
)
4503 struct hw_perf_event
*hwc
= &event
->hw
;
4504 u64 period
= hwc
->last_period
;
4508 hwc
->last_period
= hwc
->sample_period
;
4511 old
= val
= local64_read(&hwc
->period_left
);
4515 nr
= div64_u64(period
+ val
, period
);
4516 offset
= nr
* period
;
4518 if (local64_cmpxchg(&hwc
->period_left
, old
, val
) != old
)
4524 static void perf_swevent_overflow(struct perf_event
*event
, u64 overflow
,
4525 struct perf_sample_data
*data
,
4526 struct pt_regs
*regs
)
4528 struct hw_perf_event
*hwc
= &event
->hw
;
4531 data
->period
= event
->hw
.last_period
;
4533 overflow
= perf_swevent_set_period(event
);
4535 if (hwc
->interrupts
== MAX_INTERRUPTS
)
4538 for (; overflow
; overflow
--) {
4539 if (__perf_event_overflow(event
, throttle
,
4542 * We inhibit the overflow from happening when
4543 * hwc->interrupts == MAX_INTERRUPTS.
4551 static void perf_swevent_event(struct perf_event
*event
, u64 nr
,
4552 struct perf_sample_data
*data
,
4553 struct pt_regs
*regs
)
4555 struct hw_perf_event
*hwc
= &event
->hw
;
4557 local64_add(nr
, &event
->count
);
4562 if (!is_sampling_event(event
))
4565 if (nr
== 1 && hwc
->sample_period
== 1 && !event
->attr
.freq
)
4566 return perf_swevent_overflow(event
, 1, data
, regs
);
4568 if (local64_add_negative(nr
, &hwc
->period_left
))
4571 perf_swevent_overflow(event
, 0, data
, regs
);
4574 static int perf_exclude_event(struct perf_event
*event
,
4575 struct pt_regs
*regs
)
4577 if (event
->hw
.state
& PERF_HES_STOPPED
)
4581 if (event
->attr
.exclude_user
&& user_mode(regs
))
4584 if (event
->attr
.exclude_kernel
&& !user_mode(regs
))
4591 static int perf_swevent_match(struct perf_event
*event
,
4592 enum perf_type_id type
,
4594 struct perf_sample_data
*data
,
4595 struct pt_regs
*regs
)
4597 if (event
->attr
.type
!= type
)
4600 if (event
->attr
.config
!= event_id
)
4603 if (perf_exclude_event(event
, regs
))
4609 static inline u64
swevent_hash(u64 type
, u32 event_id
)
4611 u64 val
= event_id
| (type
<< 32);
4613 return hash_64(val
, SWEVENT_HLIST_BITS
);
4616 static inline struct hlist_head
*
4617 __find_swevent_head(struct swevent_hlist
*hlist
, u64 type
, u32 event_id
)
4619 u64 hash
= swevent_hash(type
, event_id
);
4621 return &hlist
->heads
[hash
];
4624 /* For the read side: events when they trigger */
4625 static inline struct hlist_head
*
4626 find_swevent_head_rcu(struct swevent_htable
*swhash
, u64 type
, u32 event_id
)
4628 struct swevent_hlist
*hlist
;
4630 hlist
= rcu_dereference(swhash
->swevent_hlist
);
4634 return __find_swevent_head(hlist
, type
, event_id
);
4637 /* For the event head insertion and removal in the hlist */
4638 static inline struct hlist_head
*
4639 find_swevent_head(struct swevent_htable
*swhash
, struct perf_event
*event
)
4641 struct swevent_hlist
*hlist
;
4642 u32 event_id
= event
->attr
.config
;
4643 u64 type
= event
->attr
.type
;
4646 * Event scheduling is always serialized against hlist allocation
4647 * and release. Which makes the protected version suitable here.
4648 * The context lock guarantees that.
4650 hlist
= rcu_dereference_protected(swhash
->swevent_hlist
,
4651 lockdep_is_held(&event
->ctx
->lock
));
4655 return __find_swevent_head(hlist
, type
, event_id
);
4658 static void do_perf_sw_event(enum perf_type_id type
, u32 event_id
,
4660 struct perf_sample_data
*data
,
4661 struct pt_regs
*regs
)
4663 struct swevent_htable
*swhash
= &__get_cpu_var(swevent_htable
);
4664 struct perf_event
*event
;
4665 struct hlist_node
*node
;
4666 struct hlist_head
*head
;
4669 head
= find_swevent_head_rcu(swhash
, type
, event_id
);
4673 hlist_for_each_entry_rcu(event
, node
, head
, hlist_entry
) {
4674 if (perf_swevent_match(event
, type
, event_id
, data
, regs
))
4675 perf_swevent_event(event
, nr
, data
, regs
);
4681 int perf_swevent_get_recursion_context(void)
4683 struct swevent_htable
*swhash
= &__get_cpu_var(swevent_htable
);
4685 return get_recursion_context(swhash
->recursion
);
4687 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context
);
4689 inline void perf_swevent_put_recursion_context(int rctx
)
4691 struct swevent_htable
*swhash
= &__get_cpu_var(swevent_htable
);
4693 put_recursion_context(swhash
->recursion
, rctx
);
4696 void __perf_sw_event(u32 event_id
, u64 nr
, struct pt_regs
*regs
, u64 addr
)
4698 struct perf_sample_data data
;
4701 preempt_disable_notrace();
4702 rctx
= perf_swevent_get_recursion_context();
4706 perf_sample_data_init(&data
, addr
);
4708 do_perf_sw_event(PERF_TYPE_SOFTWARE
, event_id
, nr
, &data
, regs
);
4710 perf_swevent_put_recursion_context(rctx
);
4711 preempt_enable_notrace();
4714 static void perf_swevent_read(struct perf_event
*event
)
4718 static int perf_swevent_add(struct perf_event
*event
, int flags
)
4720 struct swevent_htable
*swhash
= &__get_cpu_var(swevent_htable
);
4721 struct hw_perf_event
*hwc
= &event
->hw
;
4722 struct hlist_head
*head
;
4724 if (is_sampling_event(event
)) {
4725 hwc
->last_period
= hwc
->sample_period
;
4726 perf_swevent_set_period(event
);
4729 hwc
->state
= !(flags
& PERF_EF_START
);
4731 head
= find_swevent_head(swhash
, event
);
4732 if (WARN_ON_ONCE(!head
))
4735 hlist_add_head_rcu(&event
->hlist_entry
, head
);
4740 static void perf_swevent_del(struct perf_event
*event
, int flags
)
4742 hlist_del_rcu(&event
->hlist_entry
);
4745 static void perf_swevent_start(struct perf_event
*event
, int flags
)
4747 event
->hw
.state
= 0;
4750 static void perf_swevent_stop(struct perf_event
*event
, int flags
)
4752 event
->hw
.state
= PERF_HES_STOPPED
;
4755 /* Deref the hlist from the update side */
4756 static inline struct swevent_hlist
*
4757 swevent_hlist_deref(struct swevent_htable
*swhash
)
4759 return rcu_dereference_protected(swhash
->swevent_hlist
,
4760 lockdep_is_held(&swhash
->hlist_mutex
));
4763 static void swevent_hlist_release(struct swevent_htable
*swhash
)
4765 struct swevent_hlist
*hlist
= swevent_hlist_deref(swhash
);
4770 rcu_assign_pointer(swhash
->swevent_hlist
, NULL
);
4771 kfree_rcu(hlist
, rcu_head
);
4774 static void swevent_hlist_put_cpu(struct perf_event
*event
, int cpu
)
4776 struct swevent_htable
*swhash
= &per_cpu(swevent_htable
, cpu
);
4778 mutex_lock(&swhash
->hlist_mutex
);
4780 if (!--swhash
->hlist_refcount
)
4781 swevent_hlist_release(swhash
);
4783 mutex_unlock(&swhash
->hlist_mutex
);
4786 static void swevent_hlist_put(struct perf_event
*event
)
4790 if (event
->cpu
!= -1) {
4791 swevent_hlist_put_cpu(event
, event
->cpu
);
4795 for_each_possible_cpu(cpu
)
4796 swevent_hlist_put_cpu(event
, cpu
);
4799 static int swevent_hlist_get_cpu(struct perf_event
*event
, int cpu
)
4801 struct swevent_htable
*swhash
= &per_cpu(swevent_htable
, cpu
);
4804 mutex_lock(&swhash
->hlist_mutex
);
4806 if (!swevent_hlist_deref(swhash
) && cpu_online(cpu
)) {
4807 struct swevent_hlist
*hlist
;
4809 hlist
= kzalloc(sizeof(*hlist
), GFP_KERNEL
);
4814 rcu_assign_pointer(swhash
->swevent_hlist
, hlist
);
4816 swhash
->hlist_refcount
++;
4818 mutex_unlock(&swhash
->hlist_mutex
);
4823 static int swevent_hlist_get(struct perf_event
*event
)
4826 int cpu
, failed_cpu
;
4828 if (event
->cpu
!= -1)
4829 return swevent_hlist_get_cpu(event
, event
->cpu
);
4832 for_each_possible_cpu(cpu
) {
4833 err
= swevent_hlist_get_cpu(event
, cpu
);
4843 for_each_possible_cpu(cpu
) {
4844 if (cpu
== failed_cpu
)
4846 swevent_hlist_put_cpu(event
, cpu
);
4853 struct jump_label_key perf_swevent_enabled
[PERF_COUNT_SW_MAX
];
4855 static void sw_perf_event_destroy(struct perf_event
*event
)
4857 u64 event_id
= event
->attr
.config
;
4859 WARN_ON(event
->parent
);
4861 jump_label_dec(&perf_swevent_enabled
[event_id
]);
4862 swevent_hlist_put(event
);
4865 static int perf_swevent_init(struct perf_event
*event
)
4867 int event_id
= event
->attr
.config
;
4869 if (event
->attr
.type
!= PERF_TYPE_SOFTWARE
)
4873 case PERF_COUNT_SW_CPU_CLOCK
:
4874 case PERF_COUNT_SW_TASK_CLOCK
:
4881 if (event_id
>= PERF_COUNT_SW_MAX
)
4884 if (!event
->parent
) {
4887 err
= swevent_hlist_get(event
);
4891 jump_label_inc(&perf_swevent_enabled
[event_id
]);
4892 event
->destroy
= sw_perf_event_destroy
;
4898 static struct pmu perf_swevent
= {
4899 .task_ctx_nr
= perf_sw_context
,
4901 .event_init
= perf_swevent_init
,
4902 .add
= perf_swevent_add
,
4903 .del
= perf_swevent_del
,
4904 .start
= perf_swevent_start
,
4905 .stop
= perf_swevent_stop
,
4906 .read
= perf_swevent_read
,
4909 #ifdef CONFIG_EVENT_TRACING
4911 static int perf_tp_filter_match(struct perf_event
*event
,
4912 struct perf_sample_data
*data
)
4914 void *record
= data
->raw
->data
;
4916 if (likely(!event
->filter
) || filter_match_preds(event
->filter
, record
))
4921 static int perf_tp_event_match(struct perf_event
*event
,
4922 struct perf_sample_data
*data
,
4923 struct pt_regs
*regs
)
4925 if (event
->hw
.state
& PERF_HES_STOPPED
)
4928 * All tracepoints are from kernel-space.
4930 if (event
->attr
.exclude_kernel
)
4933 if (!perf_tp_filter_match(event
, data
))
4939 void perf_tp_event(u64 addr
, u64 count
, void *record
, int entry_size
,
4940 struct pt_regs
*regs
, struct hlist_head
*head
, int rctx
)
4942 struct perf_sample_data data
;
4943 struct perf_event
*event
;
4944 struct hlist_node
*node
;
4946 struct perf_raw_record raw
= {
4951 perf_sample_data_init(&data
, addr
);
4954 hlist_for_each_entry_rcu(event
, node
, head
, hlist_entry
) {
4955 if (perf_tp_event_match(event
, &data
, regs
))
4956 perf_swevent_event(event
, count
, &data
, regs
);
4959 perf_swevent_put_recursion_context(rctx
);
4961 EXPORT_SYMBOL_GPL(perf_tp_event
);
4963 static void tp_perf_event_destroy(struct perf_event
*event
)
4965 perf_trace_destroy(event
);
4968 static int perf_tp_event_init(struct perf_event
*event
)
4972 if (event
->attr
.type
!= PERF_TYPE_TRACEPOINT
)
4975 err
= perf_trace_init(event
);
4979 event
->destroy
= tp_perf_event_destroy
;
4984 static struct pmu perf_tracepoint
= {
4985 .task_ctx_nr
= perf_sw_context
,
4987 .event_init
= perf_tp_event_init
,
4988 .add
= perf_trace_add
,
4989 .del
= perf_trace_del
,
4990 .start
= perf_swevent_start
,
4991 .stop
= perf_swevent_stop
,
4992 .read
= perf_swevent_read
,
4995 static inline void perf_tp_register(void)
4997 perf_pmu_register(&perf_tracepoint
, "tracepoint", PERF_TYPE_TRACEPOINT
);
5000 static int perf_event_set_filter(struct perf_event
*event
, void __user
*arg
)
5005 if (event
->attr
.type
!= PERF_TYPE_TRACEPOINT
)
5008 filter_str
= strndup_user(arg
, PAGE_SIZE
);
5009 if (IS_ERR(filter_str
))
5010 return PTR_ERR(filter_str
);
5012 ret
= ftrace_profile_set_filter(event
, event
->attr
.config
, filter_str
);
5018 static void perf_event_free_filter(struct perf_event
*event
)
5020 ftrace_profile_free_filter(event
);
5025 static inline void perf_tp_register(void)
5029 static int perf_event_set_filter(struct perf_event
*event
, void __user
*arg
)
5034 static void perf_event_free_filter(struct perf_event
*event
)
5038 #endif /* CONFIG_EVENT_TRACING */
5040 #ifdef CONFIG_HAVE_HW_BREAKPOINT
5041 void perf_bp_event(struct perf_event
*bp
, void *data
)
5043 struct perf_sample_data sample
;
5044 struct pt_regs
*regs
= data
;
5046 perf_sample_data_init(&sample
, bp
->attr
.bp_addr
);
5048 if (!bp
->hw
.state
&& !perf_exclude_event(bp
, regs
))
5049 perf_swevent_event(bp
, 1, &sample
, regs
);
5054 * hrtimer based swevent callback
5057 static enum hrtimer_restart
perf_swevent_hrtimer(struct hrtimer
*hrtimer
)
5059 enum hrtimer_restart ret
= HRTIMER_RESTART
;
5060 struct perf_sample_data data
;
5061 struct pt_regs
*regs
;
5062 struct perf_event
*event
;
5065 event
= container_of(hrtimer
, struct perf_event
, hw
.hrtimer
);
5067 if (event
->state
!= PERF_EVENT_STATE_ACTIVE
)
5068 return HRTIMER_NORESTART
;
5070 event
->pmu
->read(event
);
5072 perf_sample_data_init(&data
, 0);
5073 data
.period
= event
->hw
.last_period
;
5074 regs
= get_irq_regs();
5076 if (regs
&& !perf_exclude_event(event
, regs
)) {
5077 if (!(event
->attr
.exclude_idle
&& current
->pid
== 0))
5078 if (perf_event_overflow(event
, &data
, regs
))
5079 ret
= HRTIMER_NORESTART
;
5082 period
= max_t(u64
, 10000, event
->hw
.sample_period
);
5083 hrtimer_forward_now(hrtimer
, ns_to_ktime(period
));
5088 static void perf_swevent_start_hrtimer(struct perf_event
*event
)
5090 struct hw_perf_event
*hwc
= &event
->hw
;
5093 if (!is_sampling_event(event
))
5096 period
= local64_read(&hwc
->period_left
);
5101 local64_set(&hwc
->period_left
, 0);
5103 period
= max_t(u64
, 10000, hwc
->sample_period
);
5105 __hrtimer_start_range_ns(&hwc
->hrtimer
,
5106 ns_to_ktime(period
), 0,
5107 HRTIMER_MODE_REL_PINNED
, 0);
5110 static void perf_swevent_cancel_hrtimer(struct perf_event
*event
)
5112 struct hw_perf_event
*hwc
= &event
->hw
;
5114 if (is_sampling_event(event
)) {
5115 ktime_t remaining
= hrtimer_get_remaining(&hwc
->hrtimer
);
5116 local64_set(&hwc
->period_left
, ktime_to_ns(remaining
));
5118 hrtimer_cancel(&hwc
->hrtimer
);
5122 static void perf_swevent_init_hrtimer(struct perf_event
*event
)
5124 struct hw_perf_event
*hwc
= &event
->hw
;
5126 if (!is_sampling_event(event
))
5129 hrtimer_init(&hwc
->hrtimer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
5130 hwc
->hrtimer
.function
= perf_swevent_hrtimer
;
5133 * Since hrtimers have a fixed rate, we can do a static freq->period
5134 * mapping and avoid the whole period adjust feedback stuff.
5136 if (event
->attr
.freq
) {
5137 long freq
= event
->attr
.sample_freq
;
5139 event
->attr
.sample_period
= NSEC_PER_SEC
/ freq
;
5140 hwc
->sample_period
= event
->attr
.sample_period
;
5141 local64_set(&hwc
->period_left
, hwc
->sample_period
);
5142 event
->attr
.freq
= 0;
5147 * Software event: cpu wall time clock
5150 static void cpu_clock_event_update(struct perf_event
*event
)
5155 now
= local_clock();
5156 prev
= local64_xchg(&event
->hw
.prev_count
, now
);
5157 local64_add(now
- prev
, &event
->count
);
5160 static void cpu_clock_event_start(struct perf_event
*event
, int flags
)
5162 local64_set(&event
->hw
.prev_count
, local_clock());
5163 perf_swevent_start_hrtimer(event
);
5166 static void cpu_clock_event_stop(struct perf_event
*event
, int flags
)
5168 perf_swevent_cancel_hrtimer(event
);
5169 cpu_clock_event_update(event
);
5172 static int cpu_clock_event_add(struct perf_event
*event
, int flags
)
5174 if (flags
& PERF_EF_START
)
5175 cpu_clock_event_start(event
, flags
);
5180 static void cpu_clock_event_del(struct perf_event
*event
, int flags
)
5182 cpu_clock_event_stop(event
, flags
);
5185 static void cpu_clock_event_read(struct perf_event
*event
)
5187 cpu_clock_event_update(event
);
5190 static int cpu_clock_event_init(struct perf_event
*event
)
5192 if (event
->attr
.type
!= PERF_TYPE_SOFTWARE
)
5195 if (event
->attr
.config
!= PERF_COUNT_SW_CPU_CLOCK
)
5198 perf_swevent_init_hrtimer(event
);
5203 static struct pmu perf_cpu_clock
= {
5204 .task_ctx_nr
= perf_sw_context
,
5206 .event_init
= cpu_clock_event_init
,
5207 .add
= cpu_clock_event_add
,
5208 .del
= cpu_clock_event_del
,
5209 .start
= cpu_clock_event_start
,
5210 .stop
= cpu_clock_event_stop
,
5211 .read
= cpu_clock_event_read
,
5215 * Software event: task time clock
5218 static void task_clock_event_update(struct perf_event
*event
, u64 now
)
5223 prev
= local64_xchg(&event
->hw
.prev_count
, now
);
5225 local64_add(delta
, &event
->count
);
5228 static void task_clock_event_start(struct perf_event
*event
, int flags
)
5230 local64_set(&event
->hw
.prev_count
, event
->ctx
->time
);
5231 perf_swevent_start_hrtimer(event
);
5234 static void task_clock_event_stop(struct perf_event
*event
, int flags
)
5236 perf_swevent_cancel_hrtimer(event
);
5237 task_clock_event_update(event
, event
->ctx
->time
);
5240 static int task_clock_event_add(struct perf_event
*event
, int flags
)
5242 if (flags
& PERF_EF_START
)
5243 task_clock_event_start(event
, flags
);
5248 static void task_clock_event_del(struct perf_event
*event
, int flags
)
5250 task_clock_event_stop(event
, PERF_EF_UPDATE
);
5253 static void task_clock_event_read(struct perf_event
*event
)
5255 u64 now
= perf_clock();
5256 u64 delta
= now
- event
->ctx
->timestamp
;
5257 u64 time
= event
->ctx
->time
+ delta
;
5259 task_clock_event_update(event
, time
);
5262 static int task_clock_event_init(struct perf_event
*event
)
5264 if (event
->attr
.type
!= PERF_TYPE_SOFTWARE
)
5267 if (event
->attr
.config
!= PERF_COUNT_SW_TASK_CLOCK
)
5270 perf_swevent_init_hrtimer(event
);
5275 static struct pmu perf_task_clock
= {
5276 .task_ctx_nr
= perf_sw_context
,
5278 .event_init
= task_clock_event_init
,
5279 .add
= task_clock_event_add
,
5280 .del
= task_clock_event_del
,
5281 .start
= task_clock_event_start
,
5282 .stop
= task_clock_event_stop
,
5283 .read
= task_clock_event_read
,
5286 static void perf_pmu_nop_void(struct pmu
*pmu
)
5290 static int perf_pmu_nop_int(struct pmu
*pmu
)
5295 static void perf_pmu_start_txn(struct pmu
*pmu
)
5297 perf_pmu_disable(pmu
);
5300 static int perf_pmu_commit_txn(struct pmu
*pmu
)
5302 perf_pmu_enable(pmu
);
5306 static void perf_pmu_cancel_txn(struct pmu
*pmu
)
5308 perf_pmu_enable(pmu
);
5312 * Ensures all contexts with the same task_ctx_nr have the same
5313 * pmu_cpu_context too.
5315 static void *find_pmu_context(int ctxn
)
5322 list_for_each_entry(pmu
, &pmus
, entry
) {
5323 if (pmu
->task_ctx_nr
== ctxn
)
5324 return pmu
->pmu_cpu_context
;
5330 static void update_pmu_context(struct pmu
*pmu
, struct pmu
*old_pmu
)
5334 for_each_possible_cpu(cpu
) {
5335 struct perf_cpu_context
*cpuctx
;
5337 cpuctx
= per_cpu_ptr(pmu
->pmu_cpu_context
, cpu
);
5339 if (cpuctx
->active_pmu
== old_pmu
)
5340 cpuctx
->active_pmu
= pmu
;
5344 static void free_pmu_context(struct pmu
*pmu
)
5348 mutex_lock(&pmus_lock
);
5350 * Like a real lame refcount.
5352 list_for_each_entry(i
, &pmus
, entry
) {
5353 if (i
->pmu_cpu_context
== pmu
->pmu_cpu_context
) {
5354 update_pmu_context(i
, pmu
);
5359 free_percpu(pmu
->pmu_cpu_context
);
5361 mutex_unlock(&pmus_lock
);
5363 static struct idr pmu_idr
;
5366 type_show(struct device
*dev
, struct device_attribute
*attr
, char *page
)
5368 struct pmu
*pmu
= dev_get_drvdata(dev
);
5370 return snprintf(page
, PAGE_SIZE
-1, "%d\n", pmu
->type
);
5373 static struct device_attribute pmu_dev_attrs
[] = {
5378 static int pmu_bus_running
;
5379 static struct bus_type pmu_bus
= {
5380 .name
= "event_source",
5381 .dev_attrs
= pmu_dev_attrs
,
5384 static void pmu_dev_release(struct device
*dev
)
5389 static int pmu_dev_alloc(struct pmu
*pmu
)
5393 pmu
->dev
= kzalloc(sizeof(struct device
), GFP_KERNEL
);
5397 device_initialize(pmu
->dev
);
5398 ret
= dev_set_name(pmu
->dev
, "%s", pmu
->name
);
5402 dev_set_drvdata(pmu
->dev
, pmu
);
5403 pmu
->dev
->bus
= &pmu_bus
;
5404 pmu
->dev
->release
= pmu_dev_release
;
5405 ret
= device_add(pmu
->dev
);
5413 put_device(pmu
->dev
);
5417 static struct lock_class_key cpuctx_mutex
;
5418 static struct lock_class_key cpuctx_lock
;
5420 int perf_pmu_register(struct pmu
*pmu
, char *name
, int type
)
5424 mutex_lock(&pmus_lock
);
5426 pmu
->pmu_disable_count
= alloc_percpu(int);
5427 if (!pmu
->pmu_disable_count
)
5436 int err
= idr_pre_get(&pmu_idr
, GFP_KERNEL
);
5440 err
= idr_get_new_above(&pmu_idr
, pmu
, PERF_TYPE_MAX
, &type
);
5448 if (pmu_bus_running
) {
5449 ret
= pmu_dev_alloc(pmu
);
5455 pmu
->pmu_cpu_context
= find_pmu_context(pmu
->task_ctx_nr
);
5456 if (pmu
->pmu_cpu_context
)
5457 goto got_cpu_context
;
5459 pmu
->pmu_cpu_context
= alloc_percpu(struct perf_cpu_context
);
5460 if (!pmu
->pmu_cpu_context
)
5463 for_each_possible_cpu(cpu
) {
5464 struct perf_cpu_context
*cpuctx
;
5466 cpuctx
= per_cpu_ptr(pmu
->pmu_cpu_context
, cpu
);
5467 __perf_event_init_context(&cpuctx
->ctx
);
5468 lockdep_set_class(&cpuctx
->ctx
.mutex
, &cpuctx_mutex
);
5469 lockdep_set_class(&cpuctx
->ctx
.lock
, &cpuctx_lock
);
5470 cpuctx
->ctx
.type
= cpu_context
;
5471 cpuctx
->ctx
.pmu
= pmu
;
5472 cpuctx
->jiffies_interval
= 1;
5473 INIT_LIST_HEAD(&cpuctx
->rotation_list
);
5474 cpuctx
->active_pmu
= pmu
;
5478 if (!pmu
->start_txn
) {
5479 if (pmu
->pmu_enable
) {
5481 * If we have pmu_enable/pmu_disable calls, install
5482 * transaction stubs that use that to try and batch
5483 * hardware accesses.
5485 pmu
->start_txn
= perf_pmu_start_txn
;
5486 pmu
->commit_txn
= perf_pmu_commit_txn
;
5487 pmu
->cancel_txn
= perf_pmu_cancel_txn
;
5489 pmu
->start_txn
= perf_pmu_nop_void
;
5490 pmu
->commit_txn
= perf_pmu_nop_int
;
5491 pmu
->cancel_txn
= perf_pmu_nop_void
;
5495 if (!pmu
->pmu_enable
) {
5496 pmu
->pmu_enable
= perf_pmu_nop_void
;
5497 pmu
->pmu_disable
= perf_pmu_nop_void
;
5500 list_add_rcu(&pmu
->entry
, &pmus
);
5503 mutex_unlock(&pmus_lock
);
5508 device_del(pmu
->dev
);
5509 put_device(pmu
->dev
);
5512 if (pmu
->type
>= PERF_TYPE_MAX
)
5513 idr_remove(&pmu_idr
, pmu
->type
);
5516 free_percpu(pmu
->pmu_disable_count
);
5520 void perf_pmu_unregister(struct pmu
*pmu
)
5522 mutex_lock(&pmus_lock
);
5523 list_del_rcu(&pmu
->entry
);
5524 mutex_unlock(&pmus_lock
);
5527 * We dereference the pmu list under both SRCU and regular RCU, so
5528 * synchronize against both of those.
5530 synchronize_srcu(&pmus_srcu
);
5533 free_percpu(pmu
->pmu_disable_count
);
5534 if (pmu
->type
>= PERF_TYPE_MAX
)
5535 idr_remove(&pmu_idr
, pmu
->type
);
5536 device_del(pmu
->dev
);
5537 put_device(pmu
->dev
);
5538 free_pmu_context(pmu
);
5541 struct pmu
*perf_init_event(struct perf_event
*event
)
5543 struct pmu
*pmu
= NULL
;
5547 idx
= srcu_read_lock(&pmus_srcu
);
5550 pmu
= idr_find(&pmu_idr
, event
->attr
.type
);
5554 ret
= pmu
->event_init(event
);
5560 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
5562 ret
= pmu
->event_init(event
);
5566 if (ret
!= -ENOENT
) {
5571 pmu
= ERR_PTR(-ENOENT
);
5573 srcu_read_unlock(&pmus_srcu
, idx
);
5579 * Allocate and initialize a event structure
5581 static struct perf_event
*
5582 perf_event_alloc(struct perf_event_attr
*attr
, int cpu
,
5583 struct task_struct
*task
,
5584 struct perf_event
*group_leader
,
5585 struct perf_event
*parent_event
,
5586 perf_overflow_handler_t overflow_handler
,
5590 struct perf_event
*event
;
5591 struct hw_perf_event
*hwc
;
5594 if ((unsigned)cpu
>= nr_cpu_ids
) {
5595 if (!task
|| cpu
!= -1)
5596 return ERR_PTR(-EINVAL
);
5599 event
= kzalloc(sizeof(*event
), GFP_KERNEL
);
5601 return ERR_PTR(-ENOMEM
);
5604 * Single events are their own group leaders, with an
5605 * empty sibling list:
5608 group_leader
= event
;
5610 mutex_init(&event
->child_mutex
);
5611 INIT_LIST_HEAD(&event
->child_list
);
5613 INIT_LIST_HEAD(&event
->group_entry
);
5614 INIT_LIST_HEAD(&event
->event_entry
);
5615 INIT_LIST_HEAD(&event
->sibling_list
);
5616 init_waitqueue_head(&event
->waitq
);
5617 init_irq_work(&event
->pending
, perf_pending_event
);
5619 mutex_init(&event
->mmap_mutex
);
5622 event
->attr
= *attr
;
5623 event
->group_leader
= group_leader
;
5627 event
->parent
= parent_event
;
5629 event
->ns
= get_pid_ns(current
->nsproxy
->pid_ns
);
5630 event
->id
= atomic64_inc_return(&perf_event_id
);
5632 event
->state
= PERF_EVENT_STATE_INACTIVE
;
5635 event
->attach_state
= PERF_ATTACH_TASK
;
5636 #ifdef CONFIG_HAVE_HW_BREAKPOINT
5638 * hw_breakpoint is a bit difficult here..
5640 if (attr
->type
== PERF_TYPE_BREAKPOINT
)
5641 event
->hw
.bp_target
= task
;
5645 if (!overflow_handler
&& parent_event
) {
5646 overflow_handler
= parent_event
->overflow_handler
;
5647 context
= parent_event
->overflow_handler_context
;
5650 event
->overflow_handler
= overflow_handler
;
5651 event
->overflow_handler_context
= context
;
5654 event
->state
= PERF_EVENT_STATE_OFF
;
5659 hwc
->sample_period
= attr
->sample_period
;
5660 if (attr
->freq
&& attr
->sample_freq
)
5661 hwc
->sample_period
= 1;
5662 hwc
->last_period
= hwc
->sample_period
;
5664 local64_set(&hwc
->period_left
, hwc
->sample_period
);
5667 * we currently do not support PERF_FORMAT_GROUP on inherited events
5669 if (attr
->inherit
&& (attr
->read_format
& PERF_FORMAT_GROUP
))
5672 pmu
= perf_init_event(event
);
5678 else if (IS_ERR(pmu
))
5683 put_pid_ns(event
->ns
);
5685 return ERR_PTR(err
);
5688 if (!event
->parent
) {
5689 if (event
->attach_state
& PERF_ATTACH_TASK
)
5690 jump_label_inc(&perf_sched_events
);
5691 if (event
->attr
.mmap
|| event
->attr
.mmap_data
)
5692 atomic_inc(&nr_mmap_events
);
5693 if (event
->attr
.comm
)
5694 atomic_inc(&nr_comm_events
);
5695 if (event
->attr
.task
)
5696 atomic_inc(&nr_task_events
);
5697 if (event
->attr
.sample_type
& PERF_SAMPLE_CALLCHAIN
) {
5698 err
= get_callchain_buffers();
5701 return ERR_PTR(err
);
5709 static int perf_copy_attr(struct perf_event_attr __user
*uattr
,
5710 struct perf_event_attr
*attr
)
5715 if (!access_ok(VERIFY_WRITE
, uattr
, PERF_ATTR_SIZE_VER0
))
5719 * zero the full structure, so that a short copy will be nice.
5721 memset(attr
, 0, sizeof(*attr
));
5723 ret
= get_user(size
, &uattr
->size
);
5727 if (size
> PAGE_SIZE
) /* silly large */
5730 if (!size
) /* abi compat */
5731 size
= PERF_ATTR_SIZE_VER0
;
5733 if (size
< PERF_ATTR_SIZE_VER0
)
5737 * If we're handed a bigger struct than we know of,
5738 * ensure all the unknown bits are 0 - i.e. new
5739 * user-space does not rely on any kernel feature
5740 * extensions we dont know about yet.
5742 if (size
> sizeof(*attr
)) {
5743 unsigned char __user
*addr
;
5744 unsigned char __user
*end
;
5747 addr
= (void __user
*)uattr
+ sizeof(*attr
);
5748 end
= (void __user
*)uattr
+ size
;
5750 for (; addr
< end
; addr
++) {
5751 ret
= get_user(val
, addr
);
5757 size
= sizeof(*attr
);
5760 ret
= copy_from_user(attr
, uattr
, size
);
5764 if (attr
->__reserved_1
)
5767 if (attr
->sample_type
& ~(PERF_SAMPLE_MAX
-1))
5770 if (attr
->read_format
& ~(PERF_FORMAT_MAX
-1))
5777 put_user(sizeof(*attr
), &uattr
->size
);
5783 perf_event_set_output(struct perf_event
*event
, struct perf_event
*output_event
)
5785 struct ring_buffer
*rb
= NULL
, *old_rb
= NULL
;
5791 /* don't allow circular references */
5792 if (event
== output_event
)
5796 * Don't allow cross-cpu buffers
5798 if (output_event
->cpu
!= event
->cpu
)
5802 * If its not a per-cpu rb, it must be the same task.
5804 if (output_event
->cpu
== -1 && output_event
->ctx
!= event
->ctx
)
5808 mutex_lock(&event
->mmap_mutex
);
5809 /* Can't redirect output if we've got an active mmap() */
5810 if (atomic_read(&event
->mmap_count
))
5814 /* get the rb we want to redirect to */
5815 rb
= ring_buffer_get(output_event
);
5821 rcu_assign_pointer(event
->rb
, rb
);
5824 mutex_unlock(&event
->mmap_mutex
);
5827 ring_buffer_put(old_rb
);
5833 * sys_perf_event_open - open a performance event, associate it to a task/cpu
5835 * @attr_uptr: event_id type attributes for monitoring/sampling
5838 * @group_fd: group leader event fd
5840 SYSCALL_DEFINE5(perf_event_open
,
5841 struct perf_event_attr __user
*, attr_uptr
,
5842 pid_t
, pid
, int, cpu
, int, group_fd
, unsigned long, flags
)
5844 struct perf_event
*group_leader
= NULL
, *output_event
= NULL
;
5845 struct perf_event
*event
, *sibling
;
5846 struct perf_event_attr attr
;
5847 struct perf_event_context
*ctx
;
5848 struct file
*event_file
= NULL
;
5849 struct file
*group_file
= NULL
;
5850 struct task_struct
*task
= NULL
;
5854 int fput_needed
= 0;
5857 /* for future expandability... */
5858 if (flags
& ~PERF_FLAG_ALL
)
5861 err
= perf_copy_attr(attr_uptr
, &attr
);
5865 if (!attr
.exclude_kernel
) {
5866 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN
))
5871 if (attr
.sample_freq
> sysctl_perf_event_sample_rate
)
5876 * In cgroup mode, the pid argument is used to pass the fd
5877 * opened to the cgroup directory in cgroupfs. The cpu argument
5878 * designates the cpu on which to monitor threads from that
5881 if ((flags
& PERF_FLAG_PID_CGROUP
) && (pid
== -1 || cpu
== -1))
5884 event_fd
= get_unused_fd_flags(O_RDWR
);
5888 if (group_fd
!= -1) {
5889 group_leader
= perf_fget_light(group_fd
, &fput_needed
);
5890 if (IS_ERR(group_leader
)) {
5891 err
= PTR_ERR(group_leader
);
5894 group_file
= group_leader
->filp
;
5895 if (flags
& PERF_FLAG_FD_OUTPUT
)
5896 output_event
= group_leader
;
5897 if (flags
& PERF_FLAG_FD_NO_GROUP
)
5898 group_leader
= NULL
;
5901 if (pid
!= -1 && !(flags
& PERF_FLAG_PID_CGROUP
)) {
5902 task
= find_lively_task_by_vpid(pid
);
5904 err
= PTR_ERR(task
);
5909 event
= perf_event_alloc(&attr
, cpu
, task
, group_leader
, NULL
,
5911 if (IS_ERR(event
)) {
5912 err
= PTR_ERR(event
);
5916 if (flags
& PERF_FLAG_PID_CGROUP
) {
5917 err
= perf_cgroup_connect(pid
, event
, &attr
, group_leader
);
5922 * - that has cgroup constraint on event->cpu
5923 * - that may need work on context switch
5925 atomic_inc(&per_cpu(perf_cgroup_events
, event
->cpu
));
5926 jump_label_inc(&perf_sched_events
);
5930 * Special case software events and allow them to be part of
5931 * any hardware group.
5936 (is_software_event(event
) != is_software_event(group_leader
))) {
5937 if (is_software_event(event
)) {
5939 * If event and group_leader are not both a software
5940 * event, and event is, then group leader is not.
5942 * Allow the addition of software events to !software
5943 * groups, this is safe because software events never
5946 pmu
= group_leader
->pmu
;
5947 } else if (is_software_event(group_leader
) &&
5948 (group_leader
->group_flags
& PERF_GROUP_SOFTWARE
)) {
5950 * In case the group is a pure software group, and we
5951 * try to add a hardware event, move the whole group to
5952 * the hardware context.
5959 * Get the target context (task or percpu):
5961 ctx
= find_get_context(pmu
, task
, cpu
);
5968 put_task_struct(task
);
5973 * Look up the group leader (we will attach this event to it):
5979 * Do not allow a recursive hierarchy (this new sibling
5980 * becoming part of another group-sibling):
5982 if (group_leader
->group_leader
!= group_leader
)
5985 * Do not allow to attach to a group in a different
5986 * task or CPU context:
5989 if (group_leader
->ctx
->type
!= ctx
->type
)
5992 if (group_leader
->ctx
!= ctx
)
5997 * Only a group leader can be exclusive or pinned
5999 if (attr
.exclusive
|| attr
.pinned
)
6004 err
= perf_event_set_output(event
, output_event
);
6009 event_file
= anon_inode_getfile("[perf_event]", &perf_fops
, event
, O_RDWR
);
6010 if (IS_ERR(event_file
)) {
6011 err
= PTR_ERR(event_file
);
6016 struct perf_event_context
*gctx
= group_leader
->ctx
;
6018 mutex_lock(&gctx
->mutex
);
6019 perf_remove_from_context(group_leader
);
6020 list_for_each_entry(sibling
, &group_leader
->sibling_list
,
6022 perf_remove_from_context(sibling
);
6025 mutex_unlock(&gctx
->mutex
);
6029 event
->filp
= event_file
;
6030 WARN_ON_ONCE(ctx
->parent_ctx
);
6031 mutex_lock(&ctx
->mutex
);
6034 perf_install_in_context(ctx
, group_leader
, cpu
);
6036 list_for_each_entry(sibling
, &group_leader
->sibling_list
,
6038 perf_install_in_context(ctx
, sibling
, cpu
);
6043 perf_install_in_context(ctx
, event
, cpu
);
6045 perf_unpin_context(ctx
);
6046 mutex_unlock(&ctx
->mutex
);
6048 event
->owner
= current
;
6050 mutex_lock(¤t
->perf_event_mutex
);
6051 list_add_tail(&event
->owner_entry
, ¤t
->perf_event_list
);
6052 mutex_unlock(¤t
->perf_event_mutex
);
6055 * Precalculate sample_data sizes
6057 perf_event__header_size(event
);
6058 perf_event__id_header_size(event
);
6061 * Drop the reference on the group_event after placing the
6062 * new event on the sibling_list. This ensures destruction
6063 * of the group leader will find the pointer to itself in
6064 * perf_group_detach().
6066 fput_light(group_file
, fput_needed
);
6067 fd_install(event_fd
, event_file
);
6071 perf_unpin_context(ctx
);
6077 put_task_struct(task
);
6079 fput_light(group_file
, fput_needed
);
6081 put_unused_fd(event_fd
);
6086 * perf_event_create_kernel_counter
6088 * @attr: attributes of the counter to create
6089 * @cpu: cpu in which the counter is bound
6090 * @task: task to profile (NULL for percpu)
6093 perf_event_create_kernel_counter(struct perf_event_attr
*attr
, int cpu
,
6094 struct task_struct
*task
,
6095 perf_overflow_handler_t overflow_handler
,
6098 struct perf_event_context
*ctx
;
6099 struct perf_event
*event
;
6103 * Get the target context (task or percpu):
6106 event
= perf_event_alloc(attr
, cpu
, task
, NULL
, NULL
,
6107 overflow_handler
, context
);
6108 if (IS_ERR(event
)) {
6109 err
= PTR_ERR(event
);
6113 ctx
= find_get_context(event
->pmu
, task
, cpu
);
6120 WARN_ON_ONCE(ctx
->parent_ctx
);
6121 mutex_lock(&ctx
->mutex
);
6122 perf_install_in_context(ctx
, event
, cpu
);
6124 perf_unpin_context(ctx
);
6125 mutex_unlock(&ctx
->mutex
);
6132 return ERR_PTR(err
);
6134 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter
);
6136 static void sync_child_event(struct perf_event
*child_event
,
6137 struct task_struct
*child
)
6139 struct perf_event
*parent_event
= child_event
->parent
;
6142 if (child_event
->attr
.inherit_stat
)
6143 perf_event_read_event(child_event
, child
);
6145 child_val
= perf_event_count(child_event
);
6148 * Add back the child's count to the parent's count:
6150 atomic64_add(child_val
, &parent_event
->child_count
);
6151 atomic64_add(child_event
->total_time_enabled
,
6152 &parent_event
->child_total_time_enabled
);
6153 atomic64_add(child_event
->total_time_running
,
6154 &parent_event
->child_total_time_running
);
6157 * Remove this event from the parent's list
6159 WARN_ON_ONCE(parent_event
->ctx
->parent_ctx
);
6160 mutex_lock(&parent_event
->child_mutex
);
6161 list_del_init(&child_event
->child_list
);
6162 mutex_unlock(&parent_event
->child_mutex
);
6165 * Release the parent event, if this was the last
6168 fput(parent_event
->filp
);
6172 __perf_event_exit_task(struct perf_event
*child_event
,
6173 struct perf_event_context
*child_ctx
,
6174 struct task_struct
*child
)
6176 if (child_event
->parent
) {
6177 raw_spin_lock_irq(&child_ctx
->lock
);
6178 perf_group_detach(child_event
);
6179 raw_spin_unlock_irq(&child_ctx
->lock
);
6182 perf_remove_from_context(child_event
);
6185 * It can happen that the parent exits first, and has events
6186 * that are still around due to the child reference. These
6187 * events need to be zapped.
6189 if (child_event
->parent
) {
6190 sync_child_event(child_event
, child
);
6191 free_event(child_event
);
6195 static void perf_event_exit_task_context(struct task_struct
*child
, int ctxn
)
6197 struct perf_event
*child_event
, *tmp
;
6198 struct perf_event_context
*child_ctx
;
6199 unsigned long flags
;
6201 if (likely(!child
->perf_event_ctxp
[ctxn
])) {
6202 perf_event_task(child
, NULL
, 0);
6206 local_irq_save(flags
);
6208 * We can't reschedule here because interrupts are disabled,
6209 * and either child is current or it is a task that can't be
6210 * scheduled, so we are now safe from rescheduling changing
6213 child_ctx
= rcu_dereference_raw(child
->perf_event_ctxp
[ctxn
]);
6216 * Take the context lock here so that if find_get_context is
6217 * reading child->perf_event_ctxp, we wait until it has
6218 * incremented the context's refcount before we do put_ctx below.
6220 raw_spin_lock(&child_ctx
->lock
);
6221 task_ctx_sched_out(child_ctx
);
6222 child
->perf_event_ctxp
[ctxn
] = NULL
;
6224 * If this context is a clone; unclone it so it can't get
6225 * swapped to another process while we're removing all
6226 * the events from it.
6228 unclone_ctx(child_ctx
);
6229 update_context_time(child_ctx
);
6230 raw_spin_unlock_irqrestore(&child_ctx
->lock
, flags
);
6233 * Report the task dead after unscheduling the events so that we
6234 * won't get any samples after PERF_RECORD_EXIT. We can however still
6235 * get a few PERF_RECORD_READ events.
6237 perf_event_task(child
, child_ctx
, 0);
6240 * We can recurse on the same lock type through:
6242 * __perf_event_exit_task()
6243 * sync_child_event()
6244 * fput(parent_event->filp)
6246 * mutex_lock(&ctx->mutex)
6248 * But since its the parent context it won't be the same instance.
6250 mutex_lock(&child_ctx
->mutex
);
6253 list_for_each_entry_safe(child_event
, tmp
, &child_ctx
->pinned_groups
,
6255 __perf_event_exit_task(child_event
, child_ctx
, child
);
6257 list_for_each_entry_safe(child_event
, tmp
, &child_ctx
->flexible_groups
,
6259 __perf_event_exit_task(child_event
, child_ctx
, child
);
6262 * If the last event was a group event, it will have appended all
6263 * its siblings to the list, but we obtained 'tmp' before that which
6264 * will still point to the list head terminating the iteration.
6266 if (!list_empty(&child_ctx
->pinned_groups
) ||
6267 !list_empty(&child_ctx
->flexible_groups
))
6270 mutex_unlock(&child_ctx
->mutex
);
6276 * When a child task exits, feed back event values to parent events.
6278 void perf_event_exit_task(struct task_struct
*child
)
6280 struct perf_event
*event
, *tmp
;
6283 mutex_lock(&child
->perf_event_mutex
);
6284 list_for_each_entry_safe(event
, tmp
, &child
->perf_event_list
,
6286 list_del_init(&event
->owner_entry
);
6289 * Ensure the list deletion is visible before we clear
6290 * the owner, closes a race against perf_release() where
6291 * we need to serialize on the owner->perf_event_mutex.
6294 event
->owner
= NULL
;
6296 mutex_unlock(&child
->perf_event_mutex
);
6298 for_each_task_context_nr(ctxn
)
6299 perf_event_exit_task_context(child
, ctxn
);
6302 static void perf_free_event(struct perf_event
*event
,
6303 struct perf_event_context
*ctx
)
6305 struct perf_event
*parent
= event
->parent
;
6307 if (WARN_ON_ONCE(!parent
))
6310 mutex_lock(&parent
->child_mutex
);
6311 list_del_init(&event
->child_list
);
6312 mutex_unlock(&parent
->child_mutex
);
6316 perf_group_detach(event
);
6317 list_del_event(event
, ctx
);
6322 * free an unexposed, unused context as created by inheritance by
6323 * perf_event_init_task below, used by fork() in case of fail.
6325 void perf_event_free_task(struct task_struct
*task
)
6327 struct perf_event_context
*ctx
;
6328 struct perf_event
*event
, *tmp
;
6331 for_each_task_context_nr(ctxn
) {
6332 ctx
= task
->perf_event_ctxp
[ctxn
];
6336 mutex_lock(&ctx
->mutex
);
6338 list_for_each_entry_safe(event
, tmp
, &ctx
->pinned_groups
,
6340 perf_free_event(event
, ctx
);
6342 list_for_each_entry_safe(event
, tmp
, &ctx
->flexible_groups
,
6344 perf_free_event(event
, ctx
);
6346 if (!list_empty(&ctx
->pinned_groups
) ||
6347 !list_empty(&ctx
->flexible_groups
))
6350 mutex_unlock(&ctx
->mutex
);
6356 void perf_event_delayed_put(struct task_struct
*task
)
6360 for_each_task_context_nr(ctxn
)
6361 WARN_ON_ONCE(task
->perf_event_ctxp
[ctxn
]);
6365 * inherit a event from parent task to child task:
6367 static struct perf_event
*
6368 inherit_event(struct perf_event
*parent_event
,
6369 struct task_struct
*parent
,
6370 struct perf_event_context
*parent_ctx
,
6371 struct task_struct
*child
,
6372 struct perf_event
*group_leader
,
6373 struct perf_event_context
*child_ctx
)
6375 struct perf_event
*child_event
;
6376 unsigned long flags
;
6379 * Instead of creating recursive hierarchies of events,
6380 * we link inherited events back to the original parent,
6381 * which has a filp for sure, which we use as the reference
6384 if (parent_event
->parent
)
6385 parent_event
= parent_event
->parent
;
6387 child_event
= perf_event_alloc(&parent_event
->attr
,
6390 group_leader
, parent_event
,
6392 if (IS_ERR(child_event
))
6397 * Make the child state follow the state of the parent event,
6398 * not its attr.disabled bit. We hold the parent's mutex,
6399 * so we won't race with perf_event_{en, dis}able_family.
6401 if (parent_event
->state
>= PERF_EVENT_STATE_INACTIVE
)
6402 child_event
->state
= PERF_EVENT_STATE_INACTIVE
;
6404 child_event
->state
= PERF_EVENT_STATE_OFF
;
6406 if (parent_event
->attr
.freq
) {
6407 u64 sample_period
= parent_event
->hw
.sample_period
;
6408 struct hw_perf_event
*hwc
= &child_event
->hw
;
6410 hwc
->sample_period
= sample_period
;
6411 hwc
->last_period
= sample_period
;
6413 local64_set(&hwc
->period_left
, sample_period
);
6416 child_event
->ctx
= child_ctx
;
6417 child_event
->overflow_handler
= parent_event
->overflow_handler
;
6418 child_event
->overflow_handler_context
6419 = parent_event
->overflow_handler_context
;
6422 * Precalculate sample_data sizes
6424 perf_event__header_size(child_event
);
6425 perf_event__id_header_size(child_event
);
6428 * Link it up in the child's context:
6430 raw_spin_lock_irqsave(&child_ctx
->lock
, flags
);
6431 add_event_to_ctx(child_event
, child_ctx
);
6432 raw_spin_unlock_irqrestore(&child_ctx
->lock
, flags
);
6435 * Get a reference to the parent filp - we will fput it
6436 * when the child event exits. This is safe to do because
6437 * we are in the parent and we know that the filp still
6438 * exists and has a nonzero count:
6440 atomic_long_inc(&parent_event
->filp
->f_count
);
6443 * Link this into the parent event's child list
6445 WARN_ON_ONCE(parent_event
->ctx
->parent_ctx
);
6446 mutex_lock(&parent_event
->child_mutex
);
6447 list_add_tail(&child_event
->child_list
, &parent_event
->child_list
);
6448 mutex_unlock(&parent_event
->child_mutex
);
6453 static int inherit_group(struct perf_event
*parent_event
,
6454 struct task_struct
*parent
,
6455 struct perf_event_context
*parent_ctx
,
6456 struct task_struct
*child
,
6457 struct perf_event_context
*child_ctx
)
6459 struct perf_event
*leader
;
6460 struct perf_event
*sub
;
6461 struct perf_event
*child_ctr
;
6463 leader
= inherit_event(parent_event
, parent
, parent_ctx
,
6464 child
, NULL
, child_ctx
);
6466 return PTR_ERR(leader
);
6467 list_for_each_entry(sub
, &parent_event
->sibling_list
, group_entry
) {
6468 child_ctr
= inherit_event(sub
, parent
, parent_ctx
,
6469 child
, leader
, child_ctx
);
6470 if (IS_ERR(child_ctr
))
6471 return PTR_ERR(child_ctr
);
6477 inherit_task_group(struct perf_event
*event
, struct task_struct
*parent
,
6478 struct perf_event_context
*parent_ctx
,
6479 struct task_struct
*child
, int ctxn
,
6483 struct perf_event_context
*child_ctx
;
6485 if (!event
->attr
.inherit
) {
6490 child_ctx
= child
->perf_event_ctxp
[ctxn
];
6493 * This is executed from the parent task context, so
6494 * inherit events that have been marked for cloning.
6495 * First allocate and initialize a context for the
6499 child_ctx
= alloc_perf_context(event
->pmu
, child
);
6503 child
->perf_event_ctxp
[ctxn
] = child_ctx
;
6506 ret
= inherit_group(event
, parent
, parent_ctx
,
6516 * Initialize the perf_event context in task_struct
6518 int perf_event_init_context(struct task_struct
*child
, int ctxn
)
6520 struct perf_event_context
*child_ctx
, *parent_ctx
;
6521 struct perf_event_context
*cloned_ctx
;
6522 struct perf_event
*event
;
6523 struct task_struct
*parent
= current
;
6524 int inherited_all
= 1;
6525 unsigned long flags
;
6528 if (likely(!parent
->perf_event_ctxp
[ctxn
]))
6532 * If the parent's context is a clone, pin it so it won't get
6535 parent_ctx
= perf_pin_task_context(parent
, ctxn
);
6538 * No need to check if parent_ctx != NULL here; since we saw
6539 * it non-NULL earlier, the only reason for it to become NULL
6540 * is if we exit, and since we're currently in the middle of
6541 * a fork we can't be exiting at the same time.
6545 * Lock the parent list. No need to lock the child - not PID
6546 * hashed yet and not running, so nobody can access it.
6548 mutex_lock(&parent_ctx
->mutex
);
6551 * We dont have to disable NMIs - we are only looking at
6552 * the list, not manipulating it:
6554 list_for_each_entry(event
, &parent_ctx
->pinned_groups
, group_entry
) {
6555 ret
= inherit_task_group(event
, parent
, parent_ctx
,
6556 child
, ctxn
, &inherited_all
);
6562 * We can't hold ctx->lock when iterating the ->flexible_group list due
6563 * to allocations, but we need to prevent rotation because
6564 * rotate_ctx() will change the list from interrupt context.
6566 raw_spin_lock_irqsave(&parent_ctx
->lock
, flags
);
6567 parent_ctx
->rotate_disable
= 1;
6568 raw_spin_unlock_irqrestore(&parent_ctx
->lock
, flags
);
6570 list_for_each_entry(event
, &parent_ctx
->flexible_groups
, group_entry
) {
6571 ret
= inherit_task_group(event
, parent
, parent_ctx
,
6572 child
, ctxn
, &inherited_all
);
6577 raw_spin_lock_irqsave(&parent_ctx
->lock
, flags
);
6578 parent_ctx
->rotate_disable
= 0;
6580 child_ctx
= child
->perf_event_ctxp
[ctxn
];
6582 if (child_ctx
&& inherited_all
) {
6584 * Mark the child context as a clone of the parent
6585 * context, or of whatever the parent is a clone of.
6587 * Note that if the parent is a clone, the holding of
6588 * parent_ctx->lock avoids it from being uncloned.
6590 cloned_ctx
= parent_ctx
->parent_ctx
;
6592 child_ctx
->parent_ctx
= cloned_ctx
;
6593 child_ctx
->parent_gen
= parent_ctx
->parent_gen
;
6595 child_ctx
->parent_ctx
= parent_ctx
;
6596 child_ctx
->parent_gen
= parent_ctx
->generation
;
6598 get_ctx(child_ctx
->parent_ctx
);
6601 raw_spin_unlock_irqrestore(&parent_ctx
->lock
, flags
);
6602 mutex_unlock(&parent_ctx
->mutex
);
6604 perf_unpin_context(parent_ctx
);
6605 put_ctx(parent_ctx
);
6611 * Initialize the perf_event context in task_struct
6613 int perf_event_init_task(struct task_struct
*child
)
6617 memset(child
->perf_event_ctxp
, 0, sizeof(child
->perf_event_ctxp
));
6618 mutex_init(&child
->perf_event_mutex
);
6619 INIT_LIST_HEAD(&child
->perf_event_list
);
6621 for_each_task_context_nr(ctxn
) {
6622 ret
= perf_event_init_context(child
, ctxn
);
6630 static void __init
perf_event_init_all_cpus(void)
6632 struct swevent_htable
*swhash
;
6635 for_each_possible_cpu(cpu
) {
6636 swhash
= &per_cpu(swevent_htable
, cpu
);
6637 mutex_init(&swhash
->hlist_mutex
);
6638 INIT_LIST_HEAD(&per_cpu(rotation_list
, cpu
));
6642 static void __cpuinit
perf_event_init_cpu(int cpu
)
6644 struct swevent_htable
*swhash
= &per_cpu(swevent_htable
, cpu
);
6646 mutex_lock(&swhash
->hlist_mutex
);
6647 if (swhash
->hlist_refcount
> 0) {
6648 struct swevent_hlist
*hlist
;
6650 hlist
= kzalloc_node(sizeof(*hlist
), GFP_KERNEL
, cpu_to_node(cpu
));
6652 rcu_assign_pointer(swhash
->swevent_hlist
, hlist
);
6654 mutex_unlock(&swhash
->hlist_mutex
);
6657 #if defined CONFIG_HOTPLUG_CPU || defined CONFIG_KEXEC
6658 static void perf_pmu_rotate_stop(struct pmu
*pmu
)
6660 struct perf_cpu_context
*cpuctx
= this_cpu_ptr(pmu
->pmu_cpu_context
);
6662 WARN_ON(!irqs_disabled());
6664 list_del_init(&cpuctx
->rotation_list
);
6667 static void __perf_event_exit_context(void *__info
)
6669 struct perf_event_context
*ctx
= __info
;
6670 struct perf_event
*event
, *tmp
;
6672 perf_pmu_rotate_stop(ctx
->pmu
);
6674 list_for_each_entry_safe(event
, tmp
, &ctx
->pinned_groups
, group_entry
)
6675 __perf_remove_from_context(event
);
6676 list_for_each_entry_safe(event
, tmp
, &ctx
->flexible_groups
, group_entry
)
6677 __perf_remove_from_context(event
);
6680 static void perf_event_exit_cpu_context(int cpu
)
6682 struct perf_event_context
*ctx
;
6686 idx
= srcu_read_lock(&pmus_srcu
);
6687 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
6688 ctx
= &per_cpu_ptr(pmu
->pmu_cpu_context
, cpu
)->ctx
;
6690 mutex_lock(&ctx
->mutex
);
6691 smp_call_function_single(cpu
, __perf_event_exit_context
, ctx
, 1);
6692 mutex_unlock(&ctx
->mutex
);
6694 srcu_read_unlock(&pmus_srcu
, idx
);
6697 static void perf_event_exit_cpu(int cpu
)
6699 struct swevent_htable
*swhash
= &per_cpu(swevent_htable
, cpu
);
6701 mutex_lock(&swhash
->hlist_mutex
);
6702 swevent_hlist_release(swhash
);
6703 mutex_unlock(&swhash
->hlist_mutex
);
6705 perf_event_exit_cpu_context(cpu
);
6708 static inline void perf_event_exit_cpu(int cpu
) { }
6712 perf_reboot(struct notifier_block
*notifier
, unsigned long val
, void *v
)
6716 for_each_online_cpu(cpu
)
6717 perf_event_exit_cpu(cpu
);
6723 * Run the perf reboot notifier at the very last possible moment so that
6724 * the generic watchdog code runs as long as possible.
6726 static struct notifier_block perf_reboot_notifier
= {
6727 .notifier_call
= perf_reboot
,
6728 .priority
= INT_MIN
,
6731 static int __cpuinit
6732 perf_cpu_notify(struct notifier_block
*self
, unsigned long action
, void *hcpu
)
6734 unsigned int cpu
= (long)hcpu
;
6736 switch (action
& ~CPU_TASKS_FROZEN
) {
6738 case CPU_UP_PREPARE
:
6739 case CPU_DOWN_FAILED
:
6740 perf_event_init_cpu(cpu
);
6743 case CPU_UP_CANCELED
:
6744 case CPU_DOWN_PREPARE
:
6745 perf_event_exit_cpu(cpu
);
6755 void __init
perf_event_init(void)
6761 perf_event_init_all_cpus();
6762 init_srcu_struct(&pmus_srcu
);
6763 perf_pmu_register(&perf_swevent
, "software", PERF_TYPE_SOFTWARE
);
6764 perf_pmu_register(&perf_cpu_clock
, NULL
, -1);
6765 perf_pmu_register(&perf_task_clock
, NULL
, -1);
6767 perf_cpu_notifier(perf_cpu_notify
);
6768 register_reboot_notifier(&perf_reboot_notifier
);
6770 ret
= init_hw_breakpoint();
6771 WARN(ret
, "hw_breakpoint initialization failed with: %d", ret
);
6774 static int __init
perf_event_sysfs_init(void)
6779 mutex_lock(&pmus_lock
);
6781 ret
= bus_register(&pmu_bus
);
6785 list_for_each_entry(pmu
, &pmus
, entry
) {
6786 if (!pmu
->name
|| pmu
->type
< 0)
6789 ret
= pmu_dev_alloc(pmu
);
6790 WARN(ret
, "Failed to register pmu: %s, reason %d\n", pmu
->name
, ret
);
6792 pmu_bus_running
= 1;
6796 mutex_unlock(&pmus_lock
);
6800 device_initcall(perf_event_sysfs_init
);
6802 #ifdef CONFIG_CGROUP_PERF
6803 static struct cgroup_subsys_state
*perf_cgroup_create(
6804 struct cgroup_subsys
*ss
, struct cgroup
*cont
)
6806 struct perf_cgroup
*jc
;
6808 jc
= kzalloc(sizeof(*jc
), GFP_KERNEL
);
6810 return ERR_PTR(-ENOMEM
);
6812 jc
->info
= alloc_percpu(struct perf_cgroup_info
);
6815 return ERR_PTR(-ENOMEM
);
6821 static void perf_cgroup_destroy(struct cgroup_subsys
*ss
,
6822 struct cgroup
*cont
)
6824 struct perf_cgroup
*jc
;
6825 jc
= container_of(cgroup_subsys_state(cont
, perf_subsys_id
),
6826 struct perf_cgroup
, css
);
6827 free_percpu(jc
->info
);
6831 static int __perf_cgroup_move(void *info
)
6833 struct task_struct
*task
= info
;
6834 perf_cgroup_switch(task
, PERF_CGROUP_SWOUT
| PERF_CGROUP_SWIN
);
6839 perf_cgroup_attach_task(struct cgroup
*cgrp
, struct task_struct
*task
)
6841 task_function_call(task
, __perf_cgroup_move
, task
);
6844 static void perf_cgroup_exit(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
6845 struct cgroup
*old_cgrp
, struct task_struct
*task
)
6848 * cgroup_exit() is called in the copy_process() failure path.
6849 * Ignore this case since the task hasn't ran yet, this avoids
6850 * trying to poke a half freed task state from generic code.
6852 if (!(task
->flags
& PF_EXITING
))
6855 perf_cgroup_attach_task(cgrp
, task
);
6858 struct cgroup_subsys perf_subsys
= {
6859 .name
= "perf_event",
6860 .subsys_id
= perf_subsys_id
,
6861 .create
= perf_cgroup_create
,
6862 .destroy
= perf_cgroup_destroy
,
6863 .exit
= perf_cgroup_exit
,
6864 .attach_task
= perf_cgroup_attach_task
,
6866 #endif /* CONFIG_CGROUP_PERF */