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/tick.h>
22 #include <linux/sysfs.h>
23 #include <linux/dcache.h>
24 #include <linux/percpu.h>
25 #include <linux/ptrace.h>
26 #include <linux/reboot.h>
27 #include <linux/vmstat.h>
28 #include <linux/device.h>
29 #include <linux/export.h>
30 #include <linux/vmalloc.h>
31 #include <linux/hardirq.h>
32 #include <linux/rculist.h>
33 #include <linux/uaccess.h>
34 #include <linux/syscalls.h>
35 #include <linux/anon_inodes.h>
36 #include <linux/kernel_stat.h>
37 #include <linux/perf_event.h>
38 #include <linux/ftrace_event.h>
39 #include <linux/hw_breakpoint.h>
40 #include <linux/mm_types.h>
41 #include <linux/cgroup.h>
42 #include <linux/compat.h>
46 #include <asm/irq_regs.h>
48 struct remote_function_call
{
49 struct task_struct
*p
;
50 int (*func
)(void *info
);
55 static void remote_function(void *data
)
57 struct remote_function_call
*tfc
= data
;
58 struct task_struct
*p
= tfc
->p
;
62 if (task_cpu(p
) != smp_processor_id() || !task_curr(p
))
66 tfc
->ret
= tfc
->func(tfc
->info
);
70 * task_function_call - call a function on the cpu on which a task runs
71 * @p: the task to evaluate
72 * @func: the function to be called
73 * @info: the function call argument
75 * Calls the function @func when the task is currently running. This might
76 * be on the current CPU, which just calls the function directly
78 * returns: @func return value, or
79 * -ESRCH - when the process isn't running
80 * -EAGAIN - when the process moved away
83 task_function_call(struct task_struct
*p
, int (*func
) (void *info
), void *info
)
85 struct remote_function_call data
= {
89 .ret
= -ESRCH
, /* No such (running) process */
93 smp_call_function_single(task_cpu(p
), remote_function
, &data
, 1);
99 * cpu_function_call - call a function on the cpu
100 * @func: the function to be called
101 * @info: the function call argument
103 * Calls the function @func on the remote cpu.
105 * returns: @func return value or -ENXIO when the cpu is offline
107 static int cpu_function_call(int cpu
, int (*func
) (void *info
), void *info
)
109 struct remote_function_call data
= {
113 .ret
= -ENXIO
, /* No such CPU */
116 smp_call_function_single(cpu
, remote_function
, &data
, 1);
121 #define PERF_FLAG_ALL (PERF_FLAG_FD_NO_GROUP |\
122 PERF_FLAG_FD_OUTPUT |\
123 PERF_FLAG_PID_CGROUP)
126 * branch priv levels that need permission checks
128 #define PERF_SAMPLE_BRANCH_PERM_PLM \
129 (PERF_SAMPLE_BRANCH_KERNEL |\
130 PERF_SAMPLE_BRANCH_HV)
133 EVENT_FLEXIBLE
= 0x1,
135 EVENT_ALL
= EVENT_FLEXIBLE
| EVENT_PINNED
,
139 * perf_sched_events : >0 events exist
140 * perf_cgroup_events: >0 per-cpu cgroup events exist on this cpu
142 struct static_key_deferred perf_sched_events __read_mostly
;
143 static DEFINE_PER_CPU(atomic_t
, perf_cgroup_events
);
144 static DEFINE_PER_CPU(atomic_t
, perf_branch_stack_events
);
146 static atomic_t nr_mmap_events __read_mostly
;
147 static atomic_t nr_comm_events __read_mostly
;
148 static atomic_t nr_task_events __read_mostly
;
150 static LIST_HEAD(pmus
);
151 static DEFINE_MUTEX(pmus_lock
);
152 static struct srcu_struct pmus_srcu
;
155 * perf event paranoia level:
156 * -1 - not paranoid at all
157 * 0 - disallow raw tracepoint access for unpriv
158 * 1 - disallow cpu events for unpriv
159 * 2 - disallow kernel profiling for unpriv
161 int sysctl_perf_event_paranoid __read_mostly
= 1;
163 /* Minimum for 512 kiB + 1 user control page */
164 int sysctl_perf_event_mlock __read_mostly
= 512 + (PAGE_SIZE
/ 1024); /* 'free' kiB per user */
167 * max perf event sample rate
169 #define DEFAULT_MAX_SAMPLE_RATE 100000
170 #define DEFAULT_SAMPLE_PERIOD_NS (NSEC_PER_SEC / DEFAULT_MAX_SAMPLE_RATE)
171 #define DEFAULT_CPU_TIME_MAX_PERCENT 25
173 int sysctl_perf_event_sample_rate __read_mostly
= DEFAULT_MAX_SAMPLE_RATE
;
175 static int max_samples_per_tick __read_mostly
= DIV_ROUND_UP(DEFAULT_MAX_SAMPLE_RATE
, HZ
);
176 static int perf_sample_period_ns __read_mostly
= DEFAULT_SAMPLE_PERIOD_NS
;
178 static atomic_t perf_sample_allowed_ns __read_mostly
=
179 ATOMIC_INIT( DEFAULT_SAMPLE_PERIOD_NS
* DEFAULT_CPU_TIME_MAX_PERCENT
/ 100);
181 void update_perf_cpu_limits(void)
183 u64 tmp
= perf_sample_period_ns
;
185 tmp
*= sysctl_perf_cpu_time_max_percent
;
187 atomic_set(&perf_sample_allowed_ns
, tmp
);
190 int perf_proc_update_handler(struct ctl_table
*table
, int write
,
191 void __user
*buffer
, size_t *lenp
,
194 int ret
= proc_dointvec_minmax(table
, write
, buffer
, lenp
, ppos
);
199 max_samples_per_tick
= DIV_ROUND_UP(sysctl_perf_event_sample_rate
, HZ
);
200 perf_sample_period_ns
= NSEC_PER_SEC
/ sysctl_perf_event_sample_rate
;
201 update_perf_cpu_limits();
206 int sysctl_perf_cpu_time_max_percent __read_mostly
= DEFAULT_CPU_TIME_MAX_PERCENT
;
208 int perf_cpu_time_max_percent_handler(struct ctl_table
*table
, int write
,
209 void __user
*buffer
, size_t *lenp
,
212 int ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
217 update_perf_cpu_limits();
223 * perf samples are done in some very critical code paths (NMIs).
224 * If they take too much CPU time, the system can lock up and not
225 * get any real work done. This will drop the sample rate when
226 * we detect that events are taking too long.
228 #define NR_ACCUMULATED_SAMPLES 128
229 DEFINE_PER_CPU(u64
, running_sample_length
);
231 void perf_sample_event_took(u64 sample_len_ns
)
233 u64 avg_local_sample_len
;
234 u64 local_samples_len
;
236 if (atomic_read(&perf_sample_allowed_ns
) == 0)
239 /* decay the counter by 1 average sample */
240 local_samples_len
= __get_cpu_var(running_sample_length
);
241 local_samples_len
-= local_samples_len
/NR_ACCUMULATED_SAMPLES
;
242 local_samples_len
+= sample_len_ns
;
243 __get_cpu_var(running_sample_length
) = local_samples_len
;
246 * note: this will be biased artifically low until we have
247 * seen NR_ACCUMULATED_SAMPLES. Doing it this way keeps us
248 * from having to maintain a count.
250 avg_local_sample_len
= local_samples_len
/NR_ACCUMULATED_SAMPLES
;
252 if (avg_local_sample_len
<= atomic_read(&perf_sample_allowed_ns
))
255 if (max_samples_per_tick
<= 1)
258 max_samples_per_tick
= DIV_ROUND_UP(max_samples_per_tick
, 2);
259 sysctl_perf_event_sample_rate
= max_samples_per_tick
* HZ
;
260 perf_sample_period_ns
= NSEC_PER_SEC
/ sysctl_perf_event_sample_rate
;
262 printk_ratelimited(KERN_WARNING
263 "perf samples too long (%lld > %d), lowering "
264 "kernel.perf_event_max_sample_rate to %d\n",
265 avg_local_sample_len
,
266 atomic_read(&perf_sample_allowed_ns
),
267 sysctl_perf_event_sample_rate
);
269 update_perf_cpu_limits();
272 static atomic64_t perf_event_id
;
274 static void cpu_ctx_sched_out(struct perf_cpu_context
*cpuctx
,
275 enum event_type_t event_type
);
277 static void cpu_ctx_sched_in(struct perf_cpu_context
*cpuctx
,
278 enum event_type_t event_type
,
279 struct task_struct
*task
);
281 static void update_context_time(struct perf_event_context
*ctx
);
282 static u64
perf_event_time(struct perf_event
*event
);
284 void __weak
perf_event_print_debug(void) { }
286 extern __weak
const char *perf_pmu_name(void)
291 static inline u64
perf_clock(void)
293 return local_clock();
296 static inline struct perf_cpu_context
*
297 __get_cpu_context(struct perf_event_context
*ctx
)
299 return this_cpu_ptr(ctx
->pmu
->pmu_cpu_context
);
302 static void perf_ctx_lock(struct perf_cpu_context
*cpuctx
,
303 struct perf_event_context
*ctx
)
305 raw_spin_lock(&cpuctx
->ctx
.lock
);
307 raw_spin_lock(&ctx
->lock
);
310 static void perf_ctx_unlock(struct perf_cpu_context
*cpuctx
,
311 struct perf_event_context
*ctx
)
314 raw_spin_unlock(&ctx
->lock
);
315 raw_spin_unlock(&cpuctx
->ctx
.lock
);
318 #ifdef CONFIG_CGROUP_PERF
321 * perf_cgroup_info keeps track of time_enabled for a cgroup.
322 * This is a per-cpu dynamically allocated data structure.
324 struct perf_cgroup_info
{
330 struct cgroup_subsys_state css
;
331 struct perf_cgroup_info __percpu
*info
;
335 * Must ensure cgroup is pinned (css_get) before calling
336 * this function. In other words, we cannot call this function
337 * if there is no cgroup event for the current CPU context.
339 static inline struct perf_cgroup
*
340 perf_cgroup_from_task(struct task_struct
*task
)
342 return container_of(task_subsys_state(task
, perf_subsys_id
),
343 struct perf_cgroup
, css
);
347 perf_cgroup_match(struct perf_event
*event
)
349 struct perf_event_context
*ctx
= event
->ctx
;
350 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
352 /* @event doesn't care about cgroup */
356 /* wants specific cgroup scope but @cpuctx isn't associated with any */
361 * Cgroup scoping is recursive. An event enabled for a cgroup is
362 * also enabled for all its descendant cgroups. If @cpuctx's
363 * cgroup is a descendant of @event's (the test covers identity
364 * case), it's a match.
366 return cgroup_is_descendant(cpuctx
->cgrp
->css
.cgroup
,
367 event
->cgrp
->css
.cgroup
);
370 static inline bool perf_tryget_cgroup(struct perf_event
*event
)
372 return css_tryget(&event
->cgrp
->css
);
375 static inline void perf_put_cgroup(struct perf_event
*event
)
377 css_put(&event
->cgrp
->css
);
380 static inline void perf_detach_cgroup(struct perf_event
*event
)
382 perf_put_cgroup(event
);
386 static inline int is_cgroup_event(struct perf_event
*event
)
388 return event
->cgrp
!= NULL
;
391 static inline u64
perf_cgroup_event_time(struct perf_event
*event
)
393 struct perf_cgroup_info
*t
;
395 t
= per_cpu_ptr(event
->cgrp
->info
, event
->cpu
);
399 static inline void __update_cgrp_time(struct perf_cgroup
*cgrp
)
401 struct perf_cgroup_info
*info
;
406 info
= this_cpu_ptr(cgrp
->info
);
408 info
->time
+= now
- info
->timestamp
;
409 info
->timestamp
= now
;
412 static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context
*cpuctx
)
414 struct perf_cgroup
*cgrp_out
= cpuctx
->cgrp
;
416 __update_cgrp_time(cgrp_out
);
419 static inline void update_cgrp_time_from_event(struct perf_event
*event
)
421 struct perf_cgroup
*cgrp
;
424 * ensure we access cgroup data only when needed and
425 * when we know the cgroup is pinned (css_get)
427 if (!is_cgroup_event(event
))
430 cgrp
= perf_cgroup_from_task(current
);
432 * Do not update time when cgroup is not active
434 if (cgrp
== event
->cgrp
)
435 __update_cgrp_time(event
->cgrp
);
439 perf_cgroup_set_timestamp(struct task_struct
*task
,
440 struct perf_event_context
*ctx
)
442 struct perf_cgroup
*cgrp
;
443 struct perf_cgroup_info
*info
;
446 * ctx->lock held by caller
447 * ensure we do not access cgroup data
448 * unless we have the cgroup pinned (css_get)
450 if (!task
|| !ctx
->nr_cgroups
)
453 cgrp
= perf_cgroup_from_task(task
);
454 info
= this_cpu_ptr(cgrp
->info
);
455 info
->timestamp
= ctx
->timestamp
;
458 #define PERF_CGROUP_SWOUT 0x1 /* cgroup switch out every event */
459 #define PERF_CGROUP_SWIN 0x2 /* cgroup switch in events based on task */
462 * reschedule events based on the cgroup constraint of task.
464 * mode SWOUT : schedule out everything
465 * mode SWIN : schedule in based on cgroup for next
467 void perf_cgroup_switch(struct task_struct
*task
, int mode
)
469 struct perf_cpu_context
*cpuctx
;
474 * disable interrupts to avoid geting nr_cgroup
475 * changes via __perf_event_disable(). Also
478 local_irq_save(flags
);
481 * we reschedule only in the presence of cgroup
482 * constrained events.
486 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
487 cpuctx
= this_cpu_ptr(pmu
->pmu_cpu_context
);
488 if (cpuctx
->unique_pmu
!= pmu
)
489 continue; /* ensure we process each cpuctx once */
492 * perf_cgroup_events says at least one
493 * context on this CPU has cgroup events.
495 * ctx->nr_cgroups reports the number of cgroup
496 * events for a context.
498 if (cpuctx
->ctx
.nr_cgroups
> 0) {
499 perf_ctx_lock(cpuctx
, cpuctx
->task_ctx
);
500 perf_pmu_disable(cpuctx
->ctx
.pmu
);
502 if (mode
& PERF_CGROUP_SWOUT
) {
503 cpu_ctx_sched_out(cpuctx
, EVENT_ALL
);
505 * must not be done before ctxswout due
506 * to event_filter_match() in event_sched_out()
511 if (mode
& PERF_CGROUP_SWIN
) {
512 WARN_ON_ONCE(cpuctx
->cgrp
);
514 * set cgrp before ctxsw in to allow
515 * event_filter_match() to not have to pass
518 cpuctx
->cgrp
= perf_cgroup_from_task(task
);
519 cpu_ctx_sched_in(cpuctx
, EVENT_ALL
, task
);
521 perf_pmu_enable(cpuctx
->ctx
.pmu
);
522 perf_ctx_unlock(cpuctx
, cpuctx
->task_ctx
);
528 local_irq_restore(flags
);
531 static inline void perf_cgroup_sched_out(struct task_struct
*task
,
532 struct task_struct
*next
)
534 struct perf_cgroup
*cgrp1
;
535 struct perf_cgroup
*cgrp2
= NULL
;
538 * we come here when we know perf_cgroup_events > 0
540 cgrp1
= perf_cgroup_from_task(task
);
543 * next is NULL when called from perf_event_enable_on_exec()
544 * that will systematically cause a cgroup_switch()
547 cgrp2
= perf_cgroup_from_task(next
);
550 * only schedule out current cgroup events if we know
551 * that we are switching to a different cgroup. Otherwise,
552 * do no touch the cgroup events.
555 perf_cgroup_switch(task
, PERF_CGROUP_SWOUT
);
558 static inline void perf_cgroup_sched_in(struct task_struct
*prev
,
559 struct task_struct
*task
)
561 struct perf_cgroup
*cgrp1
;
562 struct perf_cgroup
*cgrp2
= NULL
;
565 * we come here when we know perf_cgroup_events > 0
567 cgrp1
= perf_cgroup_from_task(task
);
569 /* prev can never be NULL */
570 cgrp2
= perf_cgroup_from_task(prev
);
573 * only need to schedule in cgroup events if we are changing
574 * cgroup during ctxsw. Cgroup events were not scheduled
575 * out of ctxsw out if that was not the case.
578 perf_cgroup_switch(task
, PERF_CGROUP_SWIN
);
581 static inline int perf_cgroup_connect(int fd
, struct perf_event
*event
,
582 struct perf_event_attr
*attr
,
583 struct perf_event
*group_leader
)
585 struct perf_cgroup
*cgrp
;
586 struct cgroup_subsys_state
*css
;
587 struct fd f
= fdget(fd
);
593 css
= cgroup_css_from_dir(f
.file
, perf_subsys_id
);
599 cgrp
= container_of(css
, struct perf_cgroup
, css
);
602 /* must be done before we fput() the file */
603 if (!perf_tryget_cgroup(event
)) {
610 * all events in a group must monitor
611 * the same cgroup because a task belongs
612 * to only one perf cgroup at a time
614 if (group_leader
&& group_leader
->cgrp
!= cgrp
) {
615 perf_detach_cgroup(event
);
624 perf_cgroup_set_shadow_time(struct perf_event
*event
, u64 now
)
626 struct perf_cgroup_info
*t
;
627 t
= per_cpu_ptr(event
->cgrp
->info
, event
->cpu
);
628 event
->shadow_ctx_time
= now
- t
->timestamp
;
632 perf_cgroup_defer_enabled(struct perf_event
*event
)
635 * when the current task's perf cgroup does not match
636 * the event's, we need to remember to call the
637 * perf_mark_enable() function the first time a task with
638 * a matching perf cgroup is scheduled in.
640 if (is_cgroup_event(event
) && !perf_cgroup_match(event
))
641 event
->cgrp_defer_enabled
= 1;
645 perf_cgroup_mark_enabled(struct perf_event
*event
,
646 struct perf_event_context
*ctx
)
648 struct perf_event
*sub
;
649 u64 tstamp
= perf_event_time(event
);
651 if (!event
->cgrp_defer_enabled
)
654 event
->cgrp_defer_enabled
= 0;
656 event
->tstamp_enabled
= tstamp
- event
->total_time_enabled
;
657 list_for_each_entry(sub
, &event
->sibling_list
, group_entry
) {
658 if (sub
->state
>= PERF_EVENT_STATE_INACTIVE
) {
659 sub
->tstamp_enabled
= tstamp
- sub
->total_time_enabled
;
660 sub
->cgrp_defer_enabled
= 0;
664 #else /* !CONFIG_CGROUP_PERF */
667 perf_cgroup_match(struct perf_event
*event
)
672 static inline void perf_detach_cgroup(struct perf_event
*event
)
675 static inline int is_cgroup_event(struct perf_event
*event
)
680 static inline u64
perf_cgroup_event_cgrp_time(struct perf_event
*event
)
685 static inline void update_cgrp_time_from_event(struct perf_event
*event
)
689 static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context
*cpuctx
)
693 static inline void perf_cgroup_sched_out(struct task_struct
*task
,
694 struct task_struct
*next
)
698 static inline void perf_cgroup_sched_in(struct task_struct
*prev
,
699 struct task_struct
*task
)
703 static inline int perf_cgroup_connect(pid_t pid
, struct perf_event
*event
,
704 struct perf_event_attr
*attr
,
705 struct perf_event
*group_leader
)
711 perf_cgroup_set_timestamp(struct task_struct
*task
,
712 struct perf_event_context
*ctx
)
717 perf_cgroup_switch(struct task_struct
*task
, struct task_struct
*next
)
722 perf_cgroup_set_shadow_time(struct perf_event
*event
, u64 now
)
726 static inline u64
perf_cgroup_event_time(struct perf_event
*event
)
732 perf_cgroup_defer_enabled(struct perf_event
*event
)
737 perf_cgroup_mark_enabled(struct perf_event
*event
,
738 struct perf_event_context
*ctx
)
743 void perf_pmu_disable(struct pmu
*pmu
)
745 int *count
= this_cpu_ptr(pmu
->pmu_disable_count
);
747 pmu
->pmu_disable(pmu
);
750 void perf_pmu_enable(struct pmu
*pmu
)
752 int *count
= this_cpu_ptr(pmu
->pmu_disable_count
);
754 pmu
->pmu_enable(pmu
);
757 static DEFINE_PER_CPU(struct list_head
, rotation_list
);
760 * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
761 * because they're strictly cpu affine and rotate_start is called with IRQs
762 * disabled, while rotate_context is called from IRQ context.
764 static void perf_pmu_rotate_start(struct pmu
*pmu
)
766 struct perf_cpu_context
*cpuctx
= this_cpu_ptr(pmu
->pmu_cpu_context
);
767 struct list_head
*head
= &__get_cpu_var(rotation_list
);
769 WARN_ON(!irqs_disabled());
771 if (list_empty(&cpuctx
->rotation_list
)) {
772 int was_empty
= list_empty(head
);
773 list_add(&cpuctx
->rotation_list
, head
);
775 tick_nohz_full_kick();
779 static void get_ctx(struct perf_event_context
*ctx
)
781 WARN_ON(!atomic_inc_not_zero(&ctx
->refcount
));
784 static void put_ctx(struct perf_event_context
*ctx
)
786 if (atomic_dec_and_test(&ctx
->refcount
)) {
788 put_ctx(ctx
->parent_ctx
);
790 put_task_struct(ctx
->task
);
791 kfree_rcu(ctx
, rcu_head
);
795 static void unclone_ctx(struct perf_event_context
*ctx
)
797 if (ctx
->parent_ctx
) {
798 put_ctx(ctx
->parent_ctx
);
799 ctx
->parent_ctx
= NULL
;
803 static u32
perf_event_pid(struct perf_event
*event
, struct task_struct
*p
)
806 * only top level events have the pid namespace they were created in
809 event
= event
->parent
;
811 return task_tgid_nr_ns(p
, event
->ns
);
814 static u32
perf_event_tid(struct perf_event
*event
, struct task_struct
*p
)
817 * only top level events have the pid namespace they were created in
820 event
= event
->parent
;
822 return task_pid_nr_ns(p
, event
->ns
);
826 * If we inherit events we want to return the parent event id
829 static u64
primary_event_id(struct perf_event
*event
)
834 id
= event
->parent
->id
;
840 * Get the perf_event_context for a task and lock it.
841 * This has to cope with with the fact that until it is locked,
842 * the context could get moved to another task.
844 static struct perf_event_context
*
845 perf_lock_task_context(struct task_struct
*task
, int ctxn
, unsigned long *flags
)
847 struct perf_event_context
*ctx
;
851 * One of the few rules of preemptible RCU is that one cannot do
852 * rcu_read_unlock() while holding a scheduler (or nested) lock when
853 * part of the read side critical section was preemptible -- see
854 * rcu_read_unlock_special().
856 * Since ctx->lock nests under rq->lock we must ensure the entire read
857 * side critical section is non-preemptible.
861 ctx
= rcu_dereference(task
->perf_event_ctxp
[ctxn
]);
864 * If this context is a clone of another, it might
865 * get swapped for another underneath us by
866 * perf_event_task_sched_out, though the
867 * rcu_read_lock() protects us from any context
868 * getting freed. Lock the context and check if it
869 * got swapped before we could get the lock, and retry
870 * if so. If we locked the right context, then it
871 * can't get swapped on us any more.
873 raw_spin_lock_irqsave(&ctx
->lock
, *flags
);
874 if (ctx
!= rcu_dereference(task
->perf_event_ctxp
[ctxn
])) {
875 raw_spin_unlock_irqrestore(&ctx
->lock
, *flags
);
881 if (!atomic_inc_not_zero(&ctx
->refcount
)) {
882 raw_spin_unlock_irqrestore(&ctx
->lock
, *flags
);
892 * Get the context for a task and increment its pin_count so it
893 * can't get swapped to another task. This also increments its
894 * reference count so that the context can't get freed.
896 static struct perf_event_context
*
897 perf_pin_task_context(struct task_struct
*task
, int ctxn
)
899 struct perf_event_context
*ctx
;
902 ctx
= perf_lock_task_context(task
, ctxn
, &flags
);
905 raw_spin_unlock_irqrestore(&ctx
->lock
, flags
);
910 static void perf_unpin_context(struct perf_event_context
*ctx
)
914 raw_spin_lock_irqsave(&ctx
->lock
, flags
);
916 raw_spin_unlock_irqrestore(&ctx
->lock
, flags
);
920 * Update the record of the current time in a context.
922 static void update_context_time(struct perf_event_context
*ctx
)
924 u64 now
= perf_clock();
926 ctx
->time
+= now
- ctx
->timestamp
;
927 ctx
->timestamp
= now
;
930 static u64
perf_event_time(struct perf_event
*event
)
932 struct perf_event_context
*ctx
= event
->ctx
;
934 if (is_cgroup_event(event
))
935 return perf_cgroup_event_time(event
);
937 return ctx
? ctx
->time
: 0;
941 * Update the total_time_enabled and total_time_running fields for a event.
942 * The caller of this function needs to hold the ctx->lock.
944 static void update_event_times(struct perf_event
*event
)
946 struct perf_event_context
*ctx
= event
->ctx
;
949 if (event
->state
< PERF_EVENT_STATE_INACTIVE
||
950 event
->group_leader
->state
< PERF_EVENT_STATE_INACTIVE
)
953 * in cgroup mode, time_enabled represents
954 * the time the event was enabled AND active
955 * tasks were in the monitored cgroup. This is
956 * independent of the activity of the context as
957 * there may be a mix of cgroup and non-cgroup events.
959 * That is why we treat cgroup events differently
962 if (is_cgroup_event(event
))
963 run_end
= perf_cgroup_event_time(event
);
964 else if (ctx
->is_active
)
967 run_end
= event
->tstamp_stopped
;
969 event
->total_time_enabled
= run_end
- event
->tstamp_enabled
;
971 if (event
->state
== PERF_EVENT_STATE_INACTIVE
)
972 run_end
= event
->tstamp_stopped
;
974 run_end
= perf_event_time(event
);
976 event
->total_time_running
= run_end
- event
->tstamp_running
;
981 * Update total_time_enabled and total_time_running for all events in a group.
983 static void update_group_times(struct perf_event
*leader
)
985 struct perf_event
*event
;
987 update_event_times(leader
);
988 list_for_each_entry(event
, &leader
->sibling_list
, group_entry
)
989 update_event_times(event
);
992 static struct list_head
*
993 ctx_group_list(struct perf_event
*event
, struct perf_event_context
*ctx
)
995 if (event
->attr
.pinned
)
996 return &ctx
->pinned_groups
;
998 return &ctx
->flexible_groups
;
1002 * Add a event from the lists for its context.
1003 * Must be called with ctx->mutex and ctx->lock held.
1006 list_add_event(struct perf_event
*event
, struct perf_event_context
*ctx
)
1008 WARN_ON_ONCE(event
->attach_state
& PERF_ATTACH_CONTEXT
);
1009 event
->attach_state
|= PERF_ATTACH_CONTEXT
;
1012 * If we're a stand alone event or group leader, we go to the context
1013 * list, group events are kept attached to the group so that
1014 * perf_group_detach can, at all times, locate all siblings.
1016 if (event
->group_leader
== event
) {
1017 struct list_head
*list
;
1019 if (is_software_event(event
))
1020 event
->group_flags
|= PERF_GROUP_SOFTWARE
;
1022 list
= ctx_group_list(event
, ctx
);
1023 list_add_tail(&event
->group_entry
, list
);
1026 if (is_cgroup_event(event
))
1029 if (has_branch_stack(event
))
1030 ctx
->nr_branch_stack
++;
1032 list_add_rcu(&event
->event_entry
, &ctx
->event_list
);
1033 if (!ctx
->nr_events
)
1034 perf_pmu_rotate_start(ctx
->pmu
);
1036 if (event
->attr
.inherit_stat
)
1041 * Initialize event state based on the perf_event_attr::disabled.
1043 static inline void perf_event__state_init(struct perf_event
*event
)
1045 event
->state
= event
->attr
.disabled
? PERF_EVENT_STATE_OFF
:
1046 PERF_EVENT_STATE_INACTIVE
;
1050 * Called at perf_event creation and when events are attached/detached from a
1053 static void perf_event__read_size(struct perf_event
*event
)
1055 int entry
= sizeof(u64
); /* value */
1059 if (event
->attr
.read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
1060 size
+= sizeof(u64
);
1062 if (event
->attr
.read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
1063 size
+= sizeof(u64
);
1065 if (event
->attr
.read_format
& PERF_FORMAT_ID
)
1066 entry
+= sizeof(u64
);
1068 if (event
->attr
.read_format
& PERF_FORMAT_GROUP
) {
1069 nr
+= event
->group_leader
->nr_siblings
;
1070 size
+= sizeof(u64
);
1074 event
->read_size
= size
;
1077 static void perf_event__header_size(struct perf_event
*event
)
1079 struct perf_sample_data
*data
;
1080 u64 sample_type
= event
->attr
.sample_type
;
1083 perf_event__read_size(event
);
1085 if (sample_type
& PERF_SAMPLE_IP
)
1086 size
+= sizeof(data
->ip
);
1088 if (sample_type
& PERF_SAMPLE_ADDR
)
1089 size
+= sizeof(data
->addr
);
1091 if (sample_type
& PERF_SAMPLE_PERIOD
)
1092 size
+= sizeof(data
->period
);
1094 if (sample_type
& PERF_SAMPLE_WEIGHT
)
1095 size
+= sizeof(data
->weight
);
1097 if (sample_type
& PERF_SAMPLE_READ
)
1098 size
+= event
->read_size
;
1100 if (sample_type
& PERF_SAMPLE_DATA_SRC
)
1101 size
+= sizeof(data
->data_src
.val
);
1103 event
->header_size
= size
;
1106 static void perf_event__id_header_size(struct perf_event
*event
)
1108 struct perf_sample_data
*data
;
1109 u64 sample_type
= event
->attr
.sample_type
;
1112 if (sample_type
& PERF_SAMPLE_TID
)
1113 size
+= sizeof(data
->tid_entry
);
1115 if (sample_type
& PERF_SAMPLE_TIME
)
1116 size
+= sizeof(data
->time
);
1118 if (sample_type
& PERF_SAMPLE_ID
)
1119 size
+= sizeof(data
->id
);
1121 if (sample_type
& PERF_SAMPLE_STREAM_ID
)
1122 size
+= sizeof(data
->stream_id
);
1124 if (sample_type
& PERF_SAMPLE_CPU
)
1125 size
+= sizeof(data
->cpu_entry
);
1127 event
->id_header_size
= size
;
1130 static void perf_group_attach(struct perf_event
*event
)
1132 struct perf_event
*group_leader
= event
->group_leader
, *pos
;
1135 * We can have double attach due to group movement in perf_event_open.
1137 if (event
->attach_state
& PERF_ATTACH_GROUP
)
1140 event
->attach_state
|= PERF_ATTACH_GROUP
;
1142 if (group_leader
== event
)
1145 if (group_leader
->group_flags
& PERF_GROUP_SOFTWARE
&&
1146 !is_software_event(event
))
1147 group_leader
->group_flags
&= ~PERF_GROUP_SOFTWARE
;
1149 list_add_tail(&event
->group_entry
, &group_leader
->sibling_list
);
1150 group_leader
->nr_siblings
++;
1152 perf_event__header_size(group_leader
);
1154 list_for_each_entry(pos
, &group_leader
->sibling_list
, group_entry
)
1155 perf_event__header_size(pos
);
1159 * Remove a event from the lists for its context.
1160 * Must be called with ctx->mutex and ctx->lock held.
1163 list_del_event(struct perf_event
*event
, struct perf_event_context
*ctx
)
1165 struct perf_cpu_context
*cpuctx
;
1167 * We can have double detach due to exit/hot-unplug + close.
1169 if (!(event
->attach_state
& PERF_ATTACH_CONTEXT
))
1172 event
->attach_state
&= ~PERF_ATTACH_CONTEXT
;
1174 if (is_cgroup_event(event
)) {
1176 cpuctx
= __get_cpu_context(ctx
);
1178 * if there are no more cgroup events
1179 * then cler cgrp to avoid stale pointer
1180 * in update_cgrp_time_from_cpuctx()
1182 if (!ctx
->nr_cgroups
)
1183 cpuctx
->cgrp
= NULL
;
1186 if (has_branch_stack(event
))
1187 ctx
->nr_branch_stack
--;
1190 if (event
->attr
.inherit_stat
)
1193 list_del_rcu(&event
->event_entry
);
1195 if (event
->group_leader
== event
)
1196 list_del_init(&event
->group_entry
);
1198 update_group_times(event
);
1201 * If event was in error state, then keep it
1202 * that way, otherwise bogus counts will be
1203 * returned on read(). The only way to get out
1204 * of error state is by explicit re-enabling
1207 if (event
->state
> PERF_EVENT_STATE_OFF
)
1208 event
->state
= PERF_EVENT_STATE_OFF
;
1211 static void perf_group_detach(struct perf_event
*event
)
1213 struct perf_event
*sibling
, *tmp
;
1214 struct list_head
*list
= NULL
;
1217 * We can have double detach due to exit/hot-unplug + close.
1219 if (!(event
->attach_state
& PERF_ATTACH_GROUP
))
1222 event
->attach_state
&= ~PERF_ATTACH_GROUP
;
1225 * If this is a sibling, remove it from its group.
1227 if (event
->group_leader
!= event
) {
1228 list_del_init(&event
->group_entry
);
1229 event
->group_leader
->nr_siblings
--;
1233 if (!list_empty(&event
->group_entry
))
1234 list
= &event
->group_entry
;
1237 * If this was a group event with sibling events then
1238 * upgrade the siblings to singleton events by adding them
1239 * to whatever list we are on.
1241 list_for_each_entry_safe(sibling
, tmp
, &event
->sibling_list
, group_entry
) {
1243 list_move_tail(&sibling
->group_entry
, list
);
1244 sibling
->group_leader
= sibling
;
1246 /* Inherit group flags from the previous leader */
1247 sibling
->group_flags
= event
->group_flags
;
1251 perf_event__header_size(event
->group_leader
);
1253 list_for_each_entry(tmp
, &event
->group_leader
->sibling_list
, group_entry
)
1254 perf_event__header_size(tmp
);
1258 event_filter_match(struct perf_event
*event
)
1260 return (event
->cpu
== -1 || event
->cpu
== smp_processor_id())
1261 && perf_cgroup_match(event
);
1265 event_sched_out(struct perf_event
*event
,
1266 struct perf_cpu_context
*cpuctx
,
1267 struct perf_event_context
*ctx
)
1269 u64 tstamp
= perf_event_time(event
);
1272 * An event which could not be activated because of
1273 * filter mismatch still needs to have its timings
1274 * maintained, otherwise bogus information is return
1275 * via read() for time_enabled, time_running:
1277 if (event
->state
== PERF_EVENT_STATE_INACTIVE
1278 && !event_filter_match(event
)) {
1279 delta
= tstamp
- event
->tstamp_stopped
;
1280 event
->tstamp_running
+= delta
;
1281 event
->tstamp_stopped
= tstamp
;
1284 if (event
->state
!= PERF_EVENT_STATE_ACTIVE
)
1287 event
->state
= PERF_EVENT_STATE_INACTIVE
;
1288 if (event
->pending_disable
) {
1289 event
->pending_disable
= 0;
1290 event
->state
= PERF_EVENT_STATE_OFF
;
1292 event
->tstamp_stopped
= tstamp
;
1293 event
->pmu
->del(event
, 0);
1296 if (!is_software_event(event
))
1297 cpuctx
->active_oncpu
--;
1299 if (event
->attr
.freq
&& event
->attr
.sample_freq
)
1301 if (event
->attr
.exclusive
|| !cpuctx
->active_oncpu
)
1302 cpuctx
->exclusive
= 0;
1306 group_sched_out(struct perf_event
*group_event
,
1307 struct perf_cpu_context
*cpuctx
,
1308 struct perf_event_context
*ctx
)
1310 struct perf_event
*event
;
1311 int state
= group_event
->state
;
1313 event_sched_out(group_event
, cpuctx
, ctx
);
1316 * Schedule out siblings (if any):
1318 list_for_each_entry(event
, &group_event
->sibling_list
, group_entry
)
1319 event_sched_out(event
, cpuctx
, ctx
);
1321 if (state
== PERF_EVENT_STATE_ACTIVE
&& group_event
->attr
.exclusive
)
1322 cpuctx
->exclusive
= 0;
1325 struct remove_event
{
1326 struct perf_event
*event
;
1331 * Cross CPU call to remove a performance event
1333 * We disable the event on the hardware level first. After that we
1334 * remove it from the context list.
1336 static int __perf_remove_from_context(void *info
)
1338 struct remove_event
*re
= info
;
1339 struct perf_event
*event
= re
->event
;
1340 struct perf_event_context
*ctx
= event
->ctx
;
1341 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
1343 raw_spin_lock(&ctx
->lock
);
1344 event_sched_out(event
, cpuctx
, ctx
);
1345 if (re
->detach_group
)
1346 perf_group_detach(event
);
1347 list_del_event(event
, ctx
);
1348 if (!ctx
->nr_events
&& cpuctx
->task_ctx
== ctx
) {
1350 cpuctx
->task_ctx
= NULL
;
1352 raw_spin_unlock(&ctx
->lock
);
1359 * Remove the event from a task's (or a CPU's) list of events.
1361 * CPU events are removed with a smp call. For task events we only
1362 * call when the task is on a CPU.
1364 * If event->ctx is a cloned context, callers must make sure that
1365 * every task struct that event->ctx->task could possibly point to
1366 * remains valid. This is OK when called from perf_release since
1367 * that only calls us on the top-level context, which can't be a clone.
1368 * When called from perf_event_exit_task, it's OK because the
1369 * context has been detached from its task.
1371 static void perf_remove_from_context(struct perf_event
*event
, bool detach_group
)
1373 struct perf_event_context
*ctx
= event
->ctx
;
1374 struct task_struct
*task
= ctx
->task
;
1375 struct remove_event re
= {
1377 .detach_group
= detach_group
,
1380 lockdep_assert_held(&ctx
->mutex
);
1384 * Per cpu events are removed via an smp call and
1385 * the removal is always successful.
1387 cpu_function_call(event
->cpu
, __perf_remove_from_context
, &re
);
1392 if (!task_function_call(task
, __perf_remove_from_context
, &re
))
1395 raw_spin_lock_irq(&ctx
->lock
);
1397 * If we failed to find a running task, but find the context active now
1398 * that we've acquired the ctx->lock, retry.
1400 if (ctx
->is_active
) {
1401 raw_spin_unlock_irq(&ctx
->lock
);
1403 * Reload the task pointer, it might have been changed by
1404 * a concurrent perf_event_context_sched_out().
1411 * Since the task isn't running, its safe to remove the event, us
1412 * holding the ctx->lock ensures the task won't get scheduled in.
1415 perf_group_detach(event
);
1416 list_del_event(event
, ctx
);
1417 raw_spin_unlock_irq(&ctx
->lock
);
1421 * Cross CPU call to disable a performance event
1423 int __perf_event_disable(void *info
)
1425 struct perf_event
*event
= info
;
1426 struct perf_event_context
*ctx
= event
->ctx
;
1427 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
1430 * If this is a per-task event, need to check whether this
1431 * event's task is the current task on this cpu.
1433 * Can trigger due to concurrent perf_event_context_sched_out()
1434 * flipping contexts around.
1436 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
)
1439 raw_spin_lock(&ctx
->lock
);
1442 * If the event is on, turn it off.
1443 * If it is in error state, leave it in error state.
1445 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
) {
1446 update_context_time(ctx
);
1447 update_cgrp_time_from_event(event
);
1448 update_group_times(event
);
1449 if (event
== event
->group_leader
)
1450 group_sched_out(event
, cpuctx
, ctx
);
1452 event_sched_out(event
, cpuctx
, ctx
);
1453 event
->state
= PERF_EVENT_STATE_OFF
;
1456 raw_spin_unlock(&ctx
->lock
);
1464 * If event->ctx is a cloned context, callers must make sure that
1465 * every task struct that event->ctx->task could possibly point to
1466 * remains valid. This condition is satisifed when called through
1467 * perf_event_for_each_child or perf_event_for_each because they
1468 * hold the top-level event's child_mutex, so any descendant that
1469 * goes to exit will block in sync_child_event.
1470 * When called from perf_pending_event it's OK because event->ctx
1471 * is the current context on this CPU and preemption is disabled,
1472 * hence we can't get into perf_event_task_sched_out for this context.
1474 void perf_event_disable(struct perf_event
*event
)
1476 struct perf_event_context
*ctx
= event
->ctx
;
1477 struct task_struct
*task
= ctx
->task
;
1481 * Disable the event on the cpu that it's on
1483 cpu_function_call(event
->cpu
, __perf_event_disable
, event
);
1488 if (!task_function_call(task
, __perf_event_disable
, event
))
1491 raw_spin_lock_irq(&ctx
->lock
);
1493 * If the event is still active, we need to retry the cross-call.
1495 if (event
->state
== PERF_EVENT_STATE_ACTIVE
) {
1496 raw_spin_unlock_irq(&ctx
->lock
);
1498 * Reload the task pointer, it might have been changed by
1499 * a concurrent perf_event_context_sched_out().
1506 * Since we have the lock this context can't be scheduled
1507 * in, so we can change the state safely.
1509 if (event
->state
== PERF_EVENT_STATE_INACTIVE
) {
1510 update_group_times(event
);
1511 event
->state
= PERF_EVENT_STATE_OFF
;
1513 raw_spin_unlock_irq(&ctx
->lock
);
1515 EXPORT_SYMBOL_GPL(perf_event_disable
);
1517 static void perf_set_shadow_time(struct perf_event
*event
,
1518 struct perf_event_context
*ctx
,
1522 * use the correct time source for the time snapshot
1524 * We could get by without this by leveraging the
1525 * fact that to get to this function, the caller
1526 * has most likely already called update_context_time()
1527 * and update_cgrp_time_xx() and thus both timestamp
1528 * are identical (or very close). Given that tstamp is,
1529 * already adjusted for cgroup, we could say that:
1530 * tstamp - ctx->timestamp
1532 * tstamp - cgrp->timestamp.
1534 * Then, in perf_output_read(), the calculation would
1535 * work with no changes because:
1536 * - event is guaranteed scheduled in
1537 * - no scheduled out in between
1538 * - thus the timestamp would be the same
1540 * But this is a bit hairy.
1542 * So instead, we have an explicit cgroup call to remain
1543 * within the time time source all along. We believe it
1544 * is cleaner and simpler to understand.
1546 if (is_cgroup_event(event
))
1547 perf_cgroup_set_shadow_time(event
, tstamp
);
1549 event
->shadow_ctx_time
= tstamp
- ctx
->timestamp
;
1552 #define MAX_INTERRUPTS (~0ULL)
1554 static void perf_log_throttle(struct perf_event
*event
, int enable
);
1557 event_sched_in(struct perf_event
*event
,
1558 struct perf_cpu_context
*cpuctx
,
1559 struct perf_event_context
*ctx
)
1561 u64 tstamp
= perf_event_time(event
);
1563 if (event
->state
<= PERF_EVENT_STATE_OFF
)
1566 event
->state
= PERF_EVENT_STATE_ACTIVE
;
1567 event
->oncpu
= smp_processor_id();
1570 * Unthrottle events, since we scheduled we might have missed several
1571 * ticks already, also for a heavily scheduling task there is little
1572 * guarantee it'll get a tick in a timely manner.
1574 if (unlikely(event
->hw
.interrupts
== MAX_INTERRUPTS
)) {
1575 perf_log_throttle(event
, 1);
1576 event
->hw
.interrupts
= 0;
1580 * The new state must be visible before we turn it on in the hardware:
1584 if (event
->pmu
->add(event
, PERF_EF_START
)) {
1585 event
->state
= PERF_EVENT_STATE_INACTIVE
;
1590 event
->tstamp_running
+= tstamp
- event
->tstamp_stopped
;
1592 perf_set_shadow_time(event
, ctx
, tstamp
);
1594 if (!is_software_event(event
))
1595 cpuctx
->active_oncpu
++;
1597 if (event
->attr
.freq
&& event
->attr
.sample_freq
)
1600 if (event
->attr
.exclusive
)
1601 cpuctx
->exclusive
= 1;
1607 group_sched_in(struct perf_event
*group_event
,
1608 struct perf_cpu_context
*cpuctx
,
1609 struct perf_event_context
*ctx
)
1611 struct perf_event
*event
, *partial_group
= NULL
;
1612 struct pmu
*pmu
= group_event
->pmu
;
1613 u64 now
= ctx
->time
;
1614 bool simulate
= false;
1616 if (group_event
->state
== PERF_EVENT_STATE_OFF
)
1619 pmu
->start_txn(pmu
);
1621 if (event_sched_in(group_event
, cpuctx
, ctx
)) {
1622 pmu
->cancel_txn(pmu
);
1627 * Schedule in siblings as one group (if any):
1629 list_for_each_entry(event
, &group_event
->sibling_list
, group_entry
) {
1630 if (event_sched_in(event
, cpuctx
, ctx
)) {
1631 partial_group
= event
;
1636 if (!pmu
->commit_txn(pmu
))
1641 * Groups can be scheduled in as one unit only, so undo any
1642 * partial group before returning:
1643 * The events up to the failed event are scheduled out normally,
1644 * tstamp_stopped will be updated.
1646 * The failed events and the remaining siblings need to have
1647 * their timings updated as if they had gone thru event_sched_in()
1648 * and event_sched_out(). This is required to get consistent timings
1649 * across the group. This also takes care of the case where the group
1650 * could never be scheduled by ensuring tstamp_stopped is set to mark
1651 * the time the event was actually stopped, such that time delta
1652 * calculation in update_event_times() is correct.
1654 list_for_each_entry(event
, &group_event
->sibling_list
, group_entry
) {
1655 if (event
== partial_group
)
1659 event
->tstamp_running
+= now
- event
->tstamp_stopped
;
1660 event
->tstamp_stopped
= now
;
1662 event_sched_out(event
, cpuctx
, ctx
);
1665 event_sched_out(group_event
, cpuctx
, ctx
);
1667 pmu
->cancel_txn(pmu
);
1673 * Work out whether we can put this event group on the CPU now.
1675 static int group_can_go_on(struct perf_event
*event
,
1676 struct perf_cpu_context
*cpuctx
,
1680 * Groups consisting entirely of software events can always go on.
1682 if (event
->group_flags
& PERF_GROUP_SOFTWARE
)
1685 * If an exclusive group is already on, no other hardware
1688 if (cpuctx
->exclusive
)
1691 * If this group is exclusive and there are already
1692 * events on the CPU, it can't go on.
1694 if (event
->attr
.exclusive
&& cpuctx
->active_oncpu
)
1697 * Otherwise, try to add it if all previous groups were able
1703 static void add_event_to_ctx(struct perf_event
*event
,
1704 struct perf_event_context
*ctx
)
1706 u64 tstamp
= perf_event_time(event
);
1708 list_add_event(event
, ctx
);
1709 perf_group_attach(event
);
1710 event
->tstamp_enabled
= tstamp
;
1711 event
->tstamp_running
= tstamp
;
1712 event
->tstamp_stopped
= tstamp
;
1715 static void task_ctx_sched_out(struct perf_event_context
*ctx
);
1717 ctx_sched_in(struct perf_event_context
*ctx
,
1718 struct perf_cpu_context
*cpuctx
,
1719 enum event_type_t event_type
,
1720 struct task_struct
*task
);
1722 static void perf_event_sched_in(struct perf_cpu_context
*cpuctx
,
1723 struct perf_event_context
*ctx
,
1724 struct task_struct
*task
)
1726 cpu_ctx_sched_in(cpuctx
, EVENT_PINNED
, task
);
1728 ctx_sched_in(ctx
, cpuctx
, EVENT_PINNED
, task
);
1729 cpu_ctx_sched_in(cpuctx
, EVENT_FLEXIBLE
, task
);
1731 ctx_sched_in(ctx
, cpuctx
, EVENT_FLEXIBLE
, task
);
1735 * Cross CPU call to install and enable a performance event
1737 * Must be called with ctx->mutex held
1739 static int __perf_install_in_context(void *info
)
1741 struct perf_event
*event
= info
;
1742 struct perf_event_context
*ctx
= event
->ctx
;
1743 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
1744 struct perf_event_context
*task_ctx
= cpuctx
->task_ctx
;
1745 struct task_struct
*task
= current
;
1747 perf_ctx_lock(cpuctx
, task_ctx
);
1748 perf_pmu_disable(cpuctx
->ctx
.pmu
);
1751 * If there was an active task_ctx schedule it out.
1754 task_ctx_sched_out(task_ctx
);
1757 * If the context we're installing events in is not the
1758 * active task_ctx, flip them.
1760 if (ctx
->task
&& task_ctx
!= ctx
) {
1762 raw_spin_unlock(&task_ctx
->lock
);
1763 raw_spin_lock(&ctx
->lock
);
1768 cpuctx
->task_ctx
= task_ctx
;
1769 task
= task_ctx
->task
;
1772 cpu_ctx_sched_out(cpuctx
, EVENT_ALL
);
1774 update_context_time(ctx
);
1776 * update cgrp time only if current cgrp
1777 * matches event->cgrp. Must be done before
1778 * calling add_event_to_ctx()
1780 update_cgrp_time_from_event(event
);
1782 add_event_to_ctx(event
, ctx
);
1785 * Schedule everything back in
1787 perf_event_sched_in(cpuctx
, task_ctx
, task
);
1789 perf_pmu_enable(cpuctx
->ctx
.pmu
);
1790 perf_ctx_unlock(cpuctx
, task_ctx
);
1796 * Attach a performance event to a context
1798 * First we add the event to the list with the hardware enable bit
1799 * in event->hw_config cleared.
1801 * If the event is attached to a task which is on a CPU we use a smp
1802 * call to enable it in the task context. The task might have been
1803 * scheduled away, but we check this in the smp call again.
1806 perf_install_in_context(struct perf_event_context
*ctx
,
1807 struct perf_event
*event
,
1810 struct task_struct
*task
= ctx
->task
;
1812 lockdep_assert_held(&ctx
->mutex
);
1815 if (event
->cpu
!= -1)
1820 * Per cpu events are installed via an smp call and
1821 * the install is always successful.
1823 cpu_function_call(cpu
, __perf_install_in_context
, event
);
1828 if (!task_function_call(task
, __perf_install_in_context
, event
))
1831 raw_spin_lock_irq(&ctx
->lock
);
1833 * If we failed to find a running task, but find the context active now
1834 * that we've acquired the ctx->lock, retry.
1836 if (ctx
->is_active
) {
1837 raw_spin_unlock_irq(&ctx
->lock
);
1839 * Reload the task pointer, it might have been changed by
1840 * a concurrent perf_event_context_sched_out().
1847 * Since the task isn't running, its safe to add the event, us holding
1848 * the ctx->lock ensures the task won't get scheduled in.
1850 add_event_to_ctx(event
, ctx
);
1851 raw_spin_unlock_irq(&ctx
->lock
);
1855 * Put a event into inactive state and update time fields.
1856 * Enabling the leader of a group effectively enables all
1857 * the group members that aren't explicitly disabled, so we
1858 * have to update their ->tstamp_enabled also.
1859 * Note: this works for group members as well as group leaders
1860 * since the non-leader members' sibling_lists will be empty.
1862 static void __perf_event_mark_enabled(struct perf_event
*event
)
1864 struct perf_event
*sub
;
1865 u64 tstamp
= perf_event_time(event
);
1867 event
->state
= PERF_EVENT_STATE_INACTIVE
;
1868 event
->tstamp_enabled
= tstamp
- event
->total_time_enabled
;
1869 list_for_each_entry(sub
, &event
->sibling_list
, group_entry
) {
1870 if (sub
->state
>= PERF_EVENT_STATE_INACTIVE
)
1871 sub
->tstamp_enabled
= tstamp
- sub
->total_time_enabled
;
1876 * Cross CPU call to enable a performance event
1878 static int __perf_event_enable(void *info
)
1880 struct perf_event
*event
= info
;
1881 struct perf_event_context
*ctx
= event
->ctx
;
1882 struct perf_event
*leader
= event
->group_leader
;
1883 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
1887 * There's a time window between 'ctx->is_active' check
1888 * in perf_event_enable function and this place having:
1890 * - ctx->lock unlocked
1892 * where the task could be killed and 'ctx' deactivated
1893 * by perf_event_exit_task.
1895 if (!ctx
->is_active
)
1898 raw_spin_lock(&ctx
->lock
);
1899 update_context_time(ctx
);
1901 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
)
1905 * set current task's cgroup time reference point
1907 perf_cgroup_set_timestamp(current
, ctx
);
1909 __perf_event_mark_enabled(event
);
1911 if (!event_filter_match(event
)) {
1912 if (is_cgroup_event(event
))
1913 perf_cgroup_defer_enabled(event
);
1918 * If the event is in a group and isn't the group leader,
1919 * then don't put it on unless the group is on.
1921 if (leader
!= event
&& leader
->state
!= PERF_EVENT_STATE_ACTIVE
)
1924 if (!group_can_go_on(event
, cpuctx
, 1)) {
1927 if (event
== leader
)
1928 err
= group_sched_in(event
, cpuctx
, ctx
);
1930 err
= event_sched_in(event
, cpuctx
, ctx
);
1935 * If this event can't go on and it's part of a
1936 * group, then the whole group has to come off.
1938 if (leader
!= event
)
1939 group_sched_out(leader
, cpuctx
, ctx
);
1940 if (leader
->attr
.pinned
) {
1941 update_group_times(leader
);
1942 leader
->state
= PERF_EVENT_STATE_ERROR
;
1947 raw_spin_unlock(&ctx
->lock
);
1955 * If event->ctx is a cloned context, callers must make sure that
1956 * every task struct that event->ctx->task could possibly point to
1957 * remains valid. This condition is satisfied when called through
1958 * perf_event_for_each_child or perf_event_for_each as described
1959 * for perf_event_disable.
1961 void perf_event_enable(struct perf_event
*event
)
1963 struct perf_event_context
*ctx
= event
->ctx
;
1964 struct task_struct
*task
= ctx
->task
;
1968 * Enable the event on the cpu that it's on
1970 cpu_function_call(event
->cpu
, __perf_event_enable
, event
);
1974 raw_spin_lock_irq(&ctx
->lock
);
1975 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
)
1979 * If the event is in error state, clear that first.
1980 * That way, if we see the event in error state below, we
1981 * know that it has gone back into error state, as distinct
1982 * from the task having been scheduled away before the
1983 * cross-call arrived.
1985 if (event
->state
== PERF_EVENT_STATE_ERROR
)
1986 event
->state
= PERF_EVENT_STATE_OFF
;
1989 if (!ctx
->is_active
) {
1990 __perf_event_mark_enabled(event
);
1994 raw_spin_unlock_irq(&ctx
->lock
);
1996 if (!task_function_call(task
, __perf_event_enable
, event
))
1999 raw_spin_lock_irq(&ctx
->lock
);
2002 * If the context is active and the event is still off,
2003 * we need to retry the cross-call.
2005 if (ctx
->is_active
&& event
->state
== PERF_EVENT_STATE_OFF
) {
2007 * task could have been flipped by a concurrent
2008 * perf_event_context_sched_out()
2015 raw_spin_unlock_irq(&ctx
->lock
);
2017 EXPORT_SYMBOL_GPL(perf_event_enable
);
2019 int perf_event_refresh(struct perf_event
*event
, int refresh
)
2022 * not supported on inherited events
2024 if (event
->attr
.inherit
|| !is_sampling_event(event
))
2027 atomic_add(refresh
, &event
->event_limit
);
2028 perf_event_enable(event
);
2032 EXPORT_SYMBOL_GPL(perf_event_refresh
);
2034 static void ctx_sched_out(struct perf_event_context
*ctx
,
2035 struct perf_cpu_context
*cpuctx
,
2036 enum event_type_t event_type
)
2038 struct perf_event
*event
;
2039 int is_active
= ctx
->is_active
;
2041 ctx
->is_active
&= ~event_type
;
2042 if (likely(!ctx
->nr_events
))
2045 update_context_time(ctx
);
2046 update_cgrp_time_from_cpuctx(cpuctx
);
2047 if (!ctx
->nr_active
)
2050 perf_pmu_disable(ctx
->pmu
);
2051 if ((is_active
& EVENT_PINNED
) && (event_type
& EVENT_PINNED
)) {
2052 list_for_each_entry(event
, &ctx
->pinned_groups
, group_entry
)
2053 group_sched_out(event
, cpuctx
, ctx
);
2056 if ((is_active
& EVENT_FLEXIBLE
) && (event_type
& EVENT_FLEXIBLE
)) {
2057 list_for_each_entry(event
, &ctx
->flexible_groups
, group_entry
)
2058 group_sched_out(event
, cpuctx
, ctx
);
2060 perf_pmu_enable(ctx
->pmu
);
2064 * Test whether two contexts are equivalent, i.e. whether they
2065 * have both been cloned from the same version of the same context
2066 * and they both have the same number of enabled events.
2067 * If the number of enabled events is the same, then the set
2068 * of enabled events should be the same, because these are both
2069 * inherited contexts, therefore we can't access individual events
2070 * in them directly with an fd; we can only enable/disable all
2071 * events via prctl, or enable/disable all events in a family
2072 * via ioctl, which will have the same effect on both contexts.
2074 static int context_equiv(struct perf_event_context
*ctx1
,
2075 struct perf_event_context
*ctx2
)
2077 return ctx1
->parent_ctx
&& ctx1
->parent_ctx
== ctx2
->parent_ctx
2078 && ctx1
->parent_gen
== ctx2
->parent_gen
2079 && !ctx1
->pin_count
&& !ctx2
->pin_count
;
2082 static void __perf_event_sync_stat(struct perf_event
*event
,
2083 struct perf_event
*next_event
)
2087 if (!event
->attr
.inherit_stat
)
2091 * Update the event value, we cannot use perf_event_read()
2092 * because we're in the middle of a context switch and have IRQs
2093 * disabled, which upsets smp_call_function_single(), however
2094 * we know the event must be on the current CPU, therefore we
2095 * don't need to use it.
2097 switch (event
->state
) {
2098 case PERF_EVENT_STATE_ACTIVE
:
2099 event
->pmu
->read(event
);
2102 case PERF_EVENT_STATE_INACTIVE
:
2103 update_event_times(event
);
2111 * In order to keep per-task stats reliable we need to flip the event
2112 * values when we flip the contexts.
2114 value
= local64_read(&next_event
->count
);
2115 value
= local64_xchg(&event
->count
, value
);
2116 local64_set(&next_event
->count
, value
);
2118 swap(event
->total_time_enabled
, next_event
->total_time_enabled
);
2119 swap(event
->total_time_running
, next_event
->total_time_running
);
2122 * Since we swizzled the values, update the user visible data too.
2124 perf_event_update_userpage(event
);
2125 perf_event_update_userpage(next_event
);
2128 static void perf_event_sync_stat(struct perf_event_context
*ctx
,
2129 struct perf_event_context
*next_ctx
)
2131 struct perf_event
*event
, *next_event
;
2136 update_context_time(ctx
);
2138 event
= list_first_entry(&ctx
->event_list
,
2139 struct perf_event
, event_entry
);
2141 next_event
= list_first_entry(&next_ctx
->event_list
,
2142 struct perf_event
, event_entry
);
2144 while (&event
->event_entry
!= &ctx
->event_list
&&
2145 &next_event
->event_entry
!= &next_ctx
->event_list
) {
2147 __perf_event_sync_stat(event
, next_event
);
2149 event
= list_next_entry(event
, event_entry
);
2150 next_event
= list_next_entry(next_event
, event_entry
);
2154 static void perf_event_context_sched_out(struct task_struct
*task
, int ctxn
,
2155 struct task_struct
*next
)
2157 struct perf_event_context
*ctx
= task
->perf_event_ctxp
[ctxn
];
2158 struct perf_event_context
*next_ctx
;
2159 struct perf_event_context
*parent
;
2160 struct perf_cpu_context
*cpuctx
;
2166 cpuctx
= __get_cpu_context(ctx
);
2167 if (!cpuctx
->task_ctx
)
2171 parent
= rcu_dereference(ctx
->parent_ctx
);
2172 next_ctx
= next
->perf_event_ctxp
[ctxn
];
2173 if (parent
&& next_ctx
&&
2174 rcu_dereference(next_ctx
->parent_ctx
) == parent
) {
2176 * Looks like the two contexts are clones, so we might be
2177 * able to optimize the context switch. We lock both
2178 * contexts and check that they are clones under the
2179 * lock (including re-checking that neither has been
2180 * uncloned in the meantime). It doesn't matter which
2181 * order we take the locks because no other cpu could
2182 * be trying to lock both of these tasks.
2184 raw_spin_lock(&ctx
->lock
);
2185 raw_spin_lock_nested(&next_ctx
->lock
, SINGLE_DEPTH_NESTING
);
2186 if (context_equiv(ctx
, next_ctx
)) {
2188 * XXX do we need a memory barrier of sorts
2189 * wrt to rcu_dereference() of perf_event_ctxp
2191 task
->perf_event_ctxp
[ctxn
] = next_ctx
;
2192 next
->perf_event_ctxp
[ctxn
] = ctx
;
2194 next_ctx
->task
= task
;
2197 perf_event_sync_stat(ctx
, next_ctx
);
2199 raw_spin_unlock(&next_ctx
->lock
);
2200 raw_spin_unlock(&ctx
->lock
);
2205 raw_spin_lock(&ctx
->lock
);
2206 ctx_sched_out(ctx
, cpuctx
, EVENT_ALL
);
2207 cpuctx
->task_ctx
= NULL
;
2208 raw_spin_unlock(&ctx
->lock
);
2212 #define for_each_task_context_nr(ctxn) \
2213 for ((ctxn) = 0; (ctxn) < perf_nr_task_contexts; (ctxn)++)
2216 * Called from scheduler to remove the events of the current task,
2217 * with interrupts disabled.
2219 * We stop each event and update the event value in event->count.
2221 * This does not protect us against NMI, but disable()
2222 * sets the disabled bit in the control field of event _before_
2223 * accessing the event control register. If a NMI hits, then it will
2224 * not restart the event.
2226 void __perf_event_task_sched_out(struct task_struct
*task
,
2227 struct task_struct
*next
)
2231 for_each_task_context_nr(ctxn
)
2232 perf_event_context_sched_out(task
, ctxn
, next
);
2235 * if cgroup events exist on this CPU, then we need
2236 * to check if we have to switch out PMU state.
2237 * cgroup event are system-wide mode only
2239 if (atomic_read(&__get_cpu_var(perf_cgroup_events
)))
2240 perf_cgroup_sched_out(task
, next
);
2243 static void task_ctx_sched_out(struct perf_event_context
*ctx
)
2245 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
2247 if (!cpuctx
->task_ctx
)
2250 if (WARN_ON_ONCE(ctx
!= cpuctx
->task_ctx
))
2253 ctx_sched_out(ctx
, cpuctx
, EVENT_ALL
);
2254 cpuctx
->task_ctx
= NULL
;
2258 * Called with IRQs disabled
2260 static void cpu_ctx_sched_out(struct perf_cpu_context
*cpuctx
,
2261 enum event_type_t event_type
)
2263 ctx_sched_out(&cpuctx
->ctx
, cpuctx
, event_type
);
2267 ctx_pinned_sched_in(struct perf_event_context
*ctx
,
2268 struct perf_cpu_context
*cpuctx
)
2270 struct perf_event
*event
;
2272 list_for_each_entry(event
, &ctx
->pinned_groups
, group_entry
) {
2273 if (event
->state
<= PERF_EVENT_STATE_OFF
)
2275 if (!event_filter_match(event
))
2278 /* may need to reset tstamp_enabled */
2279 if (is_cgroup_event(event
))
2280 perf_cgroup_mark_enabled(event
, ctx
);
2282 if (group_can_go_on(event
, cpuctx
, 1))
2283 group_sched_in(event
, cpuctx
, ctx
);
2286 * If this pinned group hasn't been scheduled,
2287 * put it in error state.
2289 if (event
->state
== PERF_EVENT_STATE_INACTIVE
) {
2290 update_group_times(event
);
2291 event
->state
= PERF_EVENT_STATE_ERROR
;
2297 ctx_flexible_sched_in(struct perf_event_context
*ctx
,
2298 struct perf_cpu_context
*cpuctx
)
2300 struct perf_event
*event
;
2303 list_for_each_entry(event
, &ctx
->flexible_groups
, group_entry
) {
2304 /* Ignore events in OFF or ERROR state */
2305 if (event
->state
<= PERF_EVENT_STATE_OFF
)
2308 * Listen to the 'cpu' scheduling filter constraint
2311 if (!event_filter_match(event
))
2314 /* may need to reset tstamp_enabled */
2315 if (is_cgroup_event(event
))
2316 perf_cgroup_mark_enabled(event
, ctx
);
2318 if (group_can_go_on(event
, cpuctx
, can_add_hw
)) {
2319 if (group_sched_in(event
, cpuctx
, ctx
))
2326 ctx_sched_in(struct perf_event_context
*ctx
,
2327 struct perf_cpu_context
*cpuctx
,
2328 enum event_type_t event_type
,
2329 struct task_struct
*task
)
2332 int is_active
= ctx
->is_active
;
2334 ctx
->is_active
|= event_type
;
2335 if (likely(!ctx
->nr_events
))
2339 ctx
->timestamp
= now
;
2340 perf_cgroup_set_timestamp(task
, ctx
);
2342 * First go through the list and put on any pinned groups
2343 * in order to give them the best chance of going on.
2345 if (!(is_active
& EVENT_PINNED
) && (event_type
& EVENT_PINNED
))
2346 ctx_pinned_sched_in(ctx
, cpuctx
);
2348 /* Then walk through the lower prio flexible groups */
2349 if (!(is_active
& EVENT_FLEXIBLE
) && (event_type
& EVENT_FLEXIBLE
))
2350 ctx_flexible_sched_in(ctx
, cpuctx
);
2353 static void cpu_ctx_sched_in(struct perf_cpu_context
*cpuctx
,
2354 enum event_type_t event_type
,
2355 struct task_struct
*task
)
2357 struct perf_event_context
*ctx
= &cpuctx
->ctx
;
2359 ctx_sched_in(ctx
, cpuctx
, event_type
, task
);
2362 static void perf_event_context_sched_in(struct perf_event_context
*ctx
,
2363 struct task_struct
*task
)
2365 struct perf_cpu_context
*cpuctx
;
2367 cpuctx
= __get_cpu_context(ctx
);
2368 if (cpuctx
->task_ctx
== ctx
)
2371 perf_ctx_lock(cpuctx
, ctx
);
2372 perf_pmu_disable(ctx
->pmu
);
2374 * We want to keep the following priority order:
2375 * cpu pinned (that don't need to move), task pinned,
2376 * cpu flexible, task flexible.
2378 cpu_ctx_sched_out(cpuctx
, EVENT_FLEXIBLE
);
2381 cpuctx
->task_ctx
= ctx
;
2383 perf_event_sched_in(cpuctx
, cpuctx
->task_ctx
, task
);
2385 perf_pmu_enable(ctx
->pmu
);
2386 perf_ctx_unlock(cpuctx
, ctx
);
2389 * Since these rotations are per-cpu, we need to ensure the
2390 * cpu-context we got scheduled on is actually rotating.
2392 perf_pmu_rotate_start(ctx
->pmu
);
2396 * When sampling the branck stack in system-wide, it may be necessary
2397 * to flush the stack on context switch. This happens when the branch
2398 * stack does not tag its entries with the pid of the current task.
2399 * Otherwise it becomes impossible to associate a branch entry with a
2400 * task. This ambiguity is more likely to appear when the branch stack
2401 * supports priv level filtering and the user sets it to monitor only
2402 * at the user level (which could be a useful measurement in system-wide
2403 * mode). In that case, the risk is high of having a branch stack with
2404 * branch from multiple tasks. Flushing may mean dropping the existing
2405 * entries or stashing them somewhere in the PMU specific code layer.
2407 * This function provides the context switch callback to the lower code
2408 * layer. It is invoked ONLY when there is at least one system-wide context
2409 * with at least one active event using taken branch sampling.
2411 static void perf_branch_stack_sched_in(struct task_struct
*prev
,
2412 struct task_struct
*task
)
2414 struct perf_cpu_context
*cpuctx
;
2416 unsigned long flags
;
2418 /* no need to flush branch stack if not changing task */
2422 local_irq_save(flags
);
2426 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
2427 cpuctx
= this_cpu_ptr(pmu
->pmu_cpu_context
);
2430 * check if the context has at least one
2431 * event using PERF_SAMPLE_BRANCH_STACK
2433 if (cpuctx
->ctx
.nr_branch_stack
> 0
2434 && pmu
->flush_branch_stack
) {
2436 pmu
= cpuctx
->ctx
.pmu
;
2438 perf_ctx_lock(cpuctx
, cpuctx
->task_ctx
);
2440 perf_pmu_disable(pmu
);
2442 pmu
->flush_branch_stack();
2444 perf_pmu_enable(pmu
);
2446 perf_ctx_unlock(cpuctx
, cpuctx
->task_ctx
);
2452 local_irq_restore(flags
);
2456 * Called from scheduler to add the events of the current task
2457 * with interrupts disabled.
2459 * We restore the event value and then enable it.
2461 * This does not protect us against NMI, but enable()
2462 * sets the enabled bit in the control field of event _before_
2463 * accessing the event control register. If a NMI hits, then it will
2464 * keep the event running.
2466 void __perf_event_task_sched_in(struct task_struct
*prev
,
2467 struct task_struct
*task
)
2469 struct perf_event_context
*ctx
;
2472 for_each_task_context_nr(ctxn
) {
2473 ctx
= task
->perf_event_ctxp
[ctxn
];
2477 perf_event_context_sched_in(ctx
, task
);
2480 * if cgroup events exist on this CPU, then we need
2481 * to check if we have to switch in PMU state.
2482 * cgroup event are system-wide mode only
2484 if (atomic_read(&__get_cpu_var(perf_cgroup_events
)))
2485 perf_cgroup_sched_in(prev
, task
);
2487 /* check for system-wide branch_stack events */
2488 if (atomic_read(&__get_cpu_var(perf_branch_stack_events
)))
2489 perf_branch_stack_sched_in(prev
, task
);
2492 static u64
perf_calculate_period(struct perf_event
*event
, u64 nsec
, u64 count
)
2494 u64 frequency
= event
->attr
.sample_freq
;
2495 u64 sec
= NSEC_PER_SEC
;
2496 u64 divisor
, dividend
;
2498 int count_fls
, nsec_fls
, frequency_fls
, sec_fls
;
2500 count_fls
= fls64(count
);
2501 nsec_fls
= fls64(nsec
);
2502 frequency_fls
= fls64(frequency
);
2506 * We got @count in @nsec, with a target of sample_freq HZ
2507 * the target period becomes:
2510 * period = -------------------
2511 * @nsec * sample_freq
2516 * Reduce accuracy by one bit such that @a and @b converge
2517 * to a similar magnitude.
2519 #define REDUCE_FLS(a, b) \
2521 if (a##_fls > b##_fls) { \
2531 * Reduce accuracy until either term fits in a u64, then proceed with
2532 * the other, so that finally we can do a u64/u64 division.
2534 while (count_fls
+ sec_fls
> 64 && nsec_fls
+ frequency_fls
> 64) {
2535 REDUCE_FLS(nsec
, frequency
);
2536 REDUCE_FLS(sec
, count
);
2539 if (count_fls
+ sec_fls
> 64) {
2540 divisor
= nsec
* frequency
;
2542 while (count_fls
+ sec_fls
> 64) {
2543 REDUCE_FLS(count
, sec
);
2547 dividend
= count
* sec
;
2549 dividend
= count
* sec
;
2551 while (nsec_fls
+ frequency_fls
> 64) {
2552 REDUCE_FLS(nsec
, frequency
);
2556 divisor
= nsec
* frequency
;
2562 return div64_u64(dividend
, divisor
);
2565 static DEFINE_PER_CPU(int, perf_throttled_count
);
2566 static DEFINE_PER_CPU(u64
, perf_throttled_seq
);
2568 static void perf_adjust_period(struct perf_event
*event
, u64 nsec
, u64 count
, bool disable
)
2570 struct hw_perf_event
*hwc
= &event
->hw
;
2571 s64 period
, sample_period
;
2574 period
= perf_calculate_period(event
, nsec
, count
);
2576 delta
= (s64
)(period
- hwc
->sample_period
);
2577 delta
= (delta
+ 7) / 8; /* low pass filter */
2579 sample_period
= hwc
->sample_period
+ delta
;
2584 hwc
->sample_period
= sample_period
;
2586 if (local64_read(&hwc
->period_left
) > 8*sample_period
) {
2588 event
->pmu
->stop(event
, PERF_EF_UPDATE
);
2590 local64_set(&hwc
->period_left
, 0);
2593 event
->pmu
->start(event
, PERF_EF_RELOAD
);
2598 * combine freq adjustment with unthrottling to avoid two passes over the
2599 * events. At the same time, make sure, having freq events does not change
2600 * the rate of unthrottling as that would introduce bias.
2602 static void perf_adjust_freq_unthr_context(struct perf_event_context
*ctx
,
2605 struct perf_event
*event
;
2606 struct hw_perf_event
*hwc
;
2607 u64 now
, period
= TICK_NSEC
;
2611 * only need to iterate over all events iff:
2612 * - context have events in frequency mode (needs freq adjust)
2613 * - there are events to unthrottle on this cpu
2615 if (!(ctx
->nr_freq
|| needs_unthr
))
2618 raw_spin_lock(&ctx
->lock
);
2619 perf_pmu_disable(ctx
->pmu
);
2621 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
2622 if (event
->state
!= PERF_EVENT_STATE_ACTIVE
)
2625 if (!event_filter_match(event
))
2630 if (needs_unthr
&& hwc
->interrupts
== MAX_INTERRUPTS
) {
2631 hwc
->interrupts
= 0;
2632 perf_log_throttle(event
, 1);
2633 event
->pmu
->start(event
, 0);
2636 if (!event
->attr
.freq
|| !event
->attr
.sample_freq
)
2640 * stop the event and update event->count
2642 event
->pmu
->stop(event
, PERF_EF_UPDATE
);
2644 now
= local64_read(&event
->count
);
2645 delta
= now
- hwc
->freq_count_stamp
;
2646 hwc
->freq_count_stamp
= now
;
2650 * reload only if value has changed
2651 * we have stopped the event so tell that
2652 * to perf_adjust_period() to avoid stopping it
2656 perf_adjust_period(event
, period
, delta
, false);
2658 event
->pmu
->start(event
, delta
> 0 ? PERF_EF_RELOAD
: 0);
2661 perf_pmu_enable(ctx
->pmu
);
2662 raw_spin_unlock(&ctx
->lock
);
2666 * Round-robin a context's events:
2668 static void rotate_ctx(struct perf_event_context
*ctx
)
2671 * Rotate the first entry last of non-pinned groups. Rotation might be
2672 * disabled by the inheritance code.
2674 if (!ctx
->rotate_disable
)
2675 list_rotate_left(&ctx
->flexible_groups
);
2679 * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
2680 * because they're strictly cpu affine and rotate_start is called with IRQs
2681 * disabled, while rotate_context is called from IRQ context.
2683 static void perf_rotate_context(struct perf_cpu_context
*cpuctx
)
2685 struct perf_event_context
*ctx
= NULL
;
2686 int rotate
= 0, remove
= 1;
2688 if (cpuctx
->ctx
.nr_events
) {
2690 if (cpuctx
->ctx
.nr_events
!= cpuctx
->ctx
.nr_active
)
2694 ctx
= cpuctx
->task_ctx
;
2695 if (ctx
&& ctx
->nr_events
) {
2697 if (ctx
->nr_events
!= ctx
->nr_active
)
2704 perf_ctx_lock(cpuctx
, cpuctx
->task_ctx
);
2705 perf_pmu_disable(cpuctx
->ctx
.pmu
);
2707 cpu_ctx_sched_out(cpuctx
, EVENT_FLEXIBLE
);
2709 ctx_sched_out(ctx
, cpuctx
, EVENT_FLEXIBLE
);
2711 rotate_ctx(&cpuctx
->ctx
);
2715 perf_event_sched_in(cpuctx
, ctx
, current
);
2717 perf_pmu_enable(cpuctx
->ctx
.pmu
);
2718 perf_ctx_unlock(cpuctx
, cpuctx
->task_ctx
);
2721 list_del_init(&cpuctx
->rotation_list
);
2724 #ifdef CONFIG_NO_HZ_FULL
2725 bool perf_event_can_stop_tick(void)
2727 if (list_empty(&__get_cpu_var(rotation_list
)))
2734 void perf_event_task_tick(void)
2736 struct list_head
*head
= &__get_cpu_var(rotation_list
);
2737 struct perf_cpu_context
*cpuctx
, *tmp
;
2738 struct perf_event_context
*ctx
;
2741 WARN_ON(!irqs_disabled());
2743 __this_cpu_inc(perf_throttled_seq
);
2744 throttled
= __this_cpu_xchg(perf_throttled_count
, 0);
2746 list_for_each_entry_safe(cpuctx
, tmp
, head
, rotation_list
) {
2748 perf_adjust_freq_unthr_context(ctx
, throttled
);
2750 ctx
= cpuctx
->task_ctx
;
2752 perf_adjust_freq_unthr_context(ctx
, throttled
);
2754 if (cpuctx
->jiffies_interval
== 1 ||
2755 !(jiffies
% cpuctx
->jiffies_interval
))
2756 perf_rotate_context(cpuctx
);
2760 static int event_enable_on_exec(struct perf_event
*event
,
2761 struct perf_event_context
*ctx
)
2763 if (!event
->attr
.enable_on_exec
)
2766 event
->attr
.enable_on_exec
= 0;
2767 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
)
2770 __perf_event_mark_enabled(event
);
2776 * Enable all of a task's events that have been marked enable-on-exec.
2777 * This expects task == current.
2779 static void perf_event_enable_on_exec(struct perf_event_context
*ctx
)
2781 struct perf_event
*event
;
2782 unsigned long flags
;
2786 local_irq_save(flags
);
2787 if (!ctx
|| !ctx
->nr_events
)
2791 * We must ctxsw out cgroup events to avoid conflict
2792 * when invoking perf_task_event_sched_in() later on
2793 * in this function. Otherwise we end up trying to
2794 * ctxswin cgroup events which are already scheduled
2797 perf_cgroup_sched_out(current
, NULL
);
2799 raw_spin_lock(&ctx
->lock
);
2800 task_ctx_sched_out(ctx
);
2802 list_for_each_entry(event
, &ctx
->event_list
, event_entry
) {
2803 ret
= event_enable_on_exec(event
, ctx
);
2809 * Unclone this context if we enabled any event.
2814 raw_spin_unlock(&ctx
->lock
);
2817 * Also calls ctxswin for cgroup events, if any:
2819 perf_event_context_sched_in(ctx
, ctx
->task
);
2821 local_irq_restore(flags
);
2825 * Cross CPU call to read the hardware event
2827 static void __perf_event_read(void *info
)
2829 struct perf_event
*event
= info
;
2830 struct perf_event_context
*ctx
= event
->ctx
;
2831 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
2834 * If this is a task context, we need to check whether it is
2835 * the current task context of this cpu. If not it has been
2836 * scheduled out before the smp call arrived. In that case
2837 * event->count would have been updated to a recent sample
2838 * when the event was scheduled out.
2840 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
)
2843 raw_spin_lock(&ctx
->lock
);
2844 if (ctx
->is_active
) {
2845 update_context_time(ctx
);
2846 update_cgrp_time_from_event(event
);
2848 update_event_times(event
);
2849 if (event
->state
== PERF_EVENT_STATE_ACTIVE
)
2850 event
->pmu
->read(event
);
2851 raw_spin_unlock(&ctx
->lock
);
2854 static inline u64
perf_event_count(struct perf_event
*event
)
2856 return local64_read(&event
->count
) + atomic64_read(&event
->child_count
);
2859 static u64
perf_event_read(struct perf_event
*event
)
2862 * If event is enabled and currently active on a CPU, update the
2863 * value in the event structure:
2865 if (event
->state
== PERF_EVENT_STATE_ACTIVE
) {
2866 smp_call_function_single(event
->oncpu
,
2867 __perf_event_read
, event
, 1);
2868 } else if (event
->state
== PERF_EVENT_STATE_INACTIVE
) {
2869 struct perf_event_context
*ctx
= event
->ctx
;
2870 unsigned long flags
;
2872 raw_spin_lock_irqsave(&ctx
->lock
, flags
);
2874 * may read while context is not active
2875 * (e.g., thread is blocked), in that case
2876 * we cannot update context time
2878 if (ctx
->is_active
) {
2879 update_context_time(ctx
);
2880 update_cgrp_time_from_event(event
);
2882 update_event_times(event
);
2883 raw_spin_unlock_irqrestore(&ctx
->lock
, flags
);
2886 return perf_event_count(event
);
2890 * Initialize the perf_event context in a task_struct:
2892 static void __perf_event_init_context(struct perf_event_context
*ctx
)
2894 raw_spin_lock_init(&ctx
->lock
);
2895 mutex_init(&ctx
->mutex
);
2896 INIT_LIST_HEAD(&ctx
->pinned_groups
);
2897 INIT_LIST_HEAD(&ctx
->flexible_groups
);
2898 INIT_LIST_HEAD(&ctx
->event_list
);
2899 atomic_set(&ctx
->refcount
, 1);
2902 static struct perf_event_context
*
2903 alloc_perf_context(struct pmu
*pmu
, struct task_struct
*task
)
2905 struct perf_event_context
*ctx
;
2907 ctx
= kzalloc(sizeof(struct perf_event_context
), GFP_KERNEL
);
2911 __perf_event_init_context(ctx
);
2914 get_task_struct(task
);
2921 static struct task_struct
*
2922 find_lively_task_by_vpid(pid_t vpid
)
2924 struct task_struct
*task
;
2931 task
= find_task_by_vpid(vpid
);
2933 get_task_struct(task
);
2937 return ERR_PTR(-ESRCH
);
2939 /* Reuse ptrace permission checks for now. */
2941 if (!ptrace_may_access(task
, PTRACE_MODE_READ
))
2946 put_task_struct(task
);
2947 return ERR_PTR(err
);
2952 * Returns a matching context with refcount and pincount.
2954 static struct perf_event_context
*
2955 find_get_context(struct pmu
*pmu
, struct task_struct
*task
, int cpu
)
2957 struct perf_event_context
*ctx
;
2958 struct perf_cpu_context
*cpuctx
;
2959 unsigned long flags
;
2963 /* Must be root to operate on a CPU event: */
2964 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN
))
2965 return ERR_PTR(-EACCES
);
2968 * We could be clever and allow to attach a event to an
2969 * offline CPU and activate it when the CPU comes up, but
2972 if (!cpu_online(cpu
))
2973 return ERR_PTR(-ENODEV
);
2975 cpuctx
= per_cpu_ptr(pmu
->pmu_cpu_context
, cpu
);
2984 ctxn
= pmu
->task_ctx_nr
;
2989 ctx
= perf_lock_task_context(task
, ctxn
, &flags
);
2993 raw_spin_unlock_irqrestore(&ctx
->lock
, flags
);
2995 ctx
= alloc_perf_context(pmu
, task
);
3001 mutex_lock(&task
->perf_event_mutex
);
3003 * If it has already passed perf_event_exit_task().
3004 * we must see PF_EXITING, it takes this mutex too.
3006 if (task
->flags
& PF_EXITING
)
3008 else if (task
->perf_event_ctxp
[ctxn
])
3013 rcu_assign_pointer(task
->perf_event_ctxp
[ctxn
], ctx
);
3015 mutex_unlock(&task
->perf_event_mutex
);
3017 if (unlikely(err
)) {
3029 return ERR_PTR(err
);
3032 static void perf_event_free_filter(struct perf_event
*event
);
3034 static void free_event_rcu(struct rcu_head
*head
)
3036 struct perf_event
*event
;
3038 event
= container_of(head
, struct perf_event
, rcu_head
);
3040 put_pid_ns(event
->ns
);
3041 perf_event_free_filter(event
);
3045 static void ring_buffer_put(struct ring_buffer
*rb
);
3046 static void ring_buffer_detach(struct perf_event
*event
, struct ring_buffer
*rb
);
3048 static void free_event(struct perf_event
*event
)
3050 irq_work_sync(&event
->pending
);
3052 if (!event
->parent
) {
3053 if (event
->attach_state
& PERF_ATTACH_TASK
)
3054 static_key_slow_dec_deferred(&perf_sched_events
);
3055 if (event
->attr
.mmap
|| event
->attr
.mmap_data
)
3056 atomic_dec(&nr_mmap_events
);
3057 if (event
->attr
.comm
)
3058 atomic_dec(&nr_comm_events
);
3059 if (event
->attr
.task
)
3060 atomic_dec(&nr_task_events
);
3061 if (event
->attr
.sample_type
& PERF_SAMPLE_CALLCHAIN
)
3062 put_callchain_buffers();
3063 if (is_cgroup_event(event
)) {
3064 atomic_dec(&per_cpu(perf_cgroup_events
, event
->cpu
));
3065 static_key_slow_dec_deferred(&perf_sched_events
);
3068 if (has_branch_stack(event
)) {
3069 static_key_slow_dec_deferred(&perf_sched_events
);
3070 /* is system-wide event */
3071 if (!(event
->attach_state
& PERF_ATTACH_TASK
)) {
3072 atomic_dec(&per_cpu(perf_branch_stack_events
,
3079 struct ring_buffer
*rb
;
3082 * Can happen when we close an event with re-directed output.
3084 * Since we have a 0 refcount, perf_mmap_close() will skip
3085 * over us; possibly making our ring_buffer_put() the last.
3087 mutex_lock(&event
->mmap_mutex
);
3090 rcu_assign_pointer(event
->rb
, NULL
);
3091 ring_buffer_detach(event
, rb
);
3092 ring_buffer_put(rb
); /* could be last */
3094 mutex_unlock(&event
->mmap_mutex
);
3097 if (is_cgroup_event(event
))
3098 perf_detach_cgroup(event
);
3101 event
->destroy(event
);
3104 put_ctx(event
->ctx
);
3106 call_rcu(&event
->rcu_head
, free_event_rcu
);
3109 int perf_event_release_kernel(struct perf_event
*event
)
3111 struct perf_event_context
*ctx
= event
->ctx
;
3113 WARN_ON_ONCE(ctx
->parent_ctx
);
3115 * There are two ways this annotation is useful:
3117 * 1) there is a lock recursion from perf_event_exit_task
3118 * see the comment there.
3120 * 2) there is a lock-inversion with mmap_sem through
3121 * perf_event_read_group(), which takes faults while
3122 * holding ctx->mutex, however this is called after
3123 * the last filedesc died, so there is no possibility
3124 * to trigger the AB-BA case.
3126 mutex_lock_nested(&ctx
->mutex
, SINGLE_DEPTH_NESTING
);
3127 perf_remove_from_context(event
, true);
3128 mutex_unlock(&ctx
->mutex
);
3134 EXPORT_SYMBOL_GPL(perf_event_release_kernel
);
3137 * Called when the last reference to the file is gone.
3139 static void put_event(struct perf_event
*event
)
3141 struct task_struct
*owner
;
3143 if (!atomic_long_dec_and_test(&event
->refcount
))
3147 owner
= ACCESS_ONCE(event
->owner
);
3149 * Matches the smp_wmb() in perf_event_exit_task(). If we observe
3150 * !owner it means the list deletion is complete and we can indeed
3151 * free this event, otherwise we need to serialize on
3152 * owner->perf_event_mutex.
3154 smp_read_barrier_depends();
3157 * Since delayed_put_task_struct() also drops the last
3158 * task reference we can safely take a new reference
3159 * while holding the rcu_read_lock().
3161 get_task_struct(owner
);
3166 mutex_lock(&owner
->perf_event_mutex
);
3168 * We have to re-check the event->owner field, if it is cleared
3169 * we raced with perf_event_exit_task(), acquiring the mutex
3170 * ensured they're done, and we can proceed with freeing the
3174 list_del_init(&event
->owner_entry
);
3175 mutex_unlock(&owner
->perf_event_mutex
);
3176 put_task_struct(owner
);
3179 perf_event_release_kernel(event
);
3182 static int perf_release(struct inode
*inode
, struct file
*file
)
3184 put_event(file
->private_data
);
3188 u64
perf_event_read_value(struct perf_event
*event
, u64
*enabled
, u64
*running
)
3190 struct perf_event
*child
;
3196 mutex_lock(&event
->child_mutex
);
3197 total
+= perf_event_read(event
);
3198 *enabled
+= event
->total_time_enabled
+
3199 atomic64_read(&event
->child_total_time_enabled
);
3200 *running
+= event
->total_time_running
+
3201 atomic64_read(&event
->child_total_time_running
);
3203 list_for_each_entry(child
, &event
->child_list
, child_list
) {
3204 total
+= perf_event_read(child
);
3205 *enabled
+= child
->total_time_enabled
;
3206 *running
+= child
->total_time_running
;
3208 mutex_unlock(&event
->child_mutex
);
3212 EXPORT_SYMBOL_GPL(perf_event_read_value
);
3214 static int perf_event_read_group(struct perf_event
*event
,
3215 u64 read_format
, char __user
*buf
)
3217 struct perf_event
*leader
= event
->group_leader
, *sub
;
3218 int n
= 0, size
= 0, ret
= -EFAULT
;
3219 struct perf_event_context
*ctx
= leader
->ctx
;
3221 u64 count
, enabled
, running
;
3223 mutex_lock(&ctx
->mutex
);
3224 count
= perf_event_read_value(leader
, &enabled
, &running
);
3226 values
[n
++] = 1 + leader
->nr_siblings
;
3227 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
3228 values
[n
++] = enabled
;
3229 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
3230 values
[n
++] = running
;
3231 values
[n
++] = count
;
3232 if (read_format
& PERF_FORMAT_ID
)
3233 values
[n
++] = primary_event_id(leader
);
3235 size
= n
* sizeof(u64
);
3237 if (copy_to_user(buf
, values
, size
))
3242 list_for_each_entry(sub
, &leader
->sibling_list
, group_entry
) {
3245 values
[n
++] = perf_event_read_value(sub
, &enabled
, &running
);
3246 if (read_format
& PERF_FORMAT_ID
)
3247 values
[n
++] = primary_event_id(sub
);
3249 size
= n
* sizeof(u64
);
3251 if (copy_to_user(buf
+ ret
, values
, size
)) {
3259 mutex_unlock(&ctx
->mutex
);
3264 static int perf_event_read_one(struct perf_event
*event
,
3265 u64 read_format
, char __user
*buf
)
3267 u64 enabled
, running
;
3271 values
[n
++] = perf_event_read_value(event
, &enabled
, &running
);
3272 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
3273 values
[n
++] = enabled
;
3274 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
3275 values
[n
++] = running
;
3276 if (read_format
& PERF_FORMAT_ID
)
3277 values
[n
++] = primary_event_id(event
);
3279 if (copy_to_user(buf
, values
, n
* sizeof(u64
)))
3282 return n
* sizeof(u64
);
3286 * Read the performance event - simple non blocking version for now
3289 perf_read_hw(struct perf_event
*event
, char __user
*buf
, size_t count
)
3291 u64 read_format
= event
->attr
.read_format
;
3295 * Return end-of-file for a read on a event that is in
3296 * error state (i.e. because it was pinned but it couldn't be
3297 * scheduled on to the CPU at some point).
3299 if (event
->state
== PERF_EVENT_STATE_ERROR
)
3302 if (count
< event
->read_size
)
3305 WARN_ON_ONCE(event
->ctx
->parent_ctx
);
3306 if (read_format
& PERF_FORMAT_GROUP
)
3307 ret
= perf_event_read_group(event
, read_format
, buf
);
3309 ret
= perf_event_read_one(event
, read_format
, buf
);
3315 perf_read(struct file
*file
, char __user
*buf
, size_t count
, loff_t
*ppos
)
3317 struct perf_event
*event
= file
->private_data
;
3319 return perf_read_hw(event
, buf
, count
);
3322 static unsigned int perf_poll(struct file
*file
, poll_table
*wait
)
3324 struct perf_event
*event
= file
->private_data
;
3325 struct ring_buffer
*rb
;
3326 unsigned int events
= POLL_HUP
;
3329 * Pin the event->rb by taking event->mmap_mutex; otherwise
3330 * perf_event_set_output() can swizzle our rb and make us miss wakeups.
3332 mutex_lock(&event
->mmap_mutex
);
3335 events
= atomic_xchg(&rb
->poll
, 0);
3336 mutex_unlock(&event
->mmap_mutex
);
3338 poll_wait(file
, &event
->waitq
, wait
);
3343 static void perf_event_reset(struct perf_event
*event
)
3345 (void)perf_event_read(event
);
3346 local64_set(&event
->count
, 0);
3347 perf_event_update_userpage(event
);
3351 * Holding the top-level event's child_mutex means that any
3352 * descendant process that has inherited this event will block
3353 * in sync_child_event if it goes to exit, thus satisfying the
3354 * task existence requirements of perf_event_enable/disable.
3356 static void perf_event_for_each_child(struct perf_event
*event
,
3357 void (*func
)(struct perf_event
*))
3359 struct perf_event
*child
;
3361 WARN_ON_ONCE(event
->ctx
->parent_ctx
);
3362 mutex_lock(&event
->child_mutex
);
3364 list_for_each_entry(child
, &event
->child_list
, child_list
)
3366 mutex_unlock(&event
->child_mutex
);
3369 static void perf_event_for_each(struct perf_event
*event
,
3370 void (*func
)(struct perf_event
*))
3372 struct perf_event_context
*ctx
= event
->ctx
;
3373 struct perf_event
*sibling
;
3375 WARN_ON_ONCE(ctx
->parent_ctx
);
3376 mutex_lock(&ctx
->mutex
);
3377 event
= event
->group_leader
;
3379 perf_event_for_each_child(event
, func
);
3380 list_for_each_entry(sibling
, &event
->sibling_list
, group_entry
)
3381 perf_event_for_each_child(sibling
, func
);
3382 mutex_unlock(&ctx
->mutex
);
3385 static int perf_event_period(struct perf_event
*event
, u64 __user
*arg
)
3387 struct perf_event_context
*ctx
= event
->ctx
;
3391 if (!is_sampling_event(event
))
3394 if (copy_from_user(&value
, arg
, sizeof(value
)))
3400 raw_spin_lock_irq(&ctx
->lock
);
3401 if (event
->attr
.freq
) {
3402 if (value
> sysctl_perf_event_sample_rate
) {
3407 event
->attr
.sample_freq
= value
;
3409 event
->attr
.sample_period
= value
;
3410 event
->hw
.sample_period
= value
;
3413 raw_spin_unlock_irq(&ctx
->lock
);
3418 static const struct file_operations perf_fops
;
3420 static inline int perf_fget_light(int fd
, struct fd
*p
)
3422 struct fd f
= fdget(fd
);
3426 if (f
.file
->f_op
!= &perf_fops
) {
3434 static int perf_event_set_output(struct perf_event
*event
,
3435 struct perf_event
*output_event
);
3436 static int perf_event_set_filter(struct perf_event
*event
, void __user
*arg
);
3438 static long perf_ioctl(struct file
*file
, unsigned int cmd
, unsigned long arg
)
3440 struct perf_event
*event
= file
->private_data
;
3441 void (*func
)(struct perf_event
*);
3445 case PERF_EVENT_IOC_ENABLE
:
3446 func
= perf_event_enable
;
3448 case PERF_EVENT_IOC_DISABLE
:
3449 func
= perf_event_disable
;
3451 case PERF_EVENT_IOC_RESET
:
3452 func
= perf_event_reset
;
3455 case PERF_EVENT_IOC_REFRESH
:
3456 return perf_event_refresh(event
, arg
);
3458 case PERF_EVENT_IOC_PERIOD
:
3459 return perf_event_period(event
, (u64 __user
*)arg
);
3461 case PERF_EVENT_IOC_SET_OUTPUT
:
3465 struct perf_event
*output_event
;
3467 ret
= perf_fget_light(arg
, &output
);
3470 output_event
= output
.file
->private_data
;
3471 ret
= perf_event_set_output(event
, output_event
);
3474 ret
= perf_event_set_output(event
, NULL
);
3479 case PERF_EVENT_IOC_SET_FILTER
:
3480 return perf_event_set_filter(event
, (void __user
*)arg
);
3486 if (flags
& PERF_IOC_FLAG_GROUP
)
3487 perf_event_for_each(event
, func
);
3489 perf_event_for_each_child(event
, func
);
3494 #ifdef CONFIG_COMPAT
3495 static long perf_compat_ioctl(struct file
*file
, unsigned int cmd
,
3498 switch (_IOC_NR(cmd
)) {
3499 case _IOC_NR(PERF_EVENT_IOC_SET_FILTER
):
3500 /* Fix up pointer size (usually 4 -> 8 in 32-on-64-bit case */
3501 if (_IOC_SIZE(cmd
) == sizeof(compat_uptr_t
)) {
3502 cmd
&= ~IOCSIZE_MASK
;
3503 cmd
|= sizeof(void *) << IOCSIZE_SHIFT
;
3507 return perf_ioctl(file
, cmd
, arg
);
3510 # define perf_compat_ioctl NULL
3513 int perf_event_task_enable(void)
3515 struct perf_event
*event
;
3517 mutex_lock(¤t
->perf_event_mutex
);
3518 list_for_each_entry(event
, ¤t
->perf_event_list
, owner_entry
)
3519 perf_event_for_each_child(event
, perf_event_enable
);
3520 mutex_unlock(¤t
->perf_event_mutex
);
3525 int perf_event_task_disable(void)
3527 struct perf_event
*event
;
3529 mutex_lock(¤t
->perf_event_mutex
);
3530 list_for_each_entry(event
, ¤t
->perf_event_list
, owner_entry
)
3531 perf_event_for_each_child(event
, perf_event_disable
);
3532 mutex_unlock(¤t
->perf_event_mutex
);
3537 static int perf_event_index(struct perf_event
*event
)
3539 if (event
->hw
.state
& PERF_HES_STOPPED
)
3542 if (event
->state
!= PERF_EVENT_STATE_ACTIVE
)
3545 return event
->pmu
->event_idx(event
);
3548 static void calc_timer_values(struct perf_event
*event
,
3555 *now
= perf_clock();
3556 ctx_time
= event
->shadow_ctx_time
+ *now
;
3557 *enabled
= ctx_time
- event
->tstamp_enabled
;
3558 *running
= ctx_time
- event
->tstamp_running
;
3561 void __weak
arch_perf_update_userpage(struct perf_event_mmap_page
*userpg
, u64 now
)
3566 * Callers need to ensure there can be no nesting of this function, otherwise
3567 * the seqlock logic goes bad. We can not serialize this because the arch
3568 * code calls this from NMI context.
3570 void perf_event_update_userpage(struct perf_event
*event
)
3572 struct perf_event_mmap_page
*userpg
;
3573 struct ring_buffer
*rb
;
3574 u64 enabled
, running
, now
;
3578 * compute total_time_enabled, total_time_running
3579 * based on snapshot values taken when the event
3580 * was last scheduled in.
3582 * we cannot simply called update_context_time()
3583 * because of locking issue as we can be called in
3586 calc_timer_values(event
, &now
, &enabled
, &running
);
3587 rb
= rcu_dereference(event
->rb
);
3591 userpg
= rb
->user_page
;
3594 * Disable preemption so as to not let the corresponding user-space
3595 * spin too long if we get preempted.
3600 userpg
->index
= perf_event_index(event
);
3601 userpg
->offset
= perf_event_count(event
);
3603 userpg
->offset
-= local64_read(&event
->hw
.prev_count
);
3605 userpg
->time_enabled
= enabled
+
3606 atomic64_read(&event
->child_total_time_enabled
);
3608 userpg
->time_running
= running
+
3609 atomic64_read(&event
->child_total_time_running
);
3611 arch_perf_update_userpage(userpg
, now
);
3620 static int perf_mmap_fault(struct vm_area_struct
*vma
, struct vm_fault
*vmf
)
3622 struct perf_event
*event
= vma
->vm_file
->private_data
;
3623 struct ring_buffer
*rb
;
3624 int ret
= VM_FAULT_SIGBUS
;
3626 if (vmf
->flags
& FAULT_FLAG_MKWRITE
) {
3627 if (vmf
->pgoff
== 0)
3633 rb
= rcu_dereference(event
->rb
);
3637 if (vmf
->pgoff
&& (vmf
->flags
& FAULT_FLAG_WRITE
))
3640 vmf
->page
= perf_mmap_to_page(rb
, vmf
->pgoff
);
3644 get_page(vmf
->page
);
3645 vmf
->page
->mapping
= vma
->vm_file
->f_mapping
;
3646 vmf
->page
->index
= vmf
->pgoff
;
3655 static void ring_buffer_attach(struct perf_event
*event
,
3656 struct ring_buffer
*rb
)
3658 unsigned long flags
;
3660 if (!list_empty(&event
->rb_entry
))
3663 spin_lock_irqsave(&rb
->event_lock
, flags
);
3664 if (list_empty(&event
->rb_entry
))
3665 list_add(&event
->rb_entry
, &rb
->event_list
);
3666 spin_unlock_irqrestore(&rb
->event_lock
, flags
);
3669 static void ring_buffer_detach(struct perf_event
*event
, struct ring_buffer
*rb
)
3671 unsigned long flags
;
3673 if (list_empty(&event
->rb_entry
))
3676 spin_lock_irqsave(&rb
->event_lock
, flags
);
3677 list_del_init(&event
->rb_entry
);
3678 wake_up_all(&event
->waitq
);
3679 spin_unlock_irqrestore(&rb
->event_lock
, flags
);
3682 static void ring_buffer_wakeup(struct perf_event
*event
)
3684 struct ring_buffer
*rb
;
3687 rb
= rcu_dereference(event
->rb
);
3689 list_for_each_entry_rcu(event
, &rb
->event_list
, rb_entry
)
3690 wake_up_all(&event
->waitq
);
3695 static void rb_free_rcu(struct rcu_head
*rcu_head
)
3697 struct ring_buffer
*rb
;
3699 rb
= container_of(rcu_head
, struct ring_buffer
, rcu_head
);
3703 static struct ring_buffer
*ring_buffer_get(struct perf_event
*event
)
3705 struct ring_buffer
*rb
;
3708 rb
= rcu_dereference(event
->rb
);
3710 if (!atomic_inc_not_zero(&rb
->refcount
))
3718 static void ring_buffer_put(struct ring_buffer
*rb
)
3720 if (!atomic_dec_and_test(&rb
->refcount
))
3723 WARN_ON_ONCE(!list_empty(&rb
->event_list
));
3725 call_rcu(&rb
->rcu_head
, rb_free_rcu
);
3728 static void perf_mmap_open(struct vm_area_struct
*vma
)
3730 struct perf_event
*event
= vma
->vm_file
->private_data
;
3732 atomic_inc(&event
->mmap_count
);
3733 atomic_inc(&event
->rb
->mmap_count
);
3737 * A buffer can be mmap()ed multiple times; either directly through the same
3738 * event, or through other events by use of perf_event_set_output().
3740 * In order to undo the VM accounting done by perf_mmap() we need to destroy
3741 * the buffer here, where we still have a VM context. This means we need
3742 * to detach all events redirecting to us.
3744 static void perf_mmap_close(struct vm_area_struct
*vma
)
3746 struct perf_event
*event
= vma
->vm_file
->private_data
;
3748 struct ring_buffer
*rb
= event
->rb
;
3749 struct user_struct
*mmap_user
= rb
->mmap_user
;
3750 int mmap_locked
= rb
->mmap_locked
;
3751 unsigned long size
= perf_data_size(rb
);
3753 atomic_dec(&rb
->mmap_count
);
3755 if (!atomic_dec_and_mutex_lock(&event
->mmap_count
, &event
->mmap_mutex
))
3758 /* Detach current event from the buffer. */
3759 rcu_assign_pointer(event
->rb
, NULL
);
3760 ring_buffer_detach(event
, rb
);
3761 mutex_unlock(&event
->mmap_mutex
);
3763 /* If there's still other mmap()s of this buffer, we're done. */
3764 if (atomic_read(&rb
->mmap_count
)) {
3765 ring_buffer_put(rb
); /* can't be last */
3770 * No other mmap()s, detach from all other events that might redirect
3771 * into the now unreachable buffer. Somewhat complicated by the
3772 * fact that rb::event_lock otherwise nests inside mmap_mutex.
3776 list_for_each_entry_rcu(event
, &rb
->event_list
, rb_entry
) {
3777 if (!atomic_long_inc_not_zero(&event
->refcount
)) {
3779 * This event is en-route to free_event() which will
3780 * detach it and remove it from the list.
3786 mutex_lock(&event
->mmap_mutex
);
3788 * Check we didn't race with perf_event_set_output() which can
3789 * swizzle the rb from under us while we were waiting to
3790 * acquire mmap_mutex.
3792 * If we find a different rb; ignore this event, a next
3793 * iteration will no longer find it on the list. We have to
3794 * still restart the iteration to make sure we're not now
3795 * iterating the wrong list.
3797 if (event
->rb
== rb
) {
3798 rcu_assign_pointer(event
->rb
, NULL
);
3799 ring_buffer_detach(event
, rb
);
3800 ring_buffer_put(rb
); /* can't be last, we still have one */
3802 mutex_unlock(&event
->mmap_mutex
);
3806 * Restart the iteration; either we're on the wrong list or
3807 * destroyed its integrity by doing a deletion.
3814 * It could be there's still a few 0-ref events on the list; they'll
3815 * get cleaned up by free_event() -- they'll also still have their
3816 * ref on the rb and will free it whenever they are done with it.
3818 * Aside from that, this buffer is 'fully' detached and unmapped,
3819 * undo the VM accounting.
3822 atomic_long_sub((size
>> PAGE_SHIFT
) + 1, &mmap_user
->locked_vm
);
3823 vma
->vm_mm
->pinned_vm
-= mmap_locked
;
3824 free_uid(mmap_user
);
3826 ring_buffer_put(rb
); /* could be last */
3829 static const struct vm_operations_struct perf_mmap_vmops
= {
3830 .open
= perf_mmap_open
,
3831 .close
= perf_mmap_close
,
3832 .fault
= perf_mmap_fault
,
3833 .page_mkwrite
= perf_mmap_fault
,
3836 static int perf_mmap(struct file
*file
, struct vm_area_struct
*vma
)
3838 struct perf_event
*event
= file
->private_data
;
3839 unsigned long user_locked
, user_lock_limit
;
3840 struct user_struct
*user
= current_user();
3841 unsigned long locked
, lock_limit
;
3842 struct ring_buffer
*rb
;
3843 unsigned long vma_size
;
3844 unsigned long nr_pages
;
3845 long user_extra
, extra
;
3846 int ret
= 0, flags
= 0;
3849 * Don't allow mmap() of inherited per-task counters. This would
3850 * create a performance issue due to all children writing to the
3853 if (event
->cpu
== -1 && event
->attr
.inherit
)
3856 if (!(vma
->vm_flags
& VM_SHARED
))
3859 vma_size
= vma
->vm_end
- vma
->vm_start
;
3860 nr_pages
= (vma_size
/ PAGE_SIZE
) - 1;
3863 * If we have rb pages ensure they're a power-of-two number, so we
3864 * can do bitmasks instead of modulo.
3866 if (nr_pages
!= 0 && !is_power_of_2(nr_pages
))
3869 if (vma_size
!= PAGE_SIZE
* (1 + nr_pages
))
3872 if (vma
->vm_pgoff
!= 0)
3875 WARN_ON_ONCE(event
->ctx
->parent_ctx
);
3877 mutex_lock(&event
->mmap_mutex
);
3879 if (event
->rb
->nr_pages
!= nr_pages
) {
3884 if (!atomic_inc_not_zero(&event
->rb
->mmap_count
)) {
3886 * Raced against perf_mmap_close() through
3887 * perf_event_set_output(). Try again, hope for better
3890 mutex_unlock(&event
->mmap_mutex
);
3897 user_extra
= nr_pages
+ 1;
3898 user_lock_limit
= sysctl_perf_event_mlock
>> (PAGE_SHIFT
- 10);
3901 * Increase the limit linearly with more CPUs:
3903 user_lock_limit
*= num_online_cpus();
3905 user_locked
= atomic_long_read(&user
->locked_vm
) + user_extra
;
3908 if (user_locked
> user_lock_limit
)
3909 extra
= user_locked
- user_lock_limit
;
3911 lock_limit
= rlimit(RLIMIT_MEMLOCK
);
3912 lock_limit
>>= PAGE_SHIFT
;
3913 locked
= vma
->vm_mm
->pinned_vm
+ extra
;
3915 if ((locked
> lock_limit
) && perf_paranoid_tracepoint_raw() &&
3916 !capable(CAP_IPC_LOCK
)) {
3923 if (vma
->vm_flags
& VM_WRITE
)
3924 flags
|= RING_BUFFER_WRITABLE
;
3926 rb
= rb_alloc(nr_pages
,
3927 event
->attr
.watermark
? event
->attr
.wakeup_watermark
: 0,
3935 atomic_set(&rb
->mmap_count
, 1);
3936 rb
->mmap_locked
= extra
;
3937 rb
->mmap_user
= get_current_user();
3939 atomic_long_add(user_extra
, &user
->locked_vm
);
3940 vma
->vm_mm
->pinned_vm
+= extra
;
3942 ring_buffer_attach(event
, rb
);
3943 rcu_assign_pointer(event
->rb
, rb
);
3945 perf_event_update_userpage(event
);
3949 atomic_inc(&event
->mmap_count
);
3950 mutex_unlock(&event
->mmap_mutex
);
3953 * Since pinned accounting is per vm we cannot allow fork() to copy our
3956 vma
->vm_flags
|= VM_DONTCOPY
| VM_DONTEXPAND
| VM_DONTDUMP
;
3957 vma
->vm_ops
= &perf_mmap_vmops
;
3962 static int perf_fasync(int fd
, struct file
*filp
, int on
)
3964 struct inode
*inode
= file_inode(filp
);
3965 struct perf_event
*event
= filp
->private_data
;
3968 mutex_lock(&inode
->i_mutex
);
3969 retval
= fasync_helper(fd
, filp
, on
, &event
->fasync
);
3970 mutex_unlock(&inode
->i_mutex
);
3978 static const struct file_operations perf_fops
= {
3979 .llseek
= no_llseek
,
3980 .release
= perf_release
,
3983 .unlocked_ioctl
= perf_ioctl
,
3984 .compat_ioctl
= perf_compat_ioctl
,
3986 .fasync
= perf_fasync
,
3992 * If there's data, ensure we set the poll() state and publish everything
3993 * to user-space before waking everybody up.
3996 void perf_event_wakeup(struct perf_event
*event
)
3998 ring_buffer_wakeup(event
);
4000 if (event
->pending_kill
) {
4001 kill_fasync(&event
->fasync
, SIGIO
, event
->pending_kill
);
4002 event
->pending_kill
= 0;
4006 static void perf_pending_event(struct irq_work
*entry
)
4008 struct perf_event
*event
= container_of(entry
,
4009 struct perf_event
, pending
);
4012 rctx
= perf_swevent_get_recursion_context();
4014 * If we 'fail' here, that's OK, it means recursion is already disabled
4015 * and we won't recurse 'further'.
4018 if (event
->pending_disable
) {
4019 event
->pending_disable
= 0;
4020 __perf_event_disable(event
);
4023 if (event
->pending_wakeup
) {
4024 event
->pending_wakeup
= 0;
4025 perf_event_wakeup(event
);
4029 perf_swevent_put_recursion_context(rctx
);
4033 * We assume there is only KVM supporting the callbacks.
4034 * Later on, we might change it to a list if there is
4035 * another virtualization implementation supporting the callbacks.
4037 struct perf_guest_info_callbacks
*perf_guest_cbs
;
4039 int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks
*cbs
)
4041 perf_guest_cbs
= cbs
;
4044 EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks
);
4046 int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks
*cbs
)
4048 perf_guest_cbs
= NULL
;
4051 EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks
);
4054 perf_output_sample_regs(struct perf_output_handle
*handle
,
4055 struct pt_regs
*regs
, u64 mask
)
4059 for_each_set_bit(bit
, (const unsigned long *) &mask
,
4060 sizeof(mask
) * BITS_PER_BYTE
) {
4063 val
= perf_reg_value(regs
, bit
);
4064 perf_output_put(handle
, val
);
4068 static void perf_sample_regs_user(struct perf_regs_user
*regs_user
,
4069 struct pt_regs
*regs
)
4071 if (!user_mode(regs
)) {
4073 regs
= task_pt_regs(current
);
4079 regs_user
->regs
= regs
;
4080 regs_user
->abi
= perf_reg_abi(current
);
4085 * Get remaining task size from user stack pointer.
4087 * It'd be better to take stack vma map and limit this more
4088 * precisly, but there's no way to get it safely under interrupt,
4089 * so using TASK_SIZE as limit.
4091 static u64
perf_ustack_task_size(struct pt_regs
*regs
)
4093 unsigned long addr
= perf_user_stack_pointer(regs
);
4095 if (!addr
|| addr
>= TASK_SIZE
)
4098 return TASK_SIZE
- addr
;
4102 perf_sample_ustack_size(u16 stack_size
, u16 header_size
,
4103 struct pt_regs
*regs
)
4107 /* No regs, no stack pointer, no dump. */
4112 * Check if we fit in with the requested stack size into the:
4114 * If we don't, we limit the size to the TASK_SIZE.
4116 * - remaining sample size
4117 * If we don't, we customize the stack size to
4118 * fit in to the remaining sample size.
4121 task_size
= min((u64
) USHRT_MAX
, perf_ustack_task_size(regs
));
4122 stack_size
= min(stack_size
, (u16
) task_size
);
4124 /* Current header size plus static size and dynamic size. */
4125 header_size
+= 2 * sizeof(u64
);
4127 /* Do we fit in with the current stack dump size? */
4128 if ((u16
) (header_size
+ stack_size
) < header_size
) {
4130 * If we overflow the maximum size for the sample,
4131 * we customize the stack dump size to fit in.
4133 stack_size
= USHRT_MAX
- header_size
- sizeof(u64
);
4134 stack_size
= round_up(stack_size
, sizeof(u64
));
4141 perf_output_sample_ustack(struct perf_output_handle
*handle
, u64 dump_size
,
4142 struct pt_regs
*regs
)
4144 /* Case of a kernel thread, nothing to dump */
4147 perf_output_put(handle
, size
);
4156 * - the size requested by user or the best one we can fit
4157 * in to the sample max size
4159 * - user stack dump data
4161 * - the actual dumped size
4165 perf_output_put(handle
, dump_size
);
4168 sp
= perf_user_stack_pointer(regs
);
4169 rem
= __output_copy_user(handle
, (void *) sp
, dump_size
);
4170 dyn_size
= dump_size
- rem
;
4172 perf_output_skip(handle
, rem
);
4175 perf_output_put(handle
, dyn_size
);
4179 static void __perf_event_header__init_id(struct perf_event_header
*header
,
4180 struct perf_sample_data
*data
,
4181 struct perf_event
*event
)
4183 u64 sample_type
= event
->attr
.sample_type
;
4185 data
->type
= sample_type
;
4186 header
->size
+= event
->id_header_size
;
4188 if (sample_type
& PERF_SAMPLE_TID
) {
4189 /* namespace issues */
4190 data
->tid_entry
.pid
= perf_event_pid(event
, current
);
4191 data
->tid_entry
.tid
= perf_event_tid(event
, current
);
4194 if (sample_type
& PERF_SAMPLE_TIME
)
4195 data
->time
= perf_clock();
4197 if (sample_type
& PERF_SAMPLE_ID
)
4198 data
->id
= primary_event_id(event
);
4200 if (sample_type
& PERF_SAMPLE_STREAM_ID
)
4201 data
->stream_id
= event
->id
;
4203 if (sample_type
& PERF_SAMPLE_CPU
) {
4204 data
->cpu_entry
.cpu
= raw_smp_processor_id();
4205 data
->cpu_entry
.reserved
= 0;
4209 void perf_event_header__init_id(struct perf_event_header
*header
,
4210 struct perf_sample_data
*data
,
4211 struct perf_event
*event
)
4213 if (event
->attr
.sample_id_all
)
4214 __perf_event_header__init_id(header
, data
, event
);
4217 static void __perf_event__output_id_sample(struct perf_output_handle
*handle
,
4218 struct perf_sample_data
*data
)
4220 u64 sample_type
= data
->type
;
4222 if (sample_type
& PERF_SAMPLE_TID
)
4223 perf_output_put(handle
, data
->tid_entry
);
4225 if (sample_type
& PERF_SAMPLE_TIME
)
4226 perf_output_put(handle
, data
->time
);
4228 if (sample_type
& PERF_SAMPLE_ID
)
4229 perf_output_put(handle
, data
->id
);
4231 if (sample_type
& PERF_SAMPLE_STREAM_ID
)
4232 perf_output_put(handle
, data
->stream_id
);
4234 if (sample_type
& PERF_SAMPLE_CPU
)
4235 perf_output_put(handle
, data
->cpu_entry
);
4238 void perf_event__output_id_sample(struct perf_event
*event
,
4239 struct perf_output_handle
*handle
,
4240 struct perf_sample_data
*sample
)
4242 if (event
->attr
.sample_id_all
)
4243 __perf_event__output_id_sample(handle
, sample
);
4246 static void perf_output_read_one(struct perf_output_handle
*handle
,
4247 struct perf_event
*event
,
4248 u64 enabled
, u64 running
)
4250 u64 read_format
= event
->attr
.read_format
;
4254 values
[n
++] = perf_event_count(event
);
4255 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
) {
4256 values
[n
++] = enabled
+
4257 atomic64_read(&event
->child_total_time_enabled
);
4259 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
) {
4260 values
[n
++] = running
+
4261 atomic64_read(&event
->child_total_time_running
);
4263 if (read_format
& PERF_FORMAT_ID
)
4264 values
[n
++] = primary_event_id(event
);
4266 __output_copy(handle
, values
, n
* sizeof(u64
));
4270 * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
4272 static void perf_output_read_group(struct perf_output_handle
*handle
,
4273 struct perf_event
*event
,
4274 u64 enabled
, u64 running
)
4276 struct perf_event
*leader
= event
->group_leader
, *sub
;
4277 u64 read_format
= event
->attr
.read_format
;
4281 values
[n
++] = 1 + leader
->nr_siblings
;
4283 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
4284 values
[n
++] = enabled
;
4286 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
4287 values
[n
++] = running
;
4289 if (leader
!= event
)
4290 leader
->pmu
->read(leader
);
4292 values
[n
++] = perf_event_count(leader
);
4293 if (read_format
& PERF_FORMAT_ID
)
4294 values
[n
++] = primary_event_id(leader
);
4296 __output_copy(handle
, values
, n
* sizeof(u64
));
4298 list_for_each_entry(sub
, &leader
->sibling_list
, group_entry
) {
4302 sub
->pmu
->read(sub
);
4304 values
[n
++] = perf_event_count(sub
);
4305 if (read_format
& PERF_FORMAT_ID
)
4306 values
[n
++] = primary_event_id(sub
);
4308 __output_copy(handle
, values
, n
* sizeof(u64
));
4312 #define PERF_FORMAT_TOTAL_TIMES (PERF_FORMAT_TOTAL_TIME_ENABLED|\
4313 PERF_FORMAT_TOTAL_TIME_RUNNING)
4315 static void perf_output_read(struct perf_output_handle
*handle
,
4316 struct perf_event
*event
)
4318 u64 enabled
= 0, running
= 0, now
;
4319 u64 read_format
= event
->attr
.read_format
;
4322 * compute total_time_enabled, total_time_running
4323 * based on snapshot values taken when the event
4324 * was last scheduled in.
4326 * we cannot simply called update_context_time()
4327 * because of locking issue as we are called in
4330 if (read_format
& PERF_FORMAT_TOTAL_TIMES
)
4331 calc_timer_values(event
, &now
, &enabled
, &running
);
4333 if (event
->attr
.read_format
& PERF_FORMAT_GROUP
)
4334 perf_output_read_group(handle
, event
, enabled
, running
);
4336 perf_output_read_one(handle
, event
, enabled
, running
);
4339 void perf_output_sample(struct perf_output_handle
*handle
,
4340 struct perf_event_header
*header
,
4341 struct perf_sample_data
*data
,
4342 struct perf_event
*event
)
4344 u64 sample_type
= data
->type
;
4346 perf_output_put(handle
, *header
);
4348 if (sample_type
& PERF_SAMPLE_IP
)
4349 perf_output_put(handle
, data
->ip
);
4351 if (sample_type
& PERF_SAMPLE_TID
)
4352 perf_output_put(handle
, data
->tid_entry
);
4354 if (sample_type
& PERF_SAMPLE_TIME
)
4355 perf_output_put(handle
, data
->time
);
4357 if (sample_type
& PERF_SAMPLE_ADDR
)
4358 perf_output_put(handle
, data
->addr
);
4360 if (sample_type
& PERF_SAMPLE_ID
)
4361 perf_output_put(handle
, data
->id
);
4363 if (sample_type
& PERF_SAMPLE_STREAM_ID
)
4364 perf_output_put(handle
, data
->stream_id
);
4366 if (sample_type
& PERF_SAMPLE_CPU
)
4367 perf_output_put(handle
, data
->cpu_entry
);
4369 if (sample_type
& PERF_SAMPLE_PERIOD
)
4370 perf_output_put(handle
, data
->period
);
4372 if (sample_type
& PERF_SAMPLE_READ
)
4373 perf_output_read(handle
, event
);
4375 if (sample_type
& PERF_SAMPLE_CALLCHAIN
) {
4376 if (data
->callchain
) {
4379 if (data
->callchain
)
4380 size
+= data
->callchain
->nr
;
4382 size
*= sizeof(u64
);
4384 __output_copy(handle
, data
->callchain
, size
);
4387 perf_output_put(handle
, nr
);
4391 if (sample_type
& PERF_SAMPLE_RAW
) {
4393 perf_output_put(handle
, data
->raw
->size
);
4394 __output_copy(handle
, data
->raw
->data
,
4401 .size
= sizeof(u32
),
4404 perf_output_put(handle
, raw
);
4408 if (!event
->attr
.watermark
) {
4409 int wakeup_events
= event
->attr
.wakeup_events
;
4411 if (wakeup_events
) {
4412 struct ring_buffer
*rb
= handle
->rb
;
4413 int events
= local_inc_return(&rb
->events
);
4415 if (events
>= wakeup_events
) {
4416 local_sub(wakeup_events
, &rb
->events
);
4417 local_inc(&rb
->wakeup
);
4422 if (sample_type
& PERF_SAMPLE_BRANCH_STACK
) {
4423 if (data
->br_stack
) {
4426 size
= data
->br_stack
->nr
4427 * sizeof(struct perf_branch_entry
);
4429 perf_output_put(handle
, data
->br_stack
->nr
);
4430 perf_output_copy(handle
, data
->br_stack
->entries
, size
);
4433 * we always store at least the value of nr
4436 perf_output_put(handle
, nr
);
4440 if (sample_type
& PERF_SAMPLE_REGS_USER
) {
4441 u64 abi
= data
->regs_user
.abi
;
4444 * If there are no regs to dump, notice it through
4445 * first u64 being zero (PERF_SAMPLE_REGS_ABI_NONE).
4447 perf_output_put(handle
, abi
);
4450 u64 mask
= event
->attr
.sample_regs_user
;
4451 perf_output_sample_regs(handle
,
4452 data
->regs_user
.regs
,
4457 if (sample_type
& PERF_SAMPLE_STACK_USER
)
4458 perf_output_sample_ustack(handle
,
4459 data
->stack_user_size
,
4460 data
->regs_user
.regs
);
4462 if (sample_type
& PERF_SAMPLE_WEIGHT
)
4463 perf_output_put(handle
, data
->weight
);
4465 if (sample_type
& PERF_SAMPLE_DATA_SRC
)
4466 perf_output_put(handle
, data
->data_src
.val
);
4469 void perf_prepare_sample(struct perf_event_header
*header
,
4470 struct perf_sample_data
*data
,
4471 struct perf_event
*event
,
4472 struct pt_regs
*regs
)
4474 u64 sample_type
= event
->attr
.sample_type
;
4476 header
->type
= PERF_RECORD_SAMPLE
;
4477 header
->size
= sizeof(*header
) + event
->header_size
;
4480 header
->misc
|= perf_misc_flags(regs
);
4482 __perf_event_header__init_id(header
, data
, event
);
4484 if (sample_type
& PERF_SAMPLE_IP
)
4485 data
->ip
= perf_instruction_pointer(regs
);
4487 if (sample_type
& PERF_SAMPLE_CALLCHAIN
) {
4490 data
->callchain
= perf_callchain(event
, regs
);
4492 if (data
->callchain
)
4493 size
+= data
->callchain
->nr
;
4495 header
->size
+= size
* sizeof(u64
);
4498 if (sample_type
& PERF_SAMPLE_RAW
) {
4499 int size
= sizeof(u32
);
4502 size
+= data
->raw
->size
;
4504 size
+= sizeof(u32
);
4506 WARN_ON_ONCE(size
& (sizeof(u64
)-1));
4507 header
->size
+= size
;
4510 if (sample_type
& PERF_SAMPLE_BRANCH_STACK
) {
4511 int size
= sizeof(u64
); /* nr */
4512 if (data
->br_stack
) {
4513 size
+= data
->br_stack
->nr
4514 * sizeof(struct perf_branch_entry
);
4516 header
->size
+= size
;
4519 if (sample_type
& PERF_SAMPLE_REGS_USER
) {
4520 /* regs dump ABI info */
4521 int size
= sizeof(u64
);
4523 perf_sample_regs_user(&data
->regs_user
, regs
);
4525 if (data
->regs_user
.regs
) {
4526 u64 mask
= event
->attr
.sample_regs_user
;
4527 size
+= hweight64(mask
) * sizeof(u64
);
4530 header
->size
+= size
;
4533 if (sample_type
& PERF_SAMPLE_STACK_USER
) {
4535 * Either we need PERF_SAMPLE_STACK_USER bit to be allways
4536 * processed as the last one or have additional check added
4537 * in case new sample type is added, because we could eat
4538 * up the rest of the sample size.
4540 struct perf_regs_user
*uregs
= &data
->regs_user
;
4541 u16 stack_size
= event
->attr
.sample_stack_user
;
4542 u16 size
= sizeof(u64
);
4545 perf_sample_regs_user(uregs
, regs
);
4547 stack_size
= perf_sample_ustack_size(stack_size
, header
->size
,
4551 * If there is something to dump, add space for the dump
4552 * itself and for the field that tells the dynamic size,
4553 * which is how many have been actually dumped.
4556 size
+= sizeof(u64
) + stack_size
;
4558 data
->stack_user_size
= stack_size
;
4559 header
->size
+= size
;
4563 static void perf_event_output(struct perf_event
*event
,
4564 struct perf_sample_data
*data
,
4565 struct pt_regs
*regs
)
4567 struct perf_output_handle handle
;
4568 struct perf_event_header header
;
4570 /* protect the callchain buffers */
4573 perf_prepare_sample(&header
, data
, event
, regs
);
4575 if (perf_output_begin(&handle
, event
, header
.size
))
4578 perf_output_sample(&handle
, &header
, data
, event
);
4580 perf_output_end(&handle
);
4590 struct perf_read_event
{
4591 struct perf_event_header header
;
4598 perf_event_read_event(struct perf_event
*event
,
4599 struct task_struct
*task
)
4601 struct perf_output_handle handle
;
4602 struct perf_sample_data sample
;
4603 struct perf_read_event read_event
= {
4605 .type
= PERF_RECORD_READ
,
4607 .size
= sizeof(read_event
) + event
->read_size
,
4609 .pid
= perf_event_pid(event
, task
),
4610 .tid
= perf_event_tid(event
, task
),
4614 perf_event_header__init_id(&read_event
.header
, &sample
, event
);
4615 ret
= perf_output_begin(&handle
, event
, read_event
.header
.size
);
4619 perf_output_put(&handle
, read_event
);
4620 perf_output_read(&handle
, event
);
4621 perf_event__output_id_sample(event
, &handle
, &sample
);
4623 perf_output_end(&handle
);
4626 typedef int (perf_event_aux_match_cb
)(struct perf_event
*event
, void *data
);
4627 typedef void (perf_event_aux_output_cb
)(struct perf_event
*event
, void *data
);
4630 perf_event_aux_ctx(struct perf_event_context
*ctx
,
4631 perf_event_aux_match_cb match
,
4632 perf_event_aux_output_cb output
,
4635 struct perf_event
*event
;
4637 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
4638 if (event
->state
< PERF_EVENT_STATE_INACTIVE
)
4640 if (!event_filter_match(event
))
4642 if (match(event
, data
))
4643 output(event
, data
);
4648 perf_event_aux(perf_event_aux_match_cb match
,
4649 perf_event_aux_output_cb output
,
4651 struct perf_event_context
*task_ctx
)
4653 struct perf_cpu_context
*cpuctx
;
4654 struct perf_event_context
*ctx
;
4659 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
4660 cpuctx
= get_cpu_ptr(pmu
->pmu_cpu_context
);
4661 if (cpuctx
->unique_pmu
!= pmu
)
4663 perf_event_aux_ctx(&cpuctx
->ctx
, match
, output
, data
);
4666 ctxn
= pmu
->task_ctx_nr
;
4669 ctx
= rcu_dereference(current
->perf_event_ctxp
[ctxn
]);
4671 perf_event_aux_ctx(ctx
, match
, output
, data
);
4673 put_cpu_ptr(pmu
->pmu_cpu_context
);
4678 perf_event_aux_ctx(task_ctx
, match
, output
, data
);
4685 * task tracking -- fork/exit
4687 * enabled by: attr.comm | attr.mmap | attr.mmap_data | attr.task
4690 struct perf_task_event
{
4691 struct task_struct
*task
;
4692 struct perf_event_context
*task_ctx
;
4695 struct perf_event_header header
;
4705 static void perf_event_task_output(struct perf_event
*event
,
4708 struct perf_task_event
*task_event
= data
;
4709 struct perf_output_handle handle
;
4710 struct perf_sample_data sample
;
4711 struct task_struct
*task
= task_event
->task
;
4712 int ret
, size
= task_event
->event_id
.header
.size
;
4714 perf_event_header__init_id(&task_event
->event_id
.header
, &sample
, event
);
4716 ret
= perf_output_begin(&handle
, event
,
4717 task_event
->event_id
.header
.size
);
4721 task_event
->event_id
.pid
= perf_event_pid(event
, task
);
4722 task_event
->event_id
.ppid
= perf_event_pid(event
, current
);
4724 task_event
->event_id
.tid
= perf_event_tid(event
, task
);
4725 task_event
->event_id
.ptid
= perf_event_tid(event
, current
);
4727 perf_output_put(&handle
, task_event
->event_id
);
4729 perf_event__output_id_sample(event
, &handle
, &sample
);
4731 perf_output_end(&handle
);
4733 task_event
->event_id
.header
.size
= size
;
4736 static int perf_event_task_match(struct perf_event
*event
,
4737 void *data __maybe_unused
)
4739 return event
->attr
.comm
|| event
->attr
.mmap
||
4740 event
->attr
.mmap_data
|| event
->attr
.task
;
4743 static void perf_event_task(struct task_struct
*task
,
4744 struct perf_event_context
*task_ctx
,
4747 struct perf_task_event task_event
;
4749 if (!atomic_read(&nr_comm_events
) &&
4750 !atomic_read(&nr_mmap_events
) &&
4751 !atomic_read(&nr_task_events
))
4754 task_event
= (struct perf_task_event
){
4756 .task_ctx
= task_ctx
,
4759 .type
= new ? PERF_RECORD_FORK
: PERF_RECORD_EXIT
,
4761 .size
= sizeof(task_event
.event_id
),
4767 .time
= perf_clock(),
4771 perf_event_aux(perf_event_task_match
,
4772 perf_event_task_output
,
4777 void perf_event_fork(struct task_struct
*task
)
4779 perf_event_task(task
, NULL
, 1);
4786 struct perf_comm_event
{
4787 struct task_struct
*task
;
4792 struct perf_event_header header
;
4799 static void perf_event_comm_output(struct perf_event
*event
,
4802 struct perf_comm_event
*comm_event
= data
;
4803 struct perf_output_handle handle
;
4804 struct perf_sample_data sample
;
4805 int size
= comm_event
->event_id
.header
.size
;
4808 perf_event_header__init_id(&comm_event
->event_id
.header
, &sample
, event
);
4809 ret
= perf_output_begin(&handle
, event
,
4810 comm_event
->event_id
.header
.size
);
4815 comm_event
->event_id
.pid
= perf_event_pid(event
, comm_event
->task
);
4816 comm_event
->event_id
.tid
= perf_event_tid(event
, comm_event
->task
);
4818 perf_output_put(&handle
, comm_event
->event_id
);
4819 __output_copy(&handle
, comm_event
->comm
,
4820 comm_event
->comm_size
);
4822 perf_event__output_id_sample(event
, &handle
, &sample
);
4824 perf_output_end(&handle
);
4826 comm_event
->event_id
.header
.size
= size
;
4829 static int perf_event_comm_match(struct perf_event
*event
,
4830 void *data __maybe_unused
)
4832 return event
->attr
.comm
;
4835 static void perf_event_comm_event(struct perf_comm_event
*comm_event
)
4837 char comm
[TASK_COMM_LEN
];
4840 memset(comm
, 0, sizeof(comm
));
4841 strlcpy(comm
, comm_event
->task
->comm
, sizeof(comm
));
4842 size
= ALIGN(strlen(comm
)+1, sizeof(u64
));
4844 comm_event
->comm
= comm
;
4845 comm_event
->comm_size
= size
;
4847 comm_event
->event_id
.header
.size
= sizeof(comm_event
->event_id
) + size
;
4849 perf_event_aux(perf_event_comm_match
,
4850 perf_event_comm_output
,
4855 void perf_event_comm(struct task_struct
*task
)
4857 struct perf_comm_event comm_event
;
4858 struct perf_event_context
*ctx
;
4862 for_each_task_context_nr(ctxn
) {
4863 ctx
= task
->perf_event_ctxp
[ctxn
];
4867 perf_event_enable_on_exec(ctx
);
4871 if (!atomic_read(&nr_comm_events
))
4874 comm_event
= (struct perf_comm_event
){
4880 .type
= PERF_RECORD_COMM
,
4889 perf_event_comm_event(&comm_event
);
4896 struct perf_mmap_event
{
4897 struct vm_area_struct
*vma
;
4899 const char *file_name
;
4903 struct perf_event_header header
;
4913 static void perf_event_mmap_output(struct perf_event
*event
,
4916 struct perf_mmap_event
*mmap_event
= data
;
4917 struct perf_output_handle handle
;
4918 struct perf_sample_data sample
;
4919 int size
= mmap_event
->event_id
.header
.size
;
4922 perf_event_header__init_id(&mmap_event
->event_id
.header
, &sample
, event
);
4923 ret
= perf_output_begin(&handle
, event
,
4924 mmap_event
->event_id
.header
.size
);
4928 mmap_event
->event_id
.pid
= perf_event_pid(event
, current
);
4929 mmap_event
->event_id
.tid
= perf_event_tid(event
, current
);
4931 perf_output_put(&handle
, mmap_event
->event_id
);
4932 __output_copy(&handle
, mmap_event
->file_name
,
4933 mmap_event
->file_size
);
4935 perf_event__output_id_sample(event
, &handle
, &sample
);
4937 perf_output_end(&handle
);
4939 mmap_event
->event_id
.header
.size
= size
;
4942 static int perf_event_mmap_match(struct perf_event
*event
,
4945 struct perf_mmap_event
*mmap_event
= data
;
4946 struct vm_area_struct
*vma
= mmap_event
->vma
;
4947 int executable
= vma
->vm_flags
& VM_EXEC
;
4949 return (!executable
&& event
->attr
.mmap_data
) ||
4950 (executable
&& event
->attr
.mmap
);
4953 static void perf_event_mmap_event(struct perf_mmap_event
*mmap_event
)
4955 struct vm_area_struct
*vma
= mmap_event
->vma
;
4956 struct file
*file
= vma
->vm_file
;
4962 memset(tmp
, 0, sizeof(tmp
));
4966 * d_path works from the end of the rb backwards, so we
4967 * need to add enough zero bytes after the string to handle
4968 * the 64bit alignment we do later.
4970 buf
= kzalloc(PATH_MAX
+ sizeof(u64
), GFP_KERNEL
);
4972 name
= strncpy(tmp
, "//enomem", sizeof(tmp
));
4975 name
= d_path(&file
->f_path
, buf
, PATH_MAX
);
4977 name
= strncpy(tmp
, "//toolong", sizeof(tmp
));
4981 if (arch_vma_name(mmap_event
->vma
)) {
4982 name
= strncpy(tmp
, arch_vma_name(mmap_event
->vma
),
4984 tmp
[sizeof(tmp
) - 1] = '\0';
4989 name
= strncpy(tmp
, "[vdso]", sizeof(tmp
));
4991 } else if (vma
->vm_start
<= vma
->vm_mm
->start_brk
&&
4992 vma
->vm_end
>= vma
->vm_mm
->brk
) {
4993 name
= strncpy(tmp
, "[heap]", sizeof(tmp
));
4995 } else if (vma
->vm_start
<= vma
->vm_mm
->start_stack
&&
4996 vma
->vm_end
>= vma
->vm_mm
->start_stack
) {
4997 name
= strncpy(tmp
, "[stack]", sizeof(tmp
));
5001 name
= strncpy(tmp
, "//anon", sizeof(tmp
));
5006 size
= ALIGN(strlen(name
)+1, sizeof(u64
));
5008 mmap_event
->file_name
= name
;
5009 mmap_event
->file_size
= size
;
5011 if (!(vma
->vm_flags
& VM_EXEC
))
5012 mmap_event
->event_id
.header
.misc
|= PERF_RECORD_MISC_MMAP_DATA
;
5014 mmap_event
->event_id
.header
.size
= sizeof(mmap_event
->event_id
) + size
;
5016 perf_event_aux(perf_event_mmap_match
,
5017 perf_event_mmap_output
,
5024 void perf_event_mmap(struct vm_area_struct
*vma
)
5026 struct perf_mmap_event mmap_event
;
5028 if (!atomic_read(&nr_mmap_events
))
5031 mmap_event
= (struct perf_mmap_event
){
5037 .type
= PERF_RECORD_MMAP
,
5038 .misc
= PERF_RECORD_MISC_USER
,
5043 .start
= vma
->vm_start
,
5044 .len
= vma
->vm_end
- vma
->vm_start
,
5045 .pgoff
= (u64
)vma
->vm_pgoff
<< PAGE_SHIFT
,
5049 perf_event_mmap_event(&mmap_event
);
5053 * IRQ throttle logging
5056 static void perf_log_throttle(struct perf_event
*event
, int enable
)
5058 struct perf_output_handle handle
;
5059 struct perf_sample_data sample
;
5063 struct perf_event_header header
;
5067 } throttle_event
= {
5069 .type
= PERF_RECORD_THROTTLE
,
5071 .size
= sizeof(throttle_event
),
5073 .time
= perf_clock(),
5074 .id
= primary_event_id(event
),
5075 .stream_id
= event
->id
,
5079 throttle_event
.header
.type
= PERF_RECORD_UNTHROTTLE
;
5081 perf_event_header__init_id(&throttle_event
.header
, &sample
, event
);
5083 ret
= perf_output_begin(&handle
, event
,
5084 throttle_event
.header
.size
);
5088 perf_output_put(&handle
, throttle_event
);
5089 perf_event__output_id_sample(event
, &handle
, &sample
);
5090 perf_output_end(&handle
);
5094 * Generic event overflow handling, sampling.
5097 static int __perf_event_overflow(struct perf_event
*event
,
5098 int throttle
, struct perf_sample_data
*data
,
5099 struct pt_regs
*regs
)
5101 int events
= atomic_read(&event
->event_limit
);
5102 struct hw_perf_event
*hwc
= &event
->hw
;
5107 * Non-sampling counters might still use the PMI to fold short
5108 * hardware counters, ignore those.
5110 if (unlikely(!is_sampling_event(event
)))
5113 seq
= __this_cpu_read(perf_throttled_seq
);
5114 if (seq
!= hwc
->interrupts_seq
) {
5115 hwc
->interrupts_seq
= seq
;
5116 hwc
->interrupts
= 1;
5119 if (unlikely(throttle
5120 && hwc
->interrupts
>= max_samples_per_tick
)) {
5121 __this_cpu_inc(perf_throttled_count
);
5122 hwc
->interrupts
= MAX_INTERRUPTS
;
5123 perf_log_throttle(event
, 0);
5128 if (event
->attr
.freq
) {
5129 u64 now
= perf_clock();
5130 s64 delta
= now
- hwc
->freq_time_stamp
;
5132 hwc
->freq_time_stamp
= now
;
5134 if (delta
> 0 && delta
< 2*TICK_NSEC
)
5135 perf_adjust_period(event
, delta
, hwc
->last_period
, true);
5139 * XXX event_limit might not quite work as expected on inherited
5143 event
->pending_kill
= POLL_IN
;
5144 if (events
&& atomic_dec_and_test(&event
->event_limit
)) {
5146 event
->pending_kill
= POLL_HUP
;
5147 event
->pending_disable
= 1;
5148 irq_work_queue(&event
->pending
);
5151 if (event
->overflow_handler
)
5152 event
->overflow_handler(event
, data
, regs
);
5154 perf_event_output(event
, data
, regs
);
5156 if (event
->fasync
&& event
->pending_kill
) {
5157 event
->pending_wakeup
= 1;
5158 irq_work_queue(&event
->pending
);
5164 int perf_event_overflow(struct perf_event
*event
,
5165 struct perf_sample_data
*data
,
5166 struct pt_regs
*regs
)
5168 return __perf_event_overflow(event
, 1, data
, regs
);
5172 * Generic software event infrastructure
5175 struct swevent_htable
{
5176 struct swevent_hlist
*swevent_hlist
;
5177 struct mutex hlist_mutex
;
5180 /* Recursion avoidance in each contexts */
5181 int recursion
[PERF_NR_CONTEXTS
];
5183 /* Keeps track of cpu being initialized/exited */
5187 static DEFINE_PER_CPU(struct swevent_htable
, swevent_htable
);
5190 * We directly increment event->count and keep a second value in
5191 * event->hw.period_left to count intervals. This period event
5192 * is kept in the range [-sample_period, 0] so that we can use the
5196 static u64
perf_swevent_set_period(struct perf_event
*event
)
5198 struct hw_perf_event
*hwc
= &event
->hw
;
5199 u64 period
= hwc
->last_period
;
5203 hwc
->last_period
= hwc
->sample_period
;
5206 old
= val
= local64_read(&hwc
->period_left
);
5210 nr
= div64_u64(period
+ val
, period
);
5211 offset
= nr
* period
;
5213 if (local64_cmpxchg(&hwc
->period_left
, old
, val
) != old
)
5219 static void perf_swevent_overflow(struct perf_event
*event
, u64 overflow
,
5220 struct perf_sample_data
*data
,
5221 struct pt_regs
*regs
)
5223 struct hw_perf_event
*hwc
= &event
->hw
;
5227 overflow
= perf_swevent_set_period(event
);
5229 if (hwc
->interrupts
== MAX_INTERRUPTS
)
5232 for (; overflow
; overflow
--) {
5233 if (__perf_event_overflow(event
, throttle
,
5236 * We inhibit the overflow from happening when
5237 * hwc->interrupts == MAX_INTERRUPTS.
5245 static void perf_swevent_event(struct perf_event
*event
, u64 nr
,
5246 struct perf_sample_data
*data
,
5247 struct pt_regs
*regs
)
5249 struct hw_perf_event
*hwc
= &event
->hw
;
5251 local64_add(nr
, &event
->count
);
5256 if (!is_sampling_event(event
))
5259 if ((event
->attr
.sample_type
& PERF_SAMPLE_PERIOD
) && !event
->attr
.freq
) {
5261 return perf_swevent_overflow(event
, 1, data
, regs
);
5263 data
->period
= event
->hw
.last_period
;
5265 if (nr
== 1 && hwc
->sample_period
== 1 && !event
->attr
.freq
)
5266 return perf_swevent_overflow(event
, 1, data
, regs
);
5268 if (local64_add_negative(nr
, &hwc
->period_left
))
5271 perf_swevent_overflow(event
, 0, data
, regs
);
5274 static int perf_exclude_event(struct perf_event
*event
,
5275 struct pt_regs
*regs
)
5277 if (event
->hw
.state
& PERF_HES_STOPPED
)
5281 if (event
->attr
.exclude_user
&& user_mode(regs
))
5284 if (event
->attr
.exclude_kernel
&& !user_mode(regs
))
5291 static int perf_swevent_match(struct perf_event
*event
,
5292 enum perf_type_id type
,
5294 struct perf_sample_data
*data
,
5295 struct pt_regs
*regs
)
5297 if (event
->attr
.type
!= type
)
5300 if (event
->attr
.config
!= event_id
)
5303 if (perf_exclude_event(event
, regs
))
5309 static inline u64
swevent_hash(u64 type
, u32 event_id
)
5311 u64 val
= event_id
| (type
<< 32);
5313 return hash_64(val
, SWEVENT_HLIST_BITS
);
5316 static inline struct hlist_head
*
5317 __find_swevent_head(struct swevent_hlist
*hlist
, u64 type
, u32 event_id
)
5319 u64 hash
= swevent_hash(type
, event_id
);
5321 return &hlist
->heads
[hash
];
5324 /* For the read side: events when they trigger */
5325 static inline struct hlist_head
*
5326 find_swevent_head_rcu(struct swevent_htable
*swhash
, u64 type
, u32 event_id
)
5328 struct swevent_hlist
*hlist
;
5330 hlist
= rcu_dereference(swhash
->swevent_hlist
);
5334 return __find_swevent_head(hlist
, type
, event_id
);
5337 /* For the event head insertion and removal in the hlist */
5338 static inline struct hlist_head
*
5339 find_swevent_head(struct swevent_htable
*swhash
, struct perf_event
*event
)
5341 struct swevent_hlist
*hlist
;
5342 u32 event_id
= event
->attr
.config
;
5343 u64 type
= event
->attr
.type
;
5346 * Event scheduling is always serialized against hlist allocation
5347 * and release. Which makes the protected version suitable here.
5348 * The context lock guarantees that.
5350 hlist
= rcu_dereference_protected(swhash
->swevent_hlist
,
5351 lockdep_is_held(&event
->ctx
->lock
));
5355 return __find_swevent_head(hlist
, type
, event_id
);
5358 static void do_perf_sw_event(enum perf_type_id type
, u32 event_id
,
5360 struct perf_sample_data
*data
,
5361 struct pt_regs
*regs
)
5363 struct swevent_htable
*swhash
= &__get_cpu_var(swevent_htable
);
5364 struct perf_event
*event
;
5365 struct hlist_head
*head
;
5368 head
= find_swevent_head_rcu(swhash
, type
, event_id
);
5372 hlist_for_each_entry_rcu(event
, head
, hlist_entry
) {
5373 if (perf_swevent_match(event
, type
, event_id
, data
, regs
))
5374 perf_swevent_event(event
, nr
, data
, regs
);
5380 int perf_swevent_get_recursion_context(void)
5382 struct swevent_htable
*swhash
= &__get_cpu_var(swevent_htable
);
5384 return get_recursion_context(swhash
->recursion
);
5386 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context
);
5388 inline void perf_swevent_put_recursion_context(int rctx
)
5390 struct swevent_htable
*swhash
= &__get_cpu_var(swevent_htable
);
5392 put_recursion_context(swhash
->recursion
, rctx
);
5395 void __perf_sw_event(u32 event_id
, u64 nr
, struct pt_regs
*regs
, u64 addr
)
5397 struct perf_sample_data data
;
5400 preempt_disable_notrace();
5401 rctx
= perf_swevent_get_recursion_context();
5405 perf_sample_data_init(&data
, addr
, 0);
5407 do_perf_sw_event(PERF_TYPE_SOFTWARE
, event_id
, nr
, &data
, regs
);
5409 perf_swevent_put_recursion_context(rctx
);
5410 preempt_enable_notrace();
5413 static void perf_swevent_read(struct perf_event
*event
)
5417 static int perf_swevent_add(struct perf_event
*event
, int flags
)
5419 struct swevent_htable
*swhash
= &__get_cpu_var(swevent_htable
);
5420 struct hw_perf_event
*hwc
= &event
->hw
;
5421 struct hlist_head
*head
;
5423 if (is_sampling_event(event
)) {
5424 hwc
->last_period
= hwc
->sample_period
;
5425 perf_swevent_set_period(event
);
5428 hwc
->state
= !(flags
& PERF_EF_START
);
5430 head
= find_swevent_head(swhash
, event
);
5433 * We can race with cpu hotplug code. Do not
5434 * WARN if the cpu just got unplugged.
5436 WARN_ON_ONCE(swhash
->online
);
5440 hlist_add_head_rcu(&event
->hlist_entry
, head
);
5445 static void perf_swevent_del(struct perf_event
*event
, int flags
)
5447 hlist_del_rcu(&event
->hlist_entry
);
5450 static void perf_swevent_start(struct perf_event
*event
, int flags
)
5452 event
->hw
.state
= 0;
5455 static void perf_swevent_stop(struct perf_event
*event
, int flags
)
5457 event
->hw
.state
= PERF_HES_STOPPED
;
5460 /* Deref the hlist from the update side */
5461 static inline struct swevent_hlist
*
5462 swevent_hlist_deref(struct swevent_htable
*swhash
)
5464 return rcu_dereference_protected(swhash
->swevent_hlist
,
5465 lockdep_is_held(&swhash
->hlist_mutex
));
5468 static void swevent_hlist_release(struct swevent_htable
*swhash
)
5470 struct swevent_hlist
*hlist
= swevent_hlist_deref(swhash
);
5475 rcu_assign_pointer(swhash
->swevent_hlist
, NULL
);
5476 kfree_rcu(hlist
, rcu_head
);
5479 static void swevent_hlist_put_cpu(struct perf_event
*event
, int cpu
)
5481 struct swevent_htable
*swhash
= &per_cpu(swevent_htable
, cpu
);
5483 mutex_lock(&swhash
->hlist_mutex
);
5485 if (!--swhash
->hlist_refcount
)
5486 swevent_hlist_release(swhash
);
5488 mutex_unlock(&swhash
->hlist_mutex
);
5491 static void swevent_hlist_put(struct perf_event
*event
)
5495 if (event
->cpu
!= -1) {
5496 swevent_hlist_put_cpu(event
, event
->cpu
);
5500 for_each_possible_cpu(cpu
)
5501 swevent_hlist_put_cpu(event
, cpu
);
5504 static int swevent_hlist_get_cpu(struct perf_event
*event
, int cpu
)
5506 struct swevent_htable
*swhash
= &per_cpu(swevent_htable
, cpu
);
5509 mutex_lock(&swhash
->hlist_mutex
);
5511 if (!swevent_hlist_deref(swhash
) && cpu_online(cpu
)) {
5512 struct swevent_hlist
*hlist
;
5514 hlist
= kzalloc(sizeof(*hlist
), GFP_KERNEL
);
5519 rcu_assign_pointer(swhash
->swevent_hlist
, hlist
);
5521 swhash
->hlist_refcount
++;
5523 mutex_unlock(&swhash
->hlist_mutex
);
5528 static int swevent_hlist_get(struct perf_event
*event
)
5531 int cpu
, failed_cpu
;
5533 if (event
->cpu
!= -1)
5534 return swevent_hlist_get_cpu(event
, event
->cpu
);
5537 for_each_possible_cpu(cpu
) {
5538 err
= swevent_hlist_get_cpu(event
, cpu
);
5548 for_each_possible_cpu(cpu
) {
5549 if (cpu
== failed_cpu
)
5551 swevent_hlist_put_cpu(event
, cpu
);
5558 struct static_key perf_swevent_enabled
[PERF_COUNT_SW_MAX
];
5560 static void sw_perf_event_destroy(struct perf_event
*event
)
5562 u64 event_id
= event
->attr
.config
;
5564 WARN_ON(event
->parent
);
5566 static_key_slow_dec(&perf_swevent_enabled
[event_id
]);
5567 swevent_hlist_put(event
);
5570 static int perf_swevent_init(struct perf_event
*event
)
5572 u64 event_id
= event
->attr
.config
;
5574 if (event
->attr
.type
!= PERF_TYPE_SOFTWARE
)
5578 * no branch sampling for software events
5580 if (has_branch_stack(event
))
5584 case PERF_COUNT_SW_CPU_CLOCK
:
5585 case PERF_COUNT_SW_TASK_CLOCK
:
5592 if (event_id
>= PERF_COUNT_SW_MAX
)
5595 if (!event
->parent
) {
5598 err
= swevent_hlist_get(event
);
5602 static_key_slow_inc(&perf_swevent_enabled
[event_id
]);
5603 event
->destroy
= sw_perf_event_destroy
;
5609 static int perf_swevent_event_idx(struct perf_event
*event
)
5614 static struct pmu perf_swevent
= {
5615 .task_ctx_nr
= perf_sw_context
,
5617 .event_init
= perf_swevent_init
,
5618 .add
= perf_swevent_add
,
5619 .del
= perf_swevent_del
,
5620 .start
= perf_swevent_start
,
5621 .stop
= perf_swevent_stop
,
5622 .read
= perf_swevent_read
,
5624 .event_idx
= perf_swevent_event_idx
,
5627 #ifdef CONFIG_EVENT_TRACING
5629 static int perf_tp_filter_match(struct perf_event
*event
,
5630 struct perf_sample_data
*data
)
5632 void *record
= data
->raw
->data
;
5634 if (likely(!event
->filter
) || filter_match_preds(event
->filter
, record
))
5639 static int perf_tp_event_match(struct perf_event
*event
,
5640 struct perf_sample_data
*data
,
5641 struct pt_regs
*regs
)
5643 if (event
->hw
.state
& PERF_HES_STOPPED
)
5646 * All tracepoints are from kernel-space.
5648 if (event
->attr
.exclude_kernel
)
5651 if (!perf_tp_filter_match(event
, data
))
5657 void perf_tp_event(u64 addr
, u64 count
, void *record
, int entry_size
,
5658 struct pt_regs
*regs
, struct hlist_head
*head
, int rctx
,
5659 struct task_struct
*task
)
5661 struct perf_sample_data data
;
5662 struct perf_event
*event
;
5664 struct perf_raw_record raw
= {
5669 perf_sample_data_init(&data
, addr
, 0);
5672 hlist_for_each_entry_rcu(event
, head
, hlist_entry
) {
5673 if (perf_tp_event_match(event
, &data
, regs
))
5674 perf_swevent_event(event
, count
, &data
, regs
);
5678 * If we got specified a target task, also iterate its context and
5679 * deliver this event there too.
5681 if (task
&& task
!= current
) {
5682 struct perf_event_context
*ctx
;
5683 struct trace_entry
*entry
= record
;
5686 ctx
= rcu_dereference(task
->perf_event_ctxp
[perf_sw_context
]);
5690 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
5691 if (event
->attr
.type
!= PERF_TYPE_TRACEPOINT
)
5693 if (event
->attr
.config
!= entry
->type
)
5695 if (perf_tp_event_match(event
, &data
, regs
))
5696 perf_swevent_event(event
, count
, &data
, regs
);
5702 perf_swevent_put_recursion_context(rctx
);
5704 EXPORT_SYMBOL_GPL(perf_tp_event
);
5706 static void tp_perf_event_destroy(struct perf_event
*event
)
5708 perf_trace_destroy(event
);
5711 static int perf_tp_event_init(struct perf_event
*event
)
5715 if (event
->attr
.type
!= PERF_TYPE_TRACEPOINT
)
5719 * no branch sampling for tracepoint events
5721 if (has_branch_stack(event
))
5724 err
= perf_trace_init(event
);
5728 event
->destroy
= tp_perf_event_destroy
;
5733 static struct pmu perf_tracepoint
= {
5734 .task_ctx_nr
= perf_sw_context
,
5736 .event_init
= perf_tp_event_init
,
5737 .add
= perf_trace_add
,
5738 .del
= perf_trace_del
,
5739 .start
= perf_swevent_start
,
5740 .stop
= perf_swevent_stop
,
5741 .read
= perf_swevent_read
,
5743 .event_idx
= perf_swevent_event_idx
,
5746 static inline void perf_tp_register(void)
5748 perf_pmu_register(&perf_tracepoint
, "tracepoint", PERF_TYPE_TRACEPOINT
);
5751 static int perf_event_set_filter(struct perf_event
*event
, void __user
*arg
)
5756 if (event
->attr
.type
!= PERF_TYPE_TRACEPOINT
)
5759 filter_str
= strndup_user(arg
, PAGE_SIZE
);
5760 if (IS_ERR(filter_str
))
5761 return PTR_ERR(filter_str
);
5763 ret
= ftrace_profile_set_filter(event
, event
->attr
.config
, filter_str
);
5769 static void perf_event_free_filter(struct perf_event
*event
)
5771 ftrace_profile_free_filter(event
);
5776 static inline void perf_tp_register(void)
5780 static int perf_event_set_filter(struct perf_event
*event
, void __user
*arg
)
5785 static void perf_event_free_filter(struct perf_event
*event
)
5789 #endif /* CONFIG_EVENT_TRACING */
5791 #ifdef CONFIG_HAVE_HW_BREAKPOINT
5792 void perf_bp_event(struct perf_event
*bp
, void *data
)
5794 struct perf_sample_data sample
;
5795 struct pt_regs
*regs
= data
;
5797 perf_sample_data_init(&sample
, bp
->attr
.bp_addr
, 0);
5799 if (!bp
->hw
.state
&& !perf_exclude_event(bp
, regs
))
5800 perf_swevent_event(bp
, 1, &sample
, regs
);
5805 * hrtimer based swevent callback
5808 static enum hrtimer_restart
perf_swevent_hrtimer(struct hrtimer
*hrtimer
)
5810 enum hrtimer_restart ret
= HRTIMER_RESTART
;
5811 struct perf_sample_data data
;
5812 struct pt_regs
*regs
;
5813 struct perf_event
*event
;
5816 event
= container_of(hrtimer
, struct perf_event
, hw
.hrtimer
);
5818 if (event
->state
!= PERF_EVENT_STATE_ACTIVE
)
5819 return HRTIMER_NORESTART
;
5821 event
->pmu
->read(event
);
5823 perf_sample_data_init(&data
, 0, event
->hw
.last_period
);
5824 regs
= get_irq_regs();
5826 if (regs
&& !perf_exclude_event(event
, regs
)) {
5827 if (!(event
->attr
.exclude_idle
&& is_idle_task(current
)))
5828 if (__perf_event_overflow(event
, 1, &data
, regs
))
5829 ret
= HRTIMER_NORESTART
;
5832 period
= max_t(u64
, 10000, event
->hw
.sample_period
);
5833 hrtimer_forward_now(hrtimer
, ns_to_ktime(period
));
5838 static void perf_swevent_start_hrtimer(struct perf_event
*event
)
5840 struct hw_perf_event
*hwc
= &event
->hw
;
5843 if (!is_sampling_event(event
))
5846 period
= local64_read(&hwc
->period_left
);
5851 local64_set(&hwc
->period_left
, 0);
5853 period
= max_t(u64
, 10000, hwc
->sample_period
);
5855 __hrtimer_start_range_ns(&hwc
->hrtimer
,
5856 ns_to_ktime(period
), 0,
5857 HRTIMER_MODE_REL_PINNED
, 0);
5860 static void perf_swevent_cancel_hrtimer(struct perf_event
*event
)
5862 struct hw_perf_event
*hwc
= &event
->hw
;
5864 if (is_sampling_event(event
)) {
5865 ktime_t remaining
= hrtimer_get_remaining(&hwc
->hrtimer
);
5866 local64_set(&hwc
->period_left
, ktime_to_ns(remaining
));
5868 hrtimer_cancel(&hwc
->hrtimer
);
5872 static void perf_swevent_init_hrtimer(struct perf_event
*event
)
5874 struct hw_perf_event
*hwc
= &event
->hw
;
5876 if (!is_sampling_event(event
))
5879 hrtimer_init(&hwc
->hrtimer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
5880 hwc
->hrtimer
.function
= perf_swevent_hrtimer
;
5883 * Since hrtimers have a fixed rate, we can do a static freq->period
5884 * mapping and avoid the whole period adjust feedback stuff.
5886 if (event
->attr
.freq
) {
5887 long freq
= event
->attr
.sample_freq
;
5889 event
->attr
.sample_period
= NSEC_PER_SEC
/ freq
;
5890 hwc
->sample_period
= event
->attr
.sample_period
;
5891 local64_set(&hwc
->period_left
, hwc
->sample_period
);
5892 hwc
->last_period
= hwc
->sample_period
;
5893 event
->attr
.freq
= 0;
5898 * Software event: cpu wall time clock
5901 static void cpu_clock_event_update(struct perf_event
*event
)
5906 now
= local_clock();
5907 prev
= local64_xchg(&event
->hw
.prev_count
, now
);
5908 local64_add(now
- prev
, &event
->count
);
5911 static void cpu_clock_event_start(struct perf_event
*event
, int flags
)
5913 local64_set(&event
->hw
.prev_count
, local_clock());
5914 perf_swevent_start_hrtimer(event
);
5917 static void cpu_clock_event_stop(struct perf_event
*event
, int flags
)
5919 perf_swevent_cancel_hrtimer(event
);
5920 cpu_clock_event_update(event
);
5923 static int cpu_clock_event_add(struct perf_event
*event
, int flags
)
5925 if (flags
& PERF_EF_START
)
5926 cpu_clock_event_start(event
, flags
);
5931 static void cpu_clock_event_del(struct perf_event
*event
, int flags
)
5933 cpu_clock_event_stop(event
, flags
);
5936 static void cpu_clock_event_read(struct perf_event
*event
)
5938 cpu_clock_event_update(event
);
5941 static int cpu_clock_event_init(struct perf_event
*event
)
5943 if (event
->attr
.type
!= PERF_TYPE_SOFTWARE
)
5946 if (event
->attr
.config
!= PERF_COUNT_SW_CPU_CLOCK
)
5950 * no branch sampling for software events
5952 if (has_branch_stack(event
))
5955 perf_swevent_init_hrtimer(event
);
5960 static struct pmu perf_cpu_clock
= {
5961 .task_ctx_nr
= perf_sw_context
,
5963 .event_init
= cpu_clock_event_init
,
5964 .add
= cpu_clock_event_add
,
5965 .del
= cpu_clock_event_del
,
5966 .start
= cpu_clock_event_start
,
5967 .stop
= cpu_clock_event_stop
,
5968 .read
= cpu_clock_event_read
,
5970 .event_idx
= perf_swevent_event_idx
,
5974 * Software event: task time clock
5977 static void task_clock_event_update(struct perf_event
*event
, u64 now
)
5982 prev
= local64_xchg(&event
->hw
.prev_count
, now
);
5984 local64_add(delta
, &event
->count
);
5987 static void task_clock_event_start(struct perf_event
*event
, int flags
)
5989 local64_set(&event
->hw
.prev_count
, event
->ctx
->time
);
5990 perf_swevent_start_hrtimer(event
);
5993 static void task_clock_event_stop(struct perf_event
*event
, int flags
)
5995 perf_swevent_cancel_hrtimer(event
);
5996 task_clock_event_update(event
, event
->ctx
->time
);
5999 static int task_clock_event_add(struct perf_event
*event
, int flags
)
6001 if (flags
& PERF_EF_START
)
6002 task_clock_event_start(event
, flags
);
6007 static void task_clock_event_del(struct perf_event
*event
, int flags
)
6009 task_clock_event_stop(event
, PERF_EF_UPDATE
);
6012 static void task_clock_event_read(struct perf_event
*event
)
6014 u64 now
= perf_clock();
6015 u64 delta
= now
- event
->ctx
->timestamp
;
6016 u64 time
= event
->ctx
->time
+ delta
;
6018 task_clock_event_update(event
, time
);
6021 static int task_clock_event_init(struct perf_event
*event
)
6023 if (event
->attr
.type
!= PERF_TYPE_SOFTWARE
)
6026 if (event
->attr
.config
!= PERF_COUNT_SW_TASK_CLOCK
)
6030 * no branch sampling for software events
6032 if (has_branch_stack(event
))
6035 perf_swevent_init_hrtimer(event
);
6040 static struct pmu perf_task_clock
= {
6041 .task_ctx_nr
= perf_sw_context
,
6043 .event_init
= task_clock_event_init
,
6044 .add
= task_clock_event_add
,
6045 .del
= task_clock_event_del
,
6046 .start
= task_clock_event_start
,
6047 .stop
= task_clock_event_stop
,
6048 .read
= task_clock_event_read
,
6050 .event_idx
= perf_swevent_event_idx
,
6053 static void perf_pmu_nop_void(struct pmu
*pmu
)
6057 static int perf_pmu_nop_int(struct pmu
*pmu
)
6062 static void perf_pmu_start_txn(struct pmu
*pmu
)
6064 perf_pmu_disable(pmu
);
6067 static int perf_pmu_commit_txn(struct pmu
*pmu
)
6069 perf_pmu_enable(pmu
);
6073 static void perf_pmu_cancel_txn(struct pmu
*pmu
)
6075 perf_pmu_enable(pmu
);
6078 static int perf_event_idx_default(struct perf_event
*event
)
6080 return event
->hw
.idx
+ 1;
6084 * Ensures all contexts with the same task_ctx_nr have the same
6085 * pmu_cpu_context too.
6087 static void *find_pmu_context(int ctxn
)
6094 list_for_each_entry(pmu
, &pmus
, entry
) {
6095 if (pmu
->task_ctx_nr
== ctxn
)
6096 return pmu
->pmu_cpu_context
;
6102 static void update_pmu_context(struct pmu
*pmu
, struct pmu
*old_pmu
)
6106 for_each_possible_cpu(cpu
) {
6107 struct perf_cpu_context
*cpuctx
;
6109 cpuctx
= per_cpu_ptr(pmu
->pmu_cpu_context
, cpu
);
6111 if (cpuctx
->unique_pmu
== old_pmu
)
6112 cpuctx
->unique_pmu
= pmu
;
6116 static void free_pmu_context(struct pmu
*pmu
)
6120 mutex_lock(&pmus_lock
);
6122 * Like a real lame refcount.
6124 list_for_each_entry(i
, &pmus
, entry
) {
6125 if (i
->pmu_cpu_context
== pmu
->pmu_cpu_context
) {
6126 update_pmu_context(i
, pmu
);
6131 free_percpu(pmu
->pmu_cpu_context
);
6133 mutex_unlock(&pmus_lock
);
6135 static struct idr pmu_idr
;
6138 type_show(struct device
*dev
, struct device_attribute
*attr
, char *page
)
6140 struct pmu
*pmu
= dev_get_drvdata(dev
);
6142 return snprintf(page
, PAGE_SIZE
-1, "%d\n", pmu
->type
);
6145 static struct device_attribute pmu_dev_attrs
[] = {
6150 static int pmu_bus_running
;
6151 static struct bus_type pmu_bus
= {
6152 .name
= "event_source",
6153 .dev_attrs
= pmu_dev_attrs
,
6156 static void pmu_dev_release(struct device
*dev
)
6161 static int pmu_dev_alloc(struct pmu
*pmu
)
6165 pmu
->dev
= kzalloc(sizeof(struct device
), GFP_KERNEL
);
6169 pmu
->dev
->groups
= pmu
->attr_groups
;
6170 device_initialize(pmu
->dev
);
6171 ret
= dev_set_name(pmu
->dev
, "%s", pmu
->name
);
6175 dev_set_drvdata(pmu
->dev
, pmu
);
6176 pmu
->dev
->bus
= &pmu_bus
;
6177 pmu
->dev
->release
= pmu_dev_release
;
6178 ret
= device_add(pmu
->dev
);
6186 put_device(pmu
->dev
);
6190 static struct lock_class_key cpuctx_mutex
;
6191 static struct lock_class_key cpuctx_lock
;
6193 int perf_pmu_register(struct pmu
*pmu
, char *name
, int type
)
6197 mutex_lock(&pmus_lock
);
6199 pmu
->pmu_disable_count
= alloc_percpu(int);
6200 if (!pmu
->pmu_disable_count
)
6209 type
= idr_alloc(&pmu_idr
, pmu
, PERF_TYPE_MAX
, 0, GFP_KERNEL
);
6217 if (pmu_bus_running
) {
6218 ret
= pmu_dev_alloc(pmu
);
6224 pmu
->pmu_cpu_context
= find_pmu_context(pmu
->task_ctx_nr
);
6225 if (pmu
->pmu_cpu_context
)
6226 goto got_cpu_context
;
6229 pmu
->pmu_cpu_context
= alloc_percpu(struct perf_cpu_context
);
6230 if (!pmu
->pmu_cpu_context
)
6233 for_each_possible_cpu(cpu
) {
6234 struct perf_cpu_context
*cpuctx
;
6236 cpuctx
= per_cpu_ptr(pmu
->pmu_cpu_context
, cpu
);
6237 __perf_event_init_context(&cpuctx
->ctx
);
6238 lockdep_set_class(&cpuctx
->ctx
.mutex
, &cpuctx_mutex
);
6239 lockdep_set_class(&cpuctx
->ctx
.lock
, &cpuctx_lock
);
6240 cpuctx
->ctx
.type
= cpu_context
;
6241 cpuctx
->ctx
.pmu
= pmu
;
6242 cpuctx
->jiffies_interval
= 1;
6243 INIT_LIST_HEAD(&cpuctx
->rotation_list
);
6244 cpuctx
->unique_pmu
= pmu
;
6248 if (!pmu
->start_txn
) {
6249 if (pmu
->pmu_enable
) {
6251 * If we have pmu_enable/pmu_disable calls, install
6252 * transaction stubs that use that to try and batch
6253 * hardware accesses.
6255 pmu
->start_txn
= perf_pmu_start_txn
;
6256 pmu
->commit_txn
= perf_pmu_commit_txn
;
6257 pmu
->cancel_txn
= perf_pmu_cancel_txn
;
6259 pmu
->start_txn
= perf_pmu_nop_void
;
6260 pmu
->commit_txn
= perf_pmu_nop_int
;
6261 pmu
->cancel_txn
= perf_pmu_nop_void
;
6265 if (!pmu
->pmu_enable
) {
6266 pmu
->pmu_enable
= perf_pmu_nop_void
;
6267 pmu
->pmu_disable
= perf_pmu_nop_void
;
6270 if (!pmu
->event_idx
)
6271 pmu
->event_idx
= perf_event_idx_default
;
6273 list_add_rcu(&pmu
->entry
, &pmus
);
6276 mutex_unlock(&pmus_lock
);
6281 device_del(pmu
->dev
);
6282 put_device(pmu
->dev
);
6285 if (pmu
->type
>= PERF_TYPE_MAX
)
6286 idr_remove(&pmu_idr
, pmu
->type
);
6289 free_percpu(pmu
->pmu_disable_count
);
6293 void perf_pmu_unregister(struct pmu
*pmu
)
6295 mutex_lock(&pmus_lock
);
6296 list_del_rcu(&pmu
->entry
);
6297 mutex_unlock(&pmus_lock
);
6300 * We dereference the pmu list under both SRCU and regular RCU, so
6301 * synchronize against both of those.
6303 synchronize_srcu(&pmus_srcu
);
6306 free_percpu(pmu
->pmu_disable_count
);
6307 if (pmu
->type
>= PERF_TYPE_MAX
)
6308 idr_remove(&pmu_idr
, pmu
->type
);
6309 device_del(pmu
->dev
);
6310 put_device(pmu
->dev
);
6311 free_pmu_context(pmu
);
6314 struct pmu
*perf_init_event(struct perf_event
*event
)
6316 struct pmu
*pmu
= NULL
;
6320 idx
= srcu_read_lock(&pmus_srcu
);
6323 pmu
= idr_find(&pmu_idr
, event
->attr
.type
);
6327 ret
= pmu
->event_init(event
);
6333 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
6335 ret
= pmu
->event_init(event
);
6339 if (ret
!= -ENOENT
) {
6344 pmu
= ERR_PTR(-ENOENT
);
6346 srcu_read_unlock(&pmus_srcu
, idx
);
6352 * Allocate and initialize a event structure
6354 static struct perf_event
*
6355 perf_event_alloc(struct perf_event_attr
*attr
, int cpu
,
6356 struct task_struct
*task
,
6357 struct perf_event
*group_leader
,
6358 struct perf_event
*parent_event
,
6359 perf_overflow_handler_t overflow_handler
,
6363 struct perf_event
*event
;
6364 struct hw_perf_event
*hwc
;
6367 if ((unsigned)cpu
>= nr_cpu_ids
) {
6368 if (!task
|| cpu
!= -1)
6369 return ERR_PTR(-EINVAL
);
6372 event
= kzalloc(sizeof(*event
), GFP_KERNEL
);
6374 return ERR_PTR(-ENOMEM
);
6377 * Single events are their own group leaders, with an
6378 * empty sibling list:
6381 group_leader
= event
;
6383 mutex_init(&event
->child_mutex
);
6384 INIT_LIST_HEAD(&event
->child_list
);
6386 INIT_LIST_HEAD(&event
->group_entry
);
6387 INIT_LIST_HEAD(&event
->event_entry
);
6388 INIT_LIST_HEAD(&event
->sibling_list
);
6389 INIT_LIST_HEAD(&event
->rb_entry
);
6391 init_waitqueue_head(&event
->waitq
);
6392 init_irq_work(&event
->pending
, perf_pending_event
);
6394 mutex_init(&event
->mmap_mutex
);
6396 atomic_long_set(&event
->refcount
, 1);
6398 event
->attr
= *attr
;
6399 event
->group_leader
= group_leader
;
6403 event
->parent
= parent_event
;
6405 event
->ns
= get_pid_ns(task_active_pid_ns(current
));
6406 event
->id
= atomic64_inc_return(&perf_event_id
);
6408 event
->state
= PERF_EVENT_STATE_INACTIVE
;
6411 event
->attach_state
= PERF_ATTACH_TASK
;
6413 if (attr
->type
== PERF_TYPE_TRACEPOINT
)
6414 event
->hw
.tp_target
= task
;
6415 #ifdef CONFIG_HAVE_HW_BREAKPOINT
6417 * hw_breakpoint is a bit difficult here..
6419 else if (attr
->type
== PERF_TYPE_BREAKPOINT
)
6420 event
->hw
.bp_target
= task
;
6424 if (!overflow_handler
&& parent_event
) {
6425 overflow_handler
= parent_event
->overflow_handler
;
6426 context
= parent_event
->overflow_handler_context
;
6429 event
->overflow_handler
= overflow_handler
;
6430 event
->overflow_handler_context
= context
;
6432 perf_event__state_init(event
);
6437 hwc
->sample_period
= attr
->sample_period
;
6438 if (attr
->freq
&& attr
->sample_freq
)
6439 hwc
->sample_period
= 1;
6440 hwc
->last_period
= hwc
->sample_period
;
6442 local64_set(&hwc
->period_left
, hwc
->sample_period
);
6445 * we currently do not support PERF_FORMAT_GROUP on inherited events
6447 if (attr
->inherit
&& (attr
->read_format
& PERF_FORMAT_GROUP
))
6450 pmu
= perf_init_event(event
);
6456 else if (IS_ERR(pmu
))
6461 put_pid_ns(event
->ns
);
6463 return ERR_PTR(err
);
6466 if (!event
->parent
) {
6467 if (event
->attach_state
& PERF_ATTACH_TASK
)
6468 static_key_slow_inc(&perf_sched_events
.key
);
6469 if (event
->attr
.mmap
|| event
->attr
.mmap_data
)
6470 atomic_inc(&nr_mmap_events
);
6471 if (event
->attr
.comm
)
6472 atomic_inc(&nr_comm_events
);
6473 if (event
->attr
.task
)
6474 atomic_inc(&nr_task_events
);
6475 if (event
->attr
.sample_type
& PERF_SAMPLE_CALLCHAIN
) {
6476 err
= get_callchain_buffers();
6479 return ERR_PTR(err
);
6482 if (has_branch_stack(event
)) {
6483 static_key_slow_inc(&perf_sched_events
.key
);
6484 if (!(event
->attach_state
& PERF_ATTACH_TASK
))
6485 atomic_inc(&per_cpu(perf_branch_stack_events
,
6493 static int perf_copy_attr(struct perf_event_attr __user
*uattr
,
6494 struct perf_event_attr
*attr
)
6499 if (!access_ok(VERIFY_WRITE
, uattr
, PERF_ATTR_SIZE_VER0
))
6503 * zero the full structure, so that a short copy will be nice.
6505 memset(attr
, 0, sizeof(*attr
));
6507 ret
= get_user(size
, &uattr
->size
);
6511 if (size
> PAGE_SIZE
) /* silly large */
6514 if (!size
) /* abi compat */
6515 size
= PERF_ATTR_SIZE_VER0
;
6517 if (size
< PERF_ATTR_SIZE_VER0
)
6521 * If we're handed a bigger struct than we know of,
6522 * ensure all the unknown bits are 0 - i.e. new
6523 * user-space does not rely on any kernel feature
6524 * extensions we dont know about yet.
6526 if (size
> sizeof(*attr
)) {
6527 unsigned char __user
*addr
;
6528 unsigned char __user
*end
;
6531 addr
= (void __user
*)uattr
+ sizeof(*attr
);
6532 end
= (void __user
*)uattr
+ size
;
6534 for (; addr
< end
; addr
++) {
6535 ret
= get_user(val
, addr
);
6541 size
= sizeof(*attr
);
6544 ret
= copy_from_user(attr
, uattr
, size
);
6548 if (attr
->__reserved_1
)
6551 if (attr
->sample_type
& ~(PERF_SAMPLE_MAX
-1))
6554 if (attr
->read_format
& ~(PERF_FORMAT_MAX
-1))
6557 if (attr
->sample_type
& PERF_SAMPLE_BRANCH_STACK
) {
6558 u64 mask
= attr
->branch_sample_type
;
6560 /* only using defined bits */
6561 if (mask
& ~(PERF_SAMPLE_BRANCH_MAX
-1))
6564 /* at least one branch bit must be set */
6565 if (!(mask
& ~PERF_SAMPLE_BRANCH_PLM_ALL
))
6568 /* kernel level capture: check permissions */
6569 if ((mask
& PERF_SAMPLE_BRANCH_PERM_PLM
)
6570 && perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN
))
6573 /* propagate priv level, when not set for branch */
6574 if (!(mask
& PERF_SAMPLE_BRANCH_PLM_ALL
)) {
6576 /* exclude_kernel checked on syscall entry */
6577 if (!attr
->exclude_kernel
)
6578 mask
|= PERF_SAMPLE_BRANCH_KERNEL
;
6580 if (!attr
->exclude_user
)
6581 mask
|= PERF_SAMPLE_BRANCH_USER
;
6583 if (!attr
->exclude_hv
)
6584 mask
|= PERF_SAMPLE_BRANCH_HV
;
6586 * adjust user setting (for HW filter setup)
6588 attr
->branch_sample_type
= mask
;
6592 if (attr
->sample_type
& PERF_SAMPLE_REGS_USER
) {
6593 ret
= perf_reg_validate(attr
->sample_regs_user
);
6598 if (attr
->sample_type
& PERF_SAMPLE_STACK_USER
) {
6599 if (!arch_perf_have_user_stack_dump())
6603 * We have __u32 type for the size, but so far
6604 * we can only use __u16 as maximum due to the
6605 * __u16 sample size limit.
6607 if (attr
->sample_stack_user
>= USHRT_MAX
)
6609 else if (!IS_ALIGNED(attr
->sample_stack_user
, sizeof(u64
)))
6617 put_user(sizeof(*attr
), &uattr
->size
);
6623 perf_event_set_output(struct perf_event
*event
, struct perf_event
*output_event
)
6625 struct ring_buffer
*rb
= NULL
, *old_rb
= NULL
;
6631 /* don't allow circular references */
6632 if (event
== output_event
)
6636 * Don't allow cross-cpu buffers
6638 if (output_event
->cpu
!= event
->cpu
)
6642 * If its not a per-cpu rb, it must be the same task.
6644 if (output_event
->cpu
== -1 && output_event
->ctx
!= event
->ctx
)
6648 mutex_lock(&event
->mmap_mutex
);
6649 /* Can't redirect output if we've got an active mmap() */
6650 if (atomic_read(&event
->mmap_count
))
6656 /* get the rb we want to redirect to */
6657 rb
= ring_buffer_get(output_event
);
6663 ring_buffer_detach(event
, old_rb
);
6666 ring_buffer_attach(event
, rb
);
6668 rcu_assign_pointer(event
->rb
, rb
);
6671 ring_buffer_put(old_rb
);
6673 * Since we detached before setting the new rb, so that we
6674 * could attach the new rb, we could have missed a wakeup.
6677 wake_up_all(&event
->waitq
);
6682 mutex_unlock(&event
->mmap_mutex
);
6689 * sys_perf_event_open - open a performance event, associate it to a task/cpu
6691 * @attr_uptr: event_id type attributes for monitoring/sampling
6694 * @group_fd: group leader event fd
6696 SYSCALL_DEFINE5(perf_event_open
,
6697 struct perf_event_attr __user
*, attr_uptr
,
6698 pid_t
, pid
, int, cpu
, int, group_fd
, unsigned long, flags
)
6700 struct perf_event
*group_leader
= NULL
, *output_event
= NULL
;
6701 struct perf_event
*event
, *sibling
;
6702 struct perf_event_attr attr
;
6703 struct perf_event_context
*ctx
;
6704 struct file
*event_file
= NULL
;
6705 struct fd group
= {NULL
, 0};
6706 struct task_struct
*task
= NULL
;
6712 /* for future expandability... */
6713 if (flags
& ~PERF_FLAG_ALL
)
6716 err
= perf_copy_attr(attr_uptr
, &attr
);
6720 if (!attr
.exclude_kernel
) {
6721 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN
))
6726 if (attr
.sample_freq
> sysctl_perf_event_sample_rate
)
6729 if (attr
.sample_period
& (1ULL << 63))
6734 * In cgroup mode, the pid argument is used to pass the fd
6735 * opened to the cgroup directory in cgroupfs. The cpu argument
6736 * designates the cpu on which to monitor threads from that
6739 if ((flags
& PERF_FLAG_PID_CGROUP
) && (pid
== -1 || cpu
== -1))
6742 event_fd
= get_unused_fd();
6746 if (group_fd
!= -1) {
6747 err
= perf_fget_light(group_fd
, &group
);
6750 group_leader
= group
.file
->private_data
;
6751 if (flags
& PERF_FLAG_FD_OUTPUT
)
6752 output_event
= group_leader
;
6753 if (flags
& PERF_FLAG_FD_NO_GROUP
)
6754 group_leader
= NULL
;
6757 if (pid
!= -1 && !(flags
& PERF_FLAG_PID_CGROUP
)) {
6758 task
= find_lively_task_by_vpid(pid
);
6760 err
= PTR_ERR(task
);
6767 event
= perf_event_alloc(&attr
, cpu
, task
, group_leader
, NULL
,
6769 if (IS_ERR(event
)) {
6770 err
= PTR_ERR(event
);
6774 if (flags
& PERF_FLAG_PID_CGROUP
) {
6775 err
= perf_cgroup_connect(pid
, event
, &attr
, group_leader
);
6780 * - that has cgroup constraint on event->cpu
6781 * - that may need work on context switch
6783 atomic_inc(&per_cpu(perf_cgroup_events
, event
->cpu
));
6784 static_key_slow_inc(&perf_sched_events
.key
);
6788 * Special case software events and allow them to be part of
6789 * any hardware group.
6794 (is_software_event(event
) != is_software_event(group_leader
))) {
6795 if (is_software_event(event
)) {
6797 * If event and group_leader are not both a software
6798 * event, and event is, then group leader is not.
6800 * Allow the addition of software events to !software
6801 * groups, this is safe because software events never
6804 pmu
= group_leader
->pmu
;
6805 } else if (is_software_event(group_leader
) &&
6806 (group_leader
->group_flags
& PERF_GROUP_SOFTWARE
)) {
6808 * In case the group is a pure software group, and we
6809 * try to add a hardware event, move the whole group to
6810 * the hardware context.
6817 * Get the target context (task or percpu):
6819 ctx
= find_get_context(pmu
, task
, event
->cpu
);
6826 put_task_struct(task
);
6831 * Look up the group leader (we will attach this event to it):
6837 * Do not allow a recursive hierarchy (this new sibling
6838 * becoming part of another group-sibling):
6840 if (group_leader
->group_leader
!= group_leader
)
6843 * Do not allow to attach to a group in a different
6844 * task or CPU context:
6847 if (group_leader
->ctx
->type
!= ctx
->type
)
6850 if (group_leader
->ctx
!= ctx
)
6855 * Only a group leader can be exclusive or pinned
6857 if (attr
.exclusive
|| attr
.pinned
)
6862 err
= perf_event_set_output(event
, output_event
);
6867 event_file
= anon_inode_getfile("[perf_event]", &perf_fops
, event
, O_RDWR
);
6868 if (IS_ERR(event_file
)) {
6869 err
= PTR_ERR(event_file
);
6874 struct perf_event_context
*gctx
= group_leader
->ctx
;
6876 mutex_lock(&gctx
->mutex
);
6877 perf_remove_from_context(group_leader
, false);
6880 * Removing from the context ends up with disabled
6881 * event. What we want here is event in the initial
6882 * startup state, ready to be add into new context.
6884 perf_event__state_init(group_leader
);
6885 list_for_each_entry(sibling
, &group_leader
->sibling_list
,
6887 perf_remove_from_context(sibling
, false);
6888 perf_event__state_init(sibling
);
6891 mutex_unlock(&gctx
->mutex
);
6895 WARN_ON_ONCE(ctx
->parent_ctx
);
6896 mutex_lock(&ctx
->mutex
);
6900 perf_install_in_context(ctx
, group_leader
, group_leader
->cpu
);
6902 list_for_each_entry(sibling
, &group_leader
->sibling_list
,
6904 perf_install_in_context(ctx
, sibling
, sibling
->cpu
);
6909 perf_install_in_context(ctx
, event
, event
->cpu
);
6911 perf_unpin_context(ctx
);
6912 mutex_unlock(&ctx
->mutex
);
6916 event
->owner
= current
;
6918 mutex_lock(¤t
->perf_event_mutex
);
6919 list_add_tail(&event
->owner_entry
, ¤t
->perf_event_list
);
6920 mutex_unlock(¤t
->perf_event_mutex
);
6923 * Precalculate sample_data sizes
6925 perf_event__header_size(event
);
6926 perf_event__id_header_size(event
);
6929 * Drop the reference on the group_event after placing the
6930 * new event on the sibling_list. This ensures destruction
6931 * of the group leader will find the pointer to itself in
6932 * perf_group_detach().
6935 fd_install(event_fd
, event_file
);
6939 perf_unpin_context(ctx
);
6946 put_task_struct(task
);
6950 put_unused_fd(event_fd
);
6955 * perf_event_create_kernel_counter
6957 * @attr: attributes of the counter to create
6958 * @cpu: cpu in which the counter is bound
6959 * @task: task to profile (NULL for percpu)
6962 perf_event_create_kernel_counter(struct perf_event_attr
*attr
, int cpu
,
6963 struct task_struct
*task
,
6964 perf_overflow_handler_t overflow_handler
,
6967 struct perf_event_context
*ctx
;
6968 struct perf_event
*event
;
6972 * Get the target context (task or percpu):
6975 event
= perf_event_alloc(attr
, cpu
, task
, NULL
, NULL
,
6976 overflow_handler
, context
);
6977 if (IS_ERR(event
)) {
6978 err
= PTR_ERR(event
);
6982 ctx
= find_get_context(event
->pmu
, task
, cpu
);
6988 WARN_ON_ONCE(ctx
->parent_ctx
);
6989 mutex_lock(&ctx
->mutex
);
6990 perf_install_in_context(ctx
, event
, cpu
);
6992 perf_unpin_context(ctx
);
6993 mutex_unlock(&ctx
->mutex
);
7000 return ERR_PTR(err
);
7002 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter
);
7004 void perf_pmu_migrate_context(struct pmu
*pmu
, int src_cpu
, int dst_cpu
)
7006 struct perf_event_context
*src_ctx
;
7007 struct perf_event_context
*dst_ctx
;
7008 struct perf_event
*event
, *tmp
;
7011 src_ctx
= &per_cpu_ptr(pmu
->pmu_cpu_context
, src_cpu
)->ctx
;
7012 dst_ctx
= &per_cpu_ptr(pmu
->pmu_cpu_context
, dst_cpu
)->ctx
;
7014 mutex_lock(&src_ctx
->mutex
);
7015 list_for_each_entry_safe(event
, tmp
, &src_ctx
->event_list
,
7017 perf_remove_from_context(event
, false);
7019 list_add(&event
->event_entry
, &events
);
7021 mutex_unlock(&src_ctx
->mutex
);
7025 mutex_lock(&dst_ctx
->mutex
);
7026 list_for_each_entry_safe(event
, tmp
, &events
, event_entry
) {
7027 list_del(&event
->event_entry
);
7028 if (event
->state
>= PERF_EVENT_STATE_OFF
)
7029 event
->state
= PERF_EVENT_STATE_INACTIVE
;
7030 perf_install_in_context(dst_ctx
, event
, dst_cpu
);
7033 mutex_unlock(&dst_ctx
->mutex
);
7035 EXPORT_SYMBOL_GPL(perf_pmu_migrate_context
);
7037 static void sync_child_event(struct perf_event
*child_event
,
7038 struct task_struct
*child
)
7040 struct perf_event
*parent_event
= child_event
->parent
;
7043 if (child_event
->attr
.inherit_stat
)
7044 perf_event_read_event(child_event
, child
);
7046 child_val
= perf_event_count(child_event
);
7049 * Add back the child's count to the parent's count:
7051 atomic64_add(child_val
, &parent_event
->child_count
);
7052 atomic64_add(child_event
->total_time_enabled
,
7053 &parent_event
->child_total_time_enabled
);
7054 atomic64_add(child_event
->total_time_running
,
7055 &parent_event
->child_total_time_running
);
7058 * Remove this event from the parent's list
7060 WARN_ON_ONCE(parent_event
->ctx
->parent_ctx
);
7061 mutex_lock(&parent_event
->child_mutex
);
7062 list_del_init(&child_event
->child_list
);
7063 mutex_unlock(&parent_event
->child_mutex
);
7066 * Release the parent event, if this was the last
7069 put_event(parent_event
);
7073 __perf_event_exit_task(struct perf_event
*child_event
,
7074 struct perf_event_context
*child_ctx
,
7075 struct task_struct
*child
)
7077 perf_remove_from_context(child_event
, !!child_event
->parent
);
7080 * It can happen that the parent exits first, and has events
7081 * that are still around due to the child reference. These
7082 * events need to be zapped.
7084 if (child_event
->parent
) {
7085 sync_child_event(child_event
, child
);
7086 free_event(child_event
);
7090 static void perf_event_exit_task_context(struct task_struct
*child
, int ctxn
)
7092 struct perf_event
*child_event
, *tmp
;
7093 struct perf_event_context
*child_ctx
;
7094 unsigned long flags
;
7096 if (likely(!child
->perf_event_ctxp
[ctxn
])) {
7097 perf_event_task(child
, NULL
, 0);
7101 local_irq_save(flags
);
7103 * We can't reschedule here because interrupts are disabled,
7104 * and either child is current or it is a task that can't be
7105 * scheduled, so we are now safe from rescheduling changing
7108 child_ctx
= rcu_dereference_raw(child
->perf_event_ctxp
[ctxn
]);
7111 * Take the context lock here so that if find_get_context is
7112 * reading child->perf_event_ctxp, we wait until it has
7113 * incremented the context's refcount before we do put_ctx below.
7115 raw_spin_lock(&child_ctx
->lock
);
7116 task_ctx_sched_out(child_ctx
);
7117 child
->perf_event_ctxp
[ctxn
] = NULL
;
7119 * If this context is a clone; unclone it so it can't get
7120 * swapped to another process while we're removing all
7121 * the events from it.
7123 unclone_ctx(child_ctx
);
7124 update_context_time(child_ctx
);
7125 raw_spin_unlock_irqrestore(&child_ctx
->lock
, flags
);
7128 * Report the task dead after unscheduling the events so that we
7129 * won't get any samples after PERF_RECORD_EXIT. We can however still
7130 * get a few PERF_RECORD_READ events.
7132 perf_event_task(child
, child_ctx
, 0);
7135 * We can recurse on the same lock type through:
7137 * __perf_event_exit_task()
7138 * sync_child_event()
7140 * mutex_lock(&ctx->mutex)
7142 * But since its the parent context it won't be the same instance.
7144 mutex_lock(&child_ctx
->mutex
);
7147 list_for_each_entry_safe(child_event
, tmp
, &child_ctx
->pinned_groups
,
7149 __perf_event_exit_task(child_event
, child_ctx
, child
);
7151 list_for_each_entry_safe(child_event
, tmp
, &child_ctx
->flexible_groups
,
7153 __perf_event_exit_task(child_event
, child_ctx
, child
);
7156 * If the last event was a group event, it will have appended all
7157 * its siblings to the list, but we obtained 'tmp' before that which
7158 * will still point to the list head terminating the iteration.
7160 if (!list_empty(&child_ctx
->pinned_groups
) ||
7161 !list_empty(&child_ctx
->flexible_groups
))
7164 mutex_unlock(&child_ctx
->mutex
);
7170 * When a child task exits, feed back event values to parent events.
7172 void perf_event_exit_task(struct task_struct
*child
)
7174 struct perf_event
*event
, *tmp
;
7177 mutex_lock(&child
->perf_event_mutex
);
7178 list_for_each_entry_safe(event
, tmp
, &child
->perf_event_list
,
7180 list_del_init(&event
->owner_entry
);
7183 * Ensure the list deletion is visible before we clear
7184 * the owner, closes a race against perf_release() where
7185 * we need to serialize on the owner->perf_event_mutex.
7188 event
->owner
= NULL
;
7190 mutex_unlock(&child
->perf_event_mutex
);
7192 for_each_task_context_nr(ctxn
)
7193 perf_event_exit_task_context(child
, ctxn
);
7196 static void perf_free_event(struct perf_event
*event
,
7197 struct perf_event_context
*ctx
)
7199 struct perf_event
*parent
= event
->parent
;
7201 if (WARN_ON_ONCE(!parent
))
7204 mutex_lock(&parent
->child_mutex
);
7205 list_del_init(&event
->child_list
);
7206 mutex_unlock(&parent
->child_mutex
);
7210 perf_group_detach(event
);
7211 list_del_event(event
, ctx
);
7216 * free an unexposed, unused context as created by inheritance by
7217 * perf_event_init_task below, used by fork() in case of fail.
7219 void perf_event_free_task(struct task_struct
*task
)
7221 struct perf_event_context
*ctx
;
7222 struct perf_event
*event
, *tmp
;
7225 for_each_task_context_nr(ctxn
) {
7226 ctx
= task
->perf_event_ctxp
[ctxn
];
7230 mutex_lock(&ctx
->mutex
);
7232 list_for_each_entry_safe(event
, tmp
, &ctx
->pinned_groups
,
7234 perf_free_event(event
, ctx
);
7236 list_for_each_entry_safe(event
, tmp
, &ctx
->flexible_groups
,
7238 perf_free_event(event
, ctx
);
7240 if (!list_empty(&ctx
->pinned_groups
) ||
7241 !list_empty(&ctx
->flexible_groups
))
7244 mutex_unlock(&ctx
->mutex
);
7250 void perf_event_delayed_put(struct task_struct
*task
)
7254 for_each_task_context_nr(ctxn
)
7255 WARN_ON_ONCE(task
->perf_event_ctxp
[ctxn
]);
7259 * inherit a event from parent task to child task:
7261 static struct perf_event
*
7262 inherit_event(struct perf_event
*parent_event
,
7263 struct task_struct
*parent
,
7264 struct perf_event_context
*parent_ctx
,
7265 struct task_struct
*child
,
7266 struct perf_event
*group_leader
,
7267 struct perf_event_context
*child_ctx
)
7269 struct perf_event
*child_event
;
7270 unsigned long flags
;
7273 * Instead of creating recursive hierarchies of events,
7274 * we link inherited events back to the original parent,
7275 * which has a filp for sure, which we use as the reference
7278 if (parent_event
->parent
)
7279 parent_event
= parent_event
->parent
;
7281 child_event
= perf_event_alloc(&parent_event
->attr
,
7284 group_leader
, parent_event
,
7286 if (IS_ERR(child_event
))
7289 if (!atomic_long_inc_not_zero(&parent_event
->refcount
)) {
7290 free_event(child_event
);
7297 * Make the child state follow the state of the parent event,
7298 * not its attr.disabled bit. We hold the parent's mutex,
7299 * so we won't race with perf_event_{en, dis}able_family.
7301 if (parent_event
->state
>= PERF_EVENT_STATE_INACTIVE
)
7302 child_event
->state
= PERF_EVENT_STATE_INACTIVE
;
7304 child_event
->state
= PERF_EVENT_STATE_OFF
;
7306 if (parent_event
->attr
.freq
) {
7307 u64 sample_period
= parent_event
->hw
.sample_period
;
7308 struct hw_perf_event
*hwc
= &child_event
->hw
;
7310 hwc
->sample_period
= sample_period
;
7311 hwc
->last_period
= sample_period
;
7313 local64_set(&hwc
->period_left
, sample_period
);
7316 child_event
->ctx
= child_ctx
;
7317 child_event
->overflow_handler
= parent_event
->overflow_handler
;
7318 child_event
->overflow_handler_context
7319 = parent_event
->overflow_handler_context
;
7322 * Precalculate sample_data sizes
7324 perf_event__header_size(child_event
);
7325 perf_event__id_header_size(child_event
);
7328 * Link it up in the child's context:
7330 raw_spin_lock_irqsave(&child_ctx
->lock
, flags
);
7331 add_event_to_ctx(child_event
, child_ctx
);
7332 raw_spin_unlock_irqrestore(&child_ctx
->lock
, flags
);
7335 * Link this into the parent event's child list
7337 WARN_ON_ONCE(parent_event
->ctx
->parent_ctx
);
7338 mutex_lock(&parent_event
->child_mutex
);
7339 list_add_tail(&child_event
->child_list
, &parent_event
->child_list
);
7340 mutex_unlock(&parent_event
->child_mutex
);
7345 static int inherit_group(struct perf_event
*parent_event
,
7346 struct task_struct
*parent
,
7347 struct perf_event_context
*parent_ctx
,
7348 struct task_struct
*child
,
7349 struct perf_event_context
*child_ctx
)
7351 struct perf_event
*leader
;
7352 struct perf_event
*sub
;
7353 struct perf_event
*child_ctr
;
7355 leader
= inherit_event(parent_event
, parent
, parent_ctx
,
7356 child
, NULL
, child_ctx
);
7358 return PTR_ERR(leader
);
7359 list_for_each_entry(sub
, &parent_event
->sibling_list
, group_entry
) {
7360 child_ctr
= inherit_event(sub
, parent
, parent_ctx
,
7361 child
, leader
, child_ctx
);
7362 if (IS_ERR(child_ctr
))
7363 return PTR_ERR(child_ctr
);
7369 inherit_task_group(struct perf_event
*event
, struct task_struct
*parent
,
7370 struct perf_event_context
*parent_ctx
,
7371 struct task_struct
*child
, int ctxn
,
7375 struct perf_event_context
*child_ctx
;
7377 if (!event
->attr
.inherit
) {
7382 child_ctx
= child
->perf_event_ctxp
[ctxn
];
7385 * This is executed from the parent task context, so
7386 * inherit events that have been marked for cloning.
7387 * First allocate and initialize a context for the
7391 child_ctx
= alloc_perf_context(parent_ctx
->pmu
, child
);
7395 child
->perf_event_ctxp
[ctxn
] = child_ctx
;
7398 ret
= inherit_group(event
, parent
, parent_ctx
,
7408 * Initialize the perf_event context in task_struct
7410 int perf_event_init_context(struct task_struct
*child
, int ctxn
)
7412 struct perf_event_context
*child_ctx
, *parent_ctx
;
7413 struct perf_event_context
*cloned_ctx
;
7414 struct perf_event
*event
;
7415 struct task_struct
*parent
= current
;
7416 int inherited_all
= 1;
7417 unsigned long flags
;
7420 if (likely(!parent
->perf_event_ctxp
[ctxn
]))
7424 * If the parent's context is a clone, pin it so it won't get
7427 parent_ctx
= perf_pin_task_context(parent
, ctxn
);
7430 * No need to check if parent_ctx != NULL here; since we saw
7431 * it non-NULL earlier, the only reason for it to become NULL
7432 * is if we exit, and since we're currently in the middle of
7433 * a fork we can't be exiting at the same time.
7437 * Lock the parent list. No need to lock the child - not PID
7438 * hashed yet and not running, so nobody can access it.
7440 mutex_lock(&parent_ctx
->mutex
);
7443 * We dont have to disable NMIs - we are only looking at
7444 * the list, not manipulating it:
7446 list_for_each_entry(event
, &parent_ctx
->pinned_groups
, group_entry
) {
7447 ret
= inherit_task_group(event
, parent
, parent_ctx
,
7448 child
, ctxn
, &inherited_all
);
7454 * We can't hold ctx->lock when iterating the ->flexible_group list due
7455 * to allocations, but we need to prevent rotation because
7456 * rotate_ctx() will change the list from interrupt context.
7458 raw_spin_lock_irqsave(&parent_ctx
->lock
, flags
);
7459 parent_ctx
->rotate_disable
= 1;
7460 raw_spin_unlock_irqrestore(&parent_ctx
->lock
, flags
);
7462 list_for_each_entry(event
, &parent_ctx
->flexible_groups
, group_entry
) {
7463 ret
= inherit_task_group(event
, parent
, parent_ctx
,
7464 child
, ctxn
, &inherited_all
);
7469 raw_spin_lock_irqsave(&parent_ctx
->lock
, flags
);
7470 parent_ctx
->rotate_disable
= 0;
7472 child_ctx
= child
->perf_event_ctxp
[ctxn
];
7474 if (child_ctx
&& inherited_all
) {
7476 * Mark the child context as a clone of the parent
7477 * context, or of whatever the parent is a clone of.
7479 * Note that if the parent is a clone, the holding of
7480 * parent_ctx->lock avoids it from being uncloned.
7482 cloned_ctx
= parent_ctx
->parent_ctx
;
7484 child_ctx
->parent_ctx
= cloned_ctx
;
7485 child_ctx
->parent_gen
= parent_ctx
->parent_gen
;
7487 child_ctx
->parent_ctx
= parent_ctx
;
7488 child_ctx
->parent_gen
= parent_ctx
->generation
;
7490 get_ctx(child_ctx
->parent_ctx
);
7493 raw_spin_unlock_irqrestore(&parent_ctx
->lock
, flags
);
7494 mutex_unlock(&parent_ctx
->mutex
);
7496 perf_unpin_context(parent_ctx
);
7497 put_ctx(parent_ctx
);
7503 * Initialize the perf_event context in task_struct
7505 int perf_event_init_task(struct task_struct
*child
)
7509 memset(child
->perf_event_ctxp
, 0, sizeof(child
->perf_event_ctxp
));
7510 mutex_init(&child
->perf_event_mutex
);
7511 INIT_LIST_HEAD(&child
->perf_event_list
);
7513 for_each_task_context_nr(ctxn
) {
7514 ret
= perf_event_init_context(child
, ctxn
);
7516 perf_event_free_task(child
);
7524 static void __init
perf_event_init_all_cpus(void)
7526 struct swevent_htable
*swhash
;
7529 for_each_possible_cpu(cpu
) {
7530 swhash
= &per_cpu(swevent_htable
, cpu
);
7531 mutex_init(&swhash
->hlist_mutex
);
7532 INIT_LIST_HEAD(&per_cpu(rotation_list
, cpu
));
7536 static void __cpuinit
perf_event_init_cpu(int cpu
)
7538 struct swevent_htable
*swhash
= &per_cpu(swevent_htable
, cpu
);
7540 mutex_lock(&swhash
->hlist_mutex
);
7541 swhash
->online
= true;
7542 if (swhash
->hlist_refcount
> 0) {
7543 struct swevent_hlist
*hlist
;
7545 hlist
= kzalloc_node(sizeof(*hlist
), GFP_KERNEL
, cpu_to_node(cpu
));
7547 rcu_assign_pointer(swhash
->swevent_hlist
, hlist
);
7549 mutex_unlock(&swhash
->hlist_mutex
);
7552 #if defined CONFIG_HOTPLUG_CPU || defined CONFIG_KEXEC
7553 static void perf_pmu_rotate_stop(struct pmu
*pmu
)
7555 struct perf_cpu_context
*cpuctx
= this_cpu_ptr(pmu
->pmu_cpu_context
);
7557 WARN_ON(!irqs_disabled());
7559 list_del_init(&cpuctx
->rotation_list
);
7562 static void __perf_event_exit_context(void *__info
)
7564 struct remove_event re
= { .detach_group
= false };
7565 struct perf_event_context
*ctx
= __info
;
7567 perf_pmu_rotate_stop(ctx
->pmu
);
7570 list_for_each_entry_rcu(re
.event
, &ctx
->event_list
, event_entry
)
7571 __perf_remove_from_context(&re
);
7575 static void perf_event_exit_cpu_context(int cpu
)
7577 struct perf_event_context
*ctx
;
7581 idx
= srcu_read_lock(&pmus_srcu
);
7582 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
7583 ctx
= &per_cpu_ptr(pmu
->pmu_cpu_context
, cpu
)->ctx
;
7585 mutex_lock(&ctx
->mutex
);
7586 smp_call_function_single(cpu
, __perf_event_exit_context
, ctx
, 1);
7587 mutex_unlock(&ctx
->mutex
);
7589 srcu_read_unlock(&pmus_srcu
, idx
);
7592 static void perf_event_exit_cpu(int cpu
)
7594 struct swevent_htable
*swhash
= &per_cpu(swevent_htable
, cpu
);
7596 perf_event_exit_cpu_context(cpu
);
7598 mutex_lock(&swhash
->hlist_mutex
);
7599 swhash
->online
= false;
7600 swevent_hlist_release(swhash
);
7601 mutex_unlock(&swhash
->hlist_mutex
);
7604 static inline void perf_event_exit_cpu(int cpu
) { }
7608 perf_reboot(struct notifier_block
*notifier
, unsigned long val
, void *v
)
7612 for_each_online_cpu(cpu
)
7613 perf_event_exit_cpu(cpu
);
7619 * Run the perf reboot notifier at the very last possible moment so that
7620 * the generic watchdog code runs as long as possible.
7622 static struct notifier_block perf_reboot_notifier
= {
7623 .notifier_call
= perf_reboot
,
7624 .priority
= INT_MIN
,
7627 static int __cpuinit
7628 perf_cpu_notify(struct notifier_block
*self
, unsigned long action
, void *hcpu
)
7630 unsigned int cpu
= (long)hcpu
;
7632 switch (action
& ~CPU_TASKS_FROZEN
) {
7634 case CPU_UP_PREPARE
:
7635 case CPU_DOWN_FAILED
:
7636 perf_event_init_cpu(cpu
);
7639 case CPU_UP_CANCELED
:
7640 case CPU_DOWN_PREPARE
:
7641 perf_event_exit_cpu(cpu
);
7651 void __init
perf_event_init(void)
7657 perf_event_init_all_cpus();
7658 init_srcu_struct(&pmus_srcu
);
7659 perf_pmu_register(&perf_swevent
, "software", PERF_TYPE_SOFTWARE
);
7660 perf_pmu_register(&perf_cpu_clock
, NULL
, -1);
7661 perf_pmu_register(&perf_task_clock
, NULL
, -1);
7663 perf_cpu_notifier(perf_cpu_notify
);
7664 register_reboot_notifier(&perf_reboot_notifier
);
7666 ret
= init_hw_breakpoint();
7667 WARN(ret
, "hw_breakpoint initialization failed with: %d", ret
);
7669 /* do not patch jump label more than once per second */
7670 jump_label_rate_limit(&perf_sched_events
, HZ
);
7673 * Build time assertion that we keep the data_head at the intended
7674 * location. IOW, validation we got the __reserved[] size right.
7676 BUILD_BUG_ON((offsetof(struct perf_event_mmap_page
, data_head
))
7680 static int __init
perf_event_sysfs_init(void)
7685 mutex_lock(&pmus_lock
);
7687 ret
= bus_register(&pmu_bus
);
7691 list_for_each_entry(pmu
, &pmus
, entry
) {
7692 if (!pmu
->name
|| pmu
->type
< 0)
7695 ret
= pmu_dev_alloc(pmu
);
7696 WARN(ret
, "Failed to register pmu: %s, reason %d\n", pmu
->name
, ret
);
7698 pmu_bus_running
= 1;
7702 mutex_unlock(&pmus_lock
);
7706 device_initcall(perf_event_sysfs_init
);
7708 #ifdef CONFIG_CGROUP_PERF
7709 static struct cgroup_subsys_state
*perf_cgroup_css_alloc(struct cgroup
*cont
)
7711 struct perf_cgroup
*jc
;
7713 jc
= kzalloc(sizeof(*jc
), GFP_KERNEL
);
7715 return ERR_PTR(-ENOMEM
);
7717 jc
->info
= alloc_percpu(struct perf_cgroup_info
);
7720 return ERR_PTR(-ENOMEM
);
7726 static void perf_cgroup_css_free(struct cgroup
*cont
)
7728 struct perf_cgroup
*jc
;
7729 jc
= container_of(cgroup_subsys_state(cont
, perf_subsys_id
),
7730 struct perf_cgroup
, css
);
7731 free_percpu(jc
->info
);
7735 static int __perf_cgroup_move(void *info
)
7737 struct task_struct
*task
= info
;
7738 perf_cgroup_switch(task
, PERF_CGROUP_SWOUT
| PERF_CGROUP_SWIN
);
7742 static void perf_cgroup_attach(struct cgroup
*cgrp
, struct cgroup_taskset
*tset
)
7744 struct task_struct
*task
;
7746 cgroup_taskset_for_each(task
, cgrp
, tset
)
7747 task_function_call(task
, __perf_cgroup_move
, task
);
7750 static void perf_cgroup_exit(struct cgroup
*cgrp
, struct cgroup
*old_cgrp
,
7751 struct task_struct
*task
)
7754 * cgroup_exit() is called in the copy_process() failure path.
7755 * Ignore this case since the task hasn't ran yet, this avoids
7756 * trying to poke a half freed task state from generic code.
7758 if (!(task
->flags
& PF_EXITING
))
7761 task_function_call(task
, __perf_cgroup_move
, task
);
7764 struct cgroup_subsys perf_subsys
= {
7765 .name
= "perf_event",
7766 .subsys_id
= perf_subsys_id
,
7767 .css_alloc
= perf_cgroup_css_alloc
,
7768 .css_free
= perf_cgroup_css_free
,
7769 .exit
= perf_cgroup_exit
,
7770 .attach
= perf_cgroup_attach
,
7772 #endif /* CONFIG_CGROUP_PERF */