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/cgroup.h>
38 #include <linux/perf_event.h>
39 #include <linux/ftrace_event.h>
40 #include <linux/hw_breakpoint.h>
41 #include <linux/mm_types.h>
42 #include <linux/module.h>
43 #include <linux/mman.h>
44 #include <linux/compat.h>
45 #include <linux/bpf.h>
46 #include <linux/filter.h>
50 #include <asm/irq_regs.h>
52 static struct workqueue_struct
*perf_wq
;
54 struct remote_function_call
{
55 struct task_struct
*p
;
56 int (*func
)(void *info
);
61 static void remote_function(void *data
)
63 struct remote_function_call
*tfc
= data
;
64 struct task_struct
*p
= tfc
->p
;
68 if (task_cpu(p
) != smp_processor_id() || !task_curr(p
))
72 tfc
->ret
= tfc
->func(tfc
->info
);
76 * task_function_call - call a function on the cpu on which a task runs
77 * @p: the task to evaluate
78 * @func: the function to be called
79 * @info: the function call argument
81 * Calls the function @func when the task is currently running. This might
82 * be on the current CPU, which just calls the function directly
84 * returns: @func return value, or
85 * -ESRCH - when the process isn't running
86 * -EAGAIN - when the process moved away
89 task_function_call(struct task_struct
*p
, int (*func
) (void *info
), void *info
)
91 struct remote_function_call data
= {
95 .ret
= -ESRCH
, /* No such (running) process */
99 smp_call_function_single(task_cpu(p
), remote_function
, &data
, 1);
105 * cpu_function_call - call a function on the cpu
106 * @func: the function to be called
107 * @info: the function call argument
109 * Calls the function @func on the remote cpu.
111 * returns: @func return value or -ENXIO when the cpu is offline
113 static int cpu_function_call(int cpu
, int (*func
) (void *info
), void *info
)
115 struct remote_function_call data
= {
119 .ret
= -ENXIO
, /* No such CPU */
122 smp_call_function_single(cpu
, remote_function
, &data
, 1);
127 #define EVENT_OWNER_KERNEL ((void *) -1)
129 static bool is_kernel_event(struct perf_event
*event
)
131 return event
->owner
== EVENT_OWNER_KERNEL
;
134 #define PERF_FLAG_ALL (PERF_FLAG_FD_NO_GROUP |\
135 PERF_FLAG_FD_OUTPUT |\
136 PERF_FLAG_PID_CGROUP |\
137 PERF_FLAG_FD_CLOEXEC)
140 * branch priv levels that need permission checks
142 #define PERF_SAMPLE_BRANCH_PERM_PLM \
143 (PERF_SAMPLE_BRANCH_KERNEL |\
144 PERF_SAMPLE_BRANCH_HV)
147 EVENT_FLEXIBLE
= 0x1,
149 EVENT_ALL
= EVENT_FLEXIBLE
| EVENT_PINNED
,
153 * perf_sched_events : >0 events exist
154 * perf_cgroup_events: >0 per-cpu cgroup events exist on this cpu
156 struct static_key_deferred perf_sched_events __read_mostly
;
157 static DEFINE_PER_CPU(atomic_t
, perf_cgroup_events
);
158 static DEFINE_PER_CPU(int, perf_sched_cb_usages
);
160 static atomic_t nr_mmap_events __read_mostly
;
161 static atomic_t nr_comm_events __read_mostly
;
162 static atomic_t nr_task_events __read_mostly
;
163 static atomic_t nr_freq_events __read_mostly
;
165 static LIST_HEAD(pmus
);
166 static DEFINE_MUTEX(pmus_lock
);
167 static struct srcu_struct pmus_srcu
;
170 * perf event paranoia level:
171 * -1 - not paranoid at all
172 * 0 - disallow raw tracepoint access for unpriv
173 * 1 - disallow cpu events for unpriv
174 * 2 - disallow kernel profiling for unpriv
176 int sysctl_perf_event_paranoid __read_mostly
= 1;
178 /* Minimum for 512 kiB + 1 user control page */
179 int sysctl_perf_event_mlock __read_mostly
= 512 + (PAGE_SIZE
/ 1024); /* 'free' kiB per user */
182 * max perf event sample rate
184 #define DEFAULT_MAX_SAMPLE_RATE 100000
185 #define DEFAULT_SAMPLE_PERIOD_NS (NSEC_PER_SEC / DEFAULT_MAX_SAMPLE_RATE)
186 #define DEFAULT_CPU_TIME_MAX_PERCENT 25
188 int sysctl_perf_event_sample_rate __read_mostly
= DEFAULT_MAX_SAMPLE_RATE
;
190 static int max_samples_per_tick __read_mostly
= DIV_ROUND_UP(DEFAULT_MAX_SAMPLE_RATE
, HZ
);
191 static int perf_sample_period_ns __read_mostly
= DEFAULT_SAMPLE_PERIOD_NS
;
193 static int perf_sample_allowed_ns __read_mostly
=
194 DEFAULT_SAMPLE_PERIOD_NS
* DEFAULT_CPU_TIME_MAX_PERCENT
/ 100;
196 void update_perf_cpu_limits(void)
198 u64 tmp
= perf_sample_period_ns
;
200 tmp
*= sysctl_perf_cpu_time_max_percent
;
202 ACCESS_ONCE(perf_sample_allowed_ns
) = tmp
;
205 static int perf_rotate_context(struct perf_cpu_context
*cpuctx
);
207 int perf_proc_update_handler(struct ctl_table
*table
, int write
,
208 void __user
*buffer
, size_t *lenp
,
211 int ret
= proc_dointvec_minmax(table
, write
, buffer
, lenp
, ppos
);
216 max_samples_per_tick
= DIV_ROUND_UP(sysctl_perf_event_sample_rate
, HZ
);
217 perf_sample_period_ns
= NSEC_PER_SEC
/ sysctl_perf_event_sample_rate
;
218 update_perf_cpu_limits();
223 int sysctl_perf_cpu_time_max_percent __read_mostly
= DEFAULT_CPU_TIME_MAX_PERCENT
;
225 int perf_cpu_time_max_percent_handler(struct ctl_table
*table
, int write
,
226 void __user
*buffer
, size_t *lenp
,
229 int ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
234 update_perf_cpu_limits();
240 * perf samples are done in some very critical code paths (NMIs).
241 * If they take too much CPU time, the system can lock up and not
242 * get any real work done. This will drop the sample rate when
243 * we detect that events are taking too long.
245 #define NR_ACCUMULATED_SAMPLES 128
246 static DEFINE_PER_CPU(u64
, running_sample_length
);
248 static void perf_duration_warn(struct irq_work
*w
)
250 u64 allowed_ns
= ACCESS_ONCE(perf_sample_allowed_ns
);
251 u64 avg_local_sample_len
;
252 u64 local_samples_len
;
254 local_samples_len
= __this_cpu_read(running_sample_length
);
255 avg_local_sample_len
= local_samples_len
/NR_ACCUMULATED_SAMPLES
;
257 printk_ratelimited(KERN_WARNING
258 "perf interrupt took too long (%lld > %lld), lowering "
259 "kernel.perf_event_max_sample_rate to %d\n",
260 avg_local_sample_len
, allowed_ns
>> 1,
261 sysctl_perf_event_sample_rate
);
264 static DEFINE_IRQ_WORK(perf_duration_work
, perf_duration_warn
);
266 void perf_sample_event_took(u64 sample_len_ns
)
268 u64 allowed_ns
= ACCESS_ONCE(perf_sample_allowed_ns
);
269 u64 avg_local_sample_len
;
270 u64 local_samples_len
;
275 /* decay the counter by 1 average sample */
276 local_samples_len
= __this_cpu_read(running_sample_length
);
277 local_samples_len
-= local_samples_len
/NR_ACCUMULATED_SAMPLES
;
278 local_samples_len
+= sample_len_ns
;
279 __this_cpu_write(running_sample_length
, local_samples_len
);
282 * note: this will be biased artifically low until we have
283 * seen NR_ACCUMULATED_SAMPLES. Doing it this way keeps us
284 * from having to maintain a count.
286 avg_local_sample_len
= local_samples_len
/NR_ACCUMULATED_SAMPLES
;
288 if (avg_local_sample_len
<= allowed_ns
)
291 if (max_samples_per_tick
<= 1)
294 max_samples_per_tick
= DIV_ROUND_UP(max_samples_per_tick
, 2);
295 sysctl_perf_event_sample_rate
= max_samples_per_tick
* HZ
;
296 perf_sample_period_ns
= NSEC_PER_SEC
/ sysctl_perf_event_sample_rate
;
298 update_perf_cpu_limits();
300 if (!irq_work_queue(&perf_duration_work
)) {
301 early_printk("perf interrupt took too long (%lld > %lld), lowering "
302 "kernel.perf_event_max_sample_rate to %d\n",
303 avg_local_sample_len
, allowed_ns
>> 1,
304 sysctl_perf_event_sample_rate
);
308 static atomic64_t perf_event_id
;
310 static void cpu_ctx_sched_out(struct perf_cpu_context
*cpuctx
,
311 enum event_type_t event_type
);
313 static void cpu_ctx_sched_in(struct perf_cpu_context
*cpuctx
,
314 enum event_type_t event_type
,
315 struct task_struct
*task
);
317 static void update_context_time(struct perf_event_context
*ctx
);
318 static u64
perf_event_time(struct perf_event
*event
);
320 void __weak
perf_event_print_debug(void) { }
322 extern __weak
const char *perf_pmu_name(void)
327 static inline u64
perf_clock(void)
329 return local_clock();
332 static inline u64
perf_event_clock(struct perf_event
*event
)
334 return event
->clock();
337 static inline struct perf_cpu_context
*
338 __get_cpu_context(struct perf_event_context
*ctx
)
340 return this_cpu_ptr(ctx
->pmu
->pmu_cpu_context
);
343 static void perf_ctx_lock(struct perf_cpu_context
*cpuctx
,
344 struct perf_event_context
*ctx
)
346 raw_spin_lock(&cpuctx
->ctx
.lock
);
348 raw_spin_lock(&ctx
->lock
);
351 static void perf_ctx_unlock(struct perf_cpu_context
*cpuctx
,
352 struct perf_event_context
*ctx
)
355 raw_spin_unlock(&ctx
->lock
);
356 raw_spin_unlock(&cpuctx
->ctx
.lock
);
359 #ifdef CONFIG_CGROUP_PERF
362 perf_cgroup_match(struct perf_event
*event
)
364 struct perf_event_context
*ctx
= event
->ctx
;
365 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
367 /* @event doesn't care about cgroup */
371 /* wants specific cgroup scope but @cpuctx isn't associated with any */
376 * Cgroup scoping is recursive. An event enabled for a cgroup is
377 * also enabled for all its descendant cgroups. If @cpuctx's
378 * cgroup is a descendant of @event's (the test covers identity
379 * case), it's a match.
381 return cgroup_is_descendant(cpuctx
->cgrp
->css
.cgroup
,
382 event
->cgrp
->css
.cgroup
);
385 static inline void perf_detach_cgroup(struct perf_event
*event
)
387 css_put(&event
->cgrp
->css
);
391 static inline int is_cgroup_event(struct perf_event
*event
)
393 return event
->cgrp
!= NULL
;
396 static inline u64
perf_cgroup_event_time(struct perf_event
*event
)
398 struct perf_cgroup_info
*t
;
400 t
= per_cpu_ptr(event
->cgrp
->info
, event
->cpu
);
404 static inline void __update_cgrp_time(struct perf_cgroup
*cgrp
)
406 struct perf_cgroup_info
*info
;
411 info
= this_cpu_ptr(cgrp
->info
);
413 info
->time
+= now
- info
->timestamp
;
414 info
->timestamp
= now
;
417 static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context
*cpuctx
)
419 struct perf_cgroup
*cgrp_out
= cpuctx
->cgrp
;
421 __update_cgrp_time(cgrp_out
);
424 static inline void update_cgrp_time_from_event(struct perf_event
*event
)
426 struct perf_cgroup
*cgrp
;
429 * ensure we access cgroup data only when needed and
430 * when we know the cgroup is pinned (css_get)
432 if (!is_cgroup_event(event
))
435 cgrp
= perf_cgroup_from_task(current
);
437 * Do not update time when cgroup is not active
439 if (cgrp
== event
->cgrp
)
440 __update_cgrp_time(event
->cgrp
);
444 perf_cgroup_set_timestamp(struct task_struct
*task
,
445 struct perf_event_context
*ctx
)
447 struct perf_cgroup
*cgrp
;
448 struct perf_cgroup_info
*info
;
451 * ctx->lock held by caller
452 * ensure we do not access cgroup data
453 * unless we have the cgroup pinned (css_get)
455 if (!task
|| !ctx
->nr_cgroups
)
458 cgrp
= perf_cgroup_from_task(task
);
459 info
= this_cpu_ptr(cgrp
->info
);
460 info
->timestamp
= ctx
->timestamp
;
463 #define PERF_CGROUP_SWOUT 0x1 /* cgroup switch out every event */
464 #define PERF_CGROUP_SWIN 0x2 /* cgroup switch in events based on task */
467 * reschedule events based on the cgroup constraint of task.
469 * mode SWOUT : schedule out everything
470 * mode SWIN : schedule in based on cgroup for next
472 void perf_cgroup_switch(struct task_struct
*task
, int mode
)
474 struct perf_cpu_context
*cpuctx
;
479 * disable interrupts to avoid geting nr_cgroup
480 * changes via __perf_event_disable(). Also
483 local_irq_save(flags
);
486 * we reschedule only in the presence of cgroup
487 * constrained events.
491 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
492 cpuctx
= this_cpu_ptr(pmu
->pmu_cpu_context
);
493 if (cpuctx
->unique_pmu
!= pmu
)
494 continue; /* ensure we process each cpuctx once */
497 * perf_cgroup_events says at least one
498 * context on this CPU has cgroup events.
500 * ctx->nr_cgroups reports the number of cgroup
501 * events for a context.
503 if (cpuctx
->ctx
.nr_cgroups
> 0) {
504 perf_ctx_lock(cpuctx
, cpuctx
->task_ctx
);
505 perf_pmu_disable(cpuctx
->ctx
.pmu
);
507 if (mode
& PERF_CGROUP_SWOUT
) {
508 cpu_ctx_sched_out(cpuctx
, EVENT_ALL
);
510 * must not be done before ctxswout due
511 * to event_filter_match() in event_sched_out()
516 if (mode
& PERF_CGROUP_SWIN
) {
517 WARN_ON_ONCE(cpuctx
->cgrp
);
519 * set cgrp before ctxsw in to allow
520 * event_filter_match() to not have to pass
523 cpuctx
->cgrp
= perf_cgroup_from_task(task
);
524 cpu_ctx_sched_in(cpuctx
, EVENT_ALL
, task
);
526 perf_pmu_enable(cpuctx
->ctx
.pmu
);
527 perf_ctx_unlock(cpuctx
, cpuctx
->task_ctx
);
533 local_irq_restore(flags
);
536 static inline void perf_cgroup_sched_out(struct task_struct
*task
,
537 struct task_struct
*next
)
539 struct perf_cgroup
*cgrp1
;
540 struct perf_cgroup
*cgrp2
= NULL
;
543 * we come here when we know perf_cgroup_events > 0
545 cgrp1
= perf_cgroup_from_task(task
);
548 * next is NULL when called from perf_event_enable_on_exec()
549 * that will systematically cause a cgroup_switch()
552 cgrp2
= perf_cgroup_from_task(next
);
555 * only schedule out current cgroup events if we know
556 * that we are switching to a different cgroup. Otherwise,
557 * do no touch the cgroup events.
560 perf_cgroup_switch(task
, PERF_CGROUP_SWOUT
);
563 static inline void perf_cgroup_sched_in(struct task_struct
*prev
,
564 struct task_struct
*task
)
566 struct perf_cgroup
*cgrp1
;
567 struct perf_cgroup
*cgrp2
= NULL
;
570 * we come here when we know perf_cgroup_events > 0
572 cgrp1
= perf_cgroup_from_task(task
);
574 /* prev can never be NULL */
575 cgrp2
= perf_cgroup_from_task(prev
);
578 * only need to schedule in cgroup events if we are changing
579 * cgroup during ctxsw. Cgroup events were not scheduled
580 * out of ctxsw out if that was not the case.
583 perf_cgroup_switch(task
, PERF_CGROUP_SWIN
);
586 static inline int perf_cgroup_connect(int fd
, struct perf_event
*event
,
587 struct perf_event_attr
*attr
,
588 struct perf_event
*group_leader
)
590 struct perf_cgroup
*cgrp
;
591 struct cgroup_subsys_state
*css
;
592 struct fd f
= fdget(fd
);
598 css
= css_tryget_online_from_dir(f
.file
->f_path
.dentry
,
599 &perf_event_cgrp_subsys
);
605 cgrp
= container_of(css
, struct perf_cgroup
, css
);
609 * all events in a group must monitor
610 * the same cgroup because a task belongs
611 * to only one perf cgroup at a time
613 if (group_leader
&& group_leader
->cgrp
!= cgrp
) {
614 perf_detach_cgroup(event
);
623 perf_cgroup_set_shadow_time(struct perf_event
*event
, u64 now
)
625 struct perf_cgroup_info
*t
;
626 t
= per_cpu_ptr(event
->cgrp
->info
, event
->cpu
);
627 event
->shadow_ctx_time
= now
- t
->timestamp
;
631 perf_cgroup_defer_enabled(struct perf_event
*event
)
634 * when the current task's perf cgroup does not match
635 * the event's, we need to remember to call the
636 * perf_mark_enable() function the first time a task with
637 * a matching perf cgroup is scheduled in.
639 if (is_cgroup_event(event
) && !perf_cgroup_match(event
))
640 event
->cgrp_defer_enabled
= 1;
644 perf_cgroup_mark_enabled(struct perf_event
*event
,
645 struct perf_event_context
*ctx
)
647 struct perf_event
*sub
;
648 u64 tstamp
= perf_event_time(event
);
650 if (!event
->cgrp_defer_enabled
)
653 event
->cgrp_defer_enabled
= 0;
655 event
->tstamp_enabled
= tstamp
- event
->total_time_enabled
;
656 list_for_each_entry(sub
, &event
->sibling_list
, group_entry
) {
657 if (sub
->state
>= PERF_EVENT_STATE_INACTIVE
) {
658 sub
->tstamp_enabled
= tstamp
- sub
->total_time_enabled
;
659 sub
->cgrp_defer_enabled
= 0;
663 #else /* !CONFIG_CGROUP_PERF */
666 perf_cgroup_match(struct perf_event
*event
)
671 static inline void perf_detach_cgroup(struct perf_event
*event
)
674 static inline int is_cgroup_event(struct perf_event
*event
)
679 static inline u64
perf_cgroup_event_cgrp_time(struct perf_event
*event
)
684 static inline void update_cgrp_time_from_event(struct perf_event
*event
)
688 static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context
*cpuctx
)
692 static inline void perf_cgroup_sched_out(struct task_struct
*task
,
693 struct task_struct
*next
)
697 static inline void perf_cgroup_sched_in(struct task_struct
*prev
,
698 struct task_struct
*task
)
702 static inline int perf_cgroup_connect(pid_t pid
, struct perf_event
*event
,
703 struct perf_event_attr
*attr
,
704 struct perf_event
*group_leader
)
710 perf_cgroup_set_timestamp(struct task_struct
*task
,
711 struct perf_event_context
*ctx
)
716 perf_cgroup_switch(struct task_struct
*task
, struct task_struct
*next
)
721 perf_cgroup_set_shadow_time(struct perf_event
*event
, u64 now
)
725 static inline u64
perf_cgroup_event_time(struct perf_event
*event
)
731 perf_cgroup_defer_enabled(struct perf_event
*event
)
736 perf_cgroup_mark_enabled(struct perf_event
*event
,
737 struct perf_event_context
*ctx
)
743 * set default to be dependent on timer tick just
746 #define PERF_CPU_HRTIMER (1000 / HZ)
748 * function must be called with interrupts disbled
750 static enum hrtimer_restart
perf_cpu_hrtimer_handler(struct hrtimer
*hr
)
752 struct perf_cpu_context
*cpuctx
;
753 enum hrtimer_restart ret
= HRTIMER_NORESTART
;
756 WARN_ON(!irqs_disabled());
758 cpuctx
= container_of(hr
, struct perf_cpu_context
, hrtimer
);
760 rotations
= perf_rotate_context(cpuctx
);
763 * arm timer if needed
766 hrtimer_forward_now(hr
, cpuctx
->hrtimer_interval
);
767 ret
= HRTIMER_RESTART
;
773 /* CPU is going down */
774 void perf_cpu_hrtimer_cancel(int cpu
)
776 struct perf_cpu_context
*cpuctx
;
780 if (WARN_ON(cpu
!= smp_processor_id()))
783 local_irq_save(flags
);
787 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
788 cpuctx
= this_cpu_ptr(pmu
->pmu_cpu_context
);
790 if (pmu
->task_ctx_nr
== perf_sw_context
)
793 hrtimer_cancel(&cpuctx
->hrtimer
);
798 local_irq_restore(flags
);
801 static void __perf_cpu_hrtimer_init(struct perf_cpu_context
*cpuctx
, int cpu
)
803 struct hrtimer
*hr
= &cpuctx
->hrtimer
;
804 struct pmu
*pmu
= cpuctx
->ctx
.pmu
;
807 /* no multiplexing needed for SW PMU */
808 if (pmu
->task_ctx_nr
== perf_sw_context
)
812 * check default is sane, if not set then force to
813 * default interval (1/tick)
815 timer
= pmu
->hrtimer_interval_ms
;
817 timer
= pmu
->hrtimer_interval_ms
= PERF_CPU_HRTIMER
;
819 cpuctx
->hrtimer_interval
= ns_to_ktime(NSEC_PER_MSEC
* timer
);
821 hrtimer_init(hr
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL_PINNED
);
822 hr
->function
= perf_cpu_hrtimer_handler
;
825 static void perf_cpu_hrtimer_restart(struct perf_cpu_context
*cpuctx
)
827 struct hrtimer
*hr
= &cpuctx
->hrtimer
;
828 struct pmu
*pmu
= cpuctx
->ctx
.pmu
;
831 if (pmu
->task_ctx_nr
== perf_sw_context
)
834 if (hrtimer_active(hr
))
837 hrtimer_start(hr
, cpuctx
->hrtimer_interval
, HRTIMER_MODE_REL_PINNED
);
840 void perf_pmu_disable(struct pmu
*pmu
)
842 int *count
= this_cpu_ptr(pmu
->pmu_disable_count
);
844 pmu
->pmu_disable(pmu
);
847 void perf_pmu_enable(struct pmu
*pmu
)
849 int *count
= this_cpu_ptr(pmu
->pmu_disable_count
);
851 pmu
->pmu_enable(pmu
);
854 static DEFINE_PER_CPU(struct list_head
, active_ctx_list
);
857 * perf_event_ctx_activate(), perf_event_ctx_deactivate(), and
858 * perf_event_task_tick() are fully serialized because they're strictly cpu
859 * affine and perf_event_ctx{activate,deactivate} are called with IRQs
860 * disabled, while perf_event_task_tick is called from IRQ context.
862 static void perf_event_ctx_activate(struct perf_event_context
*ctx
)
864 struct list_head
*head
= this_cpu_ptr(&active_ctx_list
);
866 WARN_ON(!irqs_disabled());
868 WARN_ON(!list_empty(&ctx
->active_ctx_list
));
870 list_add(&ctx
->active_ctx_list
, head
);
873 static void perf_event_ctx_deactivate(struct perf_event_context
*ctx
)
875 WARN_ON(!irqs_disabled());
877 WARN_ON(list_empty(&ctx
->active_ctx_list
));
879 list_del_init(&ctx
->active_ctx_list
);
882 static void get_ctx(struct perf_event_context
*ctx
)
884 WARN_ON(!atomic_inc_not_zero(&ctx
->refcount
));
887 static void free_ctx(struct rcu_head
*head
)
889 struct perf_event_context
*ctx
;
891 ctx
= container_of(head
, struct perf_event_context
, rcu_head
);
892 kfree(ctx
->task_ctx_data
);
896 static void put_ctx(struct perf_event_context
*ctx
)
898 if (atomic_dec_and_test(&ctx
->refcount
)) {
900 put_ctx(ctx
->parent_ctx
);
902 put_task_struct(ctx
->task
);
903 call_rcu(&ctx
->rcu_head
, free_ctx
);
908 * Because of perf_event::ctx migration in sys_perf_event_open::move_group and
909 * perf_pmu_migrate_context() we need some magic.
911 * Those places that change perf_event::ctx will hold both
912 * perf_event_ctx::mutex of the 'old' and 'new' ctx value.
914 * Lock ordering is by mutex address. There is one other site where
915 * perf_event_context::mutex nests and that is put_event(). But remember that
916 * that is a parent<->child context relation, and migration does not affect
917 * children, therefore these two orderings should not interact.
919 * The change in perf_event::ctx does not affect children (as claimed above)
920 * because the sys_perf_event_open() case will install a new event and break
921 * the ctx parent<->child relation, and perf_pmu_migrate_context() is only
922 * concerned with cpuctx and that doesn't have children.
924 * The places that change perf_event::ctx will issue:
926 * perf_remove_from_context();
928 * perf_install_in_context();
930 * to affect the change. The remove_from_context() + synchronize_rcu() should
931 * quiesce the event, after which we can install it in the new location. This
932 * means that only external vectors (perf_fops, prctl) can perturb the event
933 * while in transit. Therefore all such accessors should also acquire
934 * perf_event_context::mutex to serialize against this.
936 * However; because event->ctx can change while we're waiting to acquire
937 * ctx->mutex we must be careful and use the below perf_event_ctx_lock()
941 * task_struct::perf_event_mutex
942 * perf_event_context::mutex
943 * perf_event_context::lock
944 * perf_event::child_mutex;
945 * perf_event::mmap_mutex
948 static struct perf_event_context
*
949 perf_event_ctx_lock_nested(struct perf_event
*event
, int nesting
)
951 struct perf_event_context
*ctx
;
955 ctx
= ACCESS_ONCE(event
->ctx
);
956 if (!atomic_inc_not_zero(&ctx
->refcount
)) {
962 mutex_lock_nested(&ctx
->mutex
, nesting
);
963 if (event
->ctx
!= ctx
) {
964 mutex_unlock(&ctx
->mutex
);
972 static inline struct perf_event_context
*
973 perf_event_ctx_lock(struct perf_event
*event
)
975 return perf_event_ctx_lock_nested(event
, 0);
978 static void perf_event_ctx_unlock(struct perf_event
*event
,
979 struct perf_event_context
*ctx
)
981 mutex_unlock(&ctx
->mutex
);
986 * This must be done under the ctx->lock, such as to serialize against
987 * context_equiv(), therefore we cannot call put_ctx() since that might end up
988 * calling scheduler related locks and ctx->lock nests inside those.
990 static __must_check
struct perf_event_context
*
991 unclone_ctx(struct perf_event_context
*ctx
)
993 struct perf_event_context
*parent_ctx
= ctx
->parent_ctx
;
995 lockdep_assert_held(&ctx
->lock
);
998 ctx
->parent_ctx
= NULL
;
1004 static u32
perf_event_pid(struct perf_event
*event
, struct task_struct
*p
)
1007 * only top level events have the pid namespace they were created in
1010 event
= event
->parent
;
1012 return task_tgid_nr_ns(p
, event
->ns
);
1015 static u32
perf_event_tid(struct perf_event
*event
, struct task_struct
*p
)
1018 * only top level events have the pid namespace they were created in
1021 event
= event
->parent
;
1023 return task_pid_nr_ns(p
, event
->ns
);
1027 * If we inherit events we want to return the parent event id
1030 static u64
primary_event_id(struct perf_event
*event
)
1035 id
= event
->parent
->id
;
1041 * Get the perf_event_context for a task and lock it.
1042 * This has to cope with with the fact that until it is locked,
1043 * the context could get moved to another task.
1045 static struct perf_event_context
*
1046 perf_lock_task_context(struct task_struct
*task
, int ctxn
, unsigned long *flags
)
1048 struct perf_event_context
*ctx
;
1052 * One of the few rules of preemptible RCU is that one cannot do
1053 * rcu_read_unlock() while holding a scheduler (or nested) lock when
1054 * part of the read side critical section was preemptible -- see
1055 * rcu_read_unlock_special().
1057 * Since ctx->lock nests under rq->lock we must ensure the entire read
1058 * side critical section is non-preemptible.
1062 ctx
= rcu_dereference(task
->perf_event_ctxp
[ctxn
]);
1065 * If this context is a clone of another, it might
1066 * get swapped for another underneath us by
1067 * perf_event_task_sched_out, though the
1068 * rcu_read_lock() protects us from any context
1069 * getting freed. Lock the context and check if it
1070 * got swapped before we could get the lock, and retry
1071 * if so. If we locked the right context, then it
1072 * can't get swapped on us any more.
1074 raw_spin_lock_irqsave(&ctx
->lock
, *flags
);
1075 if (ctx
!= rcu_dereference(task
->perf_event_ctxp
[ctxn
])) {
1076 raw_spin_unlock_irqrestore(&ctx
->lock
, *flags
);
1082 if (!atomic_inc_not_zero(&ctx
->refcount
)) {
1083 raw_spin_unlock_irqrestore(&ctx
->lock
, *flags
);
1093 * Get the context for a task and increment its pin_count so it
1094 * can't get swapped to another task. This also increments its
1095 * reference count so that the context can't get freed.
1097 static struct perf_event_context
*
1098 perf_pin_task_context(struct task_struct
*task
, int ctxn
)
1100 struct perf_event_context
*ctx
;
1101 unsigned long flags
;
1103 ctx
= perf_lock_task_context(task
, ctxn
, &flags
);
1106 raw_spin_unlock_irqrestore(&ctx
->lock
, flags
);
1111 static void perf_unpin_context(struct perf_event_context
*ctx
)
1113 unsigned long flags
;
1115 raw_spin_lock_irqsave(&ctx
->lock
, flags
);
1117 raw_spin_unlock_irqrestore(&ctx
->lock
, flags
);
1121 * Update the record of the current time in a context.
1123 static void update_context_time(struct perf_event_context
*ctx
)
1125 u64 now
= perf_clock();
1127 ctx
->time
+= now
- ctx
->timestamp
;
1128 ctx
->timestamp
= now
;
1131 static u64
perf_event_time(struct perf_event
*event
)
1133 struct perf_event_context
*ctx
= event
->ctx
;
1135 if (is_cgroup_event(event
))
1136 return perf_cgroup_event_time(event
);
1138 return ctx
? ctx
->time
: 0;
1142 * Update the total_time_enabled and total_time_running fields for a event.
1143 * The caller of this function needs to hold the ctx->lock.
1145 static void update_event_times(struct perf_event
*event
)
1147 struct perf_event_context
*ctx
= event
->ctx
;
1150 if (event
->state
< PERF_EVENT_STATE_INACTIVE
||
1151 event
->group_leader
->state
< PERF_EVENT_STATE_INACTIVE
)
1154 * in cgroup mode, time_enabled represents
1155 * the time the event was enabled AND active
1156 * tasks were in the monitored cgroup. This is
1157 * independent of the activity of the context as
1158 * there may be a mix of cgroup and non-cgroup events.
1160 * That is why we treat cgroup events differently
1163 if (is_cgroup_event(event
))
1164 run_end
= perf_cgroup_event_time(event
);
1165 else if (ctx
->is_active
)
1166 run_end
= ctx
->time
;
1168 run_end
= event
->tstamp_stopped
;
1170 event
->total_time_enabled
= run_end
- event
->tstamp_enabled
;
1172 if (event
->state
== PERF_EVENT_STATE_INACTIVE
)
1173 run_end
= event
->tstamp_stopped
;
1175 run_end
= perf_event_time(event
);
1177 event
->total_time_running
= run_end
- event
->tstamp_running
;
1182 * Update total_time_enabled and total_time_running for all events in a group.
1184 static void update_group_times(struct perf_event
*leader
)
1186 struct perf_event
*event
;
1188 update_event_times(leader
);
1189 list_for_each_entry(event
, &leader
->sibling_list
, group_entry
)
1190 update_event_times(event
);
1193 static struct list_head
*
1194 ctx_group_list(struct perf_event
*event
, struct perf_event_context
*ctx
)
1196 if (event
->attr
.pinned
)
1197 return &ctx
->pinned_groups
;
1199 return &ctx
->flexible_groups
;
1203 * Add a event from the lists for its context.
1204 * Must be called with ctx->mutex and ctx->lock held.
1207 list_add_event(struct perf_event
*event
, struct perf_event_context
*ctx
)
1209 WARN_ON_ONCE(event
->attach_state
& PERF_ATTACH_CONTEXT
);
1210 event
->attach_state
|= PERF_ATTACH_CONTEXT
;
1213 * If we're a stand alone event or group leader, we go to the context
1214 * list, group events are kept attached to the group so that
1215 * perf_group_detach can, at all times, locate all siblings.
1217 if (event
->group_leader
== event
) {
1218 struct list_head
*list
;
1220 if (is_software_event(event
))
1221 event
->group_flags
|= PERF_GROUP_SOFTWARE
;
1223 list
= ctx_group_list(event
, ctx
);
1224 list_add_tail(&event
->group_entry
, list
);
1227 if (is_cgroup_event(event
))
1230 list_add_rcu(&event
->event_entry
, &ctx
->event_list
);
1232 if (event
->attr
.inherit_stat
)
1239 * Initialize event state based on the perf_event_attr::disabled.
1241 static inline void perf_event__state_init(struct perf_event
*event
)
1243 event
->state
= event
->attr
.disabled
? PERF_EVENT_STATE_OFF
:
1244 PERF_EVENT_STATE_INACTIVE
;
1248 * Called at perf_event creation and when events are attached/detached from a
1251 static void perf_event__read_size(struct perf_event
*event
)
1253 int entry
= sizeof(u64
); /* value */
1257 if (event
->attr
.read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
1258 size
+= sizeof(u64
);
1260 if (event
->attr
.read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
1261 size
+= sizeof(u64
);
1263 if (event
->attr
.read_format
& PERF_FORMAT_ID
)
1264 entry
+= sizeof(u64
);
1266 if (event
->attr
.read_format
& PERF_FORMAT_GROUP
) {
1267 nr
+= event
->group_leader
->nr_siblings
;
1268 size
+= sizeof(u64
);
1272 event
->read_size
= size
;
1275 static void perf_event__header_size(struct perf_event
*event
)
1277 struct perf_sample_data
*data
;
1278 u64 sample_type
= event
->attr
.sample_type
;
1281 perf_event__read_size(event
);
1283 if (sample_type
& PERF_SAMPLE_IP
)
1284 size
+= sizeof(data
->ip
);
1286 if (sample_type
& PERF_SAMPLE_ADDR
)
1287 size
+= sizeof(data
->addr
);
1289 if (sample_type
& PERF_SAMPLE_PERIOD
)
1290 size
+= sizeof(data
->period
);
1292 if (sample_type
& PERF_SAMPLE_WEIGHT
)
1293 size
+= sizeof(data
->weight
);
1295 if (sample_type
& PERF_SAMPLE_READ
)
1296 size
+= event
->read_size
;
1298 if (sample_type
& PERF_SAMPLE_DATA_SRC
)
1299 size
+= sizeof(data
->data_src
.val
);
1301 if (sample_type
& PERF_SAMPLE_TRANSACTION
)
1302 size
+= sizeof(data
->txn
);
1304 event
->header_size
= size
;
1307 static void perf_event__id_header_size(struct perf_event
*event
)
1309 struct perf_sample_data
*data
;
1310 u64 sample_type
= event
->attr
.sample_type
;
1313 if (sample_type
& PERF_SAMPLE_TID
)
1314 size
+= sizeof(data
->tid_entry
);
1316 if (sample_type
& PERF_SAMPLE_TIME
)
1317 size
+= sizeof(data
->time
);
1319 if (sample_type
& PERF_SAMPLE_IDENTIFIER
)
1320 size
+= sizeof(data
->id
);
1322 if (sample_type
& PERF_SAMPLE_ID
)
1323 size
+= sizeof(data
->id
);
1325 if (sample_type
& PERF_SAMPLE_STREAM_ID
)
1326 size
+= sizeof(data
->stream_id
);
1328 if (sample_type
& PERF_SAMPLE_CPU
)
1329 size
+= sizeof(data
->cpu_entry
);
1331 event
->id_header_size
= size
;
1334 static void perf_group_attach(struct perf_event
*event
)
1336 struct perf_event
*group_leader
= event
->group_leader
, *pos
;
1339 * We can have double attach due to group movement in perf_event_open.
1341 if (event
->attach_state
& PERF_ATTACH_GROUP
)
1344 event
->attach_state
|= PERF_ATTACH_GROUP
;
1346 if (group_leader
== event
)
1349 WARN_ON_ONCE(group_leader
->ctx
!= event
->ctx
);
1351 if (group_leader
->group_flags
& PERF_GROUP_SOFTWARE
&&
1352 !is_software_event(event
))
1353 group_leader
->group_flags
&= ~PERF_GROUP_SOFTWARE
;
1355 list_add_tail(&event
->group_entry
, &group_leader
->sibling_list
);
1356 group_leader
->nr_siblings
++;
1358 perf_event__header_size(group_leader
);
1360 list_for_each_entry(pos
, &group_leader
->sibling_list
, group_entry
)
1361 perf_event__header_size(pos
);
1365 * Remove a event from the lists for its context.
1366 * Must be called with ctx->mutex and ctx->lock held.
1369 list_del_event(struct perf_event
*event
, struct perf_event_context
*ctx
)
1371 struct perf_cpu_context
*cpuctx
;
1373 WARN_ON_ONCE(event
->ctx
!= ctx
);
1374 lockdep_assert_held(&ctx
->lock
);
1377 * We can have double detach due to exit/hot-unplug + close.
1379 if (!(event
->attach_state
& PERF_ATTACH_CONTEXT
))
1382 event
->attach_state
&= ~PERF_ATTACH_CONTEXT
;
1384 if (is_cgroup_event(event
)) {
1386 cpuctx
= __get_cpu_context(ctx
);
1388 * if there are no more cgroup events
1389 * then cler cgrp to avoid stale pointer
1390 * in update_cgrp_time_from_cpuctx()
1392 if (!ctx
->nr_cgroups
)
1393 cpuctx
->cgrp
= NULL
;
1397 if (event
->attr
.inherit_stat
)
1400 list_del_rcu(&event
->event_entry
);
1402 if (event
->group_leader
== event
)
1403 list_del_init(&event
->group_entry
);
1405 update_group_times(event
);
1408 * If event was in error state, then keep it
1409 * that way, otherwise bogus counts will be
1410 * returned on read(). The only way to get out
1411 * of error state is by explicit re-enabling
1414 if (event
->state
> PERF_EVENT_STATE_OFF
)
1415 event
->state
= PERF_EVENT_STATE_OFF
;
1420 static void perf_group_detach(struct perf_event
*event
)
1422 struct perf_event
*sibling
, *tmp
;
1423 struct list_head
*list
= NULL
;
1426 * We can have double detach due to exit/hot-unplug + close.
1428 if (!(event
->attach_state
& PERF_ATTACH_GROUP
))
1431 event
->attach_state
&= ~PERF_ATTACH_GROUP
;
1434 * If this is a sibling, remove it from its group.
1436 if (event
->group_leader
!= event
) {
1437 list_del_init(&event
->group_entry
);
1438 event
->group_leader
->nr_siblings
--;
1442 if (!list_empty(&event
->group_entry
))
1443 list
= &event
->group_entry
;
1446 * If this was a group event with sibling events then
1447 * upgrade the siblings to singleton events by adding them
1448 * to whatever list we are on.
1450 list_for_each_entry_safe(sibling
, tmp
, &event
->sibling_list
, group_entry
) {
1452 list_move_tail(&sibling
->group_entry
, list
);
1453 sibling
->group_leader
= sibling
;
1455 /* Inherit group flags from the previous leader */
1456 sibling
->group_flags
= event
->group_flags
;
1458 WARN_ON_ONCE(sibling
->ctx
!= event
->ctx
);
1462 perf_event__header_size(event
->group_leader
);
1464 list_for_each_entry(tmp
, &event
->group_leader
->sibling_list
, group_entry
)
1465 perf_event__header_size(tmp
);
1469 * User event without the task.
1471 static bool is_orphaned_event(struct perf_event
*event
)
1473 return event
&& !is_kernel_event(event
) && !event
->owner
;
1477 * Event has a parent but parent's task finished and it's
1478 * alive only because of children holding refference.
1480 static bool is_orphaned_child(struct perf_event
*event
)
1482 return is_orphaned_event(event
->parent
);
1485 static void orphans_remove_work(struct work_struct
*work
);
1487 static void schedule_orphans_remove(struct perf_event_context
*ctx
)
1489 if (!ctx
->task
|| ctx
->orphans_remove_sched
|| !perf_wq
)
1492 if (queue_delayed_work(perf_wq
, &ctx
->orphans_remove
, 1)) {
1494 ctx
->orphans_remove_sched
= true;
1498 static int __init
perf_workqueue_init(void)
1500 perf_wq
= create_singlethread_workqueue("perf");
1501 WARN(!perf_wq
, "failed to create perf workqueue\n");
1502 return perf_wq
? 0 : -1;
1505 core_initcall(perf_workqueue_init
);
1508 event_filter_match(struct perf_event
*event
)
1510 return (event
->cpu
== -1 || event
->cpu
== smp_processor_id())
1511 && perf_cgroup_match(event
);
1515 event_sched_out(struct perf_event
*event
,
1516 struct perf_cpu_context
*cpuctx
,
1517 struct perf_event_context
*ctx
)
1519 u64 tstamp
= perf_event_time(event
);
1522 WARN_ON_ONCE(event
->ctx
!= ctx
);
1523 lockdep_assert_held(&ctx
->lock
);
1526 * An event which could not be activated because of
1527 * filter mismatch still needs to have its timings
1528 * maintained, otherwise bogus information is return
1529 * via read() for time_enabled, time_running:
1531 if (event
->state
== PERF_EVENT_STATE_INACTIVE
1532 && !event_filter_match(event
)) {
1533 delta
= tstamp
- event
->tstamp_stopped
;
1534 event
->tstamp_running
+= delta
;
1535 event
->tstamp_stopped
= tstamp
;
1538 if (event
->state
!= PERF_EVENT_STATE_ACTIVE
)
1541 perf_pmu_disable(event
->pmu
);
1543 event
->state
= PERF_EVENT_STATE_INACTIVE
;
1544 if (event
->pending_disable
) {
1545 event
->pending_disable
= 0;
1546 event
->state
= PERF_EVENT_STATE_OFF
;
1548 event
->tstamp_stopped
= tstamp
;
1549 event
->pmu
->del(event
, 0);
1552 if (!is_software_event(event
))
1553 cpuctx
->active_oncpu
--;
1554 if (!--ctx
->nr_active
)
1555 perf_event_ctx_deactivate(ctx
);
1556 if (event
->attr
.freq
&& event
->attr
.sample_freq
)
1558 if (event
->attr
.exclusive
|| !cpuctx
->active_oncpu
)
1559 cpuctx
->exclusive
= 0;
1561 if (is_orphaned_child(event
))
1562 schedule_orphans_remove(ctx
);
1564 perf_pmu_enable(event
->pmu
);
1568 group_sched_out(struct perf_event
*group_event
,
1569 struct perf_cpu_context
*cpuctx
,
1570 struct perf_event_context
*ctx
)
1572 struct perf_event
*event
;
1573 int state
= group_event
->state
;
1575 event_sched_out(group_event
, cpuctx
, ctx
);
1578 * Schedule out siblings (if any):
1580 list_for_each_entry(event
, &group_event
->sibling_list
, group_entry
)
1581 event_sched_out(event
, cpuctx
, ctx
);
1583 if (state
== PERF_EVENT_STATE_ACTIVE
&& group_event
->attr
.exclusive
)
1584 cpuctx
->exclusive
= 0;
1587 struct remove_event
{
1588 struct perf_event
*event
;
1593 * Cross CPU call to remove a performance event
1595 * We disable the event on the hardware level first. After that we
1596 * remove it from the context list.
1598 static int __perf_remove_from_context(void *info
)
1600 struct remove_event
*re
= info
;
1601 struct perf_event
*event
= re
->event
;
1602 struct perf_event_context
*ctx
= event
->ctx
;
1603 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
1605 raw_spin_lock(&ctx
->lock
);
1606 event_sched_out(event
, cpuctx
, ctx
);
1607 if (re
->detach_group
)
1608 perf_group_detach(event
);
1609 list_del_event(event
, ctx
);
1610 if (!ctx
->nr_events
&& cpuctx
->task_ctx
== ctx
) {
1612 cpuctx
->task_ctx
= NULL
;
1614 raw_spin_unlock(&ctx
->lock
);
1621 * Remove the event from a task's (or a CPU's) list of events.
1623 * CPU events are removed with a smp call. For task events we only
1624 * call when the task is on a CPU.
1626 * If event->ctx is a cloned context, callers must make sure that
1627 * every task struct that event->ctx->task could possibly point to
1628 * remains valid. This is OK when called from perf_release since
1629 * that only calls us on the top-level context, which can't be a clone.
1630 * When called from perf_event_exit_task, it's OK because the
1631 * context has been detached from its task.
1633 static void perf_remove_from_context(struct perf_event
*event
, bool detach_group
)
1635 struct perf_event_context
*ctx
= event
->ctx
;
1636 struct task_struct
*task
= ctx
->task
;
1637 struct remove_event re
= {
1639 .detach_group
= detach_group
,
1642 lockdep_assert_held(&ctx
->mutex
);
1646 * Per cpu events are removed via an smp call. The removal can
1647 * fail if the CPU is currently offline, but in that case we
1648 * already called __perf_remove_from_context from
1649 * perf_event_exit_cpu.
1651 cpu_function_call(event
->cpu
, __perf_remove_from_context
, &re
);
1656 if (!task_function_call(task
, __perf_remove_from_context
, &re
))
1659 raw_spin_lock_irq(&ctx
->lock
);
1661 * If we failed to find a running task, but find the context active now
1662 * that we've acquired the ctx->lock, retry.
1664 if (ctx
->is_active
) {
1665 raw_spin_unlock_irq(&ctx
->lock
);
1667 * Reload the task pointer, it might have been changed by
1668 * a concurrent perf_event_context_sched_out().
1675 * Since the task isn't running, its safe to remove the event, us
1676 * holding the ctx->lock ensures the task won't get scheduled in.
1679 perf_group_detach(event
);
1680 list_del_event(event
, ctx
);
1681 raw_spin_unlock_irq(&ctx
->lock
);
1685 * Cross CPU call to disable a performance event
1687 int __perf_event_disable(void *info
)
1689 struct perf_event
*event
= info
;
1690 struct perf_event_context
*ctx
= event
->ctx
;
1691 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
1694 * If this is a per-task event, need to check whether this
1695 * event's task is the current task on this cpu.
1697 * Can trigger due to concurrent perf_event_context_sched_out()
1698 * flipping contexts around.
1700 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
)
1703 raw_spin_lock(&ctx
->lock
);
1706 * If the event is on, turn it off.
1707 * If it is in error state, leave it in error state.
1709 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
) {
1710 update_context_time(ctx
);
1711 update_cgrp_time_from_event(event
);
1712 update_group_times(event
);
1713 if (event
== event
->group_leader
)
1714 group_sched_out(event
, cpuctx
, ctx
);
1716 event_sched_out(event
, cpuctx
, ctx
);
1717 event
->state
= PERF_EVENT_STATE_OFF
;
1720 raw_spin_unlock(&ctx
->lock
);
1728 * If event->ctx is a cloned context, callers must make sure that
1729 * every task struct that event->ctx->task could possibly point to
1730 * remains valid. This condition is satisifed when called through
1731 * perf_event_for_each_child or perf_event_for_each because they
1732 * hold the top-level event's child_mutex, so any descendant that
1733 * goes to exit will block in sync_child_event.
1734 * When called from perf_pending_event it's OK because event->ctx
1735 * is the current context on this CPU and preemption is disabled,
1736 * hence we can't get into perf_event_task_sched_out for this context.
1738 static void _perf_event_disable(struct perf_event
*event
)
1740 struct perf_event_context
*ctx
= event
->ctx
;
1741 struct task_struct
*task
= ctx
->task
;
1745 * Disable the event on the cpu that it's on
1747 cpu_function_call(event
->cpu
, __perf_event_disable
, event
);
1752 if (!task_function_call(task
, __perf_event_disable
, event
))
1755 raw_spin_lock_irq(&ctx
->lock
);
1757 * If the event is still active, we need to retry the cross-call.
1759 if (event
->state
== PERF_EVENT_STATE_ACTIVE
) {
1760 raw_spin_unlock_irq(&ctx
->lock
);
1762 * Reload the task pointer, it might have been changed by
1763 * a concurrent perf_event_context_sched_out().
1770 * Since we have the lock this context can't be scheduled
1771 * in, so we can change the state safely.
1773 if (event
->state
== PERF_EVENT_STATE_INACTIVE
) {
1774 update_group_times(event
);
1775 event
->state
= PERF_EVENT_STATE_OFF
;
1777 raw_spin_unlock_irq(&ctx
->lock
);
1781 * Strictly speaking kernel users cannot create groups and therefore this
1782 * interface does not need the perf_event_ctx_lock() magic.
1784 void perf_event_disable(struct perf_event
*event
)
1786 struct perf_event_context
*ctx
;
1788 ctx
= perf_event_ctx_lock(event
);
1789 _perf_event_disable(event
);
1790 perf_event_ctx_unlock(event
, ctx
);
1792 EXPORT_SYMBOL_GPL(perf_event_disable
);
1794 static void perf_set_shadow_time(struct perf_event
*event
,
1795 struct perf_event_context
*ctx
,
1799 * use the correct time source for the time snapshot
1801 * We could get by without this by leveraging the
1802 * fact that to get to this function, the caller
1803 * has most likely already called update_context_time()
1804 * and update_cgrp_time_xx() and thus both timestamp
1805 * are identical (or very close). Given that tstamp is,
1806 * already adjusted for cgroup, we could say that:
1807 * tstamp - ctx->timestamp
1809 * tstamp - cgrp->timestamp.
1811 * Then, in perf_output_read(), the calculation would
1812 * work with no changes because:
1813 * - event is guaranteed scheduled in
1814 * - no scheduled out in between
1815 * - thus the timestamp would be the same
1817 * But this is a bit hairy.
1819 * So instead, we have an explicit cgroup call to remain
1820 * within the time time source all along. We believe it
1821 * is cleaner and simpler to understand.
1823 if (is_cgroup_event(event
))
1824 perf_cgroup_set_shadow_time(event
, tstamp
);
1826 event
->shadow_ctx_time
= tstamp
- ctx
->timestamp
;
1829 #define MAX_INTERRUPTS (~0ULL)
1831 static void perf_log_throttle(struct perf_event
*event
, int enable
);
1832 static void perf_log_itrace_start(struct perf_event
*event
);
1835 event_sched_in(struct perf_event
*event
,
1836 struct perf_cpu_context
*cpuctx
,
1837 struct perf_event_context
*ctx
)
1839 u64 tstamp
= perf_event_time(event
);
1842 lockdep_assert_held(&ctx
->lock
);
1844 if (event
->state
<= PERF_EVENT_STATE_OFF
)
1847 event
->state
= PERF_EVENT_STATE_ACTIVE
;
1848 event
->oncpu
= smp_processor_id();
1851 * Unthrottle events, since we scheduled we might have missed several
1852 * ticks already, also for a heavily scheduling task there is little
1853 * guarantee it'll get a tick in a timely manner.
1855 if (unlikely(event
->hw
.interrupts
== MAX_INTERRUPTS
)) {
1856 perf_log_throttle(event
, 1);
1857 event
->hw
.interrupts
= 0;
1861 * The new state must be visible before we turn it on in the hardware:
1865 perf_pmu_disable(event
->pmu
);
1867 event
->tstamp_running
+= tstamp
- event
->tstamp_stopped
;
1869 perf_set_shadow_time(event
, ctx
, tstamp
);
1871 perf_log_itrace_start(event
);
1873 if (event
->pmu
->add(event
, PERF_EF_START
)) {
1874 event
->state
= PERF_EVENT_STATE_INACTIVE
;
1880 if (!is_software_event(event
))
1881 cpuctx
->active_oncpu
++;
1882 if (!ctx
->nr_active
++)
1883 perf_event_ctx_activate(ctx
);
1884 if (event
->attr
.freq
&& event
->attr
.sample_freq
)
1887 if (event
->attr
.exclusive
)
1888 cpuctx
->exclusive
= 1;
1890 if (is_orphaned_child(event
))
1891 schedule_orphans_remove(ctx
);
1894 perf_pmu_enable(event
->pmu
);
1900 group_sched_in(struct perf_event
*group_event
,
1901 struct perf_cpu_context
*cpuctx
,
1902 struct perf_event_context
*ctx
)
1904 struct perf_event
*event
, *partial_group
= NULL
;
1905 struct pmu
*pmu
= ctx
->pmu
;
1906 u64 now
= ctx
->time
;
1907 bool simulate
= false;
1909 if (group_event
->state
== PERF_EVENT_STATE_OFF
)
1912 pmu
->start_txn(pmu
);
1914 if (event_sched_in(group_event
, cpuctx
, ctx
)) {
1915 pmu
->cancel_txn(pmu
);
1916 perf_cpu_hrtimer_restart(cpuctx
);
1921 * Schedule in siblings as one group (if any):
1923 list_for_each_entry(event
, &group_event
->sibling_list
, group_entry
) {
1924 if (event_sched_in(event
, cpuctx
, ctx
)) {
1925 partial_group
= event
;
1930 if (!pmu
->commit_txn(pmu
))
1935 * Groups can be scheduled in as one unit only, so undo any
1936 * partial group before returning:
1937 * The events up to the failed event are scheduled out normally,
1938 * tstamp_stopped will be updated.
1940 * The failed events and the remaining siblings need to have
1941 * their timings updated as if they had gone thru event_sched_in()
1942 * and event_sched_out(). This is required to get consistent timings
1943 * across the group. This also takes care of the case where the group
1944 * could never be scheduled by ensuring tstamp_stopped is set to mark
1945 * the time the event was actually stopped, such that time delta
1946 * calculation in update_event_times() is correct.
1948 list_for_each_entry(event
, &group_event
->sibling_list
, group_entry
) {
1949 if (event
== partial_group
)
1953 event
->tstamp_running
+= now
- event
->tstamp_stopped
;
1954 event
->tstamp_stopped
= now
;
1956 event_sched_out(event
, cpuctx
, ctx
);
1959 event_sched_out(group_event
, cpuctx
, ctx
);
1961 pmu
->cancel_txn(pmu
);
1963 perf_cpu_hrtimer_restart(cpuctx
);
1969 * Work out whether we can put this event group on the CPU now.
1971 static int group_can_go_on(struct perf_event
*event
,
1972 struct perf_cpu_context
*cpuctx
,
1976 * Groups consisting entirely of software events can always go on.
1978 if (event
->group_flags
& PERF_GROUP_SOFTWARE
)
1981 * If an exclusive group is already on, no other hardware
1984 if (cpuctx
->exclusive
)
1987 * If this group is exclusive and there are already
1988 * events on the CPU, it can't go on.
1990 if (event
->attr
.exclusive
&& cpuctx
->active_oncpu
)
1993 * Otherwise, try to add it if all previous groups were able
1999 static void add_event_to_ctx(struct perf_event
*event
,
2000 struct perf_event_context
*ctx
)
2002 u64 tstamp
= perf_event_time(event
);
2004 list_add_event(event
, ctx
);
2005 perf_group_attach(event
);
2006 event
->tstamp_enabled
= tstamp
;
2007 event
->tstamp_running
= tstamp
;
2008 event
->tstamp_stopped
= tstamp
;
2011 static void task_ctx_sched_out(struct perf_event_context
*ctx
);
2013 ctx_sched_in(struct perf_event_context
*ctx
,
2014 struct perf_cpu_context
*cpuctx
,
2015 enum event_type_t event_type
,
2016 struct task_struct
*task
);
2018 static void perf_event_sched_in(struct perf_cpu_context
*cpuctx
,
2019 struct perf_event_context
*ctx
,
2020 struct task_struct
*task
)
2022 cpu_ctx_sched_in(cpuctx
, EVENT_PINNED
, task
);
2024 ctx_sched_in(ctx
, cpuctx
, EVENT_PINNED
, task
);
2025 cpu_ctx_sched_in(cpuctx
, EVENT_FLEXIBLE
, task
);
2027 ctx_sched_in(ctx
, cpuctx
, EVENT_FLEXIBLE
, task
);
2031 * Cross CPU call to install and enable a performance event
2033 * Must be called with ctx->mutex held
2035 static int __perf_install_in_context(void *info
)
2037 struct perf_event
*event
= info
;
2038 struct perf_event_context
*ctx
= event
->ctx
;
2039 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
2040 struct perf_event_context
*task_ctx
= cpuctx
->task_ctx
;
2041 struct task_struct
*task
= current
;
2043 perf_ctx_lock(cpuctx
, task_ctx
);
2044 perf_pmu_disable(cpuctx
->ctx
.pmu
);
2047 * If there was an active task_ctx schedule it out.
2050 task_ctx_sched_out(task_ctx
);
2053 * If the context we're installing events in is not the
2054 * active task_ctx, flip them.
2056 if (ctx
->task
&& task_ctx
!= ctx
) {
2058 raw_spin_unlock(&task_ctx
->lock
);
2059 raw_spin_lock(&ctx
->lock
);
2064 cpuctx
->task_ctx
= task_ctx
;
2065 task
= task_ctx
->task
;
2068 cpu_ctx_sched_out(cpuctx
, EVENT_ALL
);
2070 update_context_time(ctx
);
2072 * update cgrp time only if current cgrp
2073 * matches event->cgrp. Must be done before
2074 * calling add_event_to_ctx()
2076 update_cgrp_time_from_event(event
);
2078 add_event_to_ctx(event
, ctx
);
2081 * Schedule everything back in
2083 perf_event_sched_in(cpuctx
, task_ctx
, task
);
2085 perf_pmu_enable(cpuctx
->ctx
.pmu
);
2086 perf_ctx_unlock(cpuctx
, task_ctx
);
2092 * Attach a performance event to a context
2094 * First we add the event to the list with the hardware enable bit
2095 * in event->hw_config cleared.
2097 * If the event is attached to a task which is on a CPU we use a smp
2098 * call to enable it in the task context. The task might have been
2099 * scheduled away, but we check this in the smp call again.
2102 perf_install_in_context(struct perf_event_context
*ctx
,
2103 struct perf_event
*event
,
2106 struct task_struct
*task
= ctx
->task
;
2108 lockdep_assert_held(&ctx
->mutex
);
2111 if (event
->cpu
!= -1)
2116 * Per cpu events are installed via an smp call and
2117 * the install is always successful.
2119 cpu_function_call(cpu
, __perf_install_in_context
, event
);
2124 if (!task_function_call(task
, __perf_install_in_context
, event
))
2127 raw_spin_lock_irq(&ctx
->lock
);
2129 * If we failed to find a running task, but find the context active now
2130 * that we've acquired the ctx->lock, retry.
2132 if (ctx
->is_active
) {
2133 raw_spin_unlock_irq(&ctx
->lock
);
2135 * Reload the task pointer, it might have been changed by
2136 * a concurrent perf_event_context_sched_out().
2143 * Since the task isn't running, its safe to add the event, us holding
2144 * the ctx->lock ensures the task won't get scheduled in.
2146 add_event_to_ctx(event
, ctx
);
2147 raw_spin_unlock_irq(&ctx
->lock
);
2151 * Put a event into inactive state and update time fields.
2152 * Enabling the leader of a group effectively enables all
2153 * the group members that aren't explicitly disabled, so we
2154 * have to update their ->tstamp_enabled also.
2155 * Note: this works for group members as well as group leaders
2156 * since the non-leader members' sibling_lists will be empty.
2158 static void __perf_event_mark_enabled(struct perf_event
*event
)
2160 struct perf_event
*sub
;
2161 u64 tstamp
= perf_event_time(event
);
2163 event
->state
= PERF_EVENT_STATE_INACTIVE
;
2164 event
->tstamp_enabled
= tstamp
- event
->total_time_enabled
;
2165 list_for_each_entry(sub
, &event
->sibling_list
, group_entry
) {
2166 if (sub
->state
>= PERF_EVENT_STATE_INACTIVE
)
2167 sub
->tstamp_enabled
= tstamp
- sub
->total_time_enabled
;
2172 * Cross CPU call to enable a performance event
2174 static int __perf_event_enable(void *info
)
2176 struct perf_event
*event
= info
;
2177 struct perf_event_context
*ctx
= event
->ctx
;
2178 struct perf_event
*leader
= event
->group_leader
;
2179 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
2183 * There's a time window between 'ctx->is_active' check
2184 * in perf_event_enable function and this place having:
2186 * - ctx->lock unlocked
2188 * where the task could be killed and 'ctx' deactivated
2189 * by perf_event_exit_task.
2191 if (!ctx
->is_active
)
2194 raw_spin_lock(&ctx
->lock
);
2195 update_context_time(ctx
);
2197 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
)
2201 * set current task's cgroup time reference point
2203 perf_cgroup_set_timestamp(current
, ctx
);
2205 __perf_event_mark_enabled(event
);
2207 if (!event_filter_match(event
)) {
2208 if (is_cgroup_event(event
))
2209 perf_cgroup_defer_enabled(event
);
2214 * If the event is in a group and isn't the group leader,
2215 * then don't put it on unless the group is on.
2217 if (leader
!= event
&& leader
->state
!= PERF_EVENT_STATE_ACTIVE
)
2220 if (!group_can_go_on(event
, cpuctx
, 1)) {
2223 if (event
== leader
)
2224 err
= group_sched_in(event
, cpuctx
, ctx
);
2226 err
= event_sched_in(event
, cpuctx
, ctx
);
2231 * If this event can't go on and it's part of a
2232 * group, then the whole group has to come off.
2234 if (leader
!= event
) {
2235 group_sched_out(leader
, cpuctx
, ctx
);
2236 perf_cpu_hrtimer_restart(cpuctx
);
2238 if (leader
->attr
.pinned
) {
2239 update_group_times(leader
);
2240 leader
->state
= PERF_EVENT_STATE_ERROR
;
2245 raw_spin_unlock(&ctx
->lock
);
2253 * If event->ctx is a cloned context, callers must make sure that
2254 * every task struct that event->ctx->task could possibly point to
2255 * remains valid. This condition is satisfied when called through
2256 * perf_event_for_each_child or perf_event_for_each as described
2257 * for perf_event_disable.
2259 static void _perf_event_enable(struct perf_event
*event
)
2261 struct perf_event_context
*ctx
= event
->ctx
;
2262 struct task_struct
*task
= ctx
->task
;
2266 * Enable the event on the cpu that it's on
2268 cpu_function_call(event
->cpu
, __perf_event_enable
, event
);
2272 raw_spin_lock_irq(&ctx
->lock
);
2273 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
)
2277 * If the event is in error state, clear that first.
2278 * That way, if we see the event in error state below, we
2279 * know that it has gone back into error state, as distinct
2280 * from the task having been scheduled away before the
2281 * cross-call arrived.
2283 if (event
->state
== PERF_EVENT_STATE_ERROR
)
2284 event
->state
= PERF_EVENT_STATE_OFF
;
2287 if (!ctx
->is_active
) {
2288 __perf_event_mark_enabled(event
);
2292 raw_spin_unlock_irq(&ctx
->lock
);
2294 if (!task_function_call(task
, __perf_event_enable
, event
))
2297 raw_spin_lock_irq(&ctx
->lock
);
2300 * If the context is active and the event is still off,
2301 * we need to retry the cross-call.
2303 if (ctx
->is_active
&& event
->state
== PERF_EVENT_STATE_OFF
) {
2305 * task could have been flipped by a concurrent
2306 * perf_event_context_sched_out()
2313 raw_spin_unlock_irq(&ctx
->lock
);
2317 * See perf_event_disable();
2319 void perf_event_enable(struct perf_event
*event
)
2321 struct perf_event_context
*ctx
;
2323 ctx
= perf_event_ctx_lock(event
);
2324 _perf_event_enable(event
);
2325 perf_event_ctx_unlock(event
, ctx
);
2327 EXPORT_SYMBOL_GPL(perf_event_enable
);
2329 static int _perf_event_refresh(struct perf_event
*event
, int refresh
)
2332 * not supported on inherited events
2334 if (event
->attr
.inherit
|| !is_sampling_event(event
))
2337 atomic_add(refresh
, &event
->event_limit
);
2338 _perf_event_enable(event
);
2344 * See perf_event_disable()
2346 int perf_event_refresh(struct perf_event
*event
, int refresh
)
2348 struct perf_event_context
*ctx
;
2351 ctx
= perf_event_ctx_lock(event
);
2352 ret
= _perf_event_refresh(event
, refresh
);
2353 perf_event_ctx_unlock(event
, ctx
);
2357 EXPORT_SYMBOL_GPL(perf_event_refresh
);
2359 static void ctx_sched_out(struct perf_event_context
*ctx
,
2360 struct perf_cpu_context
*cpuctx
,
2361 enum event_type_t event_type
)
2363 struct perf_event
*event
;
2364 int is_active
= ctx
->is_active
;
2366 ctx
->is_active
&= ~event_type
;
2367 if (likely(!ctx
->nr_events
))
2370 update_context_time(ctx
);
2371 update_cgrp_time_from_cpuctx(cpuctx
);
2372 if (!ctx
->nr_active
)
2375 perf_pmu_disable(ctx
->pmu
);
2376 if ((is_active
& EVENT_PINNED
) && (event_type
& EVENT_PINNED
)) {
2377 list_for_each_entry(event
, &ctx
->pinned_groups
, group_entry
)
2378 group_sched_out(event
, cpuctx
, ctx
);
2381 if ((is_active
& EVENT_FLEXIBLE
) && (event_type
& EVENT_FLEXIBLE
)) {
2382 list_for_each_entry(event
, &ctx
->flexible_groups
, group_entry
)
2383 group_sched_out(event
, cpuctx
, ctx
);
2385 perf_pmu_enable(ctx
->pmu
);
2389 * Test whether two contexts are equivalent, i.e. whether they have both been
2390 * cloned from the same version of the same context.
2392 * Equivalence is measured using a generation number in the context that is
2393 * incremented on each modification to it; see unclone_ctx(), list_add_event()
2394 * and list_del_event().
2396 static int context_equiv(struct perf_event_context
*ctx1
,
2397 struct perf_event_context
*ctx2
)
2399 lockdep_assert_held(&ctx1
->lock
);
2400 lockdep_assert_held(&ctx2
->lock
);
2402 /* Pinning disables the swap optimization */
2403 if (ctx1
->pin_count
|| ctx2
->pin_count
)
2406 /* If ctx1 is the parent of ctx2 */
2407 if (ctx1
== ctx2
->parent_ctx
&& ctx1
->generation
== ctx2
->parent_gen
)
2410 /* If ctx2 is the parent of ctx1 */
2411 if (ctx1
->parent_ctx
== ctx2
&& ctx1
->parent_gen
== ctx2
->generation
)
2415 * If ctx1 and ctx2 have the same parent; we flatten the parent
2416 * hierarchy, see perf_event_init_context().
2418 if (ctx1
->parent_ctx
&& ctx1
->parent_ctx
== ctx2
->parent_ctx
&&
2419 ctx1
->parent_gen
== ctx2
->parent_gen
)
2426 static void __perf_event_sync_stat(struct perf_event
*event
,
2427 struct perf_event
*next_event
)
2431 if (!event
->attr
.inherit_stat
)
2435 * Update the event value, we cannot use perf_event_read()
2436 * because we're in the middle of a context switch and have IRQs
2437 * disabled, which upsets smp_call_function_single(), however
2438 * we know the event must be on the current CPU, therefore we
2439 * don't need to use it.
2441 switch (event
->state
) {
2442 case PERF_EVENT_STATE_ACTIVE
:
2443 event
->pmu
->read(event
);
2446 case PERF_EVENT_STATE_INACTIVE
:
2447 update_event_times(event
);
2455 * In order to keep per-task stats reliable we need to flip the event
2456 * values when we flip the contexts.
2458 value
= local64_read(&next_event
->count
);
2459 value
= local64_xchg(&event
->count
, value
);
2460 local64_set(&next_event
->count
, value
);
2462 swap(event
->total_time_enabled
, next_event
->total_time_enabled
);
2463 swap(event
->total_time_running
, next_event
->total_time_running
);
2466 * Since we swizzled the values, update the user visible data too.
2468 perf_event_update_userpage(event
);
2469 perf_event_update_userpage(next_event
);
2472 static void perf_event_sync_stat(struct perf_event_context
*ctx
,
2473 struct perf_event_context
*next_ctx
)
2475 struct perf_event
*event
, *next_event
;
2480 update_context_time(ctx
);
2482 event
= list_first_entry(&ctx
->event_list
,
2483 struct perf_event
, event_entry
);
2485 next_event
= list_first_entry(&next_ctx
->event_list
,
2486 struct perf_event
, event_entry
);
2488 while (&event
->event_entry
!= &ctx
->event_list
&&
2489 &next_event
->event_entry
!= &next_ctx
->event_list
) {
2491 __perf_event_sync_stat(event
, next_event
);
2493 event
= list_next_entry(event
, event_entry
);
2494 next_event
= list_next_entry(next_event
, event_entry
);
2498 static void perf_event_context_sched_out(struct task_struct
*task
, int ctxn
,
2499 struct task_struct
*next
)
2501 struct perf_event_context
*ctx
= task
->perf_event_ctxp
[ctxn
];
2502 struct perf_event_context
*next_ctx
;
2503 struct perf_event_context
*parent
, *next_parent
;
2504 struct perf_cpu_context
*cpuctx
;
2510 cpuctx
= __get_cpu_context(ctx
);
2511 if (!cpuctx
->task_ctx
)
2515 next_ctx
= next
->perf_event_ctxp
[ctxn
];
2519 parent
= rcu_dereference(ctx
->parent_ctx
);
2520 next_parent
= rcu_dereference(next_ctx
->parent_ctx
);
2522 /* If neither context have a parent context; they cannot be clones. */
2523 if (!parent
&& !next_parent
)
2526 if (next_parent
== ctx
|| next_ctx
== parent
|| next_parent
== parent
) {
2528 * Looks like the two contexts are clones, so we might be
2529 * able to optimize the context switch. We lock both
2530 * contexts and check that they are clones under the
2531 * lock (including re-checking that neither has been
2532 * uncloned in the meantime). It doesn't matter which
2533 * order we take the locks because no other cpu could
2534 * be trying to lock both of these tasks.
2536 raw_spin_lock(&ctx
->lock
);
2537 raw_spin_lock_nested(&next_ctx
->lock
, SINGLE_DEPTH_NESTING
);
2538 if (context_equiv(ctx
, next_ctx
)) {
2540 * XXX do we need a memory barrier of sorts
2541 * wrt to rcu_dereference() of perf_event_ctxp
2543 task
->perf_event_ctxp
[ctxn
] = next_ctx
;
2544 next
->perf_event_ctxp
[ctxn
] = ctx
;
2546 next_ctx
->task
= task
;
2548 swap(ctx
->task_ctx_data
, next_ctx
->task_ctx_data
);
2552 perf_event_sync_stat(ctx
, next_ctx
);
2554 raw_spin_unlock(&next_ctx
->lock
);
2555 raw_spin_unlock(&ctx
->lock
);
2561 raw_spin_lock(&ctx
->lock
);
2562 ctx_sched_out(ctx
, cpuctx
, EVENT_ALL
);
2563 cpuctx
->task_ctx
= NULL
;
2564 raw_spin_unlock(&ctx
->lock
);
2568 void perf_sched_cb_dec(struct pmu
*pmu
)
2570 this_cpu_dec(perf_sched_cb_usages
);
2573 void perf_sched_cb_inc(struct pmu
*pmu
)
2575 this_cpu_inc(perf_sched_cb_usages
);
2579 * This function provides the context switch callback to the lower code
2580 * layer. It is invoked ONLY when the context switch callback is enabled.
2582 static void perf_pmu_sched_task(struct task_struct
*prev
,
2583 struct task_struct
*next
,
2586 struct perf_cpu_context
*cpuctx
;
2588 unsigned long flags
;
2593 local_irq_save(flags
);
2597 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
2598 if (pmu
->sched_task
) {
2599 cpuctx
= this_cpu_ptr(pmu
->pmu_cpu_context
);
2601 perf_ctx_lock(cpuctx
, cpuctx
->task_ctx
);
2603 perf_pmu_disable(pmu
);
2605 pmu
->sched_task(cpuctx
->task_ctx
, sched_in
);
2607 perf_pmu_enable(pmu
);
2609 perf_ctx_unlock(cpuctx
, cpuctx
->task_ctx
);
2615 local_irq_restore(flags
);
2618 #define for_each_task_context_nr(ctxn) \
2619 for ((ctxn) = 0; (ctxn) < perf_nr_task_contexts; (ctxn)++)
2622 * Called from scheduler to remove the events of the current task,
2623 * with interrupts disabled.
2625 * We stop each event and update the event value in event->count.
2627 * This does not protect us against NMI, but disable()
2628 * sets the disabled bit in the control field of event _before_
2629 * accessing the event control register. If a NMI hits, then it will
2630 * not restart the event.
2632 void __perf_event_task_sched_out(struct task_struct
*task
,
2633 struct task_struct
*next
)
2637 if (__this_cpu_read(perf_sched_cb_usages
))
2638 perf_pmu_sched_task(task
, next
, false);
2640 for_each_task_context_nr(ctxn
)
2641 perf_event_context_sched_out(task
, ctxn
, next
);
2644 * if cgroup events exist on this CPU, then we need
2645 * to check if we have to switch out PMU state.
2646 * cgroup event are system-wide mode only
2648 if (atomic_read(this_cpu_ptr(&perf_cgroup_events
)))
2649 perf_cgroup_sched_out(task
, next
);
2652 static void task_ctx_sched_out(struct perf_event_context
*ctx
)
2654 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
2656 if (!cpuctx
->task_ctx
)
2659 if (WARN_ON_ONCE(ctx
!= cpuctx
->task_ctx
))
2662 ctx_sched_out(ctx
, cpuctx
, EVENT_ALL
);
2663 cpuctx
->task_ctx
= NULL
;
2667 * Called with IRQs disabled
2669 static void cpu_ctx_sched_out(struct perf_cpu_context
*cpuctx
,
2670 enum event_type_t event_type
)
2672 ctx_sched_out(&cpuctx
->ctx
, cpuctx
, event_type
);
2676 ctx_pinned_sched_in(struct perf_event_context
*ctx
,
2677 struct perf_cpu_context
*cpuctx
)
2679 struct perf_event
*event
;
2681 list_for_each_entry(event
, &ctx
->pinned_groups
, group_entry
) {
2682 if (event
->state
<= PERF_EVENT_STATE_OFF
)
2684 if (!event_filter_match(event
))
2687 /* may need to reset tstamp_enabled */
2688 if (is_cgroup_event(event
))
2689 perf_cgroup_mark_enabled(event
, ctx
);
2691 if (group_can_go_on(event
, cpuctx
, 1))
2692 group_sched_in(event
, cpuctx
, ctx
);
2695 * If this pinned group hasn't been scheduled,
2696 * put it in error state.
2698 if (event
->state
== PERF_EVENT_STATE_INACTIVE
) {
2699 update_group_times(event
);
2700 event
->state
= PERF_EVENT_STATE_ERROR
;
2706 ctx_flexible_sched_in(struct perf_event_context
*ctx
,
2707 struct perf_cpu_context
*cpuctx
)
2709 struct perf_event
*event
;
2712 list_for_each_entry(event
, &ctx
->flexible_groups
, group_entry
) {
2713 /* Ignore events in OFF or ERROR state */
2714 if (event
->state
<= PERF_EVENT_STATE_OFF
)
2717 * Listen to the 'cpu' scheduling filter constraint
2720 if (!event_filter_match(event
))
2723 /* may need to reset tstamp_enabled */
2724 if (is_cgroup_event(event
))
2725 perf_cgroup_mark_enabled(event
, ctx
);
2727 if (group_can_go_on(event
, cpuctx
, can_add_hw
)) {
2728 if (group_sched_in(event
, cpuctx
, ctx
))
2735 ctx_sched_in(struct perf_event_context
*ctx
,
2736 struct perf_cpu_context
*cpuctx
,
2737 enum event_type_t event_type
,
2738 struct task_struct
*task
)
2741 int is_active
= ctx
->is_active
;
2743 ctx
->is_active
|= event_type
;
2744 if (likely(!ctx
->nr_events
))
2748 ctx
->timestamp
= now
;
2749 perf_cgroup_set_timestamp(task
, ctx
);
2751 * First go through the list and put on any pinned groups
2752 * in order to give them the best chance of going on.
2754 if (!(is_active
& EVENT_PINNED
) && (event_type
& EVENT_PINNED
))
2755 ctx_pinned_sched_in(ctx
, cpuctx
);
2757 /* Then walk through the lower prio flexible groups */
2758 if (!(is_active
& EVENT_FLEXIBLE
) && (event_type
& EVENT_FLEXIBLE
))
2759 ctx_flexible_sched_in(ctx
, cpuctx
);
2762 static void cpu_ctx_sched_in(struct perf_cpu_context
*cpuctx
,
2763 enum event_type_t event_type
,
2764 struct task_struct
*task
)
2766 struct perf_event_context
*ctx
= &cpuctx
->ctx
;
2768 ctx_sched_in(ctx
, cpuctx
, event_type
, task
);
2771 static void perf_event_context_sched_in(struct perf_event_context
*ctx
,
2772 struct task_struct
*task
)
2774 struct perf_cpu_context
*cpuctx
;
2776 cpuctx
= __get_cpu_context(ctx
);
2777 if (cpuctx
->task_ctx
== ctx
)
2780 perf_ctx_lock(cpuctx
, ctx
);
2781 perf_pmu_disable(ctx
->pmu
);
2783 * We want to keep the following priority order:
2784 * cpu pinned (that don't need to move), task pinned,
2785 * cpu flexible, task flexible.
2787 cpu_ctx_sched_out(cpuctx
, EVENT_FLEXIBLE
);
2790 cpuctx
->task_ctx
= ctx
;
2792 perf_event_sched_in(cpuctx
, cpuctx
->task_ctx
, task
);
2794 perf_pmu_enable(ctx
->pmu
);
2795 perf_ctx_unlock(cpuctx
, ctx
);
2799 * Called from scheduler to add the events of the current task
2800 * with interrupts disabled.
2802 * We restore the event value and then enable it.
2804 * This does not protect us against NMI, but enable()
2805 * sets the enabled bit in the control field of event _before_
2806 * accessing the event control register. If a NMI hits, then it will
2807 * keep the event running.
2809 void __perf_event_task_sched_in(struct task_struct
*prev
,
2810 struct task_struct
*task
)
2812 struct perf_event_context
*ctx
;
2815 for_each_task_context_nr(ctxn
) {
2816 ctx
= task
->perf_event_ctxp
[ctxn
];
2820 perf_event_context_sched_in(ctx
, task
);
2823 * if cgroup events exist on this CPU, then we need
2824 * to check if we have to switch in PMU state.
2825 * cgroup event are system-wide mode only
2827 if (atomic_read(this_cpu_ptr(&perf_cgroup_events
)))
2828 perf_cgroup_sched_in(prev
, task
);
2830 if (__this_cpu_read(perf_sched_cb_usages
))
2831 perf_pmu_sched_task(prev
, task
, true);
2834 static u64
perf_calculate_period(struct perf_event
*event
, u64 nsec
, u64 count
)
2836 u64 frequency
= event
->attr
.sample_freq
;
2837 u64 sec
= NSEC_PER_SEC
;
2838 u64 divisor
, dividend
;
2840 int count_fls
, nsec_fls
, frequency_fls
, sec_fls
;
2842 count_fls
= fls64(count
);
2843 nsec_fls
= fls64(nsec
);
2844 frequency_fls
= fls64(frequency
);
2848 * We got @count in @nsec, with a target of sample_freq HZ
2849 * the target period becomes:
2852 * period = -------------------
2853 * @nsec * sample_freq
2858 * Reduce accuracy by one bit such that @a and @b converge
2859 * to a similar magnitude.
2861 #define REDUCE_FLS(a, b) \
2863 if (a##_fls > b##_fls) { \
2873 * Reduce accuracy until either term fits in a u64, then proceed with
2874 * the other, so that finally we can do a u64/u64 division.
2876 while (count_fls
+ sec_fls
> 64 && nsec_fls
+ frequency_fls
> 64) {
2877 REDUCE_FLS(nsec
, frequency
);
2878 REDUCE_FLS(sec
, count
);
2881 if (count_fls
+ sec_fls
> 64) {
2882 divisor
= nsec
* frequency
;
2884 while (count_fls
+ sec_fls
> 64) {
2885 REDUCE_FLS(count
, sec
);
2889 dividend
= count
* sec
;
2891 dividend
= count
* sec
;
2893 while (nsec_fls
+ frequency_fls
> 64) {
2894 REDUCE_FLS(nsec
, frequency
);
2898 divisor
= nsec
* frequency
;
2904 return div64_u64(dividend
, divisor
);
2907 static DEFINE_PER_CPU(int, perf_throttled_count
);
2908 static DEFINE_PER_CPU(u64
, perf_throttled_seq
);
2910 static void perf_adjust_period(struct perf_event
*event
, u64 nsec
, u64 count
, bool disable
)
2912 struct hw_perf_event
*hwc
= &event
->hw
;
2913 s64 period
, sample_period
;
2916 period
= perf_calculate_period(event
, nsec
, count
);
2918 delta
= (s64
)(period
- hwc
->sample_period
);
2919 delta
= (delta
+ 7) / 8; /* low pass filter */
2921 sample_period
= hwc
->sample_period
+ delta
;
2926 hwc
->sample_period
= sample_period
;
2928 if (local64_read(&hwc
->period_left
) > 8*sample_period
) {
2930 event
->pmu
->stop(event
, PERF_EF_UPDATE
);
2932 local64_set(&hwc
->period_left
, 0);
2935 event
->pmu
->start(event
, PERF_EF_RELOAD
);
2940 * combine freq adjustment with unthrottling to avoid two passes over the
2941 * events. At the same time, make sure, having freq events does not change
2942 * the rate of unthrottling as that would introduce bias.
2944 static void perf_adjust_freq_unthr_context(struct perf_event_context
*ctx
,
2947 struct perf_event
*event
;
2948 struct hw_perf_event
*hwc
;
2949 u64 now
, period
= TICK_NSEC
;
2953 * only need to iterate over all events iff:
2954 * - context have events in frequency mode (needs freq adjust)
2955 * - there are events to unthrottle on this cpu
2957 if (!(ctx
->nr_freq
|| needs_unthr
))
2960 raw_spin_lock(&ctx
->lock
);
2961 perf_pmu_disable(ctx
->pmu
);
2963 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
2964 if (event
->state
!= PERF_EVENT_STATE_ACTIVE
)
2967 if (!event_filter_match(event
))
2970 perf_pmu_disable(event
->pmu
);
2974 if (hwc
->interrupts
== MAX_INTERRUPTS
) {
2975 hwc
->interrupts
= 0;
2976 perf_log_throttle(event
, 1);
2977 event
->pmu
->start(event
, 0);
2980 if (!event
->attr
.freq
|| !event
->attr
.sample_freq
)
2984 * stop the event and update event->count
2986 event
->pmu
->stop(event
, PERF_EF_UPDATE
);
2988 now
= local64_read(&event
->count
);
2989 delta
= now
- hwc
->freq_count_stamp
;
2990 hwc
->freq_count_stamp
= now
;
2994 * reload only if value has changed
2995 * we have stopped the event so tell that
2996 * to perf_adjust_period() to avoid stopping it
3000 perf_adjust_period(event
, period
, delta
, false);
3002 event
->pmu
->start(event
, delta
> 0 ? PERF_EF_RELOAD
: 0);
3004 perf_pmu_enable(event
->pmu
);
3007 perf_pmu_enable(ctx
->pmu
);
3008 raw_spin_unlock(&ctx
->lock
);
3012 * Round-robin a context's events:
3014 static void rotate_ctx(struct perf_event_context
*ctx
)
3017 * Rotate the first entry last of non-pinned groups. Rotation might be
3018 * disabled by the inheritance code.
3020 if (!ctx
->rotate_disable
)
3021 list_rotate_left(&ctx
->flexible_groups
);
3024 static int perf_rotate_context(struct perf_cpu_context
*cpuctx
)
3026 struct perf_event_context
*ctx
= NULL
;
3029 if (cpuctx
->ctx
.nr_events
) {
3030 if (cpuctx
->ctx
.nr_events
!= cpuctx
->ctx
.nr_active
)
3034 ctx
= cpuctx
->task_ctx
;
3035 if (ctx
&& ctx
->nr_events
) {
3036 if (ctx
->nr_events
!= ctx
->nr_active
)
3043 perf_ctx_lock(cpuctx
, cpuctx
->task_ctx
);
3044 perf_pmu_disable(cpuctx
->ctx
.pmu
);
3046 cpu_ctx_sched_out(cpuctx
, EVENT_FLEXIBLE
);
3048 ctx_sched_out(ctx
, cpuctx
, EVENT_FLEXIBLE
);
3050 rotate_ctx(&cpuctx
->ctx
);
3054 perf_event_sched_in(cpuctx
, ctx
, current
);
3056 perf_pmu_enable(cpuctx
->ctx
.pmu
);
3057 perf_ctx_unlock(cpuctx
, cpuctx
->task_ctx
);
3063 #ifdef CONFIG_NO_HZ_FULL
3064 bool perf_event_can_stop_tick(void)
3066 if (atomic_read(&nr_freq_events
) ||
3067 __this_cpu_read(perf_throttled_count
))
3074 void perf_event_task_tick(void)
3076 struct list_head
*head
= this_cpu_ptr(&active_ctx_list
);
3077 struct perf_event_context
*ctx
, *tmp
;
3080 WARN_ON(!irqs_disabled());
3082 __this_cpu_inc(perf_throttled_seq
);
3083 throttled
= __this_cpu_xchg(perf_throttled_count
, 0);
3085 list_for_each_entry_safe(ctx
, tmp
, head
, active_ctx_list
)
3086 perf_adjust_freq_unthr_context(ctx
, throttled
);
3089 static int event_enable_on_exec(struct perf_event
*event
,
3090 struct perf_event_context
*ctx
)
3092 if (!event
->attr
.enable_on_exec
)
3095 event
->attr
.enable_on_exec
= 0;
3096 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
)
3099 __perf_event_mark_enabled(event
);
3105 * Enable all of a task's events that have been marked enable-on-exec.
3106 * This expects task == current.
3108 static void perf_event_enable_on_exec(struct perf_event_context
*ctx
)
3110 struct perf_event_context
*clone_ctx
= NULL
;
3111 struct perf_event
*event
;
3112 unsigned long flags
;
3116 local_irq_save(flags
);
3117 if (!ctx
|| !ctx
->nr_events
)
3121 * We must ctxsw out cgroup events to avoid conflict
3122 * when invoking perf_task_event_sched_in() later on
3123 * in this function. Otherwise we end up trying to
3124 * ctxswin cgroup events which are already scheduled
3127 perf_cgroup_sched_out(current
, NULL
);
3129 raw_spin_lock(&ctx
->lock
);
3130 task_ctx_sched_out(ctx
);
3132 list_for_each_entry(event
, &ctx
->event_list
, event_entry
) {
3133 ret
= event_enable_on_exec(event
, ctx
);
3139 * Unclone this context if we enabled any event.
3142 clone_ctx
= unclone_ctx(ctx
);
3144 raw_spin_unlock(&ctx
->lock
);
3147 * Also calls ctxswin for cgroup events, if any:
3149 perf_event_context_sched_in(ctx
, ctx
->task
);
3151 local_irq_restore(flags
);
3157 void perf_event_exec(void)
3159 struct perf_event_context
*ctx
;
3163 for_each_task_context_nr(ctxn
) {
3164 ctx
= current
->perf_event_ctxp
[ctxn
];
3168 perf_event_enable_on_exec(ctx
);
3174 * Cross CPU call to read the hardware event
3176 static void __perf_event_read(void *info
)
3178 struct perf_event
*event
= info
;
3179 struct perf_event_context
*ctx
= event
->ctx
;
3180 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
3183 * If this is a task context, we need to check whether it is
3184 * the current task context of this cpu. If not it has been
3185 * scheduled out before the smp call arrived. In that case
3186 * event->count would have been updated to a recent sample
3187 * when the event was scheduled out.
3189 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
)
3192 raw_spin_lock(&ctx
->lock
);
3193 if (ctx
->is_active
) {
3194 update_context_time(ctx
);
3195 update_cgrp_time_from_event(event
);
3197 update_event_times(event
);
3198 if (event
->state
== PERF_EVENT_STATE_ACTIVE
)
3199 event
->pmu
->read(event
);
3200 raw_spin_unlock(&ctx
->lock
);
3203 static inline u64
perf_event_count(struct perf_event
*event
)
3205 if (event
->pmu
->count
)
3206 return event
->pmu
->count(event
);
3208 return __perf_event_count(event
);
3211 static u64
perf_event_read(struct perf_event
*event
)
3214 * If event is enabled and currently active on a CPU, update the
3215 * value in the event structure:
3217 if (event
->state
== PERF_EVENT_STATE_ACTIVE
) {
3218 smp_call_function_single(event
->oncpu
,
3219 __perf_event_read
, event
, 1);
3220 } else if (event
->state
== PERF_EVENT_STATE_INACTIVE
) {
3221 struct perf_event_context
*ctx
= event
->ctx
;
3222 unsigned long flags
;
3224 raw_spin_lock_irqsave(&ctx
->lock
, flags
);
3226 * may read while context is not active
3227 * (e.g., thread is blocked), in that case
3228 * we cannot update context time
3230 if (ctx
->is_active
) {
3231 update_context_time(ctx
);
3232 update_cgrp_time_from_event(event
);
3234 update_event_times(event
);
3235 raw_spin_unlock_irqrestore(&ctx
->lock
, flags
);
3238 return perf_event_count(event
);
3242 * Initialize the perf_event context in a task_struct:
3244 static void __perf_event_init_context(struct perf_event_context
*ctx
)
3246 raw_spin_lock_init(&ctx
->lock
);
3247 mutex_init(&ctx
->mutex
);
3248 INIT_LIST_HEAD(&ctx
->active_ctx_list
);
3249 INIT_LIST_HEAD(&ctx
->pinned_groups
);
3250 INIT_LIST_HEAD(&ctx
->flexible_groups
);
3251 INIT_LIST_HEAD(&ctx
->event_list
);
3252 atomic_set(&ctx
->refcount
, 1);
3253 INIT_DELAYED_WORK(&ctx
->orphans_remove
, orphans_remove_work
);
3256 static struct perf_event_context
*
3257 alloc_perf_context(struct pmu
*pmu
, struct task_struct
*task
)
3259 struct perf_event_context
*ctx
;
3261 ctx
= kzalloc(sizeof(struct perf_event_context
), GFP_KERNEL
);
3265 __perf_event_init_context(ctx
);
3268 get_task_struct(task
);
3275 static struct task_struct
*
3276 find_lively_task_by_vpid(pid_t vpid
)
3278 struct task_struct
*task
;
3285 task
= find_task_by_vpid(vpid
);
3287 get_task_struct(task
);
3291 return ERR_PTR(-ESRCH
);
3293 /* Reuse ptrace permission checks for now. */
3295 if (!ptrace_may_access(task
, PTRACE_MODE_READ
))
3300 put_task_struct(task
);
3301 return ERR_PTR(err
);
3306 * Returns a matching context with refcount and pincount.
3308 static struct perf_event_context
*
3309 find_get_context(struct pmu
*pmu
, struct task_struct
*task
,
3310 struct perf_event
*event
)
3312 struct perf_event_context
*ctx
, *clone_ctx
= NULL
;
3313 struct perf_cpu_context
*cpuctx
;
3314 void *task_ctx_data
= NULL
;
3315 unsigned long flags
;
3317 int cpu
= event
->cpu
;
3320 /* Must be root to operate on a CPU event: */
3321 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN
))
3322 return ERR_PTR(-EACCES
);
3325 * We could be clever and allow to attach a event to an
3326 * offline CPU and activate it when the CPU comes up, but
3329 if (!cpu_online(cpu
))
3330 return ERR_PTR(-ENODEV
);
3332 cpuctx
= per_cpu_ptr(pmu
->pmu_cpu_context
, cpu
);
3341 ctxn
= pmu
->task_ctx_nr
;
3345 if (event
->attach_state
& PERF_ATTACH_TASK_DATA
) {
3346 task_ctx_data
= kzalloc(pmu
->task_ctx_size
, GFP_KERNEL
);
3347 if (!task_ctx_data
) {
3354 ctx
= perf_lock_task_context(task
, ctxn
, &flags
);
3356 clone_ctx
= unclone_ctx(ctx
);
3359 if (task_ctx_data
&& !ctx
->task_ctx_data
) {
3360 ctx
->task_ctx_data
= task_ctx_data
;
3361 task_ctx_data
= NULL
;
3363 raw_spin_unlock_irqrestore(&ctx
->lock
, flags
);
3368 ctx
= alloc_perf_context(pmu
, task
);
3373 if (task_ctx_data
) {
3374 ctx
->task_ctx_data
= task_ctx_data
;
3375 task_ctx_data
= NULL
;
3379 mutex_lock(&task
->perf_event_mutex
);
3381 * If it has already passed perf_event_exit_task().
3382 * we must see PF_EXITING, it takes this mutex too.
3384 if (task
->flags
& PF_EXITING
)
3386 else if (task
->perf_event_ctxp
[ctxn
])
3391 rcu_assign_pointer(task
->perf_event_ctxp
[ctxn
], ctx
);
3393 mutex_unlock(&task
->perf_event_mutex
);
3395 if (unlikely(err
)) {
3404 kfree(task_ctx_data
);
3408 kfree(task_ctx_data
);
3409 return ERR_PTR(err
);
3412 static void perf_event_free_filter(struct perf_event
*event
);
3413 static void perf_event_free_bpf_prog(struct perf_event
*event
);
3415 static void free_event_rcu(struct rcu_head
*head
)
3417 struct perf_event
*event
;
3419 event
= container_of(head
, struct perf_event
, rcu_head
);
3421 put_pid_ns(event
->ns
);
3422 perf_event_free_filter(event
);
3423 perf_event_free_bpf_prog(event
);
3427 static void ring_buffer_attach(struct perf_event
*event
,
3428 struct ring_buffer
*rb
);
3430 static void unaccount_event_cpu(struct perf_event
*event
, int cpu
)
3435 if (is_cgroup_event(event
))
3436 atomic_dec(&per_cpu(perf_cgroup_events
, cpu
));
3439 static void unaccount_event(struct perf_event
*event
)
3444 if (event
->attach_state
& PERF_ATTACH_TASK
)
3445 static_key_slow_dec_deferred(&perf_sched_events
);
3446 if (event
->attr
.mmap
|| event
->attr
.mmap_data
)
3447 atomic_dec(&nr_mmap_events
);
3448 if (event
->attr
.comm
)
3449 atomic_dec(&nr_comm_events
);
3450 if (event
->attr
.task
)
3451 atomic_dec(&nr_task_events
);
3452 if (event
->attr
.freq
)
3453 atomic_dec(&nr_freq_events
);
3454 if (is_cgroup_event(event
))
3455 static_key_slow_dec_deferred(&perf_sched_events
);
3456 if (has_branch_stack(event
))
3457 static_key_slow_dec_deferred(&perf_sched_events
);
3459 unaccount_event_cpu(event
, event
->cpu
);
3463 * The following implement mutual exclusion of events on "exclusive" pmus
3464 * (PERF_PMU_CAP_EXCLUSIVE). Such pmus can only have one event scheduled
3465 * at a time, so we disallow creating events that might conflict, namely:
3467 * 1) cpu-wide events in the presence of per-task events,
3468 * 2) per-task events in the presence of cpu-wide events,
3469 * 3) two matching events on the same context.
3471 * The former two cases are handled in the allocation path (perf_event_alloc(),
3472 * __free_event()), the latter -- before the first perf_install_in_context().
3474 static int exclusive_event_init(struct perf_event
*event
)
3476 struct pmu
*pmu
= event
->pmu
;
3478 if (!(pmu
->capabilities
& PERF_PMU_CAP_EXCLUSIVE
))
3482 * Prevent co-existence of per-task and cpu-wide events on the
3483 * same exclusive pmu.
3485 * Negative pmu::exclusive_cnt means there are cpu-wide
3486 * events on this "exclusive" pmu, positive means there are
3489 * Since this is called in perf_event_alloc() path, event::ctx
3490 * doesn't exist yet; it is, however, safe to use PERF_ATTACH_TASK
3491 * to mean "per-task event", because unlike other attach states it
3492 * never gets cleared.
3494 if (event
->attach_state
& PERF_ATTACH_TASK
) {
3495 if (!atomic_inc_unless_negative(&pmu
->exclusive_cnt
))
3498 if (!atomic_dec_unless_positive(&pmu
->exclusive_cnt
))
3505 static void exclusive_event_destroy(struct perf_event
*event
)
3507 struct pmu
*pmu
= event
->pmu
;
3509 if (!(pmu
->capabilities
& PERF_PMU_CAP_EXCLUSIVE
))
3512 /* see comment in exclusive_event_init() */
3513 if (event
->attach_state
& PERF_ATTACH_TASK
)
3514 atomic_dec(&pmu
->exclusive_cnt
);
3516 atomic_inc(&pmu
->exclusive_cnt
);
3519 static bool exclusive_event_match(struct perf_event
*e1
, struct perf_event
*e2
)
3521 if ((e1
->pmu
->capabilities
& PERF_PMU_CAP_EXCLUSIVE
) &&
3522 (e1
->cpu
== e2
->cpu
||
3529 /* Called under the same ctx::mutex as perf_install_in_context() */
3530 static bool exclusive_event_installable(struct perf_event
*event
,
3531 struct perf_event_context
*ctx
)
3533 struct perf_event
*iter_event
;
3534 struct pmu
*pmu
= event
->pmu
;
3536 if (!(pmu
->capabilities
& PERF_PMU_CAP_EXCLUSIVE
))
3539 list_for_each_entry(iter_event
, &ctx
->event_list
, event_entry
) {
3540 if (exclusive_event_match(iter_event
, event
))
3547 static void __free_event(struct perf_event
*event
)
3549 if (!event
->parent
) {
3550 if (event
->attr
.sample_type
& PERF_SAMPLE_CALLCHAIN
)
3551 put_callchain_buffers();
3555 event
->destroy(event
);
3558 put_ctx(event
->ctx
);
3561 exclusive_event_destroy(event
);
3562 module_put(event
->pmu
->module
);
3565 call_rcu(&event
->rcu_head
, free_event_rcu
);
3568 static void _free_event(struct perf_event
*event
)
3570 irq_work_sync(&event
->pending
);
3572 unaccount_event(event
);
3576 * Can happen when we close an event with re-directed output.
3578 * Since we have a 0 refcount, perf_mmap_close() will skip
3579 * over us; possibly making our ring_buffer_put() the last.
3581 mutex_lock(&event
->mmap_mutex
);
3582 ring_buffer_attach(event
, NULL
);
3583 mutex_unlock(&event
->mmap_mutex
);
3586 if (is_cgroup_event(event
))
3587 perf_detach_cgroup(event
);
3589 __free_event(event
);
3593 * Used to free events which have a known refcount of 1, such as in error paths
3594 * where the event isn't exposed yet and inherited events.
3596 static void free_event(struct perf_event
*event
)
3598 if (WARN(atomic_long_cmpxchg(&event
->refcount
, 1, 0) != 1,
3599 "unexpected event refcount: %ld; ptr=%p\n",
3600 atomic_long_read(&event
->refcount
), event
)) {
3601 /* leak to avoid use-after-free */
3609 * Remove user event from the owner task.
3611 static void perf_remove_from_owner(struct perf_event
*event
)
3613 struct task_struct
*owner
;
3616 owner
= ACCESS_ONCE(event
->owner
);
3618 * Matches the smp_wmb() in perf_event_exit_task(). If we observe
3619 * !owner it means the list deletion is complete and we can indeed
3620 * free this event, otherwise we need to serialize on
3621 * owner->perf_event_mutex.
3623 smp_read_barrier_depends();
3626 * Since delayed_put_task_struct() also drops the last
3627 * task reference we can safely take a new reference
3628 * while holding the rcu_read_lock().
3630 get_task_struct(owner
);
3636 * If we're here through perf_event_exit_task() we're already
3637 * holding ctx->mutex which would be an inversion wrt. the
3638 * normal lock order.
3640 * However we can safely take this lock because its the child
3643 mutex_lock_nested(&owner
->perf_event_mutex
, SINGLE_DEPTH_NESTING
);
3646 * We have to re-check the event->owner field, if it is cleared
3647 * we raced with perf_event_exit_task(), acquiring the mutex
3648 * ensured they're done, and we can proceed with freeing the
3652 list_del_init(&event
->owner_entry
);
3653 mutex_unlock(&owner
->perf_event_mutex
);
3654 put_task_struct(owner
);
3659 * Called when the last reference to the file is gone.
3661 static void put_event(struct perf_event
*event
)
3663 struct perf_event_context
*ctx
;
3665 if (!atomic_long_dec_and_test(&event
->refcount
))
3668 if (!is_kernel_event(event
))
3669 perf_remove_from_owner(event
);
3672 * There are two ways this annotation is useful:
3674 * 1) there is a lock recursion from perf_event_exit_task
3675 * see the comment there.
3677 * 2) there is a lock-inversion with mmap_sem through
3678 * perf_event_read_group(), which takes faults while
3679 * holding ctx->mutex, however this is called after
3680 * the last filedesc died, so there is no possibility
3681 * to trigger the AB-BA case.
3683 ctx
= perf_event_ctx_lock_nested(event
, SINGLE_DEPTH_NESTING
);
3684 WARN_ON_ONCE(ctx
->parent_ctx
);
3685 perf_remove_from_context(event
, true);
3686 perf_event_ctx_unlock(event
, ctx
);
3691 int perf_event_release_kernel(struct perf_event
*event
)
3696 EXPORT_SYMBOL_GPL(perf_event_release_kernel
);
3698 static int perf_release(struct inode
*inode
, struct file
*file
)
3700 put_event(file
->private_data
);
3705 * Remove all orphanes events from the context.
3707 static void orphans_remove_work(struct work_struct
*work
)
3709 struct perf_event_context
*ctx
;
3710 struct perf_event
*event
, *tmp
;
3712 ctx
= container_of(work
, struct perf_event_context
,
3713 orphans_remove
.work
);
3715 mutex_lock(&ctx
->mutex
);
3716 list_for_each_entry_safe(event
, tmp
, &ctx
->event_list
, event_entry
) {
3717 struct perf_event
*parent_event
= event
->parent
;
3719 if (!is_orphaned_child(event
))
3722 perf_remove_from_context(event
, true);
3724 mutex_lock(&parent_event
->child_mutex
);
3725 list_del_init(&event
->child_list
);
3726 mutex_unlock(&parent_event
->child_mutex
);
3729 put_event(parent_event
);
3732 raw_spin_lock_irq(&ctx
->lock
);
3733 ctx
->orphans_remove_sched
= false;
3734 raw_spin_unlock_irq(&ctx
->lock
);
3735 mutex_unlock(&ctx
->mutex
);
3740 u64
perf_event_read_value(struct perf_event
*event
, u64
*enabled
, u64
*running
)
3742 struct perf_event
*child
;
3748 mutex_lock(&event
->child_mutex
);
3749 total
+= perf_event_read(event
);
3750 *enabled
+= event
->total_time_enabled
+
3751 atomic64_read(&event
->child_total_time_enabled
);
3752 *running
+= event
->total_time_running
+
3753 atomic64_read(&event
->child_total_time_running
);
3755 list_for_each_entry(child
, &event
->child_list
, child_list
) {
3756 total
+= perf_event_read(child
);
3757 *enabled
+= child
->total_time_enabled
;
3758 *running
+= child
->total_time_running
;
3760 mutex_unlock(&event
->child_mutex
);
3764 EXPORT_SYMBOL_GPL(perf_event_read_value
);
3766 static int perf_event_read_group(struct perf_event
*event
,
3767 u64 read_format
, char __user
*buf
)
3769 struct perf_event
*leader
= event
->group_leader
, *sub
;
3770 struct perf_event_context
*ctx
= leader
->ctx
;
3771 int n
= 0, size
= 0, ret
;
3772 u64 count
, enabled
, running
;
3775 lockdep_assert_held(&ctx
->mutex
);
3777 count
= perf_event_read_value(leader
, &enabled
, &running
);
3779 values
[n
++] = 1 + leader
->nr_siblings
;
3780 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
3781 values
[n
++] = enabled
;
3782 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
3783 values
[n
++] = running
;
3784 values
[n
++] = count
;
3785 if (read_format
& PERF_FORMAT_ID
)
3786 values
[n
++] = primary_event_id(leader
);
3788 size
= n
* sizeof(u64
);
3790 if (copy_to_user(buf
, values
, size
))
3795 list_for_each_entry(sub
, &leader
->sibling_list
, group_entry
) {
3798 values
[n
++] = perf_event_read_value(sub
, &enabled
, &running
);
3799 if (read_format
& PERF_FORMAT_ID
)
3800 values
[n
++] = primary_event_id(sub
);
3802 size
= n
* sizeof(u64
);
3804 if (copy_to_user(buf
+ ret
, values
, size
)) {
3814 static int perf_event_read_one(struct perf_event
*event
,
3815 u64 read_format
, char __user
*buf
)
3817 u64 enabled
, running
;
3821 values
[n
++] = perf_event_read_value(event
, &enabled
, &running
);
3822 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
3823 values
[n
++] = enabled
;
3824 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
3825 values
[n
++] = running
;
3826 if (read_format
& PERF_FORMAT_ID
)
3827 values
[n
++] = primary_event_id(event
);
3829 if (copy_to_user(buf
, values
, n
* sizeof(u64
)))
3832 return n
* sizeof(u64
);
3835 static bool is_event_hup(struct perf_event
*event
)
3839 if (event
->state
!= PERF_EVENT_STATE_EXIT
)
3842 mutex_lock(&event
->child_mutex
);
3843 no_children
= list_empty(&event
->child_list
);
3844 mutex_unlock(&event
->child_mutex
);
3849 * Read the performance event - simple non blocking version for now
3852 perf_read_hw(struct perf_event
*event
, char __user
*buf
, size_t count
)
3854 u64 read_format
= event
->attr
.read_format
;
3858 * Return end-of-file for a read on a event that is in
3859 * error state (i.e. because it was pinned but it couldn't be
3860 * scheduled on to the CPU at some point).
3862 if (event
->state
== PERF_EVENT_STATE_ERROR
)
3865 if (count
< event
->read_size
)
3868 WARN_ON_ONCE(event
->ctx
->parent_ctx
);
3869 if (read_format
& PERF_FORMAT_GROUP
)
3870 ret
= perf_event_read_group(event
, read_format
, buf
);
3872 ret
= perf_event_read_one(event
, read_format
, buf
);
3878 perf_read(struct file
*file
, char __user
*buf
, size_t count
, loff_t
*ppos
)
3880 struct perf_event
*event
= file
->private_data
;
3881 struct perf_event_context
*ctx
;
3884 ctx
= perf_event_ctx_lock(event
);
3885 ret
= perf_read_hw(event
, buf
, count
);
3886 perf_event_ctx_unlock(event
, ctx
);
3891 static unsigned int perf_poll(struct file
*file
, poll_table
*wait
)
3893 struct perf_event
*event
= file
->private_data
;
3894 struct ring_buffer
*rb
;
3895 unsigned int events
= POLLHUP
;
3897 poll_wait(file
, &event
->waitq
, wait
);
3899 if (is_event_hup(event
))
3903 * Pin the event->rb by taking event->mmap_mutex; otherwise
3904 * perf_event_set_output() can swizzle our rb and make us miss wakeups.
3906 mutex_lock(&event
->mmap_mutex
);
3909 events
= atomic_xchg(&rb
->poll
, 0);
3910 mutex_unlock(&event
->mmap_mutex
);
3914 static void _perf_event_reset(struct perf_event
*event
)
3916 (void)perf_event_read(event
);
3917 local64_set(&event
->count
, 0);
3918 perf_event_update_userpage(event
);
3922 * Holding the top-level event's child_mutex means that any
3923 * descendant process that has inherited this event will block
3924 * in sync_child_event if it goes to exit, thus satisfying the
3925 * task existence requirements of perf_event_enable/disable.
3927 static void perf_event_for_each_child(struct perf_event
*event
,
3928 void (*func
)(struct perf_event
*))
3930 struct perf_event
*child
;
3932 WARN_ON_ONCE(event
->ctx
->parent_ctx
);
3934 mutex_lock(&event
->child_mutex
);
3936 list_for_each_entry(child
, &event
->child_list
, child_list
)
3938 mutex_unlock(&event
->child_mutex
);
3941 static void perf_event_for_each(struct perf_event
*event
,
3942 void (*func
)(struct perf_event
*))
3944 struct perf_event_context
*ctx
= event
->ctx
;
3945 struct perf_event
*sibling
;
3947 lockdep_assert_held(&ctx
->mutex
);
3949 event
= event
->group_leader
;
3951 perf_event_for_each_child(event
, func
);
3952 list_for_each_entry(sibling
, &event
->sibling_list
, group_entry
)
3953 perf_event_for_each_child(sibling
, func
);
3956 static int perf_event_period(struct perf_event
*event
, u64 __user
*arg
)
3958 struct perf_event_context
*ctx
= event
->ctx
;
3959 int ret
= 0, active
;
3962 if (!is_sampling_event(event
))
3965 if (copy_from_user(&value
, arg
, sizeof(value
)))
3971 raw_spin_lock_irq(&ctx
->lock
);
3972 if (event
->attr
.freq
) {
3973 if (value
> sysctl_perf_event_sample_rate
) {
3978 event
->attr
.sample_freq
= value
;
3980 event
->attr
.sample_period
= value
;
3981 event
->hw
.sample_period
= value
;
3984 active
= (event
->state
== PERF_EVENT_STATE_ACTIVE
);
3986 perf_pmu_disable(ctx
->pmu
);
3987 event
->pmu
->stop(event
, PERF_EF_UPDATE
);
3990 local64_set(&event
->hw
.period_left
, 0);
3993 event
->pmu
->start(event
, PERF_EF_RELOAD
);
3994 perf_pmu_enable(ctx
->pmu
);
3998 raw_spin_unlock_irq(&ctx
->lock
);
4003 static const struct file_operations perf_fops
;
4005 static inline int perf_fget_light(int fd
, struct fd
*p
)
4007 struct fd f
= fdget(fd
);
4011 if (f
.file
->f_op
!= &perf_fops
) {
4019 static int perf_event_set_output(struct perf_event
*event
,
4020 struct perf_event
*output_event
);
4021 static int perf_event_set_filter(struct perf_event
*event
, void __user
*arg
);
4022 static int perf_event_set_bpf_prog(struct perf_event
*event
, u32 prog_fd
);
4024 static long _perf_ioctl(struct perf_event
*event
, unsigned int cmd
, unsigned long arg
)
4026 void (*func
)(struct perf_event
*);
4030 case PERF_EVENT_IOC_ENABLE
:
4031 func
= _perf_event_enable
;
4033 case PERF_EVENT_IOC_DISABLE
:
4034 func
= _perf_event_disable
;
4036 case PERF_EVENT_IOC_RESET
:
4037 func
= _perf_event_reset
;
4040 case PERF_EVENT_IOC_REFRESH
:
4041 return _perf_event_refresh(event
, arg
);
4043 case PERF_EVENT_IOC_PERIOD
:
4044 return perf_event_period(event
, (u64 __user
*)arg
);
4046 case PERF_EVENT_IOC_ID
:
4048 u64 id
= primary_event_id(event
);
4050 if (copy_to_user((void __user
*)arg
, &id
, sizeof(id
)))
4055 case PERF_EVENT_IOC_SET_OUTPUT
:
4059 struct perf_event
*output_event
;
4061 ret
= perf_fget_light(arg
, &output
);
4064 output_event
= output
.file
->private_data
;
4065 ret
= perf_event_set_output(event
, output_event
);
4068 ret
= perf_event_set_output(event
, NULL
);
4073 case PERF_EVENT_IOC_SET_FILTER
:
4074 return perf_event_set_filter(event
, (void __user
*)arg
);
4076 case PERF_EVENT_IOC_SET_BPF
:
4077 return perf_event_set_bpf_prog(event
, arg
);
4083 if (flags
& PERF_IOC_FLAG_GROUP
)
4084 perf_event_for_each(event
, func
);
4086 perf_event_for_each_child(event
, func
);
4091 static long perf_ioctl(struct file
*file
, unsigned int cmd
, unsigned long arg
)
4093 struct perf_event
*event
= file
->private_data
;
4094 struct perf_event_context
*ctx
;
4097 ctx
= perf_event_ctx_lock(event
);
4098 ret
= _perf_ioctl(event
, cmd
, arg
);
4099 perf_event_ctx_unlock(event
, ctx
);
4104 #ifdef CONFIG_COMPAT
4105 static long perf_compat_ioctl(struct file
*file
, unsigned int cmd
,
4108 switch (_IOC_NR(cmd
)) {
4109 case _IOC_NR(PERF_EVENT_IOC_SET_FILTER
):
4110 case _IOC_NR(PERF_EVENT_IOC_ID
):
4111 /* Fix up pointer size (usually 4 -> 8 in 32-on-64-bit case */
4112 if (_IOC_SIZE(cmd
) == sizeof(compat_uptr_t
)) {
4113 cmd
&= ~IOCSIZE_MASK
;
4114 cmd
|= sizeof(void *) << IOCSIZE_SHIFT
;
4118 return perf_ioctl(file
, cmd
, arg
);
4121 # define perf_compat_ioctl NULL
4124 int perf_event_task_enable(void)
4126 struct perf_event_context
*ctx
;
4127 struct perf_event
*event
;
4129 mutex_lock(¤t
->perf_event_mutex
);
4130 list_for_each_entry(event
, ¤t
->perf_event_list
, owner_entry
) {
4131 ctx
= perf_event_ctx_lock(event
);
4132 perf_event_for_each_child(event
, _perf_event_enable
);
4133 perf_event_ctx_unlock(event
, ctx
);
4135 mutex_unlock(¤t
->perf_event_mutex
);
4140 int perf_event_task_disable(void)
4142 struct perf_event_context
*ctx
;
4143 struct perf_event
*event
;
4145 mutex_lock(¤t
->perf_event_mutex
);
4146 list_for_each_entry(event
, ¤t
->perf_event_list
, owner_entry
) {
4147 ctx
= perf_event_ctx_lock(event
);
4148 perf_event_for_each_child(event
, _perf_event_disable
);
4149 perf_event_ctx_unlock(event
, ctx
);
4151 mutex_unlock(¤t
->perf_event_mutex
);
4156 static int perf_event_index(struct perf_event
*event
)
4158 if (event
->hw
.state
& PERF_HES_STOPPED
)
4161 if (event
->state
!= PERF_EVENT_STATE_ACTIVE
)
4164 return event
->pmu
->event_idx(event
);
4167 static void calc_timer_values(struct perf_event
*event
,
4174 *now
= perf_clock();
4175 ctx_time
= event
->shadow_ctx_time
+ *now
;
4176 *enabled
= ctx_time
- event
->tstamp_enabled
;
4177 *running
= ctx_time
- event
->tstamp_running
;
4180 static void perf_event_init_userpage(struct perf_event
*event
)
4182 struct perf_event_mmap_page
*userpg
;
4183 struct ring_buffer
*rb
;
4186 rb
= rcu_dereference(event
->rb
);
4190 userpg
= rb
->user_page
;
4192 /* Allow new userspace to detect that bit 0 is deprecated */
4193 userpg
->cap_bit0_is_deprecated
= 1;
4194 userpg
->size
= offsetof(struct perf_event_mmap_page
, __reserved
);
4195 userpg
->data_offset
= PAGE_SIZE
;
4196 userpg
->data_size
= perf_data_size(rb
);
4202 void __weak
arch_perf_update_userpage(
4203 struct perf_event
*event
, struct perf_event_mmap_page
*userpg
, u64 now
)
4208 * Callers need to ensure there can be no nesting of this function, otherwise
4209 * the seqlock logic goes bad. We can not serialize this because the arch
4210 * code calls this from NMI context.
4212 void perf_event_update_userpage(struct perf_event
*event
)
4214 struct perf_event_mmap_page
*userpg
;
4215 struct ring_buffer
*rb
;
4216 u64 enabled
, running
, now
;
4219 rb
= rcu_dereference(event
->rb
);
4224 * compute total_time_enabled, total_time_running
4225 * based on snapshot values taken when the event
4226 * was last scheduled in.
4228 * we cannot simply called update_context_time()
4229 * because of locking issue as we can be called in
4232 calc_timer_values(event
, &now
, &enabled
, &running
);
4234 userpg
= rb
->user_page
;
4236 * Disable preemption so as to not let the corresponding user-space
4237 * spin too long if we get preempted.
4242 userpg
->index
= perf_event_index(event
);
4243 userpg
->offset
= perf_event_count(event
);
4245 userpg
->offset
-= local64_read(&event
->hw
.prev_count
);
4247 userpg
->time_enabled
= enabled
+
4248 atomic64_read(&event
->child_total_time_enabled
);
4250 userpg
->time_running
= running
+
4251 atomic64_read(&event
->child_total_time_running
);
4253 arch_perf_update_userpage(event
, userpg
, now
);
4262 static int perf_mmap_fault(struct vm_area_struct
*vma
, struct vm_fault
*vmf
)
4264 struct perf_event
*event
= vma
->vm_file
->private_data
;
4265 struct ring_buffer
*rb
;
4266 int ret
= VM_FAULT_SIGBUS
;
4268 if (vmf
->flags
& FAULT_FLAG_MKWRITE
) {
4269 if (vmf
->pgoff
== 0)
4275 rb
= rcu_dereference(event
->rb
);
4279 if (vmf
->pgoff
&& (vmf
->flags
& FAULT_FLAG_WRITE
))
4282 vmf
->page
= perf_mmap_to_page(rb
, vmf
->pgoff
);
4286 get_page(vmf
->page
);
4287 vmf
->page
->mapping
= vma
->vm_file
->f_mapping
;
4288 vmf
->page
->index
= vmf
->pgoff
;
4297 static void ring_buffer_attach(struct perf_event
*event
,
4298 struct ring_buffer
*rb
)
4300 struct ring_buffer
*old_rb
= NULL
;
4301 unsigned long flags
;
4305 * Should be impossible, we set this when removing
4306 * event->rb_entry and wait/clear when adding event->rb_entry.
4308 WARN_ON_ONCE(event
->rcu_pending
);
4311 event
->rcu_batches
= get_state_synchronize_rcu();
4312 event
->rcu_pending
= 1;
4314 spin_lock_irqsave(&old_rb
->event_lock
, flags
);
4315 list_del_rcu(&event
->rb_entry
);
4316 spin_unlock_irqrestore(&old_rb
->event_lock
, flags
);
4319 if (event
->rcu_pending
&& rb
) {
4320 cond_synchronize_rcu(event
->rcu_batches
);
4321 event
->rcu_pending
= 0;
4325 spin_lock_irqsave(&rb
->event_lock
, flags
);
4326 list_add_rcu(&event
->rb_entry
, &rb
->event_list
);
4327 spin_unlock_irqrestore(&rb
->event_lock
, flags
);
4330 rcu_assign_pointer(event
->rb
, rb
);
4333 ring_buffer_put(old_rb
);
4335 * Since we detached before setting the new rb, so that we
4336 * could attach the new rb, we could have missed a wakeup.
4339 wake_up_all(&event
->waitq
);
4343 static void ring_buffer_wakeup(struct perf_event
*event
)
4345 struct ring_buffer
*rb
;
4348 rb
= rcu_dereference(event
->rb
);
4350 list_for_each_entry_rcu(event
, &rb
->event_list
, rb_entry
)
4351 wake_up_all(&event
->waitq
);
4356 static void rb_free_rcu(struct rcu_head
*rcu_head
)
4358 struct ring_buffer
*rb
;
4360 rb
= container_of(rcu_head
, struct ring_buffer
, rcu_head
);
4364 struct ring_buffer
*ring_buffer_get(struct perf_event
*event
)
4366 struct ring_buffer
*rb
;
4369 rb
= rcu_dereference(event
->rb
);
4371 if (!atomic_inc_not_zero(&rb
->refcount
))
4379 void ring_buffer_put(struct ring_buffer
*rb
)
4381 if (!atomic_dec_and_test(&rb
->refcount
))
4384 WARN_ON_ONCE(!list_empty(&rb
->event_list
));
4386 call_rcu(&rb
->rcu_head
, rb_free_rcu
);
4389 static void perf_mmap_open(struct vm_area_struct
*vma
)
4391 struct perf_event
*event
= vma
->vm_file
->private_data
;
4393 atomic_inc(&event
->mmap_count
);
4394 atomic_inc(&event
->rb
->mmap_count
);
4397 atomic_inc(&event
->rb
->aux_mmap_count
);
4399 if (event
->pmu
->event_mapped
)
4400 event
->pmu
->event_mapped(event
);
4404 * A buffer can be mmap()ed multiple times; either directly through the same
4405 * event, or through other events by use of perf_event_set_output().
4407 * In order to undo the VM accounting done by perf_mmap() we need to destroy
4408 * the buffer here, where we still have a VM context. This means we need
4409 * to detach all events redirecting to us.
4411 static void perf_mmap_close(struct vm_area_struct
*vma
)
4413 struct perf_event
*event
= vma
->vm_file
->private_data
;
4415 struct ring_buffer
*rb
= ring_buffer_get(event
);
4416 struct user_struct
*mmap_user
= rb
->mmap_user
;
4417 int mmap_locked
= rb
->mmap_locked
;
4418 unsigned long size
= perf_data_size(rb
);
4420 if (event
->pmu
->event_unmapped
)
4421 event
->pmu
->event_unmapped(event
);
4424 * rb->aux_mmap_count will always drop before rb->mmap_count and
4425 * event->mmap_count, so it is ok to use event->mmap_mutex to
4426 * serialize with perf_mmap here.
4428 if (rb_has_aux(rb
) && vma
->vm_pgoff
== rb
->aux_pgoff
&&
4429 atomic_dec_and_mutex_lock(&rb
->aux_mmap_count
, &event
->mmap_mutex
)) {
4430 atomic_long_sub(rb
->aux_nr_pages
, &mmap_user
->locked_vm
);
4431 vma
->vm_mm
->pinned_vm
-= rb
->aux_mmap_locked
;
4434 mutex_unlock(&event
->mmap_mutex
);
4437 atomic_dec(&rb
->mmap_count
);
4439 if (!atomic_dec_and_mutex_lock(&event
->mmap_count
, &event
->mmap_mutex
))
4442 ring_buffer_attach(event
, NULL
);
4443 mutex_unlock(&event
->mmap_mutex
);
4445 /* If there's still other mmap()s of this buffer, we're done. */
4446 if (atomic_read(&rb
->mmap_count
))
4450 * No other mmap()s, detach from all other events that might redirect
4451 * into the now unreachable buffer. Somewhat complicated by the
4452 * fact that rb::event_lock otherwise nests inside mmap_mutex.
4456 list_for_each_entry_rcu(event
, &rb
->event_list
, rb_entry
) {
4457 if (!atomic_long_inc_not_zero(&event
->refcount
)) {
4459 * This event is en-route to free_event() which will
4460 * detach it and remove it from the list.
4466 mutex_lock(&event
->mmap_mutex
);
4468 * Check we didn't race with perf_event_set_output() which can
4469 * swizzle the rb from under us while we were waiting to
4470 * acquire mmap_mutex.
4472 * If we find a different rb; ignore this event, a next
4473 * iteration will no longer find it on the list. We have to
4474 * still restart the iteration to make sure we're not now
4475 * iterating the wrong list.
4477 if (event
->rb
== rb
)
4478 ring_buffer_attach(event
, NULL
);
4480 mutex_unlock(&event
->mmap_mutex
);
4484 * Restart the iteration; either we're on the wrong list or
4485 * destroyed its integrity by doing a deletion.
4492 * It could be there's still a few 0-ref events on the list; they'll
4493 * get cleaned up by free_event() -- they'll also still have their
4494 * ref on the rb and will free it whenever they are done with it.
4496 * Aside from that, this buffer is 'fully' detached and unmapped,
4497 * undo the VM accounting.
4500 atomic_long_sub((size
>> PAGE_SHIFT
) + 1, &mmap_user
->locked_vm
);
4501 vma
->vm_mm
->pinned_vm
-= mmap_locked
;
4502 free_uid(mmap_user
);
4505 ring_buffer_put(rb
); /* could be last */
4508 static const struct vm_operations_struct perf_mmap_vmops
= {
4509 .open
= perf_mmap_open
,
4510 .close
= perf_mmap_close
, /* non mergable */
4511 .fault
= perf_mmap_fault
,
4512 .page_mkwrite
= perf_mmap_fault
,
4515 static int perf_mmap(struct file
*file
, struct vm_area_struct
*vma
)
4517 struct perf_event
*event
= file
->private_data
;
4518 unsigned long user_locked
, user_lock_limit
;
4519 struct user_struct
*user
= current_user();
4520 unsigned long locked
, lock_limit
;
4521 struct ring_buffer
*rb
= NULL
;
4522 unsigned long vma_size
;
4523 unsigned long nr_pages
;
4524 long user_extra
= 0, extra
= 0;
4525 int ret
= 0, flags
= 0;
4528 * Don't allow mmap() of inherited per-task counters. This would
4529 * create a performance issue due to all children writing to the
4532 if (event
->cpu
== -1 && event
->attr
.inherit
)
4535 if (!(vma
->vm_flags
& VM_SHARED
))
4538 vma_size
= vma
->vm_end
- vma
->vm_start
;
4540 if (vma
->vm_pgoff
== 0) {
4541 nr_pages
= (vma_size
/ PAGE_SIZE
) - 1;
4544 * AUX area mapping: if rb->aux_nr_pages != 0, it's already
4545 * mapped, all subsequent mappings should have the same size
4546 * and offset. Must be above the normal perf buffer.
4548 u64 aux_offset
, aux_size
;
4553 nr_pages
= vma_size
/ PAGE_SIZE
;
4555 mutex_lock(&event
->mmap_mutex
);
4562 aux_offset
= ACCESS_ONCE(rb
->user_page
->aux_offset
);
4563 aux_size
= ACCESS_ONCE(rb
->user_page
->aux_size
);
4565 if (aux_offset
< perf_data_size(rb
) + PAGE_SIZE
)
4568 if (aux_offset
!= vma
->vm_pgoff
<< PAGE_SHIFT
)
4571 /* already mapped with a different offset */
4572 if (rb_has_aux(rb
) && rb
->aux_pgoff
!= vma
->vm_pgoff
)
4575 if (aux_size
!= vma_size
|| aux_size
!= nr_pages
* PAGE_SIZE
)
4578 /* already mapped with a different size */
4579 if (rb_has_aux(rb
) && rb
->aux_nr_pages
!= nr_pages
)
4582 if (!is_power_of_2(nr_pages
))
4585 if (!atomic_inc_not_zero(&rb
->mmap_count
))
4588 if (rb_has_aux(rb
)) {
4589 atomic_inc(&rb
->aux_mmap_count
);
4594 atomic_set(&rb
->aux_mmap_count
, 1);
4595 user_extra
= nr_pages
;
4601 * If we have rb pages ensure they're a power-of-two number, so we
4602 * can do bitmasks instead of modulo.
4604 if (nr_pages
!= 0 && !is_power_of_2(nr_pages
))
4607 if (vma_size
!= PAGE_SIZE
* (1 + nr_pages
))
4610 WARN_ON_ONCE(event
->ctx
->parent_ctx
);
4612 mutex_lock(&event
->mmap_mutex
);
4614 if (event
->rb
->nr_pages
!= nr_pages
) {
4619 if (!atomic_inc_not_zero(&event
->rb
->mmap_count
)) {
4621 * Raced against perf_mmap_close() through
4622 * perf_event_set_output(). Try again, hope for better
4625 mutex_unlock(&event
->mmap_mutex
);
4632 user_extra
= nr_pages
+ 1;
4635 user_lock_limit
= sysctl_perf_event_mlock
>> (PAGE_SHIFT
- 10);
4638 * Increase the limit linearly with more CPUs:
4640 user_lock_limit
*= num_online_cpus();
4642 user_locked
= atomic_long_read(&user
->locked_vm
) + user_extra
;
4644 if (user_locked
> user_lock_limit
)
4645 extra
= user_locked
- user_lock_limit
;
4647 lock_limit
= rlimit(RLIMIT_MEMLOCK
);
4648 lock_limit
>>= PAGE_SHIFT
;
4649 locked
= vma
->vm_mm
->pinned_vm
+ extra
;
4651 if ((locked
> lock_limit
) && perf_paranoid_tracepoint_raw() &&
4652 !capable(CAP_IPC_LOCK
)) {
4657 WARN_ON(!rb
&& event
->rb
);
4659 if (vma
->vm_flags
& VM_WRITE
)
4660 flags
|= RING_BUFFER_WRITABLE
;
4663 rb
= rb_alloc(nr_pages
,
4664 event
->attr
.watermark
? event
->attr
.wakeup_watermark
: 0,
4672 atomic_set(&rb
->mmap_count
, 1);
4673 rb
->mmap_user
= get_current_user();
4674 rb
->mmap_locked
= extra
;
4676 ring_buffer_attach(event
, rb
);
4678 perf_event_init_userpage(event
);
4679 perf_event_update_userpage(event
);
4681 ret
= rb_alloc_aux(rb
, event
, vma
->vm_pgoff
, nr_pages
,
4682 event
->attr
.aux_watermark
, flags
);
4684 rb
->aux_mmap_locked
= extra
;
4689 atomic_long_add(user_extra
, &user
->locked_vm
);
4690 vma
->vm_mm
->pinned_vm
+= extra
;
4692 atomic_inc(&event
->mmap_count
);
4694 atomic_dec(&rb
->mmap_count
);
4697 mutex_unlock(&event
->mmap_mutex
);
4700 * Since pinned accounting is per vm we cannot allow fork() to copy our
4703 vma
->vm_flags
|= VM_DONTCOPY
| VM_DONTEXPAND
| VM_DONTDUMP
;
4704 vma
->vm_ops
= &perf_mmap_vmops
;
4706 if (event
->pmu
->event_mapped
)
4707 event
->pmu
->event_mapped(event
);
4712 static int perf_fasync(int fd
, struct file
*filp
, int on
)
4714 struct inode
*inode
= file_inode(filp
);
4715 struct perf_event
*event
= filp
->private_data
;
4718 mutex_lock(&inode
->i_mutex
);
4719 retval
= fasync_helper(fd
, filp
, on
, &event
->fasync
);
4720 mutex_unlock(&inode
->i_mutex
);
4728 static const struct file_operations perf_fops
= {
4729 .llseek
= no_llseek
,
4730 .release
= perf_release
,
4733 .unlocked_ioctl
= perf_ioctl
,
4734 .compat_ioctl
= perf_compat_ioctl
,
4736 .fasync
= perf_fasync
,
4742 * If there's data, ensure we set the poll() state and publish everything
4743 * to user-space before waking everybody up.
4746 void perf_event_wakeup(struct perf_event
*event
)
4748 ring_buffer_wakeup(event
);
4750 if (event
->pending_kill
) {
4751 kill_fasync(&event
->fasync
, SIGIO
, event
->pending_kill
);
4752 event
->pending_kill
= 0;
4756 static void perf_pending_event(struct irq_work
*entry
)
4758 struct perf_event
*event
= container_of(entry
,
4759 struct perf_event
, pending
);
4762 rctx
= perf_swevent_get_recursion_context();
4764 * If we 'fail' here, that's OK, it means recursion is already disabled
4765 * and we won't recurse 'further'.
4768 if (event
->pending_disable
) {
4769 event
->pending_disable
= 0;
4770 __perf_event_disable(event
);
4773 if (event
->pending_wakeup
) {
4774 event
->pending_wakeup
= 0;
4775 perf_event_wakeup(event
);
4779 perf_swevent_put_recursion_context(rctx
);
4783 * We assume there is only KVM supporting the callbacks.
4784 * Later on, we might change it to a list if there is
4785 * another virtualization implementation supporting the callbacks.
4787 struct perf_guest_info_callbacks
*perf_guest_cbs
;
4789 int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks
*cbs
)
4791 perf_guest_cbs
= cbs
;
4794 EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks
);
4796 int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks
*cbs
)
4798 perf_guest_cbs
= NULL
;
4801 EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks
);
4804 perf_output_sample_regs(struct perf_output_handle
*handle
,
4805 struct pt_regs
*regs
, u64 mask
)
4809 for_each_set_bit(bit
, (const unsigned long *) &mask
,
4810 sizeof(mask
) * BITS_PER_BYTE
) {
4813 val
= perf_reg_value(regs
, bit
);
4814 perf_output_put(handle
, val
);
4818 static void perf_sample_regs_user(struct perf_regs
*regs_user
,
4819 struct pt_regs
*regs
,
4820 struct pt_regs
*regs_user_copy
)
4822 if (user_mode(regs
)) {
4823 regs_user
->abi
= perf_reg_abi(current
);
4824 regs_user
->regs
= regs
;
4825 } else if (current
->mm
) {
4826 perf_get_regs_user(regs_user
, regs
, regs_user_copy
);
4828 regs_user
->abi
= PERF_SAMPLE_REGS_ABI_NONE
;
4829 regs_user
->regs
= NULL
;
4833 static void perf_sample_regs_intr(struct perf_regs
*regs_intr
,
4834 struct pt_regs
*regs
)
4836 regs_intr
->regs
= regs
;
4837 regs_intr
->abi
= perf_reg_abi(current
);
4842 * Get remaining task size from user stack pointer.
4844 * It'd be better to take stack vma map and limit this more
4845 * precisly, but there's no way to get it safely under interrupt,
4846 * so using TASK_SIZE as limit.
4848 static u64
perf_ustack_task_size(struct pt_regs
*regs
)
4850 unsigned long addr
= perf_user_stack_pointer(regs
);
4852 if (!addr
|| addr
>= TASK_SIZE
)
4855 return TASK_SIZE
- addr
;
4859 perf_sample_ustack_size(u16 stack_size
, u16 header_size
,
4860 struct pt_regs
*regs
)
4864 /* No regs, no stack pointer, no dump. */
4869 * Check if we fit in with the requested stack size into the:
4871 * If we don't, we limit the size to the TASK_SIZE.
4873 * - remaining sample size
4874 * If we don't, we customize the stack size to
4875 * fit in to the remaining sample size.
4878 task_size
= min((u64
) USHRT_MAX
, perf_ustack_task_size(regs
));
4879 stack_size
= min(stack_size
, (u16
) task_size
);
4881 /* Current header size plus static size and dynamic size. */
4882 header_size
+= 2 * sizeof(u64
);
4884 /* Do we fit in with the current stack dump size? */
4885 if ((u16
) (header_size
+ stack_size
) < header_size
) {
4887 * If we overflow the maximum size for the sample,
4888 * we customize the stack dump size to fit in.
4890 stack_size
= USHRT_MAX
- header_size
- sizeof(u64
);
4891 stack_size
= round_up(stack_size
, sizeof(u64
));
4898 perf_output_sample_ustack(struct perf_output_handle
*handle
, u64 dump_size
,
4899 struct pt_regs
*regs
)
4901 /* Case of a kernel thread, nothing to dump */
4904 perf_output_put(handle
, size
);
4913 * - the size requested by user or the best one we can fit
4914 * in to the sample max size
4916 * - user stack dump data
4918 * - the actual dumped size
4922 perf_output_put(handle
, dump_size
);
4925 sp
= perf_user_stack_pointer(regs
);
4926 rem
= __output_copy_user(handle
, (void *) sp
, dump_size
);
4927 dyn_size
= dump_size
- rem
;
4929 perf_output_skip(handle
, rem
);
4932 perf_output_put(handle
, dyn_size
);
4936 static void __perf_event_header__init_id(struct perf_event_header
*header
,
4937 struct perf_sample_data
*data
,
4938 struct perf_event
*event
)
4940 u64 sample_type
= event
->attr
.sample_type
;
4942 data
->type
= sample_type
;
4943 header
->size
+= event
->id_header_size
;
4945 if (sample_type
& PERF_SAMPLE_TID
) {
4946 /* namespace issues */
4947 data
->tid_entry
.pid
= perf_event_pid(event
, current
);
4948 data
->tid_entry
.tid
= perf_event_tid(event
, current
);
4951 if (sample_type
& PERF_SAMPLE_TIME
)
4952 data
->time
= perf_event_clock(event
);
4954 if (sample_type
& (PERF_SAMPLE_ID
| PERF_SAMPLE_IDENTIFIER
))
4955 data
->id
= primary_event_id(event
);
4957 if (sample_type
& PERF_SAMPLE_STREAM_ID
)
4958 data
->stream_id
= event
->id
;
4960 if (sample_type
& PERF_SAMPLE_CPU
) {
4961 data
->cpu_entry
.cpu
= raw_smp_processor_id();
4962 data
->cpu_entry
.reserved
= 0;
4966 void perf_event_header__init_id(struct perf_event_header
*header
,
4967 struct perf_sample_data
*data
,
4968 struct perf_event
*event
)
4970 if (event
->attr
.sample_id_all
)
4971 __perf_event_header__init_id(header
, data
, event
);
4974 static void __perf_event__output_id_sample(struct perf_output_handle
*handle
,
4975 struct perf_sample_data
*data
)
4977 u64 sample_type
= data
->type
;
4979 if (sample_type
& PERF_SAMPLE_TID
)
4980 perf_output_put(handle
, data
->tid_entry
);
4982 if (sample_type
& PERF_SAMPLE_TIME
)
4983 perf_output_put(handle
, data
->time
);
4985 if (sample_type
& PERF_SAMPLE_ID
)
4986 perf_output_put(handle
, data
->id
);
4988 if (sample_type
& PERF_SAMPLE_STREAM_ID
)
4989 perf_output_put(handle
, data
->stream_id
);
4991 if (sample_type
& PERF_SAMPLE_CPU
)
4992 perf_output_put(handle
, data
->cpu_entry
);
4994 if (sample_type
& PERF_SAMPLE_IDENTIFIER
)
4995 perf_output_put(handle
, data
->id
);
4998 void perf_event__output_id_sample(struct perf_event
*event
,
4999 struct perf_output_handle
*handle
,
5000 struct perf_sample_data
*sample
)
5002 if (event
->attr
.sample_id_all
)
5003 __perf_event__output_id_sample(handle
, sample
);
5006 static void perf_output_read_one(struct perf_output_handle
*handle
,
5007 struct perf_event
*event
,
5008 u64 enabled
, u64 running
)
5010 u64 read_format
= event
->attr
.read_format
;
5014 values
[n
++] = perf_event_count(event
);
5015 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
) {
5016 values
[n
++] = enabled
+
5017 atomic64_read(&event
->child_total_time_enabled
);
5019 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
) {
5020 values
[n
++] = running
+
5021 atomic64_read(&event
->child_total_time_running
);
5023 if (read_format
& PERF_FORMAT_ID
)
5024 values
[n
++] = primary_event_id(event
);
5026 __output_copy(handle
, values
, n
* sizeof(u64
));
5030 * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
5032 static void perf_output_read_group(struct perf_output_handle
*handle
,
5033 struct perf_event
*event
,
5034 u64 enabled
, u64 running
)
5036 struct perf_event
*leader
= event
->group_leader
, *sub
;
5037 u64 read_format
= event
->attr
.read_format
;
5041 values
[n
++] = 1 + leader
->nr_siblings
;
5043 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
5044 values
[n
++] = enabled
;
5046 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
5047 values
[n
++] = running
;
5049 if (leader
!= event
)
5050 leader
->pmu
->read(leader
);
5052 values
[n
++] = perf_event_count(leader
);
5053 if (read_format
& PERF_FORMAT_ID
)
5054 values
[n
++] = primary_event_id(leader
);
5056 __output_copy(handle
, values
, n
* sizeof(u64
));
5058 list_for_each_entry(sub
, &leader
->sibling_list
, group_entry
) {
5061 if ((sub
!= event
) &&
5062 (sub
->state
== PERF_EVENT_STATE_ACTIVE
))
5063 sub
->pmu
->read(sub
);
5065 values
[n
++] = perf_event_count(sub
);
5066 if (read_format
& PERF_FORMAT_ID
)
5067 values
[n
++] = primary_event_id(sub
);
5069 __output_copy(handle
, values
, n
* sizeof(u64
));
5073 #define PERF_FORMAT_TOTAL_TIMES (PERF_FORMAT_TOTAL_TIME_ENABLED|\
5074 PERF_FORMAT_TOTAL_TIME_RUNNING)
5076 static void perf_output_read(struct perf_output_handle
*handle
,
5077 struct perf_event
*event
)
5079 u64 enabled
= 0, running
= 0, now
;
5080 u64 read_format
= event
->attr
.read_format
;
5083 * compute total_time_enabled, total_time_running
5084 * based on snapshot values taken when the event
5085 * was last scheduled in.
5087 * we cannot simply called update_context_time()
5088 * because of locking issue as we are called in
5091 if (read_format
& PERF_FORMAT_TOTAL_TIMES
)
5092 calc_timer_values(event
, &now
, &enabled
, &running
);
5094 if (event
->attr
.read_format
& PERF_FORMAT_GROUP
)
5095 perf_output_read_group(handle
, event
, enabled
, running
);
5097 perf_output_read_one(handle
, event
, enabled
, running
);
5100 void perf_output_sample(struct perf_output_handle
*handle
,
5101 struct perf_event_header
*header
,
5102 struct perf_sample_data
*data
,
5103 struct perf_event
*event
)
5105 u64 sample_type
= data
->type
;
5107 perf_output_put(handle
, *header
);
5109 if (sample_type
& PERF_SAMPLE_IDENTIFIER
)
5110 perf_output_put(handle
, data
->id
);
5112 if (sample_type
& PERF_SAMPLE_IP
)
5113 perf_output_put(handle
, data
->ip
);
5115 if (sample_type
& PERF_SAMPLE_TID
)
5116 perf_output_put(handle
, data
->tid_entry
);
5118 if (sample_type
& PERF_SAMPLE_TIME
)
5119 perf_output_put(handle
, data
->time
);
5121 if (sample_type
& PERF_SAMPLE_ADDR
)
5122 perf_output_put(handle
, data
->addr
);
5124 if (sample_type
& PERF_SAMPLE_ID
)
5125 perf_output_put(handle
, data
->id
);
5127 if (sample_type
& PERF_SAMPLE_STREAM_ID
)
5128 perf_output_put(handle
, data
->stream_id
);
5130 if (sample_type
& PERF_SAMPLE_CPU
)
5131 perf_output_put(handle
, data
->cpu_entry
);
5133 if (sample_type
& PERF_SAMPLE_PERIOD
)
5134 perf_output_put(handle
, data
->period
);
5136 if (sample_type
& PERF_SAMPLE_READ
)
5137 perf_output_read(handle
, event
);
5139 if (sample_type
& PERF_SAMPLE_CALLCHAIN
) {
5140 if (data
->callchain
) {
5143 if (data
->callchain
)
5144 size
+= data
->callchain
->nr
;
5146 size
*= sizeof(u64
);
5148 __output_copy(handle
, data
->callchain
, size
);
5151 perf_output_put(handle
, nr
);
5155 if (sample_type
& PERF_SAMPLE_RAW
) {
5157 perf_output_put(handle
, data
->raw
->size
);
5158 __output_copy(handle
, data
->raw
->data
,
5165 .size
= sizeof(u32
),
5168 perf_output_put(handle
, raw
);
5172 if (sample_type
& PERF_SAMPLE_BRANCH_STACK
) {
5173 if (data
->br_stack
) {
5176 size
= data
->br_stack
->nr
5177 * sizeof(struct perf_branch_entry
);
5179 perf_output_put(handle
, data
->br_stack
->nr
);
5180 perf_output_copy(handle
, data
->br_stack
->entries
, size
);
5183 * we always store at least the value of nr
5186 perf_output_put(handle
, nr
);
5190 if (sample_type
& PERF_SAMPLE_REGS_USER
) {
5191 u64 abi
= data
->regs_user
.abi
;
5194 * If there are no regs to dump, notice it through
5195 * first u64 being zero (PERF_SAMPLE_REGS_ABI_NONE).
5197 perf_output_put(handle
, abi
);
5200 u64 mask
= event
->attr
.sample_regs_user
;
5201 perf_output_sample_regs(handle
,
5202 data
->regs_user
.regs
,
5207 if (sample_type
& PERF_SAMPLE_STACK_USER
) {
5208 perf_output_sample_ustack(handle
,
5209 data
->stack_user_size
,
5210 data
->regs_user
.regs
);
5213 if (sample_type
& PERF_SAMPLE_WEIGHT
)
5214 perf_output_put(handle
, data
->weight
);
5216 if (sample_type
& PERF_SAMPLE_DATA_SRC
)
5217 perf_output_put(handle
, data
->data_src
.val
);
5219 if (sample_type
& PERF_SAMPLE_TRANSACTION
)
5220 perf_output_put(handle
, data
->txn
);
5222 if (sample_type
& PERF_SAMPLE_REGS_INTR
) {
5223 u64 abi
= data
->regs_intr
.abi
;
5225 * If there are no regs to dump, notice it through
5226 * first u64 being zero (PERF_SAMPLE_REGS_ABI_NONE).
5228 perf_output_put(handle
, abi
);
5231 u64 mask
= event
->attr
.sample_regs_intr
;
5233 perf_output_sample_regs(handle
,
5234 data
->regs_intr
.regs
,
5239 if (!event
->attr
.watermark
) {
5240 int wakeup_events
= event
->attr
.wakeup_events
;
5242 if (wakeup_events
) {
5243 struct ring_buffer
*rb
= handle
->rb
;
5244 int events
= local_inc_return(&rb
->events
);
5246 if (events
>= wakeup_events
) {
5247 local_sub(wakeup_events
, &rb
->events
);
5248 local_inc(&rb
->wakeup
);
5254 void perf_prepare_sample(struct perf_event_header
*header
,
5255 struct perf_sample_data
*data
,
5256 struct perf_event
*event
,
5257 struct pt_regs
*regs
)
5259 u64 sample_type
= event
->attr
.sample_type
;
5261 header
->type
= PERF_RECORD_SAMPLE
;
5262 header
->size
= sizeof(*header
) + event
->header_size
;
5265 header
->misc
|= perf_misc_flags(regs
);
5267 __perf_event_header__init_id(header
, data
, event
);
5269 if (sample_type
& PERF_SAMPLE_IP
)
5270 data
->ip
= perf_instruction_pointer(regs
);
5272 if (sample_type
& PERF_SAMPLE_CALLCHAIN
) {
5275 data
->callchain
= perf_callchain(event
, regs
);
5277 if (data
->callchain
)
5278 size
+= data
->callchain
->nr
;
5280 header
->size
+= size
* sizeof(u64
);
5283 if (sample_type
& PERF_SAMPLE_RAW
) {
5284 int size
= sizeof(u32
);
5287 size
+= data
->raw
->size
;
5289 size
+= sizeof(u32
);
5291 WARN_ON_ONCE(size
& (sizeof(u64
)-1));
5292 header
->size
+= size
;
5295 if (sample_type
& PERF_SAMPLE_BRANCH_STACK
) {
5296 int size
= sizeof(u64
); /* nr */
5297 if (data
->br_stack
) {
5298 size
+= data
->br_stack
->nr
5299 * sizeof(struct perf_branch_entry
);
5301 header
->size
+= size
;
5304 if (sample_type
& (PERF_SAMPLE_REGS_USER
| PERF_SAMPLE_STACK_USER
))
5305 perf_sample_regs_user(&data
->regs_user
, regs
,
5306 &data
->regs_user_copy
);
5308 if (sample_type
& PERF_SAMPLE_REGS_USER
) {
5309 /* regs dump ABI info */
5310 int size
= sizeof(u64
);
5312 if (data
->regs_user
.regs
) {
5313 u64 mask
= event
->attr
.sample_regs_user
;
5314 size
+= hweight64(mask
) * sizeof(u64
);
5317 header
->size
+= size
;
5320 if (sample_type
& PERF_SAMPLE_STACK_USER
) {
5322 * Either we need PERF_SAMPLE_STACK_USER bit to be allways
5323 * processed as the last one or have additional check added
5324 * in case new sample type is added, because we could eat
5325 * up the rest of the sample size.
5327 u16 stack_size
= event
->attr
.sample_stack_user
;
5328 u16 size
= sizeof(u64
);
5330 stack_size
= perf_sample_ustack_size(stack_size
, header
->size
,
5331 data
->regs_user
.regs
);
5334 * If there is something to dump, add space for the dump
5335 * itself and for the field that tells the dynamic size,
5336 * which is how many have been actually dumped.
5339 size
+= sizeof(u64
) + stack_size
;
5341 data
->stack_user_size
= stack_size
;
5342 header
->size
+= size
;
5345 if (sample_type
& PERF_SAMPLE_REGS_INTR
) {
5346 /* regs dump ABI info */
5347 int size
= sizeof(u64
);
5349 perf_sample_regs_intr(&data
->regs_intr
, regs
);
5351 if (data
->regs_intr
.regs
) {
5352 u64 mask
= event
->attr
.sample_regs_intr
;
5354 size
+= hweight64(mask
) * sizeof(u64
);
5357 header
->size
+= size
;
5361 static void perf_event_output(struct perf_event
*event
,
5362 struct perf_sample_data
*data
,
5363 struct pt_regs
*regs
)
5365 struct perf_output_handle handle
;
5366 struct perf_event_header header
;
5368 /* protect the callchain buffers */
5371 perf_prepare_sample(&header
, data
, event
, regs
);
5373 if (perf_output_begin(&handle
, event
, header
.size
))
5376 perf_output_sample(&handle
, &header
, data
, event
);
5378 perf_output_end(&handle
);
5388 struct perf_read_event
{
5389 struct perf_event_header header
;
5396 perf_event_read_event(struct perf_event
*event
,
5397 struct task_struct
*task
)
5399 struct perf_output_handle handle
;
5400 struct perf_sample_data sample
;
5401 struct perf_read_event read_event
= {
5403 .type
= PERF_RECORD_READ
,
5405 .size
= sizeof(read_event
) + event
->read_size
,
5407 .pid
= perf_event_pid(event
, task
),
5408 .tid
= perf_event_tid(event
, task
),
5412 perf_event_header__init_id(&read_event
.header
, &sample
, event
);
5413 ret
= perf_output_begin(&handle
, event
, read_event
.header
.size
);
5417 perf_output_put(&handle
, read_event
);
5418 perf_output_read(&handle
, event
);
5419 perf_event__output_id_sample(event
, &handle
, &sample
);
5421 perf_output_end(&handle
);
5424 typedef void (perf_event_aux_output_cb
)(struct perf_event
*event
, void *data
);
5427 perf_event_aux_ctx(struct perf_event_context
*ctx
,
5428 perf_event_aux_output_cb output
,
5431 struct perf_event
*event
;
5433 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
5434 if (event
->state
< PERF_EVENT_STATE_INACTIVE
)
5436 if (!event_filter_match(event
))
5438 output(event
, data
);
5443 perf_event_aux(perf_event_aux_output_cb output
, void *data
,
5444 struct perf_event_context
*task_ctx
)
5446 struct perf_cpu_context
*cpuctx
;
5447 struct perf_event_context
*ctx
;
5452 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
5453 cpuctx
= get_cpu_ptr(pmu
->pmu_cpu_context
);
5454 if (cpuctx
->unique_pmu
!= pmu
)
5456 perf_event_aux_ctx(&cpuctx
->ctx
, output
, data
);
5459 ctxn
= pmu
->task_ctx_nr
;
5462 ctx
= rcu_dereference(current
->perf_event_ctxp
[ctxn
]);
5464 perf_event_aux_ctx(ctx
, output
, data
);
5466 put_cpu_ptr(pmu
->pmu_cpu_context
);
5471 perf_event_aux_ctx(task_ctx
, output
, data
);
5478 * task tracking -- fork/exit
5480 * enabled by: attr.comm | attr.mmap | attr.mmap2 | attr.mmap_data | attr.task
5483 struct perf_task_event
{
5484 struct task_struct
*task
;
5485 struct perf_event_context
*task_ctx
;
5488 struct perf_event_header header
;
5498 static int perf_event_task_match(struct perf_event
*event
)
5500 return event
->attr
.comm
|| event
->attr
.mmap
||
5501 event
->attr
.mmap2
|| event
->attr
.mmap_data
||
5505 static void perf_event_task_output(struct perf_event
*event
,
5508 struct perf_task_event
*task_event
= data
;
5509 struct perf_output_handle handle
;
5510 struct perf_sample_data sample
;
5511 struct task_struct
*task
= task_event
->task
;
5512 int ret
, size
= task_event
->event_id
.header
.size
;
5514 if (!perf_event_task_match(event
))
5517 perf_event_header__init_id(&task_event
->event_id
.header
, &sample
, event
);
5519 ret
= perf_output_begin(&handle
, event
,
5520 task_event
->event_id
.header
.size
);
5524 task_event
->event_id
.pid
= perf_event_pid(event
, task
);
5525 task_event
->event_id
.ppid
= perf_event_pid(event
, current
);
5527 task_event
->event_id
.tid
= perf_event_tid(event
, task
);
5528 task_event
->event_id
.ptid
= perf_event_tid(event
, current
);
5530 task_event
->event_id
.time
= perf_event_clock(event
);
5532 perf_output_put(&handle
, task_event
->event_id
);
5534 perf_event__output_id_sample(event
, &handle
, &sample
);
5536 perf_output_end(&handle
);
5538 task_event
->event_id
.header
.size
= size
;
5541 static void perf_event_task(struct task_struct
*task
,
5542 struct perf_event_context
*task_ctx
,
5545 struct perf_task_event task_event
;
5547 if (!atomic_read(&nr_comm_events
) &&
5548 !atomic_read(&nr_mmap_events
) &&
5549 !atomic_read(&nr_task_events
))
5552 task_event
= (struct perf_task_event
){
5554 .task_ctx
= task_ctx
,
5557 .type
= new ? PERF_RECORD_FORK
: PERF_RECORD_EXIT
,
5559 .size
= sizeof(task_event
.event_id
),
5569 perf_event_aux(perf_event_task_output
,
5574 void perf_event_fork(struct task_struct
*task
)
5576 perf_event_task(task
, NULL
, 1);
5583 struct perf_comm_event
{
5584 struct task_struct
*task
;
5589 struct perf_event_header header
;
5596 static int perf_event_comm_match(struct perf_event
*event
)
5598 return event
->attr
.comm
;
5601 static void perf_event_comm_output(struct perf_event
*event
,
5604 struct perf_comm_event
*comm_event
= data
;
5605 struct perf_output_handle handle
;
5606 struct perf_sample_data sample
;
5607 int size
= comm_event
->event_id
.header
.size
;
5610 if (!perf_event_comm_match(event
))
5613 perf_event_header__init_id(&comm_event
->event_id
.header
, &sample
, event
);
5614 ret
= perf_output_begin(&handle
, event
,
5615 comm_event
->event_id
.header
.size
);
5620 comm_event
->event_id
.pid
= perf_event_pid(event
, comm_event
->task
);
5621 comm_event
->event_id
.tid
= perf_event_tid(event
, comm_event
->task
);
5623 perf_output_put(&handle
, comm_event
->event_id
);
5624 __output_copy(&handle
, comm_event
->comm
,
5625 comm_event
->comm_size
);
5627 perf_event__output_id_sample(event
, &handle
, &sample
);
5629 perf_output_end(&handle
);
5631 comm_event
->event_id
.header
.size
= size
;
5634 static void perf_event_comm_event(struct perf_comm_event
*comm_event
)
5636 char comm
[TASK_COMM_LEN
];
5639 memset(comm
, 0, sizeof(comm
));
5640 strlcpy(comm
, comm_event
->task
->comm
, sizeof(comm
));
5641 size
= ALIGN(strlen(comm
)+1, sizeof(u64
));
5643 comm_event
->comm
= comm
;
5644 comm_event
->comm_size
= size
;
5646 comm_event
->event_id
.header
.size
= sizeof(comm_event
->event_id
) + size
;
5648 perf_event_aux(perf_event_comm_output
,
5653 void perf_event_comm(struct task_struct
*task
, bool exec
)
5655 struct perf_comm_event comm_event
;
5657 if (!atomic_read(&nr_comm_events
))
5660 comm_event
= (struct perf_comm_event
){
5666 .type
= PERF_RECORD_COMM
,
5667 .misc
= exec
? PERF_RECORD_MISC_COMM_EXEC
: 0,
5675 perf_event_comm_event(&comm_event
);
5682 struct perf_mmap_event
{
5683 struct vm_area_struct
*vma
;
5685 const char *file_name
;
5693 struct perf_event_header header
;
5703 static int perf_event_mmap_match(struct perf_event
*event
,
5706 struct perf_mmap_event
*mmap_event
= data
;
5707 struct vm_area_struct
*vma
= mmap_event
->vma
;
5708 int executable
= vma
->vm_flags
& VM_EXEC
;
5710 return (!executable
&& event
->attr
.mmap_data
) ||
5711 (executable
&& (event
->attr
.mmap
|| event
->attr
.mmap2
));
5714 static void perf_event_mmap_output(struct perf_event
*event
,
5717 struct perf_mmap_event
*mmap_event
= data
;
5718 struct perf_output_handle handle
;
5719 struct perf_sample_data sample
;
5720 int size
= mmap_event
->event_id
.header
.size
;
5723 if (!perf_event_mmap_match(event
, data
))
5726 if (event
->attr
.mmap2
) {
5727 mmap_event
->event_id
.header
.type
= PERF_RECORD_MMAP2
;
5728 mmap_event
->event_id
.header
.size
+= sizeof(mmap_event
->maj
);
5729 mmap_event
->event_id
.header
.size
+= sizeof(mmap_event
->min
);
5730 mmap_event
->event_id
.header
.size
+= sizeof(mmap_event
->ino
);
5731 mmap_event
->event_id
.header
.size
+= sizeof(mmap_event
->ino_generation
);
5732 mmap_event
->event_id
.header
.size
+= sizeof(mmap_event
->prot
);
5733 mmap_event
->event_id
.header
.size
+= sizeof(mmap_event
->flags
);
5736 perf_event_header__init_id(&mmap_event
->event_id
.header
, &sample
, event
);
5737 ret
= perf_output_begin(&handle
, event
,
5738 mmap_event
->event_id
.header
.size
);
5742 mmap_event
->event_id
.pid
= perf_event_pid(event
, current
);
5743 mmap_event
->event_id
.tid
= perf_event_tid(event
, current
);
5745 perf_output_put(&handle
, mmap_event
->event_id
);
5747 if (event
->attr
.mmap2
) {
5748 perf_output_put(&handle
, mmap_event
->maj
);
5749 perf_output_put(&handle
, mmap_event
->min
);
5750 perf_output_put(&handle
, mmap_event
->ino
);
5751 perf_output_put(&handle
, mmap_event
->ino_generation
);
5752 perf_output_put(&handle
, mmap_event
->prot
);
5753 perf_output_put(&handle
, mmap_event
->flags
);
5756 __output_copy(&handle
, mmap_event
->file_name
,
5757 mmap_event
->file_size
);
5759 perf_event__output_id_sample(event
, &handle
, &sample
);
5761 perf_output_end(&handle
);
5763 mmap_event
->event_id
.header
.size
= size
;
5766 static void perf_event_mmap_event(struct perf_mmap_event
*mmap_event
)
5768 struct vm_area_struct
*vma
= mmap_event
->vma
;
5769 struct file
*file
= vma
->vm_file
;
5770 int maj
= 0, min
= 0;
5771 u64 ino
= 0, gen
= 0;
5772 u32 prot
= 0, flags
= 0;
5779 struct inode
*inode
;
5782 buf
= kmalloc(PATH_MAX
, GFP_KERNEL
);
5788 * d_path() works from the end of the rb backwards, so we
5789 * need to add enough zero bytes after the string to handle
5790 * the 64bit alignment we do later.
5792 name
= d_path(&file
->f_path
, buf
, PATH_MAX
- sizeof(u64
));
5797 inode
= file_inode(vma
->vm_file
);
5798 dev
= inode
->i_sb
->s_dev
;
5800 gen
= inode
->i_generation
;
5804 if (vma
->vm_flags
& VM_READ
)
5806 if (vma
->vm_flags
& VM_WRITE
)
5808 if (vma
->vm_flags
& VM_EXEC
)
5811 if (vma
->vm_flags
& VM_MAYSHARE
)
5814 flags
= MAP_PRIVATE
;
5816 if (vma
->vm_flags
& VM_DENYWRITE
)
5817 flags
|= MAP_DENYWRITE
;
5818 if (vma
->vm_flags
& VM_MAYEXEC
)
5819 flags
|= MAP_EXECUTABLE
;
5820 if (vma
->vm_flags
& VM_LOCKED
)
5821 flags
|= MAP_LOCKED
;
5822 if (vma
->vm_flags
& VM_HUGETLB
)
5823 flags
|= MAP_HUGETLB
;
5827 if (vma
->vm_ops
&& vma
->vm_ops
->name
) {
5828 name
= (char *) vma
->vm_ops
->name(vma
);
5833 name
= (char *)arch_vma_name(vma
);
5837 if (vma
->vm_start
<= vma
->vm_mm
->start_brk
&&
5838 vma
->vm_end
>= vma
->vm_mm
->brk
) {
5842 if (vma
->vm_start
<= vma
->vm_mm
->start_stack
&&
5843 vma
->vm_end
>= vma
->vm_mm
->start_stack
) {
5853 strlcpy(tmp
, name
, sizeof(tmp
));
5857 * Since our buffer works in 8 byte units we need to align our string
5858 * size to a multiple of 8. However, we must guarantee the tail end is
5859 * zero'd out to avoid leaking random bits to userspace.
5861 size
= strlen(name
)+1;
5862 while (!IS_ALIGNED(size
, sizeof(u64
)))
5863 name
[size
++] = '\0';
5865 mmap_event
->file_name
= name
;
5866 mmap_event
->file_size
= size
;
5867 mmap_event
->maj
= maj
;
5868 mmap_event
->min
= min
;
5869 mmap_event
->ino
= ino
;
5870 mmap_event
->ino_generation
= gen
;
5871 mmap_event
->prot
= prot
;
5872 mmap_event
->flags
= flags
;
5874 if (!(vma
->vm_flags
& VM_EXEC
))
5875 mmap_event
->event_id
.header
.misc
|= PERF_RECORD_MISC_MMAP_DATA
;
5877 mmap_event
->event_id
.header
.size
= sizeof(mmap_event
->event_id
) + size
;
5879 perf_event_aux(perf_event_mmap_output
,
5886 void perf_event_mmap(struct vm_area_struct
*vma
)
5888 struct perf_mmap_event mmap_event
;
5890 if (!atomic_read(&nr_mmap_events
))
5893 mmap_event
= (struct perf_mmap_event
){
5899 .type
= PERF_RECORD_MMAP
,
5900 .misc
= PERF_RECORD_MISC_USER
,
5905 .start
= vma
->vm_start
,
5906 .len
= vma
->vm_end
- vma
->vm_start
,
5907 .pgoff
= (u64
)vma
->vm_pgoff
<< PAGE_SHIFT
,
5909 /* .maj (attr_mmap2 only) */
5910 /* .min (attr_mmap2 only) */
5911 /* .ino (attr_mmap2 only) */
5912 /* .ino_generation (attr_mmap2 only) */
5913 /* .prot (attr_mmap2 only) */
5914 /* .flags (attr_mmap2 only) */
5917 perf_event_mmap_event(&mmap_event
);
5920 void perf_event_aux_event(struct perf_event
*event
, unsigned long head
,
5921 unsigned long size
, u64 flags
)
5923 struct perf_output_handle handle
;
5924 struct perf_sample_data sample
;
5925 struct perf_aux_event
{
5926 struct perf_event_header header
;
5932 .type
= PERF_RECORD_AUX
,
5934 .size
= sizeof(rec
),
5942 perf_event_header__init_id(&rec
.header
, &sample
, event
);
5943 ret
= perf_output_begin(&handle
, event
, rec
.header
.size
);
5948 perf_output_put(&handle
, rec
);
5949 perf_event__output_id_sample(event
, &handle
, &sample
);
5951 perf_output_end(&handle
);
5955 * IRQ throttle logging
5958 static void perf_log_throttle(struct perf_event
*event
, int enable
)
5960 struct perf_output_handle handle
;
5961 struct perf_sample_data sample
;
5965 struct perf_event_header header
;
5969 } throttle_event
= {
5971 .type
= PERF_RECORD_THROTTLE
,
5973 .size
= sizeof(throttle_event
),
5975 .time
= perf_event_clock(event
),
5976 .id
= primary_event_id(event
),
5977 .stream_id
= event
->id
,
5981 throttle_event
.header
.type
= PERF_RECORD_UNTHROTTLE
;
5983 perf_event_header__init_id(&throttle_event
.header
, &sample
, event
);
5985 ret
= perf_output_begin(&handle
, event
,
5986 throttle_event
.header
.size
);
5990 perf_output_put(&handle
, throttle_event
);
5991 perf_event__output_id_sample(event
, &handle
, &sample
);
5992 perf_output_end(&handle
);
5995 static void perf_log_itrace_start(struct perf_event
*event
)
5997 struct perf_output_handle handle
;
5998 struct perf_sample_data sample
;
5999 struct perf_aux_event
{
6000 struct perf_event_header header
;
6007 event
= event
->parent
;
6009 if (!(event
->pmu
->capabilities
& PERF_PMU_CAP_ITRACE
) ||
6010 event
->hw
.itrace_started
)
6013 event
->hw
.itrace_started
= 1;
6015 rec
.header
.type
= PERF_RECORD_ITRACE_START
;
6016 rec
.header
.misc
= 0;
6017 rec
.header
.size
= sizeof(rec
);
6018 rec
.pid
= perf_event_pid(event
, current
);
6019 rec
.tid
= perf_event_tid(event
, current
);
6021 perf_event_header__init_id(&rec
.header
, &sample
, event
);
6022 ret
= perf_output_begin(&handle
, event
, rec
.header
.size
);
6027 perf_output_put(&handle
, rec
);
6028 perf_event__output_id_sample(event
, &handle
, &sample
);
6030 perf_output_end(&handle
);
6034 * Generic event overflow handling, sampling.
6037 static int __perf_event_overflow(struct perf_event
*event
,
6038 int throttle
, struct perf_sample_data
*data
,
6039 struct pt_regs
*regs
)
6041 int events
= atomic_read(&event
->event_limit
);
6042 struct hw_perf_event
*hwc
= &event
->hw
;
6047 * Non-sampling counters might still use the PMI to fold short
6048 * hardware counters, ignore those.
6050 if (unlikely(!is_sampling_event(event
)))
6053 seq
= __this_cpu_read(perf_throttled_seq
);
6054 if (seq
!= hwc
->interrupts_seq
) {
6055 hwc
->interrupts_seq
= seq
;
6056 hwc
->interrupts
= 1;
6059 if (unlikely(throttle
6060 && hwc
->interrupts
>= max_samples_per_tick
)) {
6061 __this_cpu_inc(perf_throttled_count
);
6062 hwc
->interrupts
= MAX_INTERRUPTS
;
6063 perf_log_throttle(event
, 0);
6064 tick_nohz_full_kick();
6069 if (event
->attr
.freq
) {
6070 u64 now
= perf_clock();
6071 s64 delta
= now
- hwc
->freq_time_stamp
;
6073 hwc
->freq_time_stamp
= now
;
6075 if (delta
> 0 && delta
< 2*TICK_NSEC
)
6076 perf_adjust_period(event
, delta
, hwc
->last_period
, true);
6080 * XXX event_limit might not quite work as expected on inherited
6084 event
->pending_kill
= POLL_IN
;
6085 if (events
&& atomic_dec_and_test(&event
->event_limit
)) {
6087 event
->pending_kill
= POLL_HUP
;
6088 event
->pending_disable
= 1;
6089 irq_work_queue(&event
->pending
);
6092 if (event
->overflow_handler
)
6093 event
->overflow_handler(event
, data
, regs
);
6095 perf_event_output(event
, data
, regs
);
6097 if (event
->fasync
&& event
->pending_kill
) {
6098 event
->pending_wakeup
= 1;
6099 irq_work_queue(&event
->pending
);
6105 int perf_event_overflow(struct perf_event
*event
,
6106 struct perf_sample_data
*data
,
6107 struct pt_regs
*regs
)
6109 return __perf_event_overflow(event
, 1, data
, regs
);
6113 * Generic software event infrastructure
6116 struct swevent_htable
{
6117 struct swevent_hlist
*swevent_hlist
;
6118 struct mutex hlist_mutex
;
6121 /* Recursion avoidance in each contexts */
6122 int recursion
[PERF_NR_CONTEXTS
];
6124 /* Keeps track of cpu being initialized/exited */
6128 static DEFINE_PER_CPU(struct swevent_htable
, swevent_htable
);
6131 * We directly increment event->count and keep a second value in
6132 * event->hw.period_left to count intervals. This period event
6133 * is kept in the range [-sample_period, 0] so that we can use the
6137 u64
perf_swevent_set_period(struct perf_event
*event
)
6139 struct hw_perf_event
*hwc
= &event
->hw
;
6140 u64 period
= hwc
->last_period
;
6144 hwc
->last_period
= hwc
->sample_period
;
6147 old
= val
= local64_read(&hwc
->period_left
);
6151 nr
= div64_u64(period
+ val
, period
);
6152 offset
= nr
* period
;
6154 if (local64_cmpxchg(&hwc
->period_left
, old
, val
) != old
)
6160 static void perf_swevent_overflow(struct perf_event
*event
, u64 overflow
,
6161 struct perf_sample_data
*data
,
6162 struct pt_regs
*regs
)
6164 struct hw_perf_event
*hwc
= &event
->hw
;
6168 overflow
= perf_swevent_set_period(event
);
6170 if (hwc
->interrupts
== MAX_INTERRUPTS
)
6173 for (; overflow
; overflow
--) {
6174 if (__perf_event_overflow(event
, throttle
,
6177 * We inhibit the overflow from happening when
6178 * hwc->interrupts == MAX_INTERRUPTS.
6186 static void perf_swevent_event(struct perf_event
*event
, u64 nr
,
6187 struct perf_sample_data
*data
,
6188 struct pt_regs
*regs
)
6190 struct hw_perf_event
*hwc
= &event
->hw
;
6192 local64_add(nr
, &event
->count
);
6197 if (!is_sampling_event(event
))
6200 if ((event
->attr
.sample_type
& PERF_SAMPLE_PERIOD
) && !event
->attr
.freq
) {
6202 return perf_swevent_overflow(event
, 1, data
, regs
);
6204 data
->period
= event
->hw
.last_period
;
6206 if (nr
== 1 && hwc
->sample_period
== 1 && !event
->attr
.freq
)
6207 return perf_swevent_overflow(event
, 1, data
, regs
);
6209 if (local64_add_negative(nr
, &hwc
->period_left
))
6212 perf_swevent_overflow(event
, 0, data
, regs
);
6215 static int perf_exclude_event(struct perf_event
*event
,
6216 struct pt_regs
*regs
)
6218 if (event
->hw
.state
& PERF_HES_STOPPED
)
6222 if (event
->attr
.exclude_user
&& user_mode(regs
))
6225 if (event
->attr
.exclude_kernel
&& !user_mode(regs
))
6232 static int perf_swevent_match(struct perf_event
*event
,
6233 enum perf_type_id type
,
6235 struct perf_sample_data
*data
,
6236 struct pt_regs
*regs
)
6238 if (event
->attr
.type
!= type
)
6241 if (event
->attr
.config
!= event_id
)
6244 if (perf_exclude_event(event
, regs
))
6250 static inline u64
swevent_hash(u64 type
, u32 event_id
)
6252 u64 val
= event_id
| (type
<< 32);
6254 return hash_64(val
, SWEVENT_HLIST_BITS
);
6257 static inline struct hlist_head
*
6258 __find_swevent_head(struct swevent_hlist
*hlist
, u64 type
, u32 event_id
)
6260 u64 hash
= swevent_hash(type
, event_id
);
6262 return &hlist
->heads
[hash
];
6265 /* For the read side: events when they trigger */
6266 static inline struct hlist_head
*
6267 find_swevent_head_rcu(struct swevent_htable
*swhash
, u64 type
, u32 event_id
)
6269 struct swevent_hlist
*hlist
;
6271 hlist
= rcu_dereference(swhash
->swevent_hlist
);
6275 return __find_swevent_head(hlist
, type
, event_id
);
6278 /* For the event head insertion and removal in the hlist */
6279 static inline struct hlist_head
*
6280 find_swevent_head(struct swevent_htable
*swhash
, struct perf_event
*event
)
6282 struct swevent_hlist
*hlist
;
6283 u32 event_id
= event
->attr
.config
;
6284 u64 type
= event
->attr
.type
;
6287 * Event scheduling is always serialized against hlist allocation
6288 * and release. Which makes the protected version suitable here.
6289 * The context lock guarantees that.
6291 hlist
= rcu_dereference_protected(swhash
->swevent_hlist
,
6292 lockdep_is_held(&event
->ctx
->lock
));
6296 return __find_swevent_head(hlist
, type
, event_id
);
6299 static void do_perf_sw_event(enum perf_type_id type
, u32 event_id
,
6301 struct perf_sample_data
*data
,
6302 struct pt_regs
*regs
)
6304 struct swevent_htable
*swhash
= this_cpu_ptr(&swevent_htable
);
6305 struct perf_event
*event
;
6306 struct hlist_head
*head
;
6309 head
= find_swevent_head_rcu(swhash
, type
, event_id
);
6313 hlist_for_each_entry_rcu(event
, head
, hlist_entry
) {
6314 if (perf_swevent_match(event
, type
, event_id
, data
, regs
))
6315 perf_swevent_event(event
, nr
, data
, regs
);
6321 DEFINE_PER_CPU(struct pt_regs
, __perf_regs
[4]);
6323 int perf_swevent_get_recursion_context(void)
6325 struct swevent_htable
*swhash
= this_cpu_ptr(&swevent_htable
);
6327 return get_recursion_context(swhash
->recursion
);
6329 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context
);
6331 inline void perf_swevent_put_recursion_context(int rctx
)
6333 struct swevent_htable
*swhash
= this_cpu_ptr(&swevent_htable
);
6335 put_recursion_context(swhash
->recursion
, rctx
);
6338 void ___perf_sw_event(u32 event_id
, u64 nr
, struct pt_regs
*regs
, u64 addr
)
6340 struct perf_sample_data data
;
6342 if (WARN_ON_ONCE(!regs
))
6345 perf_sample_data_init(&data
, addr
, 0);
6346 do_perf_sw_event(PERF_TYPE_SOFTWARE
, event_id
, nr
, &data
, regs
);
6349 void __perf_sw_event(u32 event_id
, u64 nr
, struct pt_regs
*regs
, u64 addr
)
6353 preempt_disable_notrace();
6354 rctx
= perf_swevent_get_recursion_context();
6355 if (unlikely(rctx
< 0))
6358 ___perf_sw_event(event_id
, nr
, regs
, addr
);
6360 perf_swevent_put_recursion_context(rctx
);
6362 preempt_enable_notrace();
6365 static void perf_swevent_read(struct perf_event
*event
)
6369 static int perf_swevent_add(struct perf_event
*event
, int flags
)
6371 struct swevent_htable
*swhash
= this_cpu_ptr(&swevent_htable
);
6372 struct hw_perf_event
*hwc
= &event
->hw
;
6373 struct hlist_head
*head
;
6375 if (is_sampling_event(event
)) {
6376 hwc
->last_period
= hwc
->sample_period
;
6377 perf_swevent_set_period(event
);
6380 hwc
->state
= !(flags
& PERF_EF_START
);
6382 head
= find_swevent_head(swhash
, event
);
6385 * We can race with cpu hotplug code. Do not
6386 * WARN if the cpu just got unplugged.
6388 WARN_ON_ONCE(swhash
->online
);
6392 hlist_add_head_rcu(&event
->hlist_entry
, head
);
6393 perf_event_update_userpage(event
);
6398 static void perf_swevent_del(struct perf_event
*event
, int flags
)
6400 hlist_del_rcu(&event
->hlist_entry
);
6403 static void perf_swevent_start(struct perf_event
*event
, int flags
)
6405 event
->hw
.state
= 0;
6408 static void perf_swevent_stop(struct perf_event
*event
, int flags
)
6410 event
->hw
.state
= PERF_HES_STOPPED
;
6413 /* Deref the hlist from the update side */
6414 static inline struct swevent_hlist
*
6415 swevent_hlist_deref(struct swevent_htable
*swhash
)
6417 return rcu_dereference_protected(swhash
->swevent_hlist
,
6418 lockdep_is_held(&swhash
->hlist_mutex
));
6421 static void swevent_hlist_release(struct swevent_htable
*swhash
)
6423 struct swevent_hlist
*hlist
= swevent_hlist_deref(swhash
);
6428 RCU_INIT_POINTER(swhash
->swevent_hlist
, NULL
);
6429 kfree_rcu(hlist
, rcu_head
);
6432 static void swevent_hlist_put_cpu(struct perf_event
*event
, int cpu
)
6434 struct swevent_htable
*swhash
= &per_cpu(swevent_htable
, cpu
);
6436 mutex_lock(&swhash
->hlist_mutex
);
6438 if (!--swhash
->hlist_refcount
)
6439 swevent_hlist_release(swhash
);
6441 mutex_unlock(&swhash
->hlist_mutex
);
6444 static void swevent_hlist_put(struct perf_event
*event
)
6448 for_each_possible_cpu(cpu
)
6449 swevent_hlist_put_cpu(event
, cpu
);
6452 static int swevent_hlist_get_cpu(struct perf_event
*event
, int cpu
)
6454 struct swevent_htable
*swhash
= &per_cpu(swevent_htable
, cpu
);
6457 mutex_lock(&swhash
->hlist_mutex
);
6459 if (!swevent_hlist_deref(swhash
) && cpu_online(cpu
)) {
6460 struct swevent_hlist
*hlist
;
6462 hlist
= kzalloc(sizeof(*hlist
), GFP_KERNEL
);
6467 rcu_assign_pointer(swhash
->swevent_hlist
, hlist
);
6469 swhash
->hlist_refcount
++;
6471 mutex_unlock(&swhash
->hlist_mutex
);
6476 static int swevent_hlist_get(struct perf_event
*event
)
6479 int cpu
, failed_cpu
;
6482 for_each_possible_cpu(cpu
) {
6483 err
= swevent_hlist_get_cpu(event
, cpu
);
6493 for_each_possible_cpu(cpu
) {
6494 if (cpu
== failed_cpu
)
6496 swevent_hlist_put_cpu(event
, cpu
);
6503 struct static_key perf_swevent_enabled
[PERF_COUNT_SW_MAX
];
6505 static void sw_perf_event_destroy(struct perf_event
*event
)
6507 u64 event_id
= event
->attr
.config
;
6509 WARN_ON(event
->parent
);
6511 static_key_slow_dec(&perf_swevent_enabled
[event_id
]);
6512 swevent_hlist_put(event
);
6515 static int perf_swevent_init(struct perf_event
*event
)
6517 u64 event_id
= event
->attr
.config
;
6519 if (event
->attr
.type
!= PERF_TYPE_SOFTWARE
)
6523 * no branch sampling for software events
6525 if (has_branch_stack(event
))
6529 case PERF_COUNT_SW_CPU_CLOCK
:
6530 case PERF_COUNT_SW_TASK_CLOCK
:
6537 if (event_id
>= PERF_COUNT_SW_MAX
)
6540 if (!event
->parent
) {
6543 err
= swevent_hlist_get(event
);
6547 static_key_slow_inc(&perf_swevent_enabled
[event_id
]);
6548 event
->destroy
= sw_perf_event_destroy
;
6554 static struct pmu perf_swevent
= {
6555 .task_ctx_nr
= perf_sw_context
,
6557 .capabilities
= PERF_PMU_CAP_NO_NMI
,
6559 .event_init
= perf_swevent_init
,
6560 .add
= perf_swevent_add
,
6561 .del
= perf_swevent_del
,
6562 .start
= perf_swevent_start
,
6563 .stop
= perf_swevent_stop
,
6564 .read
= perf_swevent_read
,
6567 #ifdef CONFIG_EVENT_TRACING
6569 static int perf_tp_filter_match(struct perf_event
*event
,
6570 struct perf_sample_data
*data
)
6572 void *record
= data
->raw
->data
;
6574 if (likely(!event
->filter
) || filter_match_preds(event
->filter
, record
))
6579 static int perf_tp_event_match(struct perf_event
*event
,
6580 struct perf_sample_data
*data
,
6581 struct pt_regs
*regs
)
6583 if (event
->hw
.state
& PERF_HES_STOPPED
)
6586 * All tracepoints are from kernel-space.
6588 if (event
->attr
.exclude_kernel
)
6591 if (!perf_tp_filter_match(event
, data
))
6597 void perf_tp_event(u64 addr
, u64 count
, void *record
, int entry_size
,
6598 struct pt_regs
*regs
, struct hlist_head
*head
, int rctx
,
6599 struct task_struct
*task
)
6601 struct perf_sample_data data
;
6602 struct perf_event
*event
;
6604 struct perf_raw_record raw
= {
6609 perf_sample_data_init(&data
, addr
, 0);
6612 hlist_for_each_entry_rcu(event
, head
, hlist_entry
) {
6613 if (perf_tp_event_match(event
, &data
, regs
))
6614 perf_swevent_event(event
, count
, &data
, regs
);
6618 * If we got specified a target task, also iterate its context and
6619 * deliver this event there too.
6621 if (task
&& task
!= current
) {
6622 struct perf_event_context
*ctx
;
6623 struct trace_entry
*entry
= record
;
6626 ctx
= rcu_dereference(task
->perf_event_ctxp
[perf_sw_context
]);
6630 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
6631 if (event
->attr
.type
!= PERF_TYPE_TRACEPOINT
)
6633 if (event
->attr
.config
!= entry
->type
)
6635 if (perf_tp_event_match(event
, &data
, regs
))
6636 perf_swevent_event(event
, count
, &data
, regs
);
6642 perf_swevent_put_recursion_context(rctx
);
6644 EXPORT_SYMBOL_GPL(perf_tp_event
);
6646 static void tp_perf_event_destroy(struct perf_event
*event
)
6648 perf_trace_destroy(event
);
6651 static int perf_tp_event_init(struct perf_event
*event
)
6655 if (event
->attr
.type
!= PERF_TYPE_TRACEPOINT
)
6659 * no branch sampling for tracepoint events
6661 if (has_branch_stack(event
))
6664 err
= perf_trace_init(event
);
6668 event
->destroy
= tp_perf_event_destroy
;
6673 static struct pmu perf_tracepoint
= {
6674 .task_ctx_nr
= perf_sw_context
,
6676 .event_init
= perf_tp_event_init
,
6677 .add
= perf_trace_add
,
6678 .del
= perf_trace_del
,
6679 .start
= perf_swevent_start
,
6680 .stop
= perf_swevent_stop
,
6681 .read
= perf_swevent_read
,
6684 static inline void perf_tp_register(void)
6686 perf_pmu_register(&perf_tracepoint
, "tracepoint", PERF_TYPE_TRACEPOINT
);
6689 static int perf_event_set_filter(struct perf_event
*event
, void __user
*arg
)
6694 if (event
->attr
.type
!= PERF_TYPE_TRACEPOINT
)
6697 filter_str
= strndup_user(arg
, PAGE_SIZE
);
6698 if (IS_ERR(filter_str
))
6699 return PTR_ERR(filter_str
);
6701 ret
= ftrace_profile_set_filter(event
, event
->attr
.config
, filter_str
);
6707 static void perf_event_free_filter(struct perf_event
*event
)
6709 ftrace_profile_free_filter(event
);
6712 static int perf_event_set_bpf_prog(struct perf_event
*event
, u32 prog_fd
)
6714 struct bpf_prog
*prog
;
6716 if (event
->attr
.type
!= PERF_TYPE_TRACEPOINT
)
6719 if (event
->tp_event
->prog
)
6722 if (!(event
->tp_event
->flags
& TRACE_EVENT_FL_KPROBE
))
6723 /* bpf programs can only be attached to kprobes */
6726 prog
= bpf_prog_get(prog_fd
);
6728 return PTR_ERR(prog
);
6730 if (prog
->type
!= BPF_PROG_TYPE_KPROBE
) {
6731 /* valid fd, but invalid bpf program type */
6736 event
->tp_event
->prog
= prog
;
6741 static void perf_event_free_bpf_prog(struct perf_event
*event
)
6743 struct bpf_prog
*prog
;
6745 if (!event
->tp_event
)
6748 prog
= event
->tp_event
->prog
;
6750 event
->tp_event
->prog
= NULL
;
6757 static inline void perf_tp_register(void)
6761 static int perf_event_set_filter(struct perf_event
*event
, void __user
*arg
)
6766 static void perf_event_free_filter(struct perf_event
*event
)
6770 static int perf_event_set_bpf_prog(struct perf_event
*event
, u32 prog_fd
)
6775 static void perf_event_free_bpf_prog(struct perf_event
*event
)
6778 #endif /* CONFIG_EVENT_TRACING */
6780 #ifdef CONFIG_HAVE_HW_BREAKPOINT
6781 void perf_bp_event(struct perf_event
*bp
, void *data
)
6783 struct perf_sample_data sample
;
6784 struct pt_regs
*regs
= data
;
6786 perf_sample_data_init(&sample
, bp
->attr
.bp_addr
, 0);
6788 if (!bp
->hw
.state
&& !perf_exclude_event(bp
, regs
))
6789 perf_swevent_event(bp
, 1, &sample
, regs
);
6794 * hrtimer based swevent callback
6797 static enum hrtimer_restart
perf_swevent_hrtimer(struct hrtimer
*hrtimer
)
6799 enum hrtimer_restart ret
= HRTIMER_RESTART
;
6800 struct perf_sample_data data
;
6801 struct pt_regs
*regs
;
6802 struct perf_event
*event
;
6805 event
= container_of(hrtimer
, struct perf_event
, hw
.hrtimer
);
6807 if (event
->state
!= PERF_EVENT_STATE_ACTIVE
)
6808 return HRTIMER_NORESTART
;
6810 event
->pmu
->read(event
);
6812 perf_sample_data_init(&data
, 0, event
->hw
.last_period
);
6813 regs
= get_irq_regs();
6815 if (regs
&& !perf_exclude_event(event
, regs
)) {
6816 if (!(event
->attr
.exclude_idle
&& is_idle_task(current
)))
6817 if (__perf_event_overflow(event
, 1, &data
, regs
))
6818 ret
= HRTIMER_NORESTART
;
6821 period
= max_t(u64
, 10000, event
->hw
.sample_period
);
6822 hrtimer_forward_now(hrtimer
, ns_to_ktime(period
));
6827 static void perf_swevent_start_hrtimer(struct perf_event
*event
)
6829 struct hw_perf_event
*hwc
= &event
->hw
;
6832 if (!is_sampling_event(event
))
6835 period
= local64_read(&hwc
->period_left
);
6840 local64_set(&hwc
->period_left
, 0);
6842 period
= max_t(u64
, 10000, hwc
->sample_period
);
6844 hrtimer_start(&hwc
->hrtimer
, ns_to_ktime(period
),
6845 HRTIMER_MODE_REL_PINNED
);
6848 static void perf_swevent_cancel_hrtimer(struct perf_event
*event
)
6850 struct hw_perf_event
*hwc
= &event
->hw
;
6852 if (is_sampling_event(event
)) {
6853 ktime_t remaining
= hrtimer_get_remaining(&hwc
->hrtimer
);
6854 local64_set(&hwc
->period_left
, ktime_to_ns(remaining
));
6856 hrtimer_cancel(&hwc
->hrtimer
);
6860 static void perf_swevent_init_hrtimer(struct perf_event
*event
)
6862 struct hw_perf_event
*hwc
= &event
->hw
;
6864 if (!is_sampling_event(event
))
6867 hrtimer_init(&hwc
->hrtimer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
6868 hwc
->hrtimer
.function
= perf_swevent_hrtimer
;
6871 * Since hrtimers have a fixed rate, we can do a static freq->period
6872 * mapping and avoid the whole period adjust feedback stuff.
6874 if (event
->attr
.freq
) {
6875 long freq
= event
->attr
.sample_freq
;
6877 event
->attr
.sample_period
= NSEC_PER_SEC
/ freq
;
6878 hwc
->sample_period
= event
->attr
.sample_period
;
6879 local64_set(&hwc
->period_left
, hwc
->sample_period
);
6880 hwc
->last_period
= hwc
->sample_period
;
6881 event
->attr
.freq
= 0;
6886 * Software event: cpu wall time clock
6889 static void cpu_clock_event_update(struct perf_event
*event
)
6894 now
= local_clock();
6895 prev
= local64_xchg(&event
->hw
.prev_count
, now
);
6896 local64_add(now
- prev
, &event
->count
);
6899 static void cpu_clock_event_start(struct perf_event
*event
, int flags
)
6901 local64_set(&event
->hw
.prev_count
, local_clock());
6902 perf_swevent_start_hrtimer(event
);
6905 static void cpu_clock_event_stop(struct perf_event
*event
, int flags
)
6907 perf_swevent_cancel_hrtimer(event
);
6908 cpu_clock_event_update(event
);
6911 static int cpu_clock_event_add(struct perf_event
*event
, int flags
)
6913 if (flags
& PERF_EF_START
)
6914 cpu_clock_event_start(event
, flags
);
6915 perf_event_update_userpage(event
);
6920 static void cpu_clock_event_del(struct perf_event
*event
, int flags
)
6922 cpu_clock_event_stop(event
, flags
);
6925 static void cpu_clock_event_read(struct perf_event
*event
)
6927 cpu_clock_event_update(event
);
6930 static int cpu_clock_event_init(struct perf_event
*event
)
6932 if (event
->attr
.type
!= PERF_TYPE_SOFTWARE
)
6935 if (event
->attr
.config
!= PERF_COUNT_SW_CPU_CLOCK
)
6939 * no branch sampling for software events
6941 if (has_branch_stack(event
))
6944 perf_swevent_init_hrtimer(event
);
6949 static struct pmu perf_cpu_clock
= {
6950 .task_ctx_nr
= perf_sw_context
,
6952 .capabilities
= PERF_PMU_CAP_NO_NMI
,
6954 .event_init
= cpu_clock_event_init
,
6955 .add
= cpu_clock_event_add
,
6956 .del
= cpu_clock_event_del
,
6957 .start
= cpu_clock_event_start
,
6958 .stop
= cpu_clock_event_stop
,
6959 .read
= cpu_clock_event_read
,
6963 * Software event: task time clock
6966 static void task_clock_event_update(struct perf_event
*event
, u64 now
)
6971 prev
= local64_xchg(&event
->hw
.prev_count
, now
);
6973 local64_add(delta
, &event
->count
);
6976 static void task_clock_event_start(struct perf_event
*event
, int flags
)
6978 local64_set(&event
->hw
.prev_count
, event
->ctx
->time
);
6979 perf_swevent_start_hrtimer(event
);
6982 static void task_clock_event_stop(struct perf_event
*event
, int flags
)
6984 perf_swevent_cancel_hrtimer(event
);
6985 task_clock_event_update(event
, event
->ctx
->time
);
6988 static int task_clock_event_add(struct perf_event
*event
, int flags
)
6990 if (flags
& PERF_EF_START
)
6991 task_clock_event_start(event
, flags
);
6992 perf_event_update_userpage(event
);
6997 static void task_clock_event_del(struct perf_event
*event
, int flags
)
6999 task_clock_event_stop(event
, PERF_EF_UPDATE
);
7002 static void task_clock_event_read(struct perf_event
*event
)
7004 u64 now
= perf_clock();
7005 u64 delta
= now
- event
->ctx
->timestamp
;
7006 u64 time
= event
->ctx
->time
+ delta
;
7008 task_clock_event_update(event
, time
);
7011 static int task_clock_event_init(struct perf_event
*event
)
7013 if (event
->attr
.type
!= PERF_TYPE_SOFTWARE
)
7016 if (event
->attr
.config
!= PERF_COUNT_SW_TASK_CLOCK
)
7020 * no branch sampling for software events
7022 if (has_branch_stack(event
))
7025 perf_swevent_init_hrtimer(event
);
7030 static struct pmu perf_task_clock
= {
7031 .task_ctx_nr
= perf_sw_context
,
7033 .capabilities
= PERF_PMU_CAP_NO_NMI
,
7035 .event_init
= task_clock_event_init
,
7036 .add
= task_clock_event_add
,
7037 .del
= task_clock_event_del
,
7038 .start
= task_clock_event_start
,
7039 .stop
= task_clock_event_stop
,
7040 .read
= task_clock_event_read
,
7043 static void perf_pmu_nop_void(struct pmu
*pmu
)
7047 static int perf_pmu_nop_int(struct pmu
*pmu
)
7052 static void perf_pmu_start_txn(struct pmu
*pmu
)
7054 perf_pmu_disable(pmu
);
7057 static int perf_pmu_commit_txn(struct pmu
*pmu
)
7059 perf_pmu_enable(pmu
);
7063 static void perf_pmu_cancel_txn(struct pmu
*pmu
)
7065 perf_pmu_enable(pmu
);
7068 static int perf_event_idx_default(struct perf_event
*event
)
7074 * Ensures all contexts with the same task_ctx_nr have the same
7075 * pmu_cpu_context too.
7077 static struct perf_cpu_context __percpu
*find_pmu_context(int ctxn
)
7084 list_for_each_entry(pmu
, &pmus
, entry
) {
7085 if (pmu
->task_ctx_nr
== ctxn
)
7086 return pmu
->pmu_cpu_context
;
7092 static void update_pmu_context(struct pmu
*pmu
, struct pmu
*old_pmu
)
7096 for_each_possible_cpu(cpu
) {
7097 struct perf_cpu_context
*cpuctx
;
7099 cpuctx
= per_cpu_ptr(pmu
->pmu_cpu_context
, cpu
);
7101 if (cpuctx
->unique_pmu
== old_pmu
)
7102 cpuctx
->unique_pmu
= pmu
;
7106 static void free_pmu_context(struct pmu
*pmu
)
7110 mutex_lock(&pmus_lock
);
7112 * Like a real lame refcount.
7114 list_for_each_entry(i
, &pmus
, entry
) {
7115 if (i
->pmu_cpu_context
== pmu
->pmu_cpu_context
) {
7116 update_pmu_context(i
, pmu
);
7121 free_percpu(pmu
->pmu_cpu_context
);
7123 mutex_unlock(&pmus_lock
);
7125 static struct idr pmu_idr
;
7128 type_show(struct device
*dev
, struct device_attribute
*attr
, char *page
)
7130 struct pmu
*pmu
= dev_get_drvdata(dev
);
7132 return snprintf(page
, PAGE_SIZE
-1, "%d\n", pmu
->type
);
7134 static DEVICE_ATTR_RO(type
);
7137 perf_event_mux_interval_ms_show(struct device
*dev
,
7138 struct device_attribute
*attr
,
7141 struct pmu
*pmu
= dev_get_drvdata(dev
);
7143 return snprintf(page
, PAGE_SIZE
-1, "%d\n", pmu
->hrtimer_interval_ms
);
7147 perf_event_mux_interval_ms_store(struct device
*dev
,
7148 struct device_attribute
*attr
,
7149 const char *buf
, size_t count
)
7151 struct pmu
*pmu
= dev_get_drvdata(dev
);
7152 int timer
, cpu
, ret
;
7154 ret
= kstrtoint(buf
, 0, &timer
);
7161 /* same value, noting to do */
7162 if (timer
== pmu
->hrtimer_interval_ms
)
7165 pmu
->hrtimer_interval_ms
= timer
;
7167 /* update all cpuctx for this PMU */
7168 for_each_possible_cpu(cpu
) {
7169 struct perf_cpu_context
*cpuctx
;
7170 cpuctx
= per_cpu_ptr(pmu
->pmu_cpu_context
, cpu
);
7171 cpuctx
->hrtimer_interval
= ns_to_ktime(NSEC_PER_MSEC
* timer
);
7173 if (hrtimer_active(&cpuctx
->hrtimer
))
7174 hrtimer_forward_now(&cpuctx
->hrtimer
, cpuctx
->hrtimer_interval
);
7179 static DEVICE_ATTR_RW(perf_event_mux_interval_ms
);
7181 static struct attribute
*pmu_dev_attrs
[] = {
7182 &dev_attr_type
.attr
,
7183 &dev_attr_perf_event_mux_interval_ms
.attr
,
7186 ATTRIBUTE_GROUPS(pmu_dev
);
7188 static int pmu_bus_running
;
7189 static struct bus_type pmu_bus
= {
7190 .name
= "event_source",
7191 .dev_groups
= pmu_dev_groups
,
7194 static void pmu_dev_release(struct device
*dev
)
7199 static int pmu_dev_alloc(struct pmu
*pmu
)
7203 pmu
->dev
= kzalloc(sizeof(struct device
), GFP_KERNEL
);
7207 pmu
->dev
->groups
= pmu
->attr_groups
;
7208 device_initialize(pmu
->dev
);
7209 ret
= dev_set_name(pmu
->dev
, "%s", pmu
->name
);
7213 dev_set_drvdata(pmu
->dev
, pmu
);
7214 pmu
->dev
->bus
= &pmu_bus
;
7215 pmu
->dev
->release
= pmu_dev_release
;
7216 ret
= device_add(pmu
->dev
);
7224 put_device(pmu
->dev
);
7228 static struct lock_class_key cpuctx_mutex
;
7229 static struct lock_class_key cpuctx_lock
;
7231 int perf_pmu_register(struct pmu
*pmu
, const char *name
, int type
)
7235 mutex_lock(&pmus_lock
);
7237 pmu
->pmu_disable_count
= alloc_percpu(int);
7238 if (!pmu
->pmu_disable_count
)
7247 type
= idr_alloc(&pmu_idr
, pmu
, PERF_TYPE_MAX
, 0, GFP_KERNEL
);
7255 if (pmu_bus_running
) {
7256 ret
= pmu_dev_alloc(pmu
);
7262 pmu
->pmu_cpu_context
= find_pmu_context(pmu
->task_ctx_nr
);
7263 if (pmu
->pmu_cpu_context
)
7264 goto got_cpu_context
;
7267 pmu
->pmu_cpu_context
= alloc_percpu(struct perf_cpu_context
);
7268 if (!pmu
->pmu_cpu_context
)
7271 for_each_possible_cpu(cpu
) {
7272 struct perf_cpu_context
*cpuctx
;
7274 cpuctx
= per_cpu_ptr(pmu
->pmu_cpu_context
, cpu
);
7275 __perf_event_init_context(&cpuctx
->ctx
);
7276 lockdep_set_class(&cpuctx
->ctx
.mutex
, &cpuctx_mutex
);
7277 lockdep_set_class(&cpuctx
->ctx
.lock
, &cpuctx_lock
);
7278 cpuctx
->ctx
.pmu
= pmu
;
7280 __perf_cpu_hrtimer_init(cpuctx
, cpu
);
7282 cpuctx
->unique_pmu
= pmu
;
7286 if (!pmu
->start_txn
) {
7287 if (pmu
->pmu_enable
) {
7289 * If we have pmu_enable/pmu_disable calls, install
7290 * transaction stubs that use that to try and batch
7291 * hardware accesses.
7293 pmu
->start_txn
= perf_pmu_start_txn
;
7294 pmu
->commit_txn
= perf_pmu_commit_txn
;
7295 pmu
->cancel_txn
= perf_pmu_cancel_txn
;
7297 pmu
->start_txn
= perf_pmu_nop_void
;
7298 pmu
->commit_txn
= perf_pmu_nop_int
;
7299 pmu
->cancel_txn
= perf_pmu_nop_void
;
7303 if (!pmu
->pmu_enable
) {
7304 pmu
->pmu_enable
= perf_pmu_nop_void
;
7305 pmu
->pmu_disable
= perf_pmu_nop_void
;
7308 if (!pmu
->event_idx
)
7309 pmu
->event_idx
= perf_event_idx_default
;
7311 list_add_rcu(&pmu
->entry
, &pmus
);
7312 atomic_set(&pmu
->exclusive_cnt
, 0);
7315 mutex_unlock(&pmus_lock
);
7320 device_del(pmu
->dev
);
7321 put_device(pmu
->dev
);
7324 if (pmu
->type
>= PERF_TYPE_MAX
)
7325 idr_remove(&pmu_idr
, pmu
->type
);
7328 free_percpu(pmu
->pmu_disable_count
);
7331 EXPORT_SYMBOL_GPL(perf_pmu_register
);
7333 void perf_pmu_unregister(struct pmu
*pmu
)
7335 mutex_lock(&pmus_lock
);
7336 list_del_rcu(&pmu
->entry
);
7337 mutex_unlock(&pmus_lock
);
7340 * We dereference the pmu list under both SRCU and regular RCU, so
7341 * synchronize against both of those.
7343 synchronize_srcu(&pmus_srcu
);
7346 free_percpu(pmu
->pmu_disable_count
);
7347 if (pmu
->type
>= PERF_TYPE_MAX
)
7348 idr_remove(&pmu_idr
, pmu
->type
);
7349 device_del(pmu
->dev
);
7350 put_device(pmu
->dev
);
7351 free_pmu_context(pmu
);
7353 EXPORT_SYMBOL_GPL(perf_pmu_unregister
);
7355 static int perf_try_init_event(struct pmu
*pmu
, struct perf_event
*event
)
7357 struct perf_event_context
*ctx
= NULL
;
7360 if (!try_module_get(pmu
->module
))
7363 if (event
->group_leader
!= event
) {
7364 ctx
= perf_event_ctx_lock(event
->group_leader
);
7369 ret
= pmu
->event_init(event
);
7372 perf_event_ctx_unlock(event
->group_leader
, ctx
);
7375 module_put(pmu
->module
);
7380 struct pmu
*perf_init_event(struct perf_event
*event
)
7382 struct pmu
*pmu
= NULL
;
7386 idx
= srcu_read_lock(&pmus_srcu
);
7389 pmu
= idr_find(&pmu_idr
, event
->attr
.type
);
7392 ret
= perf_try_init_event(pmu
, event
);
7398 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
7399 ret
= perf_try_init_event(pmu
, event
);
7403 if (ret
!= -ENOENT
) {
7408 pmu
= ERR_PTR(-ENOENT
);
7410 srcu_read_unlock(&pmus_srcu
, idx
);
7415 static void account_event_cpu(struct perf_event
*event
, int cpu
)
7420 if (is_cgroup_event(event
))
7421 atomic_inc(&per_cpu(perf_cgroup_events
, cpu
));
7424 static void account_event(struct perf_event
*event
)
7429 if (event
->attach_state
& PERF_ATTACH_TASK
)
7430 static_key_slow_inc(&perf_sched_events
.key
);
7431 if (event
->attr
.mmap
|| event
->attr
.mmap_data
)
7432 atomic_inc(&nr_mmap_events
);
7433 if (event
->attr
.comm
)
7434 atomic_inc(&nr_comm_events
);
7435 if (event
->attr
.task
)
7436 atomic_inc(&nr_task_events
);
7437 if (event
->attr
.freq
) {
7438 if (atomic_inc_return(&nr_freq_events
) == 1)
7439 tick_nohz_full_kick_all();
7441 if (has_branch_stack(event
))
7442 static_key_slow_inc(&perf_sched_events
.key
);
7443 if (is_cgroup_event(event
))
7444 static_key_slow_inc(&perf_sched_events
.key
);
7446 account_event_cpu(event
, event
->cpu
);
7450 * Allocate and initialize a event structure
7452 static struct perf_event
*
7453 perf_event_alloc(struct perf_event_attr
*attr
, int cpu
,
7454 struct task_struct
*task
,
7455 struct perf_event
*group_leader
,
7456 struct perf_event
*parent_event
,
7457 perf_overflow_handler_t overflow_handler
,
7458 void *context
, int cgroup_fd
)
7461 struct perf_event
*event
;
7462 struct hw_perf_event
*hwc
;
7465 if ((unsigned)cpu
>= nr_cpu_ids
) {
7466 if (!task
|| cpu
!= -1)
7467 return ERR_PTR(-EINVAL
);
7470 event
= kzalloc(sizeof(*event
), GFP_KERNEL
);
7472 return ERR_PTR(-ENOMEM
);
7475 * Single events are their own group leaders, with an
7476 * empty sibling list:
7479 group_leader
= event
;
7481 mutex_init(&event
->child_mutex
);
7482 INIT_LIST_HEAD(&event
->child_list
);
7484 INIT_LIST_HEAD(&event
->group_entry
);
7485 INIT_LIST_HEAD(&event
->event_entry
);
7486 INIT_LIST_HEAD(&event
->sibling_list
);
7487 INIT_LIST_HEAD(&event
->rb_entry
);
7488 INIT_LIST_HEAD(&event
->active_entry
);
7489 INIT_HLIST_NODE(&event
->hlist_entry
);
7492 init_waitqueue_head(&event
->waitq
);
7493 init_irq_work(&event
->pending
, perf_pending_event
);
7495 mutex_init(&event
->mmap_mutex
);
7497 atomic_long_set(&event
->refcount
, 1);
7499 event
->attr
= *attr
;
7500 event
->group_leader
= group_leader
;
7504 event
->parent
= parent_event
;
7506 event
->ns
= get_pid_ns(task_active_pid_ns(current
));
7507 event
->id
= atomic64_inc_return(&perf_event_id
);
7509 event
->state
= PERF_EVENT_STATE_INACTIVE
;
7512 event
->attach_state
= PERF_ATTACH_TASK
;
7514 * XXX pmu::event_init needs to know what task to account to
7515 * and we cannot use the ctx information because we need the
7516 * pmu before we get a ctx.
7518 event
->hw
.target
= task
;
7521 event
->clock
= &local_clock
;
7523 event
->clock
= parent_event
->clock
;
7525 if (!overflow_handler
&& parent_event
) {
7526 overflow_handler
= parent_event
->overflow_handler
;
7527 context
= parent_event
->overflow_handler_context
;
7530 event
->overflow_handler
= overflow_handler
;
7531 event
->overflow_handler_context
= context
;
7533 perf_event__state_init(event
);
7538 hwc
->sample_period
= attr
->sample_period
;
7539 if (attr
->freq
&& attr
->sample_freq
)
7540 hwc
->sample_period
= 1;
7541 hwc
->last_period
= hwc
->sample_period
;
7543 local64_set(&hwc
->period_left
, hwc
->sample_period
);
7546 * we currently do not support PERF_FORMAT_GROUP on inherited events
7548 if (attr
->inherit
&& (attr
->read_format
& PERF_FORMAT_GROUP
))
7551 if (!has_branch_stack(event
))
7552 event
->attr
.branch_sample_type
= 0;
7554 if (cgroup_fd
!= -1) {
7555 err
= perf_cgroup_connect(cgroup_fd
, event
, attr
, group_leader
);
7560 pmu
= perf_init_event(event
);
7563 else if (IS_ERR(pmu
)) {
7568 err
= exclusive_event_init(event
);
7572 if (!event
->parent
) {
7573 if (event
->attr
.sample_type
& PERF_SAMPLE_CALLCHAIN
) {
7574 err
= get_callchain_buffers();
7583 exclusive_event_destroy(event
);
7587 event
->destroy(event
);
7588 module_put(pmu
->module
);
7590 if (is_cgroup_event(event
))
7591 perf_detach_cgroup(event
);
7593 put_pid_ns(event
->ns
);
7596 return ERR_PTR(err
);
7599 static int perf_copy_attr(struct perf_event_attr __user
*uattr
,
7600 struct perf_event_attr
*attr
)
7605 if (!access_ok(VERIFY_WRITE
, uattr
, PERF_ATTR_SIZE_VER0
))
7609 * zero the full structure, so that a short copy will be nice.
7611 memset(attr
, 0, sizeof(*attr
));
7613 ret
= get_user(size
, &uattr
->size
);
7617 if (size
> PAGE_SIZE
) /* silly large */
7620 if (!size
) /* abi compat */
7621 size
= PERF_ATTR_SIZE_VER0
;
7623 if (size
< PERF_ATTR_SIZE_VER0
)
7627 * If we're handed a bigger struct than we know of,
7628 * ensure all the unknown bits are 0 - i.e. new
7629 * user-space does not rely on any kernel feature
7630 * extensions we dont know about yet.
7632 if (size
> sizeof(*attr
)) {
7633 unsigned char __user
*addr
;
7634 unsigned char __user
*end
;
7637 addr
= (void __user
*)uattr
+ sizeof(*attr
);
7638 end
= (void __user
*)uattr
+ size
;
7640 for (; addr
< end
; addr
++) {
7641 ret
= get_user(val
, addr
);
7647 size
= sizeof(*attr
);
7650 ret
= copy_from_user(attr
, uattr
, size
);
7654 if (attr
->__reserved_1
)
7657 if (attr
->sample_type
& ~(PERF_SAMPLE_MAX
-1))
7660 if (attr
->read_format
& ~(PERF_FORMAT_MAX
-1))
7663 if (attr
->sample_type
& PERF_SAMPLE_BRANCH_STACK
) {
7664 u64 mask
= attr
->branch_sample_type
;
7666 /* only using defined bits */
7667 if (mask
& ~(PERF_SAMPLE_BRANCH_MAX
-1))
7670 /* at least one branch bit must be set */
7671 if (!(mask
& ~PERF_SAMPLE_BRANCH_PLM_ALL
))
7674 /* propagate priv level, when not set for branch */
7675 if (!(mask
& PERF_SAMPLE_BRANCH_PLM_ALL
)) {
7677 /* exclude_kernel checked on syscall entry */
7678 if (!attr
->exclude_kernel
)
7679 mask
|= PERF_SAMPLE_BRANCH_KERNEL
;
7681 if (!attr
->exclude_user
)
7682 mask
|= PERF_SAMPLE_BRANCH_USER
;
7684 if (!attr
->exclude_hv
)
7685 mask
|= PERF_SAMPLE_BRANCH_HV
;
7687 * adjust user setting (for HW filter setup)
7689 attr
->branch_sample_type
= mask
;
7691 /* privileged levels capture (kernel, hv): check permissions */
7692 if ((mask
& PERF_SAMPLE_BRANCH_PERM_PLM
)
7693 && perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN
))
7697 if (attr
->sample_type
& PERF_SAMPLE_REGS_USER
) {
7698 ret
= perf_reg_validate(attr
->sample_regs_user
);
7703 if (attr
->sample_type
& PERF_SAMPLE_STACK_USER
) {
7704 if (!arch_perf_have_user_stack_dump())
7708 * We have __u32 type for the size, but so far
7709 * we can only use __u16 as maximum due to the
7710 * __u16 sample size limit.
7712 if (attr
->sample_stack_user
>= USHRT_MAX
)
7714 else if (!IS_ALIGNED(attr
->sample_stack_user
, sizeof(u64
)))
7718 if (attr
->sample_type
& PERF_SAMPLE_REGS_INTR
)
7719 ret
= perf_reg_validate(attr
->sample_regs_intr
);
7724 put_user(sizeof(*attr
), &uattr
->size
);
7730 perf_event_set_output(struct perf_event
*event
, struct perf_event
*output_event
)
7732 struct ring_buffer
*rb
= NULL
;
7738 /* don't allow circular references */
7739 if (event
== output_event
)
7743 * Don't allow cross-cpu buffers
7745 if (output_event
->cpu
!= event
->cpu
)
7749 * If its not a per-cpu rb, it must be the same task.
7751 if (output_event
->cpu
== -1 && output_event
->ctx
!= event
->ctx
)
7755 * Mixing clocks in the same buffer is trouble you don't need.
7757 if (output_event
->clock
!= event
->clock
)
7761 * If both events generate aux data, they must be on the same PMU
7763 if (has_aux(event
) && has_aux(output_event
) &&
7764 event
->pmu
!= output_event
->pmu
)
7768 mutex_lock(&event
->mmap_mutex
);
7769 /* Can't redirect output if we've got an active mmap() */
7770 if (atomic_read(&event
->mmap_count
))
7774 /* get the rb we want to redirect to */
7775 rb
= ring_buffer_get(output_event
);
7780 ring_buffer_attach(event
, rb
);
7784 mutex_unlock(&event
->mmap_mutex
);
7790 static void mutex_lock_double(struct mutex
*a
, struct mutex
*b
)
7796 mutex_lock_nested(b
, SINGLE_DEPTH_NESTING
);
7799 static int perf_event_set_clock(struct perf_event
*event
, clockid_t clk_id
)
7801 bool nmi_safe
= false;
7804 case CLOCK_MONOTONIC
:
7805 event
->clock
= &ktime_get_mono_fast_ns
;
7809 case CLOCK_MONOTONIC_RAW
:
7810 event
->clock
= &ktime_get_raw_fast_ns
;
7814 case CLOCK_REALTIME
:
7815 event
->clock
= &ktime_get_real_ns
;
7818 case CLOCK_BOOTTIME
:
7819 event
->clock
= &ktime_get_boot_ns
;
7823 event
->clock
= &ktime_get_tai_ns
;
7830 if (!nmi_safe
&& !(event
->pmu
->capabilities
& PERF_PMU_CAP_NO_NMI
))
7837 * sys_perf_event_open - open a performance event, associate it to a task/cpu
7839 * @attr_uptr: event_id type attributes for monitoring/sampling
7842 * @group_fd: group leader event fd
7844 SYSCALL_DEFINE5(perf_event_open
,
7845 struct perf_event_attr __user
*, attr_uptr
,
7846 pid_t
, pid
, int, cpu
, int, group_fd
, unsigned long, flags
)
7848 struct perf_event
*group_leader
= NULL
, *output_event
= NULL
;
7849 struct perf_event
*event
, *sibling
;
7850 struct perf_event_attr attr
;
7851 struct perf_event_context
*ctx
, *uninitialized_var(gctx
);
7852 struct file
*event_file
= NULL
;
7853 struct fd group
= {NULL
, 0};
7854 struct task_struct
*task
= NULL
;
7859 int f_flags
= O_RDWR
;
7862 /* for future expandability... */
7863 if (flags
& ~PERF_FLAG_ALL
)
7866 err
= perf_copy_attr(attr_uptr
, &attr
);
7870 if (!attr
.exclude_kernel
) {
7871 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN
))
7876 if (attr
.sample_freq
> sysctl_perf_event_sample_rate
)
7879 if (attr
.sample_period
& (1ULL << 63))
7884 * In cgroup mode, the pid argument is used to pass the fd
7885 * opened to the cgroup directory in cgroupfs. The cpu argument
7886 * designates the cpu on which to monitor threads from that
7889 if ((flags
& PERF_FLAG_PID_CGROUP
) && (pid
== -1 || cpu
== -1))
7892 if (flags
& PERF_FLAG_FD_CLOEXEC
)
7893 f_flags
|= O_CLOEXEC
;
7895 event_fd
= get_unused_fd_flags(f_flags
);
7899 if (group_fd
!= -1) {
7900 err
= perf_fget_light(group_fd
, &group
);
7903 group_leader
= group
.file
->private_data
;
7904 if (flags
& PERF_FLAG_FD_OUTPUT
)
7905 output_event
= group_leader
;
7906 if (flags
& PERF_FLAG_FD_NO_GROUP
)
7907 group_leader
= NULL
;
7910 if (pid
!= -1 && !(flags
& PERF_FLAG_PID_CGROUP
)) {
7911 task
= find_lively_task_by_vpid(pid
);
7913 err
= PTR_ERR(task
);
7918 if (task
&& group_leader
&&
7919 group_leader
->attr
.inherit
!= attr
.inherit
) {
7926 if (flags
& PERF_FLAG_PID_CGROUP
)
7929 event
= perf_event_alloc(&attr
, cpu
, task
, group_leader
, NULL
,
7930 NULL
, NULL
, cgroup_fd
);
7931 if (IS_ERR(event
)) {
7932 err
= PTR_ERR(event
);
7936 if (is_sampling_event(event
)) {
7937 if (event
->pmu
->capabilities
& PERF_PMU_CAP_NO_INTERRUPT
) {
7943 account_event(event
);
7946 * Special case software events and allow them to be part of
7947 * any hardware group.
7951 if (attr
.use_clockid
) {
7952 err
= perf_event_set_clock(event
, attr
.clockid
);
7958 (is_software_event(event
) != is_software_event(group_leader
))) {
7959 if (is_software_event(event
)) {
7961 * If event and group_leader are not both a software
7962 * event, and event is, then group leader is not.
7964 * Allow the addition of software events to !software
7965 * groups, this is safe because software events never
7968 pmu
= group_leader
->pmu
;
7969 } else if (is_software_event(group_leader
) &&
7970 (group_leader
->group_flags
& PERF_GROUP_SOFTWARE
)) {
7972 * In case the group is a pure software group, and we
7973 * try to add a hardware event, move the whole group to
7974 * the hardware context.
7981 * Get the target context (task or percpu):
7983 ctx
= find_get_context(pmu
, task
, event
);
7989 if ((pmu
->capabilities
& PERF_PMU_CAP_EXCLUSIVE
) && group_leader
) {
7995 put_task_struct(task
);
8000 * Look up the group leader (we will attach this event to it):
8006 * Do not allow a recursive hierarchy (this new sibling
8007 * becoming part of another group-sibling):
8009 if (group_leader
->group_leader
!= group_leader
)
8012 /* All events in a group should have the same clock */
8013 if (group_leader
->clock
!= event
->clock
)
8017 * Do not allow to attach to a group in a different
8018 * task or CPU context:
8022 * Make sure we're both on the same task, or both
8025 if (group_leader
->ctx
->task
!= ctx
->task
)
8029 * Make sure we're both events for the same CPU;
8030 * grouping events for different CPUs is broken; since
8031 * you can never concurrently schedule them anyhow.
8033 if (group_leader
->cpu
!= event
->cpu
)
8036 if (group_leader
->ctx
!= ctx
)
8041 * Only a group leader can be exclusive or pinned
8043 if (attr
.exclusive
|| attr
.pinned
)
8048 err
= perf_event_set_output(event
, output_event
);
8053 event_file
= anon_inode_getfile("[perf_event]", &perf_fops
, event
,
8055 if (IS_ERR(event_file
)) {
8056 err
= PTR_ERR(event_file
);
8061 gctx
= group_leader
->ctx
;
8064 * See perf_event_ctx_lock() for comments on the details
8065 * of swizzling perf_event::ctx.
8067 mutex_lock_double(&gctx
->mutex
, &ctx
->mutex
);
8069 perf_remove_from_context(group_leader
, false);
8071 list_for_each_entry(sibling
, &group_leader
->sibling_list
,
8073 perf_remove_from_context(sibling
, false);
8077 mutex_lock(&ctx
->mutex
);
8080 WARN_ON_ONCE(ctx
->parent_ctx
);
8084 * Wait for everybody to stop referencing the events through
8085 * the old lists, before installing it on new lists.
8090 * Install the group siblings before the group leader.
8092 * Because a group leader will try and install the entire group
8093 * (through the sibling list, which is still in-tact), we can
8094 * end up with siblings installed in the wrong context.
8096 * By installing siblings first we NO-OP because they're not
8097 * reachable through the group lists.
8099 list_for_each_entry(sibling
, &group_leader
->sibling_list
,
8101 perf_event__state_init(sibling
);
8102 perf_install_in_context(ctx
, sibling
, sibling
->cpu
);
8107 * Removing from the context ends up with disabled
8108 * event. What we want here is event in the initial
8109 * startup state, ready to be add into new context.
8111 perf_event__state_init(group_leader
);
8112 perf_install_in_context(ctx
, group_leader
, group_leader
->cpu
);
8116 if (!exclusive_event_installable(event
, ctx
)) {
8118 mutex_unlock(&ctx
->mutex
);
8123 perf_install_in_context(ctx
, event
, event
->cpu
);
8124 perf_unpin_context(ctx
);
8127 mutex_unlock(&gctx
->mutex
);
8130 mutex_unlock(&ctx
->mutex
);
8134 event
->owner
= current
;
8136 mutex_lock(¤t
->perf_event_mutex
);
8137 list_add_tail(&event
->owner_entry
, ¤t
->perf_event_list
);
8138 mutex_unlock(¤t
->perf_event_mutex
);
8141 * Precalculate sample_data sizes
8143 perf_event__header_size(event
);
8144 perf_event__id_header_size(event
);
8147 * Drop the reference on the group_event after placing the
8148 * new event on the sibling_list. This ensures destruction
8149 * of the group leader will find the pointer to itself in
8150 * perf_group_detach().
8153 fd_install(event_fd
, event_file
);
8157 perf_unpin_context(ctx
);
8165 put_task_struct(task
);
8169 put_unused_fd(event_fd
);
8174 * perf_event_create_kernel_counter
8176 * @attr: attributes of the counter to create
8177 * @cpu: cpu in which the counter is bound
8178 * @task: task to profile (NULL for percpu)
8181 perf_event_create_kernel_counter(struct perf_event_attr
*attr
, int cpu
,
8182 struct task_struct
*task
,
8183 perf_overflow_handler_t overflow_handler
,
8186 struct perf_event_context
*ctx
;
8187 struct perf_event
*event
;
8191 * Get the target context (task or percpu):
8194 event
= perf_event_alloc(attr
, cpu
, task
, NULL
, NULL
,
8195 overflow_handler
, context
, -1);
8196 if (IS_ERR(event
)) {
8197 err
= PTR_ERR(event
);
8201 /* Mark owner so we could distinguish it from user events. */
8202 event
->owner
= EVENT_OWNER_KERNEL
;
8204 account_event(event
);
8206 ctx
= find_get_context(event
->pmu
, task
, event
);
8212 WARN_ON_ONCE(ctx
->parent_ctx
);
8213 mutex_lock(&ctx
->mutex
);
8214 if (!exclusive_event_installable(event
, ctx
)) {
8215 mutex_unlock(&ctx
->mutex
);
8216 perf_unpin_context(ctx
);
8222 perf_install_in_context(ctx
, event
, cpu
);
8223 perf_unpin_context(ctx
);
8224 mutex_unlock(&ctx
->mutex
);
8231 return ERR_PTR(err
);
8233 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter
);
8235 void perf_pmu_migrate_context(struct pmu
*pmu
, int src_cpu
, int dst_cpu
)
8237 struct perf_event_context
*src_ctx
;
8238 struct perf_event_context
*dst_ctx
;
8239 struct perf_event
*event
, *tmp
;
8242 src_ctx
= &per_cpu_ptr(pmu
->pmu_cpu_context
, src_cpu
)->ctx
;
8243 dst_ctx
= &per_cpu_ptr(pmu
->pmu_cpu_context
, dst_cpu
)->ctx
;
8246 * See perf_event_ctx_lock() for comments on the details
8247 * of swizzling perf_event::ctx.
8249 mutex_lock_double(&src_ctx
->mutex
, &dst_ctx
->mutex
);
8250 list_for_each_entry_safe(event
, tmp
, &src_ctx
->event_list
,
8252 perf_remove_from_context(event
, false);
8253 unaccount_event_cpu(event
, src_cpu
);
8255 list_add(&event
->migrate_entry
, &events
);
8259 * Wait for the events to quiesce before re-instating them.
8264 * Re-instate events in 2 passes.
8266 * Skip over group leaders and only install siblings on this first
8267 * pass, siblings will not get enabled without a leader, however a
8268 * leader will enable its siblings, even if those are still on the old
8271 list_for_each_entry_safe(event
, tmp
, &events
, migrate_entry
) {
8272 if (event
->group_leader
== event
)
8275 list_del(&event
->migrate_entry
);
8276 if (event
->state
>= PERF_EVENT_STATE_OFF
)
8277 event
->state
= PERF_EVENT_STATE_INACTIVE
;
8278 account_event_cpu(event
, dst_cpu
);
8279 perf_install_in_context(dst_ctx
, event
, dst_cpu
);
8284 * Once all the siblings are setup properly, install the group leaders
8287 list_for_each_entry_safe(event
, tmp
, &events
, migrate_entry
) {
8288 list_del(&event
->migrate_entry
);
8289 if (event
->state
>= PERF_EVENT_STATE_OFF
)
8290 event
->state
= PERF_EVENT_STATE_INACTIVE
;
8291 account_event_cpu(event
, dst_cpu
);
8292 perf_install_in_context(dst_ctx
, event
, dst_cpu
);
8295 mutex_unlock(&dst_ctx
->mutex
);
8296 mutex_unlock(&src_ctx
->mutex
);
8298 EXPORT_SYMBOL_GPL(perf_pmu_migrate_context
);
8300 static void sync_child_event(struct perf_event
*child_event
,
8301 struct task_struct
*child
)
8303 struct perf_event
*parent_event
= child_event
->parent
;
8306 if (child_event
->attr
.inherit_stat
)
8307 perf_event_read_event(child_event
, child
);
8309 child_val
= perf_event_count(child_event
);
8312 * Add back the child's count to the parent's count:
8314 atomic64_add(child_val
, &parent_event
->child_count
);
8315 atomic64_add(child_event
->total_time_enabled
,
8316 &parent_event
->child_total_time_enabled
);
8317 atomic64_add(child_event
->total_time_running
,
8318 &parent_event
->child_total_time_running
);
8321 * Remove this event from the parent's list
8323 WARN_ON_ONCE(parent_event
->ctx
->parent_ctx
);
8324 mutex_lock(&parent_event
->child_mutex
);
8325 list_del_init(&child_event
->child_list
);
8326 mutex_unlock(&parent_event
->child_mutex
);
8329 * Make sure user/parent get notified, that we just
8332 perf_event_wakeup(parent_event
);
8335 * Release the parent event, if this was the last
8338 put_event(parent_event
);
8342 __perf_event_exit_task(struct perf_event
*child_event
,
8343 struct perf_event_context
*child_ctx
,
8344 struct task_struct
*child
)
8347 * Do not destroy the 'original' grouping; because of the context
8348 * switch optimization the original events could've ended up in a
8349 * random child task.
8351 * If we were to destroy the original group, all group related
8352 * operations would cease to function properly after this random
8355 * Do destroy all inherited groups, we don't care about those
8356 * and being thorough is better.
8358 perf_remove_from_context(child_event
, !!child_event
->parent
);
8361 * It can happen that the parent exits first, and has events
8362 * that are still around due to the child reference. These
8363 * events need to be zapped.
8365 if (child_event
->parent
) {
8366 sync_child_event(child_event
, child
);
8367 free_event(child_event
);
8369 child_event
->state
= PERF_EVENT_STATE_EXIT
;
8370 perf_event_wakeup(child_event
);
8374 static void perf_event_exit_task_context(struct task_struct
*child
, int ctxn
)
8376 struct perf_event
*child_event
, *next
;
8377 struct perf_event_context
*child_ctx
, *clone_ctx
= NULL
;
8378 unsigned long flags
;
8380 if (likely(!child
->perf_event_ctxp
[ctxn
])) {
8381 perf_event_task(child
, NULL
, 0);
8385 local_irq_save(flags
);
8387 * We can't reschedule here because interrupts are disabled,
8388 * and either child is current or it is a task that can't be
8389 * scheduled, so we are now safe from rescheduling changing
8392 child_ctx
= rcu_dereference_raw(child
->perf_event_ctxp
[ctxn
]);
8395 * Take the context lock here so that if find_get_context is
8396 * reading child->perf_event_ctxp, we wait until it has
8397 * incremented the context's refcount before we do put_ctx below.
8399 raw_spin_lock(&child_ctx
->lock
);
8400 task_ctx_sched_out(child_ctx
);
8401 child
->perf_event_ctxp
[ctxn
] = NULL
;
8404 * If this context is a clone; unclone it so it can't get
8405 * swapped to another process while we're removing all
8406 * the events from it.
8408 clone_ctx
= unclone_ctx(child_ctx
);
8409 update_context_time(child_ctx
);
8410 raw_spin_unlock_irqrestore(&child_ctx
->lock
, flags
);
8416 * Report the task dead after unscheduling the events so that we
8417 * won't get any samples after PERF_RECORD_EXIT. We can however still
8418 * get a few PERF_RECORD_READ events.
8420 perf_event_task(child
, child_ctx
, 0);
8423 * We can recurse on the same lock type through:
8425 * __perf_event_exit_task()
8426 * sync_child_event()
8428 * mutex_lock(&ctx->mutex)
8430 * But since its the parent context it won't be the same instance.
8432 mutex_lock(&child_ctx
->mutex
);
8434 list_for_each_entry_safe(child_event
, next
, &child_ctx
->event_list
, event_entry
)
8435 __perf_event_exit_task(child_event
, child_ctx
, child
);
8437 mutex_unlock(&child_ctx
->mutex
);
8443 * When a child task exits, feed back event values to parent events.
8445 void perf_event_exit_task(struct task_struct
*child
)
8447 struct perf_event
*event
, *tmp
;
8450 mutex_lock(&child
->perf_event_mutex
);
8451 list_for_each_entry_safe(event
, tmp
, &child
->perf_event_list
,
8453 list_del_init(&event
->owner_entry
);
8456 * Ensure the list deletion is visible before we clear
8457 * the owner, closes a race against perf_release() where
8458 * we need to serialize on the owner->perf_event_mutex.
8461 event
->owner
= NULL
;
8463 mutex_unlock(&child
->perf_event_mutex
);
8465 for_each_task_context_nr(ctxn
)
8466 perf_event_exit_task_context(child
, ctxn
);
8469 static void perf_free_event(struct perf_event
*event
,
8470 struct perf_event_context
*ctx
)
8472 struct perf_event
*parent
= event
->parent
;
8474 if (WARN_ON_ONCE(!parent
))
8477 mutex_lock(&parent
->child_mutex
);
8478 list_del_init(&event
->child_list
);
8479 mutex_unlock(&parent
->child_mutex
);
8483 raw_spin_lock_irq(&ctx
->lock
);
8484 perf_group_detach(event
);
8485 list_del_event(event
, ctx
);
8486 raw_spin_unlock_irq(&ctx
->lock
);
8491 * Free an unexposed, unused context as created by inheritance by
8492 * perf_event_init_task below, used by fork() in case of fail.
8494 * Not all locks are strictly required, but take them anyway to be nice and
8495 * help out with the lockdep assertions.
8497 void perf_event_free_task(struct task_struct
*task
)
8499 struct perf_event_context
*ctx
;
8500 struct perf_event
*event
, *tmp
;
8503 for_each_task_context_nr(ctxn
) {
8504 ctx
= task
->perf_event_ctxp
[ctxn
];
8508 mutex_lock(&ctx
->mutex
);
8510 list_for_each_entry_safe(event
, tmp
, &ctx
->pinned_groups
,
8512 perf_free_event(event
, ctx
);
8514 list_for_each_entry_safe(event
, tmp
, &ctx
->flexible_groups
,
8516 perf_free_event(event
, ctx
);
8518 if (!list_empty(&ctx
->pinned_groups
) ||
8519 !list_empty(&ctx
->flexible_groups
))
8522 mutex_unlock(&ctx
->mutex
);
8528 void perf_event_delayed_put(struct task_struct
*task
)
8532 for_each_task_context_nr(ctxn
)
8533 WARN_ON_ONCE(task
->perf_event_ctxp
[ctxn
]);
8537 * inherit a event from parent task to child task:
8539 static struct perf_event
*
8540 inherit_event(struct perf_event
*parent_event
,
8541 struct task_struct
*parent
,
8542 struct perf_event_context
*parent_ctx
,
8543 struct task_struct
*child
,
8544 struct perf_event
*group_leader
,
8545 struct perf_event_context
*child_ctx
)
8547 enum perf_event_active_state parent_state
= parent_event
->state
;
8548 struct perf_event
*child_event
;
8549 unsigned long flags
;
8552 * Instead of creating recursive hierarchies of events,
8553 * we link inherited events back to the original parent,
8554 * which has a filp for sure, which we use as the reference
8557 if (parent_event
->parent
)
8558 parent_event
= parent_event
->parent
;
8560 child_event
= perf_event_alloc(&parent_event
->attr
,
8563 group_leader
, parent_event
,
8565 if (IS_ERR(child_event
))
8568 if (is_orphaned_event(parent_event
) ||
8569 !atomic_long_inc_not_zero(&parent_event
->refcount
)) {
8570 free_event(child_event
);
8577 * Make the child state follow the state of the parent event,
8578 * not its attr.disabled bit. We hold the parent's mutex,
8579 * so we won't race with perf_event_{en, dis}able_family.
8581 if (parent_state
>= PERF_EVENT_STATE_INACTIVE
)
8582 child_event
->state
= PERF_EVENT_STATE_INACTIVE
;
8584 child_event
->state
= PERF_EVENT_STATE_OFF
;
8586 if (parent_event
->attr
.freq
) {
8587 u64 sample_period
= parent_event
->hw
.sample_period
;
8588 struct hw_perf_event
*hwc
= &child_event
->hw
;
8590 hwc
->sample_period
= sample_period
;
8591 hwc
->last_period
= sample_period
;
8593 local64_set(&hwc
->period_left
, sample_period
);
8596 child_event
->ctx
= child_ctx
;
8597 child_event
->overflow_handler
= parent_event
->overflow_handler
;
8598 child_event
->overflow_handler_context
8599 = parent_event
->overflow_handler_context
;
8602 * Precalculate sample_data sizes
8604 perf_event__header_size(child_event
);
8605 perf_event__id_header_size(child_event
);
8608 * Link it up in the child's context:
8610 raw_spin_lock_irqsave(&child_ctx
->lock
, flags
);
8611 add_event_to_ctx(child_event
, child_ctx
);
8612 raw_spin_unlock_irqrestore(&child_ctx
->lock
, flags
);
8615 * Link this into the parent event's child list
8617 WARN_ON_ONCE(parent_event
->ctx
->parent_ctx
);
8618 mutex_lock(&parent_event
->child_mutex
);
8619 list_add_tail(&child_event
->child_list
, &parent_event
->child_list
);
8620 mutex_unlock(&parent_event
->child_mutex
);
8625 static int inherit_group(struct perf_event
*parent_event
,
8626 struct task_struct
*parent
,
8627 struct perf_event_context
*parent_ctx
,
8628 struct task_struct
*child
,
8629 struct perf_event_context
*child_ctx
)
8631 struct perf_event
*leader
;
8632 struct perf_event
*sub
;
8633 struct perf_event
*child_ctr
;
8635 leader
= inherit_event(parent_event
, parent
, parent_ctx
,
8636 child
, NULL
, child_ctx
);
8638 return PTR_ERR(leader
);
8639 list_for_each_entry(sub
, &parent_event
->sibling_list
, group_entry
) {
8640 child_ctr
= inherit_event(sub
, parent
, parent_ctx
,
8641 child
, leader
, child_ctx
);
8642 if (IS_ERR(child_ctr
))
8643 return PTR_ERR(child_ctr
);
8649 inherit_task_group(struct perf_event
*event
, struct task_struct
*parent
,
8650 struct perf_event_context
*parent_ctx
,
8651 struct task_struct
*child
, int ctxn
,
8655 struct perf_event_context
*child_ctx
;
8657 if (!event
->attr
.inherit
) {
8662 child_ctx
= child
->perf_event_ctxp
[ctxn
];
8665 * This is executed from the parent task context, so
8666 * inherit events that have been marked for cloning.
8667 * First allocate and initialize a context for the
8671 child_ctx
= alloc_perf_context(parent_ctx
->pmu
, child
);
8675 child
->perf_event_ctxp
[ctxn
] = child_ctx
;
8678 ret
= inherit_group(event
, parent
, parent_ctx
,
8688 * Initialize the perf_event context in task_struct
8690 static int perf_event_init_context(struct task_struct
*child
, int ctxn
)
8692 struct perf_event_context
*child_ctx
, *parent_ctx
;
8693 struct perf_event_context
*cloned_ctx
;
8694 struct perf_event
*event
;
8695 struct task_struct
*parent
= current
;
8696 int inherited_all
= 1;
8697 unsigned long flags
;
8700 if (likely(!parent
->perf_event_ctxp
[ctxn
]))
8704 * If the parent's context is a clone, pin it so it won't get
8707 parent_ctx
= perf_pin_task_context(parent
, ctxn
);
8712 * No need to check if parent_ctx != NULL here; since we saw
8713 * it non-NULL earlier, the only reason for it to become NULL
8714 * is if we exit, and since we're currently in the middle of
8715 * a fork we can't be exiting at the same time.
8719 * Lock the parent list. No need to lock the child - not PID
8720 * hashed yet and not running, so nobody can access it.
8722 mutex_lock(&parent_ctx
->mutex
);
8725 * We dont have to disable NMIs - we are only looking at
8726 * the list, not manipulating it:
8728 list_for_each_entry(event
, &parent_ctx
->pinned_groups
, group_entry
) {
8729 ret
= inherit_task_group(event
, parent
, parent_ctx
,
8730 child
, ctxn
, &inherited_all
);
8736 * We can't hold ctx->lock when iterating the ->flexible_group list due
8737 * to allocations, but we need to prevent rotation because
8738 * rotate_ctx() will change the list from interrupt context.
8740 raw_spin_lock_irqsave(&parent_ctx
->lock
, flags
);
8741 parent_ctx
->rotate_disable
= 1;
8742 raw_spin_unlock_irqrestore(&parent_ctx
->lock
, flags
);
8744 list_for_each_entry(event
, &parent_ctx
->flexible_groups
, group_entry
) {
8745 ret
= inherit_task_group(event
, parent
, parent_ctx
,
8746 child
, ctxn
, &inherited_all
);
8751 raw_spin_lock_irqsave(&parent_ctx
->lock
, flags
);
8752 parent_ctx
->rotate_disable
= 0;
8754 child_ctx
= child
->perf_event_ctxp
[ctxn
];
8756 if (child_ctx
&& inherited_all
) {
8758 * Mark the child context as a clone of the parent
8759 * context, or of whatever the parent is a clone of.
8761 * Note that if the parent is a clone, the holding of
8762 * parent_ctx->lock avoids it from being uncloned.
8764 cloned_ctx
= parent_ctx
->parent_ctx
;
8766 child_ctx
->parent_ctx
= cloned_ctx
;
8767 child_ctx
->parent_gen
= parent_ctx
->parent_gen
;
8769 child_ctx
->parent_ctx
= parent_ctx
;
8770 child_ctx
->parent_gen
= parent_ctx
->generation
;
8772 get_ctx(child_ctx
->parent_ctx
);
8775 raw_spin_unlock_irqrestore(&parent_ctx
->lock
, flags
);
8776 mutex_unlock(&parent_ctx
->mutex
);
8778 perf_unpin_context(parent_ctx
);
8779 put_ctx(parent_ctx
);
8785 * Initialize the perf_event context in task_struct
8787 int perf_event_init_task(struct task_struct
*child
)
8791 memset(child
->perf_event_ctxp
, 0, sizeof(child
->perf_event_ctxp
));
8792 mutex_init(&child
->perf_event_mutex
);
8793 INIT_LIST_HEAD(&child
->perf_event_list
);
8795 for_each_task_context_nr(ctxn
) {
8796 ret
= perf_event_init_context(child
, ctxn
);
8798 perf_event_free_task(child
);
8806 static void __init
perf_event_init_all_cpus(void)
8808 struct swevent_htable
*swhash
;
8811 for_each_possible_cpu(cpu
) {
8812 swhash
= &per_cpu(swevent_htable
, cpu
);
8813 mutex_init(&swhash
->hlist_mutex
);
8814 INIT_LIST_HEAD(&per_cpu(active_ctx_list
, cpu
));
8818 static void perf_event_init_cpu(int cpu
)
8820 struct swevent_htable
*swhash
= &per_cpu(swevent_htable
, cpu
);
8822 mutex_lock(&swhash
->hlist_mutex
);
8823 swhash
->online
= true;
8824 if (swhash
->hlist_refcount
> 0) {
8825 struct swevent_hlist
*hlist
;
8827 hlist
= kzalloc_node(sizeof(*hlist
), GFP_KERNEL
, cpu_to_node(cpu
));
8829 rcu_assign_pointer(swhash
->swevent_hlist
, hlist
);
8831 mutex_unlock(&swhash
->hlist_mutex
);
8834 #if defined CONFIG_HOTPLUG_CPU || defined CONFIG_KEXEC
8835 static void __perf_event_exit_context(void *__info
)
8837 struct remove_event re
= { .detach_group
= true };
8838 struct perf_event_context
*ctx
= __info
;
8841 list_for_each_entry_rcu(re
.event
, &ctx
->event_list
, event_entry
)
8842 __perf_remove_from_context(&re
);
8846 static void perf_event_exit_cpu_context(int cpu
)
8848 struct perf_event_context
*ctx
;
8852 idx
= srcu_read_lock(&pmus_srcu
);
8853 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
8854 ctx
= &per_cpu_ptr(pmu
->pmu_cpu_context
, cpu
)->ctx
;
8856 mutex_lock(&ctx
->mutex
);
8857 smp_call_function_single(cpu
, __perf_event_exit_context
, ctx
, 1);
8858 mutex_unlock(&ctx
->mutex
);
8860 srcu_read_unlock(&pmus_srcu
, idx
);
8863 static void perf_event_exit_cpu(int cpu
)
8865 struct swevent_htable
*swhash
= &per_cpu(swevent_htable
, cpu
);
8867 perf_event_exit_cpu_context(cpu
);
8869 mutex_lock(&swhash
->hlist_mutex
);
8870 swhash
->online
= false;
8871 swevent_hlist_release(swhash
);
8872 mutex_unlock(&swhash
->hlist_mutex
);
8875 static inline void perf_event_exit_cpu(int cpu
) { }
8879 perf_reboot(struct notifier_block
*notifier
, unsigned long val
, void *v
)
8883 for_each_online_cpu(cpu
)
8884 perf_event_exit_cpu(cpu
);
8890 * Run the perf reboot notifier at the very last possible moment so that
8891 * the generic watchdog code runs as long as possible.
8893 static struct notifier_block perf_reboot_notifier
= {
8894 .notifier_call
= perf_reboot
,
8895 .priority
= INT_MIN
,
8899 perf_cpu_notify(struct notifier_block
*self
, unsigned long action
, void *hcpu
)
8901 unsigned int cpu
= (long)hcpu
;
8903 switch (action
& ~CPU_TASKS_FROZEN
) {
8905 case CPU_UP_PREPARE
:
8906 case CPU_DOWN_FAILED
:
8907 perf_event_init_cpu(cpu
);
8910 case CPU_UP_CANCELED
:
8911 case CPU_DOWN_PREPARE
:
8912 perf_event_exit_cpu(cpu
);
8921 void __init
perf_event_init(void)
8927 perf_event_init_all_cpus();
8928 init_srcu_struct(&pmus_srcu
);
8929 perf_pmu_register(&perf_swevent
, "software", PERF_TYPE_SOFTWARE
);
8930 perf_pmu_register(&perf_cpu_clock
, NULL
, -1);
8931 perf_pmu_register(&perf_task_clock
, NULL
, -1);
8933 perf_cpu_notifier(perf_cpu_notify
);
8934 register_reboot_notifier(&perf_reboot_notifier
);
8936 ret
= init_hw_breakpoint();
8937 WARN(ret
, "hw_breakpoint initialization failed with: %d", ret
);
8939 /* do not patch jump label more than once per second */
8940 jump_label_rate_limit(&perf_sched_events
, HZ
);
8943 * Build time assertion that we keep the data_head at the intended
8944 * location. IOW, validation we got the __reserved[] size right.
8946 BUILD_BUG_ON((offsetof(struct perf_event_mmap_page
, data_head
))
8950 ssize_t
perf_event_sysfs_show(struct device
*dev
, struct device_attribute
*attr
,
8953 struct perf_pmu_events_attr
*pmu_attr
=
8954 container_of(attr
, struct perf_pmu_events_attr
, attr
);
8956 if (pmu_attr
->event_str
)
8957 return sprintf(page
, "%s\n", pmu_attr
->event_str
);
8962 static int __init
perf_event_sysfs_init(void)
8967 mutex_lock(&pmus_lock
);
8969 ret
= bus_register(&pmu_bus
);
8973 list_for_each_entry(pmu
, &pmus
, entry
) {
8974 if (!pmu
->name
|| pmu
->type
< 0)
8977 ret
= pmu_dev_alloc(pmu
);
8978 WARN(ret
, "Failed to register pmu: %s, reason %d\n", pmu
->name
, ret
);
8980 pmu_bus_running
= 1;
8984 mutex_unlock(&pmus_lock
);
8988 device_initcall(perf_event_sysfs_init
);
8990 #ifdef CONFIG_CGROUP_PERF
8991 static struct cgroup_subsys_state
*
8992 perf_cgroup_css_alloc(struct cgroup_subsys_state
*parent_css
)
8994 struct perf_cgroup
*jc
;
8996 jc
= kzalloc(sizeof(*jc
), GFP_KERNEL
);
8998 return ERR_PTR(-ENOMEM
);
9000 jc
->info
= alloc_percpu(struct perf_cgroup_info
);
9003 return ERR_PTR(-ENOMEM
);
9009 static void perf_cgroup_css_free(struct cgroup_subsys_state
*css
)
9011 struct perf_cgroup
*jc
= container_of(css
, struct perf_cgroup
, css
);
9013 free_percpu(jc
->info
);
9017 static int __perf_cgroup_move(void *info
)
9019 struct task_struct
*task
= info
;
9020 perf_cgroup_switch(task
, PERF_CGROUP_SWOUT
| PERF_CGROUP_SWIN
);
9024 static void perf_cgroup_attach(struct cgroup_subsys_state
*css
,
9025 struct cgroup_taskset
*tset
)
9027 struct task_struct
*task
;
9029 cgroup_taskset_for_each(task
, tset
)
9030 task_function_call(task
, __perf_cgroup_move
, task
);
9033 static void perf_cgroup_exit(struct cgroup_subsys_state
*css
,
9034 struct cgroup_subsys_state
*old_css
,
9035 struct task_struct
*task
)
9038 * cgroup_exit() is called in the copy_process() failure path.
9039 * Ignore this case since the task hasn't ran yet, this avoids
9040 * trying to poke a half freed task state from generic code.
9042 if (!(task
->flags
& PF_EXITING
))
9045 task_function_call(task
, __perf_cgroup_move
, task
);
9048 struct cgroup_subsys perf_event_cgrp_subsys
= {
9049 .css_alloc
= perf_cgroup_css_alloc
,
9050 .css_free
= perf_cgroup_css_free
,
9051 .exit
= perf_cgroup_exit
,
9052 .attach
= perf_cgroup_attach
,
9054 #endif /* CONFIG_CGROUP_PERF */