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 if (!hrtimer_callback_running(hr
))
838 __hrtimer_start_range_ns(hr
, cpuctx
->hrtimer_interval
,
839 0, HRTIMER_MODE_REL_PINNED
, 0);
842 void perf_pmu_disable(struct pmu
*pmu
)
844 int *count
= this_cpu_ptr(pmu
->pmu_disable_count
);
846 pmu
->pmu_disable(pmu
);
849 void perf_pmu_enable(struct pmu
*pmu
)
851 int *count
= this_cpu_ptr(pmu
->pmu_disable_count
);
853 pmu
->pmu_enable(pmu
);
856 static DEFINE_PER_CPU(struct list_head
, active_ctx_list
);
859 * perf_event_ctx_activate(), perf_event_ctx_deactivate(), and
860 * perf_event_task_tick() are fully serialized because they're strictly cpu
861 * affine and perf_event_ctx{activate,deactivate} are called with IRQs
862 * disabled, while perf_event_task_tick is called from IRQ context.
864 static void perf_event_ctx_activate(struct perf_event_context
*ctx
)
866 struct list_head
*head
= this_cpu_ptr(&active_ctx_list
);
868 WARN_ON(!irqs_disabled());
870 WARN_ON(!list_empty(&ctx
->active_ctx_list
));
872 list_add(&ctx
->active_ctx_list
, head
);
875 static void perf_event_ctx_deactivate(struct perf_event_context
*ctx
)
877 WARN_ON(!irqs_disabled());
879 WARN_ON(list_empty(&ctx
->active_ctx_list
));
881 list_del_init(&ctx
->active_ctx_list
);
884 static void get_ctx(struct perf_event_context
*ctx
)
886 WARN_ON(!atomic_inc_not_zero(&ctx
->refcount
));
889 static void free_ctx(struct rcu_head
*head
)
891 struct perf_event_context
*ctx
;
893 ctx
= container_of(head
, struct perf_event_context
, rcu_head
);
894 kfree(ctx
->task_ctx_data
);
898 static void put_ctx(struct perf_event_context
*ctx
)
900 if (atomic_dec_and_test(&ctx
->refcount
)) {
902 put_ctx(ctx
->parent_ctx
);
904 put_task_struct(ctx
->task
);
905 call_rcu(&ctx
->rcu_head
, free_ctx
);
910 * Because of perf_event::ctx migration in sys_perf_event_open::move_group and
911 * perf_pmu_migrate_context() we need some magic.
913 * Those places that change perf_event::ctx will hold both
914 * perf_event_ctx::mutex of the 'old' and 'new' ctx value.
916 * Lock ordering is by mutex address. There is one other site where
917 * perf_event_context::mutex nests and that is put_event(). But remember that
918 * that is a parent<->child context relation, and migration does not affect
919 * children, therefore these two orderings should not interact.
921 * The change in perf_event::ctx does not affect children (as claimed above)
922 * because the sys_perf_event_open() case will install a new event and break
923 * the ctx parent<->child relation, and perf_pmu_migrate_context() is only
924 * concerned with cpuctx and that doesn't have children.
926 * The places that change perf_event::ctx will issue:
928 * perf_remove_from_context();
930 * perf_install_in_context();
932 * to affect the change. The remove_from_context() + synchronize_rcu() should
933 * quiesce the event, after which we can install it in the new location. This
934 * means that only external vectors (perf_fops, prctl) can perturb the event
935 * while in transit. Therefore all such accessors should also acquire
936 * perf_event_context::mutex to serialize against this.
938 * However; because event->ctx can change while we're waiting to acquire
939 * ctx->mutex we must be careful and use the below perf_event_ctx_lock()
943 * task_struct::perf_event_mutex
944 * perf_event_context::mutex
945 * perf_event_context::lock
946 * perf_event::child_mutex;
947 * perf_event::mmap_mutex
950 static struct perf_event_context
*
951 perf_event_ctx_lock_nested(struct perf_event
*event
, int nesting
)
953 struct perf_event_context
*ctx
;
957 ctx
= ACCESS_ONCE(event
->ctx
);
958 if (!atomic_inc_not_zero(&ctx
->refcount
)) {
964 mutex_lock_nested(&ctx
->mutex
, nesting
);
965 if (event
->ctx
!= ctx
) {
966 mutex_unlock(&ctx
->mutex
);
974 static inline struct perf_event_context
*
975 perf_event_ctx_lock(struct perf_event
*event
)
977 return perf_event_ctx_lock_nested(event
, 0);
980 static void perf_event_ctx_unlock(struct perf_event
*event
,
981 struct perf_event_context
*ctx
)
983 mutex_unlock(&ctx
->mutex
);
988 * This must be done under the ctx->lock, such as to serialize against
989 * context_equiv(), therefore we cannot call put_ctx() since that might end up
990 * calling scheduler related locks and ctx->lock nests inside those.
992 static __must_check
struct perf_event_context
*
993 unclone_ctx(struct perf_event_context
*ctx
)
995 struct perf_event_context
*parent_ctx
= ctx
->parent_ctx
;
997 lockdep_assert_held(&ctx
->lock
);
1000 ctx
->parent_ctx
= NULL
;
1006 static u32
perf_event_pid(struct perf_event
*event
, struct task_struct
*p
)
1009 * only top level events have the pid namespace they were created in
1012 event
= event
->parent
;
1014 return task_tgid_nr_ns(p
, event
->ns
);
1017 static u32
perf_event_tid(struct perf_event
*event
, struct task_struct
*p
)
1020 * only top level events have the pid namespace they were created in
1023 event
= event
->parent
;
1025 return task_pid_nr_ns(p
, event
->ns
);
1029 * If we inherit events we want to return the parent event id
1032 static u64
primary_event_id(struct perf_event
*event
)
1037 id
= event
->parent
->id
;
1043 * Get the perf_event_context for a task and lock it.
1044 * This has to cope with with the fact that until it is locked,
1045 * the context could get moved to another task.
1047 static struct perf_event_context
*
1048 perf_lock_task_context(struct task_struct
*task
, int ctxn
, unsigned long *flags
)
1050 struct perf_event_context
*ctx
;
1054 * One of the few rules of preemptible RCU is that one cannot do
1055 * rcu_read_unlock() while holding a scheduler (or nested) lock when
1056 * part of the read side critical section was preemptible -- see
1057 * rcu_read_unlock_special().
1059 * Since ctx->lock nests under rq->lock we must ensure the entire read
1060 * side critical section is non-preemptible.
1064 ctx
= rcu_dereference(task
->perf_event_ctxp
[ctxn
]);
1067 * If this context is a clone of another, it might
1068 * get swapped for another underneath us by
1069 * perf_event_task_sched_out, though the
1070 * rcu_read_lock() protects us from any context
1071 * getting freed. Lock the context and check if it
1072 * got swapped before we could get the lock, and retry
1073 * if so. If we locked the right context, then it
1074 * can't get swapped on us any more.
1076 raw_spin_lock_irqsave(&ctx
->lock
, *flags
);
1077 if (ctx
!= rcu_dereference(task
->perf_event_ctxp
[ctxn
])) {
1078 raw_spin_unlock_irqrestore(&ctx
->lock
, *flags
);
1084 if (!atomic_inc_not_zero(&ctx
->refcount
)) {
1085 raw_spin_unlock_irqrestore(&ctx
->lock
, *flags
);
1095 * Get the context for a task and increment its pin_count so it
1096 * can't get swapped to another task. This also increments its
1097 * reference count so that the context can't get freed.
1099 static struct perf_event_context
*
1100 perf_pin_task_context(struct task_struct
*task
, int ctxn
)
1102 struct perf_event_context
*ctx
;
1103 unsigned long flags
;
1105 ctx
= perf_lock_task_context(task
, ctxn
, &flags
);
1108 raw_spin_unlock_irqrestore(&ctx
->lock
, flags
);
1113 static void perf_unpin_context(struct perf_event_context
*ctx
)
1115 unsigned long flags
;
1117 raw_spin_lock_irqsave(&ctx
->lock
, flags
);
1119 raw_spin_unlock_irqrestore(&ctx
->lock
, flags
);
1123 * Update the record of the current time in a context.
1125 static void update_context_time(struct perf_event_context
*ctx
)
1127 u64 now
= perf_clock();
1129 ctx
->time
+= now
- ctx
->timestamp
;
1130 ctx
->timestamp
= now
;
1133 static u64
perf_event_time(struct perf_event
*event
)
1135 struct perf_event_context
*ctx
= event
->ctx
;
1137 if (is_cgroup_event(event
))
1138 return perf_cgroup_event_time(event
);
1140 return ctx
? ctx
->time
: 0;
1144 * Update the total_time_enabled and total_time_running fields for a event.
1145 * The caller of this function needs to hold the ctx->lock.
1147 static void update_event_times(struct perf_event
*event
)
1149 struct perf_event_context
*ctx
= event
->ctx
;
1152 if (event
->state
< PERF_EVENT_STATE_INACTIVE
||
1153 event
->group_leader
->state
< PERF_EVENT_STATE_INACTIVE
)
1156 * in cgroup mode, time_enabled represents
1157 * the time the event was enabled AND active
1158 * tasks were in the monitored cgroup. This is
1159 * independent of the activity of the context as
1160 * there may be a mix of cgroup and non-cgroup events.
1162 * That is why we treat cgroup events differently
1165 if (is_cgroup_event(event
))
1166 run_end
= perf_cgroup_event_time(event
);
1167 else if (ctx
->is_active
)
1168 run_end
= ctx
->time
;
1170 run_end
= event
->tstamp_stopped
;
1172 event
->total_time_enabled
= run_end
- event
->tstamp_enabled
;
1174 if (event
->state
== PERF_EVENT_STATE_INACTIVE
)
1175 run_end
= event
->tstamp_stopped
;
1177 run_end
= perf_event_time(event
);
1179 event
->total_time_running
= run_end
- event
->tstamp_running
;
1184 * Update total_time_enabled and total_time_running for all events in a group.
1186 static void update_group_times(struct perf_event
*leader
)
1188 struct perf_event
*event
;
1190 update_event_times(leader
);
1191 list_for_each_entry(event
, &leader
->sibling_list
, group_entry
)
1192 update_event_times(event
);
1195 static struct list_head
*
1196 ctx_group_list(struct perf_event
*event
, struct perf_event_context
*ctx
)
1198 if (event
->attr
.pinned
)
1199 return &ctx
->pinned_groups
;
1201 return &ctx
->flexible_groups
;
1205 * Add a event from the lists for its context.
1206 * Must be called with ctx->mutex and ctx->lock held.
1209 list_add_event(struct perf_event
*event
, struct perf_event_context
*ctx
)
1211 WARN_ON_ONCE(event
->attach_state
& PERF_ATTACH_CONTEXT
);
1212 event
->attach_state
|= PERF_ATTACH_CONTEXT
;
1215 * If we're a stand alone event or group leader, we go to the context
1216 * list, group events are kept attached to the group so that
1217 * perf_group_detach can, at all times, locate all siblings.
1219 if (event
->group_leader
== event
) {
1220 struct list_head
*list
;
1222 if (is_software_event(event
))
1223 event
->group_flags
|= PERF_GROUP_SOFTWARE
;
1225 list
= ctx_group_list(event
, ctx
);
1226 list_add_tail(&event
->group_entry
, list
);
1229 if (is_cgroup_event(event
))
1232 list_add_rcu(&event
->event_entry
, &ctx
->event_list
);
1234 if (event
->attr
.inherit_stat
)
1241 * Initialize event state based on the perf_event_attr::disabled.
1243 static inline void perf_event__state_init(struct perf_event
*event
)
1245 event
->state
= event
->attr
.disabled
? PERF_EVENT_STATE_OFF
:
1246 PERF_EVENT_STATE_INACTIVE
;
1250 * Called at perf_event creation and when events are attached/detached from a
1253 static void perf_event__read_size(struct perf_event
*event
)
1255 int entry
= sizeof(u64
); /* value */
1259 if (event
->attr
.read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
1260 size
+= sizeof(u64
);
1262 if (event
->attr
.read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
1263 size
+= sizeof(u64
);
1265 if (event
->attr
.read_format
& PERF_FORMAT_ID
)
1266 entry
+= sizeof(u64
);
1268 if (event
->attr
.read_format
& PERF_FORMAT_GROUP
) {
1269 nr
+= event
->group_leader
->nr_siblings
;
1270 size
+= sizeof(u64
);
1274 event
->read_size
= size
;
1277 static void perf_event__header_size(struct perf_event
*event
)
1279 struct perf_sample_data
*data
;
1280 u64 sample_type
= event
->attr
.sample_type
;
1283 perf_event__read_size(event
);
1285 if (sample_type
& PERF_SAMPLE_IP
)
1286 size
+= sizeof(data
->ip
);
1288 if (sample_type
& PERF_SAMPLE_ADDR
)
1289 size
+= sizeof(data
->addr
);
1291 if (sample_type
& PERF_SAMPLE_PERIOD
)
1292 size
+= sizeof(data
->period
);
1294 if (sample_type
& PERF_SAMPLE_WEIGHT
)
1295 size
+= sizeof(data
->weight
);
1297 if (sample_type
& PERF_SAMPLE_READ
)
1298 size
+= event
->read_size
;
1300 if (sample_type
& PERF_SAMPLE_DATA_SRC
)
1301 size
+= sizeof(data
->data_src
.val
);
1303 if (sample_type
& PERF_SAMPLE_TRANSACTION
)
1304 size
+= sizeof(data
->txn
);
1306 event
->header_size
= size
;
1309 static void perf_event__id_header_size(struct perf_event
*event
)
1311 struct perf_sample_data
*data
;
1312 u64 sample_type
= event
->attr
.sample_type
;
1315 if (sample_type
& PERF_SAMPLE_TID
)
1316 size
+= sizeof(data
->tid_entry
);
1318 if (sample_type
& PERF_SAMPLE_TIME
)
1319 size
+= sizeof(data
->time
);
1321 if (sample_type
& PERF_SAMPLE_IDENTIFIER
)
1322 size
+= sizeof(data
->id
);
1324 if (sample_type
& PERF_SAMPLE_ID
)
1325 size
+= sizeof(data
->id
);
1327 if (sample_type
& PERF_SAMPLE_STREAM_ID
)
1328 size
+= sizeof(data
->stream_id
);
1330 if (sample_type
& PERF_SAMPLE_CPU
)
1331 size
+= sizeof(data
->cpu_entry
);
1333 event
->id_header_size
= size
;
1336 static void perf_group_attach(struct perf_event
*event
)
1338 struct perf_event
*group_leader
= event
->group_leader
, *pos
;
1341 * We can have double attach due to group movement in perf_event_open.
1343 if (event
->attach_state
& PERF_ATTACH_GROUP
)
1346 event
->attach_state
|= PERF_ATTACH_GROUP
;
1348 if (group_leader
== event
)
1351 WARN_ON_ONCE(group_leader
->ctx
!= event
->ctx
);
1353 if (group_leader
->group_flags
& PERF_GROUP_SOFTWARE
&&
1354 !is_software_event(event
))
1355 group_leader
->group_flags
&= ~PERF_GROUP_SOFTWARE
;
1357 list_add_tail(&event
->group_entry
, &group_leader
->sibling_list
);
1358 group_leader
->nr_siblings
++;
1360 perf_event__header_size(group_leader
);
1362 list_for_each_entry(pos
, &group_leader
->sibling_list
, group_entry
)
1363 perf_event__header_size(pos
);
1367 * Remove a event from the lists for its context.
1368 * Must be called with ctx->mutex and ctx->lock held.
1371 list_del_event(struct perf_event
*event
, struct perf_event_context
*ctx
)
1373 struct perf_cpu_context
*cpuctx
;
1375 WARN_ON_ONCE(event
->ctx
!= ctx
);
1376 lockdep_assert_held(&ctx
->lock
);
1379 * We can have double detach due to exit/hot-unplug + close.
1381 if (!(event
->attach_state
& PERF_ATTACH_CONTEXT
))
1384 event
->attach_state
&= ~PERF_ATTACH_CONTEXT
;
1386 if (is_cgroup_event(event
)) {
1388 cpuctx
= __get_cpu_context(ctx
);
1390 * if there are no more cgroup events
1391 * then cler cgrp to avoid stale pointer
1392 * in update_cgrp_time_from_cpuctx()
1394 if (!ctx
->nr_cgroups
)
1395 cpuctx
->cgrp
= NULL
;
1399 if (event
->attr
.inherit_stat
)
1402 list_del_rcu(&event
->event_entry
);
1404 if (event
->group_leader
== event
)
1405 list_del_init(&event
->group_entry
);
1407 update_group_times(event
);
1410 * If event was in error state, then keep it
1411 * that way, otherwise bogus counts will be
1412 * returned on read(). The only way to get out
1413 * of error state is by explicit re-enabling
1416 if (event
->state
> PERF_EVENT_STATE_OFF
)
1417 event
->state
= PERF_EVENT_STATE_OFF
;
1422 static void perf_group_detach(struct perf_event
*event
)
1424 struct perf_event
*sibling
, *tmp
;
1425 struct list_head
*list
= NULL
;
1428 * We can have double detach due to exit/hot-unplug + close.
1430 if (!(event
->attach_state
& PERF_ATTACH_GROUP
))
1433 event
->attach_state
&= ~PERF_ATTACH_GROUP
;
1436 * If this is a sibling, remove it from its group.
1438 if (event
->group_leader
!= event
) {
1439 list_del_init(&event
->group_entry
);
1440 event
->group_leader
->nr_siblings
--;
1444 if (!list_empty(&event
->group_entry
))
1445 list
= &event
->group_entry
;
1448 * If this was a group event with sibling events then
1449 * upgrade the siblings to singleton events by adding them
1450 * to whatever list we are on.
1452 list_for_each_entry_safe(sibling
, tmp
, &event
->sibling_list
, group_entry
) {
1454 list_move_tail(&sibling
->group_entry
, list
);
1455 sibling
->group_leader
= sibling
;
1457 /* Inherit group flags from the previous leader */
1458 sibling
->group_flags
= event
->group_flags
;
1460 WARN_ON_ONCE(sibling
->ctx
!= event
->ctx
);
1464 perf_event__header_size(event
->group_leader
);
1466 list_for_each_entry(tmp
, &event
->group_leader
->sibling_list
, group_entry
)
1467 perf_event__header_size(tmp
);
1471 * User event without the task.
1473 static bool is_orphaned_event(struct perf_event
*event
)
1475 return event
&& !is_kernel_event(event
) && !event
->owner
;
1479 * Event has a parent but parent's task finished and it's
1480 * alive only because of children holding refference.
1482 static bool is_orphaned_child(struct perf_event
*event
)
1484 return is_orphaned_event(event
->parent
);
1487 static void orphans_remove_work(struct work_struct
*work
);
1489 static void schedule_orphans_remove(struct perf_event_context
*ctx
)
1491 if (!ctx
->task
|| ctx
->orphans_remove_sched
|| !perf_wq
)
1494 if (queue_delayed_work(perf_wq
, &ctx
->orphans_remove
, 1)) {
1496 ctx
->orphans_remove_sched
= true;
1500 static int __init
perf_workqueue_init(void)
1502 perf_wq
= create_singlethread_workqueue("perf");
1503 WARN(!perf_wq
, "failed to create perf workqueue\n");
1504 return perf_wq
? 0 : -1;
1507 core_initcall(perf_workqueue_init
);
1510 event_filter_match(struct perf_event
*event
)
1512 return (event
->cpu
== -1 || event
->cpu
== smp_processor_id())
1513 && perf_cgroup_match(event
);
1517 event_sched_out(struct perf_event
*event
,
1518 struct perf_cpu_context
*cpuctx
,
1519 struct perf_event_context
*ctx
)
1521 u64 tstamp
= perf_event_time(event
);
1524 WARN_ON_ONCE(event
->ctx
!= ctx
);
1525 lockdep_assert_held(&ctx
->lock
);
1528 * An event which could not be activated because of
1529 * filter mismatch still needs to have its timings
1530 * maintained, otherwise bogus information is return
1531 * via read() for time_enabled, time_running:
1533 if (event
->state
== PERF_EVENT_STATE_INACTIVE
1534 && !event_filter_match(event
)) {
1535 delta
= tstamp
- event
->tstamp_stopped
;
1536 event
->tstamp_running
+= delta
;
1537 event
->tstamp_stopped
= tstamp
;
1540 if (event
->state
!= PERF_EVENT_STATE_ACTIVE
)
1543 perf_pmu_disable(event
->pmu
);
1545 event
->state
= PERF_EVENT_STATE_INACTIVE
;
1546 if (event
->pending_disable
) {
1547 event
->pending_disable
= 0;
1548 event
->state
= PERF_EVENT_STATE_OFF
;
1550 event
->tstamp_stopped
= tstamp
;
1551 event
->pmu
->del(event
, 0);
1554 if (!is_software_event(event
))
1555 cpuctx
->active_oncpu
--;
1556 if (!--ctx
->nr_active
)
1557 perf_event_ctx_deactivate(ctx
);
1558 if (event
->attr
.freq
&& event
->attr
.sample_freq
)
1560 if (event
->attr
.exclusive
|| !cpuctx
->active_oncpu
)
1561 cpuctx
->exclusive
= 0;
1563 if (is_orphaned_child(event
))
1564 schedule_orphans_remove(ctx
);
1566 perf_pmu_enable(event
->pmu
);
1570 group_sched_out(struct perf_event
*group_event
,
1571 struct perf_cpu_context
*cpuctx
,
1572 struct perf_event_context
*ctx
)
1574 struct perf_event
*event
;
1575 int state
= group_event
->state
;
1577 event_sched_out(group_event
, cpuctx
, ctx
);
1580 * Schedule out siblings (if any):
1582 list_for_each_entry(event
, &group_event
->sibling_list
, group_entry
)
1583 event_sched_out(event
, cpuctx
, ctx
);
1585 if (state
== PERF_EVENT_STATE_ACTIVE
&& group_event
->attr
.exclusive
)
1586 cpuctx
->exclusive
= 0;
1589 struct remove_event
{
1590 struct perf_event
*event
;
1595 * Cross CPU call to remove a performance event
1597 * We disable the event on the hardware level first. After that we
1598 * remove it from the context list.
1600 static int __perf_remove_from_context(void *info
)
1602 struct remove_event
*re
= info
;
1603 struct perf_event
*event
= re
->event
;
1604 struct perf_event_context
*ctx
= event
->ctx
;
1605 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
1607 raw_spin_lock(&ctx
->lock
);
1608 event_sched_out(event
, cpuctx
, ctx
);
1609 if (re
->detach_group
)
1610 perf_group_detach(event
);
1611 list_del_event(event
, ctx
);
1612 if (!ctx
->nr_events
&& cpuctx
->task_ctx
== ctx
) {
1614 cpuctx
->task_ctx
= NULL
;
1616 raw_spin_unlock(&ctx
->lock
);
1623 * Remove the event from a task's (or a CPU's) list of events.
1625 * CPU events are removed with a smp call. For task events we only
1626 * call when the task is on a CPU.
1628 * If event->ctx is a cloned context, callers must make sure that
1629 * every task struct that event->ctx->task could possibly point to
1630 * remains valid. This is OK when called from perf_release since
1631 * that only calls us on the top-level context, which can't be a clone.
1632 * When called from perf_event_exit_task, it's OK because the
1633 * context has been detached from its task.
1635 static void perf_remove_from_context(struct perf_event
*event
, bool detach_group
)
1637 struct perf_event_context
*ctx
= event
->ctx
;
1638 struct task_struct
*task
= ctx
->task
;
1639 struct remove_event re
= {
1641 .detach_group
= detach_group
,
1644 lockdep_assert_held(&ctx
->mutex
);
1648 * Per cpu events are removed via an smp call. The removal can
1649 * fail if the CPU is currently offline, but in that case we
1650 * already called __perf_remove_from_context from
1651 * perf_event_exit_cpu.
1653 cpu_function_call(event
->cpu
, __perf_remove_from_context
, &re
);
1658 if (!task_function_call(task
, __perf_remove_from_context
, &re
))
1661 raw_spin_lock_irq(&ctx
->lock
);
1663 * If we failed to find a running task, but find the context active now
1664 * that we've acquired the ctx->lock, retry.
1666 if (ctx
->is_active
) {
1667 raw_spin_unlock_irq(&ctx
->lock
);
1669 * Reload the task pointer, it might have been changed by
1670 * a concurrent perf_event_context_sched_out().
1677 * Since the task isn't running, its safe to remove the event, us
1678 * holding the ctx->lock ensures the task won't get scheduled in.
1681 perf_group_detach(event
);
1682 list_del_event(event
, ctx
);
1683 raw_spin_unlock_irq(&ctx
->lock
);
1687 * Cross CPU call to disable a performance event
1689 int __perf_event_disable(void *info
)
1691 struct perf_event
*event
= info
;
1692 struct perf_event_context
*ctx
= event
->ctx
;
1693 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
1696 * If this is a per-task event, need to check whether this
1697 * event's task is the current task on this cpu.
1699 * Can trigger due to concurrent perf_event_context_sched_out()
1700 * flipping contexts around.
1702 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
)
1705 raw_spin_lock(&ctx
->lock
);
1708 * If the event is on, turn it off.
1709 * If it is in error state, leave it in error state.
1711 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
) {
1712 update_context_time(ctx
);
1713 update_cgrp_time_from_event(event
);
1714 update_group_times(event
);
1715 if (event
== event
->group_leader
)
1716 group_sched_out(event
, cpuctx
, ctx
);
1718 event_sched_out(event
, cpuctx
, ctx
);
1719 event
->state
= PERF_EVENT_STATE_OFF
;
1722 raw_spin_unlock(&ctx
->lock
);
1730 * If event->ctx is a cloned context, callers must make sure that
1731 * every task struct that event->ctx->task could possibly point to
1732 * remains valid. This condition is satisifed when called through
1733 * perf_event_for_each_child or perf_event_for_each because they
1734 * hold the top-level event's child_mutex, so any descendant that
1735 * goes to exit will block in sync_child_event.
1736 * When called from perf_pending_event it's OK because event->ctx
1737 * is the current context on this CPU and preemption is disabled,
1738 * hence we can't get into perf_event_task_sched_out for this context.
1740 static void _perf_event_disable(struct perf_event
*event
)
1742 struct perf_event_context
*ctx
= event
->ctx
;
1743 struct task_struct
*task
= ctx
->task
;
1747 * Disable the event on the cpu that it's on
1749 cpu_function_call(event
->cpu
, __perf_event_disable
, event
);
1754 if (!task_function_call(task
, __perf_event_disable
, event
))
1757 raw_spin_lock_irq(&ctx
->lock
);
1759 * If the event is still active, we need to retry the cross-call.
1761 if (event
->state
== PERF_EVENT_STATE_ACTIVE
) {
1762 raw_spin_unlock_irq(&ctx
->lock
);
1764 * Reload the task pointer, it might have been changed by
1765 * a concurrent perf_event_context_sched_out().
1772 * Since we have the lock this context can't be scheduled
1773 * in, so we can change the state safely.
1775 if (event
->state
== PERF_EVENT_STATE_INACTIVE
) {
1776 update_group_times(event
);
1777 event
->state
= PERF_EVENT_STATE_OFF
;
1779 raw_spin_unlock_irq(&ctx
->lock
);
1783 * Strictly speaking kernel users cannot create groups and therefore this
1784 * interface does not need the perf_event_ctx_lock() magic.
1786 void perf_event_disable(struct perf_event
*event
)
1788 struct perf_event_context
*ctx
;
1790 ctx
= perf_event_ctx_lock(event
);
1791 _perf_event_disable(event
);
1792 perf_event_ctx_unlock(event
, ctx
);
1794 EXPORT_SYMBOL_GPL(perf_event_disable
);
1796 static void perf_set_shadow_time(struct perf_event
*event
,
1797 struct perf_event_context
*ctx
,
1801 * use the correct time source for the time snapshot
1803 * We could get by without this by leveraging the
1804 * fact that to get to this function, the caller
1805 * has most likely already called update_context_time()
1806 * and update_cgrp_time_xx() and thus both timestamp
1807 * are identical (or very close). Given that tstamp is,
1808 * already adjusted for cgroup, we could say that:
1809 * tstamp - ctx->timestamp
1811 * tstamp - cgrp->timestamp.
1813 * Then, in perf_output_read(), the calculation would
1814 * work with no changes because:
1815 * - event is guaranteed scheduled in
1816 * - no scheduled out in between
1817 * - thus the timestamp would be the same
1819 * But this is a bit hairy.
1821 * So instead, we have an explicit cgroup call to remain
1822 * within the time time source all along. We believe it
1823 * is cleaner and simpler to understand.
1825 if (is_cgroup_event(event
))
1826 perf_cgroup_set_shadow_time(event
, tstamp
);
1828 event
->shadow_ctx_time
= tstamp
- ctx
->timestamp
;
1831 #define MAX_INTERRUPTS (~0ULL)
1833 static void perf_log_throttle(struct perf_event
*event
, int enable
);
1836 event_sched_in(struct perf_event
*event
,
1837 struct perf_cpu_context
*cpuctx
,
1838 struct perf_event_context
*ctx
)
1840 u64 tstamp
= perf_event_time(event
);
1843 lockdep_assert_held(&ctx
->lock
);
1845 if (event
->state
<= PERF_EVENT_STATE_OFF
)
1848 event
->state
= PERF_EVENT_STATE_ACTIVE
;
1849 event
->oncpu
= smp_processor_id();
1852 * Unthrottle events, since we scheduled we might have missed several
1853 * ticks already, also for a heavily scheduling task there is little
1854 * guarantee it'll get a tick in a timely manner.
1856 if (unlikely(event
->hw
.interrupts
== MAX_INTERRUPTS
)) {
1857 perf_log_throttle(event
, 1);
1858 event
->hw
.interrupts
= 0;
1862 * The new state must be visible before we turn it on in the hardware:
1866 perf_pmu_disable(event
->pmu
);
1868 event
->tstamp_running
+= tstamp
- event
->tstamp_stopped
;
1870 perf_set_shadow_time(event
, ctx
, tstamp
);
1872 if (event
->pmu
->add(event
, PERF_EF_START
)) {
1873 event
->state
= PERF_EVENT_STATE_INACTIVE
;
1879 if (!is_software_event(event
))
1880 cpuctx
->active_oncpu
++;
1881 if (!ctx
->nr_active
++)
1882 perf_event_ctx_activate(ctx
);
1883 if (event
->attr
.freq
&& event
->attr
.sample_freq
)
1886 if (event
->attr
.exclusive
)
1887 cpuctx
->exclusive
= 1;
1889 if (is_orphaned_child(event
))
1890 schedule_orphans_remove(ctx
);
1893 perf_pmu_enable(event
->pmu
);
1899 group_sched_in(struct perf_event
*group_event
,
1900 struct perf_cpu_context
*cpuctx
,
1901 struct perf_event_context
*ctx
)
1903 struct perf_event
*event
, *partial_group
= NULL
;
1904 struct pmu
*pmu
= ctx
->pmu
;
1905 u64 now
= ctx
->time
;
1906 bool simulate
= false;
1908 if (group_event
->state
== PERF_EVENT_STATE_OFF
)
1911 pmu
->start_txn(pmu
);
1913 if (event_sched_in(group_event
, cpuctx
, ctx
)) {
1914 pmu
->cancel_txn(pmu
);
1915 perf_cpu_hrtimer_restart(cpuctx
);
1920 * Schedule in siblings as one group (if any):
1922 list_for_each_entry(event
, &group_event
->sibling_list
, group_entry
) {
1923 if (event_sched_in(event
, cpuctx
, ctx
)) {
1924 partial_group
= event
;
1929 if (!pmu
->commit_txn(pmu
))
1934 * Groups can be scheduled in as one unit only, so undo any
1935 * partial group before returning:
1936 * The events up to the failed event are scheduled out normally,
1937 * tstamp_stopped will be updated.
1939 * The failed events and the remaining siblings need to have
1940 * their timings updated as if they had gone thru event_sched_in()
1941 * and event_sched_out(). This is required to get consistent timings
1942 * across the group. This also takes care of the case where the group
1943 * could never be scheduled by ensuring tstamp_stopped is set to mark
1944 * the time the event was actually stopped, such that time delta
1945 * calculation in update_event_times() is correct.
1947 list_for_each_entry(event
, &group_event
->sibling_list
, group_entry
) {
1948 if (event
== partial_group
)
1952 event
->tstamp_running
+= now
- event
->tstamp_stopped
;
1953 event
->tstamp_stopped
= now
;
1955 event_sched_out(event
, cpuctx
, ctx
);
1958 event_sched_out(group_event
, cpuctx
, ctx
);
1960 pmu
->cancel_txn(pmu
);
1962 perf_cpu_hrtimer_restart(cpuctx
);
1968 * Work out whether we can put this event group on the CPU now.
1970 static int group_can_go_on(struct perf_event
*event
,
1971 struct perf_cpu_context
*cpuctx
,
1975 * Groups consisting entirely of software events can always go on.
1977 if (event
->group_flags
& PERF_GROUP_SOFTWARE
)
1980 * If an exclusive group is already on, no other hardware
1983 if (cpuctx
->exclusive
)
1986 * If this group is exclusive and there are already
1987 * events on the CPU, it can't go on.
1989 if (event
->attr
.exclusive
&& cpuctx
->active_oncpu
)
1992 * Otherwise, try to add it if all previous groups were able
1998 static void add_event_to_ctx(struct perf_event
*event
,
1999 struct perf_event_context
*ctx
)
2001 u64 tstamp
= perf_event_time(event
);
2003 list_add_event(event
, ctx
);
2004 perf_group_attach(event
);
2005 event
->tstamp_enabled
= tstamp
;
2006 event
->tstamp_running
= tstamp
;
2007 event
->tstamp_stopped
= tstamp
;
2010 static void task_ctx_sched_out(struct perf_event_context
*ctx
);
2012 ctx_sched_in(struct perf_event_context
*ctx
,
2013 struct perf_cpu_context
*cpuctx
,
2014 enum event_type_t event_type
,
2015 struct task_struct
*task
);
2017 static void perf_event_sched_in(struct perf_cpu_context
*cpuctx
,
2018 struct perf_event_context
*ctx
,
2019 struct task_struct
*task
)
2021 cpu_ctx_sched_in(cpuctx
, EVENT_PINNED
, task
);
2023 ctx_sched_in(ctx
, cpuctx
, EVENT_PINNED
, task
);
2024 cpu_ctx_sched_in(cpuctx
, EVENT_FLEXIBLE
, task
);
2026 ctx_sched_in(ctx
, cpuctx
, EVENT_FLEXIBLE
, task
);
2030 * Cross CPU call to install and enable a performance event
2032 * Must be called with ctx->mutex held
2034 static int __perf_install_in_context(void *info
)
2036 struct perf_event
*event
= info
;
2037 struct perf_event_context
*ctx
= event
->ctx
;
2038 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
2039 struct perf_event_context
*task_ctx
= cpuctx
->task_ctx
;
2040 struct task_struct
*task
= current
;
2042 perf_ctx_lock(cpuctx
, task_ctx
);
2043 perf_pmu_disable(cpuctx
->ctx
.pmu
);
2046 * If there was an active task_ctx schedule it out.
2049 task_ctx_sched_out(task_ctx
);
2052 * If the context we're installing events in is not the
2053 * active task_ctx, flip them.
2055 if (ctx
->task
&& task_ctx
!= ctx
) {
2057 raw_spin_unlock(&task_ctx
->lock
);
2058 raw_spin_lock(&ctx
->lock
);
2063 cpuctx
->task_ctx
= task_ctx
;
2064 task
= task_ctx
->task
;
2067 cpu_ctx_sched_out(cpuctx
, EVENT_ALL
);
2069 update_context_time(ctx
);
2071 * update cgrp time only if current cgrp
2072 * matches event->cgrp. Must be done before
2073 * calling add_event_to_ctx()
2075 update_cgrp_time_from_event(event
);
2077 add_event_to_ctx(event
, ctx
);
2080 * Schedule everything back in
2082 perf_event_sched_in(cpuctx
, task_ctx
, task
);
2084 perf_pmu_enable(cpuctx
->ctx
.pmu
);
2085 perf_ctx_unlock(cpuctx
, task_ctx
);
2091 * Attach a performance event to a context
2093 * First we add the event to the list with the hardware enable bit
2094 * in event->hw_config cleared.
2096 * If the event is attached to a task which is on a CPU we use a smp
2097 * call to enable it in the task context. The task might have been
2098 * scheduled away, but we check this in the smp call again.
2101 perf_install_in_context(struct perf_event_context
*ctx
,
2102 struct perf_event
*event
,
2105 struct task_struct
*task
= ctx
->task
;
2107 lockdep_assert_held(&ctx
->mutex
);
2110 if (event
->cpu
!= -1)
2115 * Per cpu events are installed via an smp call and
2116 * the install is always successful.
2118 cpu_function_call(cpu
, __perf_install_in_context
, event
);
2123 if (!task_function_call(task
, __perf_install_in_context
, event
))
2126 raw_spin_lock_irq(&ctx
->lock
);
2128 * If we failed to find a running task, but find the context active now
2129 * that we've acquired the ctx->lock, retry.
2131 if (ctx
->is_active
) {
2132 raw_spin_unlock_irq(&ctx
->lock
);
2134 * Reload the task pointer, it might have been changed by
2135 * a concurrent perf_event_context_sched_out().
2142 * Since the task isn't running, its safe to add the event, us holding
2143 * the ctx->lock ensures the task won't get scheduled in.
2145 add_event_to_ctx(event
, ctx
);
2146 raw_spin_unlock_irq(&ctx
->lock
);
2150 * Put a event into inactive state and update time fields.
2151 * Enabling the leader of a group effectively enables all
2152 * the group members that aren't explicitly disabled, so we
2153 * have to update their ->tstamp_enabled also.
2154 * Note: this works for group members as well as group leaders
2155 * since the non-leader members' sibling_lists will be empty.
2157 static void __perf_event_mark_enabled(struct perf_event
*event
)
2159 struct perf_event
*sub
;
2160 u64 tstamp
= perf_event_time(event
);
2162 event
->state
= PERF_EVENT_STATE_INACTIVE
;
2163 event
->tstamp_enabled
= tstamp
- event
->total_time_enabled
;
2164 list_for_each_entry(sub
, &event
->sibling_list
, group_entry
) {
2165 if (sub
->state
>= PERF_EVENT_STATE_INACTIVE
)
2166 sub
->tstamp_enabled
= tstamp
- sub
->total_time_enabled
;
2171 * Cross CPU call to enable a performance event
2173 static int __perf_event_enable(void *info
)
2175 struct perf_event
*event
= info
;
2176 struct perf_event_context
*ctx
= event
->ctx
;
2177 struct perf_event
*leader
= event
->group_leader
;
2178 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
2182 * There's a time window between 'ctx->is_active' check
2183 * in perf_event_enable function and this place having:
2185 * - ctx->lock unlocked
2187 * where the task could be killed and 'ctx' deactivated
2188 * by perf_event_exit_task.
2190 if (!ctx
->is_active
)
2193 raw_spin_lock(&ctx
->lock
);
2194 update_context_time(ctx
);
2196 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
)
2200 * set current task's cgroup time reference point
2202 perf_cgroup_set_timestamp(current
, ctx
);
2204 __perf_event_mark_enabled(event
);
2206 if (!event_filter_match(event
)) {
2207 if (is_cgroup_event(event
))
2208 perf_cgroup_defer_enabled(event
);
2213 * If the event is in a group and isn't the group leader,
2214 * then don't put it on unless the group is on.
2216 if (leader
!= event
&& leader
->state
!= PERF_EVENT_STATE_ACTIVE
)
2219 if (!group_can_go_on(event
, cpuctx
, 1)) {
2222 if (event
== leader
)
2223 err
= group_sched_in(event
, cpuctx
, ctx
);
2225 err
= event_sched_in(event
, cpuctx
, ctx
);
2230 * If this event can't go on and it's part of a
2231 * group, then the whole group has to come off.
2233 if (leader
!= event
) {
2234 group_sched_out(leader
, cpuctx
, ctx
);
2235 perf_cpu_hrtimer_restart(cpuctx
);
2237 if (leader
->attr
.pinned
) {
2238 update_group_times(leader
);
2239 leader
->state
= PERF_EVENT_STATE_ERROR
;
2244 raw_spin_unlock(&ctx
->lock
);
2252 * If event->ctx is a cloned context, callers must make sure that
2253 * every task struct that event->ctx->task could possibly point to
2254 * remains valid. This condition is satisfied when called through
2255 * perf_event_for_each_child or perf_event_for_each as described
2256 * for perf_event_disable.
2258 static void _perf_event_enable(struct perf_event
*event
)
2260 struct perf_event_context
*ctx
= event
->ctx
;
2261 struct task_struct
*task
= ctx
->task
;
2265 * Enable the event on the cpu that it's on
2267 cpu_function_call(event
->cpu
, __perf_event_enable
, event
);
2271 raw_spin_lock_irq(&ctx
->lock
);
2272 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
)
2276 * If the event is in error state, clear that first.
2277 * That way, if we see the event in error state below, we
2278 * know that it has gone back into error state, as distinct
2279 * from the task having been scheduled away before the
2280 * cross-call arrived.
2282 if (event
->state
== PERF_EVENT_STATE_ERROR
)
2283 event
->state
= PERF_EVENT_STATE_OFF
;
2286 if (!ctx
->is_active
) {
2287 __perf_event_mark_enabled(event
);
2291 raw_spin_unlock_irq(&ctx
->lock
);
2293 if (!task_function_call(task
, __perf_event_enable
, event
))
2296 raw_spin_lock_irq(&ctx
->lock
);
2299 * If the context is active and the event is still off,
2300 * we need to retry the cross-call.
2302 if (ctx
->is_active
&& event
->state
== PERF_EVENT_STATE_OFF
) {
2304 * task could have been flipped by a concurrent
2305 * perf_event_context_sched_out()
2312 raw_spin_unlock_irq(&ctx
->lock
);
2316 * See perf_event_disable();
2318 void perf_event_enable(struct perf_event
*event
)
2320 struct perf_event_context
*ctx
;
2322 ctx
= perf_event_ctx_lock(event
);
2323 _perf_event_enable(event
);
2324 perf_event_ctx_unlock(event
, ctx
);
2326 EXPORT_SYMBOL_GPL(perf_event_enable
);
2328 static int _perf_event_refresh(struct perf_event
*event
, int refresh
)
2331 * not supported on inherited events
2333 if (event
->attr
.inherit
|| !is_sampling_event(event
))
2336 atomic_add(refresh
, &event
->event_limit
);
2337 _perf_event_enable(event
);
2343 * See perf_event_disable()
2345 int perf_event_refresh(struct perf_event
*event
, int refresh
)
2347 struct perf_event_context
*ctx
;
2350 ctx
= perf_event_ctx_lock(event
);
2351 ret
= _perf_event_refresh(event
, refresh
);
2352 perf_event_ctx_unlock(event
, ctx
);
2356 EXPORT_SYMBOL_GPL(perf_event_refresh
);
2358 static void ctx_sched_out(struct perf_event_context
*ctx
,
2359 struct perf_cpu_context
*cpuctx
,
2360 enum event_type_t event_type
)
2362 struct perf_event
*event
;
2363 int is_active
= ctx
->is_active
;
2365 ctx
->is_active
&= ~event_type
;
2366 if (likely(!ctx
->nr_events
))
2369 update_context_time(ctx
);
2370 update_cgrp_time_from_cpuctx(cpuctx
);
2371 if (!ctx
->nr_active
)
2374 perf_pmu_disable(ctx
->pmu
);
2375 if ((is_active
& EVENT_PINNED
) && (event_type
& EVENT_PINNED
)) {
2376 list_for_each_entry(event
, &ctx
->pinned_groups
, group_entry
)
2377 group_sched_out(event
, cpuctx
, ctx
);
2380 if ((is_active
& EVENT_FLEXIBLE
) && (event_type
& EVENT_FLEXIBLE
)) {
2381 list_for_each_entry(event
, &ctx
->flexible_groups
, group_entry
)
2382 group_sched_out(event
, cpuctx
, ctx
);
2384 perf_pmu_enable(ctx
->pmu
);
2388 * Test whether two contexts are equivalent, i.e. whether they have both been
2389 * cloned from the same version of the same context.
2391 * Equivalence is measured using a generation number in the context that is
2392 * incremented on each modification to it; see unclone_ctx(), list_add_event()
2393 * and list_del_event().
2395 static int context_equiv(struct perf_event_context
*ctx1
,
2396 struct perf_event_context
*ctx2
)
2398 lockdep_assert_held(&ctx1
->lock
);
2399 lockdep_assert_held(&ctx2
->lock
);
2401 /* Pinning disables the swap optimization */
2402 if (ctx1
->pin_count
|| ctx2
->pin_count
)
2405 /* If ctx1 is the parent of ctx2 */
2406 if (ctx1
== ctx2
->parent_ctx
&& ctx1
->generation
== ctx2
->parent_gen
)
2409 /* If ctx2 is the parent of ctx1 */
2410 if (ctx1
->parent_ctx
== ctx2
&& ctx1
->parent_gen
== ctx2
->generation
)
2414 * If ctx1 and ctx2 have the same parent; we flatten the parent
2415 * hierarchy, see perf_event_init_context().
2417 if (ctx1
->parent_ctx
&& ctx1
->parent_ctx
== ctx2
->parent_ctx
&&
2418 ctx1
->parent_gen
== ctx2
->parent_gen
)
2425 static void __perf_event_sync_stat(struct perf_event
*event
,
2426 struct perf_event
*next_event
)
2430 if (!event
->attr
.inherit_stat
)
2434 * Update the event value, we cannot use perf_event_read()
2435 * because we're in the middle of a context switch and have IRQs
2436 * disabled, which upsets smp_call_function_single(), however
2437 * we know the event must be on the current CPU, therefore we
2438 * don't need to use it.
2440 switch (event
->state
) {
2441 case PERF_EVENT_STATE_ACTIVE
:
2442 event
->pmu
->read(event
);
2445 case PERF_EVENT_STATE_INACTIVE
:
2446 update_event_times(event
);
2454 * In order to keep per-task stats reliable we need to flip the event
2455 * values when we flip the contexts.
2457 value
= local64_read(&next_event
->count
);
2458 value
= local64_xchg(&event
->count
, value
);
2459 local64_set(&next_event
->count
, value
);
2461 swap(event
->total_time_enabled
, next_event
->total_time_enabled
);
2462 swap(event
->total_time_running
, next_event
->total_time_running
);
2465 * Since we swizzled the values, update the user visible data too.
2467 perf_event_update_userpage(event
);
2468 perf_event_update_userpage(next_event
);
2471 static void perf_event_sync_stat(struct perf_event_context
*ctx
,
2472 struct perf_event_context
*next_ctx
)
2474 struct perf_event
*event
, *next_event
;
2479 update_context_time(ctx
);
2481 event
= list_first_entry(&ctx
->event_list
,
2482 struct perf_event
, event_entry
);
2484 next_event
= list_first_entry(&next_ctx
->event_list
,
2485 struct perf_event
, event_entry
);
2487 while (&event
->event_entry
!= &ctx
->event_list
&&
2488 &next_event
->event_entry
!= &next_ctx
->event_list
) {
2490 __perf_event_sync_stat(event
, next_event
);
2492 event
= list_next_entry(event
, event_entry
);
2493 next_event
= list_next_entry(next_event
, event_entry
);
2497 static void perf_event_context_sched_out(struct task_struct
*task
, int ctxn
,
2498 struct task_struct
*next
)
2500 struct perf_event_context
*ctx
= task
->perf_event_ctxp
[ctxn
];
2501 struct perf_event_context
*next_ctx
;
2502 struct perf_event_context
*parent
, *next_parent
;
2503 struct perf_cpu_context
*cpuctx
;
2509 cpuctx
= __get_cpu_context(ctx
);
2510 if (!cpuctx
->task_ctx
)
2514 next_ctx
= next
->perf_event_ctxp
[ctxn
];
2518 parent
= rcu_dereference(ctx
->parent_ctx
);
2519 next_parent
= rcu_dereference(next_ctx
->parent_ctx
);
2521 /* If neither context have a parent context; they cannot be clones. */
2522 if (!parent
&& !next_parent
)
2525 if (next_parent
== ctx
|| next_ctx
== parent
|| next_parent
== parent
) {
2527 * Looks like the two contexts are clones, so we might be
2528 * able to optimize the context switch. We lock both
2529 * contexts and check that they are clones under the
2530 * lock (including re-checking that neither has been
2531 * uncloned in the meantime). It doesn't matter which
2532 * order we take the locks because no other cpu could
2533 * be trying to lock both of these tasks.
2535 raw_spin_lock(&ctx
->lock
);
2536 raw_spin_lock_nested(&next_ctx
->lock
, SINGLE_DEPTH_NESTING
);
2537 if (context_equiv(ctx
, next_ctx
)) {
2539 * XXX do we need a memory barrier of sorts
2540 * wrt to rcu_dereference() of perf_event_ctxp
2542 task
->perf_event_ctxp
[ctxn
] = next_ctx
;
2543 next
->perf_event_ctxp
[ctxn
] = ctx
;
2545 next_ctx
->task
= task
;
2547 swap(ctx
->task_ctx_data
, next_ctx
->task_ctx_data
);
2551 perf_event_sync_stat(ctx
, next_ctx
);
2553 raw_spin_unlock(&next_ctx
->lock
);
2554 raw_spin_unlock(&ctx
->lock
);
2560 raw_spin_lock(&ctx
->lock
);
2561 ctx_sched_out(ctx
, cpuctx
, EVENT_ALL
);
2562 cpuctx
->task_ctx
= NULL
;
2563 raw_spin_unlock(&ctx
->lock
);
2567 void perf_sched_cb_dec(struct pmu
*pmu
)
2569 this_cpu_dec(perf_sched_cb_usages
);
2572 void perf_sched_cb_inc(struct pmu
*pmu
)
2574 this_cpu_inc(perf_sched_cb_usages
);
2578 * This function provides the context switch callback to the lower code
2579 * layer. It is invoked ONLY when the context switch callback is enabled.
2581 static void perf_pmu_sched_task(struct task_struct
*prev
,
2582 struct task_struct
*next
,
2585 struct perf_cpu_context
*cpuctx
;
2587 unsigned long flags
;
2592 local_irq_save(flags
);
2596 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
2597 if (pmu
->sched_task
) {
2598 cpuctx
= this_cpu_ptr(pmu
->pmu_cpu_context
);
2600 perf_ctx_lock(cpuctx
, cpuctx
->task_ctx
);
2602 perf_pmu_disable(pmu
);
2604 pmu
->sched_task(cpuctx
->task_ctx
, sched_in
);
2606 perf_pmu_enable(pmu
);
2608 perf_ctx_unlock(cpuctx
, cpuctx
->task_ctx
);
2614 local_irq_restore(flags
);
2617 #define for_each_task_context_nr(ctxn) \
2618 for ((ctxn) = 0; (ctxn) < perf_nr_task_contexts; (ctxn)++)
2621 * Called from scheduler to remove the events of the current task,
2622 * with interrupts disabled.
2624 * We stop each event and update the event value in event->count.
2626 * This does not protect us against NMI, but disable()
2627 * sets the disabled bit in the control field of event _before_
2628 * accessing the event control register. If a NMI hits, then it will
2629 * not restart the event.
2631 void __perf_event_task_sched_out(struct task_struct
*task
,
2632 struct task_struct
*next
)
2636 if (__this_cpu_read(perf_sched_cb_usages
))
2637 perf_pmu_sched_task(task
, next
, false);
2639 for_each_task_context_nr(ctxn
)
2640 perf_event_context_sched_out(task
, ctxn
, next
);
2643 * if cgroup events exist on this CPU, then we need
2644 * to check if we have to switch out PMU state.
2645 * cgroup event are system-wide mode only
2647 if (atomic_read(this_cpu_ptr(&perf_cgroup_events
)))
2648 perf_cgroup_sched_out(task
, next
);
2651 static void task_ctx_sched_out(struct perf_event_context
*ctx
)
2653 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
2655 if (!cpuctx
->task_ctx
)
2658 if (WARN_ON_ONCE(ctx
!= cpuctx
->task_ctx
))
2661 ctx_sched_out(ctx
, cpuctx
, EVENT_ALL
);
2662 cpuctx
->task_ctx
= NULL
;
2666 * Called with IRQs disabled
2668 static void cpu_ctx_sched_out(struct perf_cpu_context
*cpuctx
,
2669 enum event_type_t event_type
)
2671 ctx_sched_out(&cpuctx
->ctx
, cpuctx
, event_type
);
2675 ctx_pinned_sched_in(struct perf_event_context
*ctx
,
2676 struct perf_cpu_context
*cpuctx
)
2678 struct perf_event
*event
;
2680 list_for_each_entry(event
, &ctx
->pinned_groups
, group_entry
) {
2681 if (event
->state
<= PERF_EVENT_STATE_OFF
)
2683 if (!event_filter_match(event
))
2686 /* may need to reset tstamp_enabled */
2687 if (is_cgroup_event(event
))
2688 perf_cgroup_mark_enabled(event
, ctx
);
2690 if (group_can_go_on(event
, cpuctx
, 1))
2691 group_sched_in(event
, cpuctx
, ctx
);
2694 * If this pinned group hasn't been scheduled,
2695 * put it in error state.
2697 if (event
->state
== PERF_EVENT_STATE_INACTIVE
) {
2698 update_group_times(event
);
2699 event
->state
= PERF_EVENT_STATE_ERROR
;
2705 ctx_flexible_sched_in(struct perf_event_context
*ctx
,
2706 struct perf_cpu_context
*cpuctx
)
2708 struct perf_event
*event
;
2711 list_for_each_entry(event
, &ctx
->flexible_groups
, group_entry
) {
2712 /* Ignore events in OFF or ERROR state */
2713 if (event
->state
<= PERF_EVENT_STATE_OFF
)
2716 * Listen to the 'cpu' scheduling filter constraint
2719 if (!event_filter_match(event
))
2722 /* may need to reset tstamp_enabled */
2723 if (is_cgroup_event(event
))
2724 perf_cgroup_mark_enabled(event
, ctx
);
2726 if (group_can_go_on(event
, cpuctx
, can_add_hw
)) {
2727 if (group_sched_in(event
, cpuctx
, ctx
))
2734 ctx_sched_in(struct perf_event_context
*ctx
,
2735 struct perf_cpu_context
*cpuctx
,
2736 enum event_type_t event_type
,
2737 struct task_struct
*task
)
2740 int is_active
= ctx
->is_active
;
2742 ctx
->is_active
|= event_type
;
2743 if (likely(!ctx
->nr_events
))
2747 ctx
->timestamp
= now
;
2748 perf_cgroup_set_timestamp(task
, ctx
);
2750 * First go through the list and put on any pinned groups
2751 * in order to give them the best chance of going on.
2753 if (!(is_active
& EVENT_PINNED
) && (event_type
& EVENT_PINNED
))
2754 ctx_pinned_sched_in(ctx
, cpuctx
);
2756 /* Then walk through the lower prio flexible groups */
2757 if (!(is_active
& EVENT_FLEXIBLE
) && (event_type
& EVENT_FLEXIBLE
))
2758 ctx_flexible_sched_in(ctx
, cpuctx
);
2761 static void cpu_ctx_sched_in(struct perf_cpu_context
*cpuctx
,
2762 enum event_type_t event_type
,
2763 struct task_struct
*task
)
2765 struct perf_event_context
*ctx
= &cpuctx
->ctx
;
2767 ctx_sched_in(ctx
, cpuctx
, event_type
, task
);
2770 static void perf_event_context_sched_in(struct perf_event_context
*ctx
,
2771 struct task_struct
*task
)
2773 struct perf_cpu_context
*cpuctx
;
2775 cpuctx
= __get_cpu_context(ctx
);
2776 if (cpuctx
->task_ctx
== ctx
)
2779 perf_ctx_lock(cpuctx
, ctx
);
2780 perf_pmu_disable(ctx
->pmu
);
2782 * We want to keep the following priority order:
2783 * cpu pinned (that don't need to move), task pinned,
2784 * cpu flexible, task flexible.
2786 cpu_ctx_sched_out(cpuctx
, EVENT_FLEXIBLE
);
2789 cpuctx
->task_ctx
= ctx
;
2791 perf_event_sched_in(cpuctx
, cpuctx
->task_ctx
, task
);
2793 perf_pmu_enable(ctx
->pmu
);
2794 perf_ctx_unlock(cpuctx
, ctx
);
2798 * Called from scheduler to add the events of the current task
2799 * with interrupts disabled.
2801 * We restore the event value and then enable it.
2803 * This does not protect us against NMI, but enable()
2804 * sets the enabled bit in the control field of event _before_
2805 * accessing the event control register. If a NMI hits, then it will
2806 * keep the event running.
2808 void __perf_event_task_sched_in(struct task_struct
*prev
,
2809 struct task_struct
*task
)
2811 struct perf_event_context
*ctx
;
2814 for_each_task_context_nr(ctxn
) {
2815 ctx
= task
->perf_event_ctxp
[ctxn
];
2819 perf_event_context_sched_in(ctx
, task
);
2822 * if cgroup events exist on this CPU, then we need
2823 * to check if we have to switch in PMU state.
2824 * cgroup event are system-wide mode only
2826 if (atomic_read(this_cpu_ptr(&perf_cgroup_events
)))
2827 perf_cgroup_sched_in(prev
, task
);
2829 if (__this_cpu_read(perf_sched_cb_usages
))
2830 perf_pmu_sched_task(prev
, task
, true);
2833 static u64
perf_calculate_period(struct perf_event
*event
, u64 nsec
, u64 count
)
2835 u64 frequency
= event
->attr
.sample_freq
;
2836 u64 sec
= NSEC_PER_SEC
;
2837 u64 divisor
, dividend
;
2839 int count_fls
, nsec_fls
, frequency_fls
, sec_fls
;
2841 count_fls
= fls64(count
);
2842 nsec_fls
= fls64(nsec
);
2843 frequency_fls
= fls64(frequency
);
2847 * We got @count in @nsec, with a target of sample_freq HZ
2848 * the target period becomes:
2851 * period = -------------------
2852 * @nsec * sample_freq
2857 * Reduce accuracy by one bit such that @a and @b converge
2858 * to a similar magnitude.
2860 #define REDUCE_FLS(a, b) \
2862 if (a##_fls > b##_fls) { \
2872 * Reduce accuracy until either term fits in a u64, then proceed with
2873 * the other, so that finally we can do a u64/u64 division.
2875 while (count_fls
+ sec_fls
> 64 && nsec_fls
+ frequency_fls
> 64) {
2876 REDUCE_FLS(nsec
, frequency
);
2877 REDUCE_FLS(sec
, count
);
2880 if (count_fls
+ sec_fls
> 64) {
2881 divisor
= nsec
* frequency
;
2883 while (count_fls
+ sec_fls
> 64) {
2884 REDUCE_FLS(count
, sec
);
2888 dividend
= count
* sec
;
2890 dividend
= count
* sec
;
2892 while (nsec_fls
+ frequency_fls
> 64) {
2893 REDUCE_FLS(nsec
, frequency
);
2897 divisor
= nsec
* frequency
;
2903 return div64_u64(dividend
, divisor
);
2906 static DEFINE_PER_CPU(int, perf_throttled_count
);
2907 static DEFINE_PER_CPU(u64
, perf_throttled_seq
);
2909 static void perf_adjust_period(struct perf_event
*event
, u64 nsec
, u64 count
, bool disable
)
2911 struct hw_perf_event
*hwc
= &event
->hw
;
2912 s64 period
, sample_period
;
2915 period
= perf_calculate_period(event
, nsec
, count
);
2917 delta
= (s64
)(period
- hwc
->sample_period
);
2918 delta
= (delta
+ 7) / 8; /* low pass filter */
2920 sample_period
= hwc
->sample_period
+ delta
;
2925 hwc
->sample_period
= sample_period
;
2927 if (local64_read(&hwc
->period_left
) > 8*sample_period
) {
2929 event
->pmu
->stop(event
, PERF_EF_UPDATE
);
2931 local64_set(&hwc
->period_left
, 0);
2934 event
->pmu
->start(event
, PERF_EF_RELOAD
);
2939 * combine freq adjustment with unthrottling to avoid two passes over the
2940 * events. At the same time, make sure, having freq events does not change
2941 * the rate of unthrottling as that would introduce bias.
2943 static void perf_adjust_freq_unthr_context(struct perf_event_context
*ctx
,
2946 struct perf_event
*event
;
2947 struct hw_perf_event
*hwc
;
2948 u64 now
, period
= TICK_NSEC
;
2952 * only need to iterate over all events iff:
2953 * - context have events in frequency mode (needs freq adjust)
2954 * - there are events to unthrottle on this cpu
2956 if (!(ctx
->nr_freq
|| needs_unthr
))
2959 raw_spin_lock(&ctx
->lock
);
2960 perf_pmu_disable(ctx
->pmu
);
2962 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
2963 if (event
->state
!= PERF_EVENT_STATE_ACTIVE
)
2966 if (!event_filter_match(event
))
2969 perf_pmu_disable(event
->pmu
);
2973 if (hwc
->interrupts
== MAX_INTERRUPTS
) {
2974 hwc
->interrupts
= 0;
2975 perf_log_throttle(event
, 1);
2976 event
->pmu
->start(event
, 0);
2979 if (!event
->attr
.freq
|| !event
->attr
.sample_freq
)
2983 * stop the event and update event->count
2985 event
->pmu
->stop(event
, PERF_EF_UPDATE
);
2987 now
= local64_read(&event
->count
);
2988 delta
= now
- hwc
->freq_count_stamp
;
2989 hwc
->freq_count_stamp
= now
;
2993 * reload only if value has changed
2994 * we have stopped the event so tell that
2995 * to perf_adjust_period() to avoid stopping it
2999 perf_adjust_period(event
, period
, delta
, false);
3001 event
->pmu
->start(event
, delta
> 0 ? PERF_EF_RELOAD
: 0);
3003 perf_pmu_enable(event
->pmu
);
3006 perf_pmu_enable(ctx
->pmu
);
3007 raw_spin_unlock(&ctx
->lock
);
3011 * Round-robin a context's events:
3013 static void rotate_ctx(struct perf_event_context
*ctx
)
3016 * Rotate the first entry last of non-pinned groups. Rotation might be
3017 * disabled by the inheritance code.
3019 if (!ctx
->rotate_disable
)
3020 list_rotate_left(&ctx
->flexible_groups
);
3023 static int perf_rotate_context(struct perf_cpu_context
*cpuctx
)
3025 struct perf_event_context
*ctx
= NULL
;
3028 if (cpuctx
->ctx
.nr_events
) {
3029 if (cpuctx
->ctx
.nr_events
!= cpuctx
->ctx
.nr_active
)
3033 ctx
= cpuctx
->task_ctx
;
3034 if (ctx
&& ctx
->nr_events
) {
3035 if (ctx
->nr_events
!= ctx
->nr_active
)
3042 perf_ctx_lock(cpuctx
, cpuctx
->task_ctx
);
3043 perf_pmu_disable(cpuctx
->ctx
.pmu
);
3045 cpu_ctx_sched_out(cpuctx
, EVENT_FLEXIBLE
);
3047 ctx_sched_out(ctx
, cpuctx
, EVENT_FLEXIBLE
);
3049 rotate_ctx(&cpuctx
->ctx
);
3053 perf_event_sched_in(cpuctx
, ctx
, current
);
3055 perf_pmu_enable(cpuctx
->ctx
.pmu
);
3056 perf_ctx_unlock(cpuctx
, cpuctx
->task_ctx
);
3062 #ifdef CONFIG_NO_HZ_FULL
3063 bool perf_event_can_stop_tick(void)
3065 if (atomic_read(&nr_freq_events
) ||
3066 __this_cpu_read(perf_throttled_count
))
3073 void perf_event_task_tick(void)
3075 struct list_head
*head
= this_cpu_ptr(&active_ctx_list
);
3076 struct perf_event_context
*ctx
, *tmp
;
3079 WARN_ON(!irqs_disabled());
3081 __this_cpu_inc(perf_throttled_seq
);
3082 throttled
= __this_cpu_xchg(perf_throttled_count
, 0);
3084 list_for_each_entry_safe(ctx
, tmp
, head
, active_ctx_list
)
3085 perf_adjust_freq_unthr_context(ctx
, throttled
);
3088 static int event_enable_on_exec(struct perf_event
*event
,
3089 struct perf_event_context
*ctx
)
3091 if (!event
->attr
.enable_on_exec
)
3094 event
->attr
.enable_on_exec
= 0;
3095 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
)
3098 __perf_event_mark_enabled(event
);
3104 * Enable all of a task's events that have been marked enable-on-exec.
3105 * This expects task == current.
3107 static void perf_event_enable_on_exec(struct perf_event_context
*ctx
)
3109 struct perf_event_context
*clone_ctx
= NULL
;
3110 struct perf_event
*event
;
3111 unsigned long flags
;
3115 local_irq_save(flags
);
3116 if (!ctx
|| !ctx
->nr_events
)
3120 * We must ctxsw out cgroup events to avoid conflict
3121 * when invoking perf_task_event_sched_in() later on
3122 * in this function. Otherwise we end up trying to
3123 * ctxswin cgroup events which are already scheduled
3126 perf_cgroup_sched_out(current
, NULL
);
3128 raw_spin_lock(&ctx
->lock
);
3129 task_ctx_sched_out(ctx
);
3131 list_for_each_entry(event
, &ctx
->event_list
, event_entry
) {
3132 ret
= event_enable_on_exec(event
, ctx
);
3138 * Unclone this context if we enabled any event.
3141 clone_ctx
= unclone_ctx(ctx
);
3143 raw_spin_unlock(&ctx
->lock
);
3146 * Also calls ctxswin for cgroup events, if any:
3148 perf_event_context_sched_in(ctx
, ctx
->task
);
3150 local_irq_restore(flags
);
3156 void perf_event_exec(void)
3158 struct perf_event_context
*ctx
;
3162 for_each_task_context_nr(ctxn
) {
3163 ctx
= current
->perf_event_ctxp
[ctxn
];
3167 perf_event_enable_on_exec(ctx
);
3173 * Cross CPU call to read the hardware event
3175 static void __perf_event_read(void *info
)
3177 struct perf_event
*event
= info
;
3178 struct perf_event_context
*ctx
= event
->ctx
;
3179 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
3182 * If this is a task context, we need to check whether it is
3183 * the current task context of this cpu. If not it has been
3184 * scheduled out before the smp call arrived. In that case
3185 * event->count would have been updated to a recent sample
3186 * when the event was scheduled out.
3188 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
)
3191 raw_spin_lock(&ctx
->lock
);
3192 if (ctx
->is_active
) {
3193 update_context_time(ctx
);
3194 update_cgrp_time_from_event(event
);
3196 update_event_times(event
);
3197 if (event
->state
== PERF_EVENT_STATE_ACTIVE
)
3198 event
->pmu
->read(event
);
3199 raw_spin_unlock(&ctx
->lock
);
3202 static inline u64
perf_event_count(struct perf_event
*event
)
3204 if (event
->pmu
->count
)
3205 return event
->pmu
->count(event
);
3207 return __perf_event_count(event
);
3210 static u64
perf_event_read(struct perf_event
*event
)
3213 * If event is enabled and currently active on a CPU, update the
3214 * value in the event structure:
3216 if (event
->state
== PERF_EVENT_STATE_ACTIVE
) {
3217 smp_call_function_single(event
->oncpu
,
3218 __perf_event_read
, event
, 1);
3219 } else if (event
->state
== PERF_EVENT_STATE_INACTIVE
) {
3220 struct perf_event_context
*ctx
= event
->ctx
;
3221 unsigned long flags
;
3223 raw_spin_lock_irqsave(&ctx
->lock
, flags
);
3225 * may read while context is not active
3226 * (e.g., thread is blocked), in that case
3227 * we cannot update context time
3229 if (ctx
->is_active
) {
3230 update_context_time(ctx
);
3231 update_cgrp_time_from_event(event
);
3233 update_event_times(event
);
3234 raw_spin_unlock_irqrestore(&ctx
->lock
, flags
);
3237 return perf_event_count(event
);
3241 * Initialize the perf_event context in a task_struct:
3243 static void __perf_event_init_context(struct perf_event_context
*ctx
)
3245 raw_spin_lock_init(&ctx
->lock
);
3246 mutex_init(&ctx
->mutex
);
3247 INIT_LIST_HEAD(&ctx
->active_ctx_list
);
3248 INIT_LIST_HEAD(&ctx
->pinned_groups
);
3249 INIT_LIST_HEAD(&ctx
->flexible_groups
);
3250 INIT_LIST_HEAD(&ctx
->event_list
);
3251 atomic_set(&ctx
->refcount
, 1);
3252 INIT_DELAYED_WORK(&ctx
->orphans_remove
, orphans_remove_work
);
3255 static struct perf_event_context
*
3256 alloc_perf_context(struct pmu
*pmu
, struct task_struct
*task
)
3258 struct perf_event_context
*ctx
;
3260 ctx
= kzalloc(sizeof(struct perf_event_context
), GFP_KERNEL
);
3264 __perf_event_init_context(ctx
);
3267 get_task_struct(task
);
3274 static struct task_struct
*
3275 find_lively_task_by_vpid(pid_t vpid
)
3277 struct task_struct
*task
;
3284 task
= find_task_by_vpid(vpid
);
3286 get_task_struct(task
);
3290 return ERR_PTR(-ESRCH
);
3292 /* Reuse ptrace permission checks for now. */
3294 if (!ptrace_may_access(task
, PTRACE_MODE_READ
))
3299 put_task_struct(task
);
3300 return ERR_PTR(err
);
3305 * Returns a matching context with refcount and pincount.
3307 static struct perf_event_context
*
3308 find_get_context(struct pmu
*pmu
, struct task_struct
*task
,
3309 struct perf_event
*event
)
3311 struct perf_event_context
*ctx
, *clone_ctx
= NULL
;
3312 struct perf_cpu_context
*cpuctx
;
3313 void *task_ctx_data
= NULL
;
3314 unsigned long flags
;
3316 int cpu
= event
->cpu
;
3319 /* Must be root to operate on a CPU event: */
3320 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN
))
3321 return ERR_PTR(-EACCES
);
3324 * We could be clever and allow to attach a event to an
3325 * offline CPU and activate it when the CPU comes up, but
3328 if (!cpu_online(cpu
))
3329 return ERR_PTR(-ENODEV
);
3331 cpuctx
= per_cpu_ptr(pmu
->pmu_cpu_context
, cpu
);
3340 ctxn
= pmu
->task_ctx_nr
;
3344 if (event
->attach_state
& PERF_ATTACH_TASK_DATA
) {
3345 task_ctx_data
= kzalloc(pmu
->task_ctx_size
, GFP_KERNEL
);
3346 if (!task_ctx_data
) {
3353 ctx
= perf_lock_task_context(task
, ctxn
, &flags
);
3355 clone_ctx
= unclone_ctx(ctx
);
3358 if (task_ctx_data
&& !ctx
->task_ctx_data
) {
3359 ctx
->task_ctx_data
= task_ctx_data
;
3360 task_ctx_data
= NULL
;
3362 raw_spin_unlock_irqrestore(&ctx
->lock
, flags
);
3367 ctx
= alloc_perf_context(pmu
, task
);
3372 if (task_ctx_data
) {
3373 ctx
->task_ctx_data
= task_ctx_data
;
3374 task_ctx_data
= NULL
;
3378 mutex_lock(&task
->perf_event_mutex
);
3380 * If it has already passed perf_event_exit_task().
3381 * we must see PF_EXITING, it takes this mutex too.
3383 if (task
->flags
& PF_EXITING
)
3385 else if (task
->perf_event_ctxp
[ctxn
])
3390 rcu_assign_pointer(task
->perf_event_ctxp
[ctxn
], ctx
);
3392 mutex_unlock(&task
->perf_event_mutex
);
3394 if (unlikely(err
)) {
3403 kfree(task_ctx_data
);
3407 kfree(task_ctx_data
);
3408 return ERR_PTR(err
);
3411 static void perf_event_free_filter(struct perf_event
*event
);
3412 static void perf_event_free_bpf_prog(struct perf_event
*event
);
3414 static void free_event_rcu(struct rcu_head
*head
)
3416 struct perf_event
*event
;
3418 event
= container_of(head
, struct perf_event
, rcu_head
);
3420 put_pid_ns(event
->ns
);
3421 perf_event_free_filter(event
);
3422 perf_event_free_bpf_prog(event
);
3426 static void ring_buffer_put(struct ring_buffer
*rb
);
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
);
3462 static void __free_event(struct perf_event
*event
)
3464 if (!event
->parent
) {
3465 if (event
->attr
.sample_type
& PERF_SAMPLE_CALLCHAIN
)
3466 put_callchain_buffers();
3470 event
->destroy(event
);
3473 put_ctx(event
->ctx
);
3476 module_put(event
->pmu
->module
);
3478 call_rcu(&event
->rcu_head
, free_event_rcu
);
3481 static void _free_event(struct perf_event
*event
)
3483 irq_work_sync(&event
->pending
);
3485 unaccount_event(event
);
3489 * Can happen when we close an event with re-directed output.
3491 * Since we have a 0 refcount, perf_mmap_close() will skip
3492 * over us; possibly making our ring_buffer_put() the last.
3494 mutex_lock(&event
->mmap_mutex
);
3495 ring_buffer_attach(event
, NULL
);
3496 mutex_unlock(&event
->mmap_mutex
);
3499 if (is_cgroup_event(event
))
3500 perf_detach_cgroup(event
);
3502 __free_event(event
);
3506 * Used to free events which have a known refcount of 1, such as in error paths
3507 * where the event isn't exposed yet and inherited events.
3509 static void free_event(struct perf_event
*event
)
3511 if (WARN(atomic_long_cmpxchg(&event
->refcount
, 1, 0) != 1,
3512 "unexpected event refcount: %ld; ptr=%p\n",
3513 atomic_long_read(&event
->refcount
), event
)) {
3514 /* leak to avoid use-after-free */
3522 * Remove user event from the owner task.
3524 static void perf_remove_from_owner(struct perf_event
*event
)
3526 struct task_struct
*owner
;
3529 owner
= ACCESS_ONCE(event
->owner
);
3531 * Matches the smp_wmb() in perf_event_exit_task(). If we observe
3532 * !owner it means the list deletion is complete and we can indeed
3533 * free this event, otherwise we need to serialize on
3534 * owner->perf_event_mutex.
3536 smp_read_barrier_depends();
3539 * Since delayed_put_task_struct() also drops the last
3540 * task reference we can safely take a new reference
3541 * while holding the rcu_read_lock().
3543 get_task_struct(owner
);
3549 * If we're here through perf_event_exit_task() we're already
3550 * holding ctx->mutex which would be an inversion wrt. the
3551 * normal lock order.
3553 * However we can safely take this lock because its the child
3556 mutex_lock_nested(&owner
->perf_event_mutex
, SINGLE_DEPTH_NESTING
);
3559 * We have to re-check the event->owner field, if it is cleared
3560 * we raced with perf_event_exit_task(), acquiring the mutex
3561 * ensured they're done, and we can proceed with freeing the
3565 list_del_init(&event
->owner_entry
);
3566 mutex_unlock(&owner
->perf_event_mutex
);
3567 put_task_struct(owner
);
3572 * Called when the last reference to the file is gone.
3574 static void put_event(struct perf_event
*event
)
3576 struct perf_event_context
*ctx
;
3578 if (!atomic_long_dec_and_test(&event
->refcount
))
3581 if (!is_kernel_event(event
))
3582 perf_remove_from_owner(event
);
3585 * There are two ways this annotation is useful:
3587 * 1) there is a lock recursion from perf_event_exit_task
3588 * see the comment there.
3590 * 2) there is a lock-inversion with mmap_sem through
3591 * perf_event_read_group(), which takes faults while
3592 * holding ctx->mutex, however this is called after
3593 * the last filedesc died, so there is no possibility
3594 * to trigger the AB-BA case.
3596 ctx
= perf_event_ctx_lock_nested(event
, SINGLE_DEPTH_NESTING
);
3597 WARN_ON_ONCE(ctx
->parent_ctx
);
3598 perf_remove_from_context(event
, true);
3599 perf_event_ctx_unlock(event
, ctx
);
3604 int perf_event_release_kernel(struct perf_event
*event
)
3609 EXPORT_SYMBOL_GPL(perf_event_release_kernel
);
3611 static int perf_release(struct inode
*inode
, struct file
*file
)
3613 put_event(file
->private_data
);
3618 * Remove all orphanes events from the context.
3620 static void orphans_remove_work(struct work_struct
*work
)
3622 struct perf_event_context
*ctx
;
3623 struct perf_event
*event
, *tmp
;
3625 ctx
= container_of(work
, struct perf_event_context
,
3626 orphans_remove
.work
);
3628 mutex_lock(&ctx
->mutex
);
3629 list_for_each_entry_safe(event
, tmp
, &ctx
->event_list
, event_entry
) {
3630 struct perf_event
*parent_event
= event
->parent
;
3632 if (!is_orphaned_child(event
))
3635 perf_remove_from_context(event
, true);
3637 mutex_lock(&parent_event
->child_mutex
);
3638 list_del_init(&event
->child_list
);
3639 mutex_unlock(&parent_event
->child_mutex
);
3642 put_event(parent_event
);
3645 raw_spin_lock_irq(&ctx
->lock
);
3646 ctx
->orphans_remove_sched
= false;
3647 raw_spin_unlock_irq(&ctx
->lock
);
3648 mutex_unlock(&ctx
->mutex
);
3653 u64
perf_event_read_value(struct perf_event
*event
, u64
*enabled
, u64
*running
)
3655 struct perf_event
*child
;
3661 mutex_lock(&event
->child_mutex
);
3662 total
+= perf_event_read(event
);
3663 *enabled
+= event
->total_time_enabled
+
3664 atomic64_read(&event
->child_total_time_enabled
);
3665 *running
+= event
->total_time_running
+
3666 atomic64_read(&event
->child_total_time_running
);
3668 list_for_each_entry(child
, &event
->child_list
, child_list
) {
3669 total
+= perf_event_read(child
);
3670 *enabled
+= child
->total_time_enabled
;
3671 *running
+= child
->total_time_running
;
3673 mutex_unlock(&event
->child_mutex
);
3677 EXPORT_SYMBOL_GPL(perf_event_read_value
);
3679 static int perf_event_read_group(struct perf_event
*event
,
3680 u64 read_format
, char __user
*buf
)
3682 struct perf_event
*leader
= event
->group_leader
, *sub
;
3683 struct perf_event_context
*ctx
= leader
->ctx
;
3684 int n
= 0, size
= 0, ret
;
3685 u64 count
, enabled
, running
;
3688 lockdep_assert_held(&ctx
->mutex
);
3690 count
= perf_event_read_value(leader
, &enabled
, &running
);
3692 values
[n
++] = 1 + leader
->nr_siblings
;
3693 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
3694 values
[n
++] = enabled
;
3695 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
3696 values
[n
++] = running
;
3697 values
[n
++] = count
;
3698 if (read_format
& PERF_FORMAT_ID
)
3699 values
[n
++] = primary_event_id(leader
);
3701 size
= n
* sizeof(u64
);
3703 if (copy_to_user(buf
, values
, size
))
3708 list_for_each_entry(sub
, &leader
->sibling_list
, group_entry
) {
3711 values
[n
++] = perf_event_read_value(sub
, &enabled
, &running
);
3712 if (read_format
& PERF_FORMAT_ID
)
3713 values
[n
++] = primary_event_id(sub
);
3715 size
= n
* sizeof(u64
);
3717 if (copy_to_user(buf
+ ret
, values
, size
)) {
3727 static int perf_event_read_one(struct perf_event
*event
,
3728 u64 read_format
, char __user
*buf
)
3730 u64 enabled
, running
;
3734 values
[n
++] = perf_event_read_value(event
, &enabled
, &running
);
3735 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
3736 values
[n
++] = enabled
;
3737 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
3738 values
[n
++] = running
;
3739 if (read_format
& PERF_FORMAT_ID
)
3740 values
[n
++] = primary_event_id(event
);
3742 if (copy_to_user(buf
, values
, n
* sizeof(u64
)))
3745 return n
* sizeof(u64
);
3748 static bool is_event_hup(struct perf_event
*event
)
3752 if (event
->state
!= PERF_EVENT_STATE_EXIT
)
3755 mutex_lock(&event
->child_mutex
);
3756 no_children
= list_empty(&event
->child_list
);
3757 mutex_unlock(&event
->child_mutex
);
3762 * Read the performance event - simple non blocking version for now
3765 perf_read_hw(struct perf_event
*event
, char __user
*buf
, size_t count
)
3767 u64 read_format
= event
->attr
.read_format
;
3771 * Return end-of-file for a read on a event that is in
3772 * error state (i.e. because it was pinned but it couldn't be
3773 * scheduled on to the CPU at some point).
3775 if (event
->state
== PERF_EVENT_STATE_ERROR
)
3778 if (count
< event
->read_size
)
3781 WARN_ON_ONCE(event
->ctx
->parent_ctx
);
3782 if (read_format
& PERF_FORMAT_GROUP
)
3783 ret
= perf_event_read_group(event
, read_format
, buf
);
3785 ret
= perf_event_read_one(event
, read_format
, buf
);
3791 perf_read(struct file
*file
, char __user
*buf
, size_t count
, loff_t
*ppos
)
3793 struct perf_event
*event
= file
->private_data
;
3794 struct perf_event_context
*ctx
;
3797 ctx
= perf_event_ctx_lock(event
);
3798 ret
= perf_read_hw(event
, buf
, count
);
3799 perf_event_ctx_unlock(event
, ctx
);
3804 static unsigned int perf_poll(struct file
*file
, poll_table
*wait
)
3806 struct perf_event
*event
= file
->private_data
;
3807 struct ring_buffer
*rb
;
3808 unsigned int events
= POLLHUP
;
3810 poll_wait(file
, &event
->waitq
, wait
);
3812 if (is_event_hup(event
))
3816 * Pin the event->rb by taking event->mmap_mutex; otherwise
3817 * perf_event_set_output() can swizzle our rb and make us miss wakeups.
3819 mutex_lock(&event
->mmap_mutex
);
3822 events
= atomic_xchg(&rb
->poll
, 0);
3823 mutex_unlock(&event
->mmap_mutex
);
3827 static void _perf_event_reset(struct perf_event
*event
)
3829 (void)perf_event_read(event
);
3830 local64_set(&event
->count
, 0);
3831 perf_event_update_userpage(event
);
3835 * Holding the top-level event's child_mutex means that any
3836 * descendant process that has inherited this event will block
3837 * in sync_child_event if it goes to exit, thus satisfying the
3838 * task existence requirements of perf_event_enable/disable.
3840 static void perf_event_for_each_child(struct perf_event
*event
,
3841 void (*func
)(struct perf_event
*))
3843 struct perf_event
*child
;
3845 WARN_ON_ONCE(event
->ctx
->parent_ctx
);
3847 mutex_lock(&event
->child_mutex
);
3849 list_for_each_entry(child
, &event
->child_list
, child_list
)
3851 mutex_unlock(&event
->child_mutex
);
3854 static void perf_event_for_each(struct perf_event
*event
,
3855 void (*func
)(struct perf_event
*))
3857 struct perf_event_context
*ctx
= event
->ctx
;
3858 struct perf_event
*sibling
;
3860 lockdep_assert_held(&ctx
->mutex
);
3862 event
= event
->group_leader
;
3864 perf_event_for_each_child(event
, func
);
3865 list_for_each_entry(sibling
, &event
->sibling_list
, group_entry
)
3866 perf_event_for_each_child(sibling
, func
);
3869 static int perf_event_period(struct perf_event
*event
, u64 __user
*arg
)
3871 struct perf_event_context
*ctx
= event
->ctx
;
3872 int ret
= 0, active
;
3875 if (!is_sampling_event(event
))
3878 if (copy_from_user(&value
, arg
, sizeof(value
)))
3884 raw_spin_lock_irq(&ctx
->lock
);
3885 if (event
->attr
.freq
) {
3886 if (value
> sysctl_perf_event_sample_rate
) {
3891 event
->attr
.sample_freq
= value
;
3893 event
->attr
.sample_period
= value
;
3894 event
->hw
.sample_period
= value
;
3897 active
= (event
->state
== PERF_EVENT_STATE_ACTIVE
);
3899 perf_pmu_disable(ctx
->pmu
);
3900 event
->pmu
->stop(event
, PERF_EF_UPDATE
);
3903 local64_set(&event
->hw
.period_left
, 0);
3906 event
->pmu
->start(event
, PERF_EF_RELOAD
);
3907 perf_pmu_enable(ctx
->pmu
);
3911 raw_spin_unlock_irq(&ctx
->lock
);
3916 static const struct file_operations perf_fops
;
3918 static inline int perf_fget_light(int fd
, struct fd
*p
)
3920 struct fd f
= fdget(fd
);
3924 if (f
.file
->f_op
!= &perf_fops
) {
3932 static int perf_event_set_output(struct perf_event
*event
,
3933 struct perf_event
*output_event
);
3934 static int perf_event_set_filter(struct perf_event
*event
, void __user
*arg
);
3935 static int perf_event_set_bpf_prog(struct perf_event
*event
, u32 prog_fd
);
3937 static long _perf_ioctl(struct perf_event
*event
, unsigned int cmd
, unsigned long arg
)
3939 void (*func
)(struct perf_event
*);
3943 case PERF_EVENT_IOC_ENABLE
:
3944 func
= _perf_event_enable
;
3946 case PERF_EVENT_IOC_DISABLE
:
3947 func
= _perf_event_disable
;
3949 case PERF_EVENT_IOC_RESET
:
3950 func
= _perf_event_reset
;
3953 case PERF_EVENT_IOC_REFRESH
:
3954 return _perf_event_refresh(event
, arg
);
3956 case PERF_EVENT_IOC_PERIOD
:
3957 return perf_event_period(event
, (u64 __user
*)arg
);
3959 case PERF_EVENT_IOC_ID
:
3961 u64 id
= primary_event_id(event
);
3963 if (copy_to_user((void __user
*)arg
, &id
, sizeof(id
)))
3968 case PERF_EVENT_IOC_SET_OUTPUT
:
3972 struct perf_event
*output_event
;
3974 ret
= perf_fget_light(arg
, &output
);
3977 output_event
= output
.file
->private_data
;
3978 ret
= perf_event_set_output(event
, output_event
);
3981 ret
= perf_event_set_output(event
, NULL
);
3986 case PERF_EVENT_IOC_SET_FILTER
:
3987 return perf_event_set_filter(event
, (void __user
*)arg
);
3989 case PERF_EVENT_IOC_SET_BPF
:
3990 return perf_event_set_bpf_prog(event
, arg
);
3996 if (flags
& PERF_IOC_FLAG_GROUP
)
3997 perf_event_for_each(event
, func
);
3999 perf_event_for_each_child(event
, func
);
4004 static long perf_ioctl(struct file
*file
, unsigned int cmd
, unsigned long arg
)
4006 struct perf_event
*event
= file
->private_data
;
4007 struct perf_event_context
*ctx
;
4010 ctx
= perf_event_ctx_lock(event
);
4011 ret
= _perf_ioctl(event
, cmd
, arg
);
4012 perf_event_ctx_unlock(event
, ctx
);
4017 #ifdef CONFIG_COMPAT
4018 static long perf_compat_ioctl(struct file
*file
, unsigned int cmd
,
4021 switch (_IOC_NR(cmd
)) {
4022 case _IOC_NR(PERF_EVENT_IOC_SET_FILTER
):
4023 case _IOC_NR(PERF_EVENT_IOC_ID
):
4024 /* Fix up pointer size (usually 4 -> 8 in 32-on-64-bit case */
4025 if (_IOC_SIZE(cmd
) == sizeof(compat_uptr_t
)) {
4026 cmd
&= ~IOCSIZE_MASK
;
4027 cmd
|= sizeof(void *) << IOCSIZE_SHIFT
;
4031 return perf_ioctl(file
, cmd
, arg
);
4034 # define perf_compat_ioctl NULL
4037 int perf_event_task_enable(void)
4039 struct perf_event_context
*ctx
;
4040 struct perf_event
*event
;
4042 mutex_lock(¤t
->perf_event_mutex
);
4043 list_for_each_entry(event
, ¤t
->perf_event_list
, owner_entry
) {
4044 ctx
= perf_event_ctx_lock(event
);
4045 perf_event_for_each_child(event
, _perf_event_enable
);
4046 perf_event_ctx_unlock(event
, ctx
);
4048 mutex_unlock(¤t
->perf_event_mutex
);
4053 int perf_event_task_disable(void)
4055 struct perf_event_context
*ctx
;
4056 struct perf_event
*event
;
4058 mutex_lock(¤t
->perf_event_mutex
);
4059 list_for_each_entry(event
, ¤t
->perf_event_list
, owner_entry
) {
4060 ctx
= perf_event_ctx_lock(event
);
4061 perf_event_for_each_child(event
, _perf_event_disable
);
4062 perf_event_ctx_unlock(event
, ctx
);
4064 mutex_unlock(¤t
->perf_event_mutex
);
4069 static int perf_event_index(struct perf_event
*event
)
4071 if (event
->hw
.state
& PERF_HES_STOPPED
)
4074 if (event
->state
!= PERF_EVENT_STATE_ACTIVE
)
4077 return event
->pmu
->event_idx(event
);
4080 static void calc_timer_values(struct perf_event
*event
,
4087 *now
= perf_clock();
4088 ctx_time
= event
->shadow_ctx_time
+ *now
;
4089 *enabled
= ctx_time
- event
->tstamp_enabled
;
4090 *running
= ctx_time
- event
->tstamp_running
;
4093 static void perf_event_init_userpage(struct perf_event
*event
)
4095 struct perf_event_mmap_page
*userpg
;
4096 struct ring_buffer
*rb
;
4099 rb
= rcu_dereference(event
->rb
);
4103 userpg
= rb
->user_page
;
4105 /* Allow new userspace to detect that bit 0 is deprecated */
4106 userpg
->cap_bit0_is_deprecated
= 1;
4107 userpg
->size
= offsetof(struct perf_event_mmap_page
, __reserved
);
4108 userpg
->data_offset
= PAGE_SIZE
;
4109 userpg
->data_size
= perf_data_size(rb
);
4115 void __weak
arch_perf_update_userpage(
4116 struct perf_event
*event
, struct perf_event_mmap_page
*userpg
, u64 now
)
4121 * Callers need to ensure there can be no nesting of this function, otherwise
4122 * the seqlock logic goes bad. We can not serialize this because the arch
4123 * code calls this from NMI context.
4125 void perf_event_update_userpage(struct perf_event
*event
)
4127 struct perf_event_mmap_page
*userpg
;
4128 struct ring_buffer
*rb
;
4129 u64 enabled
, running
, now
;
4132 rb
= rcu_dereference(event
->rb
);
4137 * compute total_time_enabled, total_time_running
4138 * based on snapshot values taken when the event
4139 * was last scheduled in.
4141 * we cannot simply called update_context_time()
4142 * because of locking issue as we can be called in
4145 calc_timer_values(event
, &now
, &enabled
, &running
);
4147 userpg
= rb
->user_page
;
4149 * Disable preemption so as to not let the corresponding user-space
4150 * spin too long if we get preempted.
4155 userpg
->index
= perf_event_index(event
);
4156 userpg
->offset
= perf_event_count(event
);
4158 userpg
->offset
-= local64_read(&event
->hw
.prev_count
);
4160 userpg
->time_enabled
= enabled
+
4161 atomic64_read(&event
->child_total_time_enabled
);
4163 userpg
->time_running
= running
+
4164 atomic64_read(&event
->child_total_time_running
);
4166 arch_perf_update_userpage(event
, userpg
, now
);
4175 static int perf_mmap_fault(struct vm_area_struct
*vma
, struct vm_fault
*vmf
)
4177 struct perf_event
*event
= vma
->vm_file
->private_data
;
4178 struct ring_buffer
*rb
;
4179 int ret
= VM_FAULT_SIGBUS
;
4181 if (vmf
->flags
& FAULT_FLAG_MKWRITE
) {
4182 if (vmf
->pgoff
== 0)
4188 rb
= rcu_dereference(event
->rb
);
4192 if (vmf
->pgoff
&& (vmf
->flags
& FAULT_FLAG_WRITE
))
4195 vmf
->page
= perf_mmap_to_page(rb
, vmf
->pgoff
);
4199 get_page(vmf
->page
);
4200 vmf
->page
->mapping
= vma
->vm_file
->f_mapping
;
4201 vmf
->page
->index
= vmf
->pgoff
;
4210 static void ring_buffer_attach(struct perf_event
*event
,
4211 struct ring_buffer
*rb
)
4213 struct ring_buffer
*old_rb
= NULL
;
4214 unsigned long flags
;
4218 * Should be impossible, we set this when removing
4219 * event->rb_entry and wait/clear when adding event->rb_entry.
4221 WARN_ON_ONCE(event
->rcu_pending
);
4224 event
->rcu_batches
= get_state_synchronize_rcu();
4225 event
->rcu_pending
= 1;
4227 spin_lock_irqsave(&old_rb
->event_lock
, flags
);
4228 list_del_rcu(&event
->rb_entry
);
4229 spin_unlock_irqrestore(&old_rb
->event_lock
, flags
);
4232 if (event
->rcu_pending
&& rb
) {
4233 cond_synchronize_rcu(event
->rcu_batches
);
4234 event
->rcu_pending
= 0;
4238 spin_lock_irqsave(&rb
->event_lock
, flags
);
4239 list_add_rcu(&event
->rb_entry
, &rb
->event_list
);
4240 spin_unlock_irqrestore(&rb
->event_lock
, flags
);
4243 rcu_assign_pointer(event
->rb
, rb
);
4246 ring_buffer_put(old_rb
);
4248 * Since we detached before setting the new rb, so that we
4249 * could attach the new rb, we could have missed a wakeup.
4252 wake_up_all(&event
->waitq
);
4256 static void ring_buffer_wakeup(struct perf_event
*event
)
4258 struct ring_buffer
*rb
;
4261 rb
= rcu_dereference(event
->rb
);
4263 list_for_each_entry_rcu(event
, &rb
->event_list
, rb_entry
)
4264 wake_up_all(&event
->waitq
);
4269 static void rb_free_rcu(struct rcu_head
*rcu_head
)
4271 struct ring_buffer
*rb
;
4273 rb
= container_of(rcu_head
, struct ring_buffer
, rcu_head
);
4277 static struct ring_buffer
*ring_buffer_get(struct perf_event
*event
)
4279 struct ring_buffer
*rb
;
4282 rb
= rcu_dereference(event
->rb
);
4284 if (!atomic_inc_not_zero(&rb
->refcount
))
4292 static void ring_buffer_put(struct ring_buffer
*rb
)
4294 if (!atomic_dec_and_test(&rb
->refcount
))
4297 WARN_ON_ONCE(!list_empty(&rb
->event_list
));
4299 call_rcu(&rb
->rcu_head
, rb_free_rcu
);
4302 static void perf_mmap_open(struct vm_area_struct
*vma
)
4304 struct perf_event
*event
= vma
->vm_file
->private_data
;
4306 atomic_inc(&event
->mmap_count
);
4307 atomic_inc(&event
->rb
->mmap_count
);
4310 atomic_inc(&event
->rb
->aux_mmap_count
);
4312 if (event
->pmu
->event_mapped
)
4313 event
->pmu
->event_mapped(event
);
4317 * A buffer can be mmap()ed multiple times; either directly through the same
4318 * event, or through other events by use of perf_event_set_output().
4320 * In order to undo the VM accounting done by perf_mmap() we need to destroy
4321 * the buffer here, where we still have a VM context. This means we need
4322 * to detach all events redirecting to us.
4324 static void perf_mmap_close(struct vm_area_struct
*vma
)
4326 struct perf_event
*event
= vma
->vm_file
->private_data
;
4328 struct ring_buffer
*rb
= ring_buffer_get(event
);
4329 struct user_struct
*mmap_user
= rb
->mmap_user
;
4330 int mmap_locked
= rb
->mmap_locked
;
4331 unsigned long size
= perf_data_size(rb
);
4333 if (event
->pmu
->event_unmapped
)
4334 event
->pmu
->event_unmapped(event
);
4337 * rb->aux_mmap_count will always drop before rb->mmap_count and
4338 * event->mmap_count, so it is ok to use event->mmap_mutex to
4339 * serialize with perf_mmap here.
4341 if (rb_has_aux(rb
) && vma
->vm_pgoff
== rb
->aux_pgoff
&&
4342 atomic_dec_and_mutex_lock(&rb
->aux_mmap_count
, &event
->mmap_mutex
)) {
4343 atomic_long_sub(rb
->aux_nr_pages
, &mmap_user
->locked_vm
);
4344 vma
->vm_mm
->pinned_vm
-= rb
->aux_mmap_locked
;
4347 mutex_unlock(&event
->mmap_mutex
);
4350 atomic_dec(&rb
->mmap_count
);
4352 if (!atomic_dec_and_mutex_lock(&event
->mmap_count
, &event
->mmap_mutex
))
4355 ring_buffer_attach(event
, NULL
);
4356 mutex_unlock(&event
->mmap_mutex
);
4358 /* If there's still other mmap()s of this buffer, we're done. */
4359 if (atomic_read(&rb
->mmap_count
))
4363 * No other mmap()s, detach from all other events that might redirect
4364 * into the now unreachable buffer. Somewhat complicated by the
4365 * fact that rb::event_lock otherwise nests inside mmap_mutex.
4369 list_for_each_entry_rcu(event
, &rb
->event_list
, rb_entry
) {
4370 if (!atomic_long_inc_not_zero(&event
->refcount
)) {
4372 * This event is en-route to free_event() which will
4373 * detach it and remove it from the list.
4379 mutex_lock(&event
->mmap_mutex
);
4381 * Check we didn't race with perf_event_set_output() which can
4382 * swizzle the rb from under us while we were waiting to
4383 * acquire mmap_mutex.
4385 * If we find a different rb; ignore this event, a next
4386 * iteration will no longer find it on the list. We have to
4387 * still restart the iteration to make sure we're not now
4388 * iterating the wrong list.
4390 if (event
->rb
== rb
)
4391 ring_buffer_attach(event
, NULL
);
4393 mutex_unlock(&event
->mmap_mutex
);
4397 * Restart the iteration; either we're on the wrong list or
4398 * destroyed its integrity by doing a deletion.
4405 * It could be there's still a few 0-ref events on the list; they'll
4406 * get cleaned up by free_event() -- they'll also still have their
4407 * ref on the rb and will free it whenever they are done with it.
4409 * Aside from that, this buffer is 'fully' detached and unmapped,
4410 * undo the VM accounting.
4413 atomic_long_sub((size
>> PAGE_SHIFT
) + 1, &mmap_user
->locked_vm
);
4414 vma
->vm_mm
->pinned_vm
-= mmap_locked
;
4415 free_uid(mmap_user
);
4418 ring_buffer_put(rb
); /* could be last */
4421 static const struct vm_operations_struct perf_mmap_vmops
= {
4422 .open
= perf_mmap_open
,
4423 .close
= perf_mmap_close
, /* non mergable */
4424 .fault
= perf_mmap_fault
,
4425 .page_mkwrite
= perf_mmap_fault
,
4428 static int perf_mmap(struct file
*file
, struct vm_area_struct
*vma
)
4430 struct perf_event
*event
= file
->private_data
;
4431 unsigned long user_locked
, user_lock_limit
;
4432 struct user_struct
*user
= current_user();
4433 unsigned long locked
, lock_limit
;
4434 struct ring_buffer
*rb
= NULL
;
4435 unsigned long vma_size
;
4436 unsigned long nr_pages
;
4437 long user_extra
= 0, extra
= 0;
4438 int ret
= 0, flags
= 0;
4441 * Don't allow mmap() of inherited per-task counters. This would
4442 * create a performance issue due to all children writing to the
4445 if (event
->cpu
== -1 && event
->attr
.inherit
)
4448 if (!(vma
->vm_flags
& VM_SHARED
))
4451 vma_size
= vma
->vm_end
- vma
->vm_start
;
4453 if (vma
->vm_pgoff
== 0) {
4454 nr_pages
= (vma_size
/ PAGE_SIZE
) - 1;
4457 * AUX area mapping: if rb->aux_nr_pages != 0, it's already
4458 * mapped, all subsequent mappings should have the same size
4459 * and offset. Must be above the normal perf buffer.
4461 u64 aux_offset
, aux_size
;
4466 nr_pages
= vma_size
/ PAGE_SIZE
;
4468 mutex_lock(&event
->mmap_mutex
);
4475 aux_offset
= ACCESS_ONCE(rb
->user_page
->aux_offset
);
4476 aux_size
= ACCESS_ONCE(rb
->user_page
->aux_size
);
4478 if (aux_offset
< perf_data_size(rb
) + PAGE_SIZE
)
4481 if (aux_offset
!= vma
->vm_pgoff
<< PAGE_SHIFT
)
4484 /* already mapped with a different offset */
4485 if (rb_has_aux(rb
) && rb
->aux_pgoff
!= vma
->vm_pgoff
)
4488 if (aux_size
!= vma_size
|| aux_size
!= nr_pages
* PAGE_SIZE
)
4491 /* already mapped with a different size */
4492 if (rb_has_aux(rb
) && rb
->aux_nr_pages
!= nr_pages
)
4495 if (!is_power_of_2(nr_pages
))
4498 if (!atomic_inc_not_zero(&rb
->mmap_count
))
4501 if (rb_has_aux(rb
)) {
4502 atomic_inc(&rb
->aux_mmap_count
);
4507 atomic_set(&rb
->aux_mmap_count
, 1);
4508 user_extra
= nr_pages
;
4514 * If we have rb pages ensure they're a power-of-two number, so we
4515 * can do bitmasks instead of modulo.
4517 if (nr_pages
!= 0 && !is_power_of_2(nr_pages
))
4520 if (vma_size
!= PAGE_SIZE
* (1 + nr_pages
))
4523 WARN_ON_ONCE(event
->ctx
->parent_ctx
);
4525 mutex_lock(&event
->mmap_mutex
);
4527 if (event
->rb
->nr_pages
!= nr_pages
) {
4532 if (!atomic_inc_not_zero(&event
->rb
->mmap_count
)) {
4534 * Raced against perf_mmap_close() through
4535 * perf_event_set_output(). Try again, hope for better
4538 mutex_unlock(&event
->mmap_mutex
);
4545 user_extra
= nr_pages
+ 1;
4548 user_lock_limit
= sysctl_perf_event_mlock
>> (PAGE_SHIFT
- 10);
4551 * Increase the limit linearly with more CPUs:
4553 user_lock_limit
*= num_online_cpus();
4555 user_locked
= atomic_long_read(&user
->locked_vm
) + user_extra
;
4557 if (user_locked
> user_lock_limit
)
4558 extra
= user_locked
- user_lock_limit
;
4560 lock_limit
= rlimit(RLIMIT_MEMLOCK
);
4561 lock_limit
>>= PAGE_SHIFT
;
4562 locked
= vma
->vm_mm
->pinned_vm
+ extra
;
4564 if ((locked
> lock_limit
) && perf_paranoid_tracepoint_raw() &&
4565 !capable(CAP_IPC_LOCK
)) {
4570 WARN_ON(!rb
&& event
->rb
);
4572 if (vma
->vm_flags
& VM_WRITE
)
4573 flags
|= RING_BUFFER_WRITABLE
;
4576 rb
= rb_alloc(nr_pages
,
4577 event
->attr
.watermark
? event
->attr
.wakeup_watermark
: 0,
4585 atomic_set(&rb
->mmap_count
, 1);
4586 rb
->mmap_user
= get_current_user();
4587 rb
->mmap_locked
= extra
;
4589 ring_buffer_attach(event
, rb
);
4591 perf_event_init_userpage(event
);
4592 perf_event_update_userpage(event
);
4594 ret
= rb_alloc_aux(rb
, event
, vma
->vm_pgoff
, nr_pages
, flags
);
4596 rb
->aux_mmap_locked
= extra
;
4601 atomic_long_add(user_extra
, &user
->locked_vm
);
4602 vma
->vm_mm
->pinned_vm
+= extra
;
4604 atomic_inc(&event
->mmap_count
);
4606 atomic_dec(&rb
->mmap_count
);
4609 mutex_unlock(&event
->mmap_mutex
);
4612 * Since pinned accounting is per vm we cannot allow fork() to copy our
4615 vma
->vm_flags
|= VM_DONTCOPY
| VM_DONTEXPAND
| VM_DONTDUMP
;
4616 vma
->vm_ops
= &perf_mmap_vmops
;
4618 if (event
->pmu
->event_mapped
)
4619 event
->pmu
->event_mapped(event
);
4624 static int perf_fasync(int fd
, struct file
*filp
, int on
)
4626 struct inode
*inode
= file_inode(filp
);
4627 struct perf_event
*event
= filp
->private_data
;
4630 mutex_lock(&inode
->i_mutex
);
4631 retval
= fasync_helper(fd
, filp
, on
, &event
->fasync
);
4632 mutex_unlock(&inode
->i_mutex
);
4640 static const struct file_operations perf_fops
= {
4641 .llseek
= no_llseek
,
4642 .release
= perf_release
,
4645 .unlocked_ioctl
= perf_ioctl
,
4646 .compat_ioctl
= perf_compat_ioctl
,
4648 .fasync
= perf_fasync
,
4654 * If there's data, ensure we set the poll() state and publish everything
4655 * to user-space before waking everybody up.
4658 void perf_event_wakeup(struct perf_event
*event
)
4660 ring_buffer_wakeup(event
);
4662 if (event
->pending_kill
) {
4663 kill_fasync(&event
->fasync
, SIGIO
, event
->pending_kill
);
4664 event
->pending_kill
= 0;
4668 static void perf_pending_event(struct irq_work
*entry
)
4670 struct perf_event
*event
= container_of(entry
,
4671 struct perf_event
, pending
);
4674 rctx
= perf_swevent_get_recursion_context();
4676 * If we 'fail' here, that's OK, it means recursion is already disabled
4677 * and we won't recurse 'further'.
4680 if (event
->pending_disable
) {
4681 event
->pending_disable
= 0;
4682 __perf_event_disable(event
);
4685 if (event
->pending_wakeup
) {
4686 event
->pending_wakeup
= 0;
4687 perf_event_wakeup(event
);
4691 perf_swevent_put_recursion_context(rctx
);
4695 * We assume there is only KVM supporting the callbacks.
4696 * Later on, we might change it to a list if there is
4697 * another virtualization implementation supporting the callbacks.
4699 struct perf_guest_info_callbacks
*perf_guest_cbs
;
4701 int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks
*cbs
)
4703 perf_guest_cbs
= cbs
;
4706 EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks
);
4708 int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks
*cbs
)
4710 perf_guest_cbs
= NULL
;
4713 EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks
);
4716 perf_output_sample_regs(struct perf_output_handle
*handle
,
4717 struct pt_regs
*regs
, u64 mask
)
4721 for_each_set_bit(bit
, (const unsigned long *) &mask
,
4722 sizeof(mask
) * BITS_PER_BYTE
) {
4725 val
= perf_reg_value(regs
, bit
);
4726 perf_output_put(handle
, val
);
4730 static void perf_sample_regs_user(struct perf_regs
*regs_user
,
4731 struct pt_regs
*regs
,
4732 struct pt_regs
*regs_user_copy
)
4734 if (user_mode(regs
)) {
4735 regs_user
->abi
= perf_reg_abi(current
);
4736 regs_user
->regs
= regs
;
4737 } else if (current
->mm
) {
4738 perf_get_regs_user(regs_user
, regs
, regs_user_copy
);
4740 regs_user
->abi
= PERF_SAMPLE_REGS_ABI_NONE
;
4741 regs_user
->regs
= NULL
;
4745 static void perf_sample_regs_intr(struct perf_regs
*regs_intr
,
4746 struct pt_regs
*regs
)
4748 regs_intr
->regs
= regs
;
4749 regs_intr
->abi
= perf_reg_abi(current
);
4754 * Get remaining task size from user stack pointer.
4756 * It'd be better to take stack vma map and limit this more
4757 * precisly, but there's no way to get it safely under interrupt,
4758 * so using TASK_SIZE as limit.
4760 static u64
perf_ustack_task_size(struct pt_regs
*regs
)
4762 unsigned long addr
= perf_user_stack_pointer(regs
);
4764 if (!addr
|| addr
>= TASK_SIZE
)
4767 return TASK_SIZE
- addr
;
4771 perf_sample_ustack_size(u16 stack_size
, u16 header_size
,
4772 struct pt_regs
*regs
)
4776 /* No regs, no stack pointer, no dump. */
4781 * Check if we fit in with the requested stack size into the:
4783 * If we don't, we limit the size to the TASK_SIZE.
4785 * - remaining sample size
4786 * If we don't, we customize the stack size to
4787 * fit in to the remaining sample size.
4790 task_size
= min((u64
) USHRT_MAX
, perf_ustack_task_size(regs
));
4791 stack_size
= min(stack_size
, (u16
) task_size
);
4793 /* Current header size plus static size and dynamic size. */
4794 header_size
+= 2 * sizeof(u64
);
4796 /* Do we fit in with the current stack dump size? */
4797 if ((u16
) (header_size
+ stack_size
) < header_size
) {
4799 * If we overflow the maximum size for the sample,
4800 * we customize the stack dump size to fit in.
4802 stack_size
= USHRT_MAX
- header_size
- sizeof(u64
);
4803 stack_size
= round_up(stack_size
, sizeof(u64
));
4810 perf_output_sample_ustack(struct perf_output_handle
*handle
, u64 dump_size
,
4811 struct pt_regs
*regs
)
4813 /* Case of a kernel thread, nothing to dump */
4816 perf_output_put(handle
, size
);
4825 * - the size requested by user or the best one we can fit
4826 * in to the sample max size
4828 * - user stack dump data
4830 * - the actual dumped size
4834 perf_output_put(handle
, dump_size
);
4837 sp
= perf_user_stack_pointer(regs
);
4838 rem
= __output_copy_user(handle
, (void *) sp
, dump_size
);
4839 dyn_size
= dump_size
- rem
;
4841 perf_output_skip(handle
, rem
);
4844 perf_output_put(handle
, dyn_size
);
4848 static void __perf_event_header__init_id(struct perf_event_header
*header
,
4849 struct perf_sample_data
*data
,
4850 struct perf_event
*event
)
4852 u64 sample_type
= event
->attr
.sample_type
;
4854 data
->type
= sample_type
;
4855 header
->size
+= event
->id_header_size
;
4857 if (sample_type
& PERF_SAMPLE_TID
) {
4858 /* namespace issues */
4859 data
->tid_entry
.pid
= perf_event_pid(event
, current
);
4860 data
->tid_entry
.tid
= perf_event_tid(event
, current
);
4863 if (sample_type
& PERF_SAMPLE_TIME
)
4864 data
->time
= perf_event_clock(event
);
4866 if (sample_type
& (PERF_SAMPLE_ID
| PERF_SAMPLE_IDENTIFIER
))
4867 data
->id
= primary_event_id(event
);
4869 if (sample_type
& PERF_SAMPLE_STREAM_ID
)
4870 data
->stream_id
= event
->id
;
4872 if (sample_type
& PERF_SAMPLE_CPU
) {
4873 data
->cpu_entry
.cpu
= raw_smp_processor_id();
4874 data
->cpu_entry
.reserved
= 0;
4878 void perf_event_header__init_id(struct perf_event_header
*header
,
4879 struct perf_sample_data
*data
,
4880 struct perf_event
*event
)
4882 if (event
->attr
.sample_id_all
)
4883 __perf_event_header__init_id(header
, data
, event
);
4886 static void __perf_event__output_id_sample(struct perf_output_handle
*handle
,
4887 struct perf_sample_data
*data
)
4889 u64 sample_type
= data
->type
;
4891 if (sample_type
& PERF_SAMPLE_TID
)
4892 perf_output_put(handle
, data
->tid_entry
);
4894 if (sample_type
& PERF_SAMPLE_TIME
)
4895 perf_output_put(handle
, data
->time
);
4897 if (sample_type
& PERF_SAMPLE_ID
)
4898 perf_output_put(handle
, data
->id
);
4900 if (sample_type
& PERF_SAMPLE_STREAM_ID
)
4901 perf_output_put(handle
, data
->stream_id
);
4903 if (sample_type
& PERF_SAMPLE_CPU
)
4904 perf_output_put(handle
, data
->cpu_entry
);
4906 if (sample_type
& PERF_SAMPLE_IDENTIFIER
)
4907 perf_output_put(handle
, data
->id
);
4910 void perf_event__output_id_sample(struct perf_event
*event
,
4911 struct perf_output_handle
*handle
,
4912 struct perf_sample_data
*sample
)
4914 if (event
->attr
.sample_id_all
)
4915 __perf_event__output_id_sample(handle
, sample
);
4918 static void perf_output_read_one(struct perf_output_handle
*handle
,
4919 struct perf_event
*event
,
4920 u64 enabled
, u64 running
)
4922 u64 read_format
= event
->attr
.read_format
;
4926 values
[n
++] = perf_event_count(event
);
4927 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
) {
4928 values
[n
++] = enabled
+
4929 atomic64_read(&event
->child_total_time_enabled
);
4931 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
) {
4932 values
[n
++] = running
+
4933 atomic64_read(&event
->child_total_time_running
);
4935 if (read_format
& PERF_FORMAT_ID
)
4936 values
[n
++] = primary_event_id(event
);
4938 __output_copy(handle
, values
, n
* sizeof(u64
));
4942 * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
4944 static void perf_output_read_group(struct perf_output_handle
*handle
,
4945 struct perf_event
*event
,
4946 u64 enabled
, u64 running
)
4948 struct perf_event
*leader
= event
->group_leader
, *sub
;
4949 u64 read_format
= event
->attr
.read_format
;
4953 values
[n
++] = 1 + leader
->nr_siblings
;
4955 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
4956 values
[n
++] = enabled
;
4958 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
4959 values
[n
++] = running
;
4961 if (leader
!= event
)
4962 leader
->pmu
->read(leader
);
4964 values
[n
++] = perf_event_count(leader
);
4965 if (read_format
& PERF_FORMAT_ID
)
4966 values
[n
++] = primary_event_id(leader
);
4968 __output_copy(handle
, values
, n
* sizeof(u64
));
4970 list_for_each_entry(sub
, &leader
->sibling_list
, group_entry
) {
4973 if ((sub
!= event
) &&
4974 (sub
->state
== PERF_EVENT_STATE_ACTIVE
))
4975 sub
->pmu
->read(sub
);
4977 values
[n
++] = perf_event_count(sub
);
4978 if (read_format
& PERF_FORMAT_ID
)
4979 values
[n
++] = primary_event_id(sub
);
4981 __output_copy(handle
, values
, n
* sizeof(u64
));
4985 #define PERF_FORMAT_TOTAL_TIMES (PERF_FORMAT_TOTAL_TIME_ENABLED|\
4986 PERF_FORMAT_TOTAL_TIME_RUNNING)
4988 static void perf_output_read(struct perf_output_handle
*handle
,
4989 struct perf_event
*event
)
4991 u64 enabled
= 0, running
= 0, now
;
4992 u64 read_format
= event
->attr
.read_format
;
4995 * compute total_time_enabled, total_time_running
4996 * based on snapshot values taken when the event
4997 * was last scheduled in.
4999 * we cannot simply called update_context_time()
5000 * because of locking issue as we are called in
5003 if (read_format
& PERF_FORMAT_TOTAL_TIMES
)
5004 calc_timer_values(event
, &now
, &enabled
, &running
);
5006 if (event
->attr
.read_format
& PERF_FORMAT_GROUP
)
5007 perf_output_read_group(handle
, event
, enabled
, running
);
5009 perf_output_read_one(handle
, event
, enabled
, running
);
5012 void perf_output_sample(struct perf_output_handle
*handle
,
5013 struct perf_event_header
*header
,
5014 struct perf_sample_data
*data
,
5015 struct perf_event
*event
)
5017 u64 sample_type
= data
->type
;
5019 perf_output_put(handle
, *header
);
5021 if (sample_type
& PERF_SAMPLE_IDENTIFIER
)
5022 perf_output_put(handle
, data
->id
);
5024 if (sample_type
& PERF_SAMPLE_IP
)
5025 perf_output_put(handle
, data
->ip
);
5027 if (sample_type
& PERF_SAMPLE_TID
)
5028 perf_output_put(handle
, data
->tid_entry
);
5030 if (sample_type
& PERF_SAMPLE_TIME
)
5031 perf_output_put(handle
, data
->time
);
5033 if (sample_type
& PERF_SAMPLE_ADDR
)
5034 perf_output_put(handle
, data
->addr
);
5036 if (sample_type
& PERF_SAMPLE_ID
)
5037 perf_output_put(handle
, data
->id
);
5039 if (sample_type
& PERF_SAMPLE_STREAM_ID
)
5040 perf_output_put(handle
, data
->stream_id
);
5042 if (sample_type
& PERF_SAMPLE_CPU
)
5043 perf_output_put(handle
, data
->cpu_entry
);
5045 if (sample_type
& PERF_SAMPLE_PERIOD
)
5046 perf_output_put(handle
, data
->period
);
5048 if (sample_type
& PERF_SAMPLE_READ
)
5049 perf_output_read(handle
, event
);
5051 if (sample_type
& PERF_SAMPLE_CALLCHAIN
) {
5052 if (data
->callchain
) {
5055 if (data
->callchain
)
5056 size
+= data
->callchain
->nr
;
5058 size
*= sizeof(u64
);
5060 __output_copy(handle
, data
->callchain
, size
);
5063 perf_output_put(handle
, nr
);
5067 if (sample_type
& PERF_SAMPLE_RAW
) {
5069 perf_output_put(handle
, data
->raw
->size
);
5070 __output_copy(handle
, data
->raw
->data
,
5077 .size
= sizeof(u32
),
5080 perf_output_put(handle
, raw
);
5084 if (sample_type
& PERF_SAMPLE_BRANCH_STACK
) {
5085 if (data
->br_stack
) {
5088 size
= data
->br_stack
->nr
5089 * sizeof(struct perf_branch_entry
);
5091 perf_output_put(handle
, data
->br_stack
->nr
);
5092 perf_output_copy(handle
, data
->br_stack
->entries
, size
);
5095 * we always store at least the value of nr
5098 perf_output_put(handle
, nr
);
5102 if (sample_type
& PERF_SAMPLE_REGS_USER
) {
5103 u64 abi
= data
->regs_user
.abi
;
5106 * If there are no regs to dump, notice it through
5107 * first u64 being zero (PERF_SAMPLE_REGS_ABI_NONE).
5109 perf_output_put(handle
, abi
);
5112 u64 mask
= event
->attr
.sample_regs_user
;
5113 perf_output_sample_regs(handle
,
5114 data
->regs_user
.regs
,
5119 if (sample_type
& PERF_SAMPLE_STACK_USER
) {
5120 perf_output_sample_ustack(handle
,
5121 data
->stack_user_size
,
5122 data
->regs_user
.regs
);
5125 if (sample_type
& PERF_SAMPLE_WEIGHT
)
5126 perf_output_put(handle
, data
->weight
);
5128 if (sample_type
& PERF_SAMPLE_DATA_SRC
)
5129 perf_output_put(handle
, data
->data_src
.val
);
5131 if (sample_type
& PERF_SAMPLE_TRANSACTION
)
5132 perf_output_put(handle
, data
->txn
);
5134 if (sample_type
& PERF_SAMPLE_REGS_INTR
) {
5135 u64 abi
= data
->regs_intr
.abi
;
5137 * If there are no regs to dump, notice it through
5138 * first u64 being zero (PERF_SAMPLE_REGS_ABI_NONE).
5140 perf_output_put(handle
, abi
);
5143 u64 mask
= event
->attr
.sample_regs_intr
;
5145 perf_output_sample_regs(handle
,
5146 data
->regs_intr
.regs
,
5151 if (!event
->attr
.watermark
) {
5152 int wakeup_events
= event
->attr
.wakeup_events
;
5154 if (wakeup_events
) {
5155 struct ring_buffer
*rb
= handle
->rb
;
5156 int events
= local_inc_return(&rb
->events
);
5158 if (events
>= wakeup_events
) {
5159 local_sub(wakeup_events
, &rb
->events
);
5160 local_inc(&rb
->wakeup
);
5166 void perf_prepare_sample(struct perf_event_header
*header
,
5167 struct perf_sample_data
*data
,
5168 struct perf_event
*event
,
5169 struct pt_regs
*regs
)
5171 u64 sample_type
= event
->attr
.sample_type
;
5173 header
->type
= PERF_RECORD_SAMPLE
;
5174 header
->size
= sizeof(*header
) + event
->header_size
;
5177 header
->misc
|= perf_misc_flags(regs
);
5179 __perf_event_header__init_id(header
, data
, event
);
5181 if (sample_type
& PERF_SAMPLE_IP
)
5182 data
->ip
= perf_instruction_pointer(regs
);
5184 if (sample_type
& PERF_SAMPLE_CALLCHAIN
) {
5187 data
->callchain
= perf_callchain(event
, regs
);
5189 if (data
->callchain
)
5190 size
+= data
->callchain
->nr
;
5192 header
->size
+= size
* sizeof(u64
);
5195 if (sample_type
& PERF_SAMPLE_RAW
) {
5196 int size
= sizeof(u32
);
5199 size
+= data
->raw
->size
;
5201 size
+= sizeof(u32
);
5203 WARN_ON_ONCE(size
& (sizeof(u64
)-1));
5204 header
->size
+= size
;
5207 if (sample_type
& PERF_SAMPLE_BRANCH_STACK
) {
5208 int size
= sizeof(u64
); /* nr */
5209 if (data
->br_stack
) {
5210 size
+= data
->br_stack
->nr
5211 * sizeof(struct perf_branch_entry
);
5213 header
->size
+= size
;
5216 if (sample_type
& (PERF_SAMPLE_REGS_USER
| PERF_SAMPLE_STACK_USER
))
5217 perf_sample_regs_user(&data
->regs_user
, regs
,
5218 &data
->regs_user_copy
);
5220 if (sample_type
& PERF_SAMPLE_REGS_USER
) {
5221 /* regs dump ABI info */
5222 int size
= sizeof(u64
);
5224 if (data
->regs_user
.regs
) {
5225 u64 mask
= event
->attr
.sample_regs_user
;
5226 size
+= hweight64(mask
) * sizeof(u64
);
5229 header
->size
+= size
;
5232 if (sample_type
& PERF_SAMPLE_STACK_USER
) {
5234 * Either we need PERF_SAMPLE_STACK_USER bit to be allways
5235 * processed as the last one or have additional check added
5236 * in case new sample type is added, because we could eat
5237 * up the rest of the sample size.
5239 u16 stack_size
= event
->attr
.sample_stack_user
;
5240 u16 size
= sizeof(u64
);
5242 stack_size
= perf_sample_ustack_size(stack_size
, header
->size
,
5243 data
->regs_user
.regs
);
5246 * If there is something to dump, add space for the dump
5247 * itself and for the field that tells the dynamic size,
5248 * which is how many have been actually dumped.
5251 size
+= sizeof(u64
) + stack_size
;
5253 data
->stack_user_size
= stack_size
;
5254 header
->size
+= size
;
5257 if (sample_type
& PERF_SAMPLE_REGS_INTR
) {
5258 /* regs dump ABI info */
5259 int size
= sizeof(u64
);
5261 perf_sample_regs_intr(&data
->regs_intr
, regs
);
5263 if (data
->regs_intr
.regs
) {
5264 u64 mask
= event
->attr
.sample_regs_intr
;
5266 size
+= hweight64(mask
) * sizeof(u64
);
5269 header
->size
+= size
;
5273 static void perf_event_output(struct perf_event
*event
,
5274 struct perf_sample_data
*data
,
5275 struct pt_regs
*regs
)
5277 struct perf_output_handle handle
;
5278 struct perf_event_header header
;
5280 /* protect the callchain buffers */
5283 perf_prepare_sample(&header
, data
, event
, regs
);
5285 if (perf_output_begin(&handle
, event
, header
.size
))
5288 perf_output_sample(&handle
, &header
, data
, event
);
5290 perf_output_end(&handle
);
5300 struct perf_read_event
{
5301 struct perf_event_header header
;
5308 perf_event_read_event(struct perf_event
*event
,
5309 struct task_struct
*task
)
5311 struct perf_output_handle handle
;
5312 struct perf_sample_data sample
;
5313 struct perf_read_event read_event
= {
5315 .type
= PERF_RECORD_READ
,
5317 .size
= sizeof(read_event
) + event
->read_size
,
5319 .pid
= perf_event_pid(event
, task
),
5320 .tid
= perf_event_tid(event
, task
),
5324 perf_event_header__init_id(&read_event
.header
, &sample
, event
);
5325 ret
= perf_output_begin(&handle
, event
, read_event
.header
.size
);
5329 perf_output_put(&handle
, read_event
);
5330 perf_output_read(&handle
, event
);
5331 perf_event__output_id_sample(event
, &handle
, &sample
);
5333 perf_output_end(&handle
);
5336 typedef void (perf_event_aux_output_cb
)(struct perf_event
*event
, void *data
);
5339 perf_event_aux_ctx(struct perf_event_context
*ctx
,
5340 perf_event_aux_output_cb output
,
5343 struct perf_event
*event
;
5345 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
5346 if (event
->state
< PERF_EVENT_STATE_INACTIVE
)
5348 if (!event_filter_match(event
))
5350 output(event
, data
);
5355 perf_event_aux(perf_event_aux_output_cb output
, void *data
,
5356 struct perf_event_context
*task_ctx
)
5358 struct perf_cpu_context
*cpuctx
;
5359 struct perf_event_context
*ctx
;
5364 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
5365 cpuctx
= get_cpu_ptr(pmu
->pmu_cpu_context
);
5366 if (cpuctx
->unique_pmu
!= pmu
)
5368 perf_event_aux_ctx(&cpuctx
->ctx
, output
, data
);
5371 ctxn
= pmu
->task_ctx_nr
;
5374 ctx
= rcu_dereference(current
->perf_event_ctxp
[ctxn
]);
5376 perf_event_aux_ctx(ctx
, output
, data
);
5378 put_cpu_ptr(pmu
->pmu_cpu_context
);
5383 perf_event_aux_ctx(task_ctx
, output
, data
);
5390 * task tracking -- fork/exit
5392 * enabled by: attr.comm | attr.mmap | attr.mmap2 | attr.mmap_data | attr.task
5395 struct perf_task_event
{
5396 struct task_struct
*task
;
5397 struct perf_event_context
*task_ctx
;
5400 struct perf_event_header header
;
5410 static int perf_event_task_match(struct perf_event
*event
)
5412 return event
->attr
.comm
|| event
->attr
.mmap
||
5413 event
->attr
.mmap2
|| event
->attr
.mmap_data
||
5417 static void perf_event_task_output(struct perf_event
*event
,
5420 struct perf_task_event
*task_event
= data
;
5421 struct perf_output_handle handle
;
5422 struct perf_sample_data sample
;
5423 struct task_struct
*task
= task_event
->task
;
5424 int ret
, size
= task_event
->event_id
.header
.size
;
5426 if (!perf_event_task_match(event
))
5429 perf_event_header__init_id(&task_event
->event_id
.header
, &sample
, event
);
5431 ret
= perf_output_begin(&handle
, event
,
5432 task_event
->event_id
.header
.size
);
5436 task_event
->event_id
.pid
= perf_event_pid(event
, task
);
5437 task_event
->event_id
.ppid
= perf_event_pid(event
, current
);
5439 task_event
->event_id
.tid
= perf_event_tid(event
, task
);
5440 task_event
->event_id
.ptid
= perf_event_tid(event
, current
);
5442 task_event
->event_id
.time
= perf_event_clock(event
);
5444 perf_output_put(&handle
, task_event
->event_id
);
5446 perf_event__output_id_sample(event
, &handle
, &sample
);
5448 perf_output_end(&handle
);
5450 task_event
->event_id
.header
.size
= size
;
5453 static void perf_event_task(struct task_struct
*task
,
5454 struct perf_event_context
*task_ctx
,
5457 struct perf_task_event task_event
;
5459 if (!atomic_read(&nr_comm_events
) &&
5460 !atomic_read(&nr_mmap_events
) &&
5461 !atomic_read(&nr_task_events
))
5464 task_event
= (struct perf_task_event
){
5466 .task_ctx
= task_ctx
,
5469 .type
= new ? PERF_RECORD_FORK
: PERF_RECORD_EXIT
,
5471 .size
= sizeof(task_event
.event_id
),
5481 perf_event_aux(perf_event_task_output
,
5486 void perf_event_fork(struct task_struct
*task
)
5488 perf_event_task(task
, NULL
, 1);
5495 struct perf_comm_event
{
5496 struct task_struct
*task
;
5501 struct perf_event_header header
;
5508 static int perf_event_comm_match(struct perf_event
*event
)
5510 return event
->attr
.comm
;
5513 static void perf_event_comm_output(struct perf_event
*event
,
5516 struct perf_comm_event
*comm_event
= data
;
5517 struct perf_output_handle handle
;
5518 struct perf_sample_data sample
;
5519 int size
= comm_event
->event_id
.header
.size
;
5522 if (!perf_event_comm_match(event
))
5525 perf_event_header__init_id(&comm_event
->event_id
.header
, &sample
, event
);
5526 ret
= perf_output_begin(&handle
, event
,
5527 comm_event
->event_id
.header
.size
);
5532 comm_event
->event_id
.pid
= perf_event_pid(event
, comm_event
->task
);
5533 comm_event
->event_id
.tid
= perf_event_tid(event
, comm_event
->task
);
5535 perf_output_put(&handle
, comm_event
->event_id
);
5536 __output_copy(&handle
, comm_event
->comm
,
5537 comm_event
->comm_size
);
5539 perf_event__output_id_sample(event
, &handle
, &sample
);
5541 perf_output_end(&handle
);
5543 comm_event
->event_id
.header
.size
= size
;
5546 static void perf_event_comm_event(struct perf_comm_event
*comm_event
)
5548 char comm
[TASK_COMM_LEN
];
5551 memset(comm
, 0, sizeof(comm
));
5552 strlcpy(comm
, comm_event
->task
->comm
, sizeof(comm
));
5553 size
= ALIGN(strlen(comm
)+1, sizeof(u64
));
5555 comm_event
->comm
= comm
;
5556 comm_event
->comm_size
= size
;
5558 comm_event
->event_id
.header
.size
= sizeof(comm_event
->event_id
) + size
;
5560 perf_event_aux(perf_event_comm_output
,
5565 void perf_event_comm(struct task_struct
*task
, bool exec
)
5567 struct perf_comm_event comm_event
;
5569 if (!atomic_read(&nr_comm_events
))
5572 comm_event
= (struct perf_comm_event
){
5578 .type
= PERF_RECORD_COMM
,
5579 .misc
= exec
? PERF_RECORD_MISC_COMM_EXEC
: 0,
5587 perf_event_comm_event(&comm_event
);
5594 struct perf_mmap_event
{
5595 struct vm_area_struct
*vma
;
5597 const char *file_name
;
5605 struct perf_event_header header
;
5615 static int perf_event_mmap_match(struct perf_event
*event
,
5618 struct perf_mmap_event
*mmap_event
= data
;
5619 struct vm_area_struct
*vma
= mmap_event
->vma
;
5620 int executable
= vma
->vm_flags
& VM_EXEC
;
5622 return (!executable
&& event
->attr
.mmap_data
) ||
5623 (executable
&& (event
->attr
.mmap
|| event
->attr
.mmap2
));
5626 static void perf_event_mmap_output(struct perf_event
*event
,
5629 struct perf_mmap_event
*mmap_event
= data
;
5630 struct perf_output_handle handle
;
5631 struct perf_sample_data sample
;
5632 int size
= mmap_event
->event_id
.header
.size
;
5635 if (!perf_event_mmap_match(event
, data
))
5638 if (event
->attr
.mmap2
) {
5639 mmap_event
->event_id
.header
.type
= PERF_RECORD_MMAP2
;
5640 mmap_event
->event_id
.header
.size
+= sizeof(mmap_event
->maj
);
5641 mmap_event
->event_id
.header
.size
+= sizeof(mmap_event
->min
);
5642 mmap_event
->event_id
.header
.size
+= sizeof(mmap_event
->ino
);
5643 mmap_event
->event_id
.header
.size
+= sizeof(mmap_event
->ino_generation
);
5644 mmap_event
->event_id
.header
.size
+= sizeof(mmap_event
->prot
);
5645 mmap_event
->event_id
.header
.size
+= sizeof(mmap_event
->flags
);
5648 perf_event_header__init_id(&mmap_event
->event_id
.header
, &sample
, event
);
5649 ret
= perf_output_begin(&handle
, event
,
5650 mmap_event
->event_id
.header
.size
);
5654 mmap_event
->event_id
.pid
= perf_event_pid(event
, current
);
5655 mmap_event
->event_id
.tid
= perf_event_tid(event
, current
);
5657 perf_output_put(&handle
, mmap_event
->event_id
);
5659 if (event
->attr
.mmap2
) {
5660 perf_output_put(&handle
, mmap_event
->maj
);
5661 perf_output_put(&handle
, mmap_event
->min
);
5662 perf_output_put(&handle
, mmap_event
->ino
);
5663 perf_output_put(&handle
, mmap_event
->ino_generation
);
5664 perf_output_put(&handle
, mmap_event
->prot
);
5665 perf_output_put(&handle
, mmap_event
->flags
);
5668 __output_copy(&handle
, mmap_event
->file_name
,
5669 mmap_event
->file_size
);
5671 perf_event__output_id_sample(event
, &handle
, &sample
);
5673 perf_output_end(&handle
);
5675 mmap_event
->event_id
.header
.size
= size
;
5678 static void perf_event_mmap_event(struct perf_mmap_event
*mmap_event
)
5680 struct vm_area_struct
*vma
= mmap_event
->vma
;
5681 struct file
*file
= vma
->vm_file
;
5682 int maj
= 0, min
= 0;
5683 u64 ino
= 0, gen
= 0;
5684 u32 prot
= 0, flags
= 0;
5691 struct inode
*inode
;
5694 buf
= kmalloc(PATH_MAX
, GFP_KERNEL
);
5700 * d_path() works from the end of the rb backwards, so we
5701 * need to add enough zero bytes after the string to handle
5702 * the 64bit alignment we do later.
5704 name
= d_path(&file
->f_path
, buf
, PATH_MAX
- sizeof(u64
));
5709 inode
= file_inode(vma
->vm_file
);
5710 dev
= inode
->i_sb
->s_dev
;
5712 gen
= inode
->i_generation
;
5716 if (vma
->vm_flags
& VM_READ
)
5718 if (vma
->vm_flags
& VM_WRITE
)
5720 if (vma
->vm_flags
& VM_EXEC
)
5723 if (vma
->vm_flags
& VM_MAYSHARE
)
5726 flags
= MAP_PRIVATE
;
5728 if (vma
->vm_flags
& VM_DENYWRITE
)
5729 flags
|= MAP_DENYWRITE
;
5730 if (vma
->vm_flags
& VM_MAYEXEC
)
5731 flags
|= MAP_EXECUTABLE
;
5732 if (vma
->vm_flags
& VM_LOCKED
)
5733 flags
|= MAP_LOCKED
;
5734 if (vma
->vm_flags
& VM_HUGETLB
)
5735 flags
|= MAP_HUGETLB
;
5739 if (vma
->vm_ops
&& vma
->vm_ops
->name
) {
5740 name
= (char *) vma
->vm_ops
->name(vma
);
5745 name
= (char *)arch_vma_name(vma
);
5749 if (vma
->vm_start
<= vma
->vm_mm
->start_brk
&&
5750 vma
->vm_end
>= vma
->vm_mm
->brk
) {
5754 if (vma
->vm_start
<= vma
->vm_mm
->start_stack
&&
5755 vma
->vm_end
>= vma
->vm_mm
->start_stack
) {
5765 strlcpy(tmp
, name
, sizeof(tmp
));
5769 * Since our buffer works in 8 byte units we need to align our string
5770 * size to a multiple of 8. However, we must guarantee the tail end is
5771 * zero'd out to avoid leaking random bits to userspace.
5773 size
= strlen(name
)+1;
5774 while (!IS_ALIGNED(size
, sizeof(u64
)))
5775 name
[size
++] = '\0';
5777 mmap_event
->file_name
= name
;
5778 mmap_event
->file_size
= size
;
5779 mmap_event
->maj
= maj
;
5780 mmap_event
->min
= min
;
5781 mmap_event
->ino
= ino
;
5782 mmap_event
->ino_generation
= gen
;
5783 mmap_event
->prot
= prot
;
5784 mmap_event
->flags
= flags
;
5786 if (!(vma
->vm_flags
& VM_EXEC
))
5787 mmap_event
->event_id
.header
.misc
|= PERF_RECORD_MISC_MMAP_DATA
;
5789 mmap_event
->event_id
.header
.size
= sizeof(mmap_event
->event_id
) + size
;
5791 perf_event_aux(perf_event_mmap_output
,
5798 void perf_event_mmap(struct vm_area_struct
*vma
)
5800 struct perf_mmap_event mmap_event
;
5802 if (!atomic_read(&nr_mmap_events
))
5805 mmap_event
= (struct perf_mmap_event
){
5811 .type
= PERF_RECORD_MMAP
,
5812 .misc
= PERF_RECORD_MISC_USER
,
5817 .start
= vma
->vm_start
,
5818 .len
= vma
->vm_end
- vma
->vm_start
,
5819 .pgoff
= (u64
)vma
->vm_pgoff
<< PAGE_SHIFT
,
5821 /* .maj (attr_mmap2 only) */
5822 /* .min (attr_mmap2 only) */
5823 /* .ino (attr_mmap2 only) */
5824 /* .ino_generation (attr_mmap2 only) */
5825 /* .prot (attr_mmap2 only) */
5826 /* .flags (attr_mmap2 only) */
5829 perf_event_mmap_event(&mmap_event
);
5833 * IRQ throttle logging
5836 static void perf_log_throttle(struct perf_event
*event
, int enable
)
5838 struct perf_output_handle handle
;
5839 struct perf_sample_data sample
;
5843 struct perf_event_header header
;
5847 } throttle_event
= {
5849 .type
= PERF_RECORD_THROTTLE
,
5851 .size
= sizeof(throttle_event
),
5853 .time
= perf_event_clock(event
),
5854 .id
= primary_event_id(event
),
5855 .stream_id
= event
->id
,
5859 throttle_event
.header
.type
= PERF_RECORD_UNTHROTTLE
;
5861 perf_event_header__init_id(&throttle_event
.header
, &sample
, event
);
5863 ret
= perf_output_begin(&handle
, event
,
5864 throttle_event
.header
.size
);
5868 perf_output_put(&handle
, throttle_event
);
5869 perf_event__output_id_sample(event
, &handle
, &sample
);
5870 perf_output_end(&handle
);
5874 * Generic event overflow handling, sampling.
5877 static int __perf_event_overflow(struct perf_event
*event
,
5878 int throttle
, struct perf_sample_data
*data
,
5879 struct pt_regs
*regs
)
5881 int events
= atomic_read(&event
->event_limit
);
5882 struct hw_perf_event
*hwc
= &event
->hw
;
5887 * Non-sampling counters might still use the PMI to fold short
5888 * hardware counters, ignore those.
5890 if (unlikely(!is_sampling_event(event
)))
5893 seq
= __this_cpu_read(perf_throttled_seq
);
5894 if (seq
!= hwc
->interrupts_seq
) {
5895 hwc
->interrupts_seq
= seq
;
5896 hwc
->interrupts
= 1;
5899 if (unlikely(throttle
5900 && hwc
->interrupts
>= max_samples_per_tick
)) {
5901 __this_cpu_inc(perf_throttled_count
);
5902 hwc
->interrupts
= MAX_INTERRUPTS
;
5903 perf_log_throttle(event
, 0);
5904 tick_nohz_full_kick();
5909 if (event
->attr
.freq
) {
5910 u64 now
= perf_clock();
5911 s64 delta
= now
- hwc
->freq_time_stamp
;
5913 hwc
->freq_time_stamp
= now
;
5915 if (delta
> 0 && delta
< 2*TICK_NSEC
)
5916 perf_adjust_period(event
, delta
, hwc
->last_period
, true);
5920 * XXX event_limit might not quite work as expected on inherited
5924 event
->pending_kill
= POLL_IN
;
5925 if (events
&& atomic_dec_and_test(&event
->event_limit
)) {
5927 event
->pending_kill
= POLL_HUP
;
5928 event
->pending_disable
= 1;
5929 irq_work_queue(&event
->pending
);
5932 if (event
->overflow_handler
)
5933 event
->overflow_handler(event
, data
, regs
);
5935 perf_event_output(event
, data
, regs
);
5937 if (event
->fasync
&& event
->pending_kill
) {
5938 event
->pending_wakeup
= 1;
5939 irq_work_queue(&event
->pending
);
5945 int perf_event_overflow(struct perf_event
*event
,
5946 struct perf_sample_data
*data
,
5947 struct pt_regs
*regs
)
5949 return __perf_event_overflow(event
, 1, data
, regs
);
5953 * Generic software event infrastructure
5956 struct swevent_htable
{
5957 struct swevent_hlist
*swevent_hlist
;
5958 struct mutex hlist_mutex
;
5961 /* Recursion avoidance in each contexts */
5962 int recursion
[PERF_NR_CONTEXTS
];
5964 /* Keeps track of cpu being initialized/exited */
5968 static DEFINE_PER_CPU(struct swevent_htable
, swevent_htable
);
5971 * We directly increment event->count and keep a second value in
5972 * event->hw.period_left to count intervals. This period event
5973 * is kept in the range [-sample_period, 0] so that we can use the
5977 u64
perf_swevent_set_period(struct perf_event
*event
)
5979 struct hw_perf_event
*hwc
= &event
->hw
;
5980 u64 period
= hwc
->last_period
;
5984 hwc
->last_period
= hwc
->sample_period
;
5987 old
= val
= local64_read(&hwc
->period_left
);
5991 nr
= div64_u64(period
+ val
, period
);
5992 offset
= nr
* period
;
5994 if (local64_cmpxchg(&hwc
->period_left
, old
, val
) != old
)
6000 static void perf_swevent_overflow(struct perf_event
*event
, u64 overflow
,
6001 struct perf_sample_data
*data
,
6002 struct pt_regs
*regs
)
6004 struct hw_perf_event
*hwc
= &event
->hw
;
6008 overflow
= perf_swevent_set_period(event
);
6010 if (hwc
->interrupts
== MAX_INTERRUPTS
)
6013 for (; overflow
; overflow
--) {
6014 if (__perf_event_overflow(event
, throttle
,
6017 * We inhibit the overflow from happening when
6018 * hwc->interrupts == MAX_INTERRUPTS.
6026 static void perf_swevent_event(struct perf_event
*event
, u64 nr
,
6027 struct perf_sample_data
*data
,
6028 struct pt_regs
*regs
)
6030 struct hw_perf_event
*hwc
= &event
->hw
;
6032 local64_add(nr
, &event
->count
);
6037 if (!is_sampling_event(event
))
6040 if ((event
->attr
.sample_type
& PERF_SAMPLE_PERIOD
) && !event
->attr
.freq
) {
6042 return perf_swevent_overflow(event
, 1, data
, regs
);
6044 data
->period
= event
->hw
.last_period
;
6046 if (nr
== 1 && hwc
->sample_period
== 1 && !event
->attr
.freq
)
6047 return perf_swevent_overflow(event
, 1, data
, regs
);
6049 if (local64_add_negative(nr
, &hwc
->period_left
))
6052 perf_swevent_overflow(event
, 0, data
, regs
);
6055 static int perf_exclude_event(struct perf_event
*event
,
6056 struct pt_regs
*regs
)
6058 if (event
->hw
.state
& PERF_HES_STOPPED
)
6062 if (event
->attr
.exclude_user
&& user_mode(regs
))
6065 if (event
->attr
.exclude_kernel
&& !user_mode(regs
))
6072 static int perf_swevent_match(struct perf_event
*event
,
6073 enum perf_type_id type
,
6075 struct perf_sample_data
*data
,
6076 struct pt_regs
*regs
)
6078 if (event
->attr
.type
!= type
)
6081 if (event
->attr
.config
!= event_id
)
6084 if (perf_exclude_event(event
, regs
))
6090 static inline u64
swevent_hash(u64 type
, u32 event_id
)
6092 u64 val
= event_id
| (type
<< 32);
6094 return hash_64(val
, SWEVENT_HLIST_BITS
);
6097 static inline struct hlist_head
*
6098 __find_swevent_head(struct swevent_hlist
*hlist
, u64 type
, u32 event_id
)
6100 u64 hash
= swevent_hash(type
, event_id
);
6102 return &hlist
->heads
[hash
];
6105 /* For the read side: events when they trigger */
6106 static inline struct hlist_head
*
6107 find_swevent_head_rcu(struct swevent_htable
*swhash
, u64 type
, u32 event_id
)
6109 struct swevent_hlist
*hlist
;
6111 hlist
= rcu_dereference(swhash
->swevent_hlist
);
6115 return __find_swevent_head(hlist
, type
, event_id
);
6118 /* For the event head insertion and removal in the hlist */
6119 static inline struct hlist_head
*
6120 find_swevent_head(struct swevent_htable
*swhash
, struct perf_event
*event
)
6122 struct swevent_hlist
*hlist
;
6123 u32 event_id
= event
->attr
.config
;
6124 u64 type
= event
->attr
.type
;
6127 * Event scheduling is always serialized against hlist allocation
6128 * and release. Which makes the protected version suitable here.
6129 * The context lock guarantees that.
6131 hlist
= rcu_dereference_protected(swhash
->swevent_hlist
,
6132 lockdep_is_held(&event
->ctx
->lock
));
6136 return __find_swevent_head(hlist
, type
, event_id
);
6139 static void do_perf_sw_event(enum perf_type_id type
, u32 event_id
,
6141 struct perf_sample_data
*data
,
6142 struct pt_regs
*regs
)
6144 struct swevent_htable
*swhash
= this_cpu_ptr(&swevent_htable
);
6145 struct perf_event
*event
;
6146 struct hlist_head
*head
;
6149 head
= find_swevent_head_rcu(swhash
, type
, event_id
);
6153 hlist_for_each_entry_rcu(event
, head
, hlist_entry
) {
6154 if (perf_swevent_match(event
, type
, event_id
, data
, regs
))
6155 perf_swevent_event(event
, nr
, data
, regs
);
6161 DEFINE_PER_CPU(struct pt_regs
, __perf_regs
[4]);
6163 int perf_swevent_get_recursion_context(void)
6165 struct swevent_htable
*swhash
= this_cpu_ptr(&swevent_htable
);
6167 return get_recursion_context(swhash
->recursion
);
6169 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context
);
6171 inline void perf_swevent_put_recursion_context(int rctx
)
6173 struct swevent_htable
*swhash
= this_cpu_ptr(&swevent_htable
);
6175 put_recursion_context(swhash
->recursion
, rctx
);
6178 void ___perf_sw_event(u32 event_id
, u64 nr
, struct pt_regs
*regs
, u64 addr
)
6180 struct perf_sample_data data
;
6182 if (WARN_ON_ONCE(!regs
))
6185 perf_sample_data_init(&data
, addr
, 0);
6186 do_perf_sw_event(PERF_TYPE_SOFTWARE
, event_id
, nr
, &data
, regs
);
6189 void __perf_sw_event(u32 event_id
, u64 nr
, struct pt_regs
*regs
, u64 addr
)
6193 preempt_disable_notrace();
6194 rctx
= perf_swevent_get_recursion_context();
6195 if (unlikely(rctx
< 0))
6198 ___perf_sw_event(event_id
, nr
, regs
, addr
);
6200 perf_swevent_put_recursion_context(rctx
);
6202 preempt_enable_notrace();
6205 static void perf_swevent_read(struct perf_event
*event
)
6209 static int perf_swevent_add(struct perf_event
*event
, int flags
)
6211 struct swevent_htable
*swhash
= this_cpu_ptr(&swevent_htable
);
6212 struct hw_perf_event
*hwc
= &event
->hw
;
6213 struct hlist_head
*head
;
6215 if (is_sampling_event(event
)) {
6216 hwc
->last_period
= hwc
->sample_period
;
6217 perf_swevent_set_period(event
);
6220 hwc
->state
= !(flags
& PERF_EF_START
);
6222 head
= find_swevent_head(swhash
, event
);
6225 * We can race with cpu hotplug code. Do not
6226 * WARN if the cpu just got unplugged.
6228 WARN_ON_ONCE(swhash
->online
);
6232 hlist_add_head_rcu(&event
->hlist_entry
, head
);
6233 perf_event_update_userpage(event
);
6238 static void perf_swevent_del(struct perf_event
*event
, int flags
)
6240 hlist_del_rcu(&event
->hlist_entry
);
6243 static void perf_swevent_start(struct perf_event
*event
, int flags
)
6245 event
->hw
.state
= 0;
6248 static void perf_swevent_stop(struct perf_event
*event
, int flags
)
6250 event
->hw
.state
= PERF_HES_STOPPED
;
6253 /* Deref the hlist from the update side */
6254 static inline struct swevent_hlist
*
6255 swevent_hlist_deref(struct swevent_htable
*swhash
)
6257 return rcu_dereference_protected(swhash
->swevent_hlist
,
6258 lockdep_is_held(&swhash
->hlist_mutex
));
6261 static void swevent_hlist_release(struct swevent_htable
*swhash
)
6263 struct swevent_hlist
*hlist
= swevent_hlist_deref(swhash
);
6268 RCU_INIT_POINTER(swhash
->swevent_hlist
, NULL
);
6269 kfree_rcu(hlist
, rcu_head
);
6272 static void swevent_hlist_put_cpu(struct perf_event
*event
, int cpu
)
6274 struct swevent_htable
*swhash
= &per_cpu(swevent_htable
, cpu
);
6276 mutex_lock(&swhash
->hlist_mutex
);
6278 if (!--swhash
->hlist_refcount
)
6279 swevent_hlist_release(swhash
);
6281 mutex_unlock(&swhash
->hlist_mutex
);
6284 static void swevent_hlist_put(struct perf_event
*event
)
6288 for_each_possible_cpu(cpu
)
6289 swevent_hlist_put_cpu(event
, cpu
);
6292 static int swevent_hlist_get_cpu(struct perf_event
*event
, int cpu
)
6294 struct swevent_htable
*swhash
= &per_cpu(swevent_htable
, cpu
);
6297 mutex_lock(&swhash
->hlist_mutex
);
6299 if (!swevent_hlist_deref(swhash
) && cpu_online(cpu
)) {
6300 struct swevent_hlist
*hlist
;
6302 hlist
= kzalloc(sizeof(*hlist
), GFP_KERNEL
);
6307 rcu_assign_pointer(swhash
->swevent_hlist
, hlist
);
6309 swhash
->hlist_refcount
++;
6311 mutex_unlock(&swhash
->hlist_mutex
);
6316 static int swevent_hlist_get(struct perf_event
*event
)
6319 int cpu
, failed_cpu
;
6322 for_each_possible_cpu(cpu
) {
6323 err
= swevent_hlist_get_cpu(event
, cpu
);
6333 for_each_possible_cpu(cpu
) {
6334 if (cpu
== failed_cpu
)
6336 swevent_hlist_put_cpu(event
, cpu
);
6343 struct static_key perf_swevent_enabled
[PERF_COUNT_SW_MAX
];
6345 static void sw_perf_event_destroy(struct perf_event
*event
)
6347 u64 event_id
= event
->attr
.config
;
6349 WARN_ON(event
->parent
);
6351 static_key_slow_dec(&perf_swevent_enabled
[event_id
]);
6352 swevent_hlist_put(event
);
6355 static int perf_swevent_init(struct perf_event
*event
)
6357 u64 event_id
= event
->attr
.config
;
6359 if (event
->attr
.type
!= PERF_TYPE_SOFTWARE
)
6363 * no branch sampling for software events
6365 if (has_branch_stack(event
))
6369 case PERF_COUNT_SW_CPU_CLOCK
:
6370 case PERF_COUNT_SW_TASK_CLOCK
:
6377 if (event_id
>= PERF_COUNT_SW_MAX
)
6380 if (!event
->parent
) {
6383 err
= swevent_hlist_get(event
);
6387 static_key_slow_inc(&perf_swevent_enabled
[event_id
]);
6388 event
->destroy
= sw_perf_event_destroy
;
6394 static struct pmu perf_swevent
= {
6395 .task_ctx_nr
= perf_sw_context
,
6397 .capabilities
= PERF_PMU_CAP_NO_NMI
,
6399 .event_init
= perf_swevent_init
,
6400 .add
= perf_swevent_add
,
6401 .del
= perf_swevent_del
,
6402 .start
= perf_swevent_start
,
6403 .stop
= perf_swevent_stop
,
6404 .read
= perf_swevent_read
,
6407 #ifdef CONFIG_EVENT_TRACING
6409 static int perf_tp_filter_match(struct perf_event
*event
,
6410 struct perf_sample_data
*data
)
6412 void *record
= data
->raw
->data
;
6414 if (likely(!event
->filter
) || filter_match_preds(event
->filter
, record
))
6419 static int perf_tp_event_match(struct perf_event
*event
,
6420 struct perf_sample_data
*data
,
6421 struct pt_regs
*regs
)
6423 if (event
->hw
.state
& PERF_HES_STOPPED
)
6426 * All tracepoints are from kernel-space.
6428 if (event
->attr
.exclude_kernel
)
6431 if (!perf_tp_filter_match(event
, data
))
6437 void perf_tp_event(u64 addr
, u64 count
, void *record
, int entry_size
,
6438 struct pt_regs
*regs
, struct hlist_head
*head
, int rctx
,
6439 struct task_struct
*task
)
6441 struct perf_sample_data data
;
6442 struct perf_event
*event
;
6444 struct perf_raw_record raw
= {
6449 perf_sample_data_init(&data
, addr
, 0);
6452 hlist_for_each_entry_rcu(event
, head
, hlist_entry
) {
6453 if (perf_tp_event_match(event
, &data
, regs
))
6454 perf_swevent_event(event
, count
, &data
, regs
);
6458 * If we got specified a target task, also iterate its context and
6459 * deliver this event there too.
6461 if (task
&& task
!= current
) {
6462 struct perf_event_context
*ctx
;
6463 struct trace_entry
*entry
= record
;
6466 ctx
= rcu_dereference(task
->perf_event_ctxp
[perf_sw_context
]);
6470 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
6471 if (event
->attr
.type
!= PERF_TYPE_TRACEPOINT
)
6473 if (event
->attr
.config
!= entry
->type
)
6475 if (perf_tp_event_match(event
, &data
, regs
))
6476 perf_swevent_event(event
, count
, &data
, regs
);
6482 perf_swevent_put_recursion_context(rctx
);
6484 EXPORT_SYMBOL_GPL(perf_tp_event
);
6486 static void tp_perf_event_destroy(struct perf_event
*event
)
6488 perf_trace_destroy(event
);
6491 static int perf_tp_event_init(struct perf_event
*event
)
6495 if (event
->attr
.type
!= PERF_TYPE_TRACEPOINT
)
6499 * no branch sampling for tracepoint events
6501 if (has_branch_stack(event
))
6504 err
= perf_trace_init(event
);
6508 event
->destroy
= tp_perf_event_destroy
;
6513 static struct pmu perf_tracepoint
= {
6514 .task_ctx_nr
= perf_sw_context
,
6516 .event_init
= perf_tp_event_init
,
6517 .add
= perf_trace_add
,
6518 .del
= perf_trace_del
,
6519 .start
= perf_swevent_start
,
6520 .stop
= perf_swevent_stop
,
6521 .read
= perf_swevent_read
,
6524 static inline void perf_tp_register(void)
6526 perf_pmu_register(&perf_tracepoint
, "tracepoint", PERF_TYPE_TRACEPOINT
);
6529 static int perf_event_set_filter(struct perf_event
*event
, void __user
*arg
)
6534 if (event
->attr
.type
!= PERF_TYPE_TRACEPOINT
)
6537 filter_str
= strndup_user(arg
, PAGE_SIZE
);
6538 if (IS_ERR(filter_str
))
6539 return PTR_ERR(filter_str
);
6541 ret
= ftrace_profile_set_filter(event
, event
->attr
.config
, filter_str
);
6547 static void perf_event_free_filter(struct perf_event
*event
)
6549 ftrace_profile_free_filter(event
);
6552 static int perf_event_set_bpf_prog(struct perf_event
*event
, u32 prog_fd
)
6554 struct bpf_prog
*prog
;
6556 if (event
->attr
.type
!= PERF_TYPE_TRACEPOINT
)
6559 if (event
->tp_event
->prog
)
6562 if (!(event
->tp_event
->flags
& TRACE_EVENT_FL_KPROBE
))
6563 /* bpf programs can only be attached to kprobes */
6566 prog
= bpf_prog_get(prog_fd
);
6568 return PTR_ERR(prog
);
6570 if (prog
->aux
->prog_type
!= BPF_PROG_TYPE_KPROBE
) {
6571 /* valid fd, but invalid bpf program type */
6576 event
->tp_event
->prog
= prog
;
6581 static void perf_event_free_bpf_prog(struct perf_event
*event
)
6583 struct bpf_prog
*prog
;
6585 if (!event
->tp_event
)
6588 prog
= event
->tp_event
->prog
;
6590 event
->tp_event
->prog
= NULL
;
6597 static inline void perf_tp_register(void)
6601 static int perf_event_set_filter(struct perf_event
*event
, void __user
*arg
)
6606 static void perf_event_free_filter(struct perf_event
*event
)
6610 static int perf_event_set_bpf_prog(struct perf_event
*event
, u32 prog_fd
)
6615 static void perf_event_free_bpf_prog(struct perf_event
*event
)
6618 #endif /* CONFIG_EVENT_TRACING */
6620 #ifdef CONFIG_HAVE_HW_BREAKPOINT
6621 void perf_bp_event(struct perf_event
*bp
, void *data
)
6623 struct perf_sample_data sample
;
6624 struct pt_regs
*regs
= data
;
6626 perf_sample_data_init(&sample
, bp
->attr
.bp_addr
, 0);
6628 if (!bp
->hw
.state
&& !perf_exclude_event(bp
, regs
))
6629 perf_swevent_event(bp
, 1, &sample
, regs
);
6634 * hrtimer based swevent callback
6637 static enum hrtimer_restart
perf_swevent_hrtimer(struct hrtimer
*hrtimer
)
6639 enum hrtimer_restart ret
= HRTIMER_RESTART
;
6640 struct perf_sample_data data
;
6641 struct pt_regs
*regs
;
6642 struct perf_event
*event
;
6645 event
= container_of(hrtimer
, struct perf_event
, hw
.hrtimer
);
6647 if (event
->state
!= PERF_EVENT_STATE_ACTIVE
)
6648 return HRTIMER_NORESTART
;
6650 event
->pmu
->read(event
);
6652 perf_sample_data_init(&data
, 0, event
->hw
.last_period
);
6653 regs
= get_irq_regs();
6655 if (regs
&& !perf_exclude_event(event
, regs
)) {
6656 if (!(event
->attr
.exclude_idle
&& is_idle_task(current
)))
6657 if (__perf_event_overflow(event
, 1, &data
, regs
))
6658 ret
= HRTIMER_NORESTART
;
6661 period
= max_t(u64
, 10000, event
->hw
.sample_period
);
6662 hrtimer_forward_now(hrtimer
, ns_to_ktime(period
));
6667 static void perf_swevent_start_hrtimer(struct perf_event
*event
)
6669 struct hw_perf_event
*hwc
= &event
->hw
;
6672 if (!is_sampling_event(event
))
6675 period
= local64_read(&hwc
->period_left
);
6680 local64_set(&hwc
->period_left
, 0);
6682 period
= max_t(u64
, 10000, hwc
->sample_period
);
6684 __hrtimer_start_range_ns(&hwc
->hrtimer
,
6685 ns_to_ktime(period
), 0,
6686 HRTIMER_MODE_REL_PINNED
, 0);
6689 static void perf_swevent_cancel_hrtimer(struct perf_event
*event
)
6691 struct hw_perf_event
*hwc
= &event
->hw
;
6693 if (is_sampling_event(event
)) {
6694 ktime_t remaining
= hrtimer_get_remaining(&hwc
->hrtimer
);
6695 local64_set(&hwc
->period_left
, ktime_to_ns(remaining
));
6697 hrtimer_cancel(&hwc
->hrtimer
);
6701 static void perf_swevent_init_hrtimer(struct perf_event
*event
)
6703 struct hw_perf_event
*hwc
= &event
->hw
;
6705 if (!is_sampling_event(event
))
6708 hrtimer_init(&hwc
->hrtimer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
6709 hwc
->hrtimer
.function
= perf_swevent_hrtimer
;
6712 * Since hrtimers have a fixed rate, we can do a static freq->period
6713 * mapping and avoid the whole period adjust feedback stuff.
6715 if (event
->attr
.freq
) {
6716 long freq
= event
->attr
.sample_freq
;
6718 event
->attr
.sample_period
= NSEC_PER_SEC
/ freq
;
6719 hwc
->sample_period
= event
->attr
.sample_period
;
6720 local64_set(&hwc
->period_left
, hwc
->sample_period
);
6721 hwc
->last_period
= hwc
->sample_period
;
6722 event
->attr
.freq
= 0;
6727 * Software event: cpu wall time clock
6730 static void cpu_clock_event_update(struct perf_event
*event
)
6735 now
= local_clock();
6736 prev
= local64_xchg(&event
->hw
.prev_count
, now
);
6737 local64_add(now
- prev
, &event
->count
);
6740 static void cpu_clock_event_start(struct perf_event
*event
, int flags
)
6742 local64_set(&event
->hw
.prev_count
, local_clock());
6743 perf_swevent_start_hrtimer(event
);
6746 static void cpu_clock_event_stop(struct perf_event
*event
, int flags
)
6748 perf_swevent_cancel_hrtimer(event
);
6749 cpu_clock_event_update(event
);
6752 static int cpu_clock_event_add(struct perf_event
*event
, int flags
)
6754 if (flags
& PERF_EF_START
)
6755 cpu_clock_event_start(event
, flags
);
6756 perf_event_update_userpage(event
);
6761 static void cpu_clock_event_del(struct perf_event
*event
, int flags
)
6763 cpu_clock_event_stop(event
, flags
);
6766 static void cpu_clock_event_read(struct perf_event
*event
)
6768 cpu_clock_event_update(event
);
6771 static int cpu_clock_event_init(struct perf_event
*event
)
6773 if (event
->attr
.type
!= PERF_TYPE_SOFTWARE
)
6776 if (event
->attr
.config
!= PERF_COUNT_SW_CPU_CLOCK
)
6780 * no branch sampling for software events
6782 if (has_branch_stack(event
))
6785 perf_swevent_init_hrtimer(event
);
6790 static struct pmu perf_cpu_clock
= {
6791 .task_ctx_nr
= perf_sw_context
,
6793 .capabilities
= PERF_PMU_CAP_NO_NMI
,
6795 .event_init
= cpu_clock_event_init
,
6796 .add
= cpu_clock_event_add
,
6797 .del
= cpu_clock_event_del
,
6798 .start
= cpu_clock_event_start
,
6799 .stop
= cpu_clock_event_stop
,
6800 .read
= cpu_clock_event_read
,
6804 * Software event: task time clock
6807 static void task_clock_event_update(struct perf_event
*event
, u64 now
)
6812 prev
= local64_xchg(&event
->hw
.prev_count
, now
);
6814 local64_add(delta
, &event
->count
);
6817 static void task_clock_event_start(struct perf_event
*event
, int flags
)
6819 local64_set(&event
->hw
.prev_count
, event
->ctx
->time
);
6820 perf_swevent_start_hrtimer(event
);
6823 static void task_clock_event_stop(struct perf_event
*event
, int flags
)
6825 perf_swevent_cancel_hrtimer(event
);
6826 task_clock_event_update(event
, event
->ctx
->time
);
6829 static int task_clock_event_add(struct perf_event
*event
, int flags
)
6831 if (flags
& PERF_EF_START
)
6832 task_clock_event_start(event
, flags
);
6833 perf_event_update_userpage(event
);
6838 static void task_clock_event_del(struct perf_event
*event
, int flags
)
6840 task_clock_event_stop(event
, PERF_EF_UPDATE
);
6843 static void task_clock_event_read(struct perf_event
*event
)
6845 u64 now
= perf_clock();
6846 u64 delta
= now
- event
->ctx
->timestamp
;
6847 u64 time
= event
->ctx
->time
+ delta
;
6849 task_clock_event_update(event
, time
);
6852 static int task_clock_event_init(struct perf_event
*event
)
6854 if (event
->attr
.type
!= PERF_TYPE_SOFTWARE
)
6857 if (event
->attr
.config
!= PERF_COUNT_SW_TASK_CLOCK
)
6861 * no branch sampling for software events
6863 if (has_branch_stack(event
))
6866 perf_swevent_init_hrtimer(event
);
6871 static struct pmu perf_task_clock
= {
6872 .task_ctx_nr
= perf_sw_context
,
6874 .capabilities
= PERF_PMU_CAP_NO_NMI
,
6876 .event_init
= task_clock_event_init
,
6877 .add
= task_clock_event_add
,
6878 .del
= task_clock_event_del
,
6879 .start
= task_clock_event_start
,
6880 .stop
= task_clock_event_stop
,
6881 .read
= task_clock_event_read
,
6884 static void perf_pmu_nop_void(struct pmu
*pmu
)
6888 static int perf_pmu_nop_int(struct pmu
*pmu
)
6893 static void perf_pmu_start_txn(struct pmu
*pmu
)
6895 perf_pmu_disable(pmu
);
6898 static int perf_pmu_commit_txn(struct pmu
*pmu
)
6900 perf_pmu_enable(pmu
);
6904 static void perf_pmu_cancel_txn(struct pmu
*pmu
)
6906 perf_pmu_enable(pmu
);
6909 static int perf_event_idx_default(struct perf_event
*event
)
6915 * Ensures all contexts with the same task_ctx_nr have the same
6916 * pmu_cpu_context too.
6918 static struct perf_cpu_context __percpu
*find_pmu_context(int ctxn
)
6925 list_for_each_entry(pmu
, &pmus
, entry
) {
6926 if (pmu
->task_ctx_nr
== ctxn
)
6927 return pmu
->pmu_cpu_context
;
6933 static void update_pmu_context(struct pmu
*pmu
, struct pmu
*old_pmu
)
6937 for_each_possible_cpu(cpu
) {
6938 struct perf_cpu_context
*cpuctx
;
6940 cpuctx
= per_cpu_ptr(pmu
->pmu_cpu_context
, cpu
);
6942 if (cpuctx
->unique_pmu
== old_pmu
)
6943 cpuctx
->unique_pmu
= pmu
;
6947 static void free_pmu_context(struct pmu
*pmu
)
6951 mutex_lock(&pmus_lock
);
6953 * Like a real lame refcount.
6955 list_for_each_entry(i
, &pmus
, entry
) {
6956 if (i
->pmu_cpu_context
== pmu
->pmu_cpu_context
) {
6957 update_pmu_context(i
, pmu
);
6962 free_percpu(pmu
->pmu_cpu_context
);
6964 mutex_unlock(&pmus_lock
);
6966 static struct idr pmu_idr
;
6969 type_show(struct device
*dev
, struct device_attribute
*attr
, char *page
)
6971 struct pmu
*pmu
= dev_get_drvdata(dev
);
6973 return snprintf(page
, PAGE_SIZE
-1, "%d\n", pmu
->type
);
6975 static DEVICE_ATTR_RO(type
);
6978 perf_event_mux_interval_ms_show(struct device
*dev
,
6979 struct device_attribute
*attr
,
6982 struct pmu
*pmu
= dev_get_drvdata(dev
);
6984 return snprintf(page
, PAGE_SIZE
-1, "%d\n", pmu
->hrtimer_interval_ms
);
6988 perf_event_mux_interval_ms_store(struct device
*dev
,
6989 struct device_attribute
*attr
,
6990 const char *buf
, size_t count
)
6992 struct pmu
*pmu
= dev_get_drvdata(dev
);
6993 int timer
, cpu
, ret
;
6995 ret
= kstrtoint(buf
, 0, &timer
);
7002 /* same value, noting to do */
7003 if (timer
== pmu
->hrtimer_interval_ms
)
7006 pmu
->hrtimer_interval_ms
= timer
;
7008 /* update all cpuctx for this PMU */
7009 for_each_possible_cpu(cpu
) {
7010 struct perf_cpu_context
*cpuctx
;
7011 cpuctx
= per_cpu_ptr(pmu
->pmu_cpu_context
, cpu
);
7012 cpuctx
->hrtimer_interval
= ns_to_ktime(NSEC_PER_MSEC
* timer
);
7014 if (hrtimer_active(&cpuctx
->hrtimer
))
7015 hrtimer_forward_now(&cpuctx
->hrtimer
, cpuctx
->hrtimer_interval
);
7020 static DEVICE_ATTR_RW(perf_event_mux_interval_ms
);
7022 static struct attribute
*pmu_dev_attrs
[] = {
7023 &dev_attr_type
.attr
,
7024 &dev_attr_perf_event_mux_interval_ms
.attr
,
7027 ATTRIBUTE_GROUPS(pmu_dev
);
7029 static int pmu_bus_running
;
7030 static struct bus_type pmu_bus
= {
7031 .name
= "event_source",
7032 .dev_groups
= pmu_dev_groups
,
7035 static void pmu_dev_release(struct device
*dev
)
7040 static int pmu_dev_alloc(struct pmu
*pmu
)
7044 pmu
->dev
= kzalloc(sizeof(struct device
), GFP_KERNEL
);
7048 pmu
->dev
->groups
= pmu
->attr_groups
;
7049 device_initialize(pmu
->dev
);
7050 ret
= dev_set_name(pmu
->dev
, "%s", pmu
->name
);
7054 dev_set_drvdata(pmu
->dev
, pmu
);
7055 pmu
->dev
->bus
= &pmu_bus
;
7056 pmu
->dev
->release
= pmu_dev_release
;
7057 ret
= device_add(pmu
->dev
);
7065 put_device(pmu
->dev
);
7069 static struct lock_class_key cpuctx_mutex
;
7070 static struct lock_class_key cpuctx_lock
;
7072 int perf_pmu_register(struct pmu
*pmu
, const char *name
, int type
)
7076 mutex_lock(&pmus_lock
);
7078 pmu
->pmu_disable_count
= alloc_percpu(int);
7079 if (!pmu
->pmu_disable_count
)
7088 type
= idr_alloc(&pmu_idr
, pmu
, PERF_TYPE_MAX
, 0, GFP_KERNEL
);
7096 if (pmu_bus_running
) {
7097 ret
= pmu_dev_alloc(pmu
);
7103 pmu
->pmu_cpu_context
= find_pmu_context(pmu
->task_ctx_nr
);
7104 if (pmu
->pmu_cpu_context
)
7105 goto got_cpu_context
;
7108 pmu
->pmu_cpu_context
= alloc_percpu(struct perf_cpu_context
);
7109 if (!pmu
->pmu_cpu_context
)
7112 for_each_possible_cpu(cpu
) {
7113 struct perf_cpu_context
*cpuctx
;
7115 cpuctx
= per_cpu_ptr(pmu
->pmu_cpu_context
, cpu
);
7116 __perf_event_init_context(&cpuctx
->ctx
);
7117 lockdep_set_class(&cpuctx
->ctx
.mutex
, &cpuctx_mutex
);
7118 lockdep_set_class(&cpuctx
->ctx
.lock
, &cpuctx_lock
);
7119 cpuctx
->ctx
.pmu
= pmu
;
7121 __perf_cpu_hrtimer_init(cpuctx
, cpu
);
7123 cpuctx
->unique_pmu
= pmu
;
7127 if (!pmu
->start_txn
) {
7128 if (pmu
->pmu_enable
) {
7130 * If we have pmu_enable/pmu_disable calls, install
7131 * transaction stubs that use that to try and batch
7132 * hardware accesses.
7134 pmu
->start_txn
= perf_pmu_start_txn
;
7135 pmu
->commit_txn
= perf_pmu_commit_txn
;
7136 pmu
->cancel_txn
= perf_pmu_cancel_txn
;
7138 pmu
->start_txn
= perf_pmu_nop_void
;
7139 pmu
->commit_txn
= perf_pmu_nop_int
;
7140 pmu
->cancel_txn
= perf_pmu_nop_void
;
7144 if (!pmu
->pmu_enable
) {
7145 pmu
->pmu_enable
= perf_pmu_nop_void
;
7146 pmu
->pmu_disable
= perf_pmu_nop_void
;
7149 if (!pmu
->event_idx
)
7150 pmu
->event_idx
= perf_event_idx_default
;
7152 list_add_rcu(&pmu
->entry
, &pmus
);
7155 mutex_unlock(&pmus_lock
);
7160 device_del(pmu
->dev
);
7161 put_device(pmu
->dev
);
7164 if (pmu
->type
>= PERF_TYPE_MAX
)
7165 idr_remove(&pmu_idr
, pmu
->type
);
7168 free_percpu(pmu
->pmu_disable_count
);
7171 EXPORT_SYMBOL_GPL(perf_pmu_register
);
7173 void perf_pmu_unregister(struct pmu
*pmu
)
7175 mutex_lock(&pmus_lock
);
7176 list_del_rcu(&pmu
->entry
);
7177 mutex_unlock(&pmus_lock
);
7180 * We dereference the pmu list under both SRCU and regular RCU, so
7181 * synchronize against both of those.
7183 synchronize_srcu(&pmus_srcu
);
7186 free_percpu(pmu
->pmu_disable_count
);
7187 if (pmu
->type
>= PERF_TYPE_MAX
)
7188 idr_remove(&pmu_idr
, pmu
->type
);
7189 device_del(pmu
->dev
);
7190 put_device(pmu
->dev
);
7191 free_pmu_context(pmu
);
7193 EXPORT_SYMBOL_GPL(perf_pmu_unregister
);
7195 static int perf_try_init_event(struct pmu
*pmu
, struct perf_event
*event
)
7197 struct perf_event_context
*ctx
= NULL
;
7200 if (!try_module_get(pmu
->module
))
7203 if (event
->group_leader
!= event
) {
7204 ctx
= perf_event_ctx_lock(event
->group_leader
);
7209 ret
= pmu
->event_init(event
);
7212 perf_event_ctx_unlock(event
->group_leader
, ctx
);
7215 module_put(pmu
->module
);
7220 struct pmu
*perf_init_event(struct perf_event
*event
)
7222 struct pmu
*pmu
= NULL
;
7226 idx
= srcu_read_lock(&pmus_srcu
);
7229 pmu
= idr_find(&pmu_idr
, event
->attr
.type
);
7232 ret
= perf_try_init_event(pmu
, event
);
7238 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
7239 ret
= perf_try_init_event(pmu
, event
);
7243 if (ret
!= -ENOENT
) {
7248 pmu
= ERR_PTR(-ENOENT
);
7250 srcu_read_unlock(&pmus_srcu
, idx
);
7255 static void account_event_cpu(struct perf_event
*event
, int cpu
)
7260 if (is_cgroup_event(event
))
7261 atomic_inc(&per_cpu(perf_cgroup_events
, cpu
));
7264 static void account_event(struct perf_event
*event
)
7269 if (event
->attach_state
& PERF_ATTACH_TASK
)
7270 static_key_slow_inc(&perf_sched_events
.key
);
7271 if (event
->attr
.mmap
|| event
->attr
.mmap_data
)
7272 atomic_inc(&nr_mmap_events
);
7273 if (event
->attr
.comm
)
7274 atomic_inc(&nr_comm_events
);
7275 if (event
->attr
.task
)
7276 atomic_inc(&nr_task_events
);
7277 if (event
->attr
.freq
) {
7278 if (atomic_inc_return(&nr_freq_events
) == 1)
7279 tick_nohz_full_kick_all();
7281 if (has_branch_stack(event
))
7282 static_key_slow_inc(&perf_sched_events
.key
);
7283 if (is_cgroup_event(event
))
7284 static_key_slow_inc(&perf_sched_events
.key
);
7286 account_event_cpu(event
, event
->cpu
);
7290 * Allocate and initialize a event structure
7292 static struct perf_event
*
7293 perf_event_alloc(struct perf_event_attr
*attr
, int cpu
,
7294 struct task_struct
*task
,
7295 struct perf_event
*group_leader
,
7296 struct perf_event
*parent_event
,
7297 perf_overflow_handler_t overflow_handler
,
7298 void *context
, int cgroup_fd
)
7301 struct perf_event
*event
;
7302 struct hw_perf_event
*hwc
;
7305 if ((unsigned)cpu
>= nr_cpu_ids
) {
7306 if (!task
|| cpu
!= -1)
7307 return ERR_PTR(-EINVAL
);
7310 event
= kzalloc(sizeof(*event
), GFP_KERNEL
);
7312 return ERR_PTR(-ENOMEM
);
7315 * Single events are their own group leaders, with an
7316 * empty sibling list:
7319 group_leader
= event
;
7321 mutex_init(&event
->child_mutex
);
7322 INIT_LIST_HEAD(&event
->child_list
);
7324 INIT_LIST_HEAD(&event
->group_entry
);
7325 INIT_LIST_HEAD(&event
->event_entry
);
7326 INIT_LIST_HEAD(&event
->sibling_list
);
7327 INIT_LIST_HEAD(&event
->rb_entry
);
7328 INIT_LIST_HEAD(&event
->active_entry
);
7329 INIT_HLIST_NODE(&event
->hlist_entry
);
7332 init_waitqueue_head(&event
->waitq
);
7333 init_irq_work(&event
->pending
, perf_pending_event
);
7335 mutex_init(&event
->mmap_mutex
);
7337 atomic_long_set(&event
->refcount
, 1);
7339 event
->attr
= *attr
;
7340 event
->group_leader
= group_leader
;
7344 event
->parent
= parent_event
;
7346 event
->ns
= get_pid_ns(task_active_pid_ns(current
));
7347 event
->id
= atomic64_inc_return(&perf_event_id
);
7349 event
->state
= PERF_EVENT_STATE_INACTIVE
;
7352 event
->attach_state
= PERF_ATTACH_TASK
;
7354 * XXX pmu::event_init needs to know what task to account to
7355 * and we cannot use the ctx information because we need the
7356 * pmu before we get a ctx.
7358 event
->hw
.target
= task
;
7361 event
->clock
= &local_clock
;
7363 event
->clock
= parent_event
->clock
;
7365 if (!overflow_handler
&& parent_event
) {
7366 overflow_handler
= parent_event
->overflow_handler
;
7367 context
= parent_event
->overflow_handler_context
;
7370 event
->overflow_handler
= overflow_handler
;
7371 event
->overflow_handler_context
= context
;
7373 perf_event__state_init(event
);
7378 hwc
->sample_period
= attr
->sample_period
;
7379 if (attr
->freq
&& attr
->sample_freq
)
7380 hwc
->sample_period
= 1;
7381 hwc
->last_period
= hwc
->sample_period
;
7383 local64_set(&hwc
->period_left
, hwc
->sample_period
);
7386 * we currently do not support PERF_FORMAT_GROUP on inherited events
7388 if (attr
->inherit
&& (attr
->read_format
& PERF_FORMAT_GROUP
))
7391 if (!has_branch_stack(event
))
7392 event
->attr
.branch_sample_type
= 0;
7394 if (cgroup_fd
!= -1) {
7395 err
= perf_cgroup_connect(cgroup_fd
, event
, attr
, group_leader
);
7400 pmu
= perf_init_event(event
);
7403 else if (IS_ERR(pmu
)) {
7408 if (!event
->parent
) {
7409 if (event
->attr
.sample_type
& PERF_SAMPLE_CALLCHAIN
) {
7410 err
= get_callchain_buffers();
7420 event
->destroy(event
);
7421 module_put(pmu
->module
);
7423 if (is_cgroup_event(event
))
7424 perf_detach_cgroup(event
);
7426 put_pid_ns(event
->ns
);
7429 return ERR_PTR(err
);
7432 static int perf_copy_attr(struct perf_event_attr __user
*uattr
,
7433 struct perf_event_attr
*attr
)
7438 if (!access_ok(VERIFY_WRITE
, uattr
, PERF_ATTR_SIZE_VER0
))
7442 * zero the full structure, so that a short copy will be nice.
7444 memset(attr
, 0, sizeof(*attr
));
7446 ret
= get_user(size
, &uattr
->size
);
7450 if (size
> PAGE_SIZE
) /* silly large */
7453 if (!size
) /* abi compat */
7454 size
= PERF_ATTR_SIZE_VER0
;
7456 if (size
< PERF_ATTR_SIZE_VER0
)
7460 * If we're handed a bigger struct than we know of,
7461 * ensure all the unknown bits are 0 - i.e. new
7462 * user-space does not rely on any kernel feature
7463 * extensions we dont know about yet.
7465 if (size
> sizeof(*attr
)) {
7466 unsigned char __user
*addr
;
7467 unsigned char __user
*end
;
7470 addr
= (void __user
*)uattr
+ sizeof(*attr
);
7471 end
= (void __user
*)uattr
+ size
;
7473 for (; addr
< end
; addr
++) {
7474 ret
= get_user(val
, addr
);
7480 size
= sizeof(*attr
);
7483 ret
= copy_from_user(attr
, uattr
, size
);
7487 if (attr
->__reserved_1
)
7490 if (attr
->sample_type
& ~(PERF_SAMPLE_MAX
-1))
7493 if (attr
->read_format
& ~(PERF_FORMAT_MAX
-1))
7496 if (attr
->sample_type
& PERF_SAMPLE_BRANCH_STACK
) {
7497 u64 mask
= attr
->branch_sample_type
;
7499 /* only using defined bits */
7500 if (mask
& ~(PERF_SAMPLE_BRANCH_MAX
-1))
7503 /* at least one branch bit must be set */
7504 if (!(mask
& ~PERF_SAMPLE_BRANCH_PLM_ALL
))
7507 /* propagate priv level, when not set for branch */
7508 if (!(mask
& PERF_SAMPLE_BRANCH_PLM_ALL
)) {
7510 /* exclude_kernel checked on syscall entry */
7511 if (!attr
->exclude_kernel
)
7512 mask
|= PERF_SAMPLE_BRANCH_KERNEL
;
7514 if (!attr
->exclude_user
)
7515 mask
|= PERF_SAMPLE_BRANCH_USER
;
7517 if (!attr
->exclude_hv
)
7518 mask
|= PERF_SAMPLE_BRANCH_HV
;
7520 * adjust user setting (for HW filter setup)
7522 attr
->branch_sample_type
= mask
;
7524 /* privileged levels capture (kernel, hv): check permissions */
7525 if ((mask
& PERF_SAMPLE_BRANCH_PERM_PLM
)
7526 && perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN
))
7530 if (attr
->sample_type
& PERF_SAMPLE_REGS_USER
) {
7531 ret
= perf_reg_validate(attr
->sample_regs_user
);
7536 if (attr
->sample_type
& PERF_SAMPLE_STACK_USER
) {
7537 if (!arch_perf_have_user_stack_dump())
7541 * We have __u32 type for the size, but so far
7542 * we can only use __u16 as maximum due to the
7543 * __u16 sample size limit.
7545 if (attr
->sample_stack_user
>= USHRT_MAX
)
7547 else if (!IS_ALIGNED(attr
->sample_stack_user
, sizeof(u64
)))
7551 if (attr
->sample_type
& PERF_SAMPLE_REGS_INTR
)
7552 ret
= perf_reg_validate(attr
->sample_regs_intr
);
7557 put_user(sizeof(*attr
), &uattr
->size
);
7563 perf_event_set_output(struct perf_event
*event
, struct perf_event
*output_event
)
7565 struct ring_buffer
*rb
= NULL
;
7571 /* don't allow circular references */
7572 if (event
== output_event
)
7576 * Don't allow cross-cpu buffers
7578 if (output_event
->cpu
!= event
->cpu
)
7582 * If its not a per-cpu rb, it must be the same task.
7584 if (output_event
->cpu
== -1 && output_event
->ctx
!= event
->ctx
)
7588 * Mixing clocks in the same buffer is trouble you don't need.
7590 if (output_event
->clock
!= event
->clock
)
7594 * If both events generate aux data, they must be on the same PMU
7596 if (has_aux(event
) && has_aux(output_event
) &&
7597 event
->pmu
!= output_event
->pmu
)
7601 mutex_lock(&event
->mmap_mutex
);
7602 /* Can't redirect output if we've got an active mmap() */
7603 if (atomic_read(&event
->mmap_count
))
7607 /* get the rb we want to redirect to */
7608 rb
= ring_buffer_get(output_event
);
7613 ring_buffer_attach(event
, rb
);
7617 mutex_unlock(&event
->mmap_mutex
);
7623 static void mutex_lock_double(struct mutex
*a
, struct mutex
*b
)
7629 mutex_lock_nested(b
, SINGLE_DEPTH_NESTING
);
7632 static int perf_event_set_clock(struct perf_event
*event
, clockid_t clk_id
)
7634 bool nmi_safe
= false;
7637 case CLOCK_MONOTONIC
:
7638 event
->clock
= &ktime_get_mono_fast_ns
;
7642 case CLOCK_MONOTONIC_RAW
:
7643 event
->clock
= &ktime_get_raw_fast_ns
;
7647 case CLOCK_REALTIME
:
7648 event
->clock
= &ktime_get_real_ns
;
7651 case CLOCK_BOOTTIME
:
7652 event
->clock
= &ktime_get_boot_ns
;
7656 event
->clock
= &ktime_get_tai_ns
;
7663 if (!nmi_safe
&& !(event
->pmu
->capabilities
& PERF_PMU_CAP_NO_NMI
))
7670 * sys_perf_event_open - open a performance event, associate it to a task/cpu
7672 * @attr_uptr: event_id type attributes for monitoring/sampling
7675 * @group_fd: group leader event fd
7677 SYSCALL_DEFINE5(perf_event_open
,
7678 struct perf_event_attr __user
*, attr_uptr
,
7679 pid_t
, pid
, int, cpu
, int, group_fd
, unsigned long, flags
)
7681 struct perf_event
*group_leader
= NULL
, *output_event
= NULL
;
7682 struct perf_event
*event
, *sibling
;
7683 struct perf_event_attr attr
;
7684 struct perf_event_context
*ctx
, *uninitialized_var(gctx
);
7685 struct file
*event_file
= NULL
;
7686 struct fd group
= {NULL
, 0};
7687 struct task_struct
*task
= NULL
;
7692 int f_flags
= O_RDWR
;
7695 /* for future expandability... */
7696 if (flags
& ~PERF_FLAG_ALL
)
7699 err
= perf_copy_attr(attr_uptr
, &attr
);
7703 if (!attr
.exclude_kernel
) {
7704 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN
))
7709 if (attr
.sample_freq
> sysctl_perf_event_sample_rate
)
7712 if (attr
.sample_period
& (1ULL << 63))
7717 * In cgroup mode, the pid argument is used to pass the fd
7718 * opened to the cgroup directory in cgroupfs. The cpu argument
7719 * designates the cpu on which to monitor threads from that
7722 if ((flags
& PERF_FLAG_PID_CGROUP
) && (pid
== -1 || cpu
== -1))
7725 if (flags
& PERF_FLAG_FD_CLOEXEC
)
7726 f_flags
|= O_CLOEXEC
;
7728 event_fd
= get_unused_fd_flags(f_flags
);
7732 if (group_fd
!= -1) {
7733 err
= perf_fget_light(group_fd
, &group
);
7736 group_leader
= group
.file
->private_data
;
7737 if (flags
& PERF_FLAG_FD_OUTPUT
)
7738 output_event
= group_leader
;
7739 if (flags
& PERF_FLAG_FD_NO_GROUP
)
7740 group_leader
= NULL
;
7743 if (pid
!= -1 && !(flags
& PERF_FLAG_PID_CGROUP
)) {
7744 task
= find_lively_task_by_vpid(pid
);
7746 err
= PTR_ERR(task
);
7751 if (task
&& group_leader
&&
7752 group_leader
->attr
.inherit
!= attr
.inherit
) {
7759 if (flags
& PERF_FLAG_PID_CGROUP
)
7762 event
= perf_event_alloc(&attr
, cpu
, task
, group_leader
, NULL
,
7763 NULL
, NULL
, cgroup_fd
);
7764 if (IS_ERR(event
)) {
7765 err
= PTR_ERR(event
);
7769 if (is_sampling_event(event
)) {
7770 if (event
->pmu
->capabilities
& PERF_PMU_CAP_NO_INTERRUPT
) {
7776 account_event(event
);
7779 * Special case software events and allow them to be part of
7780 * any hardware group.
7784 if (attr
.use_clockid
) {
7785 err
= perf_event_set_clock(event
, attr
.clockid
);
7791 (is_software_event(event
) != is_software_event(group_leader
))) {
7792 if (is_software_event(event
)) {
7794 * If event and group_leader are not both a software
7795 * event, and event is, then group leader is not.
7797 * Allow the addition of software events to !software
7798 * groups, this is safe because software events never
7801 pmu
= group_leader
->pmu
;
7802 } else if (is_software_event(group_leader
) &&
7803 (group_leader
->group_flags
& PERF_GROUP_SOFTWARE
)) {
7805 * In case the group is a pure software group, and we
7806 * try to add a hardware event, move the whole group to
7807 * the hardware context.
7814 * Get the target context (task or percpu):
7816 ctx
= find_get_context(pmu
, task
, event
);
7823 put_task_struct(task
);
7828 * Look up the group leader (we will attach this event to it):
7834 * Do not allow a recursive hierarchy (this new sibling
7835 * becoming part of another group-sibling):
7837 if (group_leader
->group_leader
!= group_leader
)
7840 /* All events in a group should have the same clock */
7841 if (group_leader
->clock
!= event
->clock
)
7845 * Do not allow to attach to a group in a different
7846 * task or CPU context:
7850 * Make sure we're both on the same task, or both
7853 if (group_leader
->ctx
->task
!= ctx
->task
)
7857 * Make sure we're both events for the same CPU;
7858 * grouping events for different CPUs is broken; since
7859 * you can never concurrently schedule them anyhow.
7861 if (group_leader
->cpu
!= event
->cpu
)
7864 if (group_leader
->ctx
!= ctx
)
7869 * Only a group leader can be exclusive or pinned
7871 if (attr
.exclusive
|| attr
.pinned
)
7876 err
= perf_event_set_output(event
, output_event
);
7881 event_file
= anon_inode_getfile("[perf_event]", &perf_fops
, event
,
7883 if (IS_ERR(event_file
)) {
7884 err
= PTR_ERR(event_file
);
7889 gctx
= group_leader
->ctx
;
7892 * See perf_event_ctx_lock() for comments on the details
7893 * of swizzling perf_event::ctx.
7895 mutex_lock_double(&gctx
->mutex
, &ctx
->mutex
);
7897 perf_remove_from_context(group_leader
, false);
7899 list_for_each_entry(sibling
, &group_leader
->sibling_list
,
7901 perf_remove_from_context(sibling
, false);
7905 mutex_lock(&ctx
->mutex
);
7908 WARN_ON_ONCE(ctx
->parent_ctx
);
7912 * Wait for everybody to stop referencing the events through
7913 * the old lists, before installing it on new lists.
7918 * Install the group siblings before the group leader.
7920 * Because a group leader will try and install the entire group
7921 * (through the sibling list, which is still in-tact), we can
7922 * end up with siblings installed in the wrong context.
7924 * By installing siblings first we NO-OP because they're not
7925 * reachable through the group lists.
7927 list_for_each_entry(sibling
, &group_leader
->sibling_list
,
7929 perf_event__state_init(sibling
);
7930 perf_install_in_context(ctx
, sibling
, sibling
->cpu
);
7935 * Removing from the context ends up with disabled
7936 * event. What we want here is event in the initial
7937 * startup state, ready to be add into new context.
7939 perf_event__state_init(group_leader
);
7940 perf_install_in_context(ctx
, group_leader
, group_leader
->cpu
);
7944 perf_install_in_context(ctx
, event
, event
->cpu
);
7945 perf_unpin_context(ctx
);
7948 mutex_unlock(&gctx
->mutex
);
7951 mutex_unlock(&ctx
->mutex
);
7955 event
->owner
= current
;
7957 mutex_lock(¤t
->perf_event_mutex
);
7958 list_add_tail(&event
->owner_entry
, ¤t
->perf_event_list
);
7959 mutex_unlock(¤t
->perf_event_mutex
);
7962 * Precalculate sample_data sizes
7964 perf_event__header_size(event
);
7965 perf_event__id_header_size(event
);
7968 * Drop the reference on the group_event after placing the
7969 * new event on the sibling_list. This ensures destruction
7970 * of the group leader will find the pointer to itself in
7971 * perf_group_detach().
7974 fd_install(event_fd
, event_file
);
7978 perf_unpin_context(ctx
);
7986 put_task_struct(task
);
7990 put_unused_fd(event_fd
);
7995 * perf_event_create_kernel_counter
7997 * @attr: attributes of the counter to create
7998 * @cpu: cpu in which the counter is bound
7999 * @task: task to profile (NULL for percpu)
8002 perf_event_create_kernel_counter(struct perf_event_attr
*attr
, int cpu
,
8003 struct task_struct
*task
,
8004 perf_overflow_handler_t overflow_handler
,
8007 struct perf_event_context
*ctx
;
8008 struct perf_event
*event
;
8012 * Get the target context (task or percpu):
8015 event
= perf_event_alloc(attr
, cpu
, task
, NULL
, NULL
,
8016 overflow_handler
, context
, -1);
8017 if (IS_ERR(event
)) {
8018 err
= PTR_ERR(event
);
8022 /* Mark owner so we could distinguish it from user events. */
8023 event
->owner
= EVENT_OWNER_KERNEL
;
8025 account_event(event
);
8027 ctx
= find_get_context(event
->pmu
, task
, event
);
8033 WARN_ON_ONCE(ctx
->parent_ctx
);
8034 mutex_lock(&ctx
->mutex
);
8035 perf_install_in_context(ctx
, event
, cpu
);
8036 perf_unpin_context(ctx
);
8037 mutex_unlock(&ctx
->mutex
);
8044 return ERR_PTR(err
);
8046 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter
);
8048 void perf_pmu_migrate_context(struct pmu
*pmu
, int src_cpu
, int dst_cpu
)
8050 struct perf_event_context
*src_ctx
;
8051 struct perf_event_context
*dst_ctx
;
8052 struct perf_event
*event
, *tmp
;
8055 src_ctx
= &per_cpu_ptr(pmu
->pmu_cpu_context
, src_cpu
)->ctx
;
8056 dst_ctx
= &per_cpu_ptr(pmu
->pmu_cpu_context
, dst_cpu
)->ctx
;
8059 * See perf_event_ctx_lock() for comments on the details
8060 * of swizzling perf_event::ctx.
8062 mutex_lock_double(&src_ctx
->mutex
, &dst_ctx
->mutex
);
8063 list_for_each_entry_safe(event
, tmp
, &src_ctx
->event_list
,
8065 perf_remove_from_context(event
, false);
8066 unaccount_event_cpu(event
, src_cpu
);
8068 list_add(&event
->migrate_entry
, &events
);
8072 * Wait for the events to quiesce before re-instating them.
8077 * Re-instate events in 2 passes.
8079 * Skip over group leaders and only install siblings on this first
8080 * pass, siblings will not get enabled without a leader, however a
8081 * leader will enable its siblings, even if those are still on the old
8084 list_for_each_entry_safe(event
, tmp
, &events
, migrate_entry
) {
8085 if (event
->group_leader
== event
)
8088 list_del(&event
->migrate_entry
);
8089 if (event
->state
>= PERF_EVENT_STATE_OFF
)
8090 event
->state
= PERF_EVENT_STATE_INACTIVE
;
8091 account_event_cpu(event
, dst_cpu
);
8092 perf_install_in_context(dst_ctx
, event
, dst_cpu
);
8097 * Once all the siblings are setup properly, install the group leaders
8100 list_for_each_entry_safe(event
, tmp
, &events
, migrate_entry
) {
8101 list_del(&event
->migrate_entry
);
8102 if (event
->state
>= PERF_EVENT_STATE_OFF
)
8103 event
->state
= PERF_EVENT_STATE_INACTIVE
;
8104 account_event_cpu(event
, dst_cpu
);
8105 perf_install_in_context(dst_ctx
, event
, dst_cpu
);
8108 mutex_unlock(&dst_ctx
->mutex
);
8109 mutex_unlock(&src_ctx
->mutex
);
8111 EXPORT_SYMBOL_GPL(perf_pmu_migrate_context
);
8113 static void sync_child_event(struct perf_event
*child_event
,
8114 struct task_struct
*child
)
8116 struct perf_event
*parent_event
= child_event
->parent
;
8119 if (child_event
->attr
.inherit_stat
)
8120 perf_event_read_event(child_event
, child
);
8122 child_val
= perf_event_count(child_event
);
8125 * Add back the child's count to the parent's count:
8127 atomic64_add(child_val
, &parent_event
->child_count
);
8128 atomic64_add(child_event
->total_time_enabled
,
8129 &parent_event
->child_total_time_enabled
);
8130 atomic64_add(child_event
->total_time_running
,
8131 &parent_event
->child_total_time_running
);
8134 * Remove this event from the parent's list
8136 WARN_ON_ONCE(parent_event
->ctx
->parent_ctx
);
8137 mutex_lock(&parent_event
->child_mutex
);
8138 list_del_init(&child_event
->child_list
);
8139 mutex_unlock(&parent_event
->child_mutex
);
8142 * Make sure user/parent get notified, that we just
8145 perf_event_wakeup(parent_event
);
8148 * Release the parent event, if this was the last
8151 put_event(parent_event
);
8155 __perf_event_exit_task(struct perf_event
*child_event
,
8156 struct perf_event_context
*child_ctx
,
8157 struct task_struct
*child
)
8160 * Do not destroy the 'original' grouping; because of the context
8161 * switch optimization the original events could've ended up in a
8162 * random child task.
8164 * If we were to destroy the original group, all group related
8165 * operations would cease to function properly after this random
8168 * Do destroy all inherited groups, we don't care about those
8169 * and being thorough is better.
8171 perf_remove_from_context(child_event
, !!child_event
->parent
);
8174 * It can happen that the parent exits first, and has events
8175 * that are still around due to the child reference. These
8176 * events need to be zapped.
8178 if (child_event
->parent
) {
8179 sync_child_event(child_event
, child
);
8180 free_event(child_event
);
8182 child_event
->state
= PERF_EVENT_STATE_EXIT
;
8183 perf_event_wakeup(child_event
);
8187 static void perf_event_exit_task_context(struct task_struct
*child
, int ctxn
)
8189 struct perf_event
*child_event
, *next
;
8190 struct perf_event_context
*child_ctx
, *clone_ctx
= NULL
;
8191 unsigned long flags
;
8193 if (likely(!child
->perf_event_ctxp
[ctxn
])) {
8194 perf_event_task(child
, NULL
, 0);
8198 local_irq_save(flags
);
8200 * We can't reschedule here because interrupts are disabled,
8201 * and either child is current or it is a task that can't be
8202 * scheduled, so we are now safe from rescheduling changing
8205 child_ctx
= rcu_dereference_raw(child
->perf_event_ctxp
[ctxn
]);
8208 * Take the context lock here so that if find_get_context is
8209 * reading child->perf_event_ctxp, we wait until it has
8210 * incremented the context's refcount before we do put_ctx below.
8212 raw_spin_lock(&child_ctx
->lock
);
8213 task_ctx_sched_out(child_ctx
);
8214 child
->perf_event_ctxp
[ctxn
] = NULL
;
8217 * If this context is a clone; unclone it so it can't get
8218 * swapped to another process while we're removing all
8219 * the events from it.
8221 clone_ctx
= unclone_ctx(child_ctx
);
8222 update_context_time(child_ctx
);
8223 raw_spin_unlock_irqrestore(&child_ctx
->lock
, flags
);
8229 * Report the task dead after unscheduling the events so that we
8230 * won't get any samples after PERF_RECORD_EXIT. We can however still
8231 * get a few PERF_RECORD_READ events.
8233 perf_event_task(child
, child_ctx
, 0);
8236 * We can recurse on the same lock type through:
8238 * __perf_event_exit_task()
8239 * sync_child_event()
8241 * mutex_lock(&ctx->mutex)
8243 * But since its the parent context it won't be the same instance.
8245 mutex_lock(&child_ctx
->mutex
);
8247 list_for_each_entry_safe(child_event
, next
, &child_ctx
->event_list
, event_entry
)
8248 __perf_event_exit_task(child_event
, child_ctx
, child
);
8250 mutex_unlock(&child_ctx
->mutex
);
8256 * When a child task exits, feed back event values to parent events.
8258 void perf_event_exit_task(struct task_struct
*child
)
8260 struct perf_event
*event
, *tmp
;
8263 mutex_lock(&child
->perf_event_mutex
);
8264 list_for_each_entry_safe(event
, tmp
, &child
->perf_event_list
,
8266 list_del_init(&event
->owner_entry
);
8269 * Ensure the list deletion is visible before we clear
8270 * the owner, closes a race against perf_release() where
8271 * we need to serialize on the owner->perf_event_mutex.
8274 event
->owner
= NULL
;
8276 mutex_unlock(&child
->perf_event_mutex
);
8278 for_each_task_context_nr(ctxn
)
8279 perf_event_exit_task_context(child
, ctxn
);
8282 static void perf_free_event(struct perf_event
*event
,
8283 struct perf_event_context
*ctx
)
8285 struct perf_event
*parent
= event
->parent
;
8287 if (WARN_ON_ONCE(!parent
))
8290 mutex_lock(&parent
->child_mutex
);
8291 list_del_init(&event
->child_list
);
8292 mutex_unlock(&parent
->child_mutex
);
8296 raw_spin_lock_irq(&ctx
->lock
);
8297 perf_group_detach(event
);
8298 list_del_event(event
, ctx
);
8299 raw_spin_unlock_irq(&ctx
->lock
);
8304 * Free an unexposed, unused context as created by inheritance by
8305 * perf_event_init_task below, used by fork() in case of fail.
8307 * Not all locks are strictly required, but take them anyway to be nice and
8308 * help out with the lockdep assertions.
8310 void perf_event_free_task(struct task_struct
*task
)
8312 struct perf_event_context
*ctx
;
8313 struct perf_event
*event
, *tmp
;
8316 for_each_task_context_nr(ctxn
) {
8317 ctx
= task
->perf_event_ctxp
[ctxn
];
8321 mutex_lock(&ctx
->mutex
);
8323 list_for_each_entry_safe(event
, tmp
, &ctx
->pinned_groups
,
8325 perf_free_event(event
, ctx
);
8327 list_for_each_entry_safe(event
, tmp
, &ctx
->flexible_groups
,
8329 perf_free_event(event
, ctx
);
8331 if (!list_empty(&ctx
->pinned_groups
) ||
8332 !list_empty(&ctx
->flexible_groups
))
8335 mutex_unlock(&ctx
->mutex
);
8341 void perf_event_delayed_put(struct task_struct
*task
)
8345 for_each_task_context_nr(ctxn
)
8346 WARN_ON_ONCE(task
->perf_event_ctxp
[ctxn
]);
8350 * inherit a event from parent task to child task:
8352 static struct perf_event
*
8353 inherit_event(struct perf_event
*parent_event
,
8354 struct task_struct
*parent
,
8355 struct perf_event_context
*parent_ctx
,
8356 struct task_struct
*child
,
8357 struct perf_event
*group_leader
,
8358 struct perf_event_context
*child_ctx
)
8360 enum perf_event_active_state parent_state
= parent_event
->state
;
8361 struct perf_event
*child_event
;
8362 unsigned long flags
;
8365 * Instead of creating recursive hierarchies of events,
8366 * we link inherited events back to the original parent,
8367 * which has a filp for sure, which we use as the reference
8370 if (parent_event
->parent
)
8371 parent_event
= parent_event
->parent
;
8373 child_event
= perf_event_alloc(&parent_event
->attr
,
8376 group_leader
, parent_event
,
8378 if (IS_ERR(child_event
))
8381 if (is_orphaned_event(parent_event
) ||
8382 !atomic_long_inc_not_zero(&parent_event
->refcount
)) {
8383 free_event(child_event
);
8390 * Make the child state follow the state of the parent event,
8391 * not its attr.disabled bit. We hold the parent's mutex,
8392 * so we won't race with perf_event_{en, dis}able_family.
8394 if (parent_state
>= PERF_EVENT_STATE_INACTIVE
)
8395 child_event
->state
= PERF_EVENT_STATE_INACTIVE
;
8397 child_event
->state
= PERF_EVENT_STATE_OFF
;
8399 if (parent_event
->attr
.freq
) {
8400 u64 sample_period
= parent_event
->hw
.sample_period
;
8401 struct hw_perf_event
*hwc
= &child_event
->hw
;
8403 hwc
->sample_period
= sample_period
;
8404 hwc
->last_period
= sample_period
;
8406 local64_set(&hwc
->period_left
, sample_period
);
8409 child_event
->ctx
= child_ctx
;
8410 child_event
->overflow_handler
= parent_event
->overflow_handler
;
8411 child_event
->overflow_handler_context
8412 = parent_event
->overflow_handler_context
;
8415 * Precalculate sample_data sizes
8417 perf_event__header_size(child_event
);
8418 perf_event__id_header_size(child_event
);
8421 * Link it up in the child's context:
8423 raw_spin_lock_irqsave(&child_ctx
->lock
, flags
);
8424 add_event_to_ctx(child_event
, child_ctx
);
8425 raw_spin_unlock_irqrestore(&child_ctx
->lock
, flags
);
8428 * Link this into the parent event's child list
8430 WARN_ON_ONCE(parent_event
->ctx
->parent_ctx
);
8431 mutex_lock(&parent_event
->child_mutex
);
8432 list_add_tail(&child_event
->child_list
, &parent_event
->child_list
);
8433 mutex_unlock(&parent_event
->child_mutex
);
8438 static int inherit_group(struct perf_event
*parent_event
,
8439 struct task_struct
*parent
,
8440 struct perf_event_context
*parent_ctx
,
8441 struct task_struct
*child
,
8442 struct perf_event_context
*child_ctx
)
8444 struct perf_event
*leader
;
8445 struct perf_event
*sub
;
8446 struct perf_event
*child_ctr
;
8448 leader
= inherit_event(parent_event
, parent
, parent_ctx
,
8449 child
, NULL
, child_ctx
);
8451 return PTR_ERR(leader
);
8452 list_for_each_entry(sub
, &parent_event
->sibling_list
, group_entry
) {
8453 child_ctr
= inherit_event(sub
, parent
, parent_ctx
,
8454 child
, leader
, child_ctx
);
8455 if (IS_ERR(child_ctr
))
8456 return PTR_ERR(child_ctr
);
8462 inherit_task_group(struct perf_event
*event
, struct task_struct
*parent
,
8463 struct perf_event_context
*parent_ctx
,
8464 struct task_struct
*child
, int ctxn
,
8468 struct perf_event_context
*child_ctx
;
8470 if (!event
->attr
.inherit
) {
8475 child_ctx
= child
->perf_event_ctxp
[ctxn
];
8478 * This is executed from the parent task context, so
8479 * inherit events that have been marked for cloning.
8480 * First allocate and initialize a context for the
8484 child_ctx
= alloc_perf_context(parent_ctx
->pmu
, child
);
8488 child
->perf_event_ctxp
[ctxn
] = child_ctx
;
8491 ret
= inherit_group(event
, parent
, parent_ctx
,
8501 * Initialize the perf_event context in task_struct
8503 static int perf_event_init_context(struct task_struct
*child
, int ctxn
)
8505 struct perf_event_context
*child_ctx
, *parent_ctx
;
8506 struct perf_event_context
*cloned_ctx
;
8507 struct perf_event
*event
;
8508 struct task_struct
*parent
= current
;
8509 int inherited_all
= 1;
8510 unsigned long flags
;
8513 if (likely(!parent
->perf_event_ctxp
[ctxn
]))
8517 * If the parent's context is a clone, pin it so it won't get
8520 parent_ctx
= perf_pin_task_context(parent
, ctxn
);
8525 * No need to check if parent_ctx != NULL here; since we saw
8526 * it non-NULL earlier, the only reason for it to become NULL
8527 * is if we exit, and since we're currently in the middle of
8528 * a fork we can't be exiting at the same time.
8532 * Lock the parent list. No need to lock the child - not PID
8533 * hashed yet and not running, so nobody can access it.
8535 mutex_lock(&parent_ctx
->mutex
);
8538 * We dont have to disable NMIs - we are only looking at
8539 * the list, not manipulating it:
8541 list_for_each_entry(event
, &parent_ctx
->pinned_groups
, group_entry
) {
8542 ret
= inherit_task_group(event
, parent
, parent_ctx
,
8543 child
, ctxn
, &inherited_all
);
8549 * We can't hold ctx->lock when iterating the ->flexible_group list due
8550 * to allocations, but we need to prevent rotation because
8551 * rotate_ctx() will change the list from interrupt context.
8553 raw_spin_lock_irqsave(&parent_ctx
->lock
, flags
);
8554 parent_ctx
->rotate_disable
= 1;
8555 raw_spin_unlock_irqrestore(&parent_ctx
->lock
, flags
);
8557 list_for_each_entry(event
, &parent_ctx
->flexible_groups
, group_entry
) {
8558 ret
= inherit_task_group(event
, parent
, parent_ctx
,
8559 child
, ctxn
, &inherited_all
);
8564 raw_spin_lock_irqsave(&parent_ctx
->lock
, flags
);
8565 parent_ctx
->rotate_disable
= 0;
8567 child_ctx
= child
->perf_event_ctxp
[ctxn
];
8569 if (child_ctx
&& inherited_all
) {
8571 * Mark the child context as a clone of the parent
8572 * context, or of whatever the parent is a clone of.
8574 * Note that if the parent is a clone, the holding of
8575 * parent_ctx->lock avoids it from being uncloned.
8577 cloned_ctx
= parent_ctx
->parent_ctx
;
8579 child_ctx
->parent_ctx
= cloned_ctx
;
8580 child_ctx
->parent_gen
= parent_ctx
->parent_gen
;
8582 child_ctx
->parent_ctx
= parent_ctx
;
8583 child_ctx
->parent_gen
= parent_ctx
->generation
;
8585 get_ctx(child_ctx
->parent_ctx
);
8588 raw_spin_unlock_irqrestore(&parent_ctx
->lock
, flags
);
8589 mutex_unlock(&parent_ctx
->mutex
);
8591 perf_unpin_context(parent_ctx
);
8592 put_ctx(parent_ctx
);
8598 * Initialize the perf_event context in task_struct
8600 int perf_event_init_task(struct task_struct
*child
)
8604 memset(child
->perf_event_ctxp
, 0, sizeof(child
->perf_event_ctxp
));
8605 mutex_init(&child
->perf_event_mutex
);
8606 INIT_LIST_HEAD(&child
->perf_event_list
);
8608 for_each_task_context_nr(ctxn
) {
8609 ret
= perf_event_init_context(child
, ctxn
);
8611 perf_event_free_task(child
);
8619 static void __init
perf_event_init_all_cpus(void)
8621 struct swevent_htable
*swhash
;
8624 for_each_possible_cpu(cpu
) {
8625 swhash
= &per_cpu(swevent_htable
, cpu
);
8626 mutex_init(&swhash
->hlist_mutex
);
8627 INIT_LIST_HEAD(&per_cpu(active_ctx_list
, cpu
));
8631 static void perf_event_init_cpu(int cpu
)
8633 struct swevent_htable
*swhash
= &per_cpu(swevent_htable
, cpu
);
8635 mutex_lock(&swhash
->hlist_mutex
);
8636 swhash
->online
= true;
8637 if (swhash
->hlist_refcount
> 0) {
8638 struct swevent_hlist
*hlist
;
8640 hlist
= kzalloc_node(sizeof(*hlist
), GFP_KERNEL
, cpu_to_node(cpu
));
8642 rcu_assign_pointer(swhash
->swevent_hlist
, hlist
);
8644 mutex_unlock(&swhash
->hlist_mutex
);
8647 #if defined CONFIG_HOTPLUG_CPU || defined CONFIG_KEXEC
8648 static void __perf_event_exit_context(void *__info
)
8650 struct remove_event re
= { .detach_group
= true };
8651 struct perf_event_context
*ctx
= __info
;
8654 list_for_each_entry_rcu(re
.event
, &ctx
->event_list
, event_entry
)
8655 __perf_remove_from_context(&re
);
8659 static void perf_event_exit_cpu_context(int cpu
)
8661 struct perf_event_context
*ctx
;
8665 idx
= srcu_read_lock(&pmus_srcu
);
8666 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
8667 ctx
= &per_cpu_ptr(pmu
->pmu_cpu_context
, cpu
)->ctx
;
8669 mutex_lock(&ctx
->mutex
);
8670 smp_call_function_single(cpu
, __perf_event_exit_context
, ctx
, 1);
8671 mutex_unlock(&ctx
->mutex
);
8673 srcu_read_unlock(&pmus_srcu
, idx
);
8676 static void perf_event_exit_cpu(int cpu
)
8678 struct swevent_htable
*swhash
= &per_cpu(swevent_htable
, cpu
);
8680 perf_event_exit_cpu_context(cpu
);
8682 mutex_lock(&swhash
->hlist_mutex
);
8683 swhash
->online
= false;
8684 swevent_hlist_release(swhash
);
8685 mutex_unlock(&swhash
->hlist_mutex
);
8688 static inline void perf_event_exit_cpu(int cpu
) { }
8692 perf_reboot(struct notifier_block
*notifier
, unsigned long val
, void *v
)
8696 for_each_online_cpu(cpu
)
8697 perf_event_exit_cpu(cpu
);
8703 * Run the perf reboot notifier at the very last possible moment so that
8704 * the generic watchdog code runs as long as possible.
8706 static struct notifier_block perf_reboot_notifier
= {
8707 .notifier_call
= perf_reboot
,
8708 .priority
= INT_MIN
,
8712 perf_cpu_notify(struct notifier_block
*self
, unsigned long action
, void *hcpu
)
8714 unsigned int cpu
= (long)hcpu
;
8716 switch (action
& ~CPU_TASKS_FROZEN
) {
8718 case CPU_UP_PREPARE
:
8719 case CPU_DOWN_FAILED
:
8720 perf_event_init_cpu(cpu
);
8723 case CPU_UP_CANCELED
:
8724 case CPU_DOWN_PREPARE
:
8725 perf_event_exit_cpu(cpu
);
8734 void __init
perf_event_init(void)
8740 perf_event_init_all_cpus();
8741 init_srcu_struct(&pmus_srcu
);
8742 perf_pmu_register(&perf_swevent
, "software", PERF_TYPE_SOFTWARE
);
8743 perf_pmu_register(&perf_cpu_clock
, NULL
, -1);
8744 perf_pmu_register(&perf_task_clock
, NULL
, -1);
8746 perf_cpu_notifier(perf_cpu_notify
);
8747 register_reboot_notifier(&perf_reboot_notifier
);
8749 ret
= init_hw_breakpoint();
8750 WARN(ret
, "hw_breakpoint initialization failed with: %d", ret
);
8752 /* do not patch jump label more than once per second */
8753 jump_label_rate_limit(&perf_sched_events
, HZ
);
8756 * Build time assertion that we keep the data_head at the intended
8757 * location. IOW, validation we got the __reserved[] size right.
8759 BUILD_BUG_ON((offsetof(struct perf_event_mmap_page
, data_head
))
8763 ssize_t
perf_event_sysfs_show(struct device
*dev
, struct device_attribute
*attr
,
8766 struct perf_pmu_events_attr
*pmu_attr
=
8767 container_of(attr
, struct perf_pmu_events_attr
, attr
);
8769 if (pmu_attr
->event_str
)
8770 return sprintf(page
, "%s\n", pmu_attr
->event_str
);
8775 static int __init
perf_event_sysfs_init(void)
8780 mutex_lock(&pmus_lock
);
8782 ret
= bus_register(&pmu_bus
);
8786 list_for_each_entry(pmu
, &pmus
, entry
) {
8787 if (!pmu
->name
|| pmu
->type
< 0)
8790 ret
= pmu_dev_alloc(pmu
);
8791 WARN(ret
, "Failed to register pmu: %s, reason %d\n", pmu
->name
, ret
);
8793 pmu_bus_running
= 1;
8797 mutex_unlock(&pmus_lock
);
8801 device_initcall(perf_event_sysfs_init
);
8803 #ifdef CONFIG_CGROUP_PERF
8804 static struct cgroup_subsys_state
*
8805 perf_cgroup_css_alloc(struct cgroup_subsys_state
*parent_css
)
8807 struct perf_cgroup
*jc
;
8809 jc
= kzalloc(sizeof(*jc
), GFP_KERNEL
);
8811 return ERR_PTR(-ENOMEM
);
8813 jc
->info
= alloc_percpu(struct perf_cgroup_info
);
8816 return ERR_PTR(-ENOMEM
);
8822 static void perf_cgroup_css_free(struct cgroup_subsys_state
*css
)
8824 struct perf_cgroup
*jc
= container_of(css
, struct perf_cgroup
, css
);
8826 free_percpu(jc
->info
);
8830 static int __perf_cgroup_move(void *info
)
8832 struct task_struct
*task
= info
;
8833 perf_cgroup_switch(task
, PERF_CGROUP_SWOUT
| PERF_CGROUP_SWIN
);
8837 static void perf_cgroup_attach(struct cgroup_subsys_state
*css
,
8838 struct cgroup_taskset
*tset
)
8840 struct task_struct
*task
;
8842 cgroup_taskset_for_each(task
, tset
)
8843 task_function_call(task
, __perf_cgroup_move
, task
);
8846 static void perf_cgroup_exit(struct cgroup_subsys_state
*css
,
8847 struct cgroup_subsys_state
*old_css
,
8848 struct task_struct
*task
)
8851 * cgroup_exit() is called in the copy_process() failure path.
8852 * Ignore this case since the task hasn't ran yet, this avoids
8853 * trying to poke a half freed task state from generic code.
8855 if (!(task
->flags
& PF_EXITING
))
8858 task_function_call(task
, __perf_cgroup_move
, task
);
8861 struct cgroup_subsys perf_event_cgrp_subsys
= {
8862 .css_alloc
= perf_cgroup_css_alloc
,
8863 .css_free
= perf_cgroup_css_free
,
8864 .exit
= perf_cgroup_exit
,
8865 .attach
= perf_cgroup_attach
,
8867 #endif /* CONFIG_CGROUP_PERF */