2 * Performance events core code:
4 * Copyright (C) 2008 Thomas Gleixner <tglx@linutronix.de>
5 * Copyright (C) 2008-2011 Red Hat, Inc., Ingo Molnar
6 * Copyright (C) 2008-2011 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
7 * Copyright © 2009 Paul Mackerras, IBM Corp. <paulus@au1.ibm.com>
9 * For licensing details see kernel-base/COPYING
14 #include <linux/cpu.h>
15 #include <linux/smp.h>
16 #include <linux/idr.h>
17 #include <linux/file.h>
18 #include <linux/poll.h>
19 #include <linux/slab.h>
20 #include <linux/hash.h>
21 #include <linux/tick.h>
22 #include <linux/sysfs.h>
23 #include <linux/dcache.h>
24 #include <linux/percpu.h>
25 #include <linux/ptrace.h>
26 #include <linux/reboot.h>
27 #include <linux/vmstat.h>
28 #include <linux/device.h>
29 #include <linux/export.h>
30 #include <linux/vmalloc.h>
31 #include <linux/hardirq.h>
32 #include <linux/rculist.h>
33 #include <linux/uaccess.h>
34 #include <linux/syscalls.h>
35 #include <linux/anon_inodes.h>
36 #include <linux/kernel_stat.h>
37 #include <linux/perf_event.h>
38 #include <linux/ftrace_event.h>
39 #include <linux/hw_breakpoint.h>
40 #include <linux/mm_types.h>
41 #include <linux/cgroup.h>
42 #include <linux/module.h>
46 #include <asm/irq_regs.h>
48 struct remote_function_call
{
49 struct task_struct
*p
;
50 int (*func
)(void *info
);
55 static void remote_function(void *data
)
57 struct remote_function_call
*tfc
= data
;
58 struct task_struct
*p
= tfc
->p
;
62 if (task_cpu(p
) != smp_processor_id() || !task_curr(p
))
66 tfc
->ret
= tfc
->func(tfc
->info
);
70 * task_function_call - call a function on the cpu on which a task runs
71 * @p: the task to evaluate
72 * @func: the function to be called
73 * @info: the function call argument
75 * Calls the function @func when the task is currently running. This might
76 * be on the current CPU, which just calls the function directly
78 * returns: @func return value, or
79 * -ESRCH - when the process isn't running
80 * -EAGAIN - when the process moved away
83 task_function_call(struct task_struct
*p
, int (*func
) (void *info
), void *info
)
85 struct remote_function_call data
= {
89 .ret
= -ESRCH
, /* No such (running) process */
93 smp_call_function_single(task_cpu(p
), remote_function
, &data
, 1);
99 * cpu_function_call - call a function on the cpu
100 * @func: the function to be called
101 * @info: the function call argument
103 * Calls the function @func on the remote cpu.
105 * returns: @func return value or -ENXIO when the cpu is offline
107 static int cpu_function_call(int cpu
, int (*func
) (void *info
), void *info
)
109 struct remote_function_call data
= {
113 .ret
= -ENXIO
, /* No such CPU */
116 smp_call_function_single(cpu
, remote_function
, &data
, 1);
121 #define PERF_FLAG_ALL (PERF_FLAG_FD_NO_GROUP |\
122 PERF_FLAG_FD_OUTPUT |\
123 PERF_FLAG_PID_CGROUP |\
124 PERF_FLAG_FD_CLOEXEC)
127 * branch priv levels that need permission checks
129 #define PERF_SAMPLE_BRANCH_PERM_PLM \
130 (PERF_SAMPLE_BRANCH_KERNEL |\
131 PERF_SAMPLE_BRANCH_HV)
134 EVENT_FLEXIBLE
= 0x1,
136 EVENT_ALL
= EVENT_FLEXIBLE
| EVENT_PINNED
,
140 * perf_sched_events : >0 events exist
141 * perf_cgroup_events: >0 per-cpu cgroup events exist on this cpu
143 struct static_key_deferred perf_sched_events __read_mostly
;
144 static DEFINE_PER_CPU(atomic_t
, perf_cgroup_events
);
145 static DEFINE_PER_CPU(atomic_t
, perf_branch_stack_events
);
147 static atomic_t nr_mmap_events __read_mostly
;
148 static atomic_t nr_comm_events __read_mostly
;
149 static atomic_t nr_task_events __read_mostly
;
150 static atomic_t nr_freq_events __read_mostly
;
152 static LIST_HEAD(pmus
);
153 static DEFINE_MUTEX(pmus_lock
);
154 static struct srcu_struct pmus_srcu
;
157 * perf event paranoia level:
158 * -1 - not paranoid at all
159 * 0 - disallow raw tracepoint access for unpriv
160 * 1 - disallow cpu events for unpriv
161 * 2 - disallow kernel profiling for unpriv
163 int sysctl_perf_event_paranoid __read_mostly
= 1;
165 /* Minimum for 512 kiB + 1 user control page */
166 int sysctl_perf_event_mlock __read_mostly
= 512 + (PAGE_SIZE
/ 1024); /* 'free' kiB per user */
169 * max perf event sample rate
171 #define DEFAULT_MAX_SAMPLE_RATE 100000
172 #define DEFAULT_SAMPLE_PERIOD_NS (NSEC_PER_SEC / DEFAULT_MAX_SAMPLE_RATE)
173 #define DEFAULT_CPU_TIME_MAX_PERCENT 25
175 int sysctl_perf_event_sample_rate __read_mostly
= DEFAULT_MAX_SAMPLE_RATE
;
177 static int max_samples_per_tick __read_mostly
= DIV_ROUND_UP(DEFAULT_MAX_SAMPLE_RATE
, HZ
);
178 static int perf_sample_period_ns __read_mostly
= DEFAULT_SAMPLE_PERIOD_NS
;
180 static int perf_sample_allowed_ns __read_mostly
=
181 DEFAULT_SAMPLE_PERIOD_NS
* DEFAULT_CPU_TIME_MAX_PERCENT
/ 100;
183 void update_perf_cpu_limits(void)
185 u64 tmp
= perf_sample_period_ns
;
187 tmp
*= sysctl_perf_cpu_time_max_percent
;
189 ACCESS_ONCE(perf_sample_allowed_ns
) = tmp
;
192 static int perf_rotate_context(struct perf_cpu_context
*cpuctx
);
194 int perf_proc_update_handler(struct ctl_table
*table
, int write
,
195 void __user
*buffer
, size_t *lenp
,
198 int ret
= proc_dointvec_minmax(table
, write
, buffer
, lenp
, ppos
);
203 max_samples_per_tick
= DIV_ROUND_UP(sysctl_perf_event_sample_rate
, HZ
);
204 perf_sample_period_ns
= NSEC_PER_SEC
/ sysctl_perf_event_sample_rate
;
205 update_perf_cpu_limits();
210 int sysctl_perf_cpu_time_max_percent __read_mostly
= DEFAULT_CPU_TIME_MAX_PERCENT
;
212 int perf_cpu_time_max_percent_handler(struct ctl_table
*table
, int write
,
213 void __user
*buffer
, size_t *lenp
,
216 int ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
221 update_perf_cpu_limits();
227 * perf samples are done in some very critical code paths (NMIs).
228 * If they take too much CPU time, the system can lock up and not
229 * get any real work done. This will drop the sample rate when
230 * we detect that events are taking too long.
232 #define NR_ACCUMULATED_SAMPLES 128
233 static DEFINE_PER_CPU(u64
, running_sample_length
);
235 static void perf_duration_warn(struct irq_work
*w
)
237 u64 allowed_ns
= ACCESS_ONCE(perf_sample_allowed_ns
);
238 u64 avg_local_sample_len
;
239 u64 local_samples_len
;
241 local_samples_len
= __get_cpu_var(running_sample_length
);
242 avg_local_sample_len
= local_samples_len
/NR_ACCUMULATED_SAMPLES
;
244 printk_ratelimited(KERN_WARNING
245 "perf interrupt took too long (%lld > %lld), lowering "
246 "kernel.perf_event_max_sample_rate to %d\n",
247 avg_local_sample_len
, allowed_ns
>> 1,
248 sysctl_perf_event_sample_rate
);
251 static DEFINE_IRQ_WORK(perf_duration_work
, perf_duration_warn
);
253 void perf_sample_event_took(u64 sample_len_ns
)
255 u64 allowed_ns
= ACCESS_ONCE(perf_sample_allowed_ns
);
256 u64 avg_local_sample_len
;
257 u64 local_samples_len
;
262 /* decay the counter by 1 average sample */
263 local_samples_len
= __get_cpu_var(running_sample_length
);
264 local_samples_len
-= local_samples_len
/NR_ACCUMULATED_SAMPLES
;
265 local_samples_len
+= sample_len_ns
;
266 __get_cpu_var(running_sample_length
) = local_samples_len
;
269 * note: this will be biased artifically low until we have
270 * seen NR_ACCUMULATED_SAMPLES. Doing it this way keeps us
271 * from having to maintain a count.
273 avg_local_sample_len
= local_samples_len
/NR_ACCUMULATED_SAMPLES
;
275 if (avg_local_sample_len
<= allowed_ns
)
278 if (max_samples_per_tick
<= 1)
281 max_samples_per_tick
= DIV_ROUND_UP(max_samples_per_tick
, 2);
282 sysctl_perf_event_sample_rate
= max_samples_per_tick
* HZ
;
283 perf_sample_period_ns
= NSEC_PER_SEC
/ sysctl_perf_event_sample_rate
;
285 update_perf_cpu_limits();
287 if (!irq_work_queue(&perf_duration_work
)) {
288 early_printk("perf interrupt took too long (%lld > %lld), lowering "
289 "kernel.perf_event_max_sample_rate to %d\n",
290 avg_local_sample_len
, allowed_ns
>> 1,
291 sysctl_perf_event_sample_rate
);
295 static atomic64_t perf_event_id
;
297 static void cpu_ctx_sched_out(struct perf_cpu_context
*cpuctx
,
298 enum event_type_t event_type
);
300 static void cpu_ctx_sched_in(struct perf_cpu_context
*cpuctx
,
301 enum event_type_t event_type
,
302 struct task_struct
*task
);
304 static void update_context_time(struct perf_event_context
*ctx
);
305 static u64
perf_event_time(struct perf_event
*event
);
307 void __weak
perf_event_print_debug(void) { }
309 extern __weak
const char *perf_pmu_name(void)
314 static inline u64
perf_clock(void)
316 return local_clock();
319 static inline struct perf_cpu_context
*
320 __get_cpu_context(struct perf_event_context
*ctx
)
322 return this_cpu_ptr(ctx
->pmu
->pmu_cpu_context
);
325 static void perf_ctx_lock(struct perf_cpu_context
*cpuctx
,
326 struct perf_event_context
*ctx
)
328 raw_spin_lock(&cpuctx
->ctx
.lock
);
330 raw_spin_lock(&ctx
->lock
);
333 static void perf_ctx_unlock(struct perf_cpu_context
*cpuctx
,
334 struct perf_event_context
*ctx
)
337 raw_spin_unlock(&ctx
->lock
);
338 raw_spin_unlock(&cpuctx
->ctx
.lock
);
341 #ifdef CONFIG_CGROUP_PERF
344 * perf_cgroup_info keeps track of time_enabled for a cgroup.
345 * This is a per-cpu dynamically allocated data structure.
347 struct perf_cgroup_info
{
353 struct cgroup_subsys_state css
;
354 struct perf_cgroup_info __percpu
*info
;
358 * Must ensure cgroup is pinned (css_get) before calling
359 * this function. In other words, we cannot call this function
360 * if there is no cgroup event for the current CPU context.
362 static inline struct perf_cgroup
*
363 perf_cgroup_from_task(struct task_struct
*task
)
365 return container_of(task_css(task
, perf_event_cgrp_id
),
366 struct perf_cgroup
, css
);
370 perf_cgroup_match(struct perf_event
*event
)
372 struct perf_event_context
*ctx
= event
->ctx
;
373 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
375 /* @event doesn't care about cgroup */
379 /* wants specific cgroup scope but @cpuctx isn't associated with any */
384 * Cgroup scoping is recursive. An event enabled for a cgroup is
385 * also enabled for all its descendant cgroups. If @cpuctx's
386 * cgroup is a descendant of @event's (the test covers identity
387 * case), it's a match.
389 return cgroup_is_descendant(cpuctx
->cgrp
->css
.cgroup
,
390 event
->cgrp
->css
.cgroup
);
393 static inline void perf_put_cgroup(struct perf_event
*event
)
395 css_put(&event
->cgrp
->css
);
398 static inline void perf_detach_cgroup(struct perf_event
*event
)
400 perf_put_cgroup(event
);
404 static inline int is_cgroup_event(struct perf_event
*event
)
406 return event
->cgrp
!= NULL
;
409 static inline u64
perf_cgroup_event_time(struct perf_event
*event
)
411 struct perf_cgroup_info
*t
;
413 t
= per_cpu_ptr(event
->cgrp
->info
, event
->cpu
);
417 static inline void __update_cgrp_time(struct perf_cgroup
*cgrp
)
419 struct perf_cgroup_info
*info
;
424 info
= this_cpu_ptr(cgrp
->info
);
426 info
->time
+= now
- info
->timestamp
;
427 info
->timestamp
= now
;
430 static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context
*cpuctx
)
432 struct perf_cgroup
*cgrp_out
= cpuctx
->cgrp
;
434 __update_cgrp_time(cgrp_out
);
437 static inline void update_cgrp_time_from_event(struct perf_event
*event
)
439 struct perf_cgroup
*cgrp
;
442 * ensure we access cgroup data only when needed and
443 * when we know the cgroup is pinned (css_get)
445 if (!is_cgroup_event(event
))
448 cgrp
= perf_cgroup_from_task(current
);
450 * Do not update time when cgroup is not active
452 if (cgrp
== event
->cgrp
)
453 __update_cgrp_time(event
->cgrp
);
457 perf_cgroup_set_timestamp(struct task_struct
*task
,
458 struct perf_event_context
*ctx
)
460 struct perf_cgroup
*cgrp
;
461 struct perf_cgroup_info
*info
;
464 * ctx->lock held by caller
465 * ensure we do not access cgroup data
466 * unless we have the cgroup pinned (css_get)
468 if (!task
|| !ctx
->nr_cgroups
)
471 cgrp
= perf_cgroup_from_task(task
);
472 info
= this_cpu_ptr(cgrp
->info
);
473 info
->timestamp
= ctx
->timestamp
;
476 #define PERF_CGROUP_SWOUT 0x1 /* cgroup switch out every event */
477 #define PERF_CGROUP_SWIN 0x2 /* cgroup switch in events based on task */
480 * reschedule events based on the cgroup constraint of task.
482 * mode SWOUT : schedule out everything
483 * mode SWIN : schedule in based on cgroup for next
485 void perf_cgroup_switch(struct task_struct
*task
, int mode
)
487 struct perf_cpu_context
*cpuctx
;
492 * disable interrupts to avoid geting nr_cgroup
493 * changes via __perf_event_disable(). Also
496 local_irq_save(flags
);
499 * we reschedule only in the presence of cgroup
500 * constrained events.
504 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
505 cpuctx
= this_cpu_ptr(pmu
->pmu_cpu_context
);
506 if (cpuctx
->unique_pmu
!= pmu
)
507 continue; /* ensure we process each cpuctx once */
510 * perf_cgroup_events says at least one
511 * context on this CPU has cgroup events.
513 * ctx->nr_cgroups reports the number of cgroup
514 * events for a context.
516 if (cpuctx
->ctx
.nr_cgroups
> 0) {
517 perf_ctx_lock(cpuctx
, cpuctx
->task_ctx
);
518 perf_pmu_disable(cpuctx
->ctx
.pmu
);
520 if (mode
& PERF_CGROUP_SWOUT
) {
521 cpu_ctx_sched_out(cpuctx
, EVENT_ALL
);
523 * must not be done before ctxswout due
524 * to event_filter_match() in event_sched_out()
529 if (mode
& PERF_CGROUP_SWIN
) {
530 WARN_ON_ONCE(cpuctx
->cgrp
);
532 * set cgrp before ctxsw in to allow
533 * event_filter_match() to not have to pass
536 cpuctx
->cgrp
= perf_cgroup_from_task(task
);
537 cpu_ctx_sched_in(cpuctx
, EVENT_ALL
, task
);
539 perf_pmu_enable(cpuctx
->ctx
.pmu
);
540 perf_ctx_unlock(cpuctx
, cpuctx
->task_ctx
);
546 local_irq_restore(flags
);
549 static inline void perf_cgroup_sched_out(struct task_struct
*task
,
550 struct task_struct
*next
)
552 struct perf_cgroup
*cgrp1
;
553 struct perf_cgroup
*cgrp2
= NULL
;
556 * we come here when we know perf_cgroup_events > 0
558 cgrp1
= perf_cgroup_from_task(task
);
561 * next is NULL when called from perf_event_enable_on_exec()
562 * that will systematically cause a cgroup_switch()
565 cgrp2
= perf_cgroup_from_task(next
);
568 * only schedule out current cgroup events if we know
569 * that we are switching to a different cgroup. Otherwise,
570 * do no touch the cgroup events.
573 perf_cgroup_switch(task
, PERF_CGROUP_SWOUT
);
576 static inline void perf_cgroup_sched_in(struct task_struct
*prev
,
577 struct task_struct
*task
)
579 struct perf_cgroup
*cgrp1
;
580 struct perf_cgroup
*cgrp2
= NULL
;
583 * we come here when we know perf_cgroup_events > 0
585 cgrp1
= perf_cgroup_from_task(task
);
587 /* prev can never be NULL */
588 cgrp2
= perf_cgroup_from_task(prev
);
591 * only need to schedule in cgroup events if we are changing
592 * cgroup during ctxsw. Cgroup events were not scheduled
593 * out of ctxsw out if that was not the case.
596 perf_cgroup_switch(task
, PERF_CGROUP_SWIN
);
599 static inline int perf_cgroup_connect(int fd
, struct perf_event
*event
,
600 struct perf_event_attr
*attr
,
601 struct perf_event
*group_leader
)
603 struct perf_cgroup
*cgrp
;
604 struct cgroup_subsys_state
*css
;
605 struct fd f
= fdget(fd
);
611 css
= css_tryget_online_from_dir(f
.file
->f_dentry
,
612 &perf_event_cgrp_subsys
);
618 cgrp
= container_of(css
, struct perf_cgroup
, css
);
622 * all events in a group must monitor
623 * the same cgroup because a task belongs
624 * to only one perf cgroup at a time
626 if (group_leader
&& group_leader
->cgrp
!= cgrp
) {
627 perf_detach_cgroup(event
);
636 perf_cgroup_set_shadow_time(struct perf_event
*event
, u64 now
)
638 struct perf_cgroup_info
*t
;
639 t
= per_cpu_ptr(event
->cgrp
->info
, event
->cpu
);
640 event
->shadow_ctx_time
= now
- t
->timestamp
;
644 perf_cgroup_defer_enabled(struct perf_event
*event
)
647 * when the current task's perf cgroup does not match
648 * the event's, we need to remember to call the
649 * perf_mark_enable() function the first time a task with
650 * a matching perf cgroup is scheduled in.
652 if (is_cgroup_event(event
) && !perf_cgroup_match(event
))
653 event
->cgrp_defer_enabled
= 1;
657 perf_cgroup_mark_enabled(struct perf_event
*event
,
658 struct perf_event_context
*ctx
)
660 struct perf_event
*sub
;
661 u64 tstamp
= perf_event_time(event
);
663 if (!event
->cgrp_defer_enabled
)
666 event
->cgrp_defer_enabled
= 0;
668 event
->tstamp_enabled
= tstamp
- event
->total_time_enabled
;
669 list_for_each_entry(sub
, &event
->sibling_list
, group_entry
) {
670 if (sub
->state
>= PERF_EVENT_STATE_INACTIVE
) {
671 sub
->tstamp_enabled
= tstamp
- sub
->total_time_enabled
;
672 sub
->cgrp_defer_enabled
= 0;
676 #else /* !CONFIG_CGROUP_PERF */
679 perf_cgroup_match(struct perf_event
*event
)
684 static inline void perf_detach_cgroup(struct perf_event
*event
)
687 static inline int is_cgroup_event(struct perf_event
*event
)
692 static inline u64
perf_cgroup_event_cgrp_time(struct perf_event
*event
)
697 static inline void update_cgrp_time_from_event(struct perf_event
*event
)
701 static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context
*cpuctx
)
705 static inline void perf_cgroup_sched_out(struct task_struct
*task
,
706 struct task_struct
*next
)
710 static inline void perf_cgroup_sched_in(struct task_struct
*prev
,
711 struct task_struct
*task
)
715 static inline int perf_cgroup_connect(pid_t pid
, struct perf_event
*event
,
716 struct perf_event_attr
*attr
,
717 struct perf_event
*group_leader
)
723 perf_cgroup_set_timestamp(struct task_struct
*task
,
724 struct perf_event_context
*ctx
)
729 perf_cgroup_switch(struct task_struct
*task
, struct task_struct
*next
)
734 perf_cgroup_set_shadow_time(struct perf_event
*event
, u64 now
)
738 static inline u64
perf_cgroup_event_time(struct perf_event
*event
)
744 perf_cgroup_defer_enabled(struct perf_event
*event
)
749 perf_cgroup_mark_enabled(struct perf_event
*event
,
750 struct perf_event_context
*ctx
)
756 * set default to be dependent on timer tick just
759 #define PERF_CPU_HRTIMER (1000 / HZ)
761 * function must be called with interrupts disbled
763 static enum hrtimer_restart
perf_cpu_hrtimer_handler(struct hrtimer
*hr
)
765 struct perf_cpu_context
*cpuctx
;
766 enum hrtimer_restart ret
= HRTIMER_NORESTART
;
769 WARN_ON(!irqs_disabled());
771 cpuctx
= container_of(hr
, struct perf_cpu_context
, hrtimer
);
773 rotations
= perf_rotate_context(cpuctx
);
776 * arm timer if needed
779 hrtimer_forward_now(hr
, cpuctx
->hrtimer_interval
);
780 ret
= HRTIMER_RESTART
;
786 /* CPU is going down */
787 void perf_cpu_hrtimer_cancel(int cpu
)
789 struct perf_cpu_context
*cpuctx
;
793 if (WARN_ON(cpu
!= smp_processor_id()))
796 local_irq_save(flags
);
800 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
801 cpuctx
= this_cpu_ptr(pmu
->pmu_cpu_context
);
803 if (pmu
->task_ctx_nr
== perf_sw_context
)
806 hrtimer_cancel(&cpuctx
->hrtimer
);
811 local_irq_restore(flags
);
814 static void __perf_cpu_hrtimer_init(struct perf_cpu_context
*cpuctx
, int cpu
)
816 struct hrtimer
*hr
= &cpuctx
->hrtimer
;
817 struct pmu
*pmu
= cpuctx
->ctx
.pmu
;
820 /* no multiplexing needed for SW PMU */
821 if (pmu
->task_ctx_nr
== perf_sw_context
)
825 * check default is sane, if not set then force to
826 * default interval (1/tick)
828 timer
= pmu
->hrtimer_interval_ms
;
830 timer
= pmu
->hrtimer_interval_ms
= PERF_CPU_HRTIMER
;
832 cpuctx
->hrtimer_interval
= ns_to_ktime(NSEC_PER_MSEC
* timer
);
834 hrtimer_init(hr
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL_PINNED
);
835 hr
->function
= perf_cpu_hrtimer_handler
;
838 static void perf_cpu_hrtimer_restart(struct perf_cpu_context
*cpuctx
)
840 struct hrtimer
*hr
= &cpuctx
->hrtimer
;
841 struct pmu
*pmu
= cpuctx
->ctx
.pmu
;
844 if (pmu
->task_ctx_nr
== perf_sw_context
)
847 if (hrtimer_active(hr
))
850 if (!hrtimer_callback_running(hr
))
851 __hrtimer_start_range_ns(hr
, cpuctx
->hrtimer_interval
,
852 0, HRTIMER_MODE_REL_PINNED
, 0);
855 void perf_pmu_disable(struct pmu
*pmu
)
857 int *count
= this_cpu_ptr(pmu
->pmu_disable_count
);
859 pmu
->pmu_disable(pmu
);
862 void perf_pmu_enable(struct pmu
*pmu
)
864 int *count
= this_cpu_ptr(pmu
->pmu_disable_count
);
866 pmu
->pmu_enable(pmu
);
869 static DEFINE_PER_CPU(struct list_head
, rotation_list
);
872 * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
873 * because they're strictly cpu affine and rotate_start is called with IRQs
874 * disabled, while rotate_context is called from IRQ context.
876 static void perf_pmu_rotate_start(struct pmu
*pmu
)
878 struct perf_cpu_context
*cpuctx
= this_cpu_ptr(pmu
->pmu_cpu_context
);
879 struct list_head
*head
= &__get_cpu_var(rotation_list
);
881 WARN_ON(!irqs_disabled());
883 if (list_empty(&cpuctx
->rotation_list
))
884 list_add(&cpuctx
->rotation_list
, head
);
887 static void get_ctx(struct perf_event_context
*ctx
)
889 WARN_ON(!atomic_inc_not_zero(&ctx
->refcount
));
892 static void put_ctx(struct perf_event_context
*ctx
)
894 if (atomic_dec_and_test(&ctx
->refcount
)) {
896 put_ctx(ctx
->parent_ctx
);
898 put_task_struct(ctx
->task
);
899 kfree_rcu(ctx
, rcu_head
);
903 static void unclone_ctx(struct perf_event_context
*ctx
)
905 if (ctx
->parent_ctx
) {
906 put_ctx(ctx
->parent_ctx
);
907 ctx
->parent_ctx
= NULL
;
912 static u32
perf_event_pid(struct perf_event
*event
, struct task_struct
*p
)
915 * only top level events have the pid namespace they were created in
918 event
= event
->parent
;
920 return task_tgid_nr_ns(p
, event
->ns
);
923 static u32
perf_event_tid(struct perf_event
*event
, struct task_struct
*p
)
926 * only top level events have the pid namespace they were created in
929 event
= event
->parent
;
931 return task_pid_nr_ns(p
, event
->ns
);
935 * If we inherit events we want to return the parent event id
938 static u64
primary_event_id(struct perf_event
*event
)
943 id
= event
->parent
->id
;
949 * Get the perf_event_context for a task and lock it.
950 * This has to cope with with the fact that until it is locked,
951 * the context could get moved to another task.
953 static struct perf_event_context
*
954 perf_lock_task_context(struct task_struct
*task
, int ctxn
, unsigned long *flags
)
956 struct perf_event_context
*ctx
;
960 * One of the few rules of preemptible RCU is that one cannot do
961 * rcu_read_unlock() while holding a scheduler (or nested) lock when
962 * part of the read side critical section was preemptible -- see
963 * rcu_read_unlock_special().
965 * Since ctx->lock nests under rq->lock we must ensure the entire read
966 * side critical section is non-preemptible.
970 ctx
= rcu_dereference(task
->perf_event_ctxp
[ctxn
]);
973 * If this context is a clone of another, it might
974 * get swapped for another underneath us by
975 * perf_event_task_sched_out, though the
976 * rcu_read_lock() protects us from any context
977 * getting freed. Lock the context and check if it
978 * got swapped before we could get the lock, and retry
979 * if so. If we locked the right context, then it
980 * can't get swapped on us any more.
982 raw_spin_lock_irqsave(&ctx
->lock
, *flags
);
983 if (ctx
!= rcu_dereference(task
->perf_event_ctxp
[ctxn
])) {
984 raw_spin_unlock_irqrestore(&ctx
->lock
, *flags
);
990 if (!atomic_inc_not_zero(&ctx
->refcount
)) {
991 raw_spin_unlock_irqrestore(&ctx
->lock
, *flags
);
1001 * Get the context for a task and increment its pin_count so it
1002 * can't get swapped to another task. This also increments its
1003 * reference count so that the context can't get freed.
1005 static struct perf_event_context
*
1006 perf_pin_task_context(struct task_struct
*task
, int ctxn
)
1008 struct perf_event_context
*ctx
;
1009 unsigned long flags
;
1011 ctx
= perf_lock_task_context(task
, ctxn
, &flags
);
1014 raw_spin_unlock_irqrestore(&ctx
->lock
, flags
);
1019 static void perf_unpin_context(struct perf_event_context
*ctx
)
1021 unsigned long flags
;
1023 raw_spin_lock_irqsave(&ctx
->lock
, flags
);
1025 raw_spin_unlock_irqrestore(&ctx
->lock
, flags
);
1029 * Update the record of the current time in a context.
1031 static void update_context_time(struct perf_event_context
*ctx
)
1033 u64 now
= perf_clock();
1035 ctx
->time
+= now
- ctx
->timestamp
;
1036 ctx
->timestamp
= now
;
1039 static u64
perf_event_time(struct perf_event
*event
)
1041 struct perf_event_context
*ctx
= event
->ctx
;
1043 if (is_cgroup_event(event
))
1044 return perf_cgroup_event_time(event
);
1046 return ctx
? ctx
->time
: 0;
1050 * Update the total_time_enabled and total_time_running fields for a event.
1051 * The caller of this function needs to hold the ctx->lock.
1053 static void update_event_times(struct perf_event
*event
)
1055 struct perf_event_context
*ctx
= event
->ctx
;
1058 if (event
->state
< PERF_EVENT_STATE_INACTIVE
||
1059 event
->group_leader
->state
< PERF_EVENT_STATE_INACTIVE
)
1062 * in cgroup mode, time_enabled represents
1063 * the time the event was enabled AND active
1064 * tasks were in the monitored cgroup. This is
1065 * independent of the activity of the context as
1066 * there may be a mix of cgroup and non-cgroup events.
1068 * That is why we treat cgroup events differently
1071 if (is_cgroup_event(event
))
1072 run_end
= perf_cgroup_event_time(event
);
1073 else if (ctx
->is_active
)
1074 run_end
= ctx
->time
;
1076 run_end
= event
->tstamp_stopped
;
1078 event
->total_time_enabled
= run_end
- event
->tstamp_enabled
;
1080 if (event
->state
== PERF_EVENT_STATE_INACTIVE
)
1081 run_end
= event
->tstamp_stopped
;
1083 run_end
= perf_event_time(event
);
1085 event
->total_time_running
= run_end
- event
->tstamp_running
;
1090 * Update total_time_enabled and total_time_running for all events in a group.
1092 static void update_group_times(struct perf_event
*leader
)
1094 struct perf_event
*event
;
1096 update_event_times(leader
);
1097 list_for_each_entry(event
, &leader
->sibling_list
, group_entry
)
1098 update_event_times(event
);
1101 static struct list_head
*
1102 ctx_group_list(struct perf_event
*event
, struct perf_event_context
*ctx
)
1104 if (event
->attr
.pinned
)
1105 return &ctx
->pinned_groups
;
1107 return &ctx
->flexible_groups
;
1111 * Add a event from the lists for its context.
1112 * Must be called with ctx->mutex and ctx->lock held.
1115 list_add_event(struct perf_event
*event
, struct perf_event_context
*ctx
)
1117 WARN_ON_ONCE(event
->attach_state
& PERF_ATTACH_CONTEXT
);
1118 event
->attach_state
|= PERF_ATTACH_CONTEXT
;
1121 * If we're a stand alone event or group leader, we go to the context
1122 * list, group events are kept attached to the group so that
1123 * perf_group_detach can, at all times, locate all siblings.
1125 if (event
->group_leader
== event
) {
1126 struct list_head
*list
;
1128 if (is_software_event(event
))
1129 event
->group_flags
|= PERF_GROUP_SOFTWARE
;
1131 list
= ctx_group_list(event
, ctx
);
1132 list_add_tail(&event
->group_entry
, list
);
1135 if (is_cgroup_event(event
))
1138 if (has_branch_stack(event
))
1139 ctx
->nr_branch_stack
++;
1141 list_add_rcu(&event
->event_entry
, &ctx
->event_list
);
1142 if (!ctx
->nr_events
)
1143 perf_pmu_rotate_start(ctx
->pmu
);
1145 if (event
->attr
.inherit_stat
)
1152 * Initialize event state based on the perf_event_attr::disabled.
1154 static inline void perf_event__state_init(struct perf_event
*event
)
1156 event
->state
= event
->attr
.disabled
? PERF_EVENT_STATE_OFF
:
1157 PERF_EVENT_STATE_INACTIVE
;
1161 * Called at perf_event creation and when events are attached/detached from a
1164 static void perf_event__read_size(struct perf_event
*event
)
1166 int entry
= sizeof(u64
); /* value */
1170 if (event
->attr
.read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
1171 size
+= sizeof(u64
);
1173 if (event
->attr
.read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
1174 size
+= sizeof(u64
);
1176 if (event
->attr
.read_format
& PERF_FORMAT_ID
)
1177 entry
+= sizeof(u64
);
1179 if (event
->attr
.read_format
& PERF_FORMAT_GROUP
) {
1180 nr
+= event
->group_leader
->nr_siblings
;
1181 size
+= sizeof(u64
);
1185 event
->read_size
= size
;
1188 static void perf_event__header_size(struct perf_event
*event
)
1190 struct perf_sample_data
*data
;
1191 u64 sample_type
= event
->attr
.sample_type
;
1194 perf_event__read_size(event
);
1196 if (sample_type
& PERF_SAMPLE_IP
)
1197 size
+= sizeof(data
->ip
);
1199 if (sample_type
& PERF_SAMPLE_ADDR
)
1200 size
+= sizeof(data
->addr
);
1202 if (sample_type
& PERF_SAMPLE_PERIOD
)
1203 size
+= sizeof(data
->period
);
1205 if (sample_type
& PERF_SAMPLE_WEIGHT
)
1206 size
+= sizeof(data
->weight
);
1208 if (sample_type
& PERF_SAMPLE_READ
)
1209 size
+= event
->read_size
;
1211 if (sample_type
& PERF_SAMPLE_DATA_SRC
)
1212 size
+= sizeof(data
->data_src
.val
);
1214 if (sample_type
& PERF_SAMPLE_TRANSACTION
)
1215 size
+= sizeof(data
->txn
);
1217 event
->header_size
= size
;
1220 static void perf_event__id_header_size(struct perf_event
*event
)
1222 struct perf_sample_data
*data
;
1223 u64 sample_type
= event
->attr
.sample_type
;
1226 if (sample_type
& PERF_SAMPLE_TID
)
1227 size
+= sizeof(data
->tid_entry
);
1229 if (sample_type
& PERF_SAMPLE_TIME
)
1230 size
+= sizeof(data
->time
);
1232 if (sample_type
& PERF_SAMPLE_IDENTIFIER
)
1233 size
+= sizeof(data
->id
);
1235 if (sample_type
& PERF_SAMPLE_ID
)
1236 size
+= sizeof(data
->id
);
1238 if (sample_type
& PERF_SAMPLE_STREAM_ID
)
1239 size
+= sizeof(data
->stream_id
);
1241 if (sample_type
& PERF_SAMPLE_CPU
)
1242 size
+= sizeof(data
->cpu_entry
);
1244 event
->id_header_size
= size
;
1247 static void perf_group_attach(struct perf_event
*event
)
1249 struct perf_event
*group_leader
= event
->group_leader
, *pos
;
1252 * We can have double attach due to group movement in perf_event_open.
1254 if (event
->attach_state
& PERF_ATTACH_GROUP
)
1257 event
->attach_state
|= PERF_ATTACH_GROUP
;
1259 if (group_leader
== event
)
1262 if (group_leader
->group_flags
& PERF_GROUP_SOFTWARE
&&
1263 !is_software_event(event
))
1264 group_leader
->group_flags
&= ~PERF_GROUP_SOFTWARE
;
1266 list_add_tail(&event
->group_entry
, &group_leader
->sibling_list
);
1267 group_leader
->nr_siblings
++;
1269 perf_event__header_size(group_leader
);
1271 list_for_each_entry(pos
, &group_leader
->sibling_list
, group_entry
)
1272 perf_event__header_size(pos
);
1276 * Remove a event from the lists for its context.
1277 * Must be called with ctx->mutex and ctx->lock held.
1280 list_del_event(struct perf_event
*event
, struct perf_event_context
*ctx
)
1282 struct perf_cpu_context
*cpuctx
;
1284 * We can have double detach due to exit/hot-unplug + close.
1286 if (!(event
->attach_state
& PERF_ATTACH_CONTEXT
))
1289 event
->attach_state
&= ~PERF_ATTACH_CONTEXT
;
1291 if (is_cgroup_event(event
)) {
1293 cpuctx
= __get_cpu_context(ctx
);
1295 * if there are no more cgroup events
1296 * then cler cgrp to avoid stale pointer
1297 * in update_cgrp_time_from_cpuctx()
1299 if (!ctx
->nr_cgroups
)
1300 cpuctx
->cgrp
= NULL
;
1303 if (has_branch_stack(event
))
1304 ctx
->nr_branch_stack
--;
1307 if (event
->attr
.inherit_stat
)
1310 list_del_rcu(&event
->event_entry
);
1312 if (event
->group_leader
== event
)
1313 list_del_init(&event
->group_entry
);
1315 update_group_times(event
);
1318 * If event was in error state, then keep it
1319 * that way, otherwise bogus counts will be
1320 * returned on read(). The only way to get out
1321 * of error state is by explicit re-enabling
1324 if (event
->state
> PERF_EVENT_STATE_OFF
)
1325 event
->state
= PERF_EVENT_STATE_OFF
;
1330 static void perf_group_detach(struct perf_event
*event
)
1332 struct perf_event
*sibling
, *tmp
;
1333 struct list_head
*list
= NULL
;
1336 * We can have double detach due to exit/hot-unplug + close.
1338 if (!(event
->attach_state
& PERF_ATTACH_GROUP
))
1341 event
->attach_state
&= ~PERF_ATTACH_GROUP
;
1344 * If this is a sibling, remove it from its group.
1346 if (event
->group_leader
!= event
) {
1347 list_del_init(&event
->group_entry
);
1348 event
->group_leader
->nr_siblings
--;
1352 if (!list_empty(&event
->group_entry
))
1353 list
= &event
->group_entry
;
1356 * If this was a group event with sibling events then
1357 * upgrade the siblings to singleton events by adding them
1358 * to whatever list we are on.
1360 list_for_each_entry_safe(sibling
, tmp
, &event
->sibling_list
, group_entry
) {
1362 list_move_tail(&sibling
->group_entry
, list
);
1363 sibling
->group_leader
= sibling
;
1365 /* Inherit group flags from the previous leader */
1366 sibling
->group_flags
= event
->group_flags
;
1370 perf_event__header_size(event
->group_leader
);
1372 list_for_each_entry(tmp
, &event
->group_leader
->sibling_list
, group_entry
)
1373 perf_event__header_size(tmp
);
1377 event_filter_match(struct perf_event
*event
)
1379 return (event
->cpu
== -1 || event
->cpu
== smp_processor_id())
1380 && perf_cgroup_match(event
);
1384 event_sched_out(struct perf_event
*event
,
1385 struct perf_cpu_context
*cpuctx
,
1386 struct perf_event_context
*ctx
)
1388 u64 tstamp
= perf_event_time(event
);
1391 * An event which could not be activated because of
1392 * filter mismatch still needs to have its timings
1393 * maintained, otherwise bogus information is return
1394 * via read() for time_enabled, time_running:
1396 if (event
->state
== PERF_EVENT_STATE_INACTIVE
1397 && !event_filter_match(event
)) {
1398 delta
= tstamp
- event
->tstamp_stopped
;
1399 event
->tstamp_running
+= delta
;
1400 event
->tstamp_stopped
= tstamp
;
1403 if (event
->state
!= PERF_EVENT_STATE_ACTIVE
)
1406 perf_pmu_disable(event
->pmu
);
1408 event
->state
= PERF_EVENT_STATE_INACTIVE
;
1409 if (event
->pending_disable
) {
1410 event
->pending_disable
= 0;
1411 event
->state
= PERF_EVENT_STATE_OFF
;
1413 event
->tstamp_stopped
= tstamp
;
1414 event
->pmu
->del(event
, 0);
1417 if (!is_software_event(event
))
1418 cpuctx
->active_oncpu
--;
1420 if (event
->attr
.freq
&& event
->attr
.sample_freq
)
1422 if (event
->attr
.exclusive
|| !cpuctx
->active_oncpu
)
1423 cpuctx
->exclusive
= 0;
1425 perf_pmu_enable(event
->pmu
);
1429 group_sched_out(struct perf_event
*group_event
,
1430 struct perf_cpu_context
*cpuctx
,
1431 struct perf_event_context
*ctx
)
1433 struct perf_event
*event
;
1434 int state
= group_event
->state
;
1436 event_sched_out(group_event
, cpuctx
, ctx
);
1439 * Schedule out siblings (if any):
1441 list_for_each_entry(event
, &group_event
->sibling_list
, group_entry
)
1442 event_sched_out(event
, cpuctx
, ctx
);
1444 if (state
== PERF_EVENT_STATE_ACTIVE
&& group_event
->attr
.exclusive
)
1445 cpuctx
->exclusive
= 0;
1448 struct remove_event
{
1449 struct perf_event
*event
;
1454 * Cross CPU call to remove a performance event
1456 * We disable the event on the hardware level first. After that we
1457 * remove it from the context list.
1459 static int __perf_remove_from_context(void *info
)
1461 struct remove_event
*re
= info
;
1462 struct perf_event
*event
= re
->event
;
1463 struct perf_event_context
*ctx
= event
->ctx
;
1464 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
1466 raw_spin_lock(&ctx
->lock
);
1467 event_sched_out(event
, cpuctx
, ctx
);
1468 if (re
->detach_group
)
1469 perf_group_detach(event
);
1470 list_del_event(event
, ctx
);
1471 if (!ctx
->nr_events
&& cpuctx
->task_ctx
== ctx
) {
1473 cpuctx
->task_ctx
= NULL
;
1475 raw_spin_unlock(&ctx
->lock
);
1482 * Remove the event from a task's (or a CPU's) list of events.
1484 * CPU events are removed with a smp call. For task events we only
1485 * call when the task is on a CPU.
1487 * If event->ctx is a cloned context, callers must make sure that
1488 * every task struct that event->ctx->task could possibly point to
1489 * remains valid. This is OK when called from perf_release since
1490 * that only calls us on the top-level context, which can't be a clone.
1491 * When called from perf_event_exit_task, it's OK because the
1492 * context has been detached from its task.
1494 static void perf_remove_from_context(struct perf_event
*event
, bool detach_group
)
1496 struct perf_event_context
*ctx
= event
->ctx
;
1497 struct task_struct
*task
= ctx
->task
;
1498 struct remove_event re
= {
1500 .detach_group
= detach_group
,
1503 lockdep_assert_held(&ctx
->mutex
);
1507 * Per cpu events are removed via an smp call and
1508 * the removal is always successful.
1510 cpu_function_call(event
->cpu
, __perf_remove_from_context
, &re
);
1515 if (!task_function_call(task
, __perf_remove_from_context
, &re
))
1518 raw_spin_lock_irq(&ctx
->lock
);
1520 * If we failed to find a running task, but find the context active now
1521 * that we've acquired the ctx->lock, retry.
1523 if (ctx
->is_active
) {
1524 raw_spin_unlock_irq(&ctx
->lock
);
1529 * Since the task isn't running, its safe to remove the event, us
1530 * holding the ctx->lock ensures the task won't get scheduled in.
1533 perf_group_detach(event
);
1534 list_del_event(event
, ctx
);
1535 raw_spin_unlock_irq(&ctx
->lock
);
1539 * Cross CPU call to disable a performance event
1541 int __perf_event_disable(void *info
)
1543 struct perf_event
*event
= info
;
1544 struct perf_event_context
*ctx
= event
->ctx
;
1545 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
1548 * If this is a per-task event, need to check whether this
1549 * event's task is the current task on this cpu.
1551 * Can trigger due to concurrent perf_event_context_sched_out()
1552 * flipping contexts around.
1554 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
)
1557 raw_spin_lock(&ctx
->lock
);
1560 * If the event is on, turn it off.
1561 * If it is in error state, leave it in error state.
1563 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
) {
1564 update_context_time(ctx
);
1565 update_cgrp_time_from_event(event
);
1566 update_group_times(event
);
1567 if (event
== event
->group_leader
)
1568 group_sched_out(event
, cpuctx
, ctx
);
1570 event_sched_out(event
, cpuctx
, ctx
);
1571 event
->state
= PERF_EVENT_STATE_OFF
;
1574 raw_spin_unlock(&ctx
->lock
);
1582 * If event->ctx is a cloned context, callers must make sure that
1583 * every task struct that event->ctx->task could possibly point to
1584 * remains valid. This condition is satisifed when called through
1585 * perf_event_for_each_child or perf_event_for_each because they
1586 * hold the top-level event's child_mutex, so any descendant that
1587 * goes to exit will block in sync_child_event.
1588 * When called from perf_pending_event it's OK because event->ctx
1589 * is the current context on this CPU and preemption is disabled,
1590 * hence we can't get into perf_event_task_sched_out for this context.
1592 void perf_event_disable(struct perf_event
*event
)
1594 struct perf_event_context
*ctx
= event
->ctx
;
1595 struct task_struct
*task
= ctx
->task
;
1599 * Disable the event on the cpu that it's on
1601 cpu_function_call(event
->cpu
, __perf_event_disable
, event
);
1606 if (!task_function_call(task
, __perf_event_disable
, event
))
1609 raw_spin_lock_irq(&ctx
->lock
);
1611 * If the event is still active, we need to retry the cross-call.
1613 if (event
->state
== PERF_EVENT_STATE_ACTIVE
) {
1614 raw_spin_unlock_irq(&ctx
->lock
);
1616 * Reload the task pointer, it might have been changed by
1617 * a concurrent perf_event_context_sched_out().
1624 * Since we have the lock this context can't be scheduled
1625 * in, so we can change the state safely.
1627 if (event
->state
== PERF_EVENT_STATE_INACTIVE
) {
1628 update_group_times(event
);
1629 event
->state
= PERF_EVENT_STATE_OFF
;
1631 raw_spin_unlock_irq(&ctx
->lock
);
1633 EXPORT_SYMBOL_GPL(perf_event_disable
);
1635 static void perf_set_shadow_time(struct perf_event
*event
,
1636 struct perf_event_context
*ctx
,
1640 * use the correct time source for the time snapshot
1642 * We could get by without this by leveraging the
1643 * fact that to get to this function, the caller
1644 * has most likely already called update_context_time()
1645 * and update_cgrp_time_xx() and thus both timestamp
1646 * are identical (or very close). Given that tstamp is,
1647 * already adjusted for cgroup, we could say that:
1648 * tstamp - ctx->timestamp
1650 * tstamp - cgrp->timestamp.
1652 * Then, in perf_output_read(), the calculation would
1653 * work with no changes because:
1654 * - event is guaranteed scheduled in
1655 * - no scheduled out in between
1656 * - thus the timestamp would be the same
1658 * But this is a bit hairy.
1660 * So instead, we have an explicit cgroup call to remain
1661 * within the time time source all along. We believe it
1662 * is cleaner and simpler to understand.
1664 if (is_cgroup_event(event
))
1665 perf_cgroup_set_shadow_time(event
, tstamp
);
1667 event
->shadow_ctx_time
= tstamp
- ctx
->timestamp
;
1670 #define MAX_INTERRUPTS (~0ULL)
1672 static void perf_log_throttle(struct perf_event
*event
, int enable
);
1675 event_sched_in(struct perf_event
*event
,
1676 struct perf_cpu_context
*cpuctx
,
1677 struct perf_event_context
*ctx
)
1679 u64 tstamp
= perf_event_time(event
);
1682 lockdep_assert_held(&ctx
->lock
);
1684 if (event
->state
<= PERF_EVENT_STATE_OFF
)
1687 event
->state
= PERF_EVENT_STATE_ACTIVE
;
1688 event
->oncpu
= smp_processor_id();
1691 * Unthrottle events, since we scheduled we might have missed several
1692 * ticks already, also for a heavily scheduling task there is little
1693 * guarantee it'll get a tick in a timely manner.
1695 if (unlikely(event
->hw
.interrupts
== MAX_INTERRUPTS
)) {
1696 perf_log_throttle(event
, 1);
1697 event
->hw
.interrupts
= 0;
1701 * The new state must be visible before we turn it on in the hardware:
1705 perf_pmu_disable(event
->pmu
);
1707 if (event
->pmu
->add(event
, PERF_EF_START
)) {
1708 event
->state
= PERF_EVENT_STATE_INACTIVE
;
1714 event
->tstamp_running
+= tstamp
- event
->tstamp_stopped
;
1716 perf_set_shadow_time(event
, ctx
, tstamp
);
1718 if (!is_software_event(event
))
1719 cpuctx
->active_oncpu
++;
1721 if (event
->attr
.freq
&& event
->attr
.sample_freq
)
1724 if (event
->attr
.exclusive
)
1725 cpuctx
->exclusive
= 1;
1728 perf_pmu_enable(event
->pmu
);
1734 group_sched_in(struct perf_event
*group_event
,
1735 struct perf_cpu_context
*cpuctx
,
1736 struct perf_event_context
*ctx
)
1738 struct perf_event
*event
, *partial_group
= NULL
;
1739 struct pmu
*pmu
= ctx
->pmu
;
1740 u64 now
= ctx
->time
;
1741 bool simulate
= false;
1743 if (group_event
->state
== PERF_EVENT_STATE_OFF
)
1746 pmu
->start_txn(pmu
);
1748 if (event_sched_in(group_event
, cpuctx
, ctx
)) {
1749 pmu
->cancel_txn(pmu
);
1750 perf_cpu_hrtimer_restart(cpuctx
);
1755 * Schedule in siblings as one group (if any):
1757 list_for_each_entry(event
, &group_event
->sibling_list
, group_entry
) {
1758 if (event_sched_in(event
, cpuctx
, ctx
)) {
1759 partial_group
= event
;
1764 if (!pmu
->commit_txn(pmu
))
1769 * Groups can be scheduled in as one unit only, so undo any
1770 * partial group before returning:
1771 * The events up to the failed event are scheduled out normally,
1772 * tstamp_stopped will be updated.
1774 * The failed events and the remaining siblings need to have
1775 * their timings updated as if they had gone thru event_sched_in()
1776 * and event_sched_out(). This is required to get consistent timings
1777 * across the group. This also takes care of the case where the group
1778 * could never be scheduled by ensuring tstamp_stopped is set to mark
1779 * the time the event was actually stopped, such that time delta
1780 * calculation in update_event_times() is correct.
1782 list_for_each_entry(event
, &group_event
->sibling_list
, group_entry
) {
1783 if (event
== partial_group
)
1787 event
->tstamp_running
+= now
- event
->tstamp_stopped
;
1788 event
->tstamp_stopped
= now
;
1790 event_sched_out(event
, cpuctx
, ctx
);
1793 event_sched_out(group_event
, cpuctx
, ctx
);
1795 pmu
->cancel_txn(pmu
);
1797 perf_cpu_hrtimer_restart(cpuctx
);
1803 * Work out whether we can put this event group on the CPU now.
1805 static int group_can_go_on(struct perf_event
*event
,
1806 struct perf_cpu_context
*cpuctx
,
1810 * Groups consisting entirely of software events can always go on.
1812 if (event
->group_flags
& PERF_GROUP_SOFTWARE
)
1815 * If an exclusive group is already on, no other hardware
1818 if (cpuctx
->exclusive
)
1821 * If this group is exclusive and there are already
1822 * events on the CPU, it can't go on.
1824 if (event
->attr
.exclusive
&& cpuctx
->active_oncpu
)
1827 * Otherwise, try to add it if all previous groups were able
1833 static void add_event_to_ctx(struct perf_event
*event
,
1834 struct perf_event_context
*ctx
)
1836 u64 tstamp
= perf_event_time(event
);
1838 list_add_event(event
, ctx
);
1839 perf_group_attach(event
);
1840 event
->tstamp_enabled
= tstamp
;
1841 event
->tstamp_running
= tstamp
;
1842 event
->tstamp_stopped
= tstamp
;
1845 static void task_ctx_sched_out(struct perf_event_context
*ctx
);
1847 ctx_sched_in(struct perf_event_context
*ctx
,
1848 struct perf_cpu_context
*cpuctx
,
1849 enum event_type_t event_type
,
1850 struct task_struct
*task
);
1852 static void perf_event_sched_in(struct perf_cpu_context
*cpuctx
,
1853 struct perf_event_context
*ctx
,
1854 struct task_struct
*task
)
1856 cpu_ctx_sched_in(cpuctx
, EVENT_PINNED
, task
);
1858 ctx_sched_in(ctx
, cpuctx
, EVENT_PINNED
, task
);
1859 cpu_ctx_sched_in(cpuctx
, EVENT_FLEXIBLE
, task
);
1861 ctx_sched_in(ctx
, cpuctx
, EVENT_FLEXIBLE
, task
);
1865 * Cross CPU call to install and enable a performance event
1867 * Must be called with ctx->mutex held
1869 static int __perf_install_in_context(void *info
)
1871 struct perf_event
*event
= info
;
1872 struct perf_event_context
*ctx
= event
->ctx
;
1873 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
1874 struct perf_event_context
*task_ctx
= cpuctx
->task_ctx
;
1875 struct task_struct
*task
= current
;
1877 perf_ctx_lock(cpuctx
, task_ctx
);
1878 perf_pmu_disable(cpuctx
->ctx
.pmu
);
1881 * If there was an active task_ctx schedule it out.
1884 task_ctx_sched_out(task_ctx
);
1887 * If the context we're installing events in is not the
1888 * active task_ctx, flip them.
1890 if (ctx
->task
&& task_ctx
!= ctx
) {
1892 raw_spin_unlock(&task_ctx
->lock
);
1893 raw_spin_lock(&ctx
->lock
);
1898 cpuctx
->task_ctx
= task_ctx
;
1899 task
= task_ctx
->task
;
1902 cpu_ctx_sched_out(cpuctx
, EVENT_ALL
);
1904 update_context_time(ctx
);
1906 * update cgrp time only if current cgrp
1907 * matches event->cgrp. Must be done before
1908 * calling add_event_to_ctx()
1910 update_cgrp_time_from_event(event
);
1912 add_event_to_ctx(event
, ctx
);
1915 * Schedule everything back in
1917 perf_event_sched_in(cpuctx
, task_ctx
, task
);
1919 perf_pmu_enable(cpuctx
->ctx
.pmu
);
1920 perf_ctx_unlock(cpuctx
, task_ctx
);
1926 * Attach a performance event to a context
1928 * First we add the event to the list with the hardware enable bit
1929 * in event->hw_config cleared.
1931 * If the event is attached to a task which is on a CPU we use a smp
1932 * call to enable it in the task context. The task might have been
1933 * scheduled away, but we check this in the smp call again.
1936 perf_install_in_context(struct perf_event_context
*ctx
,
1937 struct perf_event
*event
,
1940 struct task_struct
*task
= ctx
->task
;
1942 lockdep_assert_held(&ctx
->mutex
);
1945 if (event
->cpu
!= -1)
1950 * Per cpu events are installed via an smp call and
1951 * the install is always successful.
1953 cpu_function_call(cpu
, __perf_install_in_context
, event
);
1958 if (!task_function_call(task
, __perf_install_in_context
, event
))
1961 raw_spin_lock_irq(&ctx
->lock
);
1963 * If we failed to find a running task, but find the context active now
1964 * that we've acquired the ctx->lock, retry.
1966 if (ctx
->is_active
) {
1967 raw_spin_unlock_irq(&ctx
->lock
);
1972 * Since the task isn't running, its safe to add the event, us holding
1973 * the ctx->lock ensures the task won't get scheduled in.
1975 add_event_to_ctx(event
, ctx
);
1976 raw_spin_unlock_irq(&ctx
->lock
);
1980 * Put a event into inactive state and update time fields.
1981 * Enabling the leader of a group effectively enables all
1982 * the group members that aren't explicitly disabled, so we
1983 * have to update their ->tstamp_enabled also.
1984 * Note: this works for group members as well as group leaders
1985 * since the non-leader members' sibling_lists will be empty.
1987 static void __perf_event_mark_enabled(struct perf_event
*event
)
1989 struct perf_event
*sub
;
1990 u64 tstamp
= perf_event_time(event
);
1992 event
->state
= PERF_EVENT_STATE_INACTIVE
;
1993 event
->tstamp_enabled
= tstamp
- event
->total_time_enabled
;
1994 list_for_each_entry(sub
, &event
->sibling_list
, group_entry
) {
1995 if (sub
->state
>= PERF_EVENT_STATE_INACTIVE
)
1996 sub
->tstamp_enabled
= tstamp
- sub
->total_time_enabled
;
2001 * Cross CPU call to enable a performance event
2003 static int __perf_event_enable(void *info
)
2005 struct perf_event
*event
= info
;
2006 struct perf_event_context
*ctx
= event
->ctx
;
2007 struct perf_event
*leader
= event
->group_leader
;
2008 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
2012 * There's a time window between 'ctx->is_active' check
2013 * in perf_event_enable function and this place having:
2015 * - ctx->lock unlocked
2017 * where the task could be killed and 'ctx' deactivated
2018 * by perf_event_exit_task.
2020 if (!ctx
->is_active
)
2023 raw_spin_lock(&ctx
->lock
);
2024 update_context_time(ctx
);
2026 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
)
2030 * set current task's cgroup time reference point
2032 perf_cgroup_set_timestamp(current
, ctx
);
2034 __perf_event_mark_enabled(event
);
2036 if (!event_filter_match(event
)) {
2037 if (is_cgroup_event(event
))
2038 perf_cgroup_defer_enabled(event
);
2043 * If the event is in a group and isn't the group leader,
2044 * then don't put it on unless the group is on.
2046 if (leader
!= event
&& leader
->state
!= PERF_EVENT_STATE_ACTIVE
)
2049 if (!group_can_go_on(event
, cpuctx
, 1)) {
2052 if (event
== leader
)
2053 err
= group_sched_in(event
, cpuctx
, ctx
);
2055 err
= event_sched_in(event
, cpuctx
, ctx
);
2060 * If this event can't go on and it's part of a
2061 * group, then the whole group has to come off.
2063 if (leader
!= event
) {
2064 group_sched_out(leader
, cpuctx
, ctx
);
2065 perf_cpu_hrtimer_restart(cpuctx
);
2067 if (leader
->attr
.pinned
) {
2068 update_group_times(leader
);
2069 leader
->state
= PERF_EVENT_STATE_ERROR
;
2074 raw_spin_unlock(&ctx
->lock
);
2082 * If event->ctx is a cloned context, callers must make sure that
2083 * every task struct that event->ctx->task could possibly point to
2084 * remains valid. This condition is satisfied when called through
2085 * perf_event_for_each_child or perf_event_for_each as described
2086 * for perf_event_disable.
2088 void perf_event_enable(struct perf_event
*event
)
2090 struct perf_event_context
*ctx
= event
->ctx
;
2091 struct task_struct
*task
= ctx
->task
;
2095 * Enable the event on the cpu that it's on
2097 cpu_function_call(event
->cpu
, __perf_event_enable
, event
);
2101 raw_spin_lock_irq(&ctx
->lock
);
2102 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
)
2106 * If the event is in error state, clear that first.
2107 * That way, if we see the event in error state below, we
2108 * know that it has gone back into error state, as distinct
2109 * from the task having been scheduled away before the
2110 * cross-call arrived.
2112 if (event
->state
== PERF_EVENT_STATE_ERROR
)
2113 event
->state
= PERF_EVENT_STATE_OFF
;
2116 if (!ctx
->is_active
) {
2117 __perf_event_mark_enabled(event
);
2121 raw_spin_unlock_irq(&ctx
->lock
);
2123 if (!task_function_call(task
, __perf_event_enable
, event
))
2126 raw_spin_lock_irq(&ctx
->lock
);
2129 * If the context is active and the event is still off,
2130 * we need to retry the cross-call.
2132 if (ctx
->is_active
&& event
->state
== PERF_EVENT_STATE_OFF
) {
2134 * task could have been flipped by a concurrent
2135 * perf_event_context_sched_out()
2142 raw_spin_unlock_irq(&ctx
->lock
);
2144 EXPORT_SYMBOL_GPL(perf_event_enable
);
2146 int perf_event_refresh(struct perf_event
*event
, int refresh
)
2149 * not supported on inherited events
2151 if (event
->attr
.inherit
|| !is_sampling_event(event
))
2154 atomic_add(refresh
, &event
->event_limit
);
2155 perf_event_enable(event
);
2159 EXPORT_SYMBOL_GPL(perf_event_refresh
);
2161 static void ctx_sched_out(struct perf_event_context
*ctx
,
2162 struct perf_cpu_context
*cpuctx
,
2163 enum event_type_t event_type
)
2165 struct perf_event
*event
;
2166 int is_active
= ctx
->is_active
;
2168 ctx
->is_active
&= ~event_type
;
2169 if (likely(!ctx
->nr_events
))
2172 update_context_time(ctx
);
2173 update_cgrp_time_from_cpuctx(cpuctx
);
2174 if (!ctx
->nr_active
)
2177 perf_pmu_disable(ctx
->pmu
);
2178 if ((is_active
& EVENT_PINNED
) && (event_type
& EVENT_PINNED
)) {
2179 list_for_each_entry(event
, &ctx
->pinned_groups
, group_entry
)
2180 group_sched_out(event
, cpuctx
, ctx
);
2183 if ((is_active
& EVENT_FLEXIBLE
) && (event_type
& EVENT_FLEXIBLE
)) {
2184 list_for_each_entry(event
, &ctx
->flexible_groups
, group_entry
)
2185 group_sched_out(event
, cpuctx
, ctx
);
2187 perf_pmu_enable(ctx
->pmu
);
2191 * Test whether two contexts are equivalent, i.e. whether they have both been
2192 * cloned from the same version of the same context.
2194 * Equivalence is measured using a generation number in the context that is
2195 * incremented on each modification to it; see unclone_ctx(), list_add_event()
2196 * and list_del_event().
2198 static int context_equiv(struct perf_event_context
*ctx1
,
2199 struct perf_event_context
*ctx2
)
2201 /* Pinning disables the swap optimization */
2202 if (ctx1
->pin_count
|| ctx2
->pin_count
)
2205 /* If ctx1 is the parent of ctx2 */
2206 if (ctx1
== ctx2
->parent_ctx
&& ctx1
->generation
== ctx2
->parent_gen
)
2209 /* If ctx2 is the parent of ctx1 */
2210 if (ctx1
->parent_ctx
== ctx2
&& ctx1
->parent_gen
== ctx2
->generation
)
2214 * If ctx1 and ctx2 have the same parent; we flatten the parent
2215 * hierarchy, see perf_event_init_context().
2217 if (ctx1
->parent_ctx
&& ctx1
->parent_ctx
== ctx2
->parent_ctx
&&
2218 ctx1
->parent_gen
== ctx2
->parent_gen
)
2225 static void __perf_event_sync_stat(struct perf_event
*event
,
2226 struct perf_event
*next_event
)
2230 if (!event
->attr
.inherit_stat
)
2234 * Update the event value, we cannot use perf_event_read()
2235 * because we're in the middle of a context switch and have IRQs
2236 * disabled, which upsets smp_call_function_single(), however
2237 * we know the event must be on the current CPU, therefore we
2238 * don't need to use it.
2240 switch (event
->state
) {
2241 case PERF_EVENT_STATE_ACTIVE
:
2242 event
->pmu
->read(event
);
2245 case PERF_EVENT_STATE_INACTIVE
:
2246 update_event_times(event
);
2254 * In order to keep per-task stats reliable we need to flip the event
2255 * values when we flip the contexts.
2257 value
= local64_read(&next_event
->count
);
2258 value
= local64_xchg(&event
->count
, value
);
2259 local64_set(&next_event
->count
, value
);
2261 swap(event
->total_time_enabled
, next_event
->total_time_enabled
);
2262 swap(event
->total_time_running
, next_event
->total_time_running
);
2265 * Since we swizzled the values, update the user visible data too.
2267 perf_event_update_userpage(event
);
2268 perf_event_update_userpage(next_event
);
2271 static void perf_event_sync_stat(struct perf_event_context
*ctx
,
2272 struct perf_event_context
*next_ctx
)
2274 struct perf_event
*event
, *next_event
;
2279 update_context_time(ctx
);
2281 event
= list_first_entry(&ctx
->event_list
,
2282 struct perf_event
, event_entry
);
2284 next_event
= list_first_entry(&next_ctx
->event_list
,
2285 struct perf_event
, event_entry
);
2287 while (&event
->event_entry
!= &ctx
->event_list
&&
2288 &next_event
->event_entry
!= &next_ctx
->event_list
) {
2290 __perf_event_sync_stat(event
, next_event
);
2292 event
= list_next_entry(event
, event_entry
);
2293 next_event
= list_next_entry(next_event
, event_entry
);
2297 static void perf_event_context_sched_out(struct task_struct
*task
, int ctxn
,
2298 struct task_struct
*next
)
2300 struct perf_event_context
*ctx
= task
->perf_event_ctxp
[ctxn
];
2301 struct perf_event_context
*next_ctx
;
2302 struct perf_event_context
*parent
, *next_parent
;
2303 struct perf_cpu_context
*cpuctx
;
2309 cpuctx
= __get_cpu_context(ctx
);
2310 if (!cpuctx
->task_ctx
)
2314 next_ctx
= next
->perf_event_ctxp
[ctxn
];
2318 parent
= rcu_dereference(ctx
->parent_ctx
);
2319 next_parent
= rcu_dereference(next_ctx
->parent_ctx
);
2321 /* If neither context have a parent context; they cannot be clones. */
2322 if (!parent
&& !next_parent
)
2325 if (next_parent
== ctx
|| next_ctx
== parent
|| next_parent
== parent
) {
2327 * Looks like the two contexts are clones, so we might be
2328 * able to optimize the context switch. We lock both
2329 * contexts and check that they are clones under the
2330 * lock (including re-checking that neither has been
2331 * uncloned in the meantime). It doesn't matter which
2332 * order we take the locks because no other cpu could
2333 * be trying to lock both of these tasks.
2335 raw_spin_lock(&ctx
->lock
);
2336 raw_spin_lock_nested(&next_ctx
->lock
, SINGLE_DEPTH_NESTING
);
2337 if (context_equiv(ctx
, next_ctx
)) {
2339 * XXX do we need a memory barrier of sorts
2340 * wrt to rcu_dereference() of perf_event_ctxp
2342 task
->perf_event_ctxp
[ctxn
] = next_ctx
;
2343 next
->perf_event_ctxp
[ctxn
] = ctx
;
2345 next_ctx
->task
= task
;
2348 perf_event_sync_stat(ctx
, next_ctx
);
2350 raw_spin_unlock(&next_ctx
->lock
);
2351 raw_spin_unlock(&ctx
->lock
);
2357 raw_spin_lock(&ctx
->lock
);
2358 ctx_sched_out(ctx
, cpuctx
, EVENT_ALL
);
2359 cpuctx
->task_ctx
= NULL
;
2360 raw_spin_unlock(&ctx
->lock
);
2364 #define for_each_task_context_nr(ctxn) \
2365 for ((ctxn) = 0; (ctxn) < perf_nr_task_contexts; (ctxn)++)
2368 * Called from scheduler to remove the events of the current task,
2369 * with interrupts disabled.
2371 * We stop each event and update the event value in event->count.
2373 * This does not protect us against NMI, but disable()
2374 * sets the disabled bit in the control field of event _before_
2375 * accessing the event control register. If a NMI hits, then it will
2376 * not restart the event.
2378 void __perf_event_task_sched_out(struct task_struct
*task
,
2379 struct task_struct
*next
)
2383 for_each_task_context_nr(ctxn
)
2384 perf_event_context_sched_out(task
, ctxn
, next
);
2387 * if cgroup events exist on this CPU, then we need
2388 * to check if we have to switch out PMU state.
2389 * cgroup event are system-wide mode only
2391 if (atomic_read(&__get_cpu_var(perf_cgroup_events
)))
2392 perf_cgroup_sched_out(task
, next
);
2395 static void task_ctx_sched_out(struct perf_event_context
*ctx
)
2397 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
2399 if (!cpuctx
->task_ctx
)
2402 if (WARN_ON_ONCE(ctx
!= cpuctx
->task_ctx
))
2405 ctx_sched_out(ctx
, cpuctx
, EVENT_ALL
);
2406 cpuctx
->task_ctx
= NULL
;
2410 * Called with IRQs disabled
2412 static void cpu_ctx_sched_out(struct perf_cpu_context
*cpuctx
,
2413 enum event_type_t event_type
)
2415 ctx_sched_out(&cpuctx
->ctx
, cpuctx
, event_type
);
2419 ctx_pinned_sched_in(struct perf_event_context
*ctx
,
2420 struct perf_cpu_context
*cpuctx
)
2422 struct perf_event
*event
;
2424 list_for_each_entry(event
, &ctx
->pinned_groups
, group_entry
) {
2425 if (event
->state
<= PERF_EVENT_STATE_OFF
)
2427 if (!event_filter_match(event
))
2430 /* may need to reset tstamp_enabled */
2431 if (is_cgroup_event(event
))
2432 perf_cgroup_mark_enabled(event
, ctx
);
2434 if (group_can_go_on(event
, cpuctx
, 1))
2435 group_sched_in(event
, cpuctx
, ctx
);
2438 * If this pinned group hasn't been scheduled,
2439 * put it in error state.
2441 if (event
->state
== PERF_EVENT_STATE_INACTIVE
) {
2442 update_group_times(event
);
2443 event
->state
= PERF_EVENT_STATE_ERROR
;
2449 ctx_flexible_sched_in(struct perf_event_context
*ctx
,
2450 struct perf_cpu_context
*cpuctx
)
2452 struct perf_event
*event
;
2455 list_for_each_entry(event
, &ctx
->flexible_groups
, group_entry
) {
2456 /* Ignore events in OFF or ERROR state */
2457 if (event
->state
<= PERF_EVENT_STATE_OFF
)
2460 * Listen to the 'cpu' scheduling filter constraint
2463 if (!event_filter_match(event
))
2466 /* may need to reset tstamp_enabled */
2467 if (is_cgroup_event(event
))
2468 perf_cgroup_mark_enabled(event
, ctx
);
2470 if (group_can_go_on(event
, cpuctx
, can_add_hw
)) {
2471 if (group_sched_in(event
, cpuctx
, ctx
))
2478 ctx_sched_in(struct perf_event_context
*ctx
,
2479 struct perf_cpu_context
*cpuctx
,
2480 enum event_type_t event_type
,
2481 struct task_struct
*task
)
2484 int is_active
= ctx
->is_active
;
2486 ctx
->is_active
|= event_type
;
2487 if (likely(!ctx
->nr_events
))
2491 ctx
->timestamp
= now
;
2492 perf_cgroup_set_timestamp(task
, ctx
);
2494 * First go through the list and put on any pinned groups
2495 * in order to give them the best chance of going on.
2497 if (!(is_active
& EVENT_PINNED
) && (event_type
& EVENT_PINNED
))
2498 ctx_pinned_sched_in(ctx
, cpuctx
);
2500 /* Then walk through the lower prio flexible groups */
2501 if (!(is_active
& EVENT_FLEXIBLE
) && (event_type
& EVENT_FLEXIBLE
))
2502 ctx_flexible_sched_in(ctx
, cpuctx
);
2505 static void cpu_ctx_sched_in(struct perf_cpu_context
*cpuctx
,
2506 enum event_type_t event_type
,
2507 struct task_struct
*task
)
2509 struct perf_event_context
*ctx
= &cpuctx
->ctx
;
2511 ctx_sched_in(ctx
, cpuctx
, event_type
, task
);
2514 static void perf_event_context_sched_in(struct perf_event_context
*ctx
,
2515 struct task_struct
*task
)
2517 struct perf_cpu_context
*cpuctx
;
2519 cpuctx
= __get_cpu_context(ctx
);
2520 if (cpuctx
->task_ctx
== ctx
)
2523 perf_ctx_lock(cpuctx
, ctx
);
2524 perf_pmu_disable(ctx
->pmu
);
2526 * We want to keep the following priority order:
2527 * cpu pinned (that don't need to move), task pinned,
2528 * cpu flexible, task flexible.
2530 cpu_ctx_sched_out(cpuctx
, EVENT_FLEXIBLE
);
2533 cpuctx
->task_ctx
= ctx
;
2535 perf_event_sched_in(cpuctx
, cpuctx
->task_ctx
, task
);
2537 perf_pmu_enable(ctx
->pmu
);
2538 perf_ctx_unlock(cpuctx
, ctx
);
2541 * Since these rotations are per-cpu, we need to ensure the
2542 * cpu-context we got scheduled on is actually rotating.
2544 perf_pmu_rotate_start(ctx
->pmu
);
2548 * When sampling the branck stack in system-wide, it may be necessary
2549 * to flush the stack on context switch. This happens when the branch
2550 * stack does not tag its entries with the pid of the current task.
2551 * Otherwise it becomes impossible to associate a branch entry with a
2552 * task. This ambiguity is more likely to appear when the branch stack
2553 * supports priv level filtering and the user sets it to monitor only
2554 * at the user level (which could be a useful measurement in system-wide
2555 * mode). In that case, the risk is high of having a branch stack with
2556 * branch from multiple tasks. Flushing may mean dropping the existing
2557 * entries or stashing them somewhere in the PMU specific code layer.
2559 * This function provides the context switch callback to the lower code
2560 * layer. It is invoked ONLY when there is at least one system-wide context
2561 * with at least one active event using taken branch sampling.
2563 static void perf_branch_stack_sched_in(struct task_struct
*prev
,
2564 struct task_struct
*task
)
2566 struct perf_cpu_context
*cpuctx
;
2568 unsigned long flags
;
2570 /* no need to flush branch stack if not changing task */
2574 local_irq_save(flags
);
2578 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
2579 cpuctx
= this_cpu_ptr(pmu
->pmu_cpu_context
);
2582 * check if the context has at least one
2583 * event using PERF_SAMPLE_BRANCH_STACK
2585 if (cpuctx
->ctx
.nr_branch_stack
> 0
2586 && pmu
->flush_branch_stack
) {
2588 perf_ctx_lock(cpuctx
, cpuctx
->task_ctx
);
2590 perf_pmu_disable(pmu
);
2592 pmu
->flush_branch_stack();
2594 perf_pmu_enable(pmu
);
2596 perf_ctx_unlock(cpuctx
, cpuctx
->task_ctx
);
2602 local_irq_restore(flags
);
2606 * Called from scheduler to add the events of the current task
2607 * with interrupts disabled.
2609 * We restore the event value and then enable it.
2611 * This does not protect us against NMI, but enable()
2612 * sets the enabled bit in the control field of event _before_
2613 * accessing the event control register. If a NMI hits, then it will
2614 * keep the event running.
2616 void __perf_event_task_sched_in(struct task_struct
*prev
,
2617 struct task_struct
*task
)
2619 struct perf_event_context
*ctx
;
2622 for_each_task_context_nr(ctxn
) {
2623 ctx
= task
->perf_event_ctxp
[ctxn
];
2627 perf_event_context_sched_in(ctx
, task
);
2630 * if cgroup events exist on this CPU, then we need
2631 * to check if we have to switch in PMU state.
2632 * cgroup event are system-wide mode only
2634 if (atomic_read(&__get_cpu_var(perf_cgroup_events
)))
2635 perf_cgroup_sched_in(prev
, task
);
2637 /* check for system-wide branch_stack events */
2638 if (atomic_read(&__get_cpu_var(perf_branch_stack_events
)))
2639 perf_branch_stack_sched_in(prev
, task
);
2642 static u64
perf_calculate_period(struct perf_event
*event
, u64 nsec
, u64 count
)
2644 u64 frequency
= event
->attr
.sample_freq
;
2645 u64 sec
= NSEC_PER_SEC
;
2646 u64 divisor
, dividend
;
2648 int count_fls
, nsec_fls
, frequency_fls
, sec_fls
;
2650 count_fls
= fls64(count
);
2651 nsec_fls
= fls64(nsec
);
2652 frequency_fls
= fls64(frequency
);
2656 * We got @count in @nsec, with a target of sample_freq HZ
2657 * the target period becomes:
2660 * period = -------------------
2661 * @nsec * sample_freq
2666 * Reduce accuracy by one bit such that @a and @b converge
2667 * to a similar magnitude.
2669 #define REDUCE_FLS(a, b) \
2671 if (a##_fls > b##_fls) { \
2681 * Reduce accuracy until either term fits in a u64, then proceed with
2682 * the other, so that finally we can do a u64/u64 division.
2684 while (count_fls
+ sec_fls
> 64 && nsec_fls
+ frequency_fls
> 64) {
2685 REDUCE_FLS(nsec
, frequency
);
2686 REDUCE_FLS(sec
, count
);
2689 if (count_fls
+ sec_fls
> 64) {
2690 divisor
= nsec
* frequency
;
2692 while (count_fls
+ sec_fls
> 64) {
2693 REDUCE_FLS(count
, sec
);
2697 dividend
= count
* sec
;
2699 dividend
= count
* sec
;
2701 while (nsec_fls
+ frequency_fls
> 64) {
2702 REDUCE_FLS(nsec
, frequency
);
2706 divisor
= nsec
* frequency
;
2712 return div64_u64(dividend
, divisor
);
2715 static DEFINE_PER_CPU(int, perf_throttled_count
);
2716 static DEFINE_PER_CPU(u64
, perf_throttled_seq
);
2718 static void perf_adjust_period(struct perf_event
*event
, u64 nsec
, u64 count
, bool disable
)
2720 struct hw_perf_event
*hwc
= &event
->hw
;
2721 s64 period
, sample_period
;
2724 period
= perf_calculate_period(event
, nsec
, count
);
2726 delta
= (s64
)(period
- hwc
->sample_period
);
2727 delta
= (delta
+ 7) / 8; /* low pass filter */
2729 sample_period
= hwc
->sample_period
+ delta
;
2734 hwc
->sample_period
= sample_period
;
2736 if (local64_read(&hwc
->period_left
) > 8*sample_period
) {
2738 event
->pmu
->stop(event
, PERF_EF_UPDATE
);
2740 local64_set(&hwc
->period_left
, 0);
2743 event
->pmu
->start(event
, PERF_EF_RELOAD
);
2748 * combine freq adjustment with unthrottling to avoid two passes over the
2749 * events. At the same time, make sure, having freq events does not change
2750 * the rate of unthrottling as that would introduce bias.
2752 static void perf_adjust_freq_unthr_context(struct perf_event_context
*ctx
,
2755 struct perf_event
*event
;
2756 struct hw_perf_event
*hwc
;
2757 u64 now
, period
= TICK_NSEC
;
2761 * only need to iterate over all events iff:
2762 * - context have events in frequency mode (needs freq adjust)
2763 * - there are events to unthrottle on this cpu
2765 if (!(ctx
->nr_freq
|| needs_unthr
))
2768 raw_spin_lock(&ctx
->lock
);
2769 perf_pmu_disable(ctx
->pmu
);
2771 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
2772 if (event
->state
!= PERF_EVENT_STATE_ACTIVE
)
2775 if (!event_filter_match(event
))
2778 perf_pmu_disable(event
->pmu
);
2782 if (hwc
->interrupts
== MAX_INTERRUPTS
) {
2783 hwc
->interrupts
= 0;
2784 perf_log_throttle(event
, 1);
2785 event
->pmu
->start(event
, 0);
2788 if (!event
->attr
.freq
|| !event
->attr
.sample_freq
)
2792 * stop the event and update event->count
2794 event
->pmu
->stop(event
, PERF_EF_UPDATE
);
2796 now
= local64_read(&event
->count
);
2797 delta
= now
- hwc
->freq_count_stamp
;
2798 hwc
->freq_count_stamp
= now
;
2802 * reload only if value has changed
2803 * we have stopped the event so tell that
2804 * to perf_adjust_period() to avoid stopping it
2808 perf_adjust_period(event
, period
, delta
, false);
2810 event
->pmu
->start(event
, delta
> 0 ? PERF_EF_RELOAD
: 0);
2812 perf_pmu_enable(event
->pmu
);
2815 perf_pmu_enable(ctx
->pmu
);
2816 raw_spin_unlock(&ctx
->lock
);
2820 * Round-robin a context's events:
2822 static void rotate_ctx(struct perf_event_context
*ctx
)
2825 * Rotate the first entry last of non-pinned groups. Rotation might be
2826 * disabled by the inheritance code.
2828 if (!ctx
->rotate_disable
)
2829 list_rotate_left(&ctx
->flexible_groups
);
2833 * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
2834 * because they're strictly cpu affine and rotate_start is called with IRQs
2835 * disabled, while rotate_context is called from IRQ context.
2837 static int perf_rotate_context(struct perf_cpu_context
*cpuctx
)
2839 struct perf_event_context
*ctx
= NULL
;
2840 int rotate
= 0, remove
= 1;
2842 if (cpuctx
->ctx
.nr_events
) {
2844 if (cpuctx
->ctx
.nr_events
!= cpuctx
->ctx
.nr_active
)
2848 ctx
= cpuctx
->task_ctx
;
2849 if (ctx
&& ctx
->nr_events
) {
2851 if (ctx
->nr_events
!= ctx
->nr_active
)
2858 perf_ctx_lock(cpuctx
, cpuctx
->task_ctx
);
2859 perf_pmu_disable(cpuctx
->ctx
.pmu
);
2861 cpu_ctx_sched_out(cpuctx
, EVENT_FLEXIBLE
);
2863 ctx_sched_out(ctx
, cpuctx
, EVENT_FLEXIBLE
);
2865 rotate_ctx(&cpuctx
->ctx
);
2869 perf_event_sched_in(cpuctx
, ctx
, current
);
2871 perf_pmu_enable(cpuctx
->ctx
.pmu
);
2872 perf_ctx_unlock(cpuctx
, cpuctx
->task_ctx
);
2875 list_del_init(&cpuctx
->rotation_list
);
2880 #ifdef CONFIG_NO_HZ_FULL
2881 bool perf_event_can_stop_tick(void)
2883 if (atomic_read(&nr_freq_events
) ||
2884 __this_cpu_read(perf_throttled_count
))
2891 void perf_event_task_tick(void)
2893 struct list_head
*head
= &__get_cpu_var(rotation_list
);
2894 struct perf_cpu_context
*cpuctx
, *tmp
;
2895 struct perf_event_context
*ctx
;
2898 WARN_ON(!irqs_disabled());
2900 __this_cpu_inc(perf_throttled_seq
);
2901 throttled
= __this_cpu_xchg(perf_throttled_count
, 0);
2903 list_for_each_entry_safe(cpuctx
, tmp
, head
, rotation_list
) {
2905 perf_adjust_freq_unthr_context(ctx
, throttled
);
2907 ctx
= cpuctx
->task_ctx
;
2909 perf_adjust_freq_unthr_context(ctx
, throttled
);
2913 static int event_enable_on_exec(struct perf_event
*event
,
2914 struct perf_event_context
*ctx
)
2916 if (!event
->attr
.enable_on_exec
)
2919 event
->attr
.enable_on_exec
= 0;
2920 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
)
2923 __perf_event_mark_enabled(event
);
2929 * Enable all of a task's events that have been marked enable-on-exec.
2930 * This expects task == current.
2932 static void perf_event_enable_on_exec(struct perf_event_context
*ctx
)
2934 struct perf_event
*event
;
2935 unsigned long flags
;
2939 local_irq_save(flags
);
2940 if (!ctx
|| !ctx
->nr_events
)
2944 * We must ctxsw out cgroup events to avoid conflict
2945 * when invoking perf_task_event_sched_in() later on
2946 * in this function. Otherwise we end up trying to
2947 * ctxswin cgroup events which are already scheduled
2950 perf_cgroup_sched_out(current
, NULL
);
2952 raw_spin_lock(&ctx
->lock
);
2953 task_ctx_sched_out(ctx
);
2955 list_for_each_entry(event
, &ctx
->event_list
, event_entry
) {
2956 ret
= event_enable_on_exec(event
, ctx
);
2962 * Unclone this context if we enabled any event.
2967 raw_spin_unlock(&ctx
->lock
);
2970 * Also calls ctxswin for cgroup events, if any:
2972 perf_event_context_sched_in(ctx
, ctx
->task
);
2974 local_irq_restore(flags
);
2977 void perf_event_exec(void)
2979 struct perf_event_context
*ctx
;
2983 for_each_task_context_nr(ctxn
) {
2984 ctx
= current
->perf_event_ctxp
[ctxn
];
2988 perf_event_enable_on_exec(ctx
);
2994 * Cross CPU call to read the hardware event
2996 static void __perf_event_read(void *info
)
2998 struct perf_event
*event
= info
;
2999 struct perf_event_context
*ctx
= event
->ctx
;
3000 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
3003 * If this is a task context, we need to check whether it is
3004 * the current task context of this cpu. If not it has been
3005 * scheduled out before the smp call arrived. In that case
3006 * event->count would have been updated to a recent sample
3007 * when the event was scheduled out.
3009 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
)
3012 raw_spin_lock(&ctx
->lock
);
3013 if (ctx
->is_active
) {
3014 update_context_time(ctx
);
3015 update_cgrp_time_from_event(event
);
3017 update_event_times(event
);
3018 if (event
->state
== PERF_EVENT_STATE_ACTIVE
)
3019 event
->pmu
->read(event
);
3020 raw_spin_unlock(&ctx
->lock
);
3023 static inline u64
perf_event_count(struct perf_event
*event
)
3025 return local64_read(&event
->count
) + atomic64_read(&event
->child_count
);
3028 static u64
perf_event_read(struct perf_event
*event
)
3031 * If event is enabled and currently active on a CPU, update the
3032 * value in the event structure:
3034 if (event
->state
== PERF_EVENT_STATE_ACTIVE
) {
3035 smp_call_function_single(event
->oncpu
,
3036 __perf_event_read
, event
, 1);
3037 } else if (event
->state
== PERF_EVENT_STATE_INACTIVE
) {
3038 struct perf_event_context
*ctx
= event
->ctx
;
3039 unsigned long flags
;
3041 raw_spin_lock_irqsave(&ctx
->lock
, flags
);
3043 * may read while context is not active
3044 * (e.g., thread is blocked), in that case
3045 * we cannot update context time
3047 if (ctx
->is_active
) {
3048 update_context_time(ctx
);
3049 update_cgrp_time_from_event(event
);
3051 update_event_times(event
);
3052 raw_spin_unlock_irqrestore(&ctx
->lock
, flags
);
3055 return perf_event_count(event
);
3059 * Initialize the perf_event context in a task_struct:
3061 static void __perf_event_init_context(struct perf_event_context
*ctx
)
3063 raw_spin_lock_init(&ctx
->lock
);
3064 mutex_init(&ctx
->mutex
);
3065 INIT_LIST_HEAD(&ctx
->pinned_groups
);
3066 INIT_LIST_HEAD(&ctx
->flexible_groups
);
3067 INIT_LIST_HEAD(&ctx
->event_list
);
3068 atomic_set(&ctx
->refcount
, 1);
3071 static struct perf_event_context
*
3072 alloc_perf_context(struct pmu
*pmu
, struct task_struct
*task
)
3074 struct perf_event_context
*ctx
;
3076 ctx
= kzalloc(sizeof(struct perf_event_context
), GFP_KERNEL
);
3080 __perf_event_init_context(ctx
);
3083 get_task_struct(task
);
3090 static struct task_struct
*
3091 find_lively_task_by_vpid(pid_t vpid
)
3093 struct task_struct
*task
;
3100 task
= find_task_by_vpid(vpid
);
3102 get_task_struct(task
);
3106 return ERR_PTR(-ESRCH
);
3108 /* Reuse ptrace permission checks for now. */
3110 if (!ptrace_may_access(task
, PTRACE_MODE_READ
))
3115 put_task_struct(task
);
3116 return ERR_PTR(err
);
3121 * Returns a matching context with refcount and pincount.
3123 static struct perf_event_context
*
3124 find_get_context(struct pmu
*pmu
, struct task_struct
*task
, int cpu
)
3126 struct perf_event_context
*ctx
;
3127 struct perf_cpu_context
*cpuctx
;
3128 unsigned long flags
;
3132 /* Must be root to operate on a CPU event: */
3133 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN
))
3134 return ERR_PTR(-EACCES
);
3137 * We could be clever and allow to attach a event to an
3138 * offline CPU and activate it when the CPU comes up, but
3141 if (!cpu_online(cpu
))
3142 return ERR_PTR(-ENODEV
);
3144 cpuctx
= per_cpu_ptr(pmu
->pmu_cpu_context
, cpu
);
3153 ctxn
= pmu
->task_ctx_nr
;
3158 ctx
= perf_lock_task_context(task
, ctxn
, &flags
);
3162 raw_spin_unlock_irqrestore(&ctx
->lock
, flags
);
3164 ctx
= alloc_perf_context(pmu
, task
);
3170 mutex_lock(&task
->perf_event_mutex
);
3172 * If it has already passed perf_event_exit_task().
3173 * we must see PF_EXITING, it takes this mutex too.
3175 if (task
->flags
& PF_EXITING
)
3177 else if (task
->perf_event_ctxp
[ctxn
])
3182 rcu_assign_pointer(task
->perf_event_ctxp
[ctxn
], ctx
);
3184 mutex_unlock(&task
->perf_event_mutex
);
3186 if (unlikely(err
)) {
3198 return ERR_PTR(err
);
3201 static void perf_event_free_filter(struct perf_event
*event
);
3203 static void free_event_rcu(struct rcu_head
*head
)
3205 struct perf_event
*event
;
3207 event
= container_of(head
, struct perf_event
, rcu_head
);
3209 put_pid_ns(event
->ns
);
3210 perf_event_free_filter(event
);
3214 static void ring_buffer_put(struct ring_buffer
*rb
);
3215 static void ring_buffer_attach(struct perf_event
*event
,
3216 struct ring_buffer
*rb
);
3218 static void unaccount_event_cpu(struct perf_event
*event
, int cpu
)
3223 if (has_branch_stack(event
)) {
3224 if (!(event
->attach_state
& PERF_ATTACH_TASK
))
3225 atomic_dec(&per_cpu(perf_branch_stack_events
, cpu
));
3227 if (is_cgroup_event(event
))
3228 atomic_dec(&per_cpu(perf_cgroup_events
, cpu
));
3231 static void unaccount_event(struct perf_event
*event
)
3236 if (event
->attach_state
& PERF_ATTACH_TASK
)
3237 static_key_slow_dec_deferred(&perf_sched_events
);
3238 if (event
->attr
.mmap
|| event
->attr
.mmap_data
)
3239 atomic_dec(&nr_mmap_events
);
3240 if (event
->attr
.comm
)
3241 atomic_dec(&nr_comm_events
);
3242 if (event
->attr
.task
)
3243 atomic_dec(&nr_task_events
);
3244 if (event
->attr
.freq
)
3245 atomic_dec(&nr_freq_events
);
3246 if (is_cgroup_event(event
))
3247 static_key_slow_dec_deferred(&perf_sched_events
);
3248 if (has_branch_stack(event
))
3249 static_key_slow_dec_deferred(&perf_sched_events
);
3251 unaccount_event_cpu(event
, event
->cpu
);
3254 static void __free_event(struct perf_event
*event
)
3256 if (!event
->parent
) {
3257 if (event
->attr
.sample_type
& PERF_SAMPLE_CALLCHAIN
)
3258 put_callchain_buffers();
3262 event
->destroy(event
);
3265 put_ctx(event
->ctx
);
3268 module_put(event
->pmu
->module
);
3270 call_rcu(&event
->rcu_head
, free_event_rcu
);
3273 static void _free_event(struct perf_event
*event
)
3275 irq_work_sync(&event
->pending
);
3277 unaccount_event(event
);
3281 * Can happen when we close an event with re-directed output.
3283 * Since we have a 0 refcount, perf_mmap_close() will skip
3284 * over us; possibly making our ring_buffer_put() the last.
3286 mutex_lock(&event
->mmap_mutex
);
3287 ring_buffer_attach(event
, NULL
);
3288 mutex_unlock(&event
->mmap_mutex
);
3291 if (is_cgroup_event(event
))
3292 perf_detach_cgroup(event
);
3294 __free_event(event
);
3298 * Used to free events which have a known refcount of 1, such as in error paths
3299 * where the event isn't exposed yet and inherited events.
3301 static void free_event(struct perf_event
*event
)
3303 if (WARN(atomic_long_cmpxchg(&event
->refcount
, 1, 0) != 1,
3304 "unexpected event refcount: %ld; ptr=%p\n",
3305 atomic_long_read(&event
->refcount
), event
)) {
3306 /* leak to avoid use-after-free */
3314 * Called when the last reference to the file is gone.
3316 static void put_event(struct perf_event
*event
)
3318 struct perf_event_context
*ctx
= event
->ctx
;
3319 struct task_struct
*owner
;
3321 if (!atomic_long_dec_and_test(&event
->refcount
))
3325 owner
= ACCESS_ONCE(event
->owner
);
3327 * Matches the smp_wmb() in perf_event_exit_task(). If we observe
3328 * !owner it means the list deletion is complete and we can indeed
3329 * free this event, otherwise we need to serialize on
3330 * owner->perf_event_mutex.
3332 smp_read_barrier_depends();
3335 * Since delayed_put_task_struct() also drops the last
3336 * task reference we can safely take a new reference
3337 * while holding the rcu_read_lock().
3339 get_task_struct(owner
);
3344 mutex_lock(&owner
->perf_event_mutex
);
3346 * We have to re-check the event->owner field, if it is cleared
3347 * we raced with perf_event_exit_task(), acquiring the mutex
3348 * ensured they're done, and we can proceed with freeing the
3352 list_del_init(&event
->owner_entry
);
3353 mutex_unlock(&owner
->perf_event_mutex
);
3354 put_task_struct(owner
);
3357 WARN_ON_ONCE(ctx
->parent_ctx
);
3359 * There are two ways this annotation is useful:
3361 * 1) there is a lock recursion from perf_event_exit_task
3362 * see the comment there.
3364 * 2) there is a lock-inversion with mmap_sem through
3365 * perf_event_read_group(), which takes faults while
3366 * holding ctx->mutex, however this is called after
3367 * the last filedesc died, so there is no possibility
3368 * to trigger the AB-BA case.
3370 mutex_lock_nested(&ctx
->mutex
, SINGLE_DEPTH_NESTING
);
3371 perf_remove_from_context(event
, true);
3372 mutex_unlock(&ctx
->mutex
);
3377 int perf_event_release_kernel(struct perf_event
*event
)
3382 EXPORT_SYMBOL_GPL(perf_event_release_kernel
);
3384 static int perf_release(struct inode
*inode
, struct file
*file
)
3386 put_event(file
->private_data
);
3390 u64
perf_event_read_value(struct perf_event
*event
, u64
*enabled
, u64
*running
)
3392 struct perf_event
*child
;
3398 mutex_lock(&event
->child_mutex
);
3399 total
+= perf_event_read(event
);
3400 *enabled
+= event
->total_time_enabled
+
3401 atomic64_read(&event
->child_total_time_enabled
);
3402 *running
+= event
->total_time_running
+
3403 atomic64_read(&event
->child_total_time_running
);
3405 list_for_each_entry(child
, &event
->child_list
, child_list
) {
3406 total
+= perf_event_read(child
);
3407 *enabled
+= child
->total_time_enabled
;
3408 *running
+= child
->total_time_running
;
3410 mutex_unlock(&event
->child_mutex
);
3414 EXPORT_SYMBOL_GPL(perf_event_read_value
);
3416 static int perf_event_read_group(struct perf_event
*event
,
3417 u64 read_format
, char __user
*buf
)
3419 struct perf_event
*leader
= event
->group_leader
, *sub
;
3420 int n
= 0, size
= 0, ret
= -EFAULT
;
3421 struct perf_event_context
*ctx
= leader
->ctx
;
3423 u64 count
, enabled
, running
;
3425 mutex_lock(&ctx
->mutex
);
3426 count
= perf_event_read_value(leader
, &enabled
, &running
);
3428 values
[n
++] = 1 + leader
->nr_siblings
;
3429 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
3430 values
[n
++] = enabled
;
3431 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
3432 values
[n
++] = running
;
3433 values
[n
++] = count
;
3434 if (read_format
& PERF_FORMAT_ID
)
3435 values
[n
++] = primary_event_id(leader
);
3437 size
= n
* sizeof(u64
);
3439 if (copy_to_user(buf
, values
, size
))
3444 list_for_each_entry(sub
, &leader
->sibling_list
, group_entry
) {
3447 values
[n
++] = perf_event_read_value(sub
, &enabled
, &running
);
3448 if (read_format
& PERF_FORMAT_ID
)
3449 values
[n
++] = primary_event_id(sub
);
3451 size
= n
* sizeof(u64
);
3453 if (copy_to_user(buf
+ ret
, values
, size
)) {
3461 mutex_unlock(&ctx
->mutex
);
3466 static int perf_event_read_one(struct perf_event
*event
,
3467 u64 read_format
, char __user
*buf
)
3469 u64 enabled
, running
;
3473 values
[n
++] = perf_event_read_value(event
, &enabled
, &running
);
3474 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
3475 values
[n
++] = enabled
;
3476 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
3477 values
[n
++] = running
;
3478 if (read_format
& PERF_FORMAT_ID
)
3479 values
[n
++] = primary_event_id(event
);
3481 if (copy_to_user(buf
, values
, n
* sizeof(u64
)))
3484 return n
* sizeof(u64
);
3488 * Read the performance event - simple non blocking version for now
3491 perf_read_hw(struct perf_event
*event
, char __user
*buf
, size_t count
)
3493 u64 read_format
= event
->attr
.read_format
;
3497 * Return end-of-file for a read on a event that is in
3498 * error state (i.e. because it was pinned but it couldn't be
3499 * scheduled on to the CPU at some point).
3501 if (event
->state
== PERF_EVENT_STATE_ERROR
)
3504 if (count
< event
->read_size
)
3507 WARN_ON_ONCE(event
->ctx
->parent_ctx
);
3508 if (read_format
& PERF_FORMAT_GROUP
)
3509 ret
= perf_event_read_group(event
, read_format
, buf
);
3511 ret
= perf_event_read_one(event
, read_format
, buf
);
3517 perf_read(struct file
*file
, char __user
*buf
, size_t count
, loff_t
*ppos
)
3519 struct perf_event
*event
= file
->private_data
;
3521 return perf_read_hw(event
, buf
, count
);
3524 static unsigned int perf_poll(struct file
*file
, poll_table
*wait
)
3526 struct perf_event
*event
= file
->private_data
;
3527 struct ring_buffer
*rb
;
3528 unsigned int events
= POLL_HUP
;
3531 * Pin the event->rb by taking event->mmap_mutex; otherwise
3532 * perf_event_set_output() can swizzle our rb and make us miss wakeups.
3534 mutex_lock(&event
->mmap_mutex
);
3537 events
= atomic_xchg(&rb
->poll
, 0);
3538 mutex_unlock(&event
->mmap_mutex
);
3540 poll_wait(file
, &event
->waitq
, wait
);
3545 static void perf_event_reset(struct perf_event
*event
)
3547 (void)perf_event_read(event
);
3548 local64_set(&event
->count
, 0);
3549 perf_event_update_userpage(event
);
3553 * Holding the top-level event's child_mutex means that any
3554 * descendant process that has inherited this event will block
3555 * in sync_child_event if it goes to exit, thus satisfying the
3556 * task existence requirements of perf_event_enable/disable.
3558 static void perf_event_for_each_child(struct perf_event
*event
,
3559 void (*func
)(struct perf_event
*))
3561 struct perf_event
*child
;
3563 WARN_ON_ONCE(event
->ctx
->parent_ctx
);
3564 mutex_lock(&event
->child_mutex
);
3566 list_for_each_entry(child
, &event
->child_list
, child_list
)
3568 mutex_unlock(&event
->child_mutex
);
3571 static void perf_event_for_each(struct perf_event
*event
,
3572 void (*func
)(struct perf_event
*))
3574 struct perf_event_context
*ctx
= event
->ctx
;
3575 struct perf_event
*sibling
;
3577 WARN_ON_ONCE(ctx
->parent_ctx
);
3578 mutex_lock(&ctx
->mutex
);
3579 event
= event
->group_leader
;
3581 perf_event_for_each_child(event
, func
);
3582 list_for_each_entry(sibling
, &event
->sibling_list
, group_entry
)
3583 perf_event_for_each_child(sibling
, func
);
3584 mutex_unlock(&ctx
->mutex
);
3587 static int perf_event_period(struct perf_event
*event
, u64 __user
*arg
)
3589 struct perf_event_context
*ctx
= event
->ctx
;
3590 int ret
= 0, active
;
3593 if (!is_sampling_event(event
))
3596 if (copy_from_user(&value
, arg
, sizeof(value
)))
3602 raw_spin_lock_irq(&ctx
->lock
);
3603 if (event
->attr
.freq
) {
3604 if (value
> sysctl_perf_event_sample_rate
) {
3609 event
->attr
.sample_freq
= value
;
3611 event
->attr
.sample_period
= value
;
3612 event
->hw
.sample_period
= value
;
3615 active
= (event
->state
== PERF_EVENT_STATE_ACTIVE
);
3617 perf_pmu_disable(ctx
->pmu
);
3618 event
->pmu
->stop(event
, PERF_EF_UPDATE
);
3621 local64_set(&event
->hw
.period_left
, 0);
3624 event
->pmu
->start(event
, PERF_EF_RELOAD
);
3625 perf_pmu_enable(ctx
->pmu
);
3629 raw_spin_unlock_irq(&ctx
->lock
);
3634 static const struct file_operations perf_fops
;
3636 static inline int perf_fget_light(int fd
, struct fd
*p
)
3638 struct fd f
= fdget(fd
);
3642 if (f
.file
->f_op
!= &perf_fops
) {
3650 static int perf_event_set_output(struct perf_event
*event
,
3651 struct perf_event
*output_event
);
3652 static int perf_event_set_filter(struct perf_event
*event
, void __user
*arg
);
3654 static long perf_ioctl(struct file
*file
, unsigned int cmd
, unsigned long arg
)
3656 struct perf_event
*event
= file
->private_data
;
3657 void (*func
)(struct perf_event
*);
3661 case PERF_EVENT_IOC_ENABLE
:
3662 func
= perf_event_enable
;
3664 case PERF_EVENT_IOC_DISABLE
:
3665 func
= perf_event_disable
;
3667 case PERF_EVENT_IOC_RESET
:
3668 func
= perf_event_reset
;
3671 case PERF_EVENT_IOC_REFRESH
:
3672 return perf_event_refresh(event
, arg
);
3674 case PERF_EVENT_IOC_PERIOD
:
3675 return perf_event_period(event
, (u64 __user
*)arg
);
3677 case PERF_EVENT_IOC_ID
:
3679 u64 id
= primary_event_id(event
);
3681 if (copy_to_user((void __user
*)arg
, &id
, sizeof(id
)))
3686 case PERF_EVENT_IOC_SET_OUTPUT
:
3690 struct perf_event
*output_event
;
3692 ret
= perf_fget_light(arg
, &output
);
3695 output_event
= output
.file
->private_data
;
3696 ret
= perf_event_set_output(event
, output_event
);
3699 ret
= perf_event_set_output(event
, NULL
);
3704 case PERF_EVENT_IOC_SET_FILTER
:
3705 return perf_event_set_filter(event
, (void __user
*)arg
);
3711 if (flags
& PERF_IOC_FLAG_GROUP
)
3712 perf_event_for_each(event
, func
);
3714 perf_event_for_each_child(event
, func
);
3719 int perf_event_task_enable(void)
3721 struct perf_event
*event
;
3723 mutex_lock(¤t
->perf_event_mutex
);
3724 list_for_each_entry(event
, ¤t
->perf_event_list
, owner_entry
)
3725 perf_event_for_each_child(event
, perf_event_enable
);
3726 mutex_unlock(¤t
->perf_event_mutex
);
3731 int perf_event_task_disable(void)
3733 struct perf_event
*event
;
3735 mutex_lock(¤t
->perf_event_mutex
);
3736 list_for_each_entry(event
, ¤t
->perf_event_list
, owner_entry
)
3737 perf_event_for_each_child(event
, perf_event_disable
);
3738 mutex_unlock(¤t
->perf_event_mutex
);
3743 static int perf_event_index(struct perf_event
*event
)
3745 if (event
->hw
.state
& PERF_HES_STOPPED
)
3748 if (event
->state
!= PERF_EVENT_STATE_ACTIVE
)
3751 return event
->pmu
->event_idx(event
);
3754 static void calc_timer_values(struct perf_event
*event
,
3761 *now
= perf_clock();
3762 ctx_time
= event
->shadow_ctx_time
+ *now
;
3763 *enabled
= ctx_time
- event
->tstamp_enabled
;
3764 *running
= ctx_time
- event
->tstamp_running
;
3767 static void perf_event_init_userpage(struct perf_event
*event
)
3769 struct perf_event_mmap_page
*userpg
;
3770 struct ring_buffer
*rb
;
3773 rb
= rcu_dereference(event
->rb
);
3777 userpg
= rb
->user_page
;
3779 /* Allow new userspace to detect that bit 0 is deprecated */
3780 userpg
->cap_bit0_is_deprecated
= 1;
3781 userpg
->size
= offsetof(struct perf_event_mmap_page
, __reserved
);
3787 void __weak
arch_perf_update_userpage(struct perf_event_mmap_page
*userpg
, u64 now
)
3792 * Callers need to ensure there can be no nesting of this function, otherwise
3793 * the seqlock logic goes bad. We can not serialize this because the arch
3794 * code calls this from NMI context.
3796 void perf_event_update_userpage(struct perf_event
*event
)
3798 struct perf_event_mmap_page
*userpg
;
3799 struct ring_buffer
*rb
;
3800 u64 enabled
, running
, now
;
3803 rb
= rcu_dereference(event
->rb
);
3808 * compute total_time_enabled, total_time_running
3809 * based on snapshot values taken when the event
3810 * was last scheduled in.
3812 * we cannot simply called update_context_time()
3813 * because of locking issue as we can be called in
3816 calc_timer_values(event
, &now
, &enabled
, &running
);
3818 userpg
= rb
->user_page
;
3820 * Disable preemption so as to not let the corresponding user-space
3821 * spin too long if we get preempted.
3826 userpg
->index
= perf_event_index(event
);
3827 userpg
->offset
= perf_event_count(event
);
3829 userpg
->offset
-= local64_read(&event
->hw
.prev_count
);
3831 userpg
->time_enabled
= enabled
+
3832 atomic64_read(&event
->child_total_time_enabled
);
3834 userpg
->time_running
= running
+
3835 atomic64_read(&event
->child_total_time_running
);
3837 arch_perf_update_userpage(userpg
, now
);
3846 static int perf_mmap_fault(struct vm_area_struct
*vma
, struct vm_fault
*vmf
)
3848 struct perf_event
*event
= vma
->vm_file
->private_data
;
3849 struct ring_buffer
*rb
;
3850 int ret
= VM_FAULT_SIGBUS
;
3852 if (vmf
->flags
& FAULT_FLAG_MKWRITE
) {
3853 if (vmf
->pgoff
== 0)
3859 rb
= rcu_dereference(event
->rb
);
3863 if (vmf
->pgoff
&& (vmf
->flags
& FAULT_FLAG_WRITE
))
3866 vmf
->page
= perf_mmap_to_page(rb
, vmf
->pgoff
);
3870 get_page(vmf
->page
);
3871 vmf
->page
->mapping
= vma
->vm_file
->f_mapping
;
3872 vmf
->page
->index
= vmf
->pgoff
;
3881 static void ring_buffer_attach(struct perf_event
*event
,
3882 struct ring_buffer
*rb
)
3884 struct ring_buffer
*old_rb
= NULL
;
3885 unsigned long flags
;
3889 * Should be impossible, we set this when removing
3890 * event->rb_entry and wait/clear when adding event->rb_entry.
3892 WARN_ON_ONCE(event
->rcu_pending
);
3895 event
->rcu_batches
= get_state_synchronize_rcu();
3896 event
->rcu_pending
= 1;
3898 spin_lock_irqsave(&old_rb
->event_lock
, flags
);
3899 list_del_rcu(&event
->rb_entry
);
3900 spin_unlock_irqrestore(&old_rb
->event_lock
, flags
);
3903 if (event
->rcu_pending
&& rb
) {
3904 cond_synchronize_rcu(event
->rcu_batches
);
3905 event
->rcu_pending
= 0;
3909 spin_lock_irqsave(&rb
->event_lock
, flags
);
3910 list_add_rcu(&event
->rb_entry
, &rb
->event_list
);
3911 spin_unlock_irqrestore(&rb
->event_lock
, flags
);
3914 rcu_assign_pointer(event
->rb
, rb
);
3917 ring_buffer_put(old_rb
);
3919 * Since we detached before setting the new rb, so that we
3920 * could attach the new rb, we could have missed a wakeup.
3923 wake_up_all(&event
->waitq
);
3927 static void ring_buffer_wakeup(struct perf_event
*event
)
3929 struct ring_buffer
*rb
;
3932 rb
= rcu_dereference(event
->rb
);
3934 list_for_each_entry_rcu(event
, &rb
->event_list
, rb_entry
)
3935 wake_up_all(&event
->waitq
);
3940 static void rb_free_rcu(struct rcu_head
*rcu_head
)
3942 struct ring_buffer
*rb
;
3944 rb
= container_of(rcu_head
, struct ring_buffer
, rcu_head
);
3948 static struct ring_buffer
*ring_buffer_get(struct perf_event
*event
)
3950 struct ring_buffer
*rb
;
3953 rb
= rcu_dereference(event
->rb
);
3955 if (!atomic_inc_not_zero(&rb
->refcount
))
3963 static void ring_buffer_put(struct ring_buffer
*rb
)
3965 if (!atomic_dec_and_test(&rb
->refcount
))
3968 WARN_ON_ONCE(!list_empty(&rb
->event_list
));
3970 call_rcu(&rb
->rcu_head
, rb_free_rcu
);
3973 static void perf_mmap_open(struct vm_area_struct
*vma
)
3975 struct perf_event
*event
= vma
->vm_file
->private_data
;
3977 atomic_inc(&event
->mmap_count
);
3978 atomic_inc(&event
->rb
->mmap_count
);
3982 * A buffer can be mmap()ed multiple times; either directly through the same
3983 * event, or through other events by use of perf_event_set_output().
3985 * In order to undo the VM accounting done by perf_mmap() we need to destroy
3986 * the buffer here, where we still have a VM context. This means we need
3987 * to detach all events redirecting to us.
3989 static void perf_mmap_close(struct vm_area_struct
*vma
)
3991 struct perf_event
*event
= vma
->vm_file
->private_data
;
3993 struct ring_buffer
*rb
= ring_buffer_get(event
);
3994 struct user_struct
*mmap_user
= rb
->mmap_user
;
3995 int mmap_locked
= rb
->mmap_locked
;
3996 unsigned long size
= perf_data_size(rb
);
3998 atomic_dec(&rb
->mmap_count
);
4000 if (!atomic_dec_and_mutex_lock(&event
->mmap_count
, &event
->mmap_mutex
))
4003 ring_buffer_attach(event
, NULL
);
4004 mutex_unlock(&event
->mmap_mutex
);
4006 /* If there's still other mmap()s of this buffer, we're done. */
4007 if (atomic_read(&rb
->mmap_count
))
4011 * No other mmap()s, detach from all other events that might redirect
4012 * into the now unreachable buffer. Somewhat complicated by the
4013 * fact that rb::event_lock otherwise nests inside mmap_mutex.
4017 list_for_each_entry_rcu(event
, &rb
->event_list
, rb_entry
) {
4018 if (!atomic_long_inc_not_zero(&event
->refcount
)) {
4020 * This event is en-route to free_event() which will
4021 * detach it and remove it from the list.
4027 mutex_lock(&event
->mmap_mutex
);
4029 * Check we didn't race with perf_event_set_output() which can
4030 * swizzle the rb from under us while we were waiting to
4031 * acquire mmap_mutex.
4033 * If we find a different rb; ignore this event, a next
4034 * iteration will no longer find it on the list. We have to
4035 * still restart the iteration to make sure we're not now
4036 * iterating the wrong list.
4038 if (event
->rb
== rb
)
4039 ring_buffer_attach(event
, NULL
);
4041 mutex_unlock(&event
->mmap_mutex
);
4045 * Restart the iteration; either we're on the wrong list or
4046 * destroyed its integrity by doing a deletion.
4053 * It could be there's still a few 0-ref events on the list; they'll
4054 * get cleaned up by free_event() -- they'll also still have their
4055 * ref on the rb and will free it whenever they are done with it.
4057 * Aside from that, this buffer is 'fully' detached and unmapped,
4058 * undo the VM accounting.
4061 atomic_long_sub((size
>> PAGE_SHIFT
) + 1, &mmap_user
->locked_vm
);
4062 vma
->vm_mm
->pinned_vm
-= mmap_locked
;
4063 free_uid(mmap_user
);
4066 ring_buffer_put(rb
); /* could be last */
4069 static const struct vm_operations_struct perf_mmap_vmops
= {
4070 .open
= perf_mmap_open
,
4071 .close
= perf_mmap_close
,
4072 .fault
= perf_mmap_fault
,
4073 .page_mkwrite
= perf_mmap_fault
,
4076 static int perf_mmap(struct file
*file
, struct vm_area_struct
*vma
)
4078 struct perf_event
*event
= file
->private_data
;
4079 unsigned long user_locked
, user_lock_limit
;
4080 struct user_struct
*user
= current_user();
4081 unsigned long locked
, lock_limit
;
4082 struct ring_buffer
*rb
;
4083 unsigned long vma_size
;
4084 unsigned long nr_pages
;
4085 long user_extra
, extra
;
4086 int ret
= 0, flags
= 0;
4089 * Don't allow mmap() of inherited per-task counters. This would
4090 * create a performance issue due to all children writing to the
4093 if (event
->cpu
== -1 && event
->attr
.inherit
)
4096 if (!(vma
->vm_flags
& VM_SHARED
))
4099 vma_size
= vma
->vm_end
- vma
->vm_start
;
4100 nr_pages
= (vma_size
/ PAGE_SIZE
) - 1;
4103 * If we have rb pages ensure they're a power-of-two number, so we
4104 * can do bitmasks instead of modulo.
4106 if (nr_pages
!= 0 && !is_power_of_2(nr_pages
))
4109 if (vma_size
!= PAGE_SIZE
* (1 + nr_pages
))
4112 if (vma
->vm_pgoff
!= 0)
4115 WARN_ON_ONCE(event
->ctx
->parent_ctx
);
4117 mutex_lock(&event
->mmap_mutex
);
4119 if (event
->rb
->nr_pages
!= nr_pages
) {
4124 if (!atomic_inc_not_zero(&event
->rb
->mmap_count
)) {
4126 * Raced against perf_mmap_close() through
4127 * perf_event_set_output(). Try again, hope for better
4130 mutex_unlock(&event
->mmap_mutex
);
4137 user_extra
= nr_pages
+ 1;
4138 user_lock_limit
= sysctl_perf_event_mlock
>> (PAGE_SHIFT
- 10);
4141 * Increase the limit linearly with more CPUs:
4143 user_lock_limit
*= num_online_cpus();
4145 user_locked
= atomic_long_read(&user
->locked_vm
) + user_extra
;
4148 if (user_locked
> user_lock_limit
)
4149 extra
= user_locked
- user_lock_limit
;
4151 lock_limit
= rlimit(RLIMIT_MEMLOCK
);
4152 lock_limit
>>= PAGE_SHIFT
;
4153 locked
= vma
->vm_mm
->pinned_vm
+ extra
;
4155 if ((locked
> lock_limit
) && perf_paranoid_tracepoint_raw() &&
4156 !capable(CAP_IPC_LOCK
)) {
4163 if (vma
->vm_flags
& VM_WRITE
)
4164 flags
|= RING_BUFFER_WRITABLE
;
4166 rb
= rb_alloc(nr_pages
,
4167 event
->attr
.watermark
? event
->attr
.wakeup_watermark
: 0,
4175 atomic_set(&rb
->mmap_count
, 1);
4176 rb
->mmap_locked
= extra
;
4177 rb
->mmap_user
= get_current_user();
4179 atomic_long_add(user_extra
, &user
->locked_vm
);
4180 vma
->vm_mm
->pinned_vm
+= extra
;
4182 ring_buffer_attach(event
, rb
);
4184 perf_event_init_userpage(event
);
4185 perf_event_update_userpage(event
);
4189 atomic_inc(&event
->mmap_count
);
4190 mutex_unlock(&event
->mmap_mutex
);
4193 * Since pinned accounting is per vm we cannot allow fork() to copy our
4196 vma
->vm_flags
|= VM_DONTCOPY
| VM_DONTEXPAND
| VM_DONTDUMP
;
4197 vma
->vm_ops
= &perf_mmap_vmops
;
4202 static int perf_fasync(int fd
, struct file
*filp
, int on
)
4204 struct inode
*inode
= file_inode(filp
);
4205 struct perf_event
*event
= filp
->private_data
;
4208 mutex_lock(&inode
->i_mutex
);
4209 retval
= fasync_helper(fd
, filp
, on
, &event
->fasync
);
4210 mutex_unlock(&inode
->i_mutex
);
4218 static const struct file_operations perf_fops
= {
4219 .llseek
= no_llseek
,
4220 .release
= perf_release
,
4223 .unlocked_ioctl
= perf_ioctl
,
4224 .compat_ioctl
= perf_ioctl
,
4226 .fasync
= perf_fasync
,
4232 * If there's data, ensure we set the poll() state and publish everything
4233 * to user-space before waking everybody up.
4236 void perf_event_wakeup(struct perf_event
*event
)
4238 ring_buffer_wakeup(event
);
4240 if (event
->pending_kill
) {
4241 kill_fasync(&event
->fasync
, SIGIO
, event
->pending_kill
);
4242 event
->pending_kill
= 0;
4246 static void perf_pending_event(struct irq_work
*entry
)
4248 struct perf_event
*event
= container_of(entry
,
4249 struct perf_event
, pending
);
4251 if (event
->pending_disable
) {
4252 event
->pending_disable
= 0;
4253 __perf_event_disable(event
);
4256 if (event
->pending_wakeup
) {
4257 event
->pending_wakeup
= 0;
4258 perf_event_wakeup(event
);
4263 * We assume there is only KVM supporting the callbacks.
4264 * Later on, we might change it to a list if there is
4265 * another virtualization implementation supporting the callbacks.
4267 struct perf_guest_info_callbacks
*perf_guest_cbs
;
4269 int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks
*cbs
)
4271 perf_guest_cbs
= cbs
;
4274 EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks
);
4276 int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks
*cbs
)
4278 perf_guest_cbs
= NULL
;
4281 EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks
);
4284 perf_output_sample_regs(struct perf_output_handle
*handle
,
4285 struct pt_regs
*regs
, u64 mask
)
4289 for_each_set_bit(bit
, (const unsigned long *) &mask
,
4290 sizeof(mask
) * BITS_PER_BYTE
) {
4293 val
= perf_reg_value(regs
, bit
);
4294 perf_output_put(handle
, val
);
4298 static void perf_sample_regs_user(struct perf_regs_user
*regs_user
,
4299 struct pt_regs
*regs
)
4301 if (!user_mode(regs
)) {
4303 regs
= task_pt_regs(current
);
4309 regs_user
->regs
= regs
;
4310 regs_user
->abi
= perf_reg_abi(current
);
4315 * Get remaining task size from user stack pointer.
4317 * It'd be better to take stack vma map and limit this more
4318 * precisly, but there's no way to get it safely under interrupt,
4319 * so using TASK_SIZE as limit.
4321 static u64
perf_ustack_task_size(struct pt_regs
*regs
)
4323 unsigned long addr
= perf_user_stack_pointer(regs
);
4325 if (!addr
|| addr
>= TASK_SIZE
)
4328 return TASK_SIZE
- addr
;
4332 perf_sample_ustack_size(u16 stack_size
, u16 header_size
,
4333 struct pt_regs
*regs
)
4337 /* No regs, no stack pointer, no dump. */
4342 * Check if we fit in with the requested stack size into the:
4344 * If we don't, we limit the size to the TASK_SIZE.
4346 * - remaining sample size
4347 * If we don't, we customize the stack size to
4348 * fit in to the remaining sample size.
4351 task_size
= min((u64
) USHRT_MAX
, perf_ustack_task_size(regs
));
4352 stack_size
= min(stack_size
, (u16
) task_size
);
4354 /* Current header size plus static size and dynamic size. */
4355 header_size
+= 2 * sizeof(u64
);
4357 /* Do we fit in with the current stack dump size? */
4358 if ((u16
) (header_size
+ stack_size
) < header_size
) {
4360 * If we overflow the maximum size for the sample,
4361 * we customize the stack dump size to fit in.
4363 stack_size
= USHRT_MAX
- header_size
- sizeof(u64
);
4364 stack_size
= round_up(stack_size
, sizeof(u64
));
4371 perf_output_sample_ustack(struct perf_output_handle
*handle
, u64 dump_size
,
4372 struct pt_regs
*regs
)
4374 /* Case of a kernel thread, nothing to dump */
4377 perf_output_put(handle
, size
);
4386 * - the size requested by user or the best one we can fit
4387 * in to the sample max size
4389 * - user stack dump data
4391 * - the actual dumped size
4395 perf_output_put(handle
, dump_size
);
4398 sp
= perf_user_stack_pointer(regs
);
4399 rem
= __output_copy_user(handle
, (void *) sp
, dump_size
);
4400 dyn_size
= dump_size
- rem
;
4402 perf_output_skip(handle
, rem
);
4405 perf_output_put(handle
, dyn_size
);
4409 static void __perf_event_header__init_id(struct perf_event_header
*header
,
4410 struct perf_sample_data
*data
,
4411 struct perf_event
*event
)
4413 u64 sample_type
= event
->attr
.sample_type
;
4415 data
->type
= sample_type
;
4416 header
->size
+= event
->id_header_size
;
4418 if (sample_type
& PERF_SAMPLE_TID
) {
4419 /* namespace issues */
4420 data
->tid_entry
.pid
= perf_event_pid(event
, current
);
4421 data
->tid_entry
.tid
= perf_event_tid(event
, current
);
4424 if (sample_type
& PERF_SAMPLE_TIME
)
4425 data
->time
= perf_clock();
4427 if (sample_type
& (PERF_SAMPLE_ID
| PERF_SAMPLE_IDENTIFIER
))
4428 data
->id
= primary_event_id(event
);
4430 if (sample_type
& PERF_SAMPLE_STREAM_ID
)
4431 data
->stream_id
= event
->id
;
4433 if (sample_type
& PERF_SAMPLE_CPU
) {
4434 data
->cpu_entry
.cpu
= raw_smp_processor_id();
4435 data
->cpu_entry
.reserved
= 0;
4439 void perf_event_header__init_id(struct perf_event_header
*header
,
4440 struct perf_sample_data
*data
,
4441 struct perf_event
*event
)
4443 if (event
->attr
.sample_id_all
)
4444 __perf_event_header__init_id(header
, data
, event
);
4447 static void __perf_event__output_id_sample(struct perf_output_handle
*handle
,
4448 struct perf_sample_data
*data
)
4450 u64 sample_type
= data
->type
;
4452 if (sample_type
& PERF_SAMPLE_TID
)
4453 perf_output_put(handle
, data
->tid_entry
);
4455 if (sample_type
& PERF_SAMPLE_TIME
)
4456 perf_output_put(handle
, data
->time
);
4458 if (sample_type
& PERF_SAMPLE_ID
)
4459 perf_output_put(handle
, data
->id
);
4461 if (sample_type
& PERF_SAMPLE_STREAM_ID
)
4462 perf_output_put(handle
, data
->stream_id
);
4464 if (sample_type
& PERF_SAMPLE_CPU
)
4465 perf_output_put(handle
, data
->cpu_entry
);
4467 if (sample_type
& PERF_SAMPLE_IDENTIFIER
)
4468 perf_output_put(handle
, data
->id
);
4471 void perf_event__output_id_sample(struct perf_event
*event
,
4472 struct perf_output_handle
*handle
,
4473 struct perf_sample_data
*sample
)
4475 if (event
->attr
.sample_id_all
)
4476 __perf_event__output_id_sample(handle
, sample
);
4479 static void perf_output_read_one(struct perf_output_handle
*handle
,
4480 struct perf_event
*event
,
4481 u64 enabled
, u64 running
)
4483 u64 read_format
= event
->attr
.read_format
;
4487 values
[n
++] = perf_event_count(event
);
4488 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
) {
4489 values
[n
++] = enabled
+
4490 atomic64_read(&event
->child_total_time_enabled
);
4492 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
) {
4493 values
[n
++] = running
+
4494 atomic64_read(&event
->child_total_time_running
);
4496 if (read_format
& PERF_FORMAT_ID
)
4497 values
[n
++] = primary_event_id(event
);
4499 __output_copy(handle
, values
, n
* sizeof(u64
));
4503 * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
4505 static void perf_output_read_group(struct perf_output_handle
*handle
,
4506 struct perf_event
*event
,
4507 u64 enabled
, u64 running
)
4509 struct perf_event
*leader
= event
->group_leader
, *sub
;
4510 u64 read_format
= event
->attr
.read_format
;
4514 values
[n
++] = 1 + leader
->nr_siblings
;
4516 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
4517 values
[n
++] = enabled
;
4519 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
4520 values
[n
++] = running
;
4522 if (leader
!= event
)
4523 leader
->pmu
->read(leader
);
4525 values
[n
++] = perf_event_count(leader
);
4526 if (read_format
& PERF_FORMAT_ID
)
4527 values
[n
++] = primary_event_id(leader
);
4529 __output_copy(handle
, values
, n
* sizeof(u64
));
4531 list_for_each_entry(sub
, &leader
->sibling_list
, group_entry
) {
4534 if ((sub
!= event
) &&
4535 (sub
->state
== PERF_EVENT_STATE_ACTIVE
))
4536 sub
->pmu
->read(sub
);
4538 values
[n
++] = perf_event_count(sub
);
4539 if (read_format
& PERF_FORMAT_ID
)
4540 values
[n
++] = primary_event_id(sub
);
4542 __output_copy(handle
, values
, n
* sizeof(u64
));
4546 #define PERF_FORMAT_TOTAL_TIMES (PERF_FORMAT_TOTAL_TIME_ENABLED|\
4547 PERF_FORMAT_TOTAL_TIME_RUNNING)
4549 static void perf_output_read(struct perf_output_handle
*handle
,
4550 struct perf_event
*event
)
4552 u64 enabled
= 0, running
= 0, now
;
4553 u64 read_format
= event
->attr
.read_format
;
4556 * compute total_time_enabled, total_time_running
4557 * based on snapshot values taken when the event
4558 * was last scheduled in.
4560 * we cannot simply called update_context_time()
4561 * because of locking issue as we are called in
4564 if (read_format
& PERF_FORMAT_TOTAL_TIMES
)
4565 calc_timer_values(event
, &now
, &enabled
, &running
);
4567 if (event
->attr
.read_format
& PERF_FORMAT_GROUP
)
4568 perf_output_read_group(handle
, event
, enabled
, running
);
4570 perf_output_read_one(handle
, event
, enabled
, running
);
4573 void perf_output_sample(struct perf_output_handle
*handle
,
4574 struct perf_event_header
*header
,
4575 struct perf_sample_data
*data
,
4576 struct perf_event
*event
)
4578 u64 sample_type
= data
->type
;
4580 perf_output_put(handle
, *header
);
4582 if (sample_type
& PERF_SAMPLE_IDENTIFIER
)
4583 perf_output_put(handle
, data
->id
);
4585 if (sample_type
& PERF_SAMPLE_IP
)
4586 perf_output_put(handle
, data
->ip
);
4588 if (sample_type
& PERF_SAMPLE_TID
)
4589 perf_output_put(handle
, data
->tid_entry
);
4591 if (sample_type
& PERF_SAMPLE_TIME
)
4592 perf_output_put(handle
, data
->time
);
4594 if (sample_type
& PERF_SAMPLE_ADDR
)
4595 perf_output_put(handle
, data
->addr
);
4597 if (sample_type
& PERF_SAMPLE_ID
)
4598 perf_output_put(handle
, data
->id
);
4600 if (sample_type
& PERF_SAMPLE_STREAM_ID
)
4601 perf_output_put(handle
, data
->stream_id
);
4603 if (sample_type
& PERF_SAMPLE_CPU
)
4604 perf_output_put(handle
, data
->cpu_entry
);
4606 if (sample_type
& PERF_SAMPLE_PERIOD
)
4607 perf_output_put(handle
, data
->period
);
4609 if (sample_type
& PERF_SAMPLE_READ
)
4610 perf_output_read(handle
, event
);
4612 if (sample_type
& PERF_SAMPLE_CALLCHAIN
) {
4613 if (data
->callchain
) {
4616 if (data
->callchain
)
4617 size
+= data
->callchain
->nr
;
4619 size
*= sizeof(u64
);
4621 __output_copy(handle
, data
->callchain
, size
);
4624 perf_output_put(handle
, nr
);
4628 if (sample_type
& PERF_SAMPLE_RAW
) {
4630 perf_output_put(handle
, data
->raw
->size
);
4631 __output_copy(handle
, data
->raw
->data
,
4638 .size
= sizeof(u32
),
4641 perf_output_put(handle
, raw
);
4645 if (sample_type
& PERF_SAMPLE_BRANCH_STACK
) {
4646 if (data
->br_stack
) {
4649 size
= data
->br_stack
->nr
4650 * sizeof(struct perf_branch_entry
);
4652 perf_output_put(handle
, data
->br_stack
->nr
);
4653 perf_output_copy(handle
, data
->br_stack
->entries
, size
);
4656 * we always store at least the value of nr
4659 perf_output_put(handle
, nr
);
4663 if (sample_type
& PERF_SAMPLE_REGS_USER
) {
4664 u64 abi
= data
->regs_user
.abi
;
4667 * If there are no regs to dump, notice it through
4668 * first u64 being zero (PERF_SAMPLE_REGS_ABI_NONE).
4670 perf_output_put(handle
, abi
);
4673 u64 mask
= event
->attr
.sample_regs_user
;
4674 perf_output_sample_regs(handle
,
4675 data
->regs_user
.regs
,
4680 if (sample_type
& PERF_SAMPLE_STACK_USER
) {
4681 perf_output_sample_ustack(handle
,
4682 data
->stack_user_size
,
4683 data
->regs_user
.regs
);
4686 if (sample_type
& PERF_SAMPLE_WEIGHT
)
4687 perf_output_put(handle
, data
->weight
);
4689 if (sample_type
& PERF_SAMPLE_DATA_SRC
)
4690 perf_output_put(handle
, data
->data_src
.val
);
4692 if (sample_type
& PERF_SAMPLE_TRANSACTION
)
4693 perf_output_put(handle
, data
->txn
);
4695 if (!event
->attr
.watermark
) {
4696 int wakeup_events
= event
->attr
.wakeup_events
;
4698 if (wakeup_events
) {
4699 struct ring_buffer
*rb
= handle
->rb
;
4700 int events
= local_inc_return(&rb
->events
);
4702 if (events
>= wakeup_events
) {
4703 local_sub(wakeup_events
, &rb
->events
);
4704 local_inc(&rb
->wakeup
);
4710 void perf_prepare_sample(struct perf_event_header
*header
,
4711 struct perf_sample_data
*data
,
4712 struct perf_event
*event
,
4713 struct pt_regs
*regs
)
4715 u64 sample_type
= event
->attr
.sample_type
;
4717 header
->type
= PERF_RECORD_SAMPLE
;
4718 header
->size
= sizeof(*header
) + event
->header_size
;
4721 header
->misc
|= perf_misc_flags(regs
);
4723 __perf_event_header__init_id(header
, data
, event
);
4725 if (sample_type
& PERF_SAMPLE_IP
)
4726 data
->ip
= perf_instruction_pointer(regs
);
4728 if (sample_type
& PERF_SAMPLE_CALLCHAIN
) {
4731 data
->callchain
= perf_callchain(event
, regs
);
4733 if (data
->callchain
)
4734 size
+= data
->callchain
->nr
;
4736 header
->size
+= size
* sizeof(u64
);
4739 if (sample_type
& PERF_SAMPLE_RAW
) {
4740 int size
= sizeof(u32
);
4743 size
+= data
->raw
->size
;
4745 size
+= sizeof(u32
);
4747 WARN_ON_ONCE(size
& (sizeof(u64
)-1));
4748 header
->size
+= size
;
4751 if (sample_type
& PERF_SAMPLE_BRANCH_STACK
) {
4752 int size
= sizeof(u64
); /* nr */
4753 if (data
->br_stack
) {
4754 size
+= data
->br_stack
->nr
4755 * sizeof(struct perf_branch_entry
);
4757 header
->size
+= size
;
4760 if (sample_type
& PERF_SAMPLE_REGS_USER
) {
4761 /* regs dump ABI info */
4762 int size
= sizeof(u64
);
4764 perf_sample_regs_user(&data
->regs_user
, regs
);
4766 if (data
->regs_user
.regs
) {
4767 u64 mask
= event
->attr
.sample_regs_user
;
4768 size
+= hweight64(mask
) * sizeof(u64
);
4771 header
->size
+= size
;
4774 if (sample_type
& PERF_SAMPLE_STACK_USER
) {
4776 * Either we need PERF_SAMPLE_STACK_USER bit to be allways
4777 * processed as the last one or have additional check added
4778 * in case new sample type is added, because we could eat
4779 * up the rest of the sample size.
4781 struct perf_regs_user
*uregs
= &data
->regs_user
;
4782 u16 stack_size
= event
->attr
.sample_stack_user
;
4783 u16 size
= sizeof(u64
);
4786 perf_sample_regs_user(uregs
, regs
);
4788 stack_size
= perf_sample_ustack_size(stack_size
, header
->size
,
4792 * If there is something to dump, add space for the dump
4793 * itself and for the field that tells the dynamic size,
4794 * which is how many have been actually dumped.
4797 size
+= sizeof(u64
) + stack_size
;
4799 data
->stack_user_size
= stack_size
;
4800 header
->size
+= size
;
4804 static void perf_event_output(struct perf_event
*event
,
4805 struct perf_sample_data
*data
,
4806 struct pt_regs
*regs
)
4808 struct perf_output_handle handle
;
4809 struct perf_event_header header
;
4811 /* protect the callchain buffers */
4814 perf_prepare_sample(&header
, data
, event
, regs
);
4816 if (perf_output_begin(&handle
, event
, header
.size
))
4819 perf_output_sample(&handle
, &header
, data
, event
);
4821 perf_output_end(&handle
);
4831 struct perf_read_event
{
4832 struct perf_event_header header
;
4839 perf_event_read_event(struct perf_event
*event
,
4840 struct task_struct
*task
)
4842 struct perf_output_handle handle
;
4843 struct perf_sample_data sample
;
4844 struct perf_read_event read_event
= {
4846 .type
= PERF_RECORD_READ
,
4848 .size
= sizeof(read_event
) + event
->read_size
,
4850 .pid
= perf_event_pid(event
, task
),
4851 .tid
= perf_event_tid(event
, task
),
4855 perf_event_header__init_id(&read_event
.header
, &sample
, event
);
4856 ret
= perf_output_begin(&handle
, event
, read_event
.header
.size
);
4860 perf_output_put(&handle
, read_event
);
4861 perf_output_read(&handle
, event
);
4862 perf_event__output_id_sample(event
, &handle
, &sample
);
4864 perf_output_end(&handle
);
4867 typedef void (perf_event_aux_output_cb
)(struct perf_event
*event
, void *data
);
4870 perf_event_aux_ctx(struct perf_event_context
*ctx
,
4871 perf_event_aux_output_cb output
,
4874 struct perf_event
*event
;
4876 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
4877 if (event
->state
< PERF_EVENT_STATE_INACTIVE
)
4879 if (!event_filter_match(event
))
4881 output(event
, data
);
4886 perf_event_aux(perf_event_aux_output_cb output
, void *data
,
4887 struct perf_event_context
*task_ctx
)
4889 struct perf_cpu_context
*cpuctx
;
4890 struct perf_event_context
*ctx
;
4895 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
4896 cpuctx
= get_cpu_ptr(pmu
->pmu_cpu_context
);
4897 if (cpuctx
->unique_pmu
!= pmu
)
4899 perf_event_aux_ctx(&cpuctx
->ctx
, output
, data
);
4902 ctxn
= pmu
->task_ctx_nr
;
4905 ctx
= rcu_dereference(current
->perf_event_ctxp
[ctxn
]);
4907 perf_event_aux_ctx(ctx
, output
, data
);
4909 put_cpu_ptr(pmu
->pmu_cpu_context
);
4914 perf_event_aux_ctx(task_ctx
, output
, data
);
4921 * task tracking -- fork/exit
4923 * enabled by: attr.comm | attr.mmap | attr.mmap2 | attr.mmap_data | attr.task
4926 struct perf_task_event
{
4927 struct task_struct
*task
;
4928 struct perf_event_context
*task_ctx
;
4931 struct perf_event_header header
;
4941 static int perf_event_task_match(struct perf_event
*event
)
4943 return event
->attr
.comm
|| event
->attr
.mmap
||
4944 event
->attr
.mmap2
|| event
->attr
.mmap_data
||
4948 static void perf_event_task_output(struct perf_event
*event
,
4951 struct perf_task_event
*task_event
= data
;
4952 struct perf_output_handle handle
;
4953 struct perf_sample_data sample
;
4954 struct task_struct
*task
= task_event
->task
;
4955 int ret
, size
= task_event
->event_id
.header
.size
;
4957 if (!perf_event_task_match(event
))
4960 perf_event_header__init_id(&task_event
->event_id
.header
, &sample
, event
);
4962 ret
= perf_output_begin(&handle
, event
,
4963 task_event
->event_id
.header
.size
);
4967 task_event
->event_id
.pid
= perf_event_pid(event
, task
);
4968 task_event
->event_id
.ppid
= perf_event_pid(event
, current
);
4970 task_event
->event_id
.tid
= perf_event_tid(event
, task
);
4971 task_event
->event_id
.ptid
= perf_event_tid(event
, current
);
4973 perf_output_put(&handle
, task_event
->event_id
);
4975 perf_event__output_id_sample(event
, &handle
, &sample
);
4977 perf_output_end(&handle
);
4979 task_event
->event_id
.header
.size
= size
;
4982 static void perf_event_task(struct task_struct
*task
,
4983 struct perf_event_context
*task_ctx
,
4986 struct perf_task_event task_event
;
4988 if (!atomic_read(&nr_comm_events
) &&
4989 !atomic_read(&nr_mmap_events
) &&
4990 !atomic_read(&nr_task_events
))
4993 task_event
= (struct perf_task_event
){
4995 .task_ctx
= task_ctx
,
4998 .type
= new ? PERF_RECORD_FORK
: PERF_RECORD_EXIT
,
5000 .size
= sizeof(task_event
.event_id
),
5006 .time
= perf_clock(),
5010 perf_event_aux(perf_event_task_output
,
5015 void perf_event_fork(struct task_struct
*task
)
5017 perf_event_task(task
, NULL
, 1);
5024 struct perf_comm_event
{
5025 struct task_struct
*task
;
5030 struct perf_event_header header
;
5037 static int perf_event_comm_match(struct perf_event
*event
)
5039 return event
->attr
.comm
;
5042 static void perf_event_comm_output(struct perf_event
*event
,
5045 struct perf_comm_event
*comm_event
= data
;
5046 struct perf_output_handle handle
;
5047 struct perf_sample_data sample
;
5048 int size
= comm_event
->event_id
.header
.size
;
5051 if (!perf_event_comm_match(event
))
5054 perf_event_header__init_id(&comm_event
->event_id
.header
, &sample
, event
);
5055 ret
= perf_output_begin(&handle
, event
,
5056 comm_event
->event_id
.header
.size
);
5061 comm_event
->event_id
.pid
= perf_event_pid(event
, comm_event
->task
);
5062 comm_event
->event_id
.tid
= perf_event_tid(event
, comm_event
->task
);
5064 perf_output_put(&handle
, comm_event
->event_id
);
5065 __output_copy(&handle
, comm_event
->comm
,
5066 comm_event
->comm_size
);
5068 perf_event__output_id_sample(event
, &handle
, &sample
);
5070 perf_output_end(&handle
);
5072 comm_event
->event_id
.header
.size
= size
;
5075 static void perf_event_comm_event(struct perf_comm_event
*comm_event
)
5077 char comm
[TASK_COMM_LEN
];
5080 memset(comm
, 0, sizeof(comm
));
5081 strlcpy(comm
, comm_event
->task
->comm
, sizeof(comm
));
5082 size
= ALIGN(strlen(comm
)+1, sizeof(u64
));
5084 comm_event
->comm
= comm
;
5085 comm_event
->comm_size
= size
;
5087 comm_event
->event_id
.header
.size
= sizeof(comm_event
->event_id
) + size
;
5089 perf_event_aux(perf_event_comm_output
,
5094 void perf_event_comm(struct task_struct
*task
, bool exec
)
5096 struct perf_comm_event comm_event
;
5098 if (!atomic_read(&nr_comm_events
))
5101 comm_event
= (struct perf_comm_event
){
5107 .type
= PERF_RECORD_COMM
,
5108 .misc
= exec
? PERF_RECORD_MISC_COMM_EXEC
: 0,
5116 perf_event_comm_event(&comm_event
);
5123 struct perf_mmap_event
{
5124 struct vm_area_struct
*vma
;
5126 const char *file_name
;
5133 struct perf_event_header header
;
5143 static int perf_event_mmap_match(struct perf_event
*event
,
5146 struct perf_mmap_event
*mmap_event
= data
;
5147 struct vm_area_struct
*vma
= mmap_event
->vma
;
5148 int executable
= vma
->vm_flags
& VM_EXEC
;
5150 return (!executable
&& event
->attr
.mmap_data
) ||
5151 (executable
&& (event
->attr
.mmap
|| event
->attr
.mmap2
));
5154 static void perf_event_mmap_output(struct perf_event
*event
,
5157 struct perf_mmap_event
*mmap_event
= data
;
5158 struct perf_output_handle handle
;
5159 struct perf_sample_data sample
;
5160 int size
= mmap_event
->event_id
.header
.size
;
5163 if (!perf_event_mmap_match(event
, data
))
5166 if (event
->attr
.mmap2
) {
5167 mmap_event
->event_id
.header
.type
= PERF_RECORD_MMAP2
;
5168 mmap_event
->event_id
.header
.size
+= sizeof(mmap_event
->maj
);
5169 mmap_event
->event_id
.header
.size
+= sizeof(mmap_event
->min
);
5170 mmap_event
->event_id
.header
.size
+= sizeof(mmap_event
->ino
);
5171 mmap_event
->event_id
.header
.size
+= sizeof(mmap_event
->ino_generation
);
5174 perf_event_header__init_id(&mmap_event
->event_id
.header
, &sample
, event
);
5175 ret
= perf_output_begin(&handle
, event
,
5176 mmap_event
->event_id
.header
.size
);
5180 mmap_event
->event_id
.pid
= perf_event_pid(event
, current
);
5181 mmap_event
->event_id
.tid
= perf_event_tid(event
, current
);
5183 perf_output_put(&handle
, mmap_event
->event_id
);
5185 if (event
->attr
.mmap2
) {
5186 perf_output_put(&handle
, mmap_event
->maj
);
5187 perf_output_put(&handle
, mmap_event
->min
);
5188 perf_output_put(&handle
, mmap_event
->ino
);
5189 perf_output_put(&handle
, mmap_event
->ino_generation
);
5192 __output_copy(&handle
, mmap_event
->file_name
,
5193 mmap_event
->file_size
);
5195 perf_event__output_id_sample(event
, &handle
, &sample
);
5197 perf_output_end(&handle
);
5199 mmap_event
->event_id
.header
.size
= size
;
5202 static void perf_event_mmap_event(struct perf_mmap_event
*mmap_event
)
5204 struct vm_area_struct
*vma
= mmap_event
->vma
;
5205 struct file
*file
= vma
->vm_file
;
5206 int maj
= 0, min
= 0;
5207 u64 ino
= 0, gen
= 0;
5214 struct inode
*inode
;
5217 buf
= kmalloc(PATH_MAX
, GFP_KERNEL
);
5223 * d_path() works from the end of the rb backwards, so we
5224 * need to add enough zero bytes after the string to handle
5225 * the 64bit alignment we do later.
5227 name
= d_path(&file
->f_path
, buf
, PATH_MAX
- sizeof(u64
));
5232 inode
= file_inode(vma
->vm_file
);
5233 dev
= inode
->i_sb
->s_dev
;
5235 gen
= inode
->i_generation
;
5240 name
= (char *)arch_vma_name(vma
);
5244 if (vma
->vm_start
<= vma
->vm_mm
->start_brk
&&
5245 vma
->vm_end
>= vma
->vm_mm
->brk
) {
5249 if (vma
->vm_start
<= vma
->vm_mm
->start_stack
&&
5250 vma
->vm_end
>= vma
->vm_mm
->start_stack
) {
5260 strlcpy(tmp
, name
, sizeof(tmp
));
5264 * Since our buffer works in 8 byte units we need to align our string
5265 * size to a multiple of 8. However, we must guarantee the tail end is
5266 * zero'd out to avoid leaking random bits to userspace.
5268 size
= strlen(name
)+1;
5269 while (!IS_ALIGNED(size
, sizeof(u64
)))
5270 name
[size
++] = '\0';
5272 mmap_event
->file_name
= name
;
5273 mmap_event
->file_size
= size
;
5274 mmap_event
->maj
= maj
;
5275 mmap_event
->min
= min
;
5276 mmap_event
->ino
= ino
;
5277 mmap_event
->ino_generation
= gen
;
5279 if (!(vma
->vm_flags
& VM_EXEC
))
5280 mmap_event
->event_id
.header
.misc
|= PERF_RECORD_MISC_MMAP_DATA
;
5282 mmap_event
->event_id
.header
.size
= sizeof(mmap_event
->event_id
) + size
;
5284 perf_event_aux(perf_event_mmap_output
,
5291 void perf_event_mmap(struct vm_area_struct
*vma
)
5293 struct perf_mmap_event mmap_event
;
5295 if (!atomic_read(&nr_mmap_events
))
5298 mmap_event
= (struct perf_mmap_event
){
5304 .type
= PERF_RECORD_MMAP
,
5305 .misc
= PERF_RECORD_MISC_USER
,
5310 .start
= vma
->vm_start
,
5311 .len
= vma
->vm_end
- vma
->vm_start
,
5312 .pgoff
= (u64
)vma
->vm_pgoff
<< PAGE_SHIFT
,
5314 /* .maj (attr_mmap2 only) */
5315 /* .min (attr_mmap2 only) */
5316 /* .ino (attr_mmap2 only) */
5317 /* .ino_generation (attr_mmap2 only) */
5320 perf_event_mmap_event(&mmap_event
);
5324 * IRQ throttle logging
5327 static void perf_log_throttle(struct perf_event
*event
, int enable
)
5329 struct perf_output_handle handle
;
5330 struct perf_sample_data sample
;
5334 struct perf_event_header header
;
5338 } throttle_event
= {
5340 .type
= PERF_RECORD_THROTTLE
,
5342 .size
= sizeof(throttle_event
),
5344 .time
= perf_clock(),
5345 .id
= primary_event_id(event
),
5346 .stream_id
= event
->id
,
5350 throttle_event
.header
.type
= PERF_RECORD_UNTHROTTLE
;
5352 perf_event_header__init_id(&throttle_event
.header
, &sample
, event
);
5354 ret
= perf_output_begin(&handle
, event
,
5355 throttle_event
.header
.size
);
5359 perf_output_put(&handle
, throttle_event
);
5360 perf_event__output_id_sample(event
, &handle
, &sample
);
5361 perf_output_end(&handle
);
5365 * Generic event overflow handling, sampling.
5368 static int __perf_event_overflow(struct perf_event
*event
,
5369 int throttle
, struct perf_sample_data
*data
,
5370 struct pt_regs
*regs
)
5372 int events
= atomic_read(&event
->event_limit
);
5373 struct hw_perf_event
*hwc
= &event
->hw
;
5378 * Non-sampling counters might still use the PMI to fold short
5379 * hardware counters, ignore those.
5381 if (unlikely(!is_sampling_event(event
)))
5384 seq
= __this_cpu_read(perf_throttled_seq
);
5385 if (seq
!= hwc
->interrupts_seq
) {
5386 hwc
->interrupts_seq
= seq
;
5387 hwc
->interrupts
= 1;
5390 if (unlikely(throttle
5391 && hwc
->interrupts
>= max_samples_per_tick
)) {
5392 __this_cpu_inc(perf_throttled_count
);
5393 hwc
->interrupts
= MAX_INTERRUPTS
;
5394 perf_log_throttle(event
, 0);
5395 tick_nohz_full_kick();
5400 if (event
->attr
.freq
) {
5401 u64 now
= perf_clock();
5402 s64 delta
= now
- hwc
->freq_time_stamp
;
5404 hwc
->freq_time_stamp
= now
;
5406 if (delta
> 0 && delta
< 2*TICK_NSEC
)
5407 perf_adjust_period(event
, delta
, hwc
->last_period
, true);
5411 * XXX event_limit might not quite work as expected on inherited
5415 event
->pending_kill
= POLL_IN
;
5416 if (events
&& atomic_dec_and_test(&event
->event_limit
)) {
5418 event
->pending_kill
= POLL_HUP
;
5419 event
->pending_disable
= 1;
5420 irq_work_queue(&event
->pending
);
5423 if (event
->overflow_handler
)
5424 event
->overflow_handler(event
, data
, regs
);
5426 perf_event_output(event
, data
, regs
);
5428 if (event
->fasync
&& event
->pending_kill
) {
5429 event
->pending_wakeup
= 1;
5430 irq_work_queue(&event
->pending
);
5436 int perf_event_overflow(struct perf_event
*event
,
5437 struct perf_sample_data
*data
,
5438 struct pt_regs
*regs
)
5440 return __perf_event_overflow(event
, 1, data
, regs
);
5444 * Generic software event infrastructure
5447 struct swevent_htable
{
5448 struct swevent_hlist
*swevent_hlist
;
5449 struct mutex hlist_mutex
;
5452 /* Recursion avoidance in each contexts */
5453 int recursion
[PERF_NR_CONTEXTS
];
5455 /* Keeps track of cpu being initialized/exited */
5459 static DEFINE_PER_CPU(struct swevent_htable
, swevent_htable
);
5462 * We directly increment event->count and keep a second value in
5463 * event->hw.period_left to count intervals. This period event
5464 * is kept in the range [-sample_period, 0] so that we can use the
5468 u64
perf_swevent_set_period(struct perf_event
*event
)
5470 struct hw_perf_event
*hwc
= &event
->hw
;
5471 u64 period
= hwc
->last_period
;
5475 hwc
->last_period
= hwc
->sample_period
;
5478 old
= val
= local64_read(&hwc
->period_left
);
5482 nr
= div64_u64(period
+ val
, period
);
5483 offset
= nr
* period
;
5485 if (local64_cmpxchg(&hwc
->period_left
, old
, val
) != old
)
5491 static void perf_swevent_overflow(struct perf_event
*event
, u64 overflow
,
5492 struct perf_sample_data
*data
,
5493 struct pt_regs
*regs
)
5495 struct hw_perf_event
*hwc
= &event
->hw
;
5499 overflow
= perf_swevent_set_period(event
);
5501 if (hwc
->interrupts
== MAX_INTERRUPTS
)
5504 for (; overflow
; overflow
--) {
5505 if (__perf_event_overflow(event
, throttle
,
5508 * We inhibit the overflow from happening when
5509 * hwc->interrupts == MAX_INTERRUPTS.
5517 static void perf_swevent_event(struct perf_event
*event
, u64 nr
,
5518 struct perf_sample_data
*data
,
5519 struct pt_regs
*regs
)
5521 struct hw_perf_event
*hwc
= &event
->hw
;
5523 local64_add(nr
, &event
->count
);
5528 if (!is_sampling_event(event
))
5531 if ((event
->attr
.sample_type
& PERF_SAMPLE_PERIOD
) && !event
->attr
.freq
) {
5533 return perf_swevent_overflow(event
, 1, data
, regs
);
5535 data
->period
= event
->hw
.last_period
;
5537 if (nr
== 1 && hwc
->sample_period
== 1 && !event
->attr
.freq
)
5538 return perf_swevent_overflow(event
, 1, data
, regs
);
5540 if (local64_add_negative(nr
, &hwc
->period_left
))
5543 perf_swevent_overflow(event
, 0, data
, regs
);
5546 static int perf_exclude_event(struct perf_event
*event
,
5547 struct pt_regs
*regs
)
5549 if (event
->hw
.state
& PERF_HES_STOPPED
)
5553 if (event
->attr
.exclude_user
&& user_mode(regs
))
5556 if (event
->attr
.exclude_kernel
&& !user_mode(regs
))
5563 static int perf_swevent_match(struct perf_event
*event
,
5564 enum perf_type_id type
,
5566 struct perf_sample_data
*data
,
5567 struct pt_regs
*regs
)
5569 if (event
->attr
.type
!= type
)
5572 if (event
->attr
.config
!= event_id
)
5575 if (perf_exclude_event(event
, regs
))
5581 static inline u64
swevent_hash(u64 type
, u32 event_id
)
5583 u64 val
= event_id
| (type
<< 32);
5585 return hash_64(val
, SWEVENT_HLIST_BITS
);
5588 static inline struct hlist_head
*
5589 __find_swevent_head(struct swevent_hlist
*hlist
, u64 type
, u32 event_id
)
5591 u64 hash
= swevent_hash(type
, event_id
);
5593 return &hlist
->heads
[hash
];
5596 /* For the read side: events when they trigger */
5597 static inline struct hlist_head
*
5598 find_swevent_head_rcu(struct swevent_htable
*swhash
, u64 type
, u32 event_id
)
5600 struct swevent_hlist
*hlist
;
5602 hlist
= rcu_dereference(swhash
->swevent_hlist
);
5606 return __find_swevent_head(hlist
, type
, event_id
);
5609 /* For the event head insertion and removal in the hlist */
5610 static inline struct hlist_head
*
5611 find_swevent_head(struct swevent_htable
*swhash
, struct perf_event
*event
)
5613 struct swevent_hlist
*hlist
;
5614 u32 event_id
= event
->attr
.config
;
5615 u64 type
= event
->attr
.type
;
5618 * Event scheduling is always serialized against hlist allocation
5619 * and release. Which makes the protected version suitable here.
5620 * The context lock guarantees that.
5622 hlist
= rcu_dereference_protected(swhash
->swevent_hlist
,
5623 lockdep_is_held(&event
->ctx
->lock
));
5627 return __find_swevent_head(hlist
, type
, event_id
);
5630 static void do_perf_sw_event(enum perf_type_id type
, u32 event_id
,
5632 struct perf_sample_data
*data
,
5633 struct pt_regs
*regs
)
5635 struct swevent_htable
*swhash
= &__get_cpu_var(swevent_htable
);
5636 struct perf_event
*event
;
5637 struct hlist_head
*head
;
5640 head
= find_swevent_head_rcu(swhash
, type
, event_id
);
5644 hlist_for_each_entry_rcu(event
, head
, hlist_entry
) {
5645 if (perf_swevent_match(event
, type
, event_id
, data
, regs
))
5646 perf_swevent_event(event
, nr
, data
, regs
);
5652 int perf_swevent_get_recursion_context(void)
5654 struct swevent_htable
*swhash
= &__get_cpu_var(swevent_htable
);
5656 return get_recursion_context(swhash
->recursion
);
5658 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context
);
5660 inline void perf_swevent_put_recursion_context(int rctx
)
5662 struct swevent_htable
*swhash
= &__get_cpu_var(swevent_htable
);
5664 put_recursion_context(swhash
->recursion
, rctx
);
5667 void __perf_sw_event(u32 event_id
, u64 nr
, struct pt_regs
*regs
, u64 addr
)
5669 struct perf_sample_data data
;
5672 preempt_disable_notrace();
5673 rctx
= perf_swevent_get_recursion_context();
5677 perf_sample_data_init(&data
, addr
, 0);
5679 do_perf_sw_event(PERF_TYPE_SOFTWARE
, event_id
, nr
, &data
, regs
);
5681 perf_swevent_put_recursion_context(rctx
);
5682 preempt_enable_notrace();
5685 static void perf_swevent_read(struct perf_event
*event
)
5689 static int perf_swevent_add(struct perf_event
*event
, int flags
)
5691 struct swevent_htable
*swhash
= &__get_cpu_var(swevent_htable
);
5692 struct hw_perf_event
*hwc
= &event
->hw
;
5693 struct hlist_head
*head
;
5695 if (is_sampling_event(event
)) {
5696 hwc
->last_period
= hwc
->sample_period
;
5697 perf_swevent_set_period(event
);
5700 hwc
->state
= !(flags
& PERF_EF_START
);
5702 head
= find_swevent_head(swhash
, event
);
5705 * We can race with cpu hotplug code. Do not
5706 * WARN if the cpu just got unplugged.
5708 WARN_ON_ONCE(swhash
->online
);
5712 hlist_add_head_rcu(&event
->hlist_entry
, head
);
5717 static void perf_swevent_del(struct perf_event
*event
, int flags
)
5719 hlist_del_rcu(&event
->hlist_entry
);
5722 static void perf_swevent_start(struct perf_event
*event
, int flags
)
5724 event
->hw
.state
= 0;
5727 static void perf_swevent_stop(struct perf_event
*event
, int flags
)
5729 event
->hw
.state
= PERF_HES_STOPPED
;
5732 /* Deref the hlist from the update side */
5733 static inline struct swevent_hlist
*
5734 swevent_hlist_deref(struct swevent_htable
*swhash
)
5736 return rcu_dereference_protected(swhash
->swevent_hlist
,
5737 lockdep_is_held(&swhash
->hlist_mutex
));
5740 static void swevent_hlist_release(struct swevent_htable
*swhash
)
5742 struct swevent_hlist
*hlist
= swevent_hlist_deref(swhash
);
5747 rcu_assign_pointer(swhash
->swevent_hlist
, NULL
);
5748 kfree_rcu(hlist
, rcu_head
);
5751 static void swevent_hlist_put_cpu(struct perf_event
*event
, int cpu
)
5753 struct swevent_htable
*swhash
= &per_cpu(swevent_htable
, cpu
);
5755 mutex_lock(&swhash
->hlist_mutex
);
5757 if (!--swhash
->hlist_refcount
)
5758 swevent_hlist_release(swhash
);
5760 mutex_unlock(&swhash
->hlist_mutex
);
5763 static void swevent_hlist_put(struct perf_event
*event
)
5767 for_each_possible_cpu(cpu
)
5768 swevent_hlist_put_cpu(event
, cpu
);
5771 static int swevent_hlist_get_cpu(struct perf_event
*event
, int cpu
)
5773 struct swevent_htable
*swhash
= &per_cpu(swevent_htable
, cpu
);
5776 mutex_lock(&swhash
->hlist_mutex
);
5778 if (!swevent_hlist_deref(swhash
) && cpu_online(cpu
)) {
5779 struct swevent_hlist
*hlist
;
5781 hlist
= kzalloc(sizeof(*hlist
), GFP_KERNEL
);
5786 rcu_assign_pointer(swhash
->swevent_hlist
, hlist
);
5788 swhash
->hlist_refcount
++;
5790 mutex_unlock(&swhash
->hlist_mutex
);
5795 static int swevent_hlist_get(struct perf_event
*event
)
5798 int cpu
, failed_cpu
;
5801 for_each_possible_cpu(cpu
) {
5802 err
= swevent_hlist_get_cpu(event
, cpu
);
5812 for_each_possible_cpu(cpu
) {
5813 if (cpu
== failed_cpu
)
5815 swevent_hlist_put_cpu(event
, cpu
);
5822 struct static_key perf_swevent_enabled
[PERF_COUNT_SW_MAX
];
5824 static void sw_perf_event_destroy(struct perf_event
*event
)
5826 u64 event_id
= event
->attr
.config
;
5828 WARN_ON(event
->parent
);
5830 static_key_slow_dec(&perf_swevent_enabled
[event_id
]);
5831 swevent_hlist_put(event
);
5834 static int perf_swevent_init(struct perf_event
*event
)
5836 u64 event_id
= event
->attr
.config
;
5838 if (event
->attr
.type
!= PERF_TYPE_SOFTWARE
)
5842 * no branch sampling for software events
5844 if (has_branch_stack(event
))
5848 case PERF_COUNT_SW_CPU_CLOCK
:
5849 case PERF_COUNT_SW_TASK_CLOCK
:
5856 if (event_id
>= PERF_COUNT_SW_MAX
)
5859 if (!event
->parent
) {
5862 err
= swevent_hlist_get(event
);
5866 static_key_slow_inc(&perf_swevent_enabled
[event_id
]);
5867 event
->destroy
= sw_perf_event_destroy
;
5873 static int perf_swevent_event_idx(struct perf_event
*event
)
5878 static struct pmu perf_swevent
= {
5879 .task_ctx_nr
= perf_sw_context
,
5881 .event_init
= perf_swevent_init
,
5882 .add
= perf_swevent_add
,
5883 .del
= perf_swevent_del
,
5884 .start
= perf_swevent_start
,
5885 .stop
= perf_swevent_stop
,
5886 .read
= perf_swevent_read
,
5888 .event_idx
= perf_swevent_event_idx
,
5891 #ifdef CONFIG_EVENT_TRACING
5893 static int perf_tp_filter_match(struct perf_event
*event
,
5894 struct perf_sample_data
*data
)
5896 void *record
= data
->raw
->data
;
5898 if (likely(!event
->filter
) || filter_match_preds(event
->filter
, record
))
5903 static int perf_tp_event_match(struct perf_event
*event
,
5904 struct perf_sample_data
*data
,
5905 struct pt_regs
*regs
)
5907 if (event
->hw
.state
& PERF_HES_STOPPED
)
5910 * All tracepoints are from kernel-space.
5912 if (event
->attr
.exclude_kernel
)
5915 if (!perf_tp_filter_match(event
, data
))
5921 void perf_tp_event(u64 addr
, u64 count
, void *record
, int entry_size
,
5922 struct pt_regs
*regs
, struct hlist_head
*head
, int rctx
,
5923 struct task_struct
*task
)
5925 struct perf_sample_data data
;
5926 struct perf_event
*event
;
5928 struct perf_raw_record raw
= {
5933 perf_sample_data_init(&data
, addr
, 0);
5936 hlist_for_each_entry_rcu(event
, head
, hlist_entry
) {
5937 if (perf_tp_event_match(event
, &data
, regs
))
5938 perf_swevent_event(event
, count
, &data
, regs
);
5942 * If we got specified a target task, also iterate its context and
5943 * deliver this event there too.
5945 if (task
&& task
!= current
) {
5946 struct perf_event_context
*ctx
;
5947 struct trace_entry
*entry
= record
;
5950 ctx
= rcu_dereference(task
->perf_event_ctxp
[perf_sw_context
]);
5954 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
5955 if (event
->attr
.type
!= PERF_TYPE_TRACEPOINT
)
5957 if (event
->attr
.config
!= entry
->type
)
5959 if (perf_tp_event_match(event
, &data
, regs
))
5960 perf_swevent_event(event
, count
, &data
, regs
);
5966 perf_swevent_put_recursion_context(rctx
);
5968 EXPORT_SYMBOL_GPL(perf_tp_event
);
5970 static void tp_perf_event_destroy(struct perf_event
*event
)
5972 perf_trace_destroy(event
);
5975 static int perf_tp_event_init(struct perf_event
*event
)
5979 if (event
->attr
.type
!= PERF_TYPE_TRACEPOINT
)
5983 * no branch sampling for tracepoint events
5985 if (has_branch_stack(event
))
5988 err
= perf_trace_init(event
);
5992 event
->destroy
= tp_perf_event_destroy
;
5997 static struct pmu perf_tracepoint
= {
5998 .task_ctx_nr
= perf_sw_context
,
6000 .event_init
= perf_tp_event_init
,
6001 .add
= perf_trace_add
,
6002 .del
= perf_trace_del
,
6003 .start
= perf_swevent_start
,
6004 .stop
= perf_swevent_stop
,
6005 .read
= perf_swevent_read
,
6007 .event_idx
= perf_swevent_event_idx
,
6010 static inline void perf_tp_register(void)
6012 perf_pmu_register(&perf_tracepoint
, "tracepoint", PERF_TYPE_TRACEPOINT
);
6015 static int perf_event_set_filter(struct perf_event
*event
, void __user
*arg
)
6020 if (event
->attr
.type
!= PERF_TYPE_TRACEPOINT
)
6023 filter_str
= strndup_user(arg
, PAGE_SIZE
);
6024 if (IS_ERR(filter_str
))
6025 return PTR_ERR(filter_str
);
6027 ret
= ftrace_profile_set_filter(event
, event
->attr
.config
, filter_str
);
6033 static void perf_event_free_filter(struct perf_event
*event
)
6035 ftrace_profile_free_filter(event
);
6040 static inline void perf_tp_register(void)
6044 static int perf_event_set_filter(struct perf_event
*event
, void __user
*arg
)
6049 static void perf_event_free_filter(struct perf_event
*event
)
6053 #endif /* CONFIG_EVENT_TRACING */
6055 #ifdef CONFIG_HAVE_HW_BREAKPOINT
6056 void perf_bp_event(struct perf_event
*bp
, void *data
)
6058 struct perf_sample_data sample
;
6059 struct pt_regs
*regs
= data
;
6061 perf_sample_data_init(&sample
, bp
->attr
.bp_addr
, 0);
6063 if (!bp
->hw
.state
&& !perf_exclude_event(bp
, regs
))
6064 perf_swevent_event(bp
, 1, &sample
, regs
);
6069 * hrtimer based swevent callback
6072 static enum hrtimer_restart
perf_swevent_hrtimer(struct hrtimer
*hrtimer
)
6074 enum hrtimer_restart ret
= HRTIMER_RESTART
;
6075 struct perf_sample_data data
;
6076 struct pt_regs
*regs
;
6077 struct perf_event
*event
;
6080 event
= container_of(hrtimer
, struct perf_event
, hw
.hrtimer
);
6082 if (event
->state
!= PERF_EVENT_STATE_ACTIVE
)
6083 return HRTIMER_NORESTART
;
6085 event
->pmu
->read(event
);
6087 perf_sample_data_init(&data
, 0, event
->hw
.last_period
);
6088 regs
= get_irq_regs();
6090 if (regs
&& !perf_exclude_event(event
, regs
)) {
6091 if (!(event
->attr
.exclude_idle
&& is_idle_task(current
)))
6092 if (__perf_event_overflow(event
, 1, &data
, regs
))
6093 ret
= HRTIMER_NORESTART
;
6096 period
= max_t(u64
, 10000, event
->hw
.sample_period
);
6097 hrtimer_forward_now(hrtimer
, ns_to_ktime(period
));
6102 static void perf_swevent_start_hrtimer(struct perf_event
*event
)
6104 struct hw_perf_event
*hwc
= &event
->hw
;
6107 if (!is_sampling_event(event
))
6110 period
= local64_read(&hwc
->period_left
);
6115 local64_set(&hwc
->period_left
, 0);
6117 period
= max_t(u64
, 10000, hwc
->sample_period
);
6119 __hrtimer_start_range_ns(&hwc
->hrtimer
,
6120 ns_to_ktime(period
), 0,
6121 HRTIMER_MODE_REL_PINNED
, 0);
6124 static void perf_swevent_cancel_hrtimer(struct perf_event
*event
)
6126 struct hw_perf_event
*hwc
= &event
->hw
;
6128 if (is_sampling_event(event
)) {
6129 ktime_t remaining
= hrtimer_get_remaining(&hwc
->hrtimer
);
6130 local64_set(&hwc
->period_left
, ktime_to_ns(remaining
));
6132 hrtimer_cancel(&hwc
->hrtimer
);
6136 static void perf_swevent_init_hrtimer(struct perf_event
*event
)
6138 struct hw_perf_event
*hwc
= &event
->hw
;
6140 if (!is_sampling_event(event
))
6143 hrtimer_init(&hwc
->hrtimer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
6144 hwc
->hrtimer
.function
= perf_swevent_hrtimer
;
6147 * Since hrtimers have a fixed rate, we can do a static freq->period
6148 * mapping and avoid the whole period adjust feedback stuff.
6150 if (event
->attr
.freq
) {
6151 long freq
= event
->attr
.sample_freq
;
6153 event
->attr
.sample_period
= NSEC_PER_SEC
/ freq
;
6154 hwc
->sample_period
= event
->attr
.sample_period
;
6155 local64_set(&hwc
->period_left
, hwc
->sample_period
);
6156 hwc
->last_period
= hwc
->sample_period
;
6157 event
->attr
.freq
= 0;
6162 * Software event: cpu wall time clock
6165 static void cpu_clock_event_update(struct perf_event
*event
)
6170 now
= local_clock();
6171 prev
= local64_xchg(&event
->hw
.prev_count
, now
);
6172 local64_add(now
- prev
, &event
->count
);
6175 static void cpu_clock_event_start(struct perf_event
*event
, int flags
)
6177 local64_set(&event
->hw
.prev_count
, local_clock());
6178 perf_swevent_start_hrtimer(event
);
6181 static void cpu_clock_event_stop(struct perf_event
*event
, int flags
)
6183 perf_swevent_cancel_hrtimer(event
);
6184 cpu_clock_event_update(event
);
6187 static int cpu_clock_event_add(struct perf_event
*event
, int flags
)
6189 if (flags
& PERF_EF_START
)
6190 cpu_clock_event_start(event
, flags
);
6195 static void cpu_clock_event_del(struct perf_event
*event
, int flags
)
6197 cpu_clock_event_stop(event
, flags
);
6200 static void cpu_clock_event_read(struct perf_event
*event
)
6202 cpu_clock_event_update(event
);
6205 static int cpu_clock_event_init(struct perf_event
*event
)
6207 if (event
->attr
.type
!= PERF_TYPE_SOFTWARE
)
6210 if (event
->attr
.config
!= PERF_COUNT_SW_CPU_CLOCK
)
6214 * no branch sampling for software events
6216 if (has_branch_stack(event
))
6219 perf_swevent_init_hrtimer(event
);
6224 static struct pmu perf_cpu_clock
= {
6225 .task_ctx_nr
= perf_sw_context
,
6227 .event_init
= cpu_clock_event_init
,
6228 .add
= cpu_clock_event_add
,
6229 .del
= cpu_clock_event_del
,
6230 .start
= cpu_clock_event_start
,
6231 .stop
= cpu_clock_event_stop
,
6232 .read
= cpu_clock_event_read
,
6234 .event_idx
= perf_swevent_event_idx
,
6238 * Software event: task time clock
6241 static void task_clock_event_update(struct perf_event
*event
, u64 now
)
6246 prev
= local64_xchg(&event
->hw
.prev_count
, now
);
6248 local64_add(delta
, &event
->count
);
6251 static void task_clock_event_start(struct perf_event
*event
, int flags
)
6253 local64_set(&event
->hw
.prev_count
, event
->ctx
->time
);
6254 perf_swevent_start_hrtimer(event
);
6257 static void task_clock_event_stop(struct perf_event
*event
, int flags
)
6259 perf_swevent_cancel_hrtimer(event
);
6260 task_clock_event_update(event
, event
->ctx
->time
);
6263 static int task_clock_event_add(struct perf_event
*event
, int flags
)
6265 if (flags
& PERF_EF_START
)
6266 task_clock_event_start(event
, flags
);
6271 static void task_clock_event_del(struct perf_event
*event
, int flags
)
6273 task_clock_event_stop(event
, PERF_EF_UPDATE
);
6276 static void task_clock_event_read(struct perf_event
*event
)
6278 u64 now
= perf_clock();
6279 u64 delta
= now
- event
->ctx
->timestamp
;
6280 u64 time
= event
->ctx
->time
+ delta
;
6282 task_clock_event_update(event
, time
);
6285 static int task_clock_event_init(struct perf_event
*event
)
6287 if (event
->attr
.type
!= PERF_TYPE_SOFTWARE
)
6290 if (event
->attr
.config
!= PERF_COUNT_SW_TASK_CLOCK
)
6294 * no branch sampling for software events
6296 if (has_branch_stack(event
))
6299 perf_swevent_init_hrtimer(event
);
6304 static struct pmu perf_task_clock
= {
6305 .task_ctx_nr
= perf_sw_context
,
6307 .event_init
= task_clock_event_init
,
6308 .add
= task_clock_event_add
,
6309 .del
= task_clock_event_del
,
6310 .start
= task_clock_event_start
,
6311 .stop
= task_clock_event_stop
,
6312 .read
= task_clock_event_read
,
6314 .event_idx
= perf_swevent_event_idx
,
6317 static void perf_pmu_nop_void(struct pmu
*pmu
)
6321 static int perf_pmu_nop_int(struct pmu
*pmu
)
6326 static void perf_pmu_start_txn(struct pmu
*pmu
)
6328 perf_pmu_disable(pmu
);
6331 static int perf_pmu_commit_txn(struct pmu
*pmu
)
6333 perf_pmu_enable(pmu
);
6337 static void perf_pmu_cancel_txn(struct pmu
*pmu
)
6339 perf_pmu_enable(pmu
);
6342 static int perf_event_idx_default(struct perf_event
*event
)
6344 return event
->hw
.idx
+ 1;
6348 * Ensures all contexts with the same task_ctx_nr have the same
6349 * pmu_cpu_context too.
6351 static struct perf_cpu_context __percpu
*find_pmu_context(int ctxn
)
6358 list_for_each_entry(pmu
, &pmus
, entry
) {
6359 if (pmu
->task_ctx_nr
== ctxn
)
6360 return pmu
->pmu_cpu_context
;
6366 static void update_pmu_context(struct pmu
*pmu
, struct pmu
*old_pmu
)
6370 for_each_possible_cpu(cpu
) {
6371 struct perf_cpu_context
*cpuctx
;
6373 cpuctx
= per_cpu_ptr(pmu
->pmu_cpu_context
, cpu
);
6375 if (cpuctx
->unique_pmu
== old_pmu
)
6376 cpuctx
->unique_pmu
= pmu
;
6380 static void free_pmu_context(struct pmu
*pmu
)
6384 mutex_lock(&pmus_lock
);
6386 * Like a real lame refcount.
6388 list_for_each_entry(i
, &pmus
, entry
) {
6389 if (i
->pmu_cpu_context
== pmu
->pmu_cpu_context
) {
6390 update_pmu_context(i
, pmu
);
6395 free_percpu(pmu
->pmu_cpu_context
);
6397 mutex_unlock(&pmus_lock
);
6399 static struct idr pmu_idr
;
6402 type_show(struct device
*dev
, struct device_attribute
*attr
, char *page
)
6404 struct pmu
*pmu
= dev_get_drvdata(dev
);
6406 return snprintf(page
, PAGE_SIZE
-1, "%d\n", pmu
->type
);
6408 static DEVICE_ATTR_RO(type
);
6411 perf_event_mux_interval_ms_show(struct device
*dev
,
6412 struct device_attribute
*attr
,
6415 struct pmu
*pmu
= dev_get_drvdata(dev
);
6417 return snprintf(page
, PAGE_SIZE
-1, "%d\n", pmu
->hrtimer_interval_ms
);
6421 perf_event_mux_interval_ms_store(struct device
*dev
,
6422 struct device_attribute
*attr
,
6423 const char *buf
, size_t count
)
6425 struct pmu
*pmu
= dev_get_drvdata(dev
);
6426 int timer
, cpu
, ret
;
6428 ret
= kstrtoint(buf
, 0, &timer
);
6435 /* same value, noting to do */
6436 if (timer
== pmu
->hrtimer_interval_ms
)
6439 pmu
->hrtimer_interval_ms
= timer
;
6441 /* update all cpuctx for this PMU */
6442 for_each_possible_cpu(cpu
) {
6443 struct perf_cpu_context
*cpuctx
;
6444 cpuctx
= per_cpu_ptr(pmu
->pmu_cpu_context
, cpu
);
6445 cpuctx
->hrtimer_interval
= ns_to_ktime(NSEC_PER_MSEC
* timer
);
6447 if (hrtimer_active(&cpuctx
->hrtimer
))
6448 hrtimer_forward_now(&cpuctx
->hrtimer
, cpuctx
->hrtimer_interval
);
6453 static DEVICE_ATTR_RW(perf_event_mux_interval_ms
);
6455 static struct attribute
*pmu_dev_attrs
[] = {
6456 &dev_attr_type
.attr
,
6457 &dev_attr_perf_event_mux_interval_ms
.attr
,
6460 ATTRIBUTE_GROUPS(pmu_dev
);
6462 static int pmu_bus_running
;
6463 static struct bus_type pmu_bus
= {
6464 .name
= "event_source",
6465 .dev_groups
= pmu_dev_groups
,
6468 static void pmu_dev_release(struct device
*dev
)
6473 static int pmu_dev_alloc(struct pmu
*pmu
)
6477 pmu
->dev
= kzalloc(sizeof(struct device
), GFP_KERNEL
);
6481 pmu
->dev
->groups
= pmu
->attr_groups
;
6482 device_initialize(pmu
->dev
);
6483 ret
= dev_set_name(pmu
->dev
, "%s", pmu
->name
);
6487 dev_set_drvdata(pmu
->dev
, pmu
);
6488 pmu
->dev
->bus
= &pmu_bus
;
6489 pmu
->dev
->release
= pmu_dev_release
;
6490 ret
= device_add(pmu
->dev
);
6498 put_device(pmu
->dev
);
6502 static struct lock_class_key cpuctx_mutex
;
6503 static struct lock_class_key cpuctx_lock
;
6505 int perf_pmu_register(struct pmu
*pmu
, const char *name
, int type
)
6509 mutex_lock(&pmus_lock
);
6511 pmu
->pmu_disable_count
= alloc_percpu(int);
6512 if (!pmu
->pmu_disable_count
)
6521 type
= idr_alloc(&pmu_idr
, pmu
, PERF_TYPE_MAX
, 0, GFP_KERNEL
);
6529 if (pmu_bus_running
) {
6530 ret
= pmu_dev_alloc(pmu
);
6536 pmu
->pmu_cpu_context
= find_pmu_context(pmu
->task_ctx_nr
);
6537 if (pmu
->pmu_cpu_context
)
6538 goto got_cpu_context
;
6541 pmu
->pmu_cpu_context
= alloc_percpu(struct perf_cpu_context
);
6542 if (!pmu
->pmu_cpu_context
)
6545 for_each_possible_cpu(cpu
) {
6546 struct perf_cpu_context
*cpuctx
;
6548 cpuctx
= per_cpu_ptr(pmu
->pmu_cpu_context
, cpu
);
6549 __perf_event_init_context(&cpuctx
->ctx
);
6550 lockdep_set_class(&cpuctx
->ctx
.mutex
, &cpuctx_mutex
);
6551 lockdep_set_class(&cpuctx
->ctx
.lock
, &cpuctx_lock
);
6552 cpuctx
->ctx
.type
= cpu_context
;
6553 cpuctx
->ctx
.pmu
= pmu
;
6555 __perf_cpu_hrtimer_init(cpuctx
, cpu
);
6557 INIT_LIST_HEAD(&cpuctx
->rotation_list
);
6558 cpuctx
->unique_pmu
= pmu
;
6562 if (!pmu
->start_txn
) {
6563 if (pmu
->pmu_enable
) {
6565 * If we have pmu_enable/pmu_disable calls, install
6566 * transaction stubs that use that to try and batch
6567 * hardware accesses.
6569 pmu
->start_txn
= perf_pmu_start_txn
;
6570 pmu
->commit_txn
= perf_pmu_commit_txn
;
6571 pmu
->cancel_txn
= perf_pmu_cancel_txn
;
6573 pmu
->start_txn
= perf_pmu_nop_void
;
6574 pmu
->commit_txn
= perf_pmu_nop_int
;
6575 pmu
->cancel_txn
= perf_pmu_nop_void
;
6579 if (!pmu
->pmu_enable
) {
6580 pmu
->pmu_enable
= perf_pmu_nop_void
;
6581 pmu
->pmu_disable
= perf_pmu_nop_void
;
6584 if (!pmu
->event_idx
)
6585 pmu
->event_idx
= perf_event_idx_default
;
6587 list_add_rcu(&pmu
->entry
, &pmus
);
6590 mutex_unlock(&pmus_lock
);
6595 device_del(pmu
->dev
);
6596 put_device(pmu
->dev
);
6599 if (pmu
->type
>= PERF_TYPE_MAX
)
6600 idr_remove(&pmu_idr
, pmu
->type
);
6603 free_percpu(pmu
->pmu_disable_count
);
6606 EXPORT_SYMBOL_GPL(perf_pmu_register
);
6608 void perf_pmu_unregister(struct pmu
*pmu
)
6610 mutex_lock(&pmus_lock
);
6611 list_del_rcu(&pmu
->entry
);
6612 mutex_unlock(&pmus_lock
);
6615 * We dereference the pmu list under both SRCU and regular RCU, so
6616 * synchronize against both of those.
6618 synchronize_srcu(&pmus_srcu
);
6621 free_percpu(pmu
->pmu_disable_count
);
6622 if (pmu
->type
>= PERF_TYPE_MAX
)
6623 idr_remove(&pmu_idr
, pmu
->type
);
6624 device_del(pmu
->dev
);
6625 put_device(pmu
->dev
);
6626 free_pmu_context(pmu
);
6628 EXPORT_SYMBOL_GPL(perf_pmu_unregister
);
6630 struct pmu
*perf_init_event(struct perf_event
*event
)
6632 struct pmu
*pmu
= NULL
;
6636 idx
= srcu_read_lock(&pmus_srcu
);
6639 pmu
= idr_find(&pmu_idr
, event
->attr
.type
);
6642 if (!try_module_get(pmu
->module
)) {
6643 pmu
= ERR_PTR(-ENODEV
);
6647 ret
= pmu
->event_init(event
);
6653 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
6654 if (!try_module_get(pmu
->module
)) {
6655 pmu
= ERR_PTR(-ENODEV
);
6659 ret
= pmu
->event_init(event
);
6663 if (ret
!= -ENOENT
) {
6668 pmu
= ERR_PTR(-ENOENT
);
6670 srcu_read_unlock(&pmus_srcu
, idx
);
6675 static void account_event_cpu(struct perf_event
*event
, int cpu
)
6680 if (has_branch_stack(event
)) {
6681 if (!(event
->attach_state
& PERF_ATTACH_TASK
))
6682 atomic_inc(&per_cpu(perf_branch_stack_events
, cpu
));
6684 if (is_cgroup_event(event
))
6685 atomic_inc(&per_cpu(perf_cgroup_events
, cpu
));
6688 static void account_event(struct perf_event
*event
)
6693 if (event
->attach_state
& PERF_ATTACH_TASK
)
6694 static_key_slow_inc(&perf_sched_events
.key
);
6695 if (event
->attr
.mmap
|| event
->attr
.mmap_data
)
6696 atomic_inc(&nr_mmap_events
);
6697 if (event
->attr
.comm
)
6698 atomic_inc(&nr_comm_events
);
6699 if (event
->attr
.task
)
6700 atomic_inc(&nr_task_events
);
6701 if (event
->attr
.freq
) {
6702 if (atomic_inc_return(&nr_freq_events
) == 1)
6703 tick_nohz_full_kick_all();
6705 if (has_branch_stack(event
))
6706 static_key_slow_inc(&perf_sched_events
.key
);
6707 if (is_cgroup_event(event
))
6708 static_key_slow_inc(&perf_sched_events
.key
);
6710 account_event_cpu(event
, event
->cpu
);
6714 * Allocate and initialize a event structure
6716 static struct perf_event
*
6717 perf_event_alloc(struct perf_event_attr
*attr
, int cpu
,
6718 struct task_struct
*task
,
6719 struct perf_event
*group_leader
,
6720 struct perf_event
*parent_event
,
6721 perf_overflow_handler_t overflow_handler
,
6725 struct perf_event
*event
;
6726 struct hw_perf_event
*hwc
;
6729 if ((unsigned)cpu
>= nr_cpu_ids
) {
6730 if (!task
|| cpu
!= -1)
6731 return ERR_PTR(-EINVAL
);
6734 event
= kzalloc(sizeof(*event
), GFP_KERNEL
);
6736 return ERR_PTR(-ENOMEM
);
6739 * Single events are their own group leaders, with an
6740 * empty sibling list:
6743 group_leader
= event
;
6745 mutex_init(&event
->child_mutex
);
6746 INIT_LIST_HEAD(&event
->child_list
);
6748 INIT_LIST_HEAD(&event
->group_entry
);
6749 INIT_LIST_HEAD(&event
->event_entry
);
6750 INIT_LIST_HEAD(&event
->sibling_list
);
6751 INIT_LIST_HEAD(&event
->rb_entry
);
6752 INIT_LIST_HEAD(&event
->active_entry
);
6753 INIT_HLIST_NODE(&event
->hlist_entry
);
6756 init_waitqueue_head(&event
->waitq
);
6757 init_irq_work(&event
->pending
, perf_pending_event
);
6759 mutex_init(&event
->mmap_mutex
);
6761 atomic_long_set(&event
->refcount
, 1);
6763 event
->attr
= *attr
;
6764 event
->group_leader
= group_leader
;
6768 event
->parent
= parent_event
;
6770 event
->ns
= get_pid_ns(task_active_pid_ns(current
));
6771 event
->id
= atomic64_inc_return(&perf_event_id
);
6773 event
->state
= PERF_EVENT_STATE_INACTIVE
;
6776 event
->attach_state
= PERF_ATTACH_TASK
;
6778 if (attr
->type
== PERF_TYPE_TRACEPOINT
)
6779 event
->hw
.tp_target
= task
;
6780 #ifdef CONFIG_HAVE_HW_BREAKPOINT
6782 * hw_breakpoint is a bit difficult here..
6784 else if (attr
->type
== PERF_TYPE_BREAKPOINT
)
6785 event
->hw
.bp_target
= task
;
6789 if (!overflow_handler
&& parent_event
) {
6790 overflow_handler
= parent_event
->overflow_handler
;
6791 context
= parent_event
->overflow_handler_context
;
6794 event
->overflow_handler
= overflow_handler
;
6795 event
->overflow_handler_context
= context
;
6797 perf_event__state_init(event
);
6802 hwc
->sample_period
= attr
->sample_period
;
6803 if (attr
->freq
&& attr
->sample_freq
)
6804 hwc
->sample_period
= 1;
6805 hwc
->last_period
= hwc
->sample_period
;
6807 local64_set(&hwc
->period_left
, hwc
->sample_period
);
6810 * we currently do not support PERF_FORMAT_GROUP on inherited events
6812 if (attr
->inherit
&& (attr
->read_format
& PERF_FORMAT_GROUP
))
6815 pmu
= perf_init_event(event
);
6818 else if (IS_ERR(pmu
)) {
6823 if (!event
->parent
) {
6824 if (event
->attr
.sample_type
& PERF_SAMPLE_CALLCHAIN
) {
6825 err
= get_callchain_buffers();
6835 event
->destroy(event
);
6836 module_put(pmu
->module
);
6839 put_pid_ns(event
->ns
);
6842 return ERR_PTR(err
);
6845 static int perf_copy_attr(struct perf_event_attr __user
*uattr
,
6846 struct perf_event_attr
*attr
)
6851 if (!access_ok(VERIFY_WRITE
, uattr
, PERF_ATTR_SIZE_VER0
))
6855 * zero the full structure, so that a short copy will be nice.
6857 memset(attr
, 0, sizeof(*attr
));
6859 ret
= get_user(size
, &uattr
->size
);
6863 if (size
> PAGE_SIZE
) /* silly large */
6866 if (!size
) /* abi compat */
6867 size
= PERF_ATTR_SIZE_VER0
;
6869 if (size
< PERF_ATTR_SIZE_VER0
)
6873 * If we're handed a bigger struct than we know of,
6874 * ensure all the unknown bits are 0 - i.e. new
6875 * user-space does not rely on any kernel feature
6876 * extensions we dont know about yet.
6878 if (size
> sizeof(*attr
)) {
6879 unsigned char __user
*addr
;
6880 unsigned char __user
*end
;
6883 addr
= (void __user
*)uattr
+ sizeof(*attr
);
6884 end
= (void __user
*)uattr
+ size
;
6886 for (; addr
< end
; addr
++) {
6887 ret
= get_user(val
, addr
);
6893 size
= sizeof(*attr
);
6896 ret
= copy_from_user(attr
, uattr
, size
);
6900 /* disabled for now */
6904 if (attr
->__reserved_1
)
6907 if (attr
->sample_type
& ~(PERF_SAMPLE_MAX
-1))
6910 if (attr
->read_format
& ~(PERF_FORMAT_MAX
-1))
6913 if (attr
->sample_type
& PERF_SAMPLE_BRANCH_STACK
) {
6914 u64 mask
= attr
->branch_sample_type
;
6916 /* only using defined bits */
6917 if (mask
& ~(PERF_SAMPLE_BRANCH_MAX
-1))
6920 /* at least one branch bit must be set */
6921 if (!(mask
& ~PERF_SAMPLE_BRANCH_PLM_ALL
))
6924 /* propagate priv level, when not set for branch */
6925 if (!(mask
& PERF_SAMPLE_BRANCH_PLM_ALL
)) {
6927 /* exclude_kernel checked on syscall entry */
6928 if (!attr
->exclude_kernel
)
6929 mask
|= PERF_SAMPLE_BRANCH_KERNEL
;
6931 if (!attr
->exclude_user
)
6932 mask
|= PERF_SAMPLE_BRANCH_USER
;
6934 if (!attr
->exclude_hv
)
6935 mask
|= PERF_SAMPLE_BRANCH_HV
;
6937 * adjust user setting (for HW filter setup)
6939 attr
->branch_sample_type
= mask
;
6941 /* privileged levels capture (kernel, hv): check permissions */
6942 if ((mask
& PERF_SAMPLE_BRANCH_PERM_PLM
)
6943 && perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN
))
6947 if (attr
->sample_type
& PERF_SAMPLE_REGS_USER
) {
6948 ret
= perf_reg_validate(attr
->sample_regs_user
);
6953 if (attr
->sample_type
& PERF_SAMPLE_STACK_USER
) {
6954 if (!arch_perf_have_user_stack_dump())
6958 * We have __u32 type for the size, but so far
6959 * we can only use __u16 as maximum due to the
6960 * __u16 sample size limit.
6962 if (attr
->sample_stack_user
>= USHRT_MAX
)
6964 else if (!IS_ALIGNED(attr
->sample_stack_user
, sizeof(u64
)))
6972 put_user(sizeof(*attr
), &uattr
->size
);
6978 perf_event_set_output(struct perf_event
*event
, struct perf_event
*output_event
)
6980 struct ring_buffer
*rb
= NULL
;
6986 /* don't allow circular references */
6987 if (event
== output_event
)
6991 * Don't allow cross-cpu buffers
6993 if (output_event
->cpu
!= event
->cpu
)
6997 * If its not a per-cpu rb, it must be the same task.
6999 if (output_event
->cpu
== -1 && output_event
->ctx
!= event
->ctx
)
7003 mutex_lock(&event
->mmap_mutex
);
7004 /* Can't redirect output if we've got an active mmap() */
7005 if (atomic_read(&event
->mmap_count
))
7009 /* get the rb we want to redirect to */
7010 rb
= ring_buffer_get(output_event
);
7015 ring_buffer_attach(event
, rb
);
7019 mutex_unlock(&event
->mmap_mutex
);
7026 * sys_perf_event_open - open a performance event, associate it to a task/cpu
7028 * @attr_uptr: event_id type attributes for monitoring/sampling
7031 * @group_fd: group leader event fd
7033 SYSCALL_DEFINE5(perf_event_open
,
7034 struct perf_event_attr __user
*, attr_uptr
,
7035 pid_t
, pid
, int, cpu
, int, group_fd
, unsigned long, flags
)
7037 struct perf_event
*group_leader
= NULL
, *output_event
= NULL
;
7038 struct perf_event
*event
, *sibling
;
7039 struct perf_event_attr attr
;
7040 struct perf_event_context
*ctx
;
7041 struct file
*event_file
= NULL
;
7042 struct fd group
= {NULL
, 0};
7043 struct task_struct
*task
= NULL
;
7048 int f_flags
= O_RDWR
;
7050 /* for future expandability... */
7051 if (flags
& ~PERF_FLAG_ALL
)
7054 err
= perf_copy_attr(attr_uptr
, &attr
);
7058 if (!attr
.exclude_kernel
) {
7059 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN
))
7064 if (attr
.sample_freq
> sysctl_perf_event_sample_rate
)
7067 if (attr
.sample_period
& (1ULL << 63))
7072 * In cgroup mode, the pid argument is used to pass the fd
7073 * opened to the cgroup directory in cgroupfs. The cpu argument
7074 * designates the cpu on which to monitor threads from that
7077 if ((flags
& PERF_FLAG_PID_CGROUP
) && (pid
== -1 || cpu
== -1))
7080 if (flags
& PERF_FLAG_FD_CLOEXEC
)
7081 f_flags
|= O_CLOEXEC
;
7083 event_fd
= get_unused_fd_flags(f_flags
);
7087 if (group_fd
!= -1) {
7088 err
= perf_fget_light(group_fd
, &group
);
7091 group_leader
= group
.file
->private_data
;
7092 if (flags
& PERF_FLAG_FD_OUTPUT
)
7093 output_event
= group_leader
;
7094 if (flags
& PERF_FLAG_FD_NO_GROUP
)
7095 group_leader
= NULL
;
7098 if (pid
!= -1 && !(flags
& PERF_FLAG_PID_CGROUP
)) {
7099 task
= find_lively_task_by_vpid(pid
);
7101 err
= PTR_ERR(task
);
7106 if (task
&& group_leader
&&
7107 group_leader
->attr
.inherit
!= attr
.inherit
) {
7114 event
= perf_event_alloc(&attr
, cpu
, task
, group_leader
, NULL
,
7116 if (IS_ERR(event
)) {
7117 err
= PTR_ERR(event
);
7121 if (flags
& PERF_FLAG_PID_CGROUP
) {
7122 err
= perf_cgroup_connect(pid
, event
, &attr
, group_leader
);
7124 __free_event(event
);
7129 if (is_sampling_event(event
)) {
7130 if (event
->pmu
->capabilities
& PERF_PMU_CAP_NO_INTERRUPT
) {
7136 account_event(event
);
7139 * Special case software events and allow them to be part of
7140 * any hardware group.
7145 (is_software_event(event
) != is_software_event(group_leader
))) {
7146 if (is_software_event(event
)) {
7148 * If event and group_leader are not both a software
7149 * event, and event is, then group leader is not.
7151 * Allow the addition of software events to !software
7152 * groups, this is safe because software events never
7155 pmu
= group_leader
->pmu
;
7156 } else if (is_software_event(group_leader
) &&
7157 (group_leader
->group_flags
& PERF_GROUP_SOFTWARE
)) {
7159 * In case the group is a pure software group, and we
7160 * try to add a hardware event, move the whole group to
7161 * the hardware context.
7168 * Get the target context (task or percpu):
7170 ctx
= find_get_context(pmu
, task
, event
->cpu
);
7177 put_task_struct(task
);
7182 * Look up the group leader (we will attach this event to it):
7188 * Do not allow a recursive hierarchy (this new sibling
7189 * becoming part of another group-sibling):
7191 if (group_leader
->group_leader
!= group_leader
)
7194 * Do not allow to attach to a group in a different
7195 * task or CPU context:
7198 if (group_leader
->ctx
->type
!= ctx
->type
)
7201 if (group_leader
->ctx
!= ctx
)
7206 * Only a group leader can be exclusive or pinned
7208 if (attr
.exclusive
|| attr
.pinned
)
7213 err
= perf_event_set_output(event
, output_event
);
7218 event_file
= anon_inode_getfile("[perf_event]", &perf_fops
, event
,
7220 if (IS_ERR(event_file
)) {
7221 err
= PTR_ERR(event_file
);
7226 struct perf_event_context
*gctx
= group_leader
->ctx
;
7228 mutex_lock(&gctx
->mutex
);
7229 perf_remove_from_context(group_leader
, false);
7232 * Removing from the context ends up with disabled
7233 * event. What we want here is event in the initial
7234 * startup state, ready to be add into new context.
7236 perf_event__state_init(group_leader
);
7237 list_for_each_entry(sibling
, &group_leader
->sibling_list
,
7239 perf_remove_from_context(sibling
, false);
7240 perf_event__state_init(sibling
);
7243 mutex_unlock(&gctx
->mutex
);
7247 WARN_ON_ONCE(ctx
->parent_ctx
);
7248 mutex_lock(&ctx
->mutex
);
7252 perf_install_in_context(ctx
, group_leader
, event
->cpu
);
7254 list_for_each_entry(sibling
, &group_leader
->sibling_list
,
7256 perf_install_in_context(ctx
, sibling
, event
->cpu
);
7261 perf_install_in_context(ctx
, event
, event
->cpu
);
7262 perf_unpin_context(ctx
);
7263 mutex_unlock(&ctx
->mutex
);
7267 event
->owner
= current
;
7269 mutex_lock(¤t
->perf_event_mutex
);
7270 list_add_tail(&event
->owner_entry
, ¤t
->perf_event_list
);
7271 mutex_unlock(¤t
->perf_event_mutex
);
7274 * Precalculate sample_data sizes
7276 perf_event__header_size(event
);
7277 perf_event__id_header_size(event
);
7280 * Drop the reference on the group_event after placing the
7281 * new event on the sibling_list. This ensures destruction
7282 * of the group leader will find the pointer to itself in
7283 * perf_group_detach().
7286 fd_install(event_fd
, event_file
);
7290 perf_unpin_context(ctx
);
7298 put_task_struct(task
);
7302 put_unused_fd(event_fd
);
7307 * perf_event_create_kernel_counter
7309 * @attr: attributes of the counter to create
7310 * @cpu: cpu in which the counter is bound
7311 * @task: task to profile (NULL for percpu)
7314 perf_event_create_kernel_counter(struct perf_event_attr
*attr
, int cpu
,
7315 struct task_struct
*task
,
7316 perf_overflow_handler_t overflow_handler
,
7319 struct perf_event_context
*ctx
;
7320 struct perf_event
*event
;
7324 * Get the target context (task or percpu):
7327 event
= perf_event_alloc(attr
, cpu
, task
, NULL
, NULL
,
7328 overflow_handler
, context
);
7329 if (IS_ERR(event
)) {
7330 err
= PTR_ERR(event
);
7334 account_event(event
);
7336 ctx
= find_get_context(event
->pmu
, task
, cpu
);
7342 WARN_ON_ONCE(ctx
->parent_ctx
);
7343 mutex_lock(&ctx
->mutex
);
7344 perf_install_in_context(ctx
, event
, cpu
);
7345 perf_unpin_context(ctx
);
7346 mutex_unlock(&ctx
->mutex
);
7353 return ERR_PTR(err
);
7355 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter
);
7357 void perf_pmu_migrate_context(struct pmu
*pmu
, int src_cpu
, int dst_cpu
)
7359 struct perf_event_context
*src_ctx
;
7360 struct perf_event_context
*dst_ctx
;
7361 struct perf_event
*event
, *tmp
;
7364 src_ctx
= &per_cpu_ptr(pmu
->pmu_cpu_context
, src_cpu
)->ctx
;
7365 dst_ctx
= &per_cpu_ptr(pmu
->pmu_cpu_context
, dst_cpu
)->ctx
;
7367 mutex_lock(&src_ctx
->mutex
);
7368 list_for_each_entry_safe(event
, tmp
, &src_ctx
->event_list
,
7370 perf_remove_from_context(event
, false);
7371 unaccount_event_cpu(event
, src_cpu
);
7373 list_add(&event
->migrate_entry
, &events
);
7375 mutex_unlock(&src_ctx
->mutex
);
7379 mutex_lock(&dst_ctx
->mutex
);
7380 list_for_each_entry_safe(event
, tmp
, &events
, migrate_entry
) {
7381 list_del(&event
->migrate_entry
);
7382 if (event
->state
>= PERF_EVENT_STATE_OFF
)
7383 event
->state
= PERF_EVENT_STATE_INACTIVE
;
7384 account_event_cpu(event
, dst_cpu
);
7385 perf_install_in_context(dst_ctx
, event
, dst_cpu
);
7388 mutex_unlock(&dst_ctx
->mutex
);
7390 EXPORT_SYMBOL_GPL(perf_pmu_migrate_context
);
7392 static void sync_child_event(struct perf_event
*child_event
,
7393 struct task_struct
*child
)
7395 struct perf_event
*parent_event
= child_event
->parent
;
7398 if (child_event
->attr
.inherit_stat
)
7399 perf_event_read_event(child_event
, child
);
7401 child_val
= perf_event_count(child_event
);
7404 * Add back the child's count to the parent's count:
7406 atomic64_add(child_val
, &parent_event
->child_count
);
7407 atomic64_add(child_event
->total_time_enabled
,
7408 &parent_event
->child_total_time_enabled
);
7409 atomic64_add(child_event
->total_time_running
,
7410 &parent_event
->child_total_time_running
);
7413 * Remove this event from the parent's list
7415 WARN_ON_ONCE(parent_event
->ctx
->parent_ctx
);
7416 mutex_lock(&parent_event
->child_mutex
);
7417 list_del_init(&child_event
->child_list
);
7418 mutex_unlock(&parent_event
->child_mutex
);
7421 * Release the parent event, if this was the last
7424 put_event(parent_event
);
7428 __perf_event_exit_task(struct perf_event
*child_event
,
7429 struct perf_event_context
*child_ctx
,
7430 struct task_struct
*child
)
7432 perf_remove_from_context(child_event
, true);
7435 * It can happen that the parent exits first, and has events
7436 * that are still around due to the child reference. These
7437 * events need to be zapped.
7439 if (child_event
->parent
) {
7440 sync_child_event(child_event
, child
);
7441 free_event(child_event
);
7445 static void perf_event_exit_task_context(struct task_struct
*child
, int ctxn
)
7447 struct perf_event
*child_event
, *next
;
7448 struct perf_event_context
*child_ctx
;
7449 unsigned long flags
;
7451 if (likely(!child
->perf_event_ctxp
[ctxn
])) {
7452 perf_event_task(child
, NULL
, 0);
7456 local_irq_save(flags
);
7458 * We can't reschedule here because interrupts are disabled,
7459 * and either child is current or it is a task that can't be
7460 * scheduled, so we are now safe from rescheduling changing
7463 child_ctx
= rcu_dereference_raw(child
->perf_event_ctxp
[ctxn
]);
7466 * Take the context lock here so that if find_get_context is
7467 * reading child->perf_event_ctxp, we wait until it has
7468 * incremented the context's refcount before we do put_ctx below.
7470 raw_spin_lock(&child_ctx
->lock
);
7471 task_ctx_sched_out(child_ctx
);
7472 child
->perf_event_ctxp
[ctxn
] = NULL
;
7474 * If this context is a clone; unclone it so it can't get
7475 * swapped to another process while we're removing all
7476 * the events from it.
7478 unclone_ctx(child_ctx
);
7479 update_context_time(child_ctx
);
7480 raw_spin_unlock_irqrestore(&child_ctx
->lock
, flags
);
7483 * Report the task dead after unscheduling the events so that we
7484 * won't get any samples after PERF_RECORD_EXIT. We can however still
7485 * get a few PERF_RECORD_READ events.
7487 perf_event_task(child
, child_ctx
, 0);
7490 * We can recurse on the same lock type through:
7492 * __perf_event_exit_task()
7493 * sync_child_event()
7495 * mutex_lock(&ctx->mutex)
7497 * But since its the parent context it won't be the same instance.
7499 mutex_lock(&child_ctx
->mutex
);
7501 list_for_each_entry_safe(child_event
, next
, &child_ctx
->event_list
, event_entry
)
7502 __perf_event_exit_task(child_event
, child_ctx
, child
);
7504 mutex_unlock(&child_ctx
->mutex
);
7510 * When a child task exits, feed back event values to parent events.
7512 void perf_event_exit_task(struct task_struct
*child
)
7514 struct perf_event
*event
, *tmp
;
7517 mutex_lock(&child
->perf_event_mutex
);
7518 list_for_each_entry_safe(event
, tmp
, &child
->perf_event_list
,
7520 list_del_init(&event
->owner_entry
);
7523 * Ensure the list deletion is visible before we clear
7524 * the owner, closes a race against perf_release() where
7525 * we need to serialize on the owner->perf_event_mutex.
7528 event
->owner
= NULL
;
7530 mutex_unlock(&child
->perf_event_mutex
);
7532 for_each_task_context_nr(ctxn
)
7533 perf_event_exit_task_context(child
, ctxn
);
7536 static void perf_free_event(struct perf_event
*event
,
7537 struct perf_event_context
*ctx
)
7539 struct perf_event
*parent
= event
->parent
;
7541 if (WARN_ON_ONCE(!parent
))
7544 mutex_lock(&parent
->child_mutex
);
7545 list_del_init(&event
->child_list
);
7546 mutex_unlock(&parent
->child_mutex
);
7550 perf_group_detach(event
);
7551 list_del_event(event
, ctx
);
7556 * free an unexposed, unused context as created by inheritance by
7557 * perf_event_init_task below, used by fork() in case of fail.
7559 void perf_event_free_task(struct task_struct
*task
)
7561 struct perf_event_context
*ctx
;
7562 struct perf_event
*event
, *tmp
;
7565 for_each_task_context_nr(ctxn
) {
7566 ctx
= task
->perf_event_ctxp
[ctxn
];
7570 mutex_lock(&ctx
->mutex
);
7572 list_for_each_entry_safe(event
, tmp
, &ctx
->pinned_groups
,
7574 perf_free_event(event
, ctx
);
7576 list_for_each_entry_safe(event
, tmp
, &ctx
->flexible_groups
,
7578 perf_free_event(event
, ctx
);
7580 if (!list_empty(&ctx
->pinned_groups
) ||
7581 !list_empty(&ctx
->flexible_groups
))
7584 mutex_unlock(&ctx
->mutex
);
7590 void perf_event_delayed_put(struct task_struct
*task
)
7594 for_each_task_context_nr(ctxn
)
7595 WARN_ON_ONCE(task
->perf_event_ctxp
[ctxn
]);
7599 * inherit a event from parent task to child task:
7601 static struct perf_event
*
7602 inherit_event(struct perf_event
*parent_event
,
7603 struct task_struct
*parent
,
7604 struct perf_event_context
*parent_ctx
,
7605 struct task_struct
*child
,
7606 struct perf_event
*group_leader
,
7607 struct perf_event_context
*child_ctx
)
7609 struct perf_event
*child_event
;
7610 unsigned long flags
;
7613 * Instead of creating recursive hierarchies of events,
7614 * we link inherited events back to the original parent,
7615 * which has a filp for sure, which we use as the reference
7618 if (parent_event
->parent
)
7619 parent_event
= parent_event
->parent
;
7621 child_event
= perf_event_alloc(&parent_event
->attr
,
7624 group_leader
, parent_event
,
7626 if (IS_ERR(child_event
))
7629 if (!atomic_long_inc_not_zero(&parent_event
->refcount
)) {
7630 free_event(child_event
);
7637 * Make the child state follow the state of the parent event,
7638 * not its attr.disabled bit. We hold the parent's mutex,
7639 * so we won't race with perf_event_{en, dis}able_family.
7641 if (parent_event
->state
>= PERF_EVENT_STATE_INACTIVE
)
7642 child_event
->state
= PERF_EVENT_STATE_INACTIVE
;
7644 child_event
->state
= PERF_EVENT_STATE_OFF
;
7646 if (parent_event
->attr
.freq
) {
7647 u64 sample_period
= parent_event
->hw
.sample_period
;
7648 struct hw_perf_event
*hwc
= &child_event
->hw
;
7650 hwc
->sample_period
= sample_period
;
7651 hwc
->last_period
= sample_period
;
7653 local64_set(&hwc
->period_left
, sample_period
);
7656 child_event
->ctx
= child_ctx
;
7657 child_event
->overflow_handler
= parent_event
->overflow_handler
;
7658 child_event
->overflow_handler_context
7659 = parent_event
->overflow_handler_context
;
7662 * Precalculate sample_data sizes
7664 perf_event__header_size(child_event
);
7665 perf_event__id_header_size(child_event
);
7668 * Link it up in the child's context:
7670 raw_spin_lock_irqsave(&child_ctx
->lock
, flags
);
7671 add_event_to_ctx(child_event
, child_ctx
);
7672 raw_spin_unlock_irqrestore(&child_ctx
->lock
, flags
);
7675 * Link this into the parent event's child list
7677 WARN_ON_ONCE(parent_event
->ctx
->parent_ctx
);
7678 mutex_lock(&parent_event
->child_mutex
);
7679 list_add_tail(&child_event
->child_list
, &parent_event
->child_list
);
7680 mutex_unlock(&parent_event
->child_mutex
);
7685 static int inherit_group(struct perf_event
*parent_event
,
7686 struct task_struct
*parent
,
7687 struct perf_event_context
*parent_ctx
,
7688 struct task_struct
*child
,
7689 struct perf_event_context
*child_ctx
)
7691 struct perf_event
*leader
;
7692 struct perf_event
*sub
;
7693 struct perf_event
*child_ctr
;
7695 leader
= inherit_event(parent_event
, parent
, parent_ctx
,
7696 child
, NULL
, child_ctx
);
7698 return PTR_ERR(leader
);
7699 list_for_each_entry(sub
, &parent_event
->sibling_list
, group_entry
) {
7700 child_ctr
= inherit_event(sub
, parent
, parent_ctx
,
7701 child
, leader
, child_ctx
);
7702 if (IS_ERR(child_ctr
))
7703 return PTR_ERR(child_ctr
);
7709 inherit_task_group(struct perf_event
*event
, struct task_struct
*parent
,
7710 struct perf_event_context
*parent_ctx
,
7711 struct task_struct
*child
, int ctxn
,
7715 struct perf_event_context
*child_ctx
;
7717 if (!event
->attr
.inherit
) {
7722 child_ctx
= child
->perf_event_ctxp
[ctxn
];
7725 * This is executed from the parent task context, so
7726 * inherit events that have been marked for cloning.
7727 * First allocate and initialize a context for the
7731 child_ctx
= alloc_perf_context(parent_ctx
->pmu
, child
);
7735 child
->perf_event_ctxp
[ctxn
] = child_ctx
;
7738 ret
= inherit_group(event
, parent
, parent_ctx
,
7748 * Initialize the perf_event context in task_struct
7750 int perf_event_init_context(struct task_struct
*child
, int ctxn
)
7752 struct perf_event_context
*child_ctx
, *parent_ctx
;
7753 struct perf_event_context
*cloned_ctx
;
7754 struct perf_event
*event
;
7755 struct task_struct
*parent
= current
;
7756 int inherited_all
= 1;
7757 unsigned long flags
;
7760 if (likely(!parent
->perf_event_ctxp
[ctxn
]))
7764 * If the parent's context is a clone, pin it so it won't get
7767 parent_ctx
= perf_pin_task_context(parent
, ctxn
);
7772 * No need to check if parent_ctx != NULL here; since we saw
7773 * it non-NULL earlier, the only reason for it to become NULL
7774 * is if we exit, and since we're currently in the middle of
7775 * a fork we can't be exiting at the same time.
7779 * Lock the parent list. No need to lock the child - not PID
7780 * hashed yet and not running, so nobody can access it.
7782 mutex_lock(&parent_ctx
->mutex
);
7785 * We dont have to disable NMIs - we are only looking at
7786 * the list, not manipulating it:
7788 list_for_each_entry(event
, &parent_ctx
->pinned_groups
, group_entry
) {
7789 ret
= inherit_task_group(event
, parent
, parent_ctx
,
7790 child
, ctxn
, &inherited_all
);
7796 * We can't hold ctx->lock when iterating the ->flexible_group list due
7797 * to allocations, but we need to prevent rotation because
7798 * rotate_ctx() will change the list from interrupt context.
7800 raw_spin_lock_irqsave(&parent_ctx
->lock
, flags
);
7801 parent_ctx
->rotate_disable
= 1;
7802 raw_spin_unlock_irqrestore(&parent_ctx
->lock
, flags
);
7804 list_for_each_entry(event
, &parent_ctx
->flexible_groups
, group_entry
) {
7805 ret
= inherit_task_group(event
, parent
, parent_ctx
,
7806 child
, ctxn
, &inherited_all
);
7811 raw_spin_lock_irqsave(&parent_ctx
->lock
, flags
);
7812 parent_ctx
->rotate_disable
= 0;
7814 child_ctx
= child
->perf_event_ctxp
[ctxn
];
7816 if (child_ctx
&& inherited_all
) {
7818 * Mark the child context as a clone of the parent
7819 * context, or of whatever the parent is a clone of.
7821 * Note that if the parent is a clone, the holding of
7822 * parent_ctx->lock avoids it from being uncloned.
7824 cloned_ctx
= parent_ctx
->parent_ctx
;
7826 child_ctx
->parent_ctx
= cloned_ctx
;
7827 child_ctx
->parent_gen
= parent_ctx
->parent_gen
;
7829 child_ctx
->parent_ctx
= parent_ctx
;
7830 child_ctx
->parent_gen
= parent_ctx
->generation
;
7832 get_ctx(child_ctx
->parent_ctx
);
7835 raw_spin_unlock_irqrestore(&parent_ctx
->lock
, flags
);
7836 mutex_unlock(&parent_ctx
->mutex
);
7838 perf_unpin_context(parent_ctx
);
7839 put_ctx(parent_ctx
);
7845 * Initialize the perf_event context in task_struct
7847 int perf_event_init_task(struct task_struct
*child
)
7851 memset(child
->perf_event_ctxp
, 0, sizeof(child
->perf_event_ctxp
));
7852 mutex_init(&child
->perf_event_mutex
);
7853 INIT_LIST_HEAD(&child
->perf_event_list
);
7855 for_each_task_context_nr(ctxn
) {
7856 ret
= perf_event_init_context(child
, ctxn
);
7864 static void __init
perf_event_init_all_cpus(void)
7866 struct swevent_htable
*swhash
;
7869 for_each_possible_cpu(cpu
) {
7870 swhash
= &per_cpu(swevent_htable
, cpu
);
7871 mutex_init(&swhash
->hlist_mutex
);
7872 INIT_LIST_HEAD(&per_cpu(rotation_list
, cpu
));
7876 static void perf_event_init_cpu(int cpu
)
7878 struct swevent_htable
*swhash
= &per_cpu(swevent_htable
, cpu
);
7880 mutex_lock(&swhash
->hlist_mutex
);
7881 swhash
->online
= true;
7882 if (swhash
->hlist_refcount
> 0) {
7883 struct swevent_hlist
*hlist
;
7885 hlist
= kzalloc_node(sizeof(*hlist
), GFP_KERNEL
, cpu_to_node(cpu
));
7887 rcu_assign_pointer(swhash
->swevent_hlist
, hlist
);
7889 mutex_unlock(&swhash
->hlist_mutex
);
7892 #if defined CONFIG_HOTPLUG_CPU || defined CONFIG_KEXEC
7893 static void perf_pmu_rotate_stop(struct pmu
*pmu
)
7895 struct perf_cpu_context
*cpuctx
= this_cpu_ptr(pmu
->pmu_cpu_context
);
7897 WARN_ON(!irqs_disabled());
7899 list_del_init(&cpuctx
->rotation_list
);
7902 static void __perf_event_exit_context(void *__info
)
7904 struct remove_event re
= { .detach_group
= false };
7905 struct perf_event_context
*ctx
= __info
;
7907 perf_pmu_rotate_stop(ctx
->pmu
);
7910 list_for_each_entry_rcu(re
.event
, &ctx
->event_list
, event_entry
)
7911 __perf_remove_from_context(&re
);
7915 static void perf_event_exit_cpu_context(int cpu
)
7917 struct perf_event_context
*ctx
;
7921 idx
= srcu_read_lock(&pmus_srcu
);
7922 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
7923 ctx
= &per_cpu_ptr(pmu
->pmu_cpu_context
, cpu
)->ctx
;
7925 mutex_lock(&ctx
->mutex
);
7926 smp_call_function_single(cpu
, __perf_event_exit_context
, ctx
, 1);
7927 mutex_unlock(&ctx
->mutex
);
7929 srcu_read_unlock(&pmus_srcu
, idx
);
7932 static void perf_event_exit_cpu(int cpu
)
7934 struct swevent_htable
*swhash
= &per_cpu(swevent_htable
, cpu
);
7936 perf_event_exit_cpu_context(cpu
);
7938 mutex_lock(&swhash
->hlist_mutex
);
7939 swhash
->online
= false;
7940 swevent_hlist_release(swhash
);
7941 mutex_unlock(&swhash
->hlist_mutex
);
7944 static inline void perf_event_exit_cpu(int cpu
) { }
7948 perf_reboot(struct notifier_block
*notifier
, unsigned long val
, void *v
)
7952 for_each_online_cpu(cpu
)
7953 perf_event_exit_cpu(cpu
);
7959 * Run the perf reboot notifier at the very last possible moment so that
7960 * the generic watchdog code runs as long as possible.
7962 static struct notifier_block perf_reboot_notifier
= {
7963 .notifier_call
= perf_reboot
,
7964 .priority
= INT_MIN
,
7968 perf_cpu_notify(struct notifier_block
*self
, unsigned long action
, void *hcpu
)
7970 unsigned int cpu
= (long)hcpu
;
7972 switch (action
& ~CPU_TASKS_FROZEN
) {
7974 case CPU_UP_PREPARE
:
7975 case CPU_DOWN_FAILED
:
7976 perf_event_init_cpu(cpu
);
7979 case CPU_UP_CANCELED
:
7980 case CPU_DOWN_PREPARE
:
7981 perf_event_exit_cpu(cpu
);
7990 void __init
perf_event_init(void)
7996 perf_event_init_all_cpus();
7997 init_srcu_struct(&pmus_srcu
);
7998 perf_pmu_register(&perf_swevent
, "software", PERF_TYPE_SOFTWARE
);
7999 perf_pmu_register(&perf_cpu_clock
, NULL
, -1);
8000 perf_pmu_register(&perf_task_clock
, NULL
, -1);
8002 perf_cpu_notifier(perf_cpu_notify
);
8003 register_reboot_notifier(&perf_reboot_notifier
);
8005 ret
= init_hw_breakpoint();
8006 WARN(ret
, "hw_breakpoint initialization failed with: %d", ret
);
8008 /* do not patch jump label more than once per second */
8009 jump_label_rate_limit(&perf_sched_events
, HZ
);
8012 * Build time assertion that we keep the data_head at the intended
8013 * location. IOW, validation we got the __reserved[] size right.
8015 BUILD_BUG_ON((offsetof(struct perf_event_mmap_page
, data_head
))
8019 static int __init
perf_event_sysfs_init(void)
8024 mutex_lock(&pmus_lock
);
8026 ret
= bus_register(&pmu_bus
);
8030 list_for_each_entry(pmu
, &pmus
, entry
) {
8031 if (!pmu
->name
|| pmu
->type
< 0)
8034 ret
= pmu_dev_alloc(pmu
);
8035 WARN(ret
, "Failed to register pmu: %s, reason %d\n", pmu
->name
, ret
);
8037 pmu_bus_running
= 1;
8041 mutex_unlock(&pmus_lock
);
8045 device_initcall(perf_event_sysfs_init
);
8047 #ifdef CONFIG_CGROUP_PERF
8048 static struct cgroup_subsys_state
*
8049 perf_cgroup_css_alloc(struct cgroup_subsys_state
*parent_css
)
8051 struct perf_cgroup
*jc
;
8053 jc
= kzalloc(sizeof(*jc
), GFP_KERNEL
);
8055 return ERR_PTR(-ENOMEM
);
8057 jc
->info
= alloc_percpu(struct perf_cgroup_info
);
8060 return ERR_PTR(-ENOMEM
);
8066 static void perf_cgroup_css_free(struct cgroup_subsys_state
*css
)
8068 struct perf_cgroup
*jc
= container_of(css
, struct perf_cgroup
, css
);
8070 free_percpu(jc
->info
);
8074 static int __perf_cgroup_move(void *info
)
8076 struct task_struct
*task
= info
;
8077 perf_cgroup_switch(task
, PERF_CGROUP_SWOUT
| PERF_CGROUP_SWIN
);
8081 static void perf_cgroup_attach(struct cgroup_subsys_state
*css
,
8082 struct cgroup_taskset
*tset
)
8084 struct task_struct
*task
;
8086 cgroup_taskset_for_each(task
, tset
)
8087 task_function_call(task
, __perf_cgroup_move
, task
);
8090 static void perf_cgroup_exit(struct cgroup_subsys_state
*css
,
8091 struct cgroup_subsys_state
*old_css
,
8092 struct task_struct
*task
)
8095 * cgroup_exit() is called in the copy_process() failure path.
8096 * Ignore this case since the task hasn't ran yet, this avoids
8097 * trying to poke a half freed task state from generic code.
8099 if (!(task
->flags
& PF_EXITING
))
8102 task_function_call(task
, __perf_cgroup_move
, task
);
8105 struct cgroup_subsys perf_event_cgrp_subsys
= {
8106 .css_alloc
= perf_cgroup_css_alloc
,
8107 .css_free
= perf_cgroup_css_free
,
8108 .exit
= perf_cgroup_exit
,
8109 .attach
= perf_cgroup_attach
,
8111 #endif /* CONFIG_CGROUP_PERF */