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>
45 #include <asm/irq_regs.h>
47 struct remote_function_call
{
48 struct task_struct
*p
;
49 int (*func
)(void *info
);
54 static void remote_function(void *data
)
56 struct remote_function_call
*tfc
= data
;
57 struct task_struct
*p
= tfc
->p
;
61 if (task_cpu(p
) != smp_processor_id() || !task_curr(p
))
65 tfc
->ret
= tfc
->func(tfc
->info
);
69 * task_function_call - call a function on the cpu on which a task runs
70 * @p: the task to evaluate
71 * @func: the function to be called
72 * @info: the function call argument
74 * Calls the function @func when the task is currently running. This might
75 * be on the current CPU, which just calls the function directly
77 * returns: @func return value, or
78 * -ESRCH - when the process isn't running
79 * -EAGAIN - when the process moved away
82 task_function_call(struct task_struct
*p
, int (*func
) (void *info
), void *info
)
84 struct remote_function_call data
= {
88 .ret
= -ESRCH
, /* No such (running) process */
92 smp_call_function_single(task_cpu(p
), remote_function
, &data
, 1);
98 * cpu_function_call - call a function on the cpu
99 * @func: the function to be called
100 * @info: the function call argument
102 * Calls the function @func on the remote cpu.
104 * returns: @func return value or -ENXIO when the cpu is offline
106 static int cpu_function_call(int cpu
, int (*func
) (void *info
), void *info
)
108 struct remote_function_call data
= {
112 .ret
= -ENXIO
, /* No such CPU */
115 smp_call_function_single(cpu
, remote_function
, &data
, 1);
120 #define PERF_FLAG_ALL (PERF_FLAG_FD_NO_GROUP |\
121 PERF_FLAG_FD_OUTPUT |\
122 PERF_FLAG_PID_CGROUP)
125 * branch priv levels that need permission checks
127 #define PERF_SAMPLE_BRANCH_PERM_PLM \
128 (PERF_SAMPLE_BRANCH_KERNEL |\
129 PERF_SAMPLE_BRANCH_HV)
132 EVENT_FLEXIBLE
= 0x1,
134 EVENT_ALL
= EVENT_FLEXIBLE
| EVENT_PINNED
,
138 * perf_sched_events : >0 events exist
139 * perf_cgroup_events: >0 per-cpu cgroup events exist on this cpu
141 struct static_key_deferred perf_sched_events __read_mostly
;
142 static DEFINE_PER_CPU(atomic_t
, perf_cgroup_events
);
143 static DEFINE_PER_CPU(atomic_t
, perf_branch_stack_events
);
145 static atomic_t nr_mmap_events __read_mostly
;
146 static atomic_t nr_comm_events __read_mostly
;
147 static atomic_t nr_task_events __read_mostly
;
148 static atomic_t nr_freq_events __read_mostly
;
150 static LIST_HEAD(pmus
);
151 static DEFINE_MUTEX(pmus_lock
);
152 static struct srcu_struct pmus_srcu
;
155 * perf event paranoia level:
156 * -1 - not paranoid at all
157 * 0 - disallow raw tracepoint access for unpriv
158 * 1 - disallow cpu events for unpriv
159 * 2 - disallow kernel profiling for unpriv
161 int sysctl_perf_event_paranoid __read_mostly
= 1;
163 /* Minimum for 512 kiB + 1 user control page */
164 int sysctl_perf_event_mlock __read_mostly
= 512 + (PAGE_SIZE
/ 1024); /* 'free' kiB per user */
167 * max perf event sample rate
169 #define DEFAULT_MAX_SAMPLE_RATE 100000
170 #define DEFAULT_SAMPLE_PERIOD_NS (NSEC_PER_SEC / DEFAULT_MAX_SAMPLE_RATE)
171 #define DEFAULT_CPU_TIME_MAX_PERCENT 25
173 int sysctl_perf_event_sample_rate __read_mostly
= DEFAULT_MAX_SAMPLE_RATE
;
175 static int max_samples_per_tick __read_mostly
= DIV_ROUND_UP(DEFAULT_MAX_SAMPLE_RATE
, HZ
);
176 static int perf_sample_period_ns __read_mostly
= DEFAULT_SAMPLE_PERIOD_NS
;
178 static atomic_t perf_sample_allowed_ns __read_mostly
=
179 ATOMIC_INIT( DEFAULT_SAMPLE_PERIOD_NS
* DEFAULT_CPU_TIME_MAX_PERCENT
/ 100);
181 void update_perf_cpu_limits(void)
183 u64 tmp
= perf_sample_period_ns
;
185 tmp
*= sysctl_perf_cpu_time_max_percent
;
187 atomic_set(&perf_sample_allowed_ns
, tmp
);
190 static int perf_rotate_context(struct perf_cpu_context
*cpuctx
);
192 int perf_proc_update_handler(struct ctl_table
*table
, int write
,
193 void __user
*buffer
, size_t *lenp
,
196 int ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
201 max_samples_per_tick
= DIV_ROUND_UP(sysctl_perf_event_sample_rate
, HZ
);
202 perf_sample_period_ns
= NSEC_PER_SEC
/ sysctl_perf_event_sample_rate
;
203 update_perf_cpu_limits();
208 int sysctl_perf_cpu_time_max_percent __read_mostly
= DEFAULT_CPU_TIME_MAX_PERCENT
;
210 int perf_cpu_time_max_percent_handler(struct ctl_table
*table
, int write
,
211 void __user
*buffer
, size_t *lenp
,
214 int ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
219 update_perf_cpu_limits();
225 * perf samples are done in some very critical code paths (NMIs).
226 * If they take too much CPU time, the system can lock up and not
227 * get any real work done. This will drop the sample rate when
228 * we detect that events are taking too long.
230 #define NR_ACCUMULATED_SAMPLES 128
231 DEFINE_PER_CPU(u64
, running_sample_length
);
233 void perf_sample_event_took(u64 sample_len_ns
)
235 u64 avg_local_sample_len
;
236 u64 local_samples_len
;
238 if (atomic_read(&perf_sample_allowed_ns
) == 0)
241 /* decay the counter by 1 average sample */
242 local_samples_len
= __get_cpu_var(running_sample_length
);
243 local_samples_len
-= local_samples_len
/NR_ACCUMULATED_SAMPLES
;
244 local_samples_len
+= sample_len_ns
;
245 __get_cpu_var(running_sample_length
) = local_samples_len
;
248 * note: this will be biased artifically low until we have
249 * seen NR_ACCUMULATED_SAMPLES. Doing it this way keeps us
250 * from having to maintain a count.
252 avg_local_sample_len
= local_samples_len
/NR_ACCUMULATED_SAMPLES
;
254 if (avg_local_sample_len
<= atomic_read(&perf_sample_allowed_ns
))
257 if (max_samples_per_tick
<= 1)
260 max_samples_per_tick
= DIV_ROUND_UP(max_samples_per_tick
, 2);
261 sysctl_perf_event_sample_rate
= max_samples_per_tick
* HZ
;
262 perf_sample_period_ns
= NSEC_PER_SEC
/ sysctl_perf_event_sample_rate
;
264 printk_ratelimited(KERN_WARNING
265 "perf samples too long (%lld > %d), lowering "
266 "kernel.perf_event_max_sample_rate to %d\n",
267 avg_local_sample_len
,
268 atomic_read(&perf_sample_allowed_ns
),
269 sysctl_perf_event_sample_rate
);
271 update_perf_cpu_limits();
274 static atomic64_t perf_event_id
;
276 static void cpu_ctx_sched_out(struct perf_cpu_context
*cpuctx
,
277 enum event_type_t event_type
);
279 static void cpu_ctx_sched_in(struct perf_cpu_context
*cpuctx
,
280 enum event_type_t event_type
,
281 struct task_struct
*task
);
283 static void update_context_time(struct perf_event_context
*ctx
);
284 static u64
perf_event_time(struct perf_event
*event
);
286 void __weak
perf_event_print_debug(void) { }
288 extern __weak
const char *perf_pmu_name(void)
293 static inline u64
perf_clock(void)
295 return local_clock();
298 static inline struct perf_cpu_context
*
299 __get_cpu_context(struct perf_event_context
*ctx
)
301 return this_cpu_ptr(ctx
->pmu
->pmu_cpu_context
);
304 static void perf_ctx_lock(struct perf_cpu_context
*cpuctx
,
305 struct perf_event_context
*ctx
)
307 raw_spin_lock(&cpuctx
->ctx
.lock
);
309 raw_spin_lock(&ctx
->lock
);
312 static void perf_ctx_unlock(struct perf_cpu_context
*cpuctx
,
313 struct perf_event_context
*ctx
)
316 raw_spin_unlock(&ctx
->lock
);
317 raw_spin_unlock(&cpuctx
->ctx
.lock
);
320 #ifdef CONFIG_CGROUP_PERF
323 * perf_cgroup_info keeps track of time_enabled for a cgroup.
324 * This is a per-cpu dynamically allocated data structure.
326 struct perf_cgroup_info
{
332 struct cgroup_subsys_state css
;
333 struct perf_cgroup_info __percpu
*info
;
337 * Must ensure cgroup is pinned (css_get) before calling
338 * this function. In other words, we cannot call this function
339 * if there is no cgroup event for the current CPU context.
341 static inline struct perf_cgroup
*
342 perf_cgroup_from_task(struct task_struct
*task
)
344 return container_of(task_subsys_state(task
, perf_subsys_id
),
345 struct perf_cgroup
, css
);
349 perf_cgroup_match(struct perf_event
*event
)
351 struct perf_event_context
*ctx
= event
->ctx
;
352 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
354 /* @event doesn't care about cgroup */
358 /* wants specific cgroup scope but @cpuctx isn't associated with any */
363 * Cgroup scoping is recursive. An event enabled for a cgroup is
364 * also enabled for all its descendant cgroups. If @cpuctx's
365 * cgroup is a descendant of @event's (the test covers identity
366 * case), it's a match.
368 return cgroup_is_descendant(cpuctx
->cgrp
->css
.cgroup
,
369 event
->cgrp
->css
.cgroup
);
372 static inline bool perf_tryget_cgroup(struct perf_event
*event
)
374 return css_tryget(&event
->cgrp
->css
);
377 static inline void perf_put_cgroup(struct perf_event
*event
)
379 css_put(&event
->cgrp
->css
);
382 static inline void perf_detach_cgroup(struct perf_event
*event
)
384 perf_put_cgroup(event
);
388 static inline int is_cgroup_event(struct perf_event
*event
)
390 return event
->cgrp
!= NULL
;
393 static inline u64
perf_cgroup_event_time(struct perf_event
*event
)
395 struct perf_cgroup_info
*t
;
397 t
= per_cpu_ptr(event
->cgrp
->info
, event
->cpu
);
401 static inline void __update_cgrp_time(struct perf_cgroup
*cgrp
)
403 struct perf_cgroup_info
*info
;
408 info
= this_cpu_ptr(cgrp
->info
);
410 info
->time
+= now
- info
->timestamp
;
411 info
->timestamp
= now
;
414 static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context
*cpuctx
)
416 struct perf_cgroup
*cgrp_out
= cpuctx
->cgrp
;
418 __update_cgrp_time(cgrp_out
);
421 static inline void update_cgrp_time_from_event(struct perf_event
*event
)
423 struct perf_cgroup
*cgrp
;
426 * ensure we access cgroup data only when needed and
427 * when we know the cgroup is pinned (css_get)
429 if (!is_cgroup_event(event
))
432 cgrp
= perf_cgroup_from_task(current
);
434 * Do not update time when cgroup is not active
436 if (cgrp
== event
->cgrp
)
437 __update_cgrp_time(event
->cgrp
);
441 perf_cgroup_set_timestamp(struct task_struct
*task
,
442 struct perf_event_context
*ctx
)
444 struct perf_cgroup
*cgrp
;
445 struct perf_cgroup_info
*info
;
448 * ctx->lock held by caller
449 * ensure we do not access cgroup data
450 * unless we have the cgroup pinned (css_get)
452 if (!task
|| !ctx
->nr_cgroups
)
455 cgrp
= perf_cgroup_from_task(task
);
456 info
= this_cpu_ptr(cgrp
->info
);
457 info
->timestamp
= ctx
->timestamp
;
460 #define PERF_CGROUP_SWOUT 0x1 /* cgroup switch out every event */
461 #define PERF_CGROUP_SWIN 0x2 /* cgroup switch in events based on task */
464 * reschedule events based on the cgroup constraint of task.
466 * mode SWOUT : schedule out everything
467 * mode SWIN : schedule in based on cgroup for next
469 void perf_cgroup_switch(struct task_struct
*task
, int mode
)
471 struct perf_cpu_context
*cpuctx
;
476 * disable interrupts to avoid geting nr_cgroup
477 * changes via __perf_event_disable(). Also
480 local_irq_save(flags
);
483 * we reschedule only in the presence of cgroup
484 * constrained events.
488 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
489 cpuctx
= this_cpu_ptr(pmu
->pmu_cpu_context
);
490 if (cpuctx
->unique_pmu
!= pmu
)
491 continue; /* ensure we process each cpuctx once */
494 * perf_cgroup_events says at least one
495 * context on this CPU has cgroup events.
497 * ctx->nr_cgroups reports the number of cgroup
498 * events for a context.
500 if (cpuctx
->ctx
.nr_cgroups
> 0) {
501 perf_ctx_lock(cpuctx
, cpuctx
->task_ctx
);
502 perf_pmu_disable(cpuctx
->ctx
.pmu
);
504 if (mode
& PERF_CGROUP_SWOUT
) {
505 cpu_ctx_sched_out(cpuctx
, EVENT_ALL
);
507 * must not be done before ctxswout due
508 * to event_filter_match() in event_sched_out()
513 if (mode
& PERF_CGROUP_SWIN
) {
514 WARN_ON_ONCE(cpuctx
->cgrp
);
516 * set cgrp before ctxsw in to allow
517 * event_filter_match() to not have to pass
520 cpuctx
->cgrp
= perf_cgroup_from_task(task
);
521 cpu_ctx_sched_in(cpuctx
, EVENT_ALL
, task
);
523 perf_pmu_enable(cpuctx
->ctx
.pmu
);
524 perf_ctx_unlock(cpuctx
, cpuctx
->task_ctx
);
530 local_irq_restore(flags
);
533 static inline void perf_cgroup_sched_out(struct task_struct
*task
,
534 struct task_struct
*next
)
536 struct perf_cgroup
*cgrp1
;
537 struct perf_cgroup
*cgrp2
= NULL
;
540 * we come here when we know perf_cgroup_events > 0
542 cgrp1
= perf_cgroup_from_task(task
);
545 * next is NULL when called from perf_event_enable_on_exec()
546 * that will systematically cause a cgroup_switch()
549 cgrp2
= perf_cgroup_from_task(next
);
552 * only schedule out current cgroup events if we know
553 * that we are switching to a different cgroup. Otherwise,
554 * do no touch the cgroup events.
557 perf_cgroup_switch(task
, PERF_CGROUP_SWOUT
);
560 static inline void perf_cgroup_sched_in(struct task_struct
*prev
,
561 struct task_struct
*task
)
563 struct perf_cgroup
*cgrp1
;
564 struct perf_cgroup
*cgrp2
= NULL
;
567 * we come here when we know perf_cgroup_events > 0
569 cgrp1
= perf_cgroup_from_task(task
);
571 /* prev can never be NULL */
572 cgrp2
= perf_cgroup_from_task(prev
);
575 * only need to schedule in cgroup events if we are changing
576 * cgroup during ctxsw. Cgroup events were not scheduled
577 * out of ctxsw out if that was not the case.
580 perf_cgroup_switch(task
, PERF_CGROUP_SWIN
);
583 static inline int perf_cgroup_connect(int fd
, struct perf_event
*event
,
584 struct perf_event_attr
*attr
,
585 struct perf_event
*group_leader
)
587 struct perf_cgroup
*cgrp
;
588 struct cgroup_subsys_state
*css
;
589 struct fd f
= fdget(fd
);
595 css
= cgroup_css_from_dir(f
.file
, perf_subsys_id
);
601 cgrp
= container_of(css
, struct perf_cgroup
, css
);
604 /* must be done before we fput() the file */
605 if (!perf_tryget_cgroup(event
)) {
612 * all events in a group must monitor
613 * the same cgroup because a task belongs
614 * to only one perf cgroup at a time
616 if (group_leader
&& group_leader
->cgrp
!= cgrp
) {
617 perf_detach_cgroup(event
);
626 perf_cgroup_set_shadow_time(struct perf_event
*event
, u64 now
)
628 struct perf_cgroup_info
*t
;
629 t
= per_cpu_ptr(event
->cgrp
->info
, event
->cpu
);
630 event
->shadow_ctx_time
= now
- t
->timestamp
;
634 perf_cgroup_defer_enabled(struct perf_event
*event
)
637 * when the current task's perf cgroup does not match
638 * the event's, we need to remember to call the
639 * perf_mark_enable() function the first time a task with
640 * a matching perf cgroup is scheduled in.
642 if (is_cgroup_event(event
) && !perf_cgroup_match(event
))
643 event
->cgrp_defer_enabled
= 1;
647 perf_cgroup_mark_enabled(struct perf_event
*event
,
648 struct perf_event_context
*ctx
)
650 struct perf_event
*sub
;
651 u64 tstamp
= perf_event_time(event
);
653 if (!event
->cgrp_defer_enabled
)
656 event
->cgrp_defer_enabled
= 0;
658 event
->tstamp_enabled
= tstamp
- event
->total_time_enabled
;
659 list_for_each_entry(sub
, &event
->sibling_list
, group_entry
) {
660 if (sub
->state
>= PERF_EVENT_STATE_INACTIVE
) {
661 sub
->tstamp_enabled
= tstamp
- sub
->total_time_enabled
;
662 sub
->cgrp_defer_enabled
= 0;
666 #else /* !CONFIG_CGROUP_PERF */
669 perf_cgroup_match(struct perf_event
*event
)
674 static inline void perf_detach_cgroup(struct perf_event
*event
)
677 static inline int is_cgroup_event(struct perf_event
*event
)
682 static inline u64
perf_cgroup_event_cgrp_time(struct perf_event
*event
)
687 static inline void update_cgrp_time_from_event(struct perf_event
*event
)
691 static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context
*cpuctx
)
695 static inline void perf_cgroup_sched_out(struct task_struct
*task
,
696 struct task_struct
*next
)
700 static inline void perf_cgroup_sched_in(struct task_struct
*prev
,
701 struct task_struct
*task
)
705 static inline int perf_cgroup_connect(pid_t pid
, struct perf_event
*event
,
706 struct perf_event_attr
*attr
,
707 struct perf_event
*group_leader
)
713 perf_cgroup_set_timestamp(struct task_struct
*task
,
714 struct perf_event_context
*ctx
)
719 perf_cgroup_switch(struct task_struct
*task
, struct task_struct
*next
)
724 perf_cgroup_set_shadow_time(struct perf_event
*event
, u64 now
)
728 static inline u64
perf_cgroup_event_time(struct perf_event
*event
)
734 perf_cgroup_defer_enabled(struct perf_event
*event
)
739 perf_cgroup_mark_enabled(struct perf_event
*event
,
740 struct perf_event_context
*ctx
)
746 * set default to be dependent on timer tick just
749 #define PERF_CPU_HRTIMER (1000 / HZ)
751 * function must be called with interrupts disbled
753 static enum hrtimer_restart
perf_cpu_hrtimer_handler(struct hrtimer
*hr
)
755 struct perf_cpu_context
*cpuctx
;
756 enum hrtimer_restart ret
= HRTIMER_NORESTART
;
759 WARN_ON(!irqs_disabled());
761 cpuctx
= container_of(hr
, struct perf_cpu_context
, hrtimer
);
763 rotations
= perf_rotate_context(cpuctx
);
766 * arm timer if needed
769 hrtimer_forward_now(hr
, cpuctx
->hrtimer_interval
);
770 ret
= HRTIMER_RESTART
;
776 /* CPU is going down */
777 void perf_cpu_hrtimer_cancel(int cpu
)
779 struct perf_cpu_context
*cpuctx
;
783 if (WARN_ON(cpu
!= smp_processor_id()))
786 local_irq_save(flags
);
790 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
791 cpuctx
= this_cpu_ptr(pmu
->pmu_cpu_context
);
793 if (pmu
->task_ctx_nr
== perf_sw_context
)
796 hrtimer_cancel(&cpuctx
->hrtimer
);
801 local_irq_restore(flags
);
804 static void __perf_cpu_hrtimer_init(struct perf_cpu_context
*cpuctx
, int cpu
)
806 struct hrtimer
*hr
= &cpuctx
->hrtimer
;
807 struct pmu
*pmu
= cpuctx
->ctx
.pmu
;
810 /* no multiplexing needed for SW PMU */
811 if (pmu
->task_ctx_nr
== perf_sw_context
)
815 * check default is sane, if not set then force to
816 * default interval (1/tick)
818 timer
= pmu
->hrtimer_interval_ms
;
820 timer
= pmu
->hrtimer_interval_ms
= PERF_CPU_HRTIMER
;
822 cpuctx
->hrtimer_interval
= ns_to_ktime(NSEC_PER_MSEC
* timer
);
824 hrtimer_init(hr
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL_PINNED
);
825 hr
->function
= perf_cpu_hrtimer_handler
;
828 static void perf_cpu_hrtimer_restart(struct perf_cpu_context
*cpuctx
)
830 struct hrtimer
*hr
= &cpuctx
->hrtimer
;
831 struct pmu
*pmu
= cpuctx
->ctx
.pmu
;
834 if (pmu
->task_ctx_nr
== perf_sw_context
)
837 if (hrtimer_active(hr
))
840 if (!hrtimer_callback_running(hr
))
841 __hrtimer_start_range_ns(hr
, cpuctx
->hrtimer_interval
,
842 0, HRTIMER_MODE_REL_PINNED
, 0);
845 void perf_pmu_disable(struct pmu
*pmu
)
847 int *count
= this_cpu_ptr(pmu
->pmu_disable_count
);
849 pmu
->pmu_disable(pmu
);
852 void perf_pmu_enable(struct pmu
*pmu
)
854 int *count
= this_cpu_ptr(pmu
->pmu_disable_count
);
856 pmu
->pmu_enable(pmu
);
859 static DEFINE_PER_CPU(struct list_head
, rotation_list
);
862 * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
863 * because they're strictly cpu affine and rotate_start is called with IRQs
864 * disabled, while rotate_context is called from IRQ context.
866 static void perf_pmu_rotate_start(struct pmu
*pmu
)
868 struct perf_cpu_context
*cpuctx
= this_cpu_ptr(pmu
->pmu_cpu_context
);
869 struct list_head
*head
= &__get_cpu_var(rotation_list
);
871 WARN_ON(!irqs_disabled());
873 if (list_empty(&cpuctx
->rotation_list
))
874 list_add(&cpuctx
->rotation_list
, head
);
877 static void get_ctx(struct perf_event_context
*ctx
)
879 WARN_ON(!atomic_inc_not_zero(&ctx
->refcount
));
882 static void put_ctx(struct perf_event_context
*ctx
)
884 if (atomic_dec_and_test(&ctx
->refcount
)) {
886 put_ctx(ctx
->parent_ctx
);
888 put_task_struct(ctx
->task
);
889 kfree_rcu(ctx
, rcu_head
);
893 static void unclone_ctx(struct perf_event_context
*ctx
)
895 if (ctx
->parent_ctx
) {
896 put_ctx(ctx
->parent_ctx
);
897 ctx
->parent_ctx
= NULL
;
901 static u32
perf_event_pid(struct perf_event
*event
, struct task_struct
*p
)
904 * only top level events have the pid namespace they were created in
907 event
= event
->parent
;
909 return task_tgid_nr_ns(p
, event
->ns
);
912 static u32
perf_event_tid(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_pid_nr_ns(p
, event
->ns
);
924 * If we inherit events we want to return the parent event id
927 static u64
primary_event_id(struct perf_event
*event
)
932 id
= event
->parent
->id
;
938 * Get the perf_event_context for a task and lock it.
939 * This has to cope with with the fact that until it is locked,
940 * the context could get moved to another task.
942 static struct perf_event_context
*
943 perf_lock_task_context(struct task_struct
*task
, int ctxn
, unsigned long *flags
)
945 struct perf_event_context
*ctx
;
949 * One of the few rules of preemptible RCU is that one cannot do
950 * rcu_read_unlock() while holding a scheduler (or nested) lock when
951 * part of the read side critical section was preemptible -- see
952 * rcu_read_unlock_special().
954 * Since ctx->lock nests under rq->lock we must ensure the entire read
955 * side critical section is non-preemptible.
959 ctx
= rcu_dereference(task
->perf_event_ctxp
[ctxn
]);
962 * If this context is a clone of another, it might
963 * get swapped for another underneath us by
964 * perf_event_task_sched_out, though the
965 * rcu_read_lock() protects us from any context
966 * getting freed. Lock the context and check if it
967 * got swapped before we could get the lock, and retry
968 * if so. If we locked the right context, then it
969 * can't get swapped on us any more.
971 raw_spin_lock_irqsave(&ctx
->lock
, *flags
);
972 if (ctx
!= rcu_dereference(task
->perf_event_ctxp
[ctxn
])) {
973 raw_spin_unlock_irqrestore(&ctx
->lock
, *flags
);
979 if (!atomic_inc_not_zero(&ctx
->refcount
)) {
980 raw_spin_unlock_irqrestore(&ctx
->lock
, *flags
);
990 * Get the context for a task and increment its pin_count so it
991 * can't get swapped to another task. This also increments its
992 * reference count so that the context can't get freed.
994 static struct perf_event_context
*
995 perf_pin_task_context(struct task_struct
*task
, int ctxn
)
997 struct perf_event_context
*ctx
;
1000 ctx
= perf_lock_task_context(task
, ctxn
, &flags
);
1003 raw_spin_unlock_irqrestore(&ctx
->lock
, flags
);
1008 static void perf_unpin_context(struct perf_event_context
*ctx
)
1010 unsigned long flags
;
1012 raw_spin_lock_irqsave(&ctx
->lock
, flags
);
1014 raw_spin_unlock_irqrestore(&ctx
->lock
, flags
);
1018 * Update the record of the current time in a context.
1020 static void update_context_time(struct perf_event_context
*ctx
)
1022 u64 now
= perf_clock();
1024 ctx
->time
+= now
- ctx
->timestamp
;
1025 ctx
->timestamp
= now
;
1028 static u64
perf_event_time(struct perf_event
*event
)
1030 struct perf_event_context
*ctx
= event
->ctx
;
1032 if (is_cgroup_event(event
))
1033 return perf_cgroup_event_time(event
);
1035 return ctx
? ctx
->time
: 0;
1039 * Update the total_time_enabled and total_time_running fields for a event.
1040 * The caller of this function needs to hold the ctx->lock.
1042 static void update_event_times(struct perf_event
*event
)
1044 struct perf_event_context
*ctx
= event
->ctx
;
1047 if (event
->state
< PERF_EVENT_STATE_INACTIVE
||
1048 event
->group_leader
->state
< PERF_EVENT_STATE_INACTIVE
)
1051 * in cgroup mode, time_enabled represents
1052 * the time the event was enabled AND active
1053 * tasks were in the monitored cgroup. This is
1054 * independent of the activity of the context as
1055 * there may be a mix of cgroup and non-cgroup events.
1057 * That is why we treat cgroup events differently
1060 if (is_cgroup_event(event
))
1061 run_end
= perf_cgroup_event_time(event
);
1062 else if (ctx
->is_active
)
1063 run_end
= ctx
->time
;
1065 run_end
= event
->tstamp_stopped
;
1067 event
->total_time_enabled
= run_end
- event
->tstamp_enabled
;
1069 if (event
->state
== PERF_EVENT_STATE_INACTIVE
)
1070 run_end
= event
->tstamp_stopped
;
1072 run_end
= perf_event_time(event
);
1074 event
->total_time_running
= run_end
- event
->tstamp_running
;
1079 * Update total_time_enabled and total_time_running for all events in a group.
1081 static void update_group_times(struct perf_event
*leader
)
1083 struct perf_event
*event
;
1085 update_event_times(leader
);
1086 list_for_each_entry(event
, &leader
->sibling_list
, group_entry
)
1087 update_event_times(event
);
1090 static struct list_head
*
1091 ctx_group_list(struct perf_event
*event
, struct perf_event_context
*ctx
)
1093 if (event
->attr
.pinned
)
1094 return &ctx
->pinned_groups
;
1096 return &ctx
->flexible_groups
;
1100 * Add a event from the lists for its context.
1101 * Must be called with ctx->mutex and ctx->lock held.
1104 list_add_event(struct perf_event
*event
, struct perf_event_context
*ctx
)
1106 WARN_ON_ONCE(event
->attach_state
& PERF_ATTACH_CONTEXT
);
1107 event
->attach_state
|= PERF_ATTACH_CONTEXT
;
1110 * If we're a stand alone event or group leader, we go to the context
1111 * list, group events are kept attached to the group so that
1112 * perf_group_detach can, at all times, locate all siblings.
1114 if (event
->group_leader
== event
) {
1115 struct list_head
*list
;
1117 if (is_software_event(event
))
1118 event
->group_flags
|= PERF_GROUP_SOFTWARE
;
1120 list
= ctx_group_list(event
, ctx
);
1121 list_add_tail(&event
->group_entry
, list
);
1124 if (is_cgroup_event(event
))
1127 if (has_branch_stack(event
))
1128 ctx
->nr_branch_stack
++;
1130 list_add_rcu(&event
->event_entry
, &ctx
->event_list
);
1131 if (!ctx
->nr_events
)
1132 perf_pmu_rotate_start(ctx
->pmu
);
1134 if (event
->attr
.inherit_stat
)
1139 * Initialize event state based on the perf_event_attr::disabled.
1141 static inline void perf_event__state_init(struct perf_event
*event
)
1143 event
->state
= event
->attr
.disabled
? PERF_EVENT_STATE_OFF
:
1144 PERF_EVENT_STATE_INACTIVE
;
1148 * Called at perf_event creation and when events are attached/detached from a
1151 static void perf_event__read_size(struct perf_event
*event
)
1153 int entry
= sizeof(u64
); /* value */
1157 if (event
->attr
.read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
1158 size
+= sizeof(u64
);
1160 if (event
->attr
.read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
1161 size
+= sizeof(u64
);
1163 if (event
->attr
.read_format
& PERF_FORMAT_ID
)
1164 entry
+= sizeof(u64
);
1166 if (event
->attr
.read_format
& PERF_FORMAT_GROUP
) {
1167 nr
+= event
->group_leader
->nr_siblings
;
1168 size
+= sizeof(u64
);
1172 event
->read_size
= size
;
1175 static void perf_event__header_size(struct perf_event
*event
)
1177 struct perf_sample_data
*data
;
1178 u64 sample_type
= event
->attr
.sample_type
;
1181 perf_event__read_size(event
);
1183 if (sample_type
& PERF_SAMPLE_IP
)
1184 size
+= sizeof(data
->ip
);
1186 if (sample_type
& PERF_SAMPLE_ADDR
)
1187 size
+= sizeof(data
->addr
);
1189 if (sample_type
& PERF_SAMPLE_PERIOD
)
1190 size
+= sizeof(data
->period
);
1192 if (sample_type
& PERF_SAMPLE_WEIGHT
)
1193 size
+= sizeof(data
->weight
);
1195 if (sample_type
& PERF_SAMPLE_READ
)
1196 size
+= event
->read_size
;
1198 if (sample_type
& PERF_SAMPLE_DATA_SRC
)
1199 size
+= sizeof(data
->data_src
.val
);
1201 event
->header_size
= size
;
1204 static void perf_event__id_header_size(struct perf_event
*event
)
1206 struct perf_sample_data
*data
;
1207 u64 sample_type
= event
->attr
.sample_type
;
1210 if (sample_type
& PERF_SAMPLE_TID
)
1211 size
+= sizeof(data
->tid_entry
);
1213 if (sample_type
& PERF_SAMPLE_TIME
)
1214 size
+= sizeof(data
->time
);
1216 if (sample_type
& PERF_SAMPLE_IDENTIFIER
)
1217 size
+= sizeof(data
->id
);
1219 if (sample_type
& PERF_SAMPLE_ID
)
1220 size
+= sizeof(data
->id
);
1222 if (sample_type
& PERF_SAMPLE_STREAM_ID
)
1223 size
+= sizeof(data
->stream_id
);
1225 if (sample_type
& PERF_SAMPLE_CPU
)
1226 size
+= sizeof(data
->cpu_entry
);
1228 event
->id_header_size
= size
;
1231 static void perf_group_attach(struct perf_event
*event
)
1233 struct perf_event
*group_leader
= event
->group_leader
, *pos
;
1236 * We can have double attach due to group movement in perf_event_open.
1238 if (event
->attach_state
& PERF_ATTACH_GROUP
)
1241 event
->attach_state
|= PERF_ATTACH_GROUP
;
1243 if (group_leader
== event
)
1246 if (group_leader
->group_flags
& PERF_GROUP_SOFTWARE
&&
1247 !is_software_event(event
))
1248 group_leader
->group_flags
&= ~PERF_GROUP_SOFTWARE
;
1250 list_add_tail(&event
->group_entry
, &group_leader
->sibling_list
);
1251 group_leader
->nr_siblings
++;
1253 perf_event__header_size(group_leader
);
1255 list_for_each_entry(pos
, &group_leader
->sibling_list
, group_entry
)
1256 perf_event__header_size(pos
);
1260 * Remove a event from the lists for its context.
1261 * Must be called with ctx->mutex and ctx->lock held.
1264 list_del_event(struct perf_event
*event
, struct perf_event_context
*ctx
)
1266 struct perf_cpu_context
*cpuctx
;
1268 * We can have double detach due to exit/hot-unplug + close.
1270 if (!(event
->attach_state
& PERF_ATTACH_CONTEXT
))
1273 event
->attach_state
&= ~PERF_ATTACH_CONTEXT
;
1275 if (is_cgroup_event(event
)) {
1277 cpuctx
= __get_cpu_context(ctx
);
1279 * if there are no more cgroup events
1280 * then cler cgrp to avoid stale pointer
1281 * in update_cgrp_time_from_cpuctx()
1283 if (!ctx
->nr_cgroups
)
1284 cpuctx
->cgrp
= NULL
;
1287 if (has_branch_stack(event
))
1288 ctx
->nr_branch_stack
--;
1291 if (event
->attr
.inherit_stat
)
1294 list_del_rcu(&event
->event_entry
);
1296 if (event
->group_leader
== event
)
1297 list_del_init(&event
->group_entry
);
1299 update_group_times(event
);
1302 * If event was in error state, then keep it
1303 * that way, otherwise bogus counts will be
1304 * returned on read(). The only way to get out
1305 * of error state is by explicit re-enabling
1308 if (event
->state
> PERF_EVENT_STATE_OFF
)
1309 event
->state
= PERF_EVENT_STATE_OFF
;
1312 static void perf_group_detach(struct perf_event
*event
)
1314 struct perf_event
*sibling
, *tmp
;
1315 struct list_head
*list
= NULL
;
1318 * We can have double detach due to exit/hot-unplug + close.
1320 if (!(event
->attach_state
& PERF_ATTACH_GROUP
))
1323 event
->attach_state
&= ~PERF_ATTACH_GROUP
;
1326 * If this is a sibling, remove it from its group.
1328 if (event
->group_leader
!= event
) {
1329 list_del_init(&event
->group_entry
);
1330 event
->group_leader
->nr_siblings
--;
1334 if (!list_empty(&event
->group_entry
))
1335 list
= &event
->group_entry
;
1338 * If this was a group event with sibling events then
1339 * upgrade the siblings to singleton events by adding them
1340 * to whatever list we are on.
1342 list_for_each_entry_safe(sibling
, tmp
, &event
->sibling_list
, group_entry
) {
1344 list_move_tail(&sibling
->group_entry
, list
);
1345 sibling
->group_leader
= sibling
;
1347 /* Inherit group flags from the previous leader */
1348 sibling
->group_flags
= event
->group_flags
;
1352 perf_event__header_size(event
->group_leader
);
1354 list_for_each_entry(tmp
, &event
->group_leader
->sibling_list
, group_entry
)
1355 perf_event__header_size(tmp
);
1359 event_filter_match(struct perf_event
*event
)
1361 return (event
->cpu
== -1 || event
->cpu
== smp_processor_id())
1362 && perf_cgroup_match(event
);
1366 event_sched_out(struct perf_event
*event
,
1367 struct perf_cpu_context
*cpuctx
,
1368 struct perf_event_context
*ctx
)
1370 u64 tstamp
= perf_event_time(event
);
1373 * An event which could not be activated because of
1374 * filter mismatch still needs to have its timings
1375 * maintained, otherwise bogus information is return
1376 * via read() for time_enabled, time_running:
1378 if (event
->state
== PERF_EVENT_STATE_INACTIVE
1379 && !event_filter_match(event
)) {
1380 delta
= tstamp
- event
->tstamp_stopped
;
1381 event
->tstamp_running
+= delta
;
1382 event
->tstamp_stopped
= tstamp
;
1385 if (event
->state
!= PERF_EVENT_STATE_ACTIVE
)
1388 event
->state
= PERF_EVENT_STATE_INACTIVE
;
1389 if (event
->pending_disable
) {
1390 event
->pending_disable
= 0;
1391 event
->state
= PERF_EVENT_STATE_OFF
;
1393 event
->tstamp_stopped
= tstamp
;
1394 event
->pmu
->del(event
, 0);
1397 if (!is_software_event(event
))
1398 cpuctx
->active_oncpu
--;
1400 if (event
->attr
.freq
&& event
->attr
.sample_freq
)
1402 if (event
->attr
.exclusive
|| !cpuctx
->active_oncpu
)
1403 cpuctx
->exclusive
= 0;
1407 group_sched_out(struct perf_event
*group_event
,
1408 struct perf_cpu_context
*cpuctx
,
1409 struct perf_event_context
*ctx
)
1411 struct perf_event
*event
;
1412 int state
= group_event
->state
;
1414 event_sched_out(group_event
, cpuctx
, ctx
);
1417 * Schedule out siblings (if any):
1419 list_for_each_entry(event
, &group_event
->sibling_list
, group_entry
)
1420 event_sched_out(event
, cpuctx
, ctx
);
1422 if (state
== PERF_EVENT_STATE_ACTIVE
&& group_event
->attr
.exclusive
)
1423 cpuctx
->exclusive
= 0;
1427 * Cross CPU call to remove a performance event
1429 * We disable the event on the hardware level first. After that we
1430 * remove it from the context list.
1432 static int __perf_remove_from_context(void *info
)
1434 struct perf_event
*event
= info
;
1435 struct perf_event_context
*ctx
= event
->ctx
;
1436 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
1438 raw_spin_lock(&ctx
->lock
);
1439 event_sched_out(event
, cpuctx
, ctx
);
1440 list_del_event(event
, ctx
);
1441 if (!ctx
->nr_events
&& cpuctx
->task_ctx
== ctx
) {
1443 cpuctx
->task_ctx
= NULL
;
1445 raw_spin_unlock(&ctx
->lock
);
1452 * Remove the event from a task's (or a CPU's) list of events.
1454 * CPU events are removed with a smp call. For task events we only
1455 * call when the task is on a CPU.
1457 * If event->ctx is a cloned context, callers must make sure that
1458 * every task struct that event->ctx->task could possibly point to
1459 * remains valid. This is OK when called from perf_release since
1460 * that only calls us on the top-level context, which can't be a clone.
1461 * When called from perf_event_exit_task, it's OK because the
1462 * context has been detached from its task.
1464 static void perf_remove_from_context(struct perf_event
*event
)
1466 struct perf_event_context
*ctx
= event
->ctx
;
1467 struct task_struct
*task
= ctx
->task
;
1469 lockdep_assert_held(&ctx
->mutex
);
1473 * Per cpu events are removed via an smp call and
1474 * the removal is always successful.
1476 cpu_function_call(event
->cpu
, __perf_remove_from_context
, event
);
1481 if (!task_function_call(task
, __perf_remove_from_context
, event
))
1484 raw_spin_lock_irq(&ctx
->lock
);
1486 * If we failed to find a running task, but find the context active now
1487 * that we've acquired the ctx->lock, retry.
1489 if (ctx
->is_active
) {
1490 raw_spin_unlock_irq(&ctx
->lock
);
1495 * Since the task isn't running, its safe to remove the event, us
1496 * holding the ctx->lock ensures the task won't get scheduled in.
1498 list_del_event(event
, ctx
);
1499 raw_spin_unlock_irq(&ctx
->lock
);
1503 * Cross CPU call to disable a performance event
1505 int __perf_event_disable(void *info
)
1507 struct perf_event
*event
= info
;
1508 struct perf_event_context
*ctx
= event
->ctx
;
1509 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
1512 * If this is a per-task event, need to check whether this
1513 * event's task is the current task on this cpu.
1515 * Can trigger due to concurrent perf_event_context_sched_out()
1516 * flipping contexts around.
1518 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
)
1521 raw_spin_lock(&ctx
->lock
);
1524 * If the event is on, turn it off.
1525 * If it is in error state, leave it in error state.
1527 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
) {
1528 update_context_time(ctx
);
1529 update_cgrp_time_from_event(event
);
1530 update_group_times(event
);
1531 if (event
== event
->group_leader
)
1532 group_sched_out(event
, cpuctx
, ctx
);
1534 event_sched_out(event
, cpuctx
, ctx
);
1535 event
->state
= PERF_EVENT_STATE_OFF
;
1538 raw_spin_unlock(&ctx
->lock
);
1546 * If event->ctx is a cloned context, callers must make sure that
1547 * every task struct that event->ctx->task could possibly point to
1548 * remains valid. This condition is satisifed when called through
1549 * perf_event_for_each_child or perf_event_for_each because they
1550 * hold the top-level event's child_mutex, so any descendant that
1551 * goes to exit will block in sync_child_event.
1552 * When called from perf_pending_event it's OK because event->ctx
1553 * is the current context on this CPU and preemption is disabled,
1554 * hence we can't get into perf_event_task_sched_out for this context.
1556 void perf_event_disable(struct perf_event
*event
)
1558 struct perf_event_context
*ctx
= event
->ctx
;
1559 struct task_struct
*task
= ctx
->task
;
1563 * Disable the event on the cpu that it's on
1565 cpu_function_call(event
->cpu
, __perf_event_disable
, event
);
1570 if (!task_function_call(task
, __perf_event_disable
, event
))
1573 raw_spin_lock_irq(&ctx
->lock
);
1575 * If the event is still active, we need to retry the cross-call.
1577 if (event
->state
== PERF_EVENT_STATE_ACTIVE
) {
1578 raw_spin_unlock_irq(&ctx
->lock
);
1580 * Reload the task pointer, it might have been changed by
1581 * a concurrent perf_event_context_sched_out().
1588 * Since we have the lock this context can't be scheduled
1589 * in, so we can change the state safely.
1591 if (event
->state
== PERF_EVENT_STATE_INACTIVE
) {
1592 update_group_times(event
);
1593 event
->state
= PERF_EVENT_STATE_OFF
;
1595 raw_spin_unlock_irq(&ctx
->lock
);
1597 EXPORT_SYMBOL_GPL(perf_event_disable
);
1599 static void perf_set_shadow_time(struct perf_event
*event
,
1600 struct perf_event_context
*ctx
,
1604 * use the correct time source for the time snapshot
1606 * We could get by without this by leveraging the
1607 * fact that to get to this function, the caller
1608 * has most likely already called update_context_time()
1609 * and update_cgrp_time_xx() and thus both timestamp
1610 * are identical (or very close). Given that tstamp is,
1611 * already adjusted for cgroup, we could say that:
1612 * tstamp - ctx->timestamp
1614 * tstamp - cgrp->timestamp.
1616 * Then, in perf_output_read(), the calculation would
1617 * work with no changes because:
1618 * - event is guaranteed scheduled in
1619 * - no scheduled out in between
1620 * - thus the timestamp would be the same
1622 * But this is a bit hairy.
1624 * So instead, we have an explicit cgroup call to remain
1625 * within the time time source all along. We believe it
1626 * is cleaner and simpler to understand.
1628 if (is_cgroup_event(event
))
1629 perf_cgroup_set_shadow_time(event
, tstamp
);
1631 event
->shadow_ctx_time
= tstamp
- ctx
->timestamp
;
1634 #define MAX_INTERRUPTS (~0ULL)
1636 static void perf_log_throttle(struct perf_event
*event
, int enable
);
1639 event_sched_in(struct perf_event
*event
,
1640 struct perf_cpu_context
*cpuctx
,
1641 struct perf_event_context
*ctx
)
1643 u64 tstamp
= perf_event_time(event
);
1645 if (event
->state
<= PERF_EVENT_STATE_OFF
)
1648 event
->state
= PERF_EVENT_STATE_ACTIVE
;
1649 event
->oncpu
= smp_processor_id();
1652 * Unthrottle events, since we scheduled we might have missed several
1653 * ticks already, also for a heavily scheduling task there is little
1654 * guarantee it'll get a tick in a timely manner.
1656 if (unlikely(event
->hw
.interrupts
== MAX_INTERRUPTS
)) {
1657 perf_log_throttle(event
, 1);
1658 event
->hw
.interrupts
= 0;
1662 * The new state must be visible before we turn it on in the hardware:
1666 if (event
->pmu
->add(event
, PERF_EF_START
)) {
1667 event
->state
= PERF_EVENT_STATE_INACTIVE
;
1672 event
->tstamp_running
+= tstamp
- event
->tstamp_stopped
;
1674 perf_set_shadow_time(event
, ctx
, tstamp
);
1676 if (!is_software_event(event
))
1677 cpuctx
->active_oncpu
++;
1679 if (event
->attr
.freq
&& event
->attr
.sample_freq
)
1682 if (event
->attr
.exclusive
)
1683 cpuctx
->exclusive
= 1;
1689 group_sched_in(struct perf_event
*group_event
,
1690 struct perf_cpu_context
*cpuctx
,
1691 struct perf_event_context
*ctx
)
1693 struct perf_event
*event
, *partial_group
= NULL
;
1694 struct pmu
*pmu
= group_event
->pmu
;
1695 u64 now
= ctx
->time
;
1696 bool simulate
= false;
1698 if (group_event
->state
== PERF_EVENT_STATE_OFF
)
1701 pmu
->start_txn(pmu
);
1703 if (event_sched_in(group_event
, cpuctx
, ctx
)) {
1704 pmu
->cancel_txn(pmu
);
1705 perf_cpu_hrtimer_restart(cpuctx
);
1710 * Schedule in siblings as one group (if any):
1712 list_for_each_entry(event
, &group_event
->sibling_list
, group_entry
) {
1713 if (event_sched_in(event
, cpuctx
, ctx
)) {
1714 partial_group
= event
;
1719 if (!pmu
->commit_txn(pmu
))
1724 * Groups can be scheduled in as one unit only, so undo any
1725 * partial group before returning:
1726 * The events up to the failed event are scheduled out normally,
1727 * tstamp_stopped will be updated.
1729 * The failed events and the remaining siblings need to have
1730 * their timings updated as if they had gone thru event_sched_in()
1731 * and event_sched_out(). This is required to get consistent timings
1732 * across the group. This also takes care of the case where the group
1733 * could never be scheduled by ensuring tstamp_stopped is set to mark
1734 * the time the event was actually stopped, such that time delta
1735 * calculation in update_event_times() is correct.
1737 list_for_each_entry(event
, &group_event
->sibling_list
, group_entry
) {
1738 if (event
== partial_group
)
1742 event
->tstamp_running
+= now
- event
->tstamp_stopped
;
1743 event
->tstamp_stopped
= now
;
1745 event_sched_out(event
, cpuctx
, ctx
);
1748 event_sched_out(group_event
, cpuctx
, ctx
);
1750 pmu
->cancel_txn(pmu
);
1752 perf_cpu_hrtimer_restart(cpuctx
);
1758 * Work out whether we can put this event group on the CPU now.
1760 static int group_can_go_on(struct perf_event
*event
,
1761 struct perf_cpu_context
*cpuctx
,
1765 * Groups consisting entirely of software events can always go on.
1767 if (event
->group_flags
& PERF_GROUP_SOFTWARE
)
1770 * If an exclusive group is already on, no other hardware
1773 if (cpuctx
->exclusive
)
1776 * If this group is exclusive and there are already
1777 * events on the CPU, it can't go on.
1779 if (event
->attr
.exclusive
&& cpuctx
->active_oncpu
)
1782 * Otherwise, try to add it if all previous groups were able
1788 static void add_event_to_ctx(struct perf_event
*event
,
1789 struct perf_event_context
*ctx
)
1791 u64 tstamp
= perf_event_time(event
);
1793 list_add_event(event
, ctx
);
1794 perf_group_attach(event
);
1795 event
->tstamp_enabled
= tstamp
;
1796 event
->tstamp_running
= tstamp
;
1797 event
->tstamp_stopped
= tstamp
;
1800 static void task_ctx_sched_out(struct perf_event_context
*ctx
);
1802 ctx_sched_in(struct perf_event_context
*ctx
,
1803 struct perf_cpu_context
*cpuctx
,
1804 enum event_type_t event_type
,
1805 struct task_struct
*task
);
1807 static void perf_event_sched_in(struct perf_cpu_context
*cpuctx
,
1808 struct perf_event_context
*ctx
,
1809 struct task_struct
*task
)
1811 cpu_ctx_sched_in(cpuctx
, EVENT_PINNED
, task
);
1813 ctx_sched_in(ctx
, cpuctx
, EVENT_PINNED
, task
);
1814 cpu_ctx_sched_in(cpuctx
, EVENT_FLEXIBLE
, task
);
1816 ctx_sched_in(ctx
, cpuctx
, EVENT_FLEXIBLE
, task
);
1820 * Cross CPU call to install and enable a performance event
1822 * Must be called with ctx->mutex held
1824 static int __perf_install_in_context(void *info
)
1826 struct perf_event
*event
= info
;
1827 struct perf_event_context
*ctx
= event
->ctx
;
1828 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
1829 struct perf_event_context
*task_ctx
= cpuctx
->task_ctx
;
1830 struct task_struct
*task
= current
;
1832 perf_ctx_lock(cpuctx
, task_ctx
);
1833 perf_pmu_disable(cpuctx
->ctx
.pmu
);
1836 * If there was an active task_ctx schedule it out.
1839 task_ctx_sched_out(task_ctx
);
1842 * If the context we're installing events in is not the
1843 * active task_ctx, flip them.
1845 if (ctx
->task
&& task_ctx
!= ctx
) {
1847 raw_spin_unlock(&task_ctx
->lock
);
1848 raw_spin_lock(&ctx
->lock
);
1853 cpuctx
->task_ctx
= task_ctx
;
1854 task
= task_ctx
->task
;
1857 cpu_ctx_sched_out(cpuctx
, EVENT_ALL
);
1859 update_context_time(ctx
);
1861 * update cgrp time only if current cgrp
1862 * matches event->cgrp. Must be done before
1863 * calling add_event_to_ctx()
1865 update_cgrp_time_from_event(event
);
1867 add_event_to_ctx(event
, ctx
);
1870 * Schedule everything back in
1872 perf_event_sched_in(cpuctx
, task_ctx
, task
);
1874 perf_pmu_enable(cpuctx
->ctx
.pmu
);
1875 perf_ctx_unlock(cpuctx
, task_ctx
);
1881 * Attach a performance event to a context
1883 * First we add the event to the list with the hardware enable bit
1884 * in event->hw_config cleared.
1886 * If the event is attached to a task which is on a CPU we use a smp
1887 * call to enable it in the task context. The task might have been
1888 * scheduled away, but we check this in the smp call again.
1891 perf_install_in_context(struct perf_event_context
*ctx
,
1892 struct perf_event
*event
,
1895 struct task_struct
*task
= ctx
->task
;
1897 lockdep_assert_held(&ctx
->mutex
);
1900 if (event
->cpu
!= -1)
1905 * Per cpu events are installed via an smp call and
1906 * the install is always successful.
1908 cpu_function_call(cpu
, __perf_install_in_context
, event
);
1913 if (!task_function_call(task
, __perf_install_in_context
, event
))
1916 raw_spin_lock_irq(&ctx
->lock
);
1918 * If we failed to find a running task, but find the context active now
1919 * that we've acquired the ctx->lock, retry.
1921 if (ctx
->is_active
) {
1922 raw_spin_unlock_irq(&ctx
->lock
);
1927 * Since the task isn't running, its safe to add the event, us holding
1928 * the ctx->lock ensures the task won't get scheduled in.
1930 add_event_to_ctx(event
, ctx
);
1931 raw_spin_unlock_irq(&ctx
->lock
);
1935 * Put a event into inactive state and update time fields.
1936 * Enabling the leader of a group effectively enables all
1937 * the group members that aren't explicitly disabled, so we
1938 * have to update their ->tstamp_enabled also.
1939 * Note: this works for group members as well as group leaders
1940 * since the non-leader members' sibling_lists will be empty.
1942 static void __perf_event_mark_enabled(struct perf_event
*event
)
1944 struct perf_event
*sub
;
1945 u64 tstamp
= perf_event_time(event
);
1947 event
->state
= PERF_EVENT_STATE_INACTIVE
;
1948 event
->tstamp_enabled
= tstamp
- event
->total_time_enabled
;
1949 list_for_each_entry(sub
, &event
->sibling_list
, group_entry
) {
1950 if (sub
->state
>= PERF_EVENT_STATE_INACTIVE
)
1951 sub
->tstamp_enabled
= tstamp
- sub
->total_time_enabled
;
1956 * Cross CPU call to enable a performance event
1958 static int __perf_event_enable(void *info
)
1960 struct perf_event
*event
= info
;
1961 struct perf_event_context
*ctx
= event
->ctx
;
1962 struct perf_event
*leader
= event
->group_leader
;
1963 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
1967 * There's a time window between 'ctx->is_active' check
1968 * in perf_event_enable function and this place having:
1970 * - ctx->lock unlocked
1972 * where the task could be killed and 'ctx' deactivated
1973 * by perf_event_exit_task.
1975 if (!ctx
->is_active
)
1978 raw_spin_lock(&ctx
->lock
);
1979 update_context_time(ctx
);
1981 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
)
1985 * set current task's cgroup time reference point
1987 perf_cgroup_set_timestamp(current
, ctx
);
1989 __perf_event_mark_enabled(event
);
1991 if (!event_filter_match(event
)) {
1992 if (is_cgroup_event(event
))
1993 perf_cgroup_defer_enabled(event
);
1998 * If the event is in a group and isn't the group leader,
1999 * then don't put it on unless the group is on.
2001 if (leader
!= event
&& leader
->state
!= PERF_EVENT_STATE_ACTIVE
)
2004 if (!group_can_go_on(event
, cpuctx
, 1)) {
2007 if (event
== leader
)
2008 err
= group_sched_in(event
, cpuctx
, ctx
);
2010 err
= event_sched_in(event
, cpuctx
, ctx
);
2015 * If this event can't go on and it's part of a
2016 * group, then the whole group has to come off.
2018 if (leader
!= event
) {
2019 group_sched_out(leader
, cpuctx
, ctx
);
2020 perf_cpu_hrtimer_restart(cpuctx
);
2022 if (leader
->attr
.pinned
) {
2023 update_group_times(leader
);
2024 leader
->state
= PERF_EVENT_STATE_ERROR
;
2029 raw_spin_unlock(&ctx
->lock
);
2037 * If event->ctx is a cloned context, callers must make sure that
2038 * every task struct that event->ctx->task could possibly point to
2039 * remains valid. This condition is satisfied when called through
2040 * perf_event_for_each_child or perf_event_for_each as described
2041 * for perf_event_disable.
2043 void perf_event_enable(struct perf_event
*event
)
2045 struct perf_event_context
*ctx
= event
->ctx
;
2046 struct task_struct
*task
= ctx
->task
;
2050 * Enable the event on the cpu that it's on
2052 cpu_function_call(event
->cpu
, __perf_event_enable
, event
);
2056 raw_spin_lock_irq(&ctx
->lock
);
2057 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
)
2061 * If the event is in error state, clear that first.
2062 * That way, if we see the event in error state below, we
2063 * know that it has gone back into error state, as distinct
2064 * from the task having been scheduled away before the
2065 * cross-call arrived.
2067 if (event
->state
== PERF_EVENT_STATE_ERROR
)
2068 event
->state
= PERF_EVENT_STATE_OFF
;
2071 if (!ctx
->is_active
) {
2072 __perf_event_mark_enabled(event
);
2076 raw_spin_unlock_irq(&ctx
->lock
);
2078 if (!task_function_call(task
, __perf_event_enable
, event
))
2081 raw_spin_lock_irq(&ctx
->lock
);
2084 * If the context is active and the event is still off,
2085 * we need to retry the cross-call.
2087 if (ctx
->is_active
&& event
->state
== PERF_EVENT_STATE_OFF
) {
2089 * task could have been flipped by a concurrent
2090 * perf_event_context_sched_out()
2097 raw_spin_unlock_irq(&ctx
->lock
);
2099 EXPORT_SYMBOL_GPL(perf_event_enable
);
2101 int perf_event_refresh(struct perf_event
*event
, int refresh
)
2104 * not supported on inherited events
2106 if (event
->attr
.inherit
|| !is_sampling_event(event
))
2109 atomic_add(refresh
, &event
->event_limit
);
2110 perf_event_enable(event
);
2114 EXPORT_SYMBOL_GPL(perf_event_refresh
);
2116 static void ctx_sched_out(struct perf_event_context
*ctx
,
2117 struct perf_cpu_context
*cpuctx
,
2118 enum event_type_t event_type
)
2120 struct perf_event
*event
;
2121 int is_active
= ctx
->is_active
;
2123 ctx
->is_active
&= ~event_type
;
2124 if (likely(!ctx
->nr_events
))
2127 update_context_time(ctx
);
2128 update_cgrp_time_from_cpuctx(cpuctx
);
2129 if (!ctx
->nr_active
)
2132 perf_pmu_disable(ctx
->pmu
);
2133 if ((is_active
& EVENT_PINNED
) && (event_type
& EVENT_PINNED
)) {
2134 list_for_each_entry(event
, &ctx
->pinned_groups
, group_entry
)
2135 group_sched_out(event
, cpuctx
, ctx
);
2138 if ((is_active
& EVENT_FLEXIBLE
) && (event_type
& EVENT_FLEXIBLE
)) {
2139 list_for_each_entry(event
, &ctx
->flexible_groups
, group_entry
)
2140 group_sched_out(event
, cpuctx
, ctx
);
2142 perf_pmu_enable(ctx
->pmu
);
2146 * Test whether two contexts are equivalent, i.e. whether they
2147 * have both been cloned from the same version of the same context
2148 * and they both have the same number of enabled events.
2149 * If the number of enabled events is the same, then the set
2150 * of enabled events should be the same, because these are both
2151 * inherited contexts, therefore we can't access individual events
2152 * in them directly with an fd; we can only enable/disable all
2153 * events via prctl, or enable/disable all events in a family
2154 * via ioctl, which will have the same effect on both contexts.
2156 static int context_equiv(struct perf_event_context
*ctx1
,
2157 struct perf_event_context
*ctx2
)
2159 return ctx1
->parent_ctx
&& ctx1
->parent_ctx
== ctx2
->parent_ctx
2160 && ctx1
->parent_gen
== ctx2
->parent_gen
2161 && !ctx1
->pin_count
&& !ctx2
->pin_count
;
2164 static void __perf_event_sync_stat(struct perf_event
*event
,
2165 struct perf_event
*next_event
)
2169 if (!event
->attr
.inherit_stat
)
2173 * Update the event value, we cannot use perf_event_read()
2174 * because we're in the middle of a context switch and have IRQs
2175 * disabled, which upsets smp_call_function_single(), however
2176 * we know the event must be on the current CPU, therefore we
2177 * don't need to use it.
2179 switch (event
->state
) {
2180 case PERF_EVENT_STATE_ACTIVE
:
2181 event
->pmu
->read(event
);
2184 case PERF_EVENT_STATE_INACTIVE
:
2185 update_event_times(event
);
2193 * In order to keep per-task stats reliable we need to flip the event
2194 * values when we flip the contexts.
2196 value
= local64_read(&next_event
->count
);
2197 value
= local64_xchg(&event
->count
, value
);
2198 local64_set(&next_event
->count
, value
);
2200 swap(event
->total_time_enabled
, next_event
->total_time_enabled
);
2201 swap(event
->total_time_running
, next_event
->total_time_running
);
2204 * Since we swizzled the values, update the user visible data too.
2206 perf_event_update_userpage(event
);
2207 perf_event_update_userpage(next_event
);
2210 #define list_next_entry(pos, member) \
2211 list_entry(pos->member.next, typeof(*pos), member)
2213 static void perf_event_sync_stat(struct perf_event_context
*ctx
,
2214 struct perf_event_context
*next_ctx
)
2216 struct perf_event
*event
, *next_event
;
2221 update_context_time(ctx
);
2223 event
= list_first_entry(&ctx
->event_list
,
2224 struct perf_event
, event_entry
);
2226 next_event
= list_first_entry(&next_ctx
->event_list
,
2227 struct perf_event
, event_entry
);
2229 while (&event
->event_entry
!= &ctx
->event_list
&&
2230 &next_event
->event_entry
!= &next_ctx
->event_list
) {
2232 __perf_event_sync_stat(event
, next_event
);
2234 event
= list_next_entry(event
, event_entry
);
2235 next_event
= list_next_entry(next_event
, event_entry
);
2239 static void perf_event_context_sched_out(struct task_struct
*task
, int ctxn
,
2240 struct task_struct
*next
)
2242 struct perf_event_context
*ctx
= task
->perf_event_ctxp
[ctxn
];
2243 struct perf_event_context
*next_ctx
;
2244 struct perf_event_context
*parent
;
2245 struct perf_cpu_context
*cpuctx
;
2251 cpuctx
= __get_cpu_context(ctx
);
2252 if (!cpuctx
->task_ctx
)
2256 parent
= rcu_dereference(ctx
->parent_ctx
);
2257 next_ctx
= next
->perf_event_ctxp
[ctxn
];
2258 if (parent
&& next_ctx
&&
2259 rcu_dereference(next_ctx
->parent_ctx
) == parent
) {
2261 * Looks like the two contexts are clones, so we might be
2262 * able to optimize the context switch. We lock both
2263 * contexts and check that they are clones under the
2264 * lock (including re-checking that neither has been
2265 * uncloned in the meantime). It doesn't matter which
2266 * order we take the locks because no other cpu could
2267 * be trying to lock both of these tasks.
2269 raw_spin_lock(&ctx
->lock
);
2270 raw_spin_lock_nested(&next_ctx
->lock
, SINGLE_DEPTH_NESTING
);
2271 if (context_equiv(ctx
, next_ctx
)) {
2273 * XXX do we need a memory barrier of sorts
2274 * wrt to rcu_dereference() of perf_event_ctxp
2276 task
->perf_event_ctxp
[ctxn
] = next_ctx
;
2277 next
->perf_event_ctxp
[ctxn
] = ctx
;
2279 next_ctx
->task
= task
;
2282 perf_event_sync_stat(ctx
, next_ctx
);
2284 raw_spin_unlock(&next_ctx
->lock
);
2285 raw_spin_unlock(&ctx
->lock
);
2290 raw_spin_lock(&ctx
->lock
);
2291 ctx_sched_out(ctx
, cpuctx
, EVENT_ALL
);
2292 cpuctx
->task_ctx
= NULL
;
2293 raw_spin_unlock(&ctx
->lock
);
2297 #define for_each_task_context_nr(ctxn) \
2298 for ((ctxn) = 0; (ctxn) < perf_nr_task_contexts; (ctxn)++)
2301 * Called from scheduler to remove the events of the current task,
2302 * with interrupts disabled.
2304 * We stop each event and update the event value in event->count.
2306 * This does not protect us against NMI, but disable()
2307 * sets the disabled bit in the control field of event _before_
2308 * accessing the event control register. If a NMI hits, then it will
2309 * not restart the event.
2311 void __perf_event_task_sched_out(struct task_struct
*task
,
2312 struct task_struct
*next
)
2316 for_each_task_context_nr(ctxn
)
2317 perf_event_context_sched_out(task
, ctxn
, next
);
2320 * if cgroup events exist on this CPU, then we need
2321 * to check if we have to switch out PMU state.
2322 * cgroup event are system-wide mode only
2324 if (atomic_read(&__get_cpu_var(perf_cgroup_events
)))
2325 perf_cgroup_sched_out(task
, next
);
2328 static void task_ctx_sched_out(struct perf_event_context
*ctx
)
2330 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
2332 if (!cpuctx
->task_ctx
)
2335 if (WARN_ON_ONCE(ctx
!= cpuctx
->task_ctx
))
2338 ctx_sched_out(ctx
, cpuctx
, EVENT_ALL
);
2339 cpuctx
->task_ctx
= NULL
;
2343 * Called with IRQs disabled
2345 static void cpu_ctx_sched_out(struct perf_cpu_context
*cpuctx
,
2346 enum event_type_t event_type
)
2348 ctx_sched_out(&cpuctx
->ctx
, cpuctx
, event_type
);
2352 ctx_pinned_sched_in(struct perf_event_context
*ctx
,
2353 struct perf_cpu_context
*cpuctx
)
2355 struct perf_event
*event
;
2357 list_for_each_entry(event
, &ctx
->pinned_groups
, group_entry
) {
2358 if (event
->state
<= PERF_EVENT_STATE_OFF
)
2360 if (!event_filter_match(event
))
2363 /* may need to reset tstamp_enabled */
2364 if (is_cgroup_event(event
))
2365 perf_cgroup_mark_enabled(event
, ctx
);
2367 if (group_can_go_on(event
, cpuctx
, 1))
2368 group_sched_in(event
, cpuctx
, ctx
);
2371 * If this pinned group hasn't been scheduled,
2372 * put it in error state.
2374 if (event
->state
== PERF_EVENT_STATE_INACTIVE
) {
2375 update_group_times(event
);
2376 event
->state
= PERF_EVENT_STATE_ERROR
;
2382 ctx_flexible_sched_in(struct perf_event_context
*ctx
,
2383 struct perf_cpu_context
*cpuctx
)
2385 struct perf_event
*event
;
2388 list_for_each_entry(event
, &ctx
->flexible_groups
, group_entry
) {
2389 /* Ignore events in OFF or ERROR state */
2390 if (event
->state
<= PERF_EVENT_STATE_OFF
)
2393 * Listen to the 'cpu' scheduling filter constraint
2396 if (!event_filter_match(event
))
2399 /* may need to reset tstamp_enabled */
2400 if (is_cgroup_event(event
))
2401 perf_cgroup_mark_enabled(event
, ctx
);
2403 if (group_can_go_on(event
, cpuctx
, can_add_hw
)) {
2404 if (group_sched_in(event
, cpuctx
, ctx
))
2411 ctx_sched_in(struct perf_event_context
*ctx
,
2412 struct perf_cpu_context
*cpuctx
,
2413 enum event_type_t event_type
,
2414 struct task_struct
*task
)
2417 int is_active
= ctx
->is_active
;
2419 ctx
->is_active
|= event_type
;
2420 if (likely(!ctx
->nr_events
))
2424 ctx
->timestamp
= now
;
2425 perf_cgroup_set_timestamp(task
, ctx
);
2427 * First go through the list and put on any pinned groups
2428 * in order to give them the best chance of going on.
2430 if (!(is_active
& EVENT_PINNED
) && (event_type
& EVENT_PINNED
))
2431 ctx_pinned_sched_in(ctx
, cpuctx
);
2433 /* Then walk through the lower prio flexible groups */
2434 if (!(is_active
& EVENT_FLEXIBLE
) && (event_type
& EVENT_FLEXIBLE
))
2435 ctx_flexible_sched_in(ctx
, cpuctx
);
2438 static void cpu_ctx_sched_in(struct perf_cpu_context
*cpuctx
,
2439 enum event_type_t event_type
,
2440 struct task_struct
*task
)
2442 struct perf_event_context
*ctx
= &cpuctx
->ctx
;
2444 ctx_sched_in(ctx
, cpuctx
, event_type
, task
);
2447 static void perf_event_context_sched_in(struct perf_event_context
*ctx
,
2448 struct task_struct
*task
)
2450 struct perf_cpu_context
*cpuctx
;
2452 cpuctx
= __get_cpu_context(ctx
);
2453 if (cpuctx
->task_ctx
== ctx
)
2456 perf_ctx_lock(cpuctx
, ctx
);
2457 perf_pmu_disable(ctx
->pmu
);
2459 * We want to keep the following priority order:
2460 * cpu pinned (that don't need to move), task pinned,
2461 * cpu flexible, task flexible.
2463 cpu_ctx_sched_out(cpuctx
, EVENT_FLEXIBLE
);
2466 cpuctx
->task_ctx
= ctx
;
2468 perf_event_sched_in(cpuctx
, cpuctx
->task_ctx
, task
);
2470 perf_pmu_enable(ctx
->pmu
);
2471 perf_ctx_unlock(cpuctx
, ctx
);
2474 * Since these rotations are per-cpu, we need to ensure the
2475 * cpu-context we got scheduled on is actually rotating.
2477 perf_pmu_rotate_start(ctx
->pmu
);
2481 * When sampling the branck stack in system-wide, it may be necessary
2482 * to flush the stack on context switch. This happens when the branch
2483 * stack does not tag its entries with the pid of the current task.
2484 * Otherwise it becomes impossible to associate a branch entry with a
2485 * task. This ambiguity is more likely to appear when the branch stack
2486 * supports priv level filtering and the user sets it to monitor only
2487 * at the user level (which could be a useful measurement in system-wide
2488 * mode). In that case, the risk is high of having a branch stack with
2489 * branch from multiple tasks. Flushing may mean dropping the existing
2490 * entries or stashing them somewhere in the PMU specific code layer.
2492 * This function provides the context switch callback to the lower code
2493 * layer. It is invoked ONLY when there is at least one system-wide context
2494 * with at least one active event using taken branch sampling.
2496 static void perf_branch_stack_sched_in(struct task_struct
*prev
,
2497 struct task_struct
*task
)
2499 struct perf_cpu_context
*cpuctx
;
2501 unsigned long flags
;
2503 /* no need to flush branch stack if not changing task */
2507 local_irq_save(flags
);
2511 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
2512 cpuctx
= this_cpu_ptr(pmu
->pmu_cpu_context
);
2515 * check if the context has at least one
2516 * event using PERF_SAMPLE_BRANCH_STACK
2518 if (cpuctx
->ctx
.nr_branch_stack
> 0
2519 && pmu
->flush_branch_stack
) {
2521 pmu
= cpuctx
->ctx
.pmu
;
2523 perf_ctx_lock(cpuctx
, cpuctx
->task_ctx
);
2525 perf_pmu_disable(pmu
);
2527 pmu
->flush_branch_stack();
2529 perf_pmu_enable(pmu
);
2531 perf_ctx_unlock(cpuctx
, cpuctx
->task_ctx
);
2537 local_irq_restore(flags
);
2541 * Called from scheduler to add the events of the current task
2542 * with interrupts disabled.
2544 * We restore the event value and then enable it.
2546 * This does not protect us against NMI, but enable()
2547 * sets the enabled bit in the control field of event _before_
2548 * accessing the event control register. If a NMI hits, then it will
2549 * keep the event running.
2551 void __perf_event_task_sched_in(struct task_struct
*prev
,
2552 struct task_struct
*task
)
2554 struct perf_event_context
*ctx
;
2557 for_each_task_context_nr(ctxn
) {
2558 ctx
= task
->perf_event_ctxp
[ctxn
];
2562 perf_event_context_sched_in(ctx
, task
);
2565 * if cgroup events exist on this CPU, then we need
2566 * to check if we have to switch in PMU state.
2567 * cgroup event are system-wide mode only
2569 if (atomic_read(&__get_cpu_var(perf_cgroup_events
)))
2570 perf_cgroup_sched_in(prev
, task
);
2572 /* check for system-wide branch_stack events */
2573 if (atomic_read(&__get_cpu_var(perf_branch_stack_events
)))
2574 perf_branch_stack_sched_in(prev
, task
);
2577 static u64
perf_calculate_period(struct perf_event
*event
, u64 nsec
, u64 count
)
2579 u64 frequency
= event
->attr
.sample_freq
;
2580 u64 sec
= NSEC_PER_SEC
;
2581 u64 divisor
, dividend
;
2583 int count_fls
, nsec_fls
, frequency_fls
, sec_fls
;
2585 count_fls
= fls64(count
);
2586 nsec_fls
= fls64(nsec
);
2587 frequency_fls
= fls64(frequency
);
2591 * We got @count in @nsec, with a target of sample_freq HZ
2592 * the target period becomes:
2595 * period = -------------------
2596 * @nsec * sample_freq
2601 * Reduce accuracy by one bit such that @a and @b converge
2602 * to a similar magnitude.
2604 #define REDUCE_FLS(a, b) \
2606 if (a##_fls > b##_fls) { \
2616 * Reduce accuracy until either term fits in a u64, then proceed with
2617 * the other, so that finally we can do a u64/u64 division.
2619 while (count_fls
+ sec_fls
> 64 && nsec_fls
+ frequency_fls
> 64) {
2620 REDUCE_FLS(nsec
, frequency
);
2621 REDUCE_FLS(sec
, count
);
2624 if (count_fls
+ sec_fls
> 64) {
2625 divisor
= nsec
* frequency
;
2627 while (count_fls
+ sec_fls
> 64) {
2628 REDUCE_FLS(count
, sec
);
2632 dividend
= count
* sec
;
2634 dividend
= count
* sec
;
2636 while (nsec_fls
+ frequency_fls
> 64) {
2637 REDUCE_FLS(nsec
, frequency
);
2641 divisor
= nsec
* frequency
;
2647 return div64_u64(dividend
, divisor
);
2650 static DEFINE_PER_CPU(int, perf_throttled_count
);
2651 static DEFINE_PER_CPU(u64
, perf_throttled_seq
);
2653 static void perf_adjust_period(struct perf_event
*event
, u64 nsec
, u64 count
, bool disable
)
2655 struct hw_perf_event
*hwc
= &event
->hw
;
2656 s64 period
, sample_period
;
2659 period
= perf_calculate_period(event
, nsec
, count
);
2661 delta
= (s64
)(period
- hwc
->sample_period
);
2662 delta
= (delta
+ 7) / 8; /* low pass filter */
2664 sample_period
= hwc
->sample_period
+ delta
;
2669 hwc
->sample_period
= sample_period
;
2671 if (local64_read(&hwc
->period_left
) > 8*sample_period
) {
2673 event
->pmu
->stop(event
, PERF_EF_UPDATE
);
2675 local64_set(&hwc
->period_left
, 0);
2678 event
->pmu
->start(event
, PERF_EF_RELOAD
);
2683 * combine freq adjustment with unthrottling to avoid two passes over the
2684 * events. At the same time, make sure, having freq events does not change
2685 * the rate of unthrottling as that would introduce bias.
2687 static void perf_adjust_freq_unthr_context(struct perf_event_context
*ctx
,
2690 struct perf_event
*event
;
2691 struct hw_perf_event
*hwc
;
2692 u64 now
, period
= TICK_NSEC
;
2696 * only need to iterate over all events iff:
2697 * - context have events in frequency mode (needs freq adjust)
2698 * - there are events to unthrottle on this cpu
2700 if (!(ctx
->nr_freq
|| needs_unthr
))
2703 raw_spin_lock(&ctx
->lock
);
2704 perf_pmu_disable(ctx
->pmu
);
2706 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
2707 if (event
->state
!= PERF_EVENT_STATE_ACTIVE
)
2710 if (!event_filter_match(event
))
2715 if (needs_unthr
&& hwc
->interrupts
== MAX_INTERRUPTS
) {
2716 hwc
->interrupts
= 0;
2717 perf_log_throttle(event
, 1);
2718 event
->pmu
->start(event
, 0);
2721 if (!event
->attr
.freq
|| !event
->attr
.sample_freq
)
2725 * stop the event and update event->count
2727 event
->pmu
->stop(event
, PERF_EF_UPDATE
);
2729 now
= local64_read(&event
->count
);
2730 delta
= now
- hwc
->freq_count_stamp
;
2731 hwc
->freq_count_stamp
= now
;
2735 * reload only if value has changed
2736 * we have stopped the event so tell that
2737 * to perf_adjust_period() to avoid stopping it
2741 perf_adjust_period(event
, period
, delta
, false);
2743 event
->pmu
->start(event
, delta
> 0 ? PERF_EF_RELOAD
: 0);
2746 perf_pmu_enable(ctx
->pmu
);
2747 raw_spin_unlock(&ctx
->lock
);
2751 * Round-robin a context's events:
2753 static void rotate_ctx(struct perf_event_context
*ctx
)
2756 * Rotate the first entry last of non-pinned groups. Rotation might be
2757 * disabled by the inheritance code.
2759 if (!ctx
->rotate_disable
)
2760 list_rotate_left(&ctx
->flexible_groups
);
2764 * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
2765 * because they're strictly cpu affine and rotate_start is called with IRQs
2766 * disabled, while rotate_context is called from IRQ context.
2768 static int perf_rotate_context(struct perf_cpu_context
*cpuctx
)
2770 struct perf_event_context
*ctx
= NULL
;
2771 int rotate
= 0, remove
= 1;
2773 if (cpuctx
->ctx
.nr_events
) {
2775 if (cpuctx
->ctx
.nr_events
!= cpuctx
->ctx
.nr_active
)
2779 ctx
= cpuctx
->task_ctx
;
2780 if (ctx
&& ctx
->nr_events
) {
2782 if (ctx
->nr_events
!= ctx
->nr_active
)
2789 perf_ctx_lock(cpuctx
, cpuctx
->task_ctx
);
2790 perf_pmu_disable(cpuctx
->ctx
.pmu
);
2792 cpu_ctx_sched_out(cpuctx
, EVENT_FLEXIBLE
);
2794 ctx_sched_out(ctx
, cpuctx
, EVENT_FLEXIBLE
);
2796 rotate_ctx(&cpuctx
->ctx
);
2800 perf_event_sched_in(cpuctx
, ctx
, current
);
2802 perf_pmu_enable(cpuctx
->ctx
.pmu
);
2803 perf_ctx_unlock(cpuctx
, cpuctx
->task_ctx
);
2806 list_del_init(&cpuctx
->rotation_list
);
2811 #ifdef CONFIG_NO_HZ_FULL
2812 bool perf_event_can_stop_tick(void)
2814 if (atomic_read(&nr_freq_events
) ||
2815 __this_cpu_read(perf_throttled_count
))
2822 void perf_event_task_tick(void)
2824 struct list_head
*head
= &__get_cpu_var(rotation_list
);
2825 struct perf_cpu_context
*cpuctx
, *tmp
;
2826 struct perf_event_context
*ctx
;
2829 WARN_ON(!irqs_disabled());
2831 __this_cpu_inc(perf_throttled_seq
);
2832 throttled
= __this_cpu_xchg(perf_throttled_count
, 0);
2834 list_for_each_entry_safe(cpuctx
, tmp
, head
, rotation_list
) {
2836 perf_adjust_freq_unthr_context(ctx
, throttled
);
2838 ctx
= cpuctx
->task_ctx
;
2840 perf_adjust_freq_unthr_context(ctx
, throttled
);
2844 static int event_enable_on_exec(struct perf_event
*event
,
2845 struct perf_event_context
*ctx
)
2847 if (!event
->attr
.enable_on_exec
)
2850 event
->attr
.enable_on_exec
= 0;
2851 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
)
2854 __perf_event_mark_enabled(event
);
2860 * Enable all of a task's events that have been marked enable-on-exec.
2861 * This expects task == current.
2863 static void perf_event_enable_on_exec(struct perf_event_context
*ctx
)
2865 struct perf_event
*event
;
2866 unsigned long flags
;
2870 local_irq_save(flags
);
2871 if (!ctx
|| !ctx
->nr_events
)
2875 * We must ctxsw out cgroup events to avoid conflict
2876 * when invoking perf_task_event_sched_in() later on
2877 * in this function. Otherwise we end up trying to
2878 * ctxswin cgroup events which are already scheduled
2881 perf_cgroup_sched_out(current
, NULL
);
2883 raw_spin_lock(&ctx
->lock
);
2884 task_ctx_sched_out(ctx
);
2886 list_for_each_entry(event
, &ctx
->event_list
, event_entry
) {
2887 ret
= event_enable_on_exec(event
, ctx
);
2893 * Unclone this context if we enabled any event.
2898 raw_spin_unlock(&ctx
->lock
);
2901 * Also calls ctxswin for cgroup events, if any:
2903 perf_event_context_sched_in(ctx
, ctx
->task
);
2905 local_irq_restore(flags
);
2909 * Cross CPU call to read the hardware event
2911 static void __perf_event_read(void *info
)
2913 struct perf_event
*event
= info
;
2914 struct perf_event_context
*ctx
= event
->ctx
;
2915 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
2918 * If this is a task context, we need to check whether it is
2919 * the current task context of this cpu. If not it has been
2920 * scheduled out before the smp call arrived. In that case
2921 * event->count would have been updated to a recent sample
2922 * when the event was scheduled out.
2924 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
)
2927 raw_spin_lock(&ctx
->lock
);
2928 if (ctx
->is_active
) {
2929 update_context_time(ctx
);
2930 update_cgrp_time_from_event(event
);
2932 update_event_times(event
);
2933 if (event
->state
== PERF_EVENT_STATE_ACTIVE
)
2934 event
->pmu
->read(event
);
2935 raw_spin_unlock(&ctx
->lock
);
2938 static inline u64
perf_event_count(struct perf_event
*event
)
2940 return local64_read(&event
->count
) + atomic64_read(&event
->child_count
);
2943 static u64
perf_event_read(struct perf_event
*event
)
2946 * If event is enabled and currently active on a CPU, update the
2947 * value in the event structure:
2949 if (event
->state
== PERF_EVENT_STATE_ACTIVE
) {
2950 smp_call_function_single(event
->oncpu
,
2951 __perf_event_read
, event
, 1);
2952 } else if (event
->state
== PERF_EVENT_STATE_INACTIVE
) {
2953 struct perf_event_context
*ctx
= event
->ctx
;
2954 unsigned long flags
;
2956 raw_spin_lock_irqsave(&ctx
->lock
, flags
);
2958 * may read while context is not active
2959 * (e.g., thread is blocked), in that case
2960 * we cannot update context time
2962 if (ctx
->is_active
) {
2963 update_context_time(ctx
);
2964 update_cgrp_time_from_event(event
);
2966 update_event_times(event
);
2967 raw_spin_unlock_irqrestore(&ctx
->lock
, flags
);
2970 return perf_event_count(event
);
2974 * Initialize the perf_event context in a task_struct:
2976 static void __perf_event_init_context(struct perf_event_context
*ctx
)
2978 raw_spin_lock_init(&ctx
->lock
);
2979 mutex_init(&ctx
->mutex
);
2980 INIT_LIST_HEAD(&ctx
->pinned_groups
);
2981 INIT_LIST_HEAD(&ctx
->flexible_groups
);
2982 INIT_LIST_HEAD(&ctx
->event_list
);
2983 atomic_set(&ctx
->refcount
, 1);
2986 static struct perf_event_context
*
2987 alloc_perf_context(struct pmu
*pmu
, struct task_struct
*task
)
2989 struct perf_event_context
*ctx
;
2991 ctx
= kzalloc(sizeof(struct perf_event_context
), GFP_KERNEL
);
2995 __perf_event_init_context(ctx
);
2998 get_task_struct(task
);
3005 static struct task_struct
*
3006 find_lively_task_by_vpid(pid_t vpid
)
3008 struct task_struct
*task
;
3015 task
= find_task_by_vpid(vpid
);
3017 get_task_struct(task
);
3021 return ERR_PTR(-ESRCH
);
3023 /* Reuse ptrace permission checks for now. */
3025 if (!ptrace_may_access(task
, PTRACE_MODE_READ
))
3030 put_task_struct(task
);
3031 return ERR_PTR(err
);
3036 * Returns a matching context with refcount and pincount.
3038 static struct perf_event_context
*
3039 find_get_context(struct pmu
*pmu
, struct task_struct
*task
, int cpu
)
3041 struct perf_event_context
*ctx
;
3042 struct perf_cpu_context
*cpuctx
;
3043 unsigned long flags
;
3047 /* Must be root to operate on a CPU event: */
3048 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN
))
3049 return ERR_PTR(-EACCES
);
3052 * We could be clever and allow to attach a event to an
3053 * offline CPU and activate it when the CPU comes up, but
3056 if (!cpu_online(cpu
))
3057 return ERR_PTR(-ENODEV
);
3059 cpuctx
= per_cpu_ptr(pmu
->pmu_cpu_context
, cpu
);
3068 ctxn
= pmu
->task_ctx_nr
;
3073 ctx
= perf_lock_task_context(task
, ctxn
, &flags
);
3077 raw_spin_unlock_irqrestore(&ctx
->lock
, flags
);
3079 ctx
= alloc_perf_context(pmu
, task
);
3085 mutex_lock(&task
->perf_event_mutex
);
3087 * If it has already passed perf_event_exit_task().
3088 * we must see PF_EXITING, it takes this mutex too.
3090 if (task
->flags
& PF_EXITING
)
3092 else if (task
->perf_event_ctxp
[ctxn
])
3097 rcu_assign_pointer(task
->perf_event_ctxp
[ctxn
], ctx
);
3099 mutex_unlock(&task
->perf_event_mutex
);
3101 if (unlikely(err
)) {
3113 return ERR_PTR(err
);
3116 static void perf_event_free_filter(struct perf_event
*event
);
3118 static void free_event_rcu(struct rcu_head
*head
)
3120 struct perf_event
*event
;
3122 event
= container_of(head
, struct perf_event
, rcu_head
);
3124 put_pid_ns(event
->ns
);
3125 perf_event_free_filter(event
);
3129 static void ring_buffer_put(struct ring_buffer
*rb
);
3130 static void ring_buffer_detach(struct perf_event
*event
, struct ring_buffer
*rb
);
3132 static void unaccount_event_cpu(struct perf_event
*event
, int cpu
)
3137 if (has_branch_stack(event
)) {
3138 if (!(event
->attach_state
& PERF_ATTACH_TASK
))
3139 atomic_dec(&per_cpu(perf_branch_stack_events
, cpu
));
3141 if (is_cgroup_event(event
))
3142 atomic_dec(&per_cpu(perf_cgroup_events
, cpu
));
3145 static void unaccount_event(struct perf_event
*event
)
3150 if (event
->attach_state
& PERF_ATTACH_TASK
)
3151 static_key_slow_dec_deferred(&perf_sched_events
);
3152 if (event
->attr
.mmap
|| event
->attr
.mmap_data
)
3153 atomic_dec(&nr_mmap_events
);
3154 if (event
->attr
.comm
)
3155 atomic_dec(&nr_comm_events
);
3156 if (event
->attr
.task
)
3157 atomic_dec(&nr_task_events
);
3158 if (event
->attr
.freq
)
3159 atomic_dec(&nr_freq_events
);
3160 if (is_cgroup_event(event
))
3161 static_key_slow_dec_deferred(&perf_sched_events
);
3162 if (has_branch_stack(event
))
3163 static_key_slow_dec_deferred(&perf_sched_events
);
3165 unaccount_event_cpu(event
, event
->cpu
);
3168 static void __free_event(struct perf_event
*event
)
3170 if (!event
->parent
) {
3171 if (event
->attr
.sample_type
& PERF_SAMPLE_CALLCHAIN
)
3172 put_callchain_buffers();
3176 event
->destroy(event
);
3179 put_ctx(event
->ctx
);
3181 call_rcu(&event
->rcu_head
, free_event_rcu
);
3183 static void free_event(struct perf_event
*event
)
3185 irq_work_sync(&event
->pending
);
3187 unaccount_event(event
);
3190 struct ring_buffer
*rb
;
3193 * Can happen when we close an event with re-directed output.
3195 * Since we have a 0 refcount, perf_mmap_close() will skip
3196 * over us; possibly making our ring_buffer_put() the last.
3198 mutex_lock(&event
->mmap_mutex
);
3201 rcu_assign_pointer(event
->rb
, NULL
);
3202 ring_buffer_detach(event
, rb
);
3203 ring_buffer_put(rb
); /* could be last */
3205 mutex_unlock(&event
->mmap_mutex
);
3208 if (is_cgroup_event(event
))
3209 perf_detach_cgroup(event
);
3212 __free_event(event
);
3215 int perf_event_release_kernel(struct perf_event
*event
)
3217 struct perf_event_context
*ctx
= event
->ctx
;
3219 WARN_ON_ONCE(ctx
->parent_ctx
);
3221 * There are two ways this annotation is useful:
3223 * 1) there is a lock recursion from perf_event_exit_task
3224 * see the comment there.
3226 * 2) there is a lock-inversion with mmap_sem through
3227 * perf_event_read_group(), which takes faults while
3228 * holding ctx->mutex, however this is called after
3229 * the last filedesc died, so there is no possibility
3230 * to trigger the AB-BA case.
3232 mutex_lock_nested(&ctx
->mutex
, SINGLE_DEPTH_NESTING
);
3233 raw_spin_lock_irq(&ctx
->lock
);
3234 perf_group_detach(event
);
3235 raw_spin_unlock_irq(&ctx
->lock
);
3236 perf_remove_from_context(event
);
3237 mutex_unlock(&ctx
->mutex
);
3243 EXPORT_SYMBOL_GPL(perf_event_release_kernel
);
3246 * Called when the last reference to the file is gone.
3248 static void put_event(struct perf_event
*event
)
3250 struct task_struct
*owner
;
3252 if (!atomic_long_dec_and_test(&event
->refcount
))
3256 owner
= ACCESS_ONCE(event
->owner
);
3258 * Matches the smp_wmb() in perf_event_exit_task(). If we observe
3259 * !owner it means the list deletion is complete and we can indeed
3260 * free this event, otherwise we need to serialize on
3261 * owner->perf_event_mutex.
3263 smp_read_barrier_depends();
3266 * Since delayed_put_task_struct() also drops the last
3267 * task reference we can safely take a new reference
3268 * while holding the rcu_read_lock().
3270 get_task_struct(owner
);
3275 mutex_lock(&owner
->perf_event_mutex
);
3277 * We have to re-check the event->owner field, if it is cleared
3278 * we raced with perf_event_exit_task(), acquiring the mutex
3279 * ensured they're done, and we can proceed with freeing the
3283 list_del_init(&event
->owner_entry
);
3284 mutex_unlock(&owner
->perf_event_mutex
);
3285 put_task_struct(owner
);
3288 perf_event_release_kernel(event
);
3291 static int perf_release(struct inode
*inode
, struct file
*file
)
3293 put_event(file
->private_data
);
3297 u64
perf_event_read_value(struct perf_event
*event
, u64
*enabled
, u64
*running
)
3299 struct perf_event
*child
;
3305 mutex_lock(&event
->child_mutex
);
3306 total
+= perf_event_read(event
);
3307 *enabled
+= event
->total_time_enabled
+
3308 atomic64_read(&event
->child_total_time_enabled
);
3309 *running
+= event
->total_time_running
+
3310 atomic64_read(&event
->child_total_time_running
);
3312 list_for_each_entry(child
, &event
->child_list
, child_list
) {
3313 total
+= perf_event_read(child
);
3314 *enabled
+= child
->total_time_enabled
;
3315 *running
+= child
->total_time_running
;
3317 mutex_unlock(&event
->child_mutex
);
3321 EXPORT_SYMBOL_GPL(perf_event_read_value
);
3323 static int perf_event_read_group(struct perf_event
*event
,
3324 u64 read_format
, char __user
*buf
)
3326 struct perf_event
*leader
= event
->group_leader
, *sub
;
3327 int n
= 0, size
= 0, ret
= -EFAULT
;
3328 struct perf_event_context
*ctx
= leader
->ctx
;
3330 u64 count
, enabled
, running
;
3332 mutex_lock(&ctx
->mutex
);
3333 count
= perf_event_read_value(leader
, &enabled
, &running
);
3335 values
[n
++] = 1 + leader
->nr_siblings
;
3336 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
3337 values
[n
++] = enabled
;
3338 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
3339 values
[n
++] = running
;
3340 values
[n
++] = count
;
3341 if (read_format
& PERF_FORMAT_ID
)
3342 values
[n
++] = primary_event_id(leader
);
3344 size
= n
* sizeof(u64
);
3346 if (copy_to_user(buf
, values
, size
))
3351 list_for_each_entry(sub
, &leader
->sibling_list
, group_entry
) {
3354 values
[n
++] = perf_event_read_value(sub
, &enabled
, &running
);
3355 if (read_format
& PERF_FORMAT_ID
)
3356 values
[n
++] = primary_event_id(sub
);
3358 size
= n
* sizeof(u64
);
3360 if (copy_to_user(buf
+ ret
, values
, size
)) {
3368 mutex_unlock(&ctx
->mutex
);
3373 static int perf_event_read_one(struct perf_event
*event
,
3374 u64 read_format
, char __user
*buf
)
3376 u64 enabled
, running
;
3380 values
[n
++] = perf_event_read_value(event
, &enabled
, &running
);
3381 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
3382 values
[n
++] = enabled
;
3383 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
3384 values
[n
++] = running
;
3385 if (read_format
& PERF_FORMAT_ID
)
3386 values
[n
++] = primary_event_id(event
);
3388 if (copy_to_user(buf
, values
, n
* sizeof(u64
)))
3391 return n
* sizeof(u64
);
3395 * Read the performance event - simple non blocking version for now
3398 perf_read_hw(struct perf_event
*event
, char __user
*buf
, size_t count
)
3400 u64 read_format
= event
->attr
.read_format
;
3404 * Return end-of-file for a read on a event that is in
3405 * error state (i.e. because it was pinned but it couldn't be
3406 * scheduled on to the CPU at some point).
3408 if (event
->state
== PERF_EVENT_STATE_ERROR
)
3411 if (count
< event
->read_size
)
3414 WARN_ON_ONCE(event
->ctx
->parent_ctx
);
3415 if (read_format
& PERF_FORMAT_GROUP
)
3416 ret
= perf_event_read_group(event
, read_format
, buf
);
3418 ret
= perf_event_read_one(event
, read_format
, buf
);
3424 perf_read(struct file
*file
, char __user
*buf
, size_t count
, loff_t
*ppos
)
3426 struct perf_event
*event
= file
->private_data
;
3428 return perf_read_hw(event
, buf
, count
);
3431 static unsigned int perf_poll(struct file
*file
, poll_table
*wait
)
3433 struct perf_event
*event
= file
->private_data
;
3434 struct ring_buffer
*rb
;
3435 unsigned int events
= POLL_HUP
;
3438 * Pin the event->rb by taking event->mmap_mutex; otherwise
3439 * perf_event_set_output() can swizzle our rb and make us miss wakeups.
3441 mutex_lock(&event
->mmap_mutex
);
3444 events
= atomic_xchg(&rb
->poll
, 0);
3445 mutex_unlock(&event
->mmap_mutex
);
3447 poll_wait(file
, &event
->waitq
, wait
);
3452 static void perf_event_reset(struct perf_event
*event
)
3454 (void)perf_event_read(event
);
3455 local64_set(&event
->count
, 0);
3456 perf_event_update_userpage(event
);
3460 * Holding the top-level event's child_mutex means that any
3461 * descendant process that has inherited this event will block
3462 * in sync_child_event if it goes to exit, thus satisfying the
3463 * task existence requirements of perf_event_enable/disable.
3465 static void perf_event_for_each_child(struct perf_event
*event
,
3466 void (*func
)(struct perf_event
*))
3468 struct perf_event
*child
;
3470 WARN_ON_ONCE(event
->ctx
->parent_ctx
);
3471 mutex_lock(&event
->child_mutex
);
3473 list_for_each_entry(child
, &event
->child_list
, child_list
)
3475 mutex_unlock(&event
->child_mutex
);
3478 static void perf_event_for_each(struct perf_event
*event
,
3479 void (*func
)(struct perf_event
*))
3481 struct perf_event_context
*ctx
= event
->ctx
;
3482 struct perf_event
*sibling
;
3484 WARN_ON_ONCE(ctx
->parent_ctx
);
3485 mutex_lock(&ctx
->mutex
);
3486 event
= event
->group_leader
;
3488 perf_event_for_each_child(event
, func
);
3489 list_for_each_entry(sibling
, &event
->sibling_list
, group_entry
)
3490 perf_event_for_each_child(sibling
, func
);
3491 mutex_unlock(&ctx
->mutex
);
3494 static int perf_event_period(struct perf_event
*event
, u64 __user
*arg
)
3496 struct perf_event_context
*ctx
= event
->ctx
;
3500 if (!is_sampling_event(event
))
3503 if (copy_from_user(&value
, arg
, sizeof(value
)))
3509 raw_spin_lock_irq(&ctx
->lock
);
3510 if (event
->attr
.freq
) {
3511 if (value
> sysctl_perf_event_sample_rate
) {
3516 event
->attr
.sample_freq
= value
;
3518 event
->attr
.sample_period
= value
;
3519 event
->hw
.sample_period
= value
;
3522 raw_spin_unlock_irq(&ctx
->lock
);
3527 static const struct file_operations perf_fops
;
3529 static inline int perf_fget_light(int fd
, struct fd
*p
)
3531 struct fd f
= fdget(fd
);
3535 if (f
.file
->f_op
!= &perf_fops
) {
3543 static int perf_event_set_output(struct perf_event
*event
,
3544 struct perf_event
*output_event
);
3545 static int perf_event_set_filter(struct perf_event
*event
, void __user
*arg
);
3547 static long perf_ioctl(struct file
*file
, unsigned int cmd
, unsigned long arg
)
3549 struct perf_event
*event
= file
->private_data
;
3550 void (*func
)(struct perf_event
*);
3554 case PERF_EVENT_IOC_ENABLE
:
3555 func
= perf_event_enable
;
3557 case PERF_EVENT_IOC_DISABLE
:
3558 func
= perf_event_disable
;
3560 case PERF_EVENT_IOC_RESET
:
3561 func
= perf_event_reset
;
3564 case PERF_EVENT_IOC_REFRESH
:
3565 return perf_event_refresh(event
, arg
);
3567 case PERF_EVENT_IOC_PERIOD
:
3568 return perf_event_period(event
, (u64 __user
*)arg
);
3570 case PERF_EVENT_IOC_ID
:
3572 u64 id
= primary_event_id(event
);
3574 if (copy_to_user((void __user
*)arg
, &id
, sizeof(id
)))
3579 case PERF_EVENT_IOC_SET_OUTPUT
:
3583 struct perf_event
*output_event
;
3585 ret
= perf_fget_light(arg
, &output
);
3588 output_event
= output
.file
->private_data
;
3589 ret
= perf_event_set_output(event
, output_event
);
3592 ret
= perf_event_set_output(event
, NULL
);
3597 case PERF_EVENT_IOC_SET_FILTER
:
3598 return perf_event_set_filter(event
, (void __user
*)arg
);
3604 if (flags
& PERF_IOC_FLAG_GROUP
)
3605 perf_event_for_each(event
, func
);
3607 perf_event_for_each_child(event
, func
);
3612 int perf_event_task_enable(void)
3614 struct perf_event
*event
;
3616 mutex_lock(¤t
->perf_event_mutex
);
3617 list_for_each_entry(event
, ¤t
->perf_event_list
, owner_entry
)
3618 perf_event_for_each_child(event
, perf_event_enable
);
3619 mutex_unlock(¤t
->perf_event_mutex
);
3624 int perf_event_task_disable(void)
3626 struct perf_event
*event
;
3628 mutex_lock(¤t
->perf_event_mutex
);
3629 list_for_each_entry(event
, ¤t
->perf_event_list
, owner_entry
)
3630 perf_event_for_each_child(event
, perf_event_disable
);
3631 mutex_unlock(¤t
->perf_event_mutex
);
3636 static int perf_event_index(struct perf_event
*event
)
3638 if (event
->hw
.state
& PERF_HES_STOPPED
)
3641 if (event
->state
!= PERF_EVENT_STATE_ACTIVE
)
3644 return event
->pmu
->event_idx(event
);
3647 static void calc_timer_values(struct perf_event
*event
,
3654 *now
= perf_clock();
3655 ctx_time
= event
->shadow_ctx_time
+ *now
;
3656 *enabled
= ctx_time
- event
->tstamp_enabled
;
3657 *running
= ctx_time
- event
->tstamp_running
;
3660 void __weak
arch_perf_update_userpage(struct perf_event_mmap_page
*userpg
, u64 now
)
3665 * Callers need to ensure there can be no nesting of this function, otherwise
3666 * the seqlock logic goes bad. We can not serialize this because the arch
3667 * code calls this from NMI context.
3669 void perf_event_update_userpage(struct perf_event
*event
)
3671 struct perf_event_mmap_page
*userpg
;
3672 struct ring_buffer
*rb
;
3673 u64 enabled
, running
, now
;
3676 rb
= rcu_dereference(event
->rb
);
3681 * compute total_time_enabled, total_time_running
3682 * based on snapshot values taken when the event
3683 * was last scheduled in.
3685 * we cannot simply called update_context_time()
3686 * because of locking issue as we can be called in
3689 calc_timer_values(event
, &now
, &enabled
, &running
);
3691 userpg
= rb
->user_page
;
3693 * Disable preemption so as to not let the corresponding user-space
3694 * spin too long if we get preempted.
3699 userpg
->index
= perf_event_index(event
);
3700 userpg
->offset
= perf_event_count(event
);
3702 userpg
->offset
-= local64_read(&event
->hw
.prev_count
);
3704 userpg
->time_enabled
= enabled
+
3705 atomic64_read(&event
->child_total_time_enabled
);
3707 userpg
->time_running
= running
+
3708 atomic64_read(&event
->child_total_time_running
);
3710 arch_perf_update_userpage(userpg
, now
);
3719 static int perf_mmap_fault(struct vm_area_struct
*vma
, struct vm_fault
*vmf
)
3721 struct perf_event
*event
= vma
->vm_file
->private_data
;
3722 struct ring_buffer
*rb
;
3723 int ret
= VM_FAULT_SIGBUS
;
3725 if (vmf
->flags
& FAULT_FLAG_MKWRITE
) {
3726 if (vmf
->pgoff
== 0)
3732 rb
= rcu_dereference(event
->rb
);
3736 if (vmf
->pgoff
&& (vmf
->flags
& FAULT_FLAG_WRITE
))
3739 vmf
->page
= perf_mmap_to_page(rb
, vmf
->pgoff
);
3743 get_page(vmf
->page
);
3744 vmf
->page
->mapping
= vma
->vm_file
->f_mapping
;
3745 vmf
->page
->index
= vmf
->pgoff
;
3754 static void ring_buffer_attach(struct perf_event
*event
,
3755 struct ring_buffer
*rb
)
3757 unsigned long flags
;
3759 if (!list_empty(&event
->rb_entry
))
3762 spin_lock_irqsave(&rb
->event_lock
, flags
);
3763 if (list_empty(&event
->rb_entry
))
3764 list_add(&event
->rb_entry
, &rb
->event_list
);
3765 spin_unlock_irqrestore(&rb
->event_lock
, flags
);
3768 static void ring_buffer_detach(struct perf_event
*event
, struct ring_buffer
*rb
)
3770 unsigned long flags
;
3772 if (list_empty(&event
->rb_entry
))
3775 spin_lock_irqsave(&rb
->event_lock
, flags
);
3776 list_del_init(&event
->rb_entry
);
3777 wake_up_all(&event
->waitq
);
3778 spin_unlock_irqrestore(&rb
->event_lock
, flags
);
3781 static void ring_buffer_wakeup(struct perf_event
*event
)
3783 struct ring_buffer
*rb
;
3786 rb
= rcu_dereference(event
->rb
);
3788 list_for_each_entry_rcu(event
, &rb
->event_list
, rb_entry
)
3789 wake_up_all(&event
->waitq
);
3794 static void rb_free_rcu(struct rcu_head
*rcu_head
)
3796 struct ring_buffer
*rb
;
3798 rb
= container_of(rcu_head
, struct ring_buffer
, rcu_head
);
3802 static struct ring_buffer
*ring_buffer_get(struct perf_event
*event
)
3804 struct ring_buffer
*rb
;
3807 rb
= rcu_dereference(event
->rb
);
3809 if (!atomic_inc_not_zero(&rb
->refcount
))
3817 static void ring_buffer_put(struct ring_buffer
*rb
)
3819 if (!atomic_dec_and_test(&rb
->refcount
))
3822 WARN_ON_ONCE(!list_empty(&rb
->event_list
));
3824 call_rcu(&rb
->rcu_head
, rb_free_rcu
);
3827 static void perf_mmap_open(struct vm_area_struct
*vma
)
3829 struct perf_event
*event
= vma
->vm_file
->private_data
;
3831 atomic_inc(&event
->mmap_count
);
3832 atomic_inc(&event
->rb
->mmap_count
);
3836 * A buffer can be mmap()ed multiple times; either directly through the same
3837 * event, or through other events by use of perf_event_set_output().
3839 * In order to undo the VM accounting done by perf_mmap() we need to destroy
3840 * the buffer here, where we still have a VM context. This means we need
3841 * to detach all events redirecting to us.
3843 static void perf_mmap_close(struct vm_area_struct
*vma
)
3845 struct perf_event
*event
= vma
->vm_file
->private_data
;
3847 struct ring_buffer
*rb
= event
->rb
;
3848 struct user_struct
*mmap_user
= rb
->mmap_user
;
3849 int mmap_locked
= rb
->mmap_locked
;
3850 unsigned long size
= perf_data_size(rb
);
3852 atomic_dec(&rb
->mmap_count
);
3854 if (!atomic_dec_and_mutex_lock(&event
->mmap_count
, &event
->mmap_mutex
))
3857 /* Detach current event from the buffer. */
3858 rcu_assign_pointer(event
->rb
, NULL
);
3859 ring_buffer_detach(event
, rb
);
3860 mutex_unlock(&event
->mmap_mutex
);
3862 /* If there's still other mmap()s of this buffer, we're done. */
3863 if (atomic_read(&rb
->mmap_count
)) {
3864 ring_buffer_put(rb
); /* can't be last */
3869 * No other mmap()s, detach from all other events that might redirect
3870 * into the now unreachable buffer. Somewhat complicated by the
3871 * fact that rb::event_lock otherwise nests inside mmap_mutex.
3875 list_for_each_entry_rcu(event
, &rb
->event_list
, rb_entry
) {
3876 if (!atomic_long_inc_not_zero(&event
->refcount
)) {
3878 * This event is en-route to free_event() which will
3879 * detach it and remove it from the list.
3885 mutex_lock(&event
->mmap_mutex
);
3887 * Check we didn't race with perf_event_set_output() which can
3888 * swizzle the rb from under us while we were waiting to
3889 * acquire mmap_mutex.
3891 * If we find a different rb; ignore this event, a next
3892 * iteration will no longer find it on the list. We have to
3893 * still restart the iteration to make sure we're not now
3894 * iterating the wrong list.
3896 if (event
->rb
== rb
) {
3897 rcu_assign_pointer(event
->rb
, NULL
);
3898 ring_buffer_detach(event
, rb
);
3899 ring_buffer_put(rb
); /* can't be last, we still have one */
3901 mutex_unlock(&event
->mmap_mutex
);
3905 * Restart the iteration; either we're on the wrong list or
3906 * destroyed its integrity by doing a deletion.
3913 * It could be there's still a few 0-ref events on the list; they'll
3914 * get cleaned up by free_event() -- they'll also still have their
3915 * ref on the rb and will free it whenever they are done with it.
3917 * Aside from that, this buffer is 'fully' detached and unmapped,
3918 * undo the VM accounting.
3921 atomic_long_sub((size
>> PAGE_SHIFT
) + 1, &mmap_user
->locked_vm
);
3922 vma
->vm_mm
->pinned_vm
-= mmap_locked
;
3923 free_uid(mmap_user
);
3925 ring_buffer_put(rb
); /* could be last */
3928 static const struct vm_operations_struct perf_mmap_vmops
= {
3929 .open
= perf_mmap_open
,
3930 .close
= perf_mmap_close
,
3931 .fault
= perf_mmap_fault
,
3932 .page_mkwrite
= perf_mmap_fault
,
3935 static int perf_mmap(struct file
*file
, struct vm_area_struct
*vma
)
3937 struct perf_event
*event
= file
->private_data
;
3938 unsigned long user_locked
, user_lock_limit
;
3939 struct user_struct
*user
= current_user();
3940 unsigned long locked
, lock_limit
;
3941 struct ring_buffer
*rb
;
3942 unsigned long vma_size
;
3943 unsigned long nr_pages
;
3944 long user_extra
, extra
;
3945 int ret
= 0, flags
= 0;
3948 * Don't allow mmap() of inherited per-task counters. This would
3949 * create a performance issue due to all children writing to the
3952 if (event
->cpu
== -1 && event
->attr
.inherit
)
3955 if (!(vma
->vm_flags
& VM_SHARED
))
3958 vma_size
= vma
->vm_end
- vma
->vm_start
;
3959 nr_pages
= (vma_size
/ PAGE_SIZE
) - 1;
3962 * If we have rb pages ensure they're a power-of-two number, so we
3963 * can do bitmasks instead of modulo.
3965 if (nr_pages
!= 0 && !is_power_of_2(nr_pages
))
3968 if (vma_size
!= PAGE_SIZE
* (1 + nr_pages
))
3971 if (vma
->vm_pgoff
!= 0)
3974 WARN_ON_ONCE(event
->ctx
->parent_ctx
);
3976 mutex_lock(&event
->mmap_mutex
);
3978 if (event
->rb
->nr_pages
!= nr_pages
) {
3983 if (!atomic_inc_not_zero(&event
->rb
->mmap_count
)) {
3985 * Raced against perf_mmap_close() through
3986 * perf_event_set_output(). Try again, hope for better
3989 mutex_unlock(&event
->mmap_mutex
);
3996 user_extra
= nr_pages
+ 1;
3997 user_lock_limit
= sysctl_perf_event_mlock
>> (PAGE_SHIFT
- 10);
4000 * Increase the limit linearly with more CPUs:
4002 user_lock_limit
*= num_online_cpus();
4004 user_locked
= atomic_long_read(&user
->locked_vm
) + user_extra
;
4007 if (user_locked
> user_lock_limit
)
4008 extra
= user_locked
- user_lock_limit
;
4010 lock_limit
= rlimit(RLIMIT_MEMLOCK
);
4011 lock_limit
>>= PAGE_SHIFT
;
4012 locked
= vma
->vm_mm
->pinned_vm
+ extra
;
4014 if ((locked
> lock_limit
) && perf_paranoid_tracepoint_raw() &&
4015 !capable(CAP_IPC_LOCK
)) {
4022 if (vma
->vm_flags
& VM_WRITE
)
4023 flags
|= RING_BUFFER_WRITABLE
;
4025 rb
= rb_alloc(nr_pages
,
4026 event
->attr
.watermark
? event
->attr
.wakeup_watermark
: 0,
4034 atomic_set(&rb
->mmap_count
, 1);
4035 rb
->mmap_locked
= extra
;
4036 rb
->mmap_user
= get_current_user();
4038 atomic_long_add(user_extra
, &user
->locked_vm
);
4039 vma
->vm_mm
->pinned_vm
+= extra
;
4041 ring_buffer_attach(event
, rb
);
4042 rcu_assign_pointer(event
->rb
, rb
);
4044 perf_event_update_userpage(event
);
4048 atomic_inc(&event
->mmap_count
);
4049 mutex_unlock(&event
->mmap_mutex
);
4052 * Since pinned accounting is per vm we cannot allow fork() to copy our
4055 vma
->vm_flags
|= VM_DONTCOPY
| VM_DONTEXPAND
| VM_DONTDUMP
;
4056 vma
->vm_ops
= &perf_mmap_vmops
;
4061 static int perf_fasync(int fd
, struct file
*filp
, int on
)
4063 struct inode
*inode
= file_inode(filp
);
4064 struct perf_event
*event
= filp
->private_data
;
4067 mutex_lock(&inode
->i_mutex
);
4068 retval
= fasync_helper(fd
, filp
, on
, &event
->fasync
);
4069 mutex_unlock(&inode
->i_mutex
);
4077 static const struct file_operations perf_fops
= {
4078 .llseek
= no_llseek
,
4079 .release
= perf_release
,
4082 .unlocked_ioctl
= perf_ioctl
,
4083 .compat_ioctl
= perf_ioctl
,
4085 .fasync
= perf_fasync
,
4091 * If there's data, ensure we set the poll() state and publish everything
4092 * to user-space before waking everybody up.
4095 void perf_event_wakeup(struct perf_event
*event
)
4097 ring_buffer_wakeup(event
);
4099 if (event
->pending_kill
) {
4100 kill_fasync(&event
->fasync
, SIGIO
, event
->pending_kill
);
4101 event
->pending_kill
= 0;
4105 static void perf_pending_event(struct irq_work
*entry
)
4107 struct perf_event
*event
= container_of(entry
,
4108 struct perf_event
, pending
);
4110 if (event
->pending_disable
) {
4111 event
->pending_disable
= 0;
4112 __perf_event_disable(event
);
4115 if (event
->pending_wakeup
) {
4116 event
->pending_wakeup
= 0;
4117 perf_event_wakeup(event
);
4122 * We assume there is only KVM supporting the callbacks.
4123 * Later on, we might change it to a list if there is
4124 * another virtualization implementation supporting the callbacks.
4126 struct perf_guest_info_callbacks
*perf_guest_cbs
;
4128 int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks
*cbs
)
4130 perf_guest_cbs
= cbs
;
4133 EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks
);
4135 int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks
*cbs
)
4137 perf_guest_cbs
= NULL
;
4140 EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks
);
4143 perf_output_sample_regs(struct perf_output_handle
*handle
,
4144 struct pt_regs
*regs
, u64 mask
)
4148 for_each_set_bit(bit
, (const unsigned long *) &mask
,
4149 sizeof(mask
) * BITS_PER_BYTE
) {
4152 val
= perf_reg_value(regs
, bit
);
4153 perf_output_put(handle
, val
);
4157 static void perf_sample_regs_user(struct perf_regs_user
*regs_user
,
4158 struct pt_regs
*regs
)
4160 if (!user_mode(regs
)) {
4162 regs
= task_pt_regs(current
);
4168 regs_user
->regs
= regs
;
4169 regs_user
->abi
= perf_reg_abi(current
);
4174 * Get remaining task size from user stack pointer.
4176 * It'd be better to take stack vma map and limit this more
4177 * precisly, but there's no way to get it safely under interrupt,
4178 * so using TASK_SIZE as limit.
4180 static u64
perf_ustack_task_size(struct pt_regs
*regs
)
4182 unsigned long addr
= perf_user_stack_pointer(regs
);
4184 if (!addr
|| addr
>= TASK_SIZE
)
4187 return TASK_SIZE
- addr
;
4191 perf_sample_ustack_size(u16 stack_size
, u16 header_size
,
4192 struct pt_regs
*regs
)
4196 /* No regs, no stack pointer, no dump. */
4201 * Check if we fit in with the requested stack size into the:
4203 * If we don't, we limit the size to the TASK_SIZE.
4205 * - remaining sample size
4206 * If we don't, we customize the stack size to
4207 * fit in to the remaining sample size.
4210 task_size
= min((u64
) USHRT_MAX
, perf_ustack_task_size(regs
));
4211 stack_size
= min(stack_size
, (u16
) task_size
);
4213 /* Current header size plus static size and dynamic size. */
4214 header_size
+= 2 * sizeof(u64
);
4216 /* Do we fit in with the current stack dump size? */
4217 if ((u16
) (header_size
+ stack_size
) < header_size
) {
4219 * If we overflow the maximum size for the sample,
4220 * we customize the stack dump size to fit in.
4222 stack_size
= USHRT_MAX
- header_size
- sizeof(u64
);
4223 stack_size
= round_up(stack_size
, sizeof(u64
));
4230 perf_output_sample_ustack(struct perf_output_handle
*handle
, u64 dump_size
,
4231 struct pt_regs
*regs
)
4233 /* Case of a kernel thread, nothing to dump */
4236 perf_output_put(handle
, size
);
4245 * - the size requested by user or the best one we can fit
4246 * in to the sample max size
4248 * - user stack dump data
4250 * - the actual dumped size
4254 perf_output_put(handle
, dump_size
);
4257 sp
= perf_user_stack_pointer(regs
);
4258 rem
= __output_copy_user(handle
, (void *) sp
, dump_size
);
4259 dyn_size
= dump_size
- rem
;
4261 perf_output_skip(handle
, rem
);
4264 perf_output_put(handle
, dyn_size
);
4268 static void __perf_event_header__init_id(struct perf_event_header
*header
,
4269 struct perf_sample_data
*data
,
4270 struct perf_event
*event
)
4272 u64 sample_type
= event
->attr
.sample_type
;
4274 data
->type
= sample_type
;
4275 header
->size
+= event
->id_header_size
;
4277 if (sample_type
& PERF_SAMPLE_TID
) {
4278 /* namespace issues */
4279 data
->tid_entry
.pid
= perf_event_pid(event
, current
);
4280 data
->tid_entry
.tid
= perf_event_tid(event
, current
);
4283 if (sample_type
& PERF_SAMPLE_TIME
)
4284 data
->time
= perf_clock();
4286 if (sample_type
& (PERF_SAMPLE_ID
| PERF_SAMPLE_IDENTIFIER
))
4287 data
->id
= primary_event_id(event
);
4289 if (sample_type
& PERF_SAMPLE_STREAM_ID
)
4290 data
->stream_id
= event
->id
;
4292 if (sample_type
& PERF_SAMPLE_CPU
) {
4293 data
->cpu_entry
.cpu
= raw_smp_processor_id();
4294 data
->cpu_entry
.reserved
= 0;
4298 void perf_event_header__init_id(struct perf_event_header
*header
,
4299 struct perf_sample_data
*data
,
4300 struct perf_event
*event
)
4302 if (event
->attr
.sample_id_all
)
4303 __perf_event_header__init_id(header
, data
, event
);
4306 static void __perf_event__output_id_sample(struct perf_output_handle
*handle
,
4307 struct perf_sample_data
*data
)
4309 u64 sample_type
= data
->type
;
4311 if (sample_type
& PERF_SAMPLE_TID
)
4312 perf_output_put(handle
, data
->tid_entry
);
4314 if (sample_type
& PERF_SAMPLE_TIME
)
4315 perf_output_put(handle
, data
->time
);
4317 if (sample_type
& PERF_SAMPLE_ID
)
4318 perf_output_put(handle
, data
->id
);
4320 if (sample_type
& PERF_SAMPLE_STREAM_ID
)
4321 perf_output_put(handle
, data
->stream_id
);
4323 if (sample_type
& PERF_SAMPLE_CPU
)
4324 perf_output_put(handle
, data
->cpu_entry
);
4326 if (sample_type
& PERF_SAMPLE_IDENTIFIER
)
4327 perf_output_put(handle
, data
->id
);
4330 void perf_event__output_id_sample(struct perf_event
*event
,
4331 struct perf_output_handle
*handle
,
4332 struct perf_sample_data
*sample
)
4334 if (event
->attr
.sample_id_all
)
4335 __perf_event__output_id_sample(handle
, sample
);
4338 static void perf_output_read_one(struct perf_output_handle
*handle
,
4339 struct perf_event
*event
,
4340 u64 enabled
, u64 running
)
4342 u64 read_format
= event
->attr
.read_format
;
4346 values
[n
++] = perf_event_count(event
);
4347 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
) {
4348 values
[n
++] = enabled
+
4349 atomic64_read(&event
->child_total_time_enabled
);
4351 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
) {
4352 values
[n
++] = running
+
4353 atomic64_read(&event
->child_total_time_running
);
4355 if (read_format
& PERF_FORMAT_ID
)
4356 values
[n
++] = primary_event_id(event
);
4358 __output_copy(handle
, values
, n
* sizeof(u64
));
4362 * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
4364 static void perf_output_read_group(struct perf_output_handle
*handle
,
4365 struct perf_event
*event
,
4366 u64 enabled
, u64 running
)
4368 struct perf_event
*leader
= event
->group_leader
, *sub
;
4369 u64 read_format
= event
->attr
.read_format
;
4373 values
[n
++] = 1 + leader
->nr_siblings
;
4375 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
4376 values
[n
++] = enabled
;
4378 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
4379 values
[n
++] = running
;
4381 if (leader
!= event
)
4382 leader
->pmu
->read(leader
);
4384 values
[n
++] = perf_event_count(leader
);
4385 if (read_format
& PERF_FORMAT_ID
)
4386 values
[n
++] = primary_event_id(leader
);
4388 __output_copy(handle
, values
, n
* sizeof(u64
));
4390 list_for_each_entry(sub
, &leader
->sibling_list
, group_entry
) {
4393 if ((sub
!= event
) &&
4394 (sub
->state
== PERF_EVENT_STATE_ACTIVE
))
4395 sub
->pmu
->read(sub
);
4397 values
[n
++] = perf_event_count(sub
);
4398 if (read_format
& PERF_FORMAT_ID
)
4399 values
[n
++] = primary_event_id(sub
);
4401 __output_copy(handle
, values
, n
* sizeof(u64
));
4405 #define PERF_FORMAT_TOTAL_TIMES (PERF_FORMAT_TOTAL_TIME_ENABLED|\
4406 PERF_FORMAT_TOTAL_TIME_RUNNING)
4408 static void perf_output_read(struct perf_output_handle
*handle
,
4409 struct perf_event
*event
)
4411 u64 enabled
= 0, running
= 0, now
;
4412 u64 read_format
= event
->attr
.read_format
;
4415 * compute total_time_enabled, total_time_running
4416 * based on snapshot values taken when the event
4417 * was last scheduled in.
4419 * we cannot simply called update_context_time()
4420 * because of locking issue as we are called in
4423 if (read_format
& PERF_FORMAT_TOTAL_TIMES
)
4424 calc_timer_values(event
, &now
, &enabled
, &running
);
4426 if (event
->attr
.read_format
& PERF_FORMAT_GROUP
)
4427 perf_output_read_group(handle
, event
, enabled
, running
);
4429 perf_output_read_one(handle
, event
, enabled
, running
);
4432 void perf_output_sample(struct perf_output_handle
*handle
,
4433 struct perf_event_header
*header
,
4434 struct perf_sample_data
*data
,
4435 struct perf_event
*event
)
4437 u64 sample_type
= data
->type
;
4439 perf_output_put(handle
, *header
);
4441 if (sample_type
& PERF_SAMPLE_IDENTIFIER
)
4442 perf_output_put(handle
, data
->id
);
4444 if (sample_type
& PERF_SAMPLE_IP
)
4445 perf_output_put(handle
, data
->ip
);
4447 if (sample_type
& PERF_SAMPLE_TID
)
4448 perf_output_put(handle
, data
->tid_entry
);
4450 if (sample_type
& PERF_SAMPLE_TIME
)
4451 perf_output_put(handle
, data
->time
);
4453 if (sample_type
& PERF_SAMPLE_ADDR
)
4454 perf_output_put(handle
, data
->addr
);
4456 if (sample_type
& PERF_SAMPLE_ID
)
4457 perf_output_put(handle
, data
->id
);
4459 if (sample_type
& PERF_SAMPLE_STREAM_ID
)
4460 perf_output_put(handle
, data
->stream_id
);
4462 if (sample_type
& PERF_SAMPLE_CPU
)
4463 perf_output_put(handle
, data
->cpu_entry
);
4465 if (sample_type
& PERF_SAMPLE_PERIOD
)
4466 perf_output_put(handle
, data
->period
);
4468 if (sample_type
& PERF_SAMPLE_READ
)
4469 perf_output_read(handle
, event
);
4471 if (sample_type
& PERF_SAMPLE_CALLCHAIN
) {
4472 if (data
->callchain
) {
4475 if (data
->callchain
)
4476 size
+= data
->callchain
->nr
;
4478 size
*= sizeof(u64
);
4480 __output_copy(handle
, data
->callchain
, size
);
4483 perf_output_put(handle
, nr
);
4487 if (sample_type
& PERF_SAMPLE_RAW
) {
4489 perf_output_put(handle
, data
->raw
->size
);
4490 __output_copy(handle
, data
->raw
->data
,
4497 .size
= sizeof(u32
),
4500 perf_output_put(handle
, raw
);
4504 if (sample_type
& PERF_SAMPLE_BRANCH_STACK
) {
4505 if (data
->br_stack
) {
4508 size
= data
->br_stack
->nr
4509 * sizeof(struct perf_branch_entry
);
4511 perf_output_put(handle
, data
->br_stack
->nr
);
4512 perf_output_copy(handle
, data
->br_stack
->entries
, size
);
4515 * we always store at least the value of nr
4518 perf_output_put(handle
, nr
);
4522 if (sample_type
& PERF_SAMPLE_REGS_USER
) {
4523 u64 abi
= data
->regs_user
.abi
;
4526 * If there are no regs to dump, notice it through
4527 * first u64 being zero (PERF_SAMPLE_REGS_ABI_NONE).
4529 perf_output_put(handle
, abi
);
4532 u64 mask
= event
->attr
.sample_regs_user
;
4533 perf_output_sample_regs(handle
,
4534 data
->regs_user
.regs
,
4539 if (sample_type
& PERF_SAMPLE_STACK_USER
) {
4540 perf_output_sample_ustack(handle
,
4541 data
->stack_user_size
,
4542 data
->regs_user
.regs
);
4545 if (sample_type
& PERF_SAMPLE_WEIGHT
)
4546 perf_output_put(handle
, data
->weight
);
4548 if (sample_type
& PERF_SAMPLE_DATA_SRC
)
4549 perf_output_put(handle
, data
->data_src
.val
);
4551 if (!event
->attr
.watermark
) {
4552 int wakeup_events
= event
->attr
.wakeup_events
;
4554 if (wakeup_events
) {
4555 struct ring_buffer
*rb
= handle
->rb
;
4556 int events
= local_inc_return(&rb
->events
);
4558 if (events
>= wakeup_events
) {
4559 local_sub(wakeup_events
, &rb
->events
);
4560 local_inc(&rb
->wakeup
);
4566 void perf_prepare_sample(struct perf_event_header
*header
,
4567 struct perf_sample_data
*data
,
4568 struct perf_event
*event
,
4569 struct pt_regs
*regs
)
4571 u64 sample_type
= event
->attr
.sample_type
;
4573 header
->type
= PERF_RECORD_SAMPLE
;
4574 header
->size
= sizeof(*header
) + event
->header_size
;
4577 header
->misc
|= perf_misc_flags(regs
);
4579 __perf_event_header__init_id(header
, data
, event
);
4581 if (sample_type
& PERF_SAMPLE_IP
)
4582 data
->ip
= perf_instruction_pointer(regs
);
4584 if (sample_type
& PERF_SAMPLE_CALLCHAIN
) {
4587 data
->callchain
= perf_callchain(event
, regs
);
4589 if (data
->callchain
)
4590 size
+= data
->callchain
->nr
;
4592 header
->size
+= size
* sizeof(u64
);
4595 if (sample_type
& PERF_SAMPLE_RAW
) {
4596 int size
= sizeof(u32
);
4599 size
+= data
->raw
->size
;
4601 size
+= sizeof(u32
);
4603 WARN_ON_ONCE(size
& (sizeof(u64
)-1));
4604 header
->size
+= size
;
4607 if (sample_type
& PERF_SAMPLE_BRANCH_STACK
) {
4608 int size
= sizeof(u64
); /* nr */
4609 if (data
->br_stack
) {
4610 size
+= data
->br_stack
->nr
4611 * sizeof(struct perf_branch_entry
);
4613 header
->size
+= size
;
4616 if (sample_type
& PERF_SAMPLE_REGS_USER
) {
4617 /* regs dump ABI info */
4618 int size
= sizeof(u64
);
4620 perf_sample_regs_user(&data
->regs_user
, regs
);
4622 if (data
->regs_user
.regs
) {
4623 u64 mask
= event
->attr
.sample_regs_user
;
4624 size
+= hweight64(mask
) * sizeof(u64
);
4627 header
->size
+= size
;
4630 if (sample_type
& PERF_SAMPLE_STACK_USER
) {
4632 * Either we need PERF_SAMPLE_STACK_USER bit to be allways
4633 * processed as the last one or have additional check added
4634 * in case new sample type is added, because we could eat
4635 * up the rest of the sample size.
4637 struct perf_regs_user
*uregs
= &data
->regs_user
;
4638 u16 stack_size
= event
->attr
.sample_stack_user
;
4639 u16 size
= sizeof(u64
);
4642 perf_sample_regs_user(uregs
, regs
);
4644 stack_size
= perf_sample_ustack_size(stack_size
, header
->size
,
4648 * If there is something to dump, add space for the dump
4649 * itself and for the field that tells the dynamic size,
4650 * which is how many have been actually dumped.
4653 size
+= sizeof(u64
) + stack_size
;
4655 data
->stack_user_size
= stack_size
;
4656 header
->size
+= size
;
4660 static void perf_event_output(struct perf_event
*event
,
4661 struct perf_sample_data
*data
,
4662 struct pt_regs
*regs
)
4664 struct perf_output_handle handle
;
4665 struct perf_event_header header
;
4667 /* protect the callchain buffers */
4670 perf_prepare_sample(&header
, data
, event
, regs
);
4672 if (perf_output_begin(&handle
, event
, header
.size
))
4675 perf_output_sample(&handle
, &header
, data
, event
);
4677 perf_output_end(&handle
);
4687 struct perf_read_event
{
4688 struct perf_event_header header
;
4695 perf_event_read_event(struct perf_event
*event
,
4696 struct task_struct
*task
)
4698 struct perf_output_handle handle
;
4699 struct perf_sample_data sample
;
4700 struct perf_read_event read_event
= {
4702 .type
= PERF_RECORD_READ
,
4704 .size
= sizeof(read_event
) + event
->read_size
,
4706 .pid
= perf_event_pid(event
, task
),
4707 .tid
= perf_event_tid(event
, task
),
4711 perf_event_header__init_id(&read_event
.header
, &sample
, event
);
4712 ret
= perf_output_begin(&handle
, event
, read_event
.header
.size
);
4716 perf_output_put(&handle
, read_event
);
4717 perf_output_read(&handle
, event
);
4718 perf_event__output_id_sample(event
, &handle
, &sample
);
4720 perf_output_end(&handle
);
4723 typedef void (perf_event_aux_output_cb
)(struct perf_event
*event
, void *data
);
4726 perf_event_aux_ctx(struct perf_event_context
*ctx
,
4727 perf_event_aux_output_cb output
,
4730 struct perf_event
*event
;
4732 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
4733 if (event
->state
< PERF_EVENT_STATE_INACTIVE
)
4735 if (!event_filter_match(event
))
4737 output(event
, data
);
4742 perf_event_aux(perf_event_aux_output_cb output
, void *data
,
4743 struct perf_event_context
*task_ctx
)
4745 struct perf_cpu_context
*cpuctx
;
4746 struct perf_event_context
*ctx
;
4751 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
4752 cpuctx
= get_cpu_ptr(pmu
->pmu_cpu_context
);
4753 if (cpuctx
->unique_pmu
!= pmu
)
4755 perf_event_aux_ctx(&cpuctx
->ctx
, output
, data
);
4758 ctxn
= pmu
->task_ctx_nr
;
4761 ctx
= rcu_dereference(current
->perf_event_ctxp
[ctxn
]);
4763 perf_event_aux_ctx(ctx
, output
, data
);
4765 put_cpu_ptr(pmu
->pmu_cpu_context
);
4770 perf_event_aux_ctx(task_ctx
, output
, data
);
4777 * task tracking -- fork/exit
4779 * enabled by: attr.comm | attr.mmap | attr.mmap2 | attr.mmap_data | attr.task
4782 struct perf_task_event
{
4783 struct task_struct
*task
;
4784 struct perf_event_context
*task_ctx
;
4787 struct perf_event_header header
;
4797 static int perf_event_task_match(struct perf_event
*event
)
4799 return event
->attr
.comm
|| event
->attr
.mmap
||
4800 event
->attr
.mmap2
|| event
->attr
.mmap_data
||
4804 static void perf_event_task_output(struct perf_event
*event
,
4807 struct perf_task_event
*task_event
= data
;
4808 struct perf_output_handle handle
;
4809 struct perf_sample_data sample
;
4810 struct task_struct
*task
= task_event
->task
;
4811 int ret
, size
= task_event
->event_id
.header
.size
;
4813 if (!perf_event_task_match(event
))
4816 perf_event_header__init_id(&task_event
->event_id
.header
, &sample
, event
);
4818 ret
= perf_output_begin(&handle
, event
,
4819 task_event
->event_id
.header
.size
);
4823 task_event
->event_id
.pid
= perf_event_pid(event
, task
);
4824 task_event
->event_id
.ppid
= perf_event_pid(event
, current
);
4826 task_event
->event_id
.tid
= perf_event_tid(event
, task
);
4827 task_event
->event_id
.ptid
= perf_event_tid(event
, current
);
4829 perf_output_put(&handle
, task_event
->event_id
);
4831 perf_event__output_id_sample(event
, &handle
, &sample
);
4833 perf_output_end(&handle
);
4835 task_event
->event_id
.header
.size
= size
;
4838 static void perf_event_task(struct task_struct
*task
,
4839 struct perf_event_context
*task_ctx
,
4842 struct perf_task_event task_event
;
4844 if (!atomic_read(&nr_comm_events
) &&
4845 !atomic_read(&nr_mmap_events
) &&
4846 !atomic_read(&nr_task_events
))
4849 task_event
= (struct perf_task_event
){
4851 .task_ctx
= task_ctx
,
4854 .type
= new ? PERF_RECORD_FORK
: PERF_RECORD_EXIT
,
4856 .size
= sizeof(task_event
.event_id
),
4862 .time
= perf_clock(),
4866 perf_event_aux(perf_event_task_output
,
4871 void perf_event_fork(struct task_struct
*task
)
4873 perf_event_task(task
, NULL
, 1);
4880 struct perf_comm_event
{
4881 struct task_struct
*task
;
4886 struct perf_event_header header
;
4893 static int perf_event_comm_match(struct perf_event
*event
)
4895 return event
->attr
.comm
;
4898 static void perf_event_comm_output(struct perf_event
*event
,
4901 struct perf_comm_event
*comm_event
= data
;
4902 struct perf_output_handle handle
;
4903 struct perf_sample_data sample
;
4904 int size
= comm_event
->event_id
.header
.size
;
4907 if (!perf_event_comm_match(event
))
4910 perf_event_header__init_id(&comm_event
->event_id
.header
, &sample
, event
);
4911 ret
= perf_output_begin(&handle
, event
,
4912 comm_event
->event_id
.header
.size
);
4917 comm_event
->event_id
.pid
= perf_event_pid(event
, comm_event
->task
);
4918 comm_event
->event_id
.tid
= perf_event_tid(event
, comm_event
->task
);
4920 perf_output_put(&handle
, comm_event
->event_id
);
4921 __output_copy(&handle
, comm_event
->comm
,
4922 comm_event
->comm_size
);
4924 perf_event__output_id_sample(event
, &handle
, &sample
);
4926 perf_output_end(&handle
);
4928 comm_event
->event_id
.header
.size
= size
;
4931 static void perf_event_comm_event(struct perf_comm_event
*comm_event
)
4933 char comm
[TASK_COMM_LEN
];
4936 memset(comm
, 0, sizeof(comm
));
4937 strlcpy(comm
, comm_event
->task
->comm
, sizeof(comm
));
4938 size
= ALIGN(strlen(comm
)+1, sizeof(u64
));
4940 comm_event
->comm
= comm
;
4941 comm_event
->comm_size
= size
;
4943 comm_event
->event_id
.header
.size
= sizeof(comm_event
->event_id
) + size
;
4945 perf_event_aux(perf_event_comm_output
,
4950 void perf_event_comm(struct task_struct
*task
)
4952 struct perf_comm_event comm_event
;
4953 struct perf_event_context
*ctx
;
4957 for_each_task_context_nr(ctxn
) {
4958 ctx
= task
->perf_event_ctxp
[ctxn
];
4962 perf_event_enable_on_exec(ctx
);
4966 if (!atomic_read(&nr_comm_events
))
4969 comm_event
= (struct perf_comm_event
){
4975 .type
= PERF_RECORD_COMM
,
4984 perf_event_comm_event(&comm_event
);
4991 struct perf_mmap_event
{
4992 struct vm_area_struct
*vma
;
4994 const char *file_name
;
5001 struct perf_event_header header
;
5011 static int perf_event_mmap_match(struct perf_event
*event
,
5014 struct perf_mmap_event
*mmap_event
= data
;
5015 struct vm_area_struct
*vma
= mmap_event
->vma
;
5016 int executable
= vma
->vm_flags
& VM_EXEC
;
5018 return (!executable
&& event
->attr
.mmap_data
) ||
5019 (executable
&& (event
->attr
.mmap
|| event
->attr
.mmap2
));
5022 static void perf_event_mmap_output(struct perf_event
*event
,
5025 struct perf_mmap_event
*mmap_event
= data
;
5026 struct perf_output_handle handle
;
5027 struct perf_sample_data sample
;
5028 int size
= mmap_event
->event_id
.header
.size
;
5031 if (!perf_event_mmap_match(event
, data
))
5034 if (event
->attr
.mmap2
) {
5035 mmap_event
->event_id
.header
.type
= PERF_RECORD_MMAP2
;
5036 mmap_event
->event_id
.header
.size
+= sizeof(mmap_event
->maj
);
5037 mmap_event
->event_id
.header
.size
+= sizeof(mmap_event
->min
);
5038 mmap_event
->event_id
.header
.size
+= sizeof(mmap_event
->ino
);
5041 perf_event_header__init_id(&mmap_event
->event_id
.header
, &sample
, event
);
5042 ret
= perf_output_begin(&handle
, event
,
5043 mmap_event
->event_id
.header
.size
);
5047 mmap_event
->event_id
.pid
= perf_event_pid(event
, current
);
5048 mmap_event
->event_id
.tid
= perf_event_tid(event
, current
);
5050 perf_output_put(&handle
, mmap_event
->event_id
);
5052 if (event
->attr
.mmap2
) {
5053 perf_output_put(&handle
, mmap_event
->maj
);
5054 perf_output_put(&handle
, mmap_event
->min
);
5055 perf_output_put(&handle
, mmap_event
->ino
);
5056 perf_output_put(&handle
, mmap_event
->ino_generation
);
5059 __output_copy(&handle
, mmap_event
->file_name
,
5060 mmap_event
->file_size
);
5062 perf_event__output_id_sample(event
, &handle
, &sample
);
5064 perf_output_end(&handle
);
5066 mmap_event
->event_id
.header
.size
= size
;
5069 static void perf_event_mmap_event(struct perf_mmap_event
*mmap_event
)
5071 struct vm_area_struct
*vma
= mmap_event
->vma
;
5072 struct file
*file
= vma
->vm_file
;
5073 int maj
= 0, min
= 0;
5074 u64 ino
= 0, gen
= 0;
5080 memset(tmp
, 0, sizeof(tmp
));
5083 struct inode
*inode
;
5086 * d_path works from the end of the rb backwards, so we
5087 * need to add enough zero bytes after the string to handle
5088 * the 64bit alignment we do later.
5090 buf
= kzalloc(PATH_MAX
+ sizeof(u64
), GFP_KERNEL
);
5092 name
= strncpy(tmp
, "//enomem", sizeof(tmp
));
5095 name
= d_path(&file
->f_path
, buf
, PATH_MAX
);
5097 name
= strncpy(tmp
, "//toolong", sizeof(tmp
));
5100 inode
= file_inode(vma
->vm_file
);
5101 dev
= inode
->i_sb
->s_dev
;
5103 gen
= inode
->i_generation
;
5108 if (arch_vma_name(mmap_event
->vma
)) {
5109 name
= strncpy(tmp
, arch_vma_name(mmap_event
->vma
),
5111 tmp
[sizeof(tmp
) - 1] = '\0';
5116 name
= strncpy(tmp
, "[vdso]", sizeof(tmp
));
5118 } else if (vma
->vm_start
<= vma
->vm_mm
->start_brk
&&
5119 vma
->vm_end
>= vma
->vm_mm
->brk
) {
5120 name
= strncpy(tmp
, "[heap]", sizeof(tmp
));
5122 } else if (vma
->vm_start
<= vma
->vm_mm
->start_stack
&&
5123 vma
->vm_end
>= vma
->vm_mm
->start_stack
) {
5124 name
= strncpy(tmp
, "[stack]", sizeof(tmp
));
5128 name
= strncpy(tmp
, "//anon", sizeof(tmp
));
5133 size
= ALIGN(strlen(name
)+1, sizeof(u64
));
5135 mmap_event
->file_name
= name
;
5136 mmap_event
->file_size
= size
;
5137 mmap_event
->maj
= maj
;
5138 mmap_event
->min
= min
;
5139 mmap_event
->ino
= ino
;
5140 mmap_event
->ino_generation
= gen
;
5142 if (!(vma
->vm_flags
& VM_EXEC
))
5143 mmap_event
->event_id
.header
.misc
|= PERF_RECORD_MISC_MMAP_DATA
;
5145 mmap_event
->event_id
.header
.size
= sizeof(mmap_event
->event_id
) + size
;
5147 perf_event_aux(perf_event_mmap_output
,
5154 void perf_event_mmap(struct vm_area_struct
*vma
)
5156 struct perf_mmap_event mmap_event
;
5158 if (!atomic_read(&nr_mmap_events
))
5161 mmap_event
= (struct perf_mmap_event
){
5167 .type
= PERF_RECORD_MMAP
,
5168 .misc
= PERF_RECORD_MISC_USER
,
5173 .start
= vma
->vm_start
,
5174 .len
= vma
->vm_end
- vma
->vm_start
,
5175 .pgoff
= (u64
)vma
->vm_pgoff
<< PAGE_SHIFT
,
5177 /* .maj (attr_mmap2 only) */
5178 /* .min (attr_mmap2 only) */
5179 /* .ino (attr_mmap2 only) */
5180 /* .ino_generation (attr_mmap2 only) */
5183 perf_event_mmap_event(&mmap_event
);
5187 * IRQ throttle logging
5190 static void perf_log_throttle(struct perf_event
*event
, int enable
)
5192 struct perf_output_handle handle
;
5193 struct perf_sample_data sample
;
5197 struct perf_event_header header
;
5201 } throttle_event
= {
5203 .type
= PERF_RECORD_THROTTLE
,
5205 .size
= sizeof(throttle_event
),
5207 .time
= perf_clock(),
5208 .id
= primary_event_id(event
),
5209 .stream_id
= event
->id
,
5213 throttle_event
.header
.type
= PERF_RECORD_UNTHROTTLE
;
5215 perf_event_header__init_id(&throttle_event
.header
, &sample
, event
);
5217 ret
= perf_output_begin(&handle
, event
,
5218 throttle_event
.header
.size
);
5222 perf_output_put(&handle
, throttle_event
);
5223 perf_event__output_id_sample(event
, &handle
, &sample
);
5224 perf_output_end(&handle
);
5228 * Generic event overflow handling, sampling.
5231 static int __perf_event_overflow(struct perf_event
*event
,
5232 int throttle
, struct perf_sample_data
*data
,
5233 struct pt_regs
*regs
)
5235 int events
= atomic_read(&event
->event_limit
);
5236 struct hw_perf_event
*hwc
= &event
->hw
;
5241 * Non-sampling counters might still use the PMI to fold short
5242 * hardware counters, ignore those.
5244 if (unlikely(!is_sampling_event(event
)))
5247 seq
= __this_cpu_read(perf_throttled_seq
);
5248 if (seq
!= hwc
->interrupts_seq
) {
5249 hwc
->interrupts_seq
= seq
;
5250 hwc
->interrupts
= 1;
5253 if (unlikely(throttle
5254 && hwc
->interrupts
>= max_samples_per_tick
)) {
5255 __this_cpu_inc(perf_throttled_count
);
5256 hwc
->interrupts
= MAX_INTERRUPTS
;
5257 perf_log_throttle(event
, 0);
5258 tick_nohz_full_kick();
5263 if (event
->attr
.freq
) {
5264 u64 now
= perf_clock();
5265 s64 delta
= now
- hwc
->freq_time_stamp
;
5267 hwc
->freq_time_stamp
= now
;
5269 if (delta
> 0 && delta
< 2*TICK_NSEC
)
5270 perf_adjust_period(event
, delta
, hwc
->last_period
, true);
5274 * XXX event_limit might not quite work as expected on inherited
5278 event
->pending_kill
= POLL_IN
;
5279 if (events
&& atomic_dec_and_test(&event
->event_limit
)) {
5281 event
->pending_kill
= POLL_HUP
;
5282 event
->pending_disable
= 1;
5283 irq_work_queue(&event
->pending
);
5286 if (event
->overflow_handler
)
5287 event
->overflow_handler(event
, data
, regs
);
5289 perf_event_output(event
, data
, regs
);
5291 if (event
->fasync
&& event
->pending_kill
) {
5292 event
->pending_wakeup
= 1;
5293 irq_work_queue(&event
->pending
);
5299 int perf_event_overflow(struct perf_event
*event
,
5300 struct perf_sample_data
*data
,
5301 struct pt_regs
*regs
)
5303 return __perf_event_overflow(event
, 1, data
, regs
);
5307 * Generic software event infrastructure
5310 struct swevent_htable
{
5311 struct swevent_hlist
*swevent_hlist
;
5312 struct mutex hlist_mutex
;
5315 /* Recursion avoidance in each contexts */
5316 int recursion
[PERF_NR_CONTEXTS
];
5319 static DEFINE_PER_CPU(struct swevent_htable
, swevent_htable
);
5322 * We directly increment event->count and keep a second value in
5323 * event->hw.period_left to count intervals. This period event
5324 * is kept in the range [-sample_period, 0] so that we can use the
5328 u64
perf_swevent_set_period(struct perf_event
*event
)
5330 struct hw_perf_event
*hwc
= &event
->hw
;
5331 u64 period
= hwc
->last_period
;
5335 hwc
->last_period
= hwc
->sample_period
;
5338 old
= val
= local64_read(&hwc
->period_left
);
5342 nr
= div64_u64(period
+ val
, period
);
5343 offset
= nr
* period
;
5345 if (local64_cmpxchg(&hwc
->period_left
, old
, val
) != old
)
5351 static void perf_swevent_overflow(struct perf_event
*event
, u64 overflow
,
5352 struct perf_sample_data
*data
,
5353 struct pt_regs
*regs
)
5355 struct hw_perf_event
*hwc
= &event
->hw
;
5359 overflow
= perf_swevent_set_period(event
);
5361 if (hwc
->interrupts
== MAX_INTERRUPTS
)
5364 for (; overflow
; overflow
--) {
5365 if (__perf_event_overflow(event
, throttle
,
5368 * We inhibit the overflow from happening when
5369 * hwc->interrupts == MAX_INTERRUPTS.
5377 static void perf_swevent_event(struct perf_event
*event
, u64 nr
,
5378 struct perf_sample_data
*data
,
5379 struct pt_regs
*regs
)
5381 struct hw_perf_event
*hwc
= &event
->hw
;
5383 local64_add(nr
, &event
->count
);
5388 if (!is_sampling_event(event
))
5391 if ((event
->attr
.sample_type
& PERF_SAMPLE_PERIOD
) && !event
->attr
.freq
) {
5393 return perf_swevent_overflow(event
, 1, data
, regs
);
5395 data
->period
= event
->hw
.last_period
;
5397 if (nr
== 1 && hwc
->sample_period
== 1 && !event
->attr
.freq
)
5398 return perf_swevent_overflow(event
, 1, data
, regs
);
5400 if (local64_add_negative(nr
, &hwc
->period_left
))
5403 perf_swevent_overflow(event
, 0, data
, regs
);
5406 static int perf_exclude_event(struct perf_event
*event
,
5407 struct pt_regs
*regs
)
5409 if (event
->hw
.state
& PERF_HES_STOPPED
)
5413 if (event
->attr
.exclude_user
&& user_mode(regs
))
5416 if (event
->attr
.exclude_kernel
&& !user_mode(regs
))
5423 static int perf_swevent_match(struct perf_event
*event
,
5424 enum perf_type_id type
,
5426 struct perf_sample_data
*data
,
5427 struct pt_regs
*regs
)
5429 if (event
->attr
.type
!= type
)
5432 if (event
->attr
.config
!= event_id
)
5435 if (perf_exclude_event(event
, regs
))
5441 static inline u64
swevent_hash(u64 type
, u32 event_id
)
5443 u64 val
= event_id
| (type
<< 32);
5445 return hash_64(val
, SWEVENT_HLIST_BITS
);
5448 static inline struct hlist_head
*
5449 __find_swevent_head(struct swevent_hlist
*hlist
, u64 type
, u32 event_id
)
5451 u64 hash
= swevent_hash(type
, event_id
);
5453 return &hlist
->heads
[hash
];
5456 /* For the read side: events when they trigger */
5457 static inline struct hlist_head
*
5458 find_swevent_head_rcu(struct swevent_htable
*swhash
, u64 type
, u32 event_id
)
5460 struct swevent_hlist
*hlist
;
5462 hlist
= rcu_dereference(swhash
->swevent_hlist
);
5466 return __find_swevent_head(hlist
, type
, event_id
);
5469 /* For the event head insertion and removal in the hlist */
5470 static inline struct hlist_head
*
5471 find_swevent_head(struct swevent_htable
*swhash
, struct perf_event
*event
)
5473 struct swevent_hlist
*hlist
;
5474 u32 event_id
= event
->attr
.config
;
5475 u64 type
= event
->attr
.type
;
5478 * Event scheduling is always serialized against hlist allocation
5479 * and release. Which makes the protected version suitable here.
5480 * The context lock guarantees that.
5482 hlist
= rcu_dereference_protected(swhash
->swevent_hlist
,
5483 lockdep_is_held(&event
->ctx
->lock
));
5487 return __find_swevent_head(hlist
, type
, event_id
);
5490 static void do_perf_sw_event(enum perf_type_id type
, u32 event_id
,
5492 struct perf_sample_data
*data
,
5493 struct pt_regs
*regs
)
5495 struct swevent_htable
*swhash
= &__get_cpu_var(swevent_htable
);
5496 struct perf_event
*event
;
5497 struct hlist_head
*head
;
5500 head
= find_swevent_head_rcu(swhash
, type
, event_id
);
5504 hlist_for_each_entry_rcu(event
, head
, hlist_entry
) {
5505 if (perf_swevent_match(event
, type
, event_id
, data
, regs
))
5506 perf_swevent_event(event
, nr
, data
, regs
);
5512 int perf_swevent_get_recursion_context(void)
5514 struct swevent_htable
*swhash
= &__get_cpu_var(swevent_htable
);
5516 return get_recursion_context(swhash
->recursion
);
5518 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context
);
5520 inline void perf_swevent_put_recursion_context(int rctx
)
5522 struct swevent_htable
*swhash
= &__get_cpu_var(swevent_htable
);
5524 put_recursion_context(swhash
->recursion
, rctx
);
5527 void __perf_sw_event(u32 event_id
, u64 nr
, struct pt_regs
*regs
, u64 addr
)
5529 struct perf_sample_data data
;
5532 preempt_disable_notrace();
5533 rctx
= perf_swevent_get_recursion_context();
5537 perf_sample_data_init(&data
, addr
, 0);
5539 do_perf_sw_event(PERF_TYPE_SOFTWARE
, event_id
, nr
, &data
, regs
);
5541 perf_swevent_put_recursion_context(rctx
);
5542 preempt_enable_notrace();
5545 static void perf_swevent_read(struct perf_event
*event
)
5549 static int perf_swevent_add(struct perf_event
*event
, int flags
)
5551 struct swevent_htable
*swhash
= &__get_cpu_var(swevent_htable
);
5552 struct hw_perf_event
*hwc
= &event
->hw
;
5553 struct hlist_head
*head
;
5555 if (is_sampling_event(event
)) {
5556 hwc
->last_period
= hwc
->sample_period
;
5557 perf_swevent_set_period(event
);
5560 hwc
->state
= !(flags
& PERF_EF_START
);
5562 head
= find_swevent_head(swhash
, event
);
5563 if (WARN_ON_ONCE(!head
))
5566 hlist_add_head_rcu(&event
->hlist_entry
, head
);
5571 static void perf_swevent_del(struct perf_event
*event
, int flags
)
5573 hlist_del_rcu(&event
->hlist_entry
);
5576 static void perf_swevent_start(struct perf_event
*event
, int flags
)
5578 event
->hw
.state
= 0;
5581 static void perf_swevent_stop(struct perf_event
*event
, int flags
)
5583 event
->hw
.state
= PERF_HES_STOPPED
;
5586 /* Deref the hlist from the update side */
5587 static inline struct swevent_hlist
*
5588 swevent_hlist_deref(struct swevent_htable
*swhash
)
5590 return rcu_dereference_protected(swhash
->swevent_hlist
,
5591 lockdep_is_held(&swhash
->hlist_mutex
));
5594 static void swevent_hlist_release(struct swevent_htable
*swhash
)
5596 struct swevent_hlist
*hlist
= swevent_hlist_deref(swhash
);
5601 rcu_assign_pointer(swhash
->swevent_hlist
, NULL
);
5602 kfree_rcu(hlist
, rcu_head
);
5605 static void swevent_hlist_put_cpu(struct perf_event
*event
, int cpu
)
5607 struct swevent_htable
*swhash
= &per_cpu(swevent_htable
, cpu
);
5609 mutex_lock(&swhash
->hlist_mutex
);
5611 if (!--swhash
->hlist_refcount
)
5612 swevent_hlist_release(swhash
);
5614 mutex_unlock(&swhash
->hlist_mutex
);
5617 static void swevent_hlist_put(struct perf_event
*event
)
5621 if (event
->cpu
!= -1) {
5622 swevent_hlist_put_cpu(event
, event
->cpu
);
5626 for_each_possible_cpu(cpu
)
5627 swevent_hlist_put_cpu(event
, cpu
);
5630 static int swevent_hlist_get_cpu(struct perf_event
*event
, int cpu
)
5632 struct swevent_htable
*swhash
= &per_cpu(swevent_htable
, cpu
);
5635 mutex_lock(&swhash
->hlist_mutex
);
5637 if (!swevent_hlist_deref(swhash
) && cpu_online(cpu
)) {
5638 struct swevent_hlist
*hlist
;
5640 hlist
= kzalloc(sizeof(*hlist
), GFP_KERNEL
);
5645 rcu_assign_pointer(swhash
->swevent_hlist
, hlist
);
5647 swhash
->hlist_refcount
++;
5649 mutex_unlock(&swhash
->hlist_mutex
);
5654 static int swevent_hlist_get(struct perf_event
*event
)
5657 int cpu
, failed_cpu
;
5659 if (event
->cpu
!= -1)
5660 return swevent_hlist_get_cpu(event
, event
->cpu
);
5663 for_each_possible_cpu(cpu
) {
5664 err
= swevent_hlist_get_cpu(event
, cpu
);
5674 for_each_possible_cpu(cpu
) {
5675 if (cpu
== failed_cpu
)
5677 swevent_hlist_put_cpu(event
, cpu
);
5684 struct static_key perf_swevent_enabled
[PERF_COUNT_SW_MAX
];
5686 static void sw_perf_event_destroy(struct perf_event
*event
)
5688 u64 event_id
= event
->attr
.config
;
5690 WARN_ON(event
->parent
);
5692 static_key_slow_dec(&perf_swevent_enabled
[event_id
]);
5693 swevent_hlist_put(event
);
5696 static int perf_swevent_init(struct perf_event
*event
)
5698 u64 event_id
= event
->attr
.config
;
5700 if (event
->attr
.type
!= PERF_TYPE_SOFTWARE
)
5704 * no branch sampling for software events
5706 if (has_branch_stack(event
))
5710 case PERF_COUNT_SW_CPU_CLOCK
:
5711 case PERF_COUNT_SW_TASK_CLOCK
:
5718 if (event_id
>= PERF_COUNT_SW_MAX
)
5721 if (!event
->parent
) {
5724 err
= swevent_hlist_get(event
);
5728 static_key_slow_inc(&perf_swevent_enabled
[event_id
]);
5729 event
->destroy
= sw_perf_event_destroy
;
5735 static int perf_swevent_event_idx(struct perf_event
*event
)
5740 static struct pmu perf_swevent
= {
5741 .task_ctx_nr
= perf_sw_context
,
5743 .event_init
= perf_swevent_init
,
5744 .add
= perf_swevent_add
,
5745 .del
= perf_swevent_del
,
5746 .start
= perf_swevent_start
,
5747 .stop
= perf_swevent_stop
,
5748 .read
= perf_swevent_read
,
5750 .event_idx
= perf_swevent_event_idx
,
5753 #ifdef CONFIG_EVENT_TRACING
5755 static int perf_tp_filter_match(struct perf_event
*event
,
5756 struct perf_sample_data
*data
)
5758 void *record
= data
->raw
->data
;
5760 if (likely(!event
->filter
) || filter_match_preds(event
->filter
, record
))
5765 static int perf_tp_event_match(struct perf_event
*event
,
5766 struct perf_sample_data
*data
,
5767 struct pt_regs
*regs
)
5769 if (event
->hw
.state
& PERF_HES_STOPPED
)
5772 * All tracepoints are from kernel-space.
5774 if (event
->attr
.exclude_kernel
)
5777 if (!perf_tp_filter_match(event
, data
))
5783 void perf_tp_event(u64 addr
, u64 count
, void *record
, int entry_size
,
5784 struct pt_regs
*regs
, struct hlist_head
*head
, int rctx
,
5785 struct task_struct
*task
)
5787 struct perf_sample_data data
;
5788 struct perf_event
*event
;
5790 struct perf_raw_record raw
= {
5795 perf_sample_data_init(&data
, addr
, 0);
5798 hlist_for_each_entry_rcu(event
, head
, hlist_entry
) {
5799 if (perf_tp_event_match(event
, &data
, regs
))
5800 perf_swevent_event(event
, count
, &data
, regs
);
5804 * If we got specified a target task, also iterate its context and
5805 * deliver this event there too.
5807 if (task
&& task
!= current
) {
5808 struct perf_event_context
*ctx
;
5809 struct trace_entry
*entry
= record
;
5812 ctx
= rcu_dereference(task
->perf_event_ctxp
[perf_sw_context
]);
5816 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
5817 if (event
->attr
.type
!= PERF_TYPE_TRACEPOINT
)
5819 if (event
->attr
.config
!= entry
->type
)
5821 if (perf_tp_event_match(event
, &data
, regs
))
5822 perf_swevent_event(event
, count
, &data
, regs
);
5828 perf_swevent_put_recursion_context(rctx
);
5830 EXPORT_SYMBOL_GPL(perf_tp_event
);
5832 static void tp_perf_event_destroy(struct perf_event
*event
)
5834 perf_trace_destroy(event
);
5837 static int perf_tp_event_init(struct perf_event
*event
)
5841 if (event
->attr
.type
!= PERF_TYPE_TRACEPOINT
)
5845 * no branch sampling for tracepoint events
5847 if (has_branch_stack(event
))
5850 err
= perf_trace_init(event
);
5854 event
->destroy
= tp_perf_event_destroy
;
5859 static struct pmu perf_tracepoint
= {
5860 .task_ctx_nr
= perf_sw_context
,
5862 .event_init
= perf_tp_event_init
,
5863 .add
= perf_trace_add
,
5864 .del
= perf_trace_del
,
5865 .start
= perf_swevent_start
,
5866 .stop
= perf_swevent_stop
,
5867 .read
= perf_swevent_read
,
5869 .event_idx
= perf_swevent_event_idx
,
5872 static inline void perf_tp_register(void)
5874 perf_pmu_register(&perf_tracepoint
, "tracepoint", PERF_TYPE_TRACEPOINT
);
5877 static int perf_event_set_filter(struct perf_event
*event
, void __user
*arg
)
5882 if (event
->attr
.type
!= PERF_TYPE_TRACEPOINT
)
5885 filter_str
= strndup_user(arg
, PAGE_SIZE
);
5886 if (IS_ERR(filter_str
))
5887 return PTR_ERR(filter_str
);
5889 ret
= ftrace_profile_set_filter(event
, event
->attr
.config
, filter_str
);
5895 static void perf_event_free_filter(struct perf_event
*event
)
5897 ftrace_profile_free_filter(event
);
5902 static inline void perf_tp_register(void)
5906 static int perf_event_set_filter(struct perf_event
*event
, void __user
*arg
)
5911 static void perf_event_free_filter(struct perf_event
*event
)
5915 #endif /* CONFIG_EVENT_TRACING */
5917 #ifdef CONFIG_HAVE_HW_BREAKPOINT
5918 void perf_bp_event(struct perf_event
*bp
, void *data
)
5920 struct perf_sample_data sample
;
5921 struct pt_regs
*regs
= data
;
5923 perf_sample_data_init(&sample
, bp
->attr
.bp_addr
, 0);
5925 if (!bp
->hw
.state
&& !perf_exclude_event(bp
, regs
))
5926 perf_swevent_event(bp
, 1, &sample
, regs
);
5931 * hrtimer based swevent callback
5934 static enum hrtimer_restart
perf_swevent_hrtimer(struct hrtimer
*hrtimer
)
5936 enum hrtimer_restart ret
= HRTIMER_RESTART
;
5937 struct perf_sample_data data
;
5938 struct pt_regs
*regs
;
5939 struct perf_event
*event
;
5942 event
= container_of(hrtimer
, struct perf_event
, hw
.hrtimer
);
5944 if (event
->state
!= PERF_EVENT_STATE_ACTIVE
)
5945 return HRTIMER_NORESTART
;
5947 event
->pmu
->read(event
);
5949 perf_sample_data_init(&data
, 0, event
->hw
.last_period
);
5950 regs
= get_irq_regs();
5952 if (regs
&& !perf_exclude_event(event
, regs
)) {
5953 if (!(event
->attr
.exclude_idle
&& is_idle_task(current
)))
5954 if (__perf_event_overflow(event
, 1, &data
, regs
))
5955 ret
= HRTIMER_NORESTART
;
5958 period
= max_t(u64
, 10000, event
->hw
.sample_period
);
5959 hrtimer_forward_now(hrtimer
, ns_to_ktime(period
));
5964 static void perf_swevent_start_hrtimer(struct perf_event
*event
)
5966 struct hw_perf_event
*hwc
= &event
->hw
;
5969 if (!is_sampling_event(event
))
5972 period
= local64_read(&hwc
->period_left
);
5977 local64_set(&hwc
->period_left
, 0);
5979 period
= max_t(u64
, 10000, hwc
->sample_period
);
5981 __hrtimer_start_range_ns(&hwc
->hrtimer
,
5982 ns_to_ktime(period
), 0,
5983 HRTIMER_MODE_REL_PINNED
, 0);
5986 static void perf_swevent_cancel_hrtimer(struct perf_event
*event
)
5988 struct hw_perf_event
*hwc
= &event
->hw
;
5990 if (is_sampling_event(event
)) {
5991 ktime_t remaining
= hrtimer_get_remaining(&hwc
->hrtimer
);
5992 local64_set(&hwc
->period_left
, ktime_to_ns(remaining
));
5994 hrtimer_cancel(&hwc
->hrtimer
);
5998 static void perf_swevent_init_hrtimer(struct perf_event
*event
)
6000 struct hw_perf_event
*hwc
= &event
->hw
;
6002 if (!is_sampling_event(event
))
6005 hrtimer_init(&hwc
->hrtimer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
6006 hwc
->hrtimer
.function
= perf_swevent_hrtimer
;
6009 * Since hrtimers have a fixed rate, we can do a static freq->period
6010 * mapping and avoid the whole period adjust feedback stuff.
6012 if (event
->attr
.freq
) {
6013 long freq
= event
->attr
.sample_freq
;
6015 event
->attr
.sample_period
= NSEC_PER_SEC
/ freq
;
6016 hwc
->sample_period
= event
->attr
.sample_period
;
6017 local64_set(&hwc
->period_left
, hwc
->sample_period
);
6018 hwc
->last_period
= hwc
->sample_period
;
6019 event
->attr
.freq
= 0;
6024 * Software event: cpu wall time clock
6027 static void cpu_clock_event_update(struct perf_event
*event
)
6032 now
= local_clock();
6033 prev
= local64_xchg(&event
->hw
.prev_count
, now
);
6034 local64_add(now
- prev
, &event
->count
);
6037 static void cpu_clock_event_start(struct perf_event
*event
, int flags
)
6039 local64_set(&event
->hw
.prev_count
, local_clock());
6040 perf_swevent_start_hrtimer(event
);
6043 static void cpu_clock_event_stop(struct perf_event
*event
, int flags
)
6045 perf_swevent_cancel_hrtimer(event
);
6046 cpu_clock_event_update(event
);
6049 static int cpu_clock_event_add(struct perf_event
*event
, int flags
)
6051 if (flags
& PERF_EF_START
)
6052 cpu_clock_event_start(event
, flags
);
6057 static void cpu_clock_event_del(struct perf_event
*event
, int flags
)
6059 cpu_clock_event_stop(event
, flags
);
6062 static void cpu_clock_event_read(struct perf_event
*event
)
6064 cpu_clock_event_update(event
);
6067 static int cpu_clock_event_init(struct perf_event
*event
)
6069 if (event
->attr
.type
!= PERF_TYPE_SOFTWARE
)
6072 if (event
->attr
.config
!= PERF_COUNT_SW_CPU_CLOCK
)
6076 * no branch sampling for software events
6078 if (has_branch_stack(event
))
6081 perf_swevent_init_hrtimer(event
);
6086 static struct pmu perf_cpu_clock
= {
6087 .task_ctx_nr
= perf_sw_context
,
6089 .event_init
= cpu_clock_event_init
,
6090 .add
= cpu_clock_event_add
,
6091 .del
= cpu_clock_event_del
,
6092 .start
= cpu_clock_event_start
,
6093 .stop
= cpu_clock_event_stop
,
6094 .read
= cpu_clock_event_read
,
6096 .event_idx
= perf_swevent_event_idx
,
6100 * Software event: task time clock
6103 static void task_clock_event_update(struct perf_event
*event
, u64 now
)
6108 prev
= local64_xchg(&event
->hw
.prev_count
, now
);
6110 local64_add(delta
, &event
->count
);
6113 static void task_clock_event_start(struct perf_event
*event
, int flags
)
6115 local64_set(&event
->hw
.prev_count
, event
->ctx
->time
);
6116 perf_swevent_start_hrtimer(event
);
6119 static void task_clock_event_stop(struct perf_event
*event
, int flags
)
6121 perf_swevent_cancel_hrtimer(event
);
6122 task_clock_event_update(event
, event
->ctx
->time
);
6125 static int task_clock_event_add(struct perf_event
*event
, int flags
)
6127 if (flags
& PERF_EF_START
)
6128 task_clock_event_start(event
, flags
);
6133 static void task_clock_event_del(struct perf_event
*event
, int flags
)
6135 task_clock_event_stop(event
, PERF_EF_UPDATE
);
6138 static void task_clock_event_read(struct perf_event
*event
)
6140 u64 now
= perf_clock();
6141 u64 delta
= now
- event
->ctx
->timestamp
;
6142 u64 time
= event
->ctx
->time
+ delta
;
6144 task_clock_event_update(event
, time
);
6147 static int task_clock_event_init(struct perf_event
*event
)
6149 if (event
->attr
.type
!= PERF_TYPE_SOFTWARE
)
6152 if (event
->attr
.config
!= PERF_COUNT_SW_TASK_CLOCK
)
6156 * no branch sampling for software events
6158 if (has_branch_stack(event
))
6161 perf_swevent_init_hrtimer(event
);
6166 static struct pmu perf_task_clock
= {
6167 .task_ctx_nr
= perf_sw_context
,
6169 .event_init
= task_clock_event_init
,
6170 .add
= task_clock_event_add
,
6171 .del
= task_clock_event_del
,
6172 .start
= task_clock_event_start
,
6173 .stop
= task_clock_event_stop
,
6174 .read
= task_clock_event_read
,
6176 .event_idx
= perf_swevent_event_idx
,
6179 static void perf_pmu_nop_void(struct pmu
*pmu
)
6183 static int perf_pmu_nop_int(struct pmu
*pmu
)
6188 static void perf_pmu_start_txn(struct pmu
*pmu
)
6190 perf_pmu_disable(pmu
);
6193 static int perf_pmu_commit_txn(struct pmu
*pmu
)
6195 perf_pmu_enable(pmu
);
6199 static void perf_pmu_cancel_txn(struct pmu
*pmu
)
6201 perf_pmu_enable(pmu
);
6204 static int perf_event_idx_default(struct perf_event
*event
)
6206 return event
->hw
.idx
+ 1;
6210 * Ensures all contexts with the same task_ctx_nr have the same
6211 * pmu_cpu_context too.
6213 static void *find_pmu_context(int ctxn
)
6220 list_for_each_entry(pmu
, &pmus
, entry
) {
6221 if (pmu
->task_ctx_nr
== ctxn
)
6222 return pmu
->pmu_cpu_context
;
6228 static void update_pmu_context(struct pmu
*pmu
, struct pmu
*old_pmu
)
6232 for_each_possible_cpu(cpu
) {
6233 struct perf_cpu_context
*cpuctx
;
6235 cpuctx
= per_cpu_ptr(pmu
->pmu_cpu_context
, cpu
);
6237 if (cpuctx
->unique_pmu
== old_pmu
)
6238 cpuctx
->unique_pmu
= pmu
;
6242 static void free_pmu_context(struct pmu
*pmu
)
6246 mutex_lock(&pmus_lock
);
6248 * Like a real lame refcount.
6250 list_for_each_entry(i
, &pmus
, entry
) {
6251 if (i
->pmu_cpu_context
== pmu
->pmu_cpu_context
) {
6252 update_pmu_context(i
, pmu
);
6257 free_percpu(pmu
->pmu_cpu_context
);
6259 mutex_unlock(&pmus_lock
);
6261 static struct idr pmu_idr
;
6264 type_show(struct device
*dev
, struct device_attribute
*attr
, char *page
)
6266 struct pmu
*pmu
= dev_get_drvdata(dev
);
6268 return snprintf(page
, PAGE_SIZE
-1, "%d\n", pmu
->type
);
6272 perf_event_mux_interval_ms_show(struct device
*dev
,
6273 struct device_attribute
*attr
,
6276 struct pmu
*pmu
= dev_get_drvdata(dev
);
6278 return snprintf(page
, PAGE_SIZE
-1, "%d\n", pmu
->hrtimer_interval_ms
);
6282 perf_event_mux_interval_ms_store(struct device
*dev
,
6283 struct device_attribute
*attr
,
6284 const char *buf
, size_t count
)
6286 struct pmu
*pmu
= dev_get_drvdata(dev
);
6287 int timer
, cpu
, ret
;
6289 ret
= kstrtoint(buf
, 0, &timer
);
6296 /* same value, noting to do */
6297 if (timer
== pmu
->hrtimer_interval_ms
)
6300 pmu
->hrtimer_interval_ms
= timer
;
6302 /* update all cpuctx for this PMU */
6303 for_each_possible_cpu(cpu
) {
6304 struct perf_cpu_context
*cpuctx
;
6305 cpuctx
= per_cpu_ptr(pmu
->pmu_cpu_context
, cpu
);
6306 cpuctx
->hrtimer_interval
= ns_to_ktime(NSEC_PER_MSEC
* timer
);
6308 if (hrtimer_active(&cpuctx
->hrtimer
))
6309 hrtimer_forward_now(&cpuctx
->hrtimer
, cpuctx
->hrtimer_interval
);
6315 static struct device_attribute pmu_dev_attrs
[] = {
6317 __ATTR_RW(perf_event_mux_interval_ms
),
6321 static int pmu_bus_running
;
6322 static struct bus_type pmu_bus
= {
6323 .name
= "event_source",
6324 .dev_attrs
= pmu_dev_attrs
,
6327 static void pmu_dev_release(struct device
*dev
)
6332 static int pmu_dev_alloc(struct pmu
*pmu
)
6336 pmu
->dev
= kzalloc(sizeof(struct device
), GFP_KERNEL
);
6340 pmu
->dev
->groups
= pmu
->attr_groups
;
6341 device_initialize(pmu
->dev
);
6342 ret
= dev_set_name(pmu
->dev
, "%s", pmu
->name
);
6346 dev_set_drvdata(pmu
->dev
, pmu
);
6347 pmu
->dev
->bus
= &pmu_bus
;
6348 pmu
->dev
->release
= pmu_dev_release
;
6349 ret
= device_add(pmu
->dev
);
6357 put_device(pmu
->dev
);
6361 static struct lock_class_key cpuctx_mutex
;
6362 static struct lock_class_key cpuctx_lock
;
6364 int perf_pmu_register(struct pmu
*pmu
, const char *name
, int type
)
6368 mutex_lock(&pmus_lock
);
6370 pmu
->pmu_disable_count
= alloc_percpu(int);
6371 if (!pmu
->pmu_disable_count
)
6380 type
= idr_alloc(&pmu_idr
, pmu
, PERF_TYPE_MAX
, 0, GFP_KERNEL
);
6388 if (pmu_bus_running
) {
6389 ret
= pmu_dev_alloc(pmu
);
6395 pmu
->pmu_cpu_context
= find_pmu_context(pmu
->task_ctx_nr
);
6396 if (pmu
->pmu_cpu_context
)
6397 goto got_cpu_context
;
6400 pmu
->pmu_cpu_context
= alloc_percpu(struct perf_cpu_context
);
6401 if (!pmu
->pmu_cpu_context
)
6404 for_each_possible_cpu(cpu
) {
6405 struct perf_cpu_context
*cpuctx
;
6407 cpuctx
= per_cpu_ptr(pmu
->pmu_cpu_context
, cpu
);
6408 __perf_event_init_context(&cpuctx
->ctx
);
6409 lockdep_set_class(&cpuctx
->ctx
.mutex
, &cpuctx_mutex
);
6410 lockdep_set_class(&cpuctx
->ctx
.lock
, &cpuctx_lock
);
6411 cpuctx
->ctx
.type
= cpu_context
;
6412 cpuctx
->ctx
.pmu
= pmu
;
6414 __perf_cpu_hrtimer_init(cpuctx
, cpu
);
6416 INIT_LIST_HEAD(&cpuctx
->rotation_list
);
6417 cpuctx
->unique_pmu
= pmu
;
6421 if (!pmu
->start_txn
) {
6422 if (pmu
->pmu_enable
) {
6424 * If we have pmu_enable/pmu_disable calls, install
6425 * transaction stubs that use that to try and batch
6426 * hardware accesses.
6428 pmu
->start_txn
= perf_pmu_start_txn
;
6429 pmu
->commit_txn
= perf_pmu_commit_txn
;
6430 pmu
->cancel_txn
= perf_pmu_cancel_txn
;
6432 pmu
->start_txn
= perf_pmu_nop_void
;
6433 pmu
->commit_txn
= perf_pmu_nop_int
;
6434 pmu
->cancel_txn
= perf_pmu_nop_void
;
6438 if (!pmu
->pmu_enable
) {
6439 pmu
->pmu_enable
= perf_pmu_nop_void
;
6440 pmu
->pmu_disable
= perf_pmu_nop_void
;
6443 if (!pmu
->event_idx
)
6444 pmu
->event_idx
= perf_event_idx_default
;
6446 list_add_rcu(&pmu
->entry
, &pmus
);
6449 mutex_unlock(&pmus_lock
);
6454 device_del(pmu
->dev
);
6455 put_device(pmu
->dev
);
6458 if (pmu
->type
>= PERF_TYPE_MAX
)
6459 idr_remove(&pmu_idr
, pmu
->type
);
6462 free_percpu(pmu
->pmu_disable_count
);
6466 void perf_pmu_unregister(struct pmu
*pmu
)
6468 mutex_lock(&pmus_lock
);
6469 list_del_rcu(&pmu
->entry
);
6470 mutex_unlock(&pmus_lock
);
6473 * We dereference the pmu list under both SRCU and regular RCU, so
6474 * synchronize against both of those.
6476 synchronize_srcu(&pmus_srcu
);
6479 free_percpu(pmu
->pmu_disable_count
);
6480 if (pmu
->type
>= PERF_TYPE_MAX
)
6481 idr_remove(&pmu_idr
, pmu
->type
);
6482 device_del(pmu
->dev
);
6483 put_device(pmu
->dev
);
6484 free_pmu_context(pmu
);
6487 struct pmu
*perf_init_event(struct perf_event
*event
)
6489 struct pmu
*pmu
= NULL
;
6493 idx
= srcu_read_lock(&pmus_srcu
);
6496 pmu
= idr_find(&pmu_idr
, event
->attr
.type
);
6500 ret
= pmu
->event_init(event
);
6506 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
6508 ret
= pmu
->event_init(event
);
6512 if (ret
!= -ENOENT
) {
6517 pmu
= ERR_PTR(-ENOENT
);
6519 srcu_read_unlock(&pmus_srcu
, idx
);
6524 static void account_event_cpu(struct perf_event
*event
, int cpu
)
6529 if (has_branch_stack(event
)) {
6530 if (!(event
->attach_state
& PERF_ATTACH_TASK
))
6531 atomic_inc(&per_cpu(perf_branch_stack_events
, cpu
));
6533 if (is_cgroup_event(event
))
6534 atomic_inc(&per_cpu(perf_cgroup_events
, cpu
));
6537 static void account_event(struct perf_event
*event
)
6542 if (event
->attach_state
& PERF_ATTACH_TASK
)
6543 static_key_slow_inc(&perf_sched_events
.key
);
6544 if (event
->attr
.mmap
|| event
->attr
.mmap_data
)
6545 atomic_inc(&nr_mmap_events
);
6546 if (event
->attr
.comm
)
6547 atomic_inc(&nr_comm_events
);
6548 if (event
->attr
.task
)
6549 atomic_inc(&nr_task_events
);
6550 if (event
->attr
.freq
) {
6551 if (atomic_inc_return(&nr_freq_events
) == 1)
6552 tick_nohz_full_kick_all();
6554 if (has_branch_stack(event
))
6555 static_key_slow_inc(&perf_sched_events
.key
);
6556 if (is_cgroup_event(event
))
6557 static_key_slow_inc(&perf_sched_events
.key
);
6559 account_event_cpu(event
, event
->cpu
);
6563 * Allocate and initialize a event structure
6565 static struct perf_event
*
6566 perf_event_alloc(struct perf_event_attr
*attr
, int cpu
,
6567 struct task_struct
*task
,
6568 struct perf_event
*group_leader
,
6569 struct perf_event
*parent_event
,
6570 perf_overflow_handler_t overflow_handler
,
6574 struct perf_event
*event
;
6575 struct hw_perf_event
*hwc
;
6578 if ((unsigned)cpu
>= nr_cpu_ids
) {
6579 if (!task
|| cpu
!= -1)
6580 return ERR_PTR(-EINVAL
);
6583 event
= kzalloc(sizeof(*event
), GFP_KERNEL
);
6585 return ERR_PTR(-ENOMEM
);
6588 * Single events are their own group leaders, with an
6589 * empty sibling list:
6592 group_leader
= event
;
6594 mutex_init(&event
->child_mutex
);
6595 INIT_LIST_HEAD(&event
->child_list
);
6597 INIT_LIST_HEAD(&event
->group_entry
);
6598 INIT_LIST_HEAD(&event
->event_entry
);
6599 INIT_LIST_HEAD(&event
->sibling_list
);
6600 INIT_LIST_HEAD(&event
->rb_entry
);
6602 init_waitqueue_head(&event
->waitq
);
6603 init_irq_work(&event
->pending
, perf_pending_event
);
6605 mutex_init(&event
->mmap_mutex
);
6607 atomic_long_set(&event
->refcount
, 1);
6609 event
->attr
= *attr
;
6610 event
->group_leader
= group_leader
;
6614 event
->parent
= parent_event
;
6616 event
->ns
= get_pid_ns(task_active_pid_ns(current
));
6617 event
->id
= atomic64_inc_return(&perf_event_id
);
6619 event
->state
= PERF_EVENT_STATE_INACTIVE
;
6622 event
->attach_state
= PERF_ATTACH_TASK
;
6624 if (attr
->type
== PERF_TYPE_TRACEPOINT
)
6625 event
->hw
.tp_target
= task
;
6626 #ifdef CONFIG_HAVE_HW_BREAKPOINT
6628 * hw_breakpoint is a bit difficult here..
6630 else if (attr
->type
== PERF_TYPE_BREAKPOINT
)
6631 event
->hw
.bp_target
= task
;
6635 if (!overflow_handler
&& parent_event
) {
6636 overflow_handler
= parent_event
->overflow_handler
;
6637 context
= parent_event
->overflow_handler_context
;
6640 event
->overflow_handler
= overflow_handler
;
6641 event
->overflow_handler_context
= context
;
6643 perf_event__state_init(event
);
6648 hwc
->sample_period
= attr
->sample_period
;
6649 if (attr
->freq
&& attr
->sample_freq
)
6650 hwc
->sample_period
= 1;
6651 hwc
->last_period
= hwc
->sample_period
;
6653 local64_set(&hwc
->period_left
, hwc
->sample_period
);
6656 * we currently do not support PERF_FORMAT_GROUP on inherited events
6658 if (attr
->inherit
&& (attr
->read_format
& PERF_FORMAT_GROUP
))
6661 pmu
= perf_init_event(event
);
6664 else if (IS_ERR(pmu
)) {
6669 if (!event
->parent
) {
6670 if (event
->attr
.sample_type
& PERF_SAMPLE_CALLCHAIN
) {
6671 err
= get_callchain_buffers();
6681 event
->destroy(event
);
6684 put_pid_ns(event
->ns
);
6687 return ERR_PTR(err
);
6690 static int perf_copy_attr(struct perf_event_attr __user
*uattr
,
6691 struct perf_event_attr
*attr
)
6696 if (!access_ok(VERIFY_WRITE
, uattr
, PERF_ATTR_SIZE_VER0
))
6700 * zero the full structure, so that a short copy will be nice.
6702 memset(attr
, 0, sizeof(*attr
));
6704 ret
= get_user(size
, &uattr
->size
);
6708 if (size
> PAGE_SIZE
) /* silly large */
6711 if (!size
) /* abi compat */
6712 size
= PERF_ATTR_SIZE_VER0
;
6714 if (size
< PERF_ATTR_SIZE_VER0
)
6718 * If we're handed a bigger struct than we know of,
6719 * ensure all the unknown bits are 0 - i.e. new
6720 * user-space does not rely on any kernel feature
6721 * extensions we dont know about yet.
6723 if (size
> sizeof(*attr
)) {
6724 unsigned char __user
*addr
;
6725 unsigned char __user
*end
;
6728 addr
= (void __user
*)uattr
+ sizeof(*attr
);
6729 end
= (void __user
*)uattr
+ size
;
6731 for (; addr
< end
; addr
++) {
6732 ret
= get_user(val
, addr
);
6738 size
= sizeof(*attr
);
6741 ret
= copy_from_user(attr
, uattr
, size
);
6745 if (attr
->__reserved_1
)
6748 if (attr
->sample_type
& ~(PERF_SAMPLE_MAX
-1))
6751 if (attr
->read_format
& ~(PERF_FORMAT_MAX
-1))
6754 if (attr
->sample_type
& PERF_SAMPLE_BRANCH_STACK
) {
6755 u64 mask
= attr
->branch_sample_type
;
6757 /* only using defined bits */
6758 if (mask
& ~(PERF_SAMPLE_BRANCH_MAX
-1))
6761 /* at least one branch bit must be set */
6762 if (!(mask
& ~PERF_SAMPLE_BRANCH_PLM_ALL
))
6765 /* propagate priv level, when not set for branch */
6766 if (!(mask
& PERF_SAMPLE_BRANCH_PLM_ALL
)) {
6768 /* exclude_kernel checked on syscall entry */
6769 if (!attr
->exclude_kernel
)
6770 mask
|= PERF_SAMPLE_BRANCH_KERNEL
;
6772 if (!attr
->exclude_user
)
6773 mask
|= PERF_SAMPLE_BRANCH_USER
;
6775 if (!attr
->exclude_hv
)
6776 mask
|= PERF_SAMPLE_BRANCH_HV
;
6778 * adjust user setting (for HW filter setup)
6780 attr
->branch_sample_type
= mask
;
6782 /* privileged levels capture (kernel, hv): check permissions */
6783 if ((mask
& PERF_SAMPLE_BRANCH_PERM_PLM
)
6784 && perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN
))
6788 if (attr
->sample_type
& PERF_SAMPLE_REGS_USER
) {
6789 ret
= perf_reg_validate(attr
->sample_regs_user
);
6794 if (attr
->sample_type
& PERF_SAMPLE_STACK_USER
) {
6795 if (!arch_perf_have_user_stack_dump())
6799 * We have __u32 type for the size, but so far
6800 * we can only use __u16 as maximum due to the
6801 * __u16 sample size limit.
6803 if (attr
->sample_stack_user
>= USHRT_MAX
)
6805 else if (!IS_ALIGNED(attr
->sample_stack_user
, sizeof(u64
)))
6813 put_user(sizeof(*attr
), &uattr
->size
);
6819 perf_event_set_output(struct perf_event
*event
, struct perf_event
*output_event
)
6821 struct ring_buffer
*rb
= NULL
, *old_rb
= NULL
;
6827 /* don't allow circular references */
6828 if (event
== output_event
)
6832 * Don't allow cross-cpu buffers
6834 if (output_event
->cpu
!= event
->cpu
)
6838 * If its not a per-cpu rb, it must be the same task.
6840 if (output_event
->cpu
== -1 && output_event
->ctx
!= event
->ctx
)
6844 mutex_lock(&event
->mmap_mutex
);
6845 /* Can't redirect output if we've got an active mmap() */
6846 if (atomic_read(&event
->mmap_count
))
6852 /* get the rb we want to redirect to */
6853 rb
= ring_buffer_get(output_event
);
6859 ring_buffer_detach(event
, old_rb
);
6862 ring_buffer_attach(event
, rb
);
6864 rcu_assign_pointer(event
->rb
, rb
);
6867 ring_buffer_put(old_rb
);
6869 * Since we detached before setting the new rb, so that we
6870 * could attach the new rb, we could have missed a wakeup.
6873 wake_up_all(&event
->waitq
);
6878 mutex_unlock(&event
->mmap_mutex
);
6885 * sys_perf_event_open - open a performance event, associate it to a task/cpu
6887 * @attr_uptr: event_id type attributes for monitoring/sampling
6890 * @group_fd: group leader event fd
6892 SYSCALL_DEFINE5(perf_event_open
,
6893 struct perf_event_attr __user
*, attr_uptr
,
6894 pid_t
, pid
, int, cpu
, int, group_fd
, unsigned long, flags
)
6896 struct perf_event
*group_leader
= NULL
, *output_event
= NULL
;
6897 struct perf_event
*event
, *sibling
;
6898 struct perf_event_attr attr
;
6899 struct perf_event_context
*ctx
;
6900 struct file
*event_file
= NULL
;
6901 struct fd group
= {NULL
, 0};
6902 struct task_struct
*task
= NULL
;
6908 /* for future expandability... */
6909 if (flags
& ~PERF_FLAG_ALL
)
6912 err
= perf_copy_attr(attr_uptr
, &attr
);
6916 if (!attr
.exclude_kernel
) {
6917 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN
))
6922 if (attr
.sample_freq
> sysctl_perf_event_sample_rate
)
6927 * In cgroup mode, the pid argument is used to pass the fd
6928 * opened to the cgroup directory in cgroupfs. The cpu argument
6929 * designates the cpu on which to monitor threads from that
6932 if ((flags
& PERF_FLAG_PID_CGROUP
) && (pid
== -1 || cpu
== -1))
6935 event_fd
= get_unused_fd();
6939 if (group_fd
!= -1) {
6940 err
= perf_fget_light(group_fd
, &group
);
6943 group_leader
= group
.file
->private_data
;
6944 if (flags
& PERF_FLAG_FD_OUTPUT
)
6945 output_event
= group_leader
;
6946 if (flags
& PERF_FLAG_FD_NO_GROUP
)
6947 group_leader
= NULL
;
6950 if (pid
!= -1 && !(flags
& PERF_FLAG_PID_CGROUP
)) {
6951 task
= find_lively_task_by_vpid(pid
);
6953 err
= PTR_ERR(task
);
6960 event
= perf_event_alloc(&attr
, cpu
, task
, group_leader
, NULL
,
6962 if (IS_ERR(event
)) {
6963 err
= PTR_ERR(event
);
6967 if (flags
& PERF_FLAG_PID_CGROUP
) {
6968 err
= perf_cgroup_connect(pid
, event
, &attr
, group_leader
);
6970 __free_event(event
);
6975 account_event(event
);
6978 * Special case software events and allow them to be part of
6979 * any hardware group.
6984 (is_software_event(event
) != is_software_event(group_leader
))) {
6985 if (is_software_event(event
)) {
6987 * If event and group_leader are not both a software
6988 * event, and event is, then group leader is not.
6990 * Allow the addition of software events to !software
6991 * groups, this is safe because software events never
6994 pmu
= group_leader
->pmu
;
6995 } else if (is_software_event(group_leader
) &&
6996 (group_leader
->group_flags
& PERF_GROUP_SOFTWARE
)) {
6998 * In case the group is a pure software group, and we
6999 * try to add a hardware event, move the whole group to
7000 * the hardware context.
7007 * Get the target context (task or percpu):
7009 ctx
= find_get_context(pmu
, task
, event
->cpu
);
7016 put_task_struct(task
);
7021 * Look up the group leader (we will attach this event to it):
7027 * Do not allow a recursive hierarchy (this new sibling
7028 * becoming part of another group-sibling):
7030 if (group_leader
->group_leader
!= group_leader
)
7033 * Do not allow to attach to a group in a different
7034 * task or CPU context:
7037 if (group_leader
->ctx
->type
!= ctx
->type
)
7040 if (group_leader
->ctx
!= ctx
)
7045 * Only a group leader can be exclusive or pinned
7047 if (attr
.exclusive
|| attr
.pinned
)
7052 err
= perf_event_set_output(event
, output_event
);
7057 event_file
= anon_inode_getfile("[perf_event]", &perf_fops
, event
, O_RDWR
);
7058 if (IS_ERR(event_file
)) {
7059 err
= PTR_ERR(event_file
);
7064 struct perf_event_context
*gctx
= group_leader
->ctx
;
7066 mutex_lock(&gctx
->mutex
);
7067 perf_remove_from_context(group_leader
);
7070 * Removing from the context ends up with disabled
7071 * event. What we want here is event in the initial
7072 * startup state, ready to be add into new context.
7074 perf_event__state_init(group_leader
);
7075 list_for_each_entry(sibling
, &group_leader
->sibling_list
,
7077 perf_remove_from_context(sibling
);
7078 perf_event__state_init(sibling
);
7081 mutex_unlock(&gctx
->mutex
);
7085 WARN_ON_ONCE(ctx
->parent_ctx
);
7086 mutex_lock(&ctx
->mutex
);
7090 perf_install_in_context(ctx
, group_leader
, event
->cpu
);
7092 list_for_each_entry(sibling
, &group_leader
->sibling_list
,
7094 perf_install_in_context(ctx
, sibling
, event
->cpu
);
7099 perf_install_in_context(ctx
, event
, event
->cpu
);
7101 perf_unpin_context(ctx
);
7102 mutex_unlock(&ctx
->mutex
);
7106 event
->owner
= current
;
7108 mutex_lock(¤t
->perf_event_mutex
);
7109 list_add_tail(&event
->owner_entry
, ¤t
->perf_event_list
);
7110 mutex_unlock(¤t
->perf_event_mutex
);
7113 * Precalculate sample_data sizes
7115 perf_event__header_size(event
);
7116 perf_event__id_header_size(event
);
7119 * Drop the reference on the group_event after placing the
7120 * new event on the sibling_list. This ensures destruction
7121 * of the group leader will find the pointer to itself in
7122 * perf_group_detach().
7125 fd_install(event_fd
, event_file
);
7129 perf_unpin_context(ctx
);
7136 put_task_struct(task
);
7140 put_unused_fd(event_fd
);
7145 * perf_event_create_kernel_counter
7147 * @attr: attributes of the counter to create
7148 * @cpu: cpu in which the counter is bound
7149 * @task: task to profile (NULL for percpu)
7152 perf_event_create_kernel_counter(struct perf_event_attr
*attr
, int cpu
,
7153 struct task_struct
*task
,
7154 perf_overflow_handler_t overflow_handler
,
7157 struct perf_event_context
*ctx
;
7158 struct perf_event
*event
;
7162 * Get the target context (task or percpu):
7165 event
= perf_event_alloc(attr
, cpu
, task
, NULL
, NULL
,
7166 overflow_handler
, context
);
7167 if (IS_ERR(event
)) {
7168 err
= PTR_ERR(event
);
7172 account_event(event
);
7174 ctx
= find_get_context(event
->pmu
, task
, cpu
);
7180 WARN_ON_ONCE(ctx
->parent_ctx
);
7181 mutex_lock(&ctx
->mutex
);
7182 perf_install_in_context(ctx
, event
, cpu
);
7184 perf_unpin_context(ctx
);
7185 mutex_unlock(&ctx
->mutex
);
7192 return ERR_PTR(err
);
7194 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter
);
7196 void perf_pmu_migrate_context(struct pmu
*pmu
, int src_cpu
, int dst_cpu
)
7198 struct perf_event_context
*src_ctx
;
7199 struct perf_event_context
*dst_ctx
;
7200 struct perf_event
*event
, *tmp
;
7203 src_ctx
= &per_cpu_ptr(pmu
->pmu_cpu_context
, src_cpu
)->ctx
;
7204 dst_ctx
= &per_cpu_ptr(pmu
->pmu_cpu_context
, dst_cpu
)->ctx
;
7206 mutex_lock(&src_ctx
->mutex
);
7207 list_for_each_entry_safe(event
, tmp
, &src_ctx
->event_list
,
7209 perf_remove_from_context(event
);
7210 unaccount_event_cpu(event
, src_cpu
);
7212 list_add(&event
->event_entry
, &events
);
7214 mutex_unlock(&src_ctx
->mutex
);
7218 mutex_lock(&dst_ctx
->mutex
);
7219 list_for_each_entry_safe(event
, tmp
, &events
, event_entry
) {
7220 list_del(&event
->event_entry
);
7221 if (event
->state
>= PERF_EVENT_STATE_OFF
)
7222 event
->state
= PERF_EVENT_STATE_INACTIVE
;
7223 account_event_cpu(event
, dst_cpu
);
7224 perf_install_in_context(dst_ctx
, event
, dst_cpu
);
7227 mutex_unlock(&dst_ctx
->mutex
);
7229 EXPORT_SYMBOL_GPL(perf_pmu_migrate_context
);
7231 static void sync_child_event(struct perf_event
*child_event
,
7232 struct task_struct
*child
)
7234 struct perf_event
*parent_event
= child_event
->parent
;
7237 if (child_event
->attr
.inherit_stat
)
7238 perf_event_read_event(child_event
, child
);
7240 child_val
= perf_event_count(child_event
);
7243 * Add back the child's count to the parent's count:
7245 atomic64_add(child_val
, &parent_event
->child_count
);
7246 atomic64_add(child_event
->total_time_enabled
,
7247 &parent_event
->child_total_time_enabled
);
7248 atomic64_add(child_event
->total_time_running
,
7249 &parent_event
->child_total_time_running
);
7252 * Remove this event from the parent's list
7254 WARN_ON_ONCE(parent_event
->ctx
->parent_ctx
);
7255 mutex_lock(&parent_event
->child_mutex
);
7256 list_del_init(&child_event
->child_list
);
7257 mutex_unlock(&parent_event
->child_mutex
);
7260 * Release the parent event, if this was the last
7263 put_event(parent_event
);
7267 __perf_event_exit_task(struct perf_event
*child_event
,
7268 struct perf_event_context
*child_ctx
,
7269 struct task_struct
*child
)
7271 if (child_event
->parent
) {
7272 raw_spin_lock_irq(&child_ctx
->lock
);
7273 perf_group_detach(child_event
);
7274 raw_spin_unlock_irq(&child_ctx
->lock
);
7277 perf_remove_from_context(child_event
);
7280 * It can happen that the parent exits first, and has events
7281 * that are still around due to the child reference. These
7282 * events need to be zapped.
7284 if (child_event
->parent
) {
7285 sync_child_event(child_event
, child
);
7286 free_event(child_event
);
7290 static void perf_event_exit_task_context(struct task_struct
*child
, int ctxn
)
7292 struct perf_event
*child_event
, *tmp
;
7293 struct perf_event_context
*child_ctx
;
7294 unsigned long flags
;
7296 if (likely(!child
->perf_event_ctxp
[ctxn
])) {
7297 perf_event_task(child
, NULL
, 0);
7301 local_irq_save(flags
);
7303 * We can't reschedule here because interrupts are disabled,
7304 * and either child is current or it is a task that can't be
7305 * scheduled, so we are now safe from rescheduling changing
7308 child_ctx
= rcu_dereference_raw(child
->perf_event_ctxp
[ctxn
]);
7311 * Take the context lock here so that if find_get_context is
7312 * reading child->perf_event_ctxp, we wait until it has
7313 * incremented the context's refcount before we do put_ctx below.
7315 raw_spin_lock(&child_ctx
->lock
);
7316 task_ctx_sched_out(child_ctx
);
7317 child
->perf_event_ctxp
[ctxn
] = NULL
;
7319 * If this context is a clone; unclone it so it can't get
7320 * swapped to another process while we're removing all
7321 * the events from it.
7323 unclone_ctx(child_ctx
);
7324 update_context_time(child_ctx
);
7325 raw_spin_unlock_irqrestore(&child_ctx
->lock
, flags
);
7328 * Report the task dead after unscheduling the events so that we
7329 * won't get any samples after PERF_RECORD_EXIT. We can however still
7330 * get a few PERF_RECORD_READ events.
7332 perf_event_task(child
, child_ctx
, 0);
7335 * We can recurse on the same lock type through:
7337 * __perf_event_exit_task()
7338 * sync_child_event()
7340 * mutex_lock(&ctx->mutex)
7342 * But since its the parent context it won't be the same instance.
7344 mutex_lock(&child_ctx
->mutex
);
7347 list_for_each_entry_safe(child_event
, tmp
, &child_ctx
->pinned_groups
,
7349 __perf_event_exit_task(child_event
, child_ctx
, child
);
7351 list_for_each_entry_safe(child_event
, tmp
, &child_ctx
->flexible_groups
,
7353 __perf_event_exit_task(child_event
, child_ctx
, child
);
7356 * If the last event was a group event, it will have appended all
7357 * its siblings to the list, but we obtained 'tmp' before that which
7358 * will still point to the list head terminating the iteration.
7360 if (!list_empty(&child_ctx
->pinned_groups
) ||
7361 !list_empty(&child_ctx
->flexible_groups
))
7364 mutex_unlock(&child_ctx
->mutex
);
7370 * When a child task exits, feed back event values to parent events.
7372 void perf_event_exit_task(struct task_struct
*child
)
7374 struct perf_event
*event
, *tmp
;
7377 mutex_lock(&child
->perf_event_mutex
);
7378 list_for_each_entry_safe(event
, tmp
, &child
->perf_event_list
,
7380 list_del_init(&event
->owner_entry
);
7383 * Ensure the list deletion is visible before we clear
7384 * the owner, closes a race against perf_release() where
7385 * we need to serialize on the owner->perf_event_mutex.
7388 event
->owner
= NULL
;
7390 mutex_unlock(&child
->perf_event_mutex
);
7392 for_each_task_context_nr(ctxn
)
7393 perf_event_exit_task_context(child
, ctxn
);
7396 static void perf_free_event(struct perf_event
*event
,
7397 struct perf_event_context
*ctx
)
7399 struct perf_event
*parent
= event
->parent
;
7401 if (WARN_ON_ONCE(!parent
))
7404 mutex_lock(&parent
->child_mutex
);
7405 list_del_init(&event
->child_list
);
7406 mutex_unlock(&parent
->child_mutex
);
7410 perf_group_detach(event
);
7411 list_del_event(event
, ctx
);
7416 * free an unexposed, unused context as created by inheritance by
7417 * perf_event_init_task below, used by fork() in case of fail.
7419 void perf_event_free_task(struct task_struct
*task
)
7421 struct perf_event_context
*ctx
;
7422 struct perf_event
*event
, *tmp
;
7425 for_each_task_context_nr(ctxn
) {
7426 ctx
= task
->perf_event_ctxp
[ctxn
];
7430 mutex_lock(&ctx
->mutex
);
7432 list_for_each_entry_safe(event
, tmp
, &ctx
->pinned_groups
,
7434 perf_free_event(event
, ctx
);
7436 list_for_each_entry_safe(event
, tmp
, &ctx
->flexible_groups
,
7438 perf_free_event(event
, ctx
);
7440 if (!list_empty(&ctx
->pinned_groups
) ||
7441 !list_empty(&ctx
->flexible_groups
))
7444 mutex_unlock(&ctx
->mutex
);
7450 void perf_event_delayed_put(struct task_struct
*task
)
7454 for_each_task_context_nr(ctxn
)
7455 WARN_ON_ONCE(task
->perf_event_ctxp
[ctxn
]);
7459 * inherit a event from parent task to child task:
7461 static struct perf_event
*
7462 inherit_event(struct perf_event
*parent_event
,
7463 struct task_struct
*parent
,
7464 struct perf_event_context
*parent_ctx
,
7465 struct task_struct
*child
,
7466 struct perf_event
*group_leader
,
7467 struct perf_event_context
*child_ctx
)
7469 struct perf_event
*child_event
;
7470 unsigned long flags
;
7473 * Instead of creating recursive hierarchies of events,
7474 * we link inherited events back to the original parent,
7475 * which has a filp for sure, which we use as the reference
7478 if (parent_event
->parent
)
7479 parent_event
= parent_event
->parent
;
7481 child_event
= perf_event_alloc(&parent_event
->attr
,
7484 group_leader
, parent_event
,
7486 if (IS_ERR(child_event
))
7489 if (!atomic_long_inc_not_zero(&parent_event
->refcount
)) {
7490 free_event(child_event
);
7497 * Make the child state follow the state of the parent event,
7498 * not its attr.disabled bit. We hold the parent's mutex,
7499 * so we won't race with perf_event_{en, dis}able_family.
7501 if (parent_event
->state
>= PERF_EVENT_STATE_INACTIVE
)
7502 child_event
->state
= PERF_EVENT_STATE_INACTIVE
;
7504 child_event
->state
= PERF_EVENT_STATE_OFF
;
7506 if (parent_event
->attr
.freq
) {
7507 u64 sample_period
= parent_event
->hw
.sample_period
;
7508 struct hw_perf_event
*hwc
= &child_event
->hw
;
7510 hwc
->sample_period
= sample_period
;
7511 hwc
->last_period
= sample_period
;
7513 local64_set(&hwc
->period_left
, sample_period
);
7516 child_event
->ctx
= child_ctx
;
7517 child_event
->overflow_handler
= parent_event
->overflow_handler
;
7518 child_event
->overflow_handler_context
7519 = parent_event
->overflow_handler_context
;
7522 * Precalculate sample_data sizes
7524 perf_event__header_size(child_event
);
7525 perf_event__id_header_size(child_event
);
7528 * Link it up in the child's context:
7530 raw_spin_lock_irqsave(&child_ctx
->lock
, flags
);
7531 add_event_to_ctx(child_event
, child_ctx
);
7532 raw_spin_unlock_irqrestore(&child_ctx
->lock
, flags
);
7535 * Link this into the parent event's child list
7537 WARN_ON_ONCE(parent_event
->ctx
->parent_ctx
);
7538 mutex_lock(&parent_event
->child_mutex
);
7539 list_add_tail(&child_event
->child_list
, &parent_event
->child_list
);
7540 mutex_unlock(&parent_event
->child_mutex
);
7545 static int inherit_group(struct perf_event
*parent_event
,
7546 struct task_struct
*parent
,
7547 struct perf_event_context
*parent_ctx
,
7548 struct task_struct
*child
,
7549 struct perf_event_context
*child_ctx
)
7551 struct perf_event
*leader
;
7552 struct perf_event
*sub
;
7553 struct perf_event
*child_ctr
;
7555 leader
= inherit_event(parent_event
, parent
, parent_ctx
,
7556 child
, NULL
, child_ctx
);
7558 return PTR_ERR(leader
);
7559 list_for_each_entry(sub
, &parent_event
->sibling_list
, group_entry
) {
7560 child_ctr
= inherit_event(sub
, parent
, parent_ctx
,
7561 child
, leader
, child_ctx
);
7562 if (IS_ERR(child_ctr
))
7563 return PTR_ERR(child_ctr
);
7569 inherit_task_group(struct perf_event
*event
, struct task_struct
*parent
,
7570 struct perf_event_context
*parent_ctx
,
7571 struct task_struct
*child
, int ctxn
,
7575 struct perf_event_context
*child_ctx
;
7577 if (!event
->attr
.inherit
) {
7582 child_ctx
= child
->perf_event_ctxp
[ctxn
];
7585 * This is executed from the parent task context, so
7586 * inherit events that have been marked for cloning.
7587 * First allocate and initialize a context for the
7591 child_ctx
= alloc_perf_context(parent_ctx
->pmu
, child
);
7595 child
->perf_event_ctxp
[ctxn
] = child_ctx
;
7598 ret
= inherit_group(event
, parent
, parent_ctx
,
7608 * Initialize the perf_event context in task_struct
7610 int perf_event_init_context(struct task_struct
*child
, int ctxn
)
7612 struct perf_event_context
*child_ctx
, *parent_ctx
;
7613 struct perf_event_context
*cloned_ctx
;
7614 struct perf_event
*event
;
7615 struct task_struct
*parent
= current
;
7616 int inherited_all
= 1;
7617 unsigned long flags
;
7620 if (likely(!parent
->perf_event_ctxp
[ctxn
]))
7624 * If the parent's context is a clone, pin it so it won't get
7627 parent_ctx
= perf_pin_task_context(parent
, ctxn
);
7630 * No need to check if parent_ctx != NULL here; since we saw
7631 * it non-NULL earlier, the only reason for it to become NULL
7632 * is if we exit, and since we're currently in the middle of
7633 * a fork we can't be exiting at the same time.
7637 * Lock the parent list. No need to lock the child - not PID
7638 * hashed yet and not running, so nobody can access it.
7640 mutex_lock(&parent_ctx
->mutex
);
7643 * We dont have to disable NMIs - we are only looking at
7644 * the list, not manipulating it:
7646 list_for_each_entry(event
, &parent_ctx
->pinned_groups
, group_entry
) {
7647 ret
= inherit_task_group(event
, parent
, parent_ctx
,
7648 child
, ctxn
, &inherited_all
);
7654 * We can't hold ctx->lock when iterating the ->flexible_group list due
7655 * to allocations, but we need to prevent rotation because
7656 * rotate_ctx() will change the list from interrupt context.
7658 raw_spin_lock_irqsave(&parent_ctx
->lock
, flags
);
7659 parent_ctx
->rotate_disable
= 1;
7660 raw_spin_unlock_irqrestore(&parent_ctx
->lock
, flags
);
7662 list_for_each_entry(event
, &parent_ctx
->flexible_groups
, group_entry
) {
7663 ret
= inherit_task_group(event
, parent
, parent_ctx
,
7664 child
, ctxn
, &inherited_all
);
7669 raw_spin_lock_irqsave(&parent_ctx
->lock
, flags
);
7670 parent_ctx
->rotate_disable
= 0;
7672 child_ctx
= child
->perf_event_ctxp
[ctxn
];
7674 if (child_ctx
&& inherited_all
) {
7676 * Mark the child context as a clone of the parent
7677 * context, or of whatever the parent is a clone of.
7679 * Note that if the parent is a clone, the holding of
7680 * parent_ctx->lock avoids it from being uncloned.
7682 cloned_ctx
= parent_ctx
->parent_ctx
;
7684 child_ctx
->parent_ctx
= cloned_ctx
;
7685 child_ctx
->parent_gen
= parent_ctx
->parent_gen
;
7687 child_ctx
->parent_ctx
= parent_ctx
;
7688 child_ctx
->parent_gen
= parent_ctx
->generation
;
7690 get_ctx(child_ctx
->parent_ctx
);
7693 raw_spin_unlock_irqrestore(&parent_ctx
->lock
, flags
);
7694 mutex_unlock(&parent_ctx
->mutex
);
7696 perf_unpin_context(parent_ctx
);
7697 put_ctx(parent_ctx
);
7703 * Initialize the perf_event context in task_struct
7705 int perf_event_init_task(struct task_struct
*child
)
7709 memset(child
->perf_event_ctxp
, 0, sizeof(child
->perf_event_ctxp
));
7710 mutex_init(&child
->perf_event_mutex
);
7711 INIT_LIST_HEAD(&child
->perf_event_list
);
7713 for_each_task_context_nr(ctxn
) {
7714 ret
= perf_event_init_context(child
, ctxn
);
7722 static void __init
perf_event_init_all_cpus(void)
7724 struct swevent_htable
*swhash
;
7727 for_each_possible_cpu(cpu
) {
7728 swhash
= &per_cpu(swevent_htable
, cpu
);
7729 mutex_init(&swhash
->hlist_mutex
);
7730 INIT_LIST_HEAD(&per_cpu(rotation_list
, cpu
));
7734 static void perf_event_init_cpu(int cpu
)
7736 struct swevent_htable
*swhash
= &per_cpu(swevent_htable
, cpu
);
7738 mutex_lock(&swhash
->hlist_mutex
);
7739 if (swhash
->hlist_refcount
> 0) {
7740 struct swevent_hlist
*hlist
;
7742 hlist
= kzalloc_node(sizeof(*hlist
), GFP_KERNEL
, cpu_to_node(cpu
));
7744 rcu_assign_pointer(swhash
->swevent_hlist
, hlist
);
7746 mutex_unlock(&swhash
->hlist_mutex
);
7749 #if defined CONFIG_HOTPLUG_CPU || defined CONFIG_KEXEC
7750 static void perf_pmu_rotate_stop(struct pmu
*pmu
)
7752 struct perf_cpu_context
*cpuctx
= this_cpu_ptr(pmu
->pmu_cpu_context
);
7754 WARN_ON(!irqs_disabled());
7756 list_del_init(&cpuctx
->rotation_list
);
7759 static void __perf_event_exit_context(void *__info
)
7761 struct perf_event_context
*ctx
= __info
;
7762 struct perf_event
*event
, *tmp
;
7764 perf_pmu_rotate_stop(ctx
->pmu
);
7766 list_for_each_entry_safe(event
, tmp
, &ctx
->pinned_groups
, group_entry
)
7767 __perf_remove_from_context(event
);
7768 list_for_each_entry_safe(event
, tmp
, &ctx
->flexible_groups
, group_entry
)
7769 __perf_remove_from_context(event
);
7772 static void perf_event_exit_cpu_context(int cpu
)
7774 struct perf_event_context
*ctx
;
7778 idx
= srcu_read_lock(&pmus_srcu
);
7779 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
7780 ctx
= &per_cpu_ptr(pmu
->pmu_cpu_context
, cpu
)->ctx
;
7782 mutex_lock(&ctx
->mutex
);
7783 smp_call_function_single(cpu
, __perf_event_exit_context
, ctx
, 1);
7784 mutex_unlock(&ctx
->mutex
);
7786 srcu_read_unlock(&pmus_srcu
, idx
);
7789 static void perf_event_exit_cpu(int cpu
)
7791 struct swevent_htable
*swhash
= &per_cpu(swevent_htable
, cpu
);
7793 mutex_lock(&swhash
->hlist_mutex
);
7794 swevent_hlist_release(swhash
);
7795 mutex_unlock(&swhash
->hlist_mutex
);
7797 perf_event_exit_cpu_context(cpu
);
7800 static inline void perf_event_exit_cpu(int cpu
) { }
7804 perf_reboot(struct notifier_block
*notifier
, unsigned long val
, void *v
)
7808 for_each_online_cpu(cpu
)
7809 perf_event_exit_cpu(cpu
);
7815 * Run the perf reboot notifier at the very last possible moment so that
7816 * the generic watchdog code runs as long as possible.
7818 static struct notifier_block perf_reboot_notifier
= {
7819 .notifier_call
= perf_reboot
,
7820 .priority
= INT_MIN
,
7824 perf_cpu_notify(struct notifier_block
*self
, unsigned long action
, void *hcpu
)
7826 unsigned int cpu
= (long)hcpu
;
7828 switch (action
& ~CPU_TASKS_FROZEN
) {
7830 case CPU_UP_PREPARE
:
7831 case CPU_DOWN_FAILED
:
7832 perf_event_init_cpu(cpu
);
7835 case CPU_UP_CANCELED
:
7836 case CPU_DOWN_PREPARE
:
7837 perf_event_exit_cpu(cpu
);
7846 void __init
perf_event_init(void)
7852 perf_event_init_all_cpus();
7853 init_srcu_struct(&pmus_srcu
);
7854 perf_pmu_register(&perf_swevent
, "software", PERF_TYPE_SOFTWARE
);
7855 perf_pmu_register(&perf_cpu_clock
, NULL
, -1);
7856 perf_pmu_register(&perf_task_clock
, NULL
, -1);
7858 perf_cpu_notifier(perf_cpu_notify
);
7859 register_reboot_notifier(&perf_reboot_notifier
);
7861 ret
= init_hw_breakpoint();
7862 WARN(ret
, "hw_breakpoint initialization failed with: %d", ret
);
7864 /* do not patch jump label more than once per second */
7865 jump_label_rate_limit(&perf_sched_events
, HZ
);
7868 * Build time assertion that we keep the data_head at the intended
7869 * location. IOW, validation we got the __reserved[] size right.
7871 BUILD_BUG_ON((offsetof(struct perf_event_mmap_page
, data_head
))
7875 static int __init
perf_event_sysfs_init(void)
7880 mutex_lock(&pmus_lock
);
7882 ret
= bus_register(&pmu_bus
);
7886 list_for_each_entry(pmu
, &pmus
, entry
) {
7887 if (!pmu
->name
|| pmu
->type
< 0)
7890 ret
= pmu_dev_alloc(pmu
);
7891 WARN(ret
, "Failed to register pmu: %s, reason %d\n", pmu
->name
, ret
);
7893 pmu_bus_running
= 1;
7897 mutex_unlock(&pmus_lock
);
7901 device_initcall(perf_event_sysfs_init
);
7903 #ifdef CONFIG_CGROUP_PERF
7904 static struct cgroup_subsys_state
*perf_cgroup_css_alloc(struct cgroup
*cont
)
7906 struct perf_cgroup
*jc
;
7908 jc
= kzalloc(sizeof(*jc
), GFP_KERNEL
);
7910 return ERR_PTR(-ENOMEM
);
7912 jc
->info
= alloc_percpu(struct perf_cgroup_info
);
7915 return ERR_PTR(-ENOMEM
);
7921 static void perf_cgroup_css_free(struct cgroup
*cont
)
7923 struct perf_cgroup
*jc
;
7924 jc
= container_of(cgroup_subsys_state(cont
, perf_subsys_id
),
7925 struct perf_cgroup
, css
);
7926 free_percpu(jc
->info
);
7930 static int __perf_cgroup_move(void *info
)
7932 struct task_struct
*task
= info
;
7933 perf_cgroup_switch(task
, PERF_CGROUP_SWOUT
| PERF_CGROUP_SWIN
);
7937 static void perf_cgroup_attach(struct cgroup
*cgrp
, struct cgroup_taskset
*tset
)
7939 struct task_struct
*task
;
7941 cgroup_taskset_for_each(task
, cgrp
, tset
)
7942 task_function_call(task
, __perf_cgroup_move
, task
);
7945 static void perf_cgroup_exit(struct cgroup
*cgrp
, struct cgroup
*old_cgrp
,
7946 struct task_struct
*task
)
7949 * cgroup_exit() is called in the copy_process() failure path.
7950 * Ignore this case since the task hasn't ran yet, this avoids
7951 * trying to poke a half freed task state from generic code.
7953 if (!(task
->flags
& PF_EXITING
))
7956 task_function_call(task
, __perf_cgroup_move
, task
);
7959 struct cgroup_subsys perf_subsys
= {
7960 .name
= "perf_event",
7961 .subsys_id
= perf_subsys_id
,
7962 .css_alloc
= perf_cgroup_css_alloc
,
7963 .css_free
= perf_cgroup_css_free
,
7964 .exit
= perf_cgroup_exit
,
7965 .attach
= perf_cgroup_attach
,
7967 #endif /* CONFIG_CGROUP_PERF */