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 |\
123 PERF_FLAG_FD_CLOEXEC)
126 * branch priv levels that need permission checks
128 #define PERF_SAMPLE_BRANCH_PERM_PLM \
129 (PERF_SAMPLE_BRANCH_KERNEL |\
130 PERF_SAMPLE_BRANCH_HV)
133 EVENT_FLEXIBLE
= 0x1,
135 EVENT_ALL
= EVENT_FLEXIBLE
| EVENT_PINNED
,
139 * perf_sched_events : >0 events exist
140 * perf_cgroup_events: >0 per-cpu cgroup events exist on this cpu
142 struct static_key_deferred perf_sched_events __read_mostly
;
143 static DEFINE_PER_CPU(atomic_t
, perf_cgroup_events
);
144 static DEFINE_PER_CPU(atomic_t
, perf_branch_stack_events
);
146 static atomic_t nr_mmap_events __read_mostly
;
147 static atomic_t nr_comm_events __read_mostly
;
148 static atomic_t nr_task_events __read_mostly
;
149 static atomic_t nr_freq_events __read_mostly
;
151 static LIST_HEAD(pmus
);
152 static DEFINE_MUTEX(pmus_lock
);
153 static struct srcu_struct pmus_srcu
;
156 * perf event paranoia level:
157 * -1 - not paranoid at all
158 * 0 - disallow raw tracepoint access for unpriv
159 * 1 - disallow cpu events for unpriv
160 * 2 - disallow kernel profiling for unpriv
162 int sysctl_perf_event_paranoid __read_mostly
= 1;
164 /* Minimum for 512 kiB + 1 user control page */
165 int sysctl_perf_event_mlock __read_mostly
= 512 + (PAGE_SIZE
/ 1024); /* 'free' kiB per user */
168 * max perf event sample rate
170 #define DEFAULT_MAX_SAMPLE_RATE 100000
171 #define DEFAULT_SAMPLE_PERIOD_NS (NSEC_PER_SEC / DEFAULT_MAX_SAMPLE_RATE)
172 #define DEFAULT_CPU_TIME_MAX_PERCENT 25
174 int sysctl_perf_event_sample_rate __read_mostly
= DEFAULT_MAX_SAMPLE_RATE
;
176 static int max_samples_per_tick __read_mostly
= DIV_ROUND_UP(DEFAULT_MAX_SAMPLE_RATE
, HZ
);
177 static int perf_sample_period_ns __read_mostly
= DEFAULT_SAMPLE_PERIOD_NS
;
179 static int perf_sample_allowed_ns __read_mostly
=
180 DEFAULT_SAMPLE_PERIOD_NS
* DEFAULT_CPU_TIME_MAX_PERCENT
/ 100;
182 void update_perf_cpu_limits(void)
184 u64 tmp
= perf_sample_period_ns
;
186 tmp
*= sysctl_perf_cpu_time_max_percent
;
188 ACCESS_ONCE(perf_sample_allowed_ns
) = tmp
;
191 static int perf_rotate_context(struct perf_cpu_context
*cpuctx
);
193 int perf_proc_update_handler(struct ctl_table
*table
, int write
,
194 void __user
*buffer
, size_t *lenp
,
197 int ret
= proc_dointvec_minmax(table
, write
, buffer
, lenp
, ppos
);
202 max_samples_per_tick
= DIV_ROUND_UP(sysctl_perf_event_sample_rate
, HZ
);
203 perf_sample_period_ns
= NSEC_PER_SEC
/ sysctl_perf_event_sample_rate
;
204 update_perf_cpu_limits();
209 int sysctl_perf_cpu_time_max_percent __read_mostly
= DEFAULT_CPU_TIME_MAX_PERCENT
;
211 int perf_cpu_time_max_percent_handler(struct ctl_table
*table
, int write
,
212 void __user
*buffer
, size_t *lenp
,
215 int ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
220 update_perf_cpu_limits();
226 * perf samples are done in some very critical code paths (NMIs).
227 * If they take too much CPU time, the system can lock up and not
228 * get any real work done. This will drop the sample rate when
229 * we detect that events are taking too long.
231 #define NR_ACCUMULATED_SAMPLES 128
232 static DEFINE_PER_CPU(u64
, running_sample_length
);
234 static void perf_duration_warn(struct irq_work
*w
)
236 u64 allowed_ns
= ACCESS_ONCE(perf_sample_allowed_ns
);
237 u64 avg_local_sample_len
;
238 u64 local_samples_len
;
240 local_samples_len
= __get_cpu_var(running_sample_length
);
241 avg_local_sample_len
= local_samples_len
/NR_ACCUMULATED_SAMPLES
;
243 printk_ratelimited(KERN_WARNING
244 "perf interrupt took too long (%lld > %lld), lowering "
245 "kernel.perf_event_max_sample_rate to %d\n",
246 avg_local_sample_len
, allowed_ns
>> 1,
247 sysctl_perf_event_sample_rate
);
250 static DEFINE_IRQ_WORK(perf_duration_work
, perf_duration_warn
);
252 void perf_sample_event_took(u64 sample_len_ns
)
254 u64 allowed_ns
= ACCESS_ONCE(perf_sample_allowed_ns
);
255 u64 avg_local_sample_len
;
256 u64 local_samples_len
;
261 /* decay the counter by 1 average sample */
262 local_samples_len
= __get_cpu_var(running_sample_length
);
263 local_samples_len
-= local_samples_len
/NR_ACCUMULATED_SAMPLES
;
264 local_samples_len
+= sample_len_ns
;
265 __get_cpu_var(running_sample_length
) = local_samples_len
;
268 * note: this will be biased artifically low until we have
269 * seen NR_ACCUMULATED_SAMPLES. Doing it this way keeps us
270 * from having to maintain a count.
272 avg_local_sample_len
= local_samples_len
/NR_ACCUMULATED_SAMPLES
;
274 if (avg_local_sample_len
<= allowed_ns
)
277 if (max_samples_per_tick
<= 1)
280 max_samples_per_tick
= DIV_ROUND_UP(max_samples_per_tick
, 2);
281 sysctl_perf_event_sample_rate
= max_samples_per_tick
* HZ
;
282 perf_sample_period_ns
= NSEC_PER_SEC
/ sysctl_perf_event_sample_rate
;
284 update_perf_cpu_limits();
286 if (!irq_work_queue(&perf_duration_work
)) {
287 early_printk("perf interrupt took too long (%lld > %lld), lowering "
288 "kernel.perf_event_max_sample_rate to %d\n",
289 avg_local_sample_len
, allowed_ns
>> 1,
290 sysctl_perf_event_sample_rate
);
294 static atomic64_t perf_event_id
;
296 static void cpu_ctx_sched_out(struct perf_cpu_context
*cpuctx
,
297 enum event_type_t event_type
);
299 static void cpu_ctx_sched_in(struct perf_cpu_context
*cpuctx
,
300 enum event_type_t event_type
,
301 struct task_struct
*task
);
303 static void update_context_time(struct perf_event_context
*ctx
);
304 static u64
perf_event_time(struct perf_event
*event
);
306 void __weak
perf_event_print_debug(void) { }
308 extern __weak
const char *perf_pmu_name(void)
313 static inline u64
perf_clock(void)
315 return local_clock();
318 static inline struct perf_cpu_context
*
319 __get_cpu_context(struct perf_event_context
*ctx
)
321 return this_cpu_ptr(ctx
->pmu
->pmu_cpu_context
);
324 static void perf_ctx_lock(struct perf_cpu_context
*cpuctx
,
325 struct perf_event_context
*ctx
)
327 raw_spin_lock(&cpuctx
->ctx
.lock
);
329 raw_spin_lock(&ctx
->lock
);
332 static void perf_ctx_unlock(struct perf_cpu_context
*cpuctx
,
333 struct perf_event_context
*ctx
)
336 raw_spin_unlock(&ctx
->lock
);
337 raw_spin_unlock(&cpuctx
->ctx
.lock
);
340 #ifdef CONFIG_CGROUP_PERF
343 * perf_cgroup_info keeps track of time_enabled for a cgroup.
344 * This is a per-cpu dynamically allocated data structure.
346 struct perf_cgroup_info
{
352 struct cgroup_subsys_state css
;
353 struct perf_cgroup_info __percpu
*info
;
357 * Must ensure cgroup is pinned (css_get) before calling
358 * this function. In other words, we cannot call this function
359 * if there is no cgroup event for the current CPU context.
361 static inline struct perf_cgroup
*
362 perf_cgroup_from_task(struct task_struct
*task
)
364 return container_of(task_css(task
, perf_subsys_id
),
365 struct perf_cgroup
, css
);
369 perf_cgroup_match(struct perf_event
*event
)
371 struct perf_event_context
*ctx
= event
->ctx
;
372 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
374 /* @event doesn't care about cgroup */
378 /* wants specific cgroup scope but @cpuctx isn't associated with any */
383 * Cgroup scoping is recursive. An event enabled for a cgroup is
384 * also enabled for all its descendant cgroups. If @cpuctx's
385 * cgroup is a descendant of @event's (the test covers identity
386 * case), it's a match.
388 return cgroup_is_descendant(cpuctx
->cgrp
->css
.cgroup
,
389 event
->cgrp
->css
.cgroup
);
392 static inline bool perf_tryget_cgroup(struct perf_event
*event
)
394 return css_tryget(&event
->cgrp
->css
);
397 static inline void perf_put_cgroup(struct perf_event
*event
)
399 css_put(&event
->cgrp
->css
);
402 static inline void perf_detach_cgroup(struct perf_event
*event
)
404 perf_put_cgroup(event
);
408 static inline int is_cgroup_event(struct perf_event
*event
)
410 return event
->cgrp
!= NULL
;
413 static inline u64
perf_cgroup_event_time(struct perf_event
*event
)
415 struct perf_cgroup_info
*t
;
417 t
= per_cpu_ptr(event
->cgrp
->info
, event
->cpu
);
421 static inline void __update_cgrp_time(struct perf_cgroup
*cgrp
)
423 struct perf_cgroup_info
*info
;
428 info
= this_cpu_ptr(cgrp
->info
);
430 info
->time
+= now
- info
->timestamp
;
431 info
->timestamp
= now
;
434 static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context
*cpuctx
)
436 struct perf_cgroup
*cgrp_out
= cpuctx
->cgrp
;
438 __update_cgrp_time(cgrp_out
);
441 static inline void update_cgrp_time_from_event(struct perf_event
*event
)
443 struct perf_cgroup
*cgrp
;
446 * ensure we access cgroup data only when needed and
447 * when we know the cgroup is pinned (css_get)
449 if (!is_cgroup_event(event
))
452 cgrp
= perf_cgroup_from_task(current
);
454 * Do not update time when cgroup is not active
456 if (cgrp
== event
->cgrp
)
457 __update_cgrp_time(event
->cgrp
);
461 perf_cgroup_set_timestamp(struct task_struct
*task
,
462 struct perf_event_context
*ctx
)
464 struct perf_cgroup
*cgrp
;
465 struct perf_cgroup_info
*info
;
468 * ctx->lock held by caller
469 * ensure we do not access cgroup data
470 * unless we have the cgroup pinned (css_get)
472 if (!task
|| !ctx
->nr_cgroups
)
475 cgrp
= perf_cgroup_from_task(task
);
476 info
= this_cpu_ptr(cgrp
->info
);
477 info
->timestamp
= ctx
->timestamp
;
480 #define PERF_CGROUP_SWOUT 0x1 /* cgroup switch out every event */
481 #define PERF_CGROUP_SWIN 0x2 /* cgroup switch in events based on task */
484 * reschedule events based on the cgroup constraint of task.
486 * mode SWOUT : schedule out everything
487 * mode SWIN : schedule in based on cgroup for next
489 void perf_cgroup_switch(struct task_struct
*task
, int mode
)
491 struct perf_cpu_context
*cpuctx
;
496 * disable interrupts to avoid geting nr_cgroup
497 * changes via __perf_event_disable(). Also
500 local_irq_save(flags
);
503 * we reschedule only in the presence of cgroup
504 * constrained events.
508 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
509 cpuctx
= this_cpu_ptr(pmu
->pmu_cpu_context
);
510 if (cpuctx
->unique_pmu
!= pmu
)
511 continue; /* ensure we process each cpuctx once */
514 * perf_cgroup_events says at least one
515 * context on this CPU has cgroup events.
517 * ctx->nr_cgroups reports the number of cgroup
518 * events for a context.
520 if (cpuctx
->ctx
.nr_cgroups
> 0) {
521 perf_ctx_lock(cpuctx
, cpuctx
->task_ctx
);
522 perf_pmu_disable(cpuctx
->ctx
.pmu
);
524 if (mode
& PERF_CGROUP_SWOUT
) {
525 cpu_ctx_sched_out(cpuctx
, EVENT_ALL
);
527 * must not be done before ctxswout due
528 * to event_filter_match() in event_sched_out()
533 if (mode
& PERF_CGROUP_SWIN
) {
534 WARN_ON_ONCE(cpuctx
->cgrp
);
536 * set cgrp before ctxsw in to allow
537 * event_filter_match() to not have to pass
540 cpuctx
->cgrp
= perf_cgroup_from_task(task
);
541 cpu_ctx_sched_in(cpuctx
, EVENT_ALL
, task
);
543 perf_pmu_enable(cpuctx
->ctx
.pmu
);
544 perf_ctx_unlock(cpuctx
, cpuctx
->task_ctx
);
550 local_irq_restore(flags
);
553 static inline void perf_cgroup_sched_out(struct task_struct
*task
,
554 struct task_struct
*next
)
556 struct perf_cgroup
*cgrp1
;
557 struct perf_cgroup
*cgrp2
= NULL
;
560 * we come here when we know perf_cgroup_events > 0
562 cgrp1
= perf_cgroup_from_task(task
);
565 * next is NULL when called from perf_event_enable_on_exec()
566 * that will systematically cause a cgroup_switch()
569 cgrp2
= perf_cgroup_from_task(next
);
572 * only schedule out current cgroup events if we know
573 * that we are switching to a different cgroup. Otherwise,
574 * do no touch the cgroup events.
577 perf_cgroup_switch(task
, PERF_CGROUP_SWOUT
);
580 static inline void perf_cgroup_sched_in(struct task_struct
*prev
,
581 struct task_struct
*task
)
583 struct perf_cgroup
*cgrp1
;
584 struct perf_cgroup
*cgrp2
= NULL
;
587 * we come here when we know perf_cgroup_events > 0
589 cgrp1
= perf_cgroup_from_task(task
);
591 /* prev can never be NULL */
592 cgrp2
= perf_cgroup_from_task(prev
);
595 * only need to schedule in cgroup events if we are changing
596 * cgroup during ctxsw. Cgroup events were not scheduled
597 * out of ctxsw out if that was not the case.
600 perf_cgroup_switch(task
, PERF_CGROUP_SWIN
);
603 static inline int perf_cgroup_connect(int fd
, struct perf_event
*event
,
604 struct perf_event_attr
*attr
,
605 struct perf_event
*group_leader
)
607 struct perf_cgroup
*cgrp
;
608 struct cgroup_subsys_state
*css
;
609 struct fd f
= fdget(fd
);
617 css
= css_from_dir(f
.file
->f_dentry
, &perf_subsys
);
623 cgrp
= container_of(css
, struct perf_cgroup
, css
);
626 /* must be done before we fput() the file */
627 if (!perf_tryget_cgroup(event
)) {
634 * all events in a group must monitor
635 * the same cgroup because a task belongs
636 * to only one perf cgroup at a time
638 if (group_leader
&& group_leader
->cgrp
!= cgrp
) {
639 perf_detach_cgroup(event
);
649 perf_cgroup_set_shadow_time(struct perf_event
*event
, u64 now
)
651 struct perf_cgroup_info
*t
;
652 t
= per_cpu_ptr(event
->cgrp
->info
, event
->cpu
);
653 event
->shadow_ctx_time
= now
- t
->timestamp
;
657 perf_cgroup_defer_enabled(struct perf_event
*event
)
660 * when the current task's perf cgroup does not match
661 * the event's, we need to remember to call the
662 * perf_mark_enable() function the first time a task with
663 * a matching perf cgroup is scheduled in.
665 if (is_cgroup_event(event
) && !perf_cgroup_match(event
))
666 event
->cgrp_defer_enabled
= 1;
670 perf_cgroup_mark_enabled(struct perf_event
*event
,
671 struct perf_event_context
*ctx
)
673 struct perf_event
*sub
;
674 u64 tstamp
= perf_event_time(event
);
676 if (!event
->cgrp_defer_enabled
)
679 event
->cgrp_defer_enabled
= 0;
681 event
->tstamp_enabled
= tstamp
- event
->total_time_enabled
;
682 list_for_each_entry(sub
, &event
->sibling_list
, group_entry
) {
683 if (sub
->state
>= PERF_EVENT_STATE_INACTIVE
) {
684 sub
->tstamp_enabled
= tstamp
- sub
->total_time_enabled
;
685 sub
->cgrp_defer_enabled
= 0;
689 #else /* !CONFIG_CGROUP_PERF */
692 perf_cgroup_match(struct perf_event
*event
)
697 static inline void perf_detach_cgroup(struct perf_event
*event
)
700 static inline int is_cgroup_event(struct perf_event
*event
)
705 static inline u64
perf_cgroup_event_cgrp_time(struct perf_event
*event
)
710 static inline void update_cgrp_time_from_event(struct perf_event
*event
)
714 static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context
*cpuctx
)
718 static inline void perf_cgroup_sched_out(struct task_struct
*task
,
719 struct task_struct
*next
)
723 static inline void perf_cgroup_sched_in(struct task_struct
*prev
,
724 struct task_struct
*task
)
728 static inline int perf_cgroup_connect(pid_t pid
, struct perf_event
*event
,
729 struct perf_event_attr
*attr
,
730 struct perf_event
*group_leader
)
736 perf_cgroup_set_timestamp(struct task_struct
*task
,
737 struct perf_event_context
*ctx
)
742 perf_cgroup_switch(struct task_struct
*task
, struct task_struct
*next
)
747 perf_cgroup_set_shadow_time(struct perf_event
*event
, u64 now
)
751 static inline u64
perf_cgroup_event_time(struct perf_event
*event
)
757 perf_cgroup_defer_enabled(struct perf_event
*event
)
762 perf_cgroup_mark_enabled(struct perf_event
*event
,
763 struct perf_event_context
*ctx
)
769 * set default to be dependent on timer tick just
772 #define PERF_CPU_HRTIMER (1000 / HZ)
774 * function must be called with interrupts disbled
776 static enum hrtimer_restart
perf_cpu_hrtimer_handler(struct hrtimer
*hr
)
778 struct perf_cpu_context
*cpuctx
;
779 enum hrtimer_restart ret
= HRTIMER_NORESTART
;
782 WARN_ON(!irqs_disabled());
784 cpuctx
= container_of(hr
, struct perf_cpu_context
, hrtimer
);
786 rotations
= perf_rotate_context(cpuctx
);
789 * arm timer if needed
792 hrtimer_forward_now(hr
, cpuctx
->hrtimer_interval
);
793 ret
= HRTIMER_RESTART
;
799 /* CPU is going down */
800 void perf_cpu_hrtimer_cancel(int cpu
)
802 struct perf_cpu_context
*cpuctx
;
806 if (WARN_ON(cpu
!= smp_processor_id()))
809 local_irq_save(flags
);
813 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
814 cpuctx
= this_cpu_ptr(pmu
->pmu_cpu_context
);
816 if (pmu
->task_ctx_nr
== perf_sw_context
)
819 hrtimer_cancel(&cpuctx
->hrtimer
);
824 local_irq_restore(flags
);
827 static void __perf_cpu_hrtimer_init(struct perf_cpu_context
*cpuctx
, int cpu
)
829 struct hrtimer
*hr
= &cpuctx
->hrtimer
;
830 struct pmu
*pmu
= cpuctx
->ctx
.pmu
;
833 /* no multiplexing needed for SW PMU */
834 if (pmu
->task_ctx_nr
== perf_sw_context
)
838 * check default is sane, if not set then force to
839 * default interval (1/tick)
841 timer
= pmu
->hrtimer_interval_ms
;
843 timer
= pmu
->hrtimer_interval_ms
= PERF_CPU_HRTIMER
;
845 cpuctx
->hrtimer_interval
= ns_to_ktime(NSEC_PER_MSEC
* timer
);
847 hrtimer_init(hr
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL_PINNED
);
848 hr
->function
= perf_cpu_hrtimer_handler
;
851 static void perf_cpu_hrtimer_restart(struct perf_cpu_context
*cpuctx
)
853 struct hrtimer
*hr
= &cpuctx
->hrtimer
;
854 struct pmu
*pmu
= cpuctx
->ctx
.pmu
;
857 if (pmu
->task_ctx_nr
== perf_sw_context
)
860 if (hrtimer_active(hr
))
863 if (!hrtimer_callback_running(hr
))
864 __hrtimer_start_range_ns(hr
, cpuctx
->hrtimer_interval
,
865 0, HRTIMER_MODE_REL_PINNED
, 0);
868 void perf_pmu_disable(struct pmu
*pmu
)
870 int *count
= this_cpu_ptr(pmu
->pmu_disable_count
);
872 pmu
->pmu_disable(pmu
);
875 void perf_pmu_enable(struct pmu
*pmu
)
877 int *count
= this_cpu_ptr(pmu
->pmu_disable_count
);
879 pmu
->pmu_enable(pmu
);
882 static DEFINE_PER_CPU(struct list_head
, rotation_list
);
885 * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
886 * because they're strictly cpu affine and rotate_start is called with IRQs
887 * disabled, while rotate_context is called from IRQ context.
889 static void perf_pmu_rotate_start(struct pmu
*pmu
)
891 struct perf_cpu_context
*cpuctx
= this_cpu_ptr(pmu
->pmu_cpu_context
);
892 struct list_head
*head
= &__get_cpu_var(rotation_list
);
894 WARN_ON(!irqs_disabled());
896 if (list_empty(&cpuctx
->rotation_list
))
897 list_add(&cpuctx
->rotation_list
, head
);
900 static void get_ctx(struct perf_event_context
*ctx
)
902 WARN_ON(!atomic_inc_not_zero(&ctx
->refcount
));
905 static void put_ctx(struct perf_event_context
*ctx
)
907 if (atomic_dec_and_test(&ctx
->refcount
)) {
909 put_ctx(ctx
->parent_ctx
);
911 put_task_struct(ctx
->task
);
912 kfree_rcu(ctx
, rcu_head
);
916 static void unclone_ctx(struct perf_event_context
*ctx
)
918 if (ctx
->parent_ctx
) {
919 put_ctx(ctx
->parent_ctx
);
920 ctx
->parent_ctx
= NULL
;
925 static u32
perf_event_pid(struct perf_event
*event
, struct task_struct
*p
)
928 * only top level events have the pid namespace they were created in
931 event
= event
->parent
;
933 return task_tgid_nr_ns(p
, event
->ns
);
936 static u32
perf_event_tid(struct perf_event
*event
, struct task_struct
*p
)
939 * only top level events have the pid namespace they were created in
942 event
= event
->parent
;
944 return task_pid_nr_ns(p
, event
->ns
);
948 * If we inherit events we want to return the parent event id
951 static u64
primary_event_id(struct perf_event
*event
)
956 id
= event
->parent
->id
;
962 * Get the perf_event_context for a task and lock it.
963 * This has to cope with with the fact that until it is locked,
964 * the context could get moved to another task.
966 static struct perf_event_context
*
967 perf_lock_task_context(struct task_struct
*task
, int ctxn
, unsigned long *flags
)
969 struct perf_event_context
*ctx
;
973 * One of the few rules of preemptible RCU is that one cannot do
974 * rcu_read_unlock() while holding a scheduler (or nested) lock when
975 * part of the read side critical section was preemptible -- see
976 * rcu_read_unlock_special().
978 * Since ctx->lock nests under rq->lock we must ensure the entire read
979 * side critical section is non-preemptible.
983 ctx
= rcu_dereference(task
->perf_event_ctxp
[ctxn
]);
986 * If this context is a clone of another, it might
987 * get swapped for another underneath us by
988 * perf_event_task_sched_out, though the
989 * rcu_read_lock() protects us from any context
990 * getting freed. Lock the context and check if it
991 * got swapped before we could get the lock, and retry
992 * if so. If we locked the right context, then it
993 * can't get swapped on us any more.
995 raw_spin_lock_irqsave(&ctx
->lock
, *flags
);
996 if (ctx
!= rcu_dereference(task
->perf_event_ctxp
[ctxn
])) {
997 raw_spin_unlock_irqrestore(&ctx
->lock
, *flags
);
1003 if (!atomic_inc_not_zero(&ctx
->refcount
)) {
1004 raw_spin_unlock_irqrestore(&ctx
->lock
, *flags
);
1014 * Get the context for a task and increment its pin_count so it
1015 * can't get swapped to another task. This also increments its
1016 * reference count so that the context can't get freed.
1018 static struct perf_event_context
*
1019 perf_pin_task_context(struct task_struct
*task
, int ctxn
)
1021 struct perf_event_context
*ctx
;
1022 unsigned long flags
;
1024 ctx
= perf_lock_task_context(task
, ctxn
, &flags
);
1027 raw_spin_unlock_irqrestore(&ctx
->lock
, flags
);
1032 static void perf_unpin_context(struct perf_event_context
*ctx
)
1034 unsigned long flags
;
1036 raw_spin_lock_irqsave(&ctx
->lock
, flags
);
1038 raw_spin_unlock_irqrestore(&ctx
->lock
, flags
);
1042 * Update the record of the current time in a context.
1044 static void update_context_time(struct perf_event_context
*ctx
)
1046 u64 now
= perf_clock();
1048 ctx
->time
+= now
- ctx
->timestamp
;
1049 ctx
->timestamp
= now
;
1052 static u64
perf_event_time(struct perf_event
*event
)
1054 struct perf_event_context
*ctx
= event
->ctx
;
1056 if (is_cgroup_event(event
))
1057 return perf_cgroup_event_time(event
);
1059 return ctx
? ctx
->time
: 0;
1063 * Update the total_time_enabled and total_time_running fields for a event.
1064 * The caller of this function needs to hold the ctx->lock.
1066 static void update_event_times(struct perf_event
*event
)
1068 struct perf_event_context
*ctx
= event
->ctx
;
1071 if (event
->state
< PERF_EVENT_STATE_INACTIVE
||
1072 event
->group_leader
->state
< PERF_EVENT_STATE_INACTIVE
)
1075 * in cgroup mode, time_enabled represents
1076 * the time the event was enabled AND active
1077 * tasks were in the monitored cgroup. This is
1078 * independent of the activity of the context as
1079 * there may be a mix of cgroup and non-cgroup events.
1081 * That is why we treat cgroup events differently
1084 if (is_cgroup_event(event
))
1085 run_end
= perf_cgroup_event_time(event
);
1086 else if (ctx
->is_active
)
1087 run_end
= ctx
->time
;
1089 run_end
= event
->tstamp_stopped
;
1091 event
->total_time_enabled
= run_end
- event
->tstamp_enabled
;
1093 if (event
->state
== PERF_EVENT_STATE_INACTIVE
)
1094 run_end
= event
->tstamp_stopped
;
1096 run_end
= perf_event_time(event
);
1098 event
->total_time_running
= run_end
- event
->tstamp_running
;
1103 * Update total_time_enabled and total_time_running for all events in a group.
1105 static void update_group_times(struct perf_event
*leader
)
1107 struct perf_event
*event
;
1109 update_event_times(leader
);
1110 list_for_each_entry(event
, &leader
->sibling_list
, group_entry
)
1111 update_event_times(event
);
1114 static struct list_head
*
1115 ctx_group_list(struct perf_event
*event
, struct perf_event_context
*ctx
)
1117 if (event
->attr
.pinned
)
1118 return &ctx
->pinned_groups
;
1120 return &ctx
->flexible_groups
;
1124 * Add a event from the lists for its context.
1125 * Must be called with ctx->mutex and ctx->lock held.
1128 list_add_event(struct perf_event
*event
, struct perf_event_context
*ctx
)
1130 WARN_ON_ONCE(event
->attach_state
& PERF_ATTACH_CONTEXT
);
1131 event
->attach_state
|= PERF_ATTACH_CONTEXT
;
1134 * If we're a stand alone event or group leader, we go to the context
1135 * list, group events are kept attached to the group so that
1136 * perf_group_detach can, at all times, locate all siblings.
1138 if (event
->group_leader
== event
) {
1139 struct list_head
*list
;
1141 if (is_software_event(event
))
1142 event
->group_flags
|= PERF_GROUP_SOFTWARE
;
1144 list
= ctx_group_list(event
, ctx
);
1145 list_add_tail(&event
->group_entry
, list
);
1148 if (is_cgroup_event(event
))
1151 if (has_branch_stack(event
))
1152 ctx
->nr_branch_stack
++;
1154 list_add_rcu(&event
->event_entry
, &ctx
->event_list
);
1155 if (!ctx
->nr_events
)
1156 perf_pmu_rotate_start(ctx
->pmu
);
1158 if (event
->attr
.inherit_stat
)
1165 * Initialize event state based on the perf_event_attr::disabled.
1167 static inline void perf_event__state_init(struct perf_event
*event
)
1169 event
->state
= event
->attr
.disabled
? PERF_EVENT_STATE_OFF
:
1170 PERF_EVENT_STATE_INACTIVE
;
1174 * Called at perf_event creation and when events are attached/detached from a
1177 static void perf_event__read_size(struct perf_event
*event
)
1179 int entry
= sizeof(u64
); /* value */
1183 if (event
->attr
.read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
1184 size
+= sizeof(u64
);
1186 if (event
->attr
.read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
1187 size
+= sizeof(u64
);
1189 if (event
->attr
.read_format
& PERF_FORMAT_ID
)
1190 entry
+= sizeof(u64
);
1192 if (event
->attr
.read_format
& PERF_FORMAT_GROUP
) {
1193 nr
+= event
->group_leader
->nr_siblings
;
1194 size
+= sizeof(u64
);
1198 event
->read_size
= size
;
1201 static void perf_event__header_size(struct perf_event
*event
)
1203 struct perf_sample_data
*data
;
1204 u64 sample_type
= event
->attr
.sample_type
;
1207 perf_event__read_size(event
);
1209 if (sample_type
& PERF_SAMPLE_IP
)
1210 size
+= sizeof(data
->ip
);
1212 if (sample_type
& PERF_SAMPLE_ADDR
)
1213 size
+= sizeof(data
->addr
);
1215 if (sample_type
& PERF_SAMPLE_PERIOD
)
1216 size
+= sizeof(data
->period
);
1218 if (sample_type
& PERF_SAMPLE_WEIGHT
)
1219 size
+= sizeof(data
->weight
);
1221 if (sample_type
& PERF_SAMPLE_READ
)
1222 size
+= event
->read_size
;
1224 if (sample_type
& PERF_SAMPLE_DATA_SRC
)
1225 size
+= sizeof(data
->data_src
.val
);
1227 if (sample_type
& PERF_SAMPLE_TRANSACTION
)
1228 size
+= sizeof(data
->txn
);
1230 event
->header_size
= size
;
1233 static void perf_event__id_header_size(struct perf_event
*event
)
1235 struct perf_sample_data
*data
;
1236 u64 sample_type
= event
->attr
.sample_type
;
1239 if (sample_type
& PERF_SAMPLE_TID
)
1240 size
+= sizeof(data
->tid_entry
);
1242 if (sample_type
& PERF_SAMPLE_TIME
)
1243 size
+= sizeof(data
->time
);
1245 if (sample_type
& PERF_SAMPLE_IDENTIFIER
)
1246 size
+= sizeof(data
->id
);
1248 if (sample_type
& PERF_SAMPLE_ID
)
1249 size
+= sizeof(data
->id
);
1251 if (sample_type
& PERF_SAMPLE_STREAM_ID
)
1252 size
+= sizeof(data
->stream_id
);
1254 if (sample_type
& PERF_SAMPLE_CPU
)
1255 size
+= sizeof(data
->cpu_entry
);
1257 event
->id_header_size
= size
;
1260 static void perf_group_attach(struct perf_event
*event
)
1262 struct perf_event
*group_leader
= event
->group_leader
, *pos
;
1265 * We can have double attach due to group movement in perf_event_open.
1267 if (event
->attach_state
& PERF_ATTACH_GROUP
)
1270 event
->attach_state
|= PERF_ATTACH_GROUP
;
1272 if (group_leader
== event
)
1275 if (group_leader
->group_flags
& PERF_GROUP_SOFTWARE
&&
1276 !is_software_event(event
))
1277 group_leader
->group_flags
&= ~PERF_GROUP_SOFTWARE
;
1279 list_add_tail(&event
->group_entry
, &group_leader
->sibling_list
);
1280 group_leader
->nr_siblings
++;
1282 perf_event__header_size(group_leader
);
1284 list_for_each_entry(pos
, &group_leader
->sibling_list
, group_entry
)
1285 perf_event__header_size(pos
);
1289 * Remove a event from the lists for its context.
1290 * Must be called with ctx->mutex and ctx->lock held.
1293 list_del_event(struct perf_event
*event
, struct perf_event_context
*ctx
)
1295 struct perf_cpu_context
*cpuctx
;
1297 * We can have double detach due to exit/hot-unplug + close.
1299 if (!(event
->attach_state
& PERF_ATTACH_CONTEXT
))
1302 event
->attach_state
&= ~PERF_ATTACH_CONTEXT
;
1304 if (is_cgroup_event(event
)) {
1306 cpuctx
= __get_cpu_context(ctx
);
1308 * if there are no more cgroup events
1309 * then cler cgrp to avoid stale pointer
1310 * in update_cgrp_time_from_cpuctx()
1312 if (!ctx
->nr_cgroups
)
1313 cpuctx
->cgrp
= NULL
;
1316 if (has_branch_stack(event
))
1317 ctx
->nr_branch_stack
--;
1320 if (event
->attr
.inherit_stat
)
1323 list_del_rcu(&event
->event_entry
);
1325 if (event
->group_leader
== event
)
1326 list_del_init(&event
->group_entry
);
1328 update_group_times(event
);
1331 * If event was in error state, then keep it
1332 * that way, otherwise bogus counts will be
1333 * returned on read(). The only way to get out
1334 * of error state is by explicit re-enabling
1337 if (event
->state
> PERF_EVENT_STATE_OFF
)
1338 event
->state
= PERF_EVENT_STATE_OFF
;
1343 static void perf_group_detach(struct perf_event
*event
)
1345 struct perf_event
*sibling
, *tmp
;
1346 struct list_head
*list
= NULL
;
1349 * We can have double detach due to exit/hot-unplug + close.
1351 if (!(event
->attach_state
& PERF_ATTACH_GROUP
))
1354 event
->attach_state
&= ~PERF_ATTACH_GROUP
;
1357 * If this is a sibling, remove it from its group.
1359 if (event
->group_leader
!= event
) {
1360 list_del_init(&event
->group_entry
);
1361 event
->group_leader
->nr_siblings
--;
1365 if (!list_empty(&event
->group_entry
))
1366 list
= &event
->group_entry
;
1369 * If this was a group event with sibling events then
1370 * upgrade the siblings to singleton events by adding them
1371 * to whatever list we are on.
1373 list_for_each_entry_safe(sibling
, tmp
, &event
->sibling_list
, group_entry
) {
1375 list_move_tail(&sibling
->group_entry
, list
);
1376 sibling
->group_leader
= sibling
;
1378 /* Inherit group flags from the previous leader */
1379 sibling
->group_flags
= event
->group_flags
;
1383 perf_event__header_size(event
->group_leader
);
1385 list_for_each_entry(tmp
, &event
->group_leader
->sibling_list
, group_entry
)
1386 perf_event__header_size(tmp
);
1390 event_filter_match(struct perf_event
*event
)
1392 return (event
->cpu
== -1 || event
->cpu
== smp_processor_id())
1393 && perf_cgroup_match(event
);
1397 event_sched_out(struct perf_event
*event
,
1398 struct perf_cpu_context
*cpuctx
,
1399 struct perf_event_context
*ctx
)
1401 u64 tstamp
= perf_event_time(event
);
1404 * An event which could not be activated because of
1405 * filter mismatch still needs to have its timings
1406 * maintained, otherwise bogus information is return
1407 * via read() for time_enabled, time_running:
1409 if (event
->state
== PERF_EVENT_STATE_INACTIVE
1410 && !event_filter_match(event
)) {
1411 delta
= tstamp
- event
->tstamp_stopped
;
1412 event
->tstamp_running
+= delta
;
1413 event
->tstamp_stopped
= tstamp
;
1416 if (event
->state
!= PERF_EVENT_STATE_ACTIVE
)
1419 perf_pmu_disable(event
->pmu
);
1421 event
->state
= PERF_EVENT_STATE_INACTIVE
;
1422 if (event
->pending_disable
) {
1423 event
->pending_disable
= 0;
1424 event
->state
= PERF_EVENT_STATE_OFF
;
1426 event
->tstamp_stopped
= tstamp
;
1427 event
->pmu
->del(event
, 0);
1430 if (!is_software_event(event
))
1431 cpuctx
->active_oncpu
--;
1433 if (event
->attr
.freq
&& event
->attr
.sample_freq
)
1435 if (event
->attr
.exclusive
|| !cpuctx
->active_oncpu
)
1436 cpuctx
->exclusive
= 0;
1438 perf_pmu_enable(event
->pmu
);
1442 group_sched_out(struct perf_event
*group_event
,
1443 struct perf_cpu_context
*cpuctx
,
1444 struct perf_event_context
*ctx
)
1446 struct perf_event
*event
;
1447 int state
= group_event
->state
;
1449 event_sched_out(group_event
, cpuctx
, ctx
);
1452 * Schedule out siblings (if any):
1454 list_for_each_entry(event
, &group_event
->sibling_list
, group_entry
)
1455 event_sched_out(event
, cpuctx
, ctx
);
1457 if (state
== PERF_EVENT_STATE_ACTIVE
&& group_event
->attr
.exclusive
)
1458 cpuctx
->exclusive
= 0;
1462 * Cross CPU call to remove a performance event
1464 * We disable the event on the hardware level first. After that we
1465 * remove it from the context list.
1467 static int __perf_remove_from_context(void *info
)
1469 struct perf_event
*event
= info
;
1470 struct perf_event_context
*ctx
= event
->ctx
;
1471 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
1473 raw_spin_lock(&ctx
->lock
);
1474 event_sched_out(event
, cpuctx
, ctx
);
1475 list_del_event(event
, ctx
);
1476 if (!ctx
->nr_events
&& cpuctx
->task_ctx
== ctx
) {
1478 cpuctx
->task_ctx
= NULL
;
1480 raw_spin_unlock(&ctx
->lock
);
1487 * Remove the event from a task's (or a CPU's) list of events.
1489 * CPU events are removed with a smp call. For task events we only
1490 * call when the task is on a CPU.
1492 * If event->ctx is a cloned context, callers must make sure that
1493 * every task struct that event->ctx->task could possibly point to
1494 * remains valid. This is OK when called from perf_release since
1495 * that only calls us on the top-level context, which can't be a clone.
1496 * When called from perf_event_exit_task, it's OK because the
1497 * context has been detached from its task.
1499 static void perf_remove_from_context(struct perf_event
*event
)
1501 struct perf_event_context
*ctx
= event
->ctx
;
1502 struct task_struct
*task
= ctx
->task
;
1504 lockdep_assert_held(&ctx
->mutex
);
1508 * Per cpu events are removed via an smp call and
1509 * the removal is always successful.
1511 cpu_function_call(event
->cpu
, __perf_remove_from_context
, event
);
1516 if (!task_function_call(task
, __perf_remove_from_context
, event
))
1519 raw_spin_lock_irq(&ctx
->lock
);
1521 * If we failed to find a running task, but find the context active now
1522 * that we've acquired the ctx->lock, retry.
1524 if (ctx
->is_active
) {
1525 raw_spin_unlock_irq(&ctx
->lock
);
1530 * Since the task isn't running, its safe to remove the event, us
1531 * holding the ctx->lock ensures the task won't get scheduled in.
1533 list_del_event(event
, ctx
);
1534 raw_spin_unlock_irq(&ctx
->lock
);
1538 * Cross CPU call to disable a performance event
1540 int __perf_event_disable(void *info
)
1542 struct perf_event
*event
= info
;
1543 struct perf_event_context
*ctx
= event
->ctx
;
1544 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
1547 * If this is a per-task event, need to check whether this
1548 * event's task is the current task on this cpu.
1550 * Can trigger due to concurrent perf_event_context_sched_out()
1551 * flipping contexts around.
1553 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
)
1556 raw_spin_lock(&ctx
->lock
);
1559 * If the event is on, turn it off.
1560 * If it is in error state, leave it in error state.
1562 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
) {
1563 update_context_time(ctx
);
1564 update_cgrp_time_from_event(event
);
1565 update_group_times(event
);
1566 if (event
== event
->group_leader
)
1567 group_sched_out(event
, cpuctx
, ctx
);
1569 event_sched_out(event
, cpuctx
, ctx
);
1570 event
->state
= PERF_EVENT_STATE_OFF
;
1573 raw_spin_unlock(&ctx
->lock
);
1581 * If event->ctx is a cloned context, callers must make sure that
1582 * every task struct that event->ctx->task could possibly point to
1583 * remains valid. This condition is satisifed when called through
1584 * perf_event_for_each_child or perf_event_for_each because they
1585 * hold the top-level event's child_mutex, so any descendant that
1586 * goes to exit will block in sync_child_event.
1587 * When called from perf_pending_event it's OK because event->ctx
1588 * is the current context on this CPU and preemption is disabled,
1589 * hence we can't get into perf_event_task_sched_out for this context.
1591 void perf_event_disable(struct perf_event
*event
)
1593 struct perf_event_context
*ctx
= event
->ctx
;
1594 struct task_struct
*task
= ctx
->task
;
1598 * Disable the event on the cpu that it's on
1600 cpu_function_call(event
->cpu
, __perf_event_disable
, event
);
1605 if (!task_function_call(task
, __perf_event_disable
, event
))
1608 raw_spin_lock_irq(&ctx
->lock
);
1610 * If the event is still active, we need to retry the cross-call.
1612 if (event
->state
== PERF_EVENT_STATE_ACTIVE
) {
1613 raw_spin_unlock_irq(&ctx
->lock
);
1615 * Reload the task pointer, it might have been changed by
1616 * a concurrent perf_event_context_sched_out().
1623 * Since we have the lock this context can't be scheduled
1624 * in, so we can change the state safely.
1626 if (event
->state
== PERF_EVENT_STATE_INACTIVE
) {
1627 update_group_times(event
);
1628 event
->state
= PERF_EVENT_STATE_OFF
;
1630 raw_spin_unlock_irq(&ctx
->lock
);
1632 EXPORT_SYMBOL_GPL(perf_event_disable
);
1634 static void perf_set_shadow_time(struct perf_event
*event
,
1635 struct perf_event_context
*ctx
,
1639 * use the correct time source for the time snapshot
1641 * We could get by without this by leveraging the
1642 * fact that to get to this function, the caller
1643 * has most likely already called update_context_time()
1644 * and update_cgrp_time_xx() and thus both timestamp
1645 * are identical (or very close). Given that tstamp is,
1646 * already adjusted for cgroup, we could say that:
1647 * tstamp - ctx->timestamp
1649 * tstamp - cgrp->timestamp.
1651 * Then, in perf_output_read(), the calculation would
1652 * work with no changes because:
1653 * - event is guaranteed scheduled in
1654 * - no scheduled out in between
1655 * - thus the timestamp would be the same
1657 * But this is a bit hairy.
1659 * So instead, we have an explicit cgroup call to remain
1660 * within the time time source all along. We believe it
1661 * is cleaner and simpler to understand.
1663 if (is_cgroup_event(event
))
1664 perf_cgroup_set_shadow_time(event
, tstamp
);
1666 event
->shadow_ctx_time
= tstamp
- ctx
->timestamp
;
1669 #define MAX_INTERRUPTS (~0ULL)
1671 static void perf_log_throttle(struct perf_event
*event
, int enable
);
1674 event_sched_in(struct perf_event
*event
,
1675 struct perf_cpu_context
*cpuctx
,
1676 struct perf_event_context
*ctx
)
1678 u64 tstamp
= perf_event_time(event
);
1681 if (event
->state
<= PERF_EVENT_STATE_OFF
)
1684 event
->state
= PERF_EVENT_STATE_ACTIVE
;
1685 event
->oncpu
= smp_processor_id();
1688 * Unthrottle events, since we scheduled we might have missed several
1689 * ticks already, also for a heavily scheduling task there is little
1690 * guarantee it'll get a tick in a timely manner.
1692 if (unlikely(event
->hw
.interrupts
== MAX_INTERRUPTS
)) {
1693 perf_log_throttle(event
, 1);
1694 event
->hw
.interrupts
= 0;
1698 * The new state must be visible before we turn it on in the hardware:
1702 perf_pmu_disable(event
->pmu
);
1704 if (event
->pmu
->add(event
, PERF_EF_START
)) {
1705 event
->state
= PERF_EVENT_STATE_INACTIVE
;
1711 event
->tstamp_running
+= tstamp
- event
->tstamp_stopped
;
1713 perf_set_shadow_time(event
, ctx
, tstamp
);
1715 if (!is_software_event(event
))
1716 cpuctx
->active_oncpu
++;
1718 if (event
->attr
.freq
&& event
->attr
.sample_freq
)
1721 if (event
->attr
.exclusive
)
1722 cpuctx
->exclusive
= 1;
1725 perf_pmu_enable(event
->pmu
);
1731 group_sched_in(struct perf_event
*group_event
,
1732 struct perf_cpu_context
*cpuctx
,
1733 struct perf_event_context
*ctx
)
1735 struct perf_event
*event
, *partial_group
= NULL
;
1736 struct pmu
*pmu
= ctx
->pmu
;
1737 u64 now
= ctx
->time
;
1738 bool simulate
= false;
1740 if (group_event
->state
== PERF_EVENT_STATE_OFF
)
1743 pmu
->start_txn(pmu
);
1745 if (event_sched_in(group_event
, cpuctx
, ctx
)) {
1746 pmu
->cancel_txn(pmu
);
1747 perf_cpu_hrtimer_restart(cpuctx
);
1752 * Schedule in siblings as one group (if any):
1754 list_for_each_entry(event
, &group_event
->sibling_list
, group_entry
) {
1755 if (event_sched_in(event
, cpuctx
, ctx
)) {
1756 partial_group
= event
;
1761 if (!pmu
->commit_txn(pmu
))
1766 * Groups can be scheduled in as one unit only, so undo any
1767 * partial group before returning:
1768 * The events up to the failed event are scheduled out normally,
1769 * tstamp_stopped will be updated.
1771 * The failed events and the remaining siblings need to have
1772 * their timings updated as if they had gone thru event_sched_in()
1773 * and event_sched_out(). This is required to get consistent timings
1774 * across the group. This also takes care of the case where the group
1775 * could never be scheduled by ensuring tstamp_stopped is set to mark
1776 * the time the event was actually stopped, such that time delta
1777 * calculation in update_event_times() is correct.
1779 list_for_each_entry(event
, &group_event
->sibling_list
, group_entry
) {
1780 if (event
== partial_group
)
1784 event
->tstamp_running
+= now
- event
->tstamp_stopped
;
1785 event
->tstamp_stopped
= now
;
1787 event_sched_out(event
, cpuctx
, ctx
);
1790 event_sched_out(group_event
, cpuctx
, ctx
);
1792 pmu
->cancel_txn(pmu
);
1794 perf_cpu_hrtimer_restart(cpuctx
);
1800 * Work out whether we can put this event group on the CPU now.
1802 static int group_can_go_on(struct perf_event
*event
,
1803 struct perf_cpu_context
*cpuctx
,
1807 * Groups consisting entirely of software events can always go on.
1809 if (event
->group_flags
& PERF_GROUP_SOFTWARE
)
1812 * If an exclusive group is already on, no other hardware
1815 if (cpuctx
->exclusive
)
1818 * If this group is exclusive and there are already
1819 * events on the CPU, it can't go on.
1821 if (event
->attr
.exclusive
&& cpuctx
->active_oncpu
)
1824 * Otherwise, try to add it if all previous groups were able
1830 static void add_event_to_ctx(struct perf_event
*event
,
1831 struct perf_event_context
*ctx
)
1833 u64 tstamp
= perf_event_time(event
);
1835 list_add_event(event
, ctx
);
1836 perf_group_attach(event
);
1837 event
->tstamp_enabled
= tstamp
;
1838 event
->tstamp_running
= tstamp
;
1839 event
->tstamp_stopped
= tstamp
;
1842 static void task_ctx_sched_out(struct perf_event_context
*ctx
);
1844 ctx_sched_in(struct perf_event_context
*ctx
,
1845 struct perf_cpu_context
*cpuctx
,
1846 enum event_type_t event_type
,
1847 struct task_struct
*task
);
1849 static void perf_event_sched_in(struct perf_cpu_context
*cpuctx
,
1850 struct perf_event_context
*ctx
,
1851 struct task_struct
*task
)
1853 cpu_ctx_sched_in(cpuctx
, EVENT_PINNED
, task
);
1855 ctx_sched_in(ctx
, cpuctx
, EVENT_PINNED
, task
);
1856 cpu_ctx_sched_in(cpuctx
, EVENT_FLEXIBLE
, task
);
1858 ctx_sched_in(ctx
, cpuctx
, EVENT_FLEXIBLE
, task
);
1862 * Cross CPU call to install and enable a performance event
1864 * Must be called with ctx->mutex held
1866 static int __perf_install_in_context(void *info
)
1868 struct perf_event
*event
= info
;
1869 struct perf_event_context
*ctx
= event
->ctx
;
1870 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
1871 struct perf_event_context
*task_ctx
= cpuctx
->task_ctx
;
1872 struct task_struct
*task
= current
;
1874 perf_ctx_lock(cpuctx
, task_ctx
);
1875 perf_pmu_disable(cpuctx
->ctx
.pmu
);
1878 * If there was an active task_ctx schedule it out.
1881 task_ctx_sched_out(task_ctx
);
1884 * If the context we're installing events in is not the
1885 * active task_ctx, flip them.
1887 if (ctx
->task
&& task_ctx
!= ctx
) {
1889 raw_spin_unlock(&task_ctx
->lock
);
1890 raw_spin_lock(&ctx
->lock
);
1895 cpuctx
->task_ctx
= task_ctx
;
1896 task
= task_ctx
->task
;
1899 cpu_ctx_sched_out(cpuctx
, EVENT_ALL
);
1901 update_context_time(ctx
);
1903 * update cgrp time only if current cgrp
1904 * matches event->cgrp. Must be done before
1905 * calling add_event_to_ctx()
1907 update_cgrp_time_from_event(event
);
1909 add_event_to_ctx(event
, ctx
);
1912 * Schedule everything back in
1914 perf_event_sched_in(cpuctx
, task_ctx
, task
);
1916 perf_pmu_enable(cpuctx
->ctx
.pmu
);
1917 perf_ctx_unlock(cpuctx
, task_ctx
);
1923 * Attach a performance event to a context
1925 * First we add the event to the list with the hardware enable bit
1926 * in event->hw_config cleared.
1928 * If the event is attached to a task which is on a CPU we use a smp
1929 * call to enable it in the task context. The task might have been
1930 * scheduled away, but we check this in the smp call again.
1933 perf_install_in_context(struct perf_event_context
*ctx
,
1934 struct perf_event
*event
,
1937 struct task_struct
*task
= ctx
->task
;
1939 lockdep_assert_held(&ctx
->mutex
);
1942 if (event
->cpu
!= -1)
1947 * Per cpu events are installed via an smp call and
1948 * the install is always successful.
1950 cpu_function_call(cpu
, __perf_install_in_context
, event
);
1955 if (!task_function_call(task
, __perf_install_in_context
, event
))
1958 raw_spin_lock_irq(&ctx
->lock
);
1960 * If we failed to find a running task, but find the context active now
1961 * that we've acquired the ctx->lock, retry.
1963 if (ctx
->is_active
) {
1964 raw_spin_unlock_irq(&ctx
->lock
);
1969 * Since the task isn't running, its safe to add the event, us holding
1970 * the ctx->lock ensures the task won't get scheduled in.
1972 add_event_to_ctx(event
, ctx
);
1973 raw_spin_unlock_irq(&ctx
->lock
);
1977 * Put a event into inactive state and update time fields.
1978 * Enabling the leader of a group effectively enables all
1979 * the group members that aren't explicitly disabled, so we
1980 * have to update their ->tstamp_enabled also.
1981 * Note: this works for group members as well as group leaders
1982 * since the non-leader members' sibling_lists will be empty.
1984 static void __perf_event_mark_enabled(struct perf_event
*event
)
1986 struct perf_event
*sub
;
1987 u64 tstamp
= perf_event_time(event
);
1989 event
->state
= PERF_EVENT_STATE_INACTIVE
;
1990 event
->tstamp_enabled
= tstamp
- event
->total_time_enabled
;
1991 list_for_each_entry(sub
, &event
->sibling_list
, group_entry
) {
1992 if (sub
->state
>= PERF_EVENT_STATE_INACTIVE
)
1993 sub
->tstamp_enabled
= tstamp
- sub
->total_time_enabled
;
1998 * Cross CPU call to enable a performance event
2000 static int __perf_event_enable(void *info
)
2002 struct perf_event
*event
= info
;
2003 struct perf_event_context
*ctx
= event
->ctx
;
2004 struct perf_event
*leader
= event
->group_leader
;
2005 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
2009 * There's a time window between 'ctx->is_active' check
2010 * in perf_event_enable function and this place having:
2012 * - ctx->lock unlocked
2014 * where the task could be killed and 'ctx' deactivated
2015 * by perf_event_exit_task.
2017 if (!ctx
->is_active
)
2020 raw_spin_lock(&ctx
->lock
);
2021 update_context_time(ctx
);
2023 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
)
2027 * set current task's cgroup time reference point
2029 perf_cgroup_set_timestamp(current
, ctx
);
2031 __perf_event_mark_enabled(event
);
2033 if (!event_filter_match(event
)) {
2034 if (is_cgroup_event(event
))
2035 perf_cgroup_defer_enabled(event
);
2040 * If the event is in a group and isn't the group leader,
2041 * then don't put it on unless the group is on.
2043 if (leader
!= event
&& leader
->state
!= PERF_EVENT_STATE_ACTIVE
)
2046 if (!group_can_go_on(event
, cpuctx
, 1)) {
2049 if (event
== leader
)
2050 err
= group_sched_in(event
, cpuctx
, ctx
);
2052 err
= event_sched_in(event
, cpuctx
, ctx
);
2057 * If this event can't go on and it's part of a
2058 * group, then the whole group has to come off.
2060 if (leader
!= event
) {
2061 group_sched_out(leader
, cpuctx
, ctx
);
2062 perf_cpu_hrtimer_restart(cpuctx
);
2064 if (leader
->attr
.pinned
) {
2065 update_group_times(leader
);
2066 leader
->state
= PERF_EVENT_STATE_ERROR
;
2071 raw_spin_unlock(&ctx
->lock
);
2079 * If event->ctx is a cloned context, callers must make sure that
2080 * every task struct that event->ctx->task could possibly point to
2081 * remains valid. This condition is satisfied when called through
2082 * perf_event_for_each_child or perf_event_for_each as described
2083 * for perf_event_disable.
2085 void perf_event_enable(struct perf_event
*event
)
2087 struct perf_event_context
*ctx
= event
->ctx
;
2088 struct task_struct
*task
= ctx
->task
;
2092 * Enable the event on the cpu that it's on
2094 cpu_function_call(event
->cpu
, __perf_event_enable
, event
);
2098 raw_spin_lock_irq(&ctx
->lock
);
2099 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
)
2103 * If the event is in error state, clear that first.
2104 * That way, if we see the event in error state below, we
2105 * know that it has gone back into error state, as distinct
2106 * from the task having been scheduled away before the
2107 * cross-call arrived.
2109 if (event
->state
== PERF_EVENT_STATE_ERROR
)
2110 event
->state
= PERF_EVENT_STATE_OFF
;
2113 if (!ctx
->is_active
) {
2114 __perf_event_mark_enabled(event
);
2118 raw_spin_unlock_irq(&ctx
->lock
);
2120 if (!task_function_call(task
, __perf_event_enable
, event
))
2123 raw_spin_lock_irq(&ctx
->lock
);
2126 * If the context is active and the event is still off,
2127 * we need to retry the cross-call.
2129 if (ctx
->is_active
&& event
->state
== PERF_EVENT_STATE_OFF
) {
2131 * task could have been flipped by a concurrent
2132 * perf_event_context_sched_out()
2139 raw_spin_unlock_irq(&ctx
->lock
);
2141 EXPORT_SYMBOL_GPL(perf_event_enable
);
2143 int perf_event_refresh(struct perf_event
*event
, int refresh
)
2146 * not supported on inherited events
2148 if (event
->attr
.inherit
|| !is_sampling_event(event
))
2151 atomic_add(refresh
, &event
->event_limit
);
2152 perf_event_enable(event
);
2156 EXPORT_SYMBOL_GPL(perf_event_refresh
);
2158 static void ctx_sched_out(struct perf_event_context
*ctx
,
2159 struct perf_cpu_context
*cpuctx
,
2160 enum event_type_t event_type
)
2162 struct perf_event
*event
;
2163 int is_active
= ctx
->is_active
;
2165 ctx
->is_active
&= ~event_type
;
2166 if (likely(!ctx
->nr_events
))
2169 update_context_time(ctx
);
2170 update_cgrp_time_from_cpuctx(cpuctx
);
2171 if (!ctx
->nr_active
)
2174 perf_pmu_disable(ctx
->pmu
);
2175 if ((is_active
& EVENT_PINNED
) && (event_type
& EVENT_PINNED
)) {
2176 list_for_each_entry(event
, &ctx
->pinned_groups
, group_entry
)
2177 group_sched_out(event
, cpuctx
, ctx
);
2180 if ((is_active
& EVENT_FLEXIBLE
) && (event_type
& EVENT_FLEXIBLE
)) {
2181 list_for_each_entry(event
, &ctx
->flexible_groups
, group_entry
)
2182 group_sched_out(event
, cpuctx
, ctx
);
2184 perf_pmu_enable(ctx
->pmu
);
2188 * Test whether two contexts are equivalent, i.e. whether they have both been
2189 * cloned from the same version of the same context.
2191 * Equivalence is measured using a generation number in the context that is
2192 * incremented on each modification to it; see unclone_ctx(), list_add_event()
2193 * and list_del_event().
2195 static int context_equiv(struct perf_event_context
*ctx1
,
2196 struct perf_event_context
*ctx2
)
2198 /* Pinning disables the swap optimization */
2199 if (ctx1
->pin_count
|| ctx2
->pin_count
)
2202 /* If ctx1 is the parent of ctx2 */
2203 if (ctx1
== ctx2
->parent_ctx
&& ctx1
->generation
== ctx2
->parent_gen
)
2206 /* If ctx2 is the parent of ctx1 */
2207 if (ctx1
->parent_ctx
== ctx2
&& ctx1
->parent_gen
== ctx2
->generation
)
2211 * If ctx1 and ctx2 have the same parent; we flatten the parent
2212 * hierarchy, see perf_event_init_context().
2214 if (ctx1
->parent_ctx
&& ctx1
->parent_ctx
== ctx2
->parent_ctx
&&
2215 ctx1
->parent_gen
== ctx2
->parent_gen
)
2222 static void __perf_event_sync_stat(struct perf_event
*event
,
2223 struct perf_event
*next_event
)
2227 if (!event
->attr
.inherit_stat
)
2231 * Update the event value, we cannot use perf_event_read()
2232 * because we're in the middle of a context switch and have IRQs
2233 * disabled, which upsets smp_call_function_single(), however
2234 * we know the event must be on the current CPU, therefore we
2235 * don't need to use it.
2237 switch (event
->state
) {
2238 case PERF_EVENT_STATE_ACTIVE
:
2239 event
->pmu
->read(event
);
2242 case PERF_EVENT_STATE_INACTIVE
:
2243 update_event_times(event
);
2251 * In order to keep per-task stats reliable we need to flip the event
2252 * values when we flip the contexts.
2254 value
= local64_read(&next_event
->count
);
2255 value
= local64_xchg(&event
->count
, value
);
2256 local64_set(&next_event
->count
, value
);
2258 swap(event
->total_time_enabled
, next_event
->total_time_enabled
);
2259 swap(event
->total_time_running
, next_event
->total_time_running
);
2262 * Since we swizzled the values, update the user visible data too.
2264 perf_event_update_userpage(event
);
2265 perf_event_update_userpage(next_event
);
2268 static void perf_event_sync_stat(struct perf_event_context
*ctx
,
2269 struct perf_event_context
*next_ctx
)
2271 struct perf_event
*event
, *next_event
;
2276 update_context_time(ctx
);
2278 event
= list_first_entry(&ctx
->event_list
,
2279 struct perf_event
, event_entry
);
2281 next_event
= list_first_entry(&next_ctx
->event_list
,
2282 struct perf_event
, event_entry
);
2284 while (&event
->event_entry
!= &ctx
->event_list
&&
2285 &next_event
->event_entry
!= &next_ctx
->event_list
) {
2287 __perf_event_sync_stat(event
, next_event
);
2289 event
= list_next_entry(event
, event_entry
);
2290 next_event
= list_next_entry(next_event
, event_entry
);
2294 static void perf_event_context_sched_out(struct task_struct
*task
, int ctxn
,
2295 struct task_struct
*next
)
2297 struct perf_event_context
*ctx
= task
->perf_event_ctxp
[ctxn
];
2298 struct perf_event_context
*next_ctx
;
2299 struct perf_event_context
*parent
, *next_parent
;
2300 struct perf_cpu_context
*cpuctx
;
2306 cpuctx
= __get_cpu_context(ctx
);
2307 if (!cpuctx
->task_ctx
)
2311 next_ctx
= next
->perf_event_ctxp
[ctxn
];
2315 parent
= rcu_dereference(ctx
->parent_ctx
);
2316 next_parent
= rcu_dereference(next_ctx
->parent_ctx
);
2318 /* If neither context have a parent context; they cannot be clones. */
2319 if (!parent
&& !next_parent
)
2322 if (next_parent
== ctx
|| next_ctx
== parent
|| next_parent
== parent
) {
2324 * Looks like the two contexts are clones, so we might be
2325 * able to optimize the context switch. We lock both
2326 * contexts and check that they are clones under the
2327 * lock (including re-checking that neither has been
2328 * uncloned in the meantime). It doesn't matter which
2329 * order we take the locks because no other cpu could
2330 * be trying to lock both of these tasks.
2332 raw_spin_lock(&ctx
->lock
);
2333 raw_spin_lock_nested(&next_ctx
->lock
, SINGLE_DEPTH_NESTING
);
2334 if (context_equiv(ctx
, next_ctx
)) {
2336 * XXX do we need a memory barrier of sorts
2337 * wrt to rcu_dereference() of perf_event_ctxp
2339 task
->perf_event_ctxp
[ctxn
] = next_ctx
;
2340 next
->perf_event_ctxp
[ctxn
] = ctx
;
2342 next_ctx
->task
= task
;
2345 perf_event_sync_stat(ctx
, next_ctx
);
2347 raw_spin_unlock(&next_ctx
->lock
);
2348 raw_spin_unlock(&ctx
->lock
);
2354 raw_spin_lock(&ctx
->lock
);
2355 ctx_sched_out(ctx
, cpuctx
, EVENT_ALL
);
2356 cpuctx
->task_ctx
= NULL
;
2357 raw_spin_unlock(&ctx
->lock
);
2361 #define for_each_task_context_nr(ctxn) \
2362 for ((ctxn) = 0; (ctxn) < perf_nr_task_contexts; (ctxn)++)
2365 * Called from scheduler to remove the events of the current task,
2366 * with interrupts disabled.
2368 * We stop each event and update the event value in event->count.
2370 * This does not protect us against NMI, but disable()
2371 * sets the disabled bit in the control field of event _before_
2372 * accessing the event control register. If a NMI hits, then it will
2373 * not restart the event.
2375 void __perf_event_task_sched_out(struct task_struct
*task
,
2376 struct task_struct
*next
)
2380 for_each_task_context_nr(ctxn
)
2381 perf_event_context_sched_out(task
, ctxn
, next
);
2384 * if cgroup events exist on this CPU, then we need
2385 * to check if we have to switch out PMU state.
2386 * cgroup event are system-wide mode only
2388 if (atomic_read(&__get_cpu_var(perf_cgroup_events
)))
2389 perf_cgroup_sched_out(task
, next
);
2392 static void task_ctx_sched_out(struct perf_event_context
*ctx
)
2394 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
2396 if (!cpuctx
->task_ctx
)
2399 if (WARN_ON_ONCE(ctx
!= cpuctx
->task_ctx
))
2402 ctx_sched_out(ctx
, cpuctx
, EVENT_ALL
);
2403 cpuctx
->task_ctx
= NULL
;
2407 * Called with IRQs disabled
2409 static void cpu_ctx_sched_out(struct perf_cpu_context
*cpuctx
,
2410 enum event_type_t event_type
)
2412 ctx_sched_out(&cpuctx
->ctx
, cpuctx
, event_type
);
2416 ctx_pinned_sched_in(struct perf_event_context
*ctx
,
2417 struct perf_cpu_context
*cpuctx
)
2419 struct perf_event
*event
;
2421 list_for_each_entry(event
, &ctx
->pinned_groups
, group_entry
) {
2422 if (event
->state
<= PERF_EVENT_STATE_OFF
)
2424 if (!event_filter_match(event
))
2427 /* may need to reset tstamp_enabled */
2428 if (is_cgroup_event(event
))
2429 perf_cgroup_mark_enabled(event
, ctx
);
2431 if (group_can_go_on(event
, cpuctx
, 1))
2432 group_sched_in(event
, cpuctx
, ctx
);
2435 * If this pinned group hasn't been scheduled,
2436 * put it in error state.
2438 if (event
->state
== PERF_EVENT_STATE_INACTIVE
) {
2439 update_group_times(event
);
2440 event
->state
= PERF_EVENT_STATE_ERROR
;
2446 ctx_flexible_sched_in(struct perf_event_context
*ctx
,
2447 struct perf_cpu_context
*cpuctx
)
2449 struct perf_event
*event
;
2452 list_for_each_entry(event
, &ctx
->flexible_groups
, group_entry
) {
2453 /* Ignore events in OFF or ERROR state */
2454 if (event
->state
<= PERF_EVENT_STATE_OFF
)
2457 * Listen to the 'cpu' scheduling filter constraint
2460 if (!event_filter_match(event
))
2463 /* may need to reset tstamp_enabled */
2464 if (is_cgroup_event(event
))
2465 perf_cgroup_mark_enabled(event
, ctx
);
2467 if (group_can_go_on(event
, cpuctx
, can_add_hw
)) {
2468 if (group_sched_in(event
, cpuctx
, ctx
))
2475 ctx_sched_in(struct perf_event_context
*ctx
,
2476 struct perf_cpu_context
*cpuctx
,
2477 enum event_type_t event_type
,
2478 struct task_struct
*task
)
2481 int is_active
= ctx
->is_active
;
2483 ctx
->is_active
|= event_type
;
2484 if (likely(!ctx
->nr_events
))
2488 ctx
->timestamp
= now
;
2489 perf_cgroup_set_timestamp(task
, ctx
);
2491 * First go through the list and put on any pinned groups
2492 * in order to give them the best chance of going on.
2494 if (!(is_active
& EVENT_PINNED
) && (event_type
& EVENT_PINNED
))
2495 ctx_pinned_sched_in(ctx
, cpuctx
);
2497 /* Then walk through the lower prio flexible groups */
2498 if (!(is_active
& EVENT_FLEXIBLE
) && (event_type
& EVENT_FLEXIBLE
))
2499 ctx_flexible_sched_in(ctx
, cpuctx
);
2502 static void cpu_ctx_sched_in(struct perf_cpu_context
*cpuctx
,
2503 enum event_type_t event_type
,
2504 struct task_struct
*task
)
2506 struct perf_event_context
*ctx
= &cpuctx
->ctx
;
2508 ctx_sched_in(ctx
, cpuctx
, event_type
, task
);
2511 static void perf_event_context_sched_in(struct perf_event_context
*ctx
,
2512 struct task_struct
*task
)
2514 struct perf_cpu_context
*cpuctx
;
2516 cpuctx
= __get_cpu_context(ctx
);
2517 if (cpuctx
->task_ctx
== ctx
)
2520 perf_ctx_lock(cpuctx
, ctx
);
2521 perf_pmu_disable(ctx
->pmu
);
2523 * We want to keep the following priority order:
2524 * cpu pinned (that don't need to move), task pinned,
2525 * cpu flexible, task flexible.
2527 cpu_ctx_sched_out(cpuctx
, EVENT_FLEXIBLE
);
2530 cpuctx
->task_ctx
= ctx
;
2532 perf_event_sched_in(cpuctx
, cpuctx
->task_ctx
, task
);
2534 perf_pmu_enable(ctx
->pmu
);
2535 perf_ctx_unlock(cpuctx
, ctx
);
2538 * Since these rotations are per-cpu, we need to ensure the
2539 * cpu-context we got scheduled on is actually rotating.
2541 perf_pmu_rotate_start(ctx
->pmu
);
2545 * When sampling the branck stack in system-wide, it may be necessary
2546 * to flush the stack on context switch. This happens when the branch
2547 * stack does not tag its entries with the pid of the current task.
2548 * Otherwise it becomes impossible to associate a branch entry with a
2549 * task. This ambiguity is more likely to appear when the branch stack
2550 * supports priv level filtering and the user sets it to monitor only
2551 * at the user level (which could be a useful measurement in system-wide
2552 * mode). In that case, the risk is high of having a branch stack with
2553 * branch from multiple tasks. Flushing may mean dropping the existing
2554 * entries or stashing them somewhere in the PMU specific code layer.
2556 * This function provides the context switch callback to the lower code
2557 * layer. It is invoked ONLY when there is at least one system-wide context
2558 * with at least one active event using taken branch sampling.
2560 static void perf_branch_stack_sched_in(struct task_struct
*prev
,
2561 struct task_struct
*task
)
2563 struct perf_cpu_context
*cpuctx
;
2565 unsigned long flags
;
2567 /* no need to flush branch stack if not changing task */
2571 local_irq_save(flags
);
2575 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
2576 cpuctx
= this_cpu_ptr(pmu
->pmu_cpu_context
);
2579 * check if the context has at least one
2580 * event using PERF_SAMPLE_BRANCH_STACK
2582 if (cpuctx
->ctx
.nr_branch_stack
> 0
2583 && pmu
->flush_branch_stack
) {
2585 perf_ctx_lock(cpuctx
, cpuctx
->task_ctx
);
2587 perf_pmu_disable(pmu
);
2589 pmu
->flush_branch_stack();
2591 perf_pmu_enable(pmu
);
2593 perf_ctx_unlock(cpuctx
, cpuctx
->task_ctx
);
2599 local_irq_restore(flags
);
2603 * Called from scheduler to add the events of the current task
2604 * with interrupts disabled.
2606 * We restore the event value and then enable it.
2608 * This does not protect us against NMI, but enable()
2609 * sets the enabled bit in the control field of event _before_
2610 * accessing the event control register. If a NMI hits, then it will
2611 * keep the event running.
2613 void __perf_event_task_sched_in(struct task_struct
*prev
,
2614 struct task_struct
*task
)
2616 struct perf_event_context
*ctx
;
2619 for_each_task_context_nr(ctxn
) {
2620 ctx
= task
->perf_event_ctxp
[ctxn
];
2624 perf_event_context_sched_in(ctx
, task
);
2627 * if cgroup events exist on this CPU, then we need
2628 * to check if we have to switch in PMU state.
2629 * cgroup event are system-wide mode only
2631 if (atomic_read(&__get_cpu_var(perf_cgroup_events
)))
2632 perf_cgroup_sched_in(prev
, task
);
2634 /* check for system-wide branch_stack events */
2635 if (atomic_read(&__get_cpu_var(perf_branch_stack_events
)))
2636 perf_branch_stack_sched_in(prev
, task
);
2639 static u64
perf_calculate_period(struct perf_event
*event
, u64 nsec
, u64 count
)
2641 u64 frequency
= event
->attr
.sample_freq
;
2642 u64 sec
= NSEC_PER_SEC
;
2643 u64 divisor
, dividend
;
2645 int count_fls
, nsec_fls
, frequency_fls
, sec_fls
;
2647 count_fls
= fls64(count
);
2648 nsec_fls
= fls64(nsec
);
2649 frequency_fls
= fls64(frequency
);
2653 * We got @count in @nsec, with a target of sample_freq HZ
2654 * the target period becomes:
2657 * period = -------------------
2658 * @nsec * sample_freq
2663 * Reduce accuracy by one bit such that @a and @b converge
2664 * to a similar magnitude.
2666 #define REDUCE_FLS(a, b) \
2668 if (a##_fls > b##_fls) { \
2678 * Reduce accuracy until either term fits in a u64, then proceed with
2679 * the other, so that finally we can do a u64/u64 division.
2681 while (count_fls
+ sec_fls
> 64 && nsec_fls
+ frequency_fls
> 64) {
2682 REDUCE_FLS(nsec
, frequency
);
2683 REDUCE_FLS(sec
, count
);
2686 if (count_fls
+ sec_fls
> 64) {
2687 divisor
= nsec
* frequency
;
2689 while (count_fls
+ sec_fls
> 64) {
2690 REDUCE_FLS(count
, sec
);
2694 dividend
= count
* sec
;
2696 dividend
= count
* sec
;
2698 while (nsec_fls
+ frequency_fls
> 64) {
2699 REDUCE_FLS(nsec
, frequency
);
2703 divisor
= nsec
* frequency
;
2709 return div64_u64(dividend
, divisor
);
2712 static DEFINE_PER_CPU(int, perf_throttled_count
);
2713 static DEFINE_PER_CPU(u64
, perf_throttled_seq
);
2715 static void perf_adjust_period(struct perf_event
*event
, u64 nsec
, u64 count
, bool disable
)
2717 struct hw_perf_event
*hwc
= &event
->hw
;
2718 s64 period
, sample_period
;
2721 period
= perf_calculate_period(event
, nsec
, count
);
2723 delta
= (s64
)(period
- hwc
->sample_period
);
2724 delta
= (delta
+ 7) / 8; /* low pass filter */
2726 sample_period
= hwc
->sample_period
+ delta
;
2731 hwc
->sample_period
= sample_period
;
2733 if (local64_read(&hwc
->period_left
) > 8*sample_period
) {
2735 event
->pmu
->stop(event
, PERF_EF_UPDATE
);
2737 local64_set(&hwc
->period_left
, 0);
2740 event
->pmu
->start(event
, PERF_EF_RELOAD
);
2745 * combine freq adjustment with unthrottling to avoid two passes over the
2746 * events. At the same time, make sure, having freq events does not change
2747 * the rate of unthrottling as that would introduce bias.
2749 static void perf_adjust_freq_unthr_context(struct perf_event_context
*ctx
,
2752 struct perf_event
*event
;
2753 struct hw_perf_event
*hwc
;
2754 u64 now
, period
= TICK_NSEC
;
2758 * only need to iterate over all events iff:
2759 * - context have events in frequency mode (needs freq adjust)
2760 * - there are events to unthrottle on this cpu
2762 if (!(ctx
->nr_freq
|| needs_unthr
))
2765 raw_spin_lock(&ctx
->lock
);
2766 perf_pmu_disable(ctx
->pmu
);
2768 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
2769 if (event
->state
!= PERF_EVENT_STATE_ACTIVE
)
2772 if (!event_filter_match(event
))
2775 perf_pmu_disable(event
->pmu
);
2779 if (hwc
->interrupts
== MAX_INTERRUPTS
) {
2780 hwc
->interrupts
= 0;
2781 perf_log_throttle(event
, 1);
2782 event
->pmu
->start(event
, 0);
2785 if (!event
->attr
.freq
|| !event
->attr
.sample_freq
)
2789 * stop the event and update event->count
2791 event
->pmu
->stop(event
, PERF_EF_UPDATE
);
2793 now
= local64_read(&event
->count
);
2794 delta
= now
- hwc
->freq_count_stamp
;
2795 hwc
->freq_count_stamp
= now
;
2799 * reload only if value has changed
2800 * we have stopped the event so tell that
2801 * to perf_adjust_period() to avoid stopping it
2805 perf_adjust_period(event
, period
, delta
, false);
2807 event
->pmu
->start(event
, delta
> 0 ? PERF_EF_RELOAD
: 0);
2809 perf_pmu_enable(event
->pmu
);
2812 perf_pmu_enable(ctx
->pmu
);
2813 raw_spin_unlock(&ctx
->lock
);
2817 * Round-robin a context's events:
2819 static void rotate_ctx(struct perf_event_context
*ctx
)
2822 * Rotate the first entry last of non-pinned groups. Rotation might be
2823 * disabled by the inheritance code.
2825 if (!ctx
->rotate_disable
)
2826 list_rotate_left(&ctx
->flexible_groups
);
2830 * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
2831 * because they're strictly cpu affine and rotate_start is called with IRQs
2832 * disabled, while rotate_context is called from IRQ context.
2834 static int perf_rotate_context(struct perf_cpu_context
*cpuctx
)
2836 struct perf_event_context
*ctx
= NULL
;
2837 int rotate
= 0, remove
= 1;
2839 if (cpuctx
->ctx
.nr_events
) {
2841 if (cpuctx
->ctx
.nr_events
!= cpuctx
->ctx
.nr_active
)
2845 ctx
= cpuctx
->task_ctx
;
2846 if (ctx
&& ctx
->nr_events
) {
2848 if (ctx
->nr_events
!= ctx
->nr_active
)
2855 perf_ctx_lock(cpuctx
, cpuctx
->task_ctx
);
2856 perf_pmu_disable(cpuctx
->ctx
.pmu
);
2858 cpu_ctx_sched_out(cpuctx
, EVENT_FLEXIBLE
);
2860 ctx_sched_out(ctx
, cpuctx
, EVENT_FLEXIBLE
);
2862 rotate_ctx(&cpuctx
->ctx
);
2866 perf_event_sched_in(cpuctx
, ctx
, current
);
2868 perf_pmu_enable(cpuctx
->ctx
.pmu
);
2869 perf_ctx_unlock(cpuctx
, cpuctx
->task_ctx
);
2872 list_del_init(&cpuctx
->rotation_list
);
2877 #ifdef CONFIG_NO_HZ_FULL
2878 bool perf_event_can_stop_tick(void)
2880 if (atomic_read(&nr_freq_events
) ||
2881 __this_cpu_read(perf_throttled_count
))
2888 void perf_event_task_tick(void)
2890 struct list_head
*head
= &__get_cpu_var(rotation_list
);
2891 struct perf_cpu_context
*cpuctx
, *tmp
;
2892 struct perf_event_context
*ctx
;
2895 WARN_ON(!irqs_disabled());
2897 __this_cpu_inc(perf_throttled_seq
);
2898 throttled
= __this_cpu_xchg(perf_throttled_count
, 0);
2900 list_for_each_entry_safe(cpuctx
, tmp
, head
, rotation_list
) {
2902 perf_adjust_freq_unthr_context(ctx
, throttled
);
2904 ctx
= cpuctx
->task_ctx
;
2906 perf_adjust_freq_unthr_context(ctx
, throttled
);
2910 static int event_enable_on_exec(struct perf_event
*event
,
2911 struct perf_event_context
*ctx
)
2913 if (!event
->attr
.enable_on_exec
)
2916 event
->attr
.enable_on_exec
= 0;
2917 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
)
2920 __perf_event_mark_enabled(event
);
2926 * Enable all of a task's events that have been marked enable-on-exec.
2927 * This expects task == current.
2929 static void perf_event_enable_on_exec(struct perf_event_context
*ctx
)
2931 struct perf_event
*event
;
2932 unsigned long flags
;
2936 local_irq_save(flags
);
2937 if (!ctx
|| !ctx
->nr_events
)
2941 * We must ctxsw out cgroup events to avoid conflict
2942 * when invoking perf_task_event_sched_in() later on
2943 * in this function. Otherwise we end up trying to
2944 * ctxswin cgroup events which are already scheduled
2947 perf_cgroup_sched_out(current
, NULL
);
2949 raw_spin_lock(&ctx
->lock
);
2950 task_ctx_sched_out(ctx
);
2952 list_for_each_entry(event
, &ctx
->event_list
, event_entry
) {
2953 ret
= event_enable_on_exec(event
, ctx
);
2959 * Unclone this context if we enabled any event.
2964 raw_spin_unlock(&ctx
->lock
);
2967 * Also calls ctxswin for cgroup events, if any:
2969 perf_event_context_sched_in(ctx
, ctx
->task
);
2971 local_irq_restore(flags
);
2975 * Cross CPU call to read the hardware event
2977 static void __perf_event_read(void *info
)
2979 struct perf_event
*event
= info
;
2980 struct perf_event_context
*ctx
= event
->ctx
;
2981 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
2984 * If this is a task context, we need to check whether it is
2985 * the current task context of this cpu. If not it has been
2986 * scheduled out before the smp call arrived. In that case
2987 * event->count would have been updated to a recent sample
2988 * when the event was scheduled out.
2990 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
)
2993 raw_spin_lock(&ctx
->lock
);
2994 if (ctx
->is_active
) {
2995 update_context_time(ctx
);
2996 update_cgrp_time_from_event(event
);
2998 update_event_times(event
);
2999 if (event
->state
== PERF_EVENT_STATE_ACTIVE
)
3000 event
->pmu
->read(event
);
3001 raw_spin_unlock(&ctx
->lock
);
3004 static inline u64
perf_event_count(struct perf_event
*event
)
3006 return local64_read(&event
->count
) + atomic64_read(&event
->child_count
);
3009 static u64
perf_event_read(struct perf_event
*event
)
3012 * If event is enabled and currently active on a CPU, update the
3013 * value in the event structure:
3015 if (event
->state
== PERF_EVENT_STATE_ACTIVE
) {
3016 smp_call_function_single(event
->oncpu
,
3017 __perf_event_read
, event
, 1);
3018 } else if (event
->state
== PERF_EVENT_STATE_INACTIVE
) {
3019 struct perf_event_context
*ctx
= event
->ctx
;
3020 unsigned long flags
;
3022 raw_spin_lock_irqsave(&ctx
->lock
, flags
);
3024 * may read while context is not active
3025 * (e.g., thread is blocked), in that case
3026 * we cannot update context time
3028 if (ctx
->is_active
) {
3029 update_context_time(ctx
);
3030 update_cgrp_time_from_event(event
);
3032 update_event_times(event
);
3033 raw_spin_unlock_irqrestore(&ctx
->lock
, flags
);
3036 return perf_event_count(event
);
3040 * Initialize the perf_event context in a task_struct:
3042 static void __perf_event_init_context(struct perf_event_context
*ctx
)
3044 raw_spin_lock_init(&ctx
->lock
);
3045 mutex_init(&ctx
->mutex
);
3046 INIT_LIST_HEAD(&ctx
->pinned_groups
);
3047 INIT_LIST_HEAD(&ctx
->flexible_groups
);
3048 INIT_LIST_HEAD(&ctx
->event_list
);
3049 atomic_set(&ctx
->refcount
, 1);
3052 static struct perf_event_context
*
3053 alloc_perf_context(struct pmu
*pmu
, struct task_struct
*task
)
3055 struct perf_event_context
*ctx
;
3057 ctx
= kzalloc(sizeof(struct perf_event_context
), GFP_KERNEL
);
3061 __perf_event_init_context(ctx
);
3064 get_task_struct(task
);
3071 static struct task_struct
*
3072 find_lively_task_by_vpid(pid_t vpid
)
3074 struct task_struct
*task
;
3081 task
= find_task_by_vpid(vpid
);
3083 get_task_struct(task
);
3087 return ERR_PTR(-ESRCH
);
3089 /* Reuse ptrace permission checks for now. */
3091 if (!ptrace_may_access(task
, PTRACE_MODE_READ
))
3096 put_task_struct(task
);
3097 return ERR_PTR(err
);
3102 * Returns a matching context with refcount and pincount.
3104 static struct perf_event_context
*
3105 find_get_context(struct pmu
*pmu
, struct task_struct
*task
, int cpu
)
3107 struct perf_event_context
*ctx
;
3108 struct perf_cpu_context
*cpuctx
;
3109 unsigned long flags
;
3113 /* Must be root to operate on a CPU event: */
3114 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN
))
3115 return ERR_PTR(-EACCES
);
3118 * We could be clever and allow to attach a event to an
3119 * offline CPU and activate it when the CPU comes up, but
3122 if (!cpu_online(cpu
))
3123 return ERR_PTR(-ENODEV
);
3125 cpuctx
= per_cpu_ptr(pmu
->pmu_cpu_context
, cpu
);
3134 ctxn
= pmu
->task_ctx_nr
;
3139 ctx
= perf_lock_task_context(task
, ctxn
, &flags
);
3143 raw_spin_unlock_irqrestore(&ctx
->lock
, flags
);
3145 ctx
= alloc_perf_context(pmu
, task
);
3151 mutex_lock(&task
->perf_event_mutex
);
3153 * If it has already passed perf_event_exit_task().
3154 * we must see PF_EXITING, it takes this mutex too.
3156 if (task
->flags
& PF_EXITING
)
3158 else if (task
->perf_event_ctxp
[ctxn
])
3163 rcu_assign_pointer(task
->perf_event_ctxp
[ctxn
], ctx
);
3165 mutex_unlock(&task
->perf_event_mutex
);
3167 if (unlikely(err
)) {
3179 return ERR_PTR(err
);
3182 static void perf_event_free_filter(struct perf_event
*event
);
3184 static void free_event_rcu(struct rcu_head
*head
)
3186 struct perf_event
*event
;
3188 event
= container_of(head
, struct perf_event
, rcu_head
);
3190 put_pid_ns(event
->ns
);
3191 perf_event_free_filter(event
);
3195 static void ring_buffer_put(struct ring_buffer
*rb
);
3196 static void ring_buffer_detach(struct perf_event
*event
, struct ring_buffer
*rb
);
3198 static void unaccount_event_cpu(struct perf_event
*event
, int cpu
)
3203 if (has_branch_stack(event
)) {
3204 if (!(event
->attach_state
& PERF_ATTACH_TASK
))
3205 atomic_dec(&per_cpu(perf_branch_stack_events
, cpu
));
3207 if (is_cgroup_event(event
))
3208 atomic_dec(&per_cpu(perf_cgroup_events
, cpu
));
3211 static void unaccount_event(struct perf_event
*event
)
3216 if (event
->attach_state
& PERF_ATTACH_TASK
)
3217 static_key_slow_dec_deferred(&perf_sched_events
);
3218 if (event
->attr
.mmap
|| event
->attr
.mmap_data
)
3219 atomic_dec(&nr_mmap_events
);
3220 if (event
->attr
.comm
)
3221 atomic_dec(&nr_comm_events
);
3222 if (event
->attr
.task
)
3223 atomic_dec(&nr_task_events
);
3224 if (event
->attr
.freq
)
3225 atomic_dec(&nr_freq_events
);
3226 if (is_cgroup_event(event
))
3227 static_key_slow_dec_deferred(&perf_sched_events
);
3228 if (has_branch_stack(event
))
3229 static_key_slow_dec_deferred(&perf_sched_events
);
3231 unaccount_event_cpu(event
, event
->cpu
);
3234 static void __free_event(struct perf_event
*event
)
3236 if (!event
->parent
) {
3237 if (event
->attr
.sample_type
& PERF_SAMPLE_CALLCHAIN
)
3238 put_callchain_buffers();
3242 event
->destroy(event
);
3245 put_ctx(event
->ctx
);
3247 call_rcu(&event
->rcu_head
, free_event_rcu
);
3249 static void free_event(struct perf_event
*event
)
3251 irq_work_sync(&event
->pending
);
3253 unaccount_event(event
);
3256 struct ring_buffer
*rb
;
3259 * Can happen when we close an event with re-directed output.
3261 * Since we have a 0 refcount, perf_mmap_close() will skip
3262 * over us; possibly making our ring_buffer_put() the last.
3264 mutex_lock(&event
->mmap_mutex
);
3267 rcu_assign_pointer(event
->rb
, NULL
);
3268 ring_buffer_detach(event
, rb
);
3269 ring_buffer_put(rb
); /* could be last */
3271 mutex_unlock(&event
->mmap_mutex
);
3274 if (is_cgroup_event(event
))
3275 perf_detach_cgroup(event
);
3278 __free_event(event
);
3281 int perf_event_release_kernel(struct perf_event
*event
)
3283 struct perf_event_context
*ctx
= event
->ctx
;
3285 WARN_ON_ONCE(ctx
->parent_ctx
);
3287 * There are two ways this annotation is useful:
3289 * 1) there is a lock recursion from perf_event_exit_task
3290 * see the comment there.
3292 * 2) there is a lock-inversion with mmap_sem through
3293 * perf_event_read_group(), which takes faults while
3294 * holding ctx->mutex, however this is called after
3295 * the last filedesc died, so there is no possibility
3296 * to trigger the AB-BA case.
3298 mutex_lock_nested(&ctx
->mutex
, SINGLE_DEPTH_NESTING
);
3299 raw_spin_lock_irq(&ctx
->lock
);
3300 perf_group_detach(event
);
3301 raw_spin_unlock_irq(&ctx
->lock
);
3302 perf_remove_from_context(event
);
3303 mutex_unlock(&ctx
->mutex
);
3309 EXPORT_SYMBOL_GPL(perf_event_release_kernel
);
3312 * Called when the last reference to the file is gone.
3314 static void put_event(struct perf_event
*event
)
3316 struct task_struct
*owner
;
3318 if (!atomic_long_dec_and_test(&event
->refcount
))
3322 owner
= ACCESS_ONCE(event
->owner
);
3324 * Matches the smp_wmb() in perf_event_exit_task(). If we observe
3325 * !owner it means the list deletion is complete and we can indeed
3326 * free this event, otherwise we need to serialize on
3327 * owner->perf_event_mutex.
3329 smp_read_barrier_depends();
3332 * Since delayed_put_task_struct() also drops the last
3333 * task reference we can safely take a new reference
3334 * while holding the rcu_read_lock().
3336 get_task_struct(owner
);
3341 mutex_lock(&owner
->perf_event_mutex
);
3343 * We have to re-check the event->owner field, if it is cleared
3344 * we raced with perf_event_exit_task(), acquiring the mutex
3345 * ensured they're done, and we can proceed with freeing the
3349 list_del_init(&event
->owner_entry
);
3350 mutex_unlock(&owner
->perf_event_mutex
);
3351 put_task_struct(owner
);
3354 perf_event_release_kernel(event
);
3357 static int perf_release(struct inode
*inode
, struct file
*file
)
3359 put_event(file
->private_data
);
3363 u64
perf_event_read_value(struct perf_event
*event
, u64
*enabled
, u64
*running
)
3365 struct perf_event
*child
;
3371 mutex_lock(&event
->child_mutex
);
3372 total
+= perf_event_read(event
);
3373 *enabled
+= event
->total_time_enabled
+
3374 atomic64_read(&event
->child_total_time_enabled
);
3375 *running
+= event
->total_time_running
+
3376 atomic64_read(&event
->child_total_time_running
);
3378 list_for_each_entry(child
, &event
->child_list
, child_list
) {
3379 total
+= perf_event_read(child
);
3380 *enabled
+= child
->total_time_enabled
;
3381 *running
+= child
->total_time_running
;
3383 mutex_unlock(&event
->child_mutex
);
3387 EXPORT_SYMBOL_GPL(perf_event_read_value
);
3389 static int perf_event_read_group(struct perf_event
*event
,
3390 u64 read_format
, char __user
*buf
)
3392 struct perf_event
*leader
= event
->group_leader
, *sub
;
3393 int n
= 0, size
= 0, ret
= -EFAULT
;
3394 struct perf_event_context
*ctx
= leader
->ctx
;
3396 u64 count
, enabled
, running
;
3398 mutex_lock(&ctx
->mutex
);
3399 count
= perf_event_read_value(leader
, &enabled
, &running
);
3401 values
[n
++] = 1 + leader
->nr_siblings
;
3402 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
3403 values
[n
++] = enabled
;
3404 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
3405 values
[n
++] = running
;
3406 values
[n
++] = count
;
3407 if (read_format
& PERF_FORMAT_ID
)
3408 values
[n
++] = primary_event_id(leader
);
3410 size
= n
* sizeof(u64
);
3412 if (copy_to_user(buf
, values
, size
))
3417 list_for_each_entry(sub
, &leader
->sibling_list
, group_entry
) {
3420 values
[n
++] = perf_event_read_value(sub
, &enabled
, &running
);
3421 if (read_format
& PERF_FORMAT_ID
)
3422 values
[n
++] = primary_event_id(sub
);
3424 size
= n
* sizeof(u64
);
3426 if (copy_to_user(buf
+ ret
, values
, size
)) {
3434 mutex_unlock(&ctx
->mutex
);
3439 static int perf_event_read_one(struct perf_event
*event
,
3440 u64 read_format
, char __user
*buf
)
3442 u64 enabled
, running
;
3446 values
[n
++] = perf_event_read_value(event
, &enabled
, &running
);
3447 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
3448 values
[n
++] = enabled
;
3449 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
3450 values
[n
++] = running
;
3451 if (read_format
& PERF_FORMAT_ID
)
3452 values
[n
++] = primary_event_id(event
);
3454 if (copy_to_user(buf
, values
, n
* sizeof(u64
)))
3457 return n
* sizeof(u64
);
3461 * Read the performance event - simple non blocking version for now
3464 perf_read_hw(struct perf_event
*event
, char __user
*buf
, size_t count
)
3466 u64 read_format
= event
->attr
.read_format
;
3470 * Return end-of-file for a read on a event that is in
3471 * error state (i.e. because it was pinned but it couldn't be
3472 * scheduled on to the CPU at some point).
3474 if (event
->state
== PERF_EVENT_STATE_ERROR
)
3477 if (count
< event
->read_size
)
3480 WARN_ON_ONCE(event
->ctx
->parent_ctx
);
3481 if (read_format
& PERF_FORMAT_GROUP
)
3482 ret
= perf_event_read_group(event
, read_format
, buf
);
3484 ret
= perf_event_read_one(event
, read_format
, buf
);
3490 perf_read(struct file
*file
, char __user
*buf
, size_t count
, loff_t
*ppos
)
3492 struct perf_event
*event
= file
->private_data
;
3494 return perf_read_hw(event
, buf
, count
);
3497 static unsigned int perf_poll(struct file
*file
, poll_table
*wait
)
3499 struct perf_event
*event
= file
->private_data
;
3500 struct ring_buffer
*rb
;
3501 unsigned int events
= POLL_HUP
;
3504 * Pin the event->rb by taking event->mmap_mutex; otherwise
3505 * perf_event_set_output() can swizzle our rb and make us miss wakeups.
3507 mutex_lock(&event
->mmap_mutex
);
3510 events
= atomic_xchg(&rb
->poll
, 0);
3511 mutex_unlock(&event
->mmap_mutex
);
3513 poll_wait(file
, &event
->waitq
, wait
);
3518 static void perf_event_reset(struct perf_event
*event
)
3520 (void)perf_event_read(event
);
3521 local64_set(&event
->count
, 0);
3522 perf_event_update_userpage(event
);
3526 * Holding the top-level event's child_mutex means that any
3527 * descendant process that has inherited this event will block
3528 * in sync_child_event if it goes to exit, thus satisfying the
3529 * task existence requirements of perf_event_enable/disable.
3531 static void perf_event_for_each_child(struct perf_event
*event
,
3532 void (*func
)(struct perf_event
*))
3534 struct perf_event
*child
;
3536 WARN_ON_ONCE(event
->ctx
->parent_ctx
);
3537 mutex_lock(&event
->child_mutex
);
3539 list_for_each_entry(child
, &event
->child_list
, child_list
)
3541 mutex_unlock(&event
->child_mutex
);
3544 static void perf_event_for_each(struct perf_event
*event
,
3545 void (*func
)(struct perf_event
*))
3547 struct perf_event_context
*ctx
= event
->ctx
;
3548 struct perf_event
*sibling
;
3550 WARN_ON_ONCE(ctx
->parent_ctx
);
3551 mutex_lock(&ctx
->mutex
);
3552 event
= event
->group_leader
;
3554 perf_event_for_each_child(event
, func
);
3555 list_for_each_entry(sibling
, &event
->sibling_list
, group_entry
)
3556 perf_event_for_each_child(sibling
, func
);
3557 mutex_unlock(&ctx
->mutex
);
3560 static int perf_event_period(struct perf_event
*event
, u64 __user
*arg
)
3562 struct perf_event_context
*ctx
= event
->ctx
;
3563 int ret
= 0, active
;
3566 if (!is_sampling_event(event
))
3569 if (copy_from_user(&value
, arg
, sizeof(value
)))
3575 raw_spin_lock_irq(&ctx
->lock
);
3576 if (event
->attr
.freq
) {
3577 if (value
> sysctl_perf_event_sample_rate
) {
3582 event
->attr
.sample_freq
= value
;
3584 event
->attr
.sample_period
= value
;
3585 event
->hw
.sample_period
= value
;
3588 active
= (event
->state
== PERF_EVENT_STATE_ACTIVE
);
3590 perf_pmu_disable(ctx
->pmu
);
3591 event
->pmu
->stop(event
, PERF_EF_UPDATE
);
3594 local64_set(&event
->hw
.period_left
, 0);
3597 event
->pmu
->start(event
, PERF_EF_RELOAD
);
3598 perf_pmu_enable(ctx
->pmu
);
3602 raw_spin_unlock_irq(&ctx
->lock
);
3607 static const struct file_operations perf_fops
;
3609 static inline int perf_fget_light(int fd
, struct fd
*p
)
3611 struct fd f
= fdget(fd
);
3615 if (f
.file
->f_op
!= &perf_fops
) {
3623 static int perf_event_set_output(struct perf_event
*event
,
3624 struct perf_event
*output_event
);
3625 static int perf_event_set_filter(struct perf_event
*event
, void __user
*arg
);
3627 static long perf_ioctl(struct file
*file
, unsigned int cmd
, unsigned long arg
)
3629 struct perf_event
*event
= file
->private_data
;
3630 void (*func
)(struct perf_event
*);
3634 case PERF_EVENT_IOC_ENABLE
:
3635 func
= perf_event_enable
;
3637 case PERF_EVENT_IOC_DISABLE
:
3638 func
= perf_event_disable
;
3640 case PERF_EVENT_IOC_RESET
:
3641 func
= perf_event_reset
;
3644 case PERF_EVENT_IOC_REFRESH
:
3645 return perf_event_refresh(event
, arg
);
3647 case PERF_EVENT_IOC_PERIOD
:
3648 return perf_event_period(event
, (u64 __user
*)arg
);
3650 case PERF_EVENT_IOC_ID
:
3652 u64 id
= primary_event_id(event
);
3654 if (copy_to_user((void __user
*)arg
, &id
, sizeof(id
)))
3659 case PERF_EVENT_IOC_SET_OUTPUT
:
3663 struct perf_event
*output_event
;
3665 ret
= perf_fget_light(arg
, &output
);
3668 output_event
= output
.file
->private_data
;
3669 ret
= perf_event_set_output(event
, output_event
);
3672 ret
= perf_event_set_output(event
, NULL
);
3677 case PERF_EVENT_IOC_SET_FILTER
:
3678 return perf_event_set_filter(event
, (void __user
*)arg
);
3684 if (flags
& PERF_IOC_FLAG_GROUP
)
3685 perf_event_for_each(event
, func
);
3687 perf_event_for_each_child(event
, func
);
3692 int perf_event_task_enable(void)
3694 struct perf_event
*event
;
3696 mutex_lock(¤t
->perf_event_mutex
);
3697 list_for_each_entry(event
, ¤t
->perf_event_list
, owner_entry
)
3698 perf_event_for_each_child(event
, perf_event_enable
);
3699 mutex_unlock(¤t
->perf_event_mutex
);
3704 int perf_event_task_disable(void)
3706 struct perf_event
*event
;
3708 mutex_lock(¤t
->perf_event_mutex
);
3709 list_for_each_entry(event
, ¤t
->perf_event_list
, owner_entry
)
3710 perf_event_for_each_child(event
, perf_event_disable
);
3711 mutex_unlock(¤t
->perf_event_mutex
);
3716 static int perf_event_index(struct perf_event
*event
)
3718 if (event
->hw
.state
& PERF_HES_STOPPED
)
3721 if (event
->state
!= PERF_EVENT_STATE_ACTIVE
)
3724 return event
->pmu
->event_idx(event
);
3727 static void calc_timer_values(struct perf_event
*event
,
3734 *now
= perf_clock();
3735 ctx_time
= event
->shadow_ctx_time
+ *now
;
3736 *enabled
= ctx_time
- event
->tstamp_enabled
;
3737 *running
= ctx_time
- event
->tstamp_running
;
3740 static void perf_event_init_userpage(struct perf_event
*event
)
3742 struct perf_event_mmap_page
*userpg
;
3743 struct ring_buffer
*rb
;
3746 rb
= rcu_dereference(event
->rb
);
3750 userpg
= rb
->user_page
;
3752 /* Allow new userspace to detect that bit 0 is deprecated */
3753 userpg
->cap_bit0_is_deprecated
= 1;
3754 userpg
->size
= offsetof(struct perf_event_mmap_page
, __reserved
);
3760 void __weak
arch_perf_update_userpage(struct perf_event_mmap_page
*userpg
, u64 now
)
3765 * Callers need to ensure there can be no nesting of this function, otherwise
3766 * the seqlock logic goes bad. We can not serialize this because the arch
3767 * code calls this from NMI context.
3769 void perf_event_update_userpage(struct perf_event
*event
)
3771 struct perf_event_mmap_page
*userpg
;
3772 struct ring_buffer
*rb
;
3773 u64 enabled
, running
, now
;
3776 rb
= rcu_dereference(event
->rb
);
3781 * compute total_time_enabled, total_time_running
3782 * based on snapshot values taken when the event
3783 * was last scheduled in.
3785 * we cannot simply called update_context_time()
3786 * because of locking issue as we can be called in
3789 calc_timer_values(event
, &now
, &enabled
, &running
);
3791 userpg
= rb
->user_page
;
3793 * Disable preemption so as to not let the corresponding user-space
3794 * spin too long if we get preempted.
3799 userpg
->index
= perf_event_index(event
);
3800 userpg
->offset
= perf_event_count(event
);
3802 userpg
->offset
-= local64_read(&event
->hw
.prev_count
);
3804 userpg
->time_enabled
= enabled
+
3805 atomic64_read(&event
->child_total_time_enabled
);
3807 userpg
->time_running
= running
+
3808 atomic64_read(&event
->child_total_time_running
);
3810 arch_perf_update_userpage(userpg
, now
);
3819 static int perf_mmap_fault(struct vm_area_struct
*vma
, struct vm_fault
*vmf
)
3821 struct perf_event
*event
= vma
->vm_file
->private_data
;
3822 struct ring_buffer
*rb
;
3823 int ret
= VM_FAULT_SIGBUS
;
3825 if (vmf
->flags
& FAULT_FLAG_MKWRITE
) {
3826 if (vmf
->pgoff
== 0)
3832 rb
= rcu_dereference(event
->rb
);
3836 if (vmf
->pgoff
&& (vmf
->flags
& FAULT_FLAG_WRITE
))
3839 vmf
->page
= perf_mmap_to_page(rb
, vmf
->pgoff
);
3843 get_page(vmf
->page
);
3844 vmf
->page
->mapping
= vma
->vm_file
->f_mapping
;
3845 vmf
->page
->index
= vmf
->pgoff
;
3854 static void ring_buffer_attach(struct perf_event
*event
,
3855 struct ring_buffer
*rb
)
3857 unsigned long flags
;
3859 if (!list_empty(&event
->rb_entry
))
3862 spin_lock_irqsave(&rb
->event_lock
, flags
);
3863 if (list_empty(&event
->rb_entry
))
3864 list_add(&event
->rb_entry
, &rb
->event_list
);
3865 spin_unlock_irqrestore(&rb
->event_lock
, flags
);
3868 static void ring_buffer_detach(struct perf_event
*event
, struct ring_buffer
*rb
)
3870 unsigned long flags
;
3872 if (list_empty(&event
->rb_entry
))
3875 spin_lock_irqsave(&rb
->event_lock
, flags
);
3876 list_del_init(&event
->rb_entry
);
3877 wake_up_all(&event
->waitq
);
3878 spin_unlock_irqrestore(&rb
->event_lock
, flags
);
3881 static void ring_buffer_wakeup(struct perf_event
*event
)
3883 struct ring_buffer
*rb
;
3886 rb
= rcu_dereference(event
->rb
);
3888 list_for_each_entry_rcu(event
, &rb
->event_list
, rb_entry
)
3889 wake_up_all(&event
->waitq
);
3894 static void rb_free_rcu(struct rcu_head
*rcu_head
)
3896 struct ring_buffer
*rb
;
3898 rb
= container_of(rcu_head
, struct ring_buffer
, rcu_head
);
3902 static struct ring_buffer
*ring_buffer_get(struct perf_event
*event
)
3904 struct ring_buffer
*rb
;
3907 rb
= rcu_dereference(event
->rb
);
3909 if (!atomic_inc_not_zero(&rb
->refcount
))
3917 static void ring_buffer_put(struct ring_buffer
*rb
)
3919 if (!atomic_dec_and_test(&rb
->refcount
))
3922 WARN_ON_ONCE(!list_empty(&rb
->event_list
));
3924 call_rcu(&rb
->rcu_head
, rb_free_rcu
);
3927 static void perf_mmap_open(struct vm_area_struct
*vma
)
3929 struct perf_event
*event
= vma
->vm_file
->private_data
;
3931 atomic_inc(&event
->mmap_count
);
3932 atomic_inc(&event
->rb
->mmap_count
);
3936 * A buffer can be mmap()ed multiple times; either directly through the same
3937 * event, or through other events by use of perf_event_set_output().
3939 * In order to undo the VM accounting done by perf_mmap() we need to destroy
3940 * the buffer here, where we still have a VM context. This means we need
3941 * to detach all events redirecting to us.
3943 static void perf_mmap_close(struct vm_area_struct
*vma
)
3945 struct perf_event
*event
= vma
->vm_file
->private_data
;
3947 struct ring_buffer
*rb
= event
->rb
;
3948 struct user_struct
*mmap_user
= rb
->mmap_user
;
3949 int mmap_locked
= rb
->mmap_locked
;
3950 unsigned long size
= perf_data_size(rb
);
3952 atomic_dec(&rb
->mmap_count
);
3954 if (!atomic_dec_and_mutex_lock(&event
->mmap_count
, &event
->mmap_mutex
))
3957 /* Detach current event from the buffer. */
3958 rcu_assign_pointer(event
->rb
, NULL
);
3959 ring_buffer_detach(event
, rb
);
3960 mutex_unlock(&event
->mmap_mutex
);
3962 /* If there's still other mmap()s of this buffer, we're done. */
3963 if (atomic_read(&rb
->mmap_count
)) {
3964 ring_buffer_put(rb
); /* can't be last */
3969 * No other mmap()s, detach from all other events that might redirect
3970 * into the now unreachable buffer. Somewhat complicated by the
3971 * fact that rb::event_lock otherwise nests inside mmap_mutex.
3975 list_for_each_entry_rcu(event
, &rb
->event_list
, rb_entry
) {
3976 if (!atomic_long_inc_not_zero(&event
->refcount
)) {
3978 * This event is en-route to free_event() which will
3979 * detach it and remove it from the list.
3985 mutex_lock(&event
->mmap_mutex
);
3987 * Check we didn't race with perf_event_set_output() which can
3988 * swizzle the rb from under us while we were waiting to
3989 * acquire mmap_mutex.
3991 * If we find a different rb; ignore this event, a next
3992 * iteration will no longer find it on the list. We have to
3993 * still restart the iteration to make sure we're not now
3994 * iterating the wrong list.
3996 if (event
->rb
== rb
) {
3997 rcu_assign_pointer(event
->rb
, NULL
);
3998 ring_buffer_detach(event
, rb
);
3999 ring_buffer_put(rb
); /* can't be last, we still have one */
4001 mutex_unlock(&event
->mmap_mutex
);
4005 * Restart the iteration; either we're on the wrong list or
4006 * destroyed its integrity by doing a deletion.
4013 * It could be there's still a few 0-ref events on the list; they'll
4014 * get cleaned up by free_event() -- they'll also still have their
4015 * ref on the rb and will free it whenever they are done with it.
4017 * Aside from that, this buffer is 'fully' detached and unmapped,
4018 * undo the VM accounting.
4021 atomic_long_sub((size
>> PAGE_SHIFT
) + 1, &mmap_user
->locked_vm
);
4022 vma
->vm_mm
->pinned_vm
-= mmap_locked
;
4023 free_uid(mmap_user
);
4025 ring_buffer_put(rb
); /* could be last */
4028 static const struct vm_operations_struct perf_mmap_vmops
= {
4029 .open
= perf_mmap_open
,
4030 .close
= perf_mmap_close
,
4031 .fault
= perf_mmap_fault
,
4032 .page_mkwrite
= perf_mmap_fault
,
4035 static int perf_mmap(struct file
*file
, struct vm_area_struct
*vma
)
4037 struct perf_event
*event
= file
->private_data
;
4038 unsigned long user_locked
, user_lock_limit
;
4039 struct user_struct
*user
= current_user();
4040 unsigned long locked
, lock_limit
;
4041 struct ring_buffer
*rb
;
4042 unsigned long vma_size
;
4043 unsigned long nr_pages
;
4044 long user_extra
, extra
;
4045 int ret
= 0, flags
= 0;
4048 * Don't allow mmap() of inherited per-task counters. This would
4049 * create a performance issue due to all children writing to the
4052 if (event
->cpu
== -1 && event
->attr
.inherit
)
4055 if (!(vma
->vm_flags
& VM_SHARED
))
4058 vma_size
= vma
->vm_end
- vma
->vm_start
;
4059 nr_pages
= (vma_size
/ PAGE_SIZE
) - 1;
4062 * If we have rb pages ensure they're a power-of-two number, so we
4063 * can do bitmasks instead of modulo.
4065 if (nr_pages
!= 0 && !is_power_of_2(nr_pages
))
4068 if (vma_size
!= PAGE_SIZE
* (1 + nr_pages
))
4071 if (vma
->vm_pgoff
!= 0)
4074 WARN_ON_ONCE(event
->ctx
->parent_ctx
);
4076 mutex_lock(&event
->mmap_mutex
);
4078 if (event
->rb
->nr_pages
!= nr_pages
) {
4083 if (!atomic_inc_not_zero(&event
->rb
->mmap_count
)) {
4085 * Raced against perf_mmap_close() through
4086 * perf_event_set_output(). Try again, hope for better
4089 mutex_unlock(&event
->mmap_mutex
);
4096 user_extra
= nr_pages
+ 1;
4097 user_lock_limit
= sysctl_perf_event_mlock
>> (PAGE_SHIFT
- 10);
4100 * Increase the limit linearly with more CPUs:
4102 user_lock_limit
*= num_online_cpus();
4104 user_locked
= atomic_long_read(&user
->locked_vm
) + user_extra
;
4107 if (user_locked
> user_lock_limit
)
4108 extra
= user_locked
- user_lock_limit
;
4110 lock_limit
= rlimit(RLIMIT_MEMLOCK
);
4111 lock_limit
>>= PAGE_SHIFT
;
4112 locked
= vma
->vm_mm
->pinned_vm
+ extra
;
4114 if ((locked
> lock_limit
) && perf_paranoid_tracepoint_raw() &&
4115 !capable(CAP_IPC_LOCK
)) {
4122 if (vma
->vm_flags
& VM_WRITE
)
4123 flags
|= RING_BUFFER_WRITABLE
;
4125 rb
= rb_alloc(nr_pages
,
4126 event
->attr
.watermark
? event
->attr
.wakeup_watermark
: 0,
4134 atomic_set(&rb
->mmap_count
, 1);
4135 rb
->mmap_locked
= extra
;
4136 rb
->mmap_user
= get_current_user();
4138 atomic_long_add(user_extra
, &user
->locked_vm
);
4139 vma
->vm_mm
->pinned_vm
+= extra
;
4141 ring_buffer_attach(event
, rb
);
4142 rcu_assign_pointer(event
->rb
, rb
);
4144 perf_event_init_userpage(event
);
4145 perf_event_update_userpage(event
);
4149 atomic_inc(&event
->mmap_count
);
4150 mutex_unlock(&event
->mmap_mutex
);
4153 * Since pinned accounting is per vm we cannot allow fork() to copy our
4156 vma
->vm_flags
|= VM_DONTCOPY
| VM_DONTEXPAND
| VM_DONTDUMP
;
4157 vma
->vm_ops
= &perf_mmap_vmops
;
4162 static int perf_fasync(int fd
, struct file
*filp
, int on
)
4164 struct inode
*inode
= file_inode(filp
);
4165 struct perf_event
*event
= filp
->private_data
;
4168 mutex_lock(&inode
->i_mutex
);
4169 retval
= fasync_helper(fd
, filp
, on
, &event
->fasync
);
4170 mutex_unlock(&inode
->i_mutex
);
4178 static const struct file_operations perf_fops
= {
4179 .llseek
= no_llseek
,
4180 .release
= perf_release
,
4183 .unlocked_ioctl
= perf_ioctl
,
4184 .compat_ioctl
= perf_ioctl
,
4186 .fasync
= perf_fasync
,
4192 * If there's data, ensure we set the poll() state and publish everything
4193 * to user-space before waking everybody up.
4196 void perf_event_wakeup(struct perf_event
*event
)
4198 ring_buffer_wakeup(event
);
4200 if (event
->pending_kill
) {
4201 kill_fasync(&event
->fasync
, SIGIO
, event
->pending_kill
);
4202 event
->pending_kill
= 0;
4206 static void perf_pending_event(struct irq_work
*entry
)
4208 struct perf_event
*event
= container_of(entry
,
4209 struct perf_event
, pending
);
4211 if (event
->pending_disable
) {
4212 event
->pending_disable
= 0;
4213 __perf_event_disable(event
);
4216 if (event
->pending_wakeup
) {
4217 event
->pending_wakeup
= 0;
4218 perf_event_wakeup(event
);
4223 * We assume there is only KVM supporting the callbacks.
4224 * Later on, we might change it to a list if there is
4225 * another virtualization implementation supporting the callbacks.
4227 struct perf_guest_info_callbacks
*perf_guest_cbs
;
4229 int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks
*cbs
)
4231 perf_guest_cbs
= cbs
;
4234 EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks
);
4236 int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks
*cbs
)
4238 perf_guest_cbs
= NULL
;
4241 EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks
);
4244 perf_output_sample_regs(struct perf_output_handle
*handle
,
4245 struct pt_regs
*regs
, u64 mask
)
4249 for_each_set_bit(bit
, (const unsigned long *) &mask
,
4250 sizeof(mask
) * BITS_PER_BYTE
) {
4253 val
= perf_reg_value(regs
, bit
);
4254 perf_output_put(handle
, val
);
4258 static void perf_sample_regs_user(struct perf_regs_user
*regs_user
,
4259 struct pt_regs
*regs
)
4261 if (!user_mode(regs
)) {
4263 regs
= task_pt_regs(current
);
4269 regs_user
->regs
= regs
;
4270 regs_user
->abi
= perf_reg_abi(current
);
4275 * Get remaining task size from user stack pointer.
4277 * It'd be better to take stack vma map and limit this more
4278 * precisly, but there's no way to get it safely under interrupt,
4279 * so using TASK_SIZE as limit.
4281 static u64
perf_ustack_task_size(struct pt_regs
*regs
)
4283 unsigned long addr
= perf_user_stack_pointer(regs
);
4285 if (!addr
|| addr
>= TASK_SIZE
)
4288 return TASK_SIZE
- addr
;
4292 perf_sample_ustack_size(u16 stack_size
, u16 header_size
,
4293 struct pt_regs
*regs
)
4297 /* No regs, no stack pointer, no dump. */
4302 * Check if we fit in with the requested stack size into the:
4304 * If we don't, we limit the size to the TASK_SIZE.
4306 * - remaining sample size
4307 * If we don't, we customize the stack size to
4308 * fit in to the remaining sample size.
4311 task_size
= min((u64
) USHRT_MAX
, perf_ustack_task_size(regs
));
4312 stack_size
= min(stack_size
, (u16
) task_size
);
4314 /* Current header size plus static size and dynamic size. */
4315 header_size
+= 2 * sizeof(u64
);
4317 /* Do we fit in with the current stack dump size? */
4318 if ((u16
) (header_size
+ stack_size
) < header_size
) {
4320 * If we overflow the maximum size for the sample,
4321 * we customize the stack dump size to fit in.
4323 stack_size
= USHRT_MAX
- header_size
- sizeof(u64
);
4324 stack_size
= round_up(stack_size
, sizeof(u64
));
4331 perf_output_sample_ustack(struct perf_output_handle
*handle
, u64 dump_size
,
4332 struct pt_regs
*regs
)
4334 /* Case of a kernel thread, nothing to dump */
4337 perf_output_put(handle
, size
);
4346 * - the size requested by user or the best one we can fit
4347 * in to the sample max size
4349 * - user stack dump data
4351 * - the actual dumped size
4355 perf_output_put(handle
, dump_size
);
4358 sp
= perf_user_stack_pointer(regs
);
4359 rem
= __output_copy_user(handle
, (void *) sp
, dump_size
);
4360 dyn_size
= dump_size
- rem
;
4362 perf_output_skip(handle
, rem
);
4365 perf_output_put(handle
, dyn_size
);
4369 static void __perf_event_header__init_id(struct perf_event_header
*header
,
4370 struct perf_sample_data
*data
,
4371 struct perf_event
*event
)
4373 u64 sample_type
= event
->attr
.sample_type
;
4375 data
->type
= sample_type
;
4376 header
->size
+= event
->id_header_size
;
4378 if (sample_type
& PERF_SAMPLE_TID
) {
4379 /* namespace issues */
4380 data
->tid_entry
.pid
= perf_event_pid(event
, current
);
4381 data
->tid_entry
.tid
= perf_event_tid(event
, current
);
4384 if (sample_type
& PERF_SAMPLE_TIME
)
4385 data
->time
= perf_clock();
4387 if (sample_type
& (PERF_SAMPLE_ID
| PERF_SAMPLE_IDENTIFIER
))
4388 data
->id
= primary_event_id(event
);
4390 if (sample_type
& PERF_SAMPLE_STREAM_ID
)
4391 data
->stream_id
= event
->id
;
4393 if (sample_type
& PERF_SAMPLE_CPU
) {
4394 data
->cpu_entry
.cpu
= raw_smp_processor_id();
4395 data
->cpu_entry
.reserved
= 0;
4399 void perf_event_header__init_id(struct perf_event_header
*header
,
4400 struct perf_sample_data
*data
,
4401 struct perf_event
*event
)
4403 if (event
->attr
.sample_id_all
)
4404 __perf_event_header__init_id(header
, data
, event
);
4407 static void __perf_event__output_id_sample(struct perf_output_handle
*handle
,
4408 struct perf_sample_data
*data
)
4410 u64 sample_type
= data
->type
;
4412 if (sample_type
& PERF_SAMPLE_TID
)
4413 perf_output_put(handle
, data
->tid_entry
);
4415 if (sample_type
& PERF_SAMPLE_TIME
)
4416 perf_output_put(handle
, data
->time
);
4418 if (sample_type
& PERF_SAMPLE_ID
)
4419 perf_output_put(handle
, data
->id
);
4421 if (sample_type
& PERF_SAMPLE_STREAM_ID
)
4422 perf_output_put(handle
, data
->stream_id
);
4424 if (sample_type
& PERF_SAMPLE_CPU
)
4425 perf_output_put(handle
, data
->cpu_entry
);
4427 if (sample_type
& PERF_SAMPLE_IDENTIFIER
)
4428 perf_output_put(handle
, data
->id
);
4431 void perf_event__output_id_sample(struct perf_event
*event
,
4432 struct perf_output_handle
*handle
,
4433 struct perf_sample_data
*sample
)
4435 if (event
->attr
.sample_id_all
)
4436 __perf_event__output_id_sample(handle
, sample
);
4439 static void perf_output_read_one(struct perf_output_handle
*handle
,
4440 struct perf_event
*event
,
4441 u64 enabled
, u64 running
)
4443 u64 read_format
= event
->attr
.read_format
;
4447 values
[n
++] = perf_event_count(event
);
4448 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
) {
4449 values
[n
++] = enabled
+
4450 atomic64_read(&event
->child_total_time_enabled
);
4452 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
) {
4453 values
[n
++] = running
+
4454 atomic64_read(&event
->child_total_time_running
);
4456 if (read_format
& PERF_FORMAT_ID
)
4457 values
[n
++] = primary_event_id(event
);
4459 __output_copy(handle
, values
, n
* sizeof(u64
));
4463 * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
4465 static void perf_output_read_group(struct perf_output_handle
*handle
,
4466 struct perf_event
*event
,
4467 u64 enabled
, u64 running
)
4469 struct perf_event
*leader
= event
->group_leader
, *sub
;
4470 u64 read_format
= event
->attr
.read_format
;
4474 values
[n
++] = 1 + leader
->nr_siblings
;
4476 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
4477 values
[n
++] = enabled
;
4479 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
4480 values
[n
++] = running
;
4482 if (leader
!= event
)
4483 leader
->pmu
->read(leader
);
4485 values
[n
++] = perf_event_count(leader
);
4486 if (read_format
& PERF_FORMAT_ID
)
4487 values
[n
++] = primary_event_id(leader
);
4489 __output_copy(handle
, values
, n
* sizeof(u64
));
4491 list_for_each_entry(sub
, &leader
->sibling_list
, group_entry
) {
4494 if ((sub
!= event
) &&
4495 (sub
->state
== PERF_EVENT_STATE_ACTIVE
))
4496 sub
->pmu
->read(sub
);
4498 values
[n
++] = perf_event_count(sub
);
4499 if (read_format
& PERF_FORMAT_ID
)
4500 values
[n
++] = primary_event_id(sub
);
4502 __output_copy(handle
, values
, n
* sizeof(u64
));
4506 #define PERF_FORMAT_TOTAL_TIMES (PERF_FORMAT_TOTAL_TIME_ENABLED|\
4507 PERF_FORMAT_TOTAL_TIME_RUNNING)
4509 static void perf_output_read(struct perf_output_handle
*handle
,
4510 struct perf_event
*event
)
4512 u64 enabled
= 0, running
= 0, now
;
4513 u64 read_format
= event
->attr
.read_format
;
4516 * compute total_time_enabled, total_time_running
4517 * based on snapshot values taken when the event
4518 * was last scheduled in.
4520 * we cannot simply called update_context_time()
4521 * because of locking issue as we are called in
4524 if (read_format
& PERF_FORMAT_TOTAL_TIMES
)
4525 calc_timer_values(event
, &now
, &enabled
, &running
);
4527 if (event
->attr
.read_format
& PERF_FORMAT_GROUP
)
4528 perf_output_read_group(handle
, event
, enabled
, running
);
4530 perf_output_read_one(handle
, event
, enabled
, running
);
4533 void perf_output_sample(struct perf_output_handle
*handle
,
4534 struct perf_event_header
*header
,
4535 struct perf_sample_data
*data
,
4536 struct perf_event
*event
)
4538 u64 sample_type
= data
->type
;
4540 perf_output_put(handle
, *header
);
4542 if (sample_type
& PERF_SAMPLE_IDENTIFIER
)
4543 perf_output_put(handle
, data
->id
);
4545 if (sample_type
& PERF_SAMPLE_IP
)
4546 perf_output_put(handle
, data
->ip
);
4548 if (sample_type
& PERF_SAMPLE_TID
)
4549 perf_output_put(handle
, data
->tid_entry
);
4551 if (sample_type
& PERF_SAMPLE_TIME
)
4552 perf_output_put(handle
, data
->time
);
4554 if (sample_type
& PERF_SAMPLE_ADDR
)
4555 perf_output_put(handle
, data
->addr
);
4557 if (sample_type
& PERF_SAMPLE_ID
)
4558 perf_output_put(handle
, data
->id
);
4560 if (sample_type
& PERF_SAMPLE_STREAM_ID
)
4561 perf_output_put(handle
, data
->stream_id
);
4563 if (sample_type
& PERF_SAMPLE_CPU
)
4564 perf_output_put(handle
, data
->cpu_entry
);
4566 if (sample_type
& PERF_SAMPLE_PERIOD
)
4567 perf_output_put(handle
, data
->period
);
4569 if (sample_type
& PERF_SAMPLE_READ
)
4570 perf_output_read(handle
, event
);
4572 if (sample_type
& PERF_SAMPLE_CALLCHAIN
) {
4573 if (data
->callchain
) {
4576 if (data
->callchain
)
4577 size
+= data
->callchain
->nr
;
4579 size
*= sizeof(u64
);
4581 __output_copy(handle
, data
->callchain
, size
);
4584 perf_output_put(handle
, nr
);
4588 if (sample_type
& PERF_SAMPLE_RAW
) {
4590 perf_output_put(handle
, data
->raw
->size
);
4591 __output_copy(handle
, data
->raw
->data
,
4598 .size
= sizeof(u32
),
4601 perf_output_put(handle
, raw
);
4605 if (sample_type
& PERF_SAMPLE_BRANCH_STACK
) {
4606 if (data
->br_stack
) {
4609 size
= data
->br_stack
->nr
4610 * sizeof(struct perf_branch_entry
);
4612 perf_output_put(handle
, data
->br_stack
->nr
);
4613 perf_output_copy(handle
, data
->br_stack
->entries
, size
);
4616 * we always store at least the value of nr
4619 perf_output_put(handle
, nr
);
4623 if (sample_type
& PERF_SAMPLE_REGS_USER
) {
4624 u64 abi
= data
->regs_user
.abi
;
4627 * If there are no regs to dump, notice it through
4628 * first u64 being zero (PERF_SAMPLE_REGS_ABI_NONE).
4630 perf_output_put(handle
, abi
);
4633 u64 mask
= event
->attr
.sample_regs_user
;
4634 perf_output_sample_regs(handle
,
4635 data
->regs_user
.regs
,
4640 if (sample_type
& PERF_SAMPLE_STACK_USER
) {
4641 perf_output_sample_ustack(handle
,
4642 data
->stack_user_size
,
4643 data
->regs_user
.regs
);
4646 if (sample_type
& PERF_SAMPLE_WEIGHT
)
4647 perf_output_put(handle
, data
->weight
);
4649 if (sample_type
& PERF_SAMPLE_DATA_SRC
)
4650 perf_output_put(handle
, data
->data_src
.val
);
4652 if (sample_type
& PERF_SAMPLE_TRANSACTION
)
4653 perf_output_put(handle
, data
->txn
);
4655 if (!event
->attr
.watermark
) {
4656 int wakeup_events
= event
->attr
.wakeup_events
;
4658 if (wakeup_events
) {
4659 struct ring_buffer
*rb
= handle
->rb
;
4660 int events
= local_inc_return(&rb
->events
);
4662 if (events
>= wakeup_events
) {
4663 local_sub(wakeup_events
, &rb
->events
);
4664 local_inc(&rb
->wakeup
);
4670 void perf_prepare_sample(struct perf_event_header
*header
,
4671 struct perf_sample_data
*data
,
4672 struct perf_event
*event
,
4673 struct pt_regs
*regs
)
4675 u64 sample_type
= event
->attr
.sample_type
;
4677 header
->type
= PERF_RECORD_SAMPLE
;
4678 header
->size
= sizeof(*header
) + event
->header_size
;
4681 header
->misc
|= perf_misc_flags(regs
);
4683 __perf_event_header__init_id(header
, data
, event
);
4685 if (sample_type
& PERF_SAMPLE_IP
)
4686 data
->ip
= perf_instruction_pointer(regs
);
4688 if (sample_type
& PERF_SAMPLE_CALLCHAIN
) {
4691 data
->callchain
= perf_callchain(event
, regs
);
4693 if (data
->callchain
)
4694 size
+= data
->callchain
->nr
;
4696 header
->size
+= size
* sizeof(u64
);
4699 if (sample_type
& PERF_SAMPLE_RAW
) {
4700 int size
= sizeof(u32
);
4703 size
+= data
->raw
->size
;
4705 size
+= sizeof(u32
);
4707 WARN_ON_ONCE(size
& (sizeof(u64
)-1));
4708 header
->size
+= size
;
4711 if (sample_type
& PERF_SAMPLE_BRANCH_STACK
) {
4712 int size
= sizeof(u64
); /* nr */
4713 if (data
->br_stack
) {
4714 size
+= data
->br_stack
->nr
4715 * sizeof(struct perf_branch_entry
);
4717 header
->size
+= size
;
4720 if (sample_type
& PERF_SAMPLE_REGS_USER
) {
4721 /* regs dump ABI info */
4722 int size
= sizeof(u64
);
4724 perf_sample_regs_user(&data
->regs_user
, regs
);
4726 if (data
->regs_user
.regs
) {
4727 u64 mask
= event
->attr
.sample_regs_user
;
4728 size
+= hweight64(mask
) * sizeof(u64
);
4731 header
->size
+= size
;
4734 if (sample_type
& PERF_SAMPLE_STACK_USER
) {
4736 * Either we need PERF_SAMPLE_STACK_USER bit to be allways
4737 * processed as the last one or have additional check added
4738 * in case new sample type is added, because we could eat
4739 * up the rest of the sample size.
4741 struct perf_regs_user
*uregs
= &data
->regs_user
;
4742 u16 stack_size
= event
->attr
.sample_stack_user
;
4743 u16 size
= sizeof(u64
);
4746 perf_sample_regs_user(uregs
, regs
);
4748 stack_size
= perf_sample_ustack_size(stack_size
, header
->size
,
4752 * If there is something to dump, add space for the dump
4753 * itself and for the field that tells the dynamic size,
4754 * which is how many have been actually dumped.
4757 size
+= sizeof(u64
) + stack_size
;
4759 data
->stack_user_size
= stack_size
;
4760 header
->size
+= size
;
4764 static void perf_event_output(struct perf_event
*event
,
4765 struct perf_sample_data
*data
,
4766 struct pt_regs
*regs
)
4768 struct perf_output_handle handle
;
4769 struct perf_event_header header
;
4771 /* protect the callchain buffers */
4774 perf_prepare_sample(&header
, data
, event
, regs
);
4776 if (perf_output_begin(&handle
, event
, header
.size
))
4779 perf_output_sample(&handle
, &header
, data
, event
);
4781 perf_output_end(&handle
);
4791 struct perf_read_event
{
4792 struct perf_event_header header
;
4799 perf_event_read_event(struct perf_event
*event
,
4800 struct task_struct
*task
)
4802 struct perf_output_handle handle
;
4803 struct perf_sample_data sample
;
4804 struct perf_read_event read_event
= {
4806 .type
= PERF_RECORD_READ
,
4808 .size
= sizeof(read_event
) + event
->read_size
,
4810 .pid
= perf_event_pid(event
, task
),
4811 .tid
= perf_event_tid(event
, task
),
4815 perf_event_header__init_id(&read_event
.header
, &sample
, event
);
4816 ret
= perf_output_begin(&handle
, event
, read_event
.header
.size
);
4820 perf_output_put(&handle
, read_event
);
4821 perf_output_read(&handle
, event
);
4822 perf_event__output_id_sample(event
, &handle
, &sample
);
4824 perf_output_end(&handle
);
4827 typedef void (perf_event_aux_output_cb
)(struct perf_event
*event
, void *data
);
4830 perf_event_aux_ctx(struct perf_event_context
*ctx
,
4831 perf_event_aux_output_cb output
,
4834 struct perf_event
*event
;
4836 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
4837 if (event
->state
< PERF_EVENT_STATE_INACTIVE
)
4839 if (!event_filter_match(event
))
4841 output(event
, data
);
4846 perf_event_aux(perf_event_aux_output_cb output
, void *data
,
4847 struct perf_event_context
*task_ctx
)
4849 struct perf_cpu_context
*cpuctx
;
4850 struct perf_event_context
*ctx
;
4855 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
4856 cpuctx
= get_cpu_ptr(pmu
->pmu_cpu_context
);
4857 if (cpuctx
->unique_pmu
!= pmu
)
4859 perf_event_aux_ctx(&cpuctx
->ctx
, output
, data
);
4862 ctxn
= pmu
->task_ctx_nr
;
4865 ctx
= rcu_dereference(current
->perf_event_ctxp
[ctxn
]);
4867 perf_event_aux_ctx(ctx
, output
, data
);
4869 put_cpu_ptr(pmu
->pmu_cpu_context
);
4874 perf_event_aux_ctx(task_ctx
, output
, data
);
4881 * task tracking -- fork/exit
4883 * enabled by: attr.comm | attr.mmap | attr.mmap2 | attr.mmap_data | attr.task
4886 struct perf_task_event
{
4887 struct task_struct
*task
;
4888 struct perf_event_context
*task_ctx
;
4891 struct perf_event_header header
;
4901 static int perf_event_task_match(struct perf_event
*event
)
4903 return event
->attr
.comm
|| event
->attr
.mmap
||
4904 event
->attr
.mmap2
|| event
->attr
.mmap_data
||
4908 static void perf_event_task_output(struct perf_event
*event
,
4911 struct perf_task_event
*task_event
= data
;
4912 struct perf_output_handle handle
;
4913 struct perf_sample_data sample
;
4914 struct task_struct
*task
= task_event
->task
;
4915 int ret
, size
= task_event
->event_id
.header
.size
;
4917 if (!perf_event_task_match(event
))
4920 perf_event_header__init_id(&task_event
->event_id
.header
, &sample
, event
);
4922 ret
= perf_output_begin(&handle
, event
,
4923 task_event
->event_id
.header
.size
);
4927 task_event
->event_id
.pid
= perf_event_pid(event
, task
);
4928 task_event
->event_id
.ppid
= perf_event_pid(event
, current
);
4930 task_event
->event_id
.tid
= perf_event_tid(event
, task
);
4931 task_event
->event_id
.ptid
= perf_event_tid(event
, current
);
4933 perf_output_put(&handle
, task_event
->event_id
);
4935 perf_event__output_id_sample(event
, &handle
, &sample
);
4937 perf_output_end(&handle
);
4939 task_event
->event_id
.header
.size
= size
;
4942 static void perf_event_task(struct task_struct
*task
,
4943 struct perf_event_context
*task_ctx
,
4946 struct perf_task_event task_event
;
4948 if (!atomic_read(&nr_comm_events
) &&
4949 !atomic_read(&nr_mmap_events
) &&
4950 !atomic_read(&nr_task_events
))
4953 task_event
= (struct perf_task_event
){
4955 .task_ctx
= task_ctx
,
4958 .type
= new ? PERF_RECORD_FORK
: PERF_RECORD_EXIT
,
4960 .size
= sizeof(task_event
.event_id
),
4966 .time
= perf_clock(),
4970 perf_event_aux(perf_event_task_output
,
4975 void perf_event_fork(struct task_struct
*task
)
4977 perf_event_task(task
, NULL
, 1);
4984 struct perf_comm_event
{
4985 struct task_struct
*task
;
4990 struct perf_event_header header
;
4997 static int perf_event_comm_match(struct perf_event
*event
)
4999 return event
->attr
.comm
;
5002 static void perf_event_comm_output(struct perf_event
*event
,
5005 struct perf_comm_event
*comm_event
= data
;
5006 struct perf_output_handle handle
;
5007 struct perf_sample_data sample
;
5008 int size
= comm_event
->event_id
.header
.size
;
5011 if (!perf_event_comm_match(event
))
5014 perf_event_header__init_id(&comm_event
->event_id
.header
, &sample
, event
);
5015 ret
= perf_output_begin(&handle
, event
,
5016 comm_event
->event_id
.header
.size
);
5021 comm_event
->event_id
.pid
= perf_event_pid(event
, comm_event
->task
);
5022 comm_event
->event_id
.tid
= perf_event_tid(event
, comm_event
->task
);
5024 perf_output_put(&handle
, comm_event
->event_id
);
5025 __output_copy(&handle
, comm_event
->comm
,
5026 comm_event
->comm_size
);
5028 perf_event__output_id_sample(event
, &handle
, &sample
);
5030 perf_output_end(&handle
);
5032 comm_event
->event_id
.header
.size
= size
;
5035 static void perf_event_comm_event(struct perf_comm_event
*comm_event
)
5037 char comm
[TASK_COMM_LEN
];
5040 memset(comm
, 0, sizeof(comm
));
5041 strlcpy(comm
, comm_event
->task
->comm
, sizeof(comm
));
5042 size
= ALIGN(strlen(comm
)+1, sizeof(u64
));
5044 comm_event
->comm
= comm
;
5045 comm_event
->comm_size
= size
;
5047 comm_event
->event_id
.header
.size
= sizeof(comm_event
->event_id
) + size
;
5049 perf_event_aux(perf_event_comm_output
,
5054 void perf_event_comm(struct task_struct
*task
)
5056 struct perf_comm_event comm_event
;
5057 struct perf_event_context
*ctx
;
5061 for_each_task_context_nr(ctxn
) {
5062 ctx
= task
->perf_event_ctxp
[ctxn
];
5066 perf_event_enable_on_exec(ctx
);
5070 if (!atomic_read(&nr_comm_events
))
5073 comm_event
= (struct perf_comm_event
){
5079 .type
= PERF_RECORD_COMM
,
5088 perf_event_comm_event(&comm_event
);
5095 struct perf_mmap_event
{
5096 struct vm_area_struct
*vma
;
5098 const char *file_name
;
5105 struct perf_event_header header
;
5115 static int perf_event_mmap_match(struct perf_event
*event
,
5118 struct perf_mmap_event
*mmap_event
= data
;
5119 struct vm_area_struct
*vma
= mmap_event
->vma
;
5120 int executable
= vma
->vm_flags
& VM_EXEC
;
5122 return (!executable
&& event
->attr
.mmap_data
) ||
5123 (executable
&& (event
->attr
.mmap
|| event
->attr
.mmap2
));
5126 static void perf_event_mmap_output(struct perf_event
*event
,
5129 struct perf_mmap_event
*mmap_event
= data
;
5130 struct perf_output_handle handle
;
5131 struct perf_sample_data sample
;
5132 int size
= mmap_event
->event_id
.header
.size
;
5135 if (!perf_event_mmap_match(event
, data
))
5138 if (event
->attr
.mmap2
) {
5139 mmap_event
->event_id
.header
.type
= PERF_RECORD_MMAP2
;
5140 mmap_event
->event_id
.header
.size
+= sizeof(mmap_event
->maj
);
5141 mmap_event
->event_id
.header
.size
+= sizeof(mmap_event
->min
);
5142 mmap_event
->event_id
.header
.size
+= sizeof(mmap_event
->ino
);
5143 mmap_event
->event_id
.header
.size
+= sizeof(mmap_event
->ino_generation
);
5146 perf_event_header__init_id(&mmap_event
->event_id
.header
, &sample
, event
);
5147 ret
= perf_output_begin(&handle
, event
,
5148 mmap_event
->event_id
.header
.size
);
5152 mmap_event
->event_id
.pid
= perf_event_pid(event
, current
);
5153 mmap_event
->event_id
.tid
= perf_event_tid(event
, current
);
5155 perf_output_put(&handle
, mmap_event
->event_id
);
5157 if (event
->attr
.mmap2
) {
5158 perf_output_put(&handle
, mmap_event
->maj
);
5159 perf_output_put(&handle
, mmap_event
->min
);
5160 perf_output_put(&handle
, mmap_event
->ino
);
5161 perf_output_put(&handle
, mmap_event
->ino_generation
);
5164 __output_copy(&handle
, mmap_event
->file_name
,
5165 mmap_event
->file_size
);
5167 perf_event__output_id_sample(event
, &handle
, &sample
);
5169 perf_output_end(&handle
);
5171 mmap_event
->event_id
.header
.size
= size
;
5174 static void perf_event_mmap_event(struct perf_mmap_event
*mmap_event
)
5176 struct vm_area_struct
*vma
= mmap_event
->vma
;
5177 struct file
*file
= vma
->vm_file
;
5178 int maj
= 0, min
= 0;
5179 u64 ino
= 0, gen
= 0;
5186 struct inode
*inode
;
5189 buf
= kmalloc(PATH_MAX
, GFP_KERNEL
);
5195 * d_path() works from the end of the rb backwards, so we
5196 * need to add enough zero bytes after the string to handle
5197 * the 64bit alignment we do later.
5199 name
= d_path(&file
->f_path
, buf
, PATH_MAX
- sizeof(u64
));
5204 inode
= file_inode(vma
->vm_file
);
5205 dev
= inode
->i_sb
->s_dev
;
5207 gen
= inode
->i_generation
;
5212 name
= (char *)arch_vma_name(vma
);
5216 if (vma
->vm_start
<= vma
->vm_mm
->start_brk
&&
5217 vma
->vm_end
>= vma
->vm_mm
->brk
) {
5221 if (vma
->vm_start
<= vma
->vm_mm
->start_stack
&&
5222 vma
->vm_end
>= vma
->vm_mm
->start_stack
) {
5232 strlcpy(tmp
, name
, sizeof(tmp
));
5236 * Since our buffer works in 8 byte units we need to align our string
5237 * size to a multiple of 8. However, we must guarantee the tail end is
5238 * zero'd out to avoid leaking random bits to userspace.
5240 size
= strlen(name
)+1;
5241 while (!IS_ALIGNED(size
, sizeof(u64
)))
5242 name
[size
++] = '\0';
5244 mmap_event
->file_name
= name
;
5245 mmap_event
->file_size
= size
;
5246 mmap_event
->maj
= maj
;
5247 mmap_event
->min
= min
;
5248 mmap_event
->ino
= ino
;
5249 mmap_event
->ino_generation
= gen
;
5251 if (!(vma
->vm_flags
& VM_EXEC
))
5252 mmap_event
->event_id
.header
.misc
|= PERF_RECORD_MISC_MMAP_DATA
;
5254 mmap_event
->event_id
.header
.size
= sizeof(mmap_event
->event_id
) + size
;
5256 perf_event_aux(perf_event_mmap_output
,
5263 void perf_event_mmap(struct vm_area_struct
*vma
)
5265 struct perf_mmap_event mmap_event
;
5267 if (!atomic_read(&nr_mmap_events
))
5270 mmap_event
= (struct perf_mmap_event
){
5276 .type
= PERF_RECORD_MMAP
,
5277 .misc
= PERF_RECORD_MISC_USER
,
5282 .start
= vma
->vm_start
,
5283 .len
= vma
->vm_end
- vma
->vm_start
,
5284 .pgoff
= (u64
)vma
->vm_pgoff
<< PAGE_SHIFT
,
5286 /* .maj (attr_mmap2 only) */
5287 /* .min (attr_mmap2 only) */
5288 /* .ino (attr_mmap2 only) */
5289 /* .ino_generation (attr_mmap2 only) */
5292 perf_event_mmap_event(&mmap_event
);
5296 * IRQ throttle logging
5299 static void perf_log_throttle(struct perf_event
*event
, int enable
)
5301 struct perf_output_handle handle
;
5302 struct perf_sample_data sample
;
5306 struct perf_event_header header
;
5310 } throttle_event
= {
5312 .type
= PERF_RECORD_THROTTLE
,
5314 .size
= sizeof(throttle_event
),
5316 .time
= perf_clock(),
5317 .id
= primary_event_id(event
),
5318 .stream_id
= event
->id
,
5322 throttle_event
.header
.type
= PERF_RECORD_UNTHROTTLE
;
5324 perf_event_header__init_id(&throttle_event
.header
, &sample
, event
);
5326 ret
= perf_output_begin(&handle
, event
,
5327 throttle_event
.header
.size
);
5331 perf_output_put(&handle
, throttle_event
);
5332 perf_event__output_id_sample(event
, &handle
, &sample
);
5333 perf_output_end(&handle
);
5337 * Generic event overflow handling, sampling.
5340 static int __perf_event_overflow(struct perf_event
*event
,
5341 int throttle
, struct perf_sample_data
*data
,
5342 struct pt_regs
*regs
)
5344 int events
= atomic_read(&event
->event_limit
);
5345 struct hw_perf_event
*hwc
= &event
->hw
;
5350 * Non-sampling counters might still use the PMI to fold short
5351 * hardware counters, ignore those.
5353 if (unlikely(!is_sampling_event(event
)))
5356 seq
= __this_cpu_read(perf_throttled_seq
);
5357 if (seq
!= hwc
->interrupts_seq
) {
5358 hwc
->interrupts_seq
= seq
;
5359 hwc
->interrupts
= 1;
5362 if (unlikely(throttle
5363 && hwc
->interrupts
>= max_samples_per_tick
)) {
5364 __this_cpu_inc(perf_throttled_count
);
5365 hwc
->interrupts
= MAX_INTERRUPTS
;
5366 perf_log_throttle(event
, 0);
5367 tick_nohz_full_kick();
5372 if (event
->attr
.freq
) {
5373 u64 now
= perf_clock();
5374 s64 delta
= now
- hwc
->freq_time_stamp
;
5376 hwc
->freq_time_stamp
= now
;
5378 if (delta
> 0 && delta
< 2*TICK_NSEC
)
5379 perf_adjust_period(event
, delta
, hwc
->last_period
, true);
5383 * XXX event_limit might not quite work as expected on inherited
5387 event
->pending_kill
= POLL_IN
;
5388 if (events
&& atomic_dec_and_test(&event
->event_limit
)) {
5390 event
->pending_kill
= POLL_HUP
;
5391 event
->pending_disable
= 1;
5392 irq_work_queue(&event
->pending
);
5395 if (event
->overflow_handler
)
5396 event
->overflow_handler(event
, data
, regs
);
5398 perf_event_output(event
, data
, regs
);
5400 if (event
->fasync
&& event
->pending_kill
) {
5401 event
->pending_wakeup
= 1;
5402 irq_work_queue(&event
->pending
);
5408 int perf_event_overflow(struct perf_event
*event
,
5409 struct perf_sample_data
*data
,
5410 struct pt_regs
*regs
)
5412 return __perf_event_overflow(event
, 1, data
, regs
);
5416 * Generic software event infrastructure
5419 struct swevent_htable
{
5420 struct swevent_hlist
*swevent_hlist
;
5421 struct mutex hlist_mutex
;
5424 /* Recursion avoidance in each contexts */
5425 int recursion
[PERF_NR_CONTEXTS
];
5428 static DEFINE_PER_CPU(struct swevent_htable
, swevent_htable
);
5431 * We directly increment event->count and keep a second value in
5432 * event->hw.period_left to count intervals. This period event
5433 * is kept in the range [-sample_period, 0] so that we can use the
5437 u64
perf_swevent_set_period(struct perf_event
*event
)
5439 struct hw_perf_event
*hwc
= &event
->hw
;
5440 u64 period
= hwc
->last_period
;
5444 hwc
->last_period
= hwc
->sample_period
;
5447 old
= val
= local64_read(&hwc
->period_left
);
5451 nr
= div64_u64(period
+ val
, period
);
5452 offset
= nr
* period
;
5454 if (local64_cmpxchg(&hwc
->period_left
, old
, val
) != old
)
5460 static void perf_swevent_overflow(struct perf_event
*event
, u64 overflow
,
5461 struct perf_sample_data
*data
,
5462 struct pt_regs
*regs
)
5464 struct hw_perf_event
*hwc
= &event
->hw
;
5468 overflow
= perf_swevent_set_period(event
);
5470 if (hwc
->interrupts
== MAX_INTERRUPTS
)
5473 for (; overflow
; overflow
--) {
5474 if (__perf_event_overflow(event
, throttle
,
5477 * We inhibit the overflow from happening when
5478 * hwc->interrupts == MAX_INTERRUPTS.
5486 static void perf_swevent_event(struct perf_event
*event
, u64 nr
,
5487 struct perf_sample_data
*data
,
5488 struct pt_regs
*regs
)
5490 struct hw_perf_event
*hwc
= &event
->hw
;
5492 local64_add(nr
, &event
->count
);
5497 if (!is_sampling_event(event
))
5500 if ((event
->attr
.sample_type
& PERF_SAMPLE_PERIOD
) && !event
->attr
.freq
) {
5502 return perf_swevent_overflow(event
, 1, data
, regs
);
5504 data
->period
= event
->hw
.last_period
;
5506 if (nr
== 1 && hwc
->sample_period
== 1 && !event
->attr
.freq
)
5507 return perf_swevent_overflow(event
, 1, data
, regs
);
5509 if (local64_add_negative(nr
, &hwc
->period_left
))
5512 perf_swevent_overflow(event
, 0, data
, regs
);
5515 static int perf_exclude_event(struct perf_event
*event
,
5516 struct pt_regs
*regs
)
5518 if (event
->hw
.state
& PERF_HES_STOPPED
)
5522 if (event
->attr
.exclude_user
&& user_mode(regs
))
5525 if (event
->attr
.exclude_kernel
&& !user_mode(regs
))
5532 static int perf_swevent_match(struct perf_event
*event
,
5533 enum perf_type_id type
,
5535 struct perf_sample_data
*data
,
5536 struct pt_regs
*regs
)
5538 if (event
->attr
.type
!= type
)
5541 if (event
->attr
.config
!= event_id
)
5544 if (perf_exclude_event(event
, regs
))
5550 static inline u64
swevent_hash(u64 type
, u32 event_id
)
5552 u64 val
= event_id
| (type
<< 32);
5554 return hash_64(val
, SWEVENT_HLIST_BITS
);
5557 static inline struct hlist_head
*
5558 __find_swevent_head(struct swevent_hlist
*hlist
, u64 type
, u32 event_id
)
5560 u64 hash
= swevent_hash(type
, event_id
);
5562 return &hlist
->heads
[hash
];
5565 /* For the read side: events when they trigger */
5566 static inline struct hlist_head
*
5567 find_swevent_head_rcu(struct swevent_htable
*swhash
, u64 type
, u32 event_id
)
5569 struct swevent_hlist
*hlist
;
5571 hlist
= rcu_dereference(swhash
->swevent_hlist
);
5575 return __find_swevent_head(hlist
, type
, event_id
);
5578 /* For the event head insertion and removal in the hlist */
5579 static inline struct hlist_head
*
5580 find_swevent_head(struct swevent_htable
*swhash
, struct perf_event
*event
)
5582 struct swevent_hlist
*hlist
;
5583 u32 event_id
= event
->attr
.config
;
5584 u64 type
= event
->attr
.type
;
5587 * Event scheduling is always serialized against hlist allocation
5588 * and release. Which makes the protected version suitable here.
5589 * The context lock guarantees that.
5591 hlist
= rcu_dereference_protected(swhash
->swevent_hlist
,
5592 lockdep_is_held(&event
->ctx
->lock
));
5596 return __find_swevent_head(hlist
, type
, event_id
);
5599 static void do_perf_sw_event(enum perf_type_id type
, u32 event_id
,
5601 struct perf_sample_data
*data
,
5602 struct pt_regs
*regs
)
5604 struct swevent_htable
*swhash
= &__get_cpu_var(swevent_htable
);
5605 struct perf_event
*event
;
5606 struct hlist_head
*head
;
5609 head
= find_swevent_head_rcu(swhash
, type
, event_id
);
5613 hlist_for_each_entry_rcu(event
, head
, hlist_entry
) {
5614 if (perf_swevent_match(event
, type
, event_id
, data
, regs
))
5615 perf_swevent_event(event
, nr
, data
, regs
);
5621 int perf_swevent_get_recursion_context(void)
5623 struct swevent_htable
*swhash
= &__get_cpu_var(swevent_htable
);
5625 return get_recursion_context(swhash
->recursion
);
5627 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context
);
5629 inline void perf_swevent_put_recursion_context(int rctx
)
5631 struct swevent_htable
*swhash
= &__get_cpu_var(swevent_htable
);
5633 put_recursion_context(swhash
->recursion
, rctx
);
5636 void __perf_sw_event(u32 event_id
, u64 nr
, struct pt_regs
*regs
, u64 addr
)
5638 struct perf_sample_data data
;
5641 preempt_disable_notrace();
5642 rctx
= perf_swevent_get_recursion_context();
5646 perf_sample_data_init(&data
, addr
, 0);
5648 do_perf_sw_event(PERF_TYPE_SOFTWARE
, event_id
, nr
, &data
, regs
);
5650 perf_swevent_put_recursion_context(rctx
);
5651 preempt_enable_notrace();
5654 static void perf_swevent_read(struct perf_event
*event
)
5658 static int perf_swevent_add(struct perf_event
*event
, int flags
)
5660 struct swevent_htable
*swhash
= &__get_cpu_var(swevent_htable
);
5661 struct hw_perf_event
*hwc
= &event
->hw
;
5662 struct hlist_head
*head
;
5664 if (is_sampling_event(event
)) {
5665 hwc
->last_period
= hwc
->sample_period
;
5666 perf_swevent_set_period(event
);
5669 hwc
->state
= !(flags
& PERF_EF_START
);
5671 head
= find_swevent_head(swhash
, event
);
5672 if (WARN_ON_ONCE(!head
))
5675 hlist_add_head_rcu(&event
->hlist_entry
, head
);
5680 static void perf_swevent_del(struct perf_event
*event
, int flags
)
5682 hlist_del_rcu(&event
->hlist_entry
);
5685 static void perf_swevent_start(struct perf_event
*event
, int flags
)
5687 event
->hw
.state
= 0;
5690 static void perf_swevent_stop(struct perf_event
*event
, int flags
)
5692 event
->hw
.state
= PERF_HES_STOPPED
;
5695 /* Deref the hlist from the update side */
5696 static inline struct swevent_hlist
*
5697 swevent_hlist_deref(struct swevent_htable
*swhash
)
5699 return rcu_dereference_protected(swhash
->swevent_hlist
,
5700 lockdep_is_held(&swhash
->hlist_mutex
));
5703 static void swevent_hlist_release(struct swevent_htable
*swhash
)
5705 struct swevent_hlist
*hlist
= swevent_hlist_deref(swhash
);
5710 rcu_assign_pointer(swhash
->swevent_hlist
, NULL
);
5711 kfree_rcu(hlist
, rcu_head
);
5714 static void swevent_hlist_put_cpu(struct perf_event
*event
, int cpu
)
5716 struct swevent_htable
*swhash
= &per_cpu(swevent_htable
, cpu
);
5718 mutex_lock(&swhash
->hlist_mutex
);
5720 if (!--swhash
->hlist_refcount
)
5721 swevent_hlist_release(swhash
);
5723 mutex_unlock(&swhash
->hlist_mutex
);
5726 static void swevent_hlist_put(struct perf_event
*event
)
5730 for_each_possible_cpu(cpu
)
5731 swevent_hlist_put_cpu(event
, cpu
);
5734 static int swevent_hlist_get_cpu(struct perf_event
*event
, int cpu
)
5736 struct swevent_htable
*swhash
= &per_cpu(swevent_htable
, cpu
);
5739 mutex_lock(&swhash
->hlist_mutex
);
5741 if (!swevent_hlist_deref(swhash
) && cpu_online(cpu
)) {
5742 struct swevent_hlist
*hlist
;
5744 hlist
= kzalloc(sizeof(*hlist
), GFP_KERNEL
);
5749 rcu_assign_pointer(swhash
->swevent_hlist
, hlist
);
5751 swhash
->hlist_refcount
++;
5753 mutex_unlock(&swhash
->hlist_mutex
);
5758 static int swevent_hlist_get(struct perf_event
*event
)
5761 int cpu
, failed_cpu
;
5764 for_each_possible_cpu(cpu
) {
5765 err
= swevent_hlist_get_cpu(event
, cpu
);
5775 for_each_possible_cpu(cpu
) {
5776 if (cpu
== failed_cpu
)
5778 swevent_hlist_put_cpu(event
, cpu
);
5785 struct static_key perf_swevent_enabled
[PERF_COUNT_SW_MAX
];
5787 static void sw_perf_event_destroy(struct perf_event
*event
)
5789 u64 event_id
= event
->attr
.config
;
5791 WARN_ON(event
->parent
);
5793 static_key_slow_dec(&perf_swevent_enabled
[event_id
]);
5794 swevent_hlist_put(event
);
5797 static int perf_swevent_init(struct perf_event
*event
)
5799 u64 event_id
= event
->attr
.config
;
5801 if (event
->attr
.type
!= PERF_TYPE_SOFTWARE
)
5805 * no branch sampling for software events
5807 if (has_branch_stack(event
))
5811 case PERF_COUNT_SW_CPU_CLOCK
:
5812 case PERF_COUNT_SW_TASK_CLOCK
:
5819 if (event_id
>= PERF_COUNT_SW_MAX
)
5822 if (!event
->parent
) {
5825 err
= swevent_hlist_get(event
);
5829 static_key_slow_inc(&perf_swevent_enabled
[event_id
]);
5830 event
->destroy
= sw_perf_event_destroy
;
5836 static int perf_swevent_event_idx(struct perf_event
*event
)
5841 static struct pmu perf_swevent
= {
5842 .task_ctx_nr
= perf_sw_context
,
5844 .event_init
= perf_swevent_init
,
5845 .add
= perf_swevent_add
,
5846 .del
= perf_swevent_del
,
5847 .start
= perf_swevent_start
,
5848 .stop
= perf_swevent_stop
,
5849 .read
= perf_swevent_read
,
5851 .event_idx
= perf_swevent_event_idx
,
5854 #ifdef CONFIG_EVENT_TRACING
5856 static int perf_tp_filter_match(struct perf_event
*event
,
5857 struct perf_sample_data
*data
)
5859 void *record
= data
->raw
->data
;
5861 if (likely(!event
->filter
) || filter_match_preds(event
->filter
, record
))
5866 static int perf_tp_event_match(struct perf_event
*event
,
5867 struct perf_sample_data
*data
,
5868 struct pt_regs
*regs
)
5870 if (event
->hw
.state
& PERF_HES_STOPPED
)
5873 * All tracepoints are from kernel-space.
5875 if (event
->attr
.exclude_kernel
)
5878 if (!perf_tp_filter_match(event
, data
))
5884 void perf_tp_event(u64 addr
, u64 count
, void *record
, int entry_size
,
5885 struct pt_regs
*regs
, struct hlist_head
*head
, int rctx
,
5886 struct task_struct
*task
)
5888 struct perf_sample_data data
;
5889 struct perf_event
*event
;
5891 struct perf_raw_record raw
= {
5896 perf_sample_data_init(&data
, addr
, 0);
5899 hlist_for_each_entry_rcu(event
, head
, hlist_entry
) {
5900 if (perf_tp_event_match(event
, &data
, regs
))
5901 perf_swevent_event(event
, count
, &data
, regs
);
5905 * If we got specified a target task, also iterate its context and
5906 * deliver this event there too.
5908 if (task
&& task
!= current
) {
5909 struct perf_event_context
*ctx
;
5910 struct trace_entry
*entry
= record
;
5913 ctx
= rcu_dereference(task
->perf_event_ctxp
[perf_sw_context
]);
5917 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
5918 if (event
->attr
.type
!= PERF_TYPE_TRACEPOINT
)
5920 if (event
->attr
.config
!= entry
->type
)
5922 if (perf_tp_event_match(event
, &data
, regs
))
5923 perf_swevent_event(event
, count
, &data
, regs
);
5929 perf_swevent_put_recursion_context(rctx
);
5931 EXPORT_SYMBOL_GPL(perf_tp_event
);
5933 static void tp_perf_event_destroy(struct perf_event
*event
)
5935 perf_trace_destroy(event
);
5938 static int perf_tp_event_init(struct perf_event
*event
)
5942 if (event
->attr
.type
!= PERF_TYPE_TRACEPOINT
)
5946 * no branch sampling for tracepoint events
5948 if (has_branch_stack(event
))
5951 err
= perf_trace_init(event
);
5955 event
->destroy
= tp_perf_event_destroy
;
5960 static struct pmu perf_tracepoint
= {
5961 .task_ctx_nr
= perf_sw_context
,
5963 .event_init
= perf_tp_event_init
,
5964 .add
= perf_trace_add
,
5965 .del
= perf_trace_del
,
5966 .start
= perf_swevent_start
,
5967 .stop
= perf_swevent_stop
,
5968 .read
= perf_swevent_read
,
5970 .event_idx
= perf_swevent_event_idx
,
5973 static inline void perf_tp_register(void)
5975 perf_pmu_register(&perf_tracepoint
, "tracepoint", PERF_TYPE_TRACEPOINT
);
5978 static int perf_event_set_filter(struct perf_event
*event
, void __user
*arg
)
5983 if (event
->attr
.type
!= PERF_TYPE_TRACEPOINT
)
5986 filter_str
= strndup_user(arg
, PAGE_SIZE
);
5987 if (IS_ERR(filter_str
))
5988 return PTR_ERR(filter_str
);
5990 ret
= ftrace_profile_set_filter(event
, event
->attr
.config
, filter_str
);
5996 static void perf_event_free_filter(struct perf_event
*event
)
5998 ftrace_profile_free_filter(event
);
6003 static inline void perf_tp_register(void)
6007 static int perf_event_set_filter(struct perf_event
*event
, void __user
*arg
)
6012 static void perf_event_free_filter(struct perf_event
*event
)
6016 #endif /* CONFIG_EVENT_TRACING */
6018 #ifdef CONFIG_HAVE_HW_BREAKPOINT
6019 void perf_bp_event(struct perf_event
*bp
, void *data
)
6021 struct perf_sample_data sample
;
6022 struct pt_regs
*regs
= data
;
6024 perf_sample_data_init(&sample
, bp
->attr
.bp_addr
, 0);
6026 if (!bp
->hw
.state
&& !perf_exclude_event(bp
, regs
))
6027 perf_swevent_event(bp
, 1, &sample
, regs
);
6032 * hrtimer based swevent callback
6035 static enum hrtimer_restart
perf_swevent_hrtimer(struct hrtimer
*hrtimer
)
6037 enum hrtimer_restart ret
= HRTIMER_RESTART
;
6038 struct perf_sample_data data
;
6039 struct pt_regs
*regs
;
6040 struct perf_event
*event
;
6043 event
= container_of(hrtimer
, struct perf_event
, hw
.hrtimer
);
6045 if (event
->state
!= PERF_EVENT_STATE_ACTIVE
)
6046 return HRTIMER_NORESTART
;
6048 event
->pmu
->read(event
);
6050 perf_sample_data_init(&data
, 0, event
->hw
.last_period
);
6051 regs
= get_irq_regs();
6053 if (regs
&& !perf_exclude_event(event
, regs
)) {
6054 if (!(event
->attr
.exclude_idle
&& is_idle_task(current
)))
6055 if (__perf_event_overflow(event
, 1, &data
, regs
))
6056 ret
= HRTIMER_NORESTART
;
6059 period
= max_t(u64
, 10000, event
->hw
.sample_period
);
6060 hrtimer_forward_now(hrtimer
, ns_to_ktime(period
));
6065 static void perf_swevent_start_hrtimer(struct perf_event
*event
)
6067 struct hw_perf_event
*hwc
= &event
->hw
;
6070 if (!is_sampling_event(event
))
6073 period
= local64_read(&hwc
->period_left
);
6078 local64_set(&hwc
->period_left
, 0);
6080 period
= max_t(u64
, 10000, hwc
->sample_period
);
6082 __hrtimer_start_range_ns(&hwc
->hrtimer
,
6083 ns_to_ktime(period
), 0,
6084 HRTIMER_MODE_REL_PINNED
, 0);
6087 static void perf_swevent_cancel_hrtimer(struct perf_event
*event
)
6089 struct hw_perf_event
*hwc
= &event
->hw
;
6091 if (is_sampling_event(event
)) {
6092 ktime_t remaining
= hrtimer_get_remaining(&hwc
->hrtimer
);
6093 local64_set(&hwc
->period_left
, ktime_to_ns(remaining
));
6095 hrtimer_cancel(&hwc
->hrtimer
);
6099 static void perf_swevent_init_hrtimer(struct perf_event
*event
)
6101 struct hw_perf_event
*hwc
= &event
->hw
;
6103 if (!is_sampling_event(event
))
6106 hrtimer_init(&hwc
->hrtimer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
6107 hwc
->hrtimer
.function
= perf_swevent_hrtimer
;
6110 * Since hrtimers have a fixed rate, we can do a static freq->period
6111 * mapping and avoid the whole period adjust feedback stuff.
6113 if (event
->attr
.freq
) {
6114 long freq
= event
->attr
.sample_freq
;
6116 event
->attr
.sample_period
= NSEC_PER_SEC
/ freq
;
6117 hwc
->sample_period
= event
->attr
.sample_period
;
6118 local64_set(&hwc
->period_left
, hwc
->sample_period
);
6119 hwc
->last_period
= hwc
->sample_period
;
6120 event
->attr
.freq
= 0;
6125 * Software event: cpu wall time clock
6128 static void cpu_clock_event_update(struct perf_event
*event
)
6133 now
= local_clock();
6134 prev
= local64_xchg(&event
->hw
.prev_count
, now
);
6135 local64_add(now
- prev
, &event
->count
);
6138 static void cpu_clock_event_start(struct perf_event
*event
, int flags
)
6140 local64_set(&event
->hw
.prev_count
, local_clock());
6141 perf_swevent_start_hrtimer(event
);
6144 static void cpu_clock_event_stop(struct perf_event
*event
, int flags
)
6146 perf_swevent_cancel_hrtimer(event
);
6147 cpu_clock_event_update(event
);
6150 static int cpu_clock_event_add(struct perf_event
*event
, int flags
)
6152 if (flags
& PERF_EF_START
)
6153 cpu_clock_event_start(event
, flags
);
6158 static void cpu_clock_event_del(struct perf_event
*event
, int flags
)
6160 cpu_clock_event_stop(event
, flags
);
6163 static void cpu_clock_event_read(struct perf_event
*event
)
6165 cpu_clock_event_update(event
);
6168 static int cpu_clock_event_init(struct perf_event
*event
)
6170 if (event
->attr
.type
!= PERF_TYPE_SOFTWARE
)
6173 if (event
->attr
.config
!= PERF_COUNT_SW_CPU_CLOCK
)
6177 * no branch sampling for software events
6179 if (has_branch_stack(event
))
6182 perf_swevent_init_hrtimer(event
);
6187 static struct pmu perf_cpu_clock
= {
6188 .task_ctx_nr
= perf_sw_context
,
6190 .event_init
= cpu_clock_event_init
,
6191 .add
= cpu_clock_event_add
,
6192 .del
= cpu_clock_event_del
,
6193 .start
= cpu_clock_event_start
,
6194 .stop
= cpu_clock_event_stop
,
6195 .read
= cpu_clock_event_read
,
6197 .event_idx
= perf_swevent_event_idx
,
6201 * Software event: task time clock
6204 static void task_clock_event_update(struct perf_event
*event
, u64 now
)
6209 prev
= local64_xchg(&event
->hw
.prev_count
, now
);
6211 local64_add(delta
, &event
->count
);
6214 static void task_clock_event_start(struct perf_event
*event
, int flags
)
6216 local64_set(&event
->hw
.prev_count
, event
->ctx
->time
);
6217 perf_swevent_start_hrtimer(event
);
6220 static void task_clock_event_stop(struct perf_event
*event
, int flags
)
6222 perf_swevent_cancel_hrtimer(event
);
6223 task_clock_event_update(event
, event
->ctx
->time
);
6226 static int task_clock_event_add(struct perf_event
*event
, int flags
)
6228 if (flags
& PERF_EF_START
)
6229 task_clock_event_start(event
, flags
);
6234 static void task_clock_event_del(struct perf_event
*event
, int flags
)
6236 task_clock_event_stop(event
, PERF_EF_UPDATE
);
6239 static void task_clock_event_read(struct perf_event
*event
)
6241 u64 now
= perf_clock();
6242 u64 delta
= now
- event
->ctx
->timestamp
;
6243 u64 time
= event
->ctx
->time
+ delta
;
6245 task_clock_event_update(event
, time
);
6248 static int task_clock_event_init(struct perf_event
*event
)
6250 if (event
->attr
.type
!= PERF_TYPE_SOFTWARE
)
6253 if (event
->attr
.config
!= PERF_COUNT_SW_TASK_CLOCK
)
6257 * no branch sampling for software events
6259 if (has_branch_stack(event
))
6262 perf_swevent_init_hrtimer(event
);
6267 static struct pmu perf_task_clock
= {
6268 .task_ctx_nr
= perf_sw_context
,
6270 .event_init
= task_clock_event_init
,
6271 .add
= task_clock_event_add
,
6272 .del
= task_clock_event_del
,
6273 .start
= task_clock_event_start
,
6274 .stop
= task_clock_event_stop
,
6275 .read
= task_clock_event_read
,
6277 .event_idx
= perf_swevent_event_idx
,
6280 static void perf_pmu_nop_void(struct pmu
*pmu
)
6284 static int perf_pmu_nop_int(struct pmu
*pmu
)
6289 static void perf_pmu_start_txn(struct pmu
*pmu
)
6291 perf_pmu_disable(pmu
);
6294 static int perf_pmu_commit_txn(struct pmu
*pmu
)
6296 perf_pmu_enable(pmu
);
6300 static void perf_pmu_cancel_txn(struct pmu
*pmu
)
6302 perf_pmu_enable(pmu
);
6305 static int perf_event_idx_default(struct perf_event
*event
)
6307 return event
->hw
.idx
+ 1;
6311 * Ensures all contexts with the same task_ctx_nr have the same
6312 * pmu_cpu_context too.
6314 static struct perf_cpu_context __percpu
*find_pmu_context(int ctxn
)
6321 list_for_each_entry(pmu
, &pmus
, entry
) {
6322 if (pmu
->task_ctx_nr
== ctxn
)
6323 return pmu
->pmu_cpu_context
;
6329 static void update_pmu_context(struct pmu
*pmu
, struct pmu
*old_pmu
)
6333 for_each_possible_cpu(cpu
) {
6334 struct perf_cpu_context
*cpuctx
;
6336 cpuctx
= per_cpu_ptr(pmu
->pmu_cpu_context
, cpu
);
6338 if (cpuctx
->unique_pmu
== old_pmu
)
6339 cpuctx
->unique_pmu
= pmu
;
6343 static void free_pmu_context(struct pmu
*pmu
)
6347 mutex_lock(&pmus_lock
);
6349 * Like a real lame refcount.
6351 list_for_each_entry(i
, &pmus
, entry
) {
6352 if (i
->pmu_cpu_context
== pmu
->pmu_cpu_context
) {
6353 update_pmu_context(i
, pmu
);
6358 free_percpu(pmu
->pmu_cpu_context
);
6360 mutex_unlock(&pmus_lock
);
6362 static struct idr pmu_idr
;
6365 type_show(struct device
*dev
, struct device_attribute
*attr
, char *page
)
6367 struct pmu
*pmu
= dev_get_drvdata(dev
);
6369 return snprintf(page
, PAGE_SIZE
-1, "%d\n", pmu
->type
);
6371 static DEVICE_ATTR_RO(type
);
6374 perf_event_mux_interval_ms_show(struct device
*dev
,
6375 struct device_attribute
*attr
,
6378 struct pmu
*pmu
= dev_get_drvdata(dev
);
6380 return snprintf(page
, PAGE_SIZE
-1, "%d\n", pmu
->hrtimer_interval_ms
);
6384 perf_event_mux_interval_ms_store(struct device
*dev
,
6385 struct device_attribute
*attr
,
6386 const char *buf
, size_t count
)
6388 struct pmu
*pmu
= dev_get_drvdata(dev
);
6389 int timer
, cpu
, ret
;
6391 ret
= kstrtoint(buf
, 0, &timer
);
6398 /* same value, noting to do */
6399 if (timer
== pmu
->hrtimer_interval_ms
)
6402 pmu
->hrtimer_interval_ms
= timer
;
6404 /* update all cpuctx for this PMU */
6405 for_each_possible_cpu(cpu
) {
6406 struct perf_cpu_context
*cpuctx
;
6407 cpuctx
= per_cpu_ptr(pmu
->pmu_cpu_context
, cpu
);
6408 cpuctx
->hrtimer_interval
= ns_to_ktime(NSEC_PER_MSEC
* timer
);
6410 if (hrtimer_active(&cpuctx
->hrtimer
))
6411 hrtimer_forward_now(&cpuctx
->hrtimer
, cpuctx
->hrtimer_interval
);
6416 static DEVICE_ATTR_RW(perf_event_mux_interval_ms
);
6418 static struct attribute
*pmu_dev_attrs
[] = {
6419 &dev_attr_type
.attr
,
6420 &dev_attr_perf_event_mux_interval_ms
.attr
,
6423 ATTRIBUTE_GROUPS(pmu_dev
);
6425 static int pmu_bus_running
;
6426 static struct bus_type pmu_bus
= {
6427 .name
= "event_source",
6428 .dev_groups
= pmu_dev_groups
,
6431 static void pmu_dev_release(struct device
*dev
)
6436 static int pmu_dev_alloc(struct pmu
*pmu
)
6440 pmu
->dev
= kzalloc(sizeof(struct device
), GFP_KERNEL
);
6444 pmu
->dev
->groups
= pmu
->attr_groups
;
6445 device_initialize(pmu
->dev
);
6446 ret
= dev_set_name(pmu
->dev
, "%s", pmu
->name
);
6450 dev_set_drvdata(pmu
->dev
, pmu
);
6451 pmu
->dev
->bus
= &pmu_bus
;
6452 pmu
->dev
->release
= pmu_dev_release
;
6453 ret
= device_add(pmu
->dev
);
6461 put_device(pmu
->dev
);
6465 static struct lock_class_key cpuctx_mutex
;
6466 static struct lock_class_key cpuctx_lock
;
6468 int perf_pmu_register(struct pmu
*pmu
, const char *name
, int type
)
6472 mutex_lock(&pmus_lock
);
6474 pmu
->pmu_disable_count
= alloc_percpu(int);
6475 if (!pmu
->pmu_disable_count
)
6484 type
= idr_alloc(&pmu_idr
, pmu
, PERF_TYPE_MAX
, 0, GFP_KERNEL
);
6492 if (pmu_bus_running
) {
6493 ret
= pmu_dev_alloc(pmu
);
6499 pmu
->pmu_cpu_context
= find_pmu_context(pmu
->task_ctx_nr
);
6500 if (pmu
->pmu_cpu_context
)
6501 goto got_cpu_context
;
6504 pmu
->pmu_cpu_context
= alloc_percpu(struct perf_cpu_context
);
6505 if (!pmu
->pmu_cpu_context
)
6508 for_each_possible_cpu(cpu
) {
6509 struct perf_cpu_context
*cpuctx
;
6511 cpuctx
= per_cpu_ptr(pmu
->pmu_cpu_context
, cpu
);
6512 __perf_event_init_context(&cpuctx
->ctx
);
6513 lockdep_set_class(&cpuctx
->ctx
.mutex
, &cpuctx_mutex
);
6514 lockdep_set_class(&cpuctx
->ctx
.lock
, &cpuctx_lock
);
6515 cpuctx
->ctx
.type
= cpu_context
;
6516 cpuctx
->ctx
.pmu
= pmu
;
6518 __perf_cpu_hrtimer_init(cpuctx
, cpu
);
6520 INIT_LIST_HEAD(&cpuctx
->rotation_list
);
6521 cpuctx
->unique_pmu
= pmu
;
6525 if (!pmu
->start_txn
) {
6526 if (pmu
->pmu_enable
) {
6528 * If we have pmu_enable/pmu_disable calls, install
6529 * transaction stubs that use that to try and batch
6530 * hardware accesses.
6532 pmu
->start_txn
= perf_pmu_start_txn
;
6533 pmu
->commit_txn
= perf_pmu_commit_txn
;
6534 pmu
->cancel_txn
= perf_pmu_cancel_txn
;
6536 pmu
->start_txn
= perf_pmu_nop_void
;
6537 pmu
->commit_txn
= perf_pmu_nop_int
;
6538 pmu
->cancel_txn
= perf_pmu_nop_void
;
6542 if (!pmu
->pmu_enable
) {
6543 pmu
->pmu_enable
= perf_pmu_nop_void
;
6544 pmu
->pmu_disable
= perf_pmu_nop_void
;
6547 if (!pmu
->event_idx
)
6548 pmu
->event_idx
= perf_event_idx_default
;
6550 list_add_rcu(&pmu
->entry
, &pmus
);
6553 mutex_unlock(&pmus_lock
);
6558 device_del(pmu
->dev
);
6559 put_device(pmu
->dev
);
6562 if (pmu
->type
>= PERF_TYPE_MAX
)
6563 idr_remove(&pmu_idr
, pmu
->type
);
6566 free_percpu(pmu
->pmu_disable_count
);
6570 void perf_pmu_unregister(struct pmu
*pmu
)
6572 mutex_lock(&pmus_lock
);
6573 list_del_rcu(&pmu
->entry
);
6574 mutex_unlock(&pmus_lock
);
6577 * We dereference the pmu list under both SRCU and regular RCU, so
6578 * synchronize against both of those.
6580 synchronize_srcu(&pmus_srcu
);
6583 free_percpu(pmu
->pmu_disable_count
);
6584 if (pmu
->type
>= PERF_TYPE_MAX
)
6585 idr_remove(&pmu_idr
, pmu
->type
);
6586 device_del(pmu
->dev
);
6587 put_device(pmu
->dev
);
6588 free_pmu_context(pmu
);
6591 struct pmu
*perf_init_event(struct perf_event
*event
)
6593 struct pmu
*pmu
= NULL
;
6597 idx
= srcu_read_lock(&pmus_srcu
);
6600 pmu
= idr_find(&pmu_idr
, event
->attr
.type
);
6604 ret
= pmu
->event_init(event
);
6610 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
6612 ret
= pmu
->event_init(event
);
6616 if (ret
!= -ENOENT
) {
6621 pmu
= ERR_PTR(-ENOENT
);
6623 srcu_read_unlock(&pmus_srcu
, idx
);
6628 static void account_event_cpu(struct perf_event
*event
, int cpu
)
6633 if (has_branch_stack(event
)) {
6634 if (!(event
->attach_state
& PERF_ATTACH_TASK
))
6635 atomic_inc(&per_cpu(perf_branch_stack_events
, cpu
));
6637 if (is_cgroup_event(event
))
6638 atomic_inc(&per_cpu(perf_cgroup_events
, cpu
));
6641 static void account_event(struct perf_event
*event
)
6646 if (event
->attach_state
& PERF_ATTACH_TASK
)
6647 static_key_slow_inc(&perf_sched_events
.key
);
6648 if (event
->attr
.mmap
|| event
->attr
.mmap_data
)
6649 atomic_inc(&nr_mmap_events
);
6650 if (event
->attr
.comm
)
6651 atomic_inc(&nr_comm_events
);
6652 if (event
->attr
.task
)
6653 atomic_inc(&nr_task_events
);
6654 if (event
->attr
.freq
) {
6655 if (atomic_inc_return(&nr_freq_events
) == 1)
6656 tick_nohz_full_kick_all();
6658 if (has_branch_stack(event
))
6659 static_key_slow_inc(&perf_sched_events
.key
);
6660 if (is_cgroup_event(event
))
6661 static_key_slow_inc(&perf_sched_events
.key
);
6663 account_event_cpu(event
, event
->cpu
);
6667 * Allocate and initialize a event structure
6669 static struct perf_event
*
6670 perf_event_alloc(struct perf_event_attr
*attr
, int cpu
,
6671 struct task_struct
*task
,
6672 struct perf_event
*group_leader
,
6673 struct perf_event
*parent_event
,
6674 perf_overflow_handler_t overflow_handler
,
6678 struct perf_event
*event
;
6679 struct hw_perf_event
*hwc
;
6682 if ((unsigned)cpu
>= nr_cpu_ids
) {
6683 if (!task
|| cpu
!= -1)
6684 return ERR_PTR(-EINVAL
);
6687 event
= kzalloc(sizeof(*event
), GFP_KERNEL
);
6689 return ERR_PTR(-ENOMEM
);
6692 * Single events are their own group leaders, with an
6693 * empty sibling list:
6696 group_leader
= event
;
6698 mutex_init(&event
->child_mutex
);
6699 INIT_LIST_HEAD(&event
->child_list
);
6701 INIT_LIST_HEAD(&event
->group_entry
);
6702 INIT_LIST_HEAD(&event
->event_entry
);
6703 INIT_LIST_HEAD(&event
->sibling_list
);
6704 INIT_LIST_HEAD(&event
->rb_entry
);
6705 INIT_LIST_HEAD(&event
->active_entry
);
6706 INIT_HLIST_NODE(&event
->hlist_entry
);
6709 init_waitqueue_head(&event
->waitq
);
6710 init_irq_work(&event
->pending
, perf_pending_event
);
6712 mutex_init(&event
->mmap_mutex
);
6714 atomic_long_set(&event
->refcount
, 1);
6716 event
->attr
= *attr
;
6717 event
->group_leader
= group_leader
;
6721 event
->parent
= parent_event
;
6723 event
->ns
= get_pid_ns(task_active_pid_ns(current
));
6724 event
->id
= atomic64_inc_return(&perf_event_id
);
6726 event
->state
= PERF_EVENT_STATE_INACTIVE
;
6729 event
->attach_state
= PERF_ATTACH_TASK
;
6731 if (attr
->type
== PERF_TYPE_TRACEPOINT
)
6732 event
->hw
.tp_target
= task
;
6733 #ifdef CONFIG_HAVE_HW_BREAKPOINT
6735 * hw_breakpoint is a bit difficult here..
6737 else if (attr
->type
== PERF_TYPE_BREAKPOINT
)
6738 event
->hw
.bp_target
= task
;
6742 if (!overflow_handler
&& parent_event
) {
6743 overflow_handler
= parent_event
->overflow_handler
;
6744 context
= parent_event
->overflow_handler_context
;
6747 event
->overflow_handler
= overflow_handler
;
6748 event
->overflow_handler_context
= context
;
6750 perf_event__state_init(event
);
6755 hwc
->sample_period
= attr
->sample_period
;
6756 if (attr
->freq
&& attr
->sample_freq
)
6757 hwc
->sample_period
= 1;
6758 hwc
->last_period
= hwc
->sample_period
;
6760 local64_set(&hwc
->period_left
, hwc
->sample_period
);
6763 * we currently do not support PERF_FORMAT_GROUP on inherited events
6765 if (attr
->inherit
&& (attr
->read_format
& PERF_FORMAT_GROUP
))
6768 pmu
= perf_init_event(event
);
6771 else if (IS_ERR(pmu
)) {
6776 if (!event
->parent
) {
6777 if (event
->attr
.sample_type
& PERF_SAMPLE_CALLCHAIN
) {
6778 err
= get_callchain_buffers();
6788 event
->destroy(event
);
6791 put_pid_ns(event
->ns
);
6794 return ERR_PTR(err
);
6797 static int perf_copy_attr(struct perf_event_attr __user
*uattr
,
6798 struct perf_event_attr
*attr
)
6803 if (!access_ok(VERIFY_WRITE
, uattr
, PERF_ATTR_SIZE_VER0
))
6807 * zero the full structure, so that a short copy will be nice.
6809 memset(attr
, 0, sizeof(*attr
));
6811 ret
= get_user(size
, &uattr
->size
);
6815 if (size
> PAGE_SIZE
) /* silly large */
6818 if (!size
) /* abi compat */
6819 size
= PERF_ATTR_SIZE_VER0
;
6821 if (size
< PERF_ATTR_SIZE_VER0
)
6825 * If we're handed a bigger struct than we know of,
6826 * ensure all the unknown bits are 0 - i.e. new
6827 * user-space does not rely on any kernel feature
6828 * extensions we dont know about yet.
6830 if (size
> sizeof(*attr
)) {
6831 unsigned char __user
*addr
;
6832 unsigned char __user
*end
;
6835 addr
= (void __user
*)uattr
+ sizeof(*attr
);
6836 end
= (void __user
*)uattr
+ size
;
6838 for (; addr
< end
; addr
++) {
6839 ret
= get_user(val
, addr
);
6845 size
= sizeof(*attr
);
6848 ret
= copy_from_user(attr
, uattr
, size
);
6852 /* disabled for now */
6856 if (attr
->__reserved_1
)
6859 if (attr
->sample_type
& ~(PERF_SAMPLE_MAX
-1))
6862 if (attr
->read_format
& ~(PERF_FORMAT_MAX
-1))
6865 if (attr
->sample_type
& PERF_SAMPLE_BRANCH_STACK
) {
6866 u64 mask
= attr
->branch_sample_type
;
6868 /* only using defined bits */
6869 if (mask
& ~(PERF_SAMPLE_BRANCH_MAX
-1))
6872 /* at least one branch bit must be set */
6873 if (!(mask
& ~PERF_SAMPLE_BRANCH_PLM_ALL
))
6876 /* propagate priv level, when not set for branch */
6877 if (!(mask
& PERF_SAMPLE_BRANCH_PLM_ALL
)) {
6879 /* exclude_kernel checked on syscall entry */
6880 if (!attr
->exclude_kernel
)
6881 mask
|= PERF_SAMPLE_BRANCH_KERNEL
;
6883 if (!attr
->exclude_user
)
6884 mask
|= PERF_SAMPLE_BRANCH_USER
;
6886 if (!attr
->exclude_hv
)
6887 mask
|= PERF_SAMPLE_BRANCH_HV
;
6889 * adjust user setting (for HW filter setup)
6891 attr
->branch_sample_type
= mask
;
6893 /* privileged levels capture (kernel, hv): check permissions */
6894 if ((mask
& PERF_SAMPLE_BRANCH_PERM_PLM
)
6895 && perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN
))
6899 if (attr
->sample_type
& PERF_SAMPLE_REGS_USER
) {
6900 ret
= perf_reg_validate(attr
->sample_regs_user
);
6905 if (attr
->sample_type
& PERF_SAMPLE_STACK_USER
) {
6906 if (!arch_perf_have_user_stack_dump())
6910 * We have __u32 type for the size, but so far
6911 * we can only use __u16 as maximum due to the
6912 * __u16 sample size limit.
6914 if (attr
->sample_stack_user
>= USHRT_MAX
)
6916 else if (!IS_ALIGNED(attr
->sample_stack_user
, sizeof(u64
)))
6924 put_user(sizeof(*attr
), &uattr
->size
);
6930 perf_event_set_output(struct perf_event
*event
, struct perf_event
*output_event
)
6932 struct ring_buffer
*rb
= NULL
, *old_rb
= NULL
;
6938 /* don't allow circular references */
6939 if (event
== output_event
)
6943 * Don't allow cross-cpu buffers
6945 if (output_event
->cpu
!= event
->cpu
)
6949 * If its not a per-cpu rb, it must be the same task.
6951 if (output_event
->cpu
== -1 && output_event
->ctx
!= event
->ctx
)
6955 mutex_lock(&event
->mmap_mutex
);
6956 /* Can't redirect output if we've got an active mmap() */
6957 if (atomic_read(&event
->mmap_count
))
6963 /* get the rb we want to redirect to */
6964 rb
= ring_buffer_get(output_event
);
6970 ring_buffer_detach(event
, old_rb
);
6973 ring_buffer_attach(event
, rb
);
6975 rcu_assign_pointer(event
->rb
, rb
);
6978 ring_buffer_put(old_rb
);
6980 * Since we detached before setting the new rb, so that we
6981 * could attach the new rb, we could have missed a wakeup.
6984 wake_up_all(&event
->waitq
);
6989 mutex_unlock(&event
->mmap_mutex
);
6996 * sys_perf_event_open - open a performance event, associate it to a task/cpu
6998 * @attr_uptr: event_id type attributes for monitoring/sampling
7001 * @group_fd: group leader event fd
7003 SYSCALL_DEFINE5(perf_event_open
,
7004 struct perf_event_attr __user
*, attr_uptr
,
7005 pid_t
, pid
, int, cpu
, int, group_fd
, unsigned long, flags
)
7007 struct perf_event
*group_leader
= NULL
, *output_event
= NULL
;
7008 struct perf_event
*event
, *sibling
;
7009 struct perf_event_attr attr
;
7010 struct perf_event_context
*ctx
;
7011 struct file
*event_file
= NULL
;
7012 struct fd group
= {NULL
, 0};
7013 struct task_struct
*task
= NULL
;
7018 int f_flags
= O_RDWR
;
7020 /* for future expandability... */
7021 if (flags
& ~PERF_FLAG_ALL
)
7024 err
= perf_copy_attr(attr_uptr
, &attr
);
7028 if (!attr
.exclude_kernel
) {
7029 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN
))
7034 if (attr
.sample_freq
> sysctl_perf_event_sample_rate
)
7039 * In cgroup mode, the pid argument is used to pass the fd
7040 * opened to the cgroup directory in cgroupfs. The cpu argument
7041 * designates the cpu on which to monitor threads from that
7044 if ((flags
& PERF_FLAG_PID_CGROUP
) && (pid
== -1 || cpu
== -1))
7047 if (flags
& PERF_FLAG_FD_CLOEXEC
)
7048 f_flags
|= O_CLOEXEC
;
7050 event_fd
= get_unused_fd_flags(f_flags
);
7054 if (group_fd
!= -1) {
7055 err
= perf_fget_light(group_fd
, &group
);
7058 group_leader
= group
.file
->private_data
;
7059 if (flags
& PERF_FLAG_FD_OUTPUT
)
7060 output_event
= group_leader
;
7061 if (flags
& PERF_FLAG_FD_NO_GROUP
)
7062 group_leader
= NULL
;
7065 if (pid
!= -1 && !(flags
& PERF_FLAG_PID_CGROUP
)) {
7066 task
= find_lively_task_by_vpid(pid
);
7068 err
= PTR_ERR(task
);
7075 event
= perf_event_alloc(&attr
, cpu
, task
, group_leader
, NULL
,
7077 if (IS_ERR(event
)) {
7078 err
= PTR_ERR(event
);
7082 if (flags
& PERF_FLAG_PID_CGROUP
) {
7083 err
= perf_cgroup_connect(pid
, event
, &attr
, group_leader
);
7085 __free_event(event
);
7090 account_event(event
);
7093 * Special case software events and allow them to be part of
7094 * any hardware group.
7099 (is_software_event(event
) != is_software_event(group_leader
))) {
7100 if (is_software_event(event
)) {
7102 * If event and group_leader are not both a software
7103 * event, and event is, then group leader is not.
7105 * Allow the addition of software events to !software
7106 * groups, this is safe because software events never
7109 pmu
= group_leader
->pmu
;
7110 } else if (is_software_event(group_leader
) &&
7111 (group_leader
->group_flags
& PERF_GROUP_SOFTWARE
)) {
7113 * In case the group is a pure software group, and we
7114 * try to add a hardware event, move the whole group to
7115 * the hardware context.
7122 * Get the target context (task or percpu):
7124 ctx
= find_get_context(pmu
, task
, event
->cpu
);
7131 put_task_struct(task
);
7136 * Look up the group leader (we will attach this event to it):
7142 * Do not allow a recursive hierarchy (this new sibling
7143 * becoming part of another group-sibling):
7145 if (group_leader
->group_leader
!= group_leader
)
7148 * Do not allow to attach to a group in a different
7149 * task or CPU context:
7152 if (group_leader
->ctx
->type
!= ctx
->type
)
7155 if (group_leader
->ctx
!= ctx
)
7160 * Only a group leader can be exclusive or pinned
7162 if (attr
.exclusive
|| attr
.pinned
)
7167 err
= perf_event_set_output(event
, output_event
);
7172 event_file
= anon_inode_getfile("[perf_event]", &perf_fops
, event
,
7174 if (IS_ERR(event_file
)) {
7175 err
= PTR_ERR(event_file
);
7180 struct perf_event_context
*gctx
= group_leader
->ctx
;
7182 mutex_lock(&gctx
->mutex
);
7183 perf_remove_from_context(group_leader
);
7186 * Removing from the context ends up with disabled
7187 * event. What we want here is event in the initial
7188 * startup state, ready to be add into new context.
7190 perf_event__state_init(group_leader
);
7191 list_for_each_entry(sibling
, &group_leader
->sibling_list
,
7193 perf_remove_from_context(sibling
);
7194 perf_event__state_init(sibling
);
7197 mutex_unlock(&gctx
->mutex
);
7201 WARN_ON_ONCE(ctx
->parent_ctx
);
7202 mutex_lock(&ctx
->mutex
);
7206 perf_install_in_context(ctx
, group_leader
, event
->cpu
);
7208 list_for_each_entry(sibling
, &group_leader
->sibling_list
,
7210 perf_install_in_context(ctx
, sibling
, event
->cpu
);
7215 perf_install_in_context(ctx
, event
, event
->cpu
);
7216 perf_unpin_context(ctx
);
7217 mutex_unlock(&ctx
->mutex
);
7221 event
->owner
= current
;
7223 mutex_lock(¤t
->perf_event_mutex
);
7224 list_add_tail(&event
->owner_entry
, ¤t
->perf_event_list
);
7225 mutex_unlock(¤t
->perf_event_mutex
);
7228 * Precalculate sample_data sizes
7230 perf_event__header_size(event
);
7231 perf_event__id_header_size(event
);
7234 * Drop the reference on the group_event after placing the
7235 * new event on the sibling_list. This ensures destruction
7236 * of the group leader will find the pointer to itself in
7237 * perf_group_detach().
7240 fd_install(event_fd
, event_file
);
7244 perf_unpin_context(ctx
);
7251 put_task_struct(task
);
7255 put_unused_fd(event_fd
);
7260 * perf_event_create_kernel_counter
7262 * @attr: attributes of the counter to create
7263 * @cpu: cpu in which the counter is bound
7264 * @task: task to profile (NULL for percpu)
7267 perf_event_create_kernel_counter(struct perf_event_attr
*attr
, int cpu
,
7268 struct task_struct
*task
,
7269 perf_overflow_handler_t overflow_handler
,
7272 struct perf_event_context
*ctx
;
7273 struct perf_event
*event
;
7277 * Get the target context (task or percpu):
7280 event
= perf_event_alloc(attr
, cpu
, task
, NULL
, NULL
,
7281 overflow_handler
, context
);
7282 if (IS_ERR(event
)) {
7283 err
= PTR_ERR(event
);
7287 account_event(event
);
7289 ctx
= find_get_context(event
->pmu
, task
, cpu
);
7295 WARN_ON_ONCE(ctx
->parent_ctx
);
7296 mutex_lock(&ctx
->mutex
);
7297 perf_install_in_context(ctx
, event
, cpu
);
7298 perf_unpin_context(ctx
);
7299 mutex_unlock(&ctx
->mutex
);
7306 return ERR_PTR(err
);
7308 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter
);
7310 void perf_pmu_migrate_context(struct pmu
*pmu
, int src_cpu
, int dst_cpu
)
7312 struct perf_event_context
*src_ctx
;
7313 struct perf_event_context
*dst_ctx
;
7314 struct perf_event
*event
, *tmp
;
7317 src_ctx
= &per_cpu_ptr(pmu
->pmu_cpu_context
, src_cpu
)->ctx
;
7318 dst_ctx
= &per_cpu_ptr(pmu
->pmu_cpu_context
, dst_cpu
)->ctx
;
7320 mutex_lock(&src_ctx
->mutex
);
7321 list_for_each_entry_safe(event
, tmp
, &src_ctx
->event_list
,
7323 perf_remove_from_context(event
);
7324 unaccount_event_cpu(event
, src_cpu
);
7326 list_add(&event
->migrate_entry
, &events
);
7328 mutex_unlock(&src_ctx
->mutex
);
7332 mutex_lock(&dst_ctx
->mutex
);
7333 list_for_each_entry_safe(event
, tmp
, &events
, migrate_entry
) {
7334 list_del(&event
->migrate_entry
);
7335 if (event
->state
>= PERF_EVENT_STATE_OFF
)
7336 event
->state
= PERF_EVENT_STATE_INACTIVE
;
7337 account_event_cpu(event
, dst_cpu
);
7338 perf_install_in_context(dst_ctx
, event
, dst_cpu
);
7341 mutex_unlock(&dst_ctx
->mutex
);
7343 EXPORT_SYMBOL_GPL(perf_pmu_migrate_context
);
7345 static void sync_child_event(struct perf_event
*child_event
,
7346 struct task_struct
*child
)
7348 struct perf_event
*parent_event
= child_event
->parent
;
7351 if (child_event
->attr
.inherit_stat
)
7352 perf_event_read_event(child_event
, child
);
7354 child_val
= perf_event_count(child_event
);
7357 * Add back the child's count to the parent's count:
7359 atomic64_add(child_val
, &parent_event
->child_count
);
7360 atomic64_add(child_event
->total_time_enabled
,
7361 &parent_event
->child_total_time_enabled
);
7362 atomic64_add(child_event
->total_time_running
,
7363 &parent_event
->child_total_time_running
);
7366 * Remove this event from the parent's list
7368 WARN_ON_ONCE(parent_event
->ctx
->parent_ctx
);
7369 mutex_lock(&parent_event
->child_mutex
);
7370 list_del_init(&child_event
->child_list
);
7371 mutex_unlock(&parent_event
->child_mutex
);
7374 * Release the parent event, if this was the last
7377 put_event(parent_event
);
7381 __perf_event_exit_task(struct perf_event
*child_event
,
7382 struct perf_event_context
*child_ctx
,
7383 struct task_struct
*child
)
7385 if (child_event
->parent
) {
7386 raw_spin_lock_irq(&child_ctx
->lock
);
7387 perf_group_detach(child_event
);
7388 raw_spin_unlock_irq(&child_ctx
->lock
);
7391 perf_remove_from_context(child_event
);
7394 * It can happen that the parent exits first, and has events
7395 * that are still around due to the child reference. These
7396 * events need to be zapped.
7398 if (child_event
->parent
) {
7399 sync_child_event(child_event
, child
);
7400 free_event(child_event
);
7404 static void perf_event_exit_task_context(struct task_struct
*child
, int ctxn
)
7406 struct perf_event
*child_event
, *tmp
;
7407 struct perf_event_context
*child_ctx
;
7408 unsigned long flags
;
7410 if (likely(!child
->perf_event_ctxp
[ctxn
])) {
7411 perf_event_task(child
, NULL
, 0);
7415 local_irq_save(flags
);
7417 * We can't reschedule here because interrupts are disabled,
7418 * and either child is current or it is a task that can't be
7419 * scheduled, so we are now safe from rescheduling changing
7422 child_ctx
= rcu_dereference_raw(child
->perf_event_ctxp
[ctxn
]);
7425 * Take the context lock here so that if find_get_context is
7426 * reading child->perf_event_ctxp, we wait until it has
7427 * incremented the context's refcount before we do put_ctx below.
7429 raw_spin_lock(&child_ctx
->lock
);
7430 task_ctx_sched_out(child_ctx
);
7431 child
->perf_event_ctxp
[ctxn
] = NULL
;
7433 * If this context is a clone; unclone it so it can't get
7434 * swapped to another process while we're removing all
7435 * the events from it.
7437 unclone_ctx(child_ctx
);
7438 update_context_time(child_ctx
);
7439 raw_spin_unlock_irqrestore(&child_ctx
->lock
, flags
);
7442 * Report the task dead after unscheduling the events so that we
7443 * won't get any samples after PERF_RECORD_EXIT. We can however still
7444 * get a few PERF_RECORD_READ events.
7446 perf_event_task(child
, child_ctx
, 0);
7449 * We can recurse on the same lock type through:
7451 * __perf_event_exit_task()
7452 * sync_child_event()
7454 * mutex_lock(&ctx->mutex)
7456 * But since its the parent context it won't be the same instance.
7458 mutex_lock(&child_ctx
->mutex
);
7461 list_for_each_entry_safe(child_event
, tmp
, &child_ctx
->pinned_groups
,
7463 __perf_event_exit_task(child_event
, child_ctx
, child
);
7465 list_for_each_entry_safe(child_event
, tmp
, &child_ctx
->flexible_groups
,
7467 __perf_event_exit_task(child_event
, child_ctx
, child
);
7470 * If the last event was a group event, it will have appended all
7471 * its siblings to the list, but we obtained 'tmp' before that which
7472 * will still point to the list head terminating the iteration.
7474 if (!list_empty(&child_ctx
->pinned_groups
) ||
7475 !list_empty(&child_ctx
->flexible_groups
))
7478 mutex_unlock(&child_ctx
->mutex
);
7484 * When a child task exits, feed back event values to parent events.
7486 void perf_event_exit_task(struct task_struct
*child
)
7488 struct perf_event
*event
, *tmp
;
7491 mutex_lock(&child
->perf_event_mutex
);
7492 list_for_each_entry_safe(event
, tmp
, &child
->perf_event_list
,
7494 list_del_init(&event
->owner_entry
);
7497 * Ensure the list deletion is visible before we clear
7498 * the owner, closes a race against perf_release() where
7499 * we need to serialize on the owner->perf_event_mutex.
7502 event
->owner
= NULL
;
7504 mutex_unlock(&child
->perf_event_mutex
);
7506 for_each_task_context_nr(ctxn
)
7507 perf_event_exit_task_context(child
, ctxn
);
7510 static void perf_free_event(struct perf_event
*event
,
7511 struct perf_event_context
*ctx
)
7513 struct perf_event
*parent
= event
->parent
;
7515 if (WARN_ON_ONCE(!parent
))
7518 mutex_lock(&parent
->child_mutex
);
7519 list_del_init(&event
->child_list
);
7520 mutex_unlock(&parent
->child_mutex
);
7524 perf_group_detach(event
);
7525 list_del_event(event
, ctx
);
7530 * free an unexposed, unused context as created by inheritance by
7531 * perf_event_init_task below, used by fork() in case of fail.
7533 void perf_event_free_task(struct task_struct
*task
)
7535 struct perf_event_context
*ctx
;
7536 struct perf_event
*event
, *tmp
;
7539 for_each_task_context_nr(ctxn
) {
7540 ctx
= task
->perf_event_ctxp
[ctxn
];
7544 mutex_lock(&ctx
->mutex
);
7546 list_for_each_entry_safe(event
, tmp
, &ctx
->pinned_groups
,
7548 perf_free_event(event
, ctx
);
7550 list_for_each_entry_safe(event
, tmp
, &ctx
->flexible_groups
,
7552 perf_free_event(event
, ctx
);
7554 if (!list_empty(&ctx
->pinned_groups
) ||
7555 !list_empty(&ctx
->flexible_groups
))
7558 mutex_unlock(&ctx
->mutex
);
7564 void perf_event_delayed_put(struct task_struct
*task
)
7568 for_each_task_context_nr(ctxn
)
7569 WARN_ON_ONCE(task
->perf_event_ctxp
[ctxn
]);
7573 * inherit a event from parent task to child task:
7575 static struct perf_event
*
7576 inherit_event(struct perf_event
*parent_event
,
7577 struct task_struct
*parent
,
7578 struct perf_event_context
*parent_ctx
,
7579 struct task_struct
*child
,
7580 struct perf_event
*group_leader
,
7581 struct perf_event_context
*child_ctx
)
7583 struct perf_event
*child_event
;
7584 unsigned long flags
;
7587 * Instead of creating recursive hierarchies of events,
7588 * we link inherited events back to the original parent,
7589 * which has a filp for sure, which we use as the reference
7592 if (parent_event
->parent
)
7593 parent_event
= parent_event
->parent
;
7595 child_event
= perf_event_alloc(&parent_event
->attr
,
7598 group_leader
, parent_event
,
7600 if (IS_ERR(child_event
))
7603 if (!atomic_long_inc_not_zero(&parent_event
->refcount
)) {
7604 free_event(child_event
);
7611 * Make the child state follow the state of the parent event,
7612 * not its attr.disabled bit. We hold the parent's mutex,
7613 * so we won't race with perf_event_{en, dis}able_family.
7615 if (parent_event
->state
>= PERF_EVENT_STATE_INACTIVE
)
7616 child_event
->state
= PERF_EVENT_STATE_INACTIVE
;
7618 child_event
->state
= PERF_EVENT_STATE_OFF
;
7620 if (parent_event
->attr
.freq
) {
7621 u64 sample_period
= parent_event
->hw
.sample_period
;
7622 struct hw_perf_event
*hwc
= &child_event
->hw
;
7624 hwc
->sample_period
= sample_period
;
7625 hwc
->last_period
= sample_period
;
7627 local64_set(&hwc
->period_left
, sample_period
);
7630 child_event
->ctx
= child_ctx
;
7631 child_event
->overflow_handler
= parent_event
->overflow_handler
;
7632 child_event
->overflow_handler_context
7633 = parent_event
->overflow_handler_context
;
7636 * Precalculate sample_data sizes
7638 perf_event__header_size(child_event
);
7639 perf_event__id_header_size(child_event
);
7642 * Link it up in the child's context:
7644 raw_spin_lock_irqsave(&child_ctx
->lock
, flags
);
7645 add_event_to_ctx(child_event
, child_ctx
);
7646 raw_spin_unlock_irqrestore(&child_ctx
->lock
, flags
);
7649 * Link this into the parent event's child list
7651 WARN_ON_ONCE(parent_event
->ctx
->parent_ctx
);
7652 mutex_lock(&parent_event
->child_mutex
);
7653 list_add_tail(&child_event
->child_list
, &parent_event
->child_list
);
7654 mutex_unlock(&parent_event
->child_mutex
);
7659 static int inherit_group(struct perf_event
*parent_event
,
7660 struct task_struct
*parent
,
7661 struct perf_event_context
*parent_ctx
,
7662 struct task_struct
*child
,
7663 struct perf_event_context
*child_ctx
)
7665 struct perf_event
*leader
;
7666 struct perf_event
*sub
;
7667 struct perf_event
*child_ctr
;
7669 leader
= inherit_event(parent_event
, parent
, parent_ctx
,
7670 child
, NULL
, child_ctx
);
7672 return PTR_ERR(leader
);
7673 list_for_each_entry(sub
, &parent_event
->sibling_list
, group_entry
) {
7674 child_ctr
= inherit_event(sub
, parent
, parent_ctx
,
7675 child
, leader
, child_ctx
);
7676 if (IS_ERR(child_ctr
))
7677 return PTR_ERR(child_ctr
);
7683 inherit_task_group(struct perf_event
*event
, struct task_struct
*parent
,
7684 struct perf_event_context
*parent_ctx
,
7685 struct task_struct
*child
, int ctxn
,
7689 struct perf_event_context
*child_ctx
;
7691 if (!event
->attr
.inherit
) {
7696 child_ctx
= child
->perf_event_ctxp
[ctxn
];
7699 * This is executed from the parent task context, so
7700 * inherit events that have been marked for cloning.
7701 * First allocate and initialize a context for the
7705 child_ctx
= alloc_perf_context(parent_ctx
->pmu
, child
);
7709 child
->perf_event_ctxp
[ctxn
] = child_ctx
;
7712 ret
= inherit_group(event
, parent
, parent_ctx
,
7722 * Initialize the perf_event context in task_struct
7724 int perf_event_init_context(struct task_struct
*child
, int ctxn
)
7726 struct perf_event_context
*child_ctx
, *parent_ctx
;
7727 struct perf_event_context
*cloned_ctx
;
7728 struct perf_event
*event
;
7729 struct task_struct
*parent
= current
;
7730 int inherited_all
= 1;
7731 unsigned long flags
;
7734 if (likely(!parent
->perf_event_ctxp
[ctxn
]))
7738 * If the parent's context is a clone, pin it so it won't get
7741 parent_ctx
= perf_pin_task_context(parent
, ctxn
);
7744 * No need to check if parent_ctx != NULL here; since we saw
7745 * it non-NULL earlier, the only reason for it to become NULL
7746 * is if we exit, and since we're currently in the middle of
7747 * a fork we can't be exiting at the same time.
7751 * Lock the parent list. No need to lock the child - not PID
7752 * hashed yet and not running, so nobody can access it.
7754 mutex_lock(&parent_ctx
->mutex
);
7757 * We dont have to disable NMIs - we are only looking at
7758 * the list, not manipulating it:
7760 list_for_each_entry(event
, &parent_ctx
->pinned_groups
, group_entry
) {
7761 ret
= inherit_task_group(event
, parent
, parent_ctx
,
7762 child
, ctxn
, &inherited_all
);
7768 * We can't hold ctx->lock when iterating the ->flexible_group list due
7769 * to allocations, but we need to prevent rotation because
7770 * rotate_ctx() will change the list from interrupt context.
7772 raw_spin_lock_irqsave(&parent_ctx
->lock
, flags
);
7773 parent_ctx
->rotate_disable
= 1;
7774 raw_spin_unlock_irqrestore(&parent_ctx
->lock
, flags
);
7776 list_for_each_entry(event
, &parent_ctx
->flexible_groups
, group_entry
) {
7777 ret
= inherit_task_group(event
, parent
, parent_ctx
,
7778 child
, ctxn
, &inherited_all
);
7783 raw_spin_lock_irqsave(&parent_ctx
->lock
, flags
);
7784 parent_ctx
->rotate_disable
= 0;
7786 child_ctx
= child
->perf_event_ctxp
[ctxn
];
7788 if (child_ctx
&& inherited_all
) {
7790 * Mark the child context as a clone of the parent
7791 * context, or of whatever the parent is a clone of.
7793 * Note that if the parent is a clone, the holding of
7794 * parent_ctx->lock avoids it from being uncloned.
7796 cloned_ctx
= parent_ctx
->parent_ctx
;
7798 child_ctx
->parent_ctx
= cloned_ctx
;
7799 child_ctx
->parent_gen
= parent_ctx
->parent_gen
;
7801 child_ctx
->parent_ctx
= parent_ctx
;
7802 child_ctx
->parent_gen
= parent_ctx
->generation
;
7804 get_ctx(child_ctx
->parent_ctx
);
7807 raw_spin_unlock_irqrestore(&parent_ctx
->lock
, flags
);
7808 mutex_unlock(&parent_ctx
->mutex
);
7810 perf_unpin_context(parent_ctx
);
7811 put_ctx(parent_ctx
);
7817 * Initialize the perf_event context in task_struct
7819 int perf_event_init_task(struct task_struct
*child
)
7823 memset(child
->perf_event_ctxp
, 0, sizeof(child
->perf_event_ctxp
));
7824 mutex_init(&child
->perf_event_mutex
);
7825 INIT_LIST_HEAD(&child
->perf_event_list
);
7827 for_each_task_context_nr(ctxn
) {
7828 ret
= perf_event_init_context(child
, ctxn
);
7836 static void __init
perf_event_init_all_cpus(void)
7838 struct swevent_htable
*swhash
;
7841 for_each_possible_cpu(cpu
) {
7842 swhash
= &per_cpu(swevent_htable
, cpu
);
7843 mutex_init(&swhash
->hlist_mutex
);
7844 INIT_LIST_HEAD(&per_cpu(rotation_list
, cpu
));
7848 static void perf_event_init_cpu(int cpu
)
7850 struct swevent_htable
*swhash
= &per_cpu(swevent_htable
, cpu
);
7852 mutex_lock(&swhash
->hlist_mutex
);
7853 if (swhash
->hlist_refcount
> 0) {
7854 struct swevent_hlist
*hlist
;
7856 hlist
= kzalloc_node(sizeof(*hlist
), GFP_KERNEL
, cpu_to_node(cpu
));
7858 rcu_assign_pointer(swhash
->swevent_hlist
, hlist
);
7860 mutex_unlock(&swhash
->hlist_mutex
);
7863 #if defined CONFIG_HOTPLUG_CPU || defined CONFIG_KEXEC
7864 static void perf_pmu_rotate_stop(struct pmu
*pmu
)
7866 struct perf_cpu_context
*cpuctx
= this_cpu_ptr(pmu
->pmu_cpu_context
);
7868 WARN_ON(!irqs_disabled());
7870 list_del_init(&cpuctx
->rotation_list
);
7873 static void __perf_event_exit_context(void *__info
)
7875 struct perf_event_context
*ctx
= __info
;
7876 struct perf_event
*event
;
7878 perf_pmu_rotate_stop(ctx
->pmu
);
7881 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
)
7882 __perf_remove_from_context(event
);
7886 static void perf_event_exit_cpu_context(int cpu
)
7888 struct perf_event_context
*ctx
;
7892 idx
= srcu_read_lock(&pmus_srcu
);
7893 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
7894 ctx
= &per_cpu_ptr(pmu
->pmu_cpu_context
, cpu
)->ctx
;
7896 mutex_lock(&ctx
->mutex
);
7897 smp_call_function_single(cpu
, __perf_event_exit_context
, ctx
, 1);
7898 mutex_unlock(&ctx
->mutex
);
7900 srcu_read_unlock(&pmus_srcu
, idx
);
7903 static void perf_event_exit_cpu(int cpu
)
7905 struct swevent_htable
*swhash
= &per_cpu(swevent_htable
, cpu
);
7907 perf_event_exit_cpu_context(cpu
);
7909 mutex_lock(&swhash
->hlist_mutex
);
7910 swevent_hlist_release(swhash
);
7911 mutex_unlock(&swhash
->hlist_mutex
);
7914 static inline void perf_event_exit_cpu(int cpu
) { }
7918 perf_reboot(struct notifier_block
*notifier
, unsigned long val
, void *v
)
7922 for_each_online_cpu(cpu
)
7923 perf_event_exit_cpu(cpu
);
7929 * Run the perf reboot notifier at the very last possible moment so that
7930 * the generic watchdog code runs as long as possible.
7932 static struct notifier_block perf_reboot_notifier
= {
7933 .notifier_call
= perf_reboot
,
7934 .priority
= INT_MIN
,
7938 perf_cpu_notify(struct notifier_block
*self
, unsigned long action
, void *hcpu
)
7940 unsigned int cpu
= (long)hcpu
;
7942 switch (action
& ~CPU_TASKS_FROZEN
) {
7944 case CPU_UP_PREPARE
:
7945 case CPU_DOWN_FAILED
:
7946 perf_event_init_cpu(cpu
);
7949 case CPU_UP_CANCELED
:
7950 case CPU_DOWN_PREPARE
:
7951 perf_event_exit_cpu(cpu
);
7960 void __init
perf_event_init(void)
7966 perf_event_init_all_cpus();
7967 init_srcu_struct(&pmus_srcu
);
7968 perf_pmu_register(&perf_swevent
, "software", PERF_TYPE_SOFTWARE
);
7969 perf_pmu_register(&perf_cpu_clock
, NULL
, -1);
7970 perf_pmu_register(&perf_task_clock
, NULL
, -1);
7972 perf_cpu_notifier(perf_cpu_notify
);
7973 register_reboot_notifier(&perf_reboot_notifier
);
7975 ret
= init_hw_breakpoint();
7976 WARN(ret
, "hw_breakpoint initialization failed with: %d", ret
);
7978 /* do not patch jump label more than once per second */
7979 jump_label_rate_limit(&perf_sched_events
, HZ
);
7982 * Build time assertion that we keep the data_head at the intended
7983 * location. IOW, validation we got the __reserved[] size right.
7985 BUILD_BUG_ON((offsetof(struct perf_event_mmap_page
, data_head
))
7989 static int __init
perf_event_sysfs_init(void)
7994 mutex_lock(&pmus_lock
);
7996 ret
= bus_register(&pmu_bus
);
8000 list_for_each_entry(pmu
, &pmus
, entry
) {
8001 if (!pmu
->name
|| pmu
->type
< 0)
8004 ret
= pmu_dev_alloc(pmu
);
8005 WARN(ret
, "Failed to register pmu: %s, reason %d\n", pmu
->name
, ret
);
8007 pmu_bus_running
= 1;
8011 mutex_unlock(&pmus_lock
);
8015 device_initcall(perf_event_sysfs_init
);
8017 #ifdef CONFIG_CGROUP_PERF
8018 static struct cgroup_subsys_state
*
8019 perf_cgroup_css_alloc(struct cgroup_subsys_state
*parent_css
)
8021 struct perf_cgroup
*jc
;
8023 jc
= kzalloc(sizeof(*jc
), GFP_KERNEL
);
8025 return ERR_PTR(-ENOMEM
);
8027 jc
->info
= alloc_percpu(struct perf_cgroup_info
);
8030 return ERR_PTR(-ENOMEM
);
8036 static void perf_cgroup_css_free(struct cgroup_subsys_state
*css
)
8038 struct perf_cgroup
*jc
= container_of(css
, struct perf_cgroup
, css
);
8040 free_percpu(jc
->info
);
8044 static int __perf_cgroup_move(void *info
)
8046 struct task_struct
*task
= info
;
8047 perf_cgroup_switch(task
, PERF_CGROUP_SWOUT
| PERF_CGROUP_SWIN
);
8051 static void perf_cgroup_attach(struct cgroup_subsys_state
*css
,
8052 struct cgroup_taskset
*tset
)
8054 struct task_struct
*task
;
8056 cgroup_taskset_for_each(task
, css
, tset
)
8057 task_function_call(task
, __perf_cgroup_move
, task
);
8060 static void perf_cgroup_exit(struct cgroup_subsys_state
*css
,
8061 struct cgroup_subsys_state
*old_css
,
8062 struct task_struct
*task
)
8065 * cgroup_exit() is called in the copy_process() failure path.
8066 * Ignore this case since the task hasn't ran yet, this avoids
8067 * trying to poke a half freed task state from generic code.
8069 if (!(task
->flags
& PF_EXITING
))
8072 task_function_call(task
, __perf_cgroup_move
, task
);
8075 struct cgroup_subsys perf_subsys
= {
8076 .name
= "perf_event",
8077 .subsys_id
= perf_subsys_id
,
8078 .css_alloc
= perf_cgroup_css_alloc
,
8079 .css_free
= perf_cgroup_css_free
,
8080 .exit
= perf_cgroup_exit
,
8081 .attach
= perf_cgroup_attach
,
8083 #endif /* CONFIG_CGROUP_PERF */