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/sysfs.h>
22 #include <linux/dcache.h>
23 #include <linux/percpu.h>
24 #include <linux/ptrace.h>
25 #include <linux/reboot.h>
26 #include <linux/vmstat.h>
27 #include <linux/device.h>
28 #include <linux/export.h>
29 #include <linux/vmalloc.h>
30 #include <linux/hardirq.h>
31 #include <linux/rculist.h>
32 #include <linux/uaccess.h>
33 #include <linux/syscalls.h>
34 #include <linux/anon_inodes.h>
35 #include <linux/kernel_stat.h>
36 #include <linux/perf_event.h>
37 #include <linux/ftrace_event.h>
38 #include <linux/hw_breakpoint.h>
39 #include <linux/mm_types.h>
43 #include <asm/irq_regs.h>
45 struct remote_function_call
{
46 struct task_struct
*p
;
47 int (*func
)(void *info
);
52 static void remote_function(void *data
)
54 struct remote_function_call
*tfc
= data
;
55 struct task_struct
*p
= tfc
->p
;
59 if (task_cpu(p
) != smp_processor_id() || !task_curr(p
))
63 tfc
->ret
= tfc
->func(tfc
->info
);
67 * task_function_call - call a function on the cpu on which a task runs
68 * @p: the task to evaluate
69 * @func: the function to be called
70 * @info: the function call argument
72 * Calls the function @func when the task is currently running. This might
73 * be on the current CPU, which just calls the function directly
75 * returns: @func return value, or
76 * -ESRCH - when the process isn't running
77 * -EAGAIN - when the process moved away
80 task_function_call(struct task_struct
*p
, int (*func
) (void *info
), void *info
)
82 struct remote_function_call data
= {
86 .ret
= -ESRCH
, /* No such (running) process */
90 smp_call_function_single(task_cpu(p
), remote_function
, &data
, 1);
96 * cpu_function_call - call a function on the cpu
97 * @func: the function to be called
98 * @info: the function call argument
100 * Calls the function @func on the remote cpu.
102 * returns: @func return value or -ENXIO when the cpu is offline
104 static int cpu_function_call(int cpu
, int (*func
) (void *info
), void *info
)
106 struct remote_function_call data
= {
110 .ret
= -ENXIO
, /* No such CPU */
113 smp_call_function_single(cpu
, remote_function
, &data
, 1);
118 #define PERF_FLAG_ALL (PERF_FLAG_FD_NO_GROUP |\
119 PERF_FLAG_FD_OUTPUT |\
120 PERF_FLAG_PID_CGROUP)
123 * branch priv levels that need permission checks
125 #define PERF_SAMPLE_BRANCH_PERM_PLM \
126 (PERF_SAMPLE_BRANCH_KERNEL |\
127 PERF_SAMPLE_BRANCH_HV)
130 EVENT_FLEXIBLE
= 0x1,
132 EVENT_ALL
= EVENT_FLEXIBLE
| EVENT_PINNED
,
136 * perf_sched_events : >0 events exist
137 * perf_cgroup_events: >0 per-cpu cgroup events exist on this cpu
139 struct static_key_deferred perf_sched_events __read_mostly
;
140 static DEFINE_PER_CPU(atomic_t
, perf_cgroup_events
);
141 static DEFINE_PER_CPU(atomic_t
, perf_branch_stack_events
);
143 static atomic_t nr_mmap_events __read_mostly
;
144 static atomic_t nr_comm_events __read_mostly
;
145 static atomic_t nr_task_events __read_mostly
;
147 static LIST_HEAD(pmus
);
148 static DEFINE_MUTEX(pmus_lock
);
149 static struct srcu_struct pmus_srcu
;
152 * perf event paranoia level:
153 * -1 - not paranoid at all
154 * 0 - disallow raw tracepoint access for unpriv
155 * 1 - disallow cpu events for unpriv
156 * 2 - disallow kernel profiling for unpriv
158 int sysctl_perf_event_paranoid __read_mostly
= 1;
160 /* Minimum for 512 kiB + 1 user control page */
161 int sysctl_perf_event_mlock __read_mostly
= 512 + (PAGE_SIZE
/ 1024); /* 'free' kiB per user */
164 * max perf event sample rate
166 #define DEFAULT_MAX_SAMPLE_RATE 100000
167 int sysctl_perf_event_sample_rate __read_mostly
= DEFAULT_MAX_SAMPLE_RATE
;
168 static int max_samples_per_tick __read_mostly
=
169 DIV_ROUND_UP(DEFAULT_MAX_SAMPLE_RATE
, HZ
);
171 int perf_proc_update_handler(struct ctl_table
*table
, int write
,
172 void __user
*buffer
, size_t *lenp
,
175 int ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
180 max_samples_per_tick
= DIV_ROUND_UP(sysctl_perf_event_sample_rate
, HZ
);
185 static atomic64_t perf_event_id
;
187 static void cpu_ctx_sched_out(struct perf_cpu_context
*cpuctx
,
188 enum event_type_t event_type
);
190 static void cpu_ctx_sched_in(struct perf_cpu_context
*cpuctx
,
191 enum event_type_t event_type
,
192 struct task_struct
*task
);
194 static void update_context_time(struct perf_event_context
*ctx
);
195 static u64
perf_event_time(struct perf_event
*event
);
197 static void ring_buffer_attach(struct perf_event
*event
,
198 struct ring_buffer
*rb
);
200 void __weak
perf_event_print_debug(void) { }
202 extern __weak
const char *perf_pmu_name(void)
207 static inline u64
perf_clock(void)
209 return local_clock();
212 static inline struct perf_cpu_context
*
213 __get_cpu_context(struct perf_event_context
*ctx
)
215 return this_cpu_ptr(ctx
->pmu
->pmu_cpu_context
);
218 static void perf_ctx_lock(struct perf_cpu_context
*cpuctx
,
219 struct perf_event_context
*ctx
)
221 raw_spin_lock(&cpuctx
->ctx
.lock
);
223 raw_spin_lock(&ctx
->lock
);
226 static void perf_ctx_unlock(struct perf_cpu_context
*cpuctx
,
227 struct perf_event_context
*ctx
)
230 raw_spin_unlock(&ctx
->lock
);
231 raw_spin_unlock(&cpuctx
->ctx
.lock
);
234 #ifdef CONFIG_CGROUP_PERF
237 * Must ensure cgroup is pinned (css_get) before calling
238 * this function. In other words, we cannot call this function
239 * if there is no cgroup event for the current CPU context.
241 static inline struct perf_cgroup
*
242 perf_cgroup_from_task(struct task_struct
*task
)
244 return container_of(task_subsys_state(task
, perf_subsys_id
),
245 struct perf_cgroup
, css
);
249 perf_cgroup_match(struct perf_event
*event
)
251 struct perf_event_context
*ctx
= event
->ctx
;
252 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
254 return !event
->cgrp
|| event
->cgrp
== cpuctx
->cgrp
;
257 static inline bool perf_tryget_cgroup(struct perf_event
*event
)
259 return css_tryget(&event
->cgrp
->css
);
262 static inline void perf_put_cgroup(struct perf_event
*event
)
264 css_put(&event
->cgrp
->css
);
267 static inline void perf_detach_cgroup(struct perf_event
*event
)
269 perf_put_cgroup(event
);
273 static inline int is_cgroup_event(struct perf_event
*event
)
275 return event
->cgrp
!= NULL
;
278 static inline u64
perf_cgroup_event_time(struct perf_event
*event
)
280 struct perf_cgroup_info
*t
;
282 t
= per_cpu_ptr(event
->cgrp
->info
, event
->cpu
);
286 static inline void __update_cgrp_time(struct perf_cgroup
*cgrp
)
288 struct perf_cgroup_info
*info
;
293 info
= this_cpu_ptr(cgrp
->info
);
295 info
->time
+= now
- info
->timestamp
;
296 info
->timestamp
= now
;
299 static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context
*cpuctx
)
301 struct perf_cgroup
*cgrp_out
= cpuctx
->cgrp
;
303 __update_cgrp_time(cgrp_out
);
306 static inline void update_cgrp_time_from_event(struct perf_event
*event
)
308 struct perf_cgroup
*cgrp
;
311 * ensure we access cgroup data only when needed and
312 * when we know the cgroup is pinned (css_get)
314 if (!is_cgroup_event(event
))
317 cgrp
= perf_cgroup_from_task(current
);
319 * Do not update time when cgroup is not active
321 if (cgrp
== event
->cgrp
)
322 __update_cgrp_time(event
->cgrp
);
326 perf_cgroup_set_timestamp(struct task_struct
*task
,
327 struct perf_event_context
*ctx
)
329 struct perf_cgroup
*cgrp
;
330 struct perf_cgroup_info
*info
;
333 * ctx->lock held by caller
334 * ensure we do not access cgroup data
335 * unless we have the cgroup pinned (css_get)
337 if (!task
|| !ctx
->nr_cgroups
)
340 cgrp
= perf_cgroup_from_task(task
);
341 info
= this_cpu_ptr(cgrp
->info
);
342 info
->timestamp
= ctx
->timestamp
;
345 #define PERF_CGROUP_SWOUT 0x1 /* cgroup switch out every event */
346 #define PERF_CGROUP_SWIN 0x2 /* cgroup switch in events based on task */
349 * reschedule events based on the cgroup constraint of task.
351 * mode SWOUT : schedule out everything
352 * mode SWIN : schedule in based on cgroup for next
354 void perf_cgroup_switch(struct task_struct
*task
, int mode
)
356 struct perf_cpu_context
*cpuctx
;
361 * disable interrupts to avoid geting nr_cgroup
362 * changes via __perf_event_disable(). Also
365 local_irq_save(flags
);
368 * we reschedule only in the presence of cgroup
369 * constrained events.
373 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
374 cpuctx
= this_cpu_ptr(pmu
->pmu_cpu_context
);
377 * perf_cgroup_events says at least one
378 * context on this CPU has cgroup events.
380 * ctx->nr_cgroups reports the number of cgroup
381 * events for a context.
383 if (cpuctx
->ctx
.nr_cgroups
> 0) {
384 perf_ctx_lock(cpuctx
, cpuctx
->task_ctx
);
385 perf_pmu_disable(cpuctx
->ctx
.pmu
);
387 if (mode
& PERF_CGROUP_SWOUT
) {
388 cpu_ctx_sched_out(cpuctx
, EVENT_ALL
);
390 * must not be done before ctxswout due
391 * to event_filter_match() in event_sched_out()
396 if (mode
& PERF_CGROUP_SWIN
) {
397 WARN_ON_ONCE(cpuctx
->cgrp
);
398 /* set cgrp before ctxsw in to
399 * allow event_filter_match() to not
400 * have to pass task around
402 cpuctx
->cgrp
= perf_cgroup_from_task(task
);
403 cpu_ctx_sched_in(cpuctx
, EVENT_ALL
, task
);
405 perf_pmu_enable(cpuctx
->ctx
.pmu
);
406 perf_ctx_unlock(cpuctx
, cpuctx
->task_ctx
);
412 local_irq_restore(flags
);
415 static inline void perf_cgroup_sched_out(struct task_struct
*task
,
416 struct task_struct
*next
)
418 struct perf_cgroup
*cgrp1
;
419 struct perf_cgroup
*cgrp2
= NULL
;
422 * we come here when we know perf_cgroup_events > 0
424 cgrp1
= perf_cgroup_from_task(task
);
427 * next is NULL when called from perf_event_enable_on_exec()
428 * that will systematically cause a cgroup_switch()
431 cgrp2
= perf_cgroup_from_task(next
);
434 * only schedule out current cgroup events if we know
435 * that we are switching to a different cgroup. Otherwise,
436 * do no touch the cgroup events.
439 perf_cgroup_switch(task
, PERF_CGROUP_SWOUT
);
442 static inline void perf_cgroup_sched_in(struct task_struct
*prev
,
443 struct task_struct
*task
)
445 struct perf_cgroup
*cgrp1
;
446 struct perf_cgroup
*cgrp2
= NULL
;
449 * we come here when we know perf_cgroup_events > 0
451 cgrp1
= perf_cgroup_from_task(task
);
453 /* prev can never be NULL */
454 cgrp2
= perf_cgroup_from_task(prev
);
457 * only need to schedule in cgroup events if we are changing
458 * cgroup during ctxsw. Cgroup events were not scheduled
459 * out of ctxsw out if that was not the case.
462 perf_cgroup_switch(task
, PERF_CGROUP_SWIN
);
465 static inline int perf_cgroup_connect(int fd
, struct perf_event
*event
,
466 struct perf_event_attr
*attr
,
467 struct perf_event
*group_leader
)
469 struct perf_cgroup
*cgrp
;
470 struct cgroup_subsys_state
*css
;
472 int ret
= 0, fput_needed
;
474 file
= fget_light(fd
, &fput_needed
);
478 css
= cgroup_css_from_dir(file
, perf_subsys_id
);
484 cgrp
= container_of(css
, struct perf_cgroup
, css
);
487 /* must be done before we fput() the file */
488 if (!perf_tryget_cgroup(event
)) {
495 * all events in a group must monitor
496 * the same cgroup because a task belongs
497 * to only one perf cgroup at a time
499 if (group_leader
&& group_leader
->cgrp
!= cgrp
) {
500 perf_detach_cgroup(event
);
504 fput_light(file
, fput_needed
);
509 perf_cgroup_set_shadow_time(struct perf_event
*event
, u64 now
)
511 struct perf_cgroup_info
*t
;
512 t
= per_cpu_ptr(event
->cgrp
->info
, event
->cpu
);
513 event
->shadow_ctx_time
= now
- t
->timestamp
;
517 perf_cgroup_defer_enabled(struct perf_event
*event
)
520 * when the current task's perf cgroup does not match
521 * the event's, we need to remember to call the
522 * perf_mark_enable() function the first time a task with
523 * a matching perf cgroup is scheduled in.
525 if (is_cgroup_event(event
) && !perf_cgroup_match(event
))
526 event
->cgrp_defer_enabled
= 1;
530 perf_cgroup_mark_enabled(struct perf_event
*event
,
531 struct perf_event_context
*ctx
)
533 struct perf_event
*sub
;
534 u64 tstamp
= perf_event_time(event
);
536 if (!event
->cgrp_defer_enabled
)
539 event
->cgrp_defer_enabled
= 0;
541 event
->tstamp_enabled
= tstamp
- event
->total_time_enabled
;
542 list_for_each_entry(sub
, &event
->sibling_list
, group_entry
) {
543 if (sub
->state
>= PERF_EVENT_STATE_INACTIVE
) {
544 sub
->tstamp_enabled
= tstamp
- sub
->total_time_enabled
;
545 sub
->cgrp_defer_enabled
= 0;
549 #else /* !CONFIG_CGROUP_PERF */
552 perf_cgroup_match(struct perf_event
*event
)
557 static inline void perf_detach_cgroup(struct perf_event
*event
)
560 static inline int is_cgroup_event(struct perf_event
*event
)
565 static inline u64
perf_cgroup_event_cgrp_time(struct perf_event
*event
)
570 static inline void update_cgrp_time_from_event(struct perf_event
*event
)
574 static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context
*cpuctx
)
578 static inline void perf_cgroup_sched_out(struct task_struct
*task
,
579 struct task_struct
*next
)
583 static inline void perf_cgroup_sched_in(struct task_struct
*prev
,
584 struct task_struct
*task
)
588 static inline int perf_cgroup_connect(pid_t pid
, struct perf_event
*event
,
589 struct perf_event_attr
*attr
,
590 struct perf_event
*group_leader
)
596 perf_cgroup_set_timestamp(struct task_struct
*task
,
597 struct perf_event_context
*ctx
)
602 perf_cgroup_switch(struct task_struct
*task
, struct task_struct
*next
)
607 perf_cgroup_set_shadow_time(struct perf_event
*event
, u64 now
)
611 static inline u64
perf_cgroup_event_time(struct perf_event
*event
)
617 perf_cgroup_defer_enabled(struct perf_event
*event
)
622 perf_cgroup_mark_enabled(struct perf_event
*event
,
623 struct perf_event_context
*ctx
)
628 void perf_pmu_disable(struct pmu
*pmu
)
630 int *count
= this_cpu_ptr(pmu
->pmu_disable_count
);
632 pmu
->pmu_disable(pmu
);
635 void perf_pmu_enable(struct pmu
*pmu
)
637 int *count
= this_cpu_ptr(pmu
->pmu_disable_count
);
639 pmu
->pmu_enable(pmu
);
642 static DEFINE_PER_CPU(struct list_head
, rotation_list
);
645 * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
646 * because they're strictly cpu affine and rotate_start is called with IRQs
647 * disabled, while rotate_context is called from IRQ context.
649 static void perf_pmu_rotate_start(struct pmu
*pmu
)
651 struct perf_cpu_context
*cpuctx
= this_cpu_ptr(pmu
->pmu_cpu_context
);
652 struct list_head
*head
= &__get_cpu_var(rotation_list
);
654 WARN_ON(!irqs_disabled());
656 if (list_empty(&cpuctx
->rotation_list
))
657 list_add(&cpuctx
->rotation_list
, head
);
660 static void get_ctx(struct perf_event_context
*ctx
)
662 WARN_ON(!atomic_inc_not_zero(&ctx
->refcount
));
665 static void put_ctx(struct perf_event_context
*ctx
)
667 if (atomic_dec_and_test(&ctx
->refcount
)) {
669 put_ctx(ctx
->parent_ctx
);
671 put_task_struct(ctx
->task
);
672 kfree_rcu(ctx
, rcu_head
);
676 static void unclone_ctx(struct perf_event_context
*ctx
)
678 if (ctx
->parent_ctx
) {
679 put_ctx(ctx
->parent_ctx
);
680 ctx
->parent_ctx
= NULL
;
684 static u32
perf_event_pid(struct perf_event
*event
, struct task_struct
*p
)
687 * only top level events have the pid namespace they were created in
690 event
= event
->parent
;
692 return task_tgid_nr_ns(p
, event
->ns
);
695 static u32
perf_event_tid(struct perf_event
*event
, struct task_struct
*p
)
698 * only top level events have the pid namespace they were created in
701 event
= event
->parent
;
703 return task_pid_nr_ns(p
, event
->ns
);
707 * If we inherit events we want to return the parent event id
710 static u64
primary_event_id(struct perf_event
*event
)
715 id
= event
->parent
->id
;
721 * Get the perf_event_context for a task and lock it.
722 * This has to cope with with the fact that until it is locked,
723 * the context could get moved to another task.
725 static struct perf_event_context
*
726 perf_lock_task_context(struct task_struct
*task
, int ctxn
, unsigned long *flags
)
728 struct perf_event_context
*ctx
;
732 ctx
= rcu_dereference(task
->perf_event_ctxp
[ctxn
]);
735 * If this context is a clone of another, it might
736 * get swapped for another underneath us by
737 * perf_event_task_sched_out, though the
738 * rcu_read_lock() protects us from any context
739 * getting freed. Lock the context and check if it
740 * got swapped before we could get the lock, and retry
741 * if so. If we locked the right context, then it
742 * can't get swapped on us any more.
744 raw_spin_lock_irqsave(&ctx
->lock
, *flags
);
745 if (ctx
!= rcu_dereference(task
->perf_event_ctxp
[ctxn
])) {
746 raw_spin_unlock_irqrestore(&ctx
->lock
, *flags
);
750 if (!atomic_inc_not_zero(&ctx
->refcount
)) {
751 raw_spin_unlock_irqrestore(&ctx
->lock
, *flags
);
760 * Get the context for a task and increment its pin_count so it
761 * can't get swapped to another task. This also increments its
762 * reference count so that the context can't get freed.
764 static struct perf_event_context
*
765 perf_pin_task_context(struct task_struct
*task
, int ctxn
)
767 struct perf_event_context
*ctx
;
770 ctx
= perf_lock_task_context(task
, ctxn
, &flags
);
773 raw_spin_unlock_irqrestore(&ctx
->lock
, flags
);
778 static void perf_unpin_context(struct perf_event_context
*ctx
)
782 raw_spin_lock_irqsave(&ctx
->lock
, flags
);
784 raw_spin_unlock_irqrestore(&ctx
->lock
, flags
);
788 * Update the record of the current time in a context.
790 static void update_context_time(struct perf_event_context
*ctx
)
792 u64 now
= perf_clock();
794 ctx
->time
+= now
- ctx
->timestamp
;
795 ctx
->timestamp
= now
;
798 static u64
perf_event_time(struct perf_event
*event
)
800 struct perf_event_context
*ctx
= event
->ctx
;
802 if (is_cgroup_event(event
))
803 return perf_cgroup_event_time(event
);
805 return ctx
? ctx
->time
: 0;
809 * Update the total_time_enabled and total_time_running fields for a event.
810 * The caller of this function needs to hold the ctx->lock.
812 static void update_event_times(struct perf_event
*event
)
814 struct perf_event_context
*ctx
= event
->ctx
;
817 if (event
->state
< PERF_EVENT_STATE_INACTIVE
||
818 event
->group_leader
->state
< PERF_EVENT_STATE_INACTIVE
)
821 * in cgroup mode, time_enabled represents
822 * the time the event was enabled AND active
823 * tasks were in the monitored cgroup. This is
824 * independent of the activity of the context as
825 * there may be a mix of cgroup and non-cgroup events.
827 * That is why we treat cgroup events differently
830 if (is_cgroup_event(event
))
831 run_end
= perf_cgroup_event_time(event
);
832 else if (ctx
->is_active
)
835 run_end
= event
->tstamp_stopped
;
837 event
->total_time_enabled
= run_end
- event
->tstamp_enabled
;
839 if (event
->state
== PERF_EVENT_STATE_INACTIVE
)
840 run_end
= event
->tstamp_stopped
;
842 run_end
= perf_event_time(event
);
844 event
->total_time_running
= run_end
- event
->tstamp_running
;
849 * Update total_time_enabled and total_time_running for all events in a group.
851 static void update_group_times(struct perf_event
*leader
)
853 struct perf_event
*event
;
855 update_event_times(leader
);
856 list_for_each_entry(event
, &leader
->sibling_list
, group_entry
)
857 update_event_times(event
);
860 static struct list_head
*
861 ctx_group_list(struct perf_event
*event
, struct perf_event_context
*ctx
)
863 if (event
->attr
.pinned
)
864 return &ctx
->pinned_groups
;
866 return &ctx
->flexible_groups
;
870 * Add a event from the lists for its context.
871 * Must be called with ctx->mutex and ctx->lock held.
874 list_add_event(struct perf_event
*event
, struct perf_event_context
*ctx
)
876 WARN_ON_ONCE(event
->attach_state
& PERF_ATTACH_CONTEXT
);
877 event
->attach_state
|= PERF_ATTACH_CONTEXT
;
880 * If we're a stand alone event or group leader, we go to the context
881 * list, group events are kept attached to the group so that
882 * perf_group_detach can, at all times, locate all siblings.
884 if (event
->group_leader
== event
) {
885 struct list_head
*list
;
887 if (is_software_event(event
))
888 event
->group_flags
|= PERF_GROUP_SOFTWARE
;
890 list
= ctx_group_list(event
, ctx
);
891 list_add_tail(&event
->group_entry
, list
);
894 if (is_cgroup_event(event
))
897 if (has_branch_stack(event
))
898 ctx
->nr_branch_stack
++;
900 list_add_rcu(&event
->event_entry
, &ctx
->event_list
);
902 perf_pmu_rotate_start(ctx
->pmu
);
904 if (event
->attr
.inherit_stat
)
909 * Called at perf_event creation and when events are attached/detached from a
912 static void perf_event__read_size(struct perf_event
*event
)
914 int entry
= sizeof(u64
); /* value */
918 if (event
->attr
.read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
921 if (event
->attr
.read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
924 if (event
->attr
.read_format
& PERF_FORMAT_ID
)
925 entry
+= sizeof(u64
);
927 if (event
->attr
.read_format
& PERF_FORMAT_GROUP
) {
928 nr
+= event
->group_leader
->nr_siblings
;
933 event
->read_size
= size
;
936 static void perf_event__header_size(struct perf_event
*event
)
938 struct perf_sample_data
*data
;
939 u64 sample_type
= event
->attr
.sample_type
;
942 perf_event__read_size(event
);
944 if (sample_type
& PERF_SAMPLE_IP
)
945 size
+= sizeof(data
->ip
);
947 if (sample_type
& PERF_SAMPLE_ADDR
)
948 size
+= sizeof(data
->addr
);
950 if (sample_type
& PERF_SAMPLE_PERIOD
)
951 size
+= sizeof(data
->period
);
953 if (sample_type
& PERF_SAMPLE_READ
)
954 size
+= event
->read_size
;
956 event
->header_size
= size
;
959 static void perf_event__id_header_size(struct perf_event
*event
)
961 struct perf_sample_data
*data
;
962 u64 sample_type
= event
->attr
.sample_type
;
965 if (sample_type
& PERF_SAMPLE_TID
)
966 size
+= sizeof(data
->tid_entry
);
968 if (sample_type
& PERF_SAMPLE_TIME
)
969 size
+= sizeof(data
->time
);
971 if (sample_type
& PERF_SAMPLE_ID
)
972 size
+= sizeof(data
->id
);
974 if (sample_type
& PERF_SAMPLE_STREAM_ID
)
975 size
+= sizeof(data
->stream_id
);
977 if (sample_type
& PERF_SAMPLE_CPU
)
978 size
+= sizeof(data
->cpu_entry
);
980 event
->id_header_size
= size
;
983 static void perf_group_attach(struct perf_event
*event
)
985 struct perf_event
*group_leader
= event
->group_leader
, *pos
;
988 * We can have double attach due to group movement in perf_event_open.
990 if (event
->attach_state
& PERF_ATTACH_GROUP
)
993 event
->attach_state
|= PERF_ATTACH_GROUP
;
995 if (group_leader
== event
)
998 if (group_leader
->group_flags
& PERF_GROUP_SOFTWARE
&&
999 !is_software_event(event
))
1000 group_leader
->group_flags
&= ~PERF_GROUP_SOFTWARE
;
1002 list_add_tail(&event
->group_entry
, &group_leader
->sibling_list
);
1003 group_leader
->nr_siblings
++;
1005 perf_event__header_size(group_leader
);
1007 list_for_each_entry(pos
, &group_leader
->sibling_list
, group_entry
)
1008 perf_event__header_size(pos
);
1012 * Remove a event from the lists for its context.
1013 * Must be called with ctx->mutex and ctx->lock held.
1016 list_del_event(struct perf_event
*event
, struct perf_event_context
*ctx
)
1018 struct perf_cpu_context
*cpuctx
;
1020 * We can have double detach due to exit/hot-unplug + close.
1022 if (!(event
->attach_state
& PERF_ATTACH_CONTEXT
))
1025 event
->attach_state
&= ~PERF_ATTACH_CONTEXT
;
1027 if (is_cgroup_event(event
)) {
1029 cpuctx
= __get_cpu_context(ctx
);
1031 * if there are no more cgroup events
1032 * then cler cgrp to avoid stale pointer
1033 * in update_cgrp_time_from_cpuctx()
1035 if (!ctx
->nr_cgroups
)
1036 cpuctx
->cgrp
= NULL
;
1039 if (has_branch_stack(event
))
1040 ctx
->nr_branch_stack
--;
1043 if (event
->attr
.inherit_stat
)
1046 list_del_rcu(&event
->event_entry
);
1048 if (event
->group_leader
== event
)
1049 list_del_init(&event
->group_entry
);
1051 update_group_times(event
);
1054 * If event was in error state, then keep it
1055 * that way, otherwise bogus counts will be
1056 * returned on read(). The only way to get out
1057 * of error state is by explicit re-enabling
1060 if (event
->state
> PERF_EVENT_STATE_OFF
)
1061 event
->state
= PERF_EVENT_STATE_OFF
;
1064 static void perf_group_detach(struct perf_event
*event
)
1066 struct perf_event
*sibling
, *tmp
;
1067 struct list_head
*list
= NULL
;
1070 * We can have double detach due to exit/hot-unplug + close.
1072 if (!(event
->attach_state
& PERF_ATTACH_GROUP
))
1075 event
->attach_state
&= ~PERF_ATTACH_GROUP
;
1078 * If this is a sibling, remove it from its group.
1080 if (event
->group_leader
!= event
) {
1081 list_del_init(&event
->group_entry
);
1082 event
->group_leader
->nr_siblings
--;
1086 if (!list_empty(&event
->group_entry
))
1087 list
= &event
->group_entry
;
1090 * If this was a group event with sibling events then
1091 * upgrade the siblings to singleton events by adding them
1092 * to whatever list we are on.
1094 list_for_each_entry_safe(sibling
, tmp
, &event
->sibling_list
, group_entry
) {
1096 list_move_tail(&sibling
->group_entry
, list
);
1097 sibling
->group_leader
= sibling
;
1099 /* Inherit group flags from the previous leader */
1100 sibling
->group_flags
= event
->group_flags
;
1104 perf_event__header_size(event
->group_leader
);
1106 list_for_each_entry(tmp
, &event
->group_leader
->sibling_list
, group_entry
)
1107 perf_event__header_size(tmp
);
1111 event_filter_match(struct perf_event
*event
)
1113 return (event
->cpu
== -1 || event
->cpu
== smp_processor_id())
1114 && perf_cgroup_match(event
);
1118 event_sched_out(struct perf_event
*event
,
1119 struct perf_cpu_context
*cpuctx
,
1120 struct perf_event_context
*ctx
)
1122 u64 tstamp
= perf_event_time(event
);
1125 * An event which could not be activated because of
1126 * filter mismatch still needs to have its timings
1127 * maintained, otherwise bogus information is return
1128 * via read() for time_enabled, time_running:
1130 if (event
->state
== PERF_EVENT_STATE_INACTIVE
1131 && !event_filter_match(event
)) {
1132 delta
= tstamp
- event
->tstamp_stopped
;
1133 event
->tstamp_running
+= delta
;
1134 event
->tstamp_stopped
= tstamp
;
1137 if (event
->state
!= PERF_EVENT_STATE_ACTIVE
)
1140 event
->state
= PERF_EVENT_STATE_INACTIVE
;
1141 if (event
->pending_disable
) {
1142 event
->pending_disable
= 0;
1143 event
->state
= PERF_EVENT_STATE_OFF
;
1145 event
->tstamp_stopped
= tstamp
;
1146 event
->pmu
->del(event
, 0);
1149 if (!is_software_event(event
))
1150 cpuctx
->active_oncpu
--;
1152 if (event
->attr
.freq
&& event
->attr
.sample_freq
)
1154 if (event
->attr
.exclusive
|| !cpuctx
->active_oncpu
)
1155 cpuctx
->exclusive
= 0;
1159 group_sched_out(struct perf_event
*group_event
,
1160 struct perf_cpu_context
*cpuctx
,
1161 struct perf_event_context
*ctx
)
1163 struct perf_event
*event
;
1164 int state
= group_event
->state
;
1166 event_sched_out(group_event
, cpuctx
, ctx
);
1169 * Schedule out siblings (if any):
1171 list_for_each_entry(event
, &group_event
->sibling_list
, group_entry
)
1172 event_sched_out(event
, cpuctx
, ctx
);
1174 if (state
== PERF_EVENT_STATE_ACTIVE
&& group_event
->attr
.exclusive
)
1175 cpuctx
->exclusive
= 0;
1179 * Cross CPU call to remove a performance event
1181 * We disable the event on the hardware level first. After that we
1182 * remove it from the context list.
1184 static int __perf_remove_from_context(void *info
)
1186 struct perf_event
*event
= info
;
1187 struct perf_event_context
*ctx
= event
->ctx
;
1188 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
1190 raw_spin_lock(&ctx
->lock
);
1191 event_sched_out(event
, cpuctx
, ctx
);
1192 list_del_event(event
, ctx
);
1193 if (!ctx
->nr_events
&& cpuctx
->task_ctx
== ctx
) {
1195 cpuctx
->task_ctx
= NULL
;
1197 raw_spin_unlock(&ctx
->lock
);
1204 * Remove the event from a task's (or a CPU's) list of events.
1206 * CPU events are removed with a smp call. For task events we only
1207 * call when the task is on a CPU.
1209 * If event->ctx is a cloned context, callers must make sure that
1210 * every task struct that event->ctx->task could possibly point to
1211 * remains valid. This is OK when called from perf_release since
1212 * that only calls us on the top-level context, which can't be a clone.
1213 * When called from perf_event_exit_task, it's OK because the
1214 * context has been detached from its task.
1216 static void perf_remove_from_context(struct perf_event
*event
)
1218 struct perf_event_context
*ctx
= event
->ctx
;
1219 struct task_struct
*task
= ctx
->task
;
1221 lockdep_assert_held(&ctx
->mutex
);
1225 * Per cpu events are removed via an smp call and
1226 * the removal is always successful.
1228 cpu_function_call(event
->cpu
, __perf_remove_from_context
, event
);
1233 if (!task_function_call(task
, __perf_remove_from_context
, event
))
1236 raw_spin_lock_irq(&ctx
->lock
);
1238 * If we failed to find a running task, but find the context active now
1239 * that we've acquired the ctx->lock, retry.
1241 if (ctx
->is_active
) {
1242 raw_spin_unlock_irq(&ctx
->lock
);
1247 * Since the task isn't running, its safe to remove the event, us
1248 * holding the ctx->lock ensures the task won't get scheduled in.
1250 list_del_event(event
, ctx
);
1251 raw_spin_unlock_irq(&ctx
->lock
);
1255 * Cross CPU call to disable a performance event
1257 int __perf_event_disable(void *info
)
1259 struct perf_event
*event
= info
;
1260 struct perf_event_context
*ctx
= event
->ctx
;
1261 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
1264 * If this is a per-task event, need to check whether this
1265 * event's task is the current task on this cpu.
1267 * Can trigger due to concurrent perf_event_context_sched_out()
1268 * flipping contexts around.
1270 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
)
1273 raw_spin_lock(&ctx
->lock
);
1276 * If the event is on, turn it off.
1277 * If it is in error state, leave it in error state.
1279 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
) {
1280 update_context_time(ctx
);
1281 update_cgrp_time_from_event(event
);
1282 update_group_times(event
);
1283 if (event
== event
->group_leader
)
1284 group_sched_out(event
, cpuctx
, ctx
);
1286 event_sched_out(event
, cpuctx
, ctx
);
1287 event
->state
= PERF_EVENT_STATE_OFF
;
1290 raw_spin_unlock(&ctx
->lock
);
1298 * If event->ctx is a cloned context, callers must make sure that
1299 * every task struct that event->ctx->task could possibly point to
1300 * remains valid. This condition is satisifed when called through
1301 * perf_event_for_each_child or perf_event_for_each because they
1302 * hold the top-level event's child_mutex, so any descendant that
1303 * goes to exit will block in sync_child_event.
1304 * When called from perf_pending_event it's OK because event->ctx
1305 * is the current context on this CPU and preemption is disabled,
1306 * hence we can't get into perf_event_task_sched_out for this context.
1308 void perf_event_disable(struct perf_event
*event
)
1310 struct perf_event_context
*ctx
= event
->ctx
;
1311 struct task_struct
*task
= ctx
->task
;
1315 * Disable the event on the cpu that it's on
1317 cpu_function_call(event
->cpu
, __perf_event_disable
, event
);
1322 if (!task_function_call(task
, __perf_event_disable
, event
))
1325 raw_spin_lock_irq(&ctx
->lock
);
1327 * If the event is still active, we need to retry the cross-call.
1329 if (event
->state
== PERF_EVENT_STATE_ACTIVE
) {
1330 raw_spin_unlock_irq(&ctx
->lock
);
1332 * Reload the task pointer, it might have been changed by
1333 * a concurrent perf_event_context_sched_out().
1340 * Since we have the lock this context can't be scheduled
1341 * in, so we can change the state safely.
1343 if (event
->state
== PERF_EVENT_STATE_INACTIVE
) {
1344 update_group_times(event
);
1345 event
->state
= PERF_EVENT_STATE_OFF
;
1347 raw_spin_unlock_irq(&ctx
->lock
);
1349 EXPORT_SYMBOL_GPL(perf_event_disable
);
1351 static void perf_set_shadow_time(struct perf_event
*event
,
1352 struct perf_event_context
*ctx
,
1356 * use the correct time source for the time snapshot
1358 * We could get by without this by leveraging the
1359 * fact that to get to this function, the caller
1360 * has most likely already called update_context_time()
1361 * and update_cgrp_time_xx() and thus both timestamp
1362 * are identical (or very close). Given that tstamp is,
1363 * already adjusted for cgroup, we could say that:
1364 * tstamp - ctx->timestamp
1366 * tstamp - cgrp->timestamp.
1368 * Then, in perf_output_read(), the calculation would
1369 * work with no changes because:
1370 * - event is guaranteed scheduled in
1371 * - no scheduled out in between
1372 * - thus the timestamp would be the same
1374 * But this is a bit hairy.
1376 * So instead, we have an explicit cgroup call to remain
1377 * within the time time source all along. We believe it
1378 * is cleaner and simpler to understand.
1380 if (is_cgroup_event(event
))
1381 perf_cgroup_set_shadow_time(event
, tstamp
);
1383 event
->shadow_ctx_time
= tstamp
- ctx
->timestamp
;
1386 #define MAX_INTERRUPTS (~0ULL)
1388 static void perf_log_throttle(struct perf_event
*event
, int enable
);
1391 event_sched_in(struct perf_event
*event
,
1392 struct perf_cpu_context
*cpuctx
,
1393 struct perf_event_context
*ctx
)
1395 u64 tstamp
= perf_event_time(event
);
1397 if (event
->state
<= PERF_EVENT_STATE_OFF
)
1400 event
->state
= PERF_EVENT_STATE_ACTIVE
;
1401 event
->oncpu
= smp_processor_id();
1404 * Unthrottle events, since we scheduled we might have missed several
1405 * ticks already, also for a heavily scheduling task there is little
1406 * guarantee it'll get a tick in a timely manner.
1408 if (unlikely(event
->hw
.interrupts
== MAX_INTERRUPTS
)) {
1409 perf_log_throttle(event
, 1);
1410 event
->hw
.interrupts
= 0;
1414 * The new state must be visible before we turn it on in the hardware:
1418 if (event
->pmu
->add(event
, PERF_EF_START
)) {
1419 event
->state
= PERF_EVENT_STATE_INACTIVE
;
1424 event
->tstamp_running
+= tstamp
- event
->tstamp_stopped
;
1426 perf_set_shadow_time(event
, ctx
, tstamp
);
1428 if (!is_software_event(event
))
1429 cpuctx
->active_oncpu
++;
1431 if (event
->attr
.freq
&& event
->attr
.sample_freq
)
1434 if (event
->attr
.exclusive
)
1435 cpuctx
->exclusive
= 1;
1441 group_sched_in(struct perf_event
*group_event
,
1442 struct perf_cpu_context
*cpuctx
,
1443 struct perf_event_context
*ctx
)
1445 struct perf_event
*event
, *partial_group
= NULL
;
1446 struct pmu
*pmu
= group_event
->pmu
;
1447 u64 now
= ctx
->time
;
1448 bool simulate
= false;
1450 if (group_event
->state
== PERF_EVENT_STATE_OFF
)
1453 pmu
->start_txn(pmu
);
1455 if (event_sched_in(group_event
, cpuctx
, ctx
)) {
1456 pmu
->cancel_txn(pmu
);
1461 * Schedule in siblings as one group (if any):
1463 list_for_each_entry(event
, &group_event
->sibling_list
, group_entry
) {
1464 if (event_sched_in(event
, cpuctx
, ctx
)) {
1465 partial_group
= event
;
1470 if (!pmu
->commit_txn(pmu
))
1475 * Groups can be scheduled in as one unit only, so undo any
1476 * partial group before returning:
1477 * The events up to the failed event are scheduled out normally,
1478 * tstamp_stopped will be updated.
1480 * The failed events and the remaining siblings need to have
1481 * their timings updated as if they had gone thru event_sched_in()
1482 * and event_sched_out(). This is required to get consistent timings
1483 * across the group. This also takes care of the case where the group
1484 * could never be scheduled by ensuring tstamp_stopped is set to mark
1485 * the time the event was actually stopped, such that time delta
1486 * calculation in update_event_times() is correct.
1488 list_for_each_entry(event
, &group_event
->sibling_list
, group_entry
) {
1489 if (event
== partial_group
)
1493 event
->tstamp_running
+= now
- event
->tstamp_stopped
;
1494 event
->tstamp_stopped
= now
;
1496 event_sched_out(event
, cpuctx
, ctx
);
1499 event_sched_out(group_event
, cpuctx
, ctx
);
1501 pmu
->cancel_txn(pmu
);
1507 * Work out whether we can put this event group on the CPU now.
1509 static int group_can_go_on(struct perf_event
*event
,
1510 struct perf_cpu_context
*cpuctx
,
1514 * Groups consisting entirely of software events can always go on.
1516 if (event
->group_flags
& PERF_GROUP_SOFTWARE
)
1519 * If an exclusive group is already on, no other hardware
1522 if (cpuctx
->exclusive
)
1525 * If this group is exclusive and there are already
1526 * events on the CPU, it can't go on.
1528 if (event
->attr
.exclusive
&& cpuctx
->active_oncpu
)
1531 * Otherwise, try to add it if all previous groups were able
1537 static void add_event_to_ctx(struct perf_event
*event
,
1538 struct perf_event_context
*ctx
)
1540 u64 tstamp
= perf_event_time(event
);
1542 list_add_event(event
, ctx
);
1543 perf_group_attach(event
);
1544 event
->tstamp_enabled
= tstamp
;
1545 event
->tstamp_running
= tstamp
;
1546 event
->tstamp_stopped
= tstamp
;
1549 static void task_ctx_sched_out(struct perf_event_context
*ctx
);
1551 ctx_sched_in(struct perf_event_context
*ctx
,
1552 struct perf_cpu_context
*cpuctx
,
1553 enum event_type_t event_type
,
1554 struct task_struct
*task
);
1556 static void perf_event_sched_in(struct perf_cpu_context
*cpuctx
,
1557 struct perf_event_context
*ctx
,
1558 struct task_struct
*task
)
1560 cpu_ctx_sched_in(cpuctx
, EVENT_PINNED
, task
);
1562 ctx_sched_in(ctx
, cpuctx
, EVENT_PINNED
, task
);
1563 cpu_ctx_sched_in(cpuctx
, EVENT_FLEXIBLE
, task
);
1565 ctx_sched_in(ctx
, cpuctx
, EVENT_FLEXIBLE
, task
);
1569 * Cross CPU call to install and enable a performance event
1571 * Must be called with ctx->mutex held
1573 static int __perf_install_in_context(void *info
)
1575 struct perf_event
*event
= info
;
1576 struct perf_event_context
*ctx
= event
->ctx
;
1577 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
1578 struct perf_event_context
*task_ctx
= cpuctx
->task_ctx
;
1579 struct task_struct
*task
= current
;
1581 perf_ctx_lock(cpuctx
, task_ctx
);
1582 perf_pmu_disable(cpuctx
->ctx
.pmu
);
1585 * If there was an active task_ctx schedule it out.
1588 task_ctx_sched_out(task_ctx
);
1591 * If the context we're installing events in is not the
1592 * active task_ctx, flip them.
1594 if (ctx
->task
&& task_ctx
!= ctx
) {
1596 raw_spin_unlock(&task_ctx
->lock
);
1597 raw_spin_lock(&ctx
->lock
);
1602 cpuctx
->task_ctx
= task_ctx
;
1603 task
= task_ctx
->task
;
1606 cpu_ctx_sched_out(cpuctx
, EVENT_ALL
);
1608 update_context_time(ctx
);
1610 * update cgrp time only if current cgrp
1611 * matches event->cgrp. Must be done before
1612 * calling add_event_to_ctx()
1614 update_cgrp_time_from_event(event
);
1616 add_event_to_ctx(event
, ctx
);
1619 * Schedule everything back in
1621 perf_event_sched_in(cpuctx
, task_ctx
, task
);
1623 perf_pmu_enable(cpuctx
->ctx
.pmu
);
1624 perf_ctx_unlock(cpuctx
, task_ctx
);
1630 * Attach a performance event to a context
1632 * First we add the event to the list with the hardware enable bit
1633 * in event->hw_config cleared.
1635 * If the event is attached to a task which is on a CPU we use a smp
1636 * call to enable it in the task context. The task might have been
1637 * scheduled away, but we check this in the smp call again.
1640 perf_install_in_context(struct perf_event_context
*ctx
,
1641 struct perf_event
*event
,
1644 struct task_struct
*task
= ctx
->task
;
1646 lockdep_assert_held(&ctx
->mutex
);
1649 if (event
->cpu
!= -1)
1654 * Per cpu events are installed via an smp call and
1655 * the install is always successful.
1657 cpu_function_call(cpu
, __perf_install_in_context
, event
);
1662 if (!task_function_call(task
, __perf_install_in_context
, event
))
1665 raw_spin_lock_irq(&ctx
->lock
);
1667 * If we failed to find a running task, but find the context active now
1668 * that we've acquired the ctx->lock, retry.
1670 if (ctx
->is_active
) {
1671 raw_spin_unlock_irq(&ctx
->lock
);
1676 * Since the task isn't running, its safe to add the event, us holding
1677 * the ctx->lock ensures the task won't get scheduled in.
1679 add_event_to_ctx(event
, ctx
);
1680 raw_spin_unlock_irq(&ctx
->lock
);
1684 * Put a event into inactive state and update time fields.
1685 * Enabling the leader of a group effectively enables all
1686 * the group members that aren't explicitly disabled, so we
1687 * have to update their ->tstamp_enabled also.
1688 * Note: this works for group members as well as group leaders
1689 * since the non-leader members' sibling_lists will be empty.
1691 static void __perf_event_mark_enabled(struct perf_event
*event
)
1693 struct perf_event
*sub
;
1694 u64 tstamp
= perf_event_time(event
);
1696 event
->state
= PERF_EVENT_STATE_INACTIVE
;
1697 event
->tstamp_enabled
= tstamp
- event
->total_time_enabled
;
1698 list_for_each_entry(sub
, &event
->sibling_list
, group_entry
) {
1699 if (sub
->state
>= PERF_EVENT_STATE_INACTIVE
)
1700 sub
->tstamp_enabled
= tstamp
- sub
->total_time_enabled
;
1705 * Cross CPU call to enable a performance event
1707 static int __perf_event_enable(void *info
)
1709 struct perf_event
*event
= info
;
1710 struct perf_event_context
*ctx
= event
->ctx
;
1711 struct perf_event
*leader
= event
->group_leader
;
1712 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
1715 if (WARN_ON_ONCE(!ctx
->is_active
))
1718 raw_spin_lock(&ctx
->lock
);
1719 update_context_time(ctx
);
1721 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
)
1725 * set current task's cgroup time reference point
1727 perf_cgroup_set_timestamp(current
, ctx
);
1729 __perf_event_mark_enabled(event
);
1731 if (!event_filter_match(event
)) {
1732 if (is_cgroup_event(event
))
1733 perf_cgroup_defer_enabled(event
);
1738 * If the event is in a group and isn't the group leader,
1739 * then don't put it on unless the group is on.
1741 if (leader
!= event
&& leader
->state
!= PERF_EVENT_STATE_ACTIVE
)
1744 if (!group_can_go_on(event
, cpuctx
, 1)) {
1747 if (event
== leader
)
1748 err
= group_sched_in(event
, cpuctx
, ctx
);
1750 err
= event_sched_in(event
, cpuctx
, ctx
);
1755 * If this event can't go on and it's part of a
1756 * group, then the whole group has to come off.
1758 if (leader
!= event
)
1759 group_sched_out(leader
, cpuctx
, ctx
);
1760 if (leader
->attr
.pinned
) {
1761 update_group_times(leader
);
1762 leader
->state
= PERF_EVENT_STATE_ERROR
;
1767 raw_spin_unlock(&ctx
->lock
);
1775 * If event->ctx is a cloned context, callers must make sure that
1776 * every task struct that event->ctx->task could possibly point to
1777 * remains valid. This condition is satisfied when called through
1778 * perf_event_for_each_child or perf_event_for_each as described
1779 * for perf_event_disable.
1781 void perf_event_enable(struct perf_event
*event
)
1783 struct perf_event_context
*ctx
= event
->ctx
;
1784 struct task_struct
*task
= ctx
->task
;
1788 * Enable the event on the cpu that it's on
1790 cpu_function_call(event
->cpu
, __perf_event_enable
, event
);
1794 raw_spin_lock_irq(&ctx
->lock
);
1795 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
)
1799 * If the event is in error state, clear that first.
1800 * That way, if we see the event in error state below, we
1801 * know that it has gone back into error state, as distinct
1802 * from the task having been scheduled away before the
1803 * cross-call arrived.
1805 if (event
->state
== PERF_EVENT_STATE_ERROR
)
1806 event
->state
= PERF_EVENT_STATE_OFF
;
1809 if (!ctx
->is_active
) {
1810 __perf_event_mark_enabled(event
);
1814 raw_spin_unlock_irq(&ctx
->lock
);
1816 if (!task_function_call(task
, __perf_event_enable
, event
))
1819 raw_spin_lock_irq(&ctx
->lock
);
1822 * If the context is active and the event is still off,
1823 * we need to retry the cross-call.
1825 if (ctx
->is_active
&& event
->state
== PERF_EVENT_STATE_OFF
) {
1827 * task could have been flipped by a concurrent
1828 * perf_event_context_sched_out()
1835 raw_spin_unlock_irq(&ctx
->lock
);
1837 EXPORT_SYMBOL_GPL(perf_event_enable
);
1839 int perf_event_refresh(struct perf_event
*event
, int refresh
)
1842 * not supported on inherited events
1844 if (event
->attr
.inherit
|| !is_sampling_event(event
))
1847 atomic_add(refresh
, &event
->event_limit
);
1848 perf_event_enable(event
);
1852 EXPORT_SYMBOL_GPL(perf_event_refresh
);
1854 static void ctx_sched_out(struct perf_event_context
*ctx
,
1855 struct perf_cpu_context
*cpuctx
,
1856 enum event_type_t event_type
)
1858 struct perf_event
*event
;
1859 int is_active
= ctx
->is_active
;
1861 ctx
->is_active
&= ~event_type
;
1862 if (likely(!ctx
->nr_events
))
1865 update_context_time(ctx
);
1866 update_cgrp_time_from_cpuctx(cpuctx
);
1867 if (!ctx
->nr_active
)
1870 perf_pmu_disable(ctx
->pmu
);
1871 if ((is_active
& EVENT_PINNED
) && (event_type
& EVENT_PINNED
)) {
1872 list_for_each_entry(event
, &ctx
->pinned_groups
, group_entry
)
1873 group_sched_out(event
, cpuctx
, ctx
);
1876 if ((is_active
& EVENT_FLEXIBLE
) && (event_type
& EVENT_FLEXIBLE
)) {
1877 list_for_each_entry(event
, &ctx
->flexible_groups
, group_entry
)
1878 group_sched_out(event
, cpuctx
, ctx
);
1880 perf_pmu_enable(ctx
->pmu
);
1884 * Test whether two contexts are equivalent, i.e. whether they
1885 * have both been cloned from the same version of the same context
1886 * and they both have the same number of enabled events.
1887 * If the number of enabled events is the same, then the set
1888 * of enabled events should be the same, because these are both
1889 * inherited contexts, therefore we can't access individual events
1890 * in them directly with an fd; we can only enable/disable all
1891 * events via prctl, or enable/disable all events in a family
1892 * via ioctl, which will have the same effect on both contexts.
1894 static int context_equiv(struct perf_event_context
*ctx1
,
1895 struct perf_event_context
*ctx2
)
1897 return ctx1
->parent_ctx
&& ctx1
->parent_ctx
== ctx2
->parent_ctx
1898 && ctx1
->parent_gen
== ctx2
->parent_gen
1899 && !ctx1
->pin_count
&& !ctx2
->pin_count
;
1902 static void __perf_event_sync_stat(struct perf_event
*event
,
1903 struct perf_event
*next_event
)
1907 if (!event
->attr
.inherit_stat
)
1911 * Update the event value, we cannot use perf_event_read()
1912 * because we're in the middle of a context switch and have IRQs
1913 * disabled, which upsets smp_call_function_single(), however
1914 * we know the event must be on the current CPU, therefore we
1915 * don't need to use it.
1917 switch (event
->state
) {
1918 case PERF_EVENT_STATE_ACTIVE
:
1919 event
->pmu
->read(event
);
1922 case PERF_EVENT_STATE_INACTIVE
:
1923 update_event_times(event
);
1931 * In order to keep per-task stats reliable we need to flip the event
1932 * values when we flip the contexts.
1934 value
= local64_read(&next_event
->count
);
1935 value
= local64_xchg(&event
->count
, value
);
1936 local64_set(&next_event
->count
, value
);
1938 swap(event
->total_time_enabled
, next_event
->total_time_enabled
);
1939 swap(event
->total_time_running
, next_event
->total_time_running
);
1942 * Since we swizzled the values, update the user visible data too.
1944 perf_event_update_userpage(event
);
1945 perf_event_update_userpage(next_event
);
1948 #define list_next_entry(pos, member) \
1949 list_entry(pos->member.next, typeof(*pos), member)
1951 static void perf_event_sync_stat(struct perf_event_context
*ctx
,
1952 struct perf_event_context
*next_ctx
)
1954 struct perf_event
*event
, *next_event
;
1959 update_context_time(ctx
);
1961 event
= list_first_entry(&ctx
->event_list
,
1962 struct perf_event
, event_entry
);
1964 next_event
= list_first_entry(&next_ctx
->event_list
,
1965 struct perf_event
, event_entry
);
1967 while (&event
->event_entry
!= &ctx
->event_list
&&
1968 &next_event
->event_entry
!= &next_ctx
->event_list
) {
1970 __perf_event_sync_stat(event
, next_event
);
1972 event
= list_next_entry(event
, event_entry
);
1973 next_event
= list_next_entry(next_event
, event_entry
);
1977 static void perf_event_context_sched_out(struct task_struct
*task
, int ctxn
,
1978 struct task_struct
*next
)
1980 struct perf_event_context
*ctx
= task
->perf_event_ctxp
[ctxn
];
1981 struct perf_event_context
*next_ctx
;
1982 struct perf_event_context
*parent
;
1983 struct perf_cpu_context
*cpuctx
;
1989 cpuctx
= __get_cpu_context(ctx
);
1990 if (!cpuctx
->task_ctx
)
1994 parent
= rcu_dereference(ctx
->parent_ctx
);
1995 next_ctx
= next
->perf_event_ctxp
[ctxn
];
1996 if (parent
&& next_ctx
&&
1997 rcu_dereference(next_ctx
->parent_ctx
) == parent
) {
1999 * Looks like the two contexts are clones, so we might be
2000 * able to optimize the context switch. We lock both
2001 * contexts and check that they are clones under the
2002 * lock (including re-checking that neither has been
2003 * uncloned in the meantime). It doesn't matter which
2004 * order we take the locks because no other cpu could
2005 * be trying to lock both of these tasks.
2007 raw_spin_lock(&ctx
->lock
);
2008 raw_spin_lock_nested(&next_ctx
->lock
, SINGLE_DEPTH_NESTING
);
2009 if (context_equiv(ctx
, next_ctx
)) {
2011 * XXX do we need a memory barrier of sorts
2012 * wrt to rcu_dereference() of perf_event_ctxp
2014 task
->perf_event_ctxp
[ctxn
] = next_ctx
;
2015 next
->perf_event_ctxp
[ctxn
] = ctx
;
2017 next_ctx
->task
= task
;
2020 perf_event_sync_stat(ctx
, next_ctx
);
2022 raw_spin_unlock(&next_ctx
->lock
);
2023 raw_spin_unlock(&ctx
->lock
);
2028 raw_spin_lock(&ctx
->lock
);
2029 ctx_sched_out(ctx
, cpuctx
, EVENT_ALL
);
2030 cpuctx
->task_ctx
= NULL
;
2031 raw_spin_unlock(&ctx
->lock
);
2035 #define for_each_task_context_nr(ctxn) \
2036 for ((ctxn) = 0; (ctxn) < perf_nr_task_contexts; (ctxn)++)
2039 * Called from scheduler to remove the events of the current task,
2040 * with interrupts disabled.
2042 * We stop each event and update the event value in event->count.
2044 * This does not protect us against NMI, but disable()
2045 * sets the disabled bit in the control field of event _before_
2046 * accessing the event control register. If a NMI hits, then it will
2047 * not restart the event.
2049 void __perf_event_task_sched_out(struct task_struct
*task
,
2050 struct task_struct
*next
)
2054 for_each_task_context_nr(ctxn
)
2055 perf_event_context_sched_out(task
, ctxn
, next
);
2058 * if cgroup events exist on this CPU, then we need
2059 * to check if we have to switch out PMU state.
2060 * cgroup event are system-wide mode only
2062 if (atomic_read(&__get_cpu_var(perf_cgroup_events
)))
2063 perf_cgroup_sched_out(task
, next
);
2066 static void task_ctx_sched_out(struct perf_event_context
*ctx
)
2068 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
2070 if (!cpuctx
->task_ctx
)
2073 if (WARN_ON_ONCE(ctx
!= cpuctx
->task_ctx
))
2076 ctx_sched_out(ctx
, cpuctx
, EVENT_ALL
);
2077 cpuctx
->task_ctx
= NULL
;
2081 * Called with IRQs disabled
2083 static void cpu_ctx_sched_out(struct perf_cpu_context
*cpuctx
,
2084 enum event_type_t event_type
)
2086 ctx_sched_out(&cpuctx
->ctx
, cpuctx
, event_type
);
2090 ctx_pinned_sched_in(struct perf_event_context
*ctx
,
2091 struct perf_cpu_context
*cpuctx
)
2093 struct perf_event
*event
;
2095 list_for_each_entry(event
, &ctx
->pinned_groups
, group_entry
) {
2096 if (event
->state
<= PERF_EVENT_STATE_OFF
)
2098 if (!event_filter_match(event
))
2101 /* may need to reset tstamp_enabled */
2102 if (is_cgroup_event(event
))
2103 perf_cgroup_mark_enabled(event
, ctx
);
2105 if (group_can_go_on(event
, cpuctx
, 1))
2106 group_sched_in(event
, cpuctx
, ctx
);
2109 * If this pinned group hasn't been scheduled,
2110 * put it in error state.
2112 if (event
->state
== PERF_EVENT_STATE_INACTIVE
) {
2113 update_group_times(event
);
2114 event
->state
= PERF_EVENT_STATE_ERROR
;
2120 ctx_flexible_sched_in(struct perf_event_context
*ctx
,
2121 struct perf_cpu_context
*cpuctx
)
2123 struct perf_event
*event
;
2126 list_for_each_entry(event
, &ctx
->flexible_groups
, group_entry
) {
2127 /* Ignore events in OFF or ERROR state */
2128 if (event
->state
<= PERF_EVENT_STATE_OFF
)
2131 * Listen to the 'cpu' scheduling filter constraint
2134 if (!event_filter_match(event
))
2137 /* may need to reset tstamp_enabled */
2138 if (is_cgroup_event(event
))
2139 perf_cgroup_mark_enabled(event
, ctx
);
2141 if (group_can_go_on(event
, cpuctx
, can_add_hw
)) {
2142 if (group_sched_in(event
, cpuctx
, ctx
))
2149 ctx_sched_in(struct perf_event_context
*ctx
,
2150 struct perf_cpu_context
*cpuctx
,
2151 enum event_type_t event_type
,
2152 struct task_struct
*task
)
2155 int is_active
= ctx
->is_active
;
2157 ctx
->is_active
|= event_type
;
2158 if (likely(!ctx
->nr_events
))
2162 ctx
->timestamp
= now
;
2163 perf_cgroup_set_timestamp(task
, ctx
);
2165 * First go through the list and put on any pinned groups
2166 * in order to give them the best chance of going on.
2168 if (!(is_active
& EVENT_PINNED
) && (event_type
& EVENT_PINNED
))
2169 ctx_pinned_sched_in(ctx
, cpuctx
);
2171 /* Then walk through the lower prio flexible groups */
2172 if (!(is_active
& EVENT_FLEXIBLE
) && (event_type
& EVENT_FLEXIBLE
))
2173 ctx_flexible_sched_in(ctx
, cpuctx
);
2176 static void cpu_ctx_sched_in(struct perf_cpu_context
*cpuctx
,
2177 enum event_type_t event_type
,
2178 struct task_struct
*task
)
2180 struct perf_event_context
*ctx
= &cpuctx
->ctx
;
2182 ctx_sched_in(ctx
, cpuctx
, event_type
, task
);
2185 static void perf_event_context_sched_in(struct perf_event_context
*ctx
,
2186 struct task_struct
*task
)
2188 struct perf_cpu_context
*cpuctx
;
2190 cpuctx
= __get_cpu_context(ctx
);
2191 if (cpuctx
->task_ctx
== ctx
)
2194 perf_ctx_lock(cpuctx
, ctx
);
2195 perf_pmu_disable(ctx
->pmu
);
2197 * We want to keep the following priority order:
2198 * cpu pinned (that don't need to move), task pinned,
2199 * cpu flexible, task flexible.
2201 cpu_ctx_sched_out(cpuctx
, EVENT_FLEXIBLE
);
2204 cpuctx
->task_ctx
= ctx
;
2206 perf_event_sched_in(cpuctx
, cpuctx
->task_ctx
, task
);
2208 perf_pmu_enable(ctx
->pmu
);
2209 perf_ctx_unlock(cpuctx
, ctx
);
2212 * Since these rotations are per-cpu, we need to ensure the
2213 * cpu-context we got scheduled on is actually rotating.
2215 perf_pmu_rotate_start(ctx
->pmu
);
2219 * When sampling the branck stack in system-wide, it may be necessary
2220 * to flush the stack on context switch. This happens when the branch
2221 * stack does not tag its entries with the pid of the current task.
2222 * Otherwise it becomes impossible to associate a branch entry with a
2223 * task. This ambiguity is more likely to appear when the branch stack
2224 * supports priv level filtering and the user sets it to monitor only
2225 * at the user level (which could be a useful measurement in system-wide
2226 * mode). In that case, the risk is high of having a branch stack with
2227 * branch from multiple tasks. Flushing may mean dropping the existing
2228 * entries or stashing them somewhere in the PMU specific code layer.
2230 * This function provides the context switch callback to the lower code
2231 * layer. It is invoked ONLY when there is at least one system-wide context
2232 * with at least one active event using taken branch sampling.
2234 static void perf_branch_stack_sched_in(struct task_struct
*prev
,
2235 struct task_struct
*task
)
2237 struct perf_cpu_context
*cpuctx
;
2239 unsigned long flags
;
2241 /* no need to flush branch stack if not changing task */
2245 local_irq_save(flags
);
2249 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
2250 cpuctx
= this_cpu_ptr(pmu
->pmu_cpu_context
);
2253 * check if the context has at least one
2254 * event using PERF_SAMPLE_BRANCH_STACK
2256 if (cpuctx
->ctx
.nr_branch_stack
> 0
2257 && pmu
->flush_branch_stack
) {
2259 pmu
= cpuctx
->ctx
.pmu
;
2261 perf_ctx_lock(cpuctx
, cpuctx
->task_ctx
);
2263 perf_pmu_disable(pmu
);
2265 pmu
->flush_branch_stack();
2267 perf_pmu_enable(pmu
);
2269 perf_ctx_unlock(cpuctx
, cpuctx
->task_ctx
);
2275 local_irq_restore(flags
);
2279 * Called from scheduler to add the events of the current task
2280 * with interrupts disabled.
2282 * We restore the event value and then enable it.
2284 * This does not protect us against NMI, but enable()
2285 * sets the enabled bit in the control field of event _before_
2286 * accessing the event control register. If a NMI hits, then it will
2287 * keep the event running.
2289 void __perf_event_task_sched_in(struct task_struct
*prev
,
2290 struct task_struct
*task
)
2292 struct perf_event_context
*ctx
;
2295 for_each_task_context_nr(ctxn
) {
2296 ctx
= task
->perf_event_ctxp
[ctxn
];
2300 perf_event_context_sched_in(ctx
, task
);
2303 * if cgroup events exist on this CPU, then we need
2304 * to check if we have to switch in PMU state.
2305 * cgroup event are system-wide mode only
2307 if (atomic_read(&__get_cpu_var(perf_cgroup_events
)))
2308 perf_cgroup_sched_in(prev
, task
);
2310 /* check for system-wide branch_stack events */
2311 if (atomic_read(&__get_cpu_var(perf_branch_stack_events
)))
2312 perf_branch_stack_sched_in(prev
, task
);
2315 static u64
perf_calculate_period(struct perf_event
*event
, u64 nsec
, u64 count
)
2317 u64 frequency
= event
->attr
.sample_freq
;
2318 u64 sec
= NSEC_PER_SEC
;
2319 u64 divisor
, dividend
;
2321 int count_fls
, nsec_fls
, frequency_fls
, sec_fls
;
2323 count_fls
= fls64(count
);
2324 nsec_fls
= fls64(nsec
);
2325 frequency_fls
= fls64(frequency
);
2329 * We got @count in @nsec, with a target of sample_freq HZ
2330 * the target period becomes:
2333 * period = -------------------
2334 * @nsec * sample_freq
2339 * Reduce accuracy by one bit such that @a and @b converge
2340 * to a similar magnitude.
2342 #define REDUCE_FLS(a, b) \
2344 if (a##_fls > b##_fls) { \
2354 * Reduce accuracy until either term fits in a u64, then proceed with
2355 * the other, so that finally we can do a u64/u64 division.
2357 while (count_fls
+ sec_fls
> 64 && nsec_fls
+ frequency_fls
> 64) {
2358 REDUCE_FLS(nsec
, frequency
);
2359 REDUCE_FLS(sec
, count
);
2362 if (count_fls
+ sec_fls
> 64) {
2363 divisor
= nsec
* frequency
;
2365 while (count_fls
+ sec_fls
> 64) {
2366 REDUCE_FLS(count
, sec
);
2370 dividend
= count
* sec
;
2372 dividend
= count
* sec
;
2374 while (nsec_fls
+ frequency_fls
> 64) {
2375 REDUCE_FLS(nsec
, frequency
);
2379 divisor
= nsec
* frequency
;
2385 return div64_u64(dividend
, divisor
);
2388 static DEFINE_PER_CPU(int, perf_throttled_count
);
2389 static DEFINE_PER_CPU(u64
, perf_throttled_seq
);
2391 static void perf_adjust_period(struct perf_event
*event
, u64 nsec
, u64 count
, bool disable
)
2393 struct hw_perf_event
*hwc
= &event
->hw
;
2394 s64 period
, sample_period
;
2397 period
= perf_calculate_period(event
, nsec
, count
);
2399 delta
= (s64
)(period
- hwc
->sample_period
);
2400 delta
= (delta
+ 7) / 8; /* low pass filter */
2402 sample_period
= hwc
->sample_period
+ delta
;
2407 hwc
->sample_period
= sample_period
;
2409 if (local64_read(&hwc
->period_left
) > 8*sample_period
) {
2411 event
->pmu
->stop(event
, PERF_EF_UPDATE
);
2413 local64_set(&hwc
->period_left
, 0);
2416 event
->pmu
->start(event
, PERF_EF_RELOAD
);
2421 * combine freq adjustment with unthrottling to avoid two passes over the
2422 * events. At the same time, make sure, having freq events does not change
2423 * the rate of unthrottling as that would introduce bias.
2425 static void perf_adjust_freq_unthr_context(struct perf_event_context
*ctx
,
2428 struct perf_event
*event
;
2429 struct hw_perf_event
*hwc
;
2430 u64 now
, period
= TICK_NSEC
;
2434 * only need to iterate over all events iff:
2435 * - context have events in frequency mode (needs freq adjust)
2436 * - there are events to unthrottle on this cpu
2438 if (!(ctx
->nr_freq
|| needs_unthr
))
2441 raw_spin_lock(&ctx
->lock
);
2442 perf_pmu_disable(ctx
->pmu
);
2444 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
2445 if (event
->state
!= PERF_EVENT_STATE_ACTIVE
)
2448 if (!event_filter_match(event
))
2453 if (needs_unthr
&& hwc
->interrupts
== MAX_INTERRUPTS
) {
2454 hwc
->interrupts
= 0;
2455 perf_log_throttle(event
, 1);
2456 event
->pmu
->start(event
, 0);
2459 if (!event
->attr
.freq
|| !event
->attr
.sample_freq
)
2463 * stop the event and update event->count
2465 event
->pmu
->stop(event
, PERF_EF_UPDATE
);
2467 now
= local64_read(&event
->count
);
2468 delta
= now
- hwc
->freq_count_stamp
;
2469 hwc
->freq_count_stamp
= now
;
2473 * reload only if value has changed
2474 * we have stopped the event so tell that
2475 * to perf_adjust_period() to avoid stopping it
2479 perf_adjust_period(event
, period
, delta
, false);
2481 event
->pmu
->start(event
, delta
> 0 ? PERF_EF_RELOAD
: 0);
2484 perf_pmu_enable(ctx
->pmu
);
2485 raw_spin_unlock(&ctx
->lock
);
2489 * Round-robin a context's events:
2491 static void rotate_ctx(struct perf_event_context
*ctx
)
2494 * Rotate the first entry last of non-pinned groups. Rotation might be
2495 * disabled by the inheritance code.
2497 if (!ctx
->rotate_disable
)
2498 list_rotate_left(&ctx
->flexible_groups
);
2502 * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
2503 * because they're strictly cpu affine and rotate_start is called with IRQs
2504 * disabled, while rotate_context is called from IRQ context.
2506 static void perf_rotate_context(struct perf_cpu_context
*cpuctx
)
2508 struct perf_event_context
*ctx
= NULL
;
2509 int rotate
= 0, remove
= 1;
2511 if (cpuctx
->ctx
.nr_events
) {
2513 if (cpuctx
->ctx
.nr_events
!= cpuctx
->ctx
.nr_active
)
2517 ctx
= cpuctx
->task_ctx
;
2518 if (ctx
&& ctx
->nr_events
) {
2520 if (ctx
->nr_events
!= ctx
->nr_active
)
2527 perf_ctx_lock(cpuctx
, cpuctx
->task_ctx
);
2528 perf_pmu_disable(cpuctx
->ctx
.pmu
);
2530 cpu_ctx_sched_out(cpuctx
, EVENT_FLEXIBLE
);
2532 ctx_sched_out(ctx
, cpuctx
, EVENT_FLEXIBLE
);
2534 rotate_ctx(&cpuctx
->ctx
);
2538 perf_event_sched_in(cpuctx
, ctx
, current
);
2540 perf_pmu_enable(cpuctx
->ctx
.pmu
);
2541 perf_ctx_unlock(cpuctx
, cpuctx
->task_ctx
);
2544 list_del_init(&cpuctx
->rotation_list
);
2547 void perf_event_task_tick(void)
2549 struct list_head
*head
= &__get_cpu_var(rotation_list
);
2550 struct perf_cpu_context
*cpuctx
, *tmp
;
2551 struct perf_event_context
*ctx
;
2554 WARN_ON(!irqs_disabled());
2556 __this_cpu_inc(perf_throttled_seq
);
2557 throttled
= __this_cpu_xchg(perf_throttled_count
, 0);
2559 list_for_each_entry_safe(cpuctx
, tmp
, head
, rotation_list
) {
2561 perf_adjust_freq_unthr_context(ctx
, throttled
);
2563 ctx
= cpuctx
->task_ctx
;
2565 perf_adjust_freq_unthr_context(ctx
, throttled
);
2567 if (cpuctx
->jiffies_interval
== 1 ||
2568 !(jiffies
% cpuctx
->jiffies_interval
))
2569 perf_rotate_context(cpuctx
);
2573 static int event_enable_on_exec(struct perf_event
*event
,
2574 struct perf_event_context
*ctx
)
2576 if (!event
->attr
.enable_on_exec
)
2579 event
->attr
.enable_on_exec
= 0;
2580 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
)
2583 __perf_event_mark_enabled(event
);
2589 * Enable all of a task's events that have been marked enable-on-exec.
2590 * This expects task == current.
2592 static void perf_event_enable_on_exec(struct perf_event_context
*ctx
)
2594 struct perf_event
*event
;
2595 unsigned long flags
;
2599 local_irq_save(flags
);
2600 if (!ctx
|| !ctx
->nr_events
)
2604 * We must ctxsw out cgroup events to avoid conflict
2605 * when invoking perf_task_event_sched_in() later on
2606 * in this function. Otherwise we end up trying to
2607 * ctxswin cgroup events which are already scheduled
2610 perf_cgroup_sched_out(current
, NULL
);
2612 raw_spin_lock(&ctx
->lock
);
2613 task_ctx_sched_out(ctx
);
2615 list_for_each_entry(event
, &ctx
->event_list
, event_entry
) {
2616 ret
= event_enable_on_exec(event
, ctx
);
2622 * Unclone this context if we enabled any event.
2627 raw_spin_unlock(&ctx
->lock
);
2630 * Also calls ctxswin for cgroup events, if any:
2632 perf_event_context_sched_in(ctx
, ctx
->task
);
2634 local_irq_restore(flags
);
2638 * Cross CPU call to read the hardware event
2640 static void __perf_event_read(void *info
)
2642 struct perf_event
*event
= info
;
2643 struct perf_event_context
*ctx
= event
->ctx
;
2644 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
2647 * If this is a task context, we need to check whether it is
2648 * the current task context of this cpu. If not it has been
2649 * scheduled out before the smp call arrived. In that case
2650 * event->count would have been updated to a recent sample
2651 * when the event was scheduled out.
2653 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
)
2656 raw_spin_lock(&ctx
->lock
);
2657 if (ctx
->is_active
) {
2658 update_context_time(ctx
);
2659 update_cgrp_time_from_event(event
);
2661 update_event_times(event
);
2662 if (event
->state
== PERF_EVENT_STATE_ACTIVE
)
2663 event
->pmu
->read(event
);
2664 raw_spin_unlock(&ctx
->lock
);
2667 static inline u64
perf_event_count(struct perf_event
*event
)
2669 return local64_read(&event
->count
) + atomic64_read(&event
->child_count
);
2672 static u64
perf_event_read(struct perf_event
*event
)
2675 * If event is enabled and currently active on a CPU, update the
2676 * value in the event structure:
2678 if (event
->state
== PERF_EVENT_STATE_ACTIVE
) {
2679 smp_call_function_single(event
->oncpu
,
2680 __perf_event_read
, event
, 1);
2681 } else if (event
->state
== PERF_EVENT_STATE_INACTIVE
) {
2682 struct perf_event_context
*ctx
= event
->ctx
;
2683 unsigned long flags
;
2685 raw_spin_lock_irqsave(&ctx
->lock
, flags
);
2687 * may read while context is not active
2688 * (e.g., thread is blocked), in that case
2689 * we cannot update context time
2691 if (ctx
->is_active
) {
2692 update_context_time(ctx
);
2693 update_cgrp_time_from_event(event
);
2695 update_event_times(event
);
2696 raw_spin_unlock_irqrestore(&ctx
->lock
, flags
);
2699 return perf_event_count(event
);
2703 * Initialize the perf_event context in a task_struct:
2705 static void __perf_event_init_context(struct perf_event_context
*ctx
)
2707 raw_spin_lock_init(&ctx
->lock
);
2708 mutex_init(&ctx
->mutex
);
2709 INIT_LIST_HEAD(&ctx
->pinned_groups
);
2710 INIT_LIST_HEAD(&ctx
->flexible_groups
);
2711 INIT_LIST_HEAD(&ctx
->event_list
);
2712 atomic_set(&ctx
->refcount
, 1);
2715 static struct perf_event_context
*
2716 alloc_perf_context(struct pmu
*pmu
, struct task_struct
*task
)
2718 struct perf_event_context
*ctx
;
2720 ctx
= kzalloc(sizeof(struct perf_event_context
), GFP_KERNEL
);
2724 __perf_event_init_context(ctx
);
2727 get_task_struct(task
);
2734 static struct task_struct
*
2735 find_lively_task_by_vpid(pid_t vpid
)
2737 struct task_struct
*task
;
2744 task
= find_task_by_vpid(vpid
);
2746 get_task_struct(task
);
2750 return ERR_PTR(-ESRCH
);
2752 /* Reuse ptrace permission checks for now. */
2754 if (!ptrace_may_access(task
, PTRACE_MODE_READ
))
2759 put_task_struct(task
);
2760 return ERR_PTR(err
);
2765 * Returns a matching context with refcount and pincount.
2767 static struct perf_event_context
*
2768 find_get_context(struct pmu
*pmu
, struct task_struct
*task
, int cpu
)
2770 struct perf_event_context
*ctx
;
2771 struct perf_cpu_context
*cpuctx
;
2772 unsigned long flags
;
2776 /* Must be root to operate on a CPU event: */
2777 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN
))
2778 return ERR_PTR(-EACCES
);
2781 * We could be clever and allow to attach a event to an
2782 * offline CPU and activate it when the CPU comes up, but
2785 if (!cpu_online(cpu
))
2786 return ERR_PTR(-ENODEV
);
2788 cpuctx
= per_cpu_ptr(pmu
->pmu_cpu_context
, cpu
);
2797 ctxn
= pmu
->task_ctx_nr
;
2802 ctx
= perf_lock_task_context(task
, ctxn
, &flags
);
2806 raw_spin_unlock_irqrestore(&ctx
->lock
, flags
);
2808 ctx
= alloc_perf_context(pmu
, task
);
2814 mutex_lock(&task
->perf_event_mutex
);
2816 * If it has already passed perf_event_exit_task().
2817 * we must see PF_EXITING, it takes this mutex too.
2819 if (task
->flags
& PF_EXITING
)
2821 else if (task
->perf_event_ctxp
[ctxn
])
2826 rcu_assign_pointer(task
->perf_event_ctxp
[ctxn
], ctx
);
2828 mutex_unlock(&task
->perf_event_mutex
);
2830 if (unlikely(err
)) {
2842 return ERR_PTR(err
);
2845 static void perf_event_free_filter(struct perf_event
*event
);
2847 static void free_event_rcu(struct rcu_head
*head
)
2849 struct perf_event
*event
;
2851 event
= container_of(head
, struct perf_event
, rcu_head
);
2853 put_pid_ns(event
->ns
);
2854 perf_event_free_filter(event
);
2858 static void ring_buffer_put(struct ring_buffer
*rb
);
2860 static void free_event(struct perf_event
*event
)
2862 irq_work_sync(&event
->pending
);
2864 if (!event
->parent
) {
2865 if (event
->attach_state
& PERF_ATTACH_TASK
)
2866 static_key_slow_dec_deferred(&perf_sched_events
);
2867 if (event
->attr
.mmap
|| event
->attr
.mmap_data
)
2868 atomic_dec(&nr_mmap_events
);
2869 if (event
->attr
.comm
)
2870 atomic_dec(&nr_comm_events
);
2871 if (event
->attr
.task
)
2872 atomic_dec(&nr_task_events
);
2873 if (event
->attr
.sample_type
& PERF_SAMPLE_CALLCHAIN
)
2874 put_callchain_buffers();
2875 if (is_cgroup_event(event
)) {
2876 atomic_dec(&per_cpu(perf_cgroup_events
, event
->cpu
));
2877 static_key_slow_dec_deferred(&perf_sched_events
);
2880 if (has_branch_stack(event
)) {
2881 static_key_slow_dec_deferred(&perf_sched_events
);
2882 /* is system-wide event */
2883 if (!(event
->attach_state
& PERF_ATTACH_TASK
))
2884 atomic_dec(&per_cpu(perf_branch_stack_events
,
2890 ring_buffer_put(event
->rb
);
2894 if (is_cgroup_event(event
))
2895 perf_detach_cgroup(event
);
2898 event
->destroy(event
);
2901 put_ctx(event
->ctx
);
2903 call_rcu(&event
->rcu_head
, free_event_rcu
);
2906 int perf_event_release_kernel(struct perf_event
*event
)
2908 struct perf_event_context
*ctx
= event
->ctx
;
2910 WARN_ON_ONCE(ctx
->parent_ctx
);
2912 * There are two ways this annotation is useful:
2914 * 1) there is a lock recursion from perf_event_exit_task
2915 * see the comment there.
2917 * 2) there is a lock-inversion with mmap_sem through
2918 * perf_event_read_group(), which takes faults while
2919 * holding ctx->mutex, however this is called after
2920 * the last filedesc died, so there is no possibility
2921 * to trigger the AB-BA case.
2923 mutex_lock_nested(&ctx
->mutex
, SINGLE_DEPTH_NESTING
);
2924 raw_spin_lock_irq(&ctx
->lock
);
2925 perf_group_detach(event
);
2926 raw_spin_unlock_irq(&ctx
->lock
);
2927 perf_remove_from_context(event
);
2928 mutex_unlock(&ctx
->mutex
);
2934 EXPORT_SYMBOL_GPL(perf_event_release_kernel
);
2937 * Called when the last reference to the file is gone.
2939 static void put_event(struct perf_event
*event
)
2941 struct task_struct
*owner
;
2943 if (!atomic_long_dec_and_test(&event
->refcount
))
2947 owner
= ACCESS_ONCE(event
->owner
);
2949 * Matches the smp_wmb() in perf_event_exit_task(). If we observe
2950 * !owner it means the list deletion is complete and we can indeed
2951 * free this event, otherwise we need to serialize on
2952 * owner->perf_event_mutex.
2954 smp_read_barrier_depends();
2957 * Since delayed_put_task_struct() also drops the last
2958 * task reference we can safely take a new reference
2959 * while holding the rcu_read_lock().
2961 get_task_struct(owner
);
2966 mutex_lock(&owner
->perf_event_mutex
);
2968 * We have to re-check the event->owner field, if it is cleared
2969 * we raced with perf_event_exit_task(), acquiring the mutex
2970 * ensured they're done, and we can proceed with freeing the
2974 list_del_init(&event
->owner_entry
);
2975 mutex_unlock(&owner
->perf_event_mutex
);
2976 put_task_struct(owner
);
2979 perf_event_release_kernel(event
);
2982 static int perf_release(struct inode
*inode
, struct file
*file
)
2984 put_event(file
->private_data
);
2988 u64
perf_event_read_value(struct perf_event
*event
, u64
*enabled
, u64
*running
)
2990 struct perf_event
*child
;
2996 mutex_lock(&event
->child_mutex
);
2997 total
+= perf_event_read(event
);
2998 *enabled
+= event
->total_time_enabled
+
2999 atomic64_read(&event
->child_total_time_enabled
);
3000 *running
+= event
->total_time_running
+
3001 atomic64_read(&event
->child_total_time_running
);
3003 list_for_each_entry(child
, &event
->child_list
, child_list
) {
3004 total
+= perf_event_read(child
);
3005 *enabled
+= child
->total_time_enabled
;
3006 *running
+= child
->total_time_running
;
3008 mutex_unlock(&event
->child_mutex
);
3012 EXPORT_SYMBOL_GPL(perf_event_read_value
);
3014 static int perf_event_read_group(struct perf_event
*event
,
3015 u64 read_format
, char __user
*buf
)
3017 struct perf_event
*leader
= event
->group_leader
, *sub
;
3018 int n
= 0, size
= 0, ret
= -EFAULT
;
3019 struct perf_event_context
*ctx
= leader
->ctx
;
3021 u64 count
, enabled
, running
;
3023 mutex_lock(&ctx
->mutex
);
3024 count
= perf_event_read_value(leader
, &enabled
, &running
);
3026 values
[n
++] = 1 + leader
->nr_siblings
;
3027 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
3028 values
[n
++] = enabled
;
3029 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
3030 values
[n
++] = running
;
3031 values
[n
++] = count
;
3032 if (read_format
& PERF_FORMAT_ID
)
3033 values
[n
++] = primary_event_id(leader
);
3035 size
= n
* sizeof(u64
);
3037 if (copy_to_user(buf
, values
, size
))
3042 list_for_each_entry(sub
, &leader
->sibling_list
, group_entry
) {
3045 values
[n
++] = perf_event_read_value(sub
, &enabled
, &running
);
3046 if (read_format
& PERF_FORMAT_ID
)
3047 values
[n
++] = primary_event_id(sub
);
3049 size
= n
* sizeof(u64
);
3051 if (copy_to_user(buf
+ ret
, values
, size
)) {
3059 mutex_unlock(&ctx
->mutex
);
3064 static int perf_event_read_one(struct perf_event
*event
,
3065 u64 read_format
, char __user
*buf
)
3067 u64 enabled
, running
;
3071 values
[n
++] = perf_event_read_value(event
, &enabled
, &running
);
3072 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
3073 values
[n
++] = enabled
;
3074 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
3075 values
[n
++] = running
;
3076 if (read_format
& PERF_FORMAT_ID
)
3077 values
[n
++] = primary_event_id(event
);
3079 if (copy_to_user(buf
, values
, n
* sizeof(u64
)))
3082 return n
* sizeof(u64
);
3086 * Read the performance event - simple non blocking version for now
3089 perf_read_hw(struct perf_event
*event
, char __user
*buf
, size_t count
)
3091 u64 read_format
= event
->attr
.read_format
;
3095 * Return end-of-file for a read on a event that is in
3096 * error state (i.e. because it was pinned but it couldn't be
3097 * scheduled on to the CPU at some point).
3099 if (event
->state
== PERF_EVENT_STATE_ERROR
)
3102 if (count
< event
->read_size
)
3105 WARN_ON_ONCE(event
->ctx
->parent_ctx
);
3106 if (read_format
& PERF_FORMAT_GROUP
)
3107 ret
= perf_event_read_group(event
, read_format
, buf
);
3109 ret
= perf_event_read_one(event
, read_format
, buf
);
3115 perf_read(struct file
*file
, char __user
*buf
, size_t count
, loff_t
*ppos
)
3117 struct perf_event
*event
= file
->private_data
;
3119 return perf_read_hw(event
, buf
, count
);
3122 static unsigned int perf_poll(struct file
*file
, poll_table
*wait
)
3124 struct perf_event
*event
= file
->private_data
;
3125 struct ring_buffer
*rb
;
3126 unsigned int events
= POLL_HUP
;
3129 * Race between perf_event_set_output() and perf_poll(): perf_poll()
3130 * grabs the rb reference but perf_event_set_output() overrides it.
3131 * Here is the timeline for two threads T1, T2:
3132 * t0: T1, rb = rcu_dereference(event->rb)
3133 * t1: T2, old_rb = event->rb
3134 * t2: T2, event->rb = new rb
3135 * t3: T2, ring_buffer_detach(old_rb)
3136 * t4: T1, ring_buffer_attach(rb1)
3137 * t5: T1, poll_wait(event->waitq)
3139 * To avoid this problem, we grab mmap_mutex in perf_poll()
3140 * thereby ensuring that the assignment of the new ring buffer
3141 * and the detachment of the old buffer appear atomic to perf_poll()
3143 mutex_lock(&event
->mmap_mutex
);
3146 rb
= rcu_dereference(event
->rb
);
3148 ring_buffer_attach(event
, rb
);
3149 events
= atomic_xchg(&rb
->poll
, 0);
3153 mutex_unlock(&event
->mmap_mutex
);
3155 poll_wait(file
, &event
->waitq
, wait
);
3160 static void perf_event_reset(struct perf_event
*event
)
3162 (void)perf_event_read(event
);
3163 local64_set(&event
->count
, 0);
3164 perf_event_update_userpage(event
);
3168 * Holding the top-level event's child_mutex means that any
3169 * descendant process that has inherited this event will block
3170 * in sync_child_event if it goes to exit, thus satisfying the
3171 * task existence requirements of perf_event_enable/disable.
3173 static void perf_event_for_each_child(struct perf_event
*event
,
3174 void (*func
)(struct perf_event
*))
3176 struct perf_event
*child
;
3178 WARN_ON_ONCE(event
->ctx
->parent_ctx
);
3179 mutex_lock(&event
->child_mutex
);
3181 list_for_each_entry(child
, &event
->child_list
, child_list
)
3183 mutex_unlock(&event
->child_mutex
);
3186 static void perf_event_for_each(struct perf_event
*event
,
3187 void (*func
)(struct perf_event
*))
3189 struct perf_event_context
*ctx
= event
->ctx
;
3190 struct perf_event
*sibling
;
3192 WARN_ON_ONCE(ctx
->parent_ctx
);
3193 mutex_lock(&ctx
->mutex
);
3194 event
= event
->group_leader
;
3196 perf_event_for_each_child(event
, func
);
3197 list_for_each_entry(sibling
, &event
->sibling_list
, group_entry
)
3198 perf_event_for_each_child(sibling
, func
);
3199 mutex_unlock(&ctx
->mutex
);
3202 static int perf_event_period(struct perf_event
*event
, u64 __user
*arg
)
3204 struct perf_event_context
*ctx
= event
->ctx
;
3208 if (!is_sampling_event(event
))
3211 if (copy_from_user(&value
, arg
, sizeof(value
)))
3217 raw_spin_lock_irq(&ctx
->lock
);
3218 if (event
->attr
.freq
) {
3219 if (value
> sysctl_perf_event_sample_rate
) {
3224 event
->attr
.sample_freq
= value
;
3226 event
->attr
.sample_period
= value
;
3227 event
->hw
.sample_period
= value
;
3230 raw_spin_unlock_irq(&ctx
->lock
);
3235 static const struct file_operations perf_fops
;
3237 static struct file
*perf_fget_light(int fd
, int *fput_needed
)
3241 file
= fget_light(fd
, fput_needed
);
3243 return ERR_PTR(-EBADF
);
3245 if (file
->f_op
!= &perf_fops
) {
3246 fput_light(file
, *fput_needed
);
3248 return ERR_PTR(-EBADF
);
3254 static int perf_event_set_output(struct perf_event
*event
,
3255 struct perf_event
*output_event
);
3256 static int perf_event_set_filter(struct perf_event
*event
, void __user
*arg
);
3258 static long perf_ioctl(struct file
*file
, unsigned int cmd
, unsigned long arg
)
3260 struct perf_event
*event
= file
->private_data
;
3261 void (*func
)(struct perf_event
*);
3265 case PERF_EVENT_IOC_ENABLE
:
3266 func
= perf_event_enable
;
3268 case PERF_EVENT_IOC_DISABLE
:
3269 func
= perf_event_disable
;
3271 case PERF_EVENT_IOC_RESET
:
3272 func
= perf_event_reset
;
3275 case PERF_EVENT_IOC_REFRESH
:
3276 return perf_event_refresh(event
, arg
);
3278 case PERF_EVENT_IOC_PERIOD
:
3279 return perf_event_period(event
, (u64 __user
*)arg
);
3281 case PERF_EVENT_IOC_SET_OUTPUT
:
3283 struct file
*output_file
= NULL
;
3284 struct perf_event
*output_event
= NULL
;
3285 int fput_needed
= 0;
3289 output_file
= perf_fget_light(arg
, &fput_needed
);
3290 if (IS_ERR(output_file
))
3291 return PTR_ERR(output_file
);
3292 output_event
= output_file
->private_data
;
3295 ret
= perf_event_set_output(event
, output_event
);
3297 fput_light(output_file
, fput_needed
);
3302 case PERF_EVENT_IOC_SET_FILTER
:
3303 return perf_event_set_filter(event
, (void __user
*)arg
);
3309 if (flags
& PERF_IOC_FLAG_GROUP
)
3310 perf_event_for_each(event
, func
);
3312 perf_event_for_each_child(event
, func
);
3317 int perf_event_task_enable(void)
3319 struct perf_event
*event
;
3321 mutex_lock(¤t
->perf_event_mutex
);
3322 list_for_each_entry(event
, ¤t
->perf_event_list
, owner_entry
)
3323 perf_event_for_each_child(event
, perf_event_enable
);
3324 mutex_unlock(¤t
->perf_event_mutex
);
3329 int perf_event_task_disable(void)
3331 struct perf_event
*event
;
3333 mutex_lock(¤t
->perf_event_mutex
);
3334 list_for_each_entry(event
, ¤t
->perf_event_list
, owner_entry
)
3335 perf_event_for_each_child(event
, perf_event_disable
);
3336 mutex_unlock(¤t
->perf_event_mutex
);
3341 static int perf_event_index(struct perf_event
*event
)
3343 if (event
->hw
.state
& PERF_HES_STOPPED
)
3346 if (event
->state
!= PERF_EVENT_STATE_ACTIVE
)
3349 return event
->pmu
->event_idx(event
);
3352 static void calc_timer_values(struct perf_event
*event
,
3359 *now
= perf_clock();
3360 ctx_time
= event
->shadow_ctx_time
+ *now
;
3361 *enabled
= ctx_time
- event
->tstamp_enabled
;
3362 *running
= ctx_time
- event
->tstamp_running
;
3365 void __weak
arch_perf_update_userpage(struct perf_event_mmap_page
*userpg
, u64 now
)
3370 * Callers need to ensure there can be no nesting of this function, otherwise
3371 * the seqlock logic goes bad. We can not serialize this because the arch
3372 * code calls this from NMI context.
3374 void perf_event_update_userpage(struct perf_event
*event
)
3376 struct perf_event_mmap_page
*userpg
;
3377 struct ring_buffer
*rb
;
3378 u64 enabled
, running
, now
;
3382 * compute total_time_enabled, total_time_running
3383 * based on snapshot values taken when the event
3384 * was last scheduled in.
3386 * we cannot simply called update_context_time()
3387 * because of locking issue as we can be called in
3390 calc_timer_values(event
, &now
, &enabled
, &running
);
3391 rb
= rcu_dereference(event
->rb
);
3395 userpg
= rb
->user_page
;
3398 * Disable preemption so as to not let the corresponding user-space
3399 * spin too long if we get preempted.
3404 userpg
->index
= perf_event_index(event
);
3405 userpg
->offset
= perf_event_count(event
);
3407 userpg
->offset
-= local64_read(&event
->hw
.prev_count
);
3409 userpg
->time_enabled
= enabled
+
3410 atomic64_read(&event
->child_total_time_enabled
);
3412 userpg
->time_running
= running
+
3413 atomic64_read(&event
->child_total_time_running
);
3415 arch_perf_update_userpage(userpg
, now
);
3424 static int perf_mmap_fault(struct vm_area_struct
*vma
, struct vm_fault
*vmf
)
3426 struct perf_event
*event
= vma
->vm_file
->private_data
;
3427 struct ring_buffer
*rb
;
3428 int ret
= VM_FAULT_SIGBUS
;
3430 if (vmf
->flags
& FAULT_FLAG_MKWRITE
) {
3431 if (vmf
->pgoff
== 0)
3437 rb
= rcu_dereference(event
->rb
);
3441 if (vmf
->pgoff
&& (vmf
->flags
& FAULT_FLAG_WRITE
))
3444 vmf
->page
= perf_mmap_to_page(rb
, vmf
->pgoff
);
3448 get_page(vmf
->page
);
3449 vmf
->page
->mapping
= vma
->vm_file
->f_mapping
;
3450 vmf
->page
->index
= vmf
->pgoff
;
3459 static void ring_buffer_attach(struct perf_event
*event
,
3460 struct ring_buffer
*rb
)
3462 unsigned long flags
;
3464 if (!list_empty(&event
->rb_entry
))
3467 spin_lock_irqsave(&rb
->event_lock
, flags
);
3468 if (!list_empty(&event
->rb_entry
))
3471 list_add(&event
->rb_entry
, &rb
->event_list
);
3473 spin_unlock_irqrestore(&rb
->event_lock
, flags
);
3476 static void ring_buffer_detach(struct perf_event
*event
,
3477 struct ring_buffer
*rb
)
3479 unsigned long flags
;
3481 if (list_empty(&event
->rb_entry
))
3484 spin_lock_irqsave(&rb
->event_lock
, flags
);
3485 list_del_init(&event
->rb_entry
);
3486 wake_up_all(&event
->waitq
);
3487 spin_unlock_irqrestore(&rb
->event_lock
, flags
);
3490 static void ring_buffer_wakeup(struct perf_event
*event
)
3492 struct ring_buffer
*rb
;
3495 rb
= rcu_dereference(event
->rb
);
3499 list_for_each_entry_rcu(event
, &rb
->event_list
, rb_entry
)
3500 wake_up_all(&event
->waitq
);
3506 static void rb_free_rcu(struct rcu_head
*rcu_head
)
3508 struct ring_buffer
*rb
;
3510 rb
= container_of(rcu_head
, struct ring_buffer
, rcu_head
);
3514 static struct ring_buffer
*ring_buffer_get(struct perf_event
*event
)
3516 struct ring_buffer
*rb
;
3519 rb
= rcu_dereference(event
->rb
);
3521 if (!atomic_inc_not_zero(&rb
->refcount
))
3529 static void ring_buffer_put(struct ring_buffer
*rb
)
3531 struct perf_event
*event
, *n
;
3532 unsigned long flags
;
3534 if (!atomic_dec_and_test(&rb
->refcount
))
3537 spin_lock_irqsave(&rb
->event_lock
, flags
);
3538 list_for_each_entry_safe(event
, n
, &rb
->event_list
, rb_entry
) {
3539 list_del_init(&event
->rb_entry
);
3540 wake_up_all(&event
->waitq
);
3542 spin_unlock_irqrestore(&rb
->event_lock
, flags
);
3544 call_rcu(&rb
->rcu_head
, rb_free_rcu
);
3547 static void perf_mmap_open(struct vm_area_struct
*vma
)
3549 struct perf_event
*event
= vma
->vm_file
->private_data
;
3551 atomic_inc(&event
->mmap_count
);
3554 static void perf_mmap_close(struct vm_area_struct
*vma
)
3556 struct perf_event
*event
= vma
->vm_file
->private_data
;
3558 if (atomic_dec_and_mutex_lock(&event
->mmap_count
, &event
->mmap_mutex
)) {
3559 unsigned long size
= perf_data_size(event
->rb
);
3560 struct user_struct
*user
= event
->mmap_user
;
3561 struct ring_buffer
*rb
= event
->rb
;
3563 atomic_long_sub((size
>> PAGE_SHIFT
) + 1, &user
->locked_vm
);
3564 vma
->vm_mm
->pinned_vm
-= event
->mmap_locked
;
3565 rcu_assign_pointer(event
->rb
, NULL
);
3566 ring_buffer_detach(event
, rb
);
3567 mutex_unlock(&event
->mmap_mutex
);
3569 ring_buffer_put(rb
);
3574 static const struct vm_operations_struct perf_mmap_vmops
= {
3575 .open
= perf_mmap_open
,
3576 .close
= perf_mmap_close
,
3577 .fault
= perf_mmap_fault
,
3578 .page_mkwrite
= perf_mmap_fault
,
3581 static int perf_mmap(struct file
*file
, struct vm_area_struct
*vma
)
3583 struct perf_event
*event
= file
->private_data
;
3584 unsigned long user_locked
, user_lock_limit
;
3585 struct user_struct
*user
= current_user();
3586 unsigned long locked
, lock_limit
;
3587 struct ring_buffer
*rb
;
3588 unsigned long vma_size
;
3589 unsigned long nr_pages
;
3590 long user_extra
, extra
;
3591 int ret
= 0, flags
= 0;
3594 * Don't allow mmap() of inherited per-task counters. This would
3595 * create a performance issue due to all children writing to the
3598 if (event
->cpu
== -1 && event
->attr
.inherit
)
3601 if (!(vma
->vm_flags
& VM_SHARED
))
3604 vma_size
= vma
->vm_end
- vma
->vm_start
;
3605 nr_pages
= (vma_size
/ PAGE_SIZE
) - 1;
3608 * If we have rb pages ensure they're a power-of-two number, so we
3609 * can do bitmasks instead of modulo.
3611 if (nr_pages
!= 0 && !is_power_of_2(nr_pages
))
3614 if (vma_size
!= PAGE_SIZE
* (1 + nr_pages
))
3617 if (vma
->vm_pgoff
!= 0)
3620 WARN_ON_ONCE(event
->ctx
->parent_ctx
);
3621 mutex_lock(&event
->mmap_mutex
);
3623 if (event
->rb
->nr_pages
== nr_pages
)
3624 atomic_inc(&event
->rb
->refcount
);
3630 user_extra
= nr_pages
+ 1;
3631 user_lock_limit
= sysctl_perf_event_mlock
>> (PAGE_SHIFT
- 10);
3634 * Increase the limit linearly with more CPUs:
3636 user_lock_limit
*= num_online_cpus();
3638 user_locked
= atomic_long_read(&user
->locked_vm
) + user_extra
;
3641 if (user_locked
> user_lock_limit
)
3642 extra
= user_locked
- user_lock_limit
;
3644 lock_limit
= rlimit(RLIMIT_MEMLOCK
);
3645 lock_limit
>>= PAGE_SHIFT
;
3646 locked
= vma
->vm_mm
->pinned_vm
+ extra
;
3648 if ((locked
> lock_limit
) && perf_paranoid_tracepoint_raw() &&
3649 !capable(CAP_IPC_LOCK
)) {
3656 if (vma
->vm_flags
& VM_WRITE
)
3657 flags
|= RING_BUFFER_WRITABLE
;
3659 rb
= rb_alloc(nr_pages
,
3660 event
->attr
.watermark
? event
->attr
.wakeup_watermark
: 0,
3667 rcu_assign_pointer(event
->rb
, rb
);
3669 atomic_long_add(user_extra
, &user
->locked_vm
);
3670 event
->mmap_locked
= extra
;
3671 event
->mmap_user
= get_current_user();
3672 vma
->vm_mm
->pinned_vm
+= event
->mmap_locked
;
3674 perf_event_update_userpage(event
);
3678 atomic_inc(&event
->mmap_count
);
3679 mutex_unlock(&event
->mmap_mutex
);
3681 vma
->vm_flags
|= VM_RESERVED
;
3682 vma
->vm_ops
= &perf_mmap_vmops
;
3687 static int perf_fasync(int fd
, struct file
*filp
, int on
)
3689 struct inode
*inode
= filp
->f_path
.dentry
->d_inode
;
3690 struct perf_event
*event
= filp
->private_data
;
3693 mutex_lock(&inode
->i_mutex
);
3694 retval
= fasync_helper(fd
, filp
, on
, &event
->fasync
);
3695 mutex_unlock(&inode
->i_mutex
);
3703 static const struct file_operations perf_fops
= {
3704 .llseek
= no_llseek
,
3705 .release
= perf_release
,
3708 .unlocked_ioctl
= perf_ioctl
,
3709 .compat_ioctl
= perf_ioctl
,
3711 .fasync
= perf_fasync
,
3717 * If there's data, ensure we set the poll() state and publish everything
3718 * to user-space before waking everybody up.
3721 void perf_event_wakeup(struct perf_event
*event
)
3723 ring_buffer_wakeup(event
);
3725 if (event
->pending_kill
) {
3726 kill_fasync(&event
->fasync
, SIGIO
, event
->pending_kill
);
3727 event
->pending_kill
= 0;
3731 static void perf_pending_event(struct irq_work
*entry
)
3733 struct perf_event
*event
= container_of(entry
,
3734 struct perf_event
, pending
);
3736 if (event
->pending_disable
) {
3737 event
->pending_disable
= 0;
3738 __perf_event_disable(event
);
3741 if (event
->pending_wakeup
) {
3742 event
->pending_wakeup
= 0;
3743 perf_event_wakeup(event
);
3748 * We assume there is only KVM supporting the callbacks.
3749 * Later on, we might change it to a list if there is
3750 * another virtualization implementation supporting the callbacks.
3752 struct perf_guest_info_callbacks
*perf_guest_cbs
;
3754 int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks
*cbs
)
3756 perf_guest_cbs
= cbs
;
3759 EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks
);
3761 int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks
*cbs
)
3763 perf_guest_cbs
= NULL
;
3766 EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks
);
3769 perf_output_sample_regs(struct perf_output_handle
*handle
,
3770 struct pt_regs
*regs
, u64 mask
)
3774 for_each_set_bit(bit
, (const unsigned long *) &mask
,
3775 sizeof(mask
) * BITS_PER_BYTE
) {
3778 val
= perf_reg_value(regs
, bit
);
3779 perf_output_put(handle
, val
);
3783 static void perf_sample_regs_user(struct perf_regs_user
*regs_user
,
3784 struct pt_regs
*regs
)
3786 if (!user_mode(regs
)) {
3788 regs
= task_pt_regs(current
);
3794 regs_user
->regs
= regs
;
3795 regs_user
->abi
= perf_reg_abi(current
);
3800 * Get remaining task size from user stack pointer.
3802 * It'd be better to take stack vma map and limit this more
3803 * precisly, but there's no way to get it safely under interrupt,
3804 * so using TASK_SIZE as limit.
3806 static u64
perf_ustack_task_size(struct pt_regs
*regs
)
3808 unsigned long addr
= perf_user_stack_pointer(regs
);
3810 if (!addr
|| addr
>= TASK_SIZE
)
3813 return TASK_SIZE
- addr
;
3817 perf_sample_ustack_size(u16 stack_size
, u16 header_size
,
3818 struct pt_regs
*regs
)
3822 /* No regs, no stack pointer, no dump. */
3827 * Check if we fit in with the requested stack size into the:
3829 * If we don't, we limit the size to the TASK_SIZE.
3831 * - remaining sample size
3832 * If we don't, we customize the stack size to
3833 * fit in to the remaining sample size.
3836 task_size
= min((u64
) USHRT_MAX
, perf_ustack_task_size(regs
));
3837 stack_size
= min(stack_size
, (u16
) task_size
);
3839 /* Current header size plus static size and dynamic size. */
3840 header_size
+= 2 * sizeof(u64
);
3842 /* Do we fit in with the current stack dump size? */
3843 if ((u16
) (header_size
+ stack_size
) < header_size
) {
3845 * If we overflow the maximum size for the sample,
3846 * we customize the stack dump size to fit in.
3848 stack_size
= USHRT_MAX
- header_size
- sizeof(u64
);
3849 stack_size
= round_up(stack_size
, sizeof(u64
));
3856 perf_output_sample_ustack(struct perf_output_handle
*handle
, u64 dump_size
,
3857 struct pt_regs
*regs
)
3859 /* Case of a kernel thread, nothing to dump */
3862 perf_output_put(handle
, size
);
3871 * - the size requested by user or the best one we can fit
3872 * in to the sample max size
3874 * - user stack dump data
3876 * - the actual dumped size
3880 perf_output_put(handle
, dump_size
);
3883 sp
= perf_user_stack_pointer(regs
);
3884 rem
= __output_copy_user(handle
, (void *) sp
, dump_size
);
3885 dyn_size
= dump_size
- rem
;
3887 perf_output_skip(handle
, rem
);
3890 perf_output_put(handle
, dyn_size
);
3894 static void __perf_event_header__init_id(struct perf_event_header
*header
,
3895 struct perf_sample_data
*data
,
3896 struct perf_event
*event
)
3898 u64 sample_type
= event
->attr
.sample_type
;
3900 data
->type
= sample_type
;
3901 header
->size
+= event
->id_header_size
;
3903 if (sample_type
& PERF_SAMPLE_TID
) {
3904 /* namespace issues */
3905 data
->tid_entry
.pid
= perf_event_pid(event
, current
);
3906 data
->tid_entry
.tid
= perf_event_tid(event
, current
);
3909 if (sample_type
& PERF_SAMPLE_TIME
)
3910 data
->time
= perf_clock();
3912 if (sample_type
& PERF_SAMPLE_ID
)
3913 data
->id
= primary_event_id(event
);
3915 if (sample_type
& PERF_SAMPLE_STREAM_ID
)
3916 data
->stream_id
= event
->id
;
3918 if (sample_type
& PERF_SAMPLE_CPU
) {
3919 data
->cpu_entry
.cpu
= raw_smp_processor_id();
3920 data
->cpu_entry
.reserved
= 0;
3924 void perf_event_header__init_id(struct perf_event_header
*header
,
3925 struct perf_sample_data
*data
,
3926 struct perf_event
*event
)
3928 if (event
->attr
.sample_id_all
)
3929 __perf_event_header__init_id(header
, data
, event
);
3932 static void __perf_event__output_id_sample(struct perf_output_handle
*handle
,
3933 struct perf_sample_data
*data
)
3935 u64 sample_type
= data
->type
;
3937 if (sample_type
& PERF_SAMPLE_TID
)
3938 perf_output_put(handle
, data
->tid_entry
);
3940 if (sample_type
& PERF_SAMPLE_TIME
)
3941 perf_output_put(handle
, data
->time
);
3943 if (sample_type
& PERF_SAMPLE_ID
)
3944 perf_output_put(handle
, data
->id
);
3946 if (sample_type
& PERF_SAMPLE_STREAM_ID
)
3947 perf_output_put(handle
, data
->stream_id
);
3949 if (sample_type
& PERF_SAMPLE_CPU
)
3950 perf_output_put(handle
, data
->cpu_entry
);
3953 void perf_event__output_id_sample(struct perf_event
*event
,
3954 struct perf_output_handle
*handle
,
3955 struct perf_sample_data
*sample
)
3957 if (event
->attr
.sample_id_all
)
3958 __perf_event__output_id_sample(handle
, sample
);
3961 static void perf_output_read_one(struct perf_output_handle
*handle
,
3962 struct perf_event
*event
,
3963 u64 enabled
, u64 running
)
3965 u64 read_format
= event
->attr
.read_format
;
3969 values
[n
++] = perf_event_count(event
);
3970 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
) {
3971 values
[n
++] = enabled
+
3972 atomic64_read(&event
->child_total_time_enabled
);
3974 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
) {
3975 values
[n
++] = running
+
3976 atomic64_read(&event
->child_total_time_running
);
3978 if (read_format
& PERF_FORMAT_ID
)
3979 values
[n
++] = primary_event_id(event
);
3981 __output_copy(handle
, values
, n
* sizeof(u64
));
3985 * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
3987 static void perf_output_read_group(struct perf_output_handle
*handle
,
3988 struct perf_event
*event
,
3989 u64 enabled
, u64 running
)
3991 struct perf_event
*leader
= event
->group_leader
, *sub
;
3992 u64 read_format
= event
->attr
.read_format
;
3996 values
[n
++] = 1 + leader
->nr_siblings
;
3998 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
3999 values
[n
++] = enabled
;
4001 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
4002 values
[n
++] = running
;
4004 if (leader
!= event
)
4005 leader
->pmu
->read(leader
);
4007 values
[n
++] = perf_event_count(leader
);
4008 if (read_format
& PERF_FORMAT_ID
)
4009 values
[n
++] = primary_event_id(leader
);
4011 __output_copy(handle
, values
, n
* sizeof(u64
));
4013 list_for_each_entry(sub
, &leader
->sibling_list
, group_entry
) {
4017 sub
->pmu
->read(sub
);
4019 values
[n
++] = perf_event_count(sub
);
4020 if (read_format
& PERF_FORMAT_ID
)
4021 values
[n
++] = primary_event_id(sub
);
4023 __output_copy(handle
, values
, n
* sizeof(u64
));
4027 #define PERF_FORMAT_TOTAL_TIMES (PERF_FORMAT_TOTAL_TIME_ENABLED|\
4028 PERF_FORMAT_TOTAL_TIME_RUNNING)
4030 static void perf_output_read(struct perf_output_handle
*handle
,
4031 struct perf_event
*event
)
4033 u64 enabled
= 0, running
= 0, now
;
4034 u64 read_format
= event
->attr
.read_format
;
4037 * compute total_time_enabled, total_time_running
4038 * based on snapshot values taken when the event
4039 * was last scheduled in.
4041 * we cannot simply called update_context_time()
4042 * because of locking issue as we are called in
4045 if (read_format
& PERF_FORMAT_TOTAL_TIMES
)
4046 calc_timer_values(event
, &now
, &enabled
, &running
);
4048 if (event
->attr
.read_format
& PERF_FORMAT_GROUP
)
4049 perf_output_read_group(handle
, event
, enabled
, running
);
4051 perf_output_read_one(handle
, event
, enabled
, running
);
4054 void perf_output_sample(struct perf_output_handle
*handle
,
4055 struct perf_event_header
*header
,
4056 struct perf_sample_data
*data
,
4057 struct perf_event
*event
)
4059 u64 sample_type
= data
->type
;
4061 perf_output_put(handle
, *header
);
4063 if (sample_type
& PERF_SAMPLE_IP
)
4064 perf_output_put(handle
, data
->ip
);
4066 if (sample_type
& PERF_SAMPLE_TID
)
4067 perf_output_put(handle
, data
->tid_entry
);
4069 if (sample_type
& PERF_SAMPLE_TIME
)
4070 perf_output_put(handle
, data
->time
);
4072 if (sample_type
& PERF_SAMPLE_ADDR
)
4073 perf_output_put(handle
, data
->addr
);
4075 if (sample_type
& PERF_SAMPLE_ID
)
4076 perf_output_put(handle
, data
->id
);
4078 if (sample_type
& PERF_SAMPLE_STREAM_ID
)
4079 perf_output_put(handle
, data
->stream_id
);
4081 if (sample_type
& PERF_SAMPLE_CPU
)
4082 perf_output_put(handle
, data
->cpu_entry
);
4084 if (sample_type
& PERF_SAMPLE_PERIOD
)
4085 perf_output_put(handle
, data
->period
);
4087 if (sample_type
& PERF_SAMPLE_READ
)
4088 perf_output_read(handle
, event
);
4090 if (sample_type
& PERF_SAMPLE_CALLCHAIN
) {
4091 if (data
->callchain
) {
4094 if (data
->callchain
)
4095 size
+= data
->callchain
->nr
;
4097 size
*= sizeof(u64
);
4099 __output_copy(handle
, data
->callchain
, size
);
4102 perf_output_put(handle
, nr
);
4106 if (sample_type
& PERF_SAMPLE_RAW
) {
4108 perf_output_put(handle
, data
->raw
->size
);
4109 __output_copy(handle
, data
->raw
->data
,
4116 .size
= sizeof(u32
),
4119 perf_output_put(handle
, raw
);
4123 if (!event
->attr
.watermark
) {
4124 int wakeup_events
= event
->attr
.wakeup_events
;
4126 if (wakeup_events
) {
4127 struct ring_buffer
*rb
= handle
->rb
;
4128 int events
= local_inc_return(&rb
->events
);
4130 if (events
>= wakeup_events
) {
4131 local_sub(wakeup_events
, &rb
->events
);
4132 local_inc(&rb
->wakeup
);
4137 if (sample_type
& PERF_SAMPLE_BRANCH_STACK
) {
4138 if (data
->br_stack
) {
4141 size
= data
->br_stack
->nr
4142 * sizeof(struct perf_branch_entry
);
4144 perf_output_put(handle
, data
->br_stack
->nr
);
4145 perf_output_copy(handle
, data
->br_stack
->entries
, size
);
4148 * we always store at least the value of nr
4151 perf_output_put(handle
, nr
);
4155 if (sample_type
& PERF_SAMPLE_REGS_USER
) {
4156 u64 abi
= data
->regs_user
.abi
;
4159 * If there are no regs to dump, notice it through
4160 * first u64 being zero (PERF_SAMPLE_REGS_ABI_NONE).
4162 perf_output_put(handle
, abi
);
4165 u64 mask
= event
->attr
.sample_regs_user
;
4166 perf_output_sample_regs(handle
,
4167 data
->regs_user
.regs
,
4172 if (sample_type
& PERF_SAMPLE_STACK_USER
)
4173 perf_output_sample_ustack(handle
,
4174 data
->stack_user_size
,
4175 data
->regs_user
.regs
);
4178 void perf_prepare_sample(struct perf_event_header
*header
,
4179 struct perf_sample_data
*data
,
4180 struct perf_event
*event
,
4181 struct pt_regs
*regs
)
4183 u64 sample_type
= event
->attr
.sample_type
;
4185 header
->type
= PERF_RECORD_SAMPLE
;
4186 header
->size
= sizeof(*header
) + event
->header_size
;
4189 header
->misc
|= perf_misc_flags(regs
);
4191 __perf_event_header__init_id(header
, data
, event
);
4193 if (sample_type
& PERF_SAMPLE_IP
)
4194 data
->ip
= perf_instruction_pointer(regs
);
4196 if (sample_type
& PERF_SAMPLE_CALLCHAIN
) {
4199 data
->callchain
= perf_callchain(event
, regs
);
4201 if (data
->callchain
)
4202 size
+= data
->callchain
->nr
;
4204 header
->size
+= size
* sizeof(u64
);
4207 if (sample_type
& PERF_SAMPLE_RAW
) {
4208 int size
= sizeof(u32
);
4211 size
+= data
->raw
->size
;
4213 size
+= sizeof(u32
);
4215 WARN_ON_ONCE(size
& (sizeof(u64
)-1));
4216 header
->size
+= size
;
4219 if (sample_type
& PERF_SAMPLE_BRANCH_STACK
) {
4220 int size
= sizeof(u64
); /* nr */
4221 if (data
->br_stack
) {
4222 size
+= data
->br_stack
->nr
4223 * sizeof(struct perf_branch_entry
);
4225 header
->size
+= size
;
4228 if (sample_type
& PERF_SAMPLE_REGS_USER
) {
4229 /* regs dump ABI info */
4230 int size
= sizeof(u64
);
4232 perf_sample_regs_user(&data
->regs_user
, regs
);
4234 if (data
->regs_user
.regs
) {
4235 u64 mask
= event
->attr
.sample_regs_user
;
4236 size
+= hweight64(mask
) * sizeof(u64
);
4239 header
->size
+= size
;
4242 if (sample_type
& PERF_SAMPLE_STACK_USER
) {
4244 * Either we need PERF_SAMPLE_STACK_USER bit to be allways
4245 * processed as the last one or have additional check added
4246 * in case new sample type is added, because we could eat
4247 * up the rest of the sample size.
4249 struct perf_regs_user
*uregs
= &data
->regs_user
;
4250 u16 stack_size
= event
->attr
.sample_stack_user
;
4251 u16 size
= sizeof(u64
);
4254 perf_sample_regs_user(uregs
, regs
);
4256 stack_size
= perf_sample_ustack_size(stack_size
, header
->size
,
4260 * If there is something to dump, add space for the dump
4261 * itself and for the field that tells the dynamic size,
4262 * which is how many have been actually dumped.
4265 size
+= sizeof(u64
) + stack_size
;
4267 data
->stack_user_size
= stack_size
;
4268 header
->size
+= size
;
4272 static void perf_event_output(struct perf_event
*event
,
4273 struct perf_sample_data
*data
,
4274 struct pt_regs
*regs
)
4276 struct perf_output_handle handle
;
4277 struct perf_event_header header
;
4279 /* protect the callchain buffers */
4282 perf_prepare_sample(&header
, data
, event
, regs
);
4284 if (perf_output_begin(&handle
, event
, header
.size
))
4287 perf_output_sample(&handle
, &header
, data
, event
);
4289 perf_output_end(&handle
);
4299 struct perf_read_event
{
4300 struct perf_event_header header
;
4307 perf_event_read_event(struct perf_event
*event
,
4308 struct task_struct
*task
)
4310 struct perf_output_handle handle
;
4311 struct perf_sample_data sample
;
4312 struct perf_read_event read_event
= {
4314 .type
= PERF_RECORD_READ
,
4316 .size
= sizeof(read_event
) + event
->read_size
,
4318 .pid
= perf_event_pid(event
, task
),
4319 .tid
= perf_event_tid(event
, task
),
4323 perf_event_header__init_id(&read_event
.header
, &sample
, event
);
4324 ret
= perf_output_begin(&handle
, event
, read_event
.header
.size
);
4328 perf_output_put(&handle
, read_event
);
4329 perf_output_read(&handle
, event
);
4330 perf_event__output_id_sample(event
, &handle
, &sample
);
4332 perf_output_end(&handle
);
4336 * task tracking -- fork/exit
4338 * enabled by: attr.comm | attr.mmap | attr.mmap_data | attr.task
4341 struct perf_task_event
{
4342 struct task_struct
*task
;
4343 struct perf_event_context
*task_ctx
;
4346 struct perf_event_header header
;
4356 static void perf_event_task_output(struct perf_event
*event
,
4357 struct perf_task_event
*task_event
)
4359 struct perf_output_handle handle
;
4360 struct perf_sample_data sample
;
4361 struct task_struct
*task
= task_event
->task
;
4362 int ret
, size
= task_event
->event_id
.header
.size
;
4364 perf_event_header__init_id(&task_event
->event_id
.header
, &sample
, event
);
4366 ret
= perf_output_begin(&handle
, event
,
4367 task_event
->event_id
.header
.size
);
4371 task_event
->event_id
.pid
= perf_event_pid(event
, task
);
4372 task_event
->event_id
.ppid
= perf_event_pid(event
, current
);
4374 task_event
->event_id
.tid
= perf_event_tid(event
, task
);
4375 task_event
->event_id
.ptid
= perf_event_tid(event
, current
);
4377 perf_output_put(&handle
, task_event
->event_id
);
4379 perf_event__output_id_sample(event
, &handle
, &sample
);
4381 perf_output_end(&handle
);
4383 task_event
->event_id
.header
.size
= size
;
4386 static int perf_event_task_match(struct perf_event
*event
)
4388 if (event
->state
< PERF_EVENT_STATE_INACTIVE
)
4391 if (!event_filter_match(event
))
4394 if (event
->attr
.comm
|| event
->attr
.mmap
||
4395 event
->attr
.mmap_data
|| event
->attr
.task
)
4401 static void perf_event_task_ctx(struct perf_event_context
*ctx
,
4402 struct perf_task_event
*task_event
)
4404 struct perf_event
*event
;
4406 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
4407 if (perf_event_task_match(event
))
4408 perf_event_task_output(event
, task_event
);
4412 static void perf_event_task_event(struct perf_task_event
*task_event
)
4414 struct perf_cpu_context
*cpuctx
;
4415 struct perf_event_context
*ctx
;
4420 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
4421 cpuctx
= get_cpu_ptr(pmu
->pmu_cpu_context
);
4422 if (cpuctx
->active_pmu
!= pmu
)
4424 perf_event_task_ctx(&cpuctx
->ctx
, task_event
);
4426 ctx
= task_event
->task_ctx
;
4428 ctxn
= pmu
->task_ctx_nr
;
4431 ctx
= rcu_dereference(current
->perf_event_ctxp
[ctxn
]);
4434 perf_event_task_ctx(ctx
, task_event
);
4436 put_cpu_ptr(pmu
->pmu_cpu_context
);
4441 static void perf_event_task(struct task_struct
*task
,
4442 struct perf_event_context
*task_ctx
,
4445 struct perf_task_event task_event
;
4447 if (!atomic_read(&nr_comm_events
) &&
4448 !atomic_read(&nr_mmap_events
) &&
4449 !atomic_read(&nr_task_events
))
4452 task_event
= (struct perf_task_event
){
4454 .task_ctx
= task_ctx
,
4457 .type
= new ? PERF_RECORD_FORK
: PERF_RECORD_EXIT
,
4459 .size
= sizeof(task_event
.event_id
),
4465 .time
= perf_clock(),
4469 perf_event_task_event(&task_event
);
4472 void perf_event_fork(struct task_struct
*task
)
4474 perf_event_task(task
, NULL
, 1);
4481 struct perf_comm_event
{
4482 struct task_struct
*task
;
4487 struct perf_event_header header
;
4494 static void perf_event_comm_output(struct perf_event
*event
,
4495 struct perf_comm_event
*comm_event
)
4497 struct perf_output_handle handle
;
4498 struct perf_sample_data sample
;
4499 int size
= comm_event
->event_id
.header
.size
;
4502 perf_event_header__init_id(&comm_event
->event_id
.header
, &sample
, event
);
4503 ret
= perf_output_begin(&handle
, event
,
4504 comm_event
->event_id
.header
.size
);
4509 comm_event
->event_id
.pid
= perf_event_pid(event
, comm_event
->task
);
4510 comm_event
->event_id
.tid
= perf_event_tid(event
, comm_event
->task
);
4512 perf_output_put(&handle
, comm_event
->event_id
);
4513 __output_copy(&handle
, comm_event
->comm
,
4514 comm_event
->comm_size
);
4516 perf_event__output_id_sample(event
, &handle
, &sample
);
4518 perf_output_end(&handle
);
4520 comm_event
->event_id
.header
.size
= size
;
4523 static int perf_event_comm_match(struct perf_event
*event
)
4525 if (event
->state
< PERF_EVENT_STATE_INACTIVE
)
4528 if (!event_filter_match(event
))
4531 if (event
->attr
.comm
)
4537 static void perf_event_comm_ctx(struct perf_event_context
*ctx
,
4538 struct perf_comm_event
*comm_event
)
4540 struct perf_event
*event
;
4542 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
4543 if (perf_event_comm_match(event
))
4544 perf_event_comm_output(event
, comm_event
);
4548 static void perf_event_comm_event(struct perf_comm_event
*comm_event
)
4550 struct perf_cpu_context
*cpuctx
;
4551 struct perf_event_context
*ctx
;
4552 char comm
[TASK_COMM_LEN
];
4557 memset(comm
, 0, sizeof(comm
));
4558 strlcpy(comm
, comm_event
->task
->comm
, sizeof(comm
));
4559 size
= ALIGN(strlen(comm
)+1, sizeof(u64
));
4561 comm_event
->comm
= comm
;
4562 comm_event
->comm_size
= size
;
4564 comm_event
->event_id
.header
.size
= sizeof(comm_event
->event_id
) + size
;
4566 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
4567 cpuctx
= get_cpu_ptr(pmu
->pmu_cpu_context
);
4568 if (cpuctx
->active_pmu
!= pmu
)
4570 perf_event_comm_ctx(&cpuctx
->ctx
, comm_event
);
4572 ctxn
= pmu
->task_ctx_nr
;
4576 ctx
= rcu_dereference(current
->perf_event_ctxp
[ctxn
]);
4578 perf_event_comm_ctx(ctx
, comm_event
);
4580 put_cpu_ptr(pmu
->pmu_cpu_context
);
4585 void perf_event_comm(struct task_struct
*task
)
4587 struct perf_comm_event comm_event
;
4588 struct perf_event_context
*ctx
;
4591 for_each_task_context_nr(ctxn
) {
4592 ctx
= task
->perf_event_ctxp
[ctxn
];
4596 perf_event_enable_on_exec(ctx
);
4599 if (!atomic_read(&nr_comm_events
))
4602 comm_event
= (struct perf_comm_event
){
4608 .type
= PERF_RECORD_COMM
,
4617 perf_event_comm_event(&comm_event
);
4624 struct perf_mmap_event
{
4625 struct vm_area_struct
*vma
;
4627 const char *file_name
;
4631 struct perf_event_header header
;
4641 static void perf_event_mmap_output(struct perf_event
*event
,
4642 struct perf_mmap_event
*mmap_event
)
4644 struct perf_output_handle handle
;
4645 struct perf_sample_data sample
;
4646 int size
= mmap_event
->event_id
.header
.size
;
4649 perf_event_header__init_id(&mmap_event
->event_id
.header
, &sample
, event
);
4650 ret
= perf_output_begin(&handle
, event
,
4651 mmap_event
->event_id
.header
.size
);
4655 mmap_event
->event_id
.pid
= perf_event_pid(event
, current
);
4656 mmap_event
->event_id
.tid
= perf_event_tid(event
, current
);
4658 perf_output_put(&handle
, mmap_event
->event_id
);
4659 __output_copy(&handle
, mmap_event
->file_name
,
4660 mmap_event
->file_size
);
4662 perf_event__output_id_sample(event
, &handle
, &sample
);
4664 perf_output_end(&handle
);
4666 mmap_event
->event_id
.header
.size
= size
;
4669 static int perf_event_mmap_match(struct perf_event
*event
,
4670 struct perf_mmap_event
*mmap_event
,
4673 if (event
->state
< PERF_EVENT_STATE_INACTIVE
)
4676 if (!event_filter_match(event
))
4679 if ((!executable
&& event
->attr
.mmap_data
) ||
4680 (executable
&& event
->attr
.mmap
))
4686 static void perf_event_mmap_ctx(struct perf_event_context
*ctx
,
4687 struct perf_mmap_event
*mmap_event
,
4690 struct perf_event
*event
;
4692 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
4693 if (perf_event_mmap_match(event
, mmap_event
, executable
))
4694 perf_event_mmap_output(event
, mmap_event
);
4698 static void perf_event_mmap_event(struct perf_mmap_event
*mmap_event
)
4700 struct perf_cpu_context
*cpuctx
;
4701 struct perf_event_context
*ctx
;
4702 struct vm_area_struct
*vma
= mmap_event
->vma
;
4703 struct file
*file
= vma
->vm_file
;
4711 memset(tmp
, 0, sizeof(tmp
));
4715 * d_path works from the end of the rb backwards, so we
4716 * need to add enough zero bytes after the string to handle
4717 * the 64bit alignment we do later.
4719 buf
= kzalloc(PATH_MAX
+ sizeof(u64
), GFP_KERNEL
);
4721 name
= strncpy(tmp
, "//enomem", sizeof(tmp
));
4724 name
= d_path(&file
->f_path
, buf
, PATH_MAX
);
4726 name
= strncpy(tmp
, "//toolong", sizeof(tmp
));
4730 if (arch_vma_name(mmap_event
->vma
)) {
4731 name
= strncpy(tmp
, arch_vma_name(mmap_event
->vma
),
4737 name
= strncpy(tmp
, "[vdso]", sizeof(tmp
));
4739 } else if (vma
->vm_start
<= vma
->vm_mm
->start_brk
&&
4740 vma
->vm_end
>= vma
->vm_mm
->brk
) {
4741 name
= strncpy(tmp
, "[heap]", sizeof(tmp
));
4743 } else if (vma
->vm_start
<= vma
->vm_mm
->start_stack
&&
4744 vma
->vm_end
>= vma
->vm_mm
->start_stack
) {
4745 name
= strncpy(tmp
, "[stack]", sizeof(tmp
));
4749 name
= strncpy(tmp
, "//anon", sizeof(tmp
));
4754 size
= ALIGN(strlen(name
)+1, sizeof(u64
));
4756 mmap_event
->file_name
= name
;
4757 mmap_event
->file_size
= size
;
4759 mmap_event
->event_id
.header
.size
= sizeof(mmap_event
->event_id
) + size
;
4762 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
4763 cpuctx
= get_cpu_ptr(pmu
->pmu_cpu_context
);
4764 if (cpuctx
->active_pmu
!= pmu
)
4766 perf_event_mmap_ctx(&cpuctx
->ctx
, mmap_event
,
4767 vma
->vm_flags
& VM_EXEC
);
4769 ctxn
= pmu
->task_ctx_nr
;
4773 ctx
= rcu_dereference(current
->perf_event_ctxp
[ctxn
]);
4775 perf_event_mmap_ctx(ctx
, mmap_event
,
4776 vma
->vm_flags
& VM_EXEC
);
4779 put_cpu_ptr(pmu
->pmu_cpu_context
);
4786 void perf_event_mmap(struct vm_area_struct
*vma
)
4788 struct perf_mmap_event mmap_event
;
4790 if (!atomic_read(&nr_mmap_events
))
4793 mmap_event
= (struct perf_mmap_event
){
4799 .type
= PERF_RECORD_MMAP
,
4800 .misc
= PERF_RECORD_MISC_USER
,
4805 .start
= vma
->vm_start
,
4806 .len
= vma
->vm_end
- vma
->vm_start
,
4807 .pgoff
= (u64
)vma
->vm_pgoff
<< PAGE_SHIFT
,
4811 perf_event_mmap_event(&mmap_event
);
4815 * IRQ throttle logging
4818 static void perf_log_throttle(struct perf_event
*event
, int enable
)
4820 struct perf_output_handle handle
;
4821 struct perf_sample_data sample
;
4825 struct perf_event_header header
;
4829 } throttle_event
= {
4831 .type
= PERF_RECORD_THROTTLE
,
4833 .size
= sizeof(throttle_event
),
4835 .time
= perf_clock(),
4836 .id
= primary_event_id(event
),
4837 .stream_id
= event
->id
,
4841 throttle_event
.header
.type
= PERF_RECORD_UNTHROTTLE
;
4843 perf_event_header__init_id(&throttle_event
.header
, &sample
, event
);
4845 ret
= perf_output_begin(&handle
, event
,
4846 throttle_event
.header
.size
);
4850 perf_output_put(&handle
, throttle_event
);
4851 perf_event__output_id_sample(event
, &handle
, &sample
);
4852 perf_output_end(&handle
);
4856 * Generic event overflow handling, sampling.
4859 static int __perf_event_overflow(struct perf_event
*event
,
4860 int throttle
, struct perf_sample_data
*data
,
4861 struct pt_regs
*regs
)
4863 int events
= atomic_read(&event
->event_limit
);
4864 struct hw_perf_event
*hwc
= &event
->hw
;
4869 * Non-sampling counters might still use the PMI to fold short
4870 * hardware counters, ignore those.
4872 if (unlikely(!is_sampling_event(event
)))
4875 seq
= __this_cpu_read(perf_throttled_seq
);
4876 if (seq
!= hwc
->interrupts_seq
) {
4877 hwc
->interrupts_seq
= seq
;
4878 hwc
->interrupts
= 1;
4881 if (unlikely(throttle
4882 && hwc
->interrupts
>= max_samples_per_tick
)) {
4883 __this_cpu_inc(perf_throttled_count
);
4884 hwc
->interrupts
= MAX_INTERRUPTS
;
4885 perf_log_throttle(event
, 0);
4890 if (event
->attr
.freq
) {
4891 u64 now
= perf_clock();
4892 s64 delta
= now
- hwc
->freq_time_stamp
;
4894 hwc
->freq_time_stamp
= now
;
4896 if (delta
> 0 && delta
< 2*TICK_NSEC
)
4897 perf_adjust_period(event
, delta
, hwc
->last_period
, true);
4901 * XXX event_limit might not quite work as expected on inherited
4905 event
->pending_kill
= POLL_IN
;
4906 if (events
&& atomic_dec_and_test(&event
->event_limit
)) {
4908 event
->pending_kill
= POLL_HUP
;
4909 event
->pending_disable
= 1;
4910 irq_work_queue(&event
->pending
);
4913 if (event
->overflow_handler
)
4914 event
->overflow_handler(event
, data
, regs
);
4916 perf_event_output(event
, data
, regs
);
4918 if (event
->fasync
&& event
->pending_kill
) {
4919 event
->pending_wakeup
= 1;
4920 irq_work_queue(&event
->pending
);
4926 int perf_event_overflow(struct perf_event
*event
,
4927 struct perf_sample_data
*data
,
4928 struct pt_regs
*regs
)
4930 return __perf_event_overflow(event
, 1, data
, regs
);
4934 * Generic software event infrastructure
4937 struct swevent_htable
{
4938 struct swevent_hlist
*swevent_hlist
;
4939 struct mutex hlist_mutex
;
4942 /* Recursion avoidance in each contexts */
4943 int recursion
[PERF_NR_CONTEXTS
];
4946 static DEFINE_PER_CPU(struct swevent_htable
, swevent_htable
);
4949 * We directly increment event->count and keep a second value in
4950 * event->hw.period_left to count intervals. This period event
4951 * is kept in the range [-sample_period, 0] so that we can use the
4955 static u64
perf_swevent_set_period(struct perf_event
*event
)
4957 struct hw_perf_event
*hwc
= &event
->hw
;
4958 u64 period
= hwc
->last_period
;
4962 hwc
->last_period
= hwc
->sample_period
;
4965 old
= val
= local64_read(&hwc
->period_left
);
4969 nr
= div64_u64(period
+ val
, period
);
4970 offset
= nr
* period
;
4972 if (local64_cmpxchg(&hwc
->period_left
, old
, val
) != old
)
4978 static void perf_swevent_overflow(struct perf_event
*event
, u64 overflow
,
4979 struct perf_sample_data
*data
,
4980 struct pt_regs
*regs
)
4982 struct hw_perf_event
*hwc
= &event
->hw
;
4986 overflow
= perf_swevent_set_period(event
);
4988 if (hwc
->interrupts
== MAX_INTERRUPTS
)
4991 for (; overflow
; overflow
--) {
4992 if (__perf_event_overflow(event
, throttle
,
4995 * We inhibit the overflow from happening when
4996 * hwc->interrupts == MAX_INTERRUPTS.
5004 static void perf_swevent_event(struct perf_event
*event
, u64 nr
,
5005 struct perf_sample_data
*data
,
5006 struct pt_regs
*regs
)
5008 struct hw_perf_event
*hwc
= &event
->hw
;
5010 local64_add(nr
, &event
->count
);
5015 if (!is_sampling_event(event
))
5018 if ((event
->attr
.sample_type
& PERF_SAMPLE_PERIOD
) && !event
->attr
.freq
) {
5020 return perf_swevent_overflow(event
, 1, data
, regs
);
5022 data
->period
= event
->hw
.last_period
;
5024 if (nr
== 1 && hwc
->sample_period
== 1 && !event
->attr
.freq
)
5025 return perf_swevent_overflow(event
, 1, data
, regs
);
5027 if (local64_add_negative(nr
, &hwc
->period_left
))
5030 perf_swevent_overflow(event
, 0, data
, regs
);
5033 static int perf_exclude_event(struct perf_event
*event
,
5034 struct pt_regs
*regs
)
5036 if (event
->hw
.state
& PERF_HES_STOPPED
)
5040 if (event
->attr
.exclude_user
&& user_mode(regs
))
5043 if (event
->attr
.exclude_kernel
&& !user_mode(regs
))
5050 static int perf_swevent_match(struct perf_event
*event
,
5051 enum perf_type_id type
,
5053 struct perf_sample_data
*data
,
5054 struct pt_regs
*regs
)
5056 if (event
->attr
.type
!= type
)
5059 if (event
->attr
.config
!= event_id
)
5062 if (perf_exclude_event(event
, regs
))
5068 static inline u64
swevent_hash(u64 type
, u32 event_id
)
5070 u64 val
= event_id
| (type
<< 32);
5072 return hash_64(val
, SWEVENT_HLIST_BITS
);
5075 static inline struct hlist_head
*
5076 __find_swevent_head(struct swevent_hlist
*hlist
, u64 type
, u32 event_id
)
5078 u64 hash
= swevent_hash(type
, event_id
);
5080 return &hlist
->heads
[hash
];
5083 /* For the read side: events when they trigger */
5084 static inline struct hlist_head
*
5085 find_swevent_head_rcu(struct swevent_htable
*swhash
, u64 type
, u32 event_id
)
5087 struct swevent_hlist
*hlist
;
5089 hlist
= rcu_dereference(swhash
->swevent_hlist
);
5093 return __find_swevent_head(hlist
, type
, event_id
);
5096 /* For the event head insertion and removal in the hlist */
5097 static inline struct hlist_head
*
5098 find_swevent_head(struct swevent_htable
*swhash
, struct perf_event
*event
)
5100 struct swevent_hlist
*hlist
;
5101 u32 event_id
= event
->attr
.config
;
5102 u64 type
= event
->attr
.type
;
5105 * Event scheduling is always serialized against hlist allocation
5106 * and release. Which makes the protected version suitable here.
5107 * The context lock guarantees that.
5109 hlist
= rcu_dereference_protected(swhash
->swevent_hlist
,
5110 lockdep_is_held(&event
->ctx
->lock
));
5114 return __find_swevent_head(hlist
, type
, event_id
);
5117 static void do_perf_sw_event(enum perf_type_id type
, u32 event_id
,
5119 struct perf_sample_data
*data
,
5120 struct pt_regs
*regs
)
5122 struct swevent_htable
*swhash
= &__get_cpu_var(swevent_htable
);
5123 struct perf_event
*event
;
5124 struct hlist_node
*node
;
5125 struct hlist_head
*head
;
5128 head
= find_swevent_head_rcu(swhash
, type
, event_id
);
5132 hlist_for_each_entry_rcu(event
, node
, head
, hlist_entry
) {
5133 if (perf_swevent_match(event
, type
, event_id
, data
, regs
))
5134 perf_swevent_event(event
, nr
, data
, regs
);
5140 int perf_swevent_get_recursion_context(void)
5142 struct swevent_htable
*swhash
= &__get_cpu_var(swevent_htable
);
5144 return get_recursion_context(swhash
->recursion
);
5146 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context
);
5148 inline void perf_swevent_put_recursion_context(int rctx
)
5150 struct swevent_htable
*swhash
= &__get_cpu_var(swevent_htable
);
5152 put_recursion_context(swhash
->recursion
, rctx
);
5155 void __perf_sw_event(u32 event_id
, u64 nr
, struct pt_regs
*regs
, u64 addr
)
5157 struct perf_sample_data data
;
5160 preempt_disable_notrace();
5161 rctx
= perf_swevent_get_recursion_context();
5165 perf_sample_data_init(&data
, addr
, 0);
5167 do_perf_sw_event(PERF_TYPE_SOFTWARE
, event_id
, nr
, &data
, regs
);
5169 perf_swevent_put_recursion_context(rctx
);
5170 preempt_enable_notrace();
5173 static void perf_swevent_read(struct perf_event
*event
)
5177 static int perf_swevent_add(struct perf_event
*event
, int flags
)
5179 struct swevent_htable
*swhash
= &__get_cpu_var(swevent_htable
);
5180 struct hw_perf_event
*hwc
= &event
->hw
;
5181 struct hlist_head
*head
;
5183 if (is_sampling_event(event
)) {
5184 hwc
->last_period
= hwc
->sample_period
;
5185 perf_swevent_set_period(event
);
5188 hwc
->state
= !(flags
& PERF_EF_START
);
5190 head
= find_swevent_head(swhash
, event
);
5191 if (WARN_ON_ONCE(!head
))
5194 hlist_add_head_rcu(&event
->hlist_entry
, head
);
5199 static void perf_swevent_del(struct perf_event
*event
, int flags
)
5201 hlist_del_rcu(&event
->hlist_entry
);
5204 static void perf_swevent_start(struct perf_event
*event
, int flags
)
5206 event
->hw
.state
= 0;
5209 static void perf_swevent_stop(struct perf_event
*event
, int flags
)
5211 event
->hw
.state
= PERF_HES_STOPPED
;
5214 /* Deref the hlist from the update side */
5215 static inline struct swevent_hlist
*
5216 swevent_hlist_deref(struct swevent_htable
*swhash
)
5218 return rcu_dereference_protected(swhash
->swevent_hlist
,
5219 lockdep_is_held(&swhash
->hlist_mutex
));
5222 static void swevent_hlist_release(struct swevent_htable
*swhash
)
5224 struct swevent_hlist
*hlist
= swevent_hlist_deref(swhash
);
5229 rcu_assign_pointer(swhash
->swevent_hlist
, NULL
);
5230 kfree_rcu(hlist
, rcu_head
);
5233 static void swevent_hlist_put_cpu(struct perf_event
*event
, int cpu
)
5235 struct swevent_htable
*swhash
= &per_cpu(swevent_htable
, cpu
);
5237 mutex_lock(&swhash
->hlist_mutex
);
5239 if (!--swhash
->hlist_refcount
)
5240 swevent_hlist_release(swhash
);
5242 mutex_unlock(&swhash
->hlist_mutex
);
5245 static void swevent_hlist_put(struct perf_event
*event
)
5249 if (event
->cpu
!= -1) {
5250 swevent_hlist_put_cpu(event
, event
->cpu
);
5254 for_each_possible_cpu(cpu
)
5255 swevent_hlist_put_cpu(event
, cpu
);
5258 static int swevent_hlist_get_cpu(struct perf_event
*event
, int cpu
)
5260 struct swevent_htable
*swhash
= &per_cpu(swevent_htable
, cpu
);
5263 mutex_lock(&swhash
->hlist_mutex
);
5265 if (!swevent_hlist_deref(swhash
) && cpu_online(cpu
)) {
5266 struct swevent_hlist
*hlist
;
5268 hlist
= kzalloc(sizeof(*hlist
), GFP_KERNEL
);
5273 rcu_assign_pointer(swhash
->swevent_hlist
, hlist
);
5275 swhash
->hlist_refcount
++;
5277 mutex_unlock(&swhash
->hlist_mutex
);
5282 static int swevent_hlist_get(struct perf_event
*event
)
5285 int cpu
, failed_cpu
;
5287 if (event
->cpu
!= -1)
5288 return swevent_hlist_get_cpu(event
, event
->cpu
);
5291 for_each_possible_cpu(cpu
) {
5292 err
= swevent_hlist_get_cpu(event
, cpu
);
5302 for_each_possible_cpu(cpu
) {
5303 if (cpu
== failed_cpu
)
5305 swevent_hlist_put_cpu(event
, cpu
);
5312 struct static_key perf_swevent_enabled
[PERF_COUNT_SW_MAX
];
5314 static void sw_perf_event_destroy(struct perf_event
*event
)
5316 u64 event_id
= event
->attr
.config
;
5318 WARN_ON(event
->parent
);
5320 static_key_slow_dec(&perf_swevent_enabled
[event_id
]);
5321 swevent_hlist_put(event
);
5324 static int perf_swevent_init(struct perf_event
*event
)
5326 int event_id
= event
->attr
.config
;
5328 if (event
->attr
.type
!= PERF_TYPE_SOFTWARE
)
5332 * no branch sampling for software events
5334 if (has_branch_stack(event
))
5338 case PERF_COUNT_SW_CPU_CLOCK
:
5339 case PERF_COUNT_SW_TASK_CLOCK
:
5346 if (event_id
>= PERF_COUNT_SW_MAX
)
5349 if (!event
->parent
) {
5352 err
= swevent_hlist_get(event
);
5356 static_key_slow_inc(&perf_swevent_enabled
[event_id
]);
5357 event
->destroy
= sw_perf_event_destroy
;
5363 static int perf_swevent_event_idx(struct perf_event
*event
)
5368 static struct pmu perf_swevent
= {
5369 .task_ctx_nr
= perf_sw_context
,
5371 .event_init
= perf_swevent_init
,
5372 .add
= perf_swevent_add
,
5373 .del
= perf_swevent_del
,
5374 .start
= perf_swevent_start
,
5375 .stop
= perf_swevent_stop
,
5376 .read
= perf_swevent_read
,
5378 .event_idx
= perf_swevent_event_idx
,
5381 #ifdef CONFIG_EVENT_TRACING
5383 static int perf_tp_filter_match(struct perf_event
*event
,
5384 struct perf_sample_data
*data
)
5386 void *record
= data
->raw
->data
;
5388 if (likely(!event
->filter
) || filter_match_preds(event
->filter
, record
))
5393 static int perf_tp_event_match(struct perf_event
*event
,
5394 struct perf_sample_data
*data
,
5395 struct pt_regs
*regs
)
5397 if (event
->hw
.state
& PERF_HES_STOPPED
)
5400 * All tracepoints are from kernel-space.
5402 if (event
->attr
.exclude_kernel
)
5405 if (!perf_tp_filter_match(event
, data
))
5411 void perf_tp_event(u64 addr
, u64 count
, void *record
, int entry_size
,
5412 struct pt_regs
*regs
, struct hlist_head
*head
, int rctx
,
5413 struct task_struct
*task
)
5415 struct perf_sample_data data
;
5416 struct perf_event
*event
;
5417 struct hlist_node
*node
;
5419 struct perf_raw_record raw
= {
5424 perf_sample_data_init(&data
, addr
, 0);
5427 hlist_for_each_entry_rcu(event
, node
, head
, hlist_entry
) {
5428 if (perf_tp_event_match(event
, &data
, regs
))
5429 perf_swevent_event(event
, count
, &data
, regs
);
5433 * If we got specified a target task, also iterate its context and
5434 * deliver this event there too.
5436 if (task
&& task
!= current
) {
5437 struct perf_event_context
*ctx
;
5438 struct trace_entry
*entry
= record
;
5441 ctx
= rcu_dereference(task
->perf_event_ctxp
[perf_sw_context
]);
5445 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
5446 if (event
->attr
.type
!= PERF_TYPE_TRACEPOINT
)
5448 if (event
->attr
.config
!= entry
->type
)
5450 if (perf_tp_event_match(event
, &data
, regs
))
5451 perf_swevent_event(event
, count
, &data
, regs
);
5457 perf_swevent_put_recursion_context(rctx
);
5459 EXPORT_SYMBOL_GPL(perf_tp_event
);
5461 static void tp_perf_event_destroy(struct perf_event
*event
)
5463 perf_trace_destroy(event
);
5466 static int perf_tp_event_init(struct perf_event
*event
)
5470 if (event
->attr
.type
!= PERF_TYPE_TRACEPOINT
)
5474 * no branch sampling for tracepoint events
5476 if (has_branch_stack(event
))
5479 err
= perf_trace_init(event
);
5483 event
->destroy
= tp_perf_event_destroy
;
5488 static struct pmu perf_tracepoint
= {
5489 .task_ctx_nr
= perf_sw_context
,
5491 .event_init
= perf_tp_event_init
,
5492 .add
= perf_trace_add
,
5493 .del
= perf_trace_del
,
5494 .start
= perf_swevent_start
,
5495 .stop
= perf_swevent_stop
,
5496 .read
= perf_swevent_read
,
5498 .event_idx
= perf_swevent_event_idx
,
5501 static inline void perf_tp_register(void)
5503 perf_pmu_register(&perf_tracepoint
, "tracepoint", PERF_TYPE_TRACEPOINT
);
5506 static int perf_event_set_filter(struct perf_event
*event
, void __user
*arg
)
5511 if (event
->attr
.type
!= PERF_TYPE_TRACEPOINT
)
5514 filter_str
= strndup_user(arg
, PAGE_SIZE
);
5515 if (IS_ERR(filter_str
))
5516 return PTR_ERR(filter_str
);
5518 ret
= ftrace_profile_set_filter(event
, event
->attr
.config
, filter_str
);
5524 static void perf_event_free_filter(struct perf_event
*event
)
5526 ftrace_profile_free_filter(event
);
5531 static inline void perf_tp_register(void)
5535 static int perf_event_set_filter(struct perf_event
*event
, void __user
*arg
)
5540 static void perf_event_free_filter(struct perf_event
*event
)
5544 #endif /* CONFIG_EVENT_TRACING */
5546 #ifdef CONFIG_HAVE_HW_BREAKPOINT
5547 void perf_bp_event(struct perf_event
*bp
, void *data
)
5549 struct perf_sample_data sample
;
5550 struct pt_regs
*regs
= data
;
5552 perf_sample_data_init(&sample
, bp
->attr
.bp_addr
, 0);
5554 if (!bp
->hw
.state
&& !perf_exclude_event(bp
, regs
))
5555 perf_swevent_event(bp
, 1, &sample
, regs
);
5560 * hrtimer based swevent callback
5563 static enum hrtimer_restart
perf_swevent_hrtimer(struct hrtimer
*hrtimer
)
5565 enum hrtimer_restart ret
= HRTIMER_RESTART
;
5566 struct perf_sample_data data
;
5567 struct pt_regs
*regs
;
5568 struct perf_event
*event
;
5571 event
= container_of(hrtimer
, struct perf_event
, hw
.hrtimer
);
5573 if (event
->state
!= PERF_EVENT_STATE_ACTIVE
)
5574 return HRTIMER_NORESTART
;
5576 event
->pmu
->read(event
);
5578 perf_sample_data_init(&data
, 0, event
->hw
.last_period
);
5579 regs
= get_irq_regs();
5581 if (regs
&& !perf_exclude_event(event
, regs
)) {
5582 if (!(event
->attr
.exclude_idle
&& is_idle_task(current
)))
5583 if (__perf_event_overflow(event
, 1, &data
, regs
))
5584 ret
= HRTIMER_NORESTART
;
5587 period
= max_t(u64
, 10000, event
->hw
.sample_period
);
5588 hrtimer_forward_now(hrtimer
, ns_to_ktime(period
));
5593 static void perf_swevent_start_hrtimer(struct perf_event
*event
)
5595 struct hw_perf_event
*hwc
= &event
->hw
;
5598 if (!is_sampling_event(event
))
5601 period
= local64_read(&hwc
->period_left
);
5606 local64_set(&hwc
->period_left
, 0);
5608 period
= max_t(u64
, 10000, hwc
->sample_period
);
5610 __hrtimer_start_range_ns(&hwc
->hrtimer
,
5611 ns_to_ktime(period
), 0,
5612 HRTIMER_MODE_REL_PINNED
, 0);
5615 static void perf_swevent_cancel_hrtimer(struct perf_event
*event
)
5617 struct hw_perf_event
*hwc
= &event
->hw
;
5619 if (is_sampling_event(event
)) {
5620 ktime_t remaining
= hrtimer_get_remaining(&hwc
->hrtimer
);
5621 local64_set(&hwc
->period_left
, ktime_to_ns(remaining
));
5623 hrtimer_cancel(&hwc
->hrtimer
);
5627 static void perf_swevent_init_hrtimer(struct perf_event
*event
)
5629 struct hw_perf_event
*hwc
= &event
->hw
;
5631 if (!is_sampling_event(event
))
5634 hrtimer_init(&hwc
->hrtimer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
5635 hwc
->hrtimer
.function
= perf_swevent_hrtimer
;
5638 * Since hrtimers have a fixed rate, we can do a static freq->period
5639 * mapping and avoid the whole period adjust feedback stuff.
5641 if (event
->attr
.freq
) {
5642 long freq
= event
->attr
.sample_freq
;
5644 event
->attr
.sample_period
= NSEC_PER_SEC
/ freq
;
5645 hwc
->sample_period
= event
->attr
.sample_period
;
5646 local64_set(&hwc
->period_left
, hwc
->sample_period
);
5647 event
->attr
.freq
= 0;
5652 * Software event: cpu wall time clock
5655 static void cpu_clock_event_update(struct perf_event
*event
)
5660 now
= local_clock();
5661 prev
= local64_xchg(&event
->hw
.prev_count
, now
);
5662 local64_add(now
- prev
, &event
->count
);
5665 static void cpu_clock_event_start(struct perf_event
*event
, int flags
)
5667 local64_set(&event
->hw
.prev_count
, local_clock());
5668 perf_swevent_start_hrtimer(event
);
5671 static void cpu_clock_event_stop(struct perf_event
*event
, int flags
)
5673 perf_swevent_cancel_hrtimer(event
);
5674 cpu_clock_event_update(event
);
5677 static int cpu_clock_event_add(struct perf_event
*event
, int flags
)
5679 if (flags
& PERF_EF_START
)
5680 cpu_clock_event_start(event
, flags
);
5685 static void cpu_clock_event_del(struct perf_event
*event
, int flags
)
5687 cpu_clock_event_stop(event
, flags
);
5690 static void cpu_clock_event_read(struct perf_event
*event
)
5692 cpu_clock_event_update(event
);
5695 static int cpu_clock_event_init(struct perf_event
*event
)
5697 if (event
->attr
.type
!= PERF_TYPE_SOFTWARE
)
5700 if (event
->attr
.config
!= PERF_COUNT_SW_CPU_CLOCK
)
5704 * no branch sampling for software events
5706 if (has_branch_stack(event
))
5709 perf_swevent_init_hrtimer(event
);
5714 static struct pmu perf_cpu_clock
= {
5715 .task_ctx_nr
= perf_sw_context
,
5717 .event_init
= cpu_clock_event_init
,
5718 .add
= cpu_clock_event_add
,
5719 .del
= cpu_clock_event_del
,
5720 .start
= cpu_clock_event_start
,
5721 .stop
= cpu_clock_event_stop
,
5722 .read
= cpu_clock_event_read
,
5724 .event_idx
= perf_swevent_event_idx
,
5728 * Software event: task time clock
5731 static void task_clock_event_update(struct perf_event
*event
, u64 now
)
5736 prev
= local64_xchg(&event
->hw
.prev_count
, now
);
5738 local64_add(delta
, &event
->count
);
5741 static void task_clock_event_start(struct perf_event
*event
, int flags
)
5743 local64_set(&event
->hw
.prev_count
, event
->ctx
->time
);
5744 perf_swevent_start_hrtimer(event
);
5747 static void task_clock_event_stop(struct perf_event
*event
, int flags
)
5749 perf_swevent_cancel_hrtimer(event
);
5750 task_clock_event_update(event
, event
->ctx
->time
);
5753 static int task_clock_event_add(struct perf_event
*event
, int flags
)
5755 if (flags
& PERF_EF_START
)
5756 task_clock_event_start(event
, flags
);
5761 static void task_clock_event_del(struct perf_event
*event
, int flags
)
5763 task_clock_event_stop(event
, PERF_EF_UPDATE
);
5766 static void task_clock_event_read(struct perf_event
*event
)
5768 u64 now
= perf_clock();
5769 u64 delta
= now
- event
->ctx
->timestamp
;
5770 u64 time
= event
->ctx
->time
+ delta
;
5772 task_clock_event_update(event
, time
);
5775 static int task_clock_event_init(struct perf_event
*event
)
5777 if (event
->attr
.type
!= PERF_TYPE_SOFTWARE
)
5780 if (event
->attr
.config
!= PERF_COUNT_SW_TASK_CLOCK
)
5784 * no branch sampling for software events
5786 if (has_branch_stack(event
))
5789 perf_swevent_init_hrtimer(event
);
5794 static struct pmu perf_task_clock
= {
5795 .task_ctx_nr
= perf_sw_context
,
5797 .event_init
= task_clock_event_init
,
5798 .add
= task_clock_event_add
,
5799 .del
= task_clock_event_del
,
5800 .start
= task_clock_event_start
,
5801 .stop
= task_clock_event_stop
,
5802 .read
= task_clock_event_read
,
5804 .event_idx
= perf_swevent_event_idx
,
5807 static void perf_pmu_nop_void(struct pmu
*pmu
)
5811 static int perf_pmu_nop_int(struct pmu
*pmu
)
5816 static void perf_pmu_start_txn(struct pmu
*pmu
)
5818 perf_pmu_disable(pmu
);
5821 static int perf_pmu_commit_txn(struct pmu
*pmu
)
5823 perf_pmu_enable(pmu
);
5827 static void perf_pmu_cancel_txn(struct pmu
*pmu
)
5829 perf_pmu_enable(pmu
);
5832 static int perf_event_idx_default(struct perf_event
*event
)
5834 return event
->hw
.idx
+ 1;
5838 * Ensures all contexts with the same task_ctx_nr have the same
5839 * pmu_cpu_context too.
5841 static void *find_pmu_context(int ctxn
)
5848 list_for_each_entry(pmu
, &pmus
, entry
) {
5849 if (pmu
->task_ctx_nr
== ctxn
)
5850 return pmu
->pmu_cpu_context
;
5856 static void update_pmu_context(struct pmu
*pmu
, struct pmu
*old_pmu
)
5860 for_each_possible_cpu(cpu
) {
5861 struct perf_cpu_context
*cpuctx
;
5863 cpuctx
= per_cpu_ptr(pmu
->pmu_cpu_context
, cpu
);
5865 if (cpuctx
->active_pmu
== old_pmu
)
5866 cpuctx
->active_pmu
= pmu
;
5870 static void free_pmu_context(struct pmu
*pmu
)
5874 mutex_lock(&pmus_lock
);
5876 * Like a real lame refcount.
5878 list_for_each_entry(i
, &pmus
, entry
) {
5879 if (i
->pmu_cpu_context
== pmu
->pmu_cpu_context
) {
5880 update_pmu_context(i
, pmu
);
5885 free_percpu(pmu
->pmu_cpu_context
);
5887 mutex_unlock(&pmus_lock
);
5889 static struct idr pmu_idr
;
5892 type_show(struct device
*dev
, struct device_attribute
*attr
, char *page
)
5894 struct pmu
*pmu
= dev_get_drvdata(dev
);
5896 return snprintf(page
, PAGE_SIZE
-1, "%d\n", pmu
->type
);
5899 static struct device_attribute pmu_dev_attrs
[] = {
5904 static int pmu_bus_running
;
5905 static struct bus_type pmu_bus
= {
5906 .name
= "event_source",
5907 .dev_attrs
= pmu_dev_attrs
,
5910 static void pmu_dev_release(struct device
*dev
)
5915 static int pmu_dev_alloc(struct pmu
*pmu
)
5919 pmu
->dev
= kzalloc(sizeof(struct device
), GFP_KERNEL
);
5923 pmu
->dev
->groups
= pmu
->attr_groups
;
5924 device_initialize(pmu
->dev
);
5925 ret
= dev_set_name(pmu
->dev
, "%s", pmu
->name
);
5929 dev_set_drvdata(pmu
->dev
, pmu
);
5930 pmu
->dev
->bus
= &pmu_bus
;
5931 pmu
->dev
->release
= pmu_dev_release
;
5932 ret
= device_add(pmu
->dev
);
5940 put_device(pmu
->dev
);
5944 static struct lock_class_key cpuctx_mutex
;
5945 static struct lock_class_key cpuctx_lock
;
5947 int perf_pmu_register(struct pmu
*pmu
, char *name
, int type
)
5951 mutex_lock(&pmus_lock
);
5953 pmu
->pmu_disable_count
= alloc_percpu(int);
5954 if (!pmu
->pmu_disable_count
)
5963 int err
= idr_pre_get(&pmu_idr
, GFP_KERNEL
);
5967 err
= idr_get_new_above(&pmu_idr
, pmu
, PERF_TYPE_MAX
, &type
);
5975 if (pmu_bus_running
) {
5976 ret
= pmu_dev_alloc(pmu
);
5982 pmu
->pmu_cpu_context
= find_pmu_context(pmu
->task_ctx_nr
);
5983 if (pmu
->pmu_cpu_context
)
5984 goto got_cpu_context
;
5986 pmu
->pmu_cpu_context
= alloc_percpu(struct perf_cpu_context
);
5987 if (!pmu
->pmu_cpu_context
)
5990 for_each_possible_cpu(cpu
) {
5991 struct perf_cpu_context
*cpuctx
;
5993 cpuctx
= per_cpu_ptr(pmu
->pmu_cpu_context
, cpu
);
5994 __perf_event_init_context(&cpuctx
->ctx
);
5995 lockdep_set_class(&cpuctx
->ctx
.mutex
, &cpuctx_mutex
);
5996 lockdep_set_class(&cpuctx
->ctx
.lock
, &cpuctx_lock
);
5997 cpuctx
->ctx
.type
= cpu_context
;
5998 cpuctx
->ctx
.pmu
= pmu
;
5999 cpuctx
->jiffies_interval
= 1;
6000 INIT_LIST_HEAD(&cpuctx
->rotation_list
);
6001 cpuctx
->active_pmu
= pmu
;
6005 if (!pmu
->start_txn
) {
6006 if (pmu
->pmu_enable
) {
6008 * If we have pmu_enable/pmu_disable calls, install
6009 * transaction stubs that use that to try and batch
6010 * hardware accesses.
6012 pmu
->start_txn
= perf_pmu_start_txn
;
6013 pmu
->commit_txn
= perf_pmu_commit_txn
;
6014 pmu
->cancel_txn
= perf_pmu_cancel_txn
;
6016 pmu
->start_txn
= perf_pmu_nop_void
;
6017 pmu
->commit_txn
= perf_pmu_nop_int
;
6018 pmu
->cancel_txn
= perf_pmu_nop_void
;
6022 if (!pmu
->pmu_enable
) {
6023 pmu
->pmu_enable
= perf_pmu_nop_void
;
6024 pmu
->pmu_disable
= perf_pmu_nop_void
;
6027 if (!pmu
->event_idx
)
6028 pmu
->event_idx
= perf_event_idx_default
;
6030 list_add_rcu(&pmu
->entry
, &pmus
);
6033 mutex_unlock(&pmus_lock
);
6038 device_del(pmu
->dev
);
6039 put_device(pmu
->dev
);
6042 if (pmu
->type
>= PERF_TYPE_MAX
)
6043 idr_remove(&pmu_idr
, pmu
->type
);
6046 free_percpu(pmu
->pmu_disable_count
);
6050 void perf_pmu_unregister(struct pmu
*pmu
)
6052 mutex_lock(&pmus_lock
);
6053 list_del_rcu(&pmu
->entry
);
6054 mutex_unlock(&pmus_lock
);
6057 * We dereference the pmu list under both SRCU and regular RCU, so
6058 * synchronize against both of those.
6060 synchronize_srcu(&pmus_srcu
);
6063 free_percpu(pmu
->pmu_disable_count
);
6064 if (pmu
->type
>= PERF_TYPE_MAX
)
6065 idr_remove(&pmu_idr
, pmu
->type
);
6066 device_del(pmu
->dev
);
6067 put_device(pmu
->dev
);
6068 free_pmu_context(pmu
);
6071 struct pmu
*perf_init_event(struct perf_event
*event
)
6073 struct pmu
*pmu
= NULL
;
6077 idx
= srcu_read_lock(&pmus_srcu
);
6080 pmu
= idr_find(&pmu_idr
, event
->attr
.type
);
6084 ret
= pmu
->event_init(event
);
6090 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
6092 ret
= pmu
->event_init(event
);
6096 if (ret
!= -ENOENT
) {
6101 pmu
= ERR_PTR(-ENOENT
);
6103 srcu_read_unlock(&pmus_srcu
, idx
);
6109 * Allocate and initialize a event structure
6111 static struct perf_event
*
6112 perf_event_alloc(struct perf_event_attr
*attr
, int cpu
,
6113 struct task_struct
*task
,
6114 struct perf_event
*group_leader
,
6115 struct perf_event
*parent_event
,
6116 perf_overflow_handler_t overflow_handler
,
6120 struct perf_event
*event
;
6121 struct hw_perf_event
*hwc
;
6124 if ((unsigned)cpu
>= nr_cpu_ids
) {
6125 if (!task
|| cpu
!= -1)
6126 return ERR_PTR(-EINVAL
);
6129 event
= kzalloc(sizeof(*event
), GFP_KERNEL
);
6131 return ERR_PTR(-ENOMEM
);
6134 * Single events are their own group leaders, with an
6135 * empty sibling list:
6138 group_leader
= event
;
6140 mutex_init(&event
->child_mutex
);
6141 INIT_LIST_HEAD(&event
->child_list
);
6143 INIT_LIST_HEAD(&event
->group_entry
);
6144 INIT_LIST_HEAD(&event
->event_entry
);
6145 INIT_LIST_HEAD(&event
->sibling_list
);
6146 INIT_LIST_HEAD(&event
->rb_entry
);
6148 init_waitqueue_head(&event
->waitq
);
6149 init_irq_work(&event
->pending
, perf_pending_event
);
6151 mutex_init(&event
->mmap_mutex
);
6153 atomic_long_set(&event
->refcount
, 1);
6155 event
->attr
= *attr
;
6156 event
->group_leader
= group_leader
;
6160 event
->parent
= parent_event
;
6162 event
->ns
= get_pid_ns(current
->nsproxy
->pid_ns
);
6163 event
->id
= atomic64_inc_return(&perf_event_id
);
6165 event
->state
= PERF_EVENT_STATE_INACTIVE
;
6168 event
->attach_state
= PERF_ATTACH_TASK
;
6169 #ifdef CONFIG_HAVE_HW_BREAKPOINT
6171 * hw_breakpoint is a bit difficult here..
6173 if (attr
->type
== PERF_TYPE_BREAKPOINT
)
6174 event
->hw
.bp_target
= task
;
6178 if (!overflow_handler
&& parent_event
) {
6179 overflow_handler
= parent_event
->overflow_handler
;
6180 context
= parent_event
->overflow_handler_context
;
6183 event
->overflow_handler
= overflow_handler
;
6184 event
->overflow_handler_context
= context
;
6187 event
->state
= PERF_EVENT_STATE_OFF
;
6192 hwc
->sample_period
= attr
->sample_period
;
6193 if (attr
->freq
&& attr
->sample_freq
)
6194 hwc
->sample_period
= 1;
6195 hwc
->last_period
= hwc
->sample_period
;
6197 local64_set(&hwc
->period_left
, hwc
->sample_period
);
6200 * we currently do not support PERF_FORMAT_GROUP on inherited events
6202 if (attr
->inherit
&& (attr
->read_format
& PERF_FORMAT_GROUP
))
6205 pmu
= perf_init_event(event
);
6211 else if (IS_ERR(pmu
))
6216 put_pid_ns(event
->ns
);
6218 return ERR_PTR(err
);
6221 if (!event
->parent
) {
6222 if (event
->attach_state
& PERF_ATTACH_TASK
)
6223 static_key_slow_inc(&perf_sched_events
.key
);
6224 if (event
->attr
.mmap
|| event
->attr
.mmap_data
)
6225 atomic_inc(&nr_mmap_events
);
6226 if (event
->attr
.comm
)
6227 atomic_inc(&nr_comm_events
);
6228 if (event
->attr
.task
)
6229 atomic_inc(&nr_task_events
);
6230 if (event
->attr
.sample_type
& PERF_SAMPLE_CALLCHAIN
) {
6231 err
= get_callchain_buffers();
6234 return ERR_PTR(err
);
6237 if (has_branch_stack(event
)) {
6238 static_key_slow_inc(&perf_sched_events
.key
);
6239 if (!(event
->attach_state
& PERF_ATTACH_TASK
))
6240 atomic_inc(&per_cpu(perf_branch_stack_events
,
6248 static int perf_copy_attr(struct perf_event_attr __user
*uattr
,
6249 struct perf_event_attr
*attr
)
6254 if (!access_ok(VERIFY_WRITE
, uattr
, PERF_ATTR_SIZE_VER0
))
6258 * zero the full structure, so that a short copy will be nice.
6260 memset(attr
, 0, sizeof(*attr
));
6262 ret
= get_user(size
, &uattr
->size
);
6266 if (size
> PAGE_SIZE
) /* silly large */
6269 if (!size
) /* abi compat */
6270 size
= PERF_ATTR_SIZE_VER0
;
6272 if (size
< PERF_ATTR_SIZE_VER0
)
6276 * If we're handed a bigger struct than we know of,
6277 * ensure all the unknown bits are 0 - i.e. new
6278 * user-space does not rely on any kernel feature
6279 * extensions we dont know about yet.
6281 if (size
> sizeof(*attr
)) {
6282 unsigned char __user
*addr
;
6283 unsigned char __user
*end
;
6286 addr
= (void __user
*)uattr
+ sizeof(*attr
);
6287 end
= (void __user
*)uattr
+ size
;
6289 for (; addr
< end
; addr
++) {
6290 ret
= get_user(val
, addr
);
6296 size
= sizeof(*attr
);
6299 ret
= copy_from_user(attr
, uattr
, size
);
6303 if (attr
->__reserved_1
)
6306 if (attr
->sample_type
& ~(PERF_SAMPLE_MAX
-1))
6309 if (attr
->read_format
& ~(PERF_FORMAT_MAX
-1))
6312 if (attr
->sample_type
& PERF_SAMPLE_BRANCH_STACK
) {
6313 u64 mask
= attr
->branch_sample_type
;
6315 /* only using defined bits */
6316 if (mask
& ~(PERF_SAMPLE_BRANCH_MAX
-1))
6319 /* at least one branch bit must be set */
6320 if (!(mask
& ~PERF_SAMPLE_BRANCH_PLM_ALL
))
6323 /* kernel level capture: check permissions */
6324 if ((mask
& PERF_SAMPLE_BRANCH_PERM_PLM
)
6325 && perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN
))
6328 /* propagate priv level, when not set for branch */
6329 if (!(mask
& PERF_SAMPLE_BRANCH_PLM_ALL
)) {
6331 /* exclude_kernel checked on syscall entry */
6332 if (!attr
->exclude_kernel
)
6333 mask
|= PERF_SAMPLE_BRANCH_KERNEL
;
6335 if (!attr
->exclude_user
)
6336 mask
|= PERF_SAMPLE_BRANCH_USER
;
6338 if (!attr
->exclude_hv
)
6339 mask
|= PERF_SAMPLE_BRANCH_HV
;
6341 * adjust user setting (for HW filter setup)
6343 attr
->branch_sample_type
= mask
;
6347 if (attr
->sample_type
& PERF_SAMPLE_REGS_USER
) {
6348 ret
= perf_reg_validate(attr
->sample_regs_user
);
6353 if (attr
->sample_type
& PERF_SAMPLE_STACK_USER
) {
6354 if (!arch_perf_have_user_stack_dump())
6358 * We have __u32 type for the size, but so far
6359 * we can only use __u16 as maximum due to the
6360 * __u16 sample size limit.
6362 if (attr
->sample_stack_user
>= USHRT_MAX
)
6364 else if (!IS_ALIGNED(attr
->sample_stack_user
, sizeof(u64
)))
6372 put_user(sizeof(*attr
), &uattr
->size
);
6378 perf_event_set_output(struct perf_event
*event
, struct perf_event
*output_event
)
6380 struct ring_buffer
*rb
= NULL
, *old_rb
= NULL
;
6386 /* don't allow circular references */
6387 if (event
== output_event
)
6391 * Don't allow cross-cpu buffers
6393 if (output_event
->cpu
!= event
->cpu
)
6397 * If its not a per-cpu rb, it must be the same task.
6399 if (output_event
->cpu
== -1 && output_event
->ctx
!= event
->ctx
)
6403 mutex_lock(&event
->mmap_mutex
);
6404 /* Can't redirect output if we've got an active mmap() */
6405 if (atomic_read(&event
->mmap_count
))
6409 /* get the rb we want to redirect to */
6410 rb
= ring_buffer_get(output_event
);
6416 rcu_assign_pointer(event
->rb
, rb
);
6418 ring_buffer_detach(event
, old_rb
);
6421 mutex_unlock(&event
->mmap_mutex
);
6424 ring_buffer_put(old_rb
);
6430 * sys_perf_event_open - open a performance event, associate it to a task/cpu
6432 * @attr_uptr: event_id type attributes for monitoring/sampling
6435 * @group_fd: group leader event fd
6437 SYSCALL_DEFINE5(perf_event_open
,
6438 struct perf_event_attr __user
*, attr_uptr
,
6439 pid_t
, pid
, int, cpu
, int, group_fd
, unsigned long, flags
)
6441 struct perf_event
*group_leader
= NULL
, *output_event
= NULL
;
6442 struct perf_event
*event
, *sibling
;
6443 struct perf_event_attr attr
;
6444 struct perf_event_context
*ctx
;
6445 struct file
*event_file
= NULL
;
6446 struct file
*group_file
= NULL
;
6447 struct task_struct
*task
= NULL
;
6451 int fput_needed
= 0;
6454 /* for future expandability... */
6455 if (flags
& ~PERF_FLAG_ALL
)
6458 err
= perf_copy_attr(attr_uptr
, &attr
);
6462 if (!attr
.exclude_kernel
) {
6463 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN
))
6468 if (attr
.sample_freq
> sysctl_perf_event_sample_rate
)
6473 * In cgroup mode, the pid argument is used to pass the fd
6474 * opened to the cgroup directory in cgroupfs. The cpu argument
6475 * designates the cpu on which to monitor threads from that
6478 if ((flags
& PERF_FLAG_PID_CGROUP
) && (pid
== -1 || cpu
== -1))
6481 event_fd
= get_unused_fd_flags(O_RDWR
);
6485 if (group_fd
!= -1) {
6486 group_file
= perf_fget_light(group_fd
, &fput_needed
);
6487 if (IS_ERR(group_file
)) {
6488 err
= PTR_ERR(group_file
);
6491 group_leader
= group_file
->private_data
;
6492 if (flags
& PERF_FLAG_FD_OUTPUT
)
6493 output_event
= group_leader
;
6494 if (flags
& PERF_FLAG_FD_NO_GROUP
)
6495 group_leader
= NULL
;
6498 if (pid
!= -1 && !(flags
& PERF_FLAG_PID_CGROUP
)) {
6499 task
= find_lively_task_by_vpid(pid
);
6501 err
= PTR_ERR(task
);
6508 event
= perf_event_alloc(&attr
, cpu
, task
, group_leader
, NULL
,
6510 if (IS_ERR(event
)) {
6511 err
= PTR_ERR(event
);
6515 if (flags
& PERF_FLAG_PID_CGROUP
) {
6516 err
= perf_cgroup_connect(pid
, event
, &attr
, group_leader
);
6521 * - that has cgroup constraint on event->cpu
6522 * - that may need work on context switch
6524 atomic_inc(&per_cpu(perf_cgroup_events
, event
->cpu
));
6525 static_key_slow_inc(&perf_sched_events
.key
);
6529 * Special case software events and allow them to be part of
6530 * any hardware group.
6535 (is_software_event(event
) != is_software_event(group_leader
))) {
6536 if (is_software_event(event
)) {
6538 * If event and group_leader are not both a software
6539 * event, and event is, then group leader is not.
6541 * Allow the addition of software events to !software
6542 * groups, this is safe because software events never
6545 pmu
= group_leader
->pmu
;
6546 } else if (is_software_event(group_leader
) &&
6547 (group_leader
->group_flags
& PERF_GROUP_SOFTWARE
)) {
6549 * In case the group is a pure software group, and we
6550 * try to add a hardware event, move the whole group to
6551 * the hardware context.
6558 * Get the target context (task or percpu):
6560 ctx
= find_get_context(pmu
, task
, event
->cpu
);
6567 put_task_struct(task
);
6572 * Look up the group leader (we will attach this event to it):
6578 * Do not allow a recursive hierarchy (this new sibling
6579 * becoming part of another group-sibling):
6581 if (group_leader
->group_leader
!= group_leader
)
6584 * Do not allow to attach to a group in a different
6585 * task or CPU context:
6588 if (group_leader
->ctx
->type
!= ctx
->type
)
6591 if (group_leader
->ctx
!= ctx
)
6596 * Only a group leader can be exclusive or pinned
6598 if (attr
.exclusive
|| attr
.pinned
)
6603 err
= perf_event_set_output(event
, output_event
);
6608 event_file
= anon_inode_getfile("[perf_event]", &perf_fops
, event
, O_RDWR
);
6609 if (IS_ERR(event_file
)) {
6610 err
= PTR_ERR(event_file
);
6615 struct perf_event_context
*gctx
= group_leader
->ctx
;
6617 mutex_lock(&gctx
->mutex
);
6618 perf_remove_from_context(group_leader
);
6619 list_for_each_entry(sibling
, &group_leader
->sibling_list
,
6621 perf_remove_from_context(sibling
);
6624 mutex_unlock(&gctx
->mutex
);
6628 WARN_ON_ONCE(ctx
->parent_ctx
);
6629 mutex_lock(&ctx
->mutex
);
6633 perf_install_in_context(ctx
, group_leader
, event
->cpu
);
6635 list_for_each_entry(sibling
, &group_leader
->sibling_list
,
6637 perf_install_in_context(ctx
, sibling
, event
->cpu
);
6642 perf_install_in_context(ctx
, event
, event
->cpu
);
6644 perf_unpin_context(ctx
);
6645 mutex_unlock(&ctx
->mutex
);
6649 event
->owner
= current
;
6651 mutex_lock(¤t
->perf_event_mutex
);
6652 list_add_tail(&event
->owner_entry
, ¤t
->perf_event_list
);
6653 mutex_unlock(¤t
->perf_event_mutex
);
6656 * Precalculate sample_data sizes
6658 perf_event__header_size(event
);
6659 perf_event__id_header_size(event
);
6662 * Drop the reference on the group_event after placing the
6663 * new event on the sibling_list. This ensures destruction
6664 * of the group leader will find the pointer to itself in
6665 * perf_group_detach().
6667 fput_light(group_file
, fput_needed
);
6668 fd_install(event_fd
, event_file
);
6672 perf_unpin_context(ctx
);
6679 put_task_struct(task
);
6681 fput_light(group_file
, fput_needed
);
6683 put_unused_fd(event_fd
);
6688 * perf_event_create_kernel_counter
6690 * @attr: attributes of the counter to create
6691 * @cpu: cpu in which the counter is bound
6692 * @task: task to profile (NULL for percpu)
6695 perf_event_create_kernel_counter(struct perf_event_attr
*attr
, int cpu
,
6696 struct task_struct
*task
,
6697 perf_overflow_handler_t overflow_handler
,
6700 struct perf_event_context
*ctx
;
6701 struct perf_event
*event
;
6705 * Get the target context (task or percpu):
6708 event
= perf_event_alloc(attr
, cpu
, task
, NULL
, NULL
,
6709 overflow_handler
, context
);
6710 if (IS_ERR(event
)) {
6711 err
= PTR_ERR(event
);
6715 ctx
= find_get_context(event
->pmu
, task
, cpu
);
6721 WARN_ON_ONCE(ctx
->parent_ctx
);
6722 mutex_lock(&ctx
->mutex
);
6723 perf_install_in_context(ctx
, event
, cpu
);
6725 perf_unpin_context(ctx
);
6726 mutex_unlock(&ctx
->mutex
);
6733 return ERR_PTR(err
);
6735 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter
);
6737 void perf_pmu_migrate_context(struct pmu
*pmu
, int src_cpu
, int dst_cpu
)
6739 struct perf_event_context
*src_ctx
;
6740 struct perf_event_context
*dst_ctx
;
6741 struct perf_event
*event
, *tmp
;
6744 src_ctx
= &per_cpu_ptr(pmu
->pmu_cpu_context
, src_cpu
)->ctx
;
6745 dst_ctx
= &per_cpu_ptr(pmu
->pmu_cpu_context
, dst_cpu
)->ctx
;
6747 mutex_lock(&src_ctx
->mutex
);
6748 list_for_each_entry_safe(event
, tmp
, &src_ctx
->event_list
,
6750 perf_remove_from_context(event
);
6752 list_add(&event
->event_entry
, &events
);
6754 mutex_unlock(&src_ctx
->mutex
);
6758 mutex_lock(&dst_ctx
->mutex
);
6759 list_for_each_entry_safe(event
, tmp
, &events
, event_entry
) {
6760 list_del(&event
->event_entry
);
6761 if (event
->state
>= PERF_EVENT_STATE_OFF
)
6762 event
->state
= PERF_EVENT_STATE_INACTIVE
;
6763 perf_install_in_context(dst_ctx
, event
, dst_cpu
);
6766 mutex_unlock(&dst_ctx
->mutex
);
6768 EXPORT_SYMBOL_GPL(perf_pmu_migrate_context
);
6770 static void sync_child_event(struct perf_event
*child_event
,
6771 struct task_struct
*child
)
6773 struct perf_event
*parent_event
= child_event
->parent
;
6776 if (child_event
->attr
.inherit_stat
)
6777 perf_event_read_event(child_event
, child
);
6779 child_val
= perf_event_count(child_event
);
6782 * Add back the child's count to the parent's count:
6784 atomic64_add(child_val
, &parent_event
->child_count
);
6785 atomic64_add(child_event
->total_time_enabled
,
6786 &parent_event
->child_total_time_enabled
);
6787 atomic64_add(child_event
->total_time_running
,
6788 &parent_event
->child_total_time_running
);
6791 * Remove this event from the parent's list
6793 WARN_ON_ONCE(parent_event
->ctx
->parent_ctx
);
6794 mutex_lock(&parent_event
->child_mutex
);
6795 list_del_init(&child_event
->child_list
);
6796 mutex_unlock(&parent_event
->child_mutex
);
6799 * Release the parent event, if this was the last
6802 put_event(parent_event
);
6806 __perf_event_exit_task(struct perf_event
*child_event
,
6807 struct perf_event_context
*child_ctx
,
6808 struct task_struct
*child
)
6810 if (child_event
->parent
) {
6811 raw_spin_lock_irq(&child_ctx
->lock
);
6812 perf_group_detach(child_event
);
6813 raw_spin_unlock_irq(&child_ctx
->lock
);
6816 perf_remove_from_context(child_event
);
6819 * It can happen that the parent exits first, and has events
6820 * that are still around due to the child reference. These
6821 * events need to be zapped.
6823 if (child_event
->parent
) {
6824 sync_child_event(child_event
, child
);
6825 free_event(child_event
);
6829 static void perf_event_exit_task_context(struct task_struct
*child
, int ctxn
)
6831 struct perf_event
*child_event
, *tmp
;
6832 struct perf_event_context
*child_ctx
;
6833 unsigned long flags
;
6835 if (likely(!child
->perf_event_ctxp
[ctxn
])) {
6836 perf_event_task(child
, NULL
, 0);
6840 local_irq_save(flags
);
6842 * We can't reschedule here because interrupts are disabled,
6843 * and either child is current or it is a task that can't be
6844 * scheduled, so we are now safe from rescheduling changing
6847 child_ctx
= rcu_dereference_raw(child
->perf_event_ctxp
[ctxn
]);
6850 * Take the context lock here so that if find_get_context is
6851 * reading child->perf_event_ctxp, we wait until it has
6852 * incremented the context's refcount before we do put_ctx below.
6854 raw_spin_lock(&child_ctx
->lock
);
6855 task_ctx_sched_out(child_ctx
);
6856 child
->perf_event_ctxp
[ctxn
] = NULL
;
6858 * If this context is a clone; unclone it so it can't get
6859 * swapped to another process while we're removing all
6860 * the events from it.
6862 unclone_ctx(child_ctx
);
6863 update_context_time(child_ctx
);
6864 raw_spin_unlock_irqrestore(&child_ctx
->lock
, flags
);
6867 * Report the task dead after unscheduling the events so that we
6868 * won't get any samples after PERF_RECORD_EXIT. We can however still
6869 * get a few PERF_RECORD_READ events.
6871 perf_event_task(child
, child_ctx
, 0);
6874 * We can recurse on the same lock type through:
6876 * __perf_event_exit_task()
6877 * sync_child_event()
6879 * mutex_lock(&ctx->mutex)
6881 * But since its the parent context it won't be the same instance.
6883 mutex_lock(&child_ctx
->mutex
);
6886 list_for_each_entry_safe(child_event
, tmp
, &child_ctx
->pinned_groups
,
6888 __perf_event_exit_task(child_event
, child_ctx
, child
);
6890 list_for_each_entry_safe(child_event
, tmp
, &child_ctx
->flexible_groups
,
6892 __perf_event_exit_task(child_event
, child_ctx
, child
);
6895 * If the last event was a group event, it will have appended all
6896 * its siblings to the list, but we obtained 'tmp' before that which
6897 * will still point to the list head terminating the iteration.
6899 if (!list_empty(&child_ctx
->pinned_groups
) ||
6900 !list_empty(&child_ctx
->flexible_groups
))
6903 mutex_unlock(&child_ctx
->mutex
);
6909 * When a child task exits, feed back event values to parent events.
6911 void perf_event_exit_task(struct task_struct
*child
)
6913 struct perf_event
*event
, *tmp
;
6916 mutex_lock(&child
->perf_event_mutex
);
6917 list_for_each_entry_safe(event
, tmp
, &child
->perf_event_list
,
6919 list_del_init(&event
->owner_entry
);
6922 * Ensure the list deletion is visible before we clear
6923 * the owner, closes a race against perf_release() where
6924 * we need to serialize on the owner->perf_event_mutex.
6927 event
->owner
= NULL
;
6929 mutex_unlock(&child
->perf_event_mutex
);
6931 for_each_task_context_nr(ctxn
)
6932 perf_event_exit_task_context(child
, ctxn
);
6935 static void perf_free_event(struct perf_event
*event
,
6936 struct perf_event_context
*ctx
)
6938 struct perf_event
*parent
= event
->parent
;
6940 if (WARN_ON_ONCE(!parent
))
6943 mutex_lock(&parent
->child_mutex
);
6944 list_del_init(&event
->child_list
);
6945 mutex_unlock(&parent
->child_mutex
);
6949 perf_group_detach(event
);
6950 list_del_event(event
, ctx
);
6955 * free an unexposed, unused context as created by inheritance by
6956 * perf_event_init_task below, used by fork() in case of fail.
6958 void perf_event_free_task(struct task_struct
*task
)
6960 struct perf_event_context
*ctx
;
6961 struct perf_event
*event
, *tmp
;
6964 for_each_task_context_nr(ctxn
) {
6965 ctx
= task
->perf_event_ctxp
[ctxn
];
6969 mutex_lock(&ctx
->mutex
);
6971 list_for_each_entry_safe(event
, tmp
, &ctx
->pinned_groups
,
6973 perf_free_event(event
, ctx
);
6975 list_for_each_entry_safe(event
, tmp
, &ctx
->flexible_groups
,
6977 perf_free_event(event
, ctx
);
6979 if (!list_empty(&ctx
->pinned_groups
) ||
6980 !list_empty(&ctx
->flexible_groups
))
6983 mutex_unlock(&ctx
->mutex
);
6989 void perf_event_delayed_put(struct task_struct
*task
)
6993 for_each_task_context_nr(ctxn
)
6994 WARN_ON_ONCE(task
->perf_event_ctxp
[ctxn
]);
6998 * inherit a event from parent task to child task:
7000 static struct perf_event
*
7001 inherit_event(struct perf_event
*parent_event
,
7002 struct task_struct
*parent
,
7003 struct perf_event_context
*parent_ctx
,
7004 struct task_struct
*child
,
7005 struct perf_event
*group_leader
,
7006 struct perf_event_context
*child_ctx
)
7008 struct perf_event
*child_event
;
7009 unsigned long flags
;
7012 * Instead of creating recursive hierarchies of events,
7013 * we link inherited events back to the original parent,
7014 * which has a filp for sure, which we use as the reference
7017 if (parent_event
->parent
)
7018 parent_event
= parent_event
->parent
;
7020 child_event
= perf_event_alloc(&parent_event
->attr
,
7023 group_leader
, parent_event
,
7025 if (IS_ERR(child_event
))
7028 if (!atomic_long_inc_not_zero(&parent_event
->refcount
)) {
7029 free_event(child_event
);
7036 * Make the child state follow the state of the parent event,
7037 * not its attr.disabled bit. We hold the parent's mutex,
7038 * so we won't race with perf_event_{en, dis}able_family.
7040 if (parent_event
->state
>= PERF_EVENT_STATE_INACTIVE
)
7041 child_event
->state
= PERF_EVENT_STATE_INACTIVE
;
7043 child_event
->state
= PERF_EVENT_STATE_OFF
;
7045 if (parent_event
->attr
.freq
) {
7046 u64 sample_period
= parent_event
->hw
.sample_period
;
7047 struct hw_perf_event
*hwc
= &child_event
->hw
;
7049 hwc
->sample_period
= sample_period
;
7050 hwc
->last_period
= sample_period
;
7052 local64_set(&hwc
->period_left
, sample_period
);
7055 child_event
->ctx
= child_ctx
;
7056 child_event
->overflow_handler
= parent_event
->overflow_handler
;
7057 child_event
->overflow_handler_context
7058 = parent_event
->overflow_handler_context
;
7061 * Precalculate sample_data sizes
7063 perf_event__header_size(child_event
);
7064 perf_event__id_header_size(child_event
);
7067 * Link it up in the child's context:
7069 raw_spin_lock_irqsave(&child_ctx
->lock
, flags
);
7070 add_event_to_ctx(child_event
, child_ctx
);
7071 raw_spin_unlock_irqrestore(&child_ctx
->lock
, flags
);
7074 * Link this into the parent event's child list
7076 WARN_ON_ONCE(parent_event
->ctx
->parent_ctx
);
7077 mutex_lock(&parent_event
->child_mutex
);
7078 list_add_tail(&child_event
->child_list
, &parent_event
->child_list
);
7079 mutex_unlock(&parent_event
->child_mutex
);
7084 static int inherit_group(struct perf_event
*parent_event
,
7085 struct task_struct
*parent
,
7086 struct perf_event_context
*parent_ctx
,
7087 struct task_struct
*child
,
7088 struct perf_event_context
*child_ctx
)
7090 struct perf_event
*leader
;
7091 struct perf_event
*sub
;
7092 struct perf_event
*child_ctr
;
7094 leader
= inherit_event(parent_event
, parent
, parent_ctx
,
7095 child
, NULL
, child_ctx
);
7097 return PTR_ERR(leader
);
7098 list_for_each_entry(sub
, &parent_event
->sibling_list
, group_entry
) {
7099 child_ctr
= inherit_event(sub
, parent
, parent_ctx
,
7100 child
, leader
, child_ctx
);
7101 if (IS_ERR(child_ctr
))
7102 return PTR_ERR(child_ctr
);
7108 inherit_task_group(struct perf_event
*event
, struct task_struct
*parent
,
7109 struct perf_event_context
*parent_ctx
,
7110 struct task_struct
*child
, int ctxn
,
7114 struct perf_event_context
*child_ctx
;
7116 if (!event
->attr
.inherit
) {
7121 child_ctx
= child
->perf_event_ctxp
[ctxn
];
7124 * This is executed from the parent task context, so
7125 * inherit events that have been marked for cloning.
7126 * First allocate and initialize a context for the
7130 child_ctx
= alloc_perf_context(event
->pmu
, child
);
7134 child
->perf_event_ctxp
[ctxn
] = child_ctx
;
7137 ret
= inherit_group(event
, parent
, parent_ctx
,
7147 * Initialize the perf_event context in task_struct
7149 int perf_event_init_context(struct task_struct
*child
, int ctxn
)
7151 struct perf_event_context
*child_ctx
, *parent_ctx
;
7152 struct perf_event_context
*cloned_ctx
;
7153 struct perf_event
*event
;
7154 struct task_struct
*parent
= current
;
7155 int inherited_all
= 1;
7156 unsigned long flags
;
7159 if (likely(!parent
->perf_event_ctxp
[ctxn
]))
7163 * If the parent's context is a clone, pin it so it won't get
7166 parent_ctx
= perf_pin_task_context(parent
, ctxn
);
7169 * No need to check if parent_ctx != NULL here; since we saw
7170 * it non-NULL earlier, the only reason for it to become NULL
7171 * is if we exit, and since we're currently in the middle of
7172 * a fork we can't be exiting at the same time.
7176 * Lock the parent list. No need to lock the child - not PID
7177 * hashed yet and not running, so nobody can access it.
7179 mutex_lock(&parent_ctx
->mutex
);
7182 * We dont have to disable NMIs - we are only looking at
7183 * the list, not manipulating it:
7185 list_for_each_entry(event
, &parent_ctx
->pinned_groups
, group_entry
) {
7186 ret
= inherit_task_group(event
, parent
, parent_ctx
,
7187 child
, ctxn
, &inherited_all
);
7193 * We can't hold ctx->lock when iterating the ->flexible_group list due
7194 * to allocations, but we need to prevent rotation because
7195 * rotate_ctx() will change the list from interrupt context.
7197 raw_spin_lock_irqsave(&parent_ctx
->lock
, flags
);
7198 parent_ctx
->rotate_disable
= 1;
7199 raw_spin_unlock_irqrestore(&parent_ctx
->lock
, flags
);
7201 list_for_each_entry(event
, &parent_ctx
->flexible_groups
, group_entry
) {
7202 ret
= inherit_task_group(event
, parent
, parent_ctx
,
7203 child
, ctxn
, &inherited_all
);
7208 raw_spin_lock_irqsave(&parent_ctx
->lock
, flags
);
7209 parent_ctx
->rotate_disable
= 0;
7211 child_ctx
= child
->perf_event_ctxp
[ctxn
];
7213 if (child_ctx
&& inherited_all
) {
7215 * Mark the child context as a clone of the parent
7216 * context, or of whatever the parent is a clone of.
7218 * Note that if the parent is a clone, the holding of
7219 * parent_ctx->lock avoids it from being uncloned.
7221 cloned_ctx
= parent_ctx
->parent_ctx
;
7223 child_ctx
->parent_ctx
= cloned_ctx
;
7224 child_ctx
->parent_gen
= parent_ctx
->parent_gen
;
7226 child_ctx
->parent_ctx
= parent_ctx
;
7227 child_ctx
->parent_gen
= parent_ctx
->generation
;
7229 get_ctx(child_ctx
->parent_ctx
);
7232 raw_spin_unlock_irqrestore(&parent_ctx
->lock
, flags
);
7233 mutex_unlock(&parent_ctx
->mutex
);
7235 perf_unpin_context(parent_ctx
);
7236 put_ctx(parent_ctx
);
7242 * Initialize the perf_event context in task_struct
7244 int perf_event_init_task(struct task_struct
*child
)
7248 memset(child
->perf_event_ctxp
, 0, sizeof(child
->perf_event_ctxp
));
7249 mutex_init(&child
->perf_event_mutex
);
7250 INIT_LIST_HEAD(&child
->perf_event_list
);
7252 for_each_task_context_nr(ctxn
) {
7253 ret
= perf_event_init_context(child
, ctxn
);
7261 static void __init
perf_event_init_all_cpus(void)
7263 struct swevent_htable
*swhash
;
7266 for_each_possible_cpu(cpu
) {
7267 swhash
= &per_cpu(swevent_htable
, cpu
);
7268 mutex_init(&swhash
->hlist_mutex
);
7269 INIT_LIST_HEAD(&per_cpu(rotation_list
, cpu
));
7273 static void __cpuinit
perf_event_init_cpu(int cpu
)
7275 struct swevent_htable
*swhash
= &per_cpu(swevent_htable
, cpu
);
7277 mutex_lock(&swhash
->hlist_mutex
);
7278 if (swhash
->hlist_refcount
> 0) {
7279 struct swevent_hlist
*hlist
;
7281 hlist
= kzalloc_node(sizeof(*hlist
), GFP_KERNEL
, cpu_to_node(cpu
));
7283 rcu_assign_pointer(swhash
->swevent_hlist
, hlist
);
7285 mutex_unlock(&swhash
->hlist_mutex
);
7288 #if defined CONFIG_HOTPLUG_CPU || defined CONFIG_KEXEC
7289 static void perf_pmu_rotate_stop(struct pmu
*pmu
)
7291 struct perf_cpu_context
*cpuctx
= this_cpu_ptr(pmu
->pmu_cpu_context
);
7293 WARN_ON(!irqs_disabled());
7295 list_del_init(&cpuctx
->rotation_list
);
7298 static void __perf_event_exit_context(void *__info
)
7300 struct perf_event_context
*ctx
= __info
;
7301 struct perf_event
*event
, *tmp
;
7303 perf_pmu_rotate_stop(ctx
->pmu
);
7305 list_for_each_entry_safe(event
, tmp
, &ctx
->pinned_groups
, group_entry
)
7306 __perf_remove_from_context(event
);
7307 list_for_each_entry_safe(event
, tmp
, &ctx
->flexible_groups
, group_entry
)
7308 __perf_remove_from_context(event
);
7311 static void perf_event_exit_cpu_context(int cpu
)
7313 struct perf_event_context
*ctx
;
7317 idx
= srcu_read_lock(&pmus_srcu
);
7318 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
7319 ctx
= &per_cpu_ptr(pmu
->pmu_cpu_context
, cpu
)->ctx
;
7321 mutex_lock(&ctx
->mutex
);
7322 smp_call_function_single(cpu
, __perf_event_exit_context
, ctx
, 1);
7323 mutex_unlock(&ctx
->mutex
);
7325 srcu_read_unlock(&pmus_srcu
, idx
);
7328 static void perf_event_exit_cpu(int cpu
)
7330 struct swevent_htable
*swhash
= &per_cpu(swevent_htable
, cpu
);
7332 mutex_lock(&swhash
->hlist_mutex
);
7333 swevent_hlist_release(swhash
);
7334 mutex_unlock(&swhash
->hlist_mutex
);
7336 perf_event_exit_cpu_context(cpu
);
7339 static inline void perf_event_exit_cpu(int cpu
) { }
7343 perf_reboot(struct notifier_block
*notifier
, unsigned long val
, void *v
)
7347 for_each_online_cpu(cpu
)
7348 perf_event_exit_cpu(cpu
);
7354 * Run the perf reboot notifier at the very last possible moment so that
7355 * the generic watchdog code runs as long as possible.
7357 static struct notifier_block perf_reboot_notifier
= {
7358 .notifier_call
= perf_reboot
,
7359 .priority
= INT_MIN
,
7362 static int __cpuinit
7363 perf_cpu_notify(struct notifier_block
*self
, unsigned long action
, void *hcpu
)
7365 unsigned int cpu
= (long)hcpu
;
7367 switch (action
& ~CPU_TASKS_FROZEN
) {
7369 case CPU_UP_PREPARE
:
7370 case CPU_DOWN_FAILED
:
7371 perf_event_init_cpu(cpu
);
7374 case CPU_UP_CANCELED
:
7375 case CPU_DOWN_PREPARE
:
7376 perf_event_exit_cpu(cpu
);
7386 void __init
perf_event_init(void)
7392 perf_event_init_all_cpus();
7393 init_srcu_struct(&pmus_srcu
);
7394 perf_pmu_register(&perf_swevent
, "software", PERF_TYPE_SOFTWARE
);
7395 perf_pmu_register(&perf_cpu_clock
, NULL
, -1);
7396 perf_pmu_register(&perf_task_clock
, NULL
, -1);
7398 perf_cpu_notifier(perf_cpu_notify
);
7399 register_reboot_notifier(&perf_reboot_notifier
);
7401 ret
= init_hw_breakpoint();
7402 WARN(ret
, "hw_breakpoint initialization failed with: %d", ret
);
7404 /* do not patch jump label more than once per second */
7405 jump_label_rate_limit(&perf_sched_events
, HZ
);
7408 * Build time assertion that we keep the data_head at the intended
7409 * location. IOW, validation we got the __reserved[] size right.
7411 BUILD_BUG_ON((offsetof(struct perf_event_mmap_page
, data_head
))
7415 static int __init
perf_event_sysfs_init(void)
7420 mutex_lock(&pmus_lock
);
7422 ret
= bus_register(&pmu_bus
);
7426 list_for_each_entry(pmu
, &pmus
, entry
) {
7427 if (!pmu
->name
|| pmu
->type
< 0)
7430 ret
= pmu_dev_alloc(pmu
);
7431 WARN(ret
, "Failed to register pmu: %s, reason %d\n", pmu
->name
, ret
);
7433 pmu_bus_running
= 1;
7437 mutex_unlock(&pmus_lock
);
7441 device_initcall(perf_event_sysfs_init
);
7443 #ifdef CONFIG_CGROUP_PERF
7444 static struct cgroup_subsys_state
*perf_cgroup_create(struct cgroup
*cont
)
7446 struct perf_cgroup
*jc
;
7448 jc
= kzalloc(sizeof(*jc
), GFP_KERNEL
);
7450 return ERR_PTR(-ENOMEM
);
7452 jc
->info
= alloc_percpu(struct perf_cgroup_info
);
7455 return ERR_PTR(-ENOMEM
);
7461 static void perf_cgroup_destroy(struct cgroup
*cont
)
7463 struct perf_cgroup
*jc
;
7464 jc
= container_of(cgroup_subsys_state(cont
, perf_subsys_id
),
7465 struct perf_cgroup
, css
);
7466 free_percpu(jc
->info
);
7470 static int __perf_cgroup_move(void *info
)
7472 struct task_struct
*task
= info
;
7473 perf_cgroup_switch(task
, PERF_CGROUP_SWOUT
| PERF_CGROUP_SWIN
);
7477 static void perf_cgroup_attach(struct cgroup
*cgrp
, struct cgroup_taskset
*tset
)
7479 struct task_struct
*task
;
7481 cgroup_taskset_for_each(task
, cgrp
, tset
)
7482 task_function_call(task
, __perf_cgroup_move
, task
);
7485 static void perf_cgroup_exit(struct cgroup
*cgrp
, struct cgroup
*old_cgrp
,
7486 struct task_struct
*task
)
7489 * cgroup_exit() is called in the copy_process() failure path.
7490 * Ignore this case since the task hasn't ran yet, this avoids
7491 * trying to poke a half freed task state from generic code.
7493 if (!(task
->flags
& PF_EXITING
))
7496 task_function_call(task
, __perf_cgroup_move
, task
);
7499 struct cgroup_subsys perf_subsys
= {
7500 .name
= "perf_event",
7501 .subsys_id
= perf_subsys_id
,
7502 .create
= perf_cgroup_create
,
7503 .destroy
= perf_cgroup_destroy
,
7504 .exit
= perf_cgroup_exit
,
7505 .attach
= perf_cgroup_attach
,
7508 * perf_event cgroup doesn't handle nesting correctly.
7509 * ctx->nr_cgroups adjustments should be propagated through the
7510 * cgroup hierarchy. Fix it and remove the following.
7512 .broken_hierarchy
= true,
7514 #endif /* CONFIG_CGROUP_PERF */