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>
42 #include <asm/irq_regs.h>
44 struct remote_function_call {
45 struct task_struct *p;
46 int (*func)(void *info);
51 static void remote_function(void *data)
53 struct remote_function_call *tfc = data;
54 struct task_struct *p = tfc->p;
58 if (task_cpu(p) != smp_processor_id() || !task_curr(p))
62 tfc->ret = tfc->func(tfc->info);
66 * task_function_call - call a function on the cpu on which a task runs
67 * @p: the task to evaluate
68 * @func: the function to be called
69 * @info: the function call argument
71 * Calls the function @func when the task is currently running. This might
72 * be on the current CPU, which just calls the function directly
74 * returns: @func return value, or
75 * -ESRCH - when the process isn't running
76 * -EAGAIN - when the process moved away
79 task_function_call(struct task_struct *p, int (*func) (void *info), void *info)
81 struct remote_function_call data = {
85 .ret = -ESRCH, /* No such (running) process */
89 smp_call_function_single(task_cpu(p), remote_function, &data, 1);
95 * cpu_function_call - call a function on the cpu
96 * @func: the function to be called
97 * @info: the function call argument
99 * Calls the function @func on the remote cpu.
101 * returns: @func return value or -ENXIO when the cpu is offline
103 static int cpu_function_call(int cpu, int (*func) (void *info), void *info)
105 struct remote_function_call data = {
109 .ret = -ENXIO, /* No such CPU */
112 smp_call_function_single(cpu, remote_function, &data, 1);
117 #define PERF_FLAG_ALL (PERF_FLAG_FD_NO_GROUP |\
118 PERF_FLAG_FD_OUTPUT |\
119 PERF_FLAG_PID_CGROUP)
122 EVENT_FLEXIBLE = 0x1,
124 EVENT_ALL = EVENT_FLEXIBLE | EVENT_PINNED,
128 * perf_sched_events : >0 events exist
129 * perf_cgroup_events: >0 per-cpu cgroup events exist on this cpu
131 struct static_key_deferred perf_sched_events __read_mostly;
132 static DEFINE_PER_CPU(atomic_t, perf_cgroup_events);
134 static atomic_t nr_mmap_events __read_mostly;
135 static atomic_t nr_comm_events __read_mostly;
136 static atomic_t nr_task_events __read_mostly;
138 static LIST_HEAD(pmus);
139 static DEFINE_MUTEX(pmus_lock);
140 static struct srcu_struct pmus_srcu;
143 * perf event paranoia level:
144 * -1 - not paranoid at all
145 * 0 - disallow raw tracepoint access for unpriv
146 * 1 - disallow cpu events for unpriv
147 * 2 - disallow kernel profiling for unpriv
149 int sysctl_perf_event_paranoid __read_mostly = 1;
151 /* Minimum for 512 kiB + 1 user control page */
152 int sysctl_perf_event_mlock __read_mostly = 512 + (PAGE_SIZE / 1024); /* 'free' kiB per user */
155 * max perf event sample rate
157 #define DEFAULT_MAX_SAMPLE_RATE 100000
158 int sysctl_perf_event_sample_rate __read_mostly = DEFAULT_MAX_SAMPLE_RATE;
159 static int max_samples_per_tick __read_mostly =
160 DIV_ROUND_UP(DEFAULT_MAX_SAMPLE_RATE, HZ);
162 int perf_proc_update_handler(struct ctl_table *table, int write,
163 void __user *buffer, size_t *lenp,
166 int ret = proc_dointvec(table, write, buffer, lenp, ppos);
171 max_samples_per_tick = DIV_ROUND_UP(sysctl_perf_event_sample_rate, HZ);
176 static atomic64_t perf_event_id;
178 static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
179 enum event_type_t event_type);
181 static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
182 enum event_type_t event_type,
183 struct task_struct *task);
185 static void update_context_time(struct perf_event_context *ctx);
186 static u64 perf_event_time(struct perf_event *event);
188 static void ring_buffer_attach(struct perf_event *event,
189 struct ring_buffer *rb);
191 void __weak perf_event_print_debug(void) { }
193 extern __weak const char *perf_pmu_name(void)
198 static inline u64 perf_clock(void)
200 return local_clock();
203 static inline struct perf_cpu_context *
204 __get_cpu_context(struct perf_event_context *ctx)
206 return this_cpu_ptr(ctx->pmu->pmu_cpu_context);
209 static void perf_ctx_lock(struct perf_cpu_context *cpuctx,
210 struct perf_event_context *ctx)
212 raw_spin_lock(&cpuctx->ctx.lock);
214 raw_spin_lock(&ctx->lock);
217 static void perf_ctx_unlock(struct perf_cpu_context *cpuctx,
218 struct perf_event_context *ctx)
221 raw_spin_unlock(&ctx->lock);
222 raw_spin_unlock(&cpuctx->ctx.lock);
225 #ifdef CONFIG_CGROUP_PERF
228 * Must ensure cgroup is pinned (css_get) before calling
229 * this function. In other words, we cannot call this function
230 * if there is no cgroup event for the current CPU context.
232 static inline struct perf_cgroup *
233 perf_cgroup_from_task(struct task_struct *task)
235 return container_of(task_subsys_state(task, perf_subsys_id),
236 struct perf_cgroup, css);
240 perf_cgroup_match(struct perf_event *event)
242 struct perf_event_context *ctx = event->ctx;
243 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
245 return !event->cgrp || event->cgrp == cpuctx->cgrp;
248 static inline void perf_get_cgroup(struct perf_event *event)
250 css_get(&event->cgrp->css);
253 static inline void perf_put_cgroup(struct perf_event *event)
255 css_put(&event->cgrp->css);
258 static inline void perf_detach_cgroup(struct perf_event *event)
260 perf_put_cgroup(event);
264 static inline int is_cgroup_event(struct perf_event *event)
266 return event->cgrp != NULL;
269 static inline u64 perf_cgroup_event_time(struct perf_event *event)
271 struct perf_cgroup_info *t;
273 t = per_cpu_ptr(event->cgrp->info, event->cpu);
277 static inline void __update_cgrp_time(struct perf_cgroup *cgrp)
279 struct perf_cgroup_info *info;
284 info = this_cpu_ptr(cgrp->info);
286 info->time += now - info->timestamp;
287 info->timestamp = now;
290 static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx)
292 struct perf_cgroup *cgrp_out = cpuctx->cgrp;
294 __update_cgrp_time(cgrp_out);
297 static inline void update_cgrp_time_from_event(struct perf_event *event)
299 struct perf_cgroup *cgrp;
302 * ensure we access cgroup data only when needed and
303 * when we know the cgroup is pinned (css_get)
305 if (!is_cgroup_event(event))
308 cgrp = perf_cgroup_from_task(current);
310 * Do not update time when cgroup is not active
312 if (cgrp == event->cgrp)
313 __update_cgrp_time(event->cgrp);
317 perf_cgroup_set_timestamp(struct task_struct *task,
318 struct perf_event_context *ctx)
320 struct perf_cgroup *cgrp;
321 struct perf_cgroup_info *info;
324 * ctx->lock held by caller
325 * ensure we do not access cgroup data
326 * unless we have the cgroup pinned (css_get)
328 if (!task || !ctx->nr_cgroups)
331 cgrp = perf_cgroup_from_task(task);
332 info = this_cpu_ptr(cgrp->info);
333 info->timestamp = ctx->timestamp;
336 #define PERF_CGROUP_SWOUT 0x1 /* cgroup switch out every event */
337 #define PERF_CGROUP_SWIN 0x2 /* cgroup switch in events based on task */
340 * reschedule events based on the cgroup constraint of task.
342 * mode SWOUT : schedule out everything
343 * mode SWIN : schedule in based on cgroup for next
345 void perf_cgroup_switch(struct task_struct *task, int mode)
347 struct perf_cpu_context *cpuctx;
352 * disable interrupts to avoid geting nr_cgroup
353 * changes via __perf_event_disable(). Also
356 local_irq_save(flags);
359 * we reschedule only in the presence of cgroup
360 * constrained events.
364 list_for_each_entry_rcu(pmu, &pmus, entry) {
365 cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
368 * perf_cgroup_events says at least one
369 * context on this CPU has cgroup events.
371 * ctx->nr_cgroups reports the number of cgroup
372 * events for a context.
374 if (cpuctx->ctx.nr_cgroups > 0) {
375 perf_ctx_lock(cpuctx, cpuctx->task_ctx);
376 perf_pmu_disable(cpuctx->ctx.pmu);
378 if (mode & PERF_CGROUP_SWOUT) {
379 cpu_ctx_sched_out(cpuctx, EVENT_ALL);
381 * must not be done before ctxswout due
382 * to event_filter_match() in event_sched_out()
387 if (mode & PERF_CGROUP_SWIN) {
388 WARN_ON_ONCE(cpuctx->cgrp);
389 /* set cgrp before ctxsw in to
390 * allow event_filter_match() to not
391 * have to pass task around
393 cpuctx->cgrp = perf_cgroup_from_task(task);
394 cpu_ctx_sched_in(cpuctx, EVENT_ALL, task);
396 perf_pmu_enable(cpuctx->ctx.pmu);
397 perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
403 local_irq_restore(flags);
406 static inline void perf_cgroup_sched_out(struct task_struct *task,
407 struct task_struct *next)
409 struct perf_cgroup *cgrp1;
410 struct perf_cgroup *cgrp2 = NULL;
413 * we come here when we know perf_cgroup_events > 0
415 cgrp1 = perf_cgroup_from_task(task);
418 * next is NULL when called from perf_event_enable_on_exec()
419 * that will systematically cause a cgroup_switch()
422 cgrp2 = perf_cgroup_from_task(next);
425 * only schedule out current cgroup events if we know
426 * that we are switching to a different cgroup. Otherwise,
427 * do no touch the cgroup events.
430 perf_cgroup_switch(task, PERF_CGROUP_SWOUT);
433 static inline void perf_cgroup_sched_in(struct task_struct *prev,
434 struct task_struct *task)
436 struct perf_cgroup *cgrp1;
437 struct perf_cgroup *cgrp2 = NULL;
440 * we come here when we know perf_cgroup_events > 0
442 cgrp1 = perf_cgroup_from_task(task);
444 /* prev can never be NULL */
445 cgrp2 = perf_cgroup_from_task(prev);
448 * only need to schedule in cgroup events if we are changing
449 * cgroup during ctxsw. Cgroup events were not scheduled
450 * out of ctxsw out if that was not the case.
453 perf_cgroup_switch(task, PERF_CGROUP_SWIN);
456 static inline int perf_cgroup_connect(int fd, struct perf_event *event,
457 struct perf_event_attr *attr,
458 struct perf_event *group_leader)
460 struct perf_cgroup *cgrp;
461 struct cgroup_subsys_state *css;
463 int ret = 0, fput_needed;
465 file = fget_light(fd, &fput_needed);
469 css = cgroup_css_from_dir(file, perf_subsys_id);
475 cgrp = container_of(css, struct perf_cgroup, css);
478 /* must be done before we fput() the file */
479 perf_get_cgroup(event);
482 * all events in a group must monitor
483 * the same cgroup because a task belongs
484 * to only one perf cgroup at a time
486 if (group_leader && group_leader->cgrp != cgrp) {
487 perf_detach_cgroup(event);
491 fput_light(file, fput_needed);
496 perf_cgroup_set_shadow_time(struct perf_event *event, u64 now)
498 struct perf_cgroup_info *t;
499 t = per_cpu_ptr(event->cgrp->info, event->cpu);
500 event->shadow_ctx_time = now - t->timestamp;
504 perf_cgroup_defer_enabled(struct perf_event *event)
507 * when the current task's perf cgroup does not match
508 * the event's, we need to remember to call the
509 * perf_mark_enable() function the first time a task with
510 * a matching perf cgroup is scheduled in.
512 if (is_cgroup_event(event) && !perf_cgroup_match(event))
513 event->cgrp_defer_enabled = 1;
517 perf_cgroup_mark_enabled(struct perf_event *event,
518 struct perf_event_context *ctx)
520 struct perf_event *sub;
521 u64 tstamp = perf_event_time(event);
523 if (!event->cgrp_defer_enabled)
526 event->cgrp_defer_enabled = 0;
528 event->tstamp_enabled = tstamp - event->total_time_enabled;
529 list_for_each_entry(sub, &event->sibling_list, group_entry) {
530 if (sub->state >= PERF_EVENT_STATE_INACTIVE) {
531 sub->tstamp_enabled = tstamp - sub->total_time_enabled;
532 sub->cgrp_defer_enabled = 0;
536 #else /* !CONFIG_CGROUP_PERF */
539 perf_cgroup_match(struct perf_event *event)
544 static inline void perf_detach_cgroup(struct perf_event *event)
547 static inline int is_cgroup_event(struct perf_event *event)
552 static inline u64 perf_cgroup_event_cgrp_time(struct perf_event *event)
557 static inline void update_cgrp_time_from_event(struct perf_event *event)
561 static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx)
565 static inline void perf_cgroup_sched_out(struct task_struct *task,
566 struct task_struct *next)
570 static inline void perf_cgroup_sched_in(struct task_struct *prev,
571 struct task_struct *task)
575 static inline int perf_cgroup_connect(pid_t pid, struct perf_event *event,
576 struct perf_event_attr *attr,
577 struct perf_event *group_leader)
583 perf_cgroup_set_timestamp(struct task_struct *task,
584 struct perf_event_context *ctx)
589 perf_cgroup_switch(struct task_struct *task, struct task_struct *next)
594 perf_cgroup_set_shadow_time(struct perf_event *event, u64 now)
598 static inline u64 perf_cgroup_event_time(struct perf_event *event)
604 perf_cgroup_defer_enabled(struct perf_event *event)
609 perf_cgroup_mark_enabled(struct perf_event *event,
610 struct perf_event_context *ctx)
615 void perf_pmu_disable(struct pmu *pmu)
617 int *count = this_cpu_ptr(pmu->pmu_disable_count);
619 pmu->pmu_disable(pmu);
622 void perf_pmu_enable(struct pmu *pmu)
624 int *count = this_cpu_ptr(pmu->pmu_disable_count);
626 pmu->pmu_enable(pmu);
629 static DEFINE_PER_CPU(struct list_head, rotation_list);
632 * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
633 * because they're strictly cpu affine and rotate_start is called with IRQs
634 * disabled, while rotate_context is called from IRQ context.
636 static void perf_pmu_rotate_start(struct pmu *pmu)
638 struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
639 struct list_head *head = &__get_cpu_var(rotation_list);
641 WARN_ON(!irqs_disabled());
643 if (list_empty(&cpuctx->rotation_list))
644 list_add(&cpuctx->rotation_list, head);
647 static void get_ctx(struct perf_event_context *ctx)
649 WARN_ON(!atomic_inc_not_zero(&ctx->refcount));
652 static void put_ctx(struct perf_event_context *ctx)
654 if (atomic_dec_and_test(&ctx->refcount)) {
656 put_ctx(ctx->parent_ctx);
658 put_task_struct(ctx->task);
659 kfree_rcu(ctx, rcu_head);
663 static void unclone_ctx(struct perf_event_context *ctx)
665 if (ctx->parent_ctx) {
666 put_ctx(ctx->parent_ctx);
667 ctx->parent_ctx = NULL;
671 static u32 perf_event_pid(struct perf_event *event, struct task_struct *p)
674 * only top level events have the pid namespace they were created in
677 event = event->parent;
679 return task_tgid_nr_ns(p, event->ns);
682 static u32 perf_event_tid(struct perf_event *event, struct task_struct *p)
685 * only top level events have the pid namespace they were created in
688 event = event->parent;
690 return task_pid_nr_ns(p, event->ns);
694 * If we inherit events we want to return the parent event id
697 static u64 primary_event_id(struct perf_event *event)
702 id = event->parent->id;
708 * Get the perf_event_context for a task and lock it.
709 * This has to cope with with the fact that until it is locked,
710 * the context could get moved to another task.
712 static struct perf_event_context *
713 perf_lock_task_context(struct task_struct *task, int ctxn, unsigned long *flags)
715 struct perf_event_context *ctx;
719 ctx = rcu_dereference(task->perf_event_ctxp[ctxn]);
722 * If this context is a clone of another, it might
723 * get swapped for another underneath us by
724 * perf_event_task_sched_out, though the
725 * rcu_read_lock() protects us from any context
726 * getting freed. Lock the context and check if it
727 * got swapped before we could get the lock, and retry
728 * if so. If we locked the right context, then it
729 * can't get swapped on us any more.
731 raw_spin_lock_irqsave(&ctx->lock, *flags);
732 if (ctx != rcu_dereference(task->perf_event_ctxp[ctxn])) {
733 raw_spin_unlock_irqrestore(&ctx->lock, *flags);
737 if (!atomic_inc_not_zero(&ctx->refcount)) {
738 raw_spin_unlock_irqrestore(&ctx->lock, *flags);
747 * Get the context for a task and increment its pin_count so it
748 * can't get swapped to another task. This also increments its
749 * reference count so that the context can't get freed.
751 static struct perf_event_context *
752 perf_pin_task_context(struct task_struct *task, int ctxn)
754 struct perf_event_context *ctx;
757 ctx = perf_lock_task_context(task, ctxn, &flags);
760 raw_spin_unlock_irqrestore(&ctx->lock, flags);
765 static void perf_unpin_context(struct perf_event_context *ctx)
769 raw_spin_lock_irqsave(&ctx->lock, flags);
771 raw_spin_unlock_irqrestore(&ctx->lock, flags);
775 * Update the record of the current time in a context.
777 static void update_context_time(struct perf_event_context *ctx)
779 u64 now = perf_clock();
781 ctx->time += now - ctx->timestamp;
782 ctx->timestamp = now;
785 static u64 perf_event_time(struct perf_event *event)
787 struct perf_event_context *ctx = event->ctx;
789 if (is_cgroup_event(event))
790 return perf_cgroup_event_time(event);
792 return ctx ? ctx->time : 0;
796 * Update the total_time_enabled and total_time_running fields for a event.
797 * The caller of this function needs to hold the ctx->lock.
799 static void update_event_times(struct perf_event *event)
801 struct perf_event_context *ctx = event->ctx;
804 if (event->state < PERF_EVENT_STATE_INACTIVE ||
805 event->group_leader->state < PERF_EVENT_STATE_INACTIVE)
808 * in cgroup mode, time_enabled represents
809 * the time the event was enabled AND active
810 * tasks were in the monitored cgroup. This is
811 * independent of the activity of the context as
812 * there may be a mix of cgroup and non-cgroup events.
814 * That is why we treat cgroup events differently
817 if (is_cgroup_event(event))
818 run_end = perf_cgroup_event_time(event);
819 else if (ctx->is_active)
822 run_end = event->tstamp_stopped;
824 event->total_time_enabled = run_end - event->tstamp_enabled;
826 if (event->state == PERF_EVENT_STATE_INACTIVE)
827 run_end = event->tstamp_stopped;
829 run_end = perf_event_time(event);
831 event->total_time_running = run_end - event->tstamp_running;
836 * Update total_time_enabled and total_time_running for all events in a group.
838 static void update_group_times(struct perf_event *leader)
840 struct perf_event *event;
842 update_event_times(leader);
843 list_for_each_entry(event, &leader->sibling_list, group_entry)
844 update_event_times(event);
847 static struct list_head *
848 ctx_group_list(struct perf_event *event, struct perf_event_context *ctx)
850 if (event->attr.pinned)
851 return &ctx->pinned_groups;
853 return &ctx->flexible_groups;
857 * Add a event from the lists for its context.
858 * Must be called with ctx->mutex and ctx->lock held.
861 list_add_event(struct perf_event *event, struct perf_event_context *ctx)
863 WARN_ON_ONCE(event->attach_state & PERF_ATTACH_CONTEXT);
864 event->attach_state |= PERF_ATTACH_CONTEXT;
867 * If we're a stand alone event or group leader, we go to the context
868 * list, group events are kept attached to the group so that
869 * perf_group_detach can, at all times, locate all siblings.
871 if (event->group_leader == event) {
872 struct list_head *list;
874 if (is_software_event(event))
875 event->group_flags |= PERF_GROUP_SOFTWARE;
877 list = ctx_group_list(event, ctx);
878 list_add_tail(&event->group_entry, list);
881 if (is_cgroup_event(event))
884 list_add_rcu(&event->event_entry, &ctx->event_list);
886 perf_pmu_rotate_start(ctx->pmu);
888 if (event->attr.inherit_stat)
893 * Called at perf_event creation and when events are attached/detached from a
896 static void perf_event__read_size(struct perf_event *event)
898 int entry = sizeof(u64); /* value */
902 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
905 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
908 if (event->attr.read_format & PERF_FORMAT_ID)
909 entry += sizeof(u64);
911 if (event->attr.read_format & PERF_FORMAT_GROUP) {
912 nr += event->group_leader->nr_siblings;
917 event->read_size = size;
920 static void perf_event__header_size(struct perf_event *event)
922 struct perf_sample_data *data;
923 u64 sample_type = event->attr.sample_type;
926 perf_event__read_size(event);
928 if (sample_type & PERF_SAMPLE_IP)
929 size += sizeof(data->ip);
931 if (sample_type & PERF_SAMPLE_ADDR)
932 size += sizeof(data->addr);
934 if (sample_type & PERF_SAMPLE_PERIOD)
935 size += sizeof(data->period);
937 if (sample_type & PERF_SAMPLE_READ)
938 size += event->read_size;
940 event->header_size = size;
943 static void perf_event__id_header_size(struct perf_event *event)
945 struct perf_sample_data *data;
946 u64 sample_type = event->attr.sample_type;
949 if (sample_type & PERF_SAMPLE_TID)
950 size += sizeof(data->tid_entry);
952 if (sample_type & PERF_SAMPLE_TIME)
953 size += sizeof(data->time);
955 if (sample_type & PERF_SAMPLE_ID)
956 size += sizeof(data->id);
958 if (sample_type & PERF_SAMPLE_STREAM_ID)
959 size += sizeof(data->stream_id);
961 if (sample_type & PERF_SAMPLE_CPU)
962 size += sizeof(data->cpu_entry);
964 event->id_header_size = size;
967 static void perf_group_attach(struct perf_event *event)
969 struct perf_event *group_leader = event->group_leader, *pos;
972 * We can have double attach due to group movement in perf_event_open.
974 if (event->attach_state & PERF_ATTACH_GROUP)
977 event->attach_state |= PERF_ATTACH_GROUP;
979 if (group_leader == event)
982 if (group_leader->group_flags & PERF_GROUP_SOFTWARE &&
983 !is_software_event(event))
984 group_leader->group_flags &= ~PERF_GROUP_SOFTWARE;
986 list_add_tail(&event->group_entry, &group_leader->sibling_list);
987 group_leader->nr_siblings++;
989 perf_event__header_size(group_leader);
991 list_for_each_entry(pos, &group_leader->sibling_list, group_entry)
992 perf_event__header_size(pos);
996 * Remove a event from the lists for its context.
997 * Must be called with ctx->mutex and ctx->lock held.
1000 list_del_event(struct perf_event *event, struct perf_event_context *ctx)
1002 struct perf_cpu_context *cpuctx;
1004 * We can have double detach due to exit/hot-unplug + close.
1006 if (!(event->attach_state & PERF_ATTACH_CONTEXT))
1009 event->attach_state &= ~PERF_ATTACH_CONTEXT;
1011 if (is_cgroup_event(event)) {
1013 cpuctx = __get_cpu_context(ctx);
1015 * if there are no more cgroup events
1016 * then cler cgrp to avoid stale pointer
1017 * in update_cgrp_time_from_cpuctx()
1019 if (!ctx->nr_cgroups)
1020 cpuctx->cgrp = NULL;
1024 if (event->attr.inherit_stat)
1027 list_del_rcu(&event->event_entry);
1029 if (event->group_leader == event)
1030 list_del_init(&event->group_entry);
1032 update_group_times(event);
1035 * If event was in error state, then keep it
1036 * that way, otherwise bogus counts will be
1037 * returned on read(). The only way to get out
1038 * of error state is by explicit re-enabling
1041 if (event->state > PERF_EVENT_STATE_OFF)
1042 event->state = PERF_EVENT_STATE_OFF;
1045 static void perf_group_detach(struct perf_event *event)
1047 struct perf_event *sibling, *tmp;
1048 struct list_head *list = NULL;
1051 * We can have double detach due to exit/hot-unplug + close.
1053 if (!(event->attach_state & PERF_ATTACH_GROUP))
1056 event->attach_state &= ~PERF_ATTACH_GROUP;
1059 * If this is a sibling, remove it from its group.
1061 if (event->group_leader != event) {
1062 list_del_init(&event->group_entry);
1063 event->group_leader->nr_siblings--;
1067 if (!list_empty(&event->group_entry))
1068 list = &event->group_entry;
1071 * If this was a group event with sibling events then
1072 * upgrade the siblings to singleton events by adding them
1073 * to whatever list we are on.
1075 list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) {
1077 list_move_tail(&sibling->group_entry, list);
1078 sibling->group_leader = sibling;
1080 /* Inherit group flags from the previous leader */
1081 sibling->group_flags = event->group_flags;
1085 perf_event__header_size(event->group_leader);
1087 list_for_each_entry(tmp, &event->group_leader->sibling_list, group_entry)
1088 perf_event__header_size(tmp);
1092 event_filter_match(struct perf_event *event)
1094 return (event->cpu == -1 || event->cpu == smp_processor_id())
1095 && perf_cgroup_match(event);
1099 event_sched_out(struct perf_event *event,
1100 struct perf_cpu_context *cpuctx,
1101 struct perf_event_context *ctx)
1103 u64 tstamp = perf_event_time(event);
1106 * An event which could not be activated because of
1107 * filter mismatch still needs to have its timings
1108 * maintained, otherwise bogus information is return
1109 * via read() for time_enabled, time_running:
1111 if (event->state == PERF_EVENT_STATE_INACTIVE
1112 && !event_filter_match(event)) {
1113 delta = tstamp - event->tstamp_stopped;
1114 event->tstamp_running += delta;
1115 event->tstamp_stopped = tstamp;
1118 if (event->state != PERF_EVENT_STATE_ACTIVE)
1121 event->state = PERF_EVENT_STATE_INACTIVE;
1122 if (event->pending_disable) {
1123 event->pending_disable = 0;
1124 event->state = PERF_EVENT_STATE_OFF;
1126 event->tstamp_stopped = tstamp;
1127 event->pmu->del(event, 0);
1130 if (!is_software_event(event))
1131 cpuctx->active_oncpu--;
1133 if (event->attr.freq && event->attr.sample_freq)
1135 if (event->attr.exclusive || !cpuctx->active_oncpu)
1136 cpuctx->exclusive = 0;
1140 group_sched_out(struct perf_event *group_event,
1141 struct perf_cpu_context *cpuctx,
1142 struct perf_event_context *ctx)
1144 struct perf_event *event;
1145 int state = group_event->state;
1147 event_sched_out(group_event, cpuctx, ctx);
1150 * Schedule out siblings (if any):
1152 list_for_each_entry(event, &group_event->sibling_list, group_entry)
1153 event_sched_out(event, cpuctx, ctx);
1155 if (state == PERF_EVENT_STATE_ACTIVE && group_event->attr.exclusive)
1156 cpuctx->exclusive = 0;
1160 * Cross CPU call to remove a performance event
1162 * We disable the event on the hardware level first. After that we
1163 * remove it from the context list.
1165 static int __perf_remove_from_context(void *info)
1167 struct perf_event *event = info;
1168 struct perf_event_context *ctx = event->ctx;
1169 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1171 raw_spin_lock(&ctx->lock);
1172 event_sched_out(event, cpuctx, ctx);
1173 list_del_event(event, ctx);
1174 if (!ctx->nr_events && cpuctx->task_ctx == ctx) {
1176 cpuctx->task_ctx = NULL;
1178 raw_spin_unlock(&ctx->lock);
1185 * Remove the event from a task's (or a CPU's) list of events.
1187 * CPU events are removed with a smp call. For task events we only
1188 * call when the task is on a CPU.
1190 * If event->ctx is a cloned context, callers must make sure that
1191 * every task struct that event->ctx->task could possibly point to
1192 * remains valid. This is OK when called from perf_release since
1193 * that only calls us on the top-level context, which can't be a clone.
1194 * When called from perf_event_exit_task, it's OK because the
1195 * context has been detached from its task.
1197 static void perf_remove_from_context(struct perf_event *event)
1199 struct perf_event_context *ctx = event->ctx;
1200 struct task_struct *task = ctx->task;
1202 lockdep_assert_held(&ctx->mutex);
1206 * Per cpu events are removed via an smp call and
1207 * the removal is always successful.
1209 cpu_function_call(event->cpu, __perf_remove_from_context, event);
1214 if (!task_function_call(task, __perf_remove_from_context, event))
1217 raw_spin_lock_irq(&ctx->lock);
1219 * If we failed to find a running task, but find the context active now
1220 * that we've acquired the ctx->lock, retry.
1222 if (ctx->is_active) {
1223 raw_spin_unlock_irq(&ctx->lock);
1228 * Since the task isn't running, its safe to remove the event, us
1229 * holding the ctx->lock ensures the task won't get scheduled in.
1231 list_del_event(event, ctx);
1232 raw_spin_unlock_irq(&ctx->lock);
1236 * Cross CPU call to disable a performance event
1238 static int __perf_event_disable(void *info)
1240 struct perf_event *event = info;
1241 struct perf_event_context *ctx = event->ctx;
1242 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1245 * If this is a per-task event, need to check whether this
1246 * event's task is the current task on this cpu.
1248 * Can trigger due to concurrent perf_event_context_sched_out()
1249 * flipping contexts around.
1251 if (ctx->task && cpuctx->task_ctx != ctx)
1254 raw_spin_lock(&ctx->lock);
1257 * If the event is on, turn it off.
1258 * If it is in error state, leave it in error state.
1260 if (event->state >= PERF_EVENT_STATE_INACTIVE) {
1261 update_context_time(ctx);
1262 update_cgrp_time_from_event(event);
1263 update_group_times(event);
1264 if (event == event->group_leader)
1265 group_sched_out(event, cpuctx, ctx);
1267 event_sched_out(event, cpuctx, ctx);
1268 event->state = PERF_EVENT_STATE_OFF;
1271 raw_spin_unlock(&ctx->lock);
1279 * If event->ctx is a cloned context, callers must make sure that
1280 * every task struct that event->ctx->task could possibly point to
1281 * remains valid. This condition is satisifed when called through
1282 * perf_event_for_each_child or perf_event_for_each because they
1283 * hold the top-level event's child_mutex, so any descendant that
1284 * goes to exit will block in sync_child_event.
1285 * When called from perf_pending_event it's OK because event->ctx
1286 * is the current context on this CPU and preemption is disabled,
1287 * hence we can't get into perf_event_task_sched_out for this context.
1289 void perf_event_disable(struct perf_event *event)
1291 struct perf_event_context *ctx = event->ctx;
1292 struct task_struct *task = ctx->task;
1296 * Disable the event on the cpu that it's on
1298 cpu_function_call(event->cpu, __perf_event_disable, event);
1303 if (!task_function_call(task, __perf_event_disable, event))
1306 raw_spin_lock_irq(&ctx->lock);
1308 * If the event is still active, we need to retry the cross-call.
1310 if (event->state == PERF_EVENT_STATE_ACTIVE) {
1311 raw_spin_unlock_irq(&ctx->lock);
1313 * Reload the task pointer, it might have been changed by
1314 * a concurrent perf_event_context_sched_out().
1321 * Since we have the lock this context can't be scheduled
1322 * in, so we can change the state safely.
1324 if (event->state == PERF_EVENT_STATE_INACTIVE) {
1325 update_group_times(event);
1326 event->state = PERF_EVENT_STATE_OFF;
1328 raw_spin_unlock_irq(&ctx->lock);
1330 EXPORT_SYMBOL_GPL(perf_event_disable);
1332 static void perf_set_shadow_time(struct perf_event *event,
1333 struct perf_event_context *ctx,
1337 * use the correct time source for the time snapshot
1339 * We could get by without this by leveraging the
1340 * fact that to get to this function, the caller
1341 * has most likely already called update_context_time()
1342 * and update_cgrp_time_xx() and thus both timestamp
1343 * are identical (or very close). Given that tstamp is,
1344 * already adjusted for cgroup, we could say that:
1345 * tstamp - ctx->timestamp
1347 * tstamp - cgrp->timestamp.
1349 * Then, in perf_output_read(), the calculation would
1350 * work with no changes because:
1351 * - event is guaranteed scheduled in
1352 * - no scheduled out in between
1353 * - thus the timestamp would be the same
1355 * But this is a bit hairy.
1357 * So instead, we have an explicit cgroup call to remain
1358 * within the time time source all along. We believe it
1359 * is cleaner and simpler to understand.
1361 if (is_cgroup_event(event))
1362 perf_cgroup_set_shadow_time(event, tstamp);
1364 event->shadow_ctx_time = tstamp - ctx->timestamp;
1367 #define MAX_INTERRUPTS (~0ULL)
1369 static void perf_log_throttle(struct perf_event *event, int enable);
1372 event_sched_in(struct perf_event *event,
1373 struct perf_cpu_context *cpuctx,
1374 struct perf_event_context *ctx)
1376 u64 tstamp = perf_event_time(event);
1378 if (event->state <= PERF_EVENT_STATE_OFF)
1381 event->state = PERF_EVENT_STATE_ACTIVE;
1382 event->oncpu = smp_processor_id();
1385 * Unthrottle events, since we scheduled we might have missed several
1386 * ticks already, also for a heavily scheduling task there is little
1387 * guarantee it'll get a tick in a timely manner.
1389 if (unlikely(event->hw.interrupts == MAX_INTERRUPTS)) {
1390 perf_log_throttle(event, 1);
1391 event->hw.interrupts = 0;
1395 * The new state must be visible before we turn it on in the hardware:
1399 if (event->pmu->add(event, PERF_EF_START)) {
1400 event->state = PERF_EVENT_STATE_INACTIVE;
1405 event->tstamp_running += tstamp - event->tstamp_stopped;
1407 perf_set_shadow_time(event, ctx, tstamp);
1409 if (!is_software_event(event))
1410 cpuctx->active_oncpu++;
1412 if (event->attr.freq && event->attr.sample_freq)
1415 if (event->attr.exclusive)
1416 cpuctx->exclusive = 1;
1422 group_sched_in(struct perf_event *group_event,
1423 struct perf_cpu_context *cpuctx,
1424 struct perf_event_context *ctx)
1426 struct perf_event *event, *partial_group = NULL;
1427 struct pmu *pmu = group_event->pmu;
1428 u64 now = ctx->time;
1429 bool simulate = false;
1431 if (group_event->state == PERF_EVENT_STATE_OFF)
1434 pmu->start_txn(pmu);
1436 if (event_sched_in(group_event, cpuctx, ctx)) {
1437 pmu->cancel_txn(pmu);
1442 * Schedule in siblings as one group (if any):
1444 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
1445 if (event_sched_in(event, cpuctx, ctx)) {
1446 partial_group = event;
1451 if (!pmu->commit_txn(pmu))
1456 * Groups can be scheduled in as one unit only, so undo any
1457 * partial group before returning:
1458 * The events up to the failed event are scheduled out normally,
1459 * tstamp_stopped will be updated.
1461 * The failed events and the remaining siblings need to have
1462 * their timings updated as if they had gone thru event_sched_in()
1463 * and event_sched_out(). This is required to get consistent timings
1464 * across the group. This also takes care of the case where the group
1465 * could never be scheduled by ensuring tstamp_stopped is set to mark
1466 * the time the event was actually stopped, such that time delta
1467 * calculation in update_event_times() is correct.
1469 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
1470 if (event == partial_group)
1474 event->tstamp_running += now - event->tstamp_stopped;
1475 event->tstamp_stopped = now;
1477 event_sched_out(event, cpuctx, ctx);
1480 event_sched_out(group_event, cpuctx, ctx);
1482 pmu->cancel_txn(pmu);
1488 * Work out whether we can put this event group on the CPU now.
1490 static int group_can_go_on(struct perf_event *event,
1491 struct perf_cpu_context *cpuctx,
1495 * Groups consisting entirely of software events can always go on.
1497 if (event->group_flags & PERF_GROUP_SOFTWARE)
1500 * If an exclusive group is already on, no other hardware
1503 if (cpuctx->exclusive)
1506 * If this group is exclusive and there are already
1507 * events on the CPU, it can't go on.
1509 if (event->attr.exclusive && cpuctx->active_oncpu)
1512 * Otherwise, try to add it if all previous groups were able
1518 static void add_event_to_ctx(struct perf_event *event,
1519 struct perf_event_context *ctx)
1521 u64 tstamp = perf_event_time(event);
1523 list_add_event(event, ctx);
1524 perf_group_attach(event);
1525 event->tstamp_enabled = tstamp;
1526 event->tstamp_running = tstamp;
1527 event->tstamp_stopped = tstamp;
1530 static void task_ctx_sched_out(struct perf_event_context *ctx);
1532 ctx_sched_in(struct perf_event_context *ctx,
1533 struct perf_cpu_context *cpuctx,
1534 enum event_type_t event_type,
1535 struct task_struct *task);
1537 static void perf_event_sched_in(struct perf_cpu_context *cpuctx,
1538 struct perf_event_context *ctx,
1539 struct task_struct *task)
1541 cpu_ctx_sched_in(cpuctx, EVENT_PINNED, task);
1543 ctx_sched_in(ctx, cpuctx, EVENT_PINNED, task);
1544 cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE, task);
1546 ctx_sched_in(ctx, cpuctx, EVENT_FLEXIBLE, task);
1550 * Cross CPU call to install and enable a performance event
1552 * Must be called with ctx->mutex held
1554 static int __perf_install_in_context(void *info)
1556 struct perf_event *event = info;
1557 struct perf_event_context *ctx = event->ctx;
1558 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1559 struct perf_event_context *task_ctx = cpuctx->task_ctx;
1560 struct task_struct *task = current;
1562 perf_ctx_lock(cpuctx, task_ctx);
1563 perf_pmu_disable(cpuctx->ctx.pmu);
1566 * If there was an active task_ctx schedule it out.
1569 task_ctx_sched_out(task_ctx);
1572 * If the context we're installing events in is not the
1573 * active task_ctx, flip them.
1575 if (ctx->task && task_ctx != ctx) {
1577 raw_spin_unlock(&task_ctx->lock);
1578 raw_spin_lock(&ctx->lock);
1583 cpuctx->task_ctx = task_ctx;
1584 task = task_ctx->task;
1587 cpu_ctx_sched_out(cpuctx, EVENT_ALL);
1589 update_context_time(ctx);
1591 * update cgrp time only if current cgrp
1592 * matches event->cgrp. Must be done before
1593 * calling add_event_to_ctx()
1595 update_cgrp_time_from_event(event);
1597 add_event_to_ctx(event, ctx);
1600 * Schedule everything back in
1602 perf_event_sched_in(cpuctx, task_ctx, task);
1604 perf_pmu_enable(cpuctx->ctx.pmu);
1605 perf_ctx_unlock(cpuctx, task_ctx);
1611 * Attach a performance event to a context
1613 * First we add the event to the list with the hardware enable bit
1614 * in event->hw_config cleared.
1616 * If the event is attached to a task which is on a CPU we use a smp
1617 * call to enable it in the task context. The task might have been
1618 * scheduled away, but we check this in the smp call again.
1621 perf_install_in_context(struct perf_event_context *ctx,
1622 struct perf_event *event,
1625 struct task_struct *task = ctx->task;
1627 lockdep_assert_held(&ctx->mutex);
1633 * Per cpu events are installed via an smp call and
1634 * the install is always successful.
1636 cpu_function_call(cpu, __perf_install_in_context, event);
1641 if (!task_function_call(task, __perf_install_in_context, event))
1644 raw_spin_lock_irq(&ctx->lock);
1646 * If we failed to find a running task, but find the context active now
1647 * that we've acquired the ctx->lock, retry.
1649 if (ctx->is_active) {
1650 raw_spin_unlock_irq(&ctx->lock);
1655 * Since the task isn't running, its safe to add the event, us holding
1656 * the ctx->lock ensures the task won't get scheduled in.
1658 add_event_to_ctx(event, ctx);
1659 raw_spin_unlock_irq(&ctx->lock);
1663 * Put a event into inactive state and update time fields.
1664 * Enabling the leader of a group effectively enables all
1665 * the group members that aren't explicitly disabled, so we
1666 * have to update their ->tstamp_enabled also.
1667 * Note: this works for group members as well as group leaders
1668 * since the non-leader members' sibling_lists will be empty.
1670 static void __perf_event_mark_enabled(struct perf_event *event)
1672 struct perf_event *sub;
1673 u64 tstamp = perf_event_time(event);
1675 event->state = PERF_EVENT_STATE_INACTIVE;
1676 event->tstamp_enabled = tstamp - event->total_time_enabled;
1677 list_for_each_entry(sub, &event->sibling_list, group_entry) {
1678 if (sub->state >= PERF_EVENT_STATE_INACTIVE)
1679 sub->tstamp_enabled = tstamp - sub->total_time_enabled;
1684 * Cross CPU call to enable a performance event
1686 static int __perf_event_enable(void *info)
1688 struct perf_event *event = info;
1689 struct perf_event_context *ctx = event->ctx;
1690 struct perf_event *leader = event->group_leader;
1691 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1694 if (WARN_ON_ONCE(!ctx->is_active))
1697 raw_spin_lock(&ctx->lock);
1698 update_context_time(ctx);
1700 if (event->state >= PERF_EVENT_STATE_INACTIVE)
1704 * set current task's cgroup time reference point
1706 perf_cgroup_set_timestamp(current, ctx);
1708 __perf_event_mark_enabled(event);
1710 if (!event_filter_match(event)) {
1711 if (is_cgroup_event(event))
1712 perf_cgroup_defer_enabled(event);
1717 * If the event is in a group and isn't the group leader,
1718 * then don't put it on unless the group is on.
1720 if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE)
1723 if (!group_can_go_on(event, cpuctx, 1)) {
1726 if (event == leader)
1727 err = group_sched_in(event, cpuctx, ctx);
1729 err = event_sched_in(event, cpuctx, ctx);
1734 * If this event can't go on and it's part of a
1735 * group, then the whole group has to come off.
1737 if (leader != event)
1738 group_sched_out(leader, cpuctx, ctx);
1739 if (leader->attr.pinned) {
1740 update_group_times(leader);
1741 leader->state = PERF_EVENT_STATE_ERROR;
1746 raw_spin_unlock(&ctx->lock);
1754 * If event->ctx is a cloned context, callers must make sure that
1755 * every task struct that event->ctx->task could possibly point to
1756 * remains valid. This condition is satisfied when called through
1757 * perf_event_for_each_child or perf_event_for_each as described
1758 * for perf_event_disable.
1760 void perf_event_enable(struct perf_event *event)
1762 struct perf_event_context *ctx = event->ctx;
1763 struct task_struct *task = ctx->task;
1767 * Enable the event on the cpu that it's on
1769 cpu_function_call(event->cpu, __perf_event_enable, event);
1773 raw_spin_lock_irq(&ctx->lock);
1774 if (event->state >= PERF_EVENT_STATE_INACTIVE)
1778 * If the event is in error state, clear that first.
1779 * That way, if we see the event in error state below, we
1780 * know that it has gone back into error state, as distinct
1781 * from the task having been scheduled away before the
1782 * cross-call arrived.
1784 if (event->state == PERF_EVENT_STATE_ERROR)
1785 event->state = PERF_EVENT_STATE_OFF;
1788 if (!ctx->is_active) {
1789 __perf_event_mark_enabled(event);
1793 raw_spin_unlock_irq(&ctx->lock);
1795 if (!task_function_call(task, __perf_event_enable, event))
1798 raw_spin_lock_irq(&ctx->lock);
1801 * If the context is active and the event is still off,
1802 * we need to retry the cross-call.
1804 if (ctx->is_active && event->state == PERF_EVENT_STATE_OFF) {
1806 * task could have been flipped by a concurrent
1807 * perf_event_context_sched_out()
1814 raw_spin_unlock_irq(&ctx->lock);
1816 EXPORT_SYMBOL_GPL(perf_event_enable);
1818 int perf_event_refresh(struct perf_event *event, int refresh)
1821 * not supported on inherited events
1823 if (event->attr.inherit || !is_sampling_event(event))
1826 atomic_add(refresh, &event->event_limit);
1827 perf_event_enable(event);
1831 EXPORT_SYMBOL_GPL(perf_event_refresh);
1833 static void ctx_sched_out(struct perf_event_context *ctx,
1834 struct perf_cpu_context *cpuctx,
1835 enum event_type_t event_type)
1837 struct perf_event *event;
1838 int is_active = ctx->is_active;
1840 ctx->is_active &= ~event_type;
1841 if (likely(!ctx->nr_events))
1844 update_context_time(ctx);
1845 update_cgrp_time_from_cpuctx(cpuctx);
1846 if (!ctx->nr_active)
1849 perf_pmu_disable(ctx->pmu);
1850 if ((is_active & EVENT_PINNED) && (event_type & EVENT_PINNED)) {
1851 list_for_each_entry(event, &ctx->pinned_groups, group_entry)
1852 group_sched_out(event, cpuctx, ctx);
1855 if ((is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE)) {
1856 list_for_each_entry(event, &ctx->flexible_groups, group_entry)
1857 group_sched_out(event, cpuctx, ctx);
1859 perf_pmu_enable(ctx->pmu);
1863 * Test whether two contexts are equivalent, i.e. whether they
1864 * have both been cloned from the same version of the same context
1865 * and they both have the same number of enabled events.
1866 * If the number of enabled events is the same, then the set
1867 * of enabled events should be the same, because these are both
1868 * inherited contexts, therefore we can't access individual events
1869 * in them directly with an fd; we can only enable/disable all
1870 * events via prctl, or enable/disable all events in a family
1871 * via ioctl, which will have the same effect on both contexts.
1873 static int context_equiv(struct perf_event_context *ctx1,
1874 struct perf_event_context *ctx2)
1876 return ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx
1877 && ctx1->parent_gen == ctx2->parent_gen
1878 && !ctx1->pin_count && !ctx2->pin_count;
1881 static void __perf_event_sync_stat(struct perf_event *event,
1882 struct perf_event *next_event)
1886 if (!event->attr.inherit_stat)
1890 * Update the event value, we cannot use perf_event_read()
1891 * because we're in the middle of a context switch and have IRQs
1892 * disabled, which upsets smp_call_function_single(), however
1893 * we know the event must be on the current CPU, therefore we
1894 * don't need to use it.
1896 switch (event->state) {
1897 case PERF_EVENT_STATE_ACTIVE:
1898 event->pmu->read(event);
1901 case PERF_EVENT_STATE_INACTIVE:
1902 update_event_times(event);
1910 * In order to keep per-task stats reliable we need to flip the event
1911 * values when we flip the contexts.
1913 value = local64_read(&next_event->count);
1914 value = local64_xchg(&event->count, value);
1915 local64_set(&next_event->count, value);
1917 swap(event->total_time_enabled, next_event->total_time_enabled);
1918 swap(event->total_time_running, next_event->total_time_running);
1921 * Since we swizzled the values, update the user visible data too.
1923 perf_event_update_userpage(event);
1924 perf_event_update_userpage(next_event);
1927 #define list_next_entry(pos, member) \
1928 list_entry(pos->member.next, typeof(*pos), member)
1930 static void perf_event_sync_stat(struct perf_event_context *ctx,
1931 struct perf_event_context *next_ctx)
1933 struct perf_event *event, *next_event;
1938 update_context_time(ctx);
1940 event = list_first_entry(&ctx->event_list,
1941 struct perf_event, event_entry);
1943 next_event = list_first_entry(&next_ctx->event_list,
1944 struct perf_event, event_entry);
1946 while (&event->event_entry != &ctx->event_list &&
1947 &next_event->event_entry != &next_ctx->event_list) {
1949 __perf_event_sync_stat(event, next_event);
1951 event = list_next_entry(event, event_entry);
1952 next_event = list_next_entry(next_event, event_entry);
1956 static void perf_event_context_sched_out(struct task_struct *task, int ctxn,
1957 struct task_struct *next)
1959 struct perf_event_context *ctx = task->perf_event_ctxp[ctxn];
1960 struct perf_event_context *next_ctx;
1961 struct perf_event_context *parent;
1962 struct perf_cpu_context *cpuctx;
1968 cpuctx = __get_cpu_context(ctx);
1969 if (!cpuctx->task_ctx)
1973 parent = rcu_dereference(ctx->parent_ctx);
1974 next_ctx = next->perf_event_ctxp[ctxn];
1975 if (parent && next_ctx &&
1976 rcu_dereference(next_ctx->parent_ctx) == parent) {
1978 * Looks like the two contexts are clones, so we might be
1979 * able to optimize the context switch. We lock both
1980 * contexts and check that they are clones under the
1981 * lock (including re-checking that neither has been
1982 * uncloned in the meantime). It doesn't matter which
1983 * order we take the locks because no other cpu could
1984 * be trying to lock both of these tasks.
1986 raw_spin_lock(&ctx->lock);
1987 raw_spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING);
1988 if (context_equiv(ctx, next_ctx)) {
1990 * XXX do we need a memory barrier of sorts
1991 * wrt to rcu_dereference() of perf_event_ctxp
1993 task->perf_event_ctxp[ctxn] = next_ctx;
1994 next->perf_event_ctxp[ctxn] = ctx;
1996 next_ctx->task = task;
1999 perf_event_sync_stat(ctx, next_ctx);
2001 raw_spin_unlock(&next_ctx->lock);
2002 raw_spin_unlock(&ctx->lock);
2007 raw_spin_lock(&ctx->lock);
2008 ctx_sched_out(ctx, cpuctx, EVENT_ALL);
2009 cpuctx->task_ctx = NULL;
2010 raw_spin_unlock(&ctx->lock);
2014 #define for_each_task_context_nr(ctxn) \
2015 for ((ctxn) = 0; (ctxn) < perf_nr_task_contexts; (ctxn)++)
2018 * Called from scheduler to remove the events of the current task,
2019 * with interrupts disabled.
2021 * We stop each event and update the event value in event->count.
2023 * This does not protect us against NMI, but disable()
2024 * sets the disabled bit in the control field of event _before_
2025 * accessing the event control register. If a NMI hits, then it will
2026 * not restart the event.
2028 void __perf_event_task_sched_out(struct task_struct *task,
2029 struct task_struct *next)
2033 for_each_task_context_nr(ctxn)
2034 perf_event_context_sched_out(task, ctxn, next);
2037 * if cgroup events exist on this CPU, then we need
2038 * to check if we have to switch out PMU state.
2039 * cgroup event are system-wide mode only
2041 if (atomic_read(&__get_cpu_var(perf_cgroup_events)))
2042 perf_cgroup_sched_out(task, next);
2045 static void task_ctx_sched_out(struct perf_event_context *ctx)
2047 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
2049 if (!cpuctx->task_ctx)
2052 if (WARN_ON_ONCE(ctx != cpuctx->task_ctx))
2055 ctx_sched_out(ctx, cpuctx, EVENT_ALL);
2056 cpuctx->task_ctx = NULL;
2060 * Called with IRQs disabled
2062 static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
2063 enum event_type_t event_type)
2065 ctx_sched_out(&cpuctx->ctx, cpuctx, event_type);
2069 ctx_pinned_sched_in(struct perf_event_context *ctx,
2070 struct perf_cpu_context *cpuctx)
2072 struct perf_event *event;
2074 list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
2075 if (event->state <= PERF_EVENT_STATE_OFF)
2077 if (!event_filter_match(event))
2080 /* may need to reset tstamp_enabled */
2081 if (is_cgroup_event(event))
2082 perf_cgroup_mark_enabled(event, ctx);
2084 if (group_can_go_on(event, cpuctx, 1))
2085 group_sched_in(event, cpuctx, ctx);
2088 * If this pinned group hasn't been scheduled,
2089 * put it in error state.
2091 if (event->state == PERF_EVENT_STATE_INACTIVE) {
2092 update_group_times(event);
2093 event->state = PERF_EVENT_STATE_ERROR;
2099 ctx_flexible_sched_in(struct perf_event_context *ctx,
2100 struct perf_cpu_context *cpuctx)
2102 struct perf_event *event;
2105 list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
2106 /* Ignore events in OFF or ERROR state */
2107 if (event->state <= PERF_EVENT_STATE_OFF)
2110 * Listen to the 'cpu' scheduling filter constraint
2113 if (!event_filter_match(event))
2116 /* may need to reset tstamp_enabled */
2117 if (is_cgroup_event(event))
2118 perf_cgroup_mark_enabled(event, ctx);
2120 if (group_can_go_on(event, cpuctx, can_add_hw)) {
2121 if (group_sched_in(event, cpuctx, ctx))
2128 ctx_sched_in(struct perf_event_context *ctx,
2129 struct perf_cpu_context *cpuctx,
2130 enum event_type_t event_type,
2131 struct task_struct *task)
2134 int is_active = ctx->is_active;
2136 ctx->is_active |= event_type;
2137 if (likely(!ctx->nr_events))
2141 ctx->timestamp = now;
2142 perf_cgroup_set_timestamp(task, ctx);
2144 * First go through the list and put on any pinned groups
2145 * in order to give them the best chance of going on.
2147 if (!(is_active & EVENT_PINNED) && (event_type & EVENT_PINNED))
2148 ctx_pinned_sched_in(ctx, cpuctx);
2150 /* Then walk through the lower prio flexible groups */
2151 if (!(is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE))
2152 ctx_flexible_sched_in(ctx, cpuctx);
2155 static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
2156 enum event_type_t event_type,
2157 struct task_struct *task)
2159 struct perf_event_context *ctx = &cpuctx->ctx;
2161 ctx_sched_in(ctx, cpuctx, event_type, task);
2164 static void perf_event_context_sched_in(struct perf_event_context *ctx,
2165 struct task_struct *task)
2167 struct perf_cpu_context *cpuctx;
2169 cpuctx = __get_cpu_context(ctx);
2170 if (cpuctx->task_ctx == ctx)
2173 perf_ctx_lock(cpuctx, ctx);
2174 perf_pmu_disable(ctx->pmu);
2176 * We want to keep the following priority order:
2177 * cpu pinned (that don't need to move), task pinned,
2178 * cpu flexible, task flexible.
2180 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
2183 cpuctx->task_ctx = ctx;
2185 perf_event_sched_in(cpuctx, cpuctx->task_ctx, task);
2187 perf_pmu_enable(ctx->pmu);
2188 perf_ctx_unlock(cpuctx, ctx);
2191 * Since these rotations are per-cpu, we need to ensure the
2192 * cpu-context we got scheduled on is actually rotating.
2194 perf_pmu_rotate_start(ctx->pmu);
2198 * Called from scheduler to add the events of the current task
2199 * with interrupts disabled.
2201 * We restore the event value and then enable it.
2203 * This does not protect us against NMI, but enable()
2204 * sets the enabled bit in the control field of event _before_
2205 * accessing the event control register. If a NMI hits, then it will
2206 * keep the event running.
2208 void __perf_event_task_sched_in(struct task_struct *prev,
2209 struct task_struct *task)
2211 struct perf_event_context *ctx;
2214 for_each_task_context_nr(ctxn) {
2215 ctx = task->perf_event_ctxp[ctxn];
2219 perf_event_context_sched_in(ctx, task);
2222 * if cgroup events exist on this CPU, then we need
2223 * to check if we have to switch in PMU state.
2224 * cgroup event are system-wide mode only
2226 if (atomic_read(&__get_cpu_var(perf_cgroup_events)))
2227 perf_cgroup_sched_in(prev, task);
2230 static u64 perf_calculate_period(struct perf_event *event, u64 nsec, u64 count)
2232 u64 frequency = event->attr.sample_freq;
2233 u64 sec = NSEC_PER_SEC;
2234 u64 divisor, dividend;
2236 int count_fls, nsec_fls, frequency_fls, sec_fls;
2238 count_fls = fls64(count);
2239 nsec_fls = fls64(nsec);
2240 frequency_fls = fls64(frequency);
2244 * We got @count in @nsec, with a target of sample_freq HZ
2245 * the target period becomes:
2248 * period = -------------------
2249 * @nsec * sample_freq
2254 * Reduce accuracy by one bit such that @a and @b converge
2255 * to a similar magnitude.
2257 #define REDUCE_FLS(a, b) \
2259 if (a##_fls > b##_fls) { \
2269 * Reduce accuracy until either term fits in a u64, then proceed with
2270 * the other, so that finally we can do a u64/u64 division.
2272 while (count_fls + sec_fls > 64 && nsec_fls + frequency_fls > 64) {
2273 REDUCE_FLS(nsec, frequency);
2274 REDUCE_FLS(sec, count);
2277 if (count_fls + sec_fls > 64) {
2278 divisor = nsec * frequency;
2280 while (count_fls + sec_fls > 64) {
2281 REDUCE_FLS(count, sec);
2285 dividend = count * sec;
2287 dividend = count * sec;
2289 while (nsec_fls + frequency_fls > 64) {
2290 REDUCE_FLS(nsec, frequency);
2294 divisor = nsec * frequency;
2300 return div64_u64(dividend, divisor);
2303 static DEFINE_PER_CPU(int, perf_throttled_count);
2304 static DEFINE_PER_CPU(u64, perf_throttled_seq);
2306 static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count, bool disable)
2308 struct hw_perf_event *hwc = &event->hw;
2309 s64 period, sample_period;
2312 period = perf_calculate_period(event, nsec, count);
2314 delta = (s64)(period - hwc->sample_period);
2315 delta = (delta + 7) / 8; /* low pass filter */
2317 sample_period = hwc->sample_period + delta;
2322 hwc->sample_period = sample_period;
2324 if (local64_read(&hwc->period_left) > 8*sample_period) {
2326 event->pmu->stop(event, PERF_EF_UPDATE);
2328 local64_set(&hwc->period_left, 0);
2331 event->pmu->start(event, PERF_EF_RELOAD);
2336 * combine freq adjustment with unthrottling to avoid two passes over the
2337 * events. At the same time, make sure, having freq events does not change
2338 * the rate of unthrottling as that would introduce bias.
2340 static void perf_adjust_freq_unthr_context(struct perf_event_context *ctx,
2343 struct perf_event *event;
2344 struct hw_perf_event *hwc;
2345 u64 now, period = TICK_NSEC;
2349 * only need to iterate over all events iff:
2350 * - context have events in frequency mode (needs freq adjust)
2351 * - there are events to unthrottle on this cpu
2353 if (!(ctx->nr_freq || needs_unthr))
2356 raw_spin_lock(&ctx->lock);
2357 perf_pmu_disable(ctx->pmu);
2359 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
2360 if (event->state != PERF_EVENT_STATE_ACTIVE)
2363 if (!event_filter_match(event))
2368 if (needs_unthr && hwc->interrupts == MAX_INTERRUPTS) {
2369 hwc->interrupts = 0;
2370 perf_log_throttle(event, 1);
2371 event->pmu->start(event, 0);
2374 if (!event->attr.freq || !event->attr.sample_freq)
2378 * stop the event and update event->count
2380 event->pmu->stop(event, PERF_EF_UPDATE);
2382 now = local64_read(&event->count);
2383 delta = now - hwc->freq_count_stamp;
2384 hwc->freq_count_stamp = now;
2388 * reload only if value has changed
2389 * we have stopped the event so tell that
2390 * to perf_adjust_period() to avoid stopping it
2394 perf_adjust_period(event, period, delta, false);
2396 event->pmu->start(event, delta > 0 ? PERF_EF_RELOAD : 0);
2399 perf_pmu_enable(ctx->pmu);
2400 raw_spin_unlock(&ctx->lock);
2404 * Round-robin a context's events:
2406 static void rotate_ctx(struct perf_event_context *ctx)
2409 * Rotate the first entry last of non-pinned groups. Rotation might be
2410 * disabled by the inheritance code.
2412 if (!ctx->rotate_disable)
2413 list_rotate_left(&ctx->flexible_groups);
2417 * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
2418 * because they're strictly cpu affine and rotate_start is called with IRQs
2419 * disabled, while rotate_context is called from IRQ context.
2421 static void perf_rotate_context(struct perf_cpu_context *cpuctx)
2423 struct perf_event_context *ctx = NULL;
2424 int rotate = 0, remove = 1;
2426 if (cpuctx->ctx.nr_events) {
2428 if (cpuctx->ctx.nr_events != cpuctx->ctx.nr_active)
2432 ctx = cpuctx->task_ctx;
2433 if (ctx && ctx->nr_events) {
2435 if (ctx->nr_events != ctx->nr_active)
2442 perf_ctx_lock(cpuctx, cpuctx->task_ctx);
2443 perf_pmu_disable(cpuctx->ctx.pmu);
2445 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
2447 ctx_sched_out(ctx, cpuctx, EVENT_FLEXIBLE);
2449 rotate_ctx(&cpuctx->ctx);
2453 perf_event_sched_in(cpuctx, ctx, current);
2455 perf_pmu_enable(cpuctx->ctx.pmu);
2456 perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
2459 list_del_init(&cpuctx->rotation_list);
2462 void perf_event_task_tick(void)
2464 struct list_head *head = &__get_cpu_var(rotation_list);
2465 struct perf_cpu_context *cpuctx, *tmp;
2466 struct perf_event_context *ctx;
2469 WARN_ON(!irqs_disabled());
2471 __this_cpu_inc(perf_throttled_seq);
2472 throttled = __this_cpu_xchg(perf_throttled_count, 0);
2474 list_for_each_entry_safe(cpuctx, tmp, head, rotation_list) {
2476 perf_adjust_freq_unthr_context(ctx, throttled);
2478 ctx = cpuctx->task_ctx;
2480 perf_adjust_freq_unthr_context(ctx, throttled);
2482 if (cpuctx->jiffies_interval == 1 ||
2483 !(jiffies % cpuctx->jiffies_interval))
2484 perf_rotate_context(cpuctx);
2488 static int event_enable_on_exec(struct perf_event *event,
2489 struct perf_event_context *ctx)
2491 if (!event->attr.enable_on_exec)
2494 event->attr.enable_on_exec = 0;
2495 if (event->state >= PERF_EVENT_STATE_INACTIVE)
2498 __perf_event_mark_enabled(event);
2504 * Enable all of a task's events that have been marked enable-on-exec.
2505 * This expects task == current.
2507 static void perf_event_enable_on_exec(struct perf_event_context *ctx)
2509 struct perf_event *event;
2510 unsigned long flags;
2514 local_irq_save(flags);
2515 if (!ctx || !ctx->nr_events)
2519 * We must ctxsw out cgroup events to avoid conflict
2520 * when invoking perf_task_event_sched_in() later on
2521 * in this function. Otherwise we end up trying to
2522 * ctxswin cgroup events which are already scheduled
2525 perf_cgroup_sched_out(current, NULL);
2527 raw_spin_lock(&ctx->lock);
2528 task_ctx_sched_out(ctx);
2530 list_for_each_entry(event, &ctx->event_list, event_entry) {
2531 ret = event_enable_on_exec(event, ctx);
2537 * Unclone this context if we enabled any event.
2542 raw_spin_unlock(&ctx->lock);
2545 * Also calls ctxswin for cgroup events, if any:
2547 perf_event_context_sched_in(ctx, ctx->task);
2549 local_irq_restore(flags);
2553 * Cross CPU call to read the hardware event
2555 static void __perf_event_read(void *info)
2557 struct perf_event *event = info;
2558 struct perf_event_context *ctx = event->ctx;
2559 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
2562 * If this is a task context, we need to check whether it is
2563 * the current task context of this cpu. If not it has been
2564 * scheduled out before the smp call arrived. In that case
2565 * event->count would have been updated to a recent sample
2566 * when the event was scheduled out.
2568 if (ctx->task && cpuctx->task_ctx != ctx)
2571 raw_spin_lock(&ctx->lock);
2572 if (ctx->is_active) {
2573 update_context_time(ctx);
2574 update_cgrp_time_from_event(event);
2576 update_event_times(event);
2577 if (event->state == PERF_EVENT_STATE_ACTIVE)
2578 event->pmu->read(event);
2579 raw_spin_unlock(&ctx->lock);
2582 static inline u64 perf_event_count(struct perf_event *event)
2584 return local64_read(&event->count) + atomic64_read(&event->child_count);
2587 static u64 perf_event_read(struct perf_event *event)
2590 * If event is enabled and currently active on a CPU, update the
2591 * value in the event structure:
2593 if (event->state == PERF_EVENT_STATE_ACTIVE) {
2594 smp_call_function_single(event->oncpu,
2595 __perf_event_read, event, 1);
2596 } else if (event->state == PERF_EVENT_STATE_INACTIVE) {
2597 struct perf_event_context *ctx = event->ctx;
2598 unsigned long flags;
2600 raw_spin_lock_irqsave(&ctx->lock, flags);
2602 * may read while context is not active
2603 * (e.g., thread is blocked), in that case
2604 * we cannot update context time
2606 if (ctx->is_active) {
2607 update_context_time(ctx);
2608 update_cgrp_time_from_event(event);
2610 update_event_times(event);
2611 raw_spin_unlock_irqrestore(&ctx->lock, flags);
2614 return perf_event_count(event);
2618 * Initialize the perf_event context in a task_struct:
2620 static void __perf_event_init_context(struct perf_event_context *ctx)
2622 raw_spin_lock_init(&ctx->lock);
2623 mutex_init(&ctx->mutex);
2624 INIT_LIST_HEAD(&ctx->pinned_groups);
2625 INIT_LIST_HEAD(&ctx->flexible_groups);
2626 INIT_LIST_HEAD(&ctx->event_list);
2627 atomic_set(&ctx->refcount, 1);
2630 static struct perf_event_context *
2631 alloc_perf_context(struct pmu *pmu, struct task_struct *task)
2633 struct perf_event_context *ctx;
2635 ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL);
2639 __perf_event_init_context(ctx);
2642 get_task_struct(task);
2649 static struct task_struct *
2650 find_lively_task_by_vpid(pid_t vpid)
2652 struct task_struct *task;
2659 task = find_task_by_vpid(vpid);
2661 get_task_struct(task);
2665 return ERR_PTR(-ESRCH);
2667 /* Reuse ptrace permission checks for now. */
2669 if (!ptrace_may_access(task, PTRACE_MODE_READ))
2674 put_task_struct(task);
2675 return ERR_PTR(err);
2680 * Returns a matching context with refcount and pincount.
2682 static struct perf_event_context *
2683 find_get_context(struct pmu *pmu, struct task_struct *task, int cpu)
2685 struct perf_event_context *ctx;
2686 struct perf_cpu_context *cpuctx;
2687 unsigned long flags;
2691 /* Must be root to operate on a CPU event: */
2692 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
2693 return ERR_PTR(-EACCES);
2696 * We could be clever and allow to attach a event to an
2697 * offline CPU and activate it when the CPU comes up, but
2700 if (!cpu_online(cpu))
2701 return ERR_PTR(-ENODEV);
2703 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
2712 ctxn = pmu->task_ctx_nr;
2717 ctx = perf_lock_task_context(task, ctxn, &flags);
2721 raw_spin_unlock_irqrestore(&ctx->lock, flags);
2723 ctx = alloc_perf_context(pmu, task);
2729 mutex_lock(&task->perf_event_mutex);
2731 * If it has already passed perf_event_exit_task().
2732 * we must see PF_EXITING, it takes this mutex too.
2734 if (task->flags & PF_EXITING)
2736 else if (task->perf_event_ctxp[ctxn])
2741 rcu_assign_pointer(task->perf_event_ctxp[ctxn], ctx);
2743 mutex_unlock(&task->perf_event_mutex);
2745 if (unlikely(err)) {
2757 return ERR_PTR(err);
2760 static void perf_event_free_filter(struct perf_event *event);
2762 static void free_event_rcu(struct rcu_head *head)
2764 struct perf_event *event;
2766 event = container_of(head, struct perf_event, rcu_head);
2768 put_pid_ns(event->ns);
2769 perf_event_free_filter(event);
2773 static void ring_buffer_put(struct ring_buffer *rb);
2775 static void free_event(struct perf_event *event)
2777 irq_work_sync(&event->pending);
2779 if (!event->parent) {
2780 if (event->attach_state & PERF_ATTACH_TASK)
2781 static_key_slow_dec_deferred(&perf_sched_events);
2782 if (event->attr.mmap || event->attr.mmap_data)
2783 atomic_dec(&nr_mmap_events);
2784 if (event->attr.comm)
2785 atomic_dec(&nr_comm_events);
2786 if (event->attr.task)
2787 atomic_dec(&nr_task_events);
2788 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN)
2789 put_callchain_buffers();
2790 if (is_cgroup_event(event)) {
2791 atomic_dec(&per_cpu(perf_cgroup_events, event->cpu));
2792 static_key_slow_dec_deferred(&perf_sched_events);
2797 ring_buffer_put(event->rb);
2801 if (is_cgroup_event(event))
2802 perf_detach_cgroup(event);
2805 event->destroy(event);
2808 put_ctx(event->ctx);
2810 call_rcu(&event->rcu_head, free_event_rcu);
2813 int perf_event_release_kernel(struct perf_event *event)
2815 struct perf_event_context *ctx = event->ctx;
2817 WARN_ON_ONCE(ctx->parent_ctx);
2819 * There are two ways this annotation is useful:
2821 * 1) there is a lock recursion from perf_event_exit_task
2822 * see the comment there.
2824 * 2) there is a lock-inversion with mmap_sem through
2825 * perf_event_read_group(), which takes faults while
2826 * holding ctx->mutex, however this is called after
2827 * the last filedesc died, so there is no possibility
2828 * to trigger the AB-BA case.
2830 mutex_lock_nested(&ctx->mutex, SINGLE_DEPTH_NESTING);
2831 raw_spin_lock_irq(&ctx->lock);
2832 perf_group_detach(event);
2833 raw_spin_unlock_irq(&ctx->lock);
2834 perf_remove_from_context(event);
2835 mutex_unlock(&ctx->mutex);
2841 EXPORT_SYMBOL_GPL(perf_event_release_kernel);
2844 * Called when the last reference to the file is gone.
2846 static int perf_release(struct inode *inode, struct file *file)
2848 struct perf_event *event = file->private_data;
2849 struct task_struct *owner;
2851 file->private_data = NULL;
2854 owner = ACCESS_ONCE(event->owner);
2856 * Matches the smp_wmb() in perf_event_exit_task(). If we observe
2857 * !owner it means the list deletion is complete and we can indeed
2858 * free this event, otherwise we need to serialize on
2859 * owner->perf_event_mutex.
2861 smp_read_barrier_depends();
2864 * Since delayed_put_task_struct() also drops the last
2865 * task reference we can safely take a new reference
2866 * while holding the rcu_read_lock().
2868 get_task_struct(owner);
2873 mutex_lock(&owner->perf_event_mutex);
2875 * We have to re-check the event->owner field, if it is cleared
2876 * we raced with perf_event_exit_task(), acquiring the mutex
2877 * ensured they're done, and we can proceed with freeing the
2881 list_del_init(&event->owner_entry);
2882 mutex_unlock(&owner->perf_event_mutex);
2883 put_task_struct(owner);
2886 return perf_event_release_kernel(event);
2889 u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running)
2891 struct perf_event *child;
2897 mutex_lock(&event->child_mutex);
2898 total += perf_event_read(event);
2899 *enabled += event->total_time_enabled +
2900 atomic64_read(&event->child_total_time_enabled);
2901 *running += event->total_time_running +
2902 atomic64_read(&event->child_total_time_running);
2904 list_for_each_entry(child, &event->child_list, child_list) {
2905 total += perf_event_read(child);
2906 *enabled += child->total_time_enabled;
2907 *running += child->total_time_running;
2909 mutex_unlock(&event->child_mutex);
2913 EXPORT_SYMBOL_GPL(perf_event_read_value);
2915 static int perf_event_read_group(struct perf_event *event,
2916 u64 read_format, char __user *buf)
2918 struct perf_event *leader = event->group_leader, *sub;
2919 int n = 0, size = 0, ret = -EFAULT;
2920 struct perf_event_context *ctx = leader->ctx;
2922 u64 count, enabled, running;
2924 mutex_lock(&ctx->mutex);
2925 count = perf_event_read_value(leader, &enabled, &running);
2927 values[n++] = 1 + leader->nr_siblings;
2928 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2929 values[n++] = enabled;
2930 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2931 values[n++] = running;
2932 values[n++] = count;
2933 if (read_format & PERF_FORMAT_ID)
2934 values[n++] = primary_event_id(leader);
2936 size = n * sizeof(u64);
2938 if (copy_to_user(buf, values, size))
2943 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
2946 values[n++] = perf_event_read_value(sub, &enabled, &running);
2947 if (read_format & PERF_FORMAT_ID)
2948 values[n++] = primary_event_id(sub);
2950 size = n * sizeof(u64);
2952 if (copy_to_user(buf + ret, values, size)) {
2960 mutex_unlock(&ctx->mutex);
2965 static int perf_event_read_one(struct perf_event *event,
2966 u64 read_format, char __user *buf)
2968 u64 enabled, running;
2972 values[n++] = perf_event_read_value(event, &enabled, &running);
2973 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2974 values[n++] = enabled;
2975 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2976 values[n++] = running;
2977 if (read_format & PERF_FORMAT_ID)
2978 values[n++] = primary_event_id(event);
2980 if (copy_to_user(buf, values, n * sizeof(u64)))
2983 return n * sizeof(u64);
2987 * Read the performance event - simple non blocking version for now
2990 perf_read_hw(struct perf_event *event, char __user *buf, size_t count)
2992 u64 read_format = event->attr.read_format;
2996 * Return end-of-file for a read on a event that is in
2997 * error state (i.e. because it was pinned but it couldn't be
2998 * scheduled on to the CPU at some point).
3000 if (event->state == PERF_EVENT_STATE_ERROR)
3003 if (count < event->read_size)
3006 WARN_ON_ONCE(event->ctx->parent_ctx);
3007 if (read_format & PERF_FORMAT_GROUP)
3008 ret = perf_event_read_group(event, read_format, buf);
3010 ret = perf_event_read_one(event, read_format, buf);
3016 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
3018 struct perf_event *event = file->private_data;
3020 return perf_read_hw(event, buf, count);
3023 static unsigned int perf_poll(struct file *file, poll_table *wait)
3025 struct perf_event *event = file->private_data;
3026 struct ring_buffer *rb;
3027 unsigned int events = POLL_HUP;
3030 * Race between perf_event_set_output() and perf_poll(): perf_poll()
3031 * grabs the rb reference but perf_event_set_output() overrides it.
3032 * Here is the timeline for two threads T1, T2:
3033 * t0: T1, rb = rcu_dereference(event->rb)
3034 * t1: T2, old_rb = event->rb
3035 * t2: T2, event->rb = new rb
3036 * t3: T2, ring_buffer_detach(old_rb)
3037 * t4: T1, ring_buffer_attach(rb1)
3038 * t5: T1, poll_wait(event->waitq)
3040 * To avoid this problem, we grab mmap_mutex in perf_poll()
3041 * thereby ensuring that the assignment of the new ring buffer
3042 * and the detachment of the old buffer appear atomic to perf_poll()
3044 mutex_lock(&event->mmap_mutex);
3047 rb = rcu_dereference(event->rb);
3049 ring_buffer_attach(event, rb);
3050 events = atomic_xchg(&rb->poll, 0);
3054 mutex_unlock(&event->mmap_mutex);
3056 poll_wait(file, &event->waitq, wait);
3061 static void perf_event_reset(struct perf_event *event)
3063 (void)perf_event_read(event);
3064 local64_set(&event->count, 0);
3065 perf_event_update_userpage(event);
3069 * Holding the top-level event's child_mutex means that any
3070 * descendant process that has inherited this event will block
3071 * in sync_child_event if it goes to exit, thus satisfying the
3072 * task existence requirements of perf_event_enable/disable.
3074 static void perf_event_for_each_child(struct perf_event *event,
3075 void (*func)(struct perf_event *))
3077 struct perf_event *child;
3079 WARN_ON_ONCE(event->ctx->parent_ctx);
3080 mutex_lock(&event->child_mutex);
3082 list_for_each_entry(child, &event->child_list, child_list)
3084 mutex_unlock(&event->child_mutex);
3087 static void perf_event_for_each(struct perf_event *event,
3088 void (*func)(struct perf_event *))
3090 struct perf_event_context *ctx = event->ctx;
3091 struct perf_event *sibling;
3093 WARN_ON_ONCE(ctx->parent_ctx);
3094 mutex_lock(&ctx->mutex);
3095 event = event->group_leader;
3097 perf_event_for_each_child(event, func);
3099 list_for_each_entry(sibling, &event->sibling_list, group_entry)
3100 perf_event_for_each_child(event, func);
3101 mutex_unlock(&ctx->mutex);
3104 static int perf_event_period(struct perf_event *event, u64 __user *arg)
3106 struct perf_event_context *ctx = event->ctx;
3110 if (!is_sampling_event(event))
3113 if (copy_from_user(&value, arg, sizeof(value)))
3119 raw_spin_lock_irq(&ctx->lock);
3120 if (event->attr.freq) {
3121 if (value > sysctl_perf_event_sample_rate) {
3126 event->attr.sample_freq = value;
3128 event->attr.sample_period = value;
3129 event->hw.sample_period = value;
3132 raw_spin_unlock_irq(&ctx->lock);
3137 static const struct file_operations perf_fops;
3139 static struct perf_event *perf_fget_light(int fd, int *fput_needed)
3143 file = fget_light(fd, fput_needed);
3145 return ERR_PTR(-EBADF);
3147 if (file->f_op != &perf_fops) {
3148 fput_light(file, *fput_needed);
3150 return ERR_PTR(-EBADF);
3153 return file->private_data;
3156 static int perf_event_set_output(struct perf_event *event,
3157 struct perf_event *output_event);
3158 static int perf_event_set_filter(struct perf_event *event, void __user *arg);
3160 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
3162 struct perf_event *event = file->private_data;
3163 void (*func)(struct perf_event *);
3167 case PERF_EVENT_IOC_ENABLE:
3168 func = perf_event_enable;
3170 case PERF_EVENT_IOC_DISABLE:
3171 func = perf_event_disable;
3173 case PERF_EVENT_IOC_RESET:
3174 func = perf_event_reset;
3177 case PERF_EVENT_IOC_REFRESH:
3178 return perf_event_refresh(event, arg);
3180 case PERF_EVENT_IOC_PERIOD:
3181 return perf_event_period(event, (u64 __user *)arg);
3183 case PERF_EVENT_IOC_SET_OUTPUT:
3185 struct perf_event *output_event = NULL;
3186 int fput_needed = 0;
3190 output_event = perf_fget_light(arg, &fput_needed);
3191 if (IS_ERR(output_event))
3192 return PTR_ERR(output_event);
3195 ret = perf_event_set_output(event, output_event);
3197 fput_light(output_event->filp, fput_needed);
3202 case PERF_EVENT_IOC_SET_FILTER:
3203 return perf_event_set_filter(event, (void __user *)arg);
3209 if (flags & PERF_IOC_FLAG_GROUP)
3210 perf_event_for_each(event, func);
3212 perf_event_for_each_child(event, func);
3217 int perf_event_task_enable(void)
3219 struct perf_event *event;
3221 mutex_lock(¤t->perf_event_mutex);
3222 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
3223 perf_event_for_each_child(event, perf_event_enable);
3224 mutex_unlock(¤t->perf_event_mutex);
3229 int perf_event_task_disable(void)
3231 struct perf_event *event;
3233 mutex_lock(¤t->perf_event_mutex);
3234 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
3235 perf_event_for_each_child(event, perf_event_disable);
3236 mutex_unlock(¤t->perf_event_mutex);
3241 static int perf_event_index(struct perf_event *event)
3243 if (event->hw.state & PERF_HES_STOPPED)
3246 if (event->state != PERF_EVENT_STATE_ACTIVE)
3249 return event->pmu->event_idx(event);
3252 static void calc_timer_values(struct perf_event *event,
3259 *now = perf_clock();
3260 ctx_time = event->shadow_ctx_time + *now;
3261 *enabled = ctx_time - event->tstamp_enabled;
3262 *running = ctx_time - event->tstamp_running;
3265 void __weak perf_update_user_clock(struct perf_event_mmap_page *userpg, u64 now)
3270 * Callers need to ensure there can be no nesting of this function, otherwise
3271 * the seqlock logic goes bad. We can not serialize this because the arch
3272 * code calls this from NMI context.
3274 void perf_event_update_userpage(struct perf_event *event)
3276 struct perf_event_mmap_page *userpg;
3277 struct ring_buffer *rb;
3278 u64 enabled, running, now;
3282 * compute total_time_enabled, total_time_running
3283 * based on snapshot values taken when the event
3284 * was last scheduled in.
3286 * we cannot simply called update_context_time()
3287 * because of locking issue as we can be called in
3290 calc_timer_values(event, &now, &enabled, &running);
3291 rb = rcu_dereference(event->rb);
3295 userpg = rb->user_page;
3298 * Disable preemption so as to not let the corresponding user-space
3299 * spin too long if we get preempted.
3304 userpg->index = perf_event_index(event);
3305 userpg->offset = perf_event_count(event);
3307 userpg->offset -= local64_read(&event->hw.prev_count);
3309 userpg->time_enabled = enabled +
3310 atomic64_read(&event->child_total_time_enabled);
3312 userpg->time_running = running +
3313 atomic64_read(&event->child_total_time_running);
3315 perf_update_user_clock(userpg, now);
3324 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
3326 struct perf_event *event = vma->vm_file->private_data;
3327 struct ring_buffer *rb;
3328 int ret = VM_FAULT_SIGBUS;
3330 if (vmf->flags & FAULT_FLAG_MKWRITE) {
3331 if (vmf->pgoff == 0)
3337 rb = rcu_dereference(event->rb);
3341 if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE))
3344 vmf->page = perf_mmap_to_page(rb, vmf->pgoff);
3348 get_page(vmf->page);
3349 vmf->page->mapping = vma->vm_file->f_mapping;
3350 vmf->page->index = vmf->pgoff;
3359 static void ring_buffer_attach(struct perf_event *event,
3360 struct ring_buffer *rb)
3362 unsigned long flags;
3364 if (!list_empty(&event->rb_entry))
3367 spin_lock_irqsave(&rb->event_lock, flags);
3368 if (!list_empty(&event->rb_entry))
3371 list_add(&event->rb_entry, &rb->event_list);
3373 spin_unlock_irqrestore(&rb->event_lock, flags);
3376 static void ring_buffer_detach(struct perf_event *event,
3377 struct ring_buffer *rb)
3379 unsigned long flags;
3381 if (list_empty(&event->rb_entry))
3384 spin_lock_irqsave(&rb->event_lock, flags);
3385 list_del_init(&event->rb_entry);
3386 wake_up_all(&event->waitq);
3387 spin_unlock_irqrestore(&rb->event_lock, flags);
3390 static void ring_buffer_wakeup(struct perf_event *event)
3392 struct ring_buffer *rb;
3395 rb = rcu_dereference(event->rb);
3399 list_for_each_entry_rcu(event, &rb->event_list, rb_entry)
3400 wake_up_all(&event->waitq);
3406 static void rb_free_rcu(struct rcu_head *rcu_head)
3408 struct ring_buffer *rb;
3410 rb = container_of(rcu_head, struct ring_buffer, rcu_head);
3414 static struct ring_buffer *ring_buffer_get(struct perf_event *event)
3416 struct ring_buffer *rb;
3419 rb = rcu_dereference(event->rb);
3421 if (!atomic_inc_not_zero(&rb->refcount))
3429 static void ring_buffer_put(struct ring_buffer *rb)
3431 struct perf_event *event, *n;
3432 unsigned long flags;
3434 if (!atomic_dec_and_test(&rb->refcount))
3437 spin_lock_irqsave(&rb->event_lock, flags);
3438 list_for_each_entry_safe(event, n, &rb->event_list, rb_entry) {
3439 list_del_init(&event->rb_entry);
3440 wake_up_all(&event->waitq);
3442 spin_unlock_irqrestore(&rb->event_lock, flags);
3444 call_rcu(&rb->rcu_head, rb_free_rcu);
3447 static void perf_mmap_open(struct vm_area_struct *vma)
3449 struct perf_event *event = vma->vm_file->private_data;
3451 atomic_inc(&event->mmap_count);
3454 static void perf_mmap_close(struct vm_area_struct *vma)
3456 struct perf_event *event = vma->vm_file->private_data;
3458 if (atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex)) {
3459 unsigned long size = perf_data_size(event->rb);
3460 struct user_struct *user = event->mmap_user;
3461 struct ring_buffer *rb = event->rb;
3463 atomic_long_sub((size >> PAGE_SHIFT) + 1, &user->locked_vm);
3464 vma->vm_mm->pinned_vm -= event->mmap_locked;
3465 rcu_assign_pointer(event->rb, NULL);
3466 ring_buffer_detach(event, rb);
3467 mutex_unlock(&event->mmap_mutex);
3469 ring_buffer_put(rb);
3474 static const struct vm_operations_struct perf_mmap_vmops = {
3475 .open = perf_mmap_open,
3476 .close = perf_mmap_close,
3477 .fault = perf_mmap_fault,
3478 .page_mkwrite = perf_mmap_fault,
3481 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
3483 struct perf_event *event = file->private_data;
3484 unsigned long user_locked, user_lock_limit;
3485 struct user_struct *user = current_user();
3486 unsigned long locked, lock_limit;
3487 struct ring_buffer *rb;
3488 unsigned long vma_size;
3489 unsigned long nr_pages;
3490 long user_extra, extra;
3491 int ret = 0, flags = 0;
3494 * Don't allow mmap() of inherited per-task counters. This would
3495 * create a performance issue due to all children writing to the
3498 if (event->cpu == -1 && event->attr.inherit)
3501 if (!(vma->vm_flags & VM_SHARED))
3504 vma_size = vma->vm_end - vma->vm_start;
3505 nr_pages = (vma_size / PAGE_SIZE) - 1;
3508 * If we have rb pages ensure they're a power-of-two number, so we
3509 * can do bitmasks instead of modulo.
3511 if (nr_pages != 0 && !is_power_of_2(nr_pages))
3514 if (vma_size != PAGE_SIZE * (1 + nr_pages))
3517 if (vma->vm_pgoff != 0)
3520 WARN_ON_ONCE(event->ctx->parent_ctx);
3521 mutex_lock(&event->mmap_mutex);
3523 if (event->rb->nr_pages == nr_pages)
3524 atomic_inc(&event->rb->refcount);
3530 user_extra = nr_pages + 1;
3531 user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10);
3534 * Increase the limit linearly with more CPUs:
3536 user_lock_limit *= num_online_cpus();
3538 user_locked = atomic_long_read(&user->locked_vm) + user_extra;
3541 if (user_locked > user_lock_limit)
3542 extra = user_locked - user_lock_limit;
3544 lock_limit = rlimit(RLIMIT_MEMLOCK);
3545 lock_limit >>= PAGE_SHIFT;
3546 locked = vma->vm_mm->pinned_vm + extra;
3548 if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() &&
3549 !capable(CAP_IPC_LOCK)) {
3556 if (vma->vm_flags & VM_WRITE)
3557 flags |= RING_BUFFER_WRITABLE;
3559 rb = rb_alloc(nr_pages,
3560 event->attr.watermark ? event->attr.wakeup_watermark : 0,
3567 rcu_assign_pointer(event->rb, rb);
3569 atomic_long_add(user_extra, &user->locked_vm);
3570 event->mmap_locked = extra;
3571 event->mmap_user = get_current_user();
3572 vma->vm_mm->pinned_vm += event->mmap_locked;
3574 perf_event_update_userpage(event);
3578 atomic_inc(&event->mmap_count);
3579 mutex_unlock(&event->mmap_mutex);
3581 vma->vm_flags |= VM_RESERVED;
3582 vma->vm_ops = &perf_mmap_vmops;
3587 static int perf_fasync(int fd, struct file *filp, int on)
3589 struct inode *inode = filp->f_path.dentry->d_inode;
3590 struct perf_event *event = filp->private_data;
3593 mutex_lock(&inode->i_mutex);
3594 retval = fasync_helper(fd, filp, on, &event->fasync);
3595 mutex_unlock(&inode->i_mutex);
3603 static const struct file_operations perf_fops = {
3604 .llseek = no_llseek,
3605 .release = perf_release,
3608 .unlocked_ioctl = perf_ioctl,
3609 .compat_ioctl = perf_ioctl,
3611 .fasync = perf_fasync,
3617 * If there's data, ensure we set the poll() state and publish everything
3618 * to user-space before waking everybody up.
3621 void perf_event_wakeup(struct perf_event *event)
3623 ring_buffer_wakeup(event);
3625 if (event->pending_kill) {
3626 kill_fasync(&event->fasync, SIGIO, event->pending_kill);
3627 event->pending_kill = 0;
3631 static void perf_pending_event(struct irq_work *entry)
3633 struct perf_event *event = container_of(entry,
3634 struct perf_event, pending);
3636 if (event->pending_disable) {
3637 event->pending_disable = 0;
3638 __perf_event_disable(event);
3641 if (event->pending_wakeup) {
3642 event->pending_wakeup = 0;
3643 perf_event_wakeup(event);
3648 * We assume there is only KVM supporting the callbacks.
3649 * Later on, we might change it to a list if there is
3650 * another virtualization implementation supporting the callbacks.
3652 struct perf_guest_info_callbacks *perf_guest_cbs;
3654 int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
3656 perf_guest_cbs = cbs;
3659 EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks);
3661 int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
3663 perf_guest_cbs = NULL;
3666 EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks);
3668 static void __perf_event_header__init_id(struct perf_event_header *header,
3669 struct perf_sample_data *data,
3670 struct perf_event *event)
3672 u64 sample_type = event->attr.sample_type;
3674 data->type = sample_type;
3675 header->size += event->id_header_size;
3677 if (sample_type & PERF_SAMPLE_TID) {
3678 /* namespace issues */
3679 data->tid_entry.pid = perf_event_pid(event, current);
3680 data->tid_entry.tid = perf_event_tid(event, current);
3683 if (sample_type & PERF_SAMPLE_TIME)
3684 data->time = perf_clock();
3686 if (sample_type & PERF_SAMPLE_ID)
3687 data->id = primary_event_id(event);
3689 if (sample_type & PERF_SAMPLE_STREAM_ID)
3690 data->stream_id = event->id;
3692 if (sample_type & PERF_SAMPLE_CPU) {
3693 data->cpu_entry.cpu = raw_smp_processor_id();
3694 data->cpu_entry.reserved = 0;
3698 void perf_event_header__init_id(struct perf_event_header *header,
3699 struct perf_sample_data *data,
3700 struct perf_event *event)
3702 if (event->attr.sample_id_all)
3703 __perf_event_header__init_id(header, data, event);
3706 static void __perf_event__output_id_sample(struct perf_output_handle *handle,
3707 struct perf_sample_data *data)
3709 u64 sample_type = data->type;
3711 if (sample_type & PERF_SAMPLE_TID)
3712 perf_output_put(handle, data->tid_entry);
3714 if (sample_type & PERF_SAMPLE_TIME)
3715 perf_output_put(handle, data->time);
3717 if (sample_type & PERF_SAMPLE_ID)
3718 perf_output_put(handle, data->id);
3720 if (sample_type & PERF_SAMPLE_STREAM_ID)
3721 perf_output_put(handle, data->stream_id);
3723 if (sample_type & PERF_SAMPLE_CPU)
3724 perf_output_put(handle, data->cpu_entry);
3727 void perf_event__output_id_sample(struct perf_event *event,
3728 struct perf_output_handle *handle,
3729 struct perf_sample_data *sample)
3731 if (event->attr.sample_id_all)
3732 __perf_event__output_id_sample(handle, sample);
3735 static void perf_output_read_one(struct perf_output_handle *handle,
3736 struct perf_event *event,
3737 u64 enabled, u64 running)
3739 u64 read_format = event->attr.read_format;
3743 values[n++] = perf_event_count(event);
3744 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
3745 values[n++] = enabled +
3746 atomic64_read(&event->child_total_time_enabled);
3748 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
3749 values[n++] = running +
3750 atomic64_read(&event->child_total_time_running);
3752 if (read_format & PERF_FORMAT_ID)
3753 values[n++] = primary_event_id(event);
3755 __output_copy(handle, values, n * sizeof(u64));
3759 * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
3761 static void perf_output_read_group(struct perf_output_handle *handle,
3762 struct perf_event *event,
3763 u64 enabled, u64 running)
3765 struct perf_event *leader = event->group_leader, *sub;
3766 u64 read_format = event->attr.read_format;
3770 values[n++] = 1 + leader->nr_siblings;
3772 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3773 values[n++] = enabled;
3775 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3776 values[n++] = running;
3778 if (leader != event)
3779 leader->pmu->read(leader);
3781 values[n++] = perf_event_count(leader);
3782 if (read_format & PERF_FORMAT_ID)
3783 values[n++] = primary_event_id(leader);
3785 __output_copy(handle, values, n * sizeof(u64));
3787 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
3791 sub->pmu->read(sub);
3793 values[n++] = perf_event_count(sub);
3794 if (read_format & PERF_FORMAT_ID)
3795 values[n++] = primary_event_id(sub);
3797 __output_copy(handle, values, n * sizeof(u64));
3801 #define PERF_FORMAT_TOTAL_TIMES (PERF_FORMAT_TOTAL_TIME_ENABLED|\
3802 PERF_FORMAT_TOTAL_TIME_RUNNING)
3804 static void perf_output_read(struct perf_output_handle *handle,
3805 struct perf_event *event)
3807 u64 enabled = 0, running = 0, now;
3808 u64 read_format = event->attr.read_format;
3811 * compute total_time_enabled, total_time_running
3812 * based on snapshot values taken when the event
3813 * was last scheduled in.
3815 * we cannot simply called update_context_time()
3816 * because of locking issue as we are called in
3819 if (read_format & PERF_FORMAT_TOTAL_TIMES)
3820 calc_timer_values(event, &now, &enabled, &running);
3822 if (event->attr.read_format & PERF_FORMAT_GROUP)
3823 perf_output_read_group(handle, event, enabled, running);
3825 perf_output_read_one(handle, event, enabled, running);
3828 void perf_output_sample(struct perf_output_handle *handle,
3829 struct perf_event_header *header,
3830 struct perf_sample_data *data,
3831 struct perf_event *event)
3833 u64 sample_type = data->type;
3835 perf_output_put(handle, *header);
3837 if (sample_type & PERF_SAMPLE_IP)
3838 perf_output_put(handle, data->ip);
3840 if (sample_type & PERF_SAMPLE_TID)
3841 perf_output_put(handle, data->tid_entry);
3843 if (sample_type & PERF_SAMPLE_TIME)
3844 perf_output_put(handle, data->time);
3846 if (sample_type & PERF_SAMPLE_ADDR)
3847 perf_output_put(handle, data->addr);
3849 if (sample_type & PERF_SAMPLE_ID)
3850 perf_output_put(handle, data->id);
3852 if (sample_type & PERF_SAMPLE_STREAM_ID)
3853 perf_output_put(handle, data->stream_id);
3855 if (sample_type & PERF_SAMPLE_CPU)
3856 perf_output_put(handle, data->cpu_entry);
3858 if (sample_type & PERF_SAMPLE_PERIOD)
3859 perf_output_put(handle, data->period);
3861 if (sample_type & PERF_SAMPLE_READ)
3862 perf_output_read(handle, event);
3864 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3865 if (data->callchain) {
3868 if (data->callchain)
3869 size += data->callchain->nr;
3871 size *= sizeof(u64);
3873 __output_copy(handle, data->callchain, size);
3876 perf_output_put(handle, nr);
3880 if (sample_type & PERF_SAMPLE_RAW) {
3882 perf_output_put(handle, data->raw->size);
3883 __output_copy(handle, data->raw->data,
3890 .size = sizeof(u32),
3893 perf_output_put(handle, raw);
3897 if (!event->attr.watermark) {
3898 int wakeup_events = event->attr.wakeup_events;
3900 if (wakeup_events) {
3901 struct ring_buffer *rb = handle->rb;
3902 int events = local_inc_return(&rb->events);
3904 if (events >= wakeup_events) {
3905 local_sub(wakeup_events, &rb->events);
3906 local_inc(&rb->wakeup);
3912 void perf_prepare_sample(struct perf_event_header *header,
3913 struct perf_sample_data *data,
3914 struct perf_event *event,
3915 struct pt_regs *regs)
3917 u64 sample_type = event->attr.sample_type;
3919 header->type = PERF_RECORD_SAMPLE;
3920 header->size = sizeof(*header) + event->header_size;
3923 header->misc |= perf_misc_flags(regs);
3925 __perf_event_header__init_id(header, data, event);
3927 if (sample_type & PERF_SAMPLE_IP)
3928 data->ip = perf_instruction_pointer(regs);
3930 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3933 data->callchain = perf_callchain(regs);
3935 if (data->callchain)
3936 size += data->callchain->nr;
3938 header->size += size * sizeof(u64);
3941 if (sample_type & PERF_SAMPLE_RAW) {
3942 int size = sizeof(u32);
3945 size += data->raw->size;
3947 size += sizeof(u32);
3949 WARN_ON_ONCE(size & (sizeof(u64)-1));
3950 header->size += size;
3954 static void perf_event_output(struct perf_event *event,
3955 struct perf_sample_data *data,
3956 struct pt_regs *regs)
3958 struct perf_output_handle handle;
3959 struct perf_event_header header;
3961 /* protect the callchain buffers */
3964 perf_prepare_sample(&header, data, event, regs);
3966 if (perf_output_begin(&handle, event, header.size))
3969 perf_output_sample(&handle, &header, data, event);
3971 perf_output_end(&handle);
3981 struct perf_read_event {
3982 struct perf_event_header header;
3989 perf_event_read_event(struct perf_event *event,
3990 struct task_struct *task)
3992 struct perf_output_handle handle;
3993 struct perf_sample_data sample;
3994 struct perf_read_event read_event = {
3996 .type = PERF_RECORD_READ,
3998 .size = sizeof(read_event) + event->read_size,
4000 .pid = perf_event_pid(event, task),
4001 .tid = perf_event_tid(event, task),
4005 perf_event_header__init_id(&read_event.header, &sample, event);
4006 ret = perf_output_begin(&handle, event, read_event.header.size);
4010 perf_output_put(&handle, read_event);
4011 perf_output_read(&handle, event);
4012 perf_event__output_id_sample(event, &handle, &sample);
4014 perf_output_end(&handle);
4018 * task tracking -- fork/exit
4020 * enabled by: attr.comm | attr.mmap | attr.mmap_data | attr.task
4023 struct perf_task_event {
4024 struct task_struct *task;
4025 struct perf_event_context *task_ctx;
4028 struct perf_event_header header;
4038 static void perf_event_task_output(struct perf_event *event,
4039 struct perf_task_event *task_event)
4041 struct perf_output_handle handle;
4042 struct perf_sample_data sample;
4043 struct task_struct *task = task_event->task;
4044 int ret, size = task_event->event_id.header.size;
4046 perf_event_header__init_id(&task_event->event_id.header, &sample, event);
4048 ret = perf_output_begin(&handle, event,
4049 task_event->event_id.header.size);
4053 task_event->event_id.pid = perf_event_pid(event, task);
4054 task_event->event_id.ppid = perf_event_pid(event, current);
4056 task_event->event_id.tid = perf_event_tid(event, task);
4057 task_event->event_id.ptid = perf_event_tid(event, current);
4059 perf_output_put(&handle, task_event->event_id);
4061 perf_event__output_id_sample(event, &handle, &sample);
4063 perf_output_end(&handle);
4065 task_event->event_id.header.size = size;
4068 static int perf_event_task_match(struct perf_event *event)
4070 if (event->state < PERF_EVENT_STATE_INACTIVE)
4073 if (!event_filter_match(event))
4076 if (event->attr.comm || event->attr.mmap ||
4077 event->attr.mmap_data || event->attr.task)
4083 static void perf_event_task_ctx(struct perf_event_context *ctx,
4084 struct perf_task_event *task_event)
4086 struct perf_event *event;
4088 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
4089 if (perf_event_task_match(event))
4090 perf_event_task_output(event, task_event);
4094 static void perf_event_task_event(struct perf_task_event *task_event)
4096 struct perf_cpu_context *cpuctx;
4097 struct perf_event_context *ctx;
4102 list_for_each_entry_rcu(pmu, &pmus, entry) {
4103 cpuctx = get_cpu_ptr(pmu->pmu_cpu_context);
4104 if (cpuctx->active_pmu != pmu)
4106 perf_event_task_ctx(&cpuctx->ctx, task_event);
4108 ctx = task_event->task_ctx;
4110 ctxn = pmu->task_ctx_nr;
4113 ctx = rcu_dereference(current->perf_event_ctxp[ctxn]);
4116 perf_event_task_ctx(ctx, task_event);
4118 put_cpu_ptr(pmu->pmu_cpu_context);
4123 static void perf_event_task(struct task_struct *task,
4124 struct perf_event_context *task_ctx,
4127 struct perf_task_event task_event;
4129 if (!atomic_read(&nr_comm_events) &&
4130 !atomic_read(&nr_mmap_events) &&
4131 !atomic_read(&nr_task_events))
4134 task_event = (struct perf_task_event){
4136 .task_ctx = task_ctx,
4139 .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT,
4141 .size = sizeof(task_event.event_id),
4147 .time = perf_clock(),
4151 perf_event_task_event(&task_event);
4154 void perf_event_fork(struct task_struct *task)
4156 perf_event_task(task, NULL, 1);
4163 struct perf_comm_event {
4164 struct task_struct *task;
4169 struct perf_event_header header;
4176 static void perf_event_comm_output(struct perf_event *event,
4177 struct perf_comm_event *comm_event)
4179 struct perf_output_handle handle;
4180 struct perf_sample_data sample;
4181 int size = comm_event->event_id.header.size;
4184 perf_event_header__init_id(&comm_event->event_id.header, &sample, event);
4185 ret = perf_output_begin(&handle, event,
4186 comm_event->event_id.header.size);
4191 comm_event->event_id.pid = perf_event_pid(event, comm_event->task);
4192 comm_event->event_id.tid = perf_event_tid(event, comm_event->task);
4194 perf_output_put(&handle, comm_event->event_id);
4195 __output_copy(&handle, comm_event->comm,
4196 comm_event->comm_size);
4198 perf_event__output_id_sample(event, &handle, &sample);
4200 perf_output_end(&handle);
4202 comm_event->event_id.header.size = size;
4205 static int perf_event_comm_match(struct perf_event *event)
4207 if (event->state < PERF_EVENT_STATE_INACTIVE)
4210 if (!event_filter_match(event))
4213 if (event->attr.comm)
4219 static void perf_event_comm_ctx(struct perf_event_context *ctx,
4220 struct perf_comm_event *comm_event)
4222 struct perf_event *event;
4224 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
4225 if (perf_event_comm_match(event))
4226 perf_event_comm_output(event, comm_event);
4230 static void perf_event_comm_event(struct perf_comm_event *comm_event)
4232 struct perf_cpu_context *cpuctx;
4233 struct perf_event_context *ctx;
4234 char comm[TASK_COMM_LEN];
4239 memset(comm, 0, sizeof(comm));
4240 strlcpy(comm, comm_event->task->comm, sizeof(comm));
4241 size = ALIGN(strlen(comm)+1, sizeof(u64));
4243 comm_event->comm = comm;
4244 comm_event->comm_size = size;
4246 comm_event->event_id.header.size = sizeof(comm_event->event_id) + size;
4248 list_for_each_entry_rcu(pmu, &pmus, entry) {
4249 cpuctx = get_cpu_ptr(pmu->pmu_cpu_context);
4250 if (cpuctx->active_pmu != pmu)
4252 perf_event_comm_ctx(&cpuctx->ctx, comm_event);
4254 ctxn = pmu->task_ctx_nr;
4258 ctx = rcu_dereference(current->perf_event_ctxp[ctxn]);
4260 perf_event_comm_ctx(ctx, comm_event);
4262 put_cpu_ptr(pmu->pmu_cpu_context);
4267 void perf_event_comm(struct task_struct *task)
4269 struct perf_comm_event comm_event;
4270 struct perf_event_context *ctx;
4273 for_each_task_context_nr(ctxn) {
4274 ctx = task->perf_event_ctxp[ctxn];
4278 perf_event_enable_on_exec(ctx);
4281 if (!atomic_read(&nr_comm_events))
4284 comm_event = (struct perf_comm_event){
4290 .type = PERF_RECORD_COMM,
4299 perf_event_comm_event(&comm_event);
4306 struct perf_mmap_event {
4307 struct vm_area_struct *vma;
4309 const char *file_name;
4313 struct perf_event_header header;
4323 static void perf_event_mmap_output(struct perf_event *event,
4324 struct perf_mmap_event *mmap_event)
4326 struct perf_output_handle handle;
4327 struct perf_sample_data sample;
4328 int size = mmap_event->event_id.header.size;
4331 perf_event_header__init_id(&mmap_event->event_id.header, &sample, event);
4332 ret = perf_output_begin(&handle, event,
4333 mmap_event->event_id.header.size);
4337 mmap_event->event_id.pid = perf_event_pid(event, current);
4338 mmap_event->event_id.tid = perf_event_tid(event, current);
4340 perf_output_put(&handle, mmap_event->event_id);
4341 __output_copy(&handle, mmap_event->file_name,
4342 mmap_event->file_size);
4344 perf_event__output_id_sample(event, &handle, &sample);
4346 perf_output_end(&handle);
4348 mmap_event->event_id.header.size = size;
4351 static int perf_event_mmap_match(struct perf_event *event,
4352 struct perf_mmap_event *mmap_event,
4355 if (event->state < PERF_EVENT_STATE_INACTIVE)
4358 if (!event_filter_match(event))
4361 if ((!executable && event->attr.mmap_data) ||
4362 (executable && event->attr.mmap))
4368 static void perf_event_mmap_ctx(struct perf_event_context *ctx,
4369 struct perf_mmap_event *mmap_event,
4372 struct perf_event *event;
4374 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
4375 if (perf_event_mmap_match(event, mmap_event, executable))
4376 perf_event_mmap_output(event, mmap_event);
4380 static void perf_event_mmap_event(struct perf_mmap_event *mmap_event)
4382 struct perf_cpu_context *cpuctx;
4383 struct perf_event_context *ctx;
4384 struct vm_area_struct *vma = mmap_event->vma;
4385 struct file *file = vma->vm_file;
4393 memset(tmp, 0, sizeof(tmp));
4397 * d_path works from the end of the rb backwards, so we
4398 * need to add enough zero bytes after the string to handle
4399 * the 64bit alignment we do later.
4401 buf = kzalloc(PATH_MAX + sizeof(u64), GFP_KERNEL);
4403 name = strncpy(tmp, "//enomem", sizeof(tmp));
4406 name = d_path(&file->f_path, buf, PATH_MAX);
4408 name = strncpy(tmp, "//toolong", sizeof(tmp));
4412 if (arch_vma_name(mmap_event->vma)) {
4413 name = strncpy(tmp, arch_vma_name(mmap_event->vma),
4419 name = strncpy(tmp, "[vdso]", sizeof(tmp));
4421 } else if (vma->vm_start <= vma->vm_mm->start_brk &&
4422 vma->vm_end >= vma->vm_mm->brk) {
4423 name = strncpy(tmp, "[heap]", sizeof(tmp));
4425 } else if (vma->vm_start <= vma->vm_mm->start_stack &&
4426 vma->vm_end >= vma->vm_mm->start_stack) {
4427 name = strncpy(tmp, "[stack]", sizeof(tmp));
4431 name = strncpy(tmp, "//anon", sizeof(tmp));
4436 size = ALIGN(strlen(name)+1, sizeof(u64));
4438 mmap_event->file_name = name;
4439 mmap_event->file_size = size;
4441 mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size;
4444 list_for_each_entry_rcu(pmu, &pmus, entry) {
4445 cpuctx = get_cpu_ptr(pmu->pmu_cpu_context);
4446 if (cpuctx->active_pmu != pmu)
4448 perf_event_mmap_ctx(&cpuctx->ctx, mmap_event,
4449 vma->vm_flags & VM_EXEC);
4451 ctxn = pmu->task_ctx_nr;
4455 ctx = rcu_dereference(current->perf_event_ctxp[ctxn]);
4457 perf_event_mmap_ctx(ctx, mmap_event,
4458 vma->vm_flags & VM_EXEC);
4461 put_cpu_ptr(pmu->pmu_cpu_context);
4468 void perf_event_mmap(struct vm_area_struct *vma)
4470 struct perf_mmap_event mmap_event;
4472 if (!atomic_read(&nr_mmap_events))
4475 mmap_event = (struct perf_mmap_event){
4481 .type = PERF_RECORD_MMAP,
4482 .misc = PERF_RECORD_MISC_USER,
4487 .start = vma->vm_start,
4488 .len = vma->vm_end - vma->vm_start,
4489 .pgoff = (u64)vma->vm_pgoff << PAGE_SHIFT,
4493 perf_event_mmap_event(&mmap_event);
4497 * IRQ throttle logging
4500 static void perf_log_throttle(struct perf_event *event, int enable)
4502 struct perf_output_handle handle;
4503 struct perf_sample_data sample;
4507 struct perf_event_header header;
4511 } throttle_event = {
4513 .type = PERF_RECORD_THROTTLE,
4515 .size = sizeof(throttle_event),
4517 .time = perf_clock(),
4518 .id = primary_event_id(event),
4519 .stream_id = event->id,
4523 throttle_event.header.type = PERF_RECORD_UNTHROTTLE;
4525 perf_event_header__init_id(&throttle_event.header, &sample, event);
4527 ret = perf_output_begin(&handle, event,
4528 throttle_event.header.size);
4532 perf_output_put(&handle, throttle_event);
4533 perf_event__output_id_sample(event, &handle, &sample);
4534 perf_output_end(&handle);
4538 * Generic event overflow handling, sampling.
4541 static int __perf_event_overflow(struct perf_event *event,
4542 int throttle, struct perf_sample_data *data,
4543 struct pt_regs *regs)
4545 int events = atomic_read(&event->event_limit);
4546 struct hw_perf_event *hwc = &event->hw;
4551 * Non-sampling counters might still use the PMI to fold short
4552 * hardware counters, ignore those.
4554 if (unlikely(!is_sampling_event(event)))
4557 seq = __this_cpu_read(perf_throttled_seq);
4558 if (seq != hwc->interrupts_seq) {
4559 hwc->interrupts_seq = seq;
4560 hwc->interrupts = 1;
4563 if (unlikely(throttle
4564 && hwc->interrupts >= max_samples_per_tick)) {
4565 __this_cpu_inc(perf_throttled_count);
4566 hwc->interrupts = MAX_INTERRUPTS;
4567 perf_log_throttle(event, 0);
4572 if (event->attr.freq) {
4573 u64 now = perf_clock();
4574 s64 delta = now - hwc->freq_time_stamp;
4576 hwc->freq_time_stamp = now;
4578 if (delta > 0 && delta < 2*TICK_NSEC)
4579 perf_adjust_period(event, delta, hwc->last_period, true);
4583 * XXX event_limit might not quite work as expected on inherited
4587 event->pending_kill = POLL_IN;
4588 if (events && atomic_dec_and_test(&event->event_limit)) {
4590 event->pending_kill = POLL_HUP;
4591 event->pending_disable = 1;
4592 irq_work_queue(&event->pending);
4595 if (event->overflow_handler)
4596 event->overflow_handler(event, data, regs);
4598 perf_event_output(event, data, regs);
4600 if (event->fasync && event->pending_kill) {
4601 event->pending_wakeup = 1;
4602 irq_work_queue(&event->pending);
4608 int perf_event_overflow(struct perf_event *event,
4609 struct perf_sample_data *data,
4610 struct pt_regs *regs)
4612 return __perf_event_overflow(event, 1, data, regs);
4616 * Generic software event infrastructure
4619 struct swevent_htable {
4620 struct swevent_hlist *swevent_hlist;
4621 struct mutex hlist_mutex;
4624 /* Recursion avoidance in each contexts */
4625 int recursion[PERF_NR_CONTEXTS];
4628 static DEFINE_PER_CPU(struct swevent_htable, swevent_htable);
4631 * We directly increment event->count and keep a second value in
4632 * event->hw.period_left to count intervals. This period event
4633 * is kept in the range [-sample_period, 0] so that we can use the
4637 static u64 perf_swevent_set_period(struct perf_event *event)
4639 struct hw_perf_event *hwc = &event->hw;
4640 u64 period = hwc->last_period;
4644 hwc->last_period = hwc->sample_period;
4647 old = val = local64_read(&hwc->period_left);
4651 nr = div64_u64(period + val, period);
4652 offset = nr * period;
4654 if (local64_cmpxchg(&hwc->period_left, old, val) != old)
4660 static void perf_swevent_overflow(struct perf_event *event, u64 overflow,
4661 struct perf_sample_data *data,
4662 struct pt_regs *regs)
4664 struct hw_perf_event *hwc = &event->hw;
4668 overflow = perf_swevent_set_period(event);
4670 if (hwc->interrupts == MAX_INTERRUPTS)
4673 for (; overflow; overflow--) {
4674 if (__perf_event_overflow(event, throttle,
4677 * We inhibit the overflow from happening when
4678 * hwc->interrupts == MAX_INTERRUPTS.
4686 static void perf_swevent_event(struct perf_event *event, u64 nr,
4687 struct perf_sample_data *data,
4688 struct pt_regs *regs)
4690 struct hw_perf_event *hwc = &event->hw;
4692 local64_add(nr, &event->count);
4697 if (!is_sampling_event(event))
4700 if ((event->attr.sample_type & PERF_SAMPLE_PERIOD) && !event->attr.freq) {
4702 return perf_swevent_overflow(event, 1, data, regs);
4704 data->period = event->hw.last_period;
4706 if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq)
4707 return perf_swevent_overflow(event, 1, data, regs);
4709 if (local64_add_negative(nr, &hwc->period_left))
4712 perf_swevent_overflow(event, 0, data, regs);
4715 static int perf_exclude_event(struct perf_event *event,
4716 struct pt_regs *regs)
4718 if (event->hw.state & PERF_HES_STOPPED)
4722 if (event->attr.exclude_user && user_mode(regs))
4725 if (event->attr.exclude_kernel && !user_mode(regs))
4732 static int perf_swevent_match(struct perf_event *event,
4733 enum perf_type_id type,
4735 struct perf_sample_data *data,
4736 struct pt_regs *regs)
4738 if (event->attr.type != type)
4741 if (event->attr.config != event_id)
4744 if (perf_exclude_event(event, regs))
4750 static inline u64 swevent_hash(u64 type, u32 event_id)
4752 u64 val = event_id | (type << 32);
4754 return hash_64(val, SWEVENT_HLIST_BITS);
4757 static inline struct hlist_head *
4758 __find_swevent_head(struct swevent_hlist *hlist, u64 type, u32 event_id)
4760 u64 hash = swevent_hash(type, event_id);
4762 return &hlist->heads[hash];
4765 /* For the read side: events when they trigger */
4766 static inline struct hlist_head *
4767 find_swevent_head_rcu(struct swevent_htable *swhash, u64 type, u32 event_id)
4769 struct swevent_hlist *hlist;
4771 hlist = rcu_dereference(swhash->swevent_hlist);
4775 return __find_swevent_head(hlist, type, event_id);
4778 /* For the event head insertion and removal in the hlist */
4779 static inline struct hlist_head *
4780 find_swevent_head(struct swevent_htable *swhash, struct perf_event *event)
4782 struct swevent_hlist *hlist;
4783 u32 event_id = event->attr.config;
4784 u64 type = event->attr.type;
4787 * Event scheduling is always serialized against hlist allocation
4788 * and release. Which makes the protected version suitable here.
4789 * The context lock guarantees that.
4791 hlist = rcu_dereference_protected(swhash->swevent_hlist,
4792 lockdep_is_held(&event->ctx->lock));
4796 return __find_swevent_head(hlist, type, event_id);
4799 static void do_perf_sw_event(enum perf_type_id type, u32 event_id,
4801 struct perf_sample_data *data,
4802 struct pt_regs *regs)
4804 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
4805 struct perf_event *event;
4806 struct hlist_node *node;
4807 struct hlist_head *head;
4810 head = find_swevent_head_rcu(swhash, type, event_id);
4814 hlist_for_each_entry_rcu(event, node, head, hlist_entry) {
4815 if (perf_swevent_match(event, type, event_id, data, regs))
4816 perf_swevent_event(event, nr, data, regs);
4822 int perf_swevent_get_recursion_context(void)
4824 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
4826 return get_recursion_context(swhash->recursion);
4828 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context);
4830 inline void perf_swevent_put_recursion_context(int rctx)
4832 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
4834 put_recursion_context(swhash->recursion, rctx);
4837 void __perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)
4839 struct perf_sample_data data;
4842 preempt_disable_notrace();
4843 rctx = perf_swevent_get_recursion_context();
4847 perf_sample_data_init(&data, addr);
4849 do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, &data, regs);
4851 perf_swevent_put_recursion_context(rctx);
4852 preempt_enable_notrace();
4855 static void perf_swevent_read(struct perf_event *event)
4859 static int perf_swevent_add(struct perf_event *event, int flags)
4861 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
4862 struct hw_perf_event *hwc = &event->hw;
4863 struct hlist_head *head;
4865 if (is_sampling_event(event)) {
4866 hwc->last_period = hwc->sample_period;
4867 perf_swevent_set_period(event);
4870 hwc->state = !(flags & PERF_EF_START);
4872 head = find_swevent_head(swhash, event);
4873 if (WARN_ON_ONCE(!head))
4876 hlist_add_head_rcu(&event->hlist_entry, head);
4881 static void perf_swevent_del(struct perf_event *event, int flags)
4883 hlist_del_rcu(&event->hlist_entry);
4886 static void perf_swevent_start(struct perf_event *event, int flags)
4888 event->hw.state = 0;
4891 static void perf_swevent_stop(struct perf_event *event, int flags)
4893 event->hw.state = PERF_HES_STOPPED;
4896 /* Deref the hlist from the update side */
4897 static inline struct swevent_hlist *
4898 swevent_hlist_deref(struct swevent_htable *swhash)
4900 return rcu_dereference_protected(swhash->swevent_hlist,
4901 lockdep_is_held(&swhash->hlist_mutex));
4904 static void swevent_hlist_release(struct swevent_htable *swhash)
4906 struct swevent_hlist *hlist = swevent_hlist_deref(swhash);
4911 rcu_assign_pointer(swhash->swevent_hlist, NULL);
4912 kfree_rcu(hlist, rcu_head);
4915 static void swevent_hlist_put_cpu(struct perf_event *event, int cpu)
4917 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
4919 mutex_lock(&swhash->hlist_mutex);
4921 if (!--swhash->hlist_refcount)
4922 swevent_hlist_release(swhash);
4924 mutex_unlock(&swhash->hlist_mutex);
4927 static void swevent_hlist_put(struct perf_event *event)
4931 if (event->cpu != -1) {
4932 swevent_hlist_put_cpu(event, event->cpu);
4936 for_each_possible_cpu(cpu)
4937 swevent_hlist_put_cpu(event, cpu);
4940 static int swevent_hlist_get_cpu(struct perf_event *event, int cpu)
4942 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
4945 mutex_lock(&swhash->hlist_mutex);
4947 if (!swevent_hlist_deref(swhash) && cpu_online(cpu)) {
4948 struct swevent_hlist *hlist;
4950 hlist = kzalloc(sizeof(*hlist), GFP_KERNEL);
4955 rcu_assign_pointer(swhash->swevent_hlist, hlist);
4957 swhash->hlist_refcount++;
4959 mutex_unlock(&swhash->hlist_mutex);
4964 static int swevent_hlist_get(struct perf_event *event)
4967 int cpu, failed_cpu;
4969 if (event->cpu != -1)
4970 return swevent_hlist_get_cpu(event, event->cpu);
4973 for_each_possible_cpu(cpu) {
4974 err = swevent_hlist_get_cpu(event, cpu);
4984 for_each_possible_cpu(cpu) {
4985 if (cpu == failed_cpu)
4987 swevent_hlist_put_cpu(event, cpu);
4994 struct static_key perf_swevent_enabled[PERF_COUNT_SW_MAX];
4996 static void sw_perf_event_destroy(struct perf_event *event)
4998 u64 event_id = event->attr.config;
5000 WARN_ON(event->parent);
5002 static_key_slow_dec(&perf_swevent_enabled[event_id]);
5003 swevent_hlist_put(event);
5006 static int perf_swevent_init(struct perf_event *event)
5008 int event_id = event->attr.config;
5010 if (event->attr.type != PERF_TYPE_SOFTWARE)
5014 case PERF_COUNT_SW_CPU_CLOCK:
5015 case PERF_COUNT_SW_TASK_CLOCK:
5022 if (event_id >= PERF_COUNT_SW_MAX)
5025 if (!event->parent) {
5028 err = swevent_hlist_get(event);
5032 static_key_slow_inc(&perf_swevent_enabled[event_id]);
5033 event->destroy = sw_perf_event_destroy;
5039 static int perf_swevent_event_idx(struct perf_event *event)
5044 static struct pmu perf_swevent = {
5045 .task_ctx_nr = perf_sw_context,
5047 .event_init = perf_swevent_init,
5048 .add = perf_swevent_add,
5049 .del = perf_swevent_del,
5050 .start = perf_swevent_start,
5051 .stop = perf_swevent_stop,
5052 .read = perf_swevent_read,
5054 .event_idx = perf_swevent_event_idx,
5057 #ifdef CONFIG_EVENT_TRACING
5059 static int perf_tp_filter_match(struct perf_event *event,
5060 struct perf_sample_data *data)
5062 void *record = data->raw->data;
5064 if (likely(!event->filter) || filter_match_preds(event->filter, record))
5069 static int perf_tp_event_match(struct perf_event *event,
5070 struct perf_sample_data *data,
5071 struct pt_regs *regs)
5073 if (event->hw.state & PERF_HES_STOPPED)
5076 * All tracepoints are from kernel-space.
5078 if (event->attr.exclude_kernel)
5081 if (!perf_tp_filter_match(event, data))
5087 void perf_tp_event(u64 addr, u64 count, void *record, int entry_size,
5088 struct pt_regs *regs, struct hlist_head *head, int rctx)
5090 struct perf_sample_data data;
5091 struct perf_event *event;
5092 struct hlist_node *node;
5094 struct perf_raw_record raw = {
5099 perf_sample_data_init(&data, addr);
5102 hlist_for_each_entry_rcu(event, node, head, hlist_entry) {
5103 if (perf_tp_event_match(event, &data, regs))
5104 perf_swevent_event(event, count, &data, regs);
5107 perf_swevent_put_recursion_context(rctx);
5109 EXPORT_SYMBOL_GPL(perf_tp_event);
5111 static void tp_perf_event_destroy(struct perf_event *event)
5113 perf_trace_destroy(event);
5116 static int perf_tp_event_init(struct perf_event *event)
5120 if (event->attr.type != PERF_TYPE_TRACEPOINT)
5123 err = perf_trace_init(event);
5127 event->destroy = tp_perf_event_destroy;
5132 static struct pmu perf_tracepoint = {
5133 .task_ctx_nr = perf_sw_context,
5135 .event_init = perf_tp_event_init,
5136 .add = perf_trace_add,
5137 .del = perf_trace_del,
5138 .start = perf_swevent_start,
5139 .stop = perf_swevent_stop,
5140 .read = perf_swevent_read,
5142 .event_idx = perf_swevent_event_idx,
5145 static inline void perf_tp_register(void)
5147 perf_pmu_register(&perf_tracepoint, "tracepoint", PERF_TYPE_TRACEPOINT);
5150 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
5155 if (event->attr.type != PERF_TYPE_TRACEPOINT)
5158 filter_str = strndup_user(arg, PAGE_SIZE);
5159 if (IS_ERR(filter_str))
5160 return PTR_ERR(filter_str);
5162 ret = ftrace_profile_set_filter(event, event->attr.config, filter_str);
5168 static void perf_event_free_filter(struct perf_event *event)
5170 ftrace_profile_free_filter(event);
5175 static inline void perf_tp_register(void)
5179 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
5184 static void perf_event_free_filter(struct perf_event *event)
5188 #endif /* CONFIG_EVENT_TRACING */
5190 #ifdef CONFIG_HAVE_HW_BREAKPOINT
5191 void perf_bp_event(struct perf_event *bp, void *data)
5193 struct perf_sample_data sample;
5194 struct pt_regs *regs = data;
5196 perf_sample_data_init(&sample, bp->attr.bp_addr);
5198 if (!bp->hw.state && !perf_exclude_event(bp, regs))
5199 perf_swevent_event(bp, 1, &sample, regs);
5204 * hrtimer based swevent callback
5207 static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer)
5209 enum hrtimer_restart ret = HRTIMER_RESTART;
5210 struct perf_sample_data data;
5211 struct pt_regs *regs;
5212 struct perf_event *event;
5215 event = container_of(hrtimer, struct perf_event, hw.hrtimer);
5217 if (event->state != PERF_EVENT_STATE_ACTIVE)
5218 return HRTIMER_NORESTART;
5220 event->pmu->read(event);
5222 perf_sample_data_init(&data, 0);
5223 data.period = event->hw.last_period;
5224 regs = get_irq_regs();
5226 if (regs && !perf_exclude_event(event, regs)) {
5227 if (!(event->attr.exclude_idle && is_idle_task(current)))
5228 if (perf_event_overflow(event, &data, regs))
5229 ret = HRTIMER_NORESTART;
5232 period = max_t(u64, 10000, event->hw.sample_period);
5233 hrtimer_forward_now(hrtimer, ns_to_ktime(period));
5238 static void perf_swevent_start_hrtimer(struct perf_event *event)
5240 struct hw_perf_event *hwc = &event->hw;
5243 if (!is_sampling_event(event))
5246 period = local64_read(&hwc->period_left);
5251 local64_set(&hwc->period_left, 0);
5253 period = max_t(u64, 10000, hwc->sample_period);
5255 __hrtimer_start_range_ns(&hwc->hrtimer,
5256 ns_to_ktime(period), 0,
5257 HRTIMER_MODE_REL_PINNED, 0);
5260 static void perf_swevent_cancel_hrtimer(struct perf_event *event)
5262 struct hw_perf_event *hwc = &event->hw;
5264 if (is_sampling_event(event)) {
5265 ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer);
5266 local64_set(&hwc->period_left, ktime_to_ns(remaining));
5268 hrtimer_cancel(&hwc->hrtimer);
5272 static void perf_swevent_init_hrtimer(struct perf_event *event)
5274 struct hw_perf_event *hwc = &event->hw;
5276 if (!is_sampling_event(event))
5279 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
5280 hwc->hrtimer.function = perf_swevent_hrtimer;
5283 * Since hrtimers have a fixed rate, we can do a static freq->period
5284 * mapping and avoid the whole period adjust feedback stuff.
5286 if (event->attr.freq) {
5287 long freq = event->attr.sample_freq;
5289 event->attr.sample_period = NSEC_PER_SEC / freq;
5290 hwc->sample_period = event->attr.sample_period;
5291 local64_set(&hwc->period_left, hwc->sample_period);
5292 event->attr.freq = 0;
5297 * Software event: cpu wall time clock
5300 static void cpu_clock_event_update(struct perf_event *event)
5305 now = local_clock();
5306 prev = local64_xchg(&event->hw.prev_count, now);
5307 local64_add(now - prev, &event->count);
5310 static void cpu_clock_event_start(struct perf_event *event, int flags)
5312 local64_set(&event->hw.prev_count, local_clock());
5313 perf_swevent_start_hrtimer(event);
5316 static void cpu_clock_event_stop(struct perf_event *event, int flags)
5318 perf_swevent_cancel_hrtimer(event);
5319 cpu_clock_event_update(event);
5322 static int cpu_clock_event_add(struct perf_event *event, int flags)
5324 if (flags & PERF_EF_START)
5325 cpu_clock_event_start(event, flags);
5330 static void cpu_clock_event_del(struct perf_event *event, int flags)
5332 cpu_clock_event_stop(event, flags);
5335 static void cpu_clock_event_read(struct perf_event *event)
5337 cpu_clock_event_update(event);
5340 static int cpu_clock_event_init(struct perf_event *event)
5342 if (event->attr.type != PERF_TYPE_SOFTWARE)
5345 if (event->attr.config != PERF_COUNT_SW_CPU_CLOCK)
5348 perf_swevent_init_hrtimer(event);
5353 static struct pmu perf_cpu_clock = {
5354 .task_ctx_nr = perf_sw_context,
5356 .event_init = cpu_clock_event_init,
5357 .add = cpu_clock_event_add,
5358 .del = cpu_clock_event_del,
5359 .start = cpu_clock_event_start,
5360 .stop = cpu_clock_event_stop,
5361 .read = cpu_clock_event_read,
5363 .event_idx = perf_swevent_event_idx,
5367 * Software event: task time clock
5370 static void task_clock_event_update(struct perf_event *event, u64 now)
5375 prev = local64_xchg(&event->hw.prev_count, now);
5377 local64_add(delta, &event->count);
5380 static void task_clock_event_start(struct perf_event *event, int flags)
5382 local64_set(&event->hw.prev_count, event->ctx->time);
5383 perf_swevent_start_hrtimer(event);
5386 static void task_clock_event_stop(struct perf_event *event, int flags)
5388 perf_swevent_cancel_hrtimer(event);
5389 task_clock_event_update(event, event->ctx->time);
5392 static int task_clock_event_add(struct perf_event *event, int flags)
5394 if (flags & PERF_EF_START)
5395 task_clock_event_start(event, flags);
5400 static void task_clock_event_del(struct perf_event *event, int flags)
5402 task_clock_event_stop(event, PERF_EF_UPDATE);
5405 static void task_clock_event_read(struct perf_event *event)
5407 u64 now = perf_clock();
5408 u64 delta = now - event->ctx->timestamp;
5409 u64 time = event->ctx->time + delta;
5411 task_clock_event_update(event, time);
5414 static int task_clock_event_init(struct perf_event *event)
5416 if (event->attr.type != PERF_TYPE_SOFTWARE)
5419 if (event->attr.config != PERF_COUNT_SW_TASK_CLOCK)
5422 perf_swevent_init_hrtimer(event);
5427 static struct pmu perf_task_clock = {
5428 .task_ctx_nr = perf_sw_context,
5430 .event_init = task_clock_event_init,
5431 .add = task_clock_event_add,
5432 .del = task_clock_event_del,
5433 .start = task_clock_event_start,
5434 .stop = task_clock_event_stop,
5435 .read = task_clock_event_read,
5437 .event_idx = perf_swevent_event_idx,
5440 static void perf_pmu_nop_void(struct pmu *pmu)
5444 static int perf_pmu_nop_int(struct pmu *pmu)
5449 static void perf_pmu_start_txn(struct pmu *pmu)
5451 perf_pmu_disable(pmu);
5454 static int perf_pmu_commit_txn(struct pmu *pmu)
5456 perf_pmu_enable(pmu);
5460 static void perf_pmu_cancel_txn(struct pmu *pmu)
5462 perf_pmu_enable(pmu);
5465 static int perf_event_idx_default(struct perf_event *event)
5467 return event->hw.idx + 1;
5471 * Ensures all contexts with the same task_ctx_nr have the same
5472 * pmu_cpu_context too.
5474 static void *find_pmu_context(int ctxn)
5481 list_for_each_entry(pmu, &pmus, entry) {
5482 if (pmu->task_ctx_nr == ctxn)
5483 return pmu->pmu_cpu_context;
5489 static void update_pmu_context(struct pmu *pmu, struct pmu *old_pmu)
5493 for_each_possible_cpu(cpu) {
5494 struct perf_cpu_context *cpuctx;
5496 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
5498 if (cpuctx->active_pmu == old_pmu)
5499 cpuctx->active_pmu = pmu;
5503 static void free_pmu_context(struct pmu *pmu)
5507 mutex_lock(&pmus_lock);
5509 * Like a real lame refcount.
5511 list_for_each_entry(i, &pmus, entry) {
5512 if (i->pmu_cpu_context == pmu->pmu_cpu_context) {
5513 update_pmu_context(i, pmu);
5518 free_percpu(pmu->pmu_cpu_context);
5520 mutex_unlock(&pmus_lock);
5522 static struct idr pmu_idr;
5525 type_show(struct device *dev, struct device_attribute *attr, char *page)
5527 struct pmu *pmu = dev_get_drvdata(dev);
5529 return snprintf(page, PAGE_SIZE-1, "%d\n", pmu->type);
5532 static struct device_attribute pmu_dev_attrs[] = {
5537 static int pmu_bus_running;
5538 static struct bus_type pmu_bus = {
5539 .name = "event_source",
5540 .dev_attrs = pmu_dev_attrs,
5543 static void pmu_dev_release(struct device *dev)
5548 static int pmu_dev_alloc(struct pmu *pmu)
5552 pmu->dev = kzalloc(sizeof(struct device), GFP_KERNEL);
5556 pmu->dev->groups = pmu->attr_groups;
5557 device_initialize(pmu->dev);
5558 ret = dev_set_name(pmu->dev, "%s", pmu->name);
5562 dev_set_drvdata(pmu->dev, pmu);
5563 pmu->dev->bus = &pmu_bus;
5564 pmu->dev->release = pmu_dev_release;
5565 ret = device_add(pmu->dev);
5573 put_device(pmu->dev);
5577 static struct lock_class_key cpuctx_mutex;
5578 static struct lock_class_key cpuctx_lock;
5580 int perf_pmu_register(struct pmu *pmu, char *name, int type)
5584 mutex_lock(&pmus_lock);
5586 pmu->pmu_disable_count = alloc_percpu(int);
5587 if (!pmu->pmu_disable_count)
5596 int err = idr_pre_get(&pmu_idr, GFP_KERNEL);
5600 err = idr_get_new_above(&pmu_idr, pmu, PERF_TYPE_MAX, &type);
5608 if (pmu_bus_running) {
5609 ret = pmu_dev_alloc(pmu);
5615 pmu->pmu_cpu_context = find_pmu_context(pmu->task_ctx_nr);
5616 if (pmu->pmu_cpu_context)
5617 goto got_cpu_context;
5619 pmu->pmu_cpu_context = alloc_percpu(struct perf_cpu_context);
5620 if (!pmu->pmu_cpu_context)
5623 for_each_possible_cpu(cpu) {
5624 struct perf_cpu_context *cpuctx;
5626 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
5627 __perf_event_init_context(&cpuctx->ctx);
5628 lockdep_set_class(&cpuctx->ctx.mutex, &cpuctx_mutex);
5629 lockdep_set_class(&cpuctx->ctx.lock, &cpuctx_lock);
5630 cpuctx->ctx.type = cpu_context;
5631 cpuctx->ctx.pmu = pmu;
5632 cpuctx->jiffies_interval = 1;
5633 INIT_LIST_HEAD(&cpuctx->rotation_list);
5634 cpuctx->active_pmu = pmu;
5638 if (!pmu->start_txn) {
5639 if (pmu->pmu_enable) {
5641 * If we have pmu_enable/pmu_disable calls, install
5642 * transaction stubs that use that to try and batch
5643 * hardware accesses.
5645 pmu->start_txn = perf_pmu_start_txn;
5646 pmu->commit_txn = perf_pmu_commit_txn;
5647 pmu->cancel_txn = perf_pmu_cancel_txn;
5649 pmu->start_txn = perf_pmu_nop_void;
5650 pmu->commit_txn = perf_pmu_nop_int;
5651 pmu->cancel_txn = perf_pmu_nop_void;
5655 if (!pmu->pmu_enable) {
5656 pmu->pmu_enable = perf_pmu_nop_void;
5657 pmu->pmu_disable = perf_pmu_nop_void;
5660 if (!pmu->event_idx)
5661 pmu->event_idx = perf_event_idx_default;
5663 list_add_rcu(&pmu->entry, &pmus);
5666 mutex_unlock(&pmus_lock);
5671 device_del(pmu->dev);
5672 put_device(pmu->dev);
5675 if (pmu->type >= PERF_TYPE_MAX)
5676 idr_remove(&pmu_idr, pmu->type);
5679 free_percpu(pmu->pmu_disable_count);
5683 void perf_pmu_unregister(struct pmu *pmu)
5685 mutex_lock(&pmus_lock);
5686 list_del_rcu(&pmu->entry);
5687 mutex_unlock(&pmus_lock);
5690 * We dereference the pmu list under both SRCU and regular RCU, so
5691 * synchronize against both of those.
5693 synchronize_srcu(&pmus_srcu);
5696 free_percpu(pmu->pmu_disable_count);
5697 if (pmu->type >= PERF_TYPE_MAX)
5698 idr_remove(&pmu_idr, pmu->type);
5699 device_del(pmu->dev);
5700 put_device(pmu->dev);
5701 free_pmu_context(pmu);
5704 struct pmu *perf_init_event(struct perf_event *event)
5706 struct pmu *pmu = NULL;
5710 idx = srcu_read_lock(&pmus_srcu);
5713 pmu = idr_find(&pmu_idr, event->attr.type);
5717 ret = pmu->event_init(event);
5723 list_for_each_entry_rcu(pmu, &pmus, entry) {
5725 ret = pmu->event_init(event);
5729 if (ret != -ENOENT) {
5734 pmu = ERR_PTR(-ENOENT);
5736 srcu_read_unlock(&pmus_srcu, idx);
5742 * Allocate and initialize a event structure
5744 static struct perf_event *
5745 perf_event_alloc(struct perf_event_attr *attr, int cpu,
5746 struct task_struct *task,
5747 struct perf_event *group_leader,
5748 struct perf_event *parent_event,
5749 perf_overflow_handler_t overflow_handler,
5753 struct perf_event *event;
5754 struct hw_perf_event *hwc;
5757 if ((unsigned)cpu >= nr_cpu_ids) {
5758 if (!task || cpu != -1)
5759 return ERR_PTR(-EINVAL);
5762 event = kzalloc(sizeof(*event), GFP_KERNEL);
5764 return ERR_PTR(-ENOMEM);
5767 * Single events are their own group leaders, with an
5768 * empty sibling list:
5771 group_leader = event;
5773 mutex_init(&event->child_mutex);
5774 INIT_LIST_HEAD(&event->child_list);
5776 INIT_LIST_HEAD(&event->group_entry);
5777 INIT_LIST_HEAD(&event->event_entry);
5778 INIT_LIST_HEAD(&event->sibling_list);
5779 INIT_LIST_HEAD(&event->rb_entry);
5781 init_waitqueue_head(&event->waitq);
5782 init_irq_work(&event->pending, perf_pending_event);
5784 mutex_init(&event->mmap_mutex);
5787 event->attr = *attr;
5788 event->group_leader = group_leader;
5792 event->parent = parent_event;
5794 event->ns = get_pid_ns(current->nsproxy->pid_ns);
5795 event->id = atomic64_inc_return(&perf_event_id);
5797 event->state = PERF_EVENT_STATE_INACTIVE;
5800 event->attach_state = PERF_ATTACH_TASK;
5801 #ifdef CONFIG_HAVE_HW_BREAKPOINT
5803 * hw_breakpoint is a bit difficult here..
5805 if (attr->type == PERF_TYPE_BREAKPOINT)
5806 event->hw.bp_target = task;
5810 if (!overflow_handler && parent_event) {
5811 overflow_handler = parent_event->overflow_handler;
5812 context = parent_event->overflow_handler_context;
5815 event->overflow_handler = overflow_handler;
5816 event->overflow_handler_context = context;
5819 event->state = PERF_EVENT_STATE_OFF;
5824 hwc->sample_period = attr->sample_period;
5825 if (attr->freq && attr->sample_freq)
5826 hwc->sample_period = 1;
5827 hwc->last_period = hwc->sample_period;
5829 local64_set(&hwc->period_left, hwc->sample_period);
5832 * we currently do not support PERF_FORMAT_GROUP on inherited events
5834 if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP))
5837 pmu = perf_init_event(event);
5843 else if (IS_ERR(pmu))
5848 put_pid_ns(event->ns);
5850 return ERR_PTR(err);
5853 if (!event->parent) {
5854 if (event->attach_state & PERF_ATTACH_TASK)
5855 static_key_slow_inc(&perf_sched_events.key);
5856 if (event->attr.mmap || event->attr.mmap_data)
5857 atomic_inc(&nr_mmap_events);
5858 if (event->attr.comm)
5859 atomic_inc(&nr_comm_events);
5860 if (event->attr.task)
5861 atomic_inc(&nr_task_events);
5862 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) {
5863 err = get_callchain_buffers();
5866 return ERR_PTR(err);
5874 static int perf_copy_attr(struct perf_event_attr __user *uattr,
5875 struct perf_event_attr *attr)
5880 if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0))
5884 * zero the full structure, so that a short copy will be nice.
5886 memset(attr, 0, sizeof(*attr));
5888 ret = get_user(size, &uattr->size);
5892 if (size > PAGE_SIZE) /* silly large */
5895 if (!size) /* abi compat */
5896 size = PERF_ATTR_SIZE_VER0;
5898 if (size < PERF_ATTR_SIZE_VER0)
5902 * If we're handed a bigger struct than we know of,
5903 * ensure all the unknown bits are 0 - i.e. new
5904 * user-space does not rely on any kernel feature
5905 * extensions we dont know about yet.
5907 if (size > sizeof(*attr)) {
5908 unsigned char __user *addr;
5909 unsigned char __user *end;
5912 addr = (void __user *)uattr + sizeof(*attr);
5913 end = (void __user *)uattr + size;
5915 for (; addr < end; addr++) {
5916 ret = get_user(val, addr);
5922 size = sizeof(*attr);
5925 ret = copy_from_user(attr, uattr, size);
5929 if (attr->__reserved_1)
5932 if (attr->sample_type & ~(PERF_SAMPLE_MAX-1))
5935 if (attr->read_format & ~(PERF_FORMAT_MAX-1))
5942 put_user(sizeof(*attr), &uattr->size);
5948 perf_event_set_output(struct perf_event *event, struct perf_event *output_event)
5950 struct ring_buffer *rb = NULL, *old_rb = NULL;
5956 /* don't allow circular references */
5957 if (event == output_event)
5961 * Don't allow cross-cpu buffers
5963 if (output_event->cpu != event->cpu)
5967 * If its not a per-cpu rb, it must be the same task.
5969 if (output_event->cpu == -1 && output_event->ctx != event->ctx)
5973 mutex_lock(&event->mmap_mutex);
5974 /* Can't redirect output if we've got an active mmap() */
5975 if (atomic_read(&event->mmap_count))
5979 /* get the rb we want to redirect to */
5980 rb = ring_buffer_get(output_event);
5986 rcu_assign_pointer(event->rb, rb);
5988 ring_buffer_detach(event, old_rb);
5991 mutex_unlock(&event->mmap_mutex);
5994 ring_buffer_put(old_rb);
6000 * sys_perf_event_open - open a performance event, associate it to a task/cpu
6002 * @attr_uptr: event_id type attributes for monitoring/sampling
6005 * @group_fd: group leader event fd
6007 SYSCALL_DEFINE5(perf_event_open,
6008 struct perf_event_attr __user *, attr_uptr,
6009 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
6011 struct perf_event *group_leader = NULL, *output_event = NULL;
6012 struct perf_event *event, *sibling;
6013 struct perf_event_attr attr;
6014 struct perf_event_context *ctx;
6015 struct file *event_file = NULL;
6016 struct file *group_file = NULL;
6017 struct task_struct *task = NULL;
6021 int fput_needed = 0;
6024 /* for future expandability... */
6025 if (flags & ~PERF_FLAG_ALL)
6028 err = perf_copy_attr(attr_uptr, &attr);
6032 if (!attr.exclude_kernel) {
6033 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
6038 if (attr.sample_freq > sysctl_perf_event_sample_rate)
6043 * In cgroup mode, the pid argument is used to pass the fd
6044 * opened to the cgroup directory in cgroupfs. The cpu argument
6045 * designates the cpu on which to monitor threads from that
6048 if ((flags & PERF_FLAG_PID_CGROUP) && (pid == -1 || cpu == -1))
6051 event_fd = get_unused_fd_flags(O_RDWR);
6055 if (group_fd != -1) {
6056 group_leader = perf_fget_light(group_fd, &fput_needed);
6057 if (IS_ERR(group_leader)) {
6058 err = PTR_ERR(group_leader);
6061 group_file = group_leader->filp;
6062 if (flags & PERF_FLAG_FD_OUTPUT)
6063 output_event = group_leader;
6064 if (flags & PERF_FLAG_FD_NO_GROUP)
6065 group_leader = NULL;
6068 if (pid != -1 && !(flags & PERF_FLAG_PID_CGROUP)) {
6069 task = find_lively_task_by_vpid(pid);
6071 err = PTR_ERR(task);
6076 event = perf_event_alloc(&attr, cpu, task, group_leader, NULL,
6078 if (IS_ERR(event)) {
6079 err = PTR_ERR(event);
6083 if (flags & PERF_FLAG_PID_CGROUP) {
6084 err = perf_cgroup_connect(pid, event, &attr, group_leader);
6089 * - that has cgroup constraint on event->cpu
6090 * - that may need work on context switch
6092 atomic_inc(&per_cpu(perf_cgroup_events, event->cpu));
6093 static_key_slow_inc(&perf_sched_events.key);
6097 * Special case software events and allow them to be part of
6098 * any hardware group.
6103 (is_software_event(event) != is_software_event(group_leader))) {
6104 if (is_software_event(event)) {
6106 * If event and group_leader are not both a software
6107 * event, and event is, then group leader is not.
6109 * Allow the addition of software events to !software
6110 * groups, this is safe because software events never
6113 pmu = group_leader->pmu;
6114 } else if (is_software_event(group_leader) &&
6115 (group_leader->group_flags & PERF_GROUP_SOFTWARE)) {
6117 * In case the group is a pure software group, and we
6118 * try to add a hardware event, move the whole group to
6119 * the hardware context.
6126 * Get the target context (task or percpu):
6128 ctx = find_get_context(pmu, task, cpu);
6135 put_task_struct(task);
6140 * Look up the group leader (we will attach this event to it):
6146 * Do not allow a recursive hierarchy (this new sibling
6147 * becoming part of another group-sibling):
6149 if (group_leader->group_leader != group_leader)
6152 * Do not allow to attach to a group in a different
6153 * task or CPU context:
6156 if (group_leader->ctx->type != ctx->type)
6159 if (group_leader->ctx != ctx)
6164 * Only a group leader can be exclusive or pinned
6166 if (attr.exclusive || attr.pinned)
6171 err = perf_event_set_output(event, output_event);
6176 event_file = anon_inode_getfile("[perf_event]", &perf_fops, event, O_RDWR);
6177 if (IS_ERR(event_file)) {
6178 err = PTR_ERR(event_file);
6183 struct perf_event_context *gctx = group_leader->ctx;
6185 mutex_lock(&gctx->mutex);
6186 perf_remove_from_context(group_leader);
6187 list_for_each_entry(sibling, &group_leader->sibling_list,
6189 perf_remove_from_context(sibling);
6192 mutex_unlock(&gctx->mutex);
6196 event->filp = event_file;
6197 WARN_ON_ONCE(ctx->parent_ctx);
6198 mutex_lock(&ctx->mutex);
6201 perf_install_in_context(ctx, group_leader, cpu);
6203 list_for_each_entry(sibling, &group_leader->sibling_list,
6205 perf_install_in_context(ctx, sibling, cpu);
6210 perf_install_in_context(ctx, event, cpu);
6212 perf_unpin_context(ctx);
6213 mutex_unlock(&ctx->mutex);
6215 event->owner = current;
6217 mutex_lock(¤t->perf_event_mutex);
6218 list_add_tail(&event->owner_entry, ¤t->perf_event_list);
6219 mutex_unlock(¤t->perf_event_mutex);
6222 * Precalculate sample_data sizes
6224 perf_event__header_size(event);
6225 perf_event__id_header_size(event);
6228 * Drop the reference on the group_event after placing the
6229 * new event on the sibling_list. This ensures destruction
6230 * of the group leader will find the pointer to itself in
6231 * perf_group_detach().
6233 fput_light(group_file, fput_needed);
6234 fd_install(event_fd, event_file);
6238 perf_unpin_context(ctx);
6244 put_task_struct(task);
6246 fput_light(group_file, fput_needed);
6248 put_unused_fd(event_fd);
6253 * perf_event_create_kernel_counter
6255 * @attr: attributes of the counter to create
6256 * @cpu: cpu in which the counter is bound
6257 * @task: task to profile (NULL for percpu)
6260 perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu,
6261 struct task_struct *task,
6262 perf_overflow_handler_t overflow_handler,
6265 struct perf_event_context *ctx;
6266 struct perf_event *event;
6270 * Get the target context (task or percpu):
6273 event = perf_event_alloc(attr, cpu, task, NULL, NULL,
6274 overflow_handler, context);
6275 if (IS_ERR(event)) {
6276 err = PTR_ERR(event);
6280 ctx = find_get_context(event->pmu, task, cpu);
6287 WARN_ON_ONCE(ctx->parent_ctx);
6288 mutex_lock(&ctx->mutex);
6289 perf_install_in_context(ctx, event, cpu);
6291 perf_unpin_context(ctx);
6292 mutex_unlock(&ctx->mutex);
6299 return ERR_PTR(err);
6301 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter);
6303 static void sync_child_event(struct perf_event *child_event,
6304 struct task_struct *child)
6306 struct perf_event *parent_event = child_event->parent;
6309 if (child_event->attr.inherit_stat)
6310 perf_event_read_event(child_event, child);
6312 child_val = perf_event_count(child_event);
6315 * Add back the child's count to the parent's count:
6317 atomic64_add(child_val, &parent_event->child_count);
6318 atomic64_add(child_event->total_time_enabled,
6319 &parent_event->child_total_time_enabled);
6320 atomic64_add(child_event->total_time_running,
6321 &parent_event->child_total_time_running);
6324 * Remove this event from the parent's list
6326 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
6327 mutex_lock(&parent_event->child_mutex);
6328 list_del_init(&child_event->child_list);
6329 mutex_unlock(&parent_event->child_mutex);
6332 * Release the parent event, if this was the last
6335 fput(parent_event->filp);
6339 __perf_event_exit_task(struct perf_event *child_event,
6340 struct perf_event_context *child_ctx,
6341 struct task_struct *child)
6343 if (child_event->parent) {
6344 raw_spin_lock_irq(&child_ctx->lock);
6345 perf_group_detach(child_event);
6346 raw_spin_unlock_irq(&child_ctx->lock);
6349 perf_remove_from_context(child_event);
6352 * It can happen that the parent exits first, and has events
6353 * that are still around due to the child reference. These
6354 * events need to be zapped.
6356 if (child_event->parent) {
6357 sync_child_event(child_event, child);
6358 free_event(child_event);
6362 static void perf_event_exit_task_context(struct task_struct *child, int ctxn)
6364 struct perf_event *child_event, *tmp;
6365 struct perf_event_context *child_ctx;
6366 unsigned long flags;
6368 if (likely(!child->perf_event_ctxp[ctxn])) {
6369 perf_event_task(child, NULL, 0);
6373 local_irq_save(flags);
6375 * We can't reschedule here because interrupts are disabled,
6376 * and either child is current or it is a task that can't be
6377 * scheduled, so we are now safe from rescheduling changing
6380 child_ctx = rcu_dereference_raw(child->perf_event_ctxp[ctxn]);
6383 * Take the context lock here so that if find_get_context is
6384 * reading child->perf_event_ctxp, we wait until it has
6385 * incremented the context's refcount before we do put_ctx below.
6387 raw_spin_lock(&child_ctx->lock);
6388 task_ctx_sched_out(child_ctx);
6389 child->perf_event_ctxp[ctxn] = NULL;
6391 * If this context is a clone; unclone it so it can't get
6392 * swapped to another process while we're removing all
6393 * the events from it.
6395 unclone_ctx(child_ctx);
6396 update_context_time(child_ctx);
6397 raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
6400 * Report the task dead after unscheduling the events so that we
6401 * won't get any samples after PERF_RECORD_EXIT. We can however still
6402 * get a few PERF_RECORD_READ events.
6404 perf_event_task(child, child_ctx, 0);
6407 * We can recurse on the same lock type through:
6409 * __perf_event_exit_task()
6410 * sync_child_event()
6411 * fput(parent_event->filp)
6413 * mutex_lock(&ctx->mutex)
6415 * But since its the parent context it won't be the same instance.
6417 mutex_lock(&child_ctx->mutex);
6420 list_for_each_entry_safe(child_event, tmp, &child_ctx->pinned_groups,
6422 __perf_event_exit_task(child_event, child_ctx, child);
6424 list_for_each_entry_safe(child_event, tmp, &child_ctx->flexible_groups,
6426 __perf_event_exit_task(child_event, child_ctx, child);
6429 * If the last event was a group event, it will have appended all
6430 * its siblings to the list, but we obtained 'tmp' before that which
6431 * will still point to the list head terminating the iteration.
6433 if (!list_empty(&child_ctx->pinned_groups) ||
6434 !list_empty(&child_ctx->flexible_groups))
6437 mutex_unlock(&child_ctx->mutex);
6443 * When a child task exits, feed back event values to parent events.
6445 void perf_event_exit_task(struct task_struct *child)
6447 struct perf_event *event, *tmp;
6450 mutex_lock(&child->perf_event_mutex);
6451 list_for_each_entry_safe(event, tmp, &child->perf_event_list,
6453 list_del_init(&event->owner_entry);
6456 * Ensure the list deletion is visible before we clear
6457 * the owner, closes a race against perf_release() where
6458 * we need to serialize on the owner->perf_event_mutex.
6461 event->owner = NULL;
6463 mutex_unlock(&child->perf_event_mutex);
6465 for_each_task_context_nr(ctxn)
6466 perf_event_exit_task_context(child, ctxn);
6469 static void perf_free_event(struct perf_event *event,
6470 struct perf_event_context *ctx)
6472 struct perf_event *parent = event->parent;
6474 if (WARN_ON_ONCE(!parent))
6477 mutex_lock(&parent->child_mutex);
6478 list_del_init(&event->child_list);
6479 mutex_unlock(&parent->child_mutex);
6483 perf_group_detach(event);
6484 list_del_event(event, ctx);
6489 * free an unexposed, unused context as created by inheritance by
6490 * perf_event_init_task below, used by fork() in case of fail.
6492 void perf_event_free_task(struct task_struct *task)
6494 struct perf_event_context *ctx;
6495 struct perf_event *event, *tmp;
6498 for_each_task_context_nr(ctxn) {
6499 ctx = task->perf_event_ctxp[ctxn];
6503 mutex_lock(&ctx->mutex);
6505 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups,
6507 perf_free_event(event, ctx);
6509 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups,
6511 perf_free_event(event, ctx);
6513 if (!list_empty(&ctx->pinned_groups) ||
6514 !list_empty(&ctx->flexible_groups))
6517 mutex_unlock(&ctx->mutex);
6523 void perf_event_delayed_put(struct task_struct *task)
6527 for_each_task_context_nr(ctxn)
6528 WARN_ON_ONCE(task->perf_event_ctxp[ctxn]);
6532 * inherit a event from parent task to child task:
6534 static struct perf_event *
6535 inherit_event(struct perf_event *parent_event,
6536 struct task_struct *parent,
6537 struct perf_event_context *parent_ctx,
6538 struct task_struct *child,
6539 struct perf_event *group_leader,
6540 struct perf_event_context *child_ctx)
6542 struct perf_event *child_event;
6543 unsigned long flags;
6546 * Instead of creating recursive hierarchies of events,
6547 * we link inherited events back to the original parent,
6548 * which has a filp for sure, which we use as the reference
6551 if (parent_event->parent)
6552 parent_event = parent_event->parent;
6554 child_event = perf_event_alloc(&parent_event->attr,
6557 group_leader, parent_event,
6559 if (IS_ERR(child_event))
6564 * Make the child state follow the state of the parent event,
6565 * not its attr.disabled bit. We hold the parent's mutex,
6566 * so we won't race with perf_event_{en, dis}able_family.
6568 if (parent_event->state >= PERF_EVENT_STATE_INACTIVE)
6569 child_event->state = PERF_EVENT_STATE_INACTIVE;
6571 child_event->state = PERF_EVENT_STATE_OFF;
6573 if (parent_event->attr.freq) {
6574 u64 sample_period = parent_event->hw.sample_period;
6575 struct hw_perf_event *hwc = &child_event->hw;
6577 hwc->sample_period = sample_period;
6578 hwc->last_period = sample_period;
6580 local64_set(&hwc->period_left, sample_period);
6583 child_event->ctx = child_ctx;
6584 child_event->overflow_handler = parent_event->overflow_handler;
6585 child_event->overflow_handler_context
6586 = parent_event->overflow_handler_context;
6589 * Precalculate sample_data sizes
6591 perf_event__header_size(child_event);
6592 perf_event__id_header_size(child_event);
6595 * Link it up in the child's context:
6597 raw_spin_lock_irqsave(&child_ctx->lock, flags);
6598 add_event_to_ctx(child_event, child_ctx);
6599 raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
6602 * Get a reference to the parent filp - we will fput it
6603 * when the child event exits. This is safe to do because
6604 * we are in the parent and we know that the filp still
6605 * exists and has a nonzero count:
6607 atomic_long_inc(&parent_event->filp->f_count);
6610 * Link this into the parent event's child list
6612 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
6613 mutex_lock(&parent_event->child_mutex);
6614 list_add_tail(&child_event->child_list, &parent_event->child_list);
6615 mutex_unlock(&parent_event->child_mutex);
6620 static int inherit_group(struct perf_event *parent_event,
6621 struct task_struct *parent,
6622 struct perf_event_context *parent_ctx,
6623 struct task_struct *child,
6624 struct perf_event_context *child_ctx)
6626 struct perf_event *leader;
6627 struct perf_event *sub;
6628 struct perf_event *child_ctr;
6630 leader = inherit_event(parent_event, parent, parent_ctx,
6631 child, NULL, child_ctx);
6633 return PTR_ERR(leader);
6634 list_for_each_entry(sub, &parent_event->sibling_list, group_entry) {
6635 child_ctr = inherit_event(sub, parent, parent_ctx,
6636 child, leader, child_ctx);
6637 if (IS_ERR(child_ctr))
6638 return PTR_ERR(child_ctr);
6644 inherit_task_group(struct perf_event *event, struct task_struct *parent,
6645 struct perf_event_context *parent_ctx,
6646 struct task_struct *child, int ctxn,
6650 struct perf_event_context *child_ctx;
6652 if (!event->attr.inherit) {
6657 child_ctx = child->perf_event_ctxp[ctxn];
6660 * This is executed from the parent task context, so
6661 * inherit events that have been marked for cloning.
6662 * First allocate and initialize a context for the
6666 child_ctx = alloc_perf_context(event->pmu, child);
6670 child->perf_event_ctxp[ctxn] = child_ctx;
6673 ret = inherit_group(event, parent, parent_ctx,
6683 * Initialize the perf_event context in task_struct
6685 int perf_event_init_context(struct task_struct *child, int ctxn)
6687 struct perf_event_context *child_ctx, *parent_ctx;
6688 struct perf_event_context *cloned_ctx;
6689 struct perf_event *event;
6690 struct task_struct *parent = current;
6691 int inherited_all = 1;
6692 unsigned long flags;
6695 if (likely(!parent->perf_event_ctxp[ctxn]))
6699 * If the parent's context is a clone, pin it so it won't get
6702 parent_ctx = perf_pin_task_context(parent, ctxn);
6705 * No need to check if parent_ctx != NULL here; since we saw
6706 * it non-NULL earlier, the only reason for it to become NULL
6707 * is if we exit, and since we're currently in the middle of
6708 * a fork we can't be exiting at the same time.
6712 * Lock the parent list. No need to lock the child - not PID
6713 * hashed yet and not running, so nobody can access it.
6715 mutex_lock(&parent_ctx->mutex);
6718 * We dont have to disable NMIs - we are only looking at
6719 * the list, not manipulating it:
6721 list_for_each_entry(event, &parent_ctx->pinned_groups, group_entry) {
6722 ret = inherit_task_group(event, parent, parent_ctx,
6723 child, ctxn, &inherited_all);
6729 * We can't hold ctx->lock when iterating the ->flexible_group list due
6730 * to allocations, but we need to prevent rotation because
6731 * rotate_ctx() will change the list from interrupt context.
6733 raw_spin_lock_irqsave(&parent_ctx->lock, flags);
6734 parent_ctx->rotate_disable = 1;
6735 raw_spin_unlock_irqrestore(&parent_ctx->lock, flags);
6737 list_for_each_entry(event, &parent_ctx->flexible_groups, group_entry) {
6738 ret = inherit_task_group(event, parent, parent_ctx,
6739 child, ctxn, &inherited_all);
6744 raw_spin_lock_irqsave(&parent_ctx->lock, flags);
6745 parent_ctx->rotate_disable = 0;
6747 child_ctx = child->perf_event_ctxp[ctxn];
6749 if (child_ctx && inherited_all) {
6751 * Mark the child context as a clone of the parent
6752 * context, or of whatever the parent is a clone of.
6754 * Note that if the parent is a clone, the holding of
6755 * parent_ctx->lock avoids it from being uncloned.
6757 cloned_ctx = parent_ctx->parent_ctx;
6759 child_ctx->parent_ctx = cloned_ctx;
6760 child_ctx->parent_gen = parent_ctx->parent_gen;
6762 child_ctx->parent_ctx = parent_ctx;
6763 child_ctx->parent_gen = parent_ctx->generation;
6765 get_ctx(child_ctx->parent_ctx);
6768 raw_spin_unlock_irqrestore(&parent_ctx->lock, flags);
6769 mutex_unlock(&parent_ctx->mutex);
6771 perf_unpin_context(parent_ctx);
6772 put_ctx(parent_ctx);
6778 * Initialize the perf_event context in task_struct
6780 int perf_event_init_task(struct task_struct *child)
6784 memset(child->perf_event_ctxp, 0, sizeof(child->perf_event_ctxp));
6785 mutex_init(&child->perf_event_mutex);
6786 INIT_LIST_HEAD(&child->perf_event_list);
6788 for_each_task_context_nr(ctxn) {
6789 ret = perf_event_init_context(child, ctxn);
6797 static void __init perf_event_init_all_cpus(void)
6799 struct swevent_htable *swhash;
6802 for_each_possible_cpu(cpu) {
6803 swhash = &per_cpu(swevent_htable, cpu);
6804 mutex_init(&swhash->hlist_mutex);
6805 INIT_LIST_HEAD(&per_cpu(rotation_list, cpu));
6809 static void __cpuinit perf_event_init_cpu(int cpu)
6811 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
6813 mutex_lock(&swhash->hlist_mutex);
6814 if (swhash->hlist_refcount > 0) {
6815 struct swevent_hlist *hlist;
6817 hlist = kzalloc_node(sizeof(*hlist), GFP_KERNEL, cpu_to_node(cpu));
6819 rcu_assign_pointer(swhash->swevent_hlist, hlist);
6821 mutex_unlock(&swhash->hlist_mutex);
6824 #if defined CONFIG_HOTPLUG_CPU || defined CONFIG_KEXEC
6825 static void perf_pmu_rotate_stop(struct pmu *pmu)
6827 struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
6829 WARN_ON(!irqs_disabled());
6831 list_del_init(&cpuctx->rotation_list);
6834 static void __perf_event_exit_context(void *__info)
6836 struct perf_event_context *ctx = __info;
6837 struct perf_event *event, *tmp;
6839 perf_pmu_rotate_stop(ctx->pmu);
6841 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry)
6842 __perf_remove_from_context(event);
6843 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups, group_entry)
6844 __perf_remove_from_context(event);
6847 static void perf_event_exit_cpu_context(int cpu)
6849 struct perf_event_context *ctx;
6853 idx = srcu_read_lock(&pmus_srcu);
6854 list_for_each_entry_rcu(pmu, &pmus, entry) {
6855 ctx = &per_cpu_ptr(pmu->pmu_cpu_context, cpu)->ctx;
6857 mutex_lock(&ctx->mutex);
6858 smp_call_function_single(cpu, __perf_event_exit_context, ctx, 1);
6859 mutex_unlock(&ctx->mutex);
6861 srcu_read_unlock(&pmus_srcu, idx);
6864 static void perf_event_exit_cpu(int cpu)
6866 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
6868 mutex_lock(&swhash->hlist_mutex);
6869 swevent_hlist_release(swhash);
6870 mutex_unlock(&swhash->hlist_mutex);
6872 perf_event_exit_cpu_context(cpu);
6875 static inline void perf_event_exit_cpu(int cpu) { }
6879 perf_reboot(struct notifier_block *notifier, unsigned long val, void *v)
6883 for_each_online_cpu(cpu)
6884 perf_event_exit_cpu(cpu);
6890 * Run the perf reboot notifier at the very last possible moment so that
6891 * the generic watchdog code runs as long as possible.
6893 static struct notifier_block perf_reboot_notifier = {
6894 .notifier_call = perf_reboot,
6895 .priority = INT_MIN,
6898 static int __cpuinit
6899 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
6901 unsigned int cpu = (long)hcpu;
6903 switch (action & ~CPU_TASKS_FROZEN) {
6905 case CPU_UP_PREPARE:
6906 case CPU_DOWN_FAILED:
6907 perf_event_init_cpu(cpu);
6910 case CPU_UP_CANCELED:
6911 case CPU_DOWN_PREPARE:
6912 perf_event_exit_cpu(cpu);
6922 void __init perf_event_init(void)
6928 perf_event_init_all_cpus();
6929 init_srcu_struct(&pmus_srcu);
6930 perf_pmu_register(&perf_swevent, "software", PERF_TYPE_SOFTWARE);
6931 perf_pmu_register(&perf_cpu_clock, NULL, -1);
6932 perf_pmu_register(&perf_task_clock, NULL, -1);
6934 perf_cpu_notifier(perf_cpu_notify);
6935 register_reboot_notifier(&perf_reboot_notifier);
6937 ret = init_hw_breakpoint();
6938 WARN(ret, "hw_breakpoint initialization failed with: %d", ret);
6940 /* do not patch jump label more than once per second */
6941 jump_label_rate_limit(&perf_sched_events, HZ);
6944 static int __init perf_event_sysfs_init(void)
6949 mutex_lock(&pmus_lock);
6951 ret = bus_register(&pmu_bus);
6955 list_for_each_entry(pmu, &pmus, entry) {
6956 if (!pmu->name || pmu->type < 0)
6959 ret = pmu_dev_alloc(pmu);
6960 WARN(ret, "Failed to register pmu: %s, reason %d\n", pmu->name, ret);
6962 pmu_bus_running = 1;
6966 mutex_unlock(&pmus_lock);
6970 device_initcall(perf_event_sysfs_init);
6972 #ifdef CONFIG_CGROUP_PERF
6973 static struct cgroup_subsys_state *perf_cgroup_create(
6974 struct cgroup_subsys *ss, struct cgroup *cont)
6976 struct perf_cgroup *jc;
6978 jc = kzalloc(sizeof(*jc), GFP_KERNEL);
6980 return ERR_PTR(-ENOMEM);
6982 jc->info = alloc_percpu(struct perf_cgroup_info);
6985 return ERR_PTR(-ENOMEM);
6991 static void perf_cgroup_destroy(struct cgroup_subsys *ss,
6992 struct cgroup *cont)
6994 struct perf_cgroup *jc;
6995 jc = container_of(cgroup_subsys_state(cont, perf_subsys_id),
6996 struct perf_cgroup, css);
6997 free_percpu(jc->info);
7001 static int __perf_cgroup_move(void *info)
7003 struct task_struct *task = info;
7004 perf_cgroup_switch(task, PERF_CGROUP_SWOUT | PERF_CGROUP_SWIN);
7008 static void perf_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
7009 struct cgroup_taskset *tset)
7011 struct task_struct *task;
7013 cgroup_taskset_for_each(task, cgrp, tset)
7014 task_function_call(task, __perf_cgroup_move, task);
7017 static void perf_cgroup_exit(struct cgroup_subsys *ss, struct cgroup *cgrp,
7018 struct cgroup *old_cgrp, struct task_struct *task)
7021 * cgroup_exit() is called in the copy_process() failure path.
7022 * Ignore this case since the task hasn't ran yet, this avoids
7023 * trying to poke a half freed task state from generic code.
7025 if (!(task->flags & PF_EXITING))
7028 task_function_call(task, __perf_cgroup_move, task);
7031 struct cgroup_subsys perf_subsys = {
7032 .name = "perf_event",
7033 .subsys_id = perf_subsys_id,
7034 .create = perf_cgroup_create,
7035 .destroy = perf_cgroup_destroy,
7036 .exit = perf_cgroup_exit,
7037 .attach = perf_cgroup_attach,
7039 #endif /* CONFIG_CGROUP_PERF */