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 jump_label_key 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_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.exclusive || !cpuctx->active_oncpu)
1134 cpuctx->exclusive = 0;
1138 group_sched_out(struct perf_event *group_event,
1139 struct perf_cpu_context *cpuctx,
1140 struct perf_event_context *ctx)
1142 struct perf_event *event;
1143 int state = group_event->state;
1145 event_sched_out(group_event, cpuctx, ctx);
1148 * Schedule out siblings (if any):
1150 list_for_each_entry(event, &group_event->sibling_list, group_entry)
1151 event_sched_out(event, cpuctx, ctx);
1153 if (state == PERF_EVENT_STATE_ACTIVE && group_event->attr.exclusive)
1154 cpuctx->exclusive = 0;
1158 * Cross CPU call to remove a performance event
1160 * We disable the event on the hardware level first. After that we
1161 * remove it from the context list.
1163 static int __perf_remove_from_context(void *info)
1165 struct perf_event *event = info;
1166 struct perf_event_context *ctx = event->ctx;
1167 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1169 raw_spin_lock(&ctx->lock);
1170 event_sched_out(event, cpuctx, ctx);
1171 list_del_event(event, ctx);
1172 if (!ctx->nr_events && cpuctx->task_ctx == ctx) {
1174 cpuctx->task_ctx = NULL;
1176 raw_spin_unlock(&ctx->lock);
1183 * Remove the event from a task's (or a CPU's) list of events.
1185 * CPU events are removed with a smp call. For task events we only
1186 * call when the task is on a CPU.
1188 * If event->ctx is a cloned context, callers must make sure that
1189 * every task struct that event->ctx->task could possibly point to
1190 * remains valid. This is OK when called from perf_release since
1191 * that only calls us on the top-level context, which can't be a clone.
1192 * When called from perf_event_exit_task, it's OK because the
1193 * context has been detached from its task.
1195 static void perf_remove_from_context(struct perf_event *event)
1197 struct perf_event_context *ctx = event->ctx;
1198 struct task_struct *task = ctx->task;
1200 lockdep_assert_held(&ctx->mutex);
1204 * Per cpu events are removed via an smp call and
1205 * the removal is always successful.
1207 cpu_function_call(event->cpu, __perf_remove_from_context, event);
1212 if (!task_function_call(task, __perf_remove_from_context, event))
1215 raw_spin_lock_irq(&ctx->lock);
1217 * If we failed to find a running task, but find the context active now
1218 * that we've acquired the ctx->lock, retry.
1220 if (ctx->is_active) {
1221 raw_spin_unlock_irq(&ctx->lock);
1226 * Since the task isn't running, its safe to remove the event, us
1227 * holding the ctx->lock ensures the task won't get scheduled in.
1229 list_del_event(event, ctx);
1230 raw_spin_unlock_irq(&ctx->lock);
1234 * Cross CPU call to disable a performance event
1236 static int __perf_event_disable(void *info)
1238 struct perf_event *event = info;
1239 struct perf_event_context *ctx = event->ctx;
1240 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1243 * If this is a per-task event, need to check whether this
1244 * event's task is the current task on this cpu.
1246 * Can trigger due to concurrent perf_event_context_sched_out()
1247 * flipping contexts around.
1249 if (ctx->task && cpuctx->task_ctx != ctx)
1252 raw_spin_lock(&ctx->lock);
1255 * If the event is on, turn it off.
1256 * If it is in error state, leave it in error state.
1258 if (event->state >= PERF_EVENT_STATE_INACTIVE) {
1259 update_context_time(ctx);
1260 update_cgrp_time_from_event(event);
1261 update_group_times(event);
1262 if (event == event->group_leader)
1263 group_sched_out(event, cpuctx, ctx);
1265 event_sched_out(event, cpuctx, ctx);
1266 event->state = PERF_EVENT_STATE_OFF;
1269 raw_spin_unlock(&ctx->lock);
1277 * If event->ctx is a cloned context, callers must make sure that
1278 * every task struct that event->ctx->task could possibly point to
1279 * remains valid. This condition is satisifed when called through
1280 * perf_event_for_each_child or perf_event_for_each because they
1281 * hold the top-level event's child_mutex, so any descendant that
1282 * goes to exit will block in sync_child_event.
1283 * When called from perf_pending_event it's OK because event->ctx
1284 * is the current context on this CPU and preemption is disabled,
1285 * hence we can't get into perf_event_task_sched_out for this context.
1287 void perf_event_disable(struct perf_event *event)
1289 struct perf_event_context *ctx = event->ctx;
1290 struct task_struct *task = ctx->task;
1294 * Disable the event on the cpu that it's on
1296 cpu_function_call(event->cpu, __perf_event_disable, event);
1301 if (!task_function_call(task, __perf_event_disable, event))
1304 raw_spin_lock_irq(&ctx->lock);
1306 * If the event is still active, we need to retry the cross-call.
1308 if (event->state == PERF_EVENT_STATE_ACTIVE) {
1309 raw_spin_unlock_irq(&ctx->lock);
1311 * Reload the task pointer, it might have been changed by
1312 * a concurrent perf_event_context_sched_out().
1319 * Since we have the lock this context can't be scheduled
1320 * in, so we can change the state safely.
1322 if (event->state == PERF_EVENT_STATE_INACTIVE) {
1323 update_group_times(event);
1324 event->state = PERF_EVENT_STATE_OFF;
1326 raw_spin_unlock_irq(&ctx->lock);
1328 EXPORT_SYMBOL_GPL(perf_event_disable);
1330 static void perf_set_shadow_time(struct perf_event *event,
1331 struct perf_event_context *ctx,
1335 * use the correct time source for the time snapshot
1337 * We could get by without this by leveraging the
1338 * fact that to get to this function, the caller
1339 * has most likely already called update_context_time()
1340 * and update_cgrp_time_xx() and thus both timestamp
1341 * are identical (or very close). Given that tstamp is,
1342 * already adjusted for cgroup, we could say that:
1343 * tstamp - ctx->timestamp
1345 * tstamp - cgrp->timestamp.
1347 * Then, in perf_output_read(), the calculation would
1348 * work with no changes because:
1349 * - event is guaranteed scheduled in
1350 * - no scheduled out in between
1351 * - thus the timestamp would be the same
1353 * But this is a bit hairy.
1355 * So instead, we have an explicit cgroup call to remain
1356 * within the time time source all along. We believe it
1357 * is cleaner and simpler to understand.
1359 if (is_cgroup_event(event))
1360 perf_cgroup_set_shadow_time(event, tstamp);
1362 event->shadow_ctx_time = tstamp - ctx->timestamp;
1365 #define MAX_INTERRUPTS (~0ULL)
1367 static void perf_log_throttle(struct perf_event *event, int enable);
1370 event_sched_in(struct perf_event *event,
1371 struct perf_cpu_context *cpuctx,
1372 struct perf_event_context *ctx)
1374 u64 tstamp = perf_event_time(event);
1376 if (event->state <= PERF_EVENT_STATE_OFF)
1379 event->state = PERF_EVENT_STATE_ACTIVE;
1380 event->oncpu = smp_processor_id();
1383 * Unthrottle events, since we scheduled we might have missed several
1384 * ticks already, also for a heavily scheduling task there is little
1385 * guarantee it'll get a tick in a timely manner.
1387 if (unlikely(event->hw.interrupts == MAX_INTERRUPTS)) {
1388 perf_log_throttle(event, 1);
1389 event->hw.interrupts = 0;
1393 * The new state must be visible before we turn it on in the hardware:
1397 if (event->pmu->add(event, PERF_EF_START)) {
1398 event->state = PERF_EVENT_STATE_INACTIVE;
1403 event->tstamp_running += tstamp - event->tstamp_stopped;
1405 perf_set_shadow_time(event, ctx, tstamp);
1407 if (!is_software_event(event))
1408 cpuctx->active_oncpu++;
1411 if (event->attr.exclusive)
1412 cpuctx->exclusive = 1;
1418 group_sched_in(struct perf_event *group_event,
1419 struct perf_cpu_context *cpuctx,
1420 struct perf_event_context *ctx)
1422 struct perf_event *event, *partial_group = NULL;
1423 struct pmu *pmu = group_event->pmu;
1424 u64 now = ctx->time;
1425 bool simulate = false;
1427 if (group_event->state == PERF_EVENT_STATE_OFF)
1430 pmu->start_txn(pmu);
1432 if (event_sched_in(group_event, cpuctx, ctx)) {
1433 pmu->cancel_txn(pmu);
1438 * Schedule in siblings as one group (if any):
1440 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
1441 if (event_sched_in(event, cpuctx, ctx)) {
1442 partial_group = event;
1447 if (!pmu->commit_txn(pmu))
1452 * Groups can be scheduled in as one unit only, so undo any
1453 * partial group before returning:
1454 * The events up to the failed event are scheduled out normally,
1455 * tstamp_stopped will be updated.
1457 * The failed events and the remaining siblings need to have
1458 * their timings updated as if they had gone thru event_sched_in()
1459 * and event_sched_out(). This is required to get consistent timings
1460 * across the group. This also takes care of the case where the group
1461 * could never be scheduled by ensuring tstamp_stopped is set to mark
1462 * the time the event was actually stopped, such that time delta
1463 * calculation in update_event_times() is correct.
1465 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
1466 if (event == partial_group)
1470 event->tstamp_running += now - event->tstamp_stopped;
1471 event->tstamp_stopped = now;
1473 event_sched_out(event, cpuctx, ctx);
1476 event_sched_out(group_event, cpuctx, ctx);
1478 pmu->cancel_txn(pmu);
1484 * Work out whether we can put this event group on the CPU now.
1486 static int group_can_go_on(struct perf_event *event,
1487 struct perf_cpu_context *cpuctx,
1491 * Groups consisting entirely of software events can always go on.
1493 if (event->group_flags & PERF_GROUP_SOFTWARE)
1496 * If an exclusive group is already on, no other hardware
1499 if (cpuctx->exclusive)
1502 * If this group is exclusive and there are already
1503 * events on the CPU, it can't go on.
1505 if (event->attr.exclusive && cpuctx->active_oncpu)
1508 * Otherwise, try to add it if all previous groups were able
1514 static void add_event_to_ctx(struct perf_event *event,
1515 struct perf_event_context *ctx)
1517 u64 tstamp = perf_event_time(event);
1519 list_add_event(event, ctx);
1520 perf_group_attach(event);
1521 event->tstamp_enabled = tstamp;
1522 event->tstamp_running = tstamp;
1523 event->tstamp_stopped = tstamp;
1526 static void task_ctx_sched_out(struct perf_event_context *ctx);
1528 ctx_sched_in(struct perf_event_context *ctx,
1529 struct perf_cpu_context *cpuctx,
1530 enum event_type_t event_type,
1531 struct task_struct *task);
1533 static void perf_event_sched_in(struct perf_cpu_context *cpuctx,
1534 struct perf_event_context *ctx,
1535 struct task_struct *task)
1537 cpu_ctx_sched_in(cpuctx, EVENT_PINNED, task);
1539 ctx_sched_in(ctx, cpuctx, EVENT_PINNED, task);
1540 cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE, task);
1542 ctx_sched_in(ctx, cpuctx, EVENT_FLEXIBLE, task);
1546 * Cross CPU call to install and enable a performance event
1548 * Must be called with ctx->mutex held
1550 static int __perf_install_in_context(void *info)
1552 struct perf_event *event = info;
1553 struct perf_event_context *ctx = event->ctx;
1554 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1555 struct perf_event_context *task_ctx = cpuctx->task_ctx;
1556 struct task_struct *task = current;
1558 perf_ctx_lock(cpuctx, task_ctx);
1559 perf_pmu_disable(cpuctx->ctx.pmu);
1562 * If there was an active task_ctx schedule it out.
1565 task_ctx_sched_out(task_ctx);
1568 * If the context we're installing events in is not the
1569 * active task_ctx, flip them.
1571 if (ctx->task && task_ctx != ctx) {
1573 raw_spin_unlock(&task_ctx->lock);
1574 raw_spin_lock(&ctx->lock);
1579 cpuctx->task_ctx = task_ctx;
1580 task = task_ctx->task;
1583 cpu_ctx_sched_out(cpuctx, EVENT_ALL);
1585 update_context_time(ctx);
1587 * update cgrp time only if current cgrp
1588 * matches event->cgrp. Must be done before
1589 * calling add_event_to_ctx()
1591 update_cgrp_time_from_event(event);
1593 add_event_to_ctx(event, ctx);
1596 * Schedule everything back in
1598 perf_event_sched_in(cpuctx, task_ctx, task);
1600 perf_pmu_enable(cpuctx->ctx.pmu);
1601 perf_ctx_unlock(cpuctx, task_ctx);
1607 * Attach a performance event to a context
1609 * First we add the event to the list with the hardware enable bit
1610 * in event->hw_config cleared.
1612 * If the event is attached to a task which is on a CPU we use a smp
1613 * call to enable it in the task context. The task might have been
1614 * scheduled away, but we check this in the smp call again.
1617 perf_install_in_context(struct perf_event_context *ctx,
1618 struct perf_event *event,
1621 struct task_struct *task = ctx->task;
1623 lockdep_assert_held(&ctx->mutex);
1629 * Per cpu events are installed via an smp call and
1630 * the install is always successful.
1632 cpu_function_call(cpu, __perf_install_in_context, event);
1637 if (!task_function_call(task, __perf_install_in_context, event))
1640 raw_spin_lock_irq(&ctx->lock);
1642 * If we failed to find a running task, but find the context active now
1643 * that we've acquired the ctx->lock, retry.
1645 if (ctx->is_active) {
1646 raw_spin_unlock_irq(&ctx->lock);
1651 * Since the task isn't running, its safe to add the event, us holding
1652 * the ctx->lock ensures the task won't get scheduled in.
1654 add_event_to_ctx(event, ctx);
1655 raw_spin_unlock_irq(&ctx->lock);
1659 * Put a event into inactive state and update time fields.
1660 * Enabling the leader of a group effectively enables all
1661 * the group members that aren't explicitly disabled, so we
1662 * have to update their ->tstamp_enabled also.
1663 * Note: this works for group members as well as group leaders
1664 * since the non-leader members' sibling_lists will be empty.
1666 static void __perf_event_mark_enabled(struct perf_event *event,
1667 struct perf_event_context *ctx)
1669 struct perf_event *sub;
1670 u64 tstamp = perf_event_time(event);
1672 event->state = PERF_EVENT_STATE_INACTIVE;
1673 event->tstamp_enabled = tstamp - event->total_time_enabled;
1674 list_for_each_entry(sub, &event->sibling_list, group_entry) {
1675 if (sub->state >= PERF_EVENT_STATE_INACTIVE)
1676 sub->tstamp_enabled = tstamp - sub->total_time_enabled;
1681 * Cross CPU call to enable a performance event
1683 static int __perf_event_enable(void *info)
1685 struct perf_event *event = info;
1686 struct perf_event_context *ctx = event->ctx;
1687 struct perf_event *leader = event->group_leader;
1688 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1691 if (WARN_ON_ONCE(!ctx->is_active))
1694 raw_spin_lock(&ctx->lock);
1695 update_context_time(ctx);
1697 if (event->state >= PERF_EVENT_STATE_INACTIVE)
1701 * set current task's cgroup time reference point
1703 perf_cgroup_set_timestamp(current, ctx);
1705 __perf_event_mark_enabled(event, ctx);
1707 if (!event_filter_match(event)) {
1708 if (is_cgroup_event(event))
1709 perf_cgroup_defer_enabled(event);
1714 * If the event is in a group and isn't the group leader,
1715 * then don't put it on unless the group is on.
1717 if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE)
1720 if (!group_can_go_on(event, cpuctx, 1)) {
1723 if (event == leader)
1724 err = group_sched_in(event, cpuctx, ctx);
1726 err = event_sched_in(event, cpuctx, ctx);
1731 * If this event can't go on and it's part of a
1732 * group, then the whole group has to come off.
1734 if (leader != event)
1735 group_sched_out(leader, cpuctx, ctx);
1736 if (leader->attr.pinned) {
1737 update_group_times(leader);
1738 leader->state = PERF_EVENT_STATE_ERROR;
1743 raw_spin_unlock(&ctx->lock);
1751 * If event->ctx is a cloned context, callers must make sure that
1752 * every task struct that event->ctx->task could possibly point to
1753 * remains valid. This condition is satisfied when called through
1754 * perf_event_for_each_child or perf_event_for_each as described
1755 * for perf_event_disable.
1757 void perf_event_enable(struct perf_event *event)
1759 struct perf_event_context *ctx = event->ctx;
1760 struct task_struct *task = ctx->task;
1764 * Enable the event on the cpu that it's on
1766 cpu_function_call(event->cpu, __perf_event_enable, event);
1770 raw_spin_lock_irq(&ctx->lock);
1771 if (event->state >= PERF_EVENT_STATE_INACTIVE)
1775 * If the event is in error state, clear that first.
1776 * That way, if we see the event in error state below, we
1777 * know that it has gone back into error state, as distinct
1778 * from the task having been scheduled away before the
1779 * cross-call arrived.
1781 if (event->state == PERF_EVENT_STATE_ERROR)
1782 event->state = PERF_EVENT_STATE_OFF;
1785 if (!ctx->is_active) {
1786 __perf_event_mark_enabled(event, ctx);
1790 raw_spin_unlock_irq(&ctx->lock);
1792 if (!task_function_call(task, __perf_event_enable, event))
1795 raw_spin_lock_irq(&ctx->lock);
1798 * If the context is active and the event is still off,
1799 * we need to retry the cross-call.
1801 if (ctx->is_active && event->state == PERF_EVENT_STATE_OFF) {
1803 * task could have been flipped by a concurrent
1804 * perf_event_context_sched_out()
1811 raw_spin_unlock_irq(&ctx->lock);
1813 EXPORT_SYMBOL_GPL(perf_event_enable);
1815 int perf_event_refresh(struct perf_event *event, int refresh)
1818 * not supported on inherited events
1820 if (event->attr.inherit || !is_sampling_event(event))
1823 atomic_add(refresh, &event->event_limit);
1824 perf_event_enable(event);
1828 EXPORT_SYMBOL_GPL(perf_event_refresh);
1830 static void ctx_sched_out(struct perf_event_context *ctx,
1831 struct perf_cpu_context *cpuctx,
1832 enum event_type_t event_type)
1834 struct perf_event *event;
1835 int is_active = ctx->is_active;
1837 ctx->is_active &= ~event_type;
1838 if (likely(!ctx->nr_events))
1841 update_context_time(ctx);
1842 update_cgrp_time_from_cpuctx(cpuctx);
1843 if (!ctx->nr_active)
1846 perf_pmu_disable(ctx->pmu);
1847 if ((is_active & EVENT_PINNED) && (event_type & EVENT_PINNED)) {
1848 list_for_each_entry(event, &ctx->pinned_groups, group_entry)
1849 group_sched_out(event, cpuctx, ctx);
1852 if ((is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE)) {
1853 list_for_each_entry(event, &ctx->flexible_groups, group_entry)
1854 group_sched_out(event, cpuctx, ctx);
1856 perf_pmu_enable(ctx->pmu);
1860 * Test whether two contexts are equivalent, i.e. whether they
1861 * have both been cloned from the same version of the same context
1862 * and they both have the same number of enabled events.
1863 * If the number of enabled events is the same, then the set
1864 * of enabled events should be the same, because these are both
1865 * inherited contexts, therefore we can't access individual events
1866 * in them directly with an fd; we can only enable/disable all
1867 * events via prctl, or enable/disable all events in a family
1868 * via ioctl, which will have the same effect on both contexts.
1870 static int context_equiv(struct perf_event_context *ctx1,
1871 struct perf_event_context *ctx2)
1873 return ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx
1874 && ctx1->parent_gen == ctx2->parent_gen
1875 && !ctx1->pin_count && !ctx2->pin_count;
1878 static void __perf_event_sync_stat(struct perf_event *event,
1879 struct perf_event *next_event)
1883 if (!event->attr.inherit_stat)
1887 * Update the event value, we cannot use perf_event_read()
1888 * because we're in the middle of a context switch and have IRQs
1889 * disabled, which upsets smp_call_function_single(), however
1890 * we know the event must be on the current CPU, therefore we
1891 * don't need to use it.
1893 switch (event->state) {
1894 case PERF_EVENT_STATE_ACTIVE:
1895 event->pmu->read(event);
1898 case PERF_EVENT_STATE_INACTIVE:
1899 update_event_times(event);
1907 * In order to keep per-task stats reliable we need to flip the event
1908 * values when we flip the contexts.
1910 value = local64_read(&next_event->count);
1911 value = local64_xchg(&event->count, value);
1912 local64_set(&next_event->count, value);
1914 swap(event->total_time_enabled, next_event->total_time_enabled);
1915 swap(event->total_time_running, next_event->total_time_running);
1918 * Since we swizzled the values, update the user visible data too.
1920 perf_event_update_userpage(event);
1921 perf_event_update_userpage(next_event);
1924 #define list_next_entry(pos, member) \
1925 list_entry(pos->member.next, typeof(*pos), member)
1927 static void perf_event_sync_stat(struct perf_event_context *ctx,
1928 struct perf_event_context *next_ctx)
1930 struct perf_event *event, *next_event;
1935 update_context_time(ctx);
1937 event = list_first_entry(&ctx->event_list,
1938 struct perf_event, event_entry);
1940 next_event = list_first_entry(&next_ctx->event_list,
1941 struct perf_event, event_entry);
1943 while (&event->event_entry != &ctx->event_list &&
1944 &next_event->event_entry != &next_ctx->event_list) {
1946 __perf_event_sync_stat(event, next_event);
1948 event = list_next_entry(event, event_entry);
1949 next_event = list_next_entry(next_event, event_entry);
1953 static void perf_event_context_sched_out(struct task_struct *task, int ctxn,
1954 struct task_struct *next)
1956 struct perf_event_context *ctx = task->perf_event_ctxp[ctxn];
1957 struct perf_event_context *next_ctx;
1958 struct perf_event_context *parent;
1959 struct perf_cpu_context *cpuctx;
1965 cpuctx = __get_cpu_context(ctx);
1966 if (!cpuctx->task_ctx)
1970 parent = rcu_dereference(ctx->parent_ctx);
1971 next_ctx = next->perf_event_ctxp[ctxn];
1972 if (parent && next_ctx &&
1973 rcu_dereference(next_ctx->parent_ctx) == parent) {
1975 * Looks like the two contexts are clones, so we might be
1976 * able to optimize the context switch. We lock both
1977 * contexts and check that they are clones under the
1978 * lock (including re-checking that neither has been
1979 * uncloned in the meantime). It doesn't matter which
1980 * order we take the locks because no other cpu could
1981 * be trying to lock both of these tasks.
1983 raw_spin_lock(&ctx->lock);
1984 raw_spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING);
1985 if (context_equiv(ctx, next_ctx)) {
1987 * XXX do we need a memory barrier of sorts
1988 * wrt to rcu_dereference() of perf_event_ctxp
1990 task->perf_event_ctxp[ctxn] = next_ctx;
1991 next->perf_event_ctxp[ctxn] = ctx;
1993 next_ctx->task = task;
1996 perf_event_sync_stat(ctx, next_ctx);
1998 raw_spin_unlock(&next_ctx->lock);
1999 raw_spin_unlock(&ctx->lock);
2004 raw_spin_lock(&ctx->lock);
2005 ctx_sched_out(ctx, cpuctx, EVENT_ALL);
2006 cpuctx->task_ctx = NULL;
2007 raw_spin_unlock(&ctx->lock);
2011 #define for_each_task_context_nr(ctxn) \
2012 for ((ctxn) = 0; (ctxn) < perf_nr_task_contexts; (ctxn)++)
2015 * Called from scheduler to remove the events of the current task,
2016 * with interrupts disabled.
2018 * We stop each event and update the event value in event->count.
2020 * This does not protect us against NMI, but disable()
2021 * sets the disabled bit in the control field of event _before_
2022 * accessing the event control register. If a NMI hits, then it will
2023 * not restart the event.
2025 void __perf_event_task_sched_out(struct task_struct *task,
2026 struct task_struct *next)
2030 for_each_task_context_nr(ctxn)
2031 perf_event_context_sched_out(task, ctxn, next);
2034 * if cgroup events exist on this CPU, then we need
2035 * to check if we have to switch out PMU state.
2036 * cgroup event are system-wide mode only
2038 if (atomic_read(&__get_cpu_var(perf_cgroup_events)))
2039 perf_cgroup_sched_out(task, next);
2042 static void task_ctx_sched_out(struct perf_event_context *ctx)
2044 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
2046 if (!cpuctx->task_ctx)
2049 if (WARN_ON_ONCE(ctx != cpuctx->task_ctx))
2052 ctx_sched_out(ctx, cpuctx, EVENT_ALL);
2053 cpuctx->task_ctx = NULL;
2057 * Called with IRQs disabled
2059 static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
2060 enum event_type_t event_type)
2062 ctx_sched_out(&cpuctx->ctx, cpuctx, event_type);
2066 ctx_pinned_sched_in(struct perf_event_context *ctx,
2067 struct perf_cpu_context *cpuctx)
2069 struct perf_event *event;
2071 list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
2072 if (event->state <= PERF_EVENT_STATE_OFF)
2074 if (!event_filter_match(event))
2077 /* may need to reset tstamp_enabled */
2078 if (is_cgroup_event(event))
2079 perf_cgroup_mark_enabled(event, ctx);
2081 if (group_can_go_on(event, cpuctx, 1))
2082 group_sched_in(event, cpuctx, ctx);
2085 * If this pinned group hasn't been scheduled,
2086 * put it in error state.
2088 if (event->state == PERF_EVENT_STATE_INACTIVE) {
2089 update_group_times(event);
2090 event->state = PERF_EVENT_STATE_ERROR;
2096 ctx_flexible_sched_in(struct perf_event_context *ctx,
2097 struct perf_cpu_context *cpuctx)
2099 struct perf_event *event;
2102 list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
2103 /* Ignore events in OFF or ERROR state */
2104 if (event->state <= PERF_EVENT_STATE_OFF)
2107 * Listen to the 'cpu' scheduling filter constraint
2110 if (!event_filter_match(event))
2113 /* may need to reset tstamp_enabled */
2114 if (is_cgroup_event(event))
2115 perf_cgroup_mark_enabled(event, ctx);
2117 if (group_can_go_on(event, cpuctx, can_add_hw)) {
2118 if (group_sched_in(event, cpuctx, ctx))
2125 ctx_sched_in(struct perf_event_context *ctx,
2126 struct perf_cpu_context *cpuctx,
2127 enum event_type_t event_type,
2128 struct task_struct *task)
2131 int is_active = ctx->is_active;
2133 ctx->is_active |= event_type;
2134 if (likely(!ctx->nr_events))
2138 ctx->timestamp = now;
2139 perf_cgroup_set_timestamp(task, ctx);
2141 * First go through the list and put on any pinned groups
2142 * in order to give them the best chance of going on.
2144 if (!(is_active & EVENT_PINNED) && (event_type & EVENT_PINNED))
2145 ctx_pinned_sched_in(ctx, cpuctx);
2147 /* Then walk through the lower prio flexible groups */
2148 if (!(is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE))
2149 ctx_flexible_sched_in(ctx, cpuctx);
2152 static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
2153 enum event_type_t event_type,
2154 struct task_struct *task)
2156 struct perf_event_context *ctx = &cpuctx->ctx;
2158 ctx_sched_in(ctx, cpuctx, event_type, task);
2161 static void perf_event_context_sched_in(struct perf_event_context *ctx,
2162 struct task_struct *task)
2164 struct perf_cpu_context *cpuctx;
2166 cpuctx = __get_cpu_context(ctx);
2167 if (cpuctx->task_ctx == ctx)
2170 perf_ctx_lock(cpuctx, ctx);
2171 perf_pmu_disable(ctx->pmu);
2173 * We want to keep the following priority order:
2174 * cpu pinned (that don't need to move), task pinned,
2175 * cpu flexible, task flexible.
2177 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
2179 perf_event_sched_in(cpuctx, ctx, task);
2182 cpuctx->task_ctx = ctx;
2184 perf_pmu_enable(ctx->pmu);
2185 perf_ctx_unlock(cpuctx, ctx);
2188 * Since these rotations are per-cpu, we need to ensure the
2189 * cpu-context we got scheduled on is actually rotating.
2191 perf_pmu_rotate_start(ctx->pmu);
2195 * Called from scheduler to add the events of the current task
2196 * with interrupts disabled.
2198 * We restore the event value and then enable it.
2200 * This does not protect us against NMI, but enable()
2201 * sets the enabled bit in the control field of event _before_
2202 * accessing the event control register. If a NMI hits, then it will
2203 * keep the event running.
2205 void __perf_event_task_sched_in(struct task_struct *prev,
2206 struct task_struct *task)
2208 struct perf_event_context *ctx;
2211 for_each_task_context_nr(ctxn) {
2212 ctx = task->perf_event_ctxp[ctxn];
2216 perf_event_context_sched_in(ctx, task);
2219 * if cgroup events exist on this CPU, then we need
2220 * to check if we have to switch in PMU state.
2221 * cgroup event are system-wide mode only
2223 if (atomic_read(&__get_cpu_var(perf_cgroup_events)))
2224 perf_cgroup_sched_in(prev, task);
2227 static u64 perf_calculate_period(struct perf_event *event, u64 nsec, u64 count)
2229 u64 frequency = event->attr.sample_freq;
2230 u64 sec = NSEC_PER_SEC;
2231 u64 divisor, dividend;
2233 int count_fls, nsec_fls, frequency_fls, sec_fls;
2235 count_fls = fls64(count);
2236 nsec_fls = fls64(nsec);
2237 frequency_fls = fls64(frequency);
2241 * We got @count in @nsec, with a target of sample_freq HZ
2242 * the target period becomes:
2245 * period = -------------------
2246 * @nsec * sample_freq
2251 * Reduce accuracy by one bit such that @a and @b converge
2252 * to a similar magnitude.
2254 #define REDUCE_FLS(a, b) \
2256 if (a##_fls > b##_fls) { \
2266 * Reduce accuracy until either term fits in a u64, then proceed with
2267 * the other, so that finally we can do a u64/u64 division.
2269 while (count_fls + sec_fls > 64 && nsec_fls + frequency_fls > 64) {
2270 REDUCE_FLS(nsec, frequency);
2271 REDUCE_FLS(sec, count);
2274 if (count_fls + sec_fls > 64) {
2275 divisor = nsec * frequency;
2277 while (count_fls + sec_fls > 64) {
2278 REDUCE_FLS(count, sec);
2282 dividend = count * sec;
2284 dividend = count * sec;
2286 while (nsec_fls + frequency_fls > 64) {
2287 REDUCE_FLS(nsec, frequency);
2291 divisor = nsec * frequency;
2297 return div64_u64(dividend, divisor);
2300 static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count)
2302 struct hw_perf_event *hwc = &event->hw;
2303 s64 period, sample_period;
2306 period = perf_calculate_period(event, nsec, count);
2308 delta = (s64)(period - hwc->sample_period);
2309 delta = (delta + 7) / 8; /* low pass filter */
2311 sample_period = hwc->sample_period + delta;
2316 hwc->sample_period = sample_period;
2318 if (local64_read(&hwc->period_left) > 8*sample_period) {
2319 event->pmu->stop(event, PERF_EF_UPDATE);
2320 local64_set(&hwc->period_left, 0);
2321 event->pmu->start(event, PERF_EF_RELOAD);
2325 static void perf_ctx_adjust_freq(struct perf_event_context *ctx, u64 period)
2327 struct perf_event *event;
2328 struct hw_perf_event *hwc;
2329 u64 interrupts, now;
2332 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
2333 if (event->state != PERF_EVENT_STATE_ACTIVE)
2336 if (!event_filter_match(event))
2341 interrupts = hwc->interrupts;
2342 hwc->interrupts = 0;
2345 * unthrottle events on the tick
2347 if (interrupts == MAX_INTERRUPTS) {
2348 perf_log_throttle(event, 1);
2349 event->pmu->start(event, 0);
2352 if (!event->attr.freq || !event->attr.sample_freq)
2355 event->pmu->read(event);
2356 now = local64_read(&event->count);
2357 delta = now - hwc->freq_count_stamp;
2358 hwc->freq_count_stamp = now;
2361 perf_adjust_period(event, period, delta);
2366 * Round-robin a context's events:
2368 static void rotate_ctx(struct perf_event_context *ctx)
2371 * Rotate the first entry last of non-pinned groups. Rotation might be
2372 * disabled by the inheritance code.
2374 if (!ctx->rotate_disable)
2375 list_rotate_left(&ctx->flexible_groups);
2379 * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
2380 * because they're strictly cpu affine and rotate_start is called with IRQs
2381 * disabled, while rotate_context is called from IRQ context.
2383 static void perf_rotate_context(struct perf_cpu_context *cpuctx)
2385 u64 interval = (u64)cpuctx->jiffies_interval * TICK_NSEC;
2386 struct perf_event_context *ctx = NULL;
2387 int rotate = 0, remove = 1;
2389 if (cpuctx->ctx.nr_events) {
2391 if (cpuctx->ctx.nr_events != cpuctx->ctx.nr_active)
2395 ctx = cpuctx->task_ctx;
2396 if (ctx && ctx->nr_events) {
2398 if (ctx->nr_events != ctx->nr_active)
2402 perf_ctx_lock(cpuctx, cpuctx->task_ctx);
2403 perf_pmu_disable(cpuctx->ctx.pmu);
2404 perf_ctx_adjust_freq(&cpuctx->ctx, interval);
2406 perf_ctx_adjust_freq(ctx, interval);
2411 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
2413 ctx_sched_out(ctx, cpuctx, EVENT_FLEXIBLE);
2415 rotate_ctx(&cpuctx->ctx);
2419 perf_event_sched_in(cpuctx, ctx, current);
2423 list_del_init(&cpuctx->rotation_list);
2425 perf_pmu_enable(cpuctx->ctx.pmu);
2426 perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
2429 void perf_event_task_tick(void)
2431 struct list_head *head = &__get_cpu_var(rotation_list);
2432 struct perf_cpu_context *cpuctx, *tmp;
2434 WARN_ON(!irqs_disabled());
2436 list_for_each_entry_safe(cpuctx, tmp, head, rotation_list) {
2437 if (cpuctx->jiffies_interval == 1 ||
2438 !(jiffies % cpuctx->jiffies_interval))
2439 perf_rotate_context(cpuctx);
2443 static int event_enable_on_exec(struct perf_event *event,
2444 struct perf_event_context *ctx)
2446 if (!event->attr.enable_on_exec)
2449 event->attr.enable_on_exec = 0;
2450 if (event->state >= PERF_EVENT_STATE_INACTIVE)
2453 __perf_event_mark_enabled(event, ctx);
2459 * Enable all of a task's events that have been marked enable-on-exec.
2460 * This expects task == current.
2462 static void perf_event_enable_on_exec(struct perf_event_context *ctx)
2464 struct perf_event *event;
2465 unsigned long flags;
2469 local_irq_save(flags);
2470 if (!ctx || !ctx->nr_events)
2474 * We must ctxsw out cgroup events to avoid conflict
2475 * when invoking perf_task_event_sched_in() later on
2476 * in this function. Otherwise we end up trying to
2477 * ctxswin cgroup events which are already scheduled
2480 perf_cgroup_sched_out(current, NULL);
2482 raw_spin_lock(&ctx->lock);
2483 task_ctx_sched_out(ctx);
2485 list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
2486 ret = event_enable_on_exec(event, ctx);
2491 list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
2492 ret = event_enable_on_exec(event, ctx);
2498 * Unclone this context if we enabled any event.
2503 raw_spin_unlock(&ctx->lock);
2506 * Also calls ctxswin for cgroup events, if any:
2508 perf_event_context_sched_in(ctx, ctx->task);
2510 local_irq_restore(flags);
2514 * Cross CPU call to read the hardware event
2516 static void __perf_event_read(void *info)
2518 struct perf_event *event = info;
2519 struct perf_event_context *ctx = event->ctx;
2520 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
2523 * If this is a task context, we need to check whether it is
2524 * the current task context of this cpu. If not it has been
2525 * scheduled out before the smp call arrived. In that case
2526 * event->count would have been updated to a recent sample
2527 * when the event was scheduled out.
2529 if (ctx->task && cpuctx->task_ctx != ctx)
2532 raw_spin_lock(&ctx->lock);
2533 if (ctx->is_active) {
2534 update_context_time(ctx);
2535 update_cgrp_time_from_event(event);
2537 update_event_times(event);
2538 if (event->state == PERF_EVENT_STATE_ACTIVE)
2539 event->pmu->read(event);
2540 raw_spin_unlock(&ctx->lock);
2543 static inline u64 perf_event_count(struct perf_event *event)
2545 return local64_read(&event->count) + atomic64_read(&event->child_count);
2548 static u64 perf_event_read(struct perf_event *event)
2551 * If event is enabled and currently active on a CPU, update the
2552 * value in the event structure:
2554 if (event->state == PERF_EVENT_STATE_ACTIVE) {
2555 smp_call_function_single(event->oncpu,
2556 __perf_event_read, event, 1);
2557 } else if (event->state == PERF_EVENT_STATE_INACTIVE) {
2558 struct perf_event_context *ctx = event->ctx;
2559 unsigned long flags;
2561 raw_spin_lock_irqsave(&ctx->lock, flags);
2563 * may read while context is not active
2564 * (e.g., thread is blocked), in that case
2565 * we cannot update context time
2567 if (ctx->is_active) {
2568 update_context_time(ctx);
2569 update_cgrp_time_from_event(event);
2571 update_event_times(event);
2572 raw_spin_unlock_irqrestore(&ctx->lock, flags);
2575 return perf_event_count(event);
2579 * Initialize the perf_event context in a task_struct:
2581 static void __perf_event_init_context(struct perf_event_context *ctx)
2583 raw_spin_lock_init(&ctx->lock);
2584 mutex_init(&ctx->mutex);
2585 INIT_LIST_HEAD(&ctx->pinned_groups);
2586 INIT_LIST_HEAD(&ctx->flexible_groups);
2587 INIT_LIST_HEAD(&ctx->event_list);
2588 atomic_set(&ctx->refcount, 1);
2591 static struct perf_event_context *
2592 alloc_perf_context(struct pmu *pmu, struct task_struct *task)
2594 struct perf_event_context *ctx;
2596 ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL);
2600 __perf_event_init_context(ctx);
2603 get_task_struct(task);
2610 static struct task_struct *
2611 find_lively_task_by_vpid(pid_t vpid)
2613 struct task_struct *task;
2620 task = find_task_by_vpid(vpid);
2622 get_task_struct(task);
2626 return ERR_PTR(-ESRCH);
2628 /* Reuse ptrace permission checks for now. */
2630 if (!ptrace_may_access(task, PTRACE_MODE_READ))
2635 put_task_struct(task);
2636 return ERR_PTR(err);
2641 * Returns a matching context with refcount and pincount.
2643 static struct perf_event_context *
2644 find_get_context(struct pmu *pmu, struct task_struct *task, int cpu)
2646 struct perf_event_context *ctx;
2647 struct perf_cpu_context *cpuctx;
2648 unsigned long flags;
2652 /* Must be root to operate on a CPU event: */
2653 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
2654 return ERR_PTR(-EACCES);
2657 * We could be clever and allow to attach a event to an
2658 * offline CPU and activate it when the CPU comes up, but
2661 if (!cpu_online(cpu))
2662 return ERR_PTR(-ENODEV);
2664 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
2673 ctxn = pmu->task_ctx_nr;
2678 ctx = perf_lock_task_context(task, ctxn, &flags);
2682 raw_spin_unlock_irqrestore(&ctx->lock, flags);
2684 ctx = alloc_perf_context(pmu, task);
2690 mutex_lock(&task->perf_event_mutex);
2692 * If it has already passed perf_event_exit_task().
2693 * we must see PF_EXITING, it takes this mutex too.
2695 if (task->flags & PF_EXITING)
2697 else if (task->perf_event_ctxp[ctxn])
2702 rcu_assign_pointer(task->perf_event_ctxp[ctxn], ctx);
2704 mutex_unlock(&task->perf_event_mutex);
2706 if (unlikely(err)) {
2718 return ERR_PTR(err);
2721 static void perf_event_free_filter(struct perf_event *event);
2723 static void free_event_rcu(struct rcu_head *head)
2725 struct perf_event *event;
2727 event = container_of(head, struct perf_event, rcu_head);
2729 put_pid_ns(event->ns);
2730 perf_event_free_filter(event);
2734 static void ring_buffer_put(struct ring_buffer *rb);
2736 static void free_event(struct perf_event *event)
2738 irq_work_sync(&event->pending);
2740 if (!event->parent) {
2741 if (event->attach_state & PERF_ATTACH_TASK)
2742 jump_label_dec(&perf_sched_events);
2743 if (event->attr.mmap || event->attr.mmap_data)
2744 atomic_dec(&nr_mmap_events);
2745 if (event->attr.comm)
2746 atomic_dec(&nr_comm_events);
2747 if (event->attr.task)
2748 atomic_dec(&nr_task_events);
2749 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN)
2750 put_callchain_buffers();
2751 if (is_cgroup_event(event)) {
2752 atomic_dec(&per_cpu(perf_cgroup_events, event->cpu));
2753 jump_label_dec(&perf_sched_events);
2758 ring_buffer_put(event->rb);
2762 if (is_cgroup_event(event))
2763 perf_detach_cgroup(event);
2766 event->destroy(event);
2769 put_ctx(event->ctx);
2771 call_rcu(&event->rcu_head, free_event_rcu);
2774 int perf_event_release_kernel(struct perf_event *event)
2776 struct perf_event_context *ctx = event->ctx;
2778 WARN_ON_ONCE(ctx->parent_ctx);
2780 * There are two ways this annotation is useful:
2782 * 1) there is a lock recursion from perf_event_exit_task
2783 * see the comment there.
2785 * 2) there is a lock-inversion with mmap_sem through
2786 * perf_event_read_group(), which takes faults while
2787 * holding ctx->mutex, however this is called after
2788 * the last filedesc died, so there is no possibility
2789 * to trigger the AB-BA case.
2791 mutex_lock_nested(&ctx->mutex, SINGLE_DEPTH_NESTING);
2792 raw_spin_lock_irq(&ctx->lock);
2793 perf_group_detach(event);
2794 raw_spin_unlock_irq(&ctx->lock);
2795 perf_remove_from_context(event);
2796 mutex_unlock(&ctx->mutex);
2802 EXPORT_SYMBOL_GPL(perf_event_release_kernel);
2805 * Called when the last reference to the file is gone.
2807 static int perf_release(struct inode *inode, struct file *file)
2809 struct perf_event *event = file->private_data;
2810 struct task_struct *owner;
2812 file->private_data = NULL;
2815 owner = ACCESS_ONCE(event->owner);
2817 * Matches the smp_wmb() in perf_event_exit_task(). If we observe
2818 * !owner it means the list deletion is complete and we can indeed
2819 * free this event, otherwise we need to serialize on
2820 * owner->perf_event_mutex.
2822 smp_read_barrier_depends();
2825 * Since delayed_put_task_struct() also drops the last
2826 * task reference we can safely take a new reference
2827 * while holding the rcu_read_lock().
2829 get_task_struct(owner);
2834 mutex_lock(&owner->perf_event_mutex);
2836 * We have to re-check the event->owner field, if it is cleared
2837 * we raced with perf_event_exit_task(), acquiring the mutex
2838 * ensured they're done, and we can proceed with freeing the
2842 list_del_init(&event->owner_entry);
2843 mutex_unlock(&owner->perf_event_mutex);
2844 put_task_struct(owner);
2847 return perf_event_release_kernel(event);
2850 u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running)
2852 struct perf_event *child;
2858 mutex_lock(&event->child_mutex);
2859 total += perf_event_read(event);
2860 *enabled += event->total_time_enabled +
2861 atomic64_read(&event->child_total_time_enabled);
2862 *running += event->total_time_running +
2863 atomic64_read(&event->child_total_time_running);
2865 list_for_each_entry(child, &event->child_list, child_list) {
2866 total += perf_event_read(child);
2867 *enabled += child->total_time_enabled;
2868 *running += child->total_time_running;
2870 mutex_unlock(&event->child_mutex);
2874 EXPORT_SYMBOL_GPL(perf_event_read_value);
2876 static int perf_event_read_group(struct perf_event *event,
2877 u64 read_format, char __user *buf)
2879 struct perf_event *leader = event->group_leader, *sub;
2880 int n = 0, size = 0, ret = -EFAULT;
2881 struct perf_event_context *ctx = leader->ctx;
2883 u64 count, enabled, running;
2885 mutex_lock(&ctx->mutex);
2886 count = perf_event_read_value(leader, &enabled, &running);
2888 values[n++] = 1 + leader->nr_siblings;
2889 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2890 values[n++] = enabled;
2891 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2892 values[n++] = running;
2893 values[n++] = count;
2894 if (read_format & PERF_FORMAT_ID)
2895 values[n++] = primary_event_id(leader);
2897 size = n * sizeof(u64);
2899 if (copy_to_user(buf, values, size))
2904 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
2907 values[n++] = perf_event_read_value(sub, &enabled, &running);
2908 if (read_format & PERF_FORMAT_ID)
2909 values[n++] = primary_event_id(sub);
2911 size = n * sizeof(u64);
2913 if (copy_to_user(buf + ret, values, size)) {
2921 mutex_unlock(&ctx->mutex);
2926 static int perf_event_read_one(struct perf_event *event,
2927 u64 read_format, char __user *buf)
2929 u64 enabled, running;
2933 values[n++] = perf_event_read_value(event, &enabled, &running);
2934 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2935 values[n++] = enabled;
2936 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2937 values[n++] = running;
2938 if (read_format & PERF_FORMAT_ID)
2939 values[n++] = primary_event_id(event);
2941 if (copy_to_user(buf, values, n * sizeof(u64)))
2944 return n * sizeof(u64);
2948 * Read the performance event - simple non blocking version for now
2951 perf_read_hw(struct perf_event *event, char __user *buf, size_t count)
2953 u64 read_format = event->attr.read_format;
2957 * Return end-of-file for a read on a event that is in
2958 * error state (i.e. because it was pinned but it couldn't be
2959 * scheduled on to the CPU at some point).
2961 if (event->state == PERF_EVENT_STATE_ERROR)
2964 if (count < event->read_size)
2967 WARN_ON_ONCE(event->ctx->parent_ctx);
2968 if (read_format & PERF_FORMAT_GROUP)
2969 ret = perf_event_read_group(event, read_format, buf);
2971 ret = perf_event_read_one(event, read_format, buf);
2977 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
2979 struct perf_event *event = file->private_data;
2981 return perf_read_hw(event, buf, count);
2984 static unsigned int perf_poll(struct file *file, poll_table *wait)
2986 struct perf_event *event = file->private_data;
2987 struct ring_buffer *rb;
2988 unsigned int events = POLL_HUP;
2991 * Race between perf_event_set_output() and perf_poll(): perf_poll()
2992 * grabs the rb reference but perf_event_set_output() overrides it.
2993 * Here is the timeline for two threads T1, T2:
2994 * t0: T1, rb = rcu_dereference(event->rb)
2995 * t1: T2, old_rb = event->rb
2996 * t2: T2, event->rb = new rb
2997 * t3: T2, ring_buffer_detach(old_rb)
2998 * t4: T1, ring_buffer_attach(rb1)
2999 * t5: T1, poll_wait(event->waitq)
3001 * To avoid this problem, we grab mmap_mutex in perf_poll()
3002 * thereby ensuring that the assignment of the new ring buffer
3003 * and the detachment of the old buffer appear atomic to perf_poll()
3005 mutex_lock(&event->mmap_mutex);
3008 rb = rcu_dereference(event->rb);
3010 ring_buffer_attach(event, rb);
3011 events = atomic_xchg(&rb->poll, 0);
3015 mutex_unlock(&event->mmap_mutex);
3017 poll_wait(file, &event->waitq, wait);
3022 static void perf_event_reset(struct perf_event *event)
3024 (void)perf_event_read(event);
3025 local64_set(&event->count, 0);
3026 perf_event_update_userpage(event);
3030 * Holding the top-level event's child_mutex means that any
3031 * descendant process that has inherited this event will block
3032 * in sync_child_event if it goes to exit, thus satisfying the
3033 * task existence requirements of perf_event_enable/disable.
3035 static void perf_event_for_each_child(struct perf_event *event,
3036 void (*func)(struct perf_event *))
3038 struct perf_event *child;
3040 WARN_ON_ONCE(event->ctx->parent_ctx);
3041 mutex_lock(&event->child_mutex);
3043 list_for_each_entry(child, &event->child_list, child_list)
3045 mutex_unlock(&event->child_mutex);
3048 static void perf_event_for_each(struct perf_event *event,
3049 void (*func)(struct perf_event *))
3051 struct perf_event_context *ctx = event->ctx;
3052 struct perf_event *sibling;
3054 WARN_ON_ONCE(ctx->parent_ctx);
3055 mutex_lock(&ctx->mutex);
3056 event = event->group_leader;
3058 perf_event_for_each_child(event, func);
3060 list_for_each_entry(sibling, &event->sibling_list, group_entry)
3061 perf_event_for_each_child(event, func);
3062 mutex_unlock(&ctx->mutex);
3065 static int perf_event_period(struct perf_event *event, u64 __user *arg)
3067 struct perf_event_context *ctx = event->ctx;
3071 if (!is_sampling_event(event))
3074 if (copy_from_user(&value, arg, sizeof(value)))
3080 raw_spin_lock_irq(&ctx->lock);
3081 if (event->attr.freq) {
3082 if (value > sysctl_perf_event_sample_rate) {
3087 event->attr.sample_freq = value;
3089 event->attr.sample_period = value;
3090 event->hw.sample_period = value;
3093 raw_spin_unlock_irq(&ctx->lock);
3098 static const struct file_operations perf_fops;
3100 static struct perf_event *perf_fget_light(int fd, int *fput_needed)
3104 file = fget_light(fd, fput_needed);
3106 return ERR_PTR(-EBADF);
3108 if (file->f_op != &perf_fops) {
3109 fput_light(file, *fput_needed);
3111 return ERR_PTR(-EBADF);
3114 return file->private_data;
3117 static int perf_event_set_output(struct perf_event *event,
3118 struct perf_event *output_event);
3119 static int perf_event_set_filter(struct perf_event *event, void __user *arg);
3121 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
3123 struct perf_event *event = file->private_data;
3124 void (*func)(struct perf_event *);
3128 case PERF_EVENT_IOC_ENABLE:
3129 func = perf_event_enable;
3131 case PERF_EVENT_IOC_DISABLE:
3132 func = perf_event_disable;
3134 case PERF_EVENT_IOC_RESET:
3135 func = perf_event_reset;
3138 case PERF_EVENT_IOC_REFRESH:
3139 return perf_event_refresh(event, arg);
3141 case PERF_EVENT_IOC_PERIOD:
3142 return perf_event_period(event, (u64 __user *)arg);
3144 case PERF_EVENT_IOC_SET_OUTPUT:
3146 struct perf_event *output_event = NULL;
3147 int fput_needed = 0;
3151 output_event = perf_fget_light(arg, &fput_needed);
3152 if (IS_ERR(output_event))
3153 return PTR_ERR(output_event);
3156 ret = perf_event_set_output(event, output_event);
3158 fput_light(output_event->filp, fput_needed);
3163 case PERF_EVENT_IOC_SET_FILTER:
3164 return perf_event_set_filter(event, (void __user *)arg);
3170 if (flags & PERF_IOC_FLAG_GROUP)
3171 perf_event_for_each(event, func);
3173 perf_event_for_each_child(event, func);
3178 int perf_event_task_enable(void)
3180 struct perf_event *event;
3182 mutex_lock(¤t->perf_event_mutex);
3183 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
3184 perf_event_for_each_child(event, perf_event_enable);
3185 mutex_unlock(¤t->perf_event_mutex);
3190 int perf_event_task_disable(void)
3192 struct perf_event *event;
3194 mutex_lock(¤t->perf_event_mutex);
3195 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
3196 perf_event_for_each_child(event, perf_event_disable);
3197 mutex_unlock(¤t->perf_event_mutex);
3202 #ifndef PERF_EVENT_INDEX_OFFSET
3203 # define PERF_EVENT_INDEX_OFFSET 0
3206 static int perf_event_index(struct perf_event *event)
3208 if (event->hw.state & PERF_HES_STOPPED)
3211 if (event->state != PERF_EVENT_STATE_ACTIVE)
3214 return event->hw.idx + 1 - PERF_EVENT_INDEX_OFFSET;
3217 static void calc_timer_values(struct perf_event *event,
3224 ctx_time = event->shadow_ctx_time + now;
3225 *enabled = ctx_time - event->tstamp_enabled;
3226 *running = ctx_time - event->tstamp_running;
3230 * Callers need to ensure there can be no nesting of this function, otherwise
3231 * the seqlock logic goes bad. We can not serialize this because the arch
3232 * code calls this from NMI context.
3234 void perf_event_update_userpage(struct perf_event *event)
3236 struct perf_event_mmap_page *userpg;
3237 struct ring_buffer *rb;
3238 u64 enabled, running;
3242 * compute total_time_enabled, total_time_running
3243 * based on snapshot values taken when the event
3244 * was last scheduled in.
3246 * we cannot simply called update_context_time()
3247 * because of locking issue as we can be called in
3250 calc_timer_values(event, &enabled, &running);
3251 rb = rcu_dereference(event->rb);
3255 userpg = rb->user_page;
3258 * Disable preemption so as to not let the corresponding user-space
3259 * spin too long if we get preempted.
3264 userpg->index = perf_event_index(event);
3265 userpg->offset = perf_event_count(event);
3266 if (event->state == PERF_EVENT_STATE_ACTIVE)
3267 userpg->offset -= local64_read(&event->hw.prev_count);
3269 userpg->time_enabled = enabled +
3270 atomic64_read(&event->child_total_time_enabled);
3272 userpg->time_running = running +
3273 atomic64_read(&event->child_total_time_running);
3282 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
3284 struct perf_event *event = vma->vm_file->private_data;
3285 struct ring_buffer *rb;
3286 int ret = VM_FAULT_SIGBUS;
3288 if (vmf->flags & FAULT_FLAG_MKWRITE) {
3289 if (vmf->pgoff == 0)
3295 rb = rcu_dereference(event->rb);
3299 if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE))
3302 vmf->page = perf_mmap_to_page(rb, vmf->pgoff);
3306 get_page(vmf->page);
3307 vmf->page->mapping = vma->vm_file->f_mapping;
3308 vmf->page->index = vmf->pgoff;
3317 static void ring_buffer_attach(struct perf_event *event,
3318 struct ring_buffer *rb)
3320 unsigned long flags;
3322 if (!list_empty(&event->rb_entry))
3325 spin_lock_irqsave(&rb->event_lock, flags);
3326 if (!list_empty(&event->rb_entry))
3329 list_add(&event->rb_entry, &rb->event_list);
3331 spin_unlock_irqrestore(&rb->event_lock, flags);
3334 static void ring_buffer_detach(struct perf_event *event,
3335 struct ring_buffer *rb)
3337 unsigned long flags;
3339 if (list_empty(&event->rb_entry))
3342 spin_lock_irqsave(&rb->event_lock, flags);
3343 list_del_init(&event->rb_entry);
3344 wake_up_all(&event->waitq);
3345 spin_unlock_irqrestore(&rb->event_lock, flags);
3348 static void ring_buffer_wakeup(struct perf_event *event)
3350 struct ring_buffer *rb;
3353 rb = rcu_dereference(event->rb);
3354 list_for_each_entry_rcu(event, &rb->event_list, rb_entry) {
3355 wake_up_all(&event->waitq);
3360 static void rb_free_rcu(struct rcu_head *rcu_head)
3362 struct ring_buffer *rb;
3364 rb = container_of(rcu_head, struct ring_buffer, rcu_head);
3368 static struct ring_buffer *ring_buffer_get(struct perf_event *event)
3370 struct ring_buffer *rb;
3373 rb = rcu_dereference(event->rb);
3375 if (!atomic_inc_not_zero(&rb->refcount))
3383 static void ring_buffer_put(struct ring_buffer *rb)
3385 struct perf_event *event, *n;
3386 unsigned long flags;
3388 if (!atomic_dec_and_test(&rb->refcount))
3391 spin_lock_irqsave(&rb->event_lock, flags);
3392 list_for_each_entry_safe(event, n, &rb->event_list, rb_entry) {
3393 list_del_init(&event->rb_entry);
3394 wake_up_all(&event->waitq);
3396 spin_unlock_irqrestore(&rb->event_lock, flags);
3398 call_rcu(&rb->rcu_head, rb_free_rcu);
3401 static void perf_mmap_open(struct vm_area_struct *vma)
3403 struct perf_event *event = vma->vm_file->private_data;
3405 atomic_inc(&event->mmap_count);
3408 static void perf_mmap_close(struct vm_area_struct *vma)
3410 struct perf_event *event = vma->vm_file->private_data;
3412 if (atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex)) {
3413 unsigned long size = perf_data_size(event->rb);
3414 struct user_struct *user = event->mmap_user;
3415 struct ring_buffer *rb = event->rb;
3417 atomic_long_sub((size >> PAGE_SHIFT) + 1, &user->locked_vm);
3418 vma->vm_mm->pinned_vm -= event->mmap_locked;
3419 rcu_assign_pointer(event->rb, NULL);
3420 ring_buffer_detach(event, rb);
3421 mutex_unlock(&event->mmap_mutex);
3423 ring_buffer_put(rb);
3428 static const struct vm_operations_struct perf_mmap_vmops = {
3429 .open = perf_mmap_open,
3430 .close = perf_mmap_close,
3431 .fault = perf_mmap_fault,
3432 .page_mkwrite = perf_mmap_fault,
3435 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
3437 struct perf_event *event = file->private_data;
3438 unsigned long user_locked, user_lock_limit;
3439 struct user_struct *user = current_user();
3440 unsigned long locked, lock_limit;
3441 struct ring_buffer *rb;
3442 unsigned long vma_size;
3443 unsigned long nr_pages;
3444 long user_extra, extra;
3445 int ret = 0, flags = 0;
3448 * Don't allow mmap() of inherited per-task counters. This would
3449 * create a performance issue due to all children writing to the
3452 if (event->cpu == -1 && event->attr.inherit)
3455 if (!(vma->vm_flags & VM_SHARED))
3458 vma_size = vma->vm_end - vma->vm_start;
3459 nr_pages = (vma_size / PAGE_SIZE) - 1;
3462 * If we have rb pages ensure they're a power-of-two number, so we
3463 * can do bitmasks instead of modulo.
3465 if (nr_pages != 0 && !is_power_of_2(nr_pages))
3468 if (vma_size != PAGE_SIZE * (1 + nr_pages))
3471 if (vma->vm_pgoff != 0)
3474 WARN_ON_ONCE(event->ctx->parent_ctx);
3475 mutex_lock(&event->mmap_mutex);
3477 if (event->rb->nr_pages == nr_pages)
3478 atomic_inc(&event->rb->refcount);
3484 user_extra = nr_pages + 1;
3485 user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10);
3488 * Increase the limit linearly with more CPUs:
3490 user_lock_limit *= num_online_cpus();
3492 user_locked = atomic_long_read(&user->locked_vm) + user_extra;
3495 if (user_locked > user_lock_limit)
3496 extra = user_locked - user_lock_limit;
3498 lock_limit = rlimit(RLIMIT_MEMLOCK);
3499 lock_limit >>= PAGE_SHIFT;
3500 locked = vma->vm_mm->pinned_vm + extra;
3502 if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() &&
3503 !capable(CAP_IPC_LOCK)) {
3510 if (vma->vm_flags & VM_WRITE)
3511 flags |= RING_BUFFER_WRITABLE;
3513 rb = rb_alloc(nr_pages,
3514 event->attr.watermark ? event->attr.wakeup_watermark : 0,
3521 rcu_assign_pointer(event->rb, rb);
3523 atomic_long_add(user_extra, &user->locked_vm);
3524 event->mmap_locked = extra;
3525 event->mmap_user = get_current_user();
3526 vma->vm_mm->pinned_vm += event->mmap_locked;
3530 atomic_inc(&event->mmap_count);
3531 mutex_unlock(&event->mmap_mutex);
3533 vma->vm_flags |= VM_RESERVED;
3534 vma->vm_ops = &perf_mmap_vmops;
3539 static int perf_fasync(int fd, struct file *filp, int on)
3541 struct inode *inode = filp->f_path.dentry->d_inode;
3542 struct perf_event *event = filp->private_data;
3545 mutex_lock(&inode->i_mutex);
3546 retval = fasync_helper(fd, filp, on, &event->fasync);
3547 mutex_unlock(&inode->i_mutex);
3555 static const struct file_operations perf_fops = {
3556 .llseek = no_llseek,
3557 .release = perf_release,
3560 .unlocked_ioctl = perf_ioctl,
3561 .compat_ioctl = perf_ioctl,
3563 .fasync = perf_fasync,
3569 * If there's data, ensure we set the poll() state and publish everything
3570 * to user-space before waking everybody up.
3573 void perf_event_wakeup(struct perf_event *event)
3575 ring_buffer_wakeup(event);
3577 if (event->pending_kill) {
3578 kill_fasync(&event->fasync, SIGIO, event->pending_kill);
3579 event->pending_kill = 0;
3583 static void perf_pending_event(struct irq_work *entry)
3585 struct perf_event *event = container_of(entry,
3586 struct perf_event, pending);
3588 if (event->pending_disable) {
3589 event->pending_disable = 0;
3590 __perf_event_disable(event);
3593 if (event->pending_wakeup) {
3594 event->pending_wakeup = 0;
3595 perf_event_wakeup(event);
3600 * We assume there is only KVM supporting the callbacks.
3601 * Later on, we might change it to a list if there is
3602 * another virtualization implementation supporting the callbacks.
3604 struct perf_guest_info_callbacks *perf_guest_cbs;
3606 int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
3608 perf_guest_cbs = cbs;
3611 EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks);
3613 int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
3615 perf_guest_cbs = NULL;
3618 EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks);
3620 static void __perf_event_header__init_id(struct perf_event_header *header,
3621 struct perf_sample_data *data,
3622 struct perf_event *event)
3624 u64 sample_type = event->attr.sample_type;
3626 data->type = sample_type;
3627 header->size += event->id_header_size;
3629 if (sample_type & PERF_SAMPLE_TID) {
3630 /* namespace issues */
3631 data->tid_entry.pid = perf_event_pid(event, current);
3632 data->tid_entry.tid = perf_event_tid(event, current);
3635 if (sample_type & PERF_SAMPLE_TIME)
3636 data->time = perf_clock();
3638 if (sample_type & PERF_SAMPLE_ID)
3639 data->id = primary_event_id(event);
3641 if (sample_type & PERF_SAMPLE_STREAM_ID)
3642 data->stream_id = event->id;
3644 if (sample_type & PERF_SAMPLE_CPU) {
3645 data->cpu_entry.cpu = raw_smp_processor_id();
3646 data->cpu_entry.reserved = 0;
3650 void perf_event_header__init_id(struct perf_event_header *header,
3651 struct perf_sample_data *data,
3652 struct perf_event *event)
3654 if (event->attr.sample_id_all)
3655 __perf_event_header__init_id(header, data, event);
3658 static void __perf_event__output_id_sample(struct perf_output_handle *handle,
3659 struct perf_sample_data *data)
3661 u64 sample_type = data->type;
3663 if (sample_type & PERF_SAMPLE_TID)
3664 perf_output_put(handle, data->tid_entry);
3666 if (sample_type & PERF_SAMPLE_TIME)
3667 perf_output_put(handle, data->time);
3669 if (sample_type & PERF_SAMPLE_ID)
3670 perf_output_put(handle, data->id);
3672 if (sample_type & PERF_SAMPLE_STREAM_ID)
3673 perf_output_put(handle, data->stream_id);
3675 if (sample_type & PERF_SAMPLE_CPU)
3676 perf_output_put(handle, data->cpu_entry);
3679 void perf_event__output_id_sample(struct perf_event *event,
3680 struct perf_output_handle *handle,
3681 struct perf_sample_data *sample)
3683 if (event->attr.sample_id_all)
3684 __perf_event__output_id_sample(handle, sample);
3687 static void perf_output_read_one(struct perf_output_handle *handle,
3688 struct perf_event *event,
3689 u64 enabled, u64 running)
3691 u64 read_format = event->attr.read_format;
3695 values[n++] = perf_event_count(event);
3696 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
3697 values[n++] = enabled +
3698 atomic64_read(&event->child_total_time_enabled);
3700 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
3701 values[n++] = running +
3702 atomic64_read(&event->child_total_time_running);
3704 if (read_format & PERF_FORMAT_ID)
3705 values[n++] = primary_event_id(event);
3707 __output_copy(handle, values, n * sizeof(u64));
3711 * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
3713 static void perf_output_read_group(struct perf_output_handle *handle,
3714 struct perf_event *event,
3715 u64 enabled, u64 running)
3717 struct perf_event *leader = event->group_leader, *sub;
3718 u64 read_format = event->attr.read_format;
3722 values[n++] = 1 + leader->nr_siblings;
3724 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3725 values[n++] = enabled;
3727 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3728 values[n++] = running;
3730 if (leader != event)
3731 leader->pmu->read(leader);
3733 values[n++] = perf_event_count(leader);
3734 if (read_format & PERF_FORMAT_ID)
3735 values[n++] = primary_event_id(leader);
3737 __output_copy(handle, values, n * sizeof(u64));
3739 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
3743 sub->pmu->read(sub);
3745 values[n++] = perf_event_count(sub);
3746 if (read_format & PERF_FORMAT_ID)
3747 values[n++] = primary_event_id(sub);
3749 __output_copy(handle, values, n * sizeof(u64));
3753 #define PERF_FORMAT_TOTAL_TIMES (PERF_FORMAT_TOTAL_TIME_ENABLED|\
3754 PERF_FORMAT_TOTAL_TIME_RUNNING)
3756 static void perf_output_read(struct perf_output_handle *handle,
3757 struct perf_event *event)
3759 u64 enabled = 0, running = 0;
3760 u64 read_format = event->attr.read_format;
3763 * compute total_time_enabled, total_time_running
3764 * based on snapshot values taken when the event
3765 * was last scheduled in.
3767 * we cannot simply called update_context_time()
3768 * because of locking issue as we are called in
3771 if (read_format & PERF_FORMAT_TOTAL_TIMES)
3772 calc_timer_values(event, &enabled, &running);
3774 if (event->attr.read_format & PERF_FORMAT_GROUP)
3775 perf_output_read_group(handle, event, enabled, running);
3777 perf_output_read_one(handle, event, enabled, running);
3780 void perf_output_sample(struct perf_output_handle *handle,
3781 struct perf_event_header *header,
3782 struct perf_sample_data *data,
3783 struct perf_event *event)
3785 u64 sample_type = data->type;
3787 perf_output_put(handle, *header);
3789 if (sample_type & PERF_SAMPLE_IP)
3790 perf_output_put(handle, data->ip);
3792 if (sample_type & PERF_SAMPLE_TID)
3793 perf_output_put(handle, data->tid_entry);
3795 if (sample_type & PERF_SAMPLE_TIME)
3796 perf_output_put(handle, data->time);
3798 if (sample_type & PERF_SAMPLE_ADDR)
3799 perf_output_put(handle, data->addr);
3801 if (sample_type & PERF_SAMPLE_ID)
3802 perf_output_put(handle, data->id);
3804 if (sample_type & PERF_SAMPLE_STREAM_ID)
3805 perf_output_put(handle, data->stream_id);
3807 if (sample_type & PERF_SAMPLE_CPU)
3808 perf_output_put(handle, data->cpu_entry);
3810 if (sample_type & PERF_SAMPLE_PERIOD)
3811 perf_output_put(handle, data->period);
3813 if (sample_type & PERF_SAMPLE_READ)
3814 perf_output_read(handle, event);
3816 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3817 if (data->callchain) {
3820 if (data->callchain)
3821 size += data->callchain->nr;
3823 size *= sizeof(u64);
3825 __output_copy(handle, data->callchain, size);
3828 perf_output_put(handle, nr);
3832 if (sample_type & PERF_SAMPLE_RAW) {
3834 perf_output_put(handle, data->raw->size);
3835 __output_copy(handle, data->raw->data,
3842 .size = sizeof(u32),
3845 perf_output_put(handle, raw);
3849 if (!event->attr.watermark) {
3850 int wakeup_events = event->attr.wakeup_events;
3852 if (wakeup_events) {
3853 struct ring_buffer *rb = handle->rb;
3854 int events = local_inc_return(&rb->events);
3856 if (events >= wakeup_events) {
3857 local_sub(wakeup_events, &rb->events);
3858 local_inc(&rb->wakeup);
3864 void perf_prepare_sample(struct perf_event_header *header,
3865 struct perf_sample_data *data,
3866 struct perf_event *event,
3867 struct pt_regs *regs)
3869 u64 sample_type = event->attr.sample_type;
3871 header->type = PERF_RECORD_SAMPLE;
3872 header->size = sizeof(*header) + event->header_size;
3875 header->misc |= perf_misc_flags(regs);
3877 __perf_event_header__init_id(header, data, event);
3879 if (sample_type & PERF_SAMPLE_IP)
3880 data->ip = perf_instruction_pointer(regs);
3882 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3885 data->callchain = perf_callchain(regs);
3887 if (data->callchain)
3888 size += data->callchain->nr;
3890 header->size += size * sizeof(u64);
3893 if (sample_type & PERF_SAMPLE_RAW) {
3894 int size = sizeof(u32);
3897 size += data->raw->size;
3899 size += sizeof(u32);
3901 WARN_ON_ONCE(size & (sizeof(u64)-1));
3902 header->size += size;
3906 static void perf_event_output(struct perf_event *event,
3907 struct perf_sample_data *data,
3908 struct pt_regs *regs)
3910 struct perf_output_handle handle;
3911 struct perf_event_header header;
3913 /* protect the callchain buffers */
3916 perf_prepare_sample(&header, data, event, regs);
3918 if (perf_output_begin(&handle, event, header.size))
3921 perf_output_sample(&handle, &header, data, event);
3923 perf_output_end(&handle);
3933 struct perf_read_event {
3934 struct perf_event_header header;
3941 perf_event_read_event(struct perf_event *event,
3942 struct task_struct *task)
3944 struct perf_output_handle handle;
3945 struct perf_sample_data sample;
3946 struct perf_read_event read_event = {
3948 .type = PERF_RECORD_READ,
3950 .size = sizeof(read_event) + event->read_size,
3952 .pid = perf_event_pid(event, task),
3953 .tid = perf_event_tid(event, task),
3957 perf_event_header__init_id(&read_event.header, &sample, event);
3958 ret = perf_output_begin(&handle, event, read_event.header.size);
3962 perf_output_put(&handle, read_event);
3963 perf_output_read(&handle, event);
3964 perf_event__output_id_sample(event, &handle, &sample);
3966 perf_output_end(&handle);
3970 * task tracking -- fork/exit
3972 * enabled by: attr.comm | attr.mmap | attr.mmap_data | attr.task
3975 struct perf_task_event {
3976 struct task_struct *task;
3977 struct perf_event_context *task_ctx;
3980 struct perf_event_header header;
3990 static void perf_event_task_output(struct perf_event *event,
3991 struct perf_task_event *task_event)
3993 struct perf_output_handle handle;
3994 struct perf_sample_data sample;
3995 struct task_struct *task = task_event->task;
3996 int ret, size = task_event->event_id.header.size;
3998 perf_event_header__init_id(&task_event->event_id.header, &sample, event);
4000 ret = perf_output_begin(&handle, event,
4001 task_event->event_id.header.size);
4005 task_event->event_id.pid = perf_event_pid(event, task);
4006 task_event->event_id.ppid = perf_event_pid(event, current);
4008 task_event->event_id.tid = perf_event_tid(event, task);
4009 task_event->event_id.ptid = perf_event_tid(event, current);
4011 perf_output_put(&handle, task_event->event_id);
4013 perf_event__output_id_sample(event, &handle, &sample);
4015 perf_output_end(&handle);
4017 task_event->event_id.header.size = size;
4020 static int perf_event_task_match(struct perf_event *event)
4022 if (event->state < PERF_EVENT_STATE_INACTIVE)
4025 if (!event_filter_match(event))
4028 if (event->attr.comm || event->attr.mmap ||
4029 event->attr.mmap_data || event->attr.task)
4035 static void perf_event_task_ctx(struct perf_event_context *ctx,
4036 struct perf_task_event *task_event)
4038 struct perf_event *event;
4040 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
4041 if (perf_event_task_match(event))
4042 perf_event_task_output(event, task_event);
4046 static void perf_event_task_event(struct perf_task_event *task_event)
4048 struct perf_cpu_context *cpuctx;
4049 struct perf_event_context *ctx;
4054 list_for_each_entry_rcu(pmu, &pmus, entry) {
4055 cpuctx = get_cpu_ptr(pmu->pmu_cpu_context);
4056 if (cpuctx->active_pmu != pmu)
4058 perf_event_task_ctx(&cpuctx->ctx, task_event);
4060 ctx = task_event->task_ctx;
4062 ctxn = pmu->task_ctx_nr;
4065 ctx = rcu_dereference(current->perf_event_ctxp[ctxn]);
4068 perf_event_task_ctx(ctx, task_event);
4070 put_cpu_ptr(pmu->pmu_cpu_context);
4075 static void perf_event_task(struct task_struct *task,
4076 struct perf_event_context *task_ctx,
4079 struct perf_task_event task_event;
4081 if (!atomic_read(&nr_comm_events) &&
4082 !atomic_read(&nr_mmap_events) &&
4083 !atomic_read(&nr_task_events))
4086 task_event = (struct perf_task_event){
4088 .task_ctx = task_ctx,
4091 .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT,
4093 .size = sizeof(task_event.event_id),
4099 .time = perf_clock(),
4103 perf_event_task_event(&task_event);
4106 void perf_event_fork(struct task_struct *task)
4108 perf_event_task(task, NULL, 1);
4115 struct perf_comm_event {
4116 struct task_struct *task;
4121 struct perf_event_header header;
4128 static void perf_event_comm_output(struct perf_event *event,
4129 struct perf_comm_event *comm_event)
4131 struct perf_output_handle handle;
4132 struct perf_sample_data sample;
4133 int size = comm_event->event_id.header.size;
4136 perf_event_header__init_id(&comm_event->event_id.header, &sample, event);
4137 ret = perf_output_begin(&handle, event,
4138 comm_event->event_id.header.size);
4143 comm_event->event_id.pid = perf_event_pid(event, comm_event->task);
4144 comm_event->event_id.tid = perf_event_tid(event, comm_event->task);
4146 perf_output_put(&handle, comm_event->event_id);
4147 __output_copy(&handle, comm_event->comm,
4148 comm_event->comm_size);
4150 perf_event__output_id_sample(event, &handle, &sample);
4152 perf_output_end(&handle);
4154 comm_event->event_id.header.size = size;
4157 static int perf_event_comm_match(struct perf_event *event)
4159 if (event->state < PERF_EVENT_STATE_INACTIVE)
4162 if (!event_filter_match(event))
4165 if (event->attr.comm)
4171 static void perf_event_comm_ctx(struct perf_event_context *ctx,
4172 struct perf_comm_event *comm_event)
4174 struct perf_event *event;
4176 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
4177 if (perf_event_comm_match(event))
4178 perf_event_comm_output(event, comm_event);
4182 static void perf_event_comm_event(struct perf_comm_event *comm_event)
4184 struct perf_cpu_context *cpuctx;
4185 struct perf_event_context *ctx;
4186 char comm[TASK_COMM_LEN];
4191 memset(comm, 0, sizeof(comm));
4192 strlcpy(comm, comm_event->task->comm, sizeof(comm));
4193 size = ALIGN(strlen(comm)+1, sizeof(u64));
4195 comm_event->comm = comm;
4196 comm_event->comm_size = size;
4198 comm_event->event_id.header.size = sizeof(comm_event->event_id) + size;
4200 list_for_each_entry_rcu(pmu, &pmus, entry) {
4201 cpuctx = get_cpu_ptr(pmu->pmu_cpu_context);
4202 if (cpuctx->active_pmu != pmu)
4204 perf_event_comm_ctx(&cpuctx->ctx, comm_event);
4206 ctxn = pmu->task_ctx_nr;
4210 ctx = rcu_dereference(current->perf_event_ctxp[ctxn]);
4212 perf_event_comm_ctx(ctx, comm_event);
4214 put_cpu_ptr(pmu->pmu_cpu_context);
4219 void perf_event_comm(struct task_struct *task)
4221 struct perf_comm_event comm_event;
4222 struct perf_event_context *ctx;
4225 for_each_task_context_nr(ctxn) {
4226 ctx = task->perf_event_ctxp[ctxn];
4230 perf_event_enable_on_exec(ctx);
4233 if (!atomic_read(&nr_comm_events))
4236 comm_event = (struct perf_comm_event){
4242 .type = PERF_RECORD_COMM,
4251 perf_event_comm_event(&comm_event);
4258 struct perf_mmap_event {
4259 struct vm_area_struct *vma;
4261 const char *file_name;
4265 struct perf_event_header header;
4275 static void perf_event_mmap_output(struct perf_event *event,
4276 struct perf_mmap_event *mmap_event)
4278 struct perf_output_handle handle;
4279 struct perf_sample_data sample;
4280 int size = mmap_event->event_id.header.size;
4283 perf_event_header__init_id(&mmap_event->event_id.header, &sample, event);
4284 ret = perf_output_begin(&handle, event,
4285 mmap_event->event_id.header.size);
4289 mmap_event->event_id.pid = perf_event_pid(event, current);
4290 mmap_event->event_id.tid = perf_event_tid(event, current);
4292 perf_output_put(&handle, mmap_event->event_id);
4293 __output_copy(&handle, mmap_event->file_name,
4294 mmap_event->file_size);
4296 perf_event__output_id_sample(event, &handle, &sample);
4298 perf_output_end(&handle);
4300 mmap_event->event_id.header.size = size;
4303 static int perf_event_mmap_match(struct perf_event *event,
4304 struct perf_mmap_event *mmap_event,
4307 if (event->state < PERF_EVENT_STATE_INACTIVE)
4310 if (!event_filter_match(event))
4313 if ((!executable && event->attr.mmap_data) ||
4314 (executable && event->attr.mmap))
4320 static void perf_event_mmap_ctx(struct perf_event_context *ctx,
4321 struct perf_mmap_event *mmap_event,
4324 struct perf_event *event;
4326 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
4327 if (perf_event_mmap_match(event, mmap_event, executable))
4328 perf_event_mmap_output(event, mmap_event);
4332 static void perf_event_mmap_event(struct perf_mmap_event *mmap_event)
4334 struct perf_cpu_context *cpuctx;
4335 struct perf_event_context *ctx;
4336 struct vm_area_struct *vma = mmap_event->vma;
4337 struct file *file = vma->vm_file;
4345 memset(tmp, 0, sizeof(tmp));
4349 * d_path works from the end of the rb backwards, so we
4350 * need to add enough zero bytes after the string to handle
4351 * the 64bit alignment we do later.
4353 buf = kzalloc(PATH_MAX + sizeof(u64), GFP_KERNEL);
4355 name = strncpy(tmp, "//enomem", sizeof(tmp));
4358 name = d_path(&file->f_path, buf, PATH_MAX);
4360 name = strncpy(tmp, "//toolong", sizeof(tmp));
4364 if (arch_vma_name(mmap_event->vma)) {
4365 name = strncpy(tmp, arch_vma_name(mmap_event->vma),
4371 name = strncpy(tmp, "[vdso]", sizeof(tmp));
4373 } else if (vma->vm_start <= vma->vm_mm->start_brk &&
4374 vma->vm_end >= vma->vm_mm->brk) {
4375 name = strncpy(tmp, "[heap]", sizeof(tmp));
4377 } else if (vma->vm_start <= vma->vm_mm->start_stack &&
4378 vma->vm_end >= vma->vm_mm->start_stack) {
4379 name = strncpy(tmp, "[stack]", sizeof(tmp));
4383 name = strncpy(tmp, "//anon", sizeof(tmp));
4388 size = ALIGN(strlen(name)+1, sizeof(u64));
4390 mmap_event->file_name = name;
4391 mmap_event->file_size = size;
4393 mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size;
4396 list_for_each_entry_rcu(pmu, &pmus, entry) {
4397 cpuctx = get_cpu_ptr(pmu->pmu_cpu_context);
4398 if (cpuctx->active_pmu != pmu)
4400 perf_event_mmap_ctx(&cpuctx->ctx, mmap_event,
4401 vma->vm_flags & VM_EXEC);
4403 ctxn = pmu->task_ctx_nr;
4407 ctx = rcu_dereference(current->perf_event_ctxp[ctxn]);
4409 perf_event_mmap_ctx(ctx, mmap_event,
4410 vma->vm_flags & VM_EXEC);
4413 put_cpu_ptr(pmu->pmu_cpu_context);
4420 void perf_event_mmap(struct vm_area_struct *vma)
4422 struct perf_mmap_event mmap_event;
4424 if (!atomic_read(&nr_mmap_events))
4427 mmap_event = (struct perf_mmap_event){
4433 .type = PERF_RECORD_MMAP,
4434 .misc = PERF_RECORD_MISC_USER,
4439 .start = vma->vm_start,
4440 .len = vma->vm_end - vma->vm_start,
4441 .pgoff = (u64)vma->vm_pgoff << PAGE_SHIFT,
4445 perf_event_mmap_event(&mmap_event);
4449 * IRQ throttle logging
4452 static void perf_log_throttle(struct perf_event *event, int enable)
4454 struct perf_output_handle handle;
4455 struct perf_sample_data sample;
4459 struct perf_event_header header;
4463 } throttle_event = {
4465 .type = PERF_RECORD_THROTTLE,
4467 .size = sizeof(throttle_event),
4469 .time = perf_clock(),
4470 .id = primary_event_id(event),
4471 .stream_id = event->id,
4475 throttle_event.header.type = PERF_RECORD_UNTHROTTLE;
4477 perf_event_header__init_id(&throttle_event.header, &sample, event);
4479 ret = perf_output_begin(&handle, event,
4480 throttle_event.header.size);
4484 perf_output_put(&handle, throttle_event);
4485 perf_event__output_id_sample(event, &handle, &sample);
4486 perf_output_end(&handle);
4490 * Generic event overflow handling, sampling.
4493 static int __perf_event_overflow(struct perf_event *event,
4494 int throttle, struct perf_sample_data *data,
4495 struct pt_regs *regs)
4497 int events = atomic_read(&event->event_limit);
4498 struct hw_perf_event *hwc = &event->hw;
4502 * Non-sampling counters might still use the PMI to fold short
4503 * hardware counters, ignore those.
4505 if (unlikely(!is_sampling_event(event)))
4508 if (unlikely(hwc->interrupts >= max_samples_per_tick)) {
4510 hwc->interrupts = MAX_INTERRUPTS;
4511 perf_log_throttle(event, 0);
4517 if (event->attr.freq) {
4518 u64 now = perf_clock();
4519 s64 delta = now - hwc->freq_time_stamp;
4521 hwc->freq_time_stamp = now;
4523 if (delta > 0 && delta < 2*TICK_NSEC)
4524 perf_adjust_period(event, delta, hwc->last_period);
4528 * XXX event_limit might not quite work as expected on inherited
4532 event->pending_kill = POLL_IN;
4533 if (events && atomic_dec_and_test(&event->event_limit)) {
4535 event->pending_kill = POLL_HUP;
4536 event->pending_disable = 1;
4537 irq_work_queue(&event->pending);
4540 if (event->overflow_handler)
4541 event->overflow_handler(event, data, regs);
4543 perf_event_output(event, data, regs);
4545 if (event->fasync && event->pending_kill) {
4546 event->pending_wakeup = 1;
4547 irq_work_queue(&event->pending);
4553 int perf_event_overflow(struct perf_event *event,
4554 struct perf_sample_data *data,
4555 struct pt_regs *regs)
4557 return __perf_event_overflow(event, 1, data, regs);
4561 * Generic software event infrastructure
4564 struct swevent_htable {
4565 struct swevent_hlist *swevent_hlist;
4566 struct mutex hlist_mutex;
4569 /* Recursion avoidance in each contexts */
4570 int recursion[PERF_NR_CONTEXTS];
4573 static DEFINE_PER_CPU(struct swevent_htable, swevent_htable);
4576 * We directly increment event->count and keep a second value in
4577 * event->hw.period_left to count intervals. This period event
4578 * is kept in the range [-sample_period, 0] so that we can use the
4582 static u64 perf_swevent_set_period(struct perf_event *event)
4584 struct hw_perf_event *hwc = &event->hw;
4585 u64 period = hwc->last_period;
4589 hwc->last_period = hwc->sample_period;
4592 old = val = local64_read(&hwc->period_left);
4596 nr = div64_u64(period + val, period);
4597 offset = nr * period;
4599 if (local64_cmpxchg(&hwc->period_left, old, val) != old)
4605 static void perf_swevent_overflow(struct perf_event *event, u64 overflow,
4606 struct perf_sample_data *data,
4607 struct pt_regs *regs)
4609 struct hw_perf_event *hwc = &event->hw;
4613 overflow = perf_swevent_set_period(event);
4615 if (hwc->interrupts == MAX_INTERRUPTS)
4618 for (; overflow; overflow--) {
4619 if (__perf_event_overflow(event, throttle,
4622 * We inhibit the overflow from happening when
4623 * hwc->interrupts == MAX_INTERRUPTS.
4631 static void perf_swevent_event(struct perf_event *event, u64 nr,
4632 struct perf_sample_data *data,
4633 struct pt_regs *regs)
4635 struct hw_perf_event *hwc = &event->hw;
4637 local64_add(nr, &event->count);
4642 if (!is_sampling_event(event))
4645 if ((event->attr.sample_type & PERF_SAMPLE_PERIOD) && !event->attr.freq) {
4647 return perf_swevent_overflow(event, 1, data, regs);
4649 data->period = event->hw.last_period;
4651 if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq)
4652 return perf_swevent_overflow(event, 1, data, regs);
4654 if (local64_add_negative(nr, &hwc->period_left))
4657 perf_swevent_overflow(event, 0, data, regs);
4660 static int perf_exclude_event(struct perf_event *event,
4661 struct pt_regs *regs)
4663 if (event->hw.state & PERF_HES_STOPPED)
4667 if (event->attr.exclude_user && user_mode(regs))
4670 if (event->attr.exclude_kernel && !user_mode(regs))
4677 static int perf_swevent_match(struct perf_event *event,
4678 enum perf_type_id type,
4680 struct perf_sample_data *data,
4681 struct pt_regs *regs)
4683 if (event->attr.type != type)
4686 if (event->attr.config != event_id)
4689 if (perf_exclude_event(event, regs))
4695 static inline u64 swevent_hash(u64 type, u32 event_id)
4697 u64 val = event_id | (type << 32);
4699 return hash_64(val, SWEVENT_HLIST_BITS);
4702 static inline struct hlist_head *
4703 __find_swevent_head(struct swevent_hlist *hlist, u64 type, u32 event_id)
4705 u64 hash = swevent_hash(type, event_id);
4707 return &hlist->heads[hash];
4710 /* For the read side: events when they trigger */
4711 static inline struct hlist_head *
4712 find_swevent_head_rcu(struct swevent_htable *swhash, u64 type, u32 event_id)
4714 struct swevent_hlist *hlist;
4716 hlist = rcu_dereference(swhash->swevent_hlist);
4720 return __find_swevent_head(hlist, type, event_id);
4723 /* For the event head insertion and removal in the hlist */
4724 static inline struct hlist_head *
4725 find_swevent_head(struct swevent_htable *swhash, struct perf_event *event)
4727 struct swevent_hlist *hlist;
4728 u32 event_id = event->attr.config;
4729 u64 type = event->attr.type;
4732 * Event scheduling is always serialized against hlist allocation
4733 * and release. Which makes the protected version suitable here.
4734 * The context lock guarantees that.
4736 hlist = rcu_dereference_protected(swhash->swevent_hlist,
4737 lockdep_is_held(&event->ctx->lock));
4741 return __find_swevent_head(hlist, type, event_id);
4744 static void do_perf_sw_event(enum perf_type_id type, u32 event_id,
4746 struct perf_sample_data *data,
4747 struct pt_regs *regs)
4749 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
4750 struct perf_event *event;
4751 struct hlist_node *node;
4752 struct hlist_head *head;
4755 head = find_swevent_head_rcu(swhash, type, event_id);
4759 hlist_for_each_entry_rcu(event, node, head, hlist_entry) {
4760 if (perf_swevent_match(event, type, event_id, data, regs))
4761 perf_swevent_event(event, nr, data, regs);
4767 int perf_swevent_get_recursion_context(void)
4769 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
4771 return get_recursion_context(swhash->recursion);
4773 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context);
4775 inline void perf_swevent_put_recursion_context(int rctx)
4777 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
4779 put_recursion_context(swhash->recursion, rctx);
4782 void __perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)
4784 struct perf_sample_data data;
4787 preempt_disable_notrace();
4788 rctx = perf_swevent_get_recursion_context();
4792 perf_sample_data_init(&data, addr);
4794 do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, &data, regs);
4796 perf_swevent_put_recursion_context(rctx);
4797 preempt_enable_notrace();
4800 static void perf_swevent_read(struct perf_event *event)
4804 static int perf_swevent_add(struct perf_event *event, int flags)
4806 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
4807 struct hw_perf_event *hwc = &event->hw;
4808 struct hlist_head *head;
4810 if (is_sampling_event(event)) {
4811 hwc->last_period = hwc->sample_period;
4812 perf_swevent_set_period(event);
4815 hwc->state = !(flags & PERF_EF_START);
4817 head = find_swevent_head(swhash, event);
4818 if (WARN_ON_ONCE(!head))
4821 hlist_add_head_rcu(&event->hlist_entry, head);
4826 static void perf_swevent_del(struct perf_event *event, int flags)
4828 hlist_del_rcu(&event->hlist_entry);
4831 static void perf_swevent_start(struct perf_event *event, int flags)
4833 event->hw.state = 0;
4836 static void perf_swevent_stop(struct perf_event *event, int flags)
4838 event->hw.state = PERF_HES_STOPPED;
4841 /* Deref the hlist from the update side */
4842 static inline struct swevent_hlist *
4843 swevent_hlist_deref(struct swevent_htable *swhash)
4845 return rcu_dereference_protected(swhash->swevent_hlist,
4846 lockdep_is_held(&swhash->hlist_mutex));
4849 static void swevent_hlist_release(struct swevent_htable *swhash)
4851 struct swevent_hlist *hlist = swevent_hlist_deref(swhash);
4856 rcu_assign_pointer(swhash->swevent_hlist, NULL);
4857 kfree_rcu(hlist, rcu_head);
4860 static void swevent_hlist_put_cpu(struct perf_event *event, int cpu)
4862 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
4864 mutex_lock(&swhash->hlist_mutex);
4866 if (!--swhash->hlist_refcount)
4867 swevent_hlist_release(swhash);
4869 mutex_unlock(&swhash->hlist_mutex);
4872 static void swevent_hlist_put(struct perf_event *event)
4876 if (event->cpu != -1) {
4877 swevent_hlist_put_cpu(event, event->cpu);
4881 for_each_possible_cpu(cpu)
4882 swevent_hlist_put_cpu(event, cpu);
4885 static int swevent_hlist_get_cpu(struct perf_event *event, int cpu)
4887 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
4890 mutex_lock(&swhash->hlist_mutex);
4892 if (!swevent_hlist_deref(swhash) && cpu_online(cpu)) {
4893 struct swevent_hlist *hlist;
4895 hlist = kzalloc(sizeof(*hlist), GFP_KERNEL);
4900 rcu_assign_pointer(swhash->swevent_hlist, hlist);
4902 swhash->hlist_refcount++;
4904 mutex_unlock(&swhash->hlist_mutex);
4909 static int swevent_hlist_get(struct perf_event *event)
4912 int cpu, failed_cpu;
4914 if (event->cpu != -1)
4915 return swevent_hlist_get_cpu(event, event->cpu);
4918 for_each_possible_cpu(cpu) {
4919 err = swevent_hlist_get_cpu(event, cpu);
4929 for_each_possible_cpu(cpu) {
4930 if (cpu == failed_cpu)
4932 swevent_hlist_put_cpu(event, cpu);
4939 struct jump_label_key perf_swevent_enabled[PERF_COUNT_SW_MAX];
4941 static void sw_perf_event_destroy(struct perf_event *event)
4943 u64 event_id = event->attr.config;
4945 WARN_ON(event->parent);
4947 jump_label_dec(&perf_swevent_enabled[event_id]);
4948 swevent_hlist_put(event);
4951 static int perf_swevent_init(struct perf_event *event)
4953 int event_id = event->attr.config;
4955 if (event->attr.type != PERF_TYPE_SOFTWARE)
4959 case PERF_COUNT_SW_CPU_CLOCK:
4960 case PERF_COUNT_SW_TASK_CLOCK:
4967 if (event_id >= PERF_COUNT_SW_MAX)
4970 if (!event->parent) {
4973 err = swevent_hlist_get(event);
4977 jump_label_inc(&perf_swevent_enabled[event_id]);
4978 event->destroy = sw_perf_event_destroy;
4984 static struct pmu perf_swevent = {
4985 .task_ctx_nr = perf_sw_context,
4987 .event_init = perf_swevent_init,
4988 .add = perf_swevent_add,
4989 .del = perf_swevent_del,
4990 .start = perf_swevent_start,
4991 .stop = perf_swevent_stop,
4992 .read = perf_swevent_read,
4995 #ifdef CONFIG_EVENT_TRACING
4997 static int perf_tp_filter_match(struct perf_event *event,
4998 struct perf_sample_data *data)
5000 void *record = data->raw->data;
5002 if (likely(!event->filter) || filter_match_preds(event->filter, record))
5007 static int perf_tp_event_match(struct perf_event *event,
5008 struct perf_sample_data *data,
5009 struct pt_regs *regs)
5011 if (event->hw.state & PERF_HES_STOPPED)
5014 * All tracepoints are from kernel-space.
5016 if (event->attr.exclude_kernel)
5019 if (!perf_tp_filter_match(event, data))
5025 void perf_tp_event(u64 addr, u64 count, void *record, int entry_size,
5026 struct pt_regs *regs, struct hlist_head *head, int rctx)
5028 struct perf_sample_data data;
5029 struct perf_event *event;
5030 struct hlist_node *node;
5032 struct perf_raw_record raw = {
5037 perf_sample_data_init(&data, addr);
5040 hlist_for_each_entry_rcu(event, node, head, hlist_entry) {
5041 if (perf_tp_event_match(event, &data, regs))
5042 perf_swevent_event(event, count, &data, regs);
5045 perf_swevent_put_recursion_context(rctx);
5047 EXPORT_SYMBOL_GPL(perf_tp_event);
5049 static void tp_perf_event_destroy(struct perf_event *event)
5051 perf_trace_destroy(event);
5054 static int perf_tp_event_init(struct perf_event *event)
5058 if (event->attr.type != PERF_TYPE_TRACEPOINT)
5061 err = perf_trace_init(event);
5065 event->destroy = tp_perf_event_destroy;
5070 static struct pmu perf_tracepoint = {
5071 .task_ctx_nr = perf_sw_context,
5073 .event_init = perf_tp_event_init,
5074 .add = perf_trace_add,
5075 .del = perf_trace_del,
5076 .start = perf_swevent_start,
5077 .stop = perf_swevent_stop,
5078 .read = perf_swevent_read,
5081 static inline void perf_tp_register(void)
5083 perf_pmu_register(&perf_tracepoint, "tracepoint", PERF_TYPE_TRACEPOINT);
5086 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
5091 if (event->attr.type != PERF_TYPE_TRACEPOINT)
5094 filter_str = strndup_user(arg, PAGE_SIZE);
5095 if (IS_ERR(filter_str))
5096 return PTR_ERR(filter_str);
5098 ret = ftrace_profile_set_filter(event, event->attr.config, filter_str);
5104 static void perf_event_free_filter(struct perf_event *event)
5106 ftrace_profile_free_filter(event);
5111 static inline void perf_tp_register(void)
5115 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
5120 static void perf_event_free_filter(struct perf_event *event)
5124 #endif /* CONFIG_EVENT_TRACING */
5126 #ifdef CONFIG_HAVE_HW_BREAKPOINT
5127 void perf_bp_event(struct perf_event *bp, void *data)
5129 struct perf_sample_data sample;
5130 struct pt_regs *regs = data;
5132 perf_sample_data_init(&sample, bp->attr.bp_addr);
5134 if (!bp->hw.state && !perf_exclude_event(bp, regs))
5135 perf_swevent_event(bp, 1, &sample, regs);
5140 * hrtimer based swevent callback
5143 static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer)
5145 enum hrtimer_restart ret = HRTIMER_RESTART;
5146 struct perf_sample_data data;
5147 struct pt_regs *regs;
5148 struct perf_event *event;
5151 event = container_of(hrtimer, struct perf_event, hw.hrtimer);
5153 if (event->state != PERF_EVENT_STATE_ACTIVE)
5154 return HRTIMER_NORESTART;
5156 event->pmu->read(event);
5158 perf_sample_data_init(&data, 0);
5159 data.period = event->hw.last_period;
5160 regs = get_irq_regs();
5162 if (regs && !perf_exclude_event(event, regs)) {
5163 if (!(event->attr.exclude_idle && current->pid == 0))
5164 if (perf_event_overflow(event, &data, regs))
5165 ret = HRTIMER_NORESTART;
5168 period = max_t(u64, 10000, event->hw.sample_period);
5169 hrtimer_forward_now(hrtimer, ns_to_ktime(period));
5174 static void perf_swevent_start_hrtimer(struct perf_event *event)
5176 struct hw_perf_event *hwc = &event->hw;
5179 if (!is_sampling_event(event))
5182 period = local64_read(&hwc->period_left);
5187 local64_set(&hwc->period_left, 0);
5189 period = max_t(u64, 10000, hwc->sample_period);
5191 __hrtimer_start_range_ns(&hwc->hrtimer,
5192 ns_to_ktime(period), 0,
5193 HRTIMER_MODE_REL_PINNED, 0);
5196 static void perf_swevent_cancel_hrtimer(struct perf_event *event)
5198 struct hw_perf_event *hwc = &event->hw;
5200 if (is_sampling_event(event)) {
5201 ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer);
5202 local64_set(&hwc->period_left, ktime_to_ns(remaining));
5204 hrtimer_cancel(&hwc->hrtimer);
5208 static void perf_swevent_init_hrtimer(struct perf_event *event)
5210 struct hw_perf_event *hwc = &event->hw;
5212 if (!is_sampling_event(event))
5215 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
5216 hwc->hrtimer.function = perf_swevent_hrtimer;
5219 * Since hrtimers have a fixed rate, we can do a static freq->period
5220 * mapping and avoid the whole period adjust feedback stuff.
5222 if (event->attr.freq) {
5223 long freq = event->attr.sample_freq;
5225 event->attr.sample_period = NSEC_PER_SEC / freq;
5226 hwc->sample_period = event->attr.sample_period;
5227 local64_set(&hwc->period_left, hwc->sample_period);
5228 event->attr.freq = 0;
5233 * Software event: cpu wall time clock
5236 static void cpu_clock_event_update(struct perf_event *event)
5241 now = local_clock();
5242 prev = local64_xchg(&event->hw.prev_count, now);
5243 local64_add(now - prev, &event->count);
5246 static void cpu_clock_event_start(struct perf_event *event, int flags)
5248 local64_set(&event->hw.prev_count, local_clock());
5249 perf_swevent_start_hrtimer(event);
5252 static void cpu_clock_event_stop(struct perf_event *event, int flags)
5254 perf_swevent_cancel_hrtimer(event);
5255 cpu_clock_event_update(event);
5258 static int cpu_clock_event_add(struct perf_event *event, int flags)
5260 if (flags & PERF_EF_START)
5261 cpu_clock_event_start(event, flags);
5266 static void cpu_clock_event_del(struct perf_event *event, int flags)
5268 cpu_clock_event_stop(event, flags);
5271 static void cpu_clock_event_read(struct perf_event *event)
5273 cpu_clock_event_update(event);
5276 static int cpu_clock_event_init(struct perf_event *event)
5278 if (event->attr.type != PERF_TYPE_SOFTWARE)
5281 if (event->attr.config != PERF_COUNT_SW_CPU_CLOCK)
5284 perf_swevent_init_hrtimer(event);
5289 static struct pmu perf_cpu_clock = {
5290 .task_ctx_nr = perf_sw_context,
5292 .event_init = cpu_clock_event_init,
5293 .add = cpu_clock_event_add,
5294 .del = cpu_clock_event_del,
5295 .start = cpu_clock_event_start,
5296 .stop = cpu_clock_event_stop,
5297 .read = cpu_clock_event_read,
5301 * Software event: task time clock
5304 static void task_clock_event_update(struct perf_event *event, u64 now)
5309 prev = local64_xchg(&event->hw.prev_count, now);
5311 local64_add(delta, &event->count);
5314 static void task_clock_event_start(struct perf_event *event, int flags)
5316 local64_set(&event->hw.prev_count, event->ctx->time);
5317 perf_swevent_start_hrtimer(event);
5320 static void task_clock_event_stop(struct perf_event *event, int flags)
5322 perf_swevent_cancel_hrtimer(event);
5323 task_clock_event_update(event, event->ctx->time);
5326 static int task_clock_event_add(struct perf_event *event, int flags)
5328 if (flags & PERF_EF_START)
5329 task_clock_event_start(event, flags);
5334 static void task_clock_event_del(struct perf_event *event, int flags)
5336 task_clock_event_stop(event, PERF_EF_UPDATE);
5339 static void task_clock_event_read(struct perf_event *event)
5341 u64 now = perf_clock();
5342 u64 delta = now - event->ctx->timestamp;
5343 u64 time = event->ctx->time + delta;
5345 task_clock_event_update(event, time);
5348 static int task_clock_event_init(struct perf_event *event)
5350 if (event->attr.type != PERF_TYPE_SOFTWARE)
5353 if (event->attr.config != PERF_COUNT_SW_TASK_CLOCK)
5356 perf_swevent_init_hrtimer(event);
5361 static struct pmu perf_task_clock = {
5362 .task_ctx_nr = perf_sw_context,
5364 .event_init = task_clock_event_init,
5365 .add = task_clock_event_add,
5366 .del = task_clock_event_del,
5367 .start = task_clock_event_start,
5368 .stop = task_clock_event_stop,
5369 .read = task_clock_event_read,
5372 static void perf_pmu_nop_void(struct pmu *pmu)
5376 static int perf_pmu_nop_int(struct pmu *pmu)
5381 static void perf_pmu_start_txn(struct pmu *pmu)
5383 perf_pmu_disable(pmu);
5386 static int perf_pmu_commit_txn(struct pmu *pmu)
5388 perf_pmu_enable(pmu);
5392 static void perf_pmu_cancel_txn(struct pmu *pmu)
5394 perf_pmu_enable(pmu);
5398 * Ensures all contexts with the same task_ctx_nr have the same
5399 * pmu_cpu_context too.
5401 static void *find_pmu_context(int ctxn)
5408 list_for_each_entry(pmu, &pmus, entry) {
5409 if (pmu->task_ctx_nr == ctxn)
5410 return pmu->pmu_cpu_context;
5416 static void update_pmu_context(struct pmu *pmu, struct pmu *old_pmu)
5420 for_each_possible_cpu(cpu) {
5421 struct perf_cpu_context *cpuctx;
5423 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
5425 if (cpuctx->active_pmu == old_pmu)
5426 cpuctx->active_pmu = pmu;
5430 static void free_pmu_context(struct pmu *pmu)
5434 mutex_lock(&pmus_lock);
5436 * Like a real lame refcount.
5438 list_for_each_entry(i, &pmus, entry) {
5439 if (i->pmu_cpu_context == pmu->pmu_cpu_context) {
5440 update_pmu_context(i, pmu);
5445 free_percpu(pmu->pmu_cpu_context);
5447 mutex_unlock(&pmus_lock);
5449 static struct idr pmu_idr;
5452 type_show(struct device *dev, struct device_attribute *attr, char *page)
5454 struct pmu *pmu = dev_get_drvdata(dev);
5456 return snprintf(page, PAGE_SIZE-1, "%d\n", pmu->type);
5459 static struct device_attribute pmu_dev_attrs[] = {
5464 static int pmu_bus_running;
5465 static struct bus_type pmu_bus = {
5466 .name = "event_source",
5467 .dev_attrs = pmu_dev_attrs,
5470 static void pmu_dev_release(struct device *dev)
5475 static int pmu_dev_alloc(struct pmu *pmu)
5479 pmu->dev = kzalloc(sizeof(struct device), GFP_KERNEL);
5483 device_initialize(pmu->dev);
5484 ret = dev_set_name(pmu->dev, "%s", pmu->name);
5488 dev_set_drvdata(pmu->dev, pmu);
5489 pmu->dev->bus = &pmu_bus;
5490 pmu->dev->release = pmu_dev_release;
5491 ret = device_add(pmu->dev);
5499 put_device(pmu->dev);
5503 static struct lock_class_key cpuctx_mutex;
5504 static struct lock_class_key cpuctx_lock;
5506 int perf_pmu_register(struct pmu *pmu, char *name, int type)
5510 mutex_lock(&pmus_lock);
5512 pmu->pmu_disable_count = alloc_percpu(int);
5513 if (!pmu->pmu_disable_count)
5522 int err = idr_pre_get(&pmu_idr, GFP_KERNEL);
5526 err = idr_get_new_above(&pmu_idr, pmu, PERF_TYPE_MAX, &type);
5534 if (pmu_bus_running) {
5535 ret = pmu_dev_alloc(pmu);
5541 pmu->pmu_cpu_context = find_pmu_context(pmu->task_ctx_nr);
5542 if (pmu->pmu_cpu_context)
5543 goto got_cpu_context;
5545 pmu->pmu_cpu_context = alloc_percpu(struct perf_cpu_context);
5546 if (!pmu->pmu_cpu_context)
5549 for_each_possible_cpu(cpu) {
5550 struct perf_cpu_context *cpuctx;
5552 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
5553 __perf_event_init_context(&cpuctx->ctx);
5554 lockdep_set_class(&cpuctx->ctx.mutex, &cpuctx_mutex);
5555 lockdep_set_class(&cpuctx->ctx.lock, &cpuctx_lock);
5556 cpuctx->ctx.type = cpu_context;
5557 cpuctx->ctx.pmu = pmu;
5558 cpuctx->jiffies_interval = 1;
5559 INIT_LIST_HEAD(&cpuctx->rotation_list);
5560 cpuctx->active_pmu = pmu;
5564 if (!pmu->start_txn) {
5565 if (pmu->pmu_enable) {
5567 * If we have pmu_enable/pmu_disable calls, install
5568 * transaction stubs that use that to try and batch
5569 * hardware accesses.
5571 pmu->start_txn = perf_pmu_start_txn;
5572 pmu->commit_txn = perf_pmu_commit_txn;
5573 pmu->cancel_txn = perf_pmu_cancel_txn;
5575 pmu->start_txn = perf_pmu_nop_void;
5576 pmu->commit_txn = perf_pmu_nop_int;
5577 pmu->cancel_txn = perf_pmu_nop_void;
5581 if (!pmu->pmu_enable) {
5582 pmu->pmu_enable = perf_pmu_nop_void;
5583 pmu->pmu_disable = perf_pmu_nop_void;
5586 list_add_rcu(&pmu->entry, &pmus);
5589 mutex_unlock(&pmus_lock);
5594 device_del(pmu->dev);
5595 put_device(pmu->dev);
5598 if (pmu->type >= PERF_TYPE_MAX)
5599 idr_remove(&pmu_idr, pmu->type);
5602 free_percpu(pmu->pmu_disable_count);
5606 void perf_pmu_unregister(struct pmu *pmu)
5608 mutex_lock(&pmus_lock);
5609 list_del_rcu(&pmu->entry);
5610 mutex_unlock(&pmus_lock);
5613 * We dereference the pmu list under both SRCU and regular RCU, so
5614 * synchronize against both of those.
5616 synchronize_srcu(&pmus_srcu);
5619 free_percpu(pmu->pmu_disable_count);
5620 if (pmu->type >= PERF_TYPE_MAX)
5621 idr_remove(&pmu_idr, pmu->type);
5622 device_del(pmu->dev);
5623 put_device(pmu->dev);
5624 free_pmu_context(pmu);
5627 struct pmu *perf_init_event(struct perf_event *event)
5629 struct pmu *pmu = NULL;
5633 idx = srcu_read_lock(&pmus_srcu);
5636 pmu = idr_find(&pmu_idr, event->attr.type);
5640 ret = pmu->event_init(event);
5646 list_for_each_entry_rcu(pmu, &pmus, entry) {
5648 ret = pmu->event_init(event);
5652 if (ret != -ENOENT) {
5657 pmu = ERR_PTR(-ENOENT);
5659 srcu_read_unlock(&pmus_srcu, idx);
5665 * Allocate and initialize a event structure
5667 static struct perf_event *
5668 perf_event_alloc(struct perf_event_attr *attr, int cpu,
5669 struct task_struct *task,
5670 struct perf_event *group_leader,
5671 struct perf_event *parent_event,
5672 perf_overflow_handler_t overflow_handler,
5676 struct perf_event *event;
5677 struct hw_perf_event *hwc;
5680 if ((unsigned)cpu >= nr_cpu_ids) {
5681 if (!task || cpu != -1)
5682 return ERR_PTR(-EINVAL);
5685 event = kzalloc(sizeof(*event), GFP_KERNEL);
5687 return ERR_PTR(-ENOMEM);
5690 * Single events are their own group leaders, with an
5691 * empty sibling list:
5694 group_leader = event;
5696 mutex_init(&event->child_mutex);
5697 INIT_LIST_HEAD(&event->child_list);
5699 INIT_LIST_HEAD(&event->group_entry);
5700 INIT_LIST_HEAD(&event->event_entry);
5701 INIT_LIST_HEAD(&event->sibling_list);
5702 INIT_LIST_HEAD(&event->rb_entry);
5704 init_waitqueue_head(&event->waitq);
5705 init_irq_work(&event->pending, perf_pending_event);
5707 mutex_init(&event->mmap_mutex);
5710 event->attr = *attr;
5711 event->group_leader = group_leader;
5715 event->parent = parent_event;
5717 event->ns = get_pid_ns(current->nsproxy->pid_ns);
5718 event->id = atomic64_inc_return(&perf_event_id);
5720 event->state = PERF_EVENT_STATE_INACTIVE;
5723 event->attach_state = PERF_ATTACH_TASK;
5724 #ifdef CONFIG_HAVE_HW_BREAKPOINT
5726 * hw_breakpoint is a bit difficult here..
5728 if (attr->type == PERF_TYPE_BREAKPOINT)
5729 event->hw.bp_target = task;
5733 if (!overflow_handler && parent_event) {
5734 overflow_handler = parent_event->overflow_handler;
5735 context = parent_event->overflow_handler_context;
5738 event->overflow_handler = overflow_handler;
5739 event->overflow_handler_context = context;
5742 event->state = PERF_EVENT_STATE_OFF;
5747 hwc->sample_period = attr->sample_period;
5748 if (attr->freq && attr->sample_freq)
5749 hwc->sample_period = 1;
5750 hwc->last_period = hwc->sample_period;
5752 local64_set(&hwc->period_left, hwc->sample_period);
5755 * we currently do not support PERF_FORMAT_GROUP on inherited events
5757 if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP))
5760 pmu = perf_init_event(event);
5766 else if (IS_ERR(pmu))
5771 put_pid_ns(event->ns);
5773 return ERR_PTR(err);
5776 if (!event->parent) {
5777 if (event->attach_state & PERF_ATTACH_TASK)
5778 jump_label_inc(&perf_sched_events);
5779 if (event->attr.mmap || event->attr.mmap_data)
5780 atomic_inc(&nr_mmap_events);
5781 if (event->attr.comm)
5782 atomic_inc(&nr_comm_events);
5783 if (event->attr.task)
5784 atomic_inc(&nr_task_events);
5785 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) {
5786 err = get_callchain_buffers();
5789 return ERR_PTR(err);
5797 static int perf_copy_attr(struct perf_event_attr __user *uattr,
5798 struct perf_event_attr *attr)
5803 if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0))
5807 * zero the full structure, so that a short copy will be nice.
5809 memset(attr, 0, sizeof(*attr));
5811 ret = get_user(size, &uattr->size);
5815 if (size > PAGE_SIZE) /* silly large */
5818 if (!size) /* abi compat */
5819 size = PERF_ATTR_SIZE_VER0;
5821 if (size < PERF_ATTR_SIZE_VER0)
5825 * If we're handed a bigger struct than we know of,
5826 * ensure all the unknown bits are 0 - i.e. new
5827 * user-space does not rely on any kernel feature
5828 * extensions we dont know about yet.
5830 if (size > sizeof(*attr)) {
5831 unsigned char __user *addr;
5832 unsigned char __user *end;
5835 addr = (void __user *)uattr + sizeof(*attr);
5836 end = (void __user *)uattr + size;
5838 for (; addr < end; addr++) {
5839 ret = get_user(val, addr);
5845 size = sizeof(*attr);
5848 ret = copy_from_user(attr, uattr, size);
5852 if (attr->__reserved_1)
5855 if (attr->sample_type & ~(PERF_SAMPLE_MAX-1))
5858 if (attr->read_format & ~(PERF_FORMAT_MAX-1))
5865 put_user(sizeof(*attr), &uattr->size);
5871 perf_event_set_output(struct perf_event *event, struct perf_event *output_event)
5873 struct ring_buffer *rb = NULL, *old_rb = NULL;
5879 /* don't allow circular references */
5880 if (event == output_event)
5884 * Don't allow cross-cpu buffers
5886 if (output_event->cpu != event->cpu)
5890 * If its not a per-cpu rb, it must be the same task.
5892 if (output_event->cpu == -1 && output_event->ctx != event->ctx)
5896 mutex_lock(&event->mmap_mutex);
5897 /* Can't redirect output if we've got an active mmap() */
5898 if (atomic_read(&event->mmap_count))
5902 /* get the rb we want to redirect to */
5903 rb = ring_buffer_get(output_event);
5909 rcu_assign_pointer(event->rb, rb);
5911 ring_buffer_detach(event, old_rb);
5914 mutex_unlock(&event->mmap_mutex);
5917 ring_buffer_put(old_rb);
5923 * sys_perf_event_open - open a performance event, associate it to a task/cpu
5925 * @attr_uptr: event_id type attributes for monitoring/sampling
5928 * @group_fd: group leader event fd
5930 SYSCALL_DEFINE5(perf_event_open,
5931 struct perf_event_attr __user *, attr_uptr,
5932 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
5934 struct perf_event *group_leader = NULL, *output_event = NULL;
5935 struct perf_event *event, *sibling;
5936 struct perf_event_attr attr;
5937 struct perf_event_context *ctx;
5938 struct file *event_file = NULL;
5939 struct file *group_file = NULL;
5940 struct task_struct *task = NULL;
5944 int fput_needed = 0;
5947 /* for future expandability... */
5948 if (flags & ~PERF_FLAG_ALL)
5951 err = perf_copy_attr(attr_uptr, &attr);
5955 if (!attr.exclude_kernel) {
5956 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
5961 if (attr.sample_freq > sysctl_perf_event_sample_rate)
5966 * In cgroup mode, the pid argument is used to pass the fd
5967 * opened to the cgroup directory in cgroupfs. The cpu argument
5968 * designates the cpu on which to monitor threads from that
5971 if ((flags & PERF_FLAG_PID_CGROUP) && (pid == -1 || cpu == -1))
5974 event_fd = get_unused_fd_flags(O_RDWR);
5978 if (group_fd != -1) {
5979 group_leader = perf_fget_light(group_fd, &fput_needed);
5980 if (IS_ERR(group_leader)) {
5981 err = PTR_ERR(group_leader);
5984 group_file = group_leader->filp;
5985 if (flags & PERF_FLAG_FD_OUTPUT)
5986 output_event = group_leader;
5987 if (flags & PERF_FLAG_FD_NO_GROUP)
5988 group_leader = NULL;
5991 if (pid != -1 && !(flags & PERF_FLAG_PID_CGROUP)) {
5992 task = find_lively_task_by_vpid(pid);
5994 err = PTR_ERR(task);
5999 event = perf_event_alloc(&attr, cpu, task, group_leader, NULL,
6001 if (IS_ERR(event)) {
6002 err = PTR_ERR(event);
6006 if (flags & PERF_FLAG_PID_CGROUP) {
6007 err = perf_cgroup_connect(pid, event, &attr, group_leader);
6012 * - that has cgroup constraint on event->cpu
6013 * - that may need work on context switch
6015 atomic_inc(&per_cpu(perf_cgroup_events, event->cpu));
6016 jump_label_inc(&perf_sched_events);
6020 * Special case software events and allow them to be part of
6021 * any hardware group.
6026 (is_software_event(event) != is_software_event(group_leader))) {
6027 if (is_software_event(event)) {
6029 * If event and group_leader are not both a software
6030 * event, and event is, then group leader is not.
6032 * Allow the addition of software events to !software
6033 * groups, this is safe because software events never
6036 pmu = group_leader->pmu;
6037 } else if (is_software_event(group_leader) &&
6038 (group_leader->group_flags & PERF_GROUP_SOFTWARE)) {
6040 * In case the group is a pure software group, and we
6041 * try to add a hardware event, move the whole group to
6042 * the hardware context.
6049 * Get the target context (task or percpu):
6051 ctx = find_get_context(pmu, task, cpu);
6058 put_task_struct(task);
6063 * Look up the group leader (we will attach this event to it):
6069 * Do not allow a recursive hierarchy (this new sibling
6070 * becoming part of another group-sibling):
6072 if (group_leader->group_leader != group_leader)
6075 * Do not allow to attach to a group in a different
6076 * task or CPU context:
6079 if (group_leader->ctx->type != ctx->type)
6082 if (group_leader->ctx != ctx)
6087 * Only a group leader can be exclusive or pinned
6089 if (attr.exclusive || attr.pinned)
6094 err = perf_event_set_output(event, output_event);
6099 event_file = anon_inode_getfile("[perf_event]", &perf_fops, event, O_RDWR);
6100 if (IS_ERR(event_file)) {
6101 err = PTR_ERR(event_file);
6106 struct perf_event_context *gctx = group_leader->ctx;
6108 mutex_lock(&gctx->mutex);
6109 perf_remove_from_context(group_leader);
6110 list_for_each_entry(sibling, &group_leader->sibling_list,
6112 perf_remove_from_context(sibling);
6115 mutex_unlock(&gctx->mutex);
6119 event->filp = event_file;
6120 WARN_ON_ONCE(ctx->parent_ctx);
6121 mutex_lock(&ctx->mutex);
6124 perf_install_in_context(ctx, group_leader, cpu);
6126 list_for_each_entry(sibling, &group_leader->sibling_list,
6128 perf_install_in_context(ctx, sibling, cpu);
6133 perf_install_in_context(ctx, event, cpu);
6135 perf_unpin_context(ctx);
6136 mutex_unlock(&ctx->mutex);
6138 event->owner = current;
6140 mutex_lock(¤t->perf_event_mutex);
6141 list_add_tail(&event->owner_entry, ¤t->perf_event_list);
6142 mutex_unlock(¤t->perf_event_mutex);
6145 * Precalculate sample_data sizes
6147 perf_event__header_size(event);
6148 perf_event__id_header_size(event);
6151 * Drop the reference on the group_event after placing the
6152 * new event on the sibling_list. This ensures destruction
6153 * of the group leader will find the pointer to itself in
6154 * perf_group_detach().
6156 fput_light(group_file, fput_needed);
6157 fd_install(event_fd, event_file);
6161 perf_unpin_context(ctx);
6167 put_task_struct(task);
6169 fput_light(group_file, fput_needed);
6171 put_unused_fd(event_fd);
6176 * perf_event_create_kernel_counter
6178 * @attr: attributes of the counter to create
6179 * @cpu: cpu in which the counter is bound
6180 * @task: task to profile (NULL for percpu)
6183 perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu,
6184 struct task_struct *task,
6185 perf_overflow_handler_t overflow_handler,
6188 struct perf_event_context *ctx;
6189 struct perf_event *event;
6193 * Get the target context (task or percpu):
6196 event = perf_event_alloc(attr, cpu, task, NULL, NULL,
6197 overflow_handler, context);
6198 if (IS_ERR(event)) {
6199 err = PTR_ERR(event);
6203 ctx = find_get_context(event->pmu, task, cpu);
6210 WARN_ON_ONCE(ctx->parent_ctx);
6211 mutex_lock(&ctx->mutex);
6212 perf_install_in_context(ctx, event, cpu);
6214 perf_unpin_context(ctx);
6215 mutex_unlock(&ctx->mutex);
6222 return ERR_PTR(err);
6224 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter);
6226 static void sync_child_event(struct perf_event *child_event,
6227 struct task_struct *child)
6229 struct perf_event *parent_event = child_event->parent;
6232 if (child_event->attr.inherit_stat)
6233 perf_event_read_event(child_event, child);
6235 child_val = perf_event_count(child_event);
6238 * Add back the child's count to the parent's count:
6240 atomic64_add(child_val, &parent_event->child_count);
6241 atomic64_add(child_event->total_time_enabled,
6242 &parent_event->child_total_time_enabled);
6243 atomic64_add(child_event->total_time_running,
6244 &parent_event->child_total_time_running);
6247 * Remove this event from the parent's list
6249 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
6250 mutex_lock(&parent_event->child_mutex);
6251 list_del_init(&child_event->child_list);
6252 mutex_unlock(&parent_event->child_mutex);
6255 * Release the parent event, if this was the last
6258 fput(parent_event->filp);
6262 __perf_event_exit_task(struct perf_event *child_event,
6263 struct perf_event_context *child_ctx,
6264 struct task_struct *child)
6266 if (child_event->parent) {
6267 raw_spin_lock_irq(&child_ctx->lock);
6268 perf_group_detach(child_event);
6269 raw_spin_unlock_irq(&child_ctx->lock);
6272 perf_remove_from_context(child_event);
6275 * It can happen that the parent exits first, and has events
6276 * that are still around due to the child reference. These
6277 * events need to be zapped.
6279 if (child_event->parent) {
6280 sync_child_event(child_event, child);
6281 free_event(child_event);
6285 static void perf_event_exit_task_context(struct task_struct *child, int ctxn)
6287 struct perf_event *child_event, *tmp;
6288 struct perf_event_context *child_ctx;
6289 unsigned long flags;
6291 if (likely(!child->perf_event_ctxp[ctxn])) {
6292 perf_event_task(child, NULL, 0);
6296 local_irq_save(flags);
6298 * We can't reschedule here because interrupts are disabled,
6299 * and either child is current or it is a task that can't be
6300 * scheduled, so we are now safe from rescheduling changing
6303 child_ctx = rcu_dereference_raw(child->perf_event_ctxp[ctxn]);
6306 * Take the context lock here so that if find_get_context is
6307 * reading child->perf_event_ctxp, we wait until it has
6308 * incremented the context's refcount before we do put_ctx below.
6310 raw_spin_lock(&child_ctx->lock);
6311 task_ctx_sched_out(child_ctx);
6312 child->perf_event_ctxp[ctxn] = NULL;
6314 * If this context is a clone; unclone it so it can't get
6315 * swapped to another process while we're removing all
6316 * the events from it.
6318 unclone_ctx(child_ctx);
6319 update_context_time(child_ctx);
6320 raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
6323 * Report the task dead after unscheduling the events so that we
6324 * won't get any samples after PERF_RECORD_EXIT. We can however still
6325 * get a few PERF_RECORD_READ events.
6327 perf_event_task(child, child_ctx, 0);
6330 * We can recurse on the same lock type through:
6332 * __perf_event_exit_task()
6333 * sync_child_event()
6334 * fput(parent_event->filp)
6336 * mutex_lock(&ctx->mutex)
6338 * But since its the parent context it won't be the same instance.
6340 mutex_lock(&child_ctx->mutex);
6343 list_for_each_entry_safe(child_event, tmp, &child_ctx->pinned_groups,
6345 __perf_event_exit_task(child_event, child_ctx, child);
6347 list_for_each_entry_safe(child_event, tmp, &child_ctx->flexible_groups,
6349 __perf_event_exit_task(child_event, child_ctx, child);
6352 * If the last event was a group event, it will have appended all
6353 * its siblings to the list, but we obtained 'tmp' before that which
6354 * will still point to the list head terminating the iteration.
6356 if (!list_empty(&child_ctx->pinned_groups) ||
6357 !list_empty(&child_ctx->flexible_groups))
6360 mutex_unlock(&child_ctx->mutex);
6366 * When a child task exits, feed back event values to parent events.
6368 void perf_event_exit_task(struct task_struct *child)
6370 struct perf_event *event, *tmp;
6373 mutex_lock(&child->perf_event_mutex);
6374 list_for_each_entry_safe(event, tmp, &child->perf_event_list,
6376 list_del_init(&event->owner_entry);
6379 * Ensure the list deletion is visible before we clear
6380 * the owner, closes a race against perf_release() where
6381 * we need to serialize on the owner->perf_event_mutex.
6384 event->owner = NULL;
6386 mutex_unlock(&child->perf_event_mutex);
6388 for_each_task_context_nr(ctxn)
6389 perf_event_exit_task_context(child, ctxn);
6392 static void perf_free_event(struct perf_event *event,
6393 struct perf_event_context *ctx)
6395 struct perf_event *parent = event->parent;
6397 if (WARN_ON_ONCE(!parent))
6400 mutex_lock(&parent->child_mutex);
6401 list_del_init(&event->child_list);
6402 mutex_unlock(&parent->child_mutex);
6406 perf_group_detach(event);
6407 list_del_event(event, ctx);
6412 * free an unexposed, unused context as created by inheritance by
6413 * perf_event_init_task below, used by fork() in case of fail.
6415 void perf_event_free_task(struct task_struct *task)
6417 struct perf_event_context *ctx;
6418 struct perf_event *event, *tmp;
6421 for_each_task_context_nr(ctxn) {
6422 ctx = task->perf_event_ctxp[ctxn];
6426 mutex_lock(&ctx->mutex);
6428 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups,
6430 perf_free_event(event, ctx);
6432 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups,
6434 perf_free_event(event, ctx);
6436 if (!list_empty(&ctx->pinned_groups) ||
6437 !list_empty(&ctx->flexible_groups))
6440 mutex_unlock(&ctx->mutex);
6446 void perf_event_delayed_put(struct task_struct *task)
6450 for_each_task_context_nr(ctxn)
6451 WARN_ON_ONCE(task->perf_event_ctxp[ctxn]);
6455 * inherit a event from parent task to child task:
6457 static struct perf_event *
6458 inherit_event(struct perf_event *parent_event,
6459 struct task_struct *parent,
6460 struct perf_event_context *parent_ctx,
6461 struct task_struct *child,
6462 struct perf_event *group_leader,
6463 struct perf_event_context *child_ctx)
6465 struct perf_event *child_event;
6466 unsigned long flags;
6469 * Instead of creating recursive hierarchies of events,
6470 * we link inherited events back to the original parent,
6471 * which has a filp for sure, which we use as the reference
6474 if (parent_event->parent)
6475 parent_event = parent_event->parent;
6477 child_event = perf_event_alloc(&parent_event->attr,
6480 group_leader, parent_event,
6482 if (IS_ERR(child_event))
6487 * Make the child state follow the state of the parent event,
6488 * not its attr.disabled bit. We hold the parent's mutex,
6489 * so we won't race with perf_event_{en, dis}able_family.
6491 if (parent_event->state >= PERF_EVENT_STATE_INACTIVE)
6492 child_event->state = PERF_EVENT_STATE_INACTIVE;
6494 child_event->state = PERF_EVENT_STATE_OFF;
6496 if (parent_event->attr.freq) {
6497 u64 sample_period = parent_event->hw.sample_period;
6498 struct hw_perf_event *hwc = &child_event->hw;
6500 hwc->sample_period = sample_period;
6501 hwc->last_period = sample_period;
6503 local64_set(&hwc->period_left, sample_period);
6506 child_event->ctx = child_ctx;
6507 child_event->overflow_handler = parent_event->overflow_handler;
6508 child_event->overflow_handler_context
6509 = parent_event->overflow_handler_context;
6512 * Precalculate sample_data sizes
6514 perf_event__header_size(child_event);
6515 perf_event__id_header_size(child_event);
6518 * Link it up in the child's context:
6520 raw_spin_lock_irqsave(&child_ctx->lock, flags);
6521 add_event_to_ctx(child_event, child_ctx);
6522 raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
6525 * Get a reference to the parent filp - we will fput it
6526 * when the child event exits. This is safe to do because
6527 * we are in the parent and we know that the filp still
6528 * exists and has a nonzero count:
6530 atomic_long_inc(&parent_event->filp->f_count);
6533 * Link this into the parent event's child list
6535 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
6536 mutex_lock(&parent_event->child_mutex);
6537 list_add_tail(&child_event->child_list, &parent_event->child_list);
6538 mutex_unlock(&parent_event->child_mutex);
6543 static int inherit_group(struct perf_event *parent_event,
6544 struct task_struct *parent,
6545 struct perf_event_context *parent_ctx,
6546 struct task_struct *child,
6547 struct perf_event_context *child_ctx)
6549 struct perf_event *leader;
6550 struct perf_event *sub;
6551 struct perf_event *child_ctr;
6553 leader = inherit_event(parent_event, parent, parent_ctx,
6554 child, NULL, child_ctx);
6556 return PTR_ERR(leader);
6557 list_for_each_entry(sub, &parent_event->sibling_list, group_entry) {
6558 child_ctr = inherit_event(sub, parent, parent_ctx,
6559 child, leader, child_ctx);
6560 if (IS_ERR(child_ctr))
6561 return PTR_ERR(child_ctr);
6567 inherit_task_group(struct perf_event *event, struct task_struct *parent,
6568 struct perf_event_context *parent_ctx,
6569 struct task_struct *child, int ctxn,
6573 struct perf_event_context *child_ctx;
6575 if (!event->attr.inherit) {
6580 child_ctx = child->perf_event_ctxp[ctxn];
6583 * This is executed from the parent task context, so
6584 * inherit events that have been marked for cloning.
6585 * First allocate and initialize a context for the
6589 child_ctx = alloc_perf_context(event->pmu, child);
6593 child->perf_event_ctxp[ctxn] = child_ctx;
6596 ret = inherit_group(event, parent, parent_ctx,
6606 * Initialize the perf_event context in task_struct
6608 int perf_event_init_context(struct task_struct *child, int ctxn)
6610 struct perf_event_context *child_ctx, *parent_ctx;
6611 struct perf_event_context *cloned_ctx;
6612 struct perf_event *event;
6613 struct task_struct *parent = current;
6614 int inherited_all = 1;
6615 unsigned long flags;
6618 if (likely(!parent->perf_event_ctxp[ctxn]))
6622 * If the parent's context is a clone, pin it so it won't get
6625 parent_ctx = perf_pin_task_context(parent, ctxn);
6628 * No need to check if parent_ctx != NULL here; since we saw
6629 * it non-NULL earlier, the only reason for it to become NULL
6630 * is if we exit, and since we're currently in the middle of
6631 * a fork we can't be exiting at the same time.
6635 * Lock the parent list. No need to lock the child - not PID
6636 * hashed yet and not running, so nobody can access it.
6638 mutex_lock(&parent_ctx->mutex);
6641 * We dont have to disable NMIs - we are only looking at
6642 * the list, not manipulating it:
6644 list_for_each_entry(event, &parent_ctx->pinned_groups, group_entry) {
6645 ret = inherit_task_group(event, parent, parent_ctx,
6646 child, ctxn, &inherited_all);
6652 * We can't hold ctx->lock when iterating the ->flexible_group list due
6653 * to allocations, but we need to prevent rotation because
6654 * rotate_ctx() will change the list from interrupt context.
6656 raw_spin_lock_irqsave(&parent_ctx->lock, flags);
6657 parent_ctx->rotate_disable = 1;
6658 raw_spin_unlock_irqrestore(&parent_ctx->lock, flags);
6660 list_for_each_entry(event, &parent_ctx->flexible_groups, group_entry) {
6661 ret = inherit_task_group(event, parent, parent_ctx,
6662 child, ctxn, &inherited_all);
6667 raw_spin_lock_irqsave(&parent_ctx->lock, flags);
6668 parent_ctx->rotate_disable = 0;
6670 child_ctx = child->perf_event_ctxp[ctxn];
6672 if (child_ctx && inherited_all) {
6674 * Mark the child context as a clone of the parent
6675 * context, or of whatever the parent is a clone of.
6677 * Note that if the parent is a clone, the holding of
6678 * parent_ctx->lock avoids it from being uncloned.
6680 cloned_ctx = parent_ctx->parent_ctx;
6682 child_ctx->parent_ctx = cloned_ctx;
6683 child_ctx->parent_gen = parent_ctx->parent_gen;
6685 child_ctx->parent_ctx = parent_ctx;
6686 child_ctx->parent_gen = parent_ctx->generation;
6688 get_ctx(child_ctx->parent_ctx);
6691 raw_spin_unlock_irqrestore(&parent_ctx->lock, flags);
6692 mutex_unlock(&parent_ctx->mutex);
6694 perf_unpin_context(parent_ctx);
6695 put_ctx(parent_ctx);
6701 * Initialize the perf_event context in task_struct
6703 int perf_event_init_task(struct task_struct *child)
6707 memset(child->perf_event_ctxp, 0, sizeof(child->perf_event_ctxp));
6708 mutex_init(&child->perf_event_mutex);
6709 INIT_LIST_HEAD(&child->perf_event_list);
6711 for_each_task_context_nr(ctxn) {
6712 ret = perf_event_init_context(child, ctxn);
6720 static void __init perf_event_init_all_cpus(void)
6722 struct swevent_htable *swhash;
6725 for_each_possible_cpu(cpu) {
6726 swhash = &per_cpu(swevent_htable, cpu);
6727 mutex_init(&swhash->hlist_mutex);
6728 INIT_LIST_HEAD(&per_cpu(rotation_list, cpu));
6732 static void __cpuinit perf_event_init_cpu(int cpu)
6734 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
6736 mutex_lock(&swhash->hlist_mutex);
6737 if (swhash->hlist_refcount > 0) {
6738 struct swevent_hlist *hlist;
6740 hlist = kzalloc_node(sizeof(*hlist), GFP_KERNEL, cpu_to_node(cpu));
6742 rcu_assign_pointer(swhash->swevent_hlist, hlist);
6744 mutex_unlock(&swhash->hlist_mutex);
6747 #if defined CONFIG_HOTPLUG_CPU || defined CONFIG_KEXEC
6748 static void perf_pmu_rotate_stop(struct pmu *pmu)
6750 struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
6752 WARN_ON(!irqs_disabled());
6754 list_del_init(&cpuctx->rotation_list);
6757 static void __perf_event_exit_context(void *__info)
6759 struct perf_event_context *ctx = __info;
6760 struct perf_event *event, *tmp;
6762 perf_pmu_rotate_stop(ctx->pmu);
6764 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry)
6765 __perf_remove_from_context(event);
6766 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups, group_entry)
6767 __perf_remove_from_context(event);
6770 static void perf_event_exit_cpu_context(int cpu)
6772 struct perf_event_context *ctx;
6776 idx = srcu_read_lock(&pmus_srcu);
6777 list_for_each_entry_rcu(pmu, &pmus, entry) {
6778 ctx = &per_cpu_ptr(pmu->pmu_cpu_context, cpu)->ctx;
6780 mutex_lock(&ctx->mutex);
6781 smp_call_function_single(cpu, __perf_event_exit_context, ctx, 1);
6782 mutex_unlock(&ctx->mutex);
6784 srcu_read_unlock(&pmus_srcu, idx);
6787 static void perf_event_exit_cpu(int cpu)
6789 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
6791 mutex_lock(&swhash->hlist_mutex);
6792 swevent_hlist_release(swhash);
6793 mutex_unlock(&swhash->hlist_mutex);
6795 perf_event_exit_cpu_context(cpu);
6798 static inline void perf_event_exit_cpu(int cpu) { }
6802 perf_reboot(struct notifier_block *notifier, unsigned long val, void *v)
6806 for_each_online_cpu(cpu)
6807 perf_event_exit_cpu(cpu);
6813 * Run the perf reboot notifier at the very last possible moment so that
6814 * the generic watchdog code runs as long as possible.
6816 static struct notifier_block perf_reboot_notifier = {
6817 .notifier_call = perf_reboot,
6818 .priority = INT_MIN,
6821 static int __cpuinit
6822 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
6824 unsigned int cpu = (long)hcpu;
6826 switch (action & ~CPU_TASKS_FROZEN) {
6828 case CPU_UP_PREPARE:
6829 case CPU_DOWN_FAILED:
6830 perf_event_init_cpu(cpu);
6833 case CPU_UP_CANCELED:
6834 case CPU_DOWN_PREPARE:
6835 perf_event_exit_cpu(cpu);
6845 void __init perf_event_init(void)
6851 perf_event_init_all_cpus();
6852 init_srcu_struct(&pmus_srcu);
6853 perf_pmu_register(&perf_swevent, "software", PERF_TYPE_SOFTWARE);
6854 perf_pmu_register(&perf_cpu_clock, NULL, -1);
6855 perf_pmu_register(&perf_task_clock, NULL, -1);
6857 perf_cpu_notifier(perf_cpu_notify);
6858 register_reboot_notifier(&perf_reboot_notifier);
6860 ret = init_hw_breakpoint();
6861 WARN(ret, "hw_breakpoint initialization failed with: %d", ret);
6864 static int __init perf_event_sysfs_init(void)
6869 mutex_lock(&pmus_lock);
6871 ret = bus_register(&pmu_bus);
6875 list_for_each_entry(pmu, &pmus, entry) {
6876 if (!pmu->name || pmu->type < 0)
6879 ret = pmu_dev_alloc(pmu);
6880 WARN(ret, "Failed to register pmu: %s, reason %d\n", pmu->name, ret);
6882 pmu_bus_running = 1;
6886 mutex_unlock(&pmus_lock);
6890 device_initcall(perf_event_sysfs_init);
6892 #ifdef CONFIG_CGROUP_PERF
6893 static struct cgroup_subsys_state *perf_cgroup_create(
6894 struct cgroup_subsys *ss, struct cgroup *cont)
6896 struct perf_cgroup *jc;
6898 jc = kzalloc(sizeof(*jc), GFP_KERNEL);
6900 return ERR_PTR(-ENOMEM);
6902 jc->info = alloc_percpu(struct perf_cgroup_info);
6905 return ERR_PTR(-ENOMEM);
6911 static void perf_cgroup_destroy(struct cgroup_subsys *ss,
6912 struct cgroup *cont)
6914 struct perf_cgroup *jc;
6915 jc = container_of(cgroup_subsys_state(cont, perf_subsys_id),
6916 struct perf_cgroup, css);
6917 free_percpu(jc->info);
6921 static int __perf_cgroup_move(void *info)
6923 struct task_struct *task = info;
6924 perf_cgroup_switch(task, PERF_CGROUP_SWOUT | PERF_CGROUP_SWIN);
6929 perf_cgroup_attach_task(struct cgroup *cgrp, struct task_struct *task)
6931 task_function_call(task, __perf_cgroup_move, task);
6934 static void perf_cgroup_exit(struct cgroup_subsys *ss, struct cgroup *cgrp,
6935 struct cgroup *old_cgrp, struct task_struct *task)
6938 * cgroup_exit() is called in the copy_process() failure path.
6939 * Ignore this case since the task hasn't ran yet, this avoids
6940 * trying to poke a half freed task state from generic code.
6942 if (!(task->flags & PF_EXITING))
6945 perf_cgroup_attach_task(cgrp, task);
6948 struct cgroup_subsys perf_subsys = {
6949 .name = "perf_event",
6950 .subsys_id = perf_subsys_id,
6951 .create = perf_cgroup_create,
6952 .destroy = perf_cgroup_destroy,
6953 .exit = perf_cgroup_exit,
6954 .attach_task = perf_cgroup_attach_task,
6956 #endif /* CONFIG_CGROUP_PERF */