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);
1329 static void perf_set_shadow_time(struct perf_event *event,
1330 struct perf_event_context *ctx,
1334 * use the correct time source for the time snapshot
1336 * We could get by without this by leveraging the
1337 * fact that to get to this function, the caller
1338 * has most likely already called update_context_time()
1339 * and update_cgrp_time_xx() and thus both timestamp
1340 * are identical (or very close). Given that tstamp is,
1341 * already adjusted for cgroup, we could say that:
1342 * tstamp - ctx->timestamp
1344 * tstamp - cgrp->timestamp.
1346 * Then, in perf_output_read(), the calculation would
1347 * work with no changes because:
1348 * - event is guaranteed scheduled in
1349 * - no scheduled out in between
1350 * - thus the timestamp would be the same
1352 * But this is a bit hairy.
1354 * So instead, we have an explicit cgroup call to remain
1355 * within the time time source all along. We believe it
1356 * is cleaner and simpler to understand.
1358 if (is_cgroup_event(event))
1359 perf_cgroup_set_shadow_time(event, tstamp);
1361 event->shadow_ctx_time = tstamp - ctx->timestamp;
1364 #define MAX_INTERRUPTS (~0ULL)
1366 static void perf_log_throttle(struct perf_event *event, int enable);
1369 event_sched_in(struct perf_event *event,
1370 struct perf_cpu_context *cpuctx,
1371 struct perf_event_context *ctx)
1373 u64 tstamp = perf_event_time(event);
1375 if (event->state <= PERF_EVENT_STATE_OFF)
1378 event->state = PERF_EVENT_STATE_ACTIVE;
1379 event->oncpu = smp_processor_id();
1382 * Unthrottle events, since we scheduled we might have missed several
1383 * ticks already, also for a heavily scheduling task there is little
1384 * guarantee it'll get a tick in a timely manner.
1386 if (unlikely(event->hw.interrupts == MAX_INTERRUPTS)) {
1387 perf_log_throttle(event, 1);
1388 event->hw.interrupts = 0;
1392 * The new state must be visible before we turn it on in the hardware:
1396 if (event->pmu->add(event, PERF_EF_START)) {
1397 event->state = PERF_EVENT_STATE_INACTIVE;
1402 event->tstamp_running += tstamp - event->tstamp_stopped;
1404 perf_set_shadow_time(event, ctx, tstamp);
1406 if (!is_software_event(event))
1407 cpuctx->active_oncpu++;
1410 if (event->attr.exclusive)
1411 cpuctx->exclusive = 1;
1417 group_sched_in(struct perf_event *group_event,
1418 struct perf_cpu_context *cpuctx,
1419 struct perf_event_context *ctx)
1421 struct perf_event *event, *partial_group = NULL;
1422 struct pmu *pmu = group_event->pmu;
1423 u64 now = ctx->time;
1424 bool simulate = false;
1426 if (group_event->state == PERF_EVENT_STATE_OFF)
1429 pmu->start_txn(pmu);
1431 if (event_sched_in(group_event, cpuctx, ctx)) {
1432 pmu->cancel_txn(pmu);
1437 * Schedule in siblings as one group (if any):
1439 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
1440 if (event_sched_in(event, cpuctx, ctx)) {
1441 partial_group = event;
1446 if (!pmu->commit_txn(pmu))
1451 * Groups can be scheduled in as one unit only, so undo any
1452 * partial group before returning:
1453 * The events up to the failed event are scheduled out normally,
1454 * tstamp_stopped will be updated.
1456 * The failed events and the remaining siblings need to have
1457 * their timings updated as if they had gone thru event_sched_in()
1458 * and event_sched_out(). This is required to get consistent timings
1459 * across the group. This also takes care of the case where the group
1460 * could never be scheduled by ensuring tstamp_stopped is set to mark
1461 * the time the event was actually stopped, such that time delta
1462 * calculation in update_event_times() is correct.
1464 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
1465 if (event == partial_group)
1469 event->tstamp_running += now - event->tstamp_stopped;
1470 event->tstamp_stopped = now;
1472 event_sched_out(event, cpuctx, ctx);
1475 event_sched_out(group_event, cpuctx, ctx);
1477 pmu->cancel_txn(pmu);
1483 * Work out whether we can put this event group on the CPU now.
1485 static int group_can_go_on(struct perf_event *event,
1486 struct perf_cpu_context *cpuctx,
1490 * Groups consisting entirely of software events can always go on.
1492 if (event->group_flags & PERF_GROUP_SOFTWARE)
1495 * If an exclusive group is already on, no other hardware
1498 if (cpuctx->exclusive)
1501 * If this group is exclusive and there are already
1502 * events on the CPU, it can't go on.
1504 if (event->attr.exclusive && cpuctx->active_oncpu)
1507 * Otherwise, try to add it if all previous groups were able
1513 static void add_event_to_ctx(struct perf_event *event,
1514 struct perf_event_context *ctx)
1516 u64 tstamp = perf_event_time(event);
1518 list_add_event(event, ctx);
1519 perf_group_attach(event);
1520 event->tstamp_enabled = tstamp;
1521 event->tstamp_running = tstamp;
1522 event->tstamp_stopped = tstamp;
1525 static void task_ctx_sched_out(struct perf_event_context *ctx);
1527 ctx_sched_in(struct perf_event_context *ctx,
1528 struct perf_cpu_context *cpuctx,
1529 enum event_type_t event_type,
1530 struct task_struct *task);
1532 static void perf_event_sched_in(struct perf_cpu_context *cpuctx,
1533 struct perf_event_context *ctx,
1534 struct task_struct *task)
1536 cpu_ctx_sched_in(cpuctx, EVENT_PINNED, task);
1538 ctx_sched_in(ctx, cpuctx, EVENT_PINNED, task);
1539 cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE, task);
1541 ctx_sched_in(ctx, cpuctx, EVENT_FLEXIBLE, task);
1545 * Cross CPU call to install and enable a performance event
1547 * Must be called with ctx->mutex held
1549 static int __perf_install_in_context(void *info)
1551 struct perf_event *event = info;
1552 struct perf_event_context *ctx = event->ctx;
1553 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1554 struct perf_event_context *task_ctx = cpuctx->task_ctx;
1555 struct task_struct *task = current;
1557 perf_ctx_lock(cpuctx, task_ctx);
1558 perf_pmu_disable(cpuctx->ctx.pmu);
1561 * If there was an active task_ctx schedule it out.
1564 task_ctx_sched_out(task_ctx);
1567 * If the context we're installing events in is not the
1568 * active task_ctx, flip them.
1570 if (ctx->task && task_ctx != ctx) {
1572 raw_spin_unlock(&task_ctx->lock);
1573 raw_spin_lock(&ctx->lock);
1578 cpuctx->task_ctx = task_ctx;
1579 task = task_ctx->task;
1582 cpu_ctx_sched_out(cpuctx, EVENT_ALL);
1584 update_context_time(ctx);
1586 * update cgrp time only if current cgrp
1587 * matches event->cgrp. Must be done before
1588 * calling add_event_to_ctx()
1590 update_cgrp_time_from_event(event);
1592 add_event_to_ctx(event, ctx);
1595 * Schedule everything back in
1597 perf_event_sched_in(cpuctx, task_ctx, task);
1599 perf_pmu_enable(cpuctx->ctx.pmu);
1600 perf_ctx_unlock(cpuctx, task_ctx);
1606 * Attach a performance event to a context
1608 * First we add the event to the list with the hardware enable bit
1609 * in event->hw_config cleared.
1611 * If the event is attached to a task which is on a CPU we use a smp
1612 * call to enable it in the task context. The task might have been
1613 * scheduled away, but we check this in the smp call again.
1616 perf_install_in_context(struct perf_event_context *ctx,
1617 struct perf_event *event,
1620 struct task_struct *task = ctx->task;
1622 lockdep_assert_held(&ctx->mutex);
1628 * Per cpu events are installed via an smp call and
1629 * the install is always successful.
1631 cpu_function_call(cpu, __perf_install_in_context, event);
1636 if (!task_function_call(task, __perf_install_in_context, event))
1639 raw_spin_lock_irq(&ctx->lock);
1641 * If we failed to find a running task, but find the context active now
1642 * that we've acquired the ctx->lock, retry.
1644 if (ctx->is_active) {
1645 raw_spin_unlock_irq(&ctx->lock);
1650 * Since the task isn't running, its safe to add the event, us holding
1651 * the ctx->lock ensures the task won't get scheduled in.
1653 add_event_to_ctx(event, ctx);
1654 raw_spin_unlock_irq(&ctx->lock);
1658 * Put a event into inactive state and update time fields.
1659 * Enabling the leader of a group effectively enables all
1660 * the group members that aren't explicitly disabled, so we
1661 * have to update their ->tstamp_enabled also.
1662 * Note: this works for group members as well as group leaders
1663 * since the non-leader members' sibling_lists will be empty.
1665 static void __perf_event_mark_enabled(struct perf_event *event,
1666 struct perf_event_context *ctx)
1668 struct perf_event *sub;
1669 u64 tstamp = perf_event_time(event);
1671 event->state = PERF_EVENT_STATE_INACTIVE;
1672 event->tstamp_enabled = tstamp - event->total_time_enabled;
1673 list_for_each_entry(sub, &event->sibling_list, group_entry) {
1674 if (sub->state >= PERF_EVENT_STATE_INACTIVE)
1675 sub->tstamp_enabled = tstamp - sub->total_time_enabled;
1680 * Cross CPU call to enable a performance event
1682 static int __perf_event_enable(void *info)
1684 struct perf_event *event = info;
1685 struct perf_event_context *ctx = event->ctx;
1686 struct perf_event *leader = event->group_leader;
1687 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1690 if (WARN_ON_ONCE(!ctx->is_active))
1693 raw_spin_lock(&ctx->lock);
1694 update_context_time(ctx);
1696 if (event->state >= PERF_EVENT_STATE_INACTIVE)
1700 * set current task's cgroup time reference point
1702 perf_cgroup_set_timestamp(current, ctx);
1704 __perf_event_mark_enabled(event, ctx);
1706 if (!event_filter_match(event)) {
1707 if (is_cgroup_event(event))
1708 perf_cgroup_defer_enabled(event);
1713 * If the event is in a group and isn't the group leader,
1714 * then don't put it on unless the group is on.
1716 if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE)
1719 if (!group_can_go_on(event, cpuctx, 1)) {
1722 if (event == leader)
1723 err = group_sched_in(event, cpuctx, ctx);
1725 err = event_sched_in(event, cpuctx, ctx);
1730 * If this event can't go on and it's part of a
1731 * group, then the whole group has to come off.
1733 if (leader != event)
1734 group_sched_out(leader, cpuctx, ctx);
1735 if (leader->attr.pinned) {
1736 update_group_times(leader);
1737 leader->state = PERF_EVENT_STATE_ERROR;
1742 raw_spin_unlock(&ctx->lock);
1750 * If event->ctx is a cloned context, callers must make sure that
1751 * every task struct that event->ctx->task could possibly point to
1752 * remains valid. This condition is satisfied when called through
1753 * perf_event_for_each_child or perf_event_for_each as described
1754 * for perf_event_disable.
1756 void perf_event_enable(struct perf_event *event)
1758 struct perf_event_context *ctx = event->ctx;
1759 struct task_struct *task = ctx->task;
1763 * Enable the event on the cpu that it's on
1765 cpu_function_call(event->cpu, __perf_event_enable, event);
1769 raw_spin_lock_irq(&ctx->lock);
1770 if (event->state >= PERF_EVENT_STATE_INACTIVE)
1774 * If the event is in error state, clear that first.
1775 * That way, if we see the event in error state below, we
1776 * know that it has gone back into error state, as distinct
1777 * from the task having been scheduled away before the
1778 * cross-call arrived.
1780 if (event->state == PERF_EVENT_STATE_ERROR)
1781 event->state = PERF_EVENT_STATE_OFF;
1784 if (!ctx->is_active) {
1785 __perf_event_mark_enabled(event, ctx);
1789 raw_spin_unlock_irq(&ctx->lock);
1791 if (!task_function_call(task, __perf_event_enable, event))
1794 raw_spin_lock_irq(&ctx->lock);
1797 * If the context is active and the event is still off,
1798 * we need to retry the cross-call.
1800 if (ctx->is_active && event->state == PERF_EVENT_STATE_OFF) {
1802 * task could have been flipped by a concurrent
1803 * perf_event_context_sched_out()
1810 raw_spin_unlock_irq(&ctx->lock);
1813 int perf_event_refresh(struct perf_event *event, int refresh)
1816 * not supported on inherited events
1818 if (event->attr.inherit || !is_sampling_event(event))
1821 atomic_add(refresh, &event->event_limit);
1822 perf_event_enable(event);
1826 EXPORT_SYMBOL_GPL(perf_event_refresh);
1828 static void ctx_sched_out(struct perf_event_context *ctx,
1829 struct perf_cpu_context *cpuctx,
1830 enum event_type_t event_type)
1832 struct perf_event *event;
1833 int is_active = ctx->is_active;
1835 ctx->is_active &= ~event_type;
1836 if (likely(!ctx->nr_events))
1839 update_context_time(ctx);
1840 update_cgrp_time_from_cpuctx(cpuctx);
1841 if (!ctx->nr_active)
1844 perf_pmu_disable(ctx->pmu);
1845 if ((is_active & EVENT_PINNED) && (event_type & EVENT_PINNED)) {
1846 list_for_each_entry(event, &ctx->pinned_groups, group_entry)
1847 group_sched_out(event, cpuctx, ctx);
1850 if ((is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE)) {
1851 list_for_each_entry(event, &ctx->flexible_groups, group_entry)
1852 group_sched_out(event, cpuctx, ctx);
1854 perf_pmu_enable(ctx->pmu);
1858 * Test whether two contexts are equivalent, i.e. whether they
1859 * have both been cloned from the same version of the same context
1860 * and they both have the same number of enabled events.
1861 * If the number of enabled events is the same, then the set
1862 * of enabled events should be the same, because these are both
1863 * inherited contexts, therefore we can't access individual events
1864 * in them directly with an fd; we can only enable/disable all
1865 * events via prctl, or enable/disable all events in a family
1866 * via ioctl, which will have the same effect on both contexts.
1868 static int context_equiv(struct perf_event_context *ctx1,
1869 struct perf_event_context *ctx2)
1871 return ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx
1872 && ctx1->parent_gen == ctx2->parent_gen
1873 && !ctx1->pin_count && !ctx2->pin_count;
1876 static void __perf_event_sync_stat(struct perf_event *event,
1877 struct perf_event *next_event)
1881 if (!event->attr.inherit_stat)
1885 * Update the event value, we cannot use perf_event_read()
1886 * because we're in the middle of a context switch and have IRQs
1887 * disabled, which upsets smp_call_function_single(), however
1888 * we know the event must be on the current CPU, therefore we
1889 * don't need to use it.
1891 switch (event->state) {
1892 case PERF_EVENT_STATE_ACTIVE:
1893 event->pmu->read(event);
1896 case PERF_EVENT_STATE_INACTIVE:
1897 update_event_times(event);
1905 * In order to keep per-task stats reliable we need to flip the event
1906 * values when we flip the contexts.
1908 value = local64_read(&next_event->count);
1909 value = local64_xchg(&event->count, value);
1910 local64_set(&next_event->count, value);
1912 swap(event->total_time_enabled, next_event->total_time_enabled);
1913 swap(event->total_time_running, next_event->total_time_running);
1916 * Since we swizzled the values, update the user visible data too.
1918 perf_event_update_userpage(event);
1919 perf_event_update_userpage(next_event);
1922 #define list_next_entry(pos, member) \
1923 list_entry(pos->member.next, typeof(*pos), member)
1925 static void perf_event_sync_stat(struct perf_event_context *ctx,
1926 struct perf_event_context *next_ctx)
1928 struct perf_event *event, *next_event;
1933 update_context_time(ctx);
1935 event = list_first_entry(&ctx->event_list,
1936 struct perf_event, event_entry);
1938 next_event = list_first_entry(&next_ctx->event_list,
1939 struct perf_event, event_entry);
1941 while (&event->event_entry != &ctx->event_list &&
1942 &next_event->event_entry != &next_ctx->event_list) {
1944 __perf_event_sync_stat(event, next_event);
1946 event = list_next_entry(event, event_entry);
1947 next_event = list_next_entry(next_event, event_entry);
1951 static void perf_event_context_sched_out(struct task_struct *task, int ctxn,
1952 struct task_struct *next)
1954 struct perf_event_context *ctx = task->perf_event_ctxp[ctxn];
1955 struct perf_event_context *next_ctx;
1956 struct perf_event_context *parent;
1957 struct perf_cpu_context *cpuctx;
1963 cpuctx = __get_cpu_context(ctx);
1964 if (!cpuctx->task_ctx)
1968 parent = rcu_dereference(ctx->parent_ctx);
1969 next_ctx = next->perf_event_ctxp[ctxn];
1970 if (parent && next_ctx &&
1971 rcu_dereference(next_ctx->parent_ctx) == parent) {
1973 * Looks like the two contexts are clones, so we might be
1974 * able to optimize the context switch. We lock both
1975 * contexts and check that they are clones under the
1976 * lock (including re-checking that neither has been
1977 * uncloned in the meantime). It doesn't matter which
1978 * order we take the locks because no other cpu could
1979 * be trying to lock both of these tasks.
1981 raw_spin_lock(&ctx->lock);
1982 raw_spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING);
1983 if (context_equiv(ctx, next_ctx)) {
1985 * XXX do we need a memory barrier of sorts
1986 * wrt to rcu_dereference() of perf_event_ctxp
1988 task->perf_event_ctxp[ctxn] = next_ctx;
1989 next->perf_event_ctxp[ctxn] = ctx;
1991 next_ctx->task = task;
1994 perf_event_sync_stat(ctx, next_ctx);
1996 raw_spin_unlock(&next_ctx->lock);
1997 raw_spin_unlock(&ctx->lock);
2002 raw_spin_lock(&ctx->lock);
2003 ctx_sched_out(ctx, cpuctx, EVENT_ALL);
2004 cpuctx->task_ctx = NULL;
2005 raw_spin_unlock(&ctx->lock);
2009 #define for_each_task_context_nr(ctxn) \
2010 for ((ctxn) = 0; (ctxn) < perf_nr_task_contexts; (ctxn)++)
2013 * Called from scheduler to remove the events of the current task,
2014 * with interrupts disabled.
2016 * We stop each event and update the event value in event->count.
2018 * This does not protect us against NMI, but disable()
2019 * sets the disabled bit in the control field of event _before_
2020 * accessing the event control register. If a NMI hits, then it will
2021 * not restart the event.
2023 void __perf_event_task_sched_out(struct task_struct *task,
2024 struct task_struct *next)
2028 for_each_task_context_nr(ctxn)
2029 perf_event_context_sched_out(task, ctxn, next);
2032 * if cgroup events exist on this CPU, then we need
2033 * to check if we have to switch out PMU state.
2034 * cgroup event are system-wide mode only
2036 if (atomic_read(&__get_cpu_var(perf_cgroup_events)))
2037 perf_cgroup_sched_out(task, next);
2040 static void task_ctx_sched_out(struct perf_event_context *ctx)
2042 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
2044 if (!cpuctx->task_ctx)
2047 if (WARN_ON_ONCE(ctx != cpuctx->task_ctx))
2050 ctx_sched_out(ctx, cpuctx, EVENT_ALL);
2051 cpuctx->task_ctx = NULL;
2055 * Called with IRQs disabled
2057 static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
2058 enum event_type_t event_type)
2060 ctx_sched_out(&cpuctx->ctx, cpuctx, event_type);
2064 ctx_pinned_sched_in(struct perf_event_context *ctx,
2065 struct perf_cpu_context *cpuctx)
2067 struct perf_event *event;
2069 list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
2070 if (event->state <= PERF_EVENT_STATE_OFF)
2072 if (!event_filter_match(event))
2075 /* may need to reset tstamp_enabled */
2076 if (is_cgroup_event(event))
2077 perf_cgroup_mark_enabled(event, ctx);
2079 if (group_can_go_on(event, cpuctx, 1))
2080 group_sched_in(event, cpuctx, ctx);
2083 * If this pinned group hasn't been scheduled,
2084 * put it in error state.
2086 if (event->state == PERF_EVENT_STATE_INACTIVE) {
2087 update_group_times(event);
2088 event->state = PERF_EVENT_STATE_ERROR;
2094 ctx_flexible_sched_in(struct perf_event_context *ctx,
2095 struct perf_cpu_context *cpuctx)
2097 struct perf_event *event;
2100 list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
2101 /* Ignore events in OFF or ERROR state */
2102 if (event->state <= PERF_EVENT_STATE_OFF)
2105 * Listen to the 'cpu' scheduling filter constraint
2108 if (!event_filter_match(event))
2111 /* may need to reset tstamp_enabled */
2112 if (is_cgroup_event(event))
2113 perf_cgroup_mark_enabled(event, ctx);
2115 if (group_can_go_on(event, cpuctx, can_add_hw)) {
2116 if (group_sched_in(event, cpuctx, ctx))
2123 ctx_sched_in(struct perf_event_context *ctx,
2124 struct perf_cpu_context *cpuctx,
2125 enum event_type_t event_type,
2126 struct task_struct *task)
2129 int is_active = ctx->is_active;
2131 ctx->is_active |= event_type;
2132 if (likely(!ctx->nr_events))
2136 ctx->timestamp = now;
2137 perf_cgroup_set_timestamp(task, ctx);
2139 * First go through the list and put on any pinned groups
2140 * in order to give them the best chance of going on.
2142 if (!(is_active & EVENT_PINNED) && (event_type & EVENT_PINNED))
2143 ctx_pinned_sched_in(ctx, cpuctx);
2145 /* Then walk through the lower prio flexible groups */
2146 if (!(is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE))
2147 ctx_flexible_sched_in(ctx, cpuctx);
2150 static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
2151 enum event_type_t event_type,
2152 struct task_struct *task)
2154 struct perf_event_context *ctx = &cpuctx->ctx;
2156 ctx_sched_in(ctx, cpuctx, event_type, task);
2159 static void perf_event_context_sched_in(struct perf_event_context *ctx,
2160 struct task_struct *task)
2162 struct perf_cpu_context *cpuctx;
2164 cpuctx = __get_cpu_context(ctx);
2165 if (cpuctx->task_ctx == ctx)
2168 perf_ctx_lock(cpuctx, ctx);
2169 perf_pmu_disable(ctx->pmu);
2171 * We want to keep the following priority order:
2172 * cpu pinned (that don't need to move), task pinned,
2173 * cpu flexible, task flexible.
2175 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
2178 cpuctx->task_ctx = ctx;
2180 perf_event_sched_in(cpuctx, cpuctx->task_ctx, task);
2182 perf_pmu_enable(ctx->pmu);
2183 perf_ctx_unlock(cpuctx, ctx);
2186 * Since these rotations are per-cpu, we need to ensure the
2187 * cpu-context we got scheduled on is actually rotating.
2189 perf_pmu_rotate_start(ctx->pmu);
2193 * Called from scheduler to add the events of the current task
2194 * with interrupts disabled.
2196 * We restore the event value and then enable it.
2198 * This does not protect us against NMI, but enable()
2199 * sets the enabled bit in the control field of event _before_
2200 * accessing the event control register. If a NMI hits, then it will
2201 * keep the event running.
2203 void __perf_event_task_sched_in(struct task_struct *prev,
2204 struct task_struct *task)
2206 struct perf_event_context *ctx;
2209 for_each_task_context_nr(ctxn) {
2210 ctx = task->perf_event_ctxp[ctxn];
2214 perf_event_context_sched_in(ctx, task);
2217 * if cgroup events exist on this CPU, then we need
2218 * to check if we have to switch in PMU state.
2219 * cgroup event are system-wide mode only
2221 if (atomic_read(&__get_cpu_var(perf_cgroup_events)))
2222 perf_cgroup_sched_in(prev, task);
2225 static u64 perf_calculate_period(struct perf_event *event, u64 nsec, u64 count)
2227 u64 frequency = event->attr.sample_freq;
2228 u64 sec = NSEC_PER_SEC;
2229 u64 divisor, dividend;
2231 int count_fls, nsec_fls, frequency_fls, sec_fls;
2233 count_fls = fls64(count);
2234 nsec_fls = fls64(nsec);
2235 frequency_fls = fls64(frequency);
2239 * We got @count in @nsec, with a target of sample_freq HZ
2240 * the target period becomes:
2243 * period = -------------------
2244 * @nsec * sample_freq
2249 * Reduce accuracy by one bit such that @a and @b converge
2250 * to a similar magnitude.
2252 #define REDUCE_FLS(a, b) \
2254 if (a##_fls > b##_fls) { \
2264 * Reduce accuracy until either term fits in a u64, then proceed with
2265 * the other, so that finally we can do a u64/u64 division.
2267 while (count_fls + sec_fls > 64 && nsec_fls + frequency_fls > 64) {
2268 REDUCE_FLS(nsec, frequency);
2269 REDUCE_FLS(sec, count);
2272 if (count_fls + sec_fls > 64) {
2273 divisor = nsec * frequency;
2275 while (count_fls + sec_fls > 64) {
2276 REDUCE_FLS(count, sec);
2280 dividend = count * sec;
2282 dividend = count * sec;
2284 while (nsec_fls + frequency_fls > 64) {
2285 REDUCE_FLS(nsec, frequency);
2289 divisor = nsec * frequency;
2295 return div64_u64(dividend, divisor);
2298 static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count)
2300 struct hw_perf_event *hwc = &event->hw;
2301 s64 period, sample_period;
2304 period = perf_calculate_period(event, nsec, count);
2306 delta = (s64)(period - hwc->sample_period);
2307 delta = (delta + 7) / 8; /* low pass filter */
2309 sample_period = hwc->sample_period + delta;
2314 hwc->sample_period = sample_period;
2316 if (local64_read(&hwc->period_left) > 8*sample_period) {
2317 event->pmu->stop(event, PERF_EF_UPDATE);
2318 local64_set(&hwc->period_left, 0);
2319 event->pmu->start(event, PERF_EF_RELOAD);
2323 static void perf_ctx_adjust_freq(struct perf_event_context *ctx, u64 period)
2325 struct perf_event *event;
2326 struct hw_perf_event *hwc;
2327 u64 interrupts, now;
2330 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
2331 if (event->state != PERF_EVENT_STATE_ACTIVE)
2334 if (!event_filter_match(event))
2339 interrupts = hwc->interrupts;
2340 hwc->interrupts = 0;
2343 * unthrottle events on the tick
2345 if (interrupts == MAX_INTERRUPTS) {
2346 perf_log_throttle(event, 1);
2347 event->pmu->start(event, 0);
2350 if (!event->attr.freq || !event->attr.sample_freq)
2353 event->pmu->read(event);
2354 now = local64_read(&event->count);
2355 delta = now - hwc->freq_count_stamp;
2356 hwc->freq_count_stamp = now;
2359 perf_adjust_period(event, period, delta);
2364 * Round-robin a context's events:
2366 static void rotate_ctx(struct perf_event_context *ctx)
2369 * Rotate the first entry last of non-pinned groups. Rotation might be
2370 * disabled by the inheritance code.
2372 if (!ctx->rotate_disable)
2373 list_rotate_left(&ctx->flexible_groups);
2377 * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
2378 * because they're strictly cpu affine and rotate_start is called with IRQs
2379 * disabled, while rotate_context is called from IRQ context.
2381 static void perf_rotate_context(struct perf_cpu_context *cpuctx)
2383 u64 interval = (u64)cpuctx->jiffies_interval * TICK_NSEC;
2384 struct perf_event_context *ctx = NULL;
2385 int rotate = 0, remove = 1;
2387 if (cpuctx->ctx.nr_events) {
2389 if (cpuctx->ctx.nr_events != cpuctx->ctx.nr_active)
2393 ctx = cpuctx->task_ctx;
2394 if (ctx && ctx->nr_events) {
2396 if (ctx->nr_events != ctx->nr_active)
2400 perf_ctx_lock(cpuctx, cpuctx->task_ctx);
2401 perf_pmu_disable(cpuctx->ctx.pmu);
2402 perf_ctx_adjust_freq(&cpuctx->ctx, interval);
2404 perf_ctx_adjust_freq(ctx, interval);
2409 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
2411 ctx_sched_out(ctx, cpuctx, EVENT_FLEXIBLE);
2413 rotate_ctx(&cpuctx->ctx);
2417 perf_event_sched_in(cpuctx, ctx, current);
2421 list_del_init(&cpuctx->rotation_list);
2423 perf_pmu_enable(cpuctx->ctx.pmu);
2424 perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
2427 void perf_event_task_tick(void)
2429 struct list_head *head = &__get_cpu_var(rotation_list);
2430 struct perf_cpu_context *cpuctx, *tmp;
2432 WARN_ON(!irqs_disabled());
2434 list_for_each_entry_safe(cpuctx, tmp, head, rotation_list) {
2435 if (cpuctx->jiffies_interval == 1 ||
2436 !(jiffies % cpuctx->jiffies_interval))
2437 perf_rotate_context(cpuctx);
2441 static int event_enable_on_exec(struct perf_event *event,
2442 struct perf_event_context *ctx)
2444 if (!event->attr.enable_on_exec)
2447 event->attr.enable_on_exec = 0;
2448 if (event->state >= PERF_EVENT_STATE_INACTIVE)
2451 __perf_event_mark_enabled(event, ctx);
2457 * Enable all of a task's events that have been marked enable-on-exec.
2458 * This expects task == current.
2460 static void perf_event_enable_on_exec(struct perf_event_context *ctx)
2462 struct perf_event *event;
2463 unsigned long flags;
2467 local_irq_save(flags);
2468 if (!ctx || !ctx->nr_events)
2472 * We must ctxsw out cgroup events to avoid conflict
2473 * when invoking perf_task_event_sched_in() later on
2474 * in this function. Otherwise we end up trying to
2475 * ctxswin cgroup events which are already scheduled
2478 perf_cgroup_sched_out(current, NULL);
2480 raw_spin_lock(&ctx->lock);
2481 task_ctx_sched_out(ctx);
2483 list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
2484 ret = event_enable_on_exec(event, ctx);
2489 list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
2490 ret = event_enable_on_exec(event, ctx);
2496 * Unclone this context if we enabled any event.
2501 raw_spin_unlock(&ctx->lock);
2504 * Also calls ctxswin for cgroup events, if any:
2506 perf_event_context_sched_in(ctx, ctx->task);
2508 local_irq_restore(flags);
2512 * Cross CPU call to read the hardware event
2514 static void __perf_event_read(void *info)
2516 struct perf_event *event = info;
2517 struct perf_event_context *ctx = event->ctx;
2518 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
2521 * If this is a task context, we need to check whether it is
2522 * the current task context of this cpu. If not it has been
2523 * scheduled out before the smp call arrived. In that case
2524 * event->count would have been updated to a recent sample
2525 * when the event was scheduled out.
2527 if (ctx->task && cpuctx->task_ctx != ctx)
2530 raw_spin_lock(&ctx->lock);
2531 if (ctx->is_active) {
2532 update_context_time(ctx);
2533 update_cgrp_time_from_event(event);
2535 update_event_times(event);
2536 if (event->state == PERF_EVENT_STATE_ACTIVE)
2537 event->pmu->read(event);
2538 raw_spin_unlock(&ctx->lock);
2541 static inline u64 perf_event_count(struct perf_event *event)
2543 return local64_read(&event->count) + atomic64_read(&event->child_count);
2546 static u64 perf_event_read(struct perf_event *event)
2549 * If event is enabled and currently active on a CPU, update the
2550 * value in the event structure:
2552 if (event->state == PERF_EVENT_STATE_ACTIVE) {
2553 smp_call_function_single(event->oncpu,
2554 __perf_event_read, event, 1);
2555 } else if (event->state == PERF_EVENT_STATE_INACTIVE) {
2556 struct perf_event_context *ctx = event->ctx;
2557 unsigned long flags;
2559 raw_spin_lock_irqsave(&ctx->lock, flags);
2561 * may read while context is not active
2562 * (e.g., thread is blocked), in that case
2563 * we cannot update context time
2565 if (ctx->is_active) {
2566 update_context_time(ctx);
2567 update_cgrp_time_from_event(event);
2569 update_event_times(event);
2570 raw_spin_unlock_irqrestore(&ctx->lock, flags);
2573 return perf_event_count(event);
2580 struct callchain_cpus_entries {
2581 struct rcu_head rcu_head;
2582 struct perf_callchain_entry *cpu_entries[0];
2585 static DEFINE_PER_CPU(int, callchain_recursion[PERF_NR_CONTEXTS]);
2586 static atomic_t nr_callchain_events;
2587 static DEFINE_MUTEX(callchain_mutex);
2588 struct callchain_cpus_entries *callchain_cpus_entries;
2591 __weak void perf_callchain_kernel(struct perf_callchain_entry *entry,
2592 struct pt_regs *regs)
2596 __weak void perf_callchain_user(struct perf_callchain_entry *entry,
2597 struct pt_regs *regs)
2601 static void release_callchain_buffers_rcu(struct rcu_head *head)
2603 struct callchain_cpus_entries *entries;
2606 entries = container_of(head, struct callchain_cpus_entries, rcu_head);
2608 for_each_possible_cpu(cpu)
2609 kfree(entries->cpu_entries[cpu]);
2614 static void release_callchain_buffers(void)
2616 struct callchain_cpus_entries *entries;
2618 entries = callchain_cpus_entries;
2619 rcu_assign_pointer(callchain_cpus_entries, NULL);
2620 call_rcu(&entries->rcu_head, release_callchain_buffers_rcu);
2623 static int alloc_callchain_buffers(void)
2627 struct callchain_cpus_entries *entries;
2630 * We can't use the percpu allocation API for data that can be
2631 * accessed from NMI. Use a temporary manual per cpu allocation
2632 * until that gets sorted out.
2634 size = offsetof(struct callchain_cpus_entries, cpu_entries[nr_cpu_ids]);
2636 entries = kzalloc(size, GFP_KERNEL);
2640 size = sizeof(struct perf_callchain_entry) * PERF_NR_CONTEXTS;
2642 for_each_possible_cpu(cpu) {
2643 entries->cpu_entries[cpu] = kmalloc_node(size, GFP_KERNEL,
2645 if (!entries->cpu_entries[cpu])
2649 rcu_assign_pointer(callchain_cpus_entries, entries);
2654 for_each_possible_cpu(cpu)
2655 kfree(entries->cpu_entries[cpu]);
2661 static int get_callchain_buffers(void)
2666 mutex_lock(&callchain_mutex);
2668 count = atomic_inc_return(&nr_callchain_events);
2669 if (WARN_ON_ONCE(count < 1)) {
2675 /* If the allocation failed, give up */
2676 if (!callchain_cpus_entries)
2681 err = alloc_callchain_buffers();
2683 release_callchain_buffers();
2685 mutex_unlock(&callchain_mutex);
2690 static void put_callchain_buffers(void)
2692 if (atomic_dec_and_mutex_lock(&nr_callchain_events, &callchain_mutex)) {
2693 release_callchain_buffers();
2694 mutex_unlock(&callchain_mutex);
2698 static int get_recursion_context(int *recursion)
2706 else if (in_softirq())
2711 if (recursion[rctx])
2720 static inline void put_recursion_context(int *recursion, int rctx)
2726 static struct perf_callchain_entry *get_callchain_entry(int *rctx)
2729 struct callchain_cpus_entries *entries;
2731 *rctx = get_recursion_context(__get_cpu_var(callchain_recursion));
2735 entries = rcu_dereference(callchain_cpus_entries);
2739 cpu = smp_processor_id();
2741 return &entries->cpu_entries[cpu][*rctx];
2745 put_callchain_entry(int rctx)
2747 put_recursion_context(__get_cpu_var(callchain_recursion), rctx);
2750 static struct perf_callchain_entry *perf_callchain(struct pt_regs *regs)
2753 struct perf_callchain_entry *entry;
2756 entry = get_callchain_entry(&rctx);
2765 if (!user_mode(regs)) {
2766 perf_callchain_store(entry, PERF_CONTEXT_KERNEL);
2767 perf_callchain_kernel(entry, regs);
2769 regs = task_pt_regs(current);
2775 perf_callchain_store(entry, PERF_CONTEXT_USER);
2776 perf_callchain_user(entry, regs);
2780 put_callchain_entry(rctx);
2786 * Initialize the perf_event context in a task_struct:
2788 static void __perf_event_init_context(struct perf_event_context *ctx)
2790 raw_spin_lock_init(&ctx->lock);
2791 mutex_init(&ctx->mutex);
2792 INIT_LIST_HEAD(&ctx->pinned_groups);
2793 INIT_LIST_HEAD(&ctx->flexible_groups);
2794 INIT_LIST_HEAD(&ctx->event_list);
2795 atomic_set(&ctx->refcount, 1);
2798 static struct perf_event_context *
2799 alloc_perf_context(struct pmu *pmu, struct task_struct *task)
2801 struct perf_event_context *ctx;
2803 ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL);
2807 __perf_event_init_context(ctx);
2810 get_task_struct(task);
2817 static struct task_struct *
2818 find_lively_task_by_vpid(pid_t vpid)
2820 struct task_struct *task;
2827 task = find_task_by_vpid(vpid);
2829 get_task_struct(task);
2833 return ERR_PTR(-ESRCH);
2835 /* Reuse ptrace permission checks for now. */
2837 if (!ptrace_may_access(task, PTRACE_MODE_READ))
2842 put_task_struct(task);
2843 return ERR_PTR(err);
2848 * Returns a matching context with refcount and pincount.
2850 static struct perf_event_context *
2851 find_get_context(struct pmu *pmu, struct task_struct *task, int cpu)
2853 struct perf_event_context *ctx;
2854 struct perf_cpu_context *cpuctx;
2855 unsigned long flags;
2859 /* Must be root to operate on a CPU event: */
2860 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
2861 return ERR_PTR(-EACCES);
2864 * We could be clever and allow to attach a event to an
2865 * offline CPU and activate it when the CPU comes up, but
2868 if (!cpu_online(cpu))
2869 return ERR_PTR(-ENODEV);
2871 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
2880 ctxn = pmu->task_ctx_nr;
2885 ctx = perf_lock_task_context(task, ctxn, &flags);
2889 raw_spin_unlock_irqrestore(&ctx->lock, flags);
2891 ctx = alloc_perf_context(pmu, task);
2897 mutex_lock(&task->perf_event_mutex);
2899 * If it has already passed perf_event_exit_task().
2900 * we must see PF_EXITING, it takes this mutex too.
2902 if (task->flags & PF_EXITING)
2904 else if (task->perf_event_ctxp[ctxn])
2909 rcu_assign_pointer(task->perf_event_ctxp[ctxn], ctx);
2911 mutex_unlock(&task->perf_event_mutex);
2913 if (unlikely(err)) {
2925 return ERR_PTR(err);
2928 static void perf_event_free_filter(struct perf_event *event);
2930 static void free_event_rcu(struct rcu_head *head)
2932 struct perf_event *event;
2934 event = container_of(head, struct perf_event, rcu_head);
2936 put_pid_ns(event->ns);
2937 perf_event_free_filter(event);
2941 static void ring_buffer_put(struct ring_buffer *rb);
2943 static void free_event(struct perf_event *event)
2945 irq_work_sync(&event->pending);
2947 if (!event->parent) {
2948 if (event->attach_state & PERF_ATTACH_TASK)
2949 jump_label_dec(&perf_sched_events);
2950 if (event->attr.mmap || event->attr.mmap_data)
2951 atomic_dec(&nr_mmap_events);
2952 if (event->attr.comm)
2953 atomic_dec(&nr_comm_events);
2954 if (event->attr.task)
2955 atomic_dec(&nr_task_events);
2956 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN)
2957 put_callchain_buffers();
2958 if (is_cgroup_event(event)) {
2959 atomic_dec(&per_cpu(perf_cgroup_events, event->cpu));
2960 jump_label_dec(&perf_sched_events);
2965 ring_buffer_put(event->rb);
2969 if (is_cgroup_event(event))
2970 perf_detach_cgroup(event);
2973 event->destroy(event);
2976 put_ctx(event->ctx);
2978 call_rcu(&event->rcu_head, free_event_rcu);
2981 int perf_event_release_kernel(struct perf_event *event)
2983 struct perf_event_context *ctx = event->ctx;
2985 WARN_ON_ONCE(ctx->parent_ctx);
2987 * There are two ways this annotation is useful:
2989 * 1) there is a lock recursion from perf_event_exit_task
2990 * see the comment there.
2992 * 2) there is a lock-inversion with mmap_sem through
2993 * perf_event_read_group(), which takes faults while
2994 * holding ctx->mutex, however this is called after
2995 * the last filedesc died, so there is no possibility
2996 * to trigger the AB-BA case.
2998 mutex_lock_nested(&ctx->mutex, SINGLE_DEPTH_NESTING);
2999 raw_spin_lock_irq(&ctx->lock);
3000 perf_group_detach(event);
3001 raw_spin_unlock_irq(&ctx->lock);
3002 perf_remove_from_context(event);
3003 mutex_unlock(&ctx->mutex);
3009 EXPORT_SYMBOL_GPL(perf_event_release_kernel);
3012 * Called when the last reference to the file is gone.
3014 static void put_event(struct perf_event *event)
3016 struct task_struct *owner;
3018 if (!atomic_long_dec_and_test(&event->refcount))
3022 owner = ACCESS_ONCE(event->owner);
3024 * Matches the smp_wmb() in perf_event_exit_task(). If we observe
3025 * !owner it means the list deletion is complete and we can indeed
3026 * free this event, otherwise we need to serialize on
3027 * owner->perf_event_mutex.
3029 smp_read_barrier_depends();
3032 * Since delayed_put_task_struct() also drops the last
3033 * task reference we can safely take a new reference
3034 * while holding the rcu_read_lock().
3036 get_task_struct(owner);
3041 mutex_lock(&owner->perf_event_mutex);
3043 * We have to re-check the event->owner field, if it is cleared
3044 * we raced with perf_event_exit_task(), acquiring the mutex
3045 * ensured they're done, and we can proceed with freeing the
3049 list_del_init(&event->owner_entry);
3050 mutex_unlock(&owner->perf_event_mutex);
3051 put_task_struct(owner);
3054 perf_event_release_kernel(event);
3057 static int perf_release(struct inode *inode, struct file *file)
3059 put_event(file->private_data);
3063 u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running)
3065 struct perf_event *child;
3071 mutex_lock(&event->child_mutex);
3072 total += perf_event_read(event);
3073 *enabled += event->total_time_enabled +
3074 atomic64_read(&event->child_total_time_enabled);
3075 *running += event->total_time_running +
3076 atomic64_read(&event->child_total_time_running);
3078 list_for_each_entry(child, &event->child_list, child_list) {
3079 total += perf_event_read(child);
3080 *enabled += child->total_time_enabled;
3081 *running += child->total_time_running;
3083 mutex_unlock(&event->child_mutex);
3087 EXPORT_SYMBOL_GPL(perf_event_read_value);
3089 static int perf_event_read_group(struct perf_event *event,
3090 u64 read_format, char __user *buf)
3092 struct perf_event *leader = event->group_leader, *sub;
3093 int n = 0, size = 0, ret = -EFAULT;
3094 struct perf_event_context *ctx = leader->ctx;
3096 u64 count, enabled, running;
3098 mutex_lock(&ctx->mutex);
3099 count = perf_event_read_value(leader, &enabled, &running);
3101 values[n++] = 1 + leader->nr_siblings;
3102 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3103 values[n++] = enabled;
3104 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3105 values[n++] = running;
3106 values[n++] = count;
3107 if (read_format & PERF_FORMAT_ID)
3108 values[n++] = primary_event_id(leader);
3110 size = n * sizeof(u64);
3112 if (copy_to_user(buf, values, size))
3117 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
3120 values[n++] = perf_event_read_value(sub, &enabled, &running);
3121 if (read_format & PERF_FORMAT_ID)
3122 values[n++] = primary_event_id(sub);
3124 size = n * sizeof(u64);
3126 if (copy_to_user(buf + ret, values, size)) {
3134 mutex_unlock(&ctx->mutex);
3139 static int perf_event_read_one(struct perf_event *event,
3140 u64 read_format, char __user *buf)
3142 u64 enabled, running;
3146 values[n++] = perf_event_read_value(event, &enabled, &running);
3147 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3148 values[n++] = enabled;
3149 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3150 values[n++] = running;
3151 if (read_format & PERF_FORMAT_ID)
3152 values[n++] = primary_event_id(event);
3154 if (copy_to_user(buf, values, n * sizeof(u64)))
3157 return n * sizeof(u64);
3161 * Read the performance event - simple non blocking version for now
3164 perf_read_hw(struct perf_event *event, char __user *buf, size_t count)
3166 u64 read_format = event->attr.read_format;
3170 * Return end-of-file for a read on a event that is in
3171 * error state (i.e. because it was pinned but it couldn't be
3172 * scheduled on to the CPU at some point).
3174 if (event->state == PERF_EVENT_STATE_ERROR)
3177 if (count < event->read_size)
3180 WARN_ON_ONCE(event->ctx->parent_ctx);
3181 if (read_format & PERF_FORMAT_GROUP)
3182 ret = perf_event_read_group(event, read_format, buf);
3184 ret = perf_event_read_one(event, read_format, buf);
3190 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
3192 struct perf_event *event = file->private_data;
3194 return perf_read_hw(event, buf, count);
3197 static unsigned int perf_poll(struct file *file, poll_table *wait)
3199 struct perf_event *event = file->private_data;
3200 struct ring_buffer *rb;
3201 unsigned int events = POLL_HUP;
3204 * Race between perf_event_set_output() and perf_poll(): perf_poll()
3205 * grabs the rb reference but perf_event_set_output() overrides it.
3206 * Here is the timeline for two threads T1, T2:
3207 * t0: T1, rb = rcu_dereference(event->rb)
3208 * t1: T2, old_rb = event->rb
3209 * t2: T2, event->rb = new rb
3210 * t3: T2, ring_buffer_detach(old_rb)
3211 * t4: T1, ring_buffer_attach(rb1)
3212 * t5: T1, poll_wait(event->waitq)
3214 * To avoid this problem, we grab mmap_mutex in perf_poll()
3215 * thereby ensuring that the assignment of the new ring buffer
3216 * and the detachment of the old buffer appear atomic to perf_poll()
3218 mutex_lock(&event->mmap_mutex);
3221 rb = rcu_dereference(event->rb);
3223 ring_buffer_attach(event, rb);
3224 events = atomic_xchg(&rb->poll, 0);
3228 mutex_unlock(&event->mmap_mutex);
3230 poll_wait(file, &event->waitq, wait);
3235 static void perf_event_reset(struct perf_event *event)
3237 (void)perf_event_read(event);
3238 local64_set(&event->count, 0);
3239 perf_event_update_userpage(event);
3243 * Holding the top-level event's child_mutex means that any
3244 * descendant process that has inherited this event will block
3245 * in sync_child_event if it goes to exit, thus satisfying the
3246 * task existence requirements of perf_event_enable/disable.
3248 static void perf_event_for_each_child(struct perf_event *event,
3249 void (*func)(struct perf_event *))
3251 struct perf_event *child;
3253 WARN_ON_ONCE(event->ctx->parent_ctx);
3254 mutex_lock(&event->child_mutex);
3256 list_for_each_entry(child, &event->child_list, child_list)
3258 mutex_unlock(&event->child_mutex);
3261 static void perf_event_for_each(struct perf_event *event,
3262 void (*func)(struct perf_event *))
3264 struct perf_event_context *ctx = event->ctx;
3265 struct perf_event *sibling;
3267 WARN_ON_ONCE(ctx->parent_ctx);
3268 mutex_lock(&ctx->mutex);
3269 event = event->group_leader;
3271 perf_event_for_each_child(event, func);
3273 list_for_each_entry(sibling, &event->sibling_list, group_entry)
3274 perf_event_for_each_child(event, func);
3275 mutex_unlock(&ctx->mutex);
3278 static int perf_event_period(struct perf_event *event, u64 __user *arg)
3280 struct perf_event_context *ctx = event->ctx;
3284 if (!is_sampling_event(event))
3287 if (copy_from_user(&value, arg, sizeof(value)))
3293 raw_spin_lock_irq(&ctx->lock);
3294 if (event->attr.freq) {
3295 if (value > sysctl_perf_event_sample_rate) {
3300 event->attr.sample_freq = value;
3302 event->attr.sample_period = value;
3303 event->hw.sample_period = value;
3306 raw_spin_unlock_irq(&ctx->lock);
3311 static const struct file_operations perf_fops;
3313 static struct file *perf_fget_light(int fd, int *fput_needed)
3317 file = fget_light(fd, fput_needed);
3319 return ERR_PTR(-EBADF);
3321 if (file->f_op != &perf_fops) {
3322 fput_light(file, *fput_needed);
3324 return ERR_PTR(-EBADF);
3330 static int perf_event_set_output(struct perf_event *event,
3331 struct perf_event *output_event);
3332 static int perf_event_set_filter(struct perf_event *event, void __user *arg);
3334 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
3336 struct perf_event *event = file->private_data;
3337 void (*func)(struct perf_event *);
3341 case PERF_EVENT_IOC_ENABLE:
3342 func = perf_event_enable;
3344 case PERF_EVENT_IOC_DISABLE:
3345 func = perf_event_disable;
3347 case PERF_EVENT_IOC_RESET:
3348 func = perf_event_reset;
3351 case PERF_EVENT_IOC_REFRESH:
3352 return perf_event_refresh(event, arg);
3354 case PERF_EVENT_IOC_PERIOD:
3355 return perf_event_period(event, (u64 __user *)arg);
3357 case PERF_EVENT_IOC_SET_OUTPUT:
3359 struct file *output_file = NULL;
3360 struct perf_event *output_event = NULL;
3361 int fput_needed = 0;
3365 output_file = perf_fget_light(arg, &fput_needed);
3366 if (IS_ERR(output_file))
3367 return PTR_ERR(output_file);
3368 output_event = output_file->private_data;
3371 ret = perf_event_set_output(event, output_event);
3373 fput_light(output_file, fput_needed);
3378 case PERF_EVENT_IOC_SET_FILTER:
3379 return perf_event_set_filter(event, (void __user *)arg);
3385 if (flags & PERF_IOC_FLAG_GROUP)
3386 perf_event_for_each(event, func);
3388 perf_event_for_each_child(event, func);
3393 int perf_event_task_enable(void)
3395 struct perf_event *event;
3397 mutex_lock(¤t->perf_event_mutex);
3398 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
3399 perf_event_for_each_child(event, perf_event_enable);
3400 mutex_unlock(¤t->perf_event_mutex);
3405 int perf_event_task_disable(void)
3407 struct perf_event *event;
3409 mutex_lock(¤t->perf_event_mutex);
3410 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
3411 perf_event_for_each_child(event, perf_event_disable);
3412 mutex_unlock(¤t->perf_event_mutex);
3417 #ifndef PERF_EVENT_INDEX_OFFSET
3418 # define PERF_EVENT_INDEX_OFFSET 0
3421 static int perf_event_index(struct perf_event *event)
3423 if (event->hw.state & PERF_HES_STOPPED)
3426 if (event->state != PERF_EVENT_STATE_ACTIVE)
3429 return event->hw.idx + 1 - PERF_EVENT_INDEX_OFFSET;
3432 static void calc_timer_values(struct perf_event *event,
3439 ctx_time = event->shadow_ctx_time + now;
3440 *enabled = ctx_time - event->tstamp_enabled;
3441 *running = ctx_time - event->tstamp_running;
3445 * Callers need to ensure there can be no nesting of this function, otherwise
3446 * the seqlock logic goes bad. We can not serialize this because the arch
3447 * code calls this from NMI context.
3449 void perf_event_update_userpage(struct perf_event *event)
3451 struct perf_event_mmap_page *userpg;
3452 struct ring_buffer *rb;
3453 u64 enabled, running;
3457 * compute total_time_enabled, total_time_running
3458 * based on snapshot values taken when the event
3459 * was last scheduled in.
3461 * we cannot simply called update_context_time()
3462 * because of locking issue as we can be called in
3465 calc_timer_values(event, &enabled, &running);
3466 rb = rcu_dereference(event->rb);
3470 userpg = rb->user_page;
3473 * Disable preemption so as to not let the corresponding user-space
3474 * spin too long if we get preempted.
3479 userpg->index = perf_event_index(event);
3480 userpg->offset = perf_event_count(event);
3481 if (event->state == PERF_EVENT_STATE_ACTIVE)
3482 userpg->offset -= local64_read(&event->hw.prev_count);
3484 userpg->time_enabled = enabled +
3485 atomic64_read(&event->child_total_time_enabled);
3487 userpg->time_running = running +
3488 atomic64_read(&event->child_total_time_running);
3497 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
3499 struct perf_event *event = vma->vm_file->private_data;
3500 struct ring_buffer *rb;
3501 int ret = VM_FAULT_SIGBUS;
3503 if (vmf->flags & FAULT_FLAG_MKWRITE) {
3504 if (vmf->pgoff == 0)
3510 rb = rcu_dereference(event->rb);
3514 if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE))
3517 vmf->page = perf_mmap_to_page(rb, vmf->pgoff);
3521 get_page(vmf->page);
3522 vmf->page->mapping = vma->vm_file->f_mapping;
3523 vmf->page->index = vmf->pgoff;
3532 static void ring_buffer_attach(struct perf_event *event,
3533 struct ring_buffer *rb)
3535 unsigned long flags;
3537 if (!list_empty(&event->rb_entry))
3540 spin_lock_irqsave(&rb->event_lock, flags);
3541 if (!list_empty(&event->rb_entry))
3544 list_add(&event->rb_entry, &rb->event_list);
3546 spin_unlock_irqrestore(&rb->event_lock, flags);
3549 static void ring_buffer_detach(struct perf_event *event,
3550 struct ring_buffer *rb)
3552 unsigned long flags;
3554 if (list_empty(&event->rb_entry))
3557 spin_lock_irqsave(&rb->event_lock, flags);
3558 list_del_init(&event->rb_entry);
3559 wake_up_all(&event->waitq);
3560 spin_unlock_irqrestore(&rb->event_lock, flags);
3563 static void ring_buffer_wakeup(struct perf_event *event)
3565 struct ring_buffer *rb;
3568 rb = rcu_dereference(event->rb);
3572 list_for_each_entry_rcu(event, &rb->event_list, rb_entry)
3573 wake_up_all(&event->waitq);
3579 static void rb_free_rcu(struct rcu_head *rcu_head)
3581 struct ring_buffer *rb;
3583 rb = container_of(rcu_head, struct ring_buffer, rcu_head);
3587 static struct ring_buffer *ring_buffer_get(struct perf_event *event)
3589 struct ring_buffer *rb;
3592 rb = rcu_dereference(event->rb);
3594 if (!atomic_inc_not_zero(&rb->refcount))
3602 static void ring_buffer_put(struct ring_buffer *rb)
3604 struct perf_event *event, *n;
3605 unsigned long flags;
3607 if (!atomic_dec_and_test(&rb->refcount))
3610 spin_lock_irqsave(&rb->event_lock, flags);
3611 list_for_each_entry_safe(event, n, &rb->event_list, rb_entry) {
3612 list_del_init(&event->rb_entry);
3613 wake_up_all(&event->waitq);
3615 spin_unlock_irqrestore(&rb->event_lock, flags);
3617 call_rcu(&rb->rcu_head, rb_free_rcu);
3620 static void perf_mmap_open(struct vm_area_struct *vma)
3622 struct perf_event *event = vma->vm_file->private_data;
3624 atomic_inc(&event->mmap_count);
3627 static void perf_mmap_close(struct vm_area_struct *vma)
3629 struct perf_event *event = vma->vm_file->private_data;
3631 if (atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex)) {
3632 unsigned long size = perf_data_size(event->rb);
3633 struct user_struct *user = event->mmap_user;
3634 struct ring_buffer *rb = event->rb;
3636 atomic_long_sub((size >> PAGE_SHIFT) + 1, &user->locked_vm);
3637 vma->vm_mm->pinned_vm -= event->mmap_locked;
3638 rcu_assign_pointer(event->rb, NULL);
3639 ring_buffer_detach(event, rb);
3640 mutex_unlock(&event->mmap_mutex);
3642 ring_buffer_put(rb);
3647 static const struct vm_operations_struct perf_mmap_vmops = {
3648 .open = perf_mmap_open,
3649 .close = perf_mmap_close,
3650 .fault = perf_mmap_fault,
3651 .page_mkwrite = perf_mmap_fault,
3654 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
3656 struct perf_event *event = file->private_data;
3657 unsigned long user_locked, user_lock_limit;
3658 struct user_struct *user = current_user();
3659 unsigned long locked, lock_limit;
3660 struct ring_buffer *rb;
3661 unsigned long vma_size;
3662 unsigned long nr_pages;
3663 long user_extra, extra;
3664 int ret = 0, flags = 0;
3667 * Don't allow mmap() of inherited per-task counters. This would
3668 * create a performance issue due to all children writing to the
3671 if (event->cpu == -1 && event->attr.inherit)
3674 if (!(vma->vm_flags & VM_SHARED))
3677 vma_size = vma->vm_end - vma->vm_start;
3678 nr_pages = (vma_size / PAGE_SIZE) - 1;
3681 * If we have rb pages ensure they're a power-of-two number, so we
3682 * can do bitmasks instead of modulo.
3684 if (nr_pages != 0 && !is_power_of_2(nr_pages))
3687 if (vma_size != PAGE_SIZE * (1 + nr_pages))
3690 if (vma->vm_pgoff != 0)
3693 WARN_ON_ONCE(event->ctx->parent_ctx);
3694 mutex_lock(&event->mmap_mutex);
3696 if (event->rb->nr_pages == nr_pages)
3697 atomic_inc(&event->rb->refcount);
3703 user_extra = nr_pages + 1;
3704 user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10);
3707 * Increase the limit linearly with more CPUs:
3709 user_lock_limit *= num_online_cpus();
3711 user_locked = atomic_long_read(&user->locked_vm) + user_extra;
3714 if (user_locked > user_lock_limit)
3715 extra = user_locked - user_lock_limit;
3717 lock_limit = rlimit(RLIMIT_MEMLOCK);
3718 lock_limit >>= PAGE_SHIFT;
3719 locked = vma->vm_mm->pinned_vm + extra;
3721 if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() &&
3722 !capable(CAP_IPC_LOCK)) {
3729 if (vma->vm_flags & VM_WRITE)
3730 flags |= RING_BUFFER_WRITABLE;
3732 rb = rb_alloc(nr_pages,
3733 event->attr.watermark ? event->attr.wakeup_watermark : 0,
3740 rcu_assign_pointer(event->rb, rb);
3742 atomic_long_add(user_extra, &user->locked_vm);
3743 event->mmap_locked = extra;
3744 event->mmap_user = get_current_user();
3745 vma->vm_mm->pinned_vm += event->mmap_locked;
3749 atomic_inc(&event->mmap_count);
3750 mutex_unlock(&event->mmap_mutex);
3752 vma->vm_flags |= VM_RESERVED;
3753 vma->vm_ops = &perf_mmap_vmops;
3758 static int perf_fasync(int fd, struct file *filp, int on)
3760 struct inode *inode = filp->f_path.dentry->d_inode;
3761 struct perf_event *event = filp->private_data;
3764 mutex_lock(&inode->i_mutex);
3765 retval = fasync_helper(fd, filp, on, &event->fasync);
3766 mutex_unlock(&inode->i_mutex);
3774 static const struct file_operations perf_fops = {
3775 .llseek = no_llseek,
3776 .release = perf_release,
3779 .unlocked_ioctl = perf_ioctl,
3780 .compat_ioctl = perf_ioctl,
3782 .fasync = perf_fasync,
3788 * If there's data, ensure we set the poll() state and publish everything
3789 * to user-space before waking everybody up.
3792 void perf_event_wakeup(struct perf_event *event)
3794 ring_buffer_wakeup(event);
3796 if (event->pending_kill) {
3797 kill_fasync(&event->fasync, SIGIO, event->pending_kill);
3798 event->pending_kill = 0;
3802 static void perf_pending_event(struct irq_work *entry)
3804 struct perf_event *event = container_of(entry,
3805 struct perf_event, pending);
3807 if (event->pending_disable) {
3808 event->pending_disable = 0;
3809 __perf_event_disable(event);
3812 if (event->pending_wakeup) {
3813 event->pending_wakeup = 0;
3814 perf_event_wakeup(event);
3819 * We assume there is only KVM supporting the callbacks.
3820 * Later on, we might change it to a list if there is
3821 * another virtualization implementation supporting the callbacks.
3823 struct perf_guest_info_callbacks *perf_guest_cbs;
3825 int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
3827 perf_guest_cbs = cbs;
3830 EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks);
3832 int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
3834 perf_guest_cbs = NULL;
3837 EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks);
3839 static void __perf_event_header__init_id(struct perf_event_header *header,
3840 struct perf_sample_data *data,
3841 struct perf_event *event)
3843 u64 sample_type = event->attr.sample_type;
3845 data->type = sample_type;
3846 header->size += event->id_header_size;
3848 if (sample_type & PERF_SAMPLE_TID) {
3849 /* namespace issues */
3850 data->tid_entry.pid = perf_event_pid(event, current);
3851 data->tid_entry.tid = perf_event_tid(event, current);
3854 if (sample_type & PERF_SAMPLE_TIME)
3855 data->time = perf_clock();
3857 if (sample_type & PERF_SAMPLE_ID)
3858 data->id = primary_event_id(event);
3860 if (sample_type & PERF_SAMPLE_STREAM_ID)
3861 data->stream_id = event->id;
3863 if (sample_type & PERF_SAMPLE_CPU) {
3864 data->cpu_entry.cpu = raw_smp_processor_id();
3865 data->cpu_entry.reserved = 0;
3869 void perf_event_header__init_id(struct perf_event_header *header,
3870 struct perf_sample_data *data,
3871 struct perf_event *event)
3873 if (event->attr.sample_id_all)
3874 __perf_event_header__init_id(header, data, event);
3877 static void __perf_event__output_id_sample(struct perf_output_handle *handle,
3878 struct perf_sample_data *data)
3880 u64 sample_type = data->type;
3882 if (sample_type & PERF_SAMPLE_TID)
3883 perf_output_put(handle, data->tid_entry);
3885 if (sample_type & PERF_SAMPLE_TIME)
3886 perf_output_put(handle, data->time);
3888 if (sample_type & PERF_SAMPLE_ID)
3889 perf_output_put(handle, data->id);
3891 if (sample_type & PERF_SAMPLE_STREAM_ID)
3892 perf_output_put(handle, data->stream_id);
3894 if (sample_type & PERF_SAMPLE_CPU)
3895 perf_output_put(handle, data->cpu_entry);
3898 void perf_event__output_id_sample(struct perf_event *event,
3899 struct perf_output_handle *handle,
3900 struct perf_sample_data *sample)
3902 if (event->attr.sample_id_all)
3903 __perf_event__output_id_sample(handle, sample);
3906 static void perf_output_read_one(struct perf_output_handle *handle,
3907 struct perf_event *event,
3908 u64 enabled, u64 running)
3910 u64 read_format = event->attr.read_format;
3914 values[n++] = perf_event_count(event);
3915 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
3916 values[n++] = enabled +
3917 atomic64_read(&event->child_total_time_enabled);
3919 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
3920 values[n++] = running +
3921 atomic64_read(&event->child_total_time_running);
3923 if (read_format & PERF_FORMAT_ID)
3924 values[n++] = primary_event_id(event);
3926 __output_copy(handle, values, n * sizeof(u64));
3930 * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
3932 static void perf_output_read_group(struct perf_output_handle *handle,
3933 struct perf_event *event,
3934 u64 enabled, u64 running)
3936 struct perf_event *leader = event->group_leader, *sub;
3937 u64 read_format = event->attr.read_format;
3941 values[n++] = 1 + leader->nr_siblings;
3943 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3944 values[n++] = enabled;
3946 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3947 values[n++] = running;
3949 if (leader != event)
3950 leader->pmu->read(leader);
3952 values[n++] = perf_event_count(leader);
3953 if (read_format & PERF_FORMAT_ID)
3954 values[n++] = primary_event_id(leader);
3956 __output_copy(handle, values, n * sizeof(u64));
3958 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
3962 sub->pmu->read(sub);
3964 values[n++] = perf_event_count(sub);
3965 if (read_format & PERF_FORMAT_ID)
3966 values[n++] = primary_event_id(sub);
3968 __output_copy(handle, values, n * sizeof(u64));
3972 #define PERF_FORMAT_TOTAL_TIMES (PERF_FORMAT_TOTAL_TIME_ENABLED|\
3973 PERF_FORMAT_TOTAL_TIME_RUNNING)
3975 static void perf_output_read(struct perf_output_handle *handle,
3976 struct perf_event *event)
3978 u64 enabled = 0, running = 0;
3979 u64 read_format = event->attr.read_format;
3982 * compute total_time_enabled, total_time_running
3983 * based on snapshot values taken when the event
3984 * was last scheduled in.
3986 * we cannot simply called update_context_time()
3987 * because of locking issue as we are called in
3990 if (read_format & PERF_FORMAT_TOTAL_TIMES)
3991 calc_timer_values(event, &enabled, &running);
3993 if (event->attr.read_format & PERF_FORMAT_GROUP)
3994 perf_output_read_group(handle, event, enabled, running);
3996 perf_output_read_one(handle, event, enabled, running);
3999 void perf_output_sample(struct perf_output_handle *handle,
4000 struct perf_event_header *header,
4001 struct perf_sample_data *data,
4002 struct perf_event *event)
4004 u64 sample_type = data->type;
4006 perf_output_put(handle, *header);
4008 if (sample_type & PERF_SAMPLE_IP)
4009 perf_output_put(handle, data->ip);
4011 if (sample_type & PERF_SAMPLE_TID)
4012 perf_output_put(handle, data->tid_entry);
4014 if (sample_type & PERF_SAMPLE_TIME)
4015 perf_output_put(handle, data->time);
4017 if (sample_type & PERF_SAMPLE_ADDR)
4018 perf_output_put(handle, data->addr);
4020 if (sample_type & PERF_SAMPLE_ID)
4021 perf_output_put(handle, data->id);
4023 if (sample_type & PERF_SAMPLE_STREAM_ID)
4024 perf_output_put(handle, data->stream_id);
4026 if (sample_type & PERF_SAMPLE_CPU)
4027 perf_output_put(handle, data->cpu_entry);
4029 if (sample_type & PERF_SAMPLE_PERIOD)
4030 perf_output_put(handle, data->period);
4032 if (sample_type & PERF_SAMPLE_READ)
4033 perf_output_read(handle, event);
4035 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
4036 if (data->callchain) {
4039 if (data->callchain)
4040 size += data->callchain->nr;
4042 size *= sizeof(u64);
4044 __output_copy(handle, data->callchain, size);
4047 perf_output_put(handle, nr);
4051 if (sample_type & PERF_SAMPLE_RAW) {
4053 perf_output_put(handle, data->raw->size);
4054 __output_copy(handle, data->raw->data,
4061 .size = sizeof(u32),
4064 perf_output_put(handle, raw);
4068 if (!event->attr.watermark) {
4069 int wakeup_events = event->attr.wakeup_events;
4071 if (wakeup_events) {
4072 struct ring_buffer *rb = handle->rb;
4073 int events = local_inc_return(&rb->events);
4075 if (events >= wakeup_events) {
4076 local_sub(wakeup_events, &rb->events);
4077 local_inc(&rb->wakeup);
4083 void perf_prepare_sample(struct perf_event_header *header,
4084 struct perf_sample_data *data,
4085 struct perf_event *event,
4086 struct pt_regs *regs)
4088 u64 sample_type = event->attr.sample_type;
4090 header->type = PERF_RECORD_SAMPLE;
4091 header->size = sizeof(*header) + event->header_size;
4094 header->misc |= perf_misc_flags(regs);
4096 __perf_event_header__init_id(header, data, event);
4098 if (sample_type & PERF_SAMPLE_IP)
4099 data->ip = perf_instruction_pointer(regs);
4101 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
4104 data->callchain = perf_callchain(regs);
4106 if (data->callchain)
4107 size += data->callchain->nr;
4109 header->size += size * sizeof(u64);
4112 if (sample_type & PERF_SAMPLE_RAW) {
4113 int size = sizeof(u32);
4116 size += data->raw->size;
4118 size += sizeof(u32);
4120 WARN_ON_ONCE(size & (sizeof(u64)-1));
4121 header->size += size;
4125 static void perf_event_output(struct perf_event *event,
4126 struct perf_sample_data *data,
4127 struct pt_regs *regs)
4129 struct perf_output_handle handle;
4130 struct perf_event_header header;
4132 /* protect the callchain buffers */
4135 perf_prepare_sample(&header, data, event, regs);
4137 if (perf_output_begin(&handle, event, header.size))
4140 perf_output_sample(&handle, &header, data, event);
4142 perf_output_end(&handle);
4152 struct perf_read_event {
4153 struct perf_event_header header;
4160 perf_event_read_event(struct perf_event *event,
4161 struct task_struct *task)
4163 struct perf_output_handle handle;
4164 struct perf_sample_data sample;
4165 struct perf_read_event read_event = {
4167 .type = PERF_RECORD_READ,
4169 .size = sizeof(read_event) + event->read_size,
4171 .pid = perf_event_pid(event, task),
4172 .tid = perf_event_tid(event, task),
4176 perf_event_header__init_id(&read_event.header, &sample, event);
4177 ret = perf_output_begin(&handle, event, read_event.header.size);
4181 perf_output_put(&handle, read_event);
4182 perf_output_read(&handle, event);
4183 perf_event__output_id_sample(event, &handle, &sample);
4185 perf_output_end(&handle);
4189 * task tracking -- fork/exit
4191 * enabled by: attr.comm | attr.mmap | attr.mmap_data | attr.task
4194 struct perf_task_event {
4195 struct task_struct *task;
4196 struct perf_event_context *task_ctx;
4199 struct perf_event_header header;
4209 static void perf_event_task_output(struct perf_event *event,
4210 struct perf_task_event *task_event)
4212 struct perf_output_handle handle;
4213 struct perf_sample_data sample;
4214 struct task_struct *task = task_event->task;
4215 int ret, size = task_event->event_id.header.size;
4217 perf_event_header__init_id(&task_event->event_id.header, &sample, event);
4219 ret = perf_output_begin(&handle, event,
4220 task_event->event_id.header.size);
4224 task_event->event_id.pid = perf_event_pid(event, task);
4225 task_event->event_id.ppid = perf_event_pid(event, current);
4227 task_event->event_id.tid = perf_event_tid(event, task);
4228 task_event->event_id.ptid = perf_event_tid(event, current);
4230 perf_output_put(&handle, task_event->event_id);
4232 perf_event__output_id_sample(event, &handle, &sample);
4234 perf_output_end(&handle);
4236 task_event->event_id.header.size = size;
4239 static int perf_event_task_match(struct perf_event *event)
4241 if (event->state < PERF_EVENT_STATE_INACTIVE)
4244 if (!event_filter_match(event))
4247 if (event->attr.comm || event->attr.mmap ||
4248 event->attr.mmap_data || event->attr.task)
4254 static void perf_event_task_ctx(struct perf_event_context *ctx,
4255 struct perf_task_event *task_event)
4257 struct perf_event *event;
4259 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
4260 if (perf_event_task_match(event))
4261 perf_event_task_output(event, task_event);
4265 static void perf_event_task_event(struct perf_task_event *task_event)
4267 struct perf_cpu_context *cpuctx;
4268 struct perf_event_context *ctx;
4273 list_for_each_entry_rcu(pmu, &pmus, entry) {
4274 cpuctx = get_cpu_ptr(pmu->pmu_cpu_context);
4275 if (cpuctx->active_pmu != pmu)
4277 perf_event_task_ctx(&cpuctx->ctx, task_event);
4279 ctx = task_event->task_ctx;
4281 ctxn = pmu->task_ctx_nr;
4284 ctx = rcu_dereference(current->perf_event_ctxp[ctxn]);
4287 perf_event_task_ctx(ctx, task_event);
4289 put_cpu_ptr(pmu->pmu_cpu_context);
4294 static void perf_event_task(struct task_struct *task,
4295 struct perf_event_context *task_ctx,
4298 struct perf_task_event task_event;
4300 if (!atomic_read(&nr_comm_events) &&
4301 !atomic_read(&nr_mmap_events) &&
4302 !atomic_read(&nr_task_events))
4305 task_event = (struct perf_task_event){
4307 .task_ctx = task_ctx,
4310 .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT,
4312 .size = sizeof(task_event.event_id),
4318 .time = perf_clock(),
4322 perf_event_task_event(&task_event);
4325 void perf_event_fork(struct task_struct *task)
4327 perf_event_task(task, NULL, 1);
4334 struct perf_comm_event {
4335 struct task_struct *task;
4340 struct perf_event_header header;
4347 static void perf_event_comm_output(struct perf_event *event,
4348 struct perf_comm_event *comm_event)
4350 struct perf_output_handle handle;
4351 struct perf_sample_data sample;
4352 int size = comm_event->event_id.header.size;
4355 perf_event_header__init_id(&comm_event->event_id.header, &sample, event);
4356 ret = perf_output_begin(&handle, event,
4357 comm_event->event_id.header.size);
4362 comm_event->event_id.pid = perf_event_pid(event, comm_event->task);
4363 comm_event->event_id.tid = perf_event_tid(event, comm_event->task);
4365 perf_output_put(&handle, comm_event->event_id);
4366 __output_copy(&handle, comm_event->comm,
4367 comm_event->comm_size);
4369 perf_event__output_id_sample(event, &handle, &sample);
4371 perf_output_end(&handle);
4373 comm_event->event_id.header.size = size;
4376 static int perf_event_comm_match(struct perf_event *event)
4378 if (event->state < PERF_EVENT_STATE_INACTIVE)
4381 if (!event_filter_match(event))
4384 if (event->attr.comm)
4390 static void perf_event_comm_ctx(struct perf_event_context *ctx,
4391 struct perf_comm_event *comm_event)
4393 struct perf_event *event;
4395 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
4396 if (perf_event_comm_match(event))
4397 perf_event_comm_output(event, comm_event);
4401 static void perf_event_comm_event(struct perf_comm_event *comm_event)
4403 struct perf_cpu_context *cpuctx;
4404 struct perf_event_context *ctx;
4405 char comm[TASK_COMM_LEN];
4410 memset(comm, 0, sizeof(comm));
4411 strlcpy(comm, comm_event->task->comm, sizeof(comm));
4412 size = ALIGN(strlen(comm)+1, sizeof(u64));
4414 comm_event->comm = comm;
4415 comm_event->comm_size = size;
4417 comm_event->event_id.header.size = sizeof(comm_event->event_id) + size;
4419 list_for_each_entry_rcu(pmu, &pmus, entry) {
4420 cpuctx = get_cpu_ptr(pmu->pmu_cpu_context);
4421 if (cpuctx->active_pmu != pmu)
4423 perf_event_comm_ctx(&cpuctx->ctx, comm_event);
4425 ctxn = pmu->task_ctx_nr;
4429 ctx = rcu_dereference(current->perf_event_ctxp[ctxn]);
4431 perf_event_comm_ctx(ctx, comm_event);
4433 put_cpu_ptr(pmu->pmu_cpu_context);
4438 void perf_event_comm(struct task_struct *task)
4440 struct perf_comm_event comm_event;
4441 struct perf_event_context *ctx;
4444 for_each_task_context_nr(ctxn) {
4445 ctx = task->perf_event_ctxp[ctxn];
4449 perf_event_enable_on_exec(ctx);
4452 if (!atomic_read(&nr_comm_events))
4455 comm_event = (struct perf_comm_event){
4461 .type = PERF_RECORD_COMM,
4470 perf_event_comm_event(&comm_event);
4477 struct perf_mmap_event {
4478 struct vm_area_struct *vma;
4480 const char *file_name;
4484 struct perf_event_header header;
4494 static void perf_event_mmap_output(struct perf_event *event,
4495 struct perf_mmap_event *mmap_event)
4497 struct perf_output_handle handle;
4498 struct perf_sample_data sample;
4499 int size = mmap_event->event_id.header.size;
4502 perf_event_header__init_id(&mmap_event->event_id.header, &sample, event);
4503 ret = perf_output_begin(&handle, event,
4504 mmap_event->event_id.header.size);
4508 mmap_event->event_id.pid = perf_event_pid(event, current);
4509 mmap_event->event_id.tid = perf_event_tid(event, current);
4511 perf_output_put(&handle, mmap_event->event_id);
4512 __output_copy(&handle, mmap_event->file_name,
4513 mmap_event->file_size);
4515 perf_event__output_id_sample(event, &handle, &sample);
4517 perf_output_end(&handle);
4519 mmap_event->event_id.header.size = size;
4522 static int perf_event_mmap_match(struct perf_event *event,
4523 struct perf_mmap_event *mmap_event,
4526 if (event->state < PERF_EVENT_STATE_INACTIVE)
4529 if (!event_filter_match(event))
4532 if ((!executable && event->attr.mmap_data) ||
4533 (executable && event->attr.mmap))
4539 static void perf_event_mmap_ctx(struct perf_event_context *ctx,
4540 struct perf_mmap_event *mmap_event,
4543 struct perf_event *event;
4545 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
4546 if (perf_event_mmap_match(event, mmap_event, executable))
4547 perf_event_mmap_output(event, mmap_event);
4551 static void perf_event_mmap_event(struct perf_mmap_event *mmap_event)
4553 struct perf_cpu_context *cpuctx;
4554 struct perf_event_context *ctx;
4555 struct vm_area_struct *vma = mmap_event->vma;
4556 struct file *file = vma->vm_file;
4564 memset(tmp, 0, sizeof(tmp));
4568 * d_path works from the end of the rb backwards, so we
4569 * need to add enough zero bytes after the string to handle
4570 * the 64bit alignment we do later.
4572 buf = kzalloc(PATH_MAX + sizeof(u64), GFP_KERNEL);
4574 name = strncpy(tmp, "//enomem", sizeof(tmp));
4577 name = d_path(&file->f_path, buf, PATH_MAX);
4579 name = strncpy(tmp, "//toolong", sizeof(tmp));
4583 if (arch_vma_name(mmap_event->vma)) {
4584 name = strncpy(tmp, arch_vma_name(mmap_event->vma),
4590 name = strncpy(tmp, "[vdso]", sizeof(tmp));
4592 } else if (vma->vm_start <= vma->vm_mm->start_brk &&
4593 vma->vm_end >= vma->vm_mm->brk) {
4594 name = strncpy(tmp, "[heap]", sizeof(tmp));
4596 } else if (vma->vm_start <= vma->vm_mm->start_stack &&
4597 vma->vm_end >= vma->vm_mm->start_stack) {
4598 name = strncpy(tmp, "[stack]", sizeof(tmp));
4602 name = strncpy(tmp, "//anon", sizeof(tmp));
4607 size = ALIGN(strlen(name)+1, sizeof(u64));
4609 mmap_event->file_name = name;
4610 mmap_event->file_size = size;
4612 mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size;
4615 list_for_each_entry_rcu(pmu, &pmus, entry) {
4616 cpuctx = get_cpu_ptr(pmu->pmu_cpu_context);
4617 if (cpuctx->active_pmu != pmu)
4619 perf_event_mmap_ctx(&cpuctx->ctx, mmap_event,
4620 vma->vm_flags & VM_EXEC);
4622 ctxn = pmu->task_ctx_nr;
4626 ctx = rcu_dereference(current->perf_event_ctxp[ctxn]);
4628 perf_event_mmap_ctx(ctx, mmap_event,
4629 vma->vm_flags & VM_EXEC);
4632 put_cpu_ptr(pmu->pmu_cpu_context);
4639 void perf_event_mmap(struct vm_area_struct *vma)
4641 struct perf_mmap_event mmap_event;
4643 if (!atomic_read(&nr_mmap_events))
4646 mmap_event = (struct perf_mmap_event){
4652 .type = PERF_RECORD_MMAP,
4653 .misc = PERF_RECORD_MISC_USER,
4658 .start = vma->vm_start,
4659 .len = vma->vm_end - vma->vm_start,
4660 .pgoff = (u64)vma->vm_pgoff << PAGE_SHIFT,
4664 perf_event_mmap_event(&mmap_event);
4668 * IRQ throttle logging
4671 static void perf_log_throttle(struct perf_event *event, int enable)
4673 struct perf_output_handle handle;
4674 struct perf_sample_data sample;
4678 struct perf_event_header header;
4682 } throttle_event = {
4684 .type = PERF_RECORD_THROTTLE,
4686 .size = sizeof(throttle_event),
4688 .time = perf_clock(),
4689 .id = primary_event_id(event),
4690 .stream_id = event->id,
4694 throttle_event.header.type = PERF_RECORD_UNTHROTTLE;
4696 perf_event_header__init_id(&throttle_event.header, &sample, event);
4698 ret = perf_output_begin(&handle, event,
4699 throttle_event.header.size);
4703 perf_output_put(&handle, throttle_event);
4704 perf_event__output_id_sample(event, &handle, &sample);
4705 perf_output_end(&handle);
4709 * Generic event overflow handling, sampling.
4712 static int __perf_event_overflow(struct perf_event *event,
4713 int throttle, struct perf_sample_data *data,
4714 struct pt_regs *regs)
4716 int events = atomic_read(&event->event_limit);
4717 struct hw_perf_event *hwc = &event->hw;
4721 * Non-sampling counters might still use the PMI to fold short
4722 * hardware counters, ignore those.
4724 if (unlikely(!is_sampling_event(event)))
4727 if (unlikely(hwc->interrupts >= max_samples_per_tick)) {
4729 hwc->interrupts = MAX_INTERRUPTS;
4730 perf_log_throttle(event, 0);
4736 if (event->attr.freq) {
4737 u64 now = perf_clock();
4738 s64 delta = now - hwc->freq_time_stamp;
4740 hwc->freq_time_stamp = now;
4742 if (delta > 0 && delta < 2*TICK_NSEC)
4743 perf_adjust_period(event, delta, hwc->last_period);
4747 * XXX event_limit might not quite work as expected on inherited
4751 event->pending_kill = POLL_IN;
4752 if (events && atomic_dec_and_test(&event->event_limit)) {
4754 event->pending_kill = POLL_HUP;
4755 event->pending_disable = 1;
4756 irq_work_queue(&event->pending);
4759 if (event->overflow_handler)
4760 event->overflow_handler(event, data, regs);
4762 perf_event_output(event, data, regs);
4764 if (event->fasync && event->pending_kill) {
4765 event->pending_wakeup = 1;
4766 irq_work_queue(&event->pending);
4772 int perf_event_overflow(struct perf_event *event,
4773 struct perf_sample_data *data,
4774 struct pt_regs *regs)
4776 return __perf_event_overflow(event, 1, data, regs);
4780 * Generic software event infrastructure
4783 struct swevent_htable {
4784 struct swevent_hlist *swevent_hlist;
4785 struct mutex hlist_mutex;
4788 /* Recursion avoidance in each contexts */
4789 int recursion[PERF_NR_CONTEXTS];
4792 static DEFINE_PER_CPU(struct swevent_htable, swevent_htable);
4795 * We directly increment event->count and keep a second value in
4796 * event->hw.period_left to count intervals. This period event
4797 * is kept in the range [-sample_period, 0] so that we can use the
4801 static u64 perf_swevent_set_period(struct perf_event *event)
4803 struct hw_perf_event *hwc = &event->hw;
4804 u64 period = hwc->last_period;
4808 hwc->last_period = hwc->sample_period;
4811 old = val = local64_read(&hwc->period_left);
4815 nr = div64_u64(period + val, period);
4816 offset = nr * period;
4818 if (local64_cmpxchg(&hwc->period_left, old, val) != old)
4824 static void perf_swevent_overflow(struct perf_event *event, u64 overflow,
4825 struct perf_sample_data *data,
4826 struct pt_regs *regs)
4828 struct hw_perf_event *hwc = &event->hw;
4831 data->period = event->hw.last_period;
4833 overflow = perf_swevent_set_period(event);
4835 if (hwc->interrupts == MAX_INTERRUPTS)
4838 for (; overflow; overflow--) {
4839 if (__perf_event_overflow(event, throttle,
4842 * We inhibit the overflow from happening when
4843 * hwc->interrupts == MAX_INTERRUPTS.
4851 static void perf_swevent_event(struct perf_event *event, u64 nr,
4852 struct perf_sample_data *data,
4853 struct pt_regs *regs)
4855 struct hw_perf_event *hwc = &event->hw;
4857 local64_add(nr, &event->count);
4862 if (!is_sampling_event(event))
4865 if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq)
4866 return perf_swevent_overflow(event, 1, data, regs);
4868 if (local64_add_negative(nr, &hwc->period_left))
4871 perf_swevent_overflow(event, 0, data, regs);
4874 static int perf_exclude_event(struct perf_event *event,
4875 struct pt_regs *regs)
4877 if (event->hw.state & PERF_HES_STOPPED)
4881 if (event->attr.exclude_user && user_mode(regs))
4884 if (event->attr.exclude_kernel && !user_mode(regs))
4891 static int perf_swevent_match(struct perf_event *event,
4892 enum perf_type_id type,
4894 struct perf_sample_data *data,
4895 struct pt_regs *regs)
4897 if (event->attr.type != type)
4900 if (event->attr.config != event_id)
4903 if (perf_exclude_event(event, regs))
4909 static inline u64 swevent_hash(u64 type, u32 event_id)
4911 u64 val = event_id | (type << 32);
4913 return hash_64(val, SWEVENT_HLIST_BITS);
4916 static inline struct hlist_head *
4917 __find_swevent_head(struct swevent_hlist *hlist, u64 type, u32 event_id)
4919 u64 hash = swevent_hash(type, event_id);
4921 return &hlist->heads[hash];
4924 /* For the read side: events when they trigger */
4925 static inline struct hlist_head *
4926 find_swevent_head_rcu(struct swevent_htable *swhash, u64 type, u32 event_id)
4928 struct swevent_hlist *hlist;
4930 hlist = rcu_dereference(swhash->swevent_hlist);
4934 return __find_swevent_head(hlist, type, event_id);
4937 /* For the event head insertion and removal in the hlist */
4938 static inline struct hlist_head *
4939 find_swevent_head(struct swevent_htable *swhash, struct perf_event *event)
4941 struct swevent_hlist *hlist;
4942 u32 event_id = event->attr.config;
4943 u64 type = event->attr.type;
4946 * Event scheduling is always serialized against hlist allocation
4947 * and release. Which makes the protected version suitable here.
4948 * The context lock guarantees that.
4950 hlist = rcu_dereference_protected(swhash->swevent_hlist,
4951 lockdep_is_held(&event->ctx->lock));
4955 return __find_swevent_head(hlist, type, event_id);
4958 static void do_perf_sw_event(enum perf_type_id type, u32 event_id,
4960 struct perf_sample_data *data,
4961 struct pt_regs *regs)
4963 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
4964 struct perf_event *event;
4965 struct hlist_node *node;
4966 struct hlist_head *head;
4969 head = find_swevent_head_rcu(swhash, type, event_id);
4973 hlist_for_each_entry_rcu(event, node, head, hlist_entry) {
4974 if (perf_swevent_match(event, type, event_id, data, regs))
4975 perf_swevent_event(event, nr, data, regs);
4981 int perf_swevent_get_recursion_context(void)
4983 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
4985 return get_recursion_context(swhash->recursion);
4987 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context);
4989 inline void perf_swevent_put_recursion_context(int rctx)
4991 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
4993 put_recursion_context(swhash->recursion, rctx);
4996 void __perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)
4998 struct perf_sample_data data;
5001 preempt_disable_notrace();
5002 rctx = perf_swevent_get_recursion_context();
5006 perf_sample_data_init(&data, addr);
5008 do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, &data, regs);
5010 perf_swevent_put_recursion_context(rctx);
5011 preempt_enable_notrace();
5014 static void perf_swevent_read(struct perf_event *event)
5018 static int perf_swevent_add(struct perf_event *event, int flags)
5020 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
5021 struct hw_perf_event *hwc = &event->hw;
5022 struct hlist_head *head;
5024 if (is_sampling_event(event)) {
5025 hwc->last_period = hwc->sample_period;
5026 perf_swevent_set_period(event);
5029 hwc->state = !(flags & PERF_EF_START);
5031 head = find_swevent_head(swhash, event);
5032 if (WARN_ON_ONCE(!head))
5035 hlist_add_head_rcu(&event->hlist_entry, head);
5040 static void perf_swevent_del(struct perf_event *event, int flags)
5042 hlist_del_rcu(&event->hlist_entry);
5045 static void perf_swevent_start(struct perf_event *event, int flags)
5047 event->hw.state = 0;
5050 static void perf_swevent_stop(struct perf_event *event, int flags)
5052 event->hw.state = PERF_HES_STOPPED;
5055 /* Deref the hlist from the update side */
5056 static inline struct swevent_hlist *
5057 swevent_hlist_deref(struct swevent_htable *swhash)
5059 return rcu_dereference_protected(swhash->swevent_hlist,
5060 lockdep_is_held(&swhash->hlist_mutex));
5063 static void swevent_hlist_release(struct swevent_htable *swhash)
5065 struct swevent_hlist *hlist = swevent_hlist_deref(swhash);
5070 rcu_assign_pointer(swhash->swevent_hlist, NULL);
5071 kfree_rcu(hlist, rcu_head);
5074 static void swevent_hlist_put_cpu(struct perf_event *event, int cpu)
5076 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
5078 mutex_lock(&swhash->hlist_mutex);
5080 if (!--swhash->hlist_refcount)
5081 swevent_hlist_release(swhash);
5083 mutex_unlock(&swhash->hlist_mutex);
5086 static void swevent_hlist_put(struct perf_event *event)
5090 if (event->cpu != -1) {
5091 swevent_hlist_put_cpu(event, event->cpu);
5095 for_each_possible_cpu(cpu)
5096 swevent_hlist_put_cpu(event, cpu);
5099 static int swevent_hlist_get_cpu(struct perf_event *event, int cpu)
5101 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
5104 mutex_lock(&swhash->hlist_mutex);
5106 if (!swevent_hlist_deref(swhash) && cpu_online(cpu)) {
5107 struct swevent_hlist *hlist;
5109 hlist = kzalloc(sizeof(*hlist), GFP_KERNEL);
5114 rcu_assign_pointer(swhash->swevent_hlist, hlist);
5116 swhash->hlist_refcount++;
5118 mutex_unlock(&swhash->hlist_mutex);
5123 static int swevent_hlist_get(struct perf_event *event)
5126 int cpu, failed_cpu;
5128 if (event->cpu != -1)
5129 return swevent_hlist_get_cpu(event, event->cpu);
5132 for_each_possible_cpu(cpu) {
5133 err = swevent_hlist_get_cpu(event, cpu);
5143 for_each_possible_cpu(cpu) {
5144 if (cpu == failed_cpu)
5146 swevent_hlist_put_cpu(event, cpu);
5153 struct jump_label_key perf_swevent_enabled[PERF_COUNT_SW_MAX];
5155 static void sw_perf_event_destroy(struct perf_event *event)
5157 u64 event_id = event->attr.config;
5159 WARN_ON(event->parent);
5161 jump_label_dec(&perf_swevent_enabled[event_id]);
5162 swevent_hlist_put(event);
5165 static int perf_swevent_init(struct perf_event *event)
5167 u64 event_id = event->attr.config;
5169 if (event->attr.type != PERF_TYPE_SOFTWARE)
5173 case PERF_COUNT_SW_CPU_CLOCK:
5174 case PERF_COUNT_SW_TASK_CLOCK:
5181 if (event_id >= PERF_COUNT_SW_MAX)
5184 if (!event->parent) {
5187 err = swevent_hlist_get(event);
5191 jump_label_inc(&perf_swevent_enabled[event_id]);
5192 event->destroy = sw_perf_event_destroy;
5198 static struct pmu perf_swevent = {
5199 .task_ctx_nr = perf_sw_context,
5201 .event_init = perf_swevent_init,
5202 .add = perf_swevent_add,
5203 .del = perf_swevent_del,
5204 .start = perf_swevent_start,
5205 .stop = perf_swevent_stop,
5206 .read = perf_swevent_read,
5209 #ifdef CONFIG_EVENT_TRACING
5211 static int perf_tp_filter_match(struct perf_event *event,
5212 struct perf_sample_data *data)
5214 void *record = data->raw->data;
5216 if (likely(!event->filter) || filter_match_preds(event->filter, record))
5221 static int perf_tp_event_match(struct perf_event *event,
5222 struct perf_sample_data *data,
5223 struct pt_regs *regs)
5225 if (event->hw.state & PERF_HES_STOPPED)
5228 * All tracepoints are from kernel-space.
5230 if (event->attr.exclude_kernel)
5233 if (!perf_tp_filter_match(event, data))
5239 void perf_tp_event(u64 addr, u64 count, void *record, int entry_size,
5240 struct pt_regs *regs, struct hlist_head *head, int rctx)
5242 struct perf_sample_data data;
5243 struct perf_event *event;
5244 struct hlist_node *node;
5246 struct perf_raw_record raw = {
5251 perf_sample_data_init(&data, addr);
5254 hlist_for_each_entry_rcu(event, node, head, hlist_entry) {
5255 if (perf_tp_event_match(event, &data, regs))
5256 perf_swevent_event(event, count, &data, regs);
5259 perf_swevent_put_recursion_context(rctx);
5261 EXPORT_SYMBOL_GPL(perf_tp_event);
5263 static void tp_perf_event_destroy(struct perf_event *event)
5265 perf_trace_destroy(event);
5268 static int perf_tp_event_init(struct perf_event *event)
5272 if (event->attr.type != PERF_TYPE_TRACEPOINT)
5275 err = perf_trace_init(event);
5279 event->destroy = tp_perf_event_destroy;
5284 static struct pmu perf_tracepoint = {
5285 .task_ctx_nr = perf_sw_context,
5287 .event_init = perf_tp_event_init,
5288 .add = perf_trace_add,
5289 .del = perf_trace_del,
5290 .start = perf_swevent_start,
5291 .stop = perf_swevent_stop,
5292 .read = perf_swevent_read,
5295 static inline void perf_tp_register(void)
5297 perf_pmu_register(&perf_tracepoint, "tracepoint", PERF_TYPE_TRACEPOINT);
5300 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
5305 if (event->attr.type != PERF_TYPE_TRACEPOINT)
5308 filter_str = strndup_user(arg, PAGE_SIZE);
5309 if (IS_ERR(filter_str))
5310 return PTR_ERR(filter_str);
5312 ret = ftrace_profile_set_filter(event, event->attr.config, filter_str);
5318 static void perf_event_free_filter(struct perf_event *event)
5320 ftrace_profile_free_filter(event);
5325 static inline void perf_tp_register(void)
5329 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
5334 static void perf_event_free_filter(struct perf_event *event)
5338 #endif /* CONFIG_EVENT_TRACING */
5340 #ifdef CONFIG_HAVE_HW_BREAKPOINT
5341 void perf_bp_event(struct perf_event *bp, void *data)
5343 struct perf_sample_data sample;
5344 struct pt_regs *regs = data;
5346 perf_sample_data_init(&sample, bp->attr.bp_addr);
5348 if (!bp->hw.state && !perf_exclude_event(bp, regs))
5349 perf_swevent_event(bp, 1, &sample, regs);
5354 * hrtimer based swevent callback
5357 static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer)
5359 enum hrtimer_restart ret = HRTIMER_RESTART;
5360 struct perf_sample_data data;
5361 struct pt_regs *regs;
5362 struct perf_event *event;
5365 event = container_of(hrtimer, struct perf_event, hw.hrtimer);
5367 if (event->state != PERF_EVENT_STATE_ACTIVE)
5368 return HRTIMER_NORESTART;
5370 event->pmu->read(event);
5372 perf_sample_data_init(&data, 0);
5373 data.period = event->hw.last_period;
5374 regs = get_irq_regs();
5376 if (regs && !perf_exclude_event(event, regs)) {
5377 if (!(event->attr.exclude_idle && current->pid == 0))
5378 if (perf_event_overflow(event, &data, regs))
5379 ret = HRTIMER_NORESTART;
5382 period = max_t(u64, 10000, event->hw.sample_period);
5383 hrtimer_forward_now(hrtimer, ns_to_ktime(period));
5388 static void perf_swevent_start_hrtimer(struct perf_event *event)
5390 struct hw_perf_event *hwc = &event->hw;
5393 if (!is_sampling_event(event))
5396 period = local64_read(&hwc->period_left);
5401 local64_set(&hwc->period_left, 0);
5403 period = max_t(u64, 10000, hwc->sample_period);
5405 __hrtimer_start_range_ns(&hwc->hrtimer,
5406 ns_to_ktime(period), 0,
5407 HRTIMER_MODE_REL_PINNED, 0);
5410 static void perf_swevent_cancel_hrtimer(struct perf_event *event)
5412 struct hw_perf_event *hwc = &event->hw;
5414 if (is_sampling_event(event)) {
5415 ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer);
5416 local64_set(&hwc->period_left, ktime_to_ns(remaining));
5418 hrtimer_cancel(&hwc->hrtimer);
5422 static void perf_swevent_init_hrtimer(struct perf_event *event)
5424 struct hw_perf_event *hwc = &event->hw;
5426 if (!is_sampling_event(event))
5429 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
5430 hwc->hrtimer.function = perf_swevent_hrtimer;
5433 * Since hrtimers have a fixed rate, we can do a static freq->period
5434 * mapping and avoid the whole period adjust feedback stuff.
5436 if (event->attr.freq) {
5437 long freq = event->attr.sample_freq;
5439 event->attr.sample_period = NSEC_PER_SEC / freq;
5440 hwc->sample_period = event->attr.sample_period;
5441 local64_set(&hwc->period_left, hwc->sample_period);
5442 event->attr.freq = 0;
5447 * Software event: cpu wall time clock
5450 static void cpu_clock_event_update(struct perf_event *event)
5455 now = local_clock();
5456 prev = local64_xchg(&event->hw.prev_count, now);
5457 local64_add(now - prev, &event->count);
5460 static void cpu_clock_event_start(struct perf_event *event, int flags)
5462 local64_set(&event->hw.prev_count, local_clock());
5463 perf_swevent_start_hrtimer(event);
5466 static void cpu_clock_event_stop(struct perf_event *event, int flags)
5468 perf_swevent_cancel_hrtimer(event);
5469 cpu_clock_event_update(event);
5472 static int cpu_clock_event_add(struct perf_event *event, int flags)
5474 if (flags & PERF_EF_START)
5475 cpu_clock_event_start(event, flags);
5480 static void cpu_clock_event_del(struct perf_event *event, int flags)
5482 cpu_clock_event_stop(event, flags);
5485 static void cpu_clock_event_read(struct perf_event *event)
5487 cpu_clock_event_update(event);
5490 static int cpu_clock_event_init(struct perf_event *event)
5492 if (event->attr.type != PERF_TYPE_SOFTWARE)
5495 if (event->attr.config != PERF_COUNT_SW_CPU_CLOCK)
5498 perf_swevent_init_hrtimer(event);
5503 static struct pmu perf_cpu_clock = {
5504 .task_ctx_nr = perf_sw_context,
5506 .event_init = cpu_clock_event_init,
5507 .add = cpu_clock_event_add,
5508 .del = cpu_clock_event_del,
5509 .start = cpu_clock_event_start,
5510 .stop = cpu_clock_event_stop,
5511 .read = cpu_clock_event_read,
5515 * Software event: task time clock
5518 static void task_clock_event_update(struct perf_event *event, u64 now)
5523 prev = local64_xchg(&event->hw.prev_count, now);
5525 local64_add(delta, &event->count);
5528 static void task_clock_event_start(struct perf_event *event, int flags)
5530 local64_set(&event->hw.prev_count, event->ctx->time);
5531 perf_swevent_start_hrtimer(event);
5534 static void task_clock_event_stop(struct perf_event *event, int flags)
5536 perf_swevent_cancel_hrtimer(event);
5537 task_clock_event_update(event, event->ctx->time);
5540 static int task_clock_event_add(struct perf_event *event, int flags)
5542 if (flags & PERF_EF_START)
5543 task_clock_event_start(event, flags);
5548 static void task_clock_event_del(struct perf_event *event, int flags)
5550 task_clock_event_stop(event, PERF_EF_UPDATE);
5553 static void task_clock_event_read(struct perf_event *event)
5555 u64 now = perf_clock();
5556 u64 delta = now - event->ctx->timestamp;
5557 u64 time = event->ctx->time + delta;
5559 task_clock_event_update(event, time);
5562 static int task_clock_event_init(struct perf_event *event)
5564 if (event->attr.type != PERF_TYPE_SOFTWARE)
5567 if (event->attr.config != PERF_COUNT_SW_TASK_CLOCK)
5570 perf_swevent_init_hrtimer(event);
5575 static struct pmu perf_task_clock = {
5576 .task_ctx_nr = perf_sw_context,
5578 .event_init = task_clock_event_init,
5579 .add = task_clock_event_add,
5580 .del = task_clock_event_del,
5581 .start = task_clock_event_start,
5582 .stop = task_clock_event_stop,
5583 .read = task_clock_event_read,
5586 static void perf_pmu_nop_void(struct pmu *pmu)
5590 static int perf_pmu_nop_int(struct pmu *pmu)
5595 static void perf_pmu_start_txn(struct pmu *pmu)
5597 perf_pmu_disable(pmu);
5600 static int perf_pmu_commit_txn(struct pmu *pmu)
5602 perf_pmu_enable(pmu);
5606 static void perf_pmu_cancel_txn(struct pmu *pmu)
5608 perf_pmu_enable(pmu);
5612 * Ensures all contexts with the same task_ctx_nr have the same
5613 * pmu_cpu_context too.
5615 static void *find_pmu_context(int ctxn)
5622 list_for_each_entry(pmu, &pmus, entry) {
5623 if (pmu->task_ctx_nr == ctxn)
5624 return pmu->pmu_cpu_context;
5630 static void update_pmu_context(struct pmu *pmu, struct pmu *old_pmu)
5634 for_each_possible_cpu(cpu) {
5635 struct perf_cpu_context *cpuctx;
5637 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
5639 if (cpuctx->active_pmu == old_pmu)
5640 cpuctx->active_pmu = pmu;
5644 static void free_pmu_context(struct pmu *pmu)
5648 mutex_lock(&pmus_lock);
5650 * Like a real lame refcount.
5652 list_for_each_entry(i, &pmus, entry) {
5653 if (i->pmu_cpu_context == pmu->pmu_cpu_context) {
5654 update_pmu_context(i, pmu);
5659 free_percpu(pmu->pmu_cpu_context);
5661 mutex_unlock(&pmus_lock);
5663 static struct idr pmu_idr;
5666 type_show(struct device *dev, struct device_attribute *attr, char *page)
5668 struct pmu *pmu = dev_get_drvdata(dev);
5670 return snprintf(page, PAGE_SIZE-1, "%d\n", pmu->type);
5673 static struct device_attribute pmu_dev_attrs[] = {
5678 static int pmu_bus_running;
5679 static struct bus_type pmu_bus = {
5680 .name = "event_source",
5681 .dev_attrs = pmu_dev_attrs,
5684 static void pmu_dev_release(struct device *dev)
5689 static int pmu_dev_alloc(struct pmu *pmu)
5693 pmu->dev = kzalloc(sizeof(struct device), GFP_KERNEL);
5697 device_initialize(pmu->dev);
5698 ret = dev_set_name(pmu->dev, "%s", pmu->name);
5702 dev_set_drvdata(pmu->dev, pmu);
5703 pmu->dev->bus = &pmu_bus;
5704 pmu->dev->release = pmu_dev_release;
5705 ret = device_add(pmu->dev);
5713 put_device(pmu->dev);
5717 static struct lock_class_key cpuctx_mutex;
5718 static struct lock_class_key cpuctx_lock;
5720 int perf_pmu_register(struct pmu *pmu, char *name, int type)
5724 mutex_lock(&pmus_lock);
5726 pmu->pmu_disable_count = alloc_percpu(int);
5727 if (!pmu->pmu_disable_count)
5736 int err = idr_pre_get(&pmu_idr, GFP_KERNEL);
5740 err = idr_get_new_above(&pmu_idr, pmu, PERF_TYPE_MAX, &type);
5748 if (pmu_bus_running) {
5749 ret = pmu_dev_alloc(pmu);
5755 pmu->pmu_cpu_context = find_pmu_context(pmu->task_ctx_nr);
5756 if (pmu->pmu_cpu_context)
5757 goto got_cpu_context;
5760 pmu->pmu_cpu_context = alloc_percpu(struct perf_cpu_context);
5761 if (!pmu->pmu_cpu_context)
5764 for_each_possible_cpu(cpu) {
5765 struct perf_cpu_context *cpuctx;
5767 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
5768 __perf_event_init_context(&cpuctx->ctx);
5769 lockdep_set_class(&cpuctx->ctx.mutex, &cpuctx_mutex);
5770 lockdep_set_class(&cpuctx->ctx.lock, &cpuctx_lock);
5771 cpuctx->ctx.type = cpu_context;
5772 cpuctx->ctx.pmu = pmu;
5773 cpuctx->jiffies_interval = 1;
5774 INIT_LIST_HEAD(&cpuctx->rotation_list);
5775 cpuctx->active_pmu = pmu;
5779 if (!pmu->start_txn) {
5780 if (pmu->pmu_enable) {
5782 * If we have pmu_enable/pmu_disable calls, install
5783 * transaction stubs that use that to try and batch
5784 * hardware accesses.
5786 pmu->start_txn = perf_pmu_start_txn;
5787 pmu->commit_txn = perf_pmu_commit_txn;
5788 pmu->cancel_txn = perf_pmu_cancel_txn;
5790 pmu->start_txn = perf_pmu_nop_void;
5791 pmu->commit_txn = perf_pmu_nop_int;
5792 pmu->cancel_txn = perf_pmu_nop_void;
5796 if (!pmu->pmu_enable) {
5797 pmu->pmu_enable = perf_pmu_nop_void;
5798 pmu->pmu_disable = perf_pmu_nop_void;
5801 list_add_rcu(&pmu->entry, &pmus);
5804 mutex_unlock(&pmus_lock);
5809 device_del(pmu->dev);
5810 put_device(pmu->dev);
5813 if (pmu->type >= PERF_TYPE_MAX)
5814 idr_remove(&pmu_idr, pmu->type);
5817 free_percpu(pmu->pmu_disable_count);
5821 void perf_pmu_unregister(struct pmu *pmu)
5823 mutex_lock(&pmus_lock);
5824 list_del_rcu(&pmu->entry);
5825 mutex_unlock(&pmus_lock);
5828 * We dereference the pmu list under both SRCU and regular RCU, so
5829 * synchronize against both of those.
5831 synchronize_srcu(&pmus_srcu);
5834 free_percpu(pmu->pmu_disable_count);
5835 if (pmu->type >= PERF_TYPE_MAX)
5836 idr_remove(&pmu_idr, pmu->type);
5837 device_del(pmu->dev);
5838 put_device(pmu->dev);
5839 free_pmu_context(pmu);
5842 struct pmu *perf_init_event(struct perf_event *event)
5844 struct pmu *pmu = NULL;
5848 idx = srcu_read_lock(&pmus_srcu);
5851 pmu = idr_find(&pmu_idr, event->attr.type);
5855 ret = pmu->event_init(event);
5861 list_for_each_entry_rcu(pmu, &pmus, entry) {
5863 ret = pmu->event_init(event);
5867 if (ret != -ENOENT) {
5872 pmu = ERR_PTR(-ENOENT);
5874 srcu_read_unlock(&pmus_srcu, idx);
5880 * Allocate and initialize a event structure
5882 static struct perf_event *
5883 perf_event_alloc(struct perf_event_attr *attr, int cpu,
5884 struct task_struct *task,
5885 struct perf_event *group_leader,
5886 struct perf_event *parent_event,
5887 perf_overflow_handler_t overflow_handler,
5891 struct perf_event *event;
5892 struct hw_perf_event *hwc;
5895 if ((unsigned)cpu >= nr_cpu_ids) {
5896 if (!task || cpu != -1)
5897 return ERR_PTR(-EINVAL);
5900 event = kzalloc(sizeof(*event), GFP_KERNEL);
5902 return ERR_PTR(-ENOMEM);
5905 * Single events are their own group leaders, with an
5906 * empty sibling list:
5909 group_leader = event;
5911 mutex_init(&event->child_mutex);
5912 INIT_LIST_HEAD(&event->child_list);
5914 INIT_LIST_HEAD(&event->group_entry);
5915 INIT_LIST_HEAD(&event->event_entry);
5916 INIT_LIST_HEAD(&event->sibling_list);
5917 INIT_LIST_HEAD(&event->rb_entry);
5919 init_waitqueue_head(&event->waitq);
5920 init_irq_work(&event->pending, perf_pending_event);
5922 mutex_init(&event->mmap_mutex);
5924 atomic_long_set(&event->refcount, 1);
5926 event->attr = *attr;
5927 event->group_leader = group_leader;
5931 event->parent = parent_event;
5933 event->ns = get_pid_ns(current->nsproxy->pid_ns);
5934 event->id = atomic64_inc_return(&perf_event_id);
5936 event->state = PERF_EVENT_STATE_INACTIVE;
5939 event->attach_state = PERF_ATTACH_TASK;
5940 #ifdef CONFIG_HAVE_HW_BREAKPOINT
5942 * hw_breakpoint is a bit difficult here..
5944 if (attr->type == PERF_TYPE_BREAKPOINT)
5945 event->hw.bp_target = task;
5949 if (!overflow_handler && parent_event) {
5950 overflow_handler = parent_event->overflow_handler;
5951 context = parent_event->overflow_handler_context;
5954 event->overflow_handler = overflow_handler;
5955 event->overflow_handler_context = context;
5958 event->state = PERF_EVENT_STATE_OFF;
5963 hwc->sample_period = attr->sample_period;
5964 if (attr->freq && attr->sample_freq)
5965 hwc->sample_period = 1;
5966 hwc->last_period = hwc->sample_period;
5968 local64_set(&hwc->period_left, hwc->sample_period);
5971 * we currently do not support PERF_FORMAT_GROUP on inherited events
5973 if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP))
5976 pmu = perf_init_event(event);
5982 else if (IS_ERR(pmu))
5987 put_pid_ns(event->ns);
5989 return ERR_PTR(err);
5992 if (!event->parent) {
5993 if (event->attach_state & PERF_ATTACH_TASK)
5994 jump_label_inc(&perf_sched_events);
5995 if (event->attr.mmap || event->attr.mmap_data)
5996 atomic_inc(&nr_mmap_events);
5997 if (event->attr.comm)
5998 atomic_inc(&nr_comm_events);
5999 if (event->attr.task)
6000 atomic_inc(&nr_task_events);
6001 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) {
6002 err = get_callchain_buffers();
6005 return ERR_PTR(err);
6013 static int perf_copy_attr(struct perf_event_attr __user *uattr,
6014 struct perf_event_attr *attr)
6019 if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0))
6023 * zero the full structure, so that a short copy will be nice.
6025 memset(attr, 0, sizeof(*attr));
6027 ret = get_user(size, &uattr->size);
6031 if (size > PAGE_SIZE) /* silly large */
6034 if (!size) /* abi compat */
6035 size = PERF_ATTR_SIZE_VER0;
6037 if (size < PERF_ATTR_SIZE_VER0)
6041 * If we're handed a bigger struct than we know of,
6042 * ensure all the unknown bits are 0 - i.e. new
6043 * user-space does not rely on any kernel feature
6044 * extensions we dont know about yet.
6046 if (size > sizeof(*attr)) {
6047 unsigned char __user *addr;
6048 unsigned char __user *end;
6051 addr = (void __user *)uattr + sizeof(*attr);
6052 end = (void __user *)uattr + size;
6054 for (; addr < end; addr++) {
6055 ret = get_user(val, addr);
6061 size = sizeof(*attr);
6064 ret = copy_from_user(attr, uattr, size);
6068 if (attr->__reserved_1)
6071 if (attr->sample_type & ~(PERF_SAMPLE_MAX-1))
6074 if (attr->read_format & ~(PERF_FORMAT_MAX-1))
6081 put_user(sizeof(*attr), &uattr->size);
6087 perf_event_set_output(struct perf_event *event, struct perf_event *output_event)
6089 struct ring_buffer *rb = NULL, *old_rb = NULL;
6095 /* don't allow circular references */
6096 if (event == output_event)
6100 * Don't allow cross-cpu buffers
6102 if (output_event->cpu != event->cpu)
6106 * If its not a per-cpu rb, it must be the same task.
6108 if (output_event->cpu == -1 && output_event->ctx != event->ctx)
6112 mutex_lock(&event->mmap_mutex);
6113 /* Can't redirect output if we've got an active mmap() */
6114 if (atomic_read(&event->mmap_count))
6118 /* get the rb we want to redirect to */
6119 rb = ring_buffer_get(output_event);
6125 rcu_assign_pointer(event->rb, rb);
6127 ring_buffer_detach(event, old_rb);
6130 mutex_unlock(&event->mmap_mutex);
6133 ring_buffer_put(old_rb);
6139 * sys_perf_event_open - open a performance event, associate it to a task/cpu
6141 * @attr_uptr: event_id type attributes for monitoring/sampling
6144 * @group_fd: group leader event fd
6146 SYSCALL_DEFINE5(perf_event_open,
6147 struct perf_event_attr __user *, attr_uptr,
6148 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
6150 struct perf_event *group_leader = NULL, *output_event = NULL;
6151 struct perf_event *event, *sibling;
6152 struct perf_event_attr attr;
6153 struct perf_event_context *ctx;
6154 struct file *event_file = NULL;
6155 struct file *group_file = NULL;
6156 struct task_struct *task = NULL;
6160 int fput_needed = 0;
6163 /* for future expandability... */
6164 if (flags & ~PERF_FLAG_ALL)
6167 err = perf_copy_attr(attr_uptr, &attr);
6171 if (!attr.exclude_kernel) {
6172 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
6177 if (attr.sample_freq > sysctl_perf_event_sample_rate)
6182 * In cgroup mode, the pid argument is used to pass the fd
6183 * opened to the cgroup directory in cgroupfs. The cpu argument
6184 * designates the cpu on which to monitor threads from that
6187 if ((flags & PERF_FLAG_PID_CGROUP) && (pid == -1 || cpu == -1))
6190 event_fd = get_unused_fd_flags(O_RDWR);
6194 if (group_fd != -1) {
6195 group_file = perf_fget_light(group_fd, &fput_needed);
6196 if (IS_ERR(group_file)) {
6197 err = PTR_ERR(group_file);
6200 group_leader = group_file->private_data;
6201 if (flags & PERF_FLAG_FD_OUTPUT)
6202 output_event = group_leader;
6203 if (flags & PERF_FLAG_FD_NO_GROUP)
6204 group_leader = NULL;
6207 if (pid != -1 && !(flags & PERF_FLAG_PID_CGROUP)) {
6208 task = find_lively_task_by_vpid(pid);
6210 err = PTR_ERR(task);
6215 event = perf_event_alloc(&attr, cpu, task, group_leader, NULL,
6217 if (IS_ERR(event)) {
6218 err = PTR_ERR(event);
6222 if (flags & PERF_FLAG_PID_CGROUP) {
6223 err = perf_cgroup_connect(pid, event, &attr, group_leader);
6228 * - that has cgroup constraint on event->cpu
6229 * - that may need work on context switch
6231 atomic_inc(&per_cpu(perf_cgroup_events, event->cpu));
6232 jump_label_inc(&perf_sched_events);
6236 * Special case software events and allow them to be part of
6237 * any hardware group.
6242 (is_software_event(event) != is_software_event(group_leader))) {
6243 if (is_software_event(event)) {
6245 * If event and group_leader are not both a software
6246 * event, and event is, then group leader is not.
6248 * Allow the addition of software events to !software
6249 * groups, this is safe because software events never
6252 pmu = group_leader->pmu;
6253 } else if (is_software_event(group_leader) &&
6254 (group_leader->group_flags & PERF_GROUP_SOFTWARE)) {
6256 * In case the group is a pure software group, and we
6257 * try to add a hardware event, move the whole group to
6258 * the hardware context.
6265 * Get the target context (task or percpu):
6267 ctx = find_get_context(pmu, task, cpu);
6274 put_task_struct(task);
6279 * Look up the group leader (we will attach this event to it):
6285 * Do not allow a recursive hierarchy (this new sibling
6286 * becoming part of another group-sibling):
6288 if (group_leader->group_leader != group_leader)
6291 * Do not allow to attach to a group in a different
6292 * task or CPU context:
6295 if (group_leader->ctx->type != ctx->type)
6298 if (group_leader->ctx != ctx)
6303 * Only a group leader can be exclusive or pinned
6305 if (attr.exclusive || attr.pinned)
6310 err = perf_event_set_output(event, output_event);
6315 event_file = anon_inode_getfile("[perf_event]", &perf_fops, event, O_RDWR);
6316 if (IS_ERR(event_file)) {
6317 err = PTR_ERR(event_file);
6322 struct perf_event_context *gctx = group_leader->ctx;
6324 mutex_lock(&gctx->mutex);
6325 perf_remove_from_context(group_leader);
6326 list_for_each_entry(sibling, &group_leader->sibling_list,
6328 perf_remove_from_context(sibling);
6331 mutex_unlock(&gctx->mutex);
6335 WARN_ON_ONCE(ctx->parent_ctx);
6336 mutex_lock(&ctx->mutex);
6339 perf_install_in_context(ctx, group_leader, cpu);
6341 list_for_each_entry(sibling, &group_leader->sibling_list,
6343 perf_install_in_context(ctx, sibling, cpu);
6348 perf_install_in_context(ctx, event, cpu);
6350 perf_unpin_context(ctx);
6351 mutex_unlock(&ctx->mutex);
6353 event->owner = current;
6355 mutex_lock(¤t->perf_event_mutex);
6356 list_add_tail(&event->owner_entry, ¤t->perf_event_list);
6357 mutex_unlock(¤t->perf_event_mutex);
6360 * Precalculate sample_data sizes
6362 perf_event__header_size(event);
6363 perf_event__id_header_size(event);
6366 * Drop the reference on the group_event after placing the
6367 * new event on the sibling_list. This ensures destruction
6368 * of the group leader will find the pointer to itself in
6369 * perf_group_detach().
6371 fput_light(group_file, fput_needed);
6372 fd_install(event_fd, event_file);
6376 perf_unpin_context(ctx);
6382 put_task_struct(task);
6384 fput_light(group_file, fput_needed);
6386 put_unused_fd(event_fd);
6391 * perf_event_create_kernel_counter
6393 * @attr: attributes of the counter to create
6394 * @cpu: cpu in which the counter is bound
6395 * @task: task to profile (NULL for percpu)
6398 perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu,
6399 struct task_struct *task,
6400 perf_overflow_handler_t overflow_handler,
6403 struct perf_event_context *ctx;
6404 struct perf_event *event;
6408 * Get the target context (task or percpu):
6411 event = perf_event_alloc(attr, cpu, task, NULL, NULL,
6412 overflow_handler, context);
6413 if (IS_ERR(event)) {
6414 err = PTR_ERR(event);
6418 ctx = find_get_context(event->pmu, task, cpu);
6424 WARN_ON_ONCE(ctx->parent_ctx);
6425 mutex_lock(&ctx->mutex);
6426 perf_install_in_context(ctx, event, cpu);
6428 perf_unpin_context(ctx);
6429 mutex_unlock(&ctx->mutex);
6436 return ERR_PTR(err);
6438 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter);
6440 static void sync_child_event(struct perf_event *child_event,
6441 struct task_struct *child)
6443 struct perf_event *parent_event = child_event->parent;
6446 if (child_event->attr.inherit_stat)
6447 perf_event_read_event(child_event, child);
6449 child_val = perf_event_count(child_event);
6452 * Add back the child's count to the parent's count:
6454 atomic64_add(child_val, &parent_event->child_count);
6455 atomic64_add(child_event->total_time_enabled,
6456 &parent_event->child_total_time_enabled);
6457 atomic64_add(child_event->total_time_running,
6458 &parent_event->child_total_time_running);
6461 * Remove this event from the parent's list
6463 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
6464 mutex_lock(&parent_event->child_mutex);
6465 list_del_init(&child_event->child_list);
6466 mutex_unlock(&parent_event->child_mutex);
6469 * Release the parent event, if this was the last
6472 put_event(parent_event);
6476 __perf_event_exit_task(struct perf_event *child_event,
6477 struct perf_event_context *child_ctx,
6478 struct task_struct *child)
6480 if (child_event->parent) {
6481 raw_spin_lock_irq(&child_ctx->lock);
6482 perf_group_detach(child_event);
6483 raw_spin_unlock_irq(&child_ctx->lock);
6486 perf_remove_from_context(child_event);
6489 * It can happen that the parent exits first, and has events
6490 * that are still around due to the child reference. These
6491 * events need to be zapped.
6493 if (child_event->parent) {
6494 sync_child_event(child_event, child);
6495 free_event(child_event);
6499 static void perf_event_exit_task_context(struct task_struct *child, int ctxn)
6501 struct perf_event *child_event, *tmp;
6502 struct perf_event_context *child_ctx;
6503 unsigned long flags;
6505 if (likely(!child->perf_event_ctxp[ctxn])) {
6506 perf_event_task(child, NULL, 0);
6510 local_irq_save(flags);
6512 * We can't reschedule here because interrupts are disabled,
6513 * and either child is current or it is a task that can't be
6514 * scheduled, so we are now safe from rescheduling changing
6517 child_ctx = rcu_dereference_raw(child->perf_event_ctxp[ctxn]);
6520 * Take the context lock here so that if find_get_context is
6521 * reading child->perf_event_ctxp, we wait until it has
6522 * incremented the context's refcount before we do put_ctx below.
6524 raw_spin_lock(&child_ctx->lock);
6525 task_ctx_sched_out(child_ctx);
6526 child->perf_event_ctxp[ctxn] = NULL;
6528 * If this context is a clone; unclone it so it can't get
6529 * swapped to another process while we're removing all
6530 * the events from it.
6532 unclone_ctx(child_ctx);
6533 update_context_time(child_ctx);
6534 raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
6537 * Report the task dead after unscheduling the events so that we
6538 * won't get any samples after PERF_RECORD_EXIT. We can however still
6539 * get a few PERF_RECORD_READ events.
6541 perf_event_task(child, child_ctx, 0);
6544 * We can recurse on the same lock type through:
6546 * __perf_event_exit_task()
6547 * sync_child_event()
6549 * mutex_lock(&ctx->mutex)
6551 * But since its the parent context it won't be the same instance.
6553 mutex_lock(&child_ctx->mutex);
6556 list_for_each_entry_safe(child_event, tmp, &child_ctx->pinned_groups,
6558 __perf_event_exit_task(child_event, child_ctx, child);
6560 list_for_each_entry_safe(child_event, tmp, &child_ctx->flexible_groups,
6562 __perf_event_exit_task(child_event, child_ctx, child);
6565 * If the last event was a group event, it will have appended all
6566 * its siblings to the list, but we obtained 'tmp' before that which
6567 * will still point to the list head terminating the iteration.
6569 if (!list_empty(&child_ctx->pinned_groups) ||
6570 !list_empty(&child_ctx->flexible_groups))
6573 mutex_unlock(&child_ctx->mutex);
6579 * When a child task exits, feed back event values to parent events.
6581 void perf_event_exit_task(struct task_struct *child)
6583 struct perf_event *event, *tmp;
6586 mutex_lock(&child->perf_event_mutex);
6587 list_for_each_entry_safe(event, tmp, &child->perf_event_list,
6589 list_del_init(&event->owner_entry);
6592 * Ensure the list deletion is visible before we clear
6593 * the owner, closes a race against perf_release() where
6594 * we need to serialize on the owner->perf_event_mutex.
6597 event->owner = NULL;
6599 mutex_unlock(&child->perf_event_mutex);
6601 for_each_task_context_nr(ctxn)
6602 perf_event_exit_task_context(child, ctxn);
6605 static void perf_free_event(struct perf_event *event,
6606 struct perf_event_context *ctx)
6608 struct perf_event *parent = event->parent;
6610 if (WARN_ON_ONCE(!parent))
6613 mutex_lock(&parent->child_mutex);
6614 list_del_init(&event->child_list);
6615 mutex_unlock(&parent->child_mutex);
6619 perf_group_detach(event);
6620 list_del_event(event, ctx);
6625 * free an unexposed, unused context as created by inheritance by
6626 * perf_event_init_task below, used by fork() in case of fail.
6628 void perf_event_free_task(struct task_struct *task)
6630 struct perf_event_context *ctx;
6631 struct perf_event *event, *tmp;
6634 for_each_task_context_nr(ctxn) {
6635 ctx = task->perf_event_ctxp[ctxn];
6639 mutex_lock(&ctx->mutex);
6641 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups,
6643 perf_free_event(event, ctx);
6645 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups,
6647 perf_free_event(event, ctx);
6649 if (!list_empty(&ctx->pinned_groups) ||
6650 !list_empty(&ctx->flexible_groups))
6653 mutex_unlock(&ctx->mutex);
6659 void perf_event_delayed_put(struct task_struct *task)
6663 for_each_task_context_nr(ctxn)
6664 WARN_ON_ONCE(task->perf_event_ctxp[ctxn]);
6668 * inherit a event from parent task to child task:
6670 static struct perf_event *
6671 inherit_event(struct perf_event *parent_event,
6672 struct task_struct *parent,
6673 struct perf_event_context *parent_ctx,
6674 struct task_struct *child,
6675 struct perf_event *group_leader,
6676 struct perf_event_context *child_ctx)
6678 struct perf_event *child_event;
6679 unsigned long flags;
6682 * Instead of creating recursive hierarchies of events,
6683 * we link inherited events back to the original parent,
6684 * which has a filp for sure, which we use as the reference
6687 if (parent_event->parent)
6688 parent_event = parent_event->parent;
6690 child_event = perf_event_alloc(&parent_event->attr,
6693 group_leader, parent_event,
6695 if (IS_ERR(child_event))
6698 if (!atomic_long_inc_not_zero(&parent_event->refcount)) {
6699 free_event(child_event);
6706 * Make the child state follow the state of the parent event,
6707 * not its attr.disabled bit. We hold the parent's mutex,
6708 * so we won't race with perf_event_{en, dis}able_family.
6710 if (parent_event->state >= PERF_EVENT_STATE_INACTIVE)
6711 child_event->state = PERF_EVENT_STATE_INACTIVE;
6713 child_event->state = PERF_EVENT_STATE_OFF;
6715 if (parent_event->attr.freq) {
6716 u64 sample_period = parent_event->hw.sample_period;
6717 struct hw_perf_event *hwc = &child_event->hw;
6719 hwc->sample_period = sample_period;
6720 hwc->last_period = sample_period;
6722 local64_set(&hwc->period_left, sample_period);
6725 child_event->ctx = child_ctx;
6726 child_event->overflow_handler = parent_event->overflow_handler;
6727 child_event->overflow_handler_context
6728 = parent_event->overflow_handler_context;
6731 * Precalculate sample_data sizes
6733 perf_event__header_size(child_event);
6734 perf_event__id_header_size(child_event);
6737 * Link it up in the child's context:
6739 raw_spin_lock_irqsave(&child_ctx->lock, flags);
6740 add_event_to_ctx(child_event, child_ctx);
6741 raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
6744 * Link this into the parent event's child list
6746 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
6747 mutex_lock(&parent_event->child_mutex);
6748 list_add_tail(&child_event->child_list, &parent_event->child_list);
6749 mutex_unlock(&parent_event->child_mutex);
6754 static int inherit_group(struct perf_event *parent_event,
6755 struct task_struct *parent,
6756 struct perf_event_context *parent_ctx,
6757 struct task_struct *child,
6758 struct perf_event_context *child_ctx)
6760 struct perf_event *leader;
6761 struct perf_event *sub;
6762 struct perf_event *child_ctr;
6764 leader = inherit_event(parent_event, parent, parent_ctx,
6765 child, NULL, child_ctx);
6767 return PTR_ERR(leader);
6768 list_for_each_entry(sub, &parent_event->sibling_list, group_entry) {
6769 child_ctr = inherit_event(sub, parent, parent_ctx,
6770 child, leader, child_ctx);
6771 if (IS_ERR(child_ctr))
6772 return PTR_ERR(child_ctr);
6778 inherit_task_group(struct perf_event *event, struct task_struct *parent,
6779 struct perf_event_context *parent_ctx,
6780 struct task_struct *child, int ctxn,
6784 struct perf_event_context *child_ctx;
6786 if (!event->attr.inherit) {
6791 child_ctx = child->perf_event_ctxp[ctxn];
6794 * This is executed from the parent task context, so
6795 * inherit events that have been marked for cloning.
6796 * First allocate and initialize a context for the
6800 child_ctx = alloc_perf_context(parent_ctx->pmu, child);
6804 child->perf_event_ctxp[ctxn] = child_ctx;
6807 ret = inherit_group(event, parent, parent_ctx,
6817 * Initialize the perf_event context in task_struct
6819 int perf_event_init_context(struct task_struct *child, int ctxn)
6821 struct perf_event_context *child_ctx, *parent_ctx;
6822 struct perf_event_context *cloned_ctx;
6823 struct perf_event *event;
6824 struct task_struct *parent = current;
6825 int inherited_all = 1;
6826 unsigned long flags;
6829 if (likely(!parent->perf_event_ctxp[ctxn]))
6833 * If the parent's context is a clone, pin it so it won't get
6836 parent_ctx = perf_pin_task_context(parent, ctxn);
6839 * No need to check if parent_ctx != NULL here; since we saw
6840 * it non-NULL earlier, the only reason for it to become NULL
6841 * is if we exit, and since we're currently in the middle of
6842 * a fork we can't be exiting at the same time.
6846 * Lock the parent list. No need to lock the child - not PID
6847 * hashed yet and not running, so nobody can access it.
6849 mutex_lock(&parent_ctx->mutex);
6852 * We dont have to disable NMIs - we are only looking at
6853 * the list, not manipulating it:
6855 list_for_each_entry(event, &parent_ctx->pinned_groups, group_entry) {
6856 ret = inherit_task_group(event, parent, parent_ctx,
6857 child, ctxn, &inherited_all);
6863 * We can't hold ctx->lock when iterating the ->flexible_group list due
6864 * to allocations, but we need to prevent rotation because
6865 * rotate_ctx() will change the list from interrupt context.
6867 raw_spin_lock_irqsave(&parent_ctx->lock, flags);
6868 parent_ctx->rotate_disable = 1;
6869 raw_spin_unlock_irqrestore(&parent_ctx->lock, flags);
6871 list_for_each_entry(event, &parent_ctx->flexible_groups, group_entry) {
6872 ret = inherit_task_group(event, parent, parent_ctx,
6873 child, ctxn, &inherited_all);
6878 raw_spin_lock_irqsave(&parent_ctx->lock, flags);
6879 parent_ctx->rotate_disable = 0;
6881 child_ctx = child->perf_event_ctxp[ctxn];
6883 if (child_ctx && inherited_all) {
6885 * Mark the child context as a clone of the parent
6886 * context, or of whatever the parent is a clone of.
6888 * Note that if the parent is a clone, the holding of
6889 * parent_ctx->lock avoids it from being uncloned.
6891 cloned_ctx = parent_ctx->parent_ctx;
6893 child_ctx->parent_ctx = cloned_ctx;
6894 child_ctx->parent_gen = parent_ctx->parent_gen;
6896 child_ctx->parent_ctx = parent_ctx;
6897 child_ctx->parent_gen = parent_ctx->generation;
6899 get_ctx(child_ctx->parent_ctx);
6902 raw_spin_unlock_irqrestore(&parent_ctx->lock, flags);
6903 mutex_unlock(&parent_ctx->mutex);
6905 perf_unpin_context(parent_ctx);
6906 put_ctx(parent_ctx);
6912 * Initialize the perf_event context in task_struct
6914 int perf_event_init_task(struct task_struct *child)
6918 memset(child->perf_event_ctxp, 0, sizeof(child->perf_event_ctxp));
6919 mutex_init(&child->perf_event_mutex);
6920 INIT_LIST_HEAD(&child->perf_event_list);
6922 for_each_task_context_nr(ctxn) {
6923 ret = perf_event_init_context(child, ctxn);
6931 static void __init perf_event_init_all_cpus(void)
6933 struct swevent_htable *swhash;
6936 for_each_possible_cpu(cpu) {
6937 swhash = &per_cpu(swevent_htable, cpu);
6938 mutex_init(&swhash->hlist_mutex);
6939 INIT_LIST_HEAD(&per_cpu(rotation_list, cpu));
6943 static void __cpuinit perf_event_init_cpu(int cpu)
6945 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
6947 mutex_lock(&swhash->hlist_mutex);
6948 if (swhash->hlist_refcount > 0) {
6949 struct swevent_hlist *hlist;
6951 hlist = kzalloc_node(sizeof(*hlist), GFP_KERNEL, cpu_to_node(cpu));
6953 rcu_assign_pointer(swhash->swevent_hlist, hlist);
6955 mutex_unlock(&swhash->hlist_mutex);
6958 #if defined CONFIG_HOTPLUG_CPU || defined CONFIG_KEXEC
6959 static void perf_pmu_rotate_stop(struct pmu *pmu)
6961 struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
6963 WARN_ON(!irqs_disabled());
6965 list_del_init(&cpuctx->rotation_list);
6968 static void __perf_event_exit_context(void *__info)
6970 struct perf_event_context *ctx = __info;
6971 struct perf_event *event, *tmp;
6973 perf_pmu_rotate_stop(ctx->pmu);
6975 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry)
6976 __perf_remove_from_context(event);
6977 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups, group_entry)
6978 __perf_remove_from_context(event);
6981 static void perf_event_exit_cpu_context(int cpu)
6983 struct perf_event_context *ctx;
6987 idx = srcu_read_lock(&pmus_srcu);
6988 list_for_each_entry_rcu(pmu, &pmus, entry) {
6989 ctx = &per_cpu_ptr(pmu->pmu_cpu_context, cpu)->ctx;
6991 mutex_lock(&ctx->mutex);
6992 smp_call_function_single(cpu, __perf_event_exit_context, ctx, 1);
6993 mutex_unlock(&ctx->mutex);
6995 srcu_read_unlock(&pmus_srcu, idx);
6998 static void perf_event_exit_cpu(int cpu)
7000 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
7002 mutex_lock(&swhash->hlist_mutex);
7003 swevent_hlist_release(swhash);
7004 mutex_unlock(&swhash->hlist_mutex);
7006 perf_event_exit_cpu_context(cpu);
7009 static inline void perf_event_exit_cpu(int cpu) { }
7013 perf_reboot(struct notifier_block *notifier, unsigned long val, void *v)
7017 for_each_online_cpu(cpu)
7018 perf_event_exit_cpu(cpu);
7024 * Run the perf reboot notifier at the very last possible moment so that
7025 * the generic watchdog code runs as long as possible.
7027 static struct notifier_block perf_reboot_notifier = {
7028 .notifier_call = perf_reboot,
7029 .priority = INT_MIN,
7032 static int __cpuinit
7033 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
7035 unsigned int cpu = (long)hcpu;
7037 switch (action & ~CPU_TASKS_FROZEN) {
7039 case CPU_UP_PREPARE:
7040 case CPU_DOWN_FAILED:
7041 perf_event_init_cpu(cpu);
7044 case CPU_UP_CANCELED:
7045 case CPU_DOWN_PREPARE:
7046 perf_event_exit_cpu(cpu);
7056 void __init perf_event_init(void)
7062 perf_event_init_all_cpus();
7063 init_srcu_struct(&pmus_srcu);
7064 perf_pmu_register(&perf_swevent, "software", PERF_TYPE_SOFTWARE);
7065 perf_pmu_register(&perf_cpu_clock, NULL, -1);
7066 perf_pmu_register(&perf_task_clock, NULL, -1);
7068 perf_cpu_notifier(perf_cpu_notify);
7069 register_reboot_notifier(&perf_reboot_notifier);
7071 ret = init_hw_breakpoint();
7072 WARN(ret, "hw_breakpoint initialization failed with: %d", ret);
7075 static int __init perf_event_sysfs_init(void)
7080 mutex_lock(&pmus_lock);
7082 ret = bus_register(&pmu_bus);
7086 list_for_each_entry(pmu, &pmus, entry) {
7087 if (!pmu->name || pmu->type < 0)
7090 ret = pmu_dev_alloc(pmu);
7091 WARN(ret, "Failed to register pmu: %s, reason %d\n", pmu->name, ret);
7093 pmu_bus_running = 1;
7097 mutex_unlock(&pmus_lock);
7101 device_initcall(perf_event_sysfs_init);
7103 #ifdef CONFIG_CGROUP_PERF
7104 static struct cgroup_subsys_state *perf_cgroup_create(
7105 struct cgroup_subsys *ss, struct cgroup *cont)
7107 struct perf_cgroup *jc;
7109 jc = kzalloc(sizeof(*jc), GFP_KERNEL);
7111 return ERR_PTR(-ENOMEM);
7113 jc->info = alloc_percpu(struct perf_cgroup_info);
7116 return ERR_PTR(-ENOMEM);
7122 static void perf_cgroup_destroy(struct cgroup_subsys *ss,
7123 struct cgroup *cont)
7125 struct perf_cgroup *jc;
7126 jc = container_of(cgroup_subsys_state(cont, perf_subsys_id),
7127 struct perf_cgroup, css);
7128 free_percpu(jc->info);
7132 static int __perf_cgroup_move(void *info)
7134 struct task_struct *task = info;
7135 perf_cgroup_switch(task, PERF_CGROUP_SWOUT | PERF_CGROUP_SWIN);
7140 perf_cgroup_attach_task(struct cgroup *cgrp, struct task_struct *task)
7142 task_function_call(task, __perf_cgroup_move, task);
7145 static void perf_cgroup_exit(struct cgroup_subsys *ss, struct cgroup *cgrp,
7146 struct cgroup *old_cgrp, struct task_struct *task)
7149 * cgroup_exit() is called in the copy_process() failure path.
7150 * Ignore this case since the task hasn't ran yet, this avoids
7151 * trying to poke a half freed task state from generic code.
7153 if (!(task->flags & PF_EXITING))
7156 perf_cgroup_attach_task(cgrp, task);
7159 struct cgroup_subsys perf_subsys = {
7160 .name = "perf_event",
7161 .subsys_id = perf_subsys_id,
7162 .create = perf_cgroup_create,
7163 .destroy = perf_cgroup_destroy,
7164 .exit = perf_cgroup_exit,
7165 .attach_task = perf_cgroup_attach_task,
7167 #endif /* CONFIG_CGROUP_PERF */