2 * Performance events core code:
4 * Copyright (C) 2008 Thomas Gleixner <tglx@linutronix.de>
5 * Copyright (C) 2008-2009 Red Hat, Inc., Ingo Molnar
6 * Copyright (C) 2008-2009 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/vmalloc.h>
29 #include <linux/hardirq.h>
30 #include <linux/rculist.h>
31 #include <linux/uaccess.h>
32 #include <linux/syscalls.h>
33 #include <linux/anon_inodes.h>
34 #include <linux/kernel_stat.h>
35 #include <linux/perf_event.h>
36 #include <linux/ftrace_event.h>
37 #include <linux/hw_breakpoint.h>
39 #include <asm/irq_regs.h>
41 struct remote_function_call {
42 struct task_struct *p;
43 int (*func)(void *info);
48 static void remote_function(void *data)
50 struct remote_function_call *tfc = data;
51 struct task_struct *p = tfc->p;
55 if (task_cpu(p) != smp_processor_id() || !task_curr(p))
59 tfc->ret = tfc->func(tfc->info);
63 * task_function_call - call a function on the cpu on which a task runs
64 * @p: the task to evaluate
65 * @func: the function to be called
66 * @info: the function call argument
68 * Calls the function @func when the task is currently running. This might
69 * be on the current CPU, which just calls the function directly
71 * returns: @func return value, or
72 * -ESRCH - when the process isn't running
73 * -EAGAIN - when the process moved away
76 task_function_call(struct task_struct *p, int (*func) (void *info), void *info)
78 struct remote_function_call data = {
82 .ret = -ESRCH, /* No such (running) process */
86 smp_call_function_single(task_cpu(p), remote_function, &data, 1);
92 * cpu_function_call - call a function on the cpu
93 * @func: the function to be called
94 * @info: the function call argument
96 * Calls the function @func on the remote cpu.
98 * returns: @func return value or -ENXIO when the cpu is offline
100 static int cpu_function_call(int cpu, int (*func) (void *info), void *info)
102 struct remote_function_call data = {
106 .ret = -ENXIO, /* No such CPU */
109 smp_call_function_single(cpu, remote_function, &data, 1);
114 #define PERF_FLAG_ALL (PERF_FLAG_FD_NO_GROUP |\
115 PERF_FLAG_FD_OUTPUT |\
116 PERF_FLAG_PID_CGROUP)
119 EVENT_FLEXIBLE = 0x1,
121 EVENT_ALL = EVENT_FLEXIBLE | EVENT_PINNED,
125 * perf_sched_events : >0 events exist
126 * perf_cgroup_events: >0 per-cpu cgroup events exist on this cpu
128 atomic_t perf_sched_events __read_mostly;
129 static DEFINE_PER_CPU(atomic_t, perf_cgroup_events);
131 static atomic_t nr_mmap_events __read_mostly;
132 static atomic_t nr_comm_events __read_mostly;
133 static atomic_t nr_task_events __read_mostly;
135 static LIST_HEAD(pmus);
136 static DEFINE_MUTEX(pmus_lock);
137 static struct srcu_struct pmus_srcu;
140 * perf event paranoia level:
141 * -1 - not paranoid at all
142 * 0 - disallow raw tracepoint access for unpriv
143 * 1 - disallow cpu events for unpriv
144 * 2 - disallow kernel profiling for unpriv
146 int sysctl_perf_event_paranoid __read_mostly = 1;
148 int sysctl_perf_event_mlock __read_mostly = 512; /* 'free' kb per user */
151 * max perf event sample rate
153 #define DEFAULT_MAX_SAMPLE_RATE 100000
154 int sysctl_perf_event_sample_rate __read_mostly = DEFAULT_MAX_SAMPLE_RATE;
155 static int max_samples_per_tick __read_mostly =
156 DIV_ROUND_UP(DEFAULT_MAX_SAMPLE_RATE, HZ);
158 int perf_proc_update_handler(struct ctl_table *table, int write,
159 void __user *buffer, size_t *lenp,
162 int ret = proc_dointvec(table, write, buffer, lenp, ppos);
167 max_samples_per_tick = DIV_ROUND_UP(sysctl_perf_event_sample_rate, HZ);
172 static atomic64_t perf_event_id;
174 static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
175 enum event_type_t event_type);
177 static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
178 enum event_type_t event_type,
179 struct task_struct *task);
181 static void update_context_time(struct perf_event_context *ctx);
182 static u64 perf_event_time(struct perf_event *event);
184 void __weak perf_event_print_debug(void) { }
186 extern __weak const char *perf_pmu_name(void)
191 static inline u64 perf_clock(void)
193 return local_clock();
196 static inline struct perf_cpu_context *
197 __get_cpu_context(struct perf_event_context *ctx)
199 return this_cpu_ptr(ctx->pmu->pmu_cpu_context);
202 #ifdef CONFIG_CGROUP_PERF
205 * Must ensure cgroup is pinned (css_get) before calling
206 * this function. In other words, we cannot call this function
207 * if there is no cgroup event for the current CPU context.
209 static inline struct perf_cgroup *
210 perf_cgroup_from_task(struct task_struct *task)
212 return container_of(task_subsys_state(task, perf_subsys_id),
213 struct perf_cgroup, css);
217 perf_cgroup_match(struct perf_event *event)
219 struct perf_event_context *ctx = event->ctx;
220 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
222 return !event->cgrp || event->cgrp == cpuctx->cgrp;
225 static inline void perf_get_cgroup(struct perf_event *event)
227 css_get(&event->cgrp->css);
230 static inline void perf_put_cgroup(struct perf_event *event)
232 css_put(&event->cgrp->css);
235 static inline void perf_detach_cgroup(struct perf_event *event)
237 perf_put_cgroup(event);
241 static inline int is_cgroup_event(struct perf_event *event)
243 return event->cgrp != NULL;
246 static inline u64 perf_cgroup_event_time(struct perf_event *event)
248 struct perf_cgroup_info *t;
250 t = per_cpu_ptr(event->cgrp->info, event->cpu);
254 static inline void __update_cgrp_time(struct perf_cgroup *cgrp)
256 struct perf_cgroup_info *info;
261 info = this_cpu_ptr(cgrp->info);
263 info->time += now - info->timestamp;
264 info->timestamp = now;
267 static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx)
269 struct perf_cgroup *cgrp_out = cpuctx->cgrp;
271 __update_cgrp_time(cgrp_out);
274 static inline void update_cgrp_time_from_event(struct perf_event *event)
276 struct perf_cgroup *cgrp;
279 * ensure we access cgroup data only when needed and
280 * when we know the cgroup is pinned (css_get)
282 if (!is_cgroup_event(event))
285 cgrp = perf_cgroup_from_task(current);
287 * Do not update time when cgroup is not active
289 if (cgrp == event->cgrp)
290 __update_cgrp_time(event->cgrp);
294 perf_cgroup_set_timestamp(struct task_struct *task,
295 struct perf_event_context *ctx)
297 struct perf_cgroup *cgrp;
298 struct perf_cgroup_info *info;
301 * ctx->lock held by caller
302 * ensure we do not access cgroup data
303 * unless we have the cgroup pinned (css_get)
305 if (!task || !ctx->nr_cgroups)
308 cgrp = perf_cgroup_from_task(task);
309 info = this_cpu_ptr(cgrp->info);
310 info->timestamp = ctx->timestamp;
313 #define PERF_CGROUP_SWOUT 0x1 /* cgroup switch out every event */
314 #define PERF_CGROUP_SWIN 0x2 /* cgroup switch in events based on task */
317 * reschedule events based on the cgroup constraint of task.
319 * mode SWOUT : schedule out everything
320 * mode SWIN : schedule in based on cgroup for next
322 void perf_cgroup_switch(struct task_struct *task, int mode)
324 struct perf_cpu_context *cpuctx;
329 * disable interrupts to avoid geting nr_cgroup
330 * changes via __perf_event_disable(). Also
333 local_irq_save(flags);
336 * we reschedule only in the presence of cgroup
337 * constrained events.
341 list_for_each_entry_rcu(pmu, &pmus, entry) {
343 cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
345 perf_pmu_disable(cpuctx->ctx.pmu);
348 * perf_cgroup_events says at least one
349 * context on this CPU has cgroup events.
351 * ctx->nr_cgroups reports the number of cgroup
352 * events for a context.
354 if (cpuctx->ctx.nr_cgroups > 0) {
356 if (mode & PERF_CGROUP_SWOUT) {
357 cpu_ctx_sched_out(cpuctx, EVENT_ALL);
359 * must not be done before ctxswout due
360 * to event_filter_match() in event_sched_out()
365 if (mode & PERF_CGROUP_SWIN) {
366 /* set cgrp before ctxsw in to
367 * allow event_filter_match() to not
368 * have to pass task around
370 cpuctx->cgrp = perf_cgroup_from_task(task);
371 cpu_ctx_sched_in(cpuctx, EVENT_ALL, task);
375 perf_pmu_enable(cpuctx->ctx.pmu);
380 local_irq_restore(flags);
383 static inline void perf_cgroup_sched_out(struct task_struct *task)
385 perf_cgroup_switch(task, PERF_CGROUP_SWOUT);
388 static inline void perf_cgroup_sched_in(struct task_struct *task)
390 perf_cgroup_switch(task, PERF_CGROUP_SWIN);
393 static inline int perf_cgroup_connect(int fd, struct perf_event *event,
394 struct perf_event_attr *attr,
395 struct perf_event *group_leader)
397 struct perf_cgroup *cgrp;
398 struct cgroup_subsys_state *css;
400 int ret = 0, fput_needed;
402 file = fget_light(fd, &fput_needed);
406 css = cgroup_css_from_dir(file, perf_subsys_id);
412 cgrp = container_of(css, struct perf_cgroup, css);
415 /* must be done before we fput() the file */
416 perf_get_cgroup(event);
419 * all events in a group must monitor
420 * the same cgroup because a task belongs
421 * to only one perf cgroup at a time
423 if (group_leader && group_leader->cgrp != cgrp) {
424 perf_detach_cgroup(event);
428 fput_light(file, fput_needed);
433 perf_cgroup_set_shadow_time(struct perf_event *event, u64 now)
435 struct perf_cgroup_info *t;
436 t = per_cpu_ptr(event->cgrp->info, event->cpu);
437 event->shadow_ctx_time = now - t->timestamp;
441 perf_cgroup_defer_enabled(struct perf_event *event)
444 * when the current task's perf cgroup does not match
445 * the event's, we need to remember to call the
446 * perf_mark_enable() function the first time a task with
447 * a matching perf cgroup is scheduled in.
449 if (is_cgroup_event(event) && !perf_cgroup_match(event))
450 event->cgrp_defer_enabled = 1;
454 perf_cgroup_mark_enabled(struct perf_event *event,
455 struct perf_event_context *ctx)
457 struct perf_event *sub;
458 u64 tstamp = perf_event_time(event);
460 if (!event->cgrp_defer_enabled)
463 event->cgrp_defer_enabled = 0;
465 event->tstamp_enabled = tstamp - event->total_time_enabled;
466 list_for_each_entry(sub, &event->sibling_list, group_entry) {
467 if (sub->state >= PERF_EVENT_STATE_INACTIVE) {
468 sub->tstamp_enabled = tstamp - sub->total_time_enabled;
469 sub->cgrp_defer_enabled = 0;
473 #else /* !CONFIG_CGROUP_PERF */
476 perf_cgroup_match(struct perf_event *event)
481 static inline void perf_detach_cgroup(struct perf_event *event)
484 static inline int is_cgroup_event(struct perf_event *event)
489 static inline u64 perf_cgroup_event_cgrp_time(struct perf_event *event)
494 static inline void update_cgrp_time_from_event(struct perf_event *event)
498 static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx)
502 static inline void perf_cgroup_sched_out(struct task_struct *task)
506 static inline void perf_cgroup_sched_in(struct task_struct *task)
510 static inline int perf_cgroup_connect(pid_t pid, struct perf_event *event,
511 struct perf_event_attr *attr,
512 struct perf_event *group_leader)
518 perf_cgroup_set_timestamp(struct task_struct *task,
519 struct perf_event_context *ctx)
524 perf_cgroup_switch(struct task_struct *task, struct task_struct *next)
529 perf_cgroup_set_shadow_time(struct perf_event *event, u64 now)
533 static inline u64 perf_cgroup_event_time(struct perf_event *event)
539 perf_cgroup_defer_enabled(struct perf_event *event)
544 perf_cgroup_mark_enabled(struct perf_event *event,
545 struct perf_event_context *ctx)
550 void perf_pmu_disable(struct pmu *pmu)
552 int *count = this_cpu_ptr(pmu->pmu_disable_count);
554 pmu->pmu_disable(pmu);
557 void perf_pmu_enable(struct pmu *pmu)
559 int *count = this_cpu_ptr(pmu->pmu_disable_count);
561 pmu->pmu_enable(pmu);
564 static DEFINE_PER_CPU(struct list_head, rotation_list);
567 * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
568 * because they're strictly cpu affine and rotate_start is called with IRQs
569 * disabled, while rotate_context is called from IRQ context.
571 static void perf_pmu_rotate_start(struct pmu *pmu)
573 struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
574 struct list_head *head = &__get_cpu_var(rotation_list);
576 WARN_ON(!irqs_disabled());
578 if (list_empty(&cpuctx->rotation_list))
579 list_add(&cpuctx->rotation_list, head);
582 static void get_ctx(struct perf_event_context *ctx)
584 WARN_ON(!atomic_inc_not_zero(&ctx->refcount));
587 static void free_ctx(struct rcu_head *head)
589 struct perf_event_context *ctx;
591 ctx = container_of(head, struct perf_event_context, rcu_head);
595 static void put_ctx(struct perf_event_context *ctx)
597 if (atomic_dec_and_test(&ctx->refcount)) {
599 put_ctx(ctx->parent_ctx);
601 put_task_struct(ctx->task);
602 call_rcu(&ctx->rcu_head, free_ctx);
606 static void unclone_ctx(struct perf_event_context *ctx)
608 if (ctx->parent_ctx) {
609 put_ctx(ctx->parent_ctx);
610 ctx->parent_ctx = NULL;
614 static u32 perf_event_pid(struct perf_event *event, struct task_struct *p)
617 * only top level events have the pid namespace they were created in
620 event = event->parent;
622 return task_tgid_nr_ns(p, event->ns);
625 static u32 perf_event_tid(struct perf_event *event, struct task_struct *p)
628 * only top level events have the pid namespace they were created in
631 event = event->parent;
633 return task_pid_nr_ns(p, event->ns);
637 * If we inherit events we want to return the parent event id
640 static u64 primary_event_id(struct perf_event *event)
645 id = event->parent->id;
651 * Get the perf_event_context for a task and lock it.
652 * This has to cope with with the fact that until it is locked,
653 * the context could get moved to another task.
655 static struct perf_event_context *
656 perf_lock_task_context(struct task_struct *task, int ctxn, unsigned long *flags)
658 struct perf_event_context *ctx;
662 ctx = rcu_dereference(task->perf_event_ctxp[ctxn]);
665 * If this context is a clone of another, it might
666 * get swapped for another underneath us by
667 * perf_event_task_sched_out, though the
668 * rcu_read_lock() protects us from any context
669 * getting freed. Lock the context and check if it
670 * got swapped before we could get the lock, and retry
671 * if so. If we locked the right context, then it
672 * can't get swapped on us any more.
674 raw_spin_lock_irqsave(&ctx->lock, *flags);
675 if (ctx != rcu_dereference(task->perf_event_ctxp[ctxn])) {
676 raw_spin_unlock_irqrestore(&ctx->lock, *flags);
680 if (!atomic_inc_not_zero(&ctx->refcount)) {
681 raw_spin_unlock_irqrestore(&ctx->lock, *flags);
690 * Get the context for a task and increment its pin_count so it
691 * can't get swapped to another task. This also increments its
692 * reference count so that the context can't get freed.
694 static struct perf_event_context *
695 perf_pin_task_context(struct task_struct *task, int ctxn)
697 struct perf_event_context *ctx;
700 ctx = perf_lock_task_context(task, ctxn, &flags);
703 raw_spin_unlock_irqrestore(&ctx->lock, flags);
708 static void perf_unpin_context(struct perf_event_context *ctx)
712 raw_spin_lock_irqsave(&ctx->lock, flags);
714 raw_spin_unlock_irqrestore(&ctx->lock, flags);
718 * Update the record of the current time in a context.
720 static void update_context_time(struct perf_event_context *ctx)
722 u64 now = perf_clock();
724 ctx->time += now - ctx->timestamp;
725 ctx->timestamp = now;
728 static u64 perf_event_time(struct perf_event *event)
730 struct perf_event_context *ctx = event->ctx;
732 if (is_cgroup_event(event))
733 return perf_cgroup_event_time(event);
735 return ctx ? ctx->time : 0;
739 * Update the total_time_enabled and total_time_running fields for a event.
741 static void update_event_times(struct perf_event *event)
743 struct perf_event_context *ctx = event->ctx;
746 if (event->state < PERF_EVENT_STATE_INACTIVE ||
747 event->group_leader->state < PERF_EVENT_STATE_INACTIVE)
750 * in cgroup mode, time_enabled represents
751 * the time the event was enabled AND active
752 * tasks were in the monitored cgroup. This is
753 * independent of the activity of the context as
754 * there may be a mix of cgroup and non-cgroup events.
756 * That is why we treat cgroup events differently
759 if (is_cgroup_event(event))
760 run_end = perf_event_time(event);
761 else if (ctx->is_active)
764 run_end = event->tstamp_stopped;
766 event->total_time_enabled = run_end - event->tstamp_enabled;
768 if (event->state == PERF_EVENT_STATE_INACTIVE)
769 run_end = event->tstamp_stopped;
771 run_end = perf_event_time(event);
773 event->total_time_running = run_end - event->tstamp_running;
778 * Update total_time_enabled and total_time_running for all events in a group.
780 static void update_group_times(struct perf_event *leader)
782 struct perf_event *event;
784 update_event_times(leader);
785 list_for_each_entry(event, &leader->sibling_list, group_entry)
786 update_event_times(event);
789 static struct list_head *
790 ctx_group_list(struct perf_event *event, struct perf_event_context *ctx)
792 if (event->attr.pinned)
793 return &ctx->pinned_groups;
795 return &ctx->flexible_groups;
799 * Add a event from the lists for its context.
800 * Must be called with ctx->mutex and ctx->lock held.
803 list_add_event(struct perf_event *event, struct perf_event_context *ctx)
805 WARN_ON_ONCE(event->attach_state & PERF_ATTACH_CONTEXT);
806 event->attach_state |= PERF_ATTACH_CONTEXT;
809 * If we're a stand alone event or group leader, we go to the context
810 * list, group events are kept attached to the group so that
811 * perf_group_detach can, at all times, locate all siblings.
813 if (event->group_leader == event) {
814 struct list_head *list;
816 if (is_software_event(event))
817 event->group_flags |= PERF_GROUP_SOFTWARE;
819 list = ctx_group_list(event, ctx);
820 list_add_tail(&event->group_entry, list);
823 if (is_cgroup_event(event)) {
827 * - that has cgroup constraint on event->cpu
828 * - that may need work on context switch
830 atomic_inc(&per_cpu(perf_cgroup_events, event->cpu));
831 jump_label_inc(&perf_sched_events);
834 list_add_rcu(&event->event_entry, &ctx->event_list);
836 perf_pmu_rotate_start(ctx->pmu);
838 if (event->attr.inherit_stat)
843 * Called at perf_event creation and when events are attached/detached from a
846 static void perf_event__read_size(struct perf_event *event)
848 int entry = sizeof(u64); /* value */
852 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
855 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
858 if (event->attr.read_format & PERF_FORMAT_ID)
859 entry += sizeof(u64);
861 if (event->attr.read_format & PERF_FORMAT_GROUP) {
862 nr += event->group_leader->nr_siblings;
867 event->read_size = size;
870 static void perf_event__header_size(struct perf_event *event)
872 struct perf_sample_data *data;
873 u64 sample_type = event->attr.sample_type;
876 perf_event__read_size(event);
878 if (sample_type & PERF_SAMPLE_IP)
879 size += sizeof(data->ip);
881 if (sample_type & PERF_SAMPLE_ADDR)
882 size += sizeof(data->addr);
884 if (sample_type & PERF_SAMPLE_PERIOD)
885 size += sizeof(data->period);
887 if (sample_type & PERF_SAMPLE_READ)
888 size += event->read_size;
890 event->header_size = size;
893 static void perf_event__id_header_size(struct perf_event *event)
895 struct perf_sample_data *data;
896 u64 sample_type = event->attr.sample_type;
899 if (sample_type & PERF_SAMPLE_TID)
900 size += sizeof(data->tid_entry);
902 if (sample_type & PERF_SAMPLE_TIME)
903 size += sizeof(data->time);
905 if (sample_type & PERF_SAMPLE_ID)
906 size += sizeof(data->id);
908 if (sample_type & PERF_SAMPLE_STREAM_ID)
909 size += sizeof(data->stream_id);
911 if (sample_type & PERF_SAMPLE_CPU)
912 size += sizeof(data->cpu_entry);
914 event->id_header_size = size;
917 static void perf_group_attach(struct perf_event *event)
919 struct perf_event *group_leader = event->group_leader, *pos;
922 * We can have double attach due to group movement in perf_event_open.
924 if (event->attach_state & PERF_ATTACH_GROUP)
927 event->attach_state |= PERF_ATTACH_GROUP;
929 if (group_leader == event)
932 if (group_leader->group_flags & PERF_GROUP_SOFTWARE &&
933 !is_software_event(event))
934 group_leader->group_flags &= ~PERF_GROUP_SOFTWARE;
936 list_add_tail(&event->group_entry, &group_leader->sibling_list);
937 group_leader->nr_siblings++;
939 perf_event__header_size(group_leader);
941 list_for_each_entry(pos, &group_leader->sibling_list, group_entry)
942 perf_event__header_size(pos);
946 * Remove a event from the lists for its context.
947 * Must be called with ctx->mutex and ctx->lock held.
950 list_del_event(struct perf_event *event, struct perf_event_context *ctx)
953 * We can have double detach due to exit/hot-unplug + close.
955 if (!(event->attach_state & PERF_ATTACH_CONTEXT))
958 event->attach_state &= ~PERF_ATTACH_CONTEXT;
960 if (is_cgroup_event(event)) {
962 atomic_dec(&per_cpu(perf_cgroup_events, event->cpu));
963 jump_label_dec(&perf_sched_events);
967 if (event->attr.inherit_stat)
970 list_del_rcu(&event->event_entry);
972 if (event->group_leader == event)
973 list_del_init(&event->group_entry);
975 update_group_times(event);
978 * If event was in error state, then keep it
979 * that way, otherwise bogus counts will be
980 * returned on read(). The only way to get out
981 * of error state is by explicit re-enabling
984 if (event->state > PERF_EVENT_STATE_OFF)
985 event->state = PERF_EVENT_STATE_OFF;
988 static void perf_group_detach(struct perf_event *event)
990 struct perf_event *sibling, *tmp;
991 struct list_head *list = NULL;
994 * We can have double detach due to exit/hot-unplug + close.
996 if (!(event->attach_state & PERF_ATTACH_GROUP))
999 event->attach_state &= ~PERF_ATTACH_GROUP;
1002 * If this is a sibling, remove it from its group.
1004 if (event->group_leader != event) {
1005 list_del_init(&event->group_entry);
1006 event->group_leader->nr_siblings--;
1010 if (!list_empty(&event->group_entry))
1011 list = &event->group_entry;
1014 * If this was a group event with sibling events then
1015 * upgrade the siblings to singleton events by adding them
1016 * to whatever list we are on.
1018 list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) {
1020 list_move_tail(&sibling->group_entry, list);
1021 sibling->group_leader = sibling;
1023 /* Inherit group flags from the previous leader */
1024 sibling->group_flags = event->group_flags;
1028 perf_event__header_size(event->group_leader);
1030 list_for_each_entry(tmp, &event->group_leader->sibling_list, group_entry)
1031 perf_event__header_size(tmp);
1035 event_filter_match(struct perf_event *event)
1037 return (event->cpu == -1 || event->cpu == smp_processor_id())
1038 && perf_cgroup_match(event);
1042 event_sched_out(struct perf_event *event,
1043 struct perf_cpu_context *cpuctx,
1044 struct perf_event_context *ctx)
1046 u64 tstamp = perf_event_time(event);
1049 * An event which could not be activated because of
1050 * filter mismatch still needs to have its timings
1051 * maintained, otherwise bogus information is return
1052 * via read() for time_enabled, time_running:
1054 if (event->state == PERF_EVENT_STATE_INACTIVE
1055 && !event_filter_match(event)) {
1056 delta = tstamp - event->tstamp_stopped;
1057 event->tstamp_running += delta;
1058 event->tstamp_stopped = tstamp;
1061 if (event->state != PERF_EVENT_STATE_ACTIVE)
1064 event->state = PERF_EVENT_STATE_INACTIVE;
1065 if (event->pending_disable) {
1066 event->pending_disable = 0;
1067 event->state = PERF_EVENT_STATE_OFF;
1069 event->tstamp_stopped = tstamp;
1070 event->pmu->del(event, 0);
1073 if (!is_software_event(event))
1074 cpuctx->active_oncpu--;
1076 if (event->attr.exclusive || !cpuctx->active_oncpu)
1077 cpuctx->exclusive = 0;
1081 group_sched_out(struct perf_event *group_event,
1082 struct perf_cpu_context *cpuctx,
1083 struct perf_event_context *ctx)
1085 struct perf_event *event;
1086 int state = group_event->state;
1088 event_sched_out(group_event, cpuctx, ctx);
1091 * Schedule out siblings (if any):
1093 list_for_each_entry(event, &group_event->sibling_list, group_entry)
1094 event_sched_out(event, cpuctx, ctx);
1096 if (state == PERF_EVENT_STATE_ACTIVE && group_event->attr.exclusive)
1097 cpuctx->exclusive = 0;
1101 * Cross CPU call to remove a performance event
1103 * We disable the event on the hardware level first. After that we
1104 * remove it from the context list.
1106 static int __perf_remove_from_context(void *info)
1108 struct perf_event *event = info;
1109 struct perf_event_context *ctx = event->ctx;
1110 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1112 raw_spin_lock(&ctx->lock);
1113 event_sched_out(event, cpuctx, ctx);
1114 list_del_event(event, ctx);
1115 raw_spin_unlock(&ctx->lock);
1122 * Remove the event from a task's (or a CPU's) list of events.
1124 * CPU events are removed with a smp call. For task events we only
1125 * call when the task is on a CPU.
1127 * If event->ctx is a cloned context, callers must make sure that
1128 * every task struct that event->ctx->task could possibly point to
1129 * remains valid. This is OK when called from perf_release since
1130 * that only calls us on the top-level context, which can't be a clone.
1131 * When called from perf_event_exit_task, it's OK because the
1132 * context has been detached from its task.
1134 static void perf_remove_from_context(struct perf_event *event)
1136 struct perf_event_context *ctx = event->ctx;
1137 struct task_struct *task = ctx->task;
1139 lockdep_assert_held(&ctx->mutex);
1143 * Per cpu events are removed via an smp call and
1144 * the removal is always successful.
1146 cpu_function_call(event->cpu, __perf_remove_from_context, event);
1151 if (!task_function_call(task, __perf_remove_from_context, event))
1154 raw_spin_lock_irq(&ctx->lock);
1156 * If we failed to find a running task, but find the context active now
1157 * that we've acquired the ctx->lock, retry.
1159 if (ctx->is_active) {
1160 raw_spin_unlock_irq(&ctx->lock);
1165 * Since the task isn't running, its safe to remove the event, us
1166 * holding the ctx->lock ensures the task won't get scheduled in.
1168 list_del_event(event, ctx);
1169 raw_spin_unlock_irq(&ctx->lock);
1173 * Cross CPU call to disable a performance event
1175 static int __perf_event_disable(void *info)
1177 struct perf_event *event = info;
1178 struct perf_event_context *ctx = event->ctx;
1179 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1182 * If this is a per-task event, need to check whether this
1183 * event's task is the current task on this cpu.
1185 * Can trigger due to concurrent perf_event_context_sched_out()
1186 * flipping contexts around.
1188 if (ctx->task && cpuctx->task_ctx != ctx)
1191 raw_spin_lock(&ctx->lock);
1194 * If the event is on, turn it off.
1195 * If it is in error state, leave it in error state.
1197 if (event->state >= PERF_EVENT_STATE_INACTIVE) {
1198 update_context_time(ctx);
1199 update_cgrp_time_from_event(event);
1200 update_group_times(event);
1201 if (event == event->group_leader)
1202 group_sched_out(event, cpuctx, ctx);
1204 event_sched_out(event, cpuctx, ctx);
1205 event->state = PERF_EVENT_STATE_OFF;
1208 raw_spin_unlock(&ctx->lock);
1216 * If event->ctx is a cloned context, callers must make sure that
1217 * every task struct that event->ctx->task could possibly point to
1218 * remains valid. This condition is satisifed when called through
1219 * perf_event_for_each_child or perf_event_for_each because they
1220 * hold the top-level event's child_mutex, so any descendant that
1221 * goes to exit will block in sync_child_event.
1222 * When called from perf_pending_event it's OK because event->ctx
1223 * is the current context on this CPU and preemption is disabled,
1224 * hence we can't get into perf_event_task_sched_out for this context.
1226 void perf_event_disable(struct perf_event *event)
1228 struct perf_event_context *ctx = event->ctx;
1229 struct task_struct *task = ctx->task;
1233 * Disable the event on the cpu that it's on
1235 cpu_function_call(event->cpu, __perf_event_disable, event);
1240 if (!task_function_call(task, __perf_event_disable, event))
1243 raw_spin_lock_irq(&ctx->lock);
1245 * If the event is still active, we need to retry the cross-call.
1247 if (event->state == PERF_EVENT_STATE_ACTIVE) {
1248 raw_spin_unlock_irq(&ctx->lock);
1250 * Reload the task pointer, it might have been changed by
1251 * a concurrent perf_event_context_sched_out().
1258 * Since we have the lock this context can't be scheduled
1259 * in, so we can change the state safely.
1261 if (event->state == PERF_EVENT_STATE_INACTIVE) {
1262 update_group_times(event);
1263 event->state = PERF_EVENT_STATE_OFF;
1265 raw_spin_unlock_irq(&ctx->lock);
1268 static void perf_set_shadow_time(struct perf_event *event,
1269 struct perf_event_context *ctx,
1273 * use the correct time source for the time snapshot
1275 * We could get by without this by leveraging the
1276 * fact that to get to this function, the caller
1277 * has most likely already called update_context_time()
1278 * and update_cgrp_time_xx() and thus both timestamp
1279 * are identical (or very close). Given that tstamp is,
1280 * already adjusted for cgroup, we could say that:
1281 * tstamp - ctx->timestamp
1283 * tstamp - cgrp->timestamp.
1285 * Then, in perf_output_read(), the calculation would
1286 * work with no changes because:
1287 * - event is guaranteed scheduled in
1288 * - no scheduled out in between
1289 * - thus the timestamp would be the same
1291 * But this is a bit hairy.
1293 * So instead, we have an explicit cgroup call to remain
1294 * within the time time source all along. We believe it
1295 * is cleaner and simpler to understand.
1297 if (is_cgroup_event(event))
1298 perf_cgroup_set_shadow_time(event, tstamp);
1300 event->shadow_ctx_time = tstamp - ctx->timestamp;
1303 #define MAX_INTERRUPTS (~0ULL)
1305 static void perf_log_throttle(struct perf_event *event, int enable);
1308 event_sched_in(struct perf_event *event,
1309 struct perf_cpu_context *cpuctx,
1310 struct perf_event_context *ctx)
1312 u64 tstamp = perf_event_time(event);
1314 if (event->state <= PERF_EVENT_STATE_OFF)
1317 event->state = PERF_EVENT_STATE_ACTIVE;
1318 event->oncpu = smp_processor_id();
1321 * Unthrottle events, since we scheduled we might have missed several
1322 * ticks already, also for a heavily scheduling task there is little
1323 * guarantee it'll get a tick in a timely manner.
1325 if (unlikely(event->hw.interrupts == MAX_INTERRUPTS)) {
1326 perf_log_throttle(event, 1);
1327 event->hw.interrupts = 0;
1331 * The new state must be visible before we turn it on in the hardware:
1335 if (event->pmu->add(event, PERF_EF_START)) {
1336 event->state = PERF_EVENT_STATE_INACTIVE;
1341 event->tstamp_running += tstamp - event->tstamp_stopped;
1343 perf_set_shadow_time(event, ctx, tstamp);
1345 if (!is_software_event(event))
1346 cpuctx->active_oncpu++;
1349 if (event->attr.exclusive)
1350 cpuctx->exclusive = 1;
1356 group_sched_in(struct perf_event *group_event,
1357 struct perf_cpu_context *cpuctx,
1358 struct perf_event_context *ctx)
1360 struct perf_event *event, *partial_group = NULL;
1361 struct pmu *pmu = group_event->pmu;
1362 u64 now = ctx->time;
1363 bool simulate = false;
1365 if (group_event->state == PERF_EVENT_STATE_OFF)
1368 pmu->start_txn(pmu);
1370 if (event_sched_in(group_event, cpuctx, ctx)) {
1371 pmu->cancel_txn(pmu);
1376 * Schedule in siblings as one group (if any):
1378 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
1379 if (event_sched_in(event, cpuctx, ctx)) {
1380 partial_group = event;
1385 if (!pmu->commit_txn(pmu))
1390 * Groups can be scheduled in as one unit only, so undo any
1391 * partial group before returning:
1392 * The events up to the failed event are scheduled out normally,
1393 * tstamp_stopped will be updated.
1395 * The failed events and the remaining siblings need to have
1396 * their timings updated as if they had gone thru event_sched_in()
1397 * and event_sched_out(). This is required to get consistent timings
1398 * across the group. This also takes care of the case where the group
1399 * could never be scheduled by ensuring tstamp_stopped is set to mark
1400 * the time the event was actually stopped, such that time delta
1401 * calculation in update_event_times() is correct.
1403 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
1404 if (event == partial_group)
1408 event->tstamp_running += now - event->tstamp_stopped;
1409 event->tstamp_stopped = now;
1411 event_sched_out(event, cpuctx, ctx);
1414 event_sched_out(group_event, cpuctx, ctx);
1416 pmu->cancel_txn(pmu);
1422 * Work out whether we can put this event group on the CPU now.
1424 static int group_can_go_on(struct perf_event *event,
1425 struct perf_cpu_context *cpuctx,
1429 * Groups consisting entirely of software events can always go on.
1431 if (event->group_flags & PERF_GROUP_SOFTWARE)
1434 * If an exclusive group is already on, no other hardware
1437 if (cpuctx->exclusive)
1440 * If this group is exclusive and there are already
1441 * events on the CPU, it can't go on.
1443 if (event->attr.exclusive && cpuctx->active_oncpu)
1446 * Otherwise, try to add it if all previous groups were able
1452 static void add_event_to_ctx(struct perf_event *event,
1453 struct perf_event_context *ctx)
1455 u64 tstamp = perf_event_time(event);
1457 list_add_event(event, ctx);
1458 perf_group_attach(event);
1459 event->tstamp_enabled = tstamp;
1460 event->tstamp_running = tstamp;
1461 event->tstamp_stopped = tstamp;
1464 static void perf_event_context_sched_in(struct perf_event_context *ctx,
1465 struct task_struct *tsk);
1468 * Cross CPU call to install and enable a performance event
1470 * Must be called with ctx->mutex held
1472 static int __perf_install_in_context(void *info)
1474 struct perf_event *event = info;
1475 struct perf_event_context *ctx = event->ctx;
1476 struct perf_event *leader = event->group_leader;
1477 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1481 * In case we're installing a new context to an already running task,
1482 * could also happen before perf_event_task_sched_in() on architectures
1483 * which do context switches with IRQs enabled.
1485 if (ctx->task && !cpuctx->task_ctx)
1486 perf_event_context_sched_in(ctx, ctx->task);
1488 raw_spin_lock(&ctx->lock);
1490 update_context_time(ctx);
1492 * update cgrp time only if current cgrp
1493 * matches event->cgrp. Must be done before
1494 * calling add_event_to_ctx()
1496 update_cgrp_time_from_event(event);
1498 add_event_to_ctx(event, ctx);
1500 if (!event_filter_match(event))
1504 * Don't put the event on if it is disabled or if
1505 * it is in a group and the group isn't on.
1507 if (event->state != PERF_EVENT_STATE_INACTIVE ||
1508 (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE))
1512 * An exclusive event can't go on if there are already active
1513 * hardware events, and no hardware event can go on if there
1514 * is already an exclusive event on.
1516 if (!group_can_go_on(event, cpuctx, 1))
1519 err = event_sched_in(event, cpuctx, ctx);
1523 * This event couldn't go on. If it is in a group
1524 * then we have to pull the whole group off.
1525 * If the event group is pinned then put it in error state.
1527 if (leader != event)
1528 group_sched_out(leader, cpuctx, ctx);
1529 if (leader->attr.pinned) {
1530 update_group_times(leader);
1531 leader->state = PERF_EVENT_STATE_ERROR;
1536 raw_spin_unlock(&ctx->lock);
1542 * Attach a performance event to a context
1544 * First we add the event to the list with the hardware enable bit
1545 * in event->hw_config cleared.
1547 * If the event is attached to a task which is on a CPU we use a smp
1548 * call to enable it in the task context. The task might have been
1549 * scheduled away, but we check this in the smp call again.
1552 perf_install_in_context(struct perf_event_context *ctx,
1553 struct perf_event *event,
1556 struct task_struct *task = ctx->task;
1558 lockdep_assert_held(&ctx->mutex);
1564 * Per cpu events are installed via an smp call and
1565 * the install is always successful.
1567 cpu_function_call(cpu, __perf_install_in_context, event);
1572 if (!task_function_call(task, __perf_install_in_context, event))
1575 raw_spin_lock_irq(&ctx->lock);
1577 * If we failed to find a running task, but find the context active now
1578 * that we've acquired the ctx->lock, retry.
1580 if (ctx->is_active) {
1581 raw_spin_unlock_irq(&ctx->lock);
1586 * Since the task isn't running, its safe to add the event, us holding
1587 * the ctx->lock ensures the task won't get scheduled in.
1589 add_event_to_ctx(event, ctx);
1590 raw_spin_unlock_irq(&ctx->lock);
1594 * Put a event into inactive state and update time fields.
1595 * Enabling the leader of a group effectively enables all
1596 * the group members that aren't explicitly disabled, so we
1597 * have to update their ->tstamp_enabled also.
1598 * Note: this works for group members as well as group leaders
1599 * since the non-leader members' sibling_lists will be empty.
1601 static void __perf_event_mark_enabled(struct perf_event *event,
1602 struct perf_event_context *ctx)
1604 struct perf_event *sub;
1605 u64 tstamp = perf_event_time(event);
1607 event->state = PERF_EVENT_STATE_INACTIVE;
1608 event->tstamp_enabled = tstamp - event->total_time_enabled;
1609 list_for_each_entry(sub, &event->sibling_list, group_entry) {
1610 if (sub->state >= PERF_EVENT_STATE_INACTIVE)
1611 sub->tstamp_enabled = tstamp - sub->total_time_enabled;
1616 * Cross CPU call to enable a performance event
1618 static int __perf_event_enable(void *info)
1620 struct perf_event *event = info;
1621 struct perf_event_context *ctx = event->ctx;
1622 struct perf_event *leader = event->group_leader;
1623 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1626 if (WARN_ON_ONCE(!ctx->is_active))
1629 raw_spin_lock(&ctx->lock);
1630 update_context_time(ctx);
1632 if (event->state >= PERF_EVENT_STATE_INACTIVE)
1636 * set current task's cgroup time reference point
1638 perf_cgroup_set_timestamp(current, ctx);
1640 __perf_event_mark_enabled(event, ctx);
1642 if (!event_filter_match(event)) {
1643 if (is_cgroup_event(event))
1644 perf_cgroup_defer_enabled(event);
1649 * If the event is in a group and isn't the group leader,
1650 * then don't put it on unless the group is on.
1652 if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE)
1655 if (!group_can_go_on(event, cpuctx, 1)) {
1658 if (event == leader)
1659 err = group_sched_in(event, cpuctx, ctx);
1661 err = event_sched_in(event, cpuctx, ctx);
1666 * If this event can't go on and it's part of a
1667 * group, then the whole group has to come off.
1669 if (leader != event)
1670 group_sched_out(leader, cpuctx, ctx);
1671 if (leader->attr.pinned) {
1672 update_group_times(leader);
1673 leader->state = PERF_EVENT_STATE_ERROR;
1678 raw_spin_unlock(&ctx->lock);
1686 * If event->ctx is a cloned context, callers must make sure that
1687 * every task struct that event->ctx->task could possibly point to
1688 * remains valid. This condition is satisfied when called through
1689 * perf_event_for_each_child or perf_event_for_each as described
1690 * for perf_event_disable.
1692 void perf_event_enable(struct perf_event *event)
1694 struct perf_event_context *ctx = event->ctx;
1695 struct task_struct *task = ctx->task;
1699 * Enable the event on the cpu that it's on
1701 cpu_function_call(event->cpu, __perf_event_enable, event);
1705 raw_spin_lock_irq(&ctx->lock);
1706 if (event->state >= PERF_EVENT_STATE_INACTIVE)
1710 * If the event is in error state, clear that first.
1711 * That way, if we see the event in error state below, we
1712 * know that it has gone back into error state, as distinct
1713 * from the task having been scheduled away before the
1714 * cross-call arrived.
1716 if (event->state == PERF_EVENT_STATE_ERROR)
1717 event->state = PERF_EVENT_STATE_OFF;
1720 if (!ctx->is_active) {
1721 __perf_event_mark_enabled(event, ctx);
1725 raw_spin_unlock_irq(&ctx->lock);
1727 if (!task_function_call(task, __perf_event_enable, event))
1730 raw_spin_lock_irq(&ctx->lock);
1733 * If the context is active and the event is still off,
1734 * we need to retry the cross-call.
1736 if (ctx->is_active && event->state == PERF_EVENT_STATE_OFF) {
1738 * task could have been flipped by a concurrent
1739 * perf_event_context_sched_out()
1746 raw_spin_unlock_irq(&ctx->lock);
1749 static int perf_event_refresh(struct perf_event *event, int refresh)
1752 * not supported on inherited events
1754 if (event->attr.inherit || !is_sampling_event(event))
1757 atomic_add(refresh, &event->event_limit);
1758 perf_event_enable(event);
1763 static void ctx_sched_out(struct perf_event_context *ctx,
1764 struct perf_cpu_context *cpuctx,
1765 enum event_type_t event_type)
1767 struct perf_event *event;
1769 raw_spin_lock(&ctx->lock);
1770 perf_pmu_disable(ctx->pmu);
1772 if (likely(!ctx->nr_events))
1774 update_context_time(ctx);
1775 update_cgrp_time_from_cpuctx(cpuctx);
1777 if (!ctx->nr_active)
1780 if (event_type & EVENT_PINNED) {
1781 list_for_each_entry(event, &ctx->pinned_groups, group_entry)
1782 group_sched_out(event, cpuctx, ctx);
1785 if (event_type & EVENT_FLEXIBLE) {
1786 list_for_each_entry(event, &ctx->flexible_groups, group_entry)
1787 group_sched_out(event, cpuctx, ctx);
1790 perf_pmu_enable(ctx->pmu);
1791 raw_spin_unlock(&ctx->lock);
1795 * Test whether two contexts are equivalent, i.e. whether they
1796 * have both been cloned from the same version of the same context
1797 * and they both have the same number of enabled events.
1798 * If the number of enabled events is the same, then the set
1799 * of enabled events should be the same, because these are both
1800 * inherited contexts, therefore we can't access individual events
1801 * in them directly with an fd; we can only enable/disable all
1802 * events via prctl, or enable/disable all events in a family
1803 * via ioctl, which will have the same effect on both contexts.
1805 static int context_equiv(struct perf_event_context *ctx1,
1806 struct perf_event_context *ctx2)
1808 return ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx
1809 && ctx1->parent_gen == ctx2->parent_gen
1810 && !ctx1->pin_count && !ctx2->pin_count;
1813 static void __perf_event_sync_stat(struct perf_event *event,
1814 struct perf_event *next_event)
1818 if (!event->attr.inherit_stat)
1822 * Update the event value, we cannot use perf_event_read()
1823 * because we're in the middle of a context switch and have IRQs
1824 * disabled, which upsets smp_call_function_single(), however
1825 * we know the event must be on the current CPU, therefore we
1826 * don't need to use it.
1828 switch (event->state) {
1829 case PERF_EVENT_STATE_ACTIVE:
1830 event->pmu->read(event);
1833 case PERF_EVENT_STATE_INACTIVE:
1834 update_event_times(event);
1842 * In order to keep per-task stats reliable we need to flip the event
1843 * values when we flip the contexts.
1845 value = local64_read(&next_event->count);
1846 value = local64_xchg(&event->count, value);
1847 local64_set(&next_event->count, value);
1849 swap(event->total_time_enabled, next_event->total_time_enabled);
1850 swap(event->total_time_running, next_event->total_time_running);
1853 * Since we swizzled the values, update the user visible data too.
1855 perf_event_update_userpage(event);
1856 perf_event_update_userpage(next_event);
1859 #define list_next_entry(pos, member) \
1860 list_entry(pos->member.next, typeof(*pos), member)
1862 static void perf_event_sync_stat(struct perf_event_context *ctx,
1863 struct perf_event_context *next_ctx)
1865 struct perf_event *event, *next_event;
1870 update_context_time(ctx);
1872 event = list_first_entry(&ctx->event_list,
1873 struct perf_event, event_entry);
1875 next_event = list_first_entry(&next_ctx->event_list,
1876 struct perf_event, event_entry);
1878 while (&event->event_entry != &ctx->event_list &&
1879 &next_event->event_entry != &next_ctx->event_list) {
1881 __perf_event_sync_stat(event, next_event);
1883 event = list_next_entry(event, event_entry);
1884 next_event = list_next_entry(next_event, event_entry);
1888 static void perf_event_context_sched_out(struct task_struct *task, int ctxn,
1889 struct task_struct *next)
1891 struct perf_event_context *ctx = task->perf_event_ctxp[ctxn];
1892 struct perf_event_context *next_ctx;
1893 struct perf_event_context *parent;
1894 struct perf_cpu_context *cpuctx;
1900 cpuctx = __get_cpu_context(ctx);
1901 if (!cpuctx->task_ctx)
1905 parent = rcu_dereference(ctx->parent_ctx);
1906 next_ctx = next->perf_event_ctxp[ctxn];
1907 if (parent && next_ctx &&
1908 rcu_dereference(next_ctx->parent_ctx) == parent) {
1910 * Looks like the two contexts are clones, so we might be
1911 * able to optimize the context switch. We lock both
1912 * contexts and check that they are clones under the
1913 * lock (including re-checking that neither has been
1914 * uncloned in the meantime). It doesn't matter which
1915 * order we take the locks because no other cpu could
1916 * be trying to lock both of these tasks.
1918 raw_spin_lock(&ctx->lock);
1919 raw_spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING);
1920 if (context_equiv(ctx, next_ctx)) {
1922 * XXX do we need a memory barrier of sorts
1923 * wrt to rcu_dereference() of perf_event_ctxp
1925 task->perf_event_ctxp[ctxn] = next_ctx;
1926 next->perf_event_ctxp[ctxn] = ctx;
1928 next_ctx->task = task;
1931 perf_event_sync_stat(ctx, next_ctx);
1933 raw_spin_unlock(&next_ctx->lock);
1934 raw_spin_unlock(&ctx->lock);
1939 ctx_sched_out(ctx, cpuctx, EVENT_ALL);
1940 cpuctx->task_ctx = NULL;
1944 #define for_each_task_context_nr(ctxn) \
1945 for ((ctxn) = 0; (ctxn) < perf_nr_task_contexts; (ctxn)++)
1948 * Called from scheduler to remove the events of the current task,
1949 * with interrupts disabled.
1951 * We stop each event and update the event value in event->count.
1953 * This does not protect us against NMI, but disable()
1954 * sets the disabled bit in the control field of event _before_
1955 * accessing the event control register. If a NMI hits, then it will
1956 * not restart the event.
1958 void __perf_event_task_sched_out(struct task_struct *task,
1959 struct task_struct *next)
1963 for_each_task_context_nr(ctxn)
1964 perf_event_context_sched_out(task, ctxn, next);
1967 * if cgroup events exist on this CPU, then we need
1968 * to check if we have to switch out PMU state.
1969 * cgroup event are system-wide mode only
1971 if (atomic_read(&__get_cpu_var(perf_cgroup_events)))
1972 perf_cgroup_sched_out(task);
1975 static void task_ctx_sched_out(struct perf_event_context *ctx,
1976 enum event_type_t event_type)
1978 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1980 if (!cpuctx->task_ctx)
1983 if (WARN_ON_ONCE(ctx != cpuctx->task_ctx))
1986 ctx_sched_out(ctx, cpuctx, event_type);
1987 cpuctx->task_ctx = NULL;
1991 * Called with IRQs disabled
1993 static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
1994 enum event_type_t event_type)
1996 ctx_sched_out(&cpuctx->ctx, cpuctx, event_type);
2000 ctx_pinned_sched_in(struct perf_event_context *ctx,
2001 struct perf_cpu_context *cpuctx)
2003 struct perf_event *event;
2005 list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
2006 if (event->state <= PERF_EVENT_STATE_OFF)
2008 if (!event_filter_match(event))
2011 /* may need to reset tstamp_enabled */
2012 if (is_cgroup_event(event))
2013 perf_cgroup_mark_enabled(event, ctx);
2015 if (group_can_go_on(event, cpuctx, 1))
2016 group_sched_in(event, cpuctx, ctx);
2019 * If this pinned group hasn't been scheduled,
2020 * put it in error state.
2022 if (event->state == PERF_EVENT_STATE_INACTIVE) {
2023 update_group_times(event);
2024 event->state = PERF_EVENT_STATE_ERROR;
2030 ctx_flexible_sched_in(struct perf_event_context *ctx,
2031 struct perf_cpu_context *cpuctx)
2033 struct perf_event *event;
2036 list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
2037 /* Ignore events in OFF or ERROR state */
2038 if (event->state <= PERF_EVENT_STATE_OFF)
2041 * Listen to the 'cpu' scheduling filter constraint
2044 if (!event_filter_match(event))
2047 /* may need to reset tstamp_enabled */
2048 if (is_cgroup_event(event))
2049 perf_cgroup_mark_enabled(event, ctx);
2051 if (group_can_go_on(event, cpuctx, can_add_hw)) {
2052 if (group_sched_in(event, cpuctx, ctx))
2059 ctx_sched_in(struct perf_event_context *ctx,
2060 struct perf_cpu_context *cpuctx,
2061 enum event_type_t event_type,
2062 struct task_struct *task)
2066 raw_spin_lock(&ctx->lock);
2068 if (likely(!ctx->nr_events))
2072 ctx->timestamp = now;
2073 perf_cgroup_set_timestamp(task, ctx);
2075 * First go through the list and put on any pinned groups
2076 * in order to give them the best chance of going on.
2078 if (event_type & EVENT_PINNED)
2079 ctx_pinned_sched_in(ctx, cpuctx);
2081 /* Then walk through the lower prio flexible groups */
2082 if (event_type & EVENT_FLEXIBLE)
2083 ctx_flexible_sched_in(ctx, cpuctx);
2086 raw_spin_unlock(&ctx->lock);
2089 static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
2090 enum event_type_t event_type,
2091 struct task_struct *task)
2093 struct perf_event_context *ctx = &cpuctx->ctx;
2095 ctx_sched_in(ctx, cpuctx, event_type, task);
2098 static void task_ctx_sched_in(struct perf_event_context *ctx,
2099 enum event_type_t event_type)
2101 struct perf_cpu_context *cpuctx;
2103 cpuctx = __get_cpu_context(ctx);
2104 if (cpuctx->task_ctx == ctx)
2107 ctx_sched_in(ctx, cpuctx, event_type, NULL);
2108 cpuctx->task_ctx = ctx;
2111 static void perf_event_context_sched_in(struct perf_event_context *ctx,
2112 struct task_struct *task)
2114 struct perf_cpu_context *cpuctx;
2116 cpuctx = __get_cpu_context(ctx);
2117 if (cpuctx->task_ctx == ctx)
2120 perf_pmu_disable(ctx->pmu);
2122 * We want to keep the following priority order:
2123 * cpu pinned (that don't need to move), task pinned,
2124 * cpu flexible, task flexible.
2126 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
2128 ctx_sched_in(ctx, cpuctx, EVENT_PINNED, task);
2129 cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE, task);
2130 ctx_sched_in(ctx, cpuctx, EVENT_FLEXIBLE, task);
2132 cpuctx->task_ctx = ctx;
2135 * Since these rotations are per-cpu, we need to ensure the
2136 * cpu-context we got scheduled on is actually rotating.
2138 perf_pmu_rotate_start(ctx->pmu);
2139 perf_pmu_enable(ctx->pmu);
2143 * Called from scheduler to add the events of the current task
2144 * with interrupts disabled.
2146 * We restore the event value and then enable it.
2148 * This does not protect us against NMI, but enable()
2149 * sets the enabled bit in the control field of event _before_
2150 * accessing the event control register. If a NMI hits, then it will
2151 * keep the event running.
2153 void __perf_event_task_sched_in(struct task_struct *task)
2155 struct perf_event_context *ctx;
2158 for_each_task_context_nr(ctxn) {
2159 ctx = task->perf_event_ctxp[ctxn];
2163 perf_event_context_sched_in(ctx, task);
2166 * if cgroup events exist on this CPU, then we need
2167 * to check if we have to switch in PMU state.
2168 * cgroup event are system-wide mode only
2170 if (atomic_read(&__get_cpu_var(perf_cgroup_events)))
2171 perf_cgroup_sched_in(task);
2174 static u64 perf_calculate_period(struct perf_event *event, u64 nsec, u64 count)
2176 u64 frequency = event->attr.sample_freq;
2177 u64 sec = NSEC_PER_SEC;
2178 u64 divisor, dividend;
2180 int count_fls, nsec_fls, frequency_fls, sec_fls;
2182 count_fls = fls64(count);
2183 nsec_fls = fls64(nsec);
2184 frequency_fls = fls64(frequency);
2188 * We got @count in @nsec, with a target of sample_freq HZ
2189 * the target period becomes:
2192 * period = -------------------
2193 * @nsec * sample_freq
2198 * Reduce accuracy by one bit such that @a and @b converge
2199 * to a similar magnitude.
2201 #define REDUCE_FLS(a, b) \
2203 if (a##_fls > b##_fls) { \
2213 * Reduce accuracy until either term fits in a u64, then proceed with
2214 * the other, so that finally we can do a u64/u64 division.
2216 while (count_fls + sec_fls > 64 && nsec_fls + frequency_fls > 64) {
2217 REDUCE_FLS(nsec, frequency);
2218 REDUCE_FLS(sec, count);
2221 if (count_fls + sec_fls > 64) {
2222 divisor = nsec * frequency;
2224 while (count_fls + sec_fls > 64) {
2225 REDUCE_FLS(count, sec);
2229 dividend = count * sec;
2231 dividend = count * sec;
2233 while (nsec_fls + frequency_fls > 64) {
2234 REDUCE_FLS(nsec, frequency);
2238 divisor = nsec * frequency;
2244 return div64_u64(dividend, divisor);
2247 static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count)
2249 struct hw_perf_event *hwc = &event->hw;
2250 s64 period, sample_period;
2253 period = perf_calculate_period(event, nsec, count);
2255 delta = (s64)(period - hwc->sample_period);
2256 delta = (delta + 7) / 8; /* low pass filter */
2258 sample_period = hwc->sample_period + delta;
2263 hwc->sample_period = sample_period;
2265 if (local64_read(&hwc->period_left) > 8*sample_period) {
2266 event->pmu->stop(event, PERF_EF_UPDATE);
2267 local64_set(&hwc->period_left, 0);
2268 event->pmu->start(event, PERF_EF_RELOAD);
2272 static void perf_ctx_adjust_freq(struct perf_event_context *ctx, u64 period)
2274 struct perf_event *event;
2275 struct hw_perf_event *hwc;
2276 u64 interrupts, now;
2279 raw_spin_lock(&ctx->lock);
2280 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
2281 if (event->state != PERF_EVENT_STATE_ACTIVE)
2284 if (!event_filter_match(event))
2289 interrupts = hwc->interrupts;
2290 hwc->interrupts = 0;
2293 * unthrottle events on the tick
2295 if (interrupts == MAX_INTERRUPTS) {
2296 perf_log_throttle(event, 1);
2297 event->pmu->start(event, 0);
2300 if (!event->attr.freq || !event->attr.sample_freq)
2303 event->pmu->read(event);
2304 now = local64_read(&event->count);
2305 delta = now - hwc->freq_count_stamp;
2306 hwc->freq_count_stamp = now;
2309 perf_adjust_period(event, period, delta);
2311 raw_spin_unlock(&ctx->lock);
2315 * Round-robin a context's events:
2317 static void rotate_ctx(struct perf_event_context *ctx)
2319 raw_spin_lock(&ctx->lock);
2322 * Rotate the first entry last of non-pinned groups. Rotation might be
2323 * disabled by the inheritance code.
2325 if (!ctx->rotate_disable)
2326 list_rotate_left(&ctx->flexible_groups);
2328 raw_spin_unlock(&ctx->lock);
2332 * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
2333 * because they're strictly cpu affine and rotate_start is called with IRQs
2334 * disabled, while rotate_context is called from IRQ context.
2336 static void perf_rotate_context(struct perf_cpu_context *cpuctx)
2338 u64 interval = (u64)cpuctx->jiffies_interval * TICK_NSEC;
2339 struct perf_event_context *ctx = NULL;
2340 int rotate = 0, remove = 1;
2342 if (cpuctx->ctx.nr_events) {
2344 if (cpuctx->ctx.nr_events != cpuctx->ctx.nr_active)
2348 ctx = cpuctx->task_ctx;
2349 if (ctx && ctx->nr_events) {
2351 if (ctx->nr_events != ctx->nr_active)
2355 perf_pmu_disable(cpuctx->ctx.pmu);
2356 perf_ctx_adjust_freq(&cpuctx->ctx, interval);
2358 perf_ctx_adjust_freq(ctx, interval);
2363 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
2365 task_ctx_sched_out(ctx, EVENT_FLEXIBLE);
2367 rotate_ctx(&cpuctx->ctx);
2371 cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE, current);
2373 task_ctx_sched_in(ctx, EVENT_FLEXIBLE);
2377 list_del_init(&cpuctx->rotation_list);
2379 perf_pmu_enable(cpuctx->ctx.pmu);
2382 void perf_event_task_tick(void)
2384 struct list_head *head = &__get_cpu_var(rotation_list);
2385 struct perf_cpu_context *cpuctx, *tmp;
2387 WARN_ON(!irqs_disabled());
2389 list_for_each_entry_safe(cpuctx, tmp, head, rotation_list) {
2390 if (cpuctx->jiffies_interval == 1 ||
2391 !(jiffies % cpuctx->jiffies_interval))
2392 perf_rotate_context(cpuctx);
2396 static int event_enable_on_exec(struct perf_event *event,
2397 struct perf_event_context *ctx)
2399 if (!event->attr.enable_on_exec)
2402 event->attr.enable_on_exec = 0;
2403 if (event->state >= PERF_EVENT_STATE_INACTIVE)
2406 __perf_event_mark_enabled(event, ctx);
2412 * Enable all of a task's events that have been marked enable-on-exec.
2413 * This expects task == current.
2415 static void perf_event_enable_on_exec(struct perf_event_context *ctx)
2417 struct perf_event *event;
2418 unsigned long flags;
2422 local_irq_save(flags);
2423 if (!ctx || !ctx->nr_events)
2426 task_ctx_sched_out(ctx, EVENT_ALL);
2428 raw_spin_lock(&ctx->lock);
2430 list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
2431 ret = event_enable_on_exec(event, ctx);
2436 list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
2437 ret = event_enable_on_exec(event, ctx);
2443 * Unclone this context if we enabled any event.
2448 raw_spin_unlock(&ctx->lock);
2450 perf_event_context_sched_in(ctx, ctx->task);
2452 local_irq_restore(flags);
2456 * Cross CPU call to read the hardware event
2458 static void __perf_event_read(void *info)
2460 struct perf_event *event = info;
2461 struct perf_event_context *ctx = event->ctx;
2462 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
2465 * If this is a task context, we need to check whether it is
2466 * the current task context of this cpu. If not it has been
2467 * scheduled out before the smp call arrived. In that case
2468 * event->count would have been updated to a recent sample
2469 * when the event was scheduled out.
2471 if (ctx->task && cpuctx->task_ctx != ctx)
2474 raw_spin_lock(&ctx->lock);
2475 if (ctx->is_active) {
2476 update_context_time(ctx);
2477 update_cgrp_time_from_event(event);
2479 update_event_times(event);
2480 if (event->state == PERF_EVENT_STATE_ACTIVE)
2481 event->pmu->read(event);
2482 raw_spin_unlock(&ctx->lock);
2485 static inline u64 perf_event_count(struct perf_event *event)
2487 return local64_read(&event->count) + atomic64_read(&event->child_count);
2490 static u64 perf_event_read(struct perf_event *event)
2493 * If event is enabled and currently active on a CPU, update the
2494 * value in the event structure:
2496 if (event->state == PERF_EVENT_STATE_ACTIVE) {
2497 smp_call_function_single(event->oncpu,
2498 __perf_event_read, event, 1);
2499 } else if (event->state == PERF_EVENT_STATE_INACTIVE) {
2500 struct perf_event_context *ctx = event->ctx;
2501 unsigned long flags;
2503 raw_spin_lock_irqsave(&ctx->lock, flags);
2505 * may read while context is not active
2506 * (e.g., thread is blocked), in that case
2507 * we cannot update context time
2509 if (ctx->is_active) {
2510 update_context_time(ctx);
2511 update_cgrp_time_from_event(event);
2513 update_event_times(event);
2514 raw_spin_unlock_irqrestore(&ctx->lock, flags);
2517 return perf_event_count(event);
2524 struct callchain_cpus_entries {
2525 struct rcu_head rcu_head;
2526 struct perf_callchain_entry *cpu_entries[0];
2529 static DEFINE_PER_CPU(int, callchain_recursion[PERF_NR_CONTEXTS]);
2530 static atomic_t nr_callchain_events;
2531 static DEFINE_MUTEX(callchain_mutex);
2532 struct callchain_cpus_entries *callchain_cpus_entries;
2535 __weak void perf_callchain_kernel(struct perf_callchain_entry *entry,
2536 struct pt_regs *regs)
2540 __weak void perf_callchain_user(struct perf_callchain_entry *entry,
2541 struct pt_regs *regs)
2545 static void release_callchain_buffers_rcu(struct rcu_head *head)
2547 struct callchain_cpus_entries *entries;
2550 entries = container_of(head, struct callchain_cpus_entries, rcu_head);
2552 for_each_possible_cpu(cpu)
2553 kfree(entries->cpu_entries[cpu]);
2558 static void release_callchain_buffers(void)
2560 struct callchain_cpus_entries *entries;
2562 entries = callchain_cpus_entries;
2563 rcu_assign_pointer(callchain_cpus_entries, NULL);
2564 call_rcu(&entries->rcu_head, release_callchain_buffers_rcu);
2567 static int alloc_callchain_buffers(void)
2571 struct callchain_cpus_entries *entries;
2574 * We can't use the percpu allocation API for data that can be
2575 * accessed from NMI. Use a temporary manual per cpu allocation
2576 * until that gets sorted out.
2578 size = offsetof(struct callchain_cpus_entries, cpu_entries[nr_cpu_ids]);
2580 entries = kzalloc(size, GFP_KERNEL);
2584 size = sizeof(struct perf_callchain_entry) * PERF_NR_CONTEXTS;
2586 for_each_possible_cpu(cpu) {
2587 entries->cpu_entries[cpu] = kmalloc_node(size, GFP_KERNEL,
2589 if (!entries->cpu_entries[cpu])
2593 rcu_assign_pointer(callchain_cpus_entries, entries);
2598 for_each_possible_cpu(cpu)
2599 kfree(entries->cpu_entries[cpu]);
2605 static int get_callchain_buffers(void)
2610 mutex_lock(&callchain_mutex);
2612 count = atomic_inc_return(&nr_callchain_events);
2613 if (WARN_ON_ONCE(count < 1)) {
2619 /* If the allocation failed, give up */
2620 if (!callchain_cpus_entries)
2625 err = alloc_callchain_buffers();
2627 release_callchain_buffers();
2629 mutex_unlock(&callchain_mutex);
2634 static void put_callchain_buffers(void)
2636 if (atomic_dec_and_mutex_lock(&nr_callchain_events, &callchain_mutex)) {
2637 release_callchain_buffers();
2638 mutex_unlock(&callchain_mutex);
2642 static int get_recursion_context(int *recursion)
2650 else if (in_softirq())
2655 if (recursion[rctx])
2664 static inline void put_recursion_context(int *recursion, int rctx)
2670 static struct perf_callchain_entry *get_callchain_entry(int *rctx)
2673 struct callchain_cpus_entries *entries;
2675 *rctx = get_recursion_context(__get_cpu_var(callchain_recursion));
2679 entries = rcu_dereference(callchain_cpus_entries);
2683 cpu = smp_processor_id();
2685 return &entries->cpu_entries[cpu][*rctx];
2689 put_callchain_entry(int rctx)
2691 put_recursion_context(__get_cpu_var(callchain_recursion), rctx);
2694 static struct perf_callchain_entry *perf_callchain(struct pt_regs *regs)
2697 struct perf_callchain_entry *entry;
2700 entry = get_callchain_entry(&rctx);
2709 if (!user_mode(regs)) {
2710 perf_callchain_store(entry, PERF_CONTEXT_KERNEL);
2711 perf_callchain_kernel(entry, regs);
2713 regs = task_pt_regs(current);
2719 perf_callchain_store(entry, PERF_CONTEXT_USER);
2720 perf_callchain_user(entry, regs);
2724 put_callchain_entry(rctx);
2730 * Initialize the perf_event context in a task_struct:
2732 static void __perf_event_init_context(struct perf_event_context *ctx)
2734 raw_spin_lock_init(&ctx->lock);
2735 mutex_init(&ctx->mutex);
2736 INIT_LIST_HEAD(&ctx->pinned_groups);
2737 INIT_LIST_HEAD(&ctx->flexible_groups);
2738 INIT_LIST_HEAD(&ctx->event_list);
2739 atomic_set(&ctx->refcount, 1);
2742 static struct perf_event_context *
2743 alloc_perf_context(struct pmu *pmu, struct task_struct *task)
2745 struct perf_event_context *ctx;
2747 ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL);
2751 __perf_event_init_context(ctx);
2754 get_task_struct(task);
2761 static struct task_struct *
2762 find_lively_task_by_vpid(pid_t vpid)
2764 struct task_struct *task;
2771 task = find_task_by_vpid(vpid);
2773 get_task_struct(task);
2777 return ERR_PTR(-ESRCH);
2779 /* Reuse ptrace permission checks for now. */
2781 if (!ptrace_may_access(task, PTRACE_MODE_READ))
2786 put_task_struct(task);
2787 return ERR_PTR(err);
2792 * Returns a matching context with refcount and pincount.
2794 static struct perf_event_context *
2795 find_get_context(struct pmu *pmu, struct task_struct *task, int cpu)
2797 struct perf_event_context *ctx;
2798 struct perf_cpu_context *cpuctx;
2799 unsigned long flags;
2803 /* Must be root to operate on a CPU event: */
2804 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
2805 return ERR_PTR(-EACCES);
2808 * We could be clever and allow to attach a event to an
2809 * offline CPU and activate it when the CPU comes up, but
2812 if (!cpu_online(cpu))
2813 return ERR_PTR(-ENODEV);
2815 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
2824 ctxn = pmu->task_ctx_nr;
2829 ctx = perf_lock_task_context(task, ctxn, &flags);
2833 raw_spin_unlock_irqrestore(&ctx->lock, flags);
2837 ctx = alloc_perf_context(pmu, task);
2845 mutex_lock(&task->perf_event_mutex);
2847 * If it has already passed perf_event_exit_task().
2848 * we must see PF_EXITING, it takes this mutex too.
2850 if (task->flags & PF_EXITING)
2852 else if (task->perf_event_ctxp[ctxn])
2856 rcu_assign_pointer(task->perf_event_ctxp[ctxn], ctx);
2858 mutex_unlock(&task->perf_event_mutex);
2860 if (unlikely(err)) {
2861 put_task_struct(task);
2873 return ERR_PTR(err);
2876 static void perf_event_free_filter(struct perf_event *event);
2878 static void free_event_rcu(struct rcu_head *head)
2880 struct perf_event *event;
2882 event = container_of(head, struct perf_event, rcu_head);
2884 put_pid_ns(event->ns);
2885 perf_event_free_filter(event);
2889 static void perf_buffer_put(struct perf_buffer *buffer);
2891 static void free_event(struct perf_event *event)
2893 irq_work_sync(&event->pending);
2895 if (!event->parent) {
2896 if (event->attach_state & PERF_ATTACH_TASK)
2897 jump_label_dec(&perf_sched_events);
2898 if (event->attr.mmap || event->attr.mmap_data)
2899 atomic_dec(&nr_mmap_events);
2900 if (event->attr.comm)
2901 atomic_dec(&nr_comm_events);
2902 if (event->attr.task)
2903 atomic_dec(&nr_task_events);
2904 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN)
2905 put_callchain_buffers();
2908 if (event->buffer) {
2909 perf_buffer_put(event->buffer);
2910 event->buffer = NULL;
2913 if (is_cgroup_event(event))
2914 perf_detach_cgroup(event);
2917 event->destroy(event);
2920 put_ctx(event->ctx);
2922 call_rcu(&event->rcu_head, free_event_rcu);
2925 int perf_event_release_kernel(struct perf_event *event)
2927 struct perf_event_context *ctx = event->ctx;
2930 * Remove from the PMU, can't get re-enabled since we got
2931 * here because the last ref went.
2933 perf_event_disable(event);
2935 WARN_ON_ONCE(ctx->parent_ctx);
2937 * There are two ways this annotation is useful:
2939 * 1) there is a lock recursion from perf_event_exit_task
2940 * see the comment there.
2942 * 2) there is a lock-inversion with mmap_sem through
2943 * perf_event_read_group(), which takes faults while
2944 * holding ctx->mutex, however this is called after
2945 * the last filedesc died, so there is no possibility
2946 * to trigger the AB-BA case.
2948 mutex_lock_nested(&ctx->mutex, SINGLE_DEPTH_NESTING);
2949 raw_spin_lock_irq(&ctx->lock);
2950 perf_group_detach(event);
2951 list_del_event(event, ctx);
2952 raw_spin_unlock_irq(&ctx->lock);
2953 mutex_unlock(&ctx->mutex);
2959 EXPORT_SYMBOL_GPL(perf_event_release_kernel);
2962 * Called when the last reference to the file is gone.
2964 static int perf_release(struct inode *inode, struct file *file)
2966 struct perf_event *event = file->private_data;
2967 struct task_struct *owner;
2969 file->private_data = NULL;
2972 owner = ACCESS_ONCE(event->owner);
2974 * Matches the smp_wmb() in perf_event_exit_task(). If we observe
2975 * !owner it means the list deletion is complete and we can indeed
2976 * free this event, otherwise we need to serialize on
2977 * owner->perf_event_mutex.
2979 smp_read_barrier_depends();
2982 * Since delayed_put_task_struct() also drops the last
2983 * task reference we can safely take a new reference
2984 * while holding the rcu_read_lock().
2986 get_task_struct(owner);
2991 mutex_lock(&owner->perf_event_mutex);
2993 * We have to re-check the event->owner field, if it is cleared
2994 * we raced with perf_event_exit_task(), acquiring the mutex
2995 * ensured they're done, and we can proceed with freeing the
2999 list_del_init(&event->owner_entry);
3000 mutex_unlock(&owner->perf_event_mutex);
3001 put_task_struct(owner);
3004 return perf_event_release_kernel(event);
3007 u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running)
3009 struct perf_event *child;
3015 mutex_lock(&event->child_mutex);
3016 total += perf_event_read(event);
3017 *enabled += event->total_time_enabled +
3018 atomic64_read(&event->child_total_time_enabled);
3019 *running += event->total_time_running +
3020 atomic64_read(&event->child_total_time_running);
3022 list_for_each_entry(child, &event->child_list, child_list) {
3023 total += perf_event_read(child);
3024 *enabled += child->total_time_enabled;
3025 *running += child->total_time_running;
3027 mutex_unlock(&event->child_mutex);
3031 EXPORT_SYMBOL_GPL(perf_event_read_value);
3033 static int perf_event_read_group(struct perf_event *event,
3034 u64 read_format, char __user *buf)
3036 struct perf_event *leader = event->group_leader, *sub;
3037 int n = 0, size = 0, ret = -EFAULT;
3038 struct perf_event_context *ctx = leader->ctx;
3040 u64 count, enabled, running;
3042 mutex_lock(&ctx->mutex);
3043 count = perf_event_read_value(leader, &enabled, &running);
3045 values[n++] = 1 + leader->nr_siblings;
3046 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3047 values[n++] = enabled;
3048 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3049 values[n++] = running;
3050 values[n++] = count;
3051 if (read_format & PERF_FORMAT_ID)
3052 values[n++] = primary_event_id(leader);
3054 size = n * sizeof(u64);
3056 if (copy_to_user(buf, values, size))
3061 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
3064 values[n++] = perf_event_read_value(sub, &enabled, &running);
3065 if (read_format & PERF_FORMAT_ID)
3066 values[n++] = primary_event_id(sub);
3068 size = n * sizeof(u64);
3070 if (copy_to_user(buf + ret, values, size)) {
3078 mutex_unlock(&ctx->mutex);
3083 static int perf_event_read_one(struct perf_event *event,
3084 u64 read_format, char __user *buf)
3086 u64 enabled, running;
3090 values[n++] = perf_event_read_value(event, &enabled, &running);
3091 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3092 values[n++] = enabled;
3093 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3094 values[n++] = running;
3095 if (read_format & PERF_FORMAT_ID)
3096 values[n++] = primary_event_id(event);
3098 if (copy_to_user(buf, values, n * sizeof(u64)))
3101 return n * sizeof(u64);
3105 * Read the performance event - simple non blocking version for now
3108 perf_read_hw(struct perf_event *event, char __user *buf, size_t count)
3110 u64 read_format = event->attr.read_format;
3114 * Return end-of-file for a read on a event that is in
3115 * error state (i.e. because it was pinned but it couldn't be
3116 * scheduled on to the CPU at some point).
3118 if (event->state == PERF_EVENT_STATE_ERROR)
3121 if (count < event->read_size)
3124 WARN_ON_ONCE(event->ctx->parent_ctx);
3125 if (read_format & PERF_FORMAT_GROUP)
3126 ret = perf_event_read_group(event, read_format, buf);
3128 ret = perf_event_read_one(event, read_format, buf);
3134 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
3136 struct perf_event *event = file->private_data;
3138 return perf_read_hw(event, buf, count);
3141 static unsigned int perf_poll(struct file *file, poll_table *wait)
3143 struct perf_event *event = file->private_data;
3144 struct perf_buffer *buffer;
3145 unsigned int events = POLL_HUP;
3148 buffer = rcu_dereference(event->buffer);
3150 events = atomic_xchg(&buffer->poll, 0);
3153 poll_wait(file, &event->waitq, wait);
3158 static void perf_event_reset(struct perf_event *event)
3160 (void)perf_event_read(event);
3161 local64_set(&event->count, 0);
3162 perf_event_update_userpage(event);
3166 * Holding the top-level event's child_mutex means that any
3167 * descendant process that has inherited this event will block
3168 * in sync_child_event if it goes to exit, thus satisfying the
3169 * task existence requirements of perf_event_enable/disable.
3171 static void perf_event_for_each_child(struct perf_event *event,
3172 void (*func)(struct perf_event *))
3174 struct perf_event *child;
3176 WARN_ON_ONCE(event->ctx->parent_ctx);
3177 mutex_lock(&event->child_mutex);
3179 list_for_each_entry(child, &event->child_list, child_list)
3181 mutex_unlock(&event->child_mutex);
3184 static void perf_event_for_each(struct perf_event *event,
3185 void (*func)(struct perf_event *))
3187 struct perf_event_context *ctx = event->ctx;
3188 struct perf_event *sibling;
3190 WARN_ON_ONCE(ctx->parent_ctx);
3191 mutex_lock(&ctx->mutex);
3192 event = event->group_leader;
3194 perf_event_for_each_child(event, func);
3196 list_for_each_entry(sibling, &event->sibling_list, group_entry)
3197 perf_event_for_each_child(event, func);
3198 mutex_unlock(&ctx->mutex);
3201 static int perf_event_period(struct perf_event *event, u64 __user *arg)
3203 struct perf_event_context *ctx = event->ctx;
3207 if (!is_sampling_event(event))
3210 if (copy_from_user(&value, arg, sizeof(value)))
3216 raw_spin_lock_irq(&ctx->lock);
3217 if (event->attr.freq) {
3218 if (value > sysctl_perf_event_sample_rate) {
3223 event->attr.sample_freq = value;
3225 event->attr.sample_period = value;
3226 event->hw.sample_period = value;
3229 raw_spin_unlock_irq(&ctx->lock);
3234 static const struct file_operations perf_fops;
3236 static struct perf_event *perf_fget_light(int fd, int *fput_needed)
3240 file = fget_light(fd, fput_needed);
3242 return ERR_PTR(-EBADF);
3244 if (file->f_op != &perf_fops) {
3245 fput_light(file, *fput_needed);
3247 return ERR_PTR(-EBADF);
3250 return file->private_data;
3253 static int perf_event_set_output(struct perf_event *event,
3254 struct perf_event *output_event);
3255 static int perf_event_set_filter(struct perf_event *event, void __user *arg);
3257 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
3259 struct perf_event *event = file->private_data;
3260 void (*func)(struct perf_event *);
3264 case PERF_EVENT_IOC_ENABLE:
3265 func = perf_event_enable;
3267 case PERF_EVENT_IOC_DISABLE:
3268 func = perf_event_disable;
3270 case PERF_EVENT_IOC_RESET:
3271 func = perf_event_reset;
3274 case PERF_EVENT_IOC_REFRESH:
3275 return perf_event_refresh(event, arg);
3277 case PERF_EVENT_IOC_PERIOD:
3278 return perf_event_period(event, (u64 __user *)arg);
3280 case PERF_EVENT_IOC_SET_OUTPUT:
3282 struct perf_event *output_event = NULL;
3283 int fput_needed = 0;
3287 output_event = perf_fget_light(arg, &fput_needed);
3288 if (IS_ERR(output_event))
3289 return PTR_ERR(output_event);
3292 ret = perf_event_set_output(event, output_event);
3294 fput_light(output_event->filp, fput_needed);
3299 case PERF_EVENT_IOC_SET_FILTER:
3300 return perf_event_set_filter(event, (void __user *)arg);
3306 if (flags & PERF_IOC_FLAG_GROUP)
3307 perf_event_for_each(event, func);
3309 perf_event_for_each_child(event, func);
3314 int perf_event_task_enable(void)
3316 struct perf_event *event;
3318 mutex_lock(¤t->perf_event_mutex);
3319 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
3320 perf_event_for_each_child(event, perf_event_enable);
3321 mutex_unlock(¤t->perf_event_mutex);
3326 int perf_event_task_disable(void)
3328 struct perf_event *event;
3330 mutex_lock(¤t->perf_event_mutex);
3331 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
3332 perf_event_for_each_child(event, perf_event_disable);
3333 mutex_unlock(¤t->perf_event_mutex);
3338 #ifndef PERF_EVENT_INDEX_OFFSET
3339 # define PERF_EVENT_INDEX_OFFSET 0
3342 static int perf_event_index(struct perf_event *event)
3344 if (event->hw.state & PERF_HES_STOPPED)
3347 if (event->state != PERF_EVENT_STATE_ACTIVE)
3350 return event->hw.idx + 1 - PERF_EVENT_INDEX_OFFSET;
3354 * Callers need to ensure there can be no nesting of this function, otherwise
3355 * the seqlock logic goes bad. We can not serialize this because the arch
3356 * code calls this from NMI context.
3358 void perf_event_update_userpage(struct perf_event *event)
3360 struct perf_event_mmap_page *userpg;
3361 struct perf_buffer *buffer;
3364 buffer = rcu_dereference(event->buffer);
3368 userpg = buffer->user_page;
3371 * Disable preemption so as to not let the corresponding user-space
3372 * spin too long if we get preempted.
3377 userpg->index = perf_event_index(event);
3378 userpg->offset = perf_event_count(event);
3379 if (event->state == PERF_EVENT_STATE_ACTIVE)
3380 userpg->offset -= local64_read(&event->hw.prev_count);
3382 userpg->time_enabled = event->total_time_enabled +
3383 atomic64_read(&event->child_total_time_enabled);
3385 userpg->time_running = event->total_time_running +
3386 atomic64_read(&event->child_total_time_running);
3395 static unsigned long perf_data_size(struct perf_buffer *buffer);
3398 perf_buffer_init(struct perf_buffer *buffer, long watermark, int flags)
3400 long max_size = perf_data_size(buffer);
3403 buffer->watermark = min(max_size, watermark);
3405 if (!buffer->watermark)
3406 buffer->watermark = max_size / 2;
3408 if (flags & PERF_BUFFER_WRITABLE)
3409 buffer->writable = 1;
3411 atomic_set(&buffer->refcount, 1);
3414 #ifndef CONFIG_PERF_USE_VMALLOC
3417 * Back perf_mmap() with regular GFP_KERNEL-0 pages.
3420 static struct page *
3421 perf_mmap_to_page(struct perf_buffer *buffer, unsigned long pgoff)
3423 if (pgoff > buffer->nr_pages)
3427 return virt_to_page(buffer->user_page);
3429 return virt_to_page(buffer->data_pages[pgoff - 1]);
3432 static void *perf_mmap_alloc_page(int cpu)
3437 node = (cpu == -1) ? cpu : cpu_to_node(cpu);
3438 page = alloc_pages_node(node, GFP_KERNEL | __GFP_ZERO, 0);
3442 return page_address(page);
3445 static struct perf_buffer *
3446 perf_buffer_alloc(int nr_pages, long watermark, int cpu, int flags)
3448 struct perf_buffer *buffer;
3452 size = sizeof(struct perf_buffer);
3453 size += nr_pages * sizeof(void *);
3455 buffer = kzalloc(size, GFP_KERNEL);
3459 buffer->user_page = perf_mmap_alloc_page(cpu);
3460 if (!buffer->user_page)
3461 goto fail_user_page;
3463 for (i = 0; i < nr_pages; i++) {
3464 buffer->data_pages[i] = perf_mmap_alloc_page(cpu);
3465 if (!buffer->data_pages[i])
3466 goto fail_data_pages;
3469 buffer->nr_pages = nr_pages;
3471 perf_buffer_init(buffer, watermark, flags);
3476 for (i--; i >= 0; i--)
3477 free_page((unsigned long)buffer->data_pages[i]);
3479 free_page((unsigned long)buffer->user_page);
3488 static void perf_mmap_free_page(unsigned long addr)
3490 struct page *page = virt_to_page((void *)addr);
3492 page->mapping = NULL;
3496 static void perf_buffer_free(struct perf_buffer *buffer)
3500 perf_mmap_free_page((unsigned long)buffer->user_page);
3501 for (i = 0; i < buffer->nr_pages; i++)
3502 perf_mmap_free_page((unsigned long)buffer->data_pages[i]);
3506 static inline int page_order(struct perf_buffer *buffer)
3514 * Back perf_mmap() with vmalloc memory.
3516 * Required for architectures that have d-cache aliasing issues.
3519 static inline int page_order(struct perf_buffer *buffer)
3521 return buffer->page_order;
3524 static struct page *
3525 perf_mmap_to_page(struct perf_buffer *buffer, unsigned long pgoff)
3527 if (pgoff > (1UL << page_order(buffer)))
3530 return vmalloc_to_page((void *)buffer->user_page + pgoff * PAGE_SIZE);
3533 static void perf_mmap_unmark_page(void *addr)
3535 struct page *page = vmalloc_to_page(addr);
3537 page->mapping = NULL;
3540 static void perf_buffer_free_work(struct work_struct *work)
3542 struct perf_buffer *buffer;
3546 buffer = container_of(work, struct perf_buffer, work);
3547 nr = 1 << page_order(buffer);
3549 base = buffer->user_page;
3550 for (i = 0; i < nr + 1; i++)
3551 perf_mmap_unmark_page(base + (i * PAGE_SIZE));
3557 static void perf_buffer_free(struct perf_buffer *buffer)
3559 schedule_work(&buffer->work);
3562 static struct perf_buffer *
3563 perf_buffer_alloc(int nr_pages, long watermark, int cpu, int flags)
3565 struct perf_buffer *buffer;
3569 size = sizeof(struct perf_buffer);
3570 size += sizeof(void *);
3572 buffer = kzalloc(size, GFP_KERNEL);
3576 INIT_WORK(&buffer->work, perf_buffer_free_work);
3578 all_buf = vmalloc_user((nr_pages + 1) * PAGE_SIZE);
3582 buffer->user_page = all_buf;
3583 buffer->data_pages[0] = all_buf + PAGE_SIZE;
3584 buffer->page_order = ilog2(nr_pages);
3585 buffer->nr_pages = 1;
3587 perf_buffer_init(buffer, watermark, flags);
3600 static unsigned long perf_data_size(struct perf_buffer *buffer)
3602 return buffer->nr_pages << (PAGE_SHIFT + page_order(buffer));
3605 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
3607 struct perf_event *event = vma->vm_file->private_data;
3608 struct perf_buffer *buffer;
3609 int ret = VM_FAULT_SIGBUS;
3611 if (vmf->flags & FAULT_FLAG_MKWRITE) {
3612 if (vmf->pgoff == 0)
3618 buffer = rcu_dereference(event->buffer);
3622 if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE))
3625 vmf->page = perf_mmap_to_page(buffer, vmf->pgoff);
3629 get_page(vmf->page);
3630 vmf->page->mapping = vma->vm_file->f_mapping;
3631 vmf->page->index = vmf->pgoff;
3640 static void perf_buffer_free_rcu(struct rcu_head *rcu_head)
3642 struct perf_buffer *buffer;
3644 buffer = container_of(rcu_head, struct perf_buffer, rcu_head);
3645 perf_buffer_free(buffer);
3648 static struct perf_buffer *perf_buffer_get(struct perf_event *event)
3650 struct perf_buffer *buffer;
3653 buffer = rcu_dereference(event->buffer);
3655 if (!atomic_inc_not_zero(&buffer->refcount))
3663 static void perf_buffer_put(struct perf_buffer *buffer)
3665 if (!atomic_dec_and_test(&buffer->refcount))
3668 call_rcu(&buffer->rcu_head, perf_buffer_free_rcu);
3671 static void perf_mmap_open(struct vm_area_struct *vma)
3673 struct perf_event *event = vma->vm_file->private_data;
3675 atomic_inc(&event->mmap_count);
3678 static void perf_mmap_close(struct vm_area_struct *vma)
3680 struct perf_event *event = vma->vm_file->private_data;
3682 if (atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex)) {
3683 unsigned long size = perf_data_size(event->buffer);
3684 struct user_struct *user = event->mmap_user;
3685 struct perf_buffer *buffer = event->buffer;
3687 atomic_long_sub((size >> PAGE_SHIFT) + 1, &user->locked_vm);
3688 vma->vm_mm->locked_vm -= event->mmap_locked;
3689 rcu_assign_pointer(event->buffer, NULL);
3690 mutex_unlock(&event->mmap_mutex);
3692 perf_buffer_put(buffer);
3697 static const struct vm_operations_struct perf_mmap_vmops = {
3698 .open = perf_mmap_open,
3699 .close = perf_mmap_close,
3700 .fault = perf_mmap_fault,
3701 .page_mkwrite = perf_mmap_fault,
3704 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
3706 struct perf_event *event = file->private_data;
3707 unsigned long user_locked, user_lock_limit;
3708 struct user_struct *user = current_user();
3709 unsigned long locked, lock_limit;
3710 struct perf_buffer *buffer;
3711 unsigned long vma_size;
3712 unsigned long nr_pages;
3713 long user_extra, extra;
3714 int ret = 0, flags = 0;
3717 * Don't allow mmap() of inherited per-task counters. This would
3718 * create a performance issue due to all children writing to the
3721 if (event->cpu == -1 && event->attr.inherit)
3724 if (!(vma->vm_flags & VM_SHARED))
3727 vma_size = vma->vm_end - vma->vm_start;
3728 nr_pages = (vma_size / PAGE_SIZE) - 1;
3731 * If we have buffer pages ensure they're a power-of-two number, so we
3732 * can do bitmasks instead of modulo.
3734 if (nr_pages != 0 && !is_power_of_2(nr_pages))
3737 if (vma_size != PAGE_SIZE * (1 + nr_pages))
3740 if (vma->vm_pgoff != 0)
3743 WARN_ON_ONCE(event->ctx->parent_ctx);
3744 mutex_lock(&event->mmap_mutex);
3745 if (event->buffer) {
3746 if (event->buffer->nr_pages == nr_pages)
3747 atomic_inc(&event->buffer->refcount);
3753 user_extra = nr_pages + 1;
3754 user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10);
3757 * Increase the limit linearly with more CPUs:
3759 user_lock_limit *= num_online_cpus();
3761 user_locked = atomic_long_read(&user->locked_vm) + user_extra;
3764 if (user_locked > user_lock_limit)
3765 extra = user_locked - user_lock_limit;
3767 lock_limit = rlimit(RLIMIT_MEMLOCK);
3768 lock_limit >>= PAGE_SHIFT;
3769 locked = vma->vm_mm->locked_vm + extra;
3771 if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() &&
3772 !capable(CAP_IPC_LOCK)) {
3777 WARN_ON(event->buffer);
3779 if (vma->vm_flags & VM_WRITE)
3780 flags |= PERF_BUFFER_WRITABLE;
3782 buffer = perf_buffer_alloc(nr_pages, event->attr.wakeup_watermark,
3788 rcu_assign_pointer(event->buffer, buffer);
3790 atomic_long_add(user_extra, &user->locked_vm);
3791 event->mmap_locked = extra;
3792 event->mmap_user = get_current_user();
3793 vma->vm_mm->locked_vm += event->mmap_locked;
3797 atomic_inc(&event->mmap_count);
3798 mutex_unlock(&event->mmap_mutex);
3800 vma->vm_flags |= VM_RESERVED;
3801 vma->vm_ops = &perf_mmap_vmops;
3806 static int perf_fasync(int fd, struct file *filp, int on)
3808 struct inode *inode = filp->f_path.dentry->d_inode;
3809 struct perf_event *event = filp->private_data;
3812 mutex_lock(&inode->i_mutex);
3813 retval = fasync_helper(fd, filp, on, &event->fasync);
3814 mutex_unlock(&inode->i_mutex);
3822 static const struct file_operations perf_fops = {
3823 .llseek = no_llseek,
3824 .release = perf_release,
3827 .unlocked_ioctl = perf_ioctl,
3828 .compat_ioctl = perf_ioctl,
3830 .fasync = perf_fasync,
3836 * If there's data, ensure we set the poll() state and publish everything
3837 * to user-space before waking everybody up.
3840 void perf_event_wakeup(struct perf_event *event)
3842 wake_up_all(&event->waitq);
3844 if (event->pending_kill) {
3845 kill_fasync(&event->fasync, SIGIO, event->pending_kill);
3846 event->pending_kill = 0;
3850 static void perf_pending_event(struct irq_work *entry)
3852 struct perf_event *event = container_of(entry,
3853 struct perf_event, pending);
3855 if (event->pending_disable) {
3856 event->pending_disable = 0;
3857 __perf_event_disable(event);
3860 if (event->pending_wakeup) {
3861 event->pending_wakeup = 0;
3862 perf_event_wakeup(event);
3867 * We assume there is only KVM supporting the callbacks.
3868 * Later on, we might change it to a list if there is
3869 * another virtualization implementation supporting the callbacks.
3871 struct perf_guest_info_callbacks *perf_guest_cbs;
3873 int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
3875 perf_guest_cbs = cbs;
3878 EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks);
3880 int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
3882 perf_guest_cbs = NULL;
3885 EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks);
3890 static bool perf_output_space(struct perf_buffer *buffer, unsigned long tail,
3891 unsigned long offset, unsigned long head)
3895 if (!buffer->writable)
3898 mask = perf_data_size(buffer) - 1;
3900 offset = (offset - tail) & mask;
3901 head = (head - tail) & mask;
3903 if ((int)(head - offset) < 0)
3909 static void perf_output_wakeup(struct perf_output_handle *handle)
3911 atomic_set(&handle->buffer->poll, POLL_IN);
3914 handle->event->pending_wakeup = 1;
3915 irq_work_queue(&handle->event->pending);
3917 perf_event_wakeup(handle->event);
3921 * We need to ensure a later event_id doesn't publish a head when a former
3922 * event isn't done writing. However since we need to deal with NMIs we
3923 * cannot fully serialize things.
3925 * We only publish the head (and generate a wakeup) when the outer-most
3928 static void perf_output_get_handle(struct perf_output_handle *handle)
3930 struct perf_buffer *buffer = handle->buffer;
3933 local_inc(&buffer->nest);
3934 handle->wakeup = local_read(&buffer->wakeup);
3937 static void perf_output_put_handle(struct perf_output_handle *handle)
3939 struct perf_buffer *buffer = handle->buffer;
3943 head = local_read(&buffer->head);
3946 * IRQ/NMI can happen here, which means we can miss a head update.
3949 if (!local_dec_and_test(&buffer->nest))
3953 * Publish the known good head. Rely on the full barrier implied
3954 * by atomic_dec_and_test() order the buffer->head read and this
3957 buffer->user_page->data_head = head;
3960 * Now check if we missed an update, rely on the (compiler)
3961 * barrier in atomic_dec_and_test() to re-read buffer->head.
3963 if (unlikely(head != local_read(&buffer->head))) {
3964 local_inc(&buffer->nest);
3968 if (handle->wakeup != local_read(&buffer->wakeup))
3969 perf_output_wakeup(handle);
3975 __always_inline void perf_output_copy(struct perf_output_handle *handle,
3976 const void *buf, unsigned int len)
3979 unsigned long size = min_t(unsigned long, handle->size, len);
3981 memcpy(handle->addr, buf, size);
3984 handle->addr += size;
3986 handle->size -= size;
3987 if (!handle->size) {
3988 struct perf_buffer *buffer = handle->buffer;
3991 handle->page &= buffer->nr_pages - 1;
3992 handle->addr = buffer->data_pages[handle->page];
3993 handle->size = PAGE_SIZE << page_order(buffer);
3998 static void __perf_event_header__init_id(struct perf_event_header *header,
3999 struct perf_sample_data *data,
4000 struct perf_event *event)
4002 u64 sample_type = event->attr.sample_type;
4004 data->type = sample_type;
4005 header->size += event->id_header_size;
4007 if (sample_type & PERF_SAMPLE_TID) {
4008 /* namespace issues */
4009 data->tid_entry.pid = perf_event_pid(event, current);
4010 data->tid_entry.tid = perf_event_tid(event, current);
4013 if (sample_type & PERF_SAMPLE_TIME)
4014 data->time = perf_clock();
4016 if (sample_type & PERF_SAMPLE_ID)
4017 data->id = primary_event_id(event);
4019 if (sample_type & PERF_SAMPLE_STREAM_ID)
4020 data->stream_id = event->id;
4022 if (sample_type & PERF_SAMPLE_CPU) {
4023 data->cpu_entry.cpu = raw_smp_processor_id();
4024 data->cpu_entry.reserved = 0;
4028 static void perf_event_header__init_id(struct perf_event_header *header,
4029 struct perf_sample_data *data,
4030 struct perf_event *event)
4032 if (event->attr.sample_id_all)
4033 __perf_event_header__init_id(header, data, event);
4036 static void __perf_event__output_id_sample(struct perf_output_handle *handle,
4037 struct perf_sample_data *data)
4039 u64 sample_type = data->type;
4041 if (sample_type & PERF_SAMPLE_TID)
4042 perf_output_put(handle, data->tid_entry);
4044 if (sample_type & PERF_SAMPLE_TIME)
4045 perf_output_put(handle, data->time);
4047 if (sample_type & PERF_SAMPLE_ID)
4048 perf_output_put(handle, data->id);
4050 if (sample_type & PERF_SAMPLE_STREAM_ID)
4051 perf_output_put(handle, data->stream_id);
4053 if (sample_type & PERF_SAMPLE_CPU)
4054 perf_output_put(handle, data->cpu_entry);
4057 static void perf_event__output_id_sample(struct perf_event *event,
4058 struct perf_output_handle *handle,
4059 struct perf_sample_data *sample)
4061 if (event->attr.sample_id_all)
4062 __perf_event__output_id_sample(handle, sample);
4065 int perf_output_begin(struct perf_output_handle *handle,
4066 struct perf_event *event, unsigned int size,
4067 int nmi, int sample)
4069 struct perf_buffer *buffer;
4070 unsigned long tail, offset, head;
4072 struct perf_sample_data sample_data;
4074 struct perf_event_header header;
4081 * For inherited events we send all the output towards the parent.
4084 event = event->parent;
4086 buffer = rcu_dereference(event->buffer);
4090 handle->buffer = buffer;
4091 handle->event = event;
4093 handle->sample = sample;
4095 if (!buffer->nr_pages)
4098 have_lost = local_read(&buffer->lost);
4100 lost_event.header.size = sizeof(lost_event);
4101 perf_event_header__init_id(&lost_event.header, &sample_data,
4103 size += lost_event.header.size;
4106 perf_output_get_handle(handle);
4110 * Userspace could choose to issue a mb() before updating the
4111 * tail pointer. So that all reads will be completed before the
4114 tail = ACCESS_ONCE(buffer->user_page->data_tail);
4116 offset = head = local_read(&buffer->head);
4118 if (unlikely(!perf_output_space(buffer, tail, offset, head)))
4120 } while (local_cmpxchg(&buffer->head, offset, head) != offset);
4122 if (head - local_read(&buffer->wakeup) > buffer->watermark)
4123 local_add(buffer->watermark, &buffer->wakeup);
4125 handle->page = offset >> (PAGE_SHIFT + page_order(buffer));
4126 handle->page &= buffer->nr_pages - 1;
4127 handle->size = offset & ((PAGE_SIZE << page_order(buffer)) - 1);
4128 handle->addr = buffer->data_pages[handle->page];
4129 handle->addr += handle->size;
4130 handle->size = (PAGE_SIZE << page_order(buffer)) - handle->size;
4133 lost_event.header.type = PERF_RECORD_LOST;
4134 lost_event.header.misc = 0;
4135 lost_event.id = event->id;
4136 lost_event.lost = local_xchg(&buffer->lost, 0);
4138 perf_output_put(handle, lost_event);
4139 perf_event__output_id_sample(event, handle, &sample_data);
4145 local_inc(&buffer->lost);
4146 perf_output_put_handle(handle);
4153 void perf_output_end(struct perf_output_handle *handle)
4155 struct perf_event *event = handle->event;
4156 struct perf_buffer *buffer = handle->buffer;
4158 int wakeup_events = event->attr.wakeup_events;
4160 if (handle->sample && wakeup_events) {
4161 int events = local_inc_return(&buffer->events);
4162 if (events >= wakeup_events) {
4163 local_sub(wakeup_events, &buffer->events);
4164 local_inc(&buffer->wakeup);
4168 perf_output_put_handle(handle);
4172 static void perf_output_read_one(struct perf_output_handle *handle,
4173 struct perf_event *event,
4174 u64 enabled, u64 running)
4176 u64 read_format = event->attr.read_format;
4180 values[n++] = perf_event_count(event);
4181 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
4182 values[n++] = enabled +
4183 atomic64_read(&event->child_total_time_enabled);
4185 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
4186 values[n++] = running +
4187 atomic64_read(&event->child_total_time_running);
4189 if (read_format & PERF_FORMAT_ID)
4190 values[n++] = primary_event_id(event);
4192 perf_output_copy(handle, values, n * sizeof(u64));
4196 * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
4198 static void perf_output_read_group(struct perf_output_handle *handle,
4199 struct perf_event *event,
4200 u64 enabled, u64 running)
4202 struct perf_event *leader = event->group_leader, *sub;
4203 u64 read_format = event->attr.read_format;
4207 values[n++] = 1 + leader->nr_siblings;
4209 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
4210 values[n++] = enabled;
4212 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
4213 values[n++] = running;
4215 if (leader != event)
4216 leader->pmu->read(leader);
4218 values[n++] = perf_event_count(leader);
4219 if (read_format & PERF_FORMAT_ID)
4220 values[n++] = primary_event_id(leader);
4222 perf_output_copy(handle, values, n * sizeof(u64));
4224 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
4228 sub->pmu->read(sub);
4230 values[n++] = perf_event_count(sub);
4231 if (read_format & PERF_FORMAT_ID)
4232 values[n++] = primary_event_id(sub);
4234 perf_output_copy(handle, values, n * sizeof(u64));
4238 #define PERF_FORMAT_TOTAL_TIMES (PERF_FORMAT_TOTAL_TIME_ENABLED|\
4239 PERF_FORMAT_TOTAL_TIME_RUNNING)
4241 static void perf_output_read(struct perf_output_handle *handle,
4242 struct perf_event *event)
4244 u64 enabled = 0, running = 0, now, ctx_time;
4245 u64 read_format = event->attr.read_format;
4248 * compute total_time_enabled, total_time_running
4249 * based on snapshot values taken when the event
4250 * was last scheduled in.
4252 * we cannot simply called update_context_time()
4253 * because of locking issue as we are called in
4256 if (read_format & PERF_FORMAT_TOTAL_TIMES) {
4258 ctx_time = event->shadow_ctx_time + now;
4259 enabled = ctx_time - event->tstamp_enabled;
4260 running = ctx_time - event->tstamp_running;
4263 if (event->attr.read_format & PERF_FORMAT_GROUP)
4264 perf_output_read_group(handle, event, enabled, running);
4266 perf_output_read_one(handle, event, enabled, running);
4269 void perf_output_sample(struct perf_output_handle *handle,
4270 struct perf_event_header *header,
4271 struct perf_sample_data *data,
4272 struct perf_event *event)
4274 u64 sample_type = data->type;
4276 perf_output_put(handle, *header);
4278 if (sample_type & PERF_SAMPLE_IP)
4279 perf_output_put(handle, data->ip);
4281 if (sample_type & PERF_SAMPLE_TID)
4282 perf_output_put(handle, data->tid_entry);
4284 if (sample_type & PERF_SAMPLE_TIME)
4285 perf_output_put(handle, data->time);
4287 if (sample_type & PERF_SAMPLE_ADDR)
4288 perf_output_put(handle, data->addr);
4290 if (sample_type & PERF_SAMPLE_ID)
4291 perf_output_put(handle, data->id);
4293 if (sample_type & PERF_SAMPLE_STREAM_ID)
4294 perf_output_put(handle, data->stream_id);
4296 if (sample_type & PERF_SAMPLE_CPU)
4297 perf_output_put(handle, data->cpu_entry);
4299 if (sample_type & PERF_SAMPLE_PERIOD)
4300 perf_output_put(handle, data->period);
4302 if (sample_type & PERF_SAMPLE_READ)
4303 perf_output_read(handle, event);
4305 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
4306 if (data->callchain) {
4309 if (data->callchain)
4310 size += data->callchain->nr;
4312 size *= sizeof(u64);
4314 perf_output_copy(handle, data->callchain, size);
4317 perf_output_put(handle, nr);
4321 if (sample_type & PERF_SAMPLE_RAW) {
4323 perf_output_put(handle, data->raw->size);
4324 perf_output_copy(handle, data->raw->data,
4331 .size = sizeof(u32),
4334 perf_output_put(handle, raw);
4339 void perf_prepare_sample(struct perf_event_header *header,
4340 struct perf_sample_data *data,
4341 struct perf_event *event,
4342 struct pt_regs *regs)
4344 u64 sample_type = event->attr.sample_type;
4346 header->type = PERF_RECORD_SAMPLE;
4347 header->size = sizeof(*header) + event->header_size;
4350 header->misc |= perf_misc_flags(regs);
4352 __perf_event_header__init_id(header, data, event);
4354 if (sample_type & PERF_SAMPLE_IP)
4355 data->ip = perf_instruction_pointer(regs);
4357 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
4360 data->callchain = perf_callchain(regs);
4362 if (data->callchain)
4363 size += data->callchain->nr;
4365 header->size += size * sizeof(u64);
4368 if (sample_type & PERF_SAMPLE_RAW) {
4369 int size = sizeof(u32);
4372 size += data->raw->size;
4374 size += sizeof(u32);
4376 WARN_ON_ONCE(size & (sizeof(u64)-1));
4377 header->size += size;
4381 static void perf_event_output(struct perf_event *event, int nmi,
4382 struct perf_sample_data *data,
4383 struct pt_regs *regs)
4385 struct perf_output_handle handle;
4386 struct perf_event_header header;
4388 /* protect the callchain buffers */
4391 perf_prepare_sample(&header, data, event, regs);
4393 if (perf_output_begin(&handle, event, header.size, nmi, 1))
4396 perf_output_sample(&handle, &header, data, event);
4398 perf_output_end(&handle);
4408 struct perf_read_event {
4409 struct perf_event_header header;
4416 perf_event_read_event(struct perf_event *event,
4417 struct task_struct *task)
4419 struct perf_output_handle handle;
4420 struct perf_sample_data sample;
4421 struct perf_read_event read_event = {
4423 .type = PERF_RECORD_READ,
4425 .size = sizeof(read_event) + event->read_size,
4427 .pid = perf_event_pid(event, task),
4428 .tid = perf_event_tid(event, task),
4432 perf_event_header__init_id(&read_event.header, &sample, event);
4433 ret = perf_output_begin(&handle, event, read_event.header.size, 0, 0);
4437 perf_output_put(&handle, read_event);
4438 perf_output_read(&handle, event);
4439 perf_event__output_id_sample(event, &handle, &sample);
4441 perf_output_end(&handle);
4445 * task tracking -- fork/exit
4447 * enabled by: attr.comm | attr.mmap | attr.mmap_data | attr.task
4450 struct perf_task_event {
4451 struct task_struct *task;
4452 struct perf_event_context *task_ctx;
4455 struct perf_event_header header;
4465 static void perf_event_task_output(struct perf_event *event,
4466 struct perf_task_event *task_event)
4468 struct perf_output_handle handle;
4469 struct perf_sample_data sample;
4470 struct task_struct *task = task_event->task;
4471 int ret, size = task_event->event_id.header.size;
4473 perf_event_header__init_id(&task_event->event_id.header, &sample, event);
4475 ret = perf_output_begin(&handle, event,
4476 task_event->event_id.header.size, 0, 0);
4480 task_event->event_id.pid = perf_event_pid(event, task);
4481 task_event->event_id.ppid = perf_event_pid(event, current);
4483 task_event->event_id.tid = perf_event_tid(event, task);
4484 task_event->event_id.ptid = perf_event_tid(event, current);
4486 perf_output_put(&handle, task_event->event_id);
4488 perf_event__output_id_sample(event, &handle, &sample);
4490 perf_output_end(&handle);
4492 task_event->event_id.header.size = size;
4495 static int perf_event_task_match(struct perf_event *event)
4497 if (event->state < PERF_EVENT_STATE_INACTIVE)
4500 if (!event_filter_match(event))
4503 if (event->attr.comm || event->attr.mmap ||
4504 event->attr.mmap_data || event->attr.task)
4510 static void perf_event_task_ctx(struct perf_event_context *ctx,
4511 struct perf_task_event *task_event)
4513 struct perf_event *event;
4515 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
4516 if (perf_event_task_match(event))
4517 perf_event_task_output(event, task_event);
4521 static void perf_event_task_event(struct perf_task_event *task_event)
4523 struct perf_cpu_context *cpuctx;
4524 struct perf_event_context *ctx;
4529 list_for_each_entry_rcu(pmu, &pmus, entry) {
4530 cpuctx = get_cpu_ptr(pmu->pmu_cpu_context);
4531 if (cpuctx->active_pmu != pmu)
4533 perf_event_task_ctx(&cpuctx->ctx, task_event);
4535 ctx = task_event->task_ctx;
4537 ctxn = pmu->task_ctx_nr;
4540 ctx = rcu_dereference(current->perf_event_ctxp[ctxn]);
4543 perf_event_task_ctx(ctx, task_event);
4545 put_cpu_ptr(pmu->pmu_cpu_context);
4550 static void perf_event_task(struct task_struct *task,
4551 struct perf_event_context *task_ctx,
4554 struct perf_task_event task_event;
4556 if (!atomic_read(&nr_comm_events) &&
4557 !atomic_read(&nr_mmap_events) &&
4558 !atomic_read(&nr_task_events))
4561 task_event = (struct perf_task_event){
4563 .task_ctx = task_ctx,
4566 .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT,
4568 .size = sizeof(task_event.event_id),
4574 .time = perf_clock(),
4578 perf_event_task_event(&task_event);
4581 void perf_event_fork(struct task_struct *task)
4583 perf_event_task(task, NULL, 1);
4590 struct perf_comm_event {
4591 struct task_struct *task;
4596 struct perf_event_header header;
4603 static void perf_event_comm_output(struct perf_event *event,
4604 struct perf_comm_event *comm_event)
4606 struct perf_output_handle handle;
4607 struct perf_sample_data sample;
4608 int size = comm_event->event_id.header.size;
4611 perf_event_header__init_id(&comm_event->event_id.header, &sample, event);
4612 ret = perf_output_begin(&handle, event,
4613 comm_event->event_id.header.size, 0, 0);
4618 comm_event->event_id.pid = perf_event_pid(event, comm_event->task);
4619 comm_event->event_id.tid = perf_event_tid(event, comm_event->task);
4621 perf_output_put(&handle, comm_event->event_id);
4622 perf_output_copy(&handle, comm_event->comm,
4623 comm_event->comm_size);
4625 perf_event__output_id_sample(event, &handle, &sample);
4627 perf_output_end(&handle);
4629 comm_event->event_id.header.size = size;
4632 static int perf_event_comm_match(struct perf_event *event)
4634 if (event->state < PERF_EVENT_STATE_INACTIVE)
4637 if (!event_filter_match(event))
4640 if (event->attr.comm)
4646 static void perf_event_comm_ctx(struct perf_event_context *ctx,
4647 struct perf_comm_event *comm_event)
4649 struct perf_event *event;
4651 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
4652 if (perf_event_comm_match(event))
4653 perf_event_comm_output(event, comm_event);
4657 static void perf_event_comm_event(struct perf_comm_event *comm_event)
4659 struct perf_cpu_context *cpuctx;
4660 struct perf_event_context *ctx;
4661 char comm[TASK_COMM_LEN];
4666 memset(comm, 0, sizeof(comm));
4667 strlcpy(comm, comm_event->task->comm, sizeof(comm));
4668 size = ALIGN(strlen(comm)+1, sizeof(u64));
4670 comm_event->comm = comm;
4671 comm_event->comm_size = size;
4673 comm_event->event_id.header.size = sizeof(comm_event->event_id) + size;
4675 list_for_each_entry_rcu(pmu, &pmus, entry) {
4676 cpuctx = get_cpu_ptr(pmu->pmu_cpu_context);
4677 if (cpuctx->active_pmu != pmu)
4679 perf_event_comm_ctx(&cpuctx->ctx, comm_event);
4681 ctxn = pmu->task_ctx_nr;
4685 ctx = rcu_dereference(current->perf_event_ctxp[ctxn]);
4687 perf_event_comm_ctx(ctx, comm_event);
4689 put_cpu_ptr(pmu->pmu_cpu_context);
4694 void perf_event_comm(struct task_struct *task)
4696 struct perf_comm_event comm_event;
4697 struct perf_event_context *ctx;
4700 for_each_task_context_nr(ctxn) {
4701 ctx = task->perf_event_ctxp[ctxn];
4705 perf_event_enable_on_exec(ctx);
4708 if (!atomic_read(&nr_comm_events))
4711 comm_event = (struct perf_comm_event){
4717 .type = PERF_RECORD_COMM,
4726 perf_event_comm_event(&comm_event);
4733 struct perf_mmap_event {
4734 struct vm_area_struct *vma;
4736 const char *file_name;
4740 struct perf_event_header header;
4750 static void perf_event_mmap_output(struct perf_event *event,
4751 struct perf_mmap_event *mmap_event)
4753 struct perf_output_handle handle;
4754 struct perf_sample_data sample;
4755 int size = mmap_event->event_id.header.size;
4758 perf_event_header__init_id(&mmap_event->event_id.header, &sample, event);
4759 ret = perf_output_begin(&handle, event,
4760 mmap_event->event_id.header.size, 0, 0);
4764 mmap_event->event_id.pid = perf_event_pid(event, current);
4765 mmap_event->event_id.tid = perf_event_tid(event, current);
4767 perf_output_put(&handle, mmap_event->event_id);
4768 perf_output_copy(&handle, mmap_event->file_name,
4769 mmap_event->file_size);
4771 perf_event__output_id_sample(event, &handle, &sample);
4773 perf_output_end(&handle);
4775 mmap_event->event_id.header.size = size;
4778 static int perf_event_mmap_match(struct perf_event *event,
4779 struct perf_mmap_event *mmap_event,
4782 if (event->state < PERF_EVENT_STATE_INACTIVE)
4785 if (!event_filter_match(event))
4788 if ((!executable && event->attr.mmap_data) ||
4789 (executable && event->attr.mmap))
4795 static void perf_event_mmap_ctx(struct perf_event_context *ctx,
4796 struct perf_mmap_event *mmap_event,
4799 struct perf_event *event;
4801 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
4802 if (perf_event_mmap_match(event, mmap_event, executable))
4803 perf_event_mmap_output(event, mmap_event);
4807 static void perf_event_mmap_event(struct perf_mmap_event *mmap_event)
4809 struct perf_cpu_context *cpuctx;
4810 struct perf_event_context *ctx;
4811 struct vm_area_struct *vma = mmap_event->vma;
4812 struct file *file = vma->vm_file;
4820 memset(tmp, 0, sizeof(tmp));
4824 * d_path works from the end of the buffer backwards, so we
4825 * need to add enough zero bytes after the string to handle
4826 * the 64bit alignment we do later.
4828 buf = kzalloc(PATH_MAX + sizeof(u64), GFP_KERNEL);
4830 name = strncpy(tmp, "//enomem", sizeof(tmp));
4833 name = d_path(&file->f_path, buf, PATH_MAX);
4835 name = strncpy(tmp, "//toolong", sizeof(tmp));
4839 if (arch_vma_name(mmap_event->vma)) {
4840 name = strncpy(tmp, arch_vma_name(mmap_event->vma),
4846 name = strncpy(tmp, "[vdso]", sizeof(tmp));
4848 } else if (vma->vm_start <= vma->vm_mm->start_brk &&
4849 vma->vm_end >= vma->vm_mm->brk) {
4850 name = strncpy(tmp, "[heap]", sizeof(tmp));
4852 } else if (vma->vm_start <= vma->vm_mm->start_stack &&
4853 vma->vm_end >= vma->vm_mm->start_stack) {
4854 name = strncpy(tmp, "[stack]", sizeof(tmp));
4858 name = strncpy(tmp, "//anon", sizeof(tmp));
4863 size = ALIGN(strlen(name)+1, sizeof(u64));
4865 mmap_event->file_name = name;
4866 mmap_event->file_size = size;
4868 mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size;
4871 list_for_each_entry_rcu(pmu, &pmus, entry) {
4872 cpuctx = get_cpu_ptr(pmu->pmu_cpu_context);
4873 if (cpuctx->active_pmu != pmu)
4875 perf_event_mmap_ctx(&cpuctx->ctx, mmap_event,
4876 vma->vm_flags & VM_EXEC);
4878 ctxn = pmu->task_ctx_nr;
4882 ctx = rcu_dereference(current->perf_event_ctxp[ctxn]);
4884 perf_event_mmap_ctx(ctx, mmap_event,
4885 vma->vm_flags & VM_EXEC);
4888 put_cpu_ptr(pmu->pmu_cpu_context);
4895 void perf_event_mmap(struct vm_area_struct *vma)
4897 struct perf_mmap_event mmap_event;
4899 if (!atomic_read(&nr_mmap_events))
4902 mmap_event = (struct perf_mmap_event){
4908 .type = PERF_RECORD_MMAP,
4909 .misc = PERF_RECORD_MISC_USER,
4914 .start = vma->vm_start,
4915 .len = vma->vm_end - vma->vm_start,
4916 .pgoff = (u64)vma->vm_pgoff << PAGE_SHIFT,
4920 perf_event_mmap_event(&mmap_event);
4924 * IRQ throttle logging
4927 static void perf_log_throttle(struct perf_event *event, int enable)
4929 struct perf_output_handle handle;
4930 struct perf_sample_data sample;
4934 struct perf_event_header header;
4938 } throttle_event = {
4940 .type = PERF_RECORD_THROTTLE,
4942 .size = sizeof(throttle_event),
4944 .time = perf_clock(),
4945 .id = primary_event_id(event),
4946 .stream_id = event->id,
4950 throttle_event.header.type = PERF_RECORD_UNTHROTTLE;
4952 perf_event_header__init_id(&throttle_event.header, &sample, event);
4954 ret = perf_output_begin(&handle, event,
4955 throttle_event.header.size, 1, 0);
4959 perf_output_put(&handle, throttle_event);
4960 perf_event__output_id_sample(event, &handle, &sample);
4961 perf_output_end(&handle);
4965 * Generic event overflow handling, sampling.
4968 static int __perf_event_overflow(struct perf_event *event, int nmi,
4969 int throttle, struct perf_sample_data *data,
4970 struct pt_regs *regs)
4972 int events = atomic_read(&event->event_limit);
4973 struct hw_perf_event *hwc = &event->hw;
4977 * Non-sampling counters might still use the PMI to fold short
4978 * hardware counters, ignore those.
4980 if (unlikely(!is_sampling_event(event)))
4983 if (unlikely(hwc->interrupts >= max_samples_per_tick)) {
4985 hwc->interrupts = MAX_INTERRUPTS;
4986 perf_log_throttle(event, 0);
4992 if (event->attr.freq) {
4993 u64 now = perf_clock();
4994 s64 delta = now - hwc->freq_time_stamp;
4996 hwc->freq_time_stamp = now;
4998 if (delta > 0 && delta < 2*TICK_NSEC)
4999 perf_adjust_period(event, delta, hwc->last_period);
5003 * XXX event_limit might not quite work as expected on inherited
5007 event->pending_kill = POLL_IN;
5008 if (events && atomic_dec_and_test(&event->event_limit)) {
5010 event->pending_kill = POLL_HUP;
5012 event->pending_disable = 1;
5013 irq_work_queue(&event->pending);
5015 perf_event_disable(event);
5018 if (event->overflow_handler)
5019 event->overflow_handler(event, nmi, data, regs);
5021 perf_event_output(event, nmi, data, regs);
5026 int perf_event_overflow(struct perf_event *event, int nmi,
5027 struct perf_sample_data *data,
5028 struct pt_regs *regs)
5030 return __perf_event_overflow(event, nmi, 1, data, regs);
5034 * Generic software event infrastructure
5037 struct swevent_htable {
5038 struct swevent_hlist *swevent_hlist;
5039 struct mutex hlist_mutex;
5042 /* Recursion avoidance in each contexts */
5043 int recursion[PERF_NR_CONTEXTS];
5046 static DEFINE_PER_CPU(struct swevent_htable, swevent_htable);
5049 * We directly increment event->count and keep a second value in
5050 * event->hw.period_left to count intervals. This period event
5051 * is kept in the range [-sample_period, 0] so that we can use the
5055 static u64 perf_swevent_set_period(struct perf_event *event)
5057 struct hw_perf_event *hwc = &event->hw;
5058 u64 period = hwc->last_period;
5062 hwc->last_period = hwc->sample_period;
5065 old = val = local64_read(&hwc->period_left);
5069 nr = div64_u64(period + val, period);
5070 offset = nr * period;
5072 if (local64_cmpxchg(&hwc->period_left, old, val) != old)
5078 static void perf_swevent_overflow(struct perf_event *event, u64 overflow,
5079 int nmi, struct perf_sample_data *data,
5080 struct pt_regs *regs)
5082 struct hw_perf_event *hwc = &event->hw;
5085 data->period = event->hw.last_period;
5087 overflow = perf_swevent_set_period(event);
5089 if (hwc->interrupts == MAX_INTERRUPTS)
5092 for (; overflow; overflow--) {
5093 if (__perf_event_overflow(event, nmi, throttle,
5096 * We inhibit the overflow from happening when
5097 * hwc->interrupts == MAX_INTERRUPTS.
5105 static void perf_swevent_event(struct perf_event *event, u64 nr,
5106 int nmi, struct perf_sample_data *data,
5107 struct pt_regs *regs)
5109 struct hw_perf_event *hwc = &event->hw;
5111 local64_add(nr, &event->count);
5116 if (!is_sampling_event(event))
5119 if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq)
5120 return perf_swevent_overflow(event, 1, nmi, data, regs);
5122 if (local64_add_negative(nr, &hwc->period_left))
5125 perf_swevent_overflow(event, 0, nmi, data, regs);
5128 static int perf_exclude_event(struct perf_event *event,
5129 struct pt_regs *regs)
5131 if (event->hw.state & PERF_HES_STOPPED)
5135 if (event->attr.exclude_user && user_mode(regs))
5138 if (event->attr.exclude_kernel && !user_mode(regs))
5145 static int perf_swevent_match(struct perf_event *event,
5146 enum perf_type_id type,
5148 struct perf_sample_data *data,
5149 struct pt_regs *regs)
5151 if (event->attr.type != type)
5154 if (event->attr.config != event_id)
5157 if (perf_exclude_event(event, regs))
5163 static inline u64 swevent_hash(u64 type, u32 event_id)
5165 u64 val = event_id | (type << 32);
5167 return hash_64(val, SWEVENT_HLIST_BITS);
5170 static inline struct hlist_head *
5171 __find_swevent_head(struct swevent_hlist *hlist, u64 type, u32 event_id)
5173 u64 hash = swevent_hash(type, event_id);
5175 return &hlist->heads[hash];
5178 /* For the read side: events when they trigger */
5179 static inline struct hlist_head *
5180 find_swevent_head_rcu(struct swevent_htable *swhash, u64 type, u32 event_id)
5182 struct swevent_hlist *hlist;
5184 hlist = rcu_dereference(swhash->swevent_hlist);
5188 return __find_swevent_head(hlist, type, event_id);
5191 /* For the event head insertion and removal in the hlist */
5192 static inline struct hlist_head *
5193 find_swevent_head(struct swevent_htable *swhash, struct perf_event *event)
5195 struct swevent_hlist *hlist;
5196 u32 event_id = event->attr.config;
5197 u64 type = event->attr.type;
5200 * Event scheduling is always serialized against hlist allocation
5201 * and release. Which makes the protected version suitable here.
5202 * The context lock guarantees that.
5204 hlist = rcu_dereference_protected(swhash->swevent_hlist,
5205 lockdep_is_held(&event->ctx->lock));
5209 return __find_swevent_head(hlist, type, event_id);
5212 static void do_perf_sw_event(enum perf_type_id type, u32 event_id,
5214 struct perf_sample_data *data,
5215 struct pt_regs *regs)
5217 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
5218 struct perf_event *event;
5219 struct hlist_node *node;
5220 struct hlist_head *head;
5223 head = find_swevent_head_rcu(swhash, type, event_id);
5227 hlist_for_each_entry_rcu(event, node, head, hlist_entry) {
5228 if (perf_swevent_match(event, type, event_id, data, regs))
5229 perf_swevent_event(event, nr, nmi, data, regs);
5235 int perf_swevent_get_recursion_context(void)
5237 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
5239 return get_recursion_context(swhash->recursion);
5241 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context);
5243 inline void perf_swevent_put_recursion_context(int rctx)
5245 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
5247 put_recursion_context(swhash->recursion, rctx);
5250 void __perf_sw_event(u32 event_id, u64 nr, int nmi,
5251 struct pt_regs *regs, u64 addr)
5253 struct perf_sample_data data;
5256 preempt_disable_notrace();
5257 rctx = perf_swevent_get_recursion_context();
5261 perf_sample_data_init(&data, addr);
5263 do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, nmi, &data, regs);
5265 perf_swevent_put_recursion_context(rctx);
5266 preempt_enable_notrace();
5269 static void perf_swevent_read(struct perf_event *event)
5273 static int perf_swevent_add(struct perf_event *event, int flags)
5275 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
5276 struct hw_perf_event *hwc = &event->hw;
5277 struct hlist_head *head;
5279 if (is_sampling_event(event)) {
5280 hwc->last_period = hwc->sample_period;
5281 perf_swevent_set_period(event);
5284 hwc->state = !(flags & PERF_EF_START);
5286 head = find_swevent_head(swhash, event);
5287 if (WARN_ON_ONCE(!head))
5290 hlist_add_head_rcu(&event->hlist_entry, head);
5295 static void perf_swevent_del(struct perf_event *event, int flags)
5297 hlist_del_rcu(&event->hlist_entry);
5300 static void perf_swevent_start(struct perf_event *event, int flags)
5302 event->hw.state = 0;
5305 static void perf_swevent_stop(struct perf_event *event, int flags)
5307 event->hw.state = PERF_HES_STOPPED;
5310 /* Deref the hlist from the update side */
5311 static inline struct swevent_hlist *
5312 swevent_hlist_deref(struct swevent_htable *swhash)
5314 return rcu_dereference_protected(swhash->swevent_hlist,
5315 lockdep_is_held(&swhash->hlist_mutex));
5318 static void swevent_hlist_release_rcu(struct rcu_head *rcu_head)
5320 struct swevent_hlist *hlist;
5322 hlist = container_of(rcu_head, struct swevent_hlist, rcu_head);
5326 static void swevent_hlist_release(struct swevent_htable *swhash)
5328 struct swevent_hlist *hlist = swevent_hlist_deref(swhash);
5333 rcu_assign_pointer(swhash->swevent_hlist, NULL);
5334 call_rcu(&hlist->rcu_head, swevent_hlist_release_rcu);
5337 static void swevent_hlist_put_cpu(struct perf_event *event, int cpu)
5339 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
5341 mutex_lock(&swhash->hlist_mutex);
5343 if (!--swhash->hlist_refcount)
5344 swevent_hlist_release(swhash);
5346 mutex_unlock(&swhash->hlist_mutex);
5349 static void swevent_hlist_put(struct perf_event *event)
5353 if (event->cpu != -1) {
5354 swevent_hlist_put_cpu(event, event->cpu);
5358 for_each_possible_cpu(cpu)
5359 swevent_hlist_put_cpu(event, cpu);
5362 static int swevent_hlist_get_cpu(struct perf_event *event, int cpu)
5364 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
5367 mutex_lock(&swhash->hlist_mutex);
5369 if (!swevent_hlist_deref(swhash) && cpu_online(cpu)) {
5370 struct swevent_hlist *hlist;
5372 hlist = kzalloc(sizeof(*hlist), GFP_KERNEL);
5377 rcu_assign_pointer(swhash->swevent_hlist, hlist);
5379 swhash->hlist_refcount++;
5381 mutex_unlock(&swhash->hlist_mutex);
5386 static int swevent_hlist_get(struct perf_event *event)
5389 int cpu, failed_cpu;
5391 if (event->cpu != -1)
5392 return swevent_hlist_get_cpu(event, event->cpu);
5395 for_each_possible_cpu(cpu) {
5396 err = swevent_hlist_get_cpu(event, cpu);
5406 for_each_possible_cpu(cpu) {
5407 if (cpu == failed_cpu)
5409 swevent_hlist_put_cpu(event, cpu);
5416 atomic_t perf_swevent_enabled[PERF_COUNT_SW_MAX];
5418 static void sw_perf_event_destroy(struct perf_event *event)
5420 u64 event_id = event->attr.config;
5422 WARN_ON(event->parent);
5424 jump_label_dec(&perf_swevent_enabled[event_id]);
5425 swevent_hlist_put(event);
5428 static int perf_swevent_init(struct perf_event *event)
5430 int event_id = event->attr.config;
5432 if (event->attr.type != PERF_TYPE_SOFTWARE)
5436 case PERF_COUNT_SW_CPU_CLOCK:
5437 case PERF_COUNT_SW_TASK_CLOCK:
5444 if (event_id >= PERF_COUNT_SW_MAX)
5447 if (!event->parent) {
5450 err = swevent_hlist_get(event);
5454 jump_label_inc(&perf_swevent_enabled[event_id]);
5455 event->destroy = sw_perf_event_destroy;
5461 static struct pmu perf_swevent = {
5462 .task_ctx_nr = perf_sw_context,
5464 .event_init = perf_swevent_init,
5465 .add = perf_swevent_add,
5466 .del = perf_swevent_del,
5467 .start = perf_swevent_start,
5468 .stop = perf_swevent_stop,
5469 .read = perf_swevent_read,
5472 #ifdef CONFIG_EVENT_TRACING
5474 static int perf_tp_filter_match(struct perf_event *event,
5475 struct perf_sample_data *data)
5477 void *record = data->raw->data;
5479 if (likely(!event->filter) || filter_match_preds(event->filter, record))
5484 static int perf_tp_event_match(struct perf_event *event,
5485 struct perf_sample_data *data,
5486 struct pt_regs *regs)
5489 * All tracepoints are from kernel-space.
5491 if (event->attr.exclude_kernel)
5494 if (!perf_tp_filter_match(event, data))
5500 void perf_tp_event(u64 addr, u64 count, void *record, int entry_size,
5501 struct pt_regs *regs, struct hlist_head *head, int rctx)
5503 struct perf_sample_data data;
5504 struct perf_event *event;
5505 struct hlist_node *node;
5507 struct perf_raw_record raw = {
5512 perf_sample_data_init(&data, addr);
5515 hlist_for_each_entry_rcu(event, node, head, hlist_entry) {
5516 if (perf_tp_event_match(event, &data, regs))
5517 perf_swevent_event(event, count, 1, &data, regs);
5520 perf_swevent_put_recursion_context(rctx);
5522 EXPORT_SYMBOL_GPL(perf_tp_event);
5524 static void tp_perf_event_destroy(struct perf_event *event)
5526 perf_trace_destroy(event);
5529 static int perf_tp_event_init(struct perf_event *event)
5533 if (event->attr.type != PERF_TYPE_TRACEPOINT)
5536 err = perf_trace_init(event);
5540 event->destroy = tp_perf_event_destroy;
5545 static struct pmu perf_tracepoint = {
5546 .task_ctx_nr = perf_sw_context,
5548 .event_init = perf_tp_event_init,
5549 .add = perf_trace_add,
5550 .del = perf_trace_del,
5551 .start = perf_swevent_start,
5552 .stop = perf_swevent_stop,
5553 .read = perf_swevent_read,
5556 static inline void perf_tp_register(void)
5558 perf_pmu_register(&perf_tracepoint, "tracepoint", PERF_TYPE_TRACEPOINT);
5561 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
5566 if (event->attr.type != PERF_TYPE_TRACEPOINT)
5569 filter_str = strndup_user(arg, PAGE_SIZE);
5570 if (IS_ERR(filter_str))
5571 return PTR_ERR(filter_str);
5573 ret = ftrace_profile_set_filter(event, event->attr.config, filter_str);
5579 static void perf_event_free_filter(struct perf_event *event)
5581 ftrace_profile_free_filter(event);
5586 static inline void perf_tp_register(void)
5590 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
5595 static void perf_event_free_filter(struct perf_event *event)
5599 #endif /* CONFIG_EVENT_TRACING */
5601 #ifdef CONFIG_HAVE_HW_BREAKPOINT
5602 void perf_bp_event(struct perf_event *bp, void *data)
5604 struct perf_sample_data sample;
5605 struct pt_regs *regs = data;
5607 perf_sample_data_init(&sample, bp->attr.bp_addr);
5609 if (!bp->hw.state && !perf_exclude_event(bp, regs))
5610 perf_swevent_event(bp, 1, 1, &sample, regs);
5615 * hrtimer based swevent callback
5618 static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer)
5620 enum hrtimer_restart ret = HRTIMER_RESTART;
5621 struct perf_sample_data data;
5622 struct pt_regs *regs;
5623 struct perf_event *event;
5626 event = container_of(hrtimer, struct perf_event, hw.hrtimer);
5628 if (event->state != PERF_EVENT_STATE_ACTIVE)
5629 return HRTIMER_NORESTART;
5631 event->pmu->read(event);
5633 perf_sample_data_init(&data, 0);
5634 data.period = event->hw.last_period;
5635 regs = get_irq_regs();
5637 if (regs && !perf_exclude_event(event, regs)) {
5638 if (!(event->attr.exclude_idle && current->pid == 0))
5639 if (perf_event_overflow(event, 0, &data, regs))
5640 ret = HRTIMER_NORESTART;
5643 period = max_t(u64, 10000, event->hw.sample_period);
5644 hrtimer_forward_now(hrtimer, ns_to_ktime(period));
5649 static void perf_swevent_start_hrtimer(struct perf_event *event)
5651 struct hw_perf_event *hwc = &event->hw;
5654 if (!is_sampling_event(event))
5657 period = local64_read(&hwc->period_left);
5662 local64_set(&hwc->period_left, 0);
5664 period = max_t(u64, 10000, hwc->sample_period);
5666 __hrtimer_start_range_ns(&hwc->hrtimer,
5667 ns_to_ktime(period), 0,
5668 HRTIMER_MODE_REL_PINNED, 0);
5671 static void perf_swevent_cancel_hrtimer(struct perf_event *event)
5673 struct hw_perf_event *hwc = &event->hw;
5675 if (is_sampling_event(event)) {
5676 ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer);
5677 local64_set(&hwc->period_left, ktime_to_ns(remaining));
5679 hrtimer_cancel(&hwc->hrtimer);
5683 static void perf_swevent_init_hrtimer(struct perf_event *event)
5685 struct hw_perf_event *hwc = &event->hw;
5687 if (!is_sampling_event(event))
5690 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
5691 hwc->hrtimer.function = perf_swevent_hrtimer;
5694 * Since hrtimers have a fixed rate, we can do a static freq->period
5695 * mapping and avoid the whole period adjust feedback stuff.
5697 if (event->attr.freq) {
5698 long freq = event->attr.sample_freq;
5700 event->attr.sample_period = NSEC_PER_SEC / freq;
5701 hwc->sample_period = event->attr.sample_period;
5702 local64_set(&hwc->period_left, hwc->sample_period);
5703 event->attr.freq = 0;
5708 * Software event: cpu wall time clock
5711 static void cpu_clock_event_update(struct perf_event *event)
5716 now = local_clock();
5717 prev = local64_xchg(&event->hw.prev_count, now);
5718 local64_add(now - prev, &event->count);
5721 static void cpu_clock_event_start(struct perf_event *event, int flags)
5723 local64_set(&event->hw.prev_count, local_clock());
5724 perf_swevent_start_hrtimer(event);
5727 static void cpu_clock_event_stop(struct perf_event *event, int flags)
5729 perf_swevent_cancel_hrtimer(event);
5730 cpu_clock_event_update(event);
5733 static int cpu_clock_event_add(struct perf_event *event, int flags)
5735 if (flags & PERF_EF_START)
5736 cpu_clock_event_start(event, flags);
5741 static void cpu_clock_event_del(struct perf_event *event, int flags)
5743 cpu_clock_event_stop(event, flags);
5746 static void cpu_clock_event_read(struct perf_event *event)
5748 cpu_clock_event_update(event);
5751 static int cpu_clock_event_init(struct perf_event *event)
5753 if (event->attr.type != PERF_TYPE_SOFTWARE)
5756 if (event->attr.config != PERF_COUNT_SW_CPU_CLOCK)
5759 perf_swevent_init_hrtimer(event);
5764 static struct pmu perf_cpu_clock = {
5765 .task_ctx_nr = perf_sw_context,
5767 .event_init = cpu_clock_event_init,
5768 .add = cpu_clock_event_add,
5769 .del = cpu_clock_event_del,
5770 .start = cpu_clock_event_start,
5771 .stop = cpu_clock_event_stop,
5772 .read = cpu_clock_event_read,
5776 * Software event: task time clock
5779 static void task_clock_event_update(struct perf_event *event, u64 now)
5784 prev = local64_xchg(&event->hw.prev_count, now);
5786 local64_add(delta, &event->count);
5789 static void task_clock_event_start(struct perf_event *event, int flags)
5791 local64_set(&event->hw.prev_count, event->ctx->time);
5792 perf_swevent_start_hrtimer(event);
5795 static void task_clock_event_stop(struct perf_event *event, int flags)
5797 perf_swevent_cancel_hrtimer(event);
5798 task_clock_event_update(event, event->ctx->time);
5801 static int task_clock_event_add(struct perf_event *event, int flags)
5803 if (flags & PERF_EF_START)
5804 task_clock_event_start(event, flags);
5809 static void task_clock_event_del(struct perf_event *event, int flags)
5811 task_clock_event_stop(event, PERF_EF_UPDATE);
5814 static void task_clock_event_read(struct perf_event *event)
5816 u64 now = perf_clock();
5817 u64 delta = now - event->ctx->timestamp;
5818 u64 time = event->ctx->time + delta;
5820 task_clock_event_update(event, time);
5823 static int task_clock_event_init(struct perf_event *event)
5825 if (event->attr.type != PERF_TYPE_SOFTWARE)
5828 if (event->attr.config != PERF_COUNT_SW_TASK_CLOCK)
5831 perf_swevent_init_hrtimer(event);
5836 static struct pmu perf_task_clock = {
5837 .task_ctx_nr = perf_sw_context,
5839 .event_init = task_clock_event_init,
5840 .add = task_clock_event_add,
5841 .del = task_clock_event_del,
5842 .start = task_clock_event_start,
5843 .stop = task_clock_event_stop,
5844 .read = task_clock_event_read,
5847 static void perf_pmu_nop_void(struct pmu *pmu)
5851 static int perf_pmu_nop_int(struct pmu *pmu)
5856 static void perf_pmu_start_txn(struct pmu *pmu)
5858 perf_pmu_disable(pmu);
5861 static int perf_pmu_commit_txn(struct pmu *pmu)
5863 perf_pmu_enable(pmu);
5867 static void perf_pmu_cancel_txn(struct pmu *pmu)
5869 perf_pmu_enable(pmu);
5873 * Ensures all contexts with the same task_ctx_nr have the same
5874 * pmu_cpu_context too.
5876 static void *find_pmu_context(int ctxn)
5883 list_for_each_entry(pmu, &pmus, entry) {
5884 if (pmu->task_ctx_nr == ctxn)
5885 return pmu->pmu_cpu_context;
5891 static void update_pmu_context(struct pmu *pmu, struct pmu *old_pmu)
5895 for_each_possible_cpu(cpu) {
5896 struct perf_cpu_context *cpuctx;
5898 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
5900 if (cpuctx->active_pmu == old_pmu)
5901 cpuctx->active_pmu = pmu;
5905 static void free_pmu_context(struct pmu *pmu)
5909 mutex_lock(&pmus_lock);
5911 * Like a real lame refcount.
5913 list_for_each_entry(i, &pmus, entry) {
5914 if (i->pmu_cpu_context == pmu->pmu_cpu_context) {
5915 update_pmu_context(i, pmu);
5920 free_percpu(pmu->pmu_cpu_context);
5922 mutex_unlock(&pmus_lock);
5924 static struct idr pmu_idr;
5927 type_show(struct device *dev, struct device_attribute *attr, char *page)
5929 struct pmu *pmu = dev_get_drvdata(dev);
5931 return snprintf(page, PAGE_SIZE-1, "%d\n", pmu->type);
5934 static struct device_attribute pmu_dev_attrs[] = {
5939 static int pmu_bus_running;
5940 static struct bus_type pmu_bus = {
5941 .name = "event_source",
5942 .dev_attrs = pmu_dev_attrs,
5945 static void pmu_dev_release(struct device *dev)
5950 static int pmu_dev_alloc(struct pmu *pmu)
5954 pmu->dev = kzalloc(sizeof(struct device), GFP_KERNEL);
5958 device_initialize(pmu->dev);
5959 ret = dev_set_name(pmu->dev, "%s", pmu->name);
5963 dev_set_drvdata(pmu->dev, pmu);
5964 pmu->dev->bus = &pmu_bus;
5965 pmu->dev->release = pmu_dev_release;
5966 ret = device_add(pmu->dev);
5974 put_device(pmu->dev);
5978 static struct lock_class_key cpuctx_mutex;
5980 int perf_pmu_register(struct pmu *pmu, char *name, int type)
5984 mutex_lock(&pmus_lock);
5986 pmu->pmu_disable_count = alloc_percpu(int);
5987 if (!pmu->pmu_disable_count)
5996 int err = idr_pre_get(&pmu_idr, GFP_KERNEL);
6000 err = idr_get_new_above(&pmu_idr, pmu, PERF_TYPE_MAX, &type);
6008 if (pmu_bus_running) {
6009 ret = pmu_dev_alloc(pmu);
6015 pmu->pmu_cpu_context = find_pmu_context(pmu->task_ctx_nr);
6016 if (pmu->pmu_cpu_context)
6017 goto got_cpu_context;
6019 pmu->pmu_cpu_context = alloc_percpu(struct perf_cpu_context);
6020 if (!pmu->pmu_cpu_context)
6023 for_each_possible_cpu(cpu) {
6024 struct perf_cpu_context *cpuctx;
6026 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
6027 __perf_event_init_context(&cpuctx->ctx);
6028 lockdep_set_class(&cpuctx->ctx.mutex, &cpuctx_mutex);
6029 cpuctx->ctx.type = cpu_context;
6030 cpuctx->ctx.pmu = pmu;
6031 cpuctx->jiffies_interval = 1;
6032 INIT_LIST_HEAD(&cpuctx->rotation_list);
6033 cpuctx->active_pmu = pmu;
6037 if (!pmu->start_txn) {
6038 if (pmu->pmu_enable) {
6040 * If we have pmu_enable/pmu_disable calls, install
6041 * transaction stubs that use that to try and batch
6042 * hardware accesses.
6044 pmu->start_txn = perf_pmu_start_txn;
6045 pmu->commit_txn = perf_pmu_commit_txn;
6046 pmu->cancel_txn = perf_pmu_cancel_txn;
6048 pmu->start_txn = perf_pmu_nop_void;
6049 pmu->commit_txn = perf_pmu_nop_int;
6050 pmu->cancel_txn = perf_pmu_nop_void;
6054 if (!pmu->pmu_enable) {
6055 pmu->pmu_enable = perf_pmu_nop_void;
6056 pmu->pmu_disable = perf_pmu_nop_void;
6059 list_add_rcu(&pmu->entry, &pmus);
6062 mutex_unlock(&pmus_lock);
6067 device_del(pmu->dev);
6068 put_device(pmu->dev);
6071 if (pmu->type >= PERF_TYPE_MAX)
6072 idr_remove(&pmu_idr, pmu->type);
6075 free_percpu(pmu->pmu_disable_count);
6079 void perf_pmu_unregister(struct pmu *pmu)
6081 mutex_lock(&pmus_lock);
6082 list_del_rcu(&pmu->entry);
6083 mutex_unlock(&pmus_lock);
6086 * We dereference the pmu list under both SRCU and regular RCU, so
6087 * synchronize against both of those.
6089 synchronize_srcu(&pmus_srcu);
6092 free_percpu(pmu->pmu_disable_count);
6093 if (pmu->type >= PERF_TYPE_MAX)
6094 idr_remove(&pmu_idr, pmu->type);
6095 device_del(pmu->dev);
6096 put_device(pmu->dev);
6097 free_pmu_context(pmu);
6100 struct pmu *perf_init_event(struct perf_event *event)
6102 struct pmu *pmu = NULL;
6106 idx = srcu_read_lock(&pmus_srcu);
6109 pmu = idr_find(&pmu_idr, event->attr.type);
6112 ret = pmu->event_init(event);
6118 list_for_each_entry_rcu(pmu, &pmus, entry) {
6119 ret = pmu->event_init(event);
6123 if (ret != -ENOENT) {
6128 pmu = ERR_PTR(-ENOENT);
6130 srcu_read_unlock(&pmus_srcu, idx);
6136 * Allocate and initialize a event structure
6138 static struct perf_event *
6139 perf_event_alloc(struct perf_event_attr *attr, int cpu,
6140 struct task_struct *task,
6141 struct perf_event *group_leader,
6142 struct perf_event *parent_event,
6143 perf_overflow_handler_t overflow_handler)
6146 struct perf_event *event;
6147 struct hw_perf_event *hwc;
6150 if ((unsigned)cpu >= nr_cpu_ids) {
6151 if (!task || cpu != -1)
6152 return ERR_PTR(-EINVAL);
6155 event = kzalloc(sizeof(*event), GFP_KERNEL);
6157 return ERR_PTR(-ENOMEM);
6160 * Single events are their own group leaders, with an
6161 * empty sibling list:
6164 group_leader = event;
6166 mutex_init(&event->child_mutex);
6167 INIT_LIST_HEAD(&event->child_list);
6169 INIT_LIST_HEAD(&event->group_entry);
6170 INIT_LIST_HEAD(&event->event_entry);
6171 INIT_LIST_HEAD(&event->sibling_list);
6172 init_waitqueue_head(&event->waitq);
6173 init_irq_work(&event->pending, perf_pending_event);
6175 mutex_init(&event->mmap_mutex);
6178 event->attr = *attr;
6179 event->group_leader = group_leader;
6183 event->parent = parent_event;
6185 event->ns = get_pid_ns(current->nsproxy->pid_ns);
6186 event->id = atomic64_inc_return(&perf_event_id);
6188 event->state = PERF_EVENT_STATE_INACTIVE;
6191 event->attach_state = PERF_ATTACH_TASK;
6192 #ifdef CONFIG_HAVE_HW_BREAKPOINT
6194 * hw_breakpoint is a bit difficult here..
6196 if (attr->type == PERF_TYPE_BREAKPOINT)
6197 event->hw.bp_target = task;
6201 if (!overflow_handler && parent_event)
6202 overflow_handler = parent_event->overflow_handler;
6204 event->overflow_handler = overflow_handler;
6207 event->state = PERF_EVENT_STATE_OFF;
6212 hwc->sample_period = attr->sample_period;
6213 if (attr->freq && attr->sample_freq)
6214 hwc->sample_period = 1;
6215 hwc->last_period = hwc->sample_period;
6217 local64_set(&hwc->period_left, hwc->sample_period);
6220 * we currently do not support PERF_FORMAT_GROUP on inherited events
6222 if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP))
6225 pmu = perf_init_event(event);
6231 else if (IS_ERR(pmu))
6236 put_pid_ns(event->ns);
6238 return ERR_PTR(err);
6243 if (!event->parent) {
6244 if (event->attach_state & PERF_ATTACH_TASK)
6245 jump_label_inc(&perf_sched_events);
6246 if (event->attr.mmap || event->attr.mmap_data)
6247 atomic_inc(&nr_mmap_events);
6248 if (event->attr.comm)
6249 atomic_inc(&nr_comm_events);
6250 if (event->attr.task)
6251 atomic_inc(&nr_task_events);
6252 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) {
6253 err = get_callchain_buffers();
6256 return ERR_PTR(err);
6264 static int perf_copy_attr(struct perf_event_attr __user *uattr,
6265 struct perf_event_attr *attr)
6270 if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0))
6274 * zero the full structure, so that a short copy will be nice.
6276 memset(attr, 0, sizeof(*attr));
6278 ret = get_user(size, &uattr->size);
6282 if (size > PAGE_SIZE) /* silly large */
6285 if (!size) /* abi compat */
6286 size = PERF_ATTR_SIZE_VER0;
6288 if (size < PERF_ATTR_SIZE_VER0)
6292 * If we're handed a bigger struct than we know of,
6293 * ensure all the unknown bits are 0 - i.e. new
6294 * user-space does not rely on any kernel feature
6295 * extensions we dont know about yet.
6297 if (size > sizeof(*attr)) {
6298 unsigned char __user *addr;
6299 unsigned char __user *end;
6302 addr = (void __user *)uattr + sizeof(*attr);
6303 end = (void __user *)uattr + size;
6305 for (; addr < end; addr++) {
6306 ret = get_user(val, addr);
6312 size = sizeof(*attr);
6315 ret = copy_from_user(attr, uattr, size);
6320 * If the type exists, the corresponding creation will verify
6323 if (attr->type >= PERF_TYPE_MAX)
6326 if (attr->__reserved_1)
6329 if (attr->sample_type & ~(PERF_SAMPLE_MAX-1))
6332 if (attr->read_format & ~(PERF_FORMAT_MAX-1))
6339 put_user(sizeof(*attr), &uattr->size);
6345 perf_event_set_output(struct perf_event *event, struct perf_event *output_event)
6347 struct perf_buffer *buffer = NULL, *old_buffer = NULL;
6353 /* don't allow circular references */
6354 if (event == output_event)
6358 * Don't allow cross-cpu buffers
6360 if (output_event->cpu != event->cpu)
6364 * If its not a per-cpu buffer, it must be the same task.
6366 if (output_event->cpu == -1 && output_event->ctx != event->ctx)
6370 mutex_lock(&event->mmap_mutex);
6371 /* Can't redirect output if we've got an active mmap() */
6372 if (atomic_read(&event->mmap_count))
6376 /* get the buffer we want to redirect to */
6377 buffer = perf_buffer_get(output_event);
6382 old_buffer = event->buffer;
6383 rcu_assign_pointer(event->buffer, buffer);
6386 mutex_unlock(&event->mmap_mutex);
6389 perf_buffer_put(old_buffer);
6395 * sys_perf_event_open - open a performance event, associate it to a task/cpu
6397 * @attr_uptr: event_id type attributes for monitoring/sampling
6400 * @group_fd: group leader event fd
6402 SYSCALL_DEFINE5(perf_event_open,
6403 struct perf_event_attr __user *, attr_uptr,
6404 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
6406 struct perf_event *group_leader = NULL, *output_event = NULL;
6407 struct perf_event *event, *sibling;
6408 struct perf_event_attr attr;
6409 struct perf_event_context *ctx;
6410 struct file *event_file = NULL;
6411 struct file *group_file = NULL;
6412 struct task_struct *task = NULL;
6416 int fput_needed = 0;
6419 /* for future expandability... */
6420 if (flags & ~PERF_FLAG_ALL)
6423 err = perf_copy_attr(attr_uptr, &attr);
6427 if (!attr.exclude_kernel) {
6428 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
6433 if (attr.sample_freq > sysctl_perf_event_sample_rate)
6438 * In cgroup mode, the pid argument is used to pass the fd
6439 * opened to the cgroup directory in cgroupfs. The cpu argument
6440 * designates the cpu on which to monitor threads from that
6443 if ((flags & PERF_FLAG_PID_CGROUP) && (pid == -1 || cpu == -1))
6446 event_fd = get_unused_fd_flags(O_RDWR);
6450 if (group_fd != -1) {
6451 group_leader = perf_fget_light(group_fd, &fput_needed);
6452 if (IS_ERR(group_leader)) {
6453 err = PTR_ERR(group_leader);
6456 group_file = group_leader->filp;
6457 if (flags & PERF_FLAG_FD_OUTPUT)
6458 output_event = group_leader;
6459 if (flags & PERF_FLAG_FD_NO_GROUP)
6460 group_leader = NULL;
6463 if (pid != -1 && !(flags & PERF_FLAG_PID_CGROUP)) {
6464 task = find_lively_task_by_vpid(pid);
6466 err = PTR_ERR(task);
6471 event = perf_event_alloc(&attr, cpu, task, group_leader, NULL, NULL);
6472 if (IS_ERR(event)) {
6473 err = PTR_ERR(event);
6477 if (flags & PERF_FLAG_PID_CGROUP) {
6478 err = perf_cgroup_connect(pid, event, &attr, group_leader);
6484 * Special case software events and allow them to be part of
6485 * any hardware group.
6490 (is_software_event(event) != is_software_event(group_leader))) {
6491 if (is_software_event(event)) {
6493 * If event and group_leader are not both a software
6494 * event, and event is, then group leader is not.
6496 * Allow the addition of software events to !software
6497 * groups, this is safe because software events never
6500 pmu = group_leader->pmu;
6501 } else if (is_software_event(group_leader) &&
6502 (group_leader->group_flags & PERF_GROUP_SOFTWARE)) {
6504 * In case the group is a pure software group, and we
6505 * try to add a hardware event, move the whole group to
6506 * the hardware context.
6513 * Get the target context (task or percpu):
6515 ctx = find_get_context(pmu, task, cpu);
6522 * Look up the group leader (we will attach this event to it):
6528 * Do not allow a recursive hierarchy (this new sibling
6529 * becoming part of another group-sibling):
6531 if (group_leader->group_leader != group_leader)
6534 * Do not allow to attach to a group in a different
6535 * task or CPU context:
6538 if (group_leader->ctx->type != ctx->type)
6541 if (group_leader->ctx != ctx)
6546 * Only a group leader can be exclusive or pinned
6548 if (attr.exclusive || attr.pinned)
6553 err = perf_event_set_output(event, output_event);
6558 event_file = anon_inode_getfile("[perf_event]", &perf_fops, event, O_RDWR);
6559 if (IS_ERR(event_file)) {
6560 err = PTR_ERR(event_file);
6565 struct perf_event_context *gctx = group_leader->ctx;
6567 mutex_lock(&gctx->mutex);
6568 perf_remove_from_context(group_leader);
6569 list_for_each_entry(sibling, &group_leader->sibling_list,
6571 perf_remove_from_context(sibling);
6574 mutex_unlock(&gctx->mutex);
6578 event->filp = event_file;
6579 WARN_ON_ONCE(ctx->parent_ctx);
6580 mutex_lock(&ctx->mutex);
6583 perf_install_in_context(ctx, group_leader, cpu);
6585 list_for_each_entry(sibling, &group_leader->sibling_list,
6587 perf_install_in_context(ctx, sibling, cpu);
6592 perf_install_in_context(ctx, event, cpu);
6594 perf_unpin_context(ctx);
6595 mutex_unlock(&ctx->mutex);
6597 event->owner = current;
6599 mutex_lock(¤t->perf_event_mutex);
6600 list_add_tail(&event->owner_entry, ¤t->perf_event_list);
6601 mutex_unlock(¤t->perf_event_mutex);
6604 * Precalculate sample_data sizes
6606 perf_event__header_size(event);
6607 perf_event__id_header_size(event);
6610 * Drop the reference on the group_event after placing the
6611 * new event on the sibling_list. This ensures destruction
6612 * of the group leader will find the pointer to itself in
6613 * perf_group_detach().
6615 fput_light(group_file, fput_needed);
6616 fd_install(event_fd, event_file);
6620 perf_unpin_context(ctx);
6626 put_task_struct(task);
6628 fput_light(group_file, fput_needed);
6630 put_unused_fd(event_fd);
6635 * perf_event_create_kernel_counter
6637 * @attr: attributes of the counter to create
6638 * @cpu: cpu in which the counter is bound
6639 * @task: task to profile (NULL for percpu)
6642 perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu,
6643 struct task_struct *task,
6644 perf_overflow_handler_t overflow_handler)
6646 struct perf_event_context *ctx;
6647 struct perf_event *event;
6651 * Get the target context (task or percpu):
6654 event = perf_event_alloc(attr, cpu, task, NULL, NULL, overflow_handler);
6655 if (IS_ERR(event)) {
6656 err = PTR_ERR(event);
6660 ctx = find_get_context(event->pmu, task, cpu);
6667 WARN_ON_ONCE(ctx->parent_ctx);
6668 mutex_lock(&ctx->mutex);
6669 perf_install_in_context(ctx, event, cpu);
6671 perf_unpin_context(ctx);
6672 mutex_unlock(&ctx->mutex);
6679 return ERR_PTR(err);
6681 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter);
6683 static void sync_child_event(struct perf_event *child_event,
6684 struct task_struct *child)
6686 struct perf_event *parent_event = child_event->parent;
6689 if (child_event->attr.inherit_stat)
6690 perf_event_read_event(child_event, child);
6692 child_val = perf_event_count(child_event);
6695 * Add back the child's count to the parent's count:
6697 atomic64_add(child_val, &parent_event->child_count);
6698 atomic64_add(child_event->total_time_enabled,
6699 &parent_event->child_total_time_enabled);
6700 atomic64_add(child_event->total_time_running,
6701 &parent_event->child_total_time_running);
6704 * Remove this event from the parent's list
6706 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
6707 mutex_lock(&parent_event->child_mutex);
6708 list_del_init(&child_event->child_list);
6709 mutex_unlock(&parent_event->child_mutex);
6712 * Release the parent event, if this was the last
6715 fput(parent_event->filp);
6719 __perf_event_exit_task(struct perf_event *child_event,
6720 struct perf_event_context *child_ctx,
6721 struct task_struct *child)
6723 struct perf_event *parent_event;
6725 perf_remove_from_context(child_event);
6727 parent_event = child_event->parent;
6729 * It can happen that parent exits first, and has events
6730 * that are still around due to the child reference. These
6731 * events need to be zapped - but otherwise linger.
6734 sync_child_event(child_event, child);
6735 free_event(child_event);
6739 static void perf_event_exit_task_context(struct task_struct *child, int ctxn)
6741 struct perf_event *child_event, *tmp;
6742 struct perf_event_context *child_ctx;
6743 unsigned long flags;
6745 if (likely(!child->perf_event_ctxp[ctxn])) {
6746 perf_event_task(child, NULL, 0);
6750 local_irq_save(flags);
6752 * We can't reschedule here because interrupts are disabled,
6753 * and either child is current or it is a task that can't be
6754 * scheduled, so we are now safe from rescheduling changing
6757 child_ctx = rcu_dereference_raw(child->perf_event_ctxp[ctxn]);
6758 task_ctx_sched_out(child_ctx, EVENT_ALL);
6761 * Take the context lock here so that if find_get_context is
6762 * reading child->perf_event_ctxp, we wait until it has
6763 * incremented the context's refcount before we do put_ctx below.
6765 raw_spin_lock(&child_ctx->lock);
6766 child->perf_event_ctxp[ctxn] = NULL;
6768 * If this context is a clone; unclone it so it can't get
6769 * swapped to another process while we're removing all
6770 * the events from it.
6772 unclone_ctx(child_ctx);
6773 update_context_time(child_ctx);
6774 raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
6777 * Report the task dead after unscheduling the events so that we
6778 * won't get any samples after PERF_RECORD_EXIT. We can however still
6779 * get a few PERF_RECORD_READ events.
6781 perf_event_task(child, child_ctx, 0);
6784 * We can recurse on the same lock type through:
6786 * __perf_event_exit_task()
6787 * sync_child_event()
6788 * fput(parent_event->filp)
6790 * mutex_lock(&ctx->mutex)
6792 * But since its the parent context it won't be the same instance.
6794 mutex_lock(&child_ctx->mutex);
6797 list_for_each_entry_safe(child_event, tmp, &child_ctx->pinned_groups,
6799 __perf_event_exit_task(child_event, child_ctx, child);
6801 list_for_each_entry_safe(child_event, tmp, &child_ctx->flexible_groups,
6803 __perf_event_exit_task(child_event, child_ctx, child);
6806 * If the last event was a group event, it will have appended all
6807 * its siblings to the list, but we obtained 'tmp' before that which
6808 * will still point to the list head terminating the iteration.
6810 if (!list_empty(&child_ctx->pinned_groups) ||
6811 !list_empty(&child_ctx->flexible_groups))
6814 mutex_unlock(&child_ctx->mutex);
6820 * When a child task exits, feed back event values to parent events.
6822 void perf_event_exit_task(struct task_struct *child)
6824 struct perf_event *event, *tmp;
6827 mutex_lock(&child->perf_event_mutex);
6828 list_for_each_entry_safe(event, tmp, &child->perf_event_list,
6830 list_del_init(&event->owner_entry);
6833 * Ensure the list deletion is visible before we clear
6834 * the owner, closes a race against perf_release() where
6835 * we need to serialize on the owner->perf_event_mutex.
6838 event->owner = NULL;
6840 mutex_unlock(&child->perf_event_mutex);
6842 for_each_task_context_nr(ctxn)
6843 perf_event_exit_task_context(child, ctxn);
6846 static void perf_free_event(struct perf_event *event,
6847 struct perf_event_context *ctx)
6849 struct perf_event *parent = event->parent;
6851 if (WARN_ON_ONCE(!parent))
6854 mutex_lock(&parent->child_mutex);
6855 list_del_init(&event->child_list);
6856 mutex_unlock(&parent->child_mutex);
6860 perf_group_detach(event);
6861 list_del_event(event, ctx);
6866 * free an unexposed, unused context as created by inheritance by
6867 * perf_event_init_task below, used by fork() in case of fail.
6869 void perf_event_free_task(struct task_struct *task)
6871 struct perf_event_context *ctx;
6872 struct perf_event *event, *tmp;
6875 for_each_task_context_nr(ctxn) {
6876 ctx = task->perf_event_ctxp[ctxn];
6880 mutex_lock(&ctx->mutex);
6882 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups,
6884 perf_free_event(event, ctx);
6886 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups,
6888 perf_free_event(event, ctx);
6890 if (!list_empty(&ctx->pinned_groups) ||
6891 !list_empty(&ctx->flexible_groups))
6894 mutex_unlock(&ctx->mutex);
6900 void perf_event_delayed_put(struct task_struct *task)
6904 for_each_task_context_nr(ctxn)
6905 WARN_ON_ONCE(task->perf_event_ctxp[ctxn]);
6909 * inherit a event from parent task to child task:
6911 static struct perf_event *
6912 inherit_event(struct perf_event *parent_event,
6913 struct task_struct *parent,
6914 struct perf_event_context *parent_ctx,
6915 struct task_struct *child,
6916 struct perf_event *group_leader,
6917 struct perf_event_context *child_ctx)
6919 struct perf_event *child_event;
6920 unsigned long flags;
6923 * Instead of creating recursive hierarchies of events,
6924 * we link inherited events back to the original parent,
6925 * which has a filp for sure, which we use as the reference
6928 if (parent_event->parent)
6929 parent_event = parent_event->parent;
6931 child_event = perf_event_alloc(&parent_event->attr,
6934 group_leader, parent_event,
6936 if (IS_ERR(child_event))
6941 * Make the child state follow the state of the parent event,
6942 * not its attr.disabled bit. We hold the parent's mutex,
6943 * so we won't race with perf_event_{en, dis}able_family.
6945 if (parent_event->state >= PERF_EVENT_STATE_INACTIVE)
6946 child_event->state = PERF_EVENT_STATE_INACTIVE;
6948 child_event->state = PERF_EVENT_STATE_OFF;
6950 if (parent_event->attr.freq) {
6951 u64 sample_period = parent_event->hw.sample_period;
6952 struct hw_perf_event *hwc = &child_event->hw;
6954 hwc->sample_period = sample_period;
6955 hwc->last_period = sample_period;
6957 local64_set(&hwc->period_left, sample_period);
6960 child_event->ctx = child_ctx;
6961 child_event->overflow_handler = parent_event->overflow_handler;
6964 * Precalculate sample_data sizes
6966 perf_event__header_size(child_event);
6967 perf_event__id_header_size(child_event);
6970 * Link it up in the child's context:
6972 raw_spin_lock_irqsave(&child_ctx->lock, flags);
6973 add_event_to_ctx(child_event, child_ctx);
6974 raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
6977 * Get a reference to the parent filp - we will fput it
6978 * when the child event exits. This is safe to do because
6979 * we are in the parent and we know that the filp still
6980 * exists and has a nonzero count:
6982 atomic_long_inc(&parent_event->filp->f_count);
6985 * Link this into the parent event's child list
6987 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
6988 mutex_lock(&parent_event->child_mutex);
6989 list_add_tail(&child_event->child_list, &parent_event->child_list);
6990 mutex_unlock(&parent_event->child_mutex);
6995 static int inherit_group(struct perf_event *parent_event,
6996 struct task_struct *parent,
6997 struct perf_event_context *parent_ctx,
6998 struct task_struct *child,
6999 struct perf_event_context *child_ctx)
7001 struct perf_event *leader;
7002 struct perf_event *sub;
7003 struct perf_event *child_ctr;
7005 leader = inherit_event(parent_event, parent, parent_ctx,
7006 child, NULL, child_ctx);
7008 return PTR_ERR(leader);
7009 list_for_each_entry(sub, &parent_event->sibling_list, group_entry) {
7010 child_ctr = inherit_event(sub, parent, parent_ctx,
7011 child, leader, child_ctx);
7012 if (IS_ERR(child_ctr))
7013 return PTR_ERR(child_ctr);
7019 inherit_task_group(struct perf_event *event, struct task_struct *parent,
7020 struct perf_event_context *parent_ctx,
7021 struct task_struct *child, int ctxn,
7025 struct perf_event_context *child_ctx;
7027 if (!event->attr.inherit) {
7032 child_ctx = child->perf_event_ctxp[ctxn];
7035 * This is executed from the parent task context, so
7036 * inherit events that have been marked for cloning.
7037 * First allocate and initialize a context for the
7041 child_ctx = alloc_perf_context(event->pmu, child);
7045 child->perf_event_ctxp[ctxn] = child_ctx;
7048 ret = inherit_group(event, parent, parent_ctx,
7058 * Initialize the perf_event context in task_struct
7060 int perf_event_init_context(struct task_struct *child, int ctxn)
7062 struct perf_event_context *child_ctx, *parent_ctx;
7063 struct perf_event_context *cloned_ctx;
7064 struct perf_event *event;
7065 struct task_struct *parent = current;
7066 int inherited_all = 1;
7067 unsigned long flags;
7070 if (likely(!parent->perf_event_ctxp[ctxn]))
7074 * If the parent's context is a clone, pin it so it won't get
7077 parent_ctx = perf_pin_task_context(parent, ctxn);
7080 * No need to check if parent_ctx != NULL here; since we saw
7081 * it non-NULL earlier, the only reason for it to become NULL
7082 * is if we exit, and since we're currently in the middle of
7083 * a fork we can't be exiting at the same time.
7087 * Lock the parent list. No need to lock the child - not PID
7088 * hashed yet and not running, so nobody can access it.
7090 mutex_lock(&parent_ctx->mutex);
7093 * We dont have to disable NMIs - we are only looking at
7094 * the list, not manipulating it:
7096 list_for_each_entry(event, &parent_ctx->pinned_groups, group_entry) {
7097 ret = inherit_task_group(event, parent, parent_ctx,
7098 child, ctxn, &inherited_all);
7104 * We can't hold ctx->lock when iterating the ->flexible_group list due
7105 * to allocations, but we need to prevent rotation because
7106 * rotate_ctx() will change the list from interrupt context.
7108 raw_spin_lock_irqsave(&parent_ctx->lock, flags);
7109 parent_ctx->rotate_disable = 1;
7110 raw_spin_unlock_irqrestore(&parent_ctx->lock, flags);
7112 list_for_each_entry(event, &parent_ctx->flexible_groups, group_entry) {
7113 ret = inherit_task_group(event, parent, parent_ctx,
7114 child, ctxn, &inherited_all);
7119 raw_spin_lock_irqsave(&parent_ctx->lock, flags);
7120 parent_ctx->rotate_disable = 0;
7122 child_ctx = child->perf_event_ctxp[ctxn];
7124 if (child_ctx && inherited_all) {
7126 * Mark the child context as a clone of the parent
7127 * context, or of whatever the parent is a clone of.
7129 * Note that if the parent is a clone, the holding of
7130 * parent_ctx->lock avoids it from being uncloned.
7132 cloned_ctx = parent_ctx->parent_ctx;
7134 child_ctx->parent_ctx = cloned_ctx;
7135 child_ctx->parent_gen = parent_ctx->parent_gen;
7137 child_ctx->parent_ctx = parent_ctx;
7138 child_ctx->parent_gen = parent_ctx->generation;
7140 get_ctx(child_ctx->parent_ctx);
7143 raw_spin_unlock_irqrestore(&parent_ctx->lock, flags);
7144 mutex_unlock(&parent_ctx->mutex);
7146 perf_unpin_context(parent_ctx);
7147 put_ctx(parent_ctx);
7153 * Initialize the perf_event context in task_struct
7155 int perf_event_init_task(struct task_struct *child)
7159 memset(child->perf_event_ctxp, 0, sizeof(child->perf_event_ctxp));
7160 mutex_init(&child->perf_event_mutex);
7161 INIT_LIST_HEAD(&child->perf_event_list);
7163 for_each_task_context_nr(ctxn) {
7164 ret = perf_event_init_context(child, ctxn);
7172 static void __init perf_event_init_all_cpus(void)
7174 struct swevent_htable *swhash;
7177 for_each_possible_cpu(cpu) {
7178 swhash = &per_cpu(swevent_htable, cpu);
7179 mutex_init(&swhash->hlist_mutex);
7180 INIT_LIST_HEAD(&per_cpu(rotation_list, cpu));
7184 static void __cpuinit perf_event_init_cpu(int cpu)
7186 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
7188 mutex_lock(&swhash->hlist_mutex);
7189 if (swhash->hlist_refcount > 0) {
7190 struct swevent_hlist *hlist;
7192 hlist = kzalloc_node(sizeof(*hlist), GFP_KERNEL, cpu_to_node(cpu));
7194 rcu_assign_pointer(swhash->swevent_hlist, hlist);
7196 mutex_unlock(&swhash->hlist_mutex);
7199 #if defined CONFIG_HOTPLUG_CPU || defined CONFIG_KEXEC
7200 static void perf_pmu_rotate_stop(struct pmu *pmu)
7202 struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
7204 WARN_ON(!irqs_disabled());
7206 list_del_init(&cpuctx->rotation_list);
7209 static void __perf_event_exit_context(void *__info)
7211 struct perf_event_context *ctx = __info;
7212 struct perf_event *event, *tmp;
7214 perf_pmu_rotate_stop(ctx->pmu);
7216 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry)
7217 __perf_remove_from_context(event);
7218 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups, group_entry)
7219 __perf_remove_from_context(event);
7222 static void perf_event_exit_cpu_context(int cpu)
7224 struct perf_event_context *ctx;
7228 idx = srcu_read_lock(&pmus_srcu);
7229 list_for_each_entry_rcu(pmu, &pmus, entry) {
7230 ctx = &per_cpu_ptr(pmu->pmu_cpu_context, cpu)->ctx;
7232 mutex_lock(&ctx->mutex);
7233 smp_call_function_single(cpu, __perf_event_exit_context, ctx, 1);
7234 mutex_unlock(&ctx->mutex);
7236 srcu_read_unlock(&pmus_srcu, idx);
7239 static void perf_event_exit_cpu(int cpu)
7241 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
7243 mutex_lock(&swhash->hlist_mutex);
7244 swevent_hlist_release(swhash);
7245 mutex_unlock(&swhash->hlist_mutex);
7247 perf_event_exit_cpu_context(cpu);
7250 static inline void perf_event_exit_cpu(int cpu) { }
7254 perf_reboot(struct notifier_block *notifier, unsigned long val, void *v)
7258 for_each_online_cpu(cpu)
7259 perf_event_exit_cpu(cpu);
7265 * Run the perf reboot notifier at the very last possible moment so that
7266 * the generic watchdog code runs as long as possible.
7268 static struct notifier_block perf_reboot_notifier = {
7269 .notifier_call = perf_reboot,
7270 .priority = INT_MIN,
7273 static int __cpuinit
7274 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
7276 unsigned int cpu = (long)hcpu;
7278 switch (action & ~CPU_TASKS_FROZEN) {
7280 case CPU_UP_PREPARE:
7281 case CPU_DOWN_FAILED:
7282 perf_event_init_cpu(cpu);
7285 case CPU_UP_CANCELED:
7286 case CPU_DOWN_PREPARE:
7287 perf_event_exit_cpu(cpu);
7297 void __init perf_event_init(void)
7303 perf_event_init_all_cpus();
7304 init_srcu_struct(&pmus_srcu);
7305 perf_pmu_register(&perf_swevent, "software", PERF_TYPE_SOFTWARE);
7306 perf_pmu_register(&perf_cpu_clock, NULL, -1);
7307 perf_pmu_register(&perf_task_clock, NULL, -1);
7309 perf_cpu_notifier(perf_cpu_notify);
7310 register_reboot_notifier(&perf_reboot_notifier);
7312 ret = init_hw_breakpoint();
7313 WARN(ret, "hw_breakpoint initialization failed with: %d", ret);
7316 static int __init perf_event_sysfs_init(void)
7321 mutex_lock(&pmus_lock);
7323 ret = bus_register(&pmu_bus);
7327 list_for_each_entry(pmu, &pmus, entry) {
7328 if (!pmu->name || pmu->type < 0)
7331 ret = pmu_dev_alloc(pmu);
7332 WARN(ret, "Failed to register pmu: %s, reason %d\n", pmu->name, ret);
7334 pmu_bus_running = 1;
7338 mutex_unlock(&pmus_lock);
7342 device_initcall(perf_event_sysfs_init);
7344 #ifdef CONFIG_CGROUP_PERF
7345 static struct cgroup_subsys_state *perf_cgroup_create(
7346 struct cgroup_subsys *ss, struct cgroup *cont)
7348 struct perf_cgroup *jc;
7349 struct perf_cgroup_info *t;
7352 jc = kmalloc(sizeof(*jc), GFP_KERNEL);
7354 return ERR_PTR(-ENOMEM);
7356 memset(jc, 0, sizeof(*jc));
7358 jc->info = alloc_percpu(struct perf_cgroup_info);
7361 return ERR_PTR(-ENOMEM);
7364 for_each_possible_cpu(c) {
7365 t = per_cpu_ptr(jc->info, c);
7372 static void perf_cgroup_destroy(struct cgroup_subsys *ss,
7373 struct cgroup *cont)
7375 struct perf_cgroup *jc;
7376 jc = container_of(cgroup_subsys_state(cont, perf_subsys_id),
7377 struct perf_cgroup, css);
7378 free_percpu(jc->info);
7382 static int __perf_cgroup_move(void *info)
7384 struct task_struct *task = info;
7385 perf_cgroup_switch(task, PERF_CGROUP_SWOUT | PERF_CGROUP_SWIN);
7389 static void perf_cgroup_move(struct task_struct *task)
7391 task_function_call(task, __perf_cgroup_move, task);
7394 static void perf_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
7395 struct cgroup *old_cgrp, struct task_struct *task,
7398 perf_cgroup_move(task);
7400 struct task_struct *c;
7402 list_for_each_entry_rcu(c, &task->thread_group, thread_group) {
7403 perf_cgroup_move(c);
7409 static void perf_cgroup_exit(struct cgroup_subsys *ss, struct cgroup *cgrp,
7410 struct cgroup *old_cgrp, struct task_struct *task)
7413 * cgroup_exit() is called in the copy_process() failure path.
7414 * Ignore this case since the task hasn't ran yet, this avoids
7415 * trying to poke a half freed task state from generic code.
7417 if (!(task->flags & PF_EXITING))
7420 perf_cgroup_move(task);
7423 struct cgroup_subsys perf_subsys = {
7424 .name = "perf_event",
7425 .subsys_id = perf_subsys_id,
7426 .create = perf_cgroup_create,
7427 .destroy = perf_cgroup_destroy,
7428 .exit = perf_cgroup_exit,
7429 .attach = perf_cgroup_attach,
7431 #endif /* CONFIG_CGROUP_PERF */