2 * Performance events core code:
4 * Copyright (C) 2008 Thomas Gleixner <tglx@linutronix.de>
5 * Copyright (C) 2008-2011 Red Hat, Inc., Ingo Molnar
6 * Copyright (C) 2008-2011 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
7 * Copyright © 2009 Paul Mackerras, IBM Corp. <paulus@au1.ibm.com>
9 * For licensing details see kernel-base/COPYING
14 #include <linux/cpu.h>
15 #include <linux/smp.h>
16 #include <linux/idr.h>
17 #include <linux/file.h>
18 #include <linux/poll.h>
19 #include <linux/slab.h>
20 #include <linux/hash.h>
21 #include <linux/sysfs.h>
22 #include <linux/dcache.h>
23 #include <linux/percpu.h>
24 #include <linux/ptrace.h>
25 #include <linux/reboot.h>
26 #include <linux/vmstat.h>
27 #include <linux/device.h>
28 #include <linux/export.h>
29 #include <linux/vmalloc.h>
30 #include <linux/hardirq.h>
31 #include <linux/rculist.h>
32 #include <linux/uaccess.h>
33 #include <linux/syscalls.h>
34 #include <linux/anon_inodes.h>
35 #include <linux/kernel_stat.h>
36 #include <linux/perf_event.h>
37 #include <linux/ftrace_event.h>
38 #include <linux/hw_breakpoint.h>
42 #include <asm/irq_regs.h>
44 struct remote_function_call {
45 struct task_struct *p;
46 int (*func)(void *info);
51 static void remote_function(void *data)
53 struct remote_function_call *tfc = data;
54 struct task_struct *p = tfc->p;
58 if (task_cpu(p) != smp_processor_id() || !task_curr(p))
62 tfc->ret = tfc->func(tfc->info);
66 * task_function_call - call a function on the cpu on which a task runs
67 * @p: the task to evaluate
68 * @func: the function to be called
69 * @info: the function call argument
71 * Calls the function @func when the task is currently running. This might
72 * be on the current CPU, which just calls the function directly
74 * returns: @func return value, or
75 * -ESRCH - when the process isn't running
76 * -EAGAIN - when the process moved away
79 task_function_call(struct task_struct *p, int (*func) (void *info), void *info)
81 struct remote_function_call data = {
85 .ret = -ESRCH, /* No such (running) process */
89 smp_call_function_single(task_cpu(p), remote_function, &data, 1);
95 * cpu_function_call - call a function on the cpu
96 * @func: the function to be called
97 * @info: the function call argument
99 * Calls the function @func on the remote cpu.
101 * returns: @func return value or -ENXIO when the cpu is offline
103 static int cpu_function_call(int cpu, int (*func) (void *info), void *info)
105 struct remote_function_call data = {
109 .ret = -ENXIO, /* No such CPU */
112 smp_call_function_single(cpu, remote_function, &data, 1);
117 #define PERF_FLAG_ALL (PERF_FLAG_FD_NO_GROUP |\
118 PERF_FLAG_FD_OUTPUT |\
119 PERF_FLAG_PID_CGROUP)
122 EVENT_FLEXIBLE = 0x1,
124 EVENT_ALL = EVENT_FLEXIBLE | EVENT_PINNED,
128 * perf_sched_events : >0 events exist
129 * perf_cgroup_events: >0 per-cpu cgroup events exist on this cpu
131 struct jump_label_key_deferred perf_sched_events __read_mostly;
132 static DEFINE_PER_CPU(atomic_t, perf_cgroup_events);
134 static atomic_t nr_mmap_events __read_mostly;
135 static atomic_t nr_comm_events __read_mostly;
136 static atomic_t nr_task_events __read_mostly;
138 static LIST_HEAD(pmus);
139 static DEFINE_MUTEX(pmus_lock);
140 static struct srcu_struct pmus_srcu;
143 * perf event paranoia level:
144 * -1 - not paranoid at all
145 * 0 - disallow raw tracepoint access for unpriv
146 * 1 - disallow cpu events for unpriv
147 * 2 - disallow kernel profiling for unpriv
149 int sysctl_perf_event_paranoid __read_mostly = 1;
151 /* Minimum for 512 kiB + 1 user control page */
152 int sysctl_perf_event_mlock __read_mostly = 512 + (PAGE_SIZE / 1024); /* 'free' kiB per user */
155 * max perf event sample rate
157 #define DEFAULT_MAX_SAMPLE_RATE 100000
158 int sysctl_perf_event_sample_rate __read_mostly = DEFAULT_MAX_SAMPLE_RATE;
159 static int max_samples_per_tick __read_mostly =
160 DIV_ROUND_UP(DEFAULT_MAX_SAMPLE_RATE, HZ);
162 int perf_proc_update_handler(struct ctl_table *table, int write,
163 void __user *buffer, size_t *lenp,
166 int ret = proc_dointvec(table, write, buffer, lenp, ppos);
171 max_samples_per_tick = DIV_ROUND_UP(sysctl_perf_event_sample_rate, HZ);
176 static atomic64_t perf_event_id;
178 static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
179 enum event_type_t event_type);
181 static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
182 enum event_type_t event_type,
183 struct task_struct *task);
185 static void update_context_time(struct perf_event_context *ctx);
186 static u64 perf_event_time(struct perf_event *event);
188 static void ring_buffer_attach(struct perf_event *event,
189 struct ring_buffer *rb);
191 void __weak perf_event_print_debug(void) { }
193 extern __weak const char *perf_pmu_name(void)
198 static inline u64 perf_clock(void)
200 return local_clock();
203 static inline struct perf_cpu_context *
204 __get_cpu_context(struct perf_event_context *ctx)
206 return this_cpu_ptr(ctx->pmu->pmu_cpu_context);
209 static void perf_ctx_lock(struct perf_cpu_context *cpuctx,
210 struct perf_event_context *ctx)
212 raw_spin_lock(&cpuctx->ctx.lock);
214 raw_spin_lock(&ctx->lock);
217 static void perf_ctx_unlock(struct perf_cpu_context *cpuctx,
218 struct perf_event_context *ctx)
221 raw_spin_unlock(&ctx->lock);
222 raw_spin_unlock(&cpuctx->ctx.lock);
225 #ifdef CONFIG_CGROUP_PERF
228 * Must ensure cgroup is pinned (css_get) before calling
229 * this function. In other words, we cannot call this function
230 * if there is no cgroup event for the current CPU context.
232 static inline struct perf_cgroup *
233 perf_cgroup_from_task(struct task_struct *task)
235 return container_of(task_subsys_state(task, perf_subsys_id),
236 struct perf_cgroup, css);
240 perf_cgroup_match(struct perf_event *event)
242 struct perf_event_context *ctx = event->ctx;
243 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
245 return !event->cgrp || event->cgrp == cpuctx->cgrp;
248 static inline void perf_get_cgroup(struct perf_event *event)
250 css_get(&event->cgrp->css);
253 static inline void perf_put_cgroup(struct perf_event *event)
255 css_put(&event->cgrp->css);
258 static inline void perf_detach_cgroup(struct perf_event *event)
260 perf_put_cgroup(event);
264 static inline int is_cgroup_event(struct perf_event *event)
266 return event->cgrp != NULL;
269 static inline u64 perf_cgroup_event_time(struct perf_event *event)
271 struct perf_cgroup_info *t;
273 t = per_cpu_ptr(event->cgrp->info, event->cpu);
277 static inline void __update_cgrp_time(struct perf_cgroup *cgrp)
279 struct perf_cgroup_info *info;
284 info = this_cpu_ptr(cgrp->info);
286 info->time += now - info->timestamp;
287 info->timestamp = now;
290 static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx)
292 struct perf_cgroup *cgrp_out = cpuctx->cgrp;
294 __update_cgrp_time(cgrp_out);
297 static inline void update_cgrp_time_from_event(struct perf_event *event)
299 struct perf_cgroup *cgrp;
302 * ensure we access cgroup data only when needed and
303 * when we know the cgroup is pinned (css_get)
305 if (!is_cgroup_event(event))
308 cgrp = perf_cgroup_from_task(current);
310 * Do not update time when cgroup is not active
312 if (cgrp == event->cgrp)
313 __update_cgrp_time(event->cgrp);
317 perf_cgroup_set_timestamp(struct task_struct *task,
318 struct perf_event_context *ctx)
320 struct perf_cgroup *cgrp;
321 struct perf_cgroup_info *info;
324 * ctx->lock held by caller
325 * ensure we do not access cgroup data
326 * unless we have the cgroup pinned (css_get)
328 if (!task || !ctx->nr_cgroups)
331 cgrp = perf_cgroup_from_task(task);
332 info = this_cpu_ptr(cgrp->info);
333 info->timestamp = ctx->timestamp;
336 #define PERF_CGROUP_SWOUT 0x1 /* cgroup switch out every event */
337 #define PERF_CGROUP_SWIN 0x2 /* cgroup switch in events based on task */
340 * reschedule events based on the cgroup constraint of task.
342 * mode SWOUT : schedule out everything
343 * mode SWIN : schedule in based on cgroup for next
345 void perf_cgroup_switch(struct task_struct *task, int mode)
347 struct perf_cpu_context *cpuctx;
352 * disable interrupts to avoid geting nr_cgroup
353 * changes via __perf_event_disable(). Also
356 local_irq_save(flags);
359 * we reschedule only in the presence of cgroup
360 * constrained events.
364 list_for_each_entry_rcu(pmu, &pmus, entry) {
365 cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
368 * perf_cgroup_events says at least one
369 * context on this CPU has cgroup events.
371 * ctx->nr_cgroups reports the number of cgroup
372 * events for a context.
374 if (cpuctx->ctx.nr_cgroups > 0) {
375 perf_ctx_lock(cpuctx, cpuctx->task_ctx);
376 perf_pmu_disable(cpuctx->ctx.pmu);
378 if (mode & PERF_CGROUP_SWOUT) {
379 cpu_ctx_sched_out(cpuctx, EVENT_ALL);
381 * must not be done before ctxswout due
382 * to event_filter_match() in event_sched_out()
387 if (mode & PERF_CGROUP_SWIN) {
388 WARN_ON_ONCE(cpuctx->cgrp);
389 /* set cgrp before ctxsw in to
390 * allow event_filter_match() to not
391 * have to pass task around
393 cpuctx->cgrp = perf_cgroup_from_task(task);
394 cpu_ctx_sched_in(cpuctx, EVENT_ALL, task);
396 perf_pmu_enable(cpuctx->ctx.pmu);
397 perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
403 local_irq_restore(flags);
406 static inline void perf_cgroup_sched_out(struct task_struct *task,
407 struct task_struct *next)
409 struct perf_cgroup *cgrp1;
410 struct perf_cgroup *cgrp2 = NULL;
413 * we come here when we know perf_cgroup_events > 0
415 cgrp1 = perf_cgroup_from_task(task);
418 * next is NULL when called from perf_event_enable_on_exec()
419 * that will systematically cause a cgroup_switch()
422 cgrp2 = perf_cgroup_from_task(next);
425 * only schedule out current cgroup events if we know
426 * that we are switching to a different cgroup. Otherwise,
427 * do no touch the cgroup events.
430 perf_cgroup_switch(task, PERF_CGROUP_SWOUT);
433 static inline void perf_cgroup_sched_in(struct task_struct *prev,
434 struct task_struct *task)
436 struct perf_cgroup *cgrp1;
437 struct perf_cgroup *cgrp2 = NULL;
440 * we come here when we know perf_cgroup_events > 0
442 cgrp1 = perf_cgroup_from_task(task);
444 /* prev can never be NULL */
445 cgrp2 = perf_cgroup_from_task(prev);
448 * only need to schedule in cgroup events if we are changing
449 * cgroup during ctxsw. Cgroup events were not scheduled
450 * out of ctxsw out if that was not the case.
453 perf_cgroup_switch(task, PERF_CGROUP_SWIN);
456 static inline int perf_cgroup_connect(int fd, struct perf_event *event,
457 struct perf_event_attr *attr,
458 struct perf_event *group_leader)
460 struct perf_cgroup *cgrp;
461 struct cgroup_subsys_state *css;
463 int ret = 0, fput_needed;
465 file = fget_light(fd, &fput_needed);
469 css = cgroup_css_from_dir(file, perf_subsys_id);
475 cgrp = container_of(css, struct perf_cgroup, css);
478 /* must be done before we fput() the file */
479 perf_get_cgroup(event);
482 * all events in a group must monitor
483 * the same cgroup because a task belongs
484 * to only one perf cgroup at a time
486 if (group_leader && group_leader->cgrp != cgrp) {
487 perf_detach_cgroup(event);
491 fput_light(file, fput_needed);
496 perf_cgroup_set_shadow_time(struct perf_event *event, u64 now)
498 struct perf_cgroup_info *t;
499 t = per_cpu_ptr(event->cgrp->info, event->cpu);
500 event->shadow_ctx_time = now - t->timestamp;
504 perf_cgroup_defer_enabled(struct perf_event *event)
507 * when the current task's perf cgroup does not match
508 * the event's, we need to remember to call the
509 * perf_mark_enable() function the first time a task with
510 * a matching perf cgroup is scheduled in.
512 if (is_cgroup_event(event) && !perf_cgroup_match(event))
513 event->cgrp_defer_enabled = 1;
517 perf_cgroup_mark_enabled(struct perf_event *event,
518 struct perf_event_context *ctx)
520 struct perf_event *sub;
521 u64 tstamp = perf_event_time(event);
523 if (!event->cgrp_defer_enabled)
526 event->cgrp_defer_enabled = 0;
528 event->tstamp_enabled = tstamp - event->total_time_enabled;
529 list_for_each_entry(sub, &event->sibling_list, group_entry) {
530 if (sub->state >= PERF_EVENT_STATE_INACTIVE) {
531 sub->tstamp_enabled = tstamp - sub->total_time_enabled;
532 sub->cgrp_defer_enabled = 0;
536 #else /* !CONFIG_CGROUP_PERF */
539 perf_cgroup_match(struct perf_event *event)
544 static inline void perf_detach_cgroup(struct perf_event *event)
547 static inline int is_cgroup_event(struct perf_event *event)
552 static inline u64 perf_cgroup_event_cgrp_time(struct perf_event *event)
557 static inline void update_cgrp_time_from_event(struct perf_event *event)
561 static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx)
565 static inline void perf_cgroup_sched_out(struct task_struct *task,
566 struct task_struct *next)
570 static inline void perf_cgroup_sched_in(struct task_struct *prev,
571 struct task_struct *task)
575 static inline int perf_cgroup_connect(pid_t pid, struct perf_event *event,
576 struct perf_event_attr *attr,
577 struct perf_event *group_leader)
583 perf_cgroup_set_timestamp(struct task_struct *task,
584 struct perf_event_context *ctx)
589 perf_cgroup_switch(struct task_struct *task, struct task_struct *next)
594 perf_cgroup_set_shadow_time(struct perf_event *event, u64 now)
598 static inline u64 perf_cgroup_event_time(struct perf_event *event)
604 perf_cgroup_defer_enabled(struct perf_event *event)
609 perf_cgroup_mark_enabled(struct perf_event *event,
610 struct perf_event_context *ctx)
615 void perf_pmu_disable(struct pmu *pmu)
617 int *count = this_cpu_ptr(pmu->pmu_disable_count);
619 pmu->pmu_disable(pmu);
622 void perf_pmu_enable(struct pmu *pmu)
624 int *count = this_cpu_ptr(pmu->pmu_disable_count);
626 pmu->pmu_enable(pmu);
629 static DEFINE_PER_CPU(struct list_head, rotation_list);
632 * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
633 * because they're strictly cpu affine and rotate_start is called with IRQs
634 * disabled, while rotate_context is called from IRQ context.
636 static void perf_pmu_rotate_start(struct pmu *pmu)
638 struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
639 struct list_head *head = &__get_cpu_var(rotation_list);
641 WARN_ON(!irqs_disabled());
643 if (list_empty(&cpuctx->rotation_list))
644 list_add(&cpuctx->rotation_list, head);
647 static void get_ctx(struct perf_event_context *ctx)
649 WARN_ON(!atomic_inc_not_zero(&ctx->refcount));
652 static void put_ctx(struct perf_event_context *ctx)
654 if (atomic_dec_and_test(&ctx->refcount)) {
656 put_ctx(ctx->parent_ctx);
658 put_task_struct(ctx->task);
659 kfree_rcu(ctx, rcu_head);
663 static void unclone_ctx(struct perf_event_context *ctx)
665 if (ctx->parent_ctx) {
666 put_ctx(ctx->parent_ctx);
667 ctx->parent_ctx = NULL;
671 static u32 perf_event_pid(struct perf_event *event, struct task_struct *p)
674 * only top level events have the pid namespace they were created in
677 event = event->parent;
679 return task_tgid_nr_ns(p, event->ns);
682 static u32 perf_event_tid(struct perf_event *event, struct task_struct *p)
685 * only top level events have the pid namespace they were created in
688 event = event->parent;
690 return task_pid_nr_ns(p, event->ns);
694 * If we inherit events we want to return the parent event id
697 static u64 primary_event_id(struct perf_event *event)
702 id = event->parent->id;
708 * Get the perf_event_context for a task and lock it.
709 * This has to cope with with the fact that until it is locked,
710 * the context could get moved to another task.
712 static struct perf_event_context *
713 perf_lock_task_context(struct task_struct *task, int ctxn, unsigned long *flags)
715 struct perf_event_context *ctx;
719 ctx = rcu_dereference(task->perf_event_ctxp[ctxn]);
722 * If this context is a clone of another, it might
723 * get swapped for another underneath us by
724 * perf_event_task_sched_out, though the
725 * rcu_read_lock() protects us from any context
726 * getting freed. Lock the context and check if it
727 * got swapped before we could get the lock, and retry
728 * if so. If we locked the right context, then it
729 * can't get swapped on us any more.
731 raw_spin_lock_irqsave(&ctx->lock, *flags);
732 if (ctx != rcu_dereference(task->perf_event_ctxp[ctxn])) {
733 raw_spin_unlock_irqrestore(&ctx->lock, *flags);
737 if (!atomic_inc_not_zero(&ctx->refcount)) {
738 raw_spin_unlock_irqrestore(&ctx->lock, *flags);
747 * Get the context for a task and increment its pin_count so it
748 * can't get swapped to another task. This also increments its
749 * reference count so that the context can't get freed.
751 static struct perf_event_context *
752 perf_pin_task_context(struct task_struct *task, int ctxn)
754 struct perf_event_context *ctx;
757 ctx = perf_lock_task_context(task, ctxn, &flags);
760 raw_spin_unlock_irqrestore(&ctx->lock, flags);
765 static void perf_unpin_context(struct perf_event_context *ctx)
769 raw_spin_lock_irqsave(&ctx->lock, flags);
771 raw_spin_unlock_irqrestore(&ctx->lock, flags);
775 * Update the record of the current time in a context.
777 static void update_context_time(struct perf_event_context *ctx)
779 u64 now = perf_clock();
781 ctx->time += now - ctx->timestamp;
782 ctx->timestamp = now;
785 static u64 perf_event_time(struct perf_event *event)
787 struct perf_event_context *ctx = event->ctx;
789 if (is_cgroup_event(event))
790 return perf_cgroup_event_time(event);
792 return ctx ? ctx->time : 0;
796 * Update the total_time_enabled and total_time_running fields for a event.
797 * The caller of this function needs to hold the ctx->lock.
799 static void update_event_times(struct perf_event *event)
801 struct perf_event_context *ctx = event->ctx;
804 if (event->state < PERF_EVENT_STATE_INACTIVE ||
805 event->group_leader->state < PERF_EVENT_STATE_INACTIVE)
808 * in cgroup mode, time_enabled represents
809 * the time the event was enabled AND active
810 * tasks were in the monitored cgroup. This is
811 * independent of the activity of the context as
812 * there may be a mix of cgroup and non-cgroup events.
814 * That is why we treat cgroup events differently
817 if (is_cgroup_event(event))
818 run_end = perf_cgroup_event_time(event);
819 else if (ctx->is_active)
822 run_end = event->tstamp_stopped;
824 event->total_time_enabled = run_end - event->tstamp_enabled;
826 if (event->state == PERF_EVENT_STATE_INACTIVE)
827 run_end = event->tstamp_stopped;
829 run_end = perf_event_time(event);
831 event->total_time_running = run_end - event->tstamp_running;
836 * Update total_time_enabled and total_time_running for all events in a group.
838 static void update_group_times(struct perf_event *leader)
840 struct perf_event *event;
842 update_event_times(leader);
843 list_for_each_entry(event, &leader->sibling_list, group_entry)
844 update_event_times(event);
847 static struct list_head *
848 ctx_group_list(struct perf_event *event, struct perf_event_context *ctx)
850 if (event->attr.pinned)
851 return &ctx->pinned_groups;
853 return &ctx->flexible_groups;
857 * Add a event from the lists for its context.
858 * Must be called with ctx->mutex and ctx->lock held.
861 list_add_event(struct perf_event *event, struct perf_event_context *ctx)
863 WARN_ON_ONCE(event->attach_state & PERF_ATTACH_CONTEXT);
864 event->attach_state |= PERF_ATTACH_CONTEXT;
867 * If we're a stand alone event or group leader, we go to the context
868 * list, group events are kept attached to the group so that
869 * perf_group_detach can, at all times, locate all siblings.
871 if (event->group_leader == event) {
872 struct list_head *list;
874 if (is_software_event(event))
875 event->group_flags |= PERF_GROUP_SOFTWARE;
877 list = ctx_group_list(event, ctx);
878 list_add_tail(&event->group_entry, list);
881 if (is_cgroup_event(event))
884 list_add_rcu(&event->event_entry, &ctx->event_list);
886 perf_pmu_rotate_start(ctx->pmu);
888 if (event->attr.inherit_stat)
893 * Called at perf_event creation and when events are attached/detached from a
896 static void perf_event__read_size(struct perf_event *event)
898 int entry = sizeof(u64); /* value */
902 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
905 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
908 if (event->attr.read_format & PERF_FORMAT_ID)
909 entry += sizeof(u64);
911 if (event->attr.read_format & PERF_FORMAT_GROUP) {
912 nr += event->group_leader->nr_siblings;
917 event->read_size = size;
920 static void perf_event__header_size(struct perf_event *event)
922 struct perf_sample_data *data;
923 u64 sample_type = event->attr.sample_type;
926 perf_event__read_size(event);
928 if (sample_type & PERF_SAMPLE_IP)
929 size += sizeof(data->ip);
931 if (sample_type & PERF_SAMPLE_ADDR)
932 size += sizeof(data->addr);
934 if (sample_type & PERF_SAMPLE_PERIOD)
935 size += sizeof(data->period);
937 if (sample_type & PERF_SAMPLE_READ)
938 size += event->read_size;
940 event->header_size = size;
943 static void perf_event__id_header_size(struct perf_event *event)
945 struct perf_sample_data *data;
946 u64 sample_type = event->attr.sample_type;
949 if (sample_type & PERF_SAMPLE_TID)
950 size += sizeof(data->tid_entry);
952 if (sample_type & PERF_SAMPLE_TIME)
953 size += sizeof(data->time);
955 if (sample_type & PERF_SAMPLE_ID)
956 size += sizeof(data->id);
958 if (sample_type & PERF_SAMPLE_STREAM_ID)
959 size += sizeof(data->stream_id);
961 if (sample_type & PERF_SAMPLE_CPU)
962 size += sizeof(data->cpu_entry);
964 event->id_header_size = size;
967 static void perf_group_attach(struct perf_event *event)
969 struct perf_event *group_leader = event->group_leader, *pos;
972 * We can have double attach due to group movement in perf_event_open.
974 if (event->attach_state & PERF_ATTACH_GROUP)
977 event->attach_state |= PERF_ATTACH_GROUP;
979 if (group_leader == event)
982 if (group_leader->group_flags & PERF_GROUP_SOFTWARE &&
983 !is_software_event(event))
984 group_leader->group_flags &= ~PERF_GROUP_SOFTWARE;
986 list_add_tail(&event->group_entry, &group_leader->sibling_list);
987 group_leader->nr_siblings++;
989 perf_event__header_size(group_leader);
991 list_for_each_entry(pos, &group_leader->sibling_list, group_entry)
992 perf_event__header_size(pos);
996 * Remove a event from the lists for its context.
997 * Must be called with ctx->mutex and ctx->lock held.
1000 list_del_event(struct perf_event *event, struct perf_event_context *ctx)
1002 struct perf_cpu_context *cpuctx;
1004 * We can have double detach due to exit/hot-unplug + close.
1006 if (!(event->attach_state & PERF_ATTACH_CONTEXT))
1009 event->attach_state &= ~PERF_ATTACH_CONTEXT;
1011 if (is_cgroup_event(event)) {
1013 cpuctx = __get_cpu_context(ctx);
1015 * if there are no more cgroup events
1016 * then cler cgrp to avoid stale pointer
1017 * in update_cgrp_time_from_cpuctx()
1019 if (!ctx->nr_cgroups)
1020 cpuctx->cgrp = NULL;
1024 if (event->attr.inherit_stat)
1027 list_del_rcu(&event->event_entry);
1029 if (event->group_leader == event)
1030 list_del_init(&event->group_entry);
1032 update_group_times(event);
1035 * If event was in error state, then keep it
1036 * that way, otherwise bogus counts will be
1037 * returned on read(). The only way to get out
1038 * of error state is by explicit re-enabling
1041 if (event->state > PERF_EVENT_STATE_OFF)
1042 event->state = PERF_EVENT_STATE_OFF;
1045 static void perf_group_detach(struct perf_event *event)
1047 struct perf_event *sibling, *tmp;
1048 struct list_head *list = NULL;
1051 * We can have double detach due to exit/hot-unplug + close.
1053 if (!(event->attach_state & PERF_ATTACH_GROUP))
1056 event->attach_state &= ~PERF_ATTACH_GROUP;
1059 * If this is a sibling, remove it from its group.
1061 if (event->group_leader != event) {
1062 list_del_init(&event->group_entry);
1063 event->group_leader->nr_siblings--;
1067 if (!list_empty(&event->group_entry))
1068 list = &event->group_entry;
1071 * If this was a group event with sibling events then
1072 * upgrade the siblings to singleton events by adding them
1073 * to whatever list we are on.
1075 list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) {
1077 list_move_tail(&sibling->group_entry, list);
1078 sibling->group_leader = sibling;
1080 /* Inherit group flags from the previous leader */
1081 sibling->group_flags = event->group_flags;
1085 perf_event__header_size(event->group_leader);
1087 list_for_each_entry(tmp, &event->group_leader->sibling_list, group_entry)
1088 perf_event__header_size(tmp);
1092 event_filter_match(struct perf_event *event)
1094 return (event->cpu == -1 || event->cpu == smp_processor_id())
1095 && perf_cgroup_match(event);
1099 event_sched_out(struct perf_event *event,
1100 struct perf_cpu_context *cpuctx,
1101 struct perf_event_context *ctx)
1103 u64 tstamp = perf_event_time(event);
1106 * An event which could not be activated because of
1107 * filter mismatch still needs to have its timings
1108 * maintained, otherwise bogus information is return
1109 * via read() for time_enabled, time_running:
1111 if (event->state == PERF_EVENT_STATE_INACTIVE
1112 && !event_filter_match(event)) {
1113 delta = tstamp - event->tstamp_stopped;
1114 event->tstamp_running += delta;
1115 event->tstamp_stopped = tstamp;
1118 if (event->state != PERF_EVENT_STATE_ACTIVE)
1121 event->state = PERF_EVENT_STATE_INACTIVE;
1122 if (event->pending_disable) {
1123 event->pending_disable = 0;
1124 event->state = PERF_EVENT_STATE_OFF;
1126 event->tstamp_stopped = tstamp;
1127 event->pmu->del(event, 0);
1130 if (!is_software_event(event))
1131 cpuctx->active_oncpu--;
1133 if (event->attr.freq && event->attr.sample_freq)
1135 if (event->attr.exclusive || !cpuctx->active_oncpu)
1136 cpuctx->exclusive = 0;
1140 group_sched_out(struct perf_event *group_event,
1141 struct perf_cpu_context *cpuctx,
1142 struct perf_event_context *ctx)
1144 struct perf_event *event;
1145 int state = group_event->state;
1147 event_sched_out(group_event, cpuctx, ctx);
1150 * Schedule out siblings (if any):
1152 list_for_each_entry(event, &group_event->sibling_list, group_entry)
1153 event_sched_out(event, cpuctx, ctx);
1155 if (state == PERF_EVENT_STATE_ACTIVE && group_event->attr.exclusive)
1156 cpuctx->exclusive = 0;
1160 * Cross CPU call to remove a performance event
1162 * We disable the event on the hardware level first. After that we
1163 * remove it from the context list.
1165 static int __perf_remove_from_context(void *info)
1167 struct perf_event *event = info;
1168 struct perf_event_context *ctx = event->ctx;
1169 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1171 raw_spin_lock(&ctx->lock);
1172 event_sched_out(event, cpuctx, ctx);
1173 list_del_event(event, ctx);
1174 if (!ctx->nr_events && cpuctx->task_ctx == ctx) {
1176 cpuctx->task_ctx = NULL;
1178 raw_spin_unlock(&ctx->lock);
1185 * Remove the event from a task's (or a CPU's) list of events.
1187 * CPU events are removed with a smp call. For task events we only
1188 * call when the task is on a CPU.
1190 * If event->ctx is a cloned context, callers must make sure that
1191 * every task struct that event->ctx->task could possibly point to
1192 * remains valid. This is OK when called from perf_release since
1193 * that only calls us on the top-level context, which can't be a clone.
1194 * When called from perf_event_exit_task, it's OK because the
1195 * context has been detached from its task.
1197 static void perf_remove_from_context(struct perf_event *event)
1199 struct perf_event_context *ctx = event->ctx;
1200 struct task_struct *task = ctx->task;
1202 lockdep_assert_held(&ctx->mutex);
1206 * Per cpu events are removed via an smp call and
1207 * the removal is always successful.
1209 cpu_function_call(event->cpu, __perf_remove_from_context, event);
1214 if (!task_function_call(task, __perf_remove_from_context, event))
1217 raw_spin_lock_irq(&ctx->lock);
1219 * If we failed to find a running task, but find the context active now
1220 * that we've acquired the ctx->lock, retry.
1222 if (ctx->is_active) {
1223 raw_spin_unlock_irq(&ctx->lock);
1228 * Since the task isn't running, its safe to remove the event, us
1229 * holding the ctx->lock ensures the task won't get scheduled in.
1231 list_del_event(event, ctx);
1232 raw_spin_unlock_irq(&ctx->lock);
1236 * Cross CPU call to disable a performance event
1238 static int __perf_event_disable(void *info)
1240 struct perf_event *event = info;
1241 struct perf_event_context *ctx = event->ctx;
1242 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1245 * If this is a per-task event, need to check whether this
1246 * event's task is the current task on this cpu.
1248 * Can trigger due to concurrent perf_event_context_sched_out()
1249 * flipping contexts around.
1251 if (ctx->task && cpuctx->task_ctx != ctx)
1254 raw_spin_lock(&ctx->lock);
1257 * If the event is on, turn it off.
1258 * If it is in error state, leave it in error state.
1260 if (event->state >= PERF_EVENT_STATE_INACTIVE) {
1261 update_context_time(ctx);
1262 update_cgrp_time_from_event(event);
1263 update_group_times(event);
1264 if (event == event->group_leader)
1265 group_sched_out(event, cpuctx, ctx);
1267 event_sched_out(event, cpuctx, ctx);
1268 event->state = PERF_EVENT_STATE_OFF;
1271 raw_spin_unlock(&ctx->lock);
1279 * If event->ctx is a cloned context, callers must make sure that
1280 * every task struct that event->ctx->task could possibly point to
1281 * remains valid. This condition is satisifed when called through
1282 * perf_event_for_each_child or perf_event_for_each because they
1283 * hold the top-level event's child_mutex, so any descendant that
1284 * goes to exit will block in sync_child_event.
1285 * When called from perf_pending_event it's OK because event->ctx
1286 * is the current context on this CPU and preemption is disabled,
1287 * hence we can't get into perf_event_task_sched_out for this context.
1289 void perf_event_disable(struct perf_event *event)
1291 struct perf_event_context *ctx = event->ctx;
1292 struct task_struct *task = ctx->task;
1296 * Disable the event on the cpu that it's on
1298 cpu_function_call(event->cpu, __perf_event_disable, event);
1303 if (!task_function_call(task, __perf_event_disable, event))
1306 raw_spin_lock_irq(&ctx->lock);
1308 * If the event is still active, we need to retry the cross-call.
1310 if (event->state == PERF_EVENT_STATE_ACTIVE) {
1311 raw_spin_unlock_irq(&ctx->lock);
1313 * Reload the task pointer, it might have been changed by
1314 * a concurrent perf_event_context_sched_out().
1321 * Since we have the lock this context can't be scheduled
1322 * in, so we can change the state safely.
1324 if (event->state == PERF_EVENT_STATE_INACTIVE) {
1325 update_group_times(event);
1326 event->state = PERF_EVENT_STATE_OFF;
1328 raw_spin_unlock_irq(&ctx->lock);
1330 EXPORT_SYMBOL_GPL(perf_event_disable);
1332 static void perf_set_shadow_time(struct perf_event *event,
1333 struct perf_event_context *ctx,
1337 * use the correct time source for the time snapshot
1339 * We could get by without this by leveraging the
1340 * fact that to get to this function, the caller
1341 * has most likely already called update_context_time()
1342 * and update_cgrp_time_xx() and thus both timestamp
1343 * are identical (or very close). Given that tstamp is,
1344 * already adjusted for cgroup, we could say that:
1345 * tstamp - ctx->timestamp
1347 * tstamp - cgrp->timestamp.
1349 * Then, in perf_output_read(), the calculation would
1350 * work with no changes because:
1351 * - event is guaranteed scheduled in
1352 * - no scheduled out in between
1353 * - thus the timestamp would be the same
1355 * But this is a bit hairy.
1357 * So instead, we have an explicit cgroup call to remain
1358 * within the time time source all along. We believe it
1359 * is cleaner and simpler to understand.
1361 if (is_cgroup_event(event))
1362 perf_cgroup_set_shadow_time(event, tstamp);
1364 event->shadow_ctx_time = tstamp - ctx->timestamp;
1367 #define MAX_INTERRUPTS (~0ULL)
1369 static void perf_log_throttle(struct perf_event *event, int enable);
1372 event_sched_in(struct perf_event *event,
1373 struct perf_cpu_context *cpuctx,
1374 struct perf_event_context *ctx)
1376 u64 tstamp = perf_event_time(event);
1378 if (event->state <= PERF_EVENT_STATE_OFF)
1381 event->state = PERF_EVENT_STATE_ACTIVE;
1382 event->oncpu = smp_processor_id();
1385 * Unthrottle events, since we scheduled we might have missed several
1386 * ticks already, also for a heavily scheduling task there is little
1387 * guarantee it'll get a tick in a timely manner.
1389 if (unlikely(event->hw.interrupts == MAX_INTERRUPTS)) {
1390 perf_log_throttle(event, 1);
1391 event->hw.interrupts = 0;
1395 * The new state must be visible before we turn it on in the hardware:
1399 if (event->pmu->add(event, PERF_EF_START)) {
1400 event->state = PERF_EVENT_STATE_INACTIVE;
1405 event->tstamp_running += tstamp - event->tstamp_stopped;
1407 perf_set_shadow_time(event, ctx, tstamp);
1409 if (!is_software_event(event))
1410 cpuctx->active_oncpu++;
1412 if (event->attr.freq && event->attr.sample_freq)
1415 if (event->attr.exclusive)
1416 cpuctx->exclusive = 1;
1422 group_sched_in(struct perf_event *group_event,
1423 struct perf_cpu_context *cpuctx,
1424 struct perf_event_context *ctx)
1426 struct perf_event *event, *partial_group = NULL;
1427 struct pmu *pmu = group_event->pmu;
1428 u64 now = ctx->time;
1429 bool simulate = false;
1431 if (group_event->state == PERF_EVENT_STATE_OFF)
1434 pmu->start_txn(pmu);
1436 if (event_sched_in(group_event, cpuctx, ctx)) {
1437 pmu->cancel_txn(pmu);
1442 * Schedule in siblings as one group (if any):
1444 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
1445 if (event_sched_in(event, cpuctx, ctx)) {
1446 partial_group = event;
1451 if (!pmu->commit_txn(pmu))
1456 * Groups can be scheduled in as one unit only, so undo any
1457 * partial group before returning:
1458 * The events up to the failed event are scheduled out normally,
1459 * tstamp_stopped will be updated.
1461 * The failed events and the remaining siblings need to have
1462 * their timings updated as if they had gone thru event_sched_in()
1463 * and event_sched_out(). This is required to get consistent timings
1464 * across the group. This also takes care of the case where the group
1465 * could never be scheduled by ensuring tstamp_stopped is set to mark
1466 * the time the event was actually stopped, such that time delta
1467 * calculation in update_event_times() is correct.
1469 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
1470 if (event == partial_group)
1474 event->tstamp_running += now - event->tstamp_stopped;
1475 event->tstamp_stopped = now;
1477 event_sched_out(event, cpuctx, ctx);
1480 event_sched_out(group_event, cpuctx, ctx);
1482 pmu->cancel_txn(pmu);
1488 * Work out whether we can put this event group on the CPU now.
1490 static int group_can_go_on(struct perf_event *event,
1491 struct perf_cpu_context *cpuctx,
1495 * Groups consisting entirely of software events can always go on.
1497 if (event->group_flags & PERF_GROUP_SOFTWARE)
1500 * If an exclusive group is already on, no other hardware
1503 if (cpuctx->exclusive)
1506 * If this group is exclusive and there are already
1507 * events on the CPU, it can't go on.
1509 if (event->attr.exclusive && cpuctx->active_oncpu)
1512 * Otherwise, try to add it if all previous groups were able
1518 static void add_event_to_ctx(struct perf_event *event,
1519 struct perf_event_context *ctx)
1521 u64 tstamp = perf_event_time(event);
1523 list_add_event(event, ctx);
1524 perf_group_attach(event);
1525 event->tstamp_enabled = tstamp;
1526 event->tstamp_running = tstamp;
1527 event->tstamp_stopped = tstamp;
1530 static void task_ctx_sched_out(struct perf_event_context *ctx);
1532 ctx_sched_in(struct perf_event_context *ctx,
1533 struct perf_cpu_context *cpuctx,
1534 enum event_type_t event_type,
1535 struct task_struct *task);
1537 static void perf_event_sched_in(struct perf_cpu_context *cpuctx,
1538 struct perf_event_context *ctx,
1539 struct task_struct *task)
1541 cpu_ctx_sched_in(cpuctx, EVENT_PINNED, task);
1543 ctx_sched_in(ctx, cpuctx, EVENT_PINNED, task);
1544 cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE, task);
1546 ctx_sched_in(ctx, cpuctx, EVENT_FLEXIBLE, task);
1550 * Cross CPU call to install and enable a performance event
1552 * Must be called with ctx->mutex held
1554 static int __perf_install_in_context(void *info)
1556 struct perf_event *event = info;
1557 struct perf_event_context *ctx = event->ctx;
1558 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1559 struct perf_event_context *task_ctx = cpuctx->task_ctx;
1560 struct task_struct *task = current;
1562 perf_ctx_lock(cpuctx, task_ctx);
1563 perf_pmu_disable(cpuctx->ctx.pmu);
1566 * If there was an active task_ctx schedule it out.
1569 task_ctx_sched_out(task_ctx);
1572 * If the context we're installing events in is not the
1573 * active task_ctx, flip them.
1575 if (ctx->task && task_ctx != ctx) {
1577 raw_spin_unlock(&task_ctx->lock);
1578 raw_spin_lock(&ctx->lock);
1583 cpuctx->task_ctx = task_ctx;
1584 task = task_ctx->task;
1587 cpu_ctx_sched_out(cpuctx, EVENT_ALL);
1589 update_context_time(ctx);
1591 * update cgrp time only if current cgrp
1592 * matches event->cgrp. Must be done before
1593 * calling add_event_to_ctx()
1595 update_cgrp_time_from_event(event);
1597 add_event_to_ctx(event, ctx);
1600 * Schedule everything back in
1602 perf_event_sched_in(cpuctx, task_ctx, task);
1604 perf_pmu_enable(cpuctx->ctx.pmu);
1605 perf_ctx_unlock(cpuctx, task_ctx);
1611 * Attach a performance event to a context
1613 * First we add the event to the list with the hardware enable bit
1614 * in event->hw_config cleared.
1616 * If the event is attached to a task which is on a CPU we use a smp
1617 * call to enable it in the task context. The task might have been
1618 * scheduled away, but we check this in the smp call again.
1621 perf_install_in_context(struct perf_event_context *ctx,
1622 struct perf_event *event,
1625 struct task_struct *task = ctx->task;
1627 lockdep_assert_held(&ctx->mutex);
1633 * Per cpu events are installed via an smp call and
1634 * the install is always successful.
1636 cpu_function_call(cpu, __perf_install_in_context, event);
1641 if (!task_function_call(task, __perf_install_in_context, event))
1644 raw_spin_lock_irq(&ctx->lock);
1646 * If we failed to find a running task, but find the context active now
1647 * that we've acquired the ctx->lock, retry.
1649 if (ctx->is_active) {
1650 raw_spin_unlock_irq(&ctx->lock);
1655 * Since the task isn't running, its safe to add the event, us holding
1656 * the ctx->lock ensures the task won't get scheduled in.
1658 add_event_to_ctx(event, ctx);
1659 raw_spin_unlock_irq(&ctx->lock);
1663 * Put a event into inactive state and update time fields.
1664 * Enabling the leader of a group effectively enables all
1665 * the group members that aren't explicitly disabled, so we
1666 * have to update their ->tstamp_enabled also.
1667 * Note: this works for group members as well as group leaders
1668 * since the non-leader members' sibling_lists will be empty.
1670 static void __perf_event_mark_enabled(struct perf_event *event)
1672 struct perf_event *sub;
1673 u64 tstamp = perf_event_time(event);
1675 event->state = PERF_EVENT_STATE_INACTIVE;
1676 event->tstamp_enabled = tstamp - event->total_time_enabled;
1677 list_for_each_entry(sub, &event->sibling_list, group_entry) {
1678 if (sub->state >= PERF_EVENT_STATE_INACTIVE)
1679 sub->tstamp_enabled = tstamp - sub->total_time_enabled;
1684 * Cross CPU call to enable a performance event
1686 static int __perf_event_enable(void *info)
1688 struct perf_event *event = info;
1689 struct perf_event_context *ctx = event->ctx;
1690 struct perf_event *leader = event->group_leader;
1691 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1694 if (WARN_ON_ONCE(!ctx->is_active))
1697 raw_spin_lock(&ctx->lock);
1698 update_context_time(ctx);
1700 if (event->state >= PERF_EVENT_STATE_INACTIVE)
1704 * set current task's cgroup time reference point
1706 perf_cgroup_set_timestamp(current, ctx);
1708 __perf_event_mark_enabled(event);
1710 if (!event_filter_match(event)) {
1711 if (is_cgroup_event(event))
1712 perf_cgroup_defer_enabled(event);
1717 * If the event is in a group and isn't the group leader,
1718 * then don't put it on unless the group is on.
1720 if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE)
1723 if (!group_can_go_on(event, cpuctx, 1)) {
1726 if (event == leader)
1727 err = group_sched_in(event, cpuctx, ctx);
1729 err = event_sched_in(event, cpuctx, ctx);
1734 * If this event can't go on and it's part of a
1735 * group, then the whole group has to come off.
1737 if (leader != event)
1738 group_sched_out(leader, cpuctx, ctx);
1739 if (leader->attr.pinned) {
1740 update_group_times(leader);
1741 leader->state = PERF_EVENT_STATE_ERROR;
1746 raw_spin_unlock(&ctx->lock);
1754 * If event->ctx is a cloned context, callers must make sure that
1755 * every task struct that event->ctx->task could possibly point to
1756 * remains valid. This condition is satisfied when called through
1757 * perf_event_for_each_child or perf_event_for_each as described
1758 * for perf_event_disable.
1760 void perf_event_enable(struct perf_event *event)
1762 struct perf_event_context *ctx = event->ctx;
1763 struct task_struct *task = ctx->task;
1767 * Enable the event on the cpu that it's on
1769 cpu_function_call(event->cpu, __perf_event_enable, event);
1773 raw_spin_lock_irq(&ctx->lock);
1774 if (event->state >= PERF_EVENT_STATE_INACTIVE)
1778 * If the event is in error state, clear that first.
1779 * That way, if we see the event in error state below, we
1780 * know that it has gone back into error state, as distinct
1781 * from the task having been scheduled away before the
1782 * cross-call arrived.
1784 if (event->state == PERF_EVENT_STATE_ERROR)
1785 event->state = PERF_EVENT_STATE_OFF;
1788 if (!ctx->is_active) {
1789 __perf_event_mark_enabled(event);
1793 raw_spin_unlock_irq(&ctx->lock);
1795 if (!task_function_call(task, __perf_event_enable, event))
1798 raw_spin_lock_irq(&ctx->lock);
1801 * If the context is active and the event is still off,
1802 * we need to retry the cross-call.
1804 if (ctx->is_active && event->state == PERF_EVENT_STATE_OFF) {
1806 * task could have been flipped by a concurrent
1807 * perf_event_context_sched_out()
1814 raw_spin_unlock_irq(&ctx->lock);
1816 EXPORT_SYMBOL_GPL(perf_event_enable);
1818 int perf_event_refresh(struct perf_event *event, int refresh)
1821 * not supported on inherited events
1823 if (event->attr.inherit || !is_sampling_event(event))
1826 atomic_add(refresh, &event->event_limit);
1827 perf_event_enable(event);
1831 EXPORT_SYMBOL_GPL(perf_event_refresh);
1833 static void ctx_sched_out(struct perf_event_context *ctx,
1834 struct perf_cpu_context *cpuctx,
1835 enum event_type_t event_type)
1837 struct perf_event *event;
1838 int is_active = ctx->is_active;
1840 ctx->is_active &= ~event_type;
1841 if (likely(!ctx->nr_events))
1844 update_context_time(ctx);
1845 update_cgrp_time_from_cpuctx(cpuctx);
1846 if (!ctx->nr_active)
1849 perf_pmu_disable(ctx->pmu);
1850 if ((is_active & EVENT_PINNED) && (event_type & EVENT_PINNED)) {
1851 list_for_each_entry(event, &ctx->pinned_groups, group_entry)
1852 group_sched_out(event, cpuctx, ctx);
1855 if ((is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE)) {
1856 list_for_each_entry(event, &ctx->flexible_groups, group_entry)
1857 group_sched_out(event, cpuctx, ctx);
1859 perf_pmu_enable(ctx->pmu);
1863 * Test whether two contexts are equivalent, i.e. whether they
1864 * have both been cloned from the same version of the same context
1865 * and they both have the same number of enabled events.
1866 * If the number of enabled events is the same, then the set
1867 * of enabled events should be the same, because these are both
1868 * inherited contexts, therefore we can't access individual events
1869 * in them directly with an fd; we can only enable/disable all
1870 * events via prctl, or enable/disable all events in a family
1871 * via ioctl, which will have the same effect on both contexts.
1873 static int context_equiv(struct perf_event_context *ctx1,
1874 struct perf_event_context *ctx2)
1876 return ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx
1877 && ctx1->parent_gen == ctx2->parent_gen
1878 && !ctx1->pin_count && !ctx2->pin_count;
1881 static void __perf_event_sync_stat(struct perf_event *event,
1882 struct perf_event *next_event)
1886 if (!event->attr.inherit_stat)
1890 * Update the event value, we cannot use perf_event_read()
1891 * because we're in the middle of a context switch and have IRQs
1892 * disabled, which upsets smp_call_function_single(), however
1893 * we know the event must be on the current CPU, therefore we
1894 * don't need to use it.
1896 switch (event->state) {
1897 case PERF_EVENT_STATE_ACTIVE:
1898 event->pmu->read(event);
1901 case PERF_EVENT_STATE_INACTIVE:
1902 update_event_times(event);
1910 * In order to keep per-task stats reliable we need to flip the event
1911 * values when we flip the contexts.
1913 value = local64_read(&next_event->count);
1914 value = local64_xchg(&event->count, value);
1915 local64_set(&next_event->count, value);
1917 swap(event->total_time_enabled, next_event->total_time_enabled);
1918 swap(event->total_time_running, next_event->total_time_running);
1921 * Since we swizzled the values, update the user visible data too.
1923 perf_event_update_userpage(event);
1924 perf_event_update_userpage(next_event);
1927 #define list_next_entry(pos, member) \
1928 list_entry(pos->member.next, typeof(*pos), member)
1930 static void perf_event_sync_stat(struct perf_event_context *ctx,
1931 struct perf_event_context *next_ctx)
1933 struct perf_event *event, *next_event;
1938 update_context_time(ctx);
1940 event = list_first_entry(&ctx->event_list,
1941 struct perf_event, event_entry);
1943 next_event = list_first_entry(&next_ctx->event_list,
1944 struct perf_event, event_entry);
1946 while (&event->event_entry != &ctx->event_list &&
1947 &next_event->event_entry != &next_ctx->event_list) {
1949 __perf_event_sync_stat(event, next_event);
1951 event = list_next_entry(event, event_entry);
1952 next_event = list_next_entry(next_event, event_entry);
1956 static void perf_event_context_sched_out(struct task_struct *task, int ctxn,
1957 struct task_struct *next)
1959 struct perf_event_context *ctx = task->perf_event_ctxp[ctxn];
1960 struct perf_event_context *next_ctx;
1961 struct perf_event_context *parent;
1962 struct perf_cpu_context *cpuctx;
1968 cpuctx = __get_cpu_context(ctx);
1969 if (!cpuctx->task_ctx)
1973 parent = rcu_dereference(ctx->parent_ctx);
1974 next_ctx = next->perf_event_ctxp[ctxn];
1975 if (parent && next_ctx &&
1976 rcu_dereference(next_ctx->parent_ctx) == parent) {
1978 * Looks like the two contexts are clones, so we might be
1979 * able to optimize the context switch. We lock both
1980 * contexts and check that they are clones under the
1981 * lock (including re-checking that neither has been
1982 * uncloned in the meantime). It doesn't matter which
1983 * order we take the locks because no other cpu could
1984 * be trying to lock both of these tasks.
1986 raw_spin_lock(&ctx->lock);
1987 raw_spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING);
1988 if (context_equiv(ctx, next_ctx)) {
1990 * XXX do we need a memory barrier of sorts
1991 * wrt to rcu_dereference() of perf_event_ctxp
1993 task->perf_event_ctxp[ctxn] = next_ctx;
1994 next->perf_event_ctxp[ctxn] = ctx;
1996 next_ctx->task = task;
1999 perf_event_sync_stat(ctx, next_ctx);
2001 raw_spin_unlock(&next_ctx->lock);
2002 raw_spin_unlock(&ctx->lock);
2007 raw_spin_lock(&ctx->lock);
2008 ctx_sched_out(ctx, cpuctx, EVENT_ALL);
2009 cpuctx->task_ctx = NULL;
2010 raw_spin_unlock(&ctx->lock);
2014 #define for_each_task_context_nr(ctxn) \
2015 for ((ctxn) = 0; (ctxn) < perf_nr_task_contexts; (ctxn)++)
2018 * Called from scheduler to remove the events of the current task,
2019 * with interrupts disabled.
2021 * We stop each event and update the event value in event->count.
2023 * This does not protect us against NMI, but disable()
2024 * sets the disabled bit in the control field of event _before_
2025 * accessing the event control register. If a NMI hits, then it will
2026 * not restart the event.
2028 void __perf_event_task_sched_out(struct task_struct *task,
2029 struct task_struct *next)
2033 for_each_task_context_nr(ctxn)
2034 perf_event_context_sched_out(task, ctxn, next);
2037 * if cgroup events exist on this CPU, then we need
2038 * to check if we have to switch out PMU state.
2039 * cgroup event are system-wide mode only
2041 if (atomic_read(&__get_cpu_var(perf_cgroup_events)))
2042 perf_cgroup_sched_out(task, next);
2045 static void task_ctx_sched_out(struct perf_event_context *ctx)
2047 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
2049 if (!cpuctx->task_ctx)
2052 if (WARN_ON_ONCE(ctx != cpuctx->task_ctx))
2055 ctx_sched_out(ctx, cpuctx, EVENT_ALL);
2056 cpuctx->task_ctx = NULL;
2060 * Called with IRQs disabled
2062 static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
2063 enum event_type_t event_type)
2065 ctx_sched_out(&cpuctx->ctx, cpuctx, event_type);
2069 ctx_pinned_sched_in(struct perf_event_context *ctx,
2070 struct perf_cpu_context *cpuctx)
2072 struct perf_event *event;
2074 list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
2075 if (event->state <= PERF_EVENT_STATE_OFF)
2077 if (!event_filter_match(event))
2080 /* may need to reset tstamp_enabled */
2081 if (is_cgroup_event(event))
2082 perf_cgroup_mark_enabled(event, ctx);
2084 if (group_can_go_on(event, cpuctx, 1))
2085 group_sched_in(event, cpuctx, ctx);
2088 * If this pinned group hasn't been scheduled,
2089 * put it in error state.
2091 if (event->state == PERF_EVENT_STATE_INACTIVE) {
2092 update_group_times(event);
2093 event->state = PERF_EVENT_STATE_ERROR;
2099 ctx_flexible_sched_in(struct perf_event_context *ctx,
2100 struct perf_cpu_context *cpuctx)
2102 struct perf_event *event;
2105 list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
2106 /* Ignore events in OFF or ERROR state */
2107 if (event->state <= PERF_EVENT_STATE_OFF)
2110 * Listen to the 'cpu' scheduling filter constraint
2113 if (!event_filter_match(event))
2116 /* may need to reset tstamp_enabled */
2117 if (is_cgroup_event(event))
2118 perf_cgroup_mark_enabled(event, ctx);
2120 if (group_can_go_on(event, cpuctx, can_add_hw)) {
2121 if (group_sched_in(event, cpuctx, ctx))
2128 ctx_sched_in(struct perf_event_context *ctx,
2129 struct perf_cpu_context *cpuctx,
2130 enum event_type_t event_type,
2131 struct task_struct *task)
2134 int is_active = ctx->is_active;
2136 ctx->is_active |= event_type;
2137 if (likely(!ctx->nr_events))
2141 ctx->timestamp = now;
2142 perf_cgroup_set_timestamp(task, ctx);
2144 * First go through the list and put on any pinned groups
2145 * in order to give them the best chance of going on.
2147 if (!(is_active & EVENT_PINNED) && (event_type & EVENT_PINNED))
2148 ctx_pinned_sched_in(ctx, cpuctx);
2150 /* Then walk through the lower prio flexible groups */
2151 if (!(is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE))
2152 ctx_flexible_sched_in(ctx, cpuctx);
2155 static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
2156 enum event_type_t event_type,
2157 struct task_struct *task)
2159 struct perf_event_context *ctx = &cpuctx->ctx;
2161 ctx_sched_in(ctx, cpuctx, event_type, task);
2164 static void perf_event_context_sched_in(struct perf_event_context *ctx,
2165 struct task_struct *task)
2167 struct perf_cpu_context *cpuctx;
2169 cpuctx = __get_cpu_context(ctx);
2170 if (cpuctx->task_ctx == ctx)
2173 perf_ctx_lock(cpuctx, ctx);
2174 perf_pmu_disable(ctx->pmu);
2176 * We want to keep the following priority order:
2177 * cpu pinned (that don't need to move), task pinned,
2178 * cpu flexible, task flexible.
2180 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
2183 cpuctx->task_ctx = ctx;
2185 perf_event_sched_in(cpuctx, cpuctx->task_ctx, task);
2187 perf_pmu_enable(ctx->pmu);
2188 perf_ctx_unlock(cpuctx, ctx);
2191 * Since these rotations are per-cpu, we need to ensure the
2192 * cpu-context we got scheduled on is actually rotating.
2194 perf_pmu_rotate_start(ctx->pmu);
2198 * Called from scheduler to add the events of the current task
2199 * with interrupts disabled.
2201 * We restore the event value and then enable it.
2203 * This does not protect us against NMI, but enable()
2204 * sets the enabled bit in the control field of event _before_
2205 * accessing the event control register. If a NMI hits, then it will
2206 * keep the event running.
2208 void __perf_event_task_sched_in(struct task_struct *prev,
2209 struct task_struct *task)
2211 struct perf_event_context *ctx;
2214 for_each_task_context_nr(ctxn) {
2215 ctx = task->perf_event_ctxp[ctxn];
2219 perf_event_context_sched_in(ctx, task);
2222 * if cgroup events exist on this CPU, then we need
2223 * to check if we have to switch in PMU state.
2224 * cgroup event are system-wide mode only
2226 if (atomic_read(&__get_cpu_var(perf_cgroup_events)))
2227 perf_cgroup_sched_in(prev, task);
2230 static u64 perf_calculate_period(struct perf_event *event, u64 nsec, u64 count)
2232 u64 frequency = event->attr.sample_freq;
2233 u64 sec = NSEC_PER_SEC;
2234 u64 divisor, dividend;
2236 int count_fls, nsec_fls, frequency_fls, sec_fls;
2238 count_fls = fls64(count);
2239 nsec_fls = fls64(nsec);
2240 frequency_fls = fls64(frequency);
2244 * We got @count in @nsec, with a target of sample_freq HZ
2245 * the target period becomes:
2248 * period = -------------------
2249 * @nsec * sample_freq
2254 * Reduce accuracy by one bit such that @a and @b converge
2255 * to a similar magnitude.
2257 #define REDUCE_FLS(a, b) \
2259 if (a##_fls > b##_fls) { \
2269 * Reduce accuracy until either term fits in a u64, then proceed with
2270 * the other, so that finally we can do a u64/u64 division.
2272 while (count_fls + sec_fls > 64 && nsec_fls + frequency_fls > 64) {
2273 REDUCE_FLS(nsec, frequency);
2274 REDUCE_FLS(sec, count);
2277 if (count_fls + sec_fls > 64) {
2278 divisor = nsec * frequency;
2280 while (count_fls + sec_fls > 64) {
2281 REDUCE_FLS(count, sec);
2285 dividend = count * sec;
2287 dividend = count * sec;
2289 while (nsec_fls + frequency_fls > 64) {
2290 REDUCE_FLS(nsec, frequency);
2294 divisor = nsec * frequency;
2300 return div64_u64(dividend, divisor);
2303 static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count)
2305 struct hw_perf_event *hwc = &event->hw;
2306 s64 period, sample_period;
2309 period = perf_calculate_period(event, nsec, count);
2311 delta = (s64)(period - hwc->sample_period);
2312 delta = (delta + 7) / 8; /* low pass filter */
2314 sample_period = hwc->sample_period + delta;
2319 hwc->sample_period = sample_period;
2321 if (local64_read(&hwc->period_left) > 8*sample_period) {
2322 event->pmu->stop(event, PERF_EF_UPDATE);
2323 local64_set(&hwc->period_left, 0);
2324 event->pmu->start(event, PERF_EF_RELOAD);
2328 static void perf_ctx_adjust_freq(struct perf_event_context *ctx, u64 period)
2330 struct perf_event *event;
2331 struct hw_perf_event *hwc;
2332 u64 interrupts, now;
2338 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
2339 if (event->state != PERF_EVENT_STATE_ACTIVE)
2342 if (!event_filter_match(event))
2347 interrupts = hwc->interrupts;
2348 hwc->interrupts = 0;
2351 * unthrottle events on the tick
2353 if (interrupts == MAX_INTERRUPTS) {
2354 perf_log_throttle(event, 1);
2355 event->pmu->start(event, 0);
2358 if (!event->attr.freq || !event->attr.sample_freq)
2361 event->pmu->read(event);
2362 now = local64_read(&event->count);
2363 delta = now - hwc->freq_count_stamp;
2364 hwc->freq_count_stamp = now;
2367 perf_adjust_period(event, period, delta);
2372 * Round-robin a context's events:
2374 static void rotate_ctx(struct perf_event_context *ctx)
2377 * Rotate the first entry last of non-pinned groups. Rotation might be
2378 * disabled by the inheritance code.
2380 if (!ctx->rotate_disable)
2381 list_rotate_left(&ctx->flexible_groups);
2385 * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
2386 * because they're strictly cpu affine and rotate_start is called with IRQs
2387 * disabled, while rotate_context is called from IRQ context.
2389 static void perf_rotate_context(struct perf_cpu_context *cpuctx)
2391 u64 interval = (u64)cpuctx->jiffies_interval * TICK_NSEC;
2392 struct perf_event_context *ctx = NULL;
2393 int rotate = 0, remove = 1, freq = 0;
2395 if (cpuctx->ctx.nr_events) {
2397 if (cpuctx->ctx.nr_events != cpuctx->ctx.nr_active)
2399 if (cpuctx->ctx.nr_freq)
2403 ctx = cpuctx->task_ctx;
2404 if (ctx && ctx->nr_events) {
2406 if (ctx->nr_events != ctx->nr_active)
2412 if (!rotate && !freq)
2415 perf_ctx_lock(cpuctx, cpuctx->task_ctx);
2416 perf_pmu_disable(cpuctx->ctx.pmu);
2419 perf_ctx_adjust_freq(&cpuctx->ctx, interval);
2421 perf_ctx_adjust_freq(ctx, interval);
2425 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
2427 ctx_sched_out(ctx, cpuctx, EVENT_FLEXIBLE);
2429 rotate_ctx(&cpuctx->ctx);
2433 perf_event_sched_in(cpuctx, ctx, current);
2436 perf_pmu_enable(cpuctx->ctx.pmu);
2437 perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
2441 list_del_init(&cpuctx->rotation_list);
2444 void perf_event_task_tick(void)
2446 struct list_head *head = &__get_cpu_var(rotation_list);
2447 struct perf_cpu_context *cpuctx, *tmp;
2449 WARN_ON(!irqs_disabled());
2451 list_for_each_entry_safe(cpuctx, tmp, head, rotation_list) {
2452 if (cpuctx->jiffies_interval == 1 ||
2453 !(jiffies % cpuctx->jiffies_interval))
2454 perf_rotate_context(cpuctx);
2458 static int event_enable_on_exec(struct perf_event *event,
2459 struct perf_event_context *ctx)
2461 if (!event->attr.enable_on_exec)
2464 event->attr.enable_on_exec = 0;
2465 if (event->state >= PERF_EVENT_STATE_INACTIVE)
2468 __perf_event_mark_enabled(event);
2474 * Enable all of a task's events that have been marked enable-on-exec.
2475 * This expects task == current.
2477 static void perf_event_enable_on_exec(struct perf_event_context *ctx)
2479 struct perf_event *event;
2480 unsigned long flags;
2484 local_irq_save(flags);
2485 if (!ctx || !ctx->nr_events)
2489 * We must ctxsw out cgroup events to avoid conflict
2490 * when invoking perf_task_event_sched_in() later on
2491 * in this function. Otherwise we end up trying to
2492 * ctxswin cgroup events which are already scheduled
2495 perf_cgroup_sched_out(current, NULL);
2497 raw_spin_lock(&ctx->lock);
2498 task_ctx_sched_out(ctx);
2500 list_for_each_entry(event, &ctx->event_list, event_entry) {
2501 ret = event_enable_on_exec(event, ctx);
2507 * Unclone this context if we enabled any event.
2512 raw_spin_unlock(&ctx->lock);
2515 * Also calls ctxswin for cgroup events, if any:
2517 perf_event_context_sched_in(ctx, ctx->task);
2519 local_irq_restore(flags);
2523 * Cross CPU call to read the hardware event
2525 static void __perf_event_read(void *info)
2527 struct perf_event *event = info;
2528 struct perf_event_context *ctx = event->ctx;
2529 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
2532 * If this is a task context, we need to check whether it is
2533 * the current task context of this cpu. If not it has been
2534 * scheduled out before the smp call arrived. In that case
2535 * event->count would have been updated to a recent sample
2536 * when the event was scheduled out.
2538 if (ctx->task && cpuctx->task_ctx != ctx)
2541 raw_spin_lock(&ctx->lock);
2542 if (ctx->is_active) {
2543 update_context_time(ctx);
2544 update_cgrp_time_from_event(event);
2546 update_event_times(event);
2547 if (event->state == PERF_EVENT_STATE_ACTIVE)
2548 event->pmu->read(event);
2549 raw_spin_unlock(&ctx->lock);
2552 static inline u64 perf_event_count(struct perf_event *event)
2554 return local64_read(&event->count) + atomic64_read(&event->child_count);
2557 static u64 perf_event_read(struct perf_event *event)
2560 * If event is enabled and currently active on a CPU, update the
2561 * value in the event structure:
2563 if (event->state == PERF_EVENT_STATE_ACTIVE) {
2564 smp_call_function_single(event->oncpu,
2565 __perf_event_read, event, 1);
2566 } else if (event->state == PERF_EVENT_STATE_INACTIVE) {
2567 struct perf_event_context *ctx = event->ctx;
2568 unsigned long flags;
2570 raw_spin_lock_irqsave(&ctx->lock, flags);
2572 * may read while context is not active
2573 * (e.g., thread is blocked), in that case
2574 * we cannot update context time
2576 if (ctx->is_active) {
2577 update_context_time(ctx);
2578 update_cgrp_time_from_event(event);
2580 update_event_times(event);
2581 raw_spin_unlock_irqrestore(&ctx->lock, flags);
2584 return perf_event_count(event);
2588 * Initialize the perf_event context in a task_struct:
2590 static void __perf_event_init_context(struct perf_event_context *ctx)
2592 raw_spin_lock_init(&ctx->lock);
2593 mutex_init(&ctx->mutex);
2594 INIT_LIST_HEAD(&ctx->pinned_groups);
2595 INIT_LIST_HEAD(&ctx->flexible_groups);
2596 INIT_LIST_HEAD(&ctx->event_list);
2597 atomic_set(&ctx->refcount, 1);
2600 static struct perf_event_context *
2601 alloc_perf_context(struct pmu *pmu, struct task_struct *task)
2603 struct perf_event_context *ctx;
2605 ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL);
2609 __perf_event_init_context(ctx);
2612 get_task_struct(task);
2619 static struct task_struct *
2620 find_lively_task_by_vpid(pid_t vpid)
2622 struct task_struct *task;
2629 task = find_task_by_vpid(vpid);
2631 get_task_struct(task);
2635 return ERR_PTR(-ESRCH);
2637 /* Reuse ptrace permission checks for now. */
2639 if (!ptrace_may_access(task, PTRACE_MODE_READ))
2644 put_task_struct(task);
2645 return ERR_PTR(err);
2650 * Returns a matching context with refcount and pincount.
2652 static struct perf_event_context *
2653 find_get_context(struct pmu *pmu, struct task_struct *task, int cpu)
2655 struct perf_event_context *ctx;
2656 struct perf_cpu_context *cpuctx;
2657 unsigned long flags;
2661 /* Must be root to operate on a CPU event: */
2662 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
2663 return ERR_PTR(-EACCES);
2666 * We could be clever and allow to attach a event to an
2667 * offline CPU and activate it when the CPU comes up, but
2670 if (!cpu_online(cpu))
2671 return ERR_PTR(-ENODEV);
2673 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
2682 ctxn = pmu->task_ctx_nr;
2687 ctx = perf_lock_task_context(task, ctxn, &flags);
2691 raw_spin_unlock_irqrestore(&ctx->lock, flags);
2693 ctx = alloc_perf_context(pmu, task);
2699 mutex_lock(&task->perf_event_mutex);
2701 * If it has already passed perf_event_exit_task().
2702 * we must see PF_EXITING, it takes this mutex too.
2704 if (task->flags & PF_EXITING)
2706 else if (task->perf_event_ctxp[ctxn])
2711 rcu_assign_pointer(task->perf_event_ctxp[ctxn], ctx);
2713 mutex_unlock(&task->perf_event_mutex);
2715 if (unlikely(err)) {
2727 return ERR_PTR(err);
2730 static void perf_event_free_filter(struct perf_event *event);
2732 static void free_event_rcu(struct rcu_head *head)
2734 struct perf_event *event;
2736 event = container_of(head, struct perf_event, rcu_head);
2738 put_pid_ns(event->ns);
2739 perf_event_free_filter(event);
2743 static void ring_buffer_put(struct ring_buffer *rb);
2745 static void free_event(struct perf_event *event)
2747 irq_work_sync(&event->pending);
2749 if (!event->parent) {
2750 if (event->attach_state & PERF_ATTACH_TASK)
2751 jump_label_dec_deferred(&perf_sched_events);
2752 if (event->attr.mmap || event->attr.mmap_data)
2753 atomic_dec(&nr_mmap_events);
2754 if (event->attr.comm)
2755 atomic_dec(&nr_comm_events);
2756 if (event->attr.task)
2757 atomic_dec(&nr_task_events);
2758 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN)
2759 put_callchain_buffers();
2760 if (is_cgroup_event(event)) {
2761 atomic_dec(&per_cpu(perf_cgroup_events, event->cpu));
2762 jump_label_dec_deferred(&perf_sched_events);
2767 ring_buffer_put(event->rb);
2771 if (is_cgroup_event(event))
2772 perf_detach_cgroup(event);
2775 event->destroy(event);
2778 put_ctx(event->ctx);
2780 call_rcu(&event->rcu_head, free_event_rcu);
2783 int perf_event_release_kernel(struct perf_event *event)
2785 struct perf_event_context *ctx = event->ctx;
2787 WARN_ON_ONCE(ctx->parent_ctx);
2789 * There are two ways this annotation is useful:
2791 * 1) there is a lock recursion from perf_event_exit_task
2792 * see the comment there.
2794 * 2) there is a lock-inversion with mmap_sem through
2795 * perf_event_read_group(), which takes faults while
2796 * holding ctx->mutex, however this is called after
2797 * the last filedesc died, so there is no possibility
2798 * to trigger the AB-BA case.
2800 mutex_lock_nested(&ctx->mutex, SINGLE_DEPTH_NESTING);
2801 raw_spin_lock_irq(&ctx->lock);
2802 perf_group_detach(event);
2803 raw_spin_unlock_irq(&ctx->lock);
2804 perf_remove_from_context(event);
2805 mutex_unlock(&ctx->mutex);
2811 EXPORT_SYMBOL_GPL(perf_event_release_kernel);
2814 * Called when the last reference to the file is gone.
2816 static int perf_release(struct inode *inode, struct file *file)
2818 struct perf_event *event = file->private_data;
2819 struct task_struct *owner;
2821 file->private_data = NULL;
2824 owner = ACCESS_ONCE(event->owner);
2826 * Matches the smp_wmb() in perf_event_exit_task(). If we observe
2827 * !owner it means the list deletion is complete and we can indeed
2828 * free this event, otherwise we need to serialize on
2829 * owner->perf_event_mutex.
2831 smp_read_barrier_depends();
2834 * Since delayed_put_task_struct() also drops the last
2835 * task reference we can safely take a new reference
2836 * while holding the rcu_read_lock().
2838 get_task_struct(owner);
2843 mutex_lock(&owner->perf_event_mutex);
2845 * We have to re-check the event->owner field, if it is cleared
2846 * we raced with perf_event_exit_task(), acquiring the mutex
2847 * ensured they're done, and we can proceed with freeing the
2851 list_del_init(&event->owner_entry);
2852 mutex_unlock(&owner->perf_event_mutex);
2853 put_task_struct(owner);
2856 return perf_event_release_kernel(event);
2859 u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running)
2861 struct perf_event *child;
2867 mutex_lock(&event->child_mutex);
2868 total += perf_event_read(event);
2869 *enabled += event->total_time_enabled +
2870 atomic64_read(&event->child_total_time_enabled);
2871 *running += event->total_time_running +
2872 atomic64_read(&event->child_total_time_running);
2874 list_for_each_entry(child, &event->child_list, child_list) {
2875 total += perf_event_read(child);
2876 *enabled += child->total_time_enabled;
2877 *running += child->total_time_running;
2879 mutex_unlock(&event->child_mutex);
2883 EXPORT_SYMBOL_GPL(perf_event_read_value);
2885 static int perf_event_read_group(struct perf_event *event,
2886 u64 read_format, char __user *buf)
2888 struct perf_event *leader = event->group_leader, *sub;
2889 int n = 0, size = 0, ret = -EFAULT;
2890 struct perf_event_context *ctx = leader->ctx;
2892 u64 count, enabled, running;
2894 mutex_lock(&ctx->mutex);
2895 count = perf_event_read_value(leader, &enabled, &running);
2897 values[n++] = 1 + leader->nr_siblings;
2898 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2899 values[n++] = enabled;
2900 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2901 values[n++] = running;
2902 values[n++] = count;
2903 if (read_format & PERF_FORMAT_ID)
2904 values[n++] = primary_event_id(leader);
2906 size = n * sizeof(u64);
2908 if (copy_to_user(buf, values, size))
2913 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
2916 values[n++] = perf_event_read_value(sub, &enabled, &running);
2917 if (read_format & PERF_FORMAT_ID)
2918 values[n++] = primary_event_id(sub);
2920 size = n * sizeof(u64);
2922 if (copy_to_user(buf + ret, values, size)) {
2930 mutex_unlock(&ctx->mutex);
2935 static int perf_event_read_one(struct perf_event *event,
2936 u64 read_format, char __user *buf)
2938 u64 enabled, running;
2942 values[n++] = perf_event_read_value(event, &enabled, &running);
2943 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2944 values[n++] = enabled;
2945 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2946 values[n++] = running;
2947 if (read_format & PERF_FORMAT_ID)
2948 values[n++] = primary_event_id(event);
2950 if (copy_to_user(buf, values, n * sizeof(u64)))
2953 return n * sizeof(u64);
2957 * Read the performance event - simple non blocking version for now
2960 perf_read_hw(struct perf_event *event, char __user *buf, size_t count)
2962 u64 read_format = event->attr.read_format;
2966 * Return end-of-file for a read on a event that is in
2967 * error state (i.e. because it was pinned but it couldn't be
2968 * scheduled on to the CPU at some point).
2970 if (event->state == PERF_EVENT_STATE_ERROR)
2973 if (count < event->read_size)
2976 WARN_ON_ONCE(event->ctx->parent_ctx);
2977 if (read_format & PERF_FORMAT_GROUP)
2978 ret = perf_event_read_group(event, read_format, buf);
2980 ret = perf_event_read_one(event, read_format, buf);
2986 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
2988 struct perf_event *event = file->private_data;
2990 return perf_read_hw(event, buf, count);
2993 static unsigned int perf_poll(struct file *file, poll_table *wait)
2995 struct perf_event *event = file->private_data;
2996 struct ring_buffer *rb;
2997 unsigned int events = POLL_HUP;
3000 * Race between perf_event_set_output() and perf_poll(): perf_poll()
3001 * grabs the rb reference but perf_event_set_output() overrides it.
3002 * Here is the timeline for two threads T1, T2:
3003 * t0: T1, rb = rcu_dereference(event->rb)
3004 * t1: T2, old_rb = event->rb
3005 * t2: T2, event->rb = new rb
3006 * t3: T2, ring_buffer_detach(old_rb)
3007 * t4: T1, ring_buffer_attach(rb1)
3008 * t5: T1, poll_wait(event->waitq)
3010 * To avoid this problem, we grab mmap_mutex in perf_poll()
3011 * thereby ensuring that the assignment of the new ring buffer
3012 * and the detachment of the old buffer appear atomic to perf_poll()
3014 mutex_lock(&event->mmap_mutex);
3017 rb = rcu_dereference(event->rb);
3019 ring_buffer_attach(event, rb);
3020 events = atomic_xchg(&rb->poll, 0);
3024 mutex_unlock(&event->mmap_mutex);
3026 poll_wait(file, &event->waitq, wait);
3031 static void perf_event_reset(struct perf_event *event)
3033 (void)perf_event_read(event);
3034 local64_set(&event->count, 0);
3035 perf_event_update_userpage(event);
3039 * Holding the top-level event's child_mutex means that any
3040 * descendant process that has inherited this event will block
3041 * in sync_child_event if it goes to exit, thus satisfying the
3042 * task existence requirements of perf_event_enable/disable.
3044 static void perf_event_for_each_child(struct perf_event *event,
3045 void (*func)(struct perf_event *))
3047 struct perf_event *child;
3049 WARN_ON_ONCE(event->ctx->parent_ctx);
3050 mutex_lock(&event->child_mutex);
3052 list_for_each_entry(child, &event->child_list, child_list)
3054 mutex_unlock(&event->child_mutex);
3057 static void perf_event_for_each(struct perf_event *event,
3058 void (*func)(struct perf_event *))
3060 struct perf_event_context *ctx = event->ctx;
3061 struct perf_event *sibling;
3063 WARN_ON_ONCE(ctx->parent_ctx);
3064 mutex_lock(&ctx->mutex);
3065 event = event->group_leader;
3067 perf_event_for_each_child(event, func);
3069 list_for_each_entry(sibling, &event->sibling_list, group_entry)
3070 perf_event_for_each_child(event, func);
3071 mutex_unlock(&ctx->mutex);
3074 static int perf_event_period(struct perf_event *event, u64 __user *arg)
3076 struct perf_event_context *ctx = event->ctx;
3080 if (!is_sampling_event(event))
3083 if (copy_from_user(&value, arg, sizeof(value)))
3089 raw_spin_lock_irq(&ctx->lock);
3090 if (event->attr.freq) {
3091 if (value > sysctl_perf_event_sample_rate) {
3096 event->attr.sample_freq = value;
3098 event->attr.sample_period = value;
3099 event->hw.sample_period = value;
3102 raw_spin_unlock_irq(&ctx->lock);
3107 static const struct file_operations perf_fops;
3109 static struct perf_event *perf_fget_light(int fd, int *fput_needed)
3113 file = fget_light(fd, fput_needed);
3115 return ERR_PTR(-EBADF);
3117 if (file->f_op != &perf_fops) {
3118 fput_light(file, *fput_needed);
3120 return ERR_PTR(-EBADF);
3123 return file->private_data;
3126 static int perf_event_set_output(struct perf_event *event,
3127 struct perf_event *output_event);
3128 static int perf_event_set_filter(struct perf_event *event, void __user *arg);
3130 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
3132 struct perf_event *event = file->private_data;
3133 void (*func)(struct perf_event *);
3137 case PERF_EVENT_IOC_ENABLE:
3138 func = perf_event_enable;
3140 case PERF_EVENT_IOC_DISABLE:
3141 func = perf_event_disable;
3143 case PERF_EVENT_IOC_RESET:
3144 func = perf_event_reset;
3147 case PERF_EVENT_IOC_REFRESH:
3148 return perf_event_refresh(event, arg);
3150 case PERF_EVENT_IOC_PERIOD:
3151 return perf_event_period(event, (u64 __user *)arg);
3153 case PERF_EVENT_IOC_SET_OUTPUT:
3155 struct perf_event *output_event = NULL;
3156 int fput_needed = 0;
3160 output_event = perf_fget_light(arg, &fput_needed);
3161 if (IS_ERR(output_event))
3162 return PTR_ERR(output_event);
3165 ret = perf_event_set_output(event, output_event);
3167 fput_light(output_event->filp, fput_needed);
3172 case PERF_EVENT_IOC_SET_FILTER:
3173 return perf_event_set_filter(event, (void __user *)arg);
3179 if (flags & PERF_IOC_FLAG_GROUP)
3180 perf_event_for_each(event, func);
3182 perf_event_for_each_child(event, func);
3187 int perf_event_task_enable(void)
3189 struct perf_event *event;
3191 mutex_lock(¤t->perf_event_mutex);
3192 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
3193 perf_event_for_each_child(event, perf_event_enable);
3194 mutex_unlock(¤t->perf_event_mutex);
3199 int perf_event_task_disable(void)
3201 struct perf_event *event;
3203 mutex_lock(¤t->perf_event_mutex);
3204 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
3205 perf_event_for_each_child(event, perf_event_disable);
3206 mutex_unlock(¤t->perf_event_mutex);
3211 static int perf_event_index(struct perf_event *event)
3213 if (event->hw.state & PERF_HES_STOPPED)
3216 if (event->state != PERF_EVENT_STATE_ACTIVE)
3219 return event->pmu->event_idx(event);
3222 static void calc_timer_values(struct perf_event *event,
3229 *now = perf_clock();
3230 ctx_time = event->shadow_ctx_time + *now;
3231 *enabled = ctx_time - event->tstamp_enabled;
3232 *running = ctx_time - event->tstamp_running;
3235 void __weak perf_update_user_clock(struct perf_event_mmap_page *userpg, u64 now)
3240 * Callers need to ensure there can be no nesting of this function, otherwise
3241 * the seqlock logic goes bad. We can not serialize this because the arch
3242 * code calls this from NMI context.
3244 void perf_event_update_userpage(struct perf_event *event)
3246 struct perf_event_mmap_page *userpg;
3247 struct ring_buffer *rb;
3248 u64 enabled, running, now;
3252 * compute total_time_enabled, total_time_running
3253 * based on snapshot values taken when the event
3254 * was last scheduled in.
3256 * we cannot simply called update_context_time()
3257 * because of locking issue as we can be called in
3260 calc_timer_values(event, &now, &enabled, &running);
3261 rb = rcu_dereference(event->rb);
3265 userpg = rb->user_page;
3268 * Disable preemption so as to not let the corresponding user-space
3269 * spin too long if we get preempted.
3274 userpg->index = perf_event_index(event);
3275 userpg->offset = perf_event_count(event);
3277 userpg->offset -= local64_read(&event->hw.prev_count);
3279 userpg->time_enabled = enabled +
3280 atomic64_read(&event->child_total_time_enabled);
3282 userpg->time_running = running +
3283 atomic64_read(&event->child_total_time_running);
3285 perf_update_user_clock(userpg, now);
3294 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
3296 struct perf_event *event = vma->vm_file->private_data;
3297 struct ring_buffer *rb;
3298 int ret = VM_FAULT_SIGBUS;
3300 if (vmf->flags & FAULT_FLAG_MKWRITE) {
3301 if (vmf->pgoff == 0)
3307 rb = rcu_dereference(event->rb);
3311 if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE))
3314 vmf->page = perf_mmap_to_page(rb, vmf->pgoff);
3318 get_page(vmf->page);
3319 vmf->page->mapping = vma->vm_file->f_mapping;
3320 vmf->page->index = vmf->pgoff;
3329 static void ring_buffer_attach(struct perf_event *event,
3330 struct ring_buffer *rb)
3332 unsigned long flags;
3334 if (!list_empty(&event->rb_entry))
3337 spin_lock_irqsave(&rb->event_lock, flags);
3338 if (!list_empty(&event->rb_entry))
3341 list_add(&event->rb_entry, &rb->event_list);
3343 spin_unlock_irqrestore(&rb->event_lock, flags);
3346 static void ring_buffer_detach(struct perf_event *event,
3347 struct ring_buffer *rb)
3349 unsigned long flags;
3351 if (list_empty(&event->rb_entry))
3354 spin_lock_irqsave(&rb->event_lock, flags);
3355 list_del_init(&event->rb_entry);
3356 wake_up_all(&event->waitq);
3357 spin_unlock_irqrestore(&rb->event_lock, flags);
3360 static void ring_buffer_wakeup(struct perf_event *event)
3362 struct ring_buffer *rb;
3365 rb = rcu_dereference(event->rb);
3369 list_for_each_entry_rcu(event, &rb->event_list, rb_entry)
3370 wake_up_all(&event->waitq);
3376 static void rb_free_rcu(struct rcu_head *rcu_head)
3378 struct ring_buffer *rb;
3380 rb = container_of(rcu_head, struct ring_buffer, rcu_head);
3384 static struct ring_buffer *ring_buffer_get(struct perf_event *event)
3386 struct ring_buffer *rb;
3389 rb = rcu_dereference(event->rb);
3391 if (!atomic_inc_not_zero(&rb->refcount))
3399 static void ring_buffer_put(struct ring_buffer *rb)
3401 struct perf_event *event, *n;
3402 unsigned long flags;
3404 if (!atomic_dec_and_test(&rb->refcount))
3407 spin_lock_irqsave(&rb->event_lock, flags);
3408 list_for_each_entry_safe(event, n, &rb->event_list, rb_entry) {
3409 list_del_init(&event->rb_entry);
3410 wake_up_all(&event->waitq);
3412 spin_unlock_irqrestore(&rb->event_lock, flags);
3414 call_rcu(&rb->rcu_head, rb_free_rcu);
3417 static void perf_mmap_open(struct vm_area_struct *vma)
3419 struct perf_event *event = vma->vm_file->private_data;
3421 atomic_inc(&event->mmap_count);
3424 static void perf_mmap_close(struct vm_area_struct *vma)
3426 struct perf_event *event = vma->vm_file->private_data;
3428 if (atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex)) {
3429 unsigned long size = perf_data_size(event->rb);
3430 struct user_struct *user = event->mmap_user;
3431 struct ring_buffer *rb = event->rb;
3433 atomic_long_sub((size >> PAGE_SHIFT) + 1, &user->locked_vm);
3434 vma->vm_mm->pinned_vm -= event->mmap_locked;
3435 rcu_assign_pointer(event->rb, NULL);
3436 ring_buffer_detach(event, rb);
3437 mutex_unlock(&event->mmap_mutex);
3439 ring_buffer_put(rb);
3444 static const struct vm_operations_struct perf_mmap_vmops = {
3445 .open = perf_mmap_open,
3446 .close = perf_mmap_close,
3447 .fault = perf_mmap_fault,
3448 .page_mkwrite = perf_mmap_fault,
3451 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
3453 struct perf_event *event = file->private_data;
3454 unsigned long user_locked, user_lock_limit;
3455 struct user_struct *user = current_user();
3456 unsigned long locked, lock_limit;
3457 struct ring_buffer *rb;
3458 unsigned long vma_size;
3459 unsigned long nr_pages;
3460 long user_extra, extra;
3461 int ret = 0, flags = 0;
3464 * Don't allow mmap() of inherited per-task counters. This would
3465 * create a performance issue due to all children writing to the
3468 if (event->cpu == -1 && event->attr.inherit)
3471 if (!(vma->vm_flags & VM_SHARED))
3474 vma_size = vma->vm_end - vma->vm_start;
3475 nr_pages = (vma_size / PAGE_SIZE) - 1;
3478 * If we have rb pages ensure they're a power-of-two number, so we
3479 * can do bitmasks instead of modulo.
3481 if (nr_pages != 0 && !is_power_of_2(nr_pages))
3484 if (vma_size != PAGE_SIZE * (1 + nr_pages))
3487 if (vma->vm_pgoff != 0)
3490 WARN_ON_ONCE(event->ctx->parent_ctx);
3491 mutex_lock(&event->mmap_mutex);
3493 if (event->rb->nr_pages == nr_pages)
3494 atomic_inc(&event->rb->refcount);
3500 user_extra = nr_pages + 1;
3501 user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10);
3504 * Increase the limit linearly with more CPUs:
3506 user_lock_limit *= num_online_cpus();
3508 user_locked = atomic_long_read(&user->locked_vm) + user_extra;
3511 if (user_locked > user_lock_limit)
3512 extra = user_locked - user_lock_limit;
3514 lock_limit = rlimit(RLIMIT_MEMLOCK);
3515 lock_limit >>= PAGE_SHIFT;
3516 locked = vma->vm_mm->pinned_vm + extra;
3518 if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() &&
3519 !capable(CAP_IPC_LOCK)) {
3526 if (vma->vm_flags & VM_WRITE)
3527 flags |= RING_BUFFER_WRITABLE;
3529 rb = rb_alloc(nr_pages,
3530 event->attr.watermark ? event->attr.wakeup_watermark : 0,
3537 rcu_assign_pointer(event->rb, rb);
3539 atomic_long_add(user_extra, &user->locked_vm);
3540 event->mmap_locked = extra;
3541 event->mmap_user = get_current_user();
3542 vma->vm_mm->pinned_vm += event->mmap_locked;
3544 perf_event_update_userpage(event);
3548 atomic_inc(&event->mmap_count);
3549 mutex_unlock(&event->mmap_mutex);
3551 vma->vm_flags |= VM_RESERVED;
3552 vma->vm_ops = &perf_mmap_vmops;
3557 static int perf_fasync(int fd, struct file *filp, int on)
3559 struct inode *inode = filp->f_path.dentry->d_inode;
3560 struct perf_event *event = filp->private_data;
3563 mutex_lock(&inode->i_mutex);
3564 retval = fasync_helper(fd, filp, on, &event->fasync);
3565 mutex_unlock(&inode->i_mutex);
3573 static const struct file_operations perf_fops = {
3574 .llseek = no_llseek,
3575 .release = perf_release,
3578 .unlocked_ioctl = perf_ioctl,
3579 .compat_ioctl = perf_ioctl,
3581 .fasync = perf_fasync,
3587 * If there's data, ensure we set the poll() state and publish everything
3588 * to user-space before waking everybody up.
3591 void perf_event_wakeup(struct perf_event *event)
3593 ring_buffer_wakeup(event);
3595 if (event->pending_kill) {
3596 kill_fasync(&event->fasync, SIGIO, event->pending_kill);
3597 event->pending_kill = 0;
3601 static void perf_pending_event(struct irq_work *entry)
3603 struct perf_event *event = container_of(entry,
3604 struct perf_event, pending);
3606 if (event->pending_disable) {
3607 event->pending_disable = 0;
3608 __perf_event_disable(event);
3611 if (event->pending_wakeup) {
3612 event->pending_wakeup = 0;
3613 perf_event_wakeup(event);
3618 * We assume there is only KVM supporting the callbacks.
3619 * Later on, we might change it to a list if there is
3620 * another virtualization implementation supporting the callbacks.
3622 struct perf_guest_info_callbacks *perf_guest_cbs;
3624 int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
3626 perf_guest_cbs = cbs;
3629 EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks);
3631 int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
3633 perf_guest_cbs = NULL;
3636 EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks);
3638 static void __perf_event_header__init_id(struct perf_event_header *header,
3639 struct perf_sample_data *data,
3640 struct perf_event *event)
3642 u64 sample_type = event->attr.sample_type;
3644 data->type = sample_type;
3645 header->size += event->id_header_size;
3647 if (sample_type & PERF_SAMPLE_TID) {
3648 /* namespace issues */
3649 data->tid_entry.pid = perf_event_pid(event, current);
3650 data->tid_entry.tid = perf_event_tid(event, current);
3653 if (sample_type & PERF_SAMPLE_TIME)
3654 data->time = perf_clock();
3656 if (sample_type & PERF_SAMPLE_ID)
3657 data->id = primary_event_id(event);
3659 if (sample_type & PERF_SAMPLE_STREAM_ID)
3660 data->stream_id = event->id;
3662 if (sample_type & PERF_SAMPLE_CPU) {
3663 data->cpu_entry.cpu = raw_smp_processor_id();
3664 data->cpu_entry.reserved = 0;
3668 void perf_event_header__init_id(struct perf_event_header *header,
3669 struct perf_sample_data *data,
3670 struct perf_event *event)
3672 if (event->attr.sample_id_all)
3673 __perf_event_header__init_id(header, data, event);
3676 static void __perf_event__output_id_sample(struct perf_output_handle *handle,
3677 struct perf_sample_data *data)
3679 u64 sample_type = data->type;
3681 if (sample_type & PERF_SAMPLE_TID)
3682 perf_output_put(handle, data->tid_entry);
3684 if (sample_type & PERF_SAMPLE_TIME)
3685 perf_output_put(handle, data->time);
3687 if (sample_type & PERF_SAMPLE_ID)
3688 perf_output_put(handle, data->id);
3690 if (sample_type & PERF_SAMPLE_STREAM_ID)
3691 perf_output_put(handle, data->stream_id);
3693 if (sample_type & PERF_SAMPLE_CPU)
3694 perf_output_put(handle, data->cpu_entry);
3697 void perf_event__output_id_sample(struct perf_event *event,
3698 struct perf_output_handle *handle,
3699 struct perf_sample_data *sample)
3701 if (event->attr.sample_id_all)
3702 __perf_event__output_id_sample(handle, sample);
3705 static void perf_output_read_one(struct perf_output_handle *handle,
3706 struct perf_event *event,
3707 u64 enabled, u64 running)
3709 u64 read_format = event->attr.read_format;
3713 values[n++] = perf_event_count(event);
3714 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
3715 values[n++] = enabled +
3716 atomic64_read(&event->child_total_time_enabled);
3718 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
3719 values[n++] = running +
3720 atomic64_read(&event->child_total_time_running);
3722 if (read_format & PERF_FORMAT_ID)
3723 values[n++] = primary_event_id(event);
3725 __output_copy(handle, values, n * sizeof(u64));
3729 * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
3731 static void perf_output_read_group(struct perf_output_handle *handle,
3732 struct perf_event *event,
3733 u64 enabled, u64 running)
3735 struct perf_event *leader = event->group_leader, *sub;
3736 u64 read_format = event->attr.read_format;
3740 values[n++] = 1 + leader->nr_siblings;
3742 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3743 values[n++] = enabled;
3745 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3746 values[n++] = running;
3748 if (leader != event)
3749 leader->pmu->read(leader);
3751 values[n++] = perf_event_count(leader);
3752 if (read_format & PERF_FORMAT_ID)
3753 values[n++] = primary_event_id(leader);
3755 __output_copy(handle, values, n * sizeof(u64));
3757 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
3761 sub->pmu->read(sub);
3763 values[n++] = perf_event_count(sub);
3764 if (read_format & PERF_FORMAT_ID)
3765 values[n++] = primary_event_id(sub);
3767 __output_copy(handle, values, n * sizeof(u64));
3771 #define PERF_FORMAT_TOTAL_TIMES (PERF_FORMAT_TOTAL_TIME_ENABLED|\
3772 PERF_FORMAT_TOTAL_TIME_RUNNING)
3774 static void perf_output_read(struct perf_output_handle *handle,
3775 struct perf_event *event)
3777 u64 enabled = 0, running = 0, now;
3778 u64 read_format = event->attr.read_format;
3781 * compute total_time_enabled, total_time_running
3782 * based on snapshot values taken when the event
3783 * was last scheduled in.
3785 * we cannot simply called update_context_time()
3786 * because of locking issue as we are called in
3789 if (read_format & PERF_FORMAT_TOTAL_TIMES)
3790 calc_timer_values(event, &now, &enabled, &running);
3792 if (event->attr.read_format & PERF_FORMAT_GROUP)
3793 perf_output_read_group(handle, event, enabled, running);
3795 perf_output_read_one(handle, event, enabled, running);
3798 void perf_output_sample(struct perf_output_handle *handle,
3799 struct perf_event_header *header,
3800 struct perf_sample_data *data,
3801 struct perf_event *event)
3803 u64 sample_type = data->type;
3805 perf_output_put(handle, *header);
3807 if (sample_type & PERF_SAMPLE_IP)
3808 perf_output_put(handle, data->ip);
3810 if (sample_type & PERF_SAMPLE_TID)
3811 perf_output_put(handle, data->tid_entry);
3813 if (sample_type & PERF_SAMPLE_TIME)
3814 perf_output_put(handle, data->time);
3816 if (sample_type & PERF_SAMPLE_ADDR)
3817 perf_output_put(handle, data->addr);
3819 if (sample_type & PERF_SAMPLE_ID)
3820 perf_output_put(handle, data->id);
3822 if (sample_type & PERF_SAMPLE_STREAM_ID)
3823 perf_output_put(handle, data->stream_id);
3825 if (sample_type & PERF_SAMPLE_CPU)
3826 perf_output_put(handle, data->cpu_entry);
3828 if (sample_type & PERF_SAMPLE_PERIOD)
3829 perf_output_put(handle, data->period);
3831 if (sample_type & PERF_SAMPLE_READ)
3832 perf_output_read(handle, event);
3834 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3835 if (data->callchain) {
3838 if (data->callchain)
3839 size += data->callchain->nr;
3841 size *= sizeof(u64);
3843 __output_copy(handle, data->callchain, size);
3846 perf_output_put(handle, nr);
3850 if (sample_type & PERF_SAMPLE_RAW) {
3852 perf_output_put(handle, data->raw->size);
3853 __output_copy(handle, data->raw->data,
3860 .size = sizeof(u32),
3863 perf_output_put(handle, raw);
3867 if (!event->attr.watermark) {
3868 int wakeup_events = event->attr.wakeup_events;
3870 if (wakeup_events) {
3871 struct ring_buffer *rb = handle->rb;
3872 int events = local_inc_return(&rb->events);
3874 if (events >= wakeup_events) {
3875 local_sub(wakeup_events, &rb->events);
3876 local_inc(&rb->wakeup);
3882 void perf_prepare_sample(struct perf_event_header *header,
3883 struct perf_sample_data *data,
3884 struct perf_event *event,
3885 struct pt_regs *regs)
3887 u64 sample_type = event->attr.sample_type;
3889 header->type = PERF_RECORD_SAMPLE;
3890 header->size = sizeof(*header) + event->header_size;
3893 header->misc |= perf_misc_flags(regs);
3895 __perf_event_header__init_id(header, data, event);
3897 if (sample_type & PERF_SAMPLE_IP)
3898 data->ip = perf_instruction_pointer(regs);
3900 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3903 data->callchain = perf_callchain(regs);
3905 if (data->callchain)
3906 size += data->callchain->nr;
3908 header->size += size * sizeof(u64);
3911 if (sample_type & PERF_SAMPLE_RAW) {
3912 int size = sizeof(u32);
3915 size += data->raw->size;
3917 size += sizeof(u32);
3919 WARN_ON_ONCE(size & (sizeof(u64)-1));
3920 header->size += size;
3924 static void perf_event_output(struct perf_event *event,
3925 struct perf_sample_data *data,
3926 struct pt_regs *regs)
3928 struct perf_output_handle handle;
3929 struct perf_event_header header;
3931 /* protect the callchain buffers */
3934 perf_prepare_sample(&header, data, event, regs);
3936 if (perf_output_begin(&handle, event, header.size))
3939 perf_output_sample(&handle, &header, data, event);
3941 perf_output_end(&handle);
3951 struct perf_read_event {
3952 struct perf_event_header header;
3959 perf_event_read_event(struct perf_event *event,
3960 struct task_struct *task)
3962 struct perf_output_handle handle;
3963 struct perf_sample_data sample;
3964 struct perf_read_event read_event = {
3966 .type = PERF_RECORD_READ,
3968 .size = sizeof(read_event) + event->read_size,
3970 .pid = perf_event_pid(event, task),
3971 .tid = perf_event_tid(event, task),
3975 perf_event_header__init_id(&read_event.header, &sample, event);
3976 ret = perf_output_begin(&handle, event, read_event.header.size);
3980 perf_output_put(&handle, read_event);
3981 perf_output_read(&handle, event);
3982 perf_event__output_id_sample(event, &handle, &sample);
3984 perf_output_end(&handle);
3988 * task tracking -- fork/exit
3990 * enabled by: attr.comm | attr.mmap | attr.mmap_data | attr.task
3993 struct perf_task_event {
3994 struct task_struct *task;
3995 struct perf_event_context *task_ctx;
3998 struct perf_event_header header;
4008 static void perf_event_task_output(struct perf_event *event,
4009 struct perf_task_event *task_event)
4011 struct perf_output_handle handle;
4012 struct perf_sample_data sample;
4013 struct task_struct *task = task_event->task;
4014 int ret, size = task_event->event_id.header.size;
4016 perf_event_header__init_id(&task_event->event_id.header, &sample, event);
4018 ret = perf_output_begin(&handle, event,
4019 task_event->event_id.header.size);
4023 task_event->event_id.pid = perf_event_pid(event, task);
4024 task_event->event_id.ppid = perf_event_pid(event, current);
4026 task_event->event_id.tid = perf_event_tid(event, task);
4027 task_event->event_id.ptid = perf_event_tid(event, current);
4029 perf_output_put(&handle, task_event->event_id);
4031 perf_event__output_id_sample(event, &handle, &sample);
4033 perf_output_end(&handle);
4035 task_event->event_id.header.size = size;
4038 static int perf_event_task_match(struct perf_event *event)
4040 if (event->state < PERF_EVENT_STATE_INACTIVE)
4043 if (!event_filter_match(event))
4046 if (event->attr.comm || event->attr.mmap ||
4047 event->attr.mmap_data || event->attr.task)
4053 static void perf_event_task_ctx(struct perf_event_context *ctx,
4054 struct perf_task_event *task_event)
4056 struct perf_event *event;
4058 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
4059 if (perf_event_task_match(event))
4060 perf_event_task_output(event, task_event);
4064 static void perf_event_task_event(struct perf_task_event *task_event)
4066 struct perf_cpu_context *cpuctx;
4067 struct perf_event_context *ctx;
4072 list_for_each_entry_rcu(pmu, &pmus, entry) {
4073 cpuctx = get_cpu_ptr(pmu->pmu_cpu_context);
4074 if (cpuctx->active_pmu != pmu)
4076 perf_event_task_ctx(&cpuctx->ctx, task_event);
4078 ctx = task_event->task_ctx;
4080 ctxn = pmu->task_ctx_nr;
4083 ctx = rcu_dereference(current->perf_event_ctxp[ctxn]);
4086 perf_event_task_ctx(ctx, task_event);
4088 put_cpu_ptr(pmu->pmu_cpu_context);
4093 static void perf_event_task(struct task_struct *task,
4094 struct perf_event_context *task_ctx,
4097 struct perf_task_event task_event;
4099 if (!atomic_read(&nr_comm_events) &&
4100 !atomic_read(&nr_mmap_events) &&
4101 !atomic_read(&nr_task_events))
4104 task_event = (struct perf_task_event){
4106 .task_ctx = task_ctx,
4109 .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT,
4111 .size = sizeof(task_event.event_id),
4117 .time = perf_clock(),
4121 perf_event_task_event(&task_event);
4124 void perf_event_fork(struct task_struct *task)
4126 perf_event_task(task, NULL, 1);
4133 struct perf_comm_event {
4134 struct task_struct *task;
4139 struct perf_event_header header;
4146 static void perf_event_comm_output(struct perf_event *event,
4147 struct perf_comm_event *comm_event)
4149 struct perf_output_handle handle;
4150 struct perf_sample_data sample;
4151 int size = comm_event->event_id.header.size;
4154 perf_event_header__init_id(&comm_event->event_id.header, &sample, event);
4155 ret = perf_output_begin(&handle, event,
4156 comm_event->event_id.header.size);
4161 comm_event->event_id.pid = perf_event_pid(event, comm_event->task);
4162 comm_event->event_id.tid = perf_event_tid(event, comm_event->task);
4164 perf_output_put(&handle, comm_event->event_id);
4165 __output_copy(&handle, comm_event->comm,
4166 comm_event->comm_size);
4168 perf_event__output_id_sample(event, &handle, &sample);
4170 perf_output_end(&handle);
4172 comm_event->event_id.header.size = size;
4175 static int perf_event_comm_match(struct perf_event *event)
4177 if (event->state < PERF_EVENT_STATE_INACTIVE)
4180 if (!event_filter_match(event))
4183 if (event->attr.comm)
4189 static void perf_event_comm_ctx(struct perf_event_context *ctx,
4190 struct perf_comm_event *comm_event)
4192 struct perf_event *event;
4194 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
4195 if (perf_event_comm_match(event))
4196 perf_event_comm_output(event, comm_event);
4200 static void perf_event_comm_event(struct perf_comm_event *comm_event)
4202 struct perf_cpu_context *cpuctx;
4203 struct perf_event_context *ctx;
4204 char comm[TASK_COMM_LEN];
4209 memset(comm, 0, sizeof(comm));
4210 strlcpy(comm, comm_event->task->comm, sizeof(comm));
4211 size = ALIGN(strlen(comm)+1, sizeof(u64));
4213 comm_event->comm = comm;
4214 comm_event->comm_size = size;
4216 comm_event->event_id.header.size = sizeof(comm_event->event_id) + size;
4218 list_for_each_entry_rcu(pmu, &pmus, entry) {
4219 cpuctx = get_cpu_ptr(pmu->pmu_cpu_context);
4220 if (cpuctx->active_pmu != pmu)
4222 perf_event_comm_ctx(&cpuctx->ctx, comm_event);
4224 ctxn = pmu->task_ctx_nr;
4228 ctx = rcu_dereference(current->perf_event_ctxp[ctxn]);
4230 perf_event_comm_ctx(ctx, comm_event);
4232 put_cpu_ptr(pmu->pmu_cpu_context);
4237 void perf_event_comm(struct task_struct *task)
4239 struct perf_comm_event comm_event;
4240 struct perf_event_context *ctx;
4243 for_each_task_context_nr(ctxn) {
4244 ctx = task->perf_event_ctxp[ctxn];
4248 perf_event_enable_on_exec(ctx);
4251 if (!atomic_read(&nr_comm_events))
4254 comm_event = (struct perf_comm_event){
4260 .type = PERF_RECORD_COMM,
4269 perf_event_comm_event(&comm_event);
4276 struct perf_mmap_event {
4277 struct vm_area_struct *vma;
4279 const char *file_name;
4283 struct perf_event_header header;
4293 static void perf_event_mmap_output(struct perf_event *event,
4294 struct perf_mmap_event *mmap_event)
4296 struct perf_output_handle handle;
4297 struct perf_sample_data sample;
4298 int size = mmap_event->event_id.header.size;
4301 perf_event_header__init_id(&mmap_event->event_id.header, &sample, event);
4302 ret = perf_output_begin(&handle, event,
4303 mmap_event->event_id.header.size);
4307 mmap_event->event_id.pid = perf_event_pid(event, current);
4308 mmap_event->event_id.tid = perf_event_tid(event, current);
4310 perf_output_put(&handle, mmap_event->event_id);
4311 __output_copy(&handle, mmap_event->file_name,
4312 mmap_event->file_size);
4314 perf_event__output_id_sample(event, &handle, &sample);
4316 perf_output_end(&handle);
4318 mmap_event->event_id.header.size = size;
4321 static int perf_event_mmap_match(struct perf_event *event,
4322 struct perf_mmap_event *mmap_event,
4325 if (event->state < PERF_EVENT_STATE_INACTIVE)
4328 if (!event_filter_match(event))
4331 if ((!executable && event->attr.mmap_data) ||
4332 (executable && event->attr.mmap))
4338 static void perf_event_mmap_ctx(struct perf_event_context *ctx,
4339 struct perf_mmap_event *mmap_event,
4342 struct perf_event *event;
4344 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
4345 if (perf_event_mmap_match(event, mmap_event, executable))
4346 perf_event_mmap_output(event, mmap_event);
4350 static void perf_event_mmap_event(struct perf_mmap_event *mmap_event)
4352 struct perf_cpu_context *cpuctx;
4353 struct perf_event_context *ctx;
4354 struct vm_area_struct *vma = mmap_event->vma;
4355 struct file *file = vma->vm_file;
4363 memset(tmp, 0, sizeof(tmp));
4367 * d_path works from the end of the rb backwards, so we
4368 * need to add enough zero bytes after the string to handle
4369 * the 64bit alignment we do later.
4371 buf = kzalloc(PATH_MAX + sizeof(u64), GFP_KERNEL);
4373 name = strncpy(tmp, "//enomem", sizeof(tmp));
4376 name = d_path(&file->f_path, buf, PATH_MAX);
4378 name = strncpy(tmp, "//toolong", sizeof(tmp));
4382 if (arch_vma_name(mmap_event->vma)) {
4383 name = strncpy(tmp, arch_vma_name(mmap_event->vma),
4389 name = strncpy(tmp, "[vdso]", sizeof(tmp));
4391 } else if (vma->vm_start <= vma->vm_mm->start_brk &&
4392 vma->vm_end >= vma->vm_mm->brk) {
4393 name = strncpy(tmp, "[heap]", sizeof(tmp));
4395 } else if (vma->vm_start <= vma->vm_mm->start_stack &&
4396 vma->vm_end >= vma->vm_mm->start_stack) {
4397 name = strncpy(tmp, "[stack]", sizeof(tmp));
4401 name = strncpy(tmp, "//anon", sizeof(tmp));
4406 size = ALIGN(strlen(name)+1, sizeof(u64));
4408 mmap_event->file_name = name;
4409 mmap_event->file_size = size;
4411 mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size;
4414 list_for_each_entry_rcu(pmu, &pmus, entry) {
4415 cpuctx = get_cpu_ptr(pmu->pmu_cpu_context);
4416 if (cpuctx->active_pmu != pmu)
4418 perf_event_mmap_ctx(&cpuctx->ctx, mmap_event,
4419 vma->vm_flags & VM_EXEC);
4421 ctxn = pmu->task_ctx_nr;
4425 ctx = rcu_dereference(current->perf_event_ctxp[ctxn]);
4427 perf_event_mmap_ctx(ctx, mmap_event,
4428 vma->vm_flags & VM_EXEC);
4431 put_cpu_ptr(pmu->pmu_cpu_context);
4438 void perf_event_mmap(struct vm_area_struct *vma)
4440 struct perf_mmap_event mmap_event;
4442 if (!atomic_read(&nr_mmap_events))
4445 mmap_event = (struct perf_mmap_event){
4451 .type = PERF_RECORD_MMAP,
4452 .misc = PERF_RECORD_MISC_USER,
4457 .start = vma->vm_start,
4458 .len = vma->vm_end - vma->vm_start,
4459 .pgoff = (u64)vma->vm_pgoff << PAGE_SHIFT,
4463 perf_event_mmap_event(&mmap_event);
4467 * IRQ throttle logging
4470 static void perf_log_throttle(struct perf_event *event, int enable)
4472 struct perf_output_handle handle;
4473 struct perf_sample_data sample;
4477 struct perf_event_header header;
4481 } throttle_event = {
4483 .type = PERF_RECORD_THROTTLE,
4485 .size = sizeof(throttle_event),
4487 .time = perf_clock(),
4488 .id = primary_event_id(event),
4489 .stream_id = event->id,
4493 throttle_event.header.type = PERF_RECORD_UNTHROTTLE;
4495 perf_event_header__init_id(&throttle_event.header, &sample, event);
4497 ret = perf_output_begin(&handle, event,
4498 throttle_event.header.size);
4502 perf_output_put(&handle, throttle_event);
4503 perf_event__output_id_sample(event, &handle, &sample);
4504 perf_output_end(&handle);
4508 * Generic event overflow handling, sampling.
4511 static int __perf_event_overflow(struct perf_event *event,
4512 int throttle, struct perf_sample_data *data,
4513 struct pt_regs *regs)
4515 int events = atomic_read(&event->event_limit);
4516 struct hw_perf_event *hwc = &event->hw;
4520 * Non-sampling counters might still use the PMI to fold short
4521 * hardware counters, ignore those.
4523 if (unlikely(!is_sampling_event(event)))
4526 if (unlikely(hwc->interrupts >= max_samples_per_tick)) {
4528 hwc->interrupts = MAX_INTERRUPTS;
4529 perf_log_throttle(event, 0);
4535 if (event->attr.freq) {
4536 u64 now = perf_clock();
4537 s64 delta = now - hwc->freq_time_stamp;
4539 hwc->freq_time_stamp = now;
4541 if (delta > 0 && delta < 2*TICK_NSEC)
4542 perf_adjust_period(event, delta, hwc->last_period);
4546 * XXX event_limit might not quite work as expected on inherited
4550 event->pending_kill = POLL_IN;
4551 if (events && atomic_dec_and_test(&event->event_limit)) {
4553 event->pending_kill = POLL_HUP;
4554 event->pending_disable = 1;
4555 irq_work_queue(&event->pending);
4558 if (event->overflow_handler)
4559 event->overflow_handler(event, data, regs);
4561 perf_event_output(event, data, regs);
4563 if (event->fasync && event->pending_kill) {
4564 event->pending_wakeup = 1;
4565 irq_work_queue(&event->pending);
4571 int perf_event_overflow(struct perf_event *event,
4572 struct perf_sample_data *data,
4573 struct pt_regs *regs)
4575 return __perf_event_overflow(event, 1, data, regs);
4579 * Generic software event infrastructure
4582 struct swevent_htable {
4583 struct swevent_hlist *swevent_hlist;
4584 struct mutex hlist_mutex;
4587 /* Recursion avoidance in each contexts */
4588 int recursion[PERF_NR_CONTEXTS];
4591 static DEFINE_PER_CPU(struct swevent_htable, swevent_htable);
4594 * We directly increment event->count and keep a second value in
4595 * event->hw.period_left to count intervals. This period event
4596 * is kept in the range [-sample_period, 0] so that we can use the
4600 static u64 perf_swevent_set_period(struct perf_event *event)
4602 struct hw_perf_event *hwc = &event->hw;
4603 u64 period = hwc->last_period;
4607 hwc->last_period = hwc->sample_period;
4610 old = val = local64_read(&hwc->period_left);
4614 nr = div64_u64(period + val, period);
4615 offset = nr * period;
4617 if (local64_cmpxchg(&hwc->period_left, old, val) != old)
4623 static void perf_swevent_overflow(struct perf_event *event, u64 overflow,
4624 struct perf_sample_data *data,
4625 struct pt_regs *regs)
4627 struct hw_perf_event *hwc = &event->hw;
4631 overflow = perf_swevent_set_period(event);
4633 if (hwc->interrupts == MAX_INTERRUPTS)
4636 for (; overflow; overflow--) {
4637 if (__perf_event_overflow(event, throttle,
4640 * We inhibit the overflow from happening when
4641 * hwc->interrupts == MAX_INTERRUPTS.
4649 static void perf_swevent_event(struct perf_event *event, u64 nr,
4650 struct perf_sample_data *data,
4651 struct pt_regs *regs)
4653 struct hw_perf_event *hwc = &event->hw;
4655 local64_add(nr, &event->count);
4660 if (!is_sampling_event(event))
4663 if ((event->attr.sample_type & PERF_SAMPLE_PERIOD) && !event->attr.freq) {
4665 return perf_swevent_overflow(event, 1, data, regs);
4667 data->period = event->hw.last_period;
4669 if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq)
4670 return perf_swevent_overflow(event, 1, data, regs);
4672 if (local64_add_negative(nr, &hwc->period_left))
4675 perf_swevent_overflow(event, 0, data, regs);
4678 static int perf_exclude_event(struct perf_event *event,
4679 struct pt_regs *regs)
4681 if (event->hw.state & PERF_HES_STOPPED)
4685 if (event->attr.exclude_user && user_mode(regs))
4688 if (event->attr.exclude_kernel && !user_mode(regs))
4695 static int perf_swevent_match(struct perf_event *event,
4696 enum perf_type_id type,
4698 struct perf_sample_data *data,
4699 struct pt_regs *regs)
4701 if (event->attr.type != type)
4704 if (event->attr.config != event_id)
4707 if (perf_exclude_event(event, regs))
4713 static inline u64 swevent_hash(u64 type, u32 event_id)
4715 u64 val = event_id | (type << 32);
4717 return hash_64(val, SWEVENT_HLIST_BITS);
4720 static inline struct hlist_head *
4721 __find_swevent_head(struct swevent_hlist *hlist, u64 type, u32 event_id)
4723 u64 hash = swevent_hash(type, event_id);
4725 return &hlist->heads[hash];
4728 /* For the read side: events when they trigger */
4729 static inline struct hlist_head *
4730 find_swevent_head_rcu(struct swevent_htable *swhash, u64 type, u32 event_id)
4732 struct swevent_hlist *hlist;
4734 hlist = rcu_dereference(swhash->swevent_hlist);
4738 return __find_swevent_head(hlist, type, event_id);
4741 /* For the event head insertion and removal in the hlist */
4742 static inline struct hlist_head *
4743 find_swevent_head(struct swevent_htable *swhash, struct perf_event *event)
4745 struct swevent_hlist *hlist;
4746 u32 event_id = event->attr.config;
4747 u64 type = event->attr.type;
4750 * Event scheduling is always serialized against hlist allocation
4751 * and release. Which makes the protected version suitable here.
4752 * The context lock guarantees that.
4754 hlist = rcu_dereference_protected(swhash->swevent_hlist,
4755 lockdep_is_held(&event->ctx->lock));
4759 return __find_swevent_head(hlist, type, event_id);
4762 static void do_perf_sw_event(enum perf_type_id type, u32 event_id,
4764 struct perf_sample_data *data,
4765 struct pt_regs *regs)
4767 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
4768 struct perf_event *event;
4769 struct hlist_node *node;
4770 struct hlist_head *head;
4773 head = find_swevent_head_rcu(swhash, type, event_id);
4777 hlist_for_each_entry_rcu(event, node, head, hlist_entry) {
4778 if (perf_swevent_match(event, type, event_id, data, regs))
4779 perf_swevent_event(event, nr, data, regs);
4785 int perf_swevent_get_recursion_context(void)
4787 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
4789 return get_recursion_context(swhash->recursion);
4791 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context);
4793 inline void perf_swevent_put_recursion_context(int rctx)
4795 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
4797 put_recursion_context(swhash->recursion, rctx);
4800 void __perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)
4802 struct perf_sample_data data;
4805 preempt_disable_notrace();
4806 rctx = perf_swevent_get_recursion_context();
4810 perf_sample_data_init(&data, addr);
4812 do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, &data, regs);
4814 perf_swevent_put_recursion_context(rctx);
4815 preempt_enable_notrace();
4818 static void perf_swevent_read(struct perf_event *event)
4822 static int perf_swevent_add(struct perf_event *event, int flags)
4824 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
4825 struct hw_perf_event *hwc = &event->hw;
4826 struct hlist_head *head;
4828 if (is_sampling_event(event)) {
4829 hwc->last_period = hwc->sample_period;
4830 perf_swevent_set_period(event);
4833 hwc->state = !(flags & PERF_EF_START);
4835 head = find_swevent_head(swhash, event);
4836 if (WARN_ON_ONCE(!head))
4839 hlist_add_head_rcu(&event->hlist_entry, head);
4844 static void perf_swevent_del(struct perf_event *event, int flags)
4846 hlist_del_rcu(&event->hlist_entry);
4849 static void perf_swevent_start(struct perf_event *event, int flags)
4851 event->hw.state = 0;
4854 static void perf_swevent_stop(struct perf_event *event, int flags)
4856 event->hw.state = PERF_HES_STOPPED;
4859 /* Deref the hlist from the update side */
4860 static inline struct swevent_hlist *
4861 swevent_hlist_deref(struct swevent_htable *swhash)
4863 return rcu_dereference_protected(swhash->swevent_hlist,
4864 lockdep_is_held(&swhash->hlist_mutex));
4867 static void swevent_hlist_release(struct swevent_htable *swhash)
4869 struct swevent_hlist *hlist = swevent_hlist_deref(swhash);
4874 rcu_assign_pointer(swhash->swevent_hlist, NULL);
4875 kfree_rcu(hlist, rcu_head);
4878 static void swevent_hlist_put_cpu(struct perf_event *event, int cpu)
4880 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
4882 mutex_lock(&swhash->hlist_mutex);
4884 if (!--swhash->hlist_refcount)
4885 swevent_hlist_release(swhash);
4887 mutex_unlock(&swhash->hlist_mutex);
4890 static void swevent_hlist_put(struct perf_event *event)
4894 if (event->cpu != -1) {
4895 swevent_hlist_put_cpu(event, event->cpu);
4899 for_each_possible_cpu(cpu)
4900 swevent_hlist_put_cpu(event, cpu);
4903 static int swevent_hlist_get_cpu(struct perf_event *event, int cpu)
4905 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
4908 mutex_lock(&swhash->hlist_mutex);
4910 if (!swevent_hlist_deref(swhash) && cpu_online(cpu)) {
4911 struct swevent_hlist *hlist;
4913 hlist = kzalloc(sizeof(*hlist), GFP_KERNEL);
4918 rcu_assign_pointer(swhash->swevent_hlist, hlist);
4920 swhash->hlist_refcount++;
4922 mutex_unlock(&swhash->hlist_mutex);
4927 static int swevent_hlist_get(struct perf_event *event)
4930 int cpu, failed_cpu;
4932 if (event->cpu != -1)
4933 return swevent_hlist_get_cpu(event, event->cpu);
4936 for_each_possible_cpu(cpu) {
4937 err = swevent_hlist_get_cpu(event, cpu);
4947 for_each_possible_cpu(cpu) {
4948 if (cpu == failed_cpu)
4950 swevent_hlist_put_cpu(event, cpu);
4957 struct jump_label_key perf_swevent_enabled[PERF_COUNT_SW_MAX];
4959 static void sw_perf_event_destroy(struct perf_event *event)
4961 u64 event_id = event->attr.config;
4963 WARN_ON(event->parent);
4965 jump_label_dec(&perf_swevent_enabled[event_id]);
4966 swevent_hlist_put(event);
4969 static int perf_swevent_init(struct perf_event *event)
4971 int event_id = event->attr.config;
4973 if (event->attr.type != PERF_TYPE_SOFTWARE)
4977 case PERF_COUNT_SW_CPU_CLOCK:
4978 case PERF_COUNT_SW_TASK_CLOCK:
4985 if (event_id >= PERF_COUNT_SW_MAX)
4988 if (!event->parent) {
4991 err = swevent_hlist_get(event);
4995 jump_label_inc(&perf_swevent_enabled[event_id]);
4996 event->destroy = sw_perf_event_destroy;
5002 static int perf_swevent_event_idx(struct perf_event *event)
5007 static struct pmu perf_swevent = {
5008 .task_ctx_nr = perf_sw_context,
5010 .event_init = perf_swevent_init,
5011 .add = perf_swevent_add,
5012 .del = perf_swevent_del,
5013 .start = perf_swevent_start,
5014 .stop = perf_swevent_stop,
5015 .read = perf_swevent_read,
5017 .event_idx = perf_swevent_event_idx,
5020 #ifdef CONFIG_EVENT_TRACING
5022 static int perf_tp_filter_match(struct perf_event *event,
5023 struct perf_sample_data *data)
5025 void *record = data->raw->data;
5027 if (likely(!event->filter) || filter_match_preds(event->filter, record))
5032 static int perf_tp_event_match(struct perf_event *event,
5033 struct perf_sample_data *data,
5034 struct pt_regs *regs)
5036 if (event->hw.state & PERF_HES_STOPPED)
5039 * All tracepoints are from kernel-space.
5041 if (event->attr.exclude_kernel)
5044 if (!perf_tp_filter_match(event, data))
5050 void perf_tp_event(u64 addr, u64 count, void *record, int entry_size,
5051 struct pt_regs *regs, struct hlist_head *head, int rctx)
5053 struct perf_sample_data data;
5054 struct perf_event *event;
5055 struct hlist_node *node;
5057 struct perf_raw_record raw = {
5062 perf_sample_data_init(&data, addr);
5065 hlist_for_each_entry_rcu(event, node, head, hlist_entry) {
5066 if (perf_tp_event_match(event, &data, regs))
5067 perf_swevent_event(event, count, &data, regs);
5070 perf_swevent_put_recursion_context(rctx);
5072 EXPORT_SYMBOL_GPL(perf_tp_event);
5074 static void tp_perf_event_destroy(struct perf_event *event)
5076 perf_trace_destroy(event);
5079 static int perf_tp_event_init(struct perf_event *event)
5083 if (event->attr.type != PERF_TYPE_TRACEPOINT)
5086 err = perf_trace_init(event);
5090 event->destroy = tp_perf_event_destroy;
5095 static struct pmu perf_tracepoint = {
5096 .task_ctx_nr = perf_sw_context,
5098 .event_init = perf_tp_event_init,
5099 .add = perf_trace_add,
5100 .del = perf_trace_del,
5101 .start = perf_swevent_start,
5102 .stop = perf_swevent_stop,
5103 .read = perf_swevent_read,
5105 .event_idx = perf_swevent_event_idx,
5108 static inline void perf_tp_register(void)
5110 perf_pmu_register(&perf_tracepoint, "tracepoint", PERF_TYPE_TRACEPOINT);
5113 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
5118 if (event->attr.type != PERF_TYPE_TRACEPOINT)
5121 filter_str = strndup_user(arg, PAGE_SIZE);
5122 if (IS_ERR(filter_str))
5123 return PTR_ERR(filter_str);
5125 ret = ftrace_profile_set_filter(event, event->attr.config, filter_str);
5131 static void perf_event_free_filter(struct perf_event *event)
5133 ftrace_profile_free_filter(event);
5138 static inline void perf_tp_register(void)
5142 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
5147 static void perf_event_free_filter(struct perf_event *event)
5151 #endif /* CONFIG_EVENT_TRACING */
5153 #ifdef CONFIG_HAVE_HW_BREAKPOINT
5154 void perf_bp_event(struct perf_event *bp, void *data)
5156 struct perf_sample_data sample;
5157 struct pt_regs *regs = data;
5159 perf_sample_data_init(&sample, bp->attr.bp_addr);
5161 if (!bp->hw.state && !perf_exclude_event(bp, regs))
5162 perf_swevent_event(bp, 1, &sample, regs);
5167 * hrtimer based swevent callback
5170 static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer)
5172 enum hrtimer_restart ret = HRTIMER_RESTART;
5173 struct perf_sample_data data;
5174 struct pt_regs *regs;
5175 struct perf_event *event;
5178 event = container_of(hrtimer, struct perf_event, hw.hrtimer);
5180 if (event->state != PERF_EVENT_STATE_ACTIVE)
5181 return HRTIMER_NORESTART;
5183 event->pmu->read(event);
5185 perf_sample_data_init(&data, 0);
5186 data.period = event->hw.last_period;
5187 regs = get_irq_regs();
5189 if (regs && !perf_exclude_event(event, regs)) {
5190 if (!(event->attr.exclude_idle && is_idle_task(current)))
5191 if (perf_event_overflow(event, &data, regs))
5192 ret = HRTIMER_NORESTART;
5195 period = max_t(u64, 10000, event->hw.sample_period);
5196 hrtimer_forward_now(hrtimer, ns_to_ktime(period));
5201 static void perf_swevent_start_hrtimer(struct perf_event *event)
5203 struct hw_perf_event *hwc = &event->hw;
5206 if (!is_sampling_event(event))
5209 period = local64_read(&hwc->period_left);
5214 local64_set(&hwc->period_left, 0);
5216 period = max_t(u64, 10000, hwc->sample_period);
5218 __hrtimer_start_range_ns(&hwc->hrtimer,
5219 ns_to_ktime(period), 0,
5220 HRTIMER_MODE_REL_PINNED, 0);
5223 static void perf_swevent_cancel_hrtimer(struct perf_event *event)
5225 struct hw_perf_event *hwc = &event->hw;
5227 if (is_sampling_event(event)) {
5228 ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer);
5229 local64_set(&hwc->period_left, ktime_to_ns(remaining));
5231 hrtimer_cancel(&hwc->hrtimer);
5235 static void perf_swevent_init_hrtimer(struct perf_event *event)
5237 struct hw_perf_event *hwc = &event->hw;
5239 if (!is_sampling_event(event))
5242 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
5243 hwc->hrtimer.function = perf_swevent_hrtimer;
5246 * Since hrtimers have a fixed rate, we can do a static freq->period
5247 * mapping and avoid the whole period adjust feedback stuff.
5249 if (event->attr.freq) {
5250 long freq = event->attr.sample_freq;
5252 event->attr.sample_period = NSEC_PER_SEC / freq;
5253 hwc->sample_period = event->attr.sample_period;
5254 local64_set(&hwc->period_left, hwc->sample_period);
5255 event->attr.freq = 0;
5260 * Software event: cpu wall time clock
5263 static void cpu_clock_event_update(struct perf_event *event)
5268 now = local_clock();
5269 prev = local64_xchg(&event->hw.prev_count, now);
5270 local64_add(now - prev, &event->count);
5273 static void cpu_clock_event_start(struct perf_event *event, int flags)
5275 local64_set(&event->hw.prev_count, local_clock());
5276 perf_swevent_start_hrtimer(event);
5279 static void cpu_clock_event_stop(struct perf_event *event, int flags)
5281 perf_swevent_cancel_hrtimer(event);
5282 cpu_clock_event_update(event);
5285 static int cpu_clock_event_add(struct perf_event *event, int flags)
5287 if (flags & PERF_EF_START)
5288 cpu_clock_event_start(event, flags);
5293 static void cpu_clock_event_del(struct perf_event *event, int flags)
5295 cpu_clock_event_stop(event, flags);
5298 static void cpu_clock_event_read(struct perf_event *event)
5300 cpu_clock_event_update(event);
5303 static int cpu_clock_event_init(struct perf_event *event)
5305 if (event->attr.type != PERF_TYPE_SOFTWARE)
5308 if (event->attr.config != PERF_COUNT_SW_CPU_CLOCK)
5311 perf_swevent_init_hrtimer(event);
5316 static struct pmu perf_cpu_clock = {
5317 .task_ctx_nr = perf_sw_context,
5319 .event_init = cpu_clock_event_init,
5320 .add = cpu_clock_event_add,
5321 .del = cpu_clock_event_del,
5322 .start = cpu_clock_event_start,
5323 .stop = cpu_clock_event_stop,
5324 .read = cpu_clock_event_read,
5326 .event_idx = perf_swevent_event_idx,
5330 * Software event: task time clock
5333 static void task_clock_event_update(struct perf_event *event, u64 now)
5338 prev = local64_xchg(&event->hw.prev_count, now);
5340 local64_add(delta, &event->count);
5343 static void task_clock_event_start(struct perf_event *event, int flags)
5345 local64_set(&event->hw.prev_count, event->ctx->time);
5346 perf_swevent_start_hrtimer(event);
5349 static void task_clock_event_stop(struct perf_event *event, int flags)
5351 perf_swevent_cancel_hrtimer(event);
5352 task_clock_event_update(event, event->ctx->time);
5355 static int task_clock_event_add(struct perf_event *event, int flags)
5357 if (flags & PERF_EF_START)
5358 task_clock_event_start(event, flags);
5363 static void task_clock_event_del(struct perf_event *event, int flags)
5365 task_clock_event_stop(event, PERF_EF_UPDATE);
5368 static void task_clock_event_read(struct perf_event *event)
5370 u64 now = perf_clock();
5371 u64 delta = now - event->ctx->timestamp;
5372 u64 time = event->ctx->time + delta;
5374 task_clock_event_update(event, time);
5377 static int task_clock_event_init(struct perf_event *event)
5379 if (event->attr.type != PERF_TYPE_SOFTWARE)
5382 if (event->attr.config != PERF_COUNT_SW_TASK_CLOCK)
5385 perf_swevent_init_hrtimer(event);
5390 static struct pmu perf_task_clock = {
5391 .task_ctx_nr = perf_sw_context,
5393 .event_init = task_clock_event_init,
5394 .add = task_clock_event_add,
5395 .del = task_clock_event_del,
5396 .start = task_clock_event_start,
5397 .stop = task_clock_event_stop,
5398 .read = task_clock_event_read,
5400 .event_idx = perf_swevent_event_idx,
5403 static void perf_pmu_nop_void(struct pmu *pmu)
5407 static int perf_pmu_nop_int(struct pmu *pmu)
5412 static void perf_pmu_start_txn(struct pmu *pmu)
5414 perf_pmu_disable(pmu);
5417 static int perf_pmu_commit_txn(struct pmu *pmu)
5419 perf_pmu_enable(pmu);
5423 static void perf_pmu_cancel_txn(struct pmu *pmu)
5425 perf_pmu_enable(pmu);
5428 static int perf_event_idx_default(struct perf_event *event)
5430 return event->hw.idx + 1;
5434 * Ensures all contexts with the same task_ctx_nr have the same
5435 * pmu_cpu_context too.
5437 static void *find_pmu_context(int ctxn)
5444 list_for_each_entry(pmu, &pmus, entry) {
5445 if (pmu->task_ctx_nr == ctxn)
5446 return pmu->pmu_cpu_context;
5452 static void update_pmu_context(struct pmu *pmu, struct pmu *old_pmu)
5456 for_each_possible_cpu(cpu) {
5457 struct perf_cpu_context *cpuctx;
5459 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
5461 if (cpuctx->active_pmu == old_pmu)
5462 cpuctx->active_pmu = pmu;
5466 static void free_pmu_context(struct pmu *pmu)
5470 mutex_lock(&pmus_lock);
5472 * Like a real lame refcount.
5474 list_for_each_entry(i, &pmus, entry) {
5475 if (i->pmu_cpu_context == pmu->pmu_cpu_context) {
5476 update_pmu_context(i, pmu);
5481 free_percpu(pmu->pmu_cpu_context);
5483 mutex_unlock(&pmus_lock);
5485 static struct idr pmu_idr;
5488 type_show(struct device *dev, struct device_attribute *attr, char *page)
5490 struct pmu *pmu = dev_get_drvdata(dev);
5492 return snprintf(page, PAGE_SIZE-1, "%d\n", pmu->type);
5495 static struct device_attribute pmu_dev_attrs[] = {
5500 static int pmu_bus_running;
5501 static struct bus_type pmu_bus = {
5502 .name = "event_source",
5503 .dev_attrs = pmu_dev_attrs,
5506 static void pmu_dev_release(struct device *dev)
5511 static int pmu_dev_alloc(struct pmu *pmu)
5515 pmu->dev = kzalloc(sizeof(struct device), GFP_KERNEL);
5519 pmu->dev->groups = pmu->attr_groups;
5520 device_initialize(pmu->dev);
5521 ret = dev_set_name(pmu->dev, "%s", pmu->name);
5525 dev_set_drvdata(pmu->dev, pmu);
5526 pmu->dev->bus = &pmu_bus;
5527 pmu->dev->release = pmu_dev_release;
5528 ret = device_add(pmu->dev);
5536 put_device(pmu->dev);
5540 static struct lock_class_key cpuctx_mutex;
5541 static struct lock_class_key cpuctx_lock;
5543 int perf_pmu_register(struct pmu *pmu, char *name, int type)
5547 mutex_lock(&pmus_lock);
5549 pmu->pmu_disable_count = alloc_percpu(int);
5550 if (!pmu->pmu_disable_count)
5559 int err = idr_pre_get(&pmu_idr, GFP_KERNEL);
5563 err = idr_get_new_above(&pmu_idr, pmu, PERF_TYPE_MAX, &type);
5571 if (pmu_bus_running) {
5572 ret = pmu_dev_alloc(pmu);
5578 pmu->pmu_cpu_context = find_pmu_context(pmu->task_ctx_nr);
5579 if (pmu->pmu_cpu_context)
5580 goto got_cpu_context;
5582 pmu->pmu_cpu_context = alloc_percpu(struct perf_cpu_context);
5583 if (!pmu->pmu_cpu_context)
5586 for_each_possible_cpu(cpu) {
5587 struct perf_cpu_context *cpuctx;
5589 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
5590 __perf_event_init_context(&cpuctx->ctx);
5591 lockdep_set_class(&cpuctx->ctx.mutex, &cpuctx_mutex);
5592 lockdep_set_class(&cpuctx->ctx.lock, &cpuctx_lock);
5593 cpuctx->ctx.type = cpu_context;
5594 cpuctx->ctx.pmu = pmu;
5595 cpuctx->jiffies_interval = 1;
5596 INIT_LIST_HEAD(&cpuctx->rotation_list);
5597 cpuctx->active_pmu = pmu;
5601 if (!pmu->start_txn) {
5602 if (pmu->pmu_enable) {
5604 * If we have pmu_enable/pmu_disable calls, install
5605 * transaction stubs that use that to try and batch
5606 * hardware accesses.
5608 pmu->start_txn = perf_pmu_start_txn;
5609 pmu->commit_txn = perf_pmu_commit_txn;
5610 pmu->cancel_txn = perf_pmu_cancel_txn;
5612 pmu->start_txn = perf_pmu_nop_void;
5613 pmu->commit_txn = perf_pmu_nop_int;
5614 pmu->cancel_txn = perf_pmu_nop_void;
5618 if (!pmu->pmu_enable) {
5619 pmu->pmu_enable = perf_pmu_nop_void;
5620 pmu->pmu_disable = perf_pmu_nop_void;
5623 if (!pmu->event_idx)
5624 pmu->event_idx = perf_event_idx_default;
5626 list_add_rcu(&pmu->entry, &pmus);
5629 mutex_unlock(&pmus_lock);
5634 device_del(pmu->dev);
5635 put_device(pmu->dev);
5638 if (pmu->type >= PERF_TYPE_MAX)
5639 idr_remove(&pmu_idr, pmu->type);
5642 free_percpu(pmu->pmu_disable_count);
5646 void perf_pmu_unregister(struct pmu *pmu)
5648 mutex_lock(&pmus_lock);
5649 list_del_rcu(&pmu->entry);
5650 mutex_unlock(&pmus_lock);
5653 * We dereference the pmu list under both SRCU and regular RCU, so
5654 * synchronize against both of those.
5656 synchronize_srcu(&pmus_srcu);
5659 free_percpu(pmu->pmu_disable_count);
5660 if (pmu->type >= PERF_TYPE_MAX)
5661 idr_remove(&pmu_idr, pmu->type);
5662 device_del(pmu->dev);
5663 put_device(pmu->dev);
5664 free_pmu_context(pmu);
5667 struct pmu *perf_init_event(struct perf_event *event)
5669 struct pmu *pmu = NULL;
5673 idx = srcu_read_lock(&pmus_srcu);
5676 pmu = idr_find(&pmu_idr, event->attr.type);
5680 ret = pmu->event_init(event);
5686 list_for_each_entry_rcu(pmu, &pmus, entry) {
5688 ret = pmu->event_init(event);
5692 if (ret != -ENOENT) {
5697 pmu = ERR_PTR(-ENOENT);
5699 srcu_read_unlock(&pmus_srcu, idx);
5705 * Allocate and initialize a event structure
5707 static struct perf_event *
5708 perf_event_alloc(struct perf_event_attr *attr, int cpu,
5709 struct task_struct *task,
5710 struct perf_event *group_leader,
5711 struct perf_event *parent_event,
5712 perf_overflow_handler_t overflow_handler,
5716 struct perf_event *event;
5717 struct hw_perf_event *hwc;
5720 if ((unsigned)cpu >= nr_cpu_ids) {
5721 if (!task || cpu != -1)
5722 return ERR_PTR(-EINVAL);
5725 event = kzalloc(sizeof(*event), GFP_KERNEL);
5727 return ERR_PTR(-ENOMEM);
5730 * Single events are their own group leaders, with an
5731 * empty sibling list:
5734 group_leader = event;
5736 mutex_init(&event->child_mutex);
5737 INIT_LIST_HEAD(&event->child_list);
5739 INIT_LIST_HEAD(&event->group_entry);
5740 INIT_LIST_HEAD(&event->event_entry);
5741 INIT_LIST_HEAD(&event->sibling_list);
5742 INIT_LIST_HEAD(&event->rb_entry);
5744 init_waitqueue_head(&event->waitq);
5745 init_irq_work(&event->pending, perf_pending_event);
5747 mutex_init(&event->mmap_mutex);
5750 event->attr = *attr;
5751 event->group_leader = group_leader;
5755 event->parent = parent_event;
5757 event->ns = get_pid_ns(current->nsproxy->pid_ns);
5758 event->id = atomic64_inc_return(&perf_event_id);
5760 event->state = PERF_EVENT_STATE_INACTIVE;
5763 event->attach_state = PERF_ATTACH_TASK;
5764 #ifdef CONFIG_HAVE_HW_BREAKPOINT
5766 * hw_breakpoint is a bit difficult here..
5768 if (attr->type == PERF_TYPE_BREAKPOINT)
5769 event->hw.bp_target = task;
5773 if (!overflow_handler && parent_event) {
5774 overflow_handler = parent_event->overflow_handler;
5775 context = parent_event->overflow_handler_context;
5778 event->overflow_handler = overflow_handler;
5779 event->overflow_handler_context = context;
5782 event->state = PERF_EVENT_STATE_OFF;
5787 hwc->sample_period = attr->sample_period;
5788 if (attr->freq && attr->sample_freq)
5789 hwc->sample_period = 1;
5790 hwc->last_period = hwc->sample_period;
5792 local64_set(&hwc->period_left, hwc->sample_period);
5795 * we currently do not support PERF_FORMAT_GROUP on inherited events
5797 if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP))
5800 pmu = perf_init_event(event);
5806 else if (IS_ERR(pmu))
5811 put_pid_ns(event->ns);
5813 return ERR_PTR(err);
5816 if (!event->parent) {
5817 if (event->attach_state & PERF_ATTACH_TASK)
5818 jump_label_inc(&perf_sched_events.key);
5819 if (event->attr.mmap || event->attr.mmap_data)
5820 atomic_inc(&nr_mmap_events);
5821 if (event->attr.comm)
5822 atomic_inc(&nr_comm_events);
5823 if (event->attr.task)
5824 atomic_inc(&nr_task_events);
5825 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) {
5826 err = get_callchain_buffers();
5829 return ERR_PTR(err);
5837 static int perf_copy_attr(struct perf_event_attr __user *uattr,
5838 struct perf_event_attr *attr)
5843 if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0))
5847 * zero the full structure, so that a short copy will be nice.
5849 memset(attr, 0, sizeof(*attr));
5851 ret = get_user(size, &uattr->size);
5855 if (size > PAGE_SIZE) /* silly large */
5858 if (!size) /* abi compat */
5859 size = PERF_ATTR_SIZE_VER0;
5861 if (size < PERF_ATTR_SIZE_VER0)
5865 * If we're handed a bigger struct than we know of,
5866 * ensure all the unknown bits are 0 - i.e. new
5867 * user-space does not rely on any kernel feature
5868 * extensions we dont know about yet.
5870 if (size > sizeof(*attr)) {
5871 unsigned char __user *addr;
5872 unsigned char __user *end;
5875 addr = (void __user *)uattr + sizeof(*attr);
5876 end = (void __user *)uattr + size;
5878 for (; addr < end; addr++) {
5879 ret = get_user(val, addr);
5885 size = sizeof(*attr);
5888 ret = copy_from_user(attr, uattr, size);
5892 if (attr->__reserved_1)
5895 if (attr->sample_type & ~(PERF_SAMPLE_MAX-1))
5898 if (attr->read_format & ~(PERF_FORMAT_MAX-1))
5905 put_user(sizeof(*attr), &uattr->size);
5911 perf_event_set_output(struct perf_event *event, struct perf_event *output_event)
5913 struct ring_buffer *rb = NULL, *old_rb = NULL;
5919 /* don't allow circular references */
5920 if (event == output_event)
5924 * Don't allow cross-cpu buffers
5926 if (output_event->cpu != event->cpu)
5930 * If its not a per-cpu rb, it must be the same task.
5932 if (output_event->cpu == -1 && output_event->ctx != event->ctx)
5936 mutex_lock(&event->mmap_mutex);
5937 /* Can't redirect output if we've got an active mmap() */
5938 if (atomic_read(&event->mmap_count))
5942 /* get the rb we want to redirect to */
5943 rb = ring_buffer_get(output_event);
5949 rcu_assign_pointer(event->rb, rb);
5951 ring_buffer_detach(event, old_rb);
5954 mutex_unlock(&event->mmap_mutex);
5957 ring_buffer_put(old_rb);
5963 * sys_perf_event_open - open a performance event, associate it to a task/cpu
5965 * @attr_uptr: event_id type attributes for monitoring/sampling
5968 * @group_fd: group leader event fd
5970 SYSCALL_DEFINE5(perf_event_open,
5971 struct perf_event_attr __user *, attr_uptr,
5972 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
5974 struct perf_event *group_leader = NULL, *output_event = NULL;
5975 struct perf_event *event, *sibling;
5976 struct perf_event_attr attr;
5977 struct perf_event_context *ctx;
5978 struct file *event_file = NULL;
5979 struct file *group_file = NULL;
5980 struct task_struct *task = NULL;
5984 int fput_needed = 0;
5987 /* for future expandability... */
5988 if (flags & ~PERF_FLAG_ALL)
5991 err = perf_copy_attr(attr_uptr, &attr);
5995 if (!attr.exclude_kernel) {
5996 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
6001 if (attr.sample_freq > sysctl_perf_event_sample_rate)
6006 * In cgroup mode, the pid argument is used to pass the fd
6007 * opened to the cgroup directory in cgroupfs. The cpu argument
6008 * designates the cpu on which to monitor threads from that
6011 if ((flags & PERF_FLAG_PID_CGROUP) && (pid == -1 || cpu == -1))
6014 event_fd = get_unused_fd_flags(O_RDWR);
6018 if (group_fd != -1) {
6019 group_leader = perf_fget_light(group_fd, &fput_needed);
6020 if (IS_ERR(group_leader)) {
6021 err = PTR_ERR(group_leader);
6024 group_file = group_leader->filp;
6025 if (flags & PERF_FLAG_FD_OUTPUT)
6026 output_event = group_leader;
6027 if (flags & PERF_FLAG_FD_NO_GROUP)
6028 group_leader = NULL;
6031 if (pid != -1 && !(flags & PERF_FLAG_PID_CGROUP)) {
6032 task = find_lively_task_by_vpid(pid);
6034 err = PTR_ERR(task);
6039 event = perf_event_alloc(&attr, cpu, task, group_leader, NULL,
6041 if (IS_ERR(event)) {
6042 err = PTR_ERR(event);
6046 if (flags & PERF_FLAG_PID_CGROUP) {
6047 err = perf_cgroup_connect(pid, event, &attr, group_leader);
6052 * - that has cgroup constraint on event->cpu
6053 * - that may need work on context switch
6055 atomic_inc(&per_cpu(perf_cgroup_events, event->cpu));
6056 jump_label_inc(&perf_sched_events.key);
6060 * Special case software events and allow them to be part of
6061 * any hardware group.
6066 (is_software_event(event) != is_software_event(group_leader))) {
6067 if (is_software_event(event)) {
6069 * If event and group_leader are not both a software
6070 * event, and event is, then group leader is not.
6072 * Allow the addition of software events to !software
6073 * groups, this is safe because software events never
6076 pmu = group_leader->pmu;
6077 } else if (is_software_event(group_leader) &&
6078 (group_leader->group_flags & PERF_GROUP_SOFTWARE)) {
6080 * In case the group is a pure software group, and we
6081 * try to add a hardware event, move the whole group to
6082 * the hardware context.
6089 * Get the target context (task or percpu):
6091 ctx = find_get_context(pmu, task, cpu);
6098 put_task_struct(task);
6103 * Look up the group leader (we will attach this event to it):
6109 * Do not allow a recursive hierarchy (this new sibling
6110 * becoming part of another group-sibling):
6112 if (group_leader->group_leader != group_leader)
6115 * Do not allow to attach to a group in a different
6116 * task or CPU context:
6119 if (group_leader->ctx->type != ctx->type)
6122 if (group_leader->ctx != ctx)
6127 * Only a group leader can be exclusive or pinned
6129 if (attr.exclusive || attr.pinned)
6134 err = perf_event_set_output(event, output_event);
6139 event_file = anon_inode_getfile("[perf_event]", &perf_fops, event, O_RDWR);
6140 if (IS_ERR(event_file)) {
6141 err = PTR_ERR(event_file);
6146 struct perf_event_context *gctx = group_leader->ctx;
6148 mutex_lock(&gctx->mutex);
6149 perf_remove_from_context(group_leader);
6150 list_for_each_entry(sibling, &group_leader->sibling_list,
6152 perf_remove_from_context(sibling);
6155 mutex_unlock(&gctx->mutex);
6159 event->filp = event_file;
6160 WARN_ON_ONCE(ctx->parent_ctx);
6161 mutex_lock(&ctx->mutex);
6164 perf_install_in_context(ctx, group_leader, cpu);
6166 list_for_each_entry(sibling, &group_leader->sibling_list,
6168 perf_install_in_context(ctx, sibling, cpu);
6173 perf_install_in_context(ctx, event, cpu);
6175 perf_unpin_context(ctx);
6176 mutex_unlock(&ctx->mutex);
6178 event->owner = current;
6180 mutex_lock(¤t->perf_event_mutex);
6181 list_add_tail(&event->owner_entry, ¤t->perf_event_list);
6182 mutex_unlock(¤t->perf_event_mutex);
6185 * Precalculate sample_data sizes
6187 perf_event__header_size(event);
6188 perf_event__id_header_size(event);
6191 * Drop the reference on the group_event after placing the
6192 * new event on the sibling_list. This ensures destruction
6193 * of the group leader will find the pointer to itself in
6194 * perf_group_detach().
6196 fput_light(group_file, fput_needed);
6197 fd_install(event_fd, event_file);
6201 perf_unpin_context(ctx);
6207 put_task_struct(task);
6209 fput_light(group_file, fput_needed);
6211 put_unused_fd(event_fd);
6216 * perf_event_create_kernel_counter
6218 * @attr: attributes of the counter to create
6219 * @cpu: cpu in which the counter is bound
6220 * @task: task to profile (NULL for percpu)
6223 perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu,
6224 struct task_struct *task,
6225 perf_overflow_handler_t overflow_handler,
6228 struct perf_event_context *ctx;
6229 struct perf_event *event;
6233 * Get the target context (task or percpu):
6236 event = perf_event_alloc(attr, cpu, task, NULL, NULL,
6237 overflow_handler, context);
6238 if (IS_ERR(event)) {
6239 err = PTR_ERR(event);
6243 ctx = find_get_context(event->pmu, task, cpu);
6250 WARN_ON_ONCE(ctx->parent_ctx);
6251 mutex_lock(&ctx->mutex);
6252 perf_install_in_context(ctx, event, cpu);
6254 perf_unpin_context(ctx);
6255 mutex_unlock(&ctx->mutex);
6262 return ERR_PTR(err);
6264 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter);
6266 static void sync_child_event(struct perf_event *child_event,
6267 struct task_struct *child)
6269 struct perf_event *parent_event = child_event->parent;
6272 if (child_event->attr.inherit_stat)
6273 perf_event_read_event(child_event, child);
6275 child_val = perf_event_count(child_event);
6278 * Add back the child's count to the parent's count:
6280 atomic64_add(child_val, &parent_event->child_count);
6281 atomic64_add(child_event->total_time_enabled,
6282 &parent_event->child_total_time_enabled);
6283 atomic64_add(child_event->total_time_running,
6284 &parent_event->child_total_time_running);
6287 * Remove this event from the parent's list
6289 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
6290 mutex_lock(&parent_event->child_mutex);
6291 list_del_init(&child_event->child_list);
6292 mutex_unlock(&parent_event->child_mutex);
6295 * Release the parent event, if this was the last
6298 fput(parent_event->filp);
6302 __perf_event_exit_task(struct perf_event *child_event,
6303 struct perf_event_context *child_ctx,
6304 struct task_struct *child)
6306 if (child_event->parent) {
6307 raw_spin_lock_irq(&child_ctx->lock);
6308 perf_group_detach(child_event);
6309 raw_spin_unlock_irq(&child_ctx->lock);
6312 perf_remove_from_context(child_event);
6315 * It can happen that the parent exits first, and has events
6316 * that are still around due to the child reference. These
6317 * events need to be zapped.
6319 if (child_event->parent) {
6320 sync_child_event(child_event, child);
6321 free_event(child_event);
6325 static void perf_event_exit_task_context(struct task_struct *child, int ctxn)
6327 struct perf_event *child_event, *tmp;
6328 struct perf_event_context *child_ctx;
6329 unsigned long flags;
6331 if (likely(!child->perf_event_ctxp[ctxn])) {
6332 perf_event_task(child, NULL, 0);
6336 local_irq_save(flags);
6338 * We can't reschedule here because interrupts are disabled,
6339 * and either child is current or it is a task that can't be
6340 * scheduled, so we are now safe from rescheduling changing
6343 child_ctx = rcu_dereference_raw(child->perf_event_ctxp[ctxn]);
6346 * Take the context lock here so that if find_get_context is
6347 * reading child->perf_event_ctxp, we wait until it has
6348 * incremented the context's refcount before we do put_ctx below.
6350 raw_spin_lock(&child_ctx->lock);
6351 task_ctx_sched_out(child_ctx);
6352 child->perf_event_ctxp[ctxn] = NULL;
6354 * If this context is a clone; unclone it so it can't get
6355 * swapped to another process while we're removing all
6356 * the events from it.
6358 unclone_ctx(child_ctx);
6359 update_context_time(child_ctx);
6360 raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
6363 * Report the task dead after unscheduling the events so that we
6364 * won't get any samples after PERF_RECORD_EXIT. We can however still
6365 * get a few PERF_RECORD_READ events.
6367 perf_event_task(child, child_ctx, 0);
6370 * We can recurse on the same lock type through:
6372 * __perf_event_exit_task()
6373 * sync_child_event()
6374 * fput(parent_event->filp)
6376 * mutex_lock(&ctx->mutex)
6378 * But since its the parent context it won't be the same instance.
6380 mutex_lock(&child_ctx->mutex);
6383 list_for_each_entry_safe(child_event, tmp, &child_ctx->pinned_groups,
6385 __perf_event_exit_task(child_event, child_ctx, child);
6387 list_for_each_entry_safe(child_event, tmp, &child_ctx->flexible_groups,
6389 __perf_event_exit_task(child_event, child_ctx, child);
6392 * If the last event was a group event, it will have appended all
6393 * its siblings to the list, but we obtained 'tmp' before that which
6394 * will still point to the list head terminating the iteration.
6396 if (!list_empty(&child_ctx->pinned_groups) ||
6397 !list_empty(&child_ctx->flexible_groups))
6400 mutex_unlock(&child_ctx->mutex);
6406 * When a child task exits, feed back event values to parent events.
6408 void perf_event_exit_task(struct task_struct *child)
6410 struct perf_event *event, *tmp;
6413 mutex_lock(&child->perf_event_mutex);
6414 list_for_each_entry_safe(event, tmp, &child->perf_event_list,
6416 list_del_init(&event->owner_entry);
6419 * Ensure the list deletion is visible before we clear
6420 * the owner, closes a race against perf_release() where
6421 * we need to serialize on the owner->perf_event_mutex.
6424 event->owner = NULL;
6426 mutex_unlock(&child->perf_event_mutex);
6428 for_each_task_context_nr(ctxn)
6429 perf_event_exit_task_context(child, ctxn);
6432 static void perf_free_event(struct perf_event *event,
6433 struct perf_event_context *ctx)
6435 struct perf_event *parent = event->parent;
6437 if (WARN_ON_ONCE(!parent))
6440 mutex_lock(&parent->child_mutex);
6441 list_del_init(&event->child_list);
6442 mutex_unlock(&parent->child_mutex);
6446 perf_group_detach(event);
6447 list_del_event(event, ctx);
6452 * free an unexposed, unused context as created by inheritance by
6453 * perf_event_init_task below, used by fork() in case of fail.
6455 void perf_event_free_task(struct task_struct *task)
6457 struct perf_event_context *ctx;
6458 struct perf_event *event, *tmp;
6461 for_each_task_context_nr(ctxn) {
6462 ctx = task->perf_event_ctxp[ctxn];
6466 mutex_lock(&ctx->mutex);
6468 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups,
6470 perf_free_event(event, ctx);
6472 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups,
6474 perf_free_event(event, ctx);
6476 if (!list_empty(&ctx->pinned_groups) ||
6477 !list_empty(&ctx->flexible_groups))
6480 mutex_unlock(&ctx->mutex);
6486 void perf_event_delayed_put(struct task_struct *task)
6490 for_each_task_context_nr(ctxn)
6491 WARN_ON_ONCE(task->perf_event_ctxp[ctxn]);
6495 * inherit a event from parent task to child task:
6497 static struct perf_event *
6498 inherit_event(struct perf_event *parent_event,
6499 struct task_struct *parent,
6500 struct perf_event_context *parent_ctx,
6501 struct task_struct *child,
6502 struct perf_event *group_leader,
6503 struct perf_event_context *child_ctx)
6505 struct perf_event *child_event;
6506 unsigned long flags;
6509 * Instead of creating recursive hierarchies of events,
6510 * we link inherited events back to the original parent,
6511 * which has a filp for sure, which we use as the reference
6514 if (parent_event->parent)
6515 parent_event = parent_event->parent;
6517 child_event = perf_event_alloc(&parent_event->attr,
6520 group_leader, parent_event,
6522 if (IS_ERR(child_event))
6527 * Make the child state follow the state of the parent event,
6528 * not its attr.disabled bit. We hold the parent's mutex,
6529 * so we won't race with perf_event_{en, dis}able_family.
6531 if (parent_event->state >= PERF_EVENT_STATE_INACTIVE)
6532 child_event->state = PERF_EVENT_STATE_INACTIVE;
6534 child_event->state = PERF_EVENT_STATE_OFF;
6536 if (parent_event->attr.freq) {
6537 u64 sample_period = parent_event->hw.sample_period;
6538 struct hw_perf_event *hwc = &child_event->hw;
6540 hwc->sample_period = sample_period;
6541 hwc->last_period = sample_period;
6543 local64_set(&hwc->period_left, sample_period);
6546 child_event->ctx = child_ctx;
6547 child_event->overflow_handler = parent_event->overflow_handler;
6548 child_event->overflow_handler_context
6549 = parent_event->overflow_handler_context;
6552 * Precalculate sample_data sizes
6554 perf_event__header_size(child_event);
6555 perf_event__id_header_size(child_event);
6558 * Link it up in the child's context:
6560 raw_spin_lock_irqsave(&child_ctx->lock, flags);
6561 add_event_to_ctx(child_event, child_ctx);
6562 raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
6565 * Get a reference to the parent filp - we will fput it
6566 * when the child event exits. This is safe to do because
6567 * we are in the parent and we know that the filp still
6568 * exists and has a nonzero count:
6570 atomic_long_inc(&parent_event->filp->f_count);
6573 * Link this into the parent event's child list
6575 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
6576 mutex_lock(&parent_event->child_mutex);
6577 list_add_tail(&child_event->child_list, &parent_event->child_list);
6578 mutex_unlock(&parent_event->child_mutex);
6583 static int inherit_group(struct perf_event *parent_event,
6584 struct task_struct *parent,
6585 struct perf_event_context *parent_ctx,
6586 struct task_struct *child,
6587 struct perf_event_context *child_ctx)
6589 struct perf_event *leader;
6590 struct perf_event *sub;
6591 struct perf_event *child_ctr;
6593 leader = inherit_event(parent_event, parent, parent_ctx,
6594 child, NULL, child_ctx);
6596 return PTR_ERR(leader);
6597 list_for_each_entry(sub, &parent_event->sibling_list, group_entry) {
6598 child_ctr = inherit_event(sub, parent, parent_ctx,
6599 child, leader, child_ctx);
6600 if (IS_ERR(child_ctr))
6601 return PTR_ERR(child_ctr);
6607 inherit_task_group(struct perf_event *event, struct task_struct *parent,
6608 struct perf_event_context *parent_ctx,
6609 struct task_struct *child, int ctxn,
6613 struct perf_event_context *child_ctx;
6615 if (!event->attr.inherit) {
6620 child_ctx = child->perf_event_ctxp[ctxn];
6623 * This is executed from the parent task context, so
6624 * inherit events that have been marked for cloning.
6625 * First allocate and initialize a context for the
6629 child_ctx = alloc_perf_context(event->pmu, child);
6633 child->perf_event_ctxp[ctxn] = child_ctx;
6636 ret = inherit_group(event, parent, parent_ctx,
6646 * Initialize the perf_event context in task_struct
6648 int perf_event_init_context(struct task_struct *child, int ctxn)
6650 struct perf_event_context *child_ctx, *parent_ctx;
6651 struct perf_event_context *cloned_ctx;
6652 struct perf_event *event;
6653 struct task_struct *parent = current;
6654 int inherited_all = 1;
6655 unsigned long flags;
6658 if (likely(!parent->perf_event_ctxp[ctxn]))
6662 * If the parent's context is a clone, pin it so it won't get
6665 parent_ctx = perf_pin_task_context(parent, ctxn);
6668 * No need to check if parent_ctx != NULL here; since we saw
6669 * it non-NULL earlier, the only reason for it to become NULL
6670 * is if we exit, and since we're currently in the middle of
6671 * a fork we can't be exiting at the same time.
6675 * Lock the parent list. No need to lock the child - not PID
6676 * hashed yet and not running, so nobody can access it.
6678 mutex_lock(&parent_ctx->mutex);
6681 * We dont have to disable NMIs - we are only looking at
6682 * the list, not manipulating it:
6684 list_for_each_entry(event, &parent_ctx->pinned_groups, group_entry) {
6685 ret = inherit_task_group(event, parent, parent_ctx,
6686 child, ctxn, &inherited_all);
6692 * We can't hold ctx->lock when iterating the ->flexible_group list due
6693 * to allocations, but we need to prevent rotation because
6694 * rotate_ctx() will change the list from interrupt context.
6696 raw_spin_lock_irqsave(&parent_ctx->lock, flags);
6697 parent_ctx->rotate_disable = 1;
6698 raw_spin_unlock_irqrestore(&parent_ctx->lock, flags);
6700 list_for_each_entry(event, &parent_ctx->flexible_groups, group_entry) {
6701 ret = inherit_task_group(event, parent, parent_ctx,
6702 child, ctxn, &inherited_all);
6707 raw_spin_lock_irqsave(&parent_ctx->lock, flags);
6708 parent_ctx->rotate_disable = 0;
6710 child_ctx = child->perf_event_ctxp[ctxn];
6712 if (child_ctx && inherited_all) {
6714 * Mark the child context as a clone of the parent
6715 * context, or of whatever the parent is a clone of.
6717 * Note that if the parent is a clone, the holding of
6718 * parent_ctx->lock avoids it from being uncloned.
6720 cloned_ctx = parent_ctx->parent_ctx;
6722 child_ctx->parent_ctx = cloned_ctx;
6723 child_ctx->parent_gen = parent_ctx->parent_gen;
6725 child_ctx->parent_ctx = parent_ctx;
6726 child_ctx->parent_gen = parent_ctx->generation;
6728 get_ctx(child_ctx->parent_ctx);
6731 raw_spin_unlock_irqrestore(&parent_ctx->lock, flags);
6732 mutex_unlock(&parent_ctx->mutex);
6734 perf_unpin_context(parent_ctx);
6735 put_ctx(parent_ctx);
6741 * Initialize the perf_event context in task_struct
6743 int perf_event_init_task(struct task_struct *child)
6747 memset(child->perf_event_ctxp, 0, sizeof(child->perf_event_ctxp));
6748 mutex_init(&child->perf_event_mutex);
6749 INIT_LIST_HEAD(&child->perf_event_list);
6751 for_each_task_context_nr(ctxn) {
6752 ret = perf_event_init_context(child, ctxn);
6760 static void __init perf_event_init_all_cpus(void)
6762 struct swevent_htable *swhash;
6765 for_each_possible_cpu(cpu) {
6766 swhash = &per_cpu(swevent_htable, cpu);
6767 mutex_init(&swhash->hlist_mutex);
6768 INIT_LIST_HEAD(&per_cpu(rotation_list, cpu));
6772 static void __cpuinit perf_event_init_cpu(int cpu)
6774 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
6776 mutex_lock(&swhash->hlist_mutex);
6777 if (swhash->hlist_refcount > 0) {
6778 struct swevent_hlist *hlist;
6780 hlist = kzalloc_node(sizeof(*hlist), GFP_KERNEL, cpu_to_node(cpu));
6782 rcu_assign_pointer(swhash->swevent_hlist, hlist);
6784 mutex_unlock(&swhash->hlist_mutex);
6787 #if defined CONFIG_HOTPLUG_CPU || defined CONFIG_KEXEC
6788 static void perf_pmu_rotate_stop(struct pmu *pmu)
6790 struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
6792 WARN_ON(!irqs_disabled());
6794 list_del_init(&cpuctx->rotation_list);
6797 static void __perf_event_exit_context(void *__info)
6799 struct perf_event_context *ctx = __info;
6800 struct perf_event *event, *tmp;
6802 perf_pmu_rotate_stop(ctx->pmu);
6804 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry)
6805 __perf_remove_from_context(event);
6806 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups, group_entry)
6807 __perf_remove_from_context(event);
6810 static void perf_event_exit_cpu_context(int cpu)
6812 struct perf_event_context *ctx;
6816 idx = srcu_read_lock(&pmus_srcu);
6817 list_for_each_entry_rcu(pmu, &pmus, entry) {
6818 ctx = &per_cpu_ptr(pmu->pmu_cpu_context, cpu)->ctx;
6820 mutex_lock(&ctx->mutex);
6821 smp_call_function_single(cpu, __perf_event_exit_context, ctx, 1);
6822 mutex_unlock(&ctx->mutex);
6824 srcu_read_unlock(&pmus_srcu, idx);
6827 static void perf_event_exit_cpu(int cpu)
6829 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
6831 mutex_lock(&swhash->hlist_mutex);
6832 swevent_hlist_release(swhash);
6833 mutex_unlock(&swhash->hlist_mutex);
6835 perf_event_exit_cpu_context(cpu);
6838 static inline void perf_event_exit_cpu(int cpu) { }
6842 perf_reboot(struct notifier_block *notifier, unsigned long val, void *v)
6846 for_each_online_cpu(cpu)
6847 perf_event_exit_cpu(cpu);
6853 * Run the perf reboot notifier at the very last possible moment so that
6854 * the generic watchdog code runs as long as possible.
6856 static struct notifier_block perf_reboot_notifier = {
6857 .notifier_call = perf_reboot,
6858 .priority = INT_MIN,
6861 static int __cpuinit
6862 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
6864 unsigned int cpu = (long)hcpu;
6866 switch (action & ~CPU_TASKS_FROZEN) {
6868 case CPU_UP_PREPARE:
6869 case CPU_DOWN_FAILED:
6870 perf_event_init_cpu(cpu);
6873 case CPU_UP_CANCELED:
6874 case CPU_DOWN_PREPARE:
6875 perf_event_exit_cpu(cpu);
6885 void __init perf_event_init(void)
6891 perf_event_init_all_cpus();
6892 init_srcu_struct(&pmus_srcu);
6893 perf_pmu_register(&perf_swevent, "software", PERF_TYPE_SOFTWARE);
6894 perf_pmu_register(&perf_cpu_clock, NULL, -1);
6895 perf_pmu_register(&perf_task_clock, NULL, -1);
6897 perf_cpu_notifier(perf_cpu_notify);
6898 register_reboot_notifier(&perf_reboot_notifier);
6900 ret = init_hw_breakpoint();
6901 WARN(ret, "hw_breakpoint initialization failed with: %d", ret);
6903 /* do not patch jump label more than once per second */
6904 jump_label_rate_limit(&perf_sched_events, HZ);
6907 static int __init perf_event_sysfs_init(void)
6912 mutex_lock(&pmus_lock);
6914 ret = bus_register(&pmu_bus);
6918 list_for_each_entry(pmu, &pmus, entry) {
6919 if (!pmu->name || pmu->type < 0)
6922 ret = pmu_dev_alloc(pmu);
6923 WARN(ret, "Failed to register pmu: %s, reason %d\n", pmu->name, ret);
6925 pmu_bus_running = 1;
6929 mutex_unlock(&pmus_lock);
6933 device_initcall(perf_event_sysfs_init);
6935 #ifdef CONFIG_CGROUP_PERF
6936 static struct cgroup_subsys_state *perf_cgroup_create(
6937 struct cgroup_subsys *ss, struct cgroup *cont)
6939 struct perf_cgroup *jc;
6941 jc = kzalloc(sizeof(*jc), GFP_KERNEL);
6943 return ERR_PTR(-ENOMEM);
6945 jc->info = alloc_percpu(struct perf_cgroup_info);
6948 return ERR_PTR(-ENOMEM);
6954 static void perf_cgroup_destroy(struct cgroup_subsys *ss,
6955 struct cgroup *cont)
6957 struct perf_cgroup *jc;
6958 jc = container_of(cgroup_subsys_state(cont, perf_subsys_id),
6959 struct perf_cgroup, css);
6960 free_percpu(jc->info);
6964 static int __perf_cgroup_move(void *info)
6966 struct task_struct *task = info;
6967 perf_cgroup_switch(task, PERF_CGROUP_SWOUT | PERF_CGROUP_SWIN);
6971 static void perf_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
6972 struct cgroup_taskset *tset)
6974 struct task_struct *task;
6976 cgroup_taskset_for_each(task, cgrp, tset)
6977 task_function_call(task, __perf_cgroup_move, task);
6980 static void perf_cgroup_exit(struct cgroup_subsys *ss, struct cgroup *cgrp,
6981 struct cgroup *old_cgrp, struct task_struct *task)
6984 * cgroup_exit() is called in the copy_process() failure path.
6985 * Ignore this case since the task hasn't ran yet, this avoids
6986 * trying to poke a half freed task state from generic code.
6988 if (!(task->flags & PF_EXITING))
6991 task_function_call(task, __perf_cgroup_move, task);
6994 struct cgroup_subsys perf_subsys = {
6995 .name = "perf_event",
6996 .subsys_id = perf_subsys_id,
6997 .create = perf_cgroup_create,
6998 .destroy = perf_cgroup_destroy,
6999 .exit = perf_cgroup_exit,
7000 .attach = perf_cgroup_attach,
7002 #endif /* CONFIG_CGROUP_PERF */