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/vmalloc.h>
29 #include <linux/hardirq.h>
30 #include <linux/rculist.h>
31 #include <linux/uaccess.h>
32 #include <linux/syscalls.h>
33 #include <linux/anon_inodes.h>
34 #include <linux/kernel_stat.h>
35 #include <linux/perf_event.h>
36 #include <linux/ftrace_event.h>
37 #include <linux/hw_breakpoint.h>
39 #include <asm/irq_regs.h>
41 struct remote_function_call {
42 struct task_struct *p;
43 int (*func)(void *info);
48 static void remote_function(void *data)
50 struct remote_function_call *tfc = data;
51 struct task_struct *p = tfc->p;
55 if (task_cpu(p) != smp_processor_id() || !task_curr(p))
59 tfc->ret = tfc->func(tfc->info);
63 * task_function_call - call a function on the cpu on which a task runs
64 * @p: the task to evaluate
65 * @func: the function to be called
66 * @info: the function call argument
68 * Calls the function @func when the task is currently running. This might
69 * be on the current CPU, which just calls the function directly
71 * returns: @func return value, or
72 * -ESRCH - when the process isn't running
73 * -EAGAIN - when the process moved away
76 task_function_call(struct task_struct *p, int (*func) (void *info), void *info)
78 struct remote_function_call data = {
82 .ret = -ESRCH, /* No such (running) process */
86 smp_call_function_single(task_cpu(p), remote_function, &data, 1);
92 * cpu_function_call - call a function on the cpu
93 * @func: the function to be called
94 * @info: the function call argument
96 * Calls the function @func on the remote cpu.
98 * returns: @func return value or -ENXIO when the cpu is offline
100 static int cpu_function_call(int cpu, int (*func) (void *info), void *info)
102 struct remote_function_call data = {
106 .ret = -ENXIO, /* No such CPU */
109 smp_call_function_single(cpu, remote_function, &data, 1);
114 #define PERF_FLAG_ALL (PERF_FLAG_FD_NO_GROUP |\
115 PERF_FLAG_FD_OUTPUT |\
116 PERF_FLAG_PID_CGROUP)
119 EVENT_FLEXIBLE = 0x1,
121 EVENT_ALL = EVENT_FLEXIBLE | EVENT_PINNED,
125 * perf_sched_events : >0 events exist
126 * perf_cgroup_events: >0 per-cpu cgroup events exist on this cpu
128 struct jump_label_key perf_sched_events __read_mostly;
129 static DEFINE_PER_CPU(atomic_t, perf_cgroup_events);
131 static atomic_t nr_mmap_events __read_mostly;
132 static atomic_t nr_comm_events __read_mostly;
133 static atomic_t nr_task_events __read_mostly;
135 static LIST_HEAD(pmus);
136 static DEFINE_MUTEX(pmus_lock);
137 static struct srcu_struct pmus_srcu;
140 * perf event paranoia level:
141 * -1 - not paranoid at all
142 * 0 - disallow raw tracepoint access for unpriv
143 * 1 - disallow cpu events for unpriv
144 * 2 - disallow kernel profiling for unpriv
146 int sysctl_perf_event_paranoid __read_mostly = 1;
148 /* Minimum for 512 kiB + 1 user control page */
149 int sysctl_perf_event_mlock __read_mostly = 512 + (PAGE_SIZE / 1024); /* 'free' kiB per user */
152 * max perf event sample rate
154 #define DEFAULT_MAX_SAMPLE_RATE 100000
155 int sysctl_perf_event_sample_rate __read_mostly = DEFAULT_MAX_SAMPLE_RATE;
156 static int max_samples_per_tick __read_mostly =
157 DIV_ROUND_UP(DEFAULT_MAX_SAMPLE_RATE, HZ);
159 int perf_proc_update_handler(struct ctl_table *table, int write,
160 void __user *buffer, size_t *lenp,
163 int ret = proc_dointvec(table, write, buffer, lenp, ppos);
168 max_samples_per_tick = DIV_ROUND_UP(sysctl_perf_event_sample_rate, HZ);
173 static atomic64_t perf_event_id;
175 static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
176 enum event_type_t event_type);
178 static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
179 enum event_type_t event_type,
180 struct task_struct *task);
182 static void update_context_time(struct perf_event_context *ctx);
183 static u64 perf_event_time(struct perf_event *event);
185 void __weak perf_event_print_debug(void) { }
187 extern __weak const char *perf_pmu_name(void)
192 static inline u64 perf_clock(void)
194 return local_clock();
197 static inline struct perf_cpu_context *
198 __get_cpu_context(struct perf_event_context *ctx)
200 return this_cpu_ptr(ctx->pmu->pmu_cpu_context);
203 static void perf_ctx_lock(struct perf_cpu_context *cpuctx,
204 struct perf_event_context *ctx)
206 raw_spin_lock(&cpuctx->ctx.lock);
208 raw_spin_lock(&ctx->lock);
211 static void perf_ctx_unlock(struct perf_cpu_context *cpuctx,
212 struct perf_event_context *ctx)
215 raw_spin_unlock(&ctx->lock);
216 raw_spin_unlock(&cpuctx->ctx.lock);
219 #ifdef CONFIG_CGROUP_PERF
222 * Must ensure cgroup is pinned (css_get) before calling
223 * this function. In other words, we cannot call this function
224 * if there is no cgroup event for the current CPU context.
226 static inline struct perf_cgroup *
227 perf_cgroup_from_task(struct task_struct *task)
229 return container_of(task_subsys_state(task, perf_subsys_id),
230 struct perf_cgroup, css);
234 perf_cgroup_match(struct perf_event *event)
236 struct perf_event_context *ctx = event->ctx;
237 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
239 return !event->cgrp || event->cgrp == cpuctx->cgrp;
242 static inline void perf_get_cgroup(struct perf_event *event)
244 css_get(&event->cgrp->css);
247 static inline void perf_put_cgroup(struct perf_event *event)
249 css_put(&event->cgrp->css);
252 static inline void perf_detach_cgroup(struct perf_event *event)
254 perf_put_cgroup(event);
258 static inline int is_cgroup_event(struct perf_event *event)
260 return event->cgrp != NULL;
263 static inline u64 perf_cgroup_event_time(struct perf_event *event)
265 struct perf_cgroup_info *t;
267 t = per_cpu_ptr(event->cgrp->info, event->cpu);
271 static inline void __update_cgrp_time(struct perf_cgroup *cgrp)
273 struct perf_cgroup_info *info;
278 info = this_cpu_ptr(cgrp->info);
280 info->time += now - info->timestamp;
281 info->timestamp = now;
284 static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx)
286 struct perf_cgroup *cgrp_out = cpuctx->cgrp;
288 __update_cgrp_time(cgrp_out);
291 static inline void update_cgrp_time_from_event(struct perf_event *event)
293 struct perf_cgroup *cgrp;
296 * ensure we access cgroup data only when needed and
297 * when we know the cgroup is pinned (css_get)
299 if (!is_cgroup_event(event))
302 cgrp = perf_cgroup_from_task(current);
304 * Do not update time when cgroup is not active
306 if (cgrp == event->cgrp)
307 __update_cgrp_time(event->cgrp);
311 perf_cgroup_set_timestamp(struct task_struct *task,
312 struct perf_event_context *ctx)
314 struct perf_cgroup *cgrp;
315 struct perf_cgroup_info *info;
318 * ctx->lock held by caller
319 * ensure we do not access cgroup data
320 * unless we have the cgroup pinned (css_get)
322 if (!task || !ctx->nr_cgroups)
325 cgrp = perf_cgroup_from_task(task);
326 info = this_cpu_ptr(cgrp->info);
327 info->timestamp = ctx->timestamp;
330 #define PERF_CGROUP_SWOUT 0x1 /* cgroup switch out every event */
331 #define PERF_CGROUP_SWIN 0x2 /* cgroup switch in events based on task */
334 * reschedule events based on the cgroup constraint of task.
336 * mode SWOUT : schedule out everything
337 * mode SWIN : schedule in based on cgroup for next
339 void perf_cgroup_switch(struct task_struct *task, int mode)
341 struct perf_cpu_context *cpuctx;
346 * disable interrupts to avoid geting nr_cgroup
347 * changes via __perf_event_disable(). Also
350 local_irq_save(flags);
353 * we reschedule only in the presence of cgroup
354 * constrained events.
358 list_for_each_entry_rcu(pmu, &pmus, entry) {
359 cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
362 * perf_cgroup_events says at least one
363 * context on this CPU has cgroup events.
365 * ctx->nr_cgroups reports the number of cgroup
366 * events for a context.
368 if (cpuctx->ctx.nr_cgroups > 0) {
369 perf_ctx_lock(cpuctx, cpuctx->task_ctx);
370 perf_pmu_disable(cpuctx->ctx.pmu);
372 if (mode & PERF_CGROUP_SWOUT) {
373 cpu_ctx_sched_out(cpuctx, EVENT_ALL);
375 * must not be done before ctxswout due
376 * to event_filter_match() in event_sched_out()
381 if (mode & PERF_CGROUP_SWIN) {
382 WARN_ON_ONCE(cpuctx->cgrp);
383 /* set cgrp before ctxsw in to
384 * allow event_filter_match() to not
385 * have to pass task around
387 cpuctx->cgrp = perf_cgroup_from_task(task);
388 cpu_ctx_sched_in(cpuctx, EVENT_ALL, task);
390 perf_pmu_enable(cpuctx->ctx.pmu);
391 perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
397 local_irq_restore(flags);
400 static inline void perf_cgroup_sched_out(struct task_struct *task)
402 perf_cgroup_switch(task, PERF_CGROUP_SWOUT);
405 static inline void perf_cgroup_sched_in(struct task_struct *task)
407 perf_cgroup_switch(task, PERF_CGROUP_SWIN);
410 static inline int perf_cgroup_connect(int fd, struct perf_event *event,
411 struct perf_event_attr *attr,
412 struct perf_event *group_leader)
414 struct perf_cgroup *cgrp;
415 struct cgroup_subsys_state *css;
417 int ret = 0, fput_needed;
419 file = fget_light(fd, &fput_needed);
423 css = cgroup_css_from_dir(file, perf_subsys_id);
429 cgrp = container_of(css, struct perf_cgroup, css);
432 /* must be done before we fput() the file */
433 perf_get_cgroup(event);
436 * all events in a group must monitor
437 * the same cgroup because a task belongs
438 * to only one perf cgroup at a time
440 if (group_leader && group_leader->cgrp != cgrp) {
441 perf_detach_cgroup(event);
445 fput_light(file, fput_needed);
450 perf_cgroup_set_shadow_time(struct perf_event *event, u64 now)
452 struct perf_cgroup_info *t;
453 t = per_cpu_ptr(event->cgrp->info, event->cpu);
454 event->shadow_ctx_time = now - t->timestamp;
458 perf_cgroup_defer_enabled(struct perf_event *event)
461 * when the current task's perf cgroup does not match
462 * the event's, we need to remember to call the
463 * perf_mark_enable() function the first time a task with
464 * a matching perf cgroup is scheduled in.
466 if (is_cgroup_event(event) && !perf_cgroup_match(event))
467 event->cgrp_defer_enabled = 1;
471 perf_cgroup_mark_enabled(struct perf_event *event,
472 struct perf_event_context *ctx)
474 struct perf_event *sub;
475 u64 tstamp = perf_event_time(event);
477 if (!event->cgrp_defer_enabled)
480 event->cgrp_defer_enabled = 0;
482 event->tstamp_enabled = tstamp - event->total_time_enabled;
483 list_for_each_entry(sub, &event->sibling_list, group_entry) {
484 if (sub->state >= PERF_EVENT_STATE_INACTIVE) {
485 sub->tstamp_enabled = tstamp - sub->total_time_enabled;
486 sub->cgrp_defer_enabled = 0;
490 #else /* !CONFIG_CGROUP_PERF */
493 perf_cgroup_match(struct perf_event *event)
498 static inline void perf_detach_cgroup(struct perf_event *event)
501 static inline int is_cgroup_event(struct perf_event *event)
506 static inline u64 perf_cgroup_event_cgrp_time(struct perf_event *event)
511 static inline void update_cgrp_time_from_event(struct perf_event *event)
515 static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx)
519 static inline void perf_cgroup_sched_out(struct task_struct *task)
523 static inline void perf_cgroup_sched_in(struct task_struct *task)
527 static inline int perf_cgroup_connect(pid_t pid, struct perf_event *event,
528 struct perf_event_attr *attr,
529 struct perf_event *group_leader)
535 perf_cgroup_set_timestamp(struct task_struct *task,
536 struct perf_event_context *ctx)
541 perf_cgroup_switch(struct task_struct *task, struct task_struct *next)
546 perf_cgroup_set_shadow_time(struct perf_event *event, u64 now)
550 static inline u64 perf_cgroup_event_time(struct perf_event *event)
556 perf_cgroup_defer_enabled(struct perf_event *event)
561 perf_cgroup_mark_enabled(struct perf_event *event,
562 struct perf_event_context *ctx)
567 void perf_pmu_disable(struct pmu *pmu)
569 int *count = this_cpu_ptr(pmu->pmu_disable_count);
571 pmu->pmu_disable(pmu);
574 void perf_pmu_enable(struct pmu *pmu)
576 int *count = this_cpu_ptr(pmu->pmu_disable_count);
578 pmu->pmu_enable(pmu);
581 static DEFINE_PER_CPU(struct list_head, rotation_list);
584 * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
585 * because they're strictly cpu affine and rotate_start is called with IRQs
586 * disabled, while rotate_context is called from IRQ context.
588 static void perf_pmu_rotate_start(struct pmu *pmu)
590 struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
591 struct list_head *head = &__get_cpu_var(rotation_list);
593 WARN_ON(!irqs_disabled());
595 if (list_empty(&cpuctx->rotation_list))
596 list_add(&cpuctx->rotation_list, head);
599 static void get_ctx(struct perf_event_context *ctx)
601 WARN_ON(!atomic_inc_not_zero(&ctx->refcount));
604 static void put_ctx(struct perf_event_context *ctx)
606 if (atomic_dec_and_test(&ctx->refcount)) {
608 put_ctx(ctx->parent_ctx);
610 put_task_struct(ctx->task);
611 kfree_rcu(ctx, rcu_head);
615 static void unclone_ctx(struct perf_event_context *ctx)
617 if (ctx->parent_ctx) {
618 put_ctx(ctx->parent_ctx);
619 ctx->parent_ctx = NULL;
623 static u32 perf_event_pid(struct perf_event *event, struct task_struct *p)
626 * only top level events have the pid namespace they were created in
629 event = event->parent;
631 return task_tgid_nr_ns(p, event->ns);
634 static u32 perf_event_tid(struct perf_event *event, struct task_struct *p)
637 * only top level events have the pid namespace they were created in
640 event = event->parent;
642 return task_pid_nr_ns(p, event->ns);
646 * If we inherit events we want to return the parent event id
649 static u64 primary_event_id(struct perf_event *event)
654 id = event->parent->id;
660 * Get the perf_event_context for a task and lock it.
661 * This has to cope with with the fact that until it is locked,
662 * the context could get moved to another task.
664 static struct perf_event_context *
665 perf_lock_task_context(struct task_struct *task, int ctxn, unsigned long *flags)
667 struct perf_event_context *ctx;
671 ctx = rcu_dereference(task->perf_event_ctxp[ctxn]);
674 * If this context is a clone of another, it might
675 * get swapped for another underneath us by
676 * perf_event_task_sched_out, though the
677 * rcu_read_lock() protects us from any context
678 * getting freed. Lock the context and check if it
679 * got swapped before we could get the lock, and retry
680 * if so. If we locked the right context, then it
681 * can't get swapped on us any more.
683 raw_spin_lock_irqsave(&ctx->lock, *flags);
684 if (ctx != rcu_dereference(task->perf_event_ctxp[ctxn])) {
685 raw_spin_unlock_irqrestore(&ctx->lock, *flags);
689 if (!atomic_inc_not_zero(&ctx->refcount)) {
690 raw_spin_unlock_irqrestore(&ctx->lock, *flags);
699 * Get the context for a task and increment its pin_count so it
700 * can't get swapped to another task. This also increments its
701 * reference count so that the context can't get freed.
703 static struct perf_event_context *
704 perf_pin_task_context(struct task_struct *task, int ctxn)
706 struct perf_event_context *ctx;
709 ctx = perf_lock_task_context(task, ctxn, &flags);
712 raw_spin_unlock_irqrestore(&ctx->lock, flags);
717 static void perf_unpin_context(struct perf_event_context *ctx)
721 raw_spin_lock_irqsave(&ctx->lock, flags);
723 raw_spin_unlock_irqrestore(&ctx->lock, flags);
727 * Update the record of the current time in a context.
729 static void update_context_time(struct perf_event_context *ctx)
731 u64 now = perf_clock();
733 ctx->time += now - ctx->timestamp;
734 ctx->timestamp = now;
737 static u64 perf_event_time(struct perf_event *event)
739 struct perf_event_context *ctx = event->ctx;
741 if (is_cgroup_event(event))
742 return perf_cgroup_event_time(event);
744 return ctx ? ctx->time : 0;
748 * Update the total_time_enabled and total_time_running fields for a event.
750 static void update_event_times(struct perf_event *event)
752 struct perf_event_context *ctx = event->ctx;
755 if (event->state < PERF_EVENT_STATE_INACTIVE ||
756 event->group_leader->state < PERF_EVENT_STATE_INACTIVE)
759 * in cgroup mode, time_enabled represents
760 * the time the event was enabled AND active
761 * tasks were in the monitored cgroup. This is
762 * independent of the activity of the context as
763 * there may be a mix of cgroup and non-cgroup events.
765 * That is why we treat cgroup events differently
768 if (is_cgroup_event(event))
769 run_end = perf_event_time(event);
770 else if (ctx->is_active)
773 run_end = event->tstamp_stopped;
775 event->total_time_enabled = run_end - event->tstamp_enabled;
777 if (event->state == PERF_EVENT_STATE_INACTIVE)
778 run_end = event->tstamp_stopped;
780 run_end = perf_event_time(event);
782 event->total_time_running = run_end - event->tstamp_running;
787 * Update total_time_enabled and total_time_running for all events in a group.
789 static void update_group_times(struct perf_event *leader)
791 struct perf_event *event;
793 update_event_times(leader);
794 list_for_each_entry(event, &leader->sibling_list, group_entry)
795 update_event_times(event);
798 static struct list_head *
799 ctx_group_list(struct perf_event *event, struct perf_event_context *ctx)
801 if (event->attr.pinned)
802 return &ctx->pinned_groups;
804 return &ctx->flexible_groups;
808 * Add a event from the lists for its context.
809 * Must be called with ctx->mutex and ctx->lock held.
812 list_add_event(struct perf_event *event, struct perf_event_context *ctx)
814 WARN_ON_ONCE(event->attach_state & PERF_ATTACH_CONTEXT);
815 event->attach_state |= PERF_ATTACH_CONTEXT;
818 * If we're a stand alone event or group leader, we go to the context
819 * list, group events are kept attached to the group so that
820 * perf_group_detach can, at all times, locate all siblings.
822 if (event->group_leader == event) {
823 struct list_head *list;
825 if (is_software_event(event))
826 event->group_flags |= PERF_GROUP_SOFTWARE;
828 list = ctx_group_list(event, ctx);
829 list_add_tail(&event->group_entry, list);
832 if (is_cgroup_event(event))
835 list_add_rcu(&event->event_entry, &ctx->event_list);
837 perf_pmu_rotate_start(ctx->pmu);
839 if (event->attr.inherit_stat)
844 * Called at perf_event creation and when events are attached/detached from a
847 static void perf_event__read_size(struct perf_event *event)
849 int entry = sizeof(u64); /* value */
853 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
856 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
859 if (event->attr.read_format & PERF_FORMAT_ID)
860 entry += sizeof(u64);
862 if (event->attr.read_format & PERF_FORMAT_GROUP) {
863 nr += event->group_leader->nr_siblings;
868 event->read_size = size;
871 static void perf_event__header_size(struct perf_event *event)
873 struct perf_sample_data *data;
874 u64 sample_type = event->attr.sample_type;
877 perf_event__read_size(event);
879 if (sample_type & PERF_SAMPLE_IP)
880 size += sizeof(data->ip);
882 if (sample_type & PERF_SAMPLE_ADDR)
883 size += sizeof(data->addr);
885 if (sample_type & PERF_SAMPLE_PERIOD)
886 size += sizeof(data->period);
888 if (sample_type & PERF_SAMPLE_READ)
889 size += event->read_size;
891 event->header_size = size;
894 static void perf_event__id_header_size(struct perf_event *event)
896 struct perf_sample_data *data;
897 u64 sample_type = event->attr.sample_type;
900 if (sample_type & PERF_SAMPLE_TID)
901 size += sizeof(data->tid_entry);
903 if (sample_type & PERF_SAMPLE_TIME)
904 size += sizeof(data->time);
906 if (sample_type & PERF_SAMPLE_ID)
907 size += sizeof(data->id);
909 if (sample_type & PERF_SAMPLE_STREAM_ID)
910 size += sizeof(data->stream_id);
912 if (sample_type & PERF_SAMPLE_CPU)
913 size += sizeof(data->cpu_entry);
915 event->id_header_size = size;
918 static void perf_group_attach(struct perf_event *event)
920 struct perf_event *group_leader = event->group_leader, *pos;
923 * We can have double attach due to group movement in perf_event_open.
925 if (event->attach_state & PERF_ATTACH_GROUP)
928 event->attach_state |= PERF_ATTACH_GROUP;
930 if (group_leader == event)
933 if (group_leader->group_flags & PERF_GROUP_SOFTWARE &&
934 !is_software_event(event))
935 group_leader->group_flags &= ~PERF_GROUP_SOFTWARE;
937 list_add_tail(&event->group_entry, &group_leader->sibling_list);
938 group_leader->nr_siblings++;
940 perf_event__header_size(group_leader);
942 list_for_each_entry(pos, &group_leader->sibling_list, group_entry)
943 perf_event__header_size(pos);
947 * Remove a event from the lists for its context.
948 * Must be called with ctx->mutex and ctx->lock held.
951 list_del_event(struct perf_event *event, struct perf_event_context *ctx)
953 struct perf_cpu_context *cpuctx;
955 * We can have double detach due to exit/hot-unplug + close.
957 if (!(event->attach_state & PERF_ATTACH_CONTEXT))
960 event->attach_state &= ~PERF_ATTACH_CONTEXT;
962 if (is_cgroup_event(event)) {
964 cpuctx = __get_cpu_context(ctx);
966 * if there are no more cgroup events
967 * then cler cgrp to avoid stale pointer
968 * in update_cgrp_time_from_cpuctx()
970 if (!ctx->nr_cgroups)
975 if (event->attr.inherit_stat)
978 list_del_rcu(&event->event_entry);
980 if (event->group_leader == event)
981 list_del_init(&event->group_entry);
983 update_group_times(event);
986 * If event was in error state, then keep it
987 * that way, otherwise bogus counts will be
988 * returned on read(). The only way to get out
989 * of error state is by explicit re-enabling
992 if (event->state > PERF_EVENT_STATE_OFF)
993 event->state = PERF_EVENT_STATE_OFF;
996 static void perf_group_detach(struct perf_event *event)
998 struct perf_event *sibling, *tmp;
999 struct list_head *list = NULL;
1002 * We can have double detach due to exit/hot-unplug + close.
1004 if (!(event->attach_state & PERF_ATTACH_GROUP))
1007 event->attach_state &= ~PERF_ATTACH_GROUP;
1010 * If this is a sibling, remove it from its group.
1012 if (event->group_leader != event) {
1013 list_del_init(&event->group_entry);
1014 event->group_leader->nr_siblings--;
1018 if (!list_empty(&event->group_entry))
1019 list = &event->group_entry;
1022 * If this was a group event with sibling events then
1023 * upgrade the siblings to singleton events by adding them
1024 * to whatever list we are on.
1026 list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) {
1028 list_move_tail(&sibling->group_entry, list);
1029 sibling->group_leader = sibling;
1031 /* Inherit group flags from the previous leader */
1032 sibling->group_flags = event->group_flags;
1036 perf_event__header_size(event->group_leader);
1038 list_for_each_entry(tmp, &event->group_leader->sibling_list, group_entry)
1039 perf_event__header_size(tmp);
1043 event_filter_match(struct perf_event *event)
1045 return (event->cpu == -1 || event->cpu == smp_processor_id())
1046 && perf_cgroup_match(event);
1050 event_sched_out(struct perf_event *event,
1051 struct perf_cpu_context *cpuctx,
1052 struct perf_event_context *ctx)
1054 u64 tstamp = perf_event_time(event);
1057 * An event which could not be activated because of
1058 * filter mismatch still needs to have its timings
1059 * maintained, otherwise bogus information is return
1060 * via read() for time_enabled, time_running:
1062 if (event->state == PERF_EVENT_STATE_INACTIVE
1063 && !event_filter_match(event)) {
1064 delta = tstamp - event->tstamp_stopped;
1065 event->tstamp_running += delta;
1066 event->tstamp_stopped = tstamp;
1069 if (event->state != PERF_EVENT_STATE_ACTIVE)
1072 event->state = PERF_EVENT_STATE_INACTIVE;
1073 if (event->pending_disable) {
1074 event->pending_disable = 0;
1075 event->state = PERF_EVENT_STATE_OFF;
1077 event->tstamp_stopped = tstamp;
1078 event->pmu->del(event, 0);
1081 if (!is_software_event(event))
1082 cpuctx->active_oncpu--;
1084 if (event->attr.exclusive || !cpuctx->active_oncpu)
1085 cpuctx->exclusive = 0;
1089 group_sched_out(struct perf_event *group_event,
1090 struct perf_cpu_context *cpuctx,
1091 struct perf_event_context *ctx)
1093 struct perf_event *event;
1094 int state = group_event->state;
1096 event_sched_out(group_event, cpuctx, ctx);
1099 * Schedule out siblings (if any):
1101 list_for_each_entry(event, &group_event->sibling_list, group_entry)
1102 event_sched_out(event, cpuctx, ctx);
1104 if (state == PERF_EVENT_STATE_ACTIVE && group_event->attr.exclusive)
1105 cpuctx->exclusive = 0;
1109 * Cross CPU call to remove a performance event
1111 * We disable the event on the hardware level first. After that we
1112 * remove it from the context list.
1114 static int __perf_remove_from_context(void *info)
1116 struct perf_event *event = info;
1117 struct perf_event_context *ctx = event->ctx;
1118 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1120 raw_spin_lock(&ctx->lock);
1121 event_sched_out(event, cpuctx, ctx);
1122 list_del_event(event, ctx);
1123 raw_spin_unlock(&ctx->lock);
1130 * Remove the event from a task's (or a CPU's) list of events.
1132 * CPU events are removed with a smp call. For task events we only
1133 * call when the task is on a CPU.
1135 * If event->ctx is a cloned context, callers must make sure that
1136 * every task struct that event->ctx->task could possibly point to
1137 * remains valid. This is OK when called from perf_release since
1138 * that only calls us on the top-level context, which can't be a clone.
1139 * When called from perf_event_exit_task, it's OK because the
1140 * context has been detached from its task.
1142 static void perf_remove_from_context(struct perf_event *event)
1144 struct perf_event_context *ctx = event->ctx;
1145 struct task_struct *task = ctx->task;
1147 lockdep_assert_held(&ctx->mutex);
1151 * Per cpu events are removed via an smp call and
1152 * the removal is always successful.
1154 cpu_function_call(event->cpu, __perf_remove_from_context, event);
1159 if (!task_function_call(task, __perf_remove_from_context, event))
1162 raw_spin_lock_irq(&ctx->lock);
1164 * If we failed to find a running task, but find the context active now
1165 * that we've acquired the ctx->lock, retry.
1167 if (ctx->is_active) {
1168 raw_spin_unlock_irq(&ctx->lock);
1173 * Since the task isn't running, its safe to remove the event, us
1174 * holding the ctx->lock ensures the task won't get scheduled in.
1176 list_del_event(event, ctx);
1177 raw_spin_unlock_irq(&ctx->lock);
1181 * Cross CPU call to disable a performance event
1183 static int __perf_event_disable(void *info)
1185 struct perf_event *event = info;
1186 struct perf_event_context *ctx = event->ctx;
1187 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1190 * If this is a per-task event, need to check whether this
1191 * event's task is the current task on this cpu.
1193 * Can trigger due to concurrent perf_event_context_sched_out()
1194 * flipping contexts around.
1196 if (ctx->task && cpuctx->task_ctx != ctx)
1199 raw_spin_lock(&ctx->lock);
1202 * If the event is on, turn it off.
1203 * If it is in error state, leave it in error state.
1205 if (event->state >= PERF_EVENT_STATE_INACTIVE) {
1206 update_context_time(ctx);
1207 update_cgrp_time_from_event(event);
1208 update_group_times(event);
1209 if (event == event->group_leader)
1210 group_sched_out(event, cpuctx, ctx);
1212 event_sched_out(event, cpuctx, ctx);
1213 event->state = PERF_EVENT_STATE_OFF;
1216 raw_spin_unlock(&ctx->lock);
1224 * If event->ctx is a cloned context, callers must make sure that
1225 * every task struct that event->ctx->task could possibly point to
1226 * remains valid. This condition is satisifed when called through
1227 * perf_event_for_each_child or perf_event_for_each because they
1228 * hold the top-level event's child_mutex, so any descendant that
1229 * goes to exit will block in sync_child_event.
1230 * When called from perf_pending_event it's OK because event->ctx
1231 * is the current context on this CPU and preemption is disabled,
1232 * hence we can't get into perf_event_task_sched_out for this context.
1234 void perf_event_disable(struct perf_event *event)
1236 struct perf_event_context *ctx = event->ctx;
1237 struct task_struct *task = ctx->task;
1241 * Disable the event on the cpu that it's on
1243 cpu_function_call(event->cpu, __perf_event_disable, event);
1248 if (!task_function_call(task, __perf_event_disable, event))
1251 raw_spin_lock_irq(&ctx->lock);
1253 * If the event is still active, we need to retry the cross-call.
1255 if (event->state == PERF_EVENT_STATE_ACTIVE) {
1256 raw_spin_unlock_irq(&ctx->lock);
1258 * Reload the task pointer, it might have been changed by
1259 * a concurrent perf_event_context_sched_out().
1266 * Since we have the lock this context can't be scheduled
1267 * in, so we can change the state safely.
1269 if (event->state == PERF_EVENT_STATE_INACTIVE) {
1270 update_group_times(event);
1271 event->state = PERF_EVENT_STATE_OFF;
1273 raw_spin_unlock_irq(&ctx->lock);
1276 static void perf_set_shadow_time(struct perf_event *event,
1277 struct perf_event_context *ctx,
1281 * use the correct time source for the time snapshot
1283 * We could get by without this by leveraging the
1284 * fact that to get to this function, the caller
1285 * has most likely already called update_context_time()
1286 * and update_cgrp_time_xx() and thus both timestamp
1287 * are identical (or very close). Given that tstamp is,
1288 * already adjusted for cgroup, we could say that:
1289 * tstamp - ctx->timestamp
1291 * tstamp - cgrp->timestamp.
1293 * Then, in perf_output_read(), the calculation would
1294 * work with no changes because:
1295 * - event is guaranteed scheduled in
1296 * - no scheduled out in between
1297 * - thus the timestamp would be the same
1299 * But this is a bit hairy.
1301 * So instead, we have an explicit cgroup call to remain
1302 * within the time time source all along. We believe it
1303 * is cleaner and simpler to understand.
1305 if (is_cgroup_event(event))
1306 perf_cgroup_set_shadow_time(event, tstamp);
1308 event->shadow_ctx_time = tstamp - ctx->timestamp;
1311 #define MAX_INTERRUPTS (~0ULL)
1313 static void perf_log_throttle(struct perf_event *event, int enable);
1316 event_sched_in(struct perf_event *event,
1317 struct perf_cpu_context *cpuctx,
1318 struct perf_event_context *ctx)
1320 u64 tstamp = perf_event_time(event);
1322 if (event->state <= PERF_EVENT_STATE_OFF)
1325 event->state = PERF_EVENT_STATE_ACTIVE;
1326 event->oncpu = smp_processor_id();
1329 * Unthrottle events, since we scheduled we might have missed several
1330 * ticks already, also for a heavily scheduling task there is little
1331 * guarantee it'll get a tick in a timely manner.
1333 if (unlikely(event->hw.interrupts == MAX_INTERRUPTS)) {
1334 perf_log_throttle(event, 1);
1335 event->hw.interrupts = 0;
1339 * The new state must be visible before we turn it on in the hardware:
1343 if (event->pmu->add(event, PERF_EF_START)) {
1344 event->state = PERF_EVENT_STATE_INACTIVE;
1349 event->tstamp_running += tstamp - event->tstamp_stopped;
1351 perf_set_shadow_time(event, ctx, tstamp);
1353 if (!is_software_event(event))
1354 cpuctx->active_oncpu++;
1357 if (event->attr.exclusive)
1358 cpuctx->exclusive = 1;
1364 group_sched_in(struct perf_event *group_event,
1365 struct perf_cpu_context *cpuctx,
1366 struct perf_event_context *ctx)
1368 struct perf_event *event, *partial_group = NULL;
1369 struct pmu *pmu = group_event->pmu;
1370 u64 now = ctx->time;
1371 bool simulate = false;
1373 if (group_event->state == PERF_EVENT_STATE_OFF)
1376 pmu->start_txn(pmu);
1378 if (event_sched_in(group_event, cpuctx, ctx)) {
1379 pmu->cancel_txn(pmu);
1384 * Schedule in siblings as one group (if any):
1386 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
1387 if (event_sched_in(event, cpuctx, ctx)) {
1388 partial_group = event;
1393 if (!pmu->commit_txn(pmu))
1398 * Groups can be scheduled in as one unit only, so undo any
1399 * partial group before returning:
1400 * The events up to the failed event are scheduled out normally,
1401 * tstamp_stopped will be updated.
1403 * The failed events and the remaining siblings need to have
1404 * their timings updated as if they had gone thru event_sched_in()
1405 * and event_sched_out(). This is required to get consistent timings
1406 * across the group. This also takes care of the case where the group
1407 * could never be scheduled by ensuring tstamp_stopped is set to mark
1408 * the time the event was actually stopped, such that time delta
1409 * calculation in update_event_times() is correct.
1411 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
1412 if (event == partial_group)
1416 event->tstamp_running += now - event->tstamp_stopped;
1417 event->tstamp_stopped = now;
1419 event_sched_out(event, cpuctx, ctx);
1422 event_sched_out(group_event, cpuctx, ctx);
1424 pmu->cancel_txn(pmu);
1430 * Work out whether we can put this event group on the CPU now.
1432 static int group_can_go_on(struct perf_event *event,
1433 struct perf_cpu_context *cpuctx,
1437 * Groups consisting entirely of software events can always go on.
1439 if (event->group_flags & PERF_GROUP_SOFTWARE)
1442 * If an exclusive group is already on, no other hardware
1445 if (cpuctx->exclusive)
1448 * If this group is exclusive and there are already
1449 * events on the CPU, it can't go on.
1451 if (event->attr.exclusive && cpuctx->active_oncpu)
1454 * Otherwise, try to add it if all previous groups were able
1460 static void add_event_to_ctx(struct perf_event *event,
1461 struct perf_event_context *ctx)
1463 u64 tstamp = perf_event_time(event);
1465 list_add_event(event, ctx);
1466 perf_group_attach(event);
1467 event->tstamp_enabled = tstamp;
1468 event->tstamp_running = tstamp;
1469 event->tstamp_stopped = tstamp;
1472 static void task_ctx_sched_out(struct perf_event_context *ctx);
1474 ctx_sched_in(struct perf_event_context *ctx,
1475 struct perf_cpu_context *cpuctx,
1476 enum event_type_t event_type,
1477 struct task_struct *task);
1479 static void perf_event_sched_in(struct perf_cpu_context *cpuctx,
1480 struct perf_event_context *ctx,
1481 struct task_struct *task)
1483 cpu_ctx_sched_in(cpuctx, EVENT_PINNED, task);
1485 ctx_sched_in(ctx, cpuctx, EVENT_PINNED, task);
1486 cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE, task);
1488 ctx_sched_in(ctx, cpuctx, EVENT_FLEXIBLE, task);
1492 * Cross CPU call to install and enable a performance event
1494 * Must be called with ctx->mutex held
1496 static int __perf_install_in_context(void *info)
1498 struct perf_event *event = info;
1499 struct perf_event_context *ctx = event->ctx;
1500 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1501 struct perf_event_context *task_ctx = cpuctx->task_ctx;
1502 struct task_struct *task = current;
1504 perf_ctx_lock(cpuctx, cpuctx->task_ctx);
1505 perf_pmu_disable(cpuctx->ctx.pmu);
1508 * If there was an active task_ctx schedule it out.
1511 task_ctx_sched_out(task_ctx);
1513 * If the context we're installing events in is not the
1514 * active task_ctx, flip them.
1516 if (ctx->task && task_ctx != ctx) {
1517 raw_spin_unlock(&cpuctx->ctx.lock);
1518 raw_spin_lock(&ctx->lock);
1519 cpuctx->task_ctx = task_ctx = ctx;
1521 task = task_ctx->task;
1523 cpu_ctx_sched_out(cpuctx, EVENT_ALL);
1525 update_context_time(ctx);
1527 * update cgrp time only if current cgrp
1528 * matches event->cgrp. Must be done before
1529 * calling add_event_to_ctx()
1531 update_cgrp_time_from_event(event);
1533 add_event_to_ctx(event, ctx);
1536 * Schedule everything back in
1538 perf_event_sched_in(cpuctx, task_ctx, task);
1540 perf_pmu_enable(cpuctx->ctx.pmu);
1541 perf_ctx_unlock(cpuctx, task_ctx);
1547 * Attach a performance event to a context
1549 * First we add the event to the list with the hardware enable bit
1550 * in event->hw_config cleared.
1552 * If the event is attached to a task which is on a CPU we use a smp
1553 * call to enable it in the task context. The task might have been
1554 * scheduled away, but we check this in the smp call again.
1557 perf_install_in_context(struct perf_event_context *ctx,
1558 struct perf_event *event,
1561 struct task_struct *task = ctx->task;
1563 lockdep_assert_held(&ctx->mutex);
1569 * Per cpu events are installed via an smp call and
1570 * the install is always successful.
1572 cpu_function_call(cpu, __perf_install_in_context, event);
1577 if (!task_function_call(task, __perf_install_in_context, event))
1580 raw_spin_lock_irq(&ctx->lock);
1582 * If we failed to find a running task, but find the context active now
1583 * that we've acquired the ctx->lock, retry.
1585 if (ctx->is_active) {
1586 raw_spin_unlock_irq(&ctx->lock);
1591 * Since the task isn't running, its safe to add the event, us holding
1592 * the ctx->lock ensures the task won't get scheduled in.
1594 add_event_to_ctx(event, ctx);
1595 raw_spin_unlock_irq(&ctx->lock);
1599 * Put a event into inactive state and update time fields.
1600 * Enabling the leader of a group effectively enables all
1601 * the group members that aren't explicitly disabled, so we
1602 * have to update their ->tstamp_enabled also.
1603 * Note: this works for group members as well as group leaders
1604 * since the non-leader members' sibling_lists will be empty.
1606 static void __perf_event_mark_enabled(struct perf_event *event,
1607 struct perf_event_context *ctx)
1609 struct perf_event *sub;
1610 u64 tstamp = perf_event_time(event);
1612 event->state = PERF_EVENT_STATE_INACTIVE;
1613 event->tstamp_enabled = tstamp - event->total_time_enabled;
1614 list_for_each_entry(sub, &event->sibling_list, group_entry) {
1615 if (sub->state >= PERF_EVENT_STATE_INACTIVE)
1616 sub->tstamp_enabled = tstamp - sub->total_time_enabled;
1621 * Cross CPU call to enable a performance event
1623 static int __perf_event_enable(void *info)
1625 struct perf_event *event = info;
1626 struct perf_event_context *ctx = event->ctx;
1627 struct perf_event *leader = event->group_leader;
1628 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1631 if (WARN_ON_ONCE(!ctx->is_active))
1634 raw_spin_lock(&ctx->lock);
1635 update_context_time(ctx);
1637 if (event->state >= PERF_EVENT_STATE_INACTIVE)
1641 * set current task's cgroup time reference point
1643 perf_cgroup_set_timestamp(current, ctx);
1645 __perf_event_mark_enabled(event, ctx);
1647 if (!event_filter_match(event)) {
1648 if (is_cgroup_event(event))
1649 perf_cgroup_defer_enabled(event);
1654 * If the event is in a group and isn't the group leader,
1655 * then don't put it on unless the group is on.
1657 if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE)
1660 if (!group_can_go_on(event, cpuctx, 1)) {
1663 if (event == leader)
1664 err = group_sched_in(event, cpuctx, ctx);
1666 err = event_sched_in(event, cpuctx, ctx);
1671 * If this event can't go on and it's part of a
1672 * group, then the whole group has to come off.
1674 if (leader != event)
1675 group_sched_out(leader, cpuctx, ctx);
1676 if (leader->attr.pinned) {
1677 update_group_times(leader);
1678 leader->state = PERF_EVENT_STATE_ERROR;
1683 raw_spin_unlock(&ctx->lock);
1691 * If event->ctx is a cloned context, callers must make sure that
1692 * every task struct that event->ctx->task could possibly point to
1693 * remains valid. This condition is satisfied when called through
1694 * perf_event_for_each_child or perf_event_for_each as described
1695 * for perf_event_disable.
1697 void perf_event_enable(struct perf_event *event)
1699 struct perf_event_context *ctx = event->ctx;
1700 struct task_struct *task = ctx->task;
1704 * Enable the event on the cpu that it's on
1706 cpu_function_call(event->cpu, __perf_event_enable, event);
1710 raw_spin_lock_irq(&ctx->lock);
1711 if (event->state >= PERF_EVENT_STATE_INACTIVE)
1715 * If the event is in error state, clear that first.
1716 * That way, if we see the event in error state below, we
1717 * know that it has gone back into error state, as distinct
1718 * from the task having been scheduled away before the
1719 * cross-call arrived.
1721 if (event->state == PERF_EVENT_STATE_ERROR)
1722 event->state = PERF_EVENT_STATE_OFF;
1725 if (!ctx->is_active) {
1726 __perf_event_mark_enabled(event, ctx);
1730 raw_spin_unlock_irq(&ctx->lock);
1732 if (!task_function_call(task, __perf_event_enable, event))
1735 raw_spin_lock_irq(&ctx->lock);
1738 * If the context is active and the event is still off,
1739 * we need to retry the cross-call.
1741 if (ctx->is_active && event->state == PERF_EVENT_STATE_OFF) {
1743 * task could have been flipped by a concurrent
1744 * perf_event_context_sched_out()
1751 raw_spin_unlock_irq(&ctx->lock);
1754 static int perf_event_refresh(struct perf_event *event, int refresh)
1757 * not supported on inherited events
1759 if (event->attr.inherit || !is_sampling_event(event))
1762 atomic_add(refresh, &event->event_limit);
1763 perf_event_enable(event);
1768 static void ctx_sched_out(struct perf_event_context *ctx,
1769 struct perf_cpu_context *cpuctx,
1770 enum event_type_t event_type)
1772 struct perf_event *event;
1773 int is_active = ctx->is_active;
1775 ctx->is_active &= ~event_type;
1776 if (likely(!ctx->nr_events))
1779 update_context_time(ctx);
1780 update_cgrp_time_from_cpuctx(cpuctx);
1781 if (!ctx->nr_active)
1784 perf_pmu_disable(ctx->pmu);
1785 if ((is_active & EVENT_PINNED) && (event_type & EVENT_PINNED)) {
1786 list_for_each_entry(event, &ctx->pinned_groups, group_entry)
1787 group_sched_out(event, cpuctx, ctx);
1790 if ((is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE)) {
1791 list_for_each_entry(event, &ctx->flexible_groups, group_entry)
1792 group_sched_out(event, cpuctx, ctx);
1794 perf_pmu_enable(ctx->pmu);
1798 * Test whether two contexts are equivalent, i.e. whether they
1799 * have both been cloned from the same version of the same context
1800 * and they both have the same number of enabled events.
1801 * If the number of enabled events is the same, then the set
1802 * of enabled events should be the same, because these are both
1803 * inherited contexts, therefore we can't access individual events
1804 * in them directly with an fd; we can only enable/disable all
1805 * events via prctl, or enable/disable all events in a family
1806 * via ioctl, which will have the same effect on both contexts.
1808 static int context_equiv(struct perf_event_context *ctx1,
1809 struct perf_event_context *ctx2)
1811 return ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx
1812 && ctx1->parent_gen == ctx2->parent_gen
1813 && !ctx1->pin_count && !ctx2->pin_count;
1816 static void __perf_event_sync_stat(struct perf_event *event,
1817 struct perf_event *next_event)
1821 if (!event->attr.inherit_stat)
1825 * Update the event value, we cannot use perf_event_read()
1826 * because we're in the middle of a context switch and have IRQs
1827 * disabled, which upsets smp_call_function_single(), however
1828 * we know the event must be on the current CPU, therefore we
1829 * don't need to use it.
1831 switch (event->state) {
1832 case PERF_EVENT_STATE_ACTIVE:
1833 event->pmu->read(event);
1836 case PERF_EVENT_STATE_INACTIVE:
1837 update_event_times(event);
1845 * In order to keep per-task stats reliable we need to flip the event
1846 * values when we flip the contexts.
1848 value = local64_read(&next_event->count);
1849 value = local64_xchg(&event->count, value);
1850 local64_set(&next_event->count, value);
1852 swap(event->total_time_enabled, next_event->total_time_enabled);
1853 swap(event->total_time_running, next_event->total_time_running);
1856 * Since we swizzled the values, update the user visible data too.
1858 perf_event_update_userpage(event);
1859 perf_event_update_userpage(next_event);
1862 #define list_next_entry(pos, member) \
1863 list_entry(pos->member.next, typeof(*pos), member)
1865 static void perf_event_sync_stat(struct perf_event_context *ctx,
1866 struct perf_event_context *next_ctx)
1868 struct perf_event *event, *next_event;
1873 update_context_time(ctx);
1875 event = list_first_entry(&ctx->event_list,
1876 struct perf_event, event_entry);
1878 next_event = list_first_entry(&next_ctx->event_list,
1879 struct perf_event, event_entry);
1881 while (&event->event_entry != &ctx->event_list &&
1882 &next_event->event_entry != &next_ctx->event_list) {
1884 __perf_event_sync_stat(event, next_event);
1886 event = list_next_entry(event, event_entry);
1887 next_event = list_next_entry(next_event, event_entry);
1891 static void perf_event_context_sched_out(struct task_struct *task, int ctxn,
1892 struct task_struct *next)
1894 struct perf_event_context *ctx = task->perf_event_ctxp[ctxn];
1895 struct perf_event_context *next_ctx;
1896 struct perf_event_context *parent;
1897 struct perf_cpu_context *cpuctx;
1903 cpuctx = __get_cpu_context(ctx);
1904 if (!cpuctx->task_ctx)
1908 parent = rcu_dereference(ctx->parent_ctx);
1909 next_ctx = next->perf_event_ctxp[ctxn];
1910 if (parent && next_ctx &&
1911 rcu_dereference(next_ctx->parent_ctx) == parent) {
1913 * Looks like the two contexts are clones, so we might be
1914 * able to optimize the context switch. We lock both
1915 * contexts and check that they are clones under the
1916 * lock (including re-checking that neither has been
1917 * uncloned in the meantime). It doesn't matter which
1918 * order we take the locks because no other cpu could
1919 * be trying to lock both of these tasks.
1921 raw_spin_lock(&ctx->lock);
1922 raw_spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING);
1923 if (context_equiv(ctx, next_ctx)) {
1925 * XXX do we need a memory barrier of sorts
1926 * wrt to rcu_dereference() of perf_event_ctxp
1928 task->perf_event_ctxp[ctxn] = next_ctx;
1929 next->perf_event_ctxp[ctxn] = ctx;
1931 next_ctx->task = task;
1934 perf_event_sync_stat(ctx, next_ctx);
1936 raw_spin_unlock(&next_ctx->lock);
1937 raw_spin_unlock(&ctx->lock);
1942 raw_spin_lock(&ctx->lock);
1943 ctx_sched_out(ctx, cpuctx, EVENT_ALL);
1944 cpuctx->task_ctx = NULL;
1945 raw_spin_unlock(&ctx->lock);
1949 #define for_each_task_context_nr(ctxn) \
1950 for ((ctxn) = 0; (ctxn) < perf_nr_task_contexts; (ctxn)++)
1953 * Called from scheduler to remove the events of the current task,
1954 * with interrupts disabled.
1956 * We stop each event and update the event value in event->count.
1958 * This does not protect us against NMI, but disable()
1959 * sets the disabled bit in the control field of event _before_
1960 * accessing the event control register. If a NMI hits, then it will
1961 * not restart the event.
1963 void __perf_event_task_sched_out(struct task_struct *task,
1964 struct task_struct *next)
1968 for_each_task_context_nr(ctxn)
1969 perf_event_context_sched_out(task, ctxn, next);
1972 * if cgroup events exist on this CPU, then we need
1973 * to check if we have to switch out PMU state.
1974 * cgroup event are system-wide mode only
1976 if (atomic_read(&__get_cpu_var(perf_cgroup_events)))
1977 perf_cgroup_sched_out(task);
1980 static void task_ctx_sched_out(struct perf_event_context *ctx)
1982 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1984 if (!cpuctx->task_ctx)
1987 if (WARN_ON_ONCE(ctx != cpuctx->task_ctx))
1990 ctx_sched_out(ctx, cpuctx, EVENT_ALL);
1991 cpuctx->task_ctx = NULL;
1995 * Called with IRQs disabled
1997 static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
1998 enum event_type_t event_type)
2000 ctx_sched_out(&cpuctx->ctx, cpuctx, event_type);
2004 ctx_pinned_sched_in(struct perf_event_context *ctx,
2005 struct perf_cpu_context *cpuctx)
2007 struct perf_event *event;
2009 list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
2010 if (event->state <= PERF_EVENT_STATE_OFF)
2012 if (!event_filter_match(event))
2015 /* may need to reset tstamp_enabled */
2016 if (is_cgroup_event(event))
2017 perf_cgroup_mark_enabled(event, ctx);
2019 if (group_can_go_on(event, cpuctx, 1))
2020 group_sched_in(event, cpuctx, ctx);
2023 * If this pinned group hasn't been scheduled,
2024 * put it in error state.
2026 if (event->state == PERF_EVENT_STATE_INACTIVE) {
2027 update_group_times(event);
2028 event->state = PERF_EVENT_STATE_ERROR;
2034 ctx_flexible_sched_in(struct perf_event_context *ctx,
2035 struct perf_cpu_context *cpuctx)
2037 struct perf_event *event;
2040 list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
2041 /* Ignore events in OFF or ERROR state */
2042 if (event->state <= PERF_EVENT_STATE_OFF)
2045 * Listen to the 'cpu' scheduling filter constraint
2048 if (!event_filter_match(event))
2051 /* may need to reset tstamp_enabled */
2052 if (is_cgroup_event(event))
2053 perf_cgroup_mark_enabled(event, ctx);
2055 if (group_can_go_on(event, cpuctx, can_add_hw)) {
2056 if (group_sched_in(event, cpuctx, ctx))
2063 ctx_sched_in(struct perf_event_context *ctx,
2064 struct perf_cpu_context *cpuctx,
2065 enum event_type_t event_type,
2066 struct task_struct *task)
2069 int is_active = ctx->is_active;
2071 ctx->is_active |= event_type;
2072 if (likely(!ctx->nr_events))
2076 ctx->timestamp = now;
2077 perf_cgroup_set_timestamp(task, ctx);
2079 * First go through the list and put on any pinned groups
2080 * in order to give them the best chance of going on.
2082 if (!(is_active & EVENT_PINNED) && (event_type & EVENT_PINNED))
2083 ctx_pinned_sched_in(ctx, cpuctx);
2085 /* Then walk through the lower prio flexible groups */
2086 if (!(is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE))
2087 ctx_flexible_sched_in(ctx, cpuctx);
2090 static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
2091 enum event_type_t event_type,
2092 struct task_struct *task)
2094 struct perf_event_context *ctx = &cpuctx->ctx;
2096 ctx_sched_in(ctx, cpuctx, event_type, task);
2099 static void perf_event_context_sched_in(struct perf_event_context *ctx,
2100 struct task_struct *task)
2102 struct perf_cpu_context *cpuctx;
2104 cpuctx = __get_cpu_context(ctx);
2105 if (cpuctx->task_ctx == ctx)
2108 perf_ctx_lock(cpuctx, ctx);
2109 perf_pmu_disable(ctx->pmu);
2111 * We want to keep the following priority order:
2112 * cpu pinned (that don't need to move), task pinned,
2113 * cpu flexible, task flexible.
2115 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
2117 perf_event_sched_in(cpuctx, ctx, task);
2119 cpuctx->task_ctx = ctx;
2121 perf_pmu_enable(ctx->pmu);
2122 perf_ctx_unlock(cpuctx, ctx);
2125 * Since these rotations are per-cpu, we need to ensure the
2126 * cpu-context we got scheduled on is actually rotating.
2128 perf_pmu_rotate_start(ctx->pmu);
2132 * Called from scheduler to add the events of the current task
2133 * with interrupts disabled.
2135 * We restore the event value and then enable it.
2137 * This does not protect us against NMI, but enable()
2138 * sets the enabled bit in the control field of event _before_
2139 * accessing the event control register. If a NMI hits, then it will
2140 * keep the event running.
2142 void __perf_event_task_sched_in(struct task_struct *task)
2144 struct perf_event_context *ctx;
2147 for_each_task_context_nr(ctxn) {
2148 ctx = task->perf_event_ctxp[ctxn];
2152 perf_event_context_sched_in(ctx, task);
2155 * if cgroup events exist on this CPU, then we need
2156 * to check if we have to switch in PMU state.
2157 * cgroup event are system-wide mode only
2159 if (atomic_read(&__get_cpu_var(perf_cgroup_events)))
2160 perf_cgroup_sched_in(task);
2163 static u64 perf_calculate_period(struct perf_event *event, u64 nsec, u64 count)
2165 u64 frequency = event->attr.sample_freq;
2166 u64 sec = NSEC_PER_SEC;
2167 u64 divisor, dividend;
2169 int count_fls, nsec_fls, frequency_fls, sec_fls;
2171 count_fls = fls64(count);
2172 nsec_fls = fls64(nsec);
2173 frequency_fls = fls64(frequency);
2177 * We got @count in @nsec, with a target of sample_freq HZ
2178 * the target period becomes:
2181 * period = -------------------
2182 * @nsec * sample_freq
2187 * Reduce accuracy by one bit such that @a and @b converge
2188 * to a similar magnitude.
2190 #define REDUCE_FLS(a, b) \
2192 if (a##_fls > b##_fls) { \
2202 * Reduce accuracy until either term fits in a u64, then proceed with
2203 * the other, so that finally we can do a u64/u64 division.
2205 while (count_fls + sec_fls > 64 && nsec_fls + frequency_fls > 64) {
2206 REDUCE_FLS(nsec, frequency);
2207 REDUCE_FLS(sec, count);
2210 if (count_fls + sec_fls > 64) {
2211 divisor = nsec * frequency;
2213 while (count_fls + sec_fls > 64) {
2214 REDUCE_FLS(count, sec);
2218 dividend = count * sec;
2220 dividend = count * sec;
2222 while (nsec_fls + frequency_fls > 64) {
2223 REDUCE_FLS(nsec, frequency);
2227 divisor = nsec * frequency;
2233 return div64_u64(dividend, divisor);
2236 static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count)
2238 struct hw_perf_event *hwc = &event->hw;
2239 s64 period, sample_period;
2242 period = perf_calculate_period(event, nsec, count);
2244 delta = (s64)(period - hwc->sample_period);
2245 delta = (delta + 7) / 8; /* low pass filter */
2247 sample_period = hwc->sample_period + delta;
2252 hwc->sample_period = sample_period;
2254 if (local64_read(&hwc->period_left) > 8*sample_period) {
2255 event->pmu->stop(event, PERF_EF_UPDATE);
2256 local64_set(&hwc->period_left, 0);
2257 event->pmu->start(event, PERF_EF_RELOAD);
2261 static void perf_ctx_adjust_freq(struct perf_event_context *ctx, u64 period)
2263 struct perf_event *event;
2264 struct hw_perf_event *hwc;
2265 u64 interrupts, now;
2268 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
2269 if (event->state != PERF_EVENT_STATE_ACTIVE)
2272 if (!event_filter_match(event))
2277 interrupts = hwc->interrupts;
2278 hwc->interrupts = 0;
2281 * unthrottle events on the tick
2283 if (interrupts == MAX_INTERRUPTS) {
2284 perf_log_throttle(event, 1);
2285 event->pmu->start(event, 0);
2288 if (!event->attr.freq || !event->attr.sample_freq)
2291 event->pmu->read(event);
2292 now = local64_read(&event->count);
2293 delta = now - hwc->freq_count_stamp;
2294 hwc->freq_count_stamp = now;
2297 perf_adjust_period(event, period, delta);
2302 * Round-robin a context's events:
2304 static void rotate_ctx(struct perf_event_context *ctx)
2307 * Rotate the first entry last of non-pinned groups. Rotation might be
2308 * disabled by the inheritance code.
2310 if (!ctx->rotate_disable)
2311 list_rotate_left(&ctx->flexible_groups);
2315 * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
2316 * because they're strictly cpu affine and rotate_start is called with IRQs
2317 * disabled, while rotate_context is called from IRQ context.
2319 static void perf_rotate_context(struct perf_cpu_context *cpuctx)
2321 u64 interval = (u64)cpuctx->jiffies_interval * TICK_NSEC;
2322 struct perf_event_context *ctx = NULL;
2323 int rotate = 0, remove = 1;
2325 if (cpuctx->ctx.nr_events) {
2327 if (cpuctx->ctx.nr_events != cpuctx->ctx.nr_active)
2331 ctx = cpuctx->task_ctx;
2332 if (ctx && ctx->nr_events) {
2334 if (ctx->nr_events != ctx->nr_active)
2338 perf_ctx_lock(cpuctx, cpuctx->task_ctx);
2339 perf_pmu_disable(cpuctx->ctx.pmu);
2340 perf_ctx_adjust_freq(&cpuctx->ctx, interval);
2342 perf_ctx_adjust_freq(ctx, interval);
2347 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
2349 ctx_sched_out(ctx, cpuctx, EVENT_FLEXIBLE);
2351 rotate_ctx(&cpuctx->ctx);
2355 perf_event_sched_in(cpuctx, ctx, current);
2359 list_del_init(&cpuctx->rotation_list);
2361 perf_pmu_enable(cpuctx->ctx.pmu);
2362 perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
2365 void perf_event_task_tick(void)
2367 struct list_head *head = &__get_cpu_var(rotation_list);
2368 struct perf_cpu_context *cpuctx, *tmp;
2370 WARN_ON(!irqs_disabled());
2372 list_for_each_entry_safe(cpuctx, tmp, head, rotation_list) {
2373 if (cpuctx->jiffies_interval == 1 ||
2374 !(jiffies % cpuctx->jiffies_interval))
2375 perf_rotate_context(cpuctx);
2379 static int event_enable_on_exec(struct perf_event *event,
2380 struct perf_event_context *ctx)
2382 if (!event->attr.enable_on_exec)
2385 event->attr.enable_on_exec = 0;
2386 if (event->state >= PERF_EVENT_STATE_INACTIVE)
2389 __perf_event_mark_enabled(event, ctx);
2395 * Enable all of a task's events that have been marked enable-on-exec.
2396 * This expects task == current.
2398 static void perf_event_enable_on_exec(struct perf_event_context *ctx)
2400 struct perf_event *event;
2401 unsigned long flags;
2405 local_irq_save(flags);
2406 if (!ctx || !ctx->nr_events)
2410 * We must ctxsw out cgroup events to avoid conflict
2411 * when invoking perf_task_event_sched_in() later on
2412 * in this function. Otherwise we end up trying to
2413 * ctxswin cgroup events which are already scheduled
2416 perf_cgroup_sched_out(current);
2418 raw_spin_lock(&ctx->lock);
2419 task_ctx_sched_out(ctx);
2421 list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
2422 ret = event_enable_on_exec(event, ctx);
2427 list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
2428 ret = event_enable_on_exec(event, ctx);
2434 * Unclone this context if we enabled any event.
2439 raw_spin_unlock(&ctx->lock);
2442 * Also calls ctxswin for cgroup events, if any:
2444 perf_event_context_sched_in(ctx, ctx->task);
2446 local_irq_restore(flags);
2450 * Cross CPU call to read the hardware event
2452 static void __perf_event_read(void *info)
2454 struct perf_event *event = info;
2455 struct perf_event_context *ctx = event->ctx;
2456 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
2459 * If this is a task context, we need to check whether it is
2460 * the current task context of this cpu. If not it has been
2461 * scheduled out before the smp call arrived. In that case
2462 * event->count would have been updated to a recent sample
2463 * when the event was scheduled out.
2465 if (ctx->task && cpuctx->task_ctx != ctx)
2468 raw_spin_lock(&ctx->lock);
2469 if (ctx->is_active) {
2470 update_context_time(ctx);
2471 update_cgrp_time_from_event(event);
2473 update_event_times(event);
2474 if (event->state == PERF_EVENT_STATE_ACTIVE)
2475 event->pmu->read(event);
2476 raw_spin_unlock(&ctx->lock);
2479 static inline u64 perf_event_count(struct perf_event *event)
2481 return local64_read(&event->count) + atomic64_read(&event->child_count);
2484 static u64 perf_event_read(struct perf_event *event)
2487 * If event is enabled and currently active on a CPU, update the
2488 * value in the event structure:
2490 if (event->state == PERF_EVENT_STATE_ACTIVE) {
2491 smp_call_function_single(event->oncpu,
2492 __perf_event_read, event, 1);
2493 } else if (event->state == PERF_EVENT_STATE_INACTIVE) {
2494 struct perf_event_context *ctx = event->ctx;
2495 unsigned long flags;
2497 raw_spin_lock_irqsave(&ctx->lock, flags);
2499 * may read while context is not active
2500 * (e.g., thread is blocked), in that case
2501 * we cannot update context time
2503 if (ctx->is_active) {
2504 update_context_time(ctx);
2505 update_cgrp_time_from_event(event);
2507 update_event_times(event);
2508 raw_spin_unlock_irqrestore(&ctx->lock, flags);
2511 return perf_event_count(event);
2518 struct callchain_cpus_entries {
2519 struct rcu_head rcu_head;
2520 struct perf_callchain_entry *cpu_entries[0];
2523 static DEFINE_PER_CPU(int, callchain_recursion[PERF_NR_CONTEXTS]);
2524 static atomic_t nr_callchain_events;
2525 static DEFINE_MUTEX(callchain_mutex);
2526 struct callchain_cpus_entries *callchain_cpus_entries;
2529 __weak void perf_callchain_kernel(struct perf_callchain_entry *entry,
2530 struct pt_regs *regs)
2534 __weak void perf_callchain_user(struct perf_callchain_entry *entry,
2535 struct pt_regs *regs)
2539 static void release_callchain_buffers_rcu(struct rcu_head *head)
2541 struct callchain_cpus_entries *entries;
2544 entries = container_of(head, struct callchain_cpus_entries, rcu_head);
2546 for_each_possible_cpu(cpu)
2547 kfree(entries->cpu_entries[cpu]);
2552 static void release_callchain_buffers(void)
2554 struct callchain_cpus_entries *entries;
2556 entries = callchain_cpus_entries;
2557 rcu_assign_pointer(callchain_cpus_entries, NULL);
2558 call_rcu(&entries->rcu_head, release_callchain_buffers_rcu);
2561 static int alloc_callchain_buffers(void)
2565 struct callchain_cpus_entries *entries;
2568 * We can't use the percpu allocation API for data that can be
2569 * accessed from NMI. Use a temporary manual per cpu allocation
2570 * until that gets sorted out.
2572 size = offsetof(struct callchain_cpus_entries, cpu_entries[nr_cpu_ids]);
2574 entries = kzalloc(size, GFP_KERNEL);
2578 size = sizeof(struct perf_callchain_entry) * PERF_NR_CONTEXTS;
2580 for_each_possible_cpu(cpu) {
2581 entries->cpu_entries[cpu] = kmalloc_node(size, GFP_KERNEL,
2583 if (!entries->cpu_entries[cpu])
2587 rcu_assign_pointer(callchain_cpus_entries, entries);
2592 for_each_possible_cpu(cpu)
2593 kfree(entries->cpu_entries[cpu]);
2599 static int get_callchain_buffers(void)
2604 mutex_lock(&callchain_mutex);
2606 count = atomic_inc_return(&nr_callchain_events);
2607 if (WARN_ON_ONCE(count < 1)) {
2613 /* If the allocation failed, give up */
2614 if (!callchain_cpus_entries)
2619 err = alloc_callchain_buffers();
2621 release_callchain_buffers();
2623 mutex_unlock(&callchain_mutex);
2628 static void put_callchain_buffers(void)
2630 if (atomic_dec_and_mutex_lock(&nr_callchain_events, &callchain_mutex)) {
2631 release_callchain_buffers();
2632 mutex_unlock(&callchain_mutex);
2636 static int get_recursion_context(int *recursion)
2644 else if (in_softirq())
2649 if (recursion[rctx])
2658 static inline void put_recursion_context(int *recursion, int rctx)
2664 static struct perf_callchain_entry *get_callchain_entry(int *rctx)
2667 struct callchain_cpus_entries *entries;
2669 *rctx = get_recursion_context(__get_cpu_var(callchain_recursion));
2673 entries = rcu_dereference(callchain_cpus_entries);
2677 cpu = smp_processor_id();
2679 return &entries->cpu_entries[cpu][*rctx];
2683 put_callchain_entry(int rctx)
2685 put_recursion_context(__get_cpu_var(callchain_recursion), rctx);
2688 static struct perf_callchain_entry *perf_callchain(struct pt_regs *regs)
2691 struct perf_callchain_entry *entry;
2694 entry = get_callchain_entry(&rctx);
2703 if (!user_mode(regs)) {
2704 perf_callchain_store(entry, PERF_CONTEXT_KERNEL);
2705 perf_callchain_kernel(entry, regs);
2707 regs = task_pt_regs(current);
2713 perf_callchain_store(entry, PERF_CONTEXT_USER);
2714 perf_callchain_user(entry, regs);
2718 put_callchain_entry(rctx);
2724 * Initialize the perf_event context in a task_struct:
2726 static void __perf_event_init_context(struct perf_event_context *ctx)
2728 raw_spin_lock_init(&ctx->lock);
2729 mutex_init(&ctx->mutex);
2730 INIT_LIST_HEAD(&ctx->pinned_groups);
2731 INIT_LIST_HEAD(&ctx->flexible_groups);
2732 INIT_LIST_HEAD(&ctx->event_list);
2733 atomic_set(&ctx->refcount, 1);
2736 static struct perf_event_context *
2737 alloc_perf_context(struct pmu *pmu, struct task_struct *task)
2739 struct perf_event_context *ctx;
2741 ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL);
2745 __perf_event_init_context(ctx);
2748 get_task_struct(task);
2755 static struct task_struct *
2756 find_lively_task_by_vpid(pid_t vpid)
2758 struct task_struct *task;
2765 task = find_task_by_vpid(vpid);
2767 get_task_struct(task);
2771 return ERR_PTR(-ESRCH);
2773 /* Reuse ptrace permission checks for now. */
2775 if (!ptrace_may_access(task, PTRACE_MODE_READ))
2780 put_task_struct(task);
2781 return ERR_PTR(err);
2786 * Returns a matching context with refcount and pincount.
2788 static struct perf_event_context *
2789 find_get_context(struct pmu *pmu, struct task_struct *task, int cpu)
2791 struct perf_event_context *ctx;
2792 struct perf_cpu_context *cpuctx;
2793 unsigned long flags;
2797 /* Must be root to operate on a CPU event: */
2798 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
2799 return ERR_PTR(-EACCES);
2802 * We could be clever and allow to attach a event to an
2803 * offline CPU and activate it when the CPU comes up, but
2806 if (!cpu_online(cpu))
2807 return ERR_PTR(-ENODEV);
2809 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
2818 ctxn = pmu->task_ctx_nr;
2823 ctx = perf_lock_task_context(task, ctxn, &flags);
2827 raw_spin_unlock_irqrestore(&ctx->lock, flags);
2829 ctx = alloc_perf_context(pmu, task);
2835 mutex_lock(&task->perf_event_mutex);
2837 * If it has already passed perf_event_exit_task().
2838 * we must see PF_EXITING, it takes this mutex too.
2840 if (task->flags & PF_EXITING)
2842 else if (task->perf_event_ctxp[ctxn])
2847 rcu_assign_pointer(task->perf_event_ctxp[ctxn], ctx);
2849 mutex_unlock(&task->perf_event_mutex);
2851 if (unlikely(err)) {
2863 return ERR_PTR(err);
2866 static void perf_event_free_filter(struct perf_event *event);
2868 static void free_event_rcu(struct rcu_head *head)
2870 struct perf_event *event;
2872 event = container_of(head, struct perf_event, rcu_head);
2874 put_pid_ns(event->ns);
2875 perf_event_free_filter(event);
2879 static void perf_buffer_put(struct perf_buffer *buffer);
2881 static void free_event(struct perf_event *event)
2883 irq_work_sync(&event->pending);
2885 if (!event->parent) {
2886 if (event->attach_state & PERF_ATTACH_TASK)
2887 jump_label_dec(&perf_sched_events);
2888 if (event->attr.mmap || event->attr.mmap_data)
2889 atomic_dec(&nr_mmap_events);
2890 if (event->attr.comm)
2891 atomic_dec(&nr_comm_events);
2892 if (event->attr.task)
2893 atomic_dec(&nr_task_events);
2894 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN)
2895 put_callchain_buffers();
2896 if (is_cgroup_event(event)) {
2897 atomic_dec(&per_cpu(perf_cgroup_events, event->cpu));
2898 jump_label_dec(&perf_sched_events);
2902 if (event->buffer) {
2903 perf_buffer_put(event->buffer);
2904 event->buffer = NULL;
2907 if (is_cgroup_event(event))
2908 perf_detach_cgroup(event);
2911 event->destroy(event);
2914 put_ctx(event->ctx);
2916 call_rcu(&event->rcu_head, free_event_rcu);
2919 int perf_event_release_kernel(struct perf_event *event)
2921 struct perf_event_context *ctx = event->ctx;
2923 WARN_ON_ONCE(ctx->parent_ctx);
2925 * There are two ways this annotation is useful:
2927 * 1) there is a lock recursion from perf_event_exit_task
2928 * see the comment there.
2930 * 2) there is a lock-inversion with mmap_sem through
2931 * perf_event_read_group(), which takes faults while
2932 * holding ctx->mutex, however this is called after
2933 * the last filedesc died, so there is no possibility
2934 * to trigger the AB-BA case.
2936 mutex_lock_nested(&ctx->mutex, SINGLE_DEPTH_NESTING);
2937 raw_spin_lock_irq(&ctx->lock);
2938 perf_group_detach(event);
2939 raw_spin_unlock_irq(&ctx->lock);
2940 perf_remove_from_context(event);
2941 mutex_unlock(&ctx->mutex);
2947 EXPORT_SYMBOL_GPL(perf_event_release_kernel);
2950 * Called when the last reference to the file is gone.
2952 static int perf_release(struct inode *inode, struct file *file)
2954 struct perf_event *event = file->private_data;
2955 struct task_struct *owner;
2957 file->private_data = NULL;
2960 owner = ACCESS_ONCE(event->owner);
2962 * Matches the smp_wmb() in perf_event_exit_task(). If we observe
2963 * !owner it means the list deletion is complete and we can indeed
2964 * free this event, otherwise we need to serialize on
2965 * owner->perf_event_mutex.
2967 smp_read_barrier_depends();
2970 * Since delayed_put_task_struct() also drops the last
2971 * task reference we can safely take a new reference
2972 * while holding the rcu_read_lock().
2974 get_task_struct(owner);
2979 mutex_lock(&owner->perf_event_mutex);
2981 * We have to re-check the event->owner field, if it is cleared
2982 * we raced with perf_event_exit_task(), acquiring the mutex
2983 * ensured they're done, and we can proceed with freeing the
2987 list_del_init(&event->owner_entry);
2988 mutex_unlock(&owner->perf_event_mutex);
2989 put_task_struct(owner);
2992 return perf_event_release_kernel(event);
2995 u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running)
2997 struct perf_event *child;
3003 mutex_lock(&event->child_mutex);
3004 total += perf_event_read(event);
3005 *enabled += event->total_time_enabled +
3006 atomic64_read(&event->child_total_time_enabled);
3007 *running += event->total_time_running +
3008 atomic64_read(&event->child_total_time_running);
3010 list_for_each_entry(child, &event->child_list, child_list) {
3011 total += perf_event_read(child);
3012 *enabled += child->total_time_enabled;
3013 *running += child->total_time_running;
3015 mutex_unlock(&event->child_mutex);
3019 EXPORT_SYMBOL_GPL(perf_event_read_value);
3021 static int perf_event_read_group(struct perf_event *event,
3022 u64 read_format, char __user *buf)
3024 struct perf_event *leader = event->group_leader, *sub;
3025 int n = 0, size = 0, ret = -EFAULT;
3026 struct perf_event_context *ctx = leader->ctx;
3028 u64 count, enabled, running;
3030 mutex_lock(&ctx->mutex);
3031 count = perf_event_read_value(leader, &enabled, &running);
3033 values[n++] = 1 + leader->nr_siblings;
3034 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3035 values[n++] = enabled;
3036 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3037 values[n++] = running;
3038 values[n++] = count;
3039 if (read_format & PERF_FORMAT_ID)
3040 values[n++] = primary_event_id(leader);
3042 size = n * sizeof(u64);
3044 if (copy_to_user(buf, values, size))
3049 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
3052 values[n++] = perf_event_read_value(sub, &enabled, &running);
3053 if (read_format & PERF_FORMAT_ID)
3054 values[n++] = primary_event_id(sub);
3056 size = n * sizeof(u64);
3058 if (copy_to_user(buf + ret, values, size)) {
3066 mutex_unlock(&ctx->mutex);
3071 static int perf_event_read_one(struct perf_event *event,
3072 u64 read_format, char __user *buf)
3074 u64 enabled, running;
3078 values[n++] = perf_event_read_value(event, &enabled, &running);
3079 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3080 values[n++] = enabled;
3081 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3082 values[n++] = running;
3083 if (read_format & PERF_FORMAT_ID)
3084 values[n++] = primary_event_id(event);
3086 if (copy_to_user(buf, values, n * sizeof(u64)))
3089 return n * sizeof(u64);
3093 * Read the performance event - simple non blocking version for now
3096 perf_read_hw(struct perf_event *event, char __user *buf, size_t count)
3098 u64 read_format = event->attr.read_format;
3102 * Return end-of-file for a read on a event that is in
3103 * error state (i.e. because it was pinned but it couldn't be
3104 * scheduled on to the CPU at some point).
3106 if (event->state == PERF_EVENT_STATE_ERROR)
3109 if (count < event->read_size)
3112 WARN_ON_ONCE(event->ctx->parent_ctx);
3113 if (read_format & PERF_FORMAT_GROUP)
3114 ret = perf_event_read_group(event, read_format, buf);
3116 ret = perf_event_read_one(event, read_format, buf);
3122 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
3124 struct perf_event *event = file->private_data;
3126 return perf_read_hw(event, buf, count);
3129 static unsigned int perf_poll(struct file *file, poll_table *wait)
3131 struct perf_event *event = file->private_data;
3132 struct perf_buffer *buffer;
3133 unsigned int events = POLL_HUP;
3136 buffer = rcu_dereference(event->buffer);
3138 events = atomic_xchg(&buffer->poll, 0);
3141 poll_wait(file, &event->waitq, wait);
3146 static void perf_event_reset(struct perf_event *event)
3148 (void)perf_event_read(event);
3149 local64_set(&event->count, 0);
3150 perf_event_update_userpage(event);
3154 * Holding the top-level event's child_mutex means that any
3155 * descendant process that has inherited this event will block
3156 * in sync_child_event if it goes to exit, thus satisfying the
3157 * task existence requirements of perf_event_enable/disable.
3159 static void perf_event_for_each_child(struct perf_event *event,
3160 void (*func)(struct perf_event *))
3162 struct perf_event *child;
3164 WARN_ON_ONCE(event->ctx->parent_ctx);
3165 mutex_lock(&event->child_mutex);
3167 list_for_each_entry(child, &event->child_list, child_list)
3169 mutex_unlock(&event->child_mutex);
3172 static void perf_event_for_each(struct perf_event *event,
3173 void (*func)(struct perf_event *))
3175 struct perf_event_context *ctx = event->ctx;
3176 struct perf_event *sibling;
3178 WARN_ON_ONCE(ctx->parent_ctx);
3179 mutex_lock(&ctx->mutex);
3180 event = event->group_leader;
3182 perf_event_for_each_child(event, func);
3184 list_for_each_entry(sibling, &event->sibling_list, group_entry)
3185 perf_event_for_each_child(event, func);
3186 mutex_unlock(&ctx->mutex);
3189 static int perf_event_period(struct perf_event *event, u64 __user *arg)
3191 struct perf_event_context *ctx = event->ctx;
3195 if (!is_sampling_event(event))
3198 if (copy_from_user(&value, arg, sizeof(value)))
3204 raw_spin_lock_irq(&ctx->lock);
3205 if (event->attr.freq) {
3206 if (value > sysctl_perf_event_sample_rate) {
3211 event->attr.sample_freq = value;
3213 event->attr.sample_period = value;
3214 event->hw.sample_period = value;
3217 raw_spin_unlock_irq(&ctx->lock);
3222 static const struct file_operations perf_fops;
3224 static struct perf_event *perf_fget_light(int fd, int *fput_needed)
3228 file = fget_light(fd, fput_needed);
3230 return ERR_PTR(-EBADF);
3232 if (file->f_op != &perf_fops) {
3233 fput_light(file, *fput_needed);
3235 return ERR_PTR(-EBADF);
3238 return file->private_data;
3241 static int perf_event_set_output(struct perf_event *event,
3242 struct perf_event *output_event);
3243 static int perf_event_set_filter(struct perf_event *event, void __user *arg);
3245 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
3247 struct perf_event *event = file->private_data;
3248 void (*func)(struct perf_event *);
3252 case PERF_EVENT_IOC_ENABLE:
3253 func = perf_event_enable;
3255 case PERF_EVENT_IOC_DISABLE:
3256 func = perf_event_disable;
3258 case PERF_EVENT_IOC_RESET:
3259 func = perf_event_reset;
3262 case PERF_EVENT_IOC_REFRESH:
3263 return perf_event_refresh(event, arg);
3265 case PERF_EVENT_IOC_PERIOD:
3266 return perf_event_period(event, (u64 __user *)arg);
3268 case PERF_EVENT_IOC_SET_OUTPUT:
3270 struct perf_event *output_event = NULL;
3271 int fput_needed = 0;
3275 output_event = perf_fget_light(arg, &fput_needed);
3276 if (IS_ERR(output_event))
3277 return PTR_ERR(output_event);
3280 ret = perf_event_set_output(event, output_event);
3282 fput_light(output_event->filp, fput_needed);
3287 case PERF_EVENT_IOC_SET_FILTER:
3288 return perf_event_set_filter(event, (void __user *)arg);
3294 if (flags & PERF_IOC_FLAG_GROUP)
3295 perf_event_for_each(event, func);
3297 perf_event_for_each_child(event, func);
3302 int perf_event_task_enable(void)
3304 struct perf_event *event;
3306 mutex_lock(¤t->perf_event_mutex);
3307 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
3308 perf_event_for_each_child(event, perf_event_enable);
3309 mutex_unlock(¤t->perf_event_mutex);
3314 int perf_event_task_disable(void)
3316 struct perf_event *event;
3318 mutex_lock(¤t->perf_event_mutex);
3319 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
3320 perf_event_for_each_child(event, perf_event_disable);
3321 mutex_unlock(¤t->perf_event_mutex);
3326 #ifndef PERF_EVENT_INDEX_OFFSET
3327 # define PERF_EVENT_INDEX_OFFSET 0
3330 static int perf_event_index(struct perf_event *event)
3332 if (event->hw.state & PERF_HES_STOPPED)
3335 if (event->state != PERF_EVENT_STATE_ACTIVE)
3338 return event->hw.idx + 1 - PERF_EVENT_INDEX_OFFSET;
3342 * Callers need to ensure there can be no nesting of this function, otherwise
3343 * the seqlock logic goes bad. We can not serialize this because the arch
3344 * code calls this from NMI context.
3346 void perf_event_update_userpage(struct perf_event *event)
3348 struct perf_event_mmap_page *userpg;
3349 struct perf_buffer *buffer;
3352 buffer = rcu_dereference(event->buffer);
3356 userpg = buffer->user_page;
3359 * Disable preemption so as to not let the corresponding user-space
3360 * spin too long if we get preempted.
3365 userpg->index = perf_event_index(event);
3366 userpg->offset = perf_event_count(event);
3367 if (event->state == PERF_EVENT_STATE_ACTIVE)
3368 userpg->offset -= local64_read(&event->hw.prev_count);
3370 userpg->time_enabled = event->total_time_enabled +
3371 atomic64_read(&event->child_total_time_enabled);
3373 userpg->time_running = event->total_time_running +
3374 atomic64_read(&event->child_total_time_running);
3383 static unsigned long perf_data_size(struct perf_buffer *buffer);
3386 perf_buffer_init(struct perf_buffer *buffer, long watermark, int flags)
3388 long max_size = perf_data_size(buffer);
3391 buffer->watermark = min(max_size, watermark);
3393 if (!buffer->watermark)
3394 buffer->watermark = max_size / 2;
3396 if (flags & PERF_BUFFER_WRITABLE)
3397 buffer->writable = 1;
3399 atomic_set(&buffer->refcount, 1);
3402 #ifndef CONFIG_PERF_USE_VMALLOC
3405 * Back perf_mmap() with regular GFP_KERNEL-0 pages.
3408 static struct page *
3409 perf_mmap_to_page(struct perf_buffer *buffer, unsigned long pgoff)
3411 if (pgoff > buffer->nr_pages)
3415 return virt_to_page(buffer->user_page);
3417 return virt_to_page(buffer->data_pages[pgoff - 1]);
3420 static void *perf_mmap_alloc_page(int cpu)
3425 node = (cpu == -1) ? cpu : cpu_to_node(cpu);
3426 page = alloc_pages_node(node, GFP_KERNEL | __GFP_ZERO, 0);
3430 return page_address(page);
3433 static struct perf_buffer *
3434 perf_buffer_alloc(int nr_pages, long watermark, int cpu, int flags)
3436 struct perf_buffer *buffer;
3440 size = sizeof(struct perf_buffer);
3441 size += nr_pages * sizeof(void *);
3443 buffer = kzalloc(size, GFP_KERNEL);
3447 buffer->user_page = perf_mmap_alloc_page(cpu);
3448 if (!buffer->user_page)
3449 goto fail_user_page;
3451 for (i = 0; i < nr_pages; i++) {
3452 buffer->data_pages[i] = perf_mmap_alloc_page(cpu);
3453 if (!buffer->data_pages[i])
3454 goto fail_data_pages;
3457 buffer->nr_pages = nr_pages;
3459 perf_buffer_init(buffer, watermark, flags);
3464 for (i--; i >= 0; i--)
3465 free_page((unsigned long)buffer->data_pages[i]);
3467 free_page((unsigned long)buffer->user_page);
3476 static void perf_mmap_free_page(unsigned long addr)
3478 struct page *page = virt_to_page((void *)addr);
3480 page->mapping = NULL;
3484 static void perf_buffer_free(struct perf_buffer *buffer)
3488 perf_mmap_free_page((unsigned long)buffer->user_page);
3489 for (i = 0; i < buffer->nr_pages; i++)
3490 perf_mmap_free_page((unsigned long)buffer->data_pages[i]);
3494 static inline int page_order(struct perf_buffer *buffer)
3502 * Back perf_mmap() with vmalloc memory.
3504 * Required for architectures that have d-cache aliasing issues.
3507 static inline int page_order(struct perf_buffer *buffer)
3509 return buffer->page_order;
3512 static struct page *
3513 perf_mmap_to_page(struct perf_buffer *buffer, unsigned long pgoff)
3515 if (pgoff > (1UL << page_order(buffer)))
3518 return vmalloc_to_page((void *)buffer->user_page + pgoff * PAGE_SIZE);
3521 static void perf_mmap_unmark_page(void *addr)
3523 struct page *page = vmalloc_to_page(addr);
3525 page->mapping = NULL;
3528 static void perf_buffer_free_work(struct work_struct *work)
3530 struct perf_buffer *buffer;
3534 buffer = container_of(work, struct perf_buffer, work);
3535 nr = 1 << page_order(buffer);
3537 base = buffer->user_page;
3538 for (i = 0; i < nr + 1; i++)
3539 perf_mmap_unmark_page(base + (i * PAGE_SIZE));
3545 static void perf_buffer_free(struct perf_buffer *buffer)
3547 schedule_work(&buffer->work);
3550 static struct perf_buffer *
3551 perf_buffer_alloc(int nr_pages, long watermark, int cpu, int flags)
3553 struct perf_buffer *buffer;
3557 size = sizeof(struct perf_buffer);
3558 size += sizeof(void *);
3560 buffer = kzalloc(size, GFP_KERNEL);
3564 INIT_WORK(&buffer->work, perf_buffer_free_work);
3566 all_buf = vmalloc_user((nr_pages + 1) * PAGE_SIZE);
3570 buffer->user_page = all_buf;
3571 buffer->data_pages[0] = all_buf + PAGE_SIZE;
3572 buffer->page_order = ilog2(nr_pages);
3573 buffer->nr_pages = 1;
3575 perf_buffer_init(buffer, watermark, flags);
3588 static unsigned long perf_data_size(struct perf_buffer *buffer)
3590 return buffer->nr_pages << (PAGE_SHIFT + page_order(buffer));
3593 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
3595 struct perf_event *event = vma->vm_file->private_data;
3596 struct perf_buffer *buffer;
3597 int ret = VM_FAULT_SIGBUS;
3599 if (vmf->flags & FAULT_FLAG_MKWRITE) {
3600 if (vmf->pgoff == 0)
3606 buffer = rcu_dereference(event->buffer);
3610 if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE))
3613 vmf->page = perf_mmap_to_page(buffer, vmf->pgoff);
3617 get_page(vmf->page);
3618 vmf->page->mapping = vma->vm_file->f_mapping;
3619 vmf->page->index = vmf->pgoff;
3628 static void perf_buffer_free_rcu(struct rcu_head *rcu_head)
3630 struct perf_buffer *buffer;
3632 buffer = container_of(rcu_head, struct perf_buffer, rcu_head);
3633 perf_buffer_free(buffer);
3636 static struct perf_buffer *perf_buffer_get(struct perf_event *event)
3638 struct perf_buffer *buffer;
3641 buffer = rcu_dereference(event->buffer);
3643 if (!atomic_inc_not_zero(&buffer->refcount))
3651 static void perf_buffer_put(struct perf_buffer *buffer)
3653 if (!atomic_dec_and_test(&buffer->refcount))
3656 call_rcu(&buffer->rcu_head, perf_buffer_free_rcu);
3659 static void perf_mmap_open(struct vm_area_struct *vma)
3661 struct perf_event *event = vma->vm_file->private_data;
3663 atomic_inc(&event->mmap_count);
3666 static void perf_mmap_close(struct vm_area_struct *vma)
3668 struct perf_event *event = vma->vm_file->private_data;
3670 if (atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex)) {
3671 unsigned long size = perf_data_size(event->buffer);
3672 struct user_struct *user = event->mmap_user;
3673 struct perf_buffer *buffer = event->buffer;
3675 atomic_long_sub((size >> PAGE_SHIFT) + 1, &user->locked_vm);
3676 vma->vm_mm->locked_vm -= event->mmap_locked;
3677 rcu_assign_pointer(event->buffer, NULL);
3678 mutex_unlock(&event->mmap_mutex);
3680 perf_buffer_put(buffer);
3685 static const struct vm_operations_struct perf_mmap_vmops = {
3686 .open = perf_mmap_open,
3687 .close = perf_mmap_close,
3688 .fault = perf_mmap_fault,
3689 .page_mkwrite = perf_mmap_fault,
3692 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
3694 struct perf_event *event = file->private_data;
3695 unsigned long user_locked, user_lock_limit;
3696 struct user_struct *user = current_user();
3697 unsigned long locked, lock_limit;
3698 struct perf_buffer *buffer;
3699 unsigned long vma_size;
3700 unsigned long nr_pages;
3701 long user_extra, extra;
3702 int ret = 0, flags = 0;
3705 * Don't allow mmap() of inherited per-task counters. This would
3706 * create a performance issue due to all children writing to the
3709 if (event->cpu == -1 && event->attr.inherit)
3712 if (!(vma->vm_flags & VM_SHARED))
3715 vma_size = vma->vm_end - vma->vm_start;
3716 nr_pages = (vma_size / PAGE_SIZE) - 1;
3719 * If we have buffer pages ensure they're a power-of-two number, so we
3720 * can do bitmasks instead of modulo.
3722 if (nr_pages != 0 && !is_power_of_2(nr_pages))
3725 if (vma_size != PAGE_SIZE * (1 + nr_pages))
3728 if (vma->vm_pgoff != 0)
3731 WARN_ON_ONCE(event->ctx->parent_ctx);
3732 mutex_lock(&event->mmap_mutex);
3733 if (event->buffer) {
3734 if (event->buffer->nr_pages == nr_pages)
3735 atomic_inc(&event->buffer->refcount);
3741 user_extra = nr_pages + 1;
3742 user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10);
3745 * Increase the limit linearly with more CPUs:
3747 user_lock_limit *= num_online_cpus();
3749 user_locked = atomic_long_read(&user->locked_vm) + user_extra;
3752 if (user_locked > user_lock_limit)
3753 extra = user_locked - user_lock_limit;
3755 lock_limit = rlimit(RLIMIT_MEMLOCK);
3756 lock_limit >>= PAGE_SHIFT;
3757 locked = vma->vm_mm->locked_vm + extra;
3759 if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() &&
3760 !capable(CAP_IPC_LOCK)) {
3765 WARN_ON(event->buffer);
3767 if (vma->vm_flags & VM_WRITE)
3768 flags |= PERF_BUFFER_WRITABLE;
3770 buffer = perf_buffer_alloc(nr_pages, event->attr.wakeup_watermark,
3776 rcu_assign_pointer(event->buffer, buffer);
3778 atomic_long_add(user_extra, &user->locked_vm);
3779 event->mmap_locked = extra;
3780 event->mmap_user = get_current_user();
3781 vma->vm_mm->locked_vm += event->mmap_locked;
3785 atomic_inc(&event->mmap_count);
3786 mutex_unlock(&event->mmap_mutex);
3788 vma->vm_flags |= VM_RESERVED;
3789 vma->vm_ops = &perf_mmap_vmops;
3794 static int perf_fasync(int fd, struct file *filp, int on)
3796 struct inode *inode = filp->f_path.dentry->d_inode;
3797 struct perf_event *event = filp->private_data;
3800 mutex_lock(&inode->i_mutex);
3801 retval = fasync_helper(fd, filp, on, &event->fasync);
3802 mutex_unlock(&inode->i_mutex);
3810 static const struct file_operations perf_fops = {
3811 .llseek = no_llseek,
3812 .release = perf_release,
3815 .unlocked_ioctl = perf_ioctl,
3816 .compat_ioctl = perf_ioctl,
3818 .fasync = perf_fasync,
3824 * If there's data, ensure we set the poll() state and publish everything
3825 * to user-space before waking everybody up.
3828 void perf_event_wakeup(struct perf_event *event)
3830 wake_up_all(&event->waitq);
3832 if (event->pending_kill) {
3833 kill_fasync(&event->fasync, SIGIO, event->pending_kill);
3834 event->pending_kill = 0;
3838 static void perf_pending_event(struct irq_work *entry)
3840 struct perf_event *event = container_of(entry,
3841 struct perf_event, pending);
3843 if (event->pending_disable) {
3844 event->pending_disable = 0;
3845 __perf_event_disable(event);
3848 if (event->pending_wakeup) {
3849 event->pending_wakeup = 0;
3850 perf_event_wakeup(event);
3855 * We assume there is only KVM supporting the callbacks.
3856 * Later on, we might change it to a list if there is
3857 * another virtualization implementation supporting the callbacks.
3859 struct perf_guest_info_callbacks *perf_guest_cbs;
3861 int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
3863 perf_guest_cbs = cbs;
3866 EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks);
3868 int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
3870 perf_guest_cbs = NULL;
3873 EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks);
3878 static bool perf_output_space(struct perf_buffer *buffer, unsigned long tail,
3879 unsigned long offset, unsigned long head)
3883 if (!buffer->writable)
3886 mask = perf_data_size(buffer) - 1;
3888 offset = (offset - tail) & mask;
3889 head = (head - tail) & mask;
3891 if ((int)(head - offset) < 0)
3897 static void perf_output_wakeup(struct perf_output_handle *handle)
3899 atomic_set(&handle->buffer->poll, POLL_IN);
3902 handle->event->pending_wakeup = 1;
3903 irq_work_queue(&handle->event->pending);
3905 perf_event_wakeup(handle->event);
3909 * We need to ensure a later event_id doesn't publish a head when a former
3910 * event isn't done writing. However since we need to deal with NMIs we
3911 * cannot fully serialize things.
3913 * We only publish the head (and generate a wakeup) when the outer-most
3916 static void perf_output_get_handle(struct perf_output_handle *handle)
3918 struct perf_buffer *buffer = handle->buffer;
3921 local_inc(&buffer->nest);
3922 handle->wakeup = local_read(&buffer->wakeup);
3925 static void perf_output_put_handle(struct perf_output_handle *handle)
3927 struct perf_buffer *buffer = handle->buffer;
3931 head = local_read(&buffer->head);
3934 * IRQ/NMI can happen here, which means we can miss a head update.
3937 if (!local_dec_and_test(&buffer->nest))
3941 * Publish the known good head. Rely on the full barrier implied
3942 * by atomic_dec_and_test() order the buffer->head read and this
3945 buffer->user_page->data_head = head;
3948 * Now check if we missed an update, rely on the (compiler)
3949 * barrier in atomic_dec_and_test() to re-read buffer->head.
3951 if (unlikely(head != local_read(&buffer->head))) {
3952 local_inc(&buffer->nest);
3956 if (handle->wakeup != local_read(&buffer->wakeup))
3957 perf_output_wakeup(handle);
3963 __always_inline void perf_output_copy(struct perf_output_handle *handle,
3964 const void *buf, unsigned int len)
3967 unsigned long size = min_t(unsigned long, handle->size, len);
3969 memcpy(handle->addr, buf, size);
3972 handle->addr += size;
3974 handle->size -= size;
3975 if (!handle->size) {
3976 struct perf_buffer *buffer = handle->buffer;
3979 handle->page &= buffer->nr_pages - 1;
3980 handle->addr = buffer->data_pages[handle->page];
3981 handle->size = PAGE_SIZE << page_order(buffer);
3986 static void __perf_event_header__init_id(struct perf_event_header *header,
3987 struct perf_sample_data *data,
3988 struct perf_event *event)
3990 u64 sample_type = event->attr.sample_type;
3992 data->type = sample_type;
3993 header->size += event->id_header_size;
3995 if (sample_type & PERF_SAMPLE_TID) {
3996 /* namespace issues */
3997 data->tid_entry.pid = perf_event_pid(event, current);
3998 data->tid_entry.tid = perf_event_tid(event, current);
4001 if (sample_type & PERF_SAMPLE_TIME)
4002 data->time = perf_clock();
4004 if (sample_type & PERF_SAMPLE_ID)
4005 data->id = primary_event_id(event);
4007 if (sample_type & PERF_SAMPLE_STREAM_ID)
4008 data->stream_id = event->id;
4010 if (sample_type & PERF_SAMPLE_CPU) {
4011 data->cpu_entry.cpu = raw_smp_processor_id();
4012 data->cpu_entry.reserved = 0;
4016 static void perf_event_header__init_id(struct perf_event_header *header,
4017 struct perf_sample_data *data,
4018 struct perf_event *event)
4020 if (event->attr.sample_id_all)
4021 __perf_event_header__init_id(header, data, event);
4024 static void __perf_event__output_id_sample(struct perf_output_handle *handle,
4025 struct perf_sample_data *data)
4027 u64 sample_type = data->type;
4029 if (sample_type & PERF_SAMPLE_TID)
4030 perf_output_put(handle, data->tid_entry);
4032 if (sample_type & PERF_SAMPLE_TIME)
4033 perf_output_put(handle, data->time);
4035 if (sample_type & PERF_SAMPLE_ID)
4036 perf_output_put(handle, data->id);
4038 if (sample_type & PERF_SAMPLE_STREAM_ID)
4039 perf_output_put(handle, data->stream_id);
4041 if (sample_type & PERF_SAMPLE_CPU)
4042 perf_output_put(handle, data->cpu_entry);
4045 static void perf_event__output_id_sample(struct perf_event *event,
4046 struct perf_output_handle *handle,
4047 struct perf_sample_data *sample)
4049 if (event->attr.sample_id_all)
4050 __perf_event__output_id_sample(handle, sample);
4053 int perf_output_begin(struct perf_output_handle *handle,
4054 struct perf_event *event, unsigned int size,
4055 int nmi, int sample)
4057 struct perf_buffer *buffer;
4058 unsigned long tail, offset, head;
4060 struct perf_sample_data sample_data;
4062 struct perf_event_header header;
4069 * For inherited events we send all the output towards the parent.
4072 event = event->parent;
4074 buffer = rcu_dereference(event->buffer);
4078 handle->buffer = buffer;
4079 handle->event = event;
4081 handle->sample = sample;
4083 if (!buffer->nr_pages)
4086 have_lost = local_read(&buffer->lost);
4088 lost_event.header.size = sizeof(lost_event);
4089 perf_event_header__init_id(&lost_event.header, &sample_data,
4091 size += lost_event.header.size;
4094 perf_output_get_handle(handle);
4098 * Userspace could choose to issue a mb() before updating the
4099 * tail pointer. So that all reads will be completed before the
4102 tail = ACCESS_ONCE(buffer->user_page->data_tail);
4104 offset = head = local_read(&buffer->head);
4106 if (unlikely(!perf_output_space(buffer, tail, offset, head)))
4108 } while (local_cmpxchg(&buffer->head, offset, head) != offset);
4110 if (head - local_read(&buffer->wakeup) > buffer->watermark)
4111 local_add(buffer->watermark, &buffer->wakeup);
4113 handle->page = offset >> (PAGE_SHIFT + page_order(buffer));
4114 handle->page &= buffer->nr_pages - 1;
4115 handle->size = offset & ((PAGE_SIZE << page_order(buffer)) - 1);
4116 handle->addr = buffer->data_pages[handle->page];
4117 handle->addr += handle->size;
4118 handle->size = (PAGE_SIZE << page_order(buffer)) - handle->size;
4121 lost_event.header.type = PERF_RECORD_LOST;
4122 lost_event.header.misc = 0;
4123 lost_event.id = event->id;
4124 lost_event.lost = local_xchg(&buffer->lost, 0);
4126 perf_output_put(handle, lost_event);
4127 perf_event__output_id_sample(event, handle, &sample_data);
4133 local_inc(&buffer->lost);
4134 perf_output_put_handle(handle);
4141 void perf_output_end(struct perf_output_handle *handle)
4143 struct perf_event *event = handle->event;
4144 struct perf_buffer *buffer = handle->buffer;
4146 int wakeup_events = event->attr.wakeup_events;
4148 if (handle->sample && wakeup_events) {
4149 int events = local_inc_return(&buffer->events);
4150 if (events >= wakeup_events) {
4151 local_sub(wakeup_events, &buffer->events);
4152 local_inc(&buffer->wakeup);
4156 perf_output_put_handle(handle);
4160 static void perf_output_read_one(struct perf_output_handle *handle,
4161 struct perf_event *event,
4162 u64 enabled, u64 running)
4164 u64 read_format = event->attr.read_format;
4168 values[n++] = perf_event_count(event);
4169 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
4170 values[n++] = enabled +
4171 atomic64_read(&event->child_total_time_enabled);
4173 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
4174 values[n++] = running +
4175 atomic64_read(&event->child_total_time_running);
4177 if (read_format & PERF_FORMAT_ID)
4178 values[n++] = primary_event_id(event);
4180 perf_output_copy(handle, values, n * sizeof(u64));
4184 * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
4186 static void perf_output_read_group(struct perf_output_handle *handle,
4187 struct perf_event *event,
4188 u64 enabled, u64 running)
4190 struct perf_event *leader = event->group_leader, *sub;
4191 u64 read_format = event->attr.read_format;
4195 values[n++] = 1 + leader->nr_siblings;
4197 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
4198 values[n++] = enabled;
4200 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
4201 values[n++] = running;
4203 if (leader != event)
4204 leader->pmu->read(leader);
4206 values[n++] = perf_event_count(leader);
4207 if (read_format & PERF_FORMAT_ID)
4208 values[n++] = primary_event_id(leader);
4210 perf_output_copy(handle, values, n * sizeof(u64));
4212 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
4216 sub->pmu->read(sub);
4218 values[n++] = perf_event_count(sub);
4219 if (read_format & PERF_FORMAT_ID)
4220 values[n++] = primary_event_id(sub);
4222 perf_output_copy(handle, values, n * sizeof(u64));
4226 #define PERF_FORMAT_TOTAL_TIMES (PERF_FORMAT_TOTAL_TIME_ENABLED|\
4227 PERF_FORMAT_TOTAL_TIME_RUNNING)
4229 static void perf_output_read(struct perf_output_handle *handle,
4230 struct perf_event *event)
4232 u64 enabled = 0, running = 0, now, ctx_time;
4233 u64 read_format = event->attr.read_format;
4236 * compute total_time_enabled, total_time_running
4237 * based on snapshot values taken when the event
4238 * was last scheduled in.
4240 * we cannot simply called update_context_time()
4241 * because of locking issue as we are called in
4244 if (read_format & PERF_FORMAT_TOTAL_TIMES) {
4246 ctx_time = event->shadow_ctx_time + now;
4247 enabled = ctx_time - event->tstamp_enabled;
4248 running = ctx_time - event->tstamp_running;
4251 if (event->attr.read_format & PERF_FORMAT_GROUP)
4252 perf_output_read_group(handle, event, enabled, running);
4254 perf_output_read_one(handle, event, enabled, running);
4257 void perf_output_sample(struct perf_output_handle *handle,
4258 struct perf_event_header *header,
4259 struct perf_sample_data *data,
4260 struct perf_event *event)
4262 u64 sample_type = data->type;
4264 perf_output_put(handle, *header);
4266 if (sample_type & PERF_SAMPLE_IP)
4267 perf_output_put(handle, data->ip);
4269 if (sample_type & PERF_SAMPLE_TID)
4270 perf_output_put(handle, data->tid_entry);
4272 if (sample_type & PERF_SAMPLE_TIME)
4273 perf_output_put(handle, data->time);
4275 if (sample_type & PERF_SAMPLE_ADDR)
4276 perf_output_put(handle, data->addr);
4278 if (sample_type & PERF_SAMPLE_ID)
4279 perf_output_put(handle, data->id);
4281 if (sample_type & PERF_SAMPLE_STREAM_ID)
4282 perf_output_put(handle, data->stream_id);
4284 if (sample_type & PERF_SAMPLE_CPU)
4285 perf_output_put(handle, data->cpu_entry);
4287 if (sample_type & PERF_SAMPLE_PERIOD)
4288 perf_output_put(handle, data->period);
4290 if (sample_type & PERF_SAMPLE_READ)
4291 perf_output_read(handle, event);
4293 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
4294 if (data->callchain) {
4297 if (data->callchain)
4298 size += data->callchain->nr;
4300 size *= sizeof(u64);
4302 perf_output_copy(handle, data->callchain, size);
4305 perf_output_put(handle, nr);
4309 if (sample_type & PERF_SAMPLE_RAW) {
4311 perf_output_put(handle, data->raw->size);
4312 perf_output_copy(handle, data->raw->data,
4319 .size = sizeof(u32),
4322 perf_output_put(handle, raw);
4327 void perf_prepare_sample(struct perf_event_header *header,
4328 struct perf_sample_data *data,
4329 struct perf_event *event,
4330 struct pt_regs *regs)
4332 u64 sample_type = event->attr.sample_type;
4334 header->type = PERF_RECORD_SAMPLE;
4335 header->size = sizeof(*header) + event->header_size;
4338 header->misc |= perf_misc_flags(regs);
4340 __perf_event_header__init_id(header, data, event);
4342 if (sample_type & PERF_SAMPLE_IP)
4343 data->ip = perf_instruction_pointer(regs);
4345 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
4348 data->callchain = perf_callchain(regs);
4350 if (data->callchain)
4351 size += data->callchain->nr;
4353 header->size += size * sizeof(u64);
4356 if (sample_type & PERF_SAMPLE_RAW) {
4357 int size = sizeof(u32);
4360 size += data->raw->size;
4362 size += sizeof(u32);
4364 WARN_ON_ONCE(size & (sizeof(u64)-1));
4365 header->size += size;
4369 static void perf_event_output(struct perf_event *event, int nmi,
4370 struct perf_sample_data *data,
4371 struct pt_regs *regs)
4373 struct perf_output_handle handle;
4374 struct perf_event_header header;
4376 /* protect the callchain buffers */
4379 perf_prepare_sample(&header, data, event, regs);
4381 if (perf_output_begin(&handle, event, header.size, nmi, 1))
4384 perf_output_sample(&handle, &header, data, event);
4386 perf_output_end(&handle);
4396 struct perf_read_event {
4397 struct perf_event_header header;
4404 perf_event_read_event(struct perf_event *event,
4405 struct task_struct *task)
4407 struct perf_output_handle handle;
4408 struct perf_sample_data sample;
4409 struct perf_read_event read_event = {
4411 .type = PERF_RECORD_READ,
4413 .size = sizeof(read_event) + event->read_size,
4415 .pid = perf_event_pid(event, task),
4416 .tid = perf_event_tid(event, task),
4420 perf_event_header__init_id(&read_event.header, &sample, event);
4421 ret = perf_output_begin(&handle, event, read_event.header.size, 0, 0);
4425 perf_output_put(&handle, read_event);
4426 perf_output_read(&handle, event);
4427 perf_event__output_id_sample(event, &handle, &sample);
4429 perf_output_end(&handle);
4433 * task tracking -- fork/exit
4435 * enabled by: attr.comm | attr.mmap | attr.mmap_data | attr.task
4438 struct perf_task_event {
4439 struct task_struct *task;
4440 struct perf_event_context *task_ctx;
4443 struct perf_event_header header;
4453 static void perf_event_task_output(struct perf_event *event,
4454 struct perf_task_event *task_event)
4456 struct perf_output_handle handle;
4457 struct perf_sample_data sample;
4458 struct task_struct *task = task_event->task;
4459 int ret, size = task_event->event_id.header.size;
4461 perf_event_header__init_id(&task_event->event_id.header, &sample, event);
4463 ret = perf_output_begin(&handle, event,
4464 task_event->event_id.header.size, 0, 0);
4468 task_event->event_id.pid = perf_event_pid(event, task);
4469 task_event->event_id.ppid = perf_event_pid(event, current);
4471 task_event->event_id.tid = perf_event_tid(event, task);
4472 task_event->event_id.ptid = perf_event_tid(event, current);
4474 perf_output_put(&handle, task_event->event_id);
4476 perf_event__output_id_sample(event, &handle, &sample);
4478 perf_output_end(&handle);
4480 task_event->event_id.header.size = size;
4483 static int perf_event_task_match(struct perf_event *event)
4485 if (event->state < PERF_EVENT_STATE_INACTIVE)
4488 if (!event_filter_match(event))
4491 if (event->attr.comm || event->attr.mmap ||
4492 event->attr.mmap_data || event->attr.task)
4498 static void perf_event_task_ctx(struct perf_event_context *ctx,
4499 struct perf_task_event *task_event)
4501 struct perf_event *event;
4503 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
4504 if (perf_event_task_match(event))
4505 perf_event_task_output(event, task_event);
4509 static void perf_event_task_event(struct perf_task_event *task_event)
4511 struct perf_cpu_context *cpuctx;
4512 struct perf_event_context *ctx;
4517 list_for_each_entry_rcu(pmu, &pmus, entry) {
4518 cpuctx = get_cpu_ptr(pmu->pmu_cpu_context);
4519 if (cpuctx->active_pmu != pmu)
4521 perf_event_task_ctx(&cpuctx->ctx, task_event);
4523 ctx = task_event->task_ctx;
4525 ctxn = pmu->task_ctx_nr;
4528 ctx = rcu_dereference(current->perf_event_ctxp[ctxn]);
4531 perf_event_task_ctx(ctx, task_event);
4533 put_cpu_ptr(pmu->pmu_cpu_context);
4538 static void perf_event_task(struct task_struct *task,
4539 struct perf_event_context *task_ctx,
4542 struct perf_task_event task_event;
4544 if (!atomic_read(&nr_comm_events) &&
4545 !atomic_read(&nr_mmap_events) &&
4546 !atomic_read(&nr_task_events))
4549 task_event = (struct perf_task_event){
4551 .task_ctx = task_ctx,
4554 .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT,
4556 .size = sizeof(task_event.event_id),
4562 .time = perf_clock(),
4566 perf_event_task_event(&task_event);
4569 void perf_event_fork(struct task_struct *task)
4571 perf_event_task(task, NULL, 1);
4578 struct perf_comm_event {
4579 struct task_struct *task;
4584 struct perf_event_header header;
4591 static void perf_event_comm_output(struct perf_event *event,
4592 struct perf_comm_event *comm_event)
4594 struct perf_output_handle handle;
4595 struct perf_sample_data sample;
4596 int size = comm_event->event_id.header.size;
4599 perf_event_header__init_id(&comm_event->event_id.header, &sample, event);
4600 ret = perf_output_begin(&handle, event,
4601 comm_event->event_id.header.size, 0, 0);
4606 comm_event->event_id.pid = perf_event_pid(event, comm_event->task);
4607 comm_event->event_id.tid = perf_event_tid(event, comm_event->task);
4609 perf_output_put(&handle, comm_event->event_id);
4610 perf_output_copy(&handle, comm_event->comm,
4611 comm_event->comm_size);
4613 perf_event__output_id_sample(event, &handle, &sample);
4615 perf_output_end(&handle);
4617 comm_event->event_id.header.size = size;
4620 static int perf_event_comm_match(struct perf_event *event)
4622 if (event->state < PERF_EVENT_STATE_INACTIVE)
4625 if (!event_filter_match(event))
4628 if (event->attr.comm)
4634 static void perf_event_comm_ctx(struct perf_event_context *ctx,
4635 struct perf_comm_event *comm_event)
4637 struct perf_event *event;
4639 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
4640 if (perf_event_comm_match(event))
4641 perf_event_comm_output(event, comm_event);
4645 static void perf_event_comm_event(struct perf_comm_event *comm_event)
4647 struct perf_cpu_context *cpuctx;
4648 struct perf_event_context *ctx;
4649 char comm[TASK_COMM_LEN];
4654 memset(comm, 0, sizeof(comm));
4655 strlcpy(comm, comm_event->task->comm, sizeof(comm));
4656 size = ALIGN(strlen(comm)+1, sizeof(u64));
4658 comm_event->comm = comm;
4659 comm_event->comm_size = size;
4661 comm_event->event_id.header.size = sizeof(comm_event->event_id) + size;
4663 list_for_each_entry_rcu(pmu, &pmus, entry) {
4664 cpuctx = get_cpu_ptr(pmu->pmu_cpu_context);
4665 if (cpuctx->active_pmu != pmu)
4667 perf_event_comm_ctx(&cpuctx->ctx, comm_event);
4669 ctxn = pmu->task_ctx_nr;
4673 ctx = rcu_dereference(current->perf_event_ctxp[ctxn]);
4675 perf_event_comm_ctx(ctx, comm_event);
4677 put_cpu_ptr(pmu->pmu_cpu_context);
4682 void perf_event_comm(struct task_struct *task)
4684 struct perf_comm_event comm_event;
4685 struct perf_event_context *ctx;
4688 for_each_task_context_nr(ctxn) {
4689 ctx = task->perf_event_ctxp[ctxn];
4693 perf_event_enable_on_exec(ctx);
4696 if (!atomic_read(&nr_comm_events))
4699 comm_event = (struct perf_comm_event){
4705 .type = PERF_RECORD_COMM,
4714 perf_event_comm_event(&comm_event);
4721 struct perf_mmap_event {
4722 struct vm_area_struct *vma;
4724 const char *file_name;
4728 struct perf_event_header header;
4738 static void perf_event_mmap_output(struct perf_event *event,
4739 struct perf_mmap_event *mmap_event)
4741 struct perf_output_handle handle;
4742 struct perf_sample_data sample;
4743 int size = mmap_event->event_id.header.size;
4746 perf_event_header__init_id(&mmap_event->event_id.header, &sample, event);
4747 ret = perf_output_begin(&handle, event,
4748 mmap_event->event_id.header.size, 0, 0);
4752 mmap_event->event_id.pid = perf_event_pid(event, current);
4753 mmap_event->event_id.tid = perf_event_tid(event, current);
4755 perf_output_put(&handle, mmap_event->event_id);
4756 perf_output_copy(&handle, mmap_event->file_name,
4757 mmap_event->file_size);
4759 perf_event__output_id_sample(event, &handle, &sample);
4761 perf_output_end(&handle);
4763 mmap_event->event_id.header.size = size;
4766 static int perf_event_mmap_match(struct perf_event *event,
4767 struct perf_mmap_event *mmap_event,
4770 if (event->state < PERF_EVENT_STATE_INACTIVE)
4773 if (!event_filter_match(event))
4776 if ((!executable && event->attr.mmap_data) ||
4777 (executable && event->attr.mmap))
4783 static void perf_event_mmap_ctx(struct perf_event_context *ctx,
4784 struct perf_mmap_event *mmap_event,
4787 struct perf_event *event;
4789 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
4790 if (perf_event_mmap_match(event, mmap_event, executable))
4791 perf_event_mmap_output(event, mmap_event);
4795 static void perf_event_mmap_event(struct perf_mmap_event *mmap_event)
4797 struct perf_cpu_context *cpuctx;
4798 struct perf_event_context *ctx;
4799 struct vm_area_struct *vma = mmap_event->vma;
4800 struct file *file = vma->vm_file;
4808 memset(tmp, 0, sizeof(tmp));
4812 * d_path works from the end of the buffer backwards, so we
4813 * need to add enough zero bytes after the string to handle
4814 * the 64bit alignment we do later.
4816 buf = kzalloc(PATH_MAX + sizeof(u64), GFP_KERNEL);
4818 name = strncpy(tmp, "//enomem", sizeof(tmp));
4821 name = d_path(&file->f_path, buf, PATH_MAX);
4823 name = strncpy(tmp, "//toolong", sizeof(tmp));
4827 if (arch_vma_name(mmap_event->vma)) {
4828 name = strncpy(tmp, arch_vma_name(mmap_event->vma),
4834 name = strncpy(tmp, "[vdso]", sizeof(tmp));
4836 } else if (vma->vm_start <= vma->vm_mm->start_brk &&
4837 vma->vm_end >= vma->vm_mm->brk) {
4838 name = strncpy(tmp, "[heap]", sizeof(tmp));
4840 } else if (vma->vm_start <= vma->vm_mm->start_stack &&
4841 vma->vm_end >= vma->vm_mm->start_stack) {
4842 name = strncpy(tmp, "[stack]", sizeof(tmp));
4846 name = strncpy(tmp, "//anon", sizeof(tmp));
4851 size = ALIGN(strlen(name)+1, sizeof(u64));
4853 mmap_event->file_name = name;
4854 mmap_event->file_size = size;
4856 mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size;
4859 list_for_each_entry_rcu(pmu, &pmus, entry) {
4860 cpuctx = get_cpu_ptr(pmu->pmu_cpu_context);
4861 if (cpuctx->active_pmu != pmu)
4863 perf_event_mmap_ctx(&cpuctx->ctx, mmap_event,
4864 vma->vm_flags & VM_EXEC);
4866 ctxn = pmu->task_ctx_nr;
4870 ctx = rcu_dereference(current->perf_event_ctxp[ctxn]);
4872 perf_event_mmap_ctx(ctx, mmap_event,
4873 vma->vm_flags & VM_EXEC);
4876 put_cpu_ptr(pmu->pmu_cpu_context);
4883 void perf_event_mmap(struct vm_area_struct *vma)
4885 struct perf_mmap_event mmap_event;
4887 if (!atomic_read(&nr_mmap_events))
4890 mmap_event = (struct perf_mmap_event){
4896 .type = PERF_RECORD_MMAP,
4897 .misc = PERF_RECORD_MISC_USER,
4902 .start = vma->vm_start,
4903 .len = vma->vm_end - vma->vm_start,
4904 .pgoff = (u64)vma->vm_pgoff << PAGE_SHIFT,
4908 perf_event_mmap_event(&mmap_event);
4912 * IRQ throttle logging
4915 static void perf_log_throttle(struct perf_event *event, int enable)
4917 struct perf_output_handle handle;
4918 struct perf_sample_data sample;
4922 struct perf_event_header header;
4926 } throttle_event = {
4928 .type = PERF_RECORD_THROTTLE,
4930 .size = sizeof(throttle_event),
4932 .time = perf_clock(),
4933 .id = primary_event_id(event),
4934 .stream_id = event->id,
4938 throttle_event.header.type = PERF_RECORD_UNTHROTTLE;
4940 perf_event_header__init_id(&throttle_event.header, &sample, event);
4942 ret = perf_output_begin(&handle, event,
4943 throttle_event.header.size, 1, 0);
4947 perf_output_put(&handle, throttle_event);
4948 perf_event__output_id_sample(event, &handle, &sample);
4949 perf_output_end(&handle);
4953 * Generic event overflow handling, sampling.
4956 static int __perf_event_overflow(struct perf_event *event, int nmi,
4957 int throttle, struct perf_sample_data *data,
4958 struct pt_regs *regs)
4960 int events = atomic_read(&event->event_limit);
4961 struct hw_perf_event *hwc = &event->hw;
4965 * Non-sampling counters might still use the PMI to fold short
4966 * hardware counters, ignore those.
4968 if (unlikely(!is_sampling_event(event)))
4971 if (unlikely(hwc->interrupts >= max_samples_per_tick)) {
4973 hwc->interrupts = MAX_INTERRUPTS;
4974 perf_log_throttle(event, 0);
4980 if (event->attr.freq) {
4981 u64 now = perf_clock();
4982 s64 delta = now - hwc->freq_time_stamp;
4984 hwc->freq_time_stamp = now;
4986 if (delta > 0 && delta < 2*TICK_NSEC)
4987 perf_adjust_period(event, delta, hwc->last_period);
4991 * XXX event_limit might not quite work as expected on inherited
4995 event->pending_kill = POLL_IN;
4996 if (events && atomic_dec_and_test(&event->event_limit)) {
4998 event->pending_kill = POLL_HUP;
5000 event->pending_disable = 1;
5001 irq_work_queue(&event->pending);
5003 perf_event_disable(event);
5006 if (event->overflow_handler)
5007 event->overflow_handler(event, nmi, data, regs);
5009 perf_event_output(event, nmi, data, regs);
5011 if (event->fasync && event->pending_kill) {
5013 event->pending_wakeup = 1;
5014 irq_work_queue(&event->pending);
5016 perf_event_wakeup(event);
5022 int perf_event_overflow(struct perf_event *event, int nmi,
5023 struct perf_sample_data *data,
5024 struct pt_regs *regs)
5026 return __perf_event_overflow(event, nmi, 1, data, regs);
5030 * Generic software event infrastructure
5033 struct swevent_htable {
5034 struct swevent_hlist *swevent_hlist;
5035 struct mutex hlist_mutex;
5038 /* Recursion avoidance in each contexts */
5039 int recursion[PERF_NR_CONTEXTS];
5042 static DEFINE_PER_CPU(struct swevent_htable, swevent_htable);
5045 * We directly increment event->count and keep a second value in
5046 * event->hw.period_left to count intervals. This period event
5047 * is kept in the range [-sample_period, 0] so that we can use the
5051 static u64 perf_swevent_set_period(struct perf_event *event)
5053 struct hw_perf_event *hwc = &event->hw;
5054 u64 period = hwc->last_period;
5058 hwc->last_period = hwc->sample_period;
5061 old = val = local64_read(&hwc->period_left);
5065 nr = div64_u64(period + val, period);
5066 offset = nr * period;
5068 if (local64_cmpxchg(&hwc->period_left, old, val) != old)
5074 static void perf_swevent_overflow(struct perf_event *event, u64 overflow,
5075 int nmi, struct perf_sample_data *data,
5076 struct pt_regs *regs)
5078 struct hw_perf_event *hwc = &event->hw;
5081 data->period = event->hw.last_period;
5083 overflow = perf_swevent_set_period(event);
5085 if (hwc->interrupts == MAX_INTERRUPTS)
5088 for (; overflow; overflow--) {
5089 if (__perf_event_overflow(event, nmi, throttle,
5092 * We inhibit the overflow from happening when
5093 * hwc->interrupts == MAX_INTERRUPTS.
5101 static void perf_swevent_event(struct perf_event *event, u64 nr,
5102 int nmi, struct perf_sample_data *data,
5103 struct pt_regs *regs)
5105 struct hw_perf_event *hwc = &event->hw;
5107 local64_add(nr, &event->count);
5112 if (!is_sampling_event(event))
5115 if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq)
5116 return perf_swevent_overflow(event, 1, nmi, data, regs);
5118 if (local64_add_negative(nr, &hwc->period_left))
5121 perf_swevent_overflow(event, 0, nmi, data, regs);
5124 static int perf_exclude_event(struct perf_event *event,
5125 struct pt_regs *regs)
5127 if (event->hw.state & PERF_HES_STOPPED)
5131 if (event->attr.exclude_user && user_mode(regs))
5134 if (event->attr.exclude_kernel && !user_mode(regs))
5141 static int perf_swevent_match(struct perf_event *event,
5142 enum perf_type_id type,
5144 struct perf_sample_data *data,
5145 struct pt_regs *regs)
5147 if (event->attr.type != type)
5150 if (event->attr.config != event_id)
5153 if (perf_exclude_event(event, regs))
5159 static inline u64 swevent_hash(u64 type, u32 event_id)
5161 u64 val = event_id | (type << 32);
5163 return hash_64(val, SWEVENT_HLIST_BITS);
5166 static inline struct hlist_head *
5167 __find_swevent_head(struct swevent_hlist *hlist, u64 type, u32 event_id)
5169 u64 hash = swevent_hash(type, event_id);
5171 return &hlist->heads[hash];
5174 /* For the read side: events when they trigger */
5175 static inline struct hlist_head *
5176 find_swevent_head_rcu(struct swevent_htable *swhash, u64 type, u32 event_id)
5178 struct swevent_hlist *hlist;
5180 hlist = rcu_dereference(swhash->swevent_hlist);
5184 return __find_swevent_head(hlist, type, event_id);
5187 /* For the event head insertion and removal in the hlist */
5188 static inline struct hlist_head *
5189 find_swevent_head(struct swevent_htable *swhash, struct perf_event *event)
5191 struct swevent_hlist *hlist;
5192 u32 event_id = event->attr.config;
5193 u64 type = event->attr.type;
5196 * Event scheduling is always serialized against hlist allocation
5197 * and release. Which makes the protected version suitable here.
5198 * The context lock guarantees that.
5200 hlist = rcu_dereference_protected(swhash->swevent_hlist,
5201 lockdep_is_held(&event->ctx->lock));
5205 return __find_swevent_head(hlist, type, event_id);
5208 static void do_perf_sw_event(enum perf_type_id type, u32 event_id,
5210 struct perf_sample_data *data,
5211 struct pt_regs *regs)
5213 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
5214 struct perf_event *event;
5215 struct hlist_node *node;
5216 struct hlist_head *head;
5219 head = find_swevent_head_rcu(swhash, type, event_id);
5223 hlist_for_each_entry_rcu(event, node, head, hlist_entry) {
5224 if (perf_swevent_match(event, type, event_id, data, regs))
5225 perf_swevent_event(event, nr, nmi, data, regs);
5231 int perf_swevent_get_recursion_context(void)
5233 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
5235 return get_recursion_context(swhash->recursion);
5237 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context);
5239 inline void perf_swevent_put_recursion_context(int rctx)
5241 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
5243 put_recursion_context(swhash->recursion, rctx);
5246 void __perf_sw_event(u32 event_id, u64 nr, int nmi,
5247 struct pt_regs *regs, u64 addr)
5249 struct perf_sample_data data;
5252 preempt_disable_notrace();
5253 rctx = perf_swevent_get_recursion_context();
5257 perf_sample_data_init(&data, addr);
5259 do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, nmi, &data, regs);
5261 perf_swevent_put_recursion_context(rctx);
5262 preempt_enable_notrace();
5265 static void perf_swevent_read(struct perf_event *event)
5269 static int perf_swevent_add(struct perf_event *event, int flags)
5271 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
5272 struct hw_perf_event *hwc = &event->hw;
5273 struct hlist_head *head;
5275 if (is_sampling_event(event)) {
5276 hwc->last_period = hwc->sample_period;
5277 perf_swevent_set_period(event);
5280 hwc->state = !(flags & PERF_EF_START);
5282 head = find_swevent_head(swhash, event);
5283 if (WARN_ON_ONCE(!head))
5286 hlist_add_head_rcu(&event->hlist_entry, head);
5291 static void perf_swevent_del(struct perf_event *event, int flags)
5293 hlist_del_rcu(&event->hlist_entry);
5296 static void perf_swevent_start(struct perf_event *event, int flags)
5298 event->hw.state = 0;
5301 static void perf_swevent_stop(struct perf_event *event, int flags)
5303 event->hw.state = PERF_HES_STOPPED;
5306 /* Deref the hlist from the update side */
5307 static inline struct swevent_hlist *
5308 swevent_hlist_deref(struct swevent_htable *swhash)
5310 return rcu_dereference_protected(swhash->swevent_hlist,
5311 lockdep_is_held(&swhash->hlist_mutex));
5314 static void swevent_hlist_release(struct swevent_htable *swhash)
5316 struct swevent_hlist *hlist = swevent_hlist_deref(swhash);
5321 rcu_assign_pointer(swhash->swevent_hlist, NULL);
5322 kfree_rcu(hlist, rcu_head);
5325 static void swevent_hlist_put_cpu(struct perf_event *event, int cpu)
5327 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
5329 mutex_lock(&swhash->hlist_mutex);
5331 if (!--swhash->hlist_refcount)
5332 swevent_hlist_release(swhash);
5334 mutex_unlock(&swhash->hlist_mutex);
5337 static void swevent_hlist_put(struct perf_event *event)
5341 if (event->cpu != -1) {
5342 swevent_hlist_put_cpu(event, event->cpu);
5346 for_each_possible_cpu(cpu)
5347 swevent_hlist_put_cpu(event, cpu);
5350 static int swevent_hlist_get_cpu(struct perf_event *event, int cpu)
5352 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
5355 mutex_lock(&swhash->hlist_mutex);
5357 if (!swevent_hlist_deref(swhash) && cpu_online(cpu)) {
5358 struct swevent_hlist *hlist;
5360 hlist = kzalloc(sizeof(*hlist), GFP_KERNEL);
5365 rcu_assign_pointer(swhash->swevent_hlist, hlist);
5367 swhash->hlist_refcount++;
5369 mutex_unlock(&swhash->hlist_mutex);
5374 static int swevent_hlist_get(struct perf_event *event)
5377 int cpu, failed_cpu;
5379 if (event->cpu != -1)
5380 return swevent_hlist_get_cpu(event, event->cpu);
5383 for_each_possible_cpu(cpu) {
5384 err = swevent_hlist_get_cpu(event, cpu);
5394 for_each_possible_cpu(cpu) {
5395 if (cpu == failed_cpu)
5397 swevent_hlist_put_cpu(event, cpu);
5404 struct jump_label_key perf_swevent_enabled[PERF_COUNT_SW_MAX];
5406 static void sw_perf_event_destroy(struct perf_event *event)
5408 u64 event_id = event->attr.config;
5410 WARN_ON(event->parent);
5412 jump_label_dec(&perf_swevent_enabled[event_id]);
5413 swevent_hlist_put(event);
5416 static int perf_swevent_init(struct perf_event *event)
5418 int event_id = event->attr.config;
5420 if (event->attr.type != PERF_TYPE_SOFTWARE)
5424 case PERF_COUNT_SW_CPU_CLOCK:
5425 case PERF_COUNT_SW_TASK_CLOCK:
5432 if (event_id >= PERF_COUNT_SW_MAX)
5435 if (!event->parent) {
5438 err = swevent_hlist_get(event);
5442 jump_label_inc(&perf_swevent_enabled[event_id]);
5443 event->destroy = sw_perf_event_destroy;
5449 static struct pmu perf_swevent = {
5450 .task_ctx_nr = perf_sw_context,
5452 .event_init = perf_swevent_init,
5453 .add = perf_swevent_add,
5454 .del = perf_swevent_del,
5455 .start = perf_swevent_start,
5456 .stop = perf_swevent_stop,
5457 .read = perf_swevent_read,
5460 #ifdef CONFIG_EVENT_TRACING
5462 static int perf_tp_filter_match(struct perf_event *event,
5463 struct perf_sample_data *data)
5465 void *record = data->raw->data;
5467 if (likely(!event->filter) || filter_match_preds(event->filter, record))
5472 static int perf_tp_event_match(struct perf_event *event,
5473 struct perf_sample_data *data,
5474 struct pt_regs *regs)
5476 if (event->hw.state & PERF_HES_STOPPED)
5479 * All tracepoints are from kernel-space.
5481 if (event->attr.exclude_kernel)
5484 if (!perf_tp_filter_match(event, data))
5490 void perf_tp_event(u64 addr, u64 count, void *record, int entry_size,
5491 struct pt_regs *regs, struct hlist_head *head, int rctx)
5493 struct perf_sample_data data;
5494 struct perf_event *event;
5495 struct hlist_node *node;
5497 struct perf_raw_record raw = {
5502 perf_sample_data_init(&data, addr);
5505 hlist_for_each_entry_rcu(event, node, head, hlist_entry) {
5506 if (perf_tp_event_match(event, &data, regs))
5507 perf_swevent_event(event, count, 1, &data, regs);
5510 perf_swevent_put_recursion_context(rctx);
5512 EXPORT_SYMBOL_GPL(perf_tp_event);
5514 static void tp_perf_event_destroy(struct perf_event *event)
5516 perf_trace_destroy(event);
5519 static int perf_tp_event_init(struct perf_event *event)
5523 if (event->attr.type != PERF_TYPE_TRACEPOINT)
5526 err = perf_trace_init(event);
5530 event->destroy = tp_perf_event_destroy;
5535 static struct pmu perf_tracepoint = {
5536 .task_ctx_nr = perf_sw_context,
5538 .event_init = perf_tp_event_init,
5539 .add = perf_trace_add,
5540 .del = perf_trace_del,
5541 .start = perf_swevent_start,
5542 .stop = perf_swevent_stop,
5543 .read = perf_swevent_read,
5546 static inline void perf_tp_register(void)
5548 perf_pmu_register(&perf_tracepoint, "tracepoint", PERF_TYPE_TRACEPOINT);
5551 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
5556 if (event->attr.type != PERF_TYPE_TRACEPOINT)
5559 filter_str = strndup_user(arg, PAGE_SIZE);
5560 if (IS_ERR(filter_str))
5561 return PTR_ERR(filter_str);
5563 ret = ftrace_profile_set_filter(event, event->attr.config, filter_str);
5569 static void perf_event_free_filter(struct perf_event *event)
5571 ftrace_profile_free_filter(event);
5576 static inline void perf_tp_register(void)
5580 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
5585 static void perf_event_free_filter(struct perf_event *event)
5589 #endif /* CONFIG_EVENT_TRACING */
5591 #ifdef CONFIG_HAVE_HW_BREAKPOINT
5592 void perf_bp_event(struct perf_event *bp, void *data)
5594 struct perf_sample_data sample;
5595 struct pt_regs *regs = data;
5597 perf_sample_data_init(&sample, bp->attr.bp_addr);
5599 if (!bp->hw.state && !perf_exclude_event(bp, regs))
5600 perf_swevent_event(bp, 1, 1, &sample, regs);
5605 * hrtimer based swevent callback
5608 static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer)
5610 enum hrtimer_restart ret = HRTIMER_RESTART;
5611 struct perf_sample_data data;
5612 struct pt_regs *regs;
5613 struct perf_event *event;
5616 event = container_of(hrtimer, struct perf_event, hw.hrtimer);
5618 if (event->state != PERF_EVENT_STATE_ACTIVE)
5619 return HRTIMER_NORESTART;
5621 event->pmu->read(event);
5623 perf_sample_data_init(&data, 0);
5624 data.period = event->hw.last_period;
5625 regs = get_irq_regs();
5627 if (regs && !perf_exclude_event(event, regs)) {
5628 if (!(event->attr.exclude_idle && current->pid == 0))
5629 if (perf_event_overflow(event, 0, &data, regs))
5630 ret = HRTIMER_NORESTART;
5633 period = max_t(u64, 10000, event->hw.sample_period);
5634 hrtimer_forward_now(hrtimer, ns_to_ktime(period));
5639 static void perf_swevent_start_hrtimer(struct perf_event *event)
5641 struct hw_perf_event *hwc = &event->hw;
5644 if (!is_sampling_event(event))
5647 period = local64_read(&hwc->period_left);
5652 local64_set(&hwc->period_left, 0);
5654 period = max_t(u64, 10000, hwc->sample_period);
5656 __hrtimer_start_range_ns(&hwc->hrtimer,
5657 ns_to_ktime(period), 0,
5658 HRTIMER_MODE_REL_PINNED, 0);
5661 static void perf_swevent_cancel_hrtimer(struct perf_event *event)
5663 struct hw_perf_event *hwc = &event->hw;
5665 if (is_sampling_event(event)) {
5666 ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer);
5667 local64_set(&hwc->period_left, ktime_to_ns(remaining));
5669 hrtimer_cancel(&hwc->hrtimer);
5673 static void perf_swevent_init_hrtimer(struct perf_event *event)
5675 struct hw_perf_event *hwc = &event->hw;
5677 if (!is_sampling_event(event))
5680 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
5681 hwc->hrtimer.function = perf_swevent_hrtimer;
5684 * Since hrtimers have a fixed rate, we can do a static freq->period
5685 * mapping and avoid the whole period adjust feedback stuff.
5687 if (event->attr.freq) {
5688 long freq = event->attr.sample_freq;
5690 event->attr.sample_period = NSEC_PER_SEC / freq;
5691 hwc->sample_period = event->attr.sample_period;
5692 local64_set(&hwc->period_left, hwc->sample_period);
5693 event->attr.freq = 0;
5698 * Software event: cpu wall time clock
5701 static void cpu_clock_event_update(struct perf_event *event)
5706 now = local_clock();
5707 prev = local64_xchg(&event->hw.prev_count, now);
5708 local64_add(now - prev, &event->count);
5711 static void cpu_clock_event_start(struct perf_event *event, int flags)
5713 local64_set(&event->hw.prev_count, local_clock());
5714 perf_swevent_start_hrtimer(event);
5717 static void cpu_clock_event_stop(struct perf_event *event, int flags)
5719 perf_swevent_cancel_hrtimer(event);
5720 cpu_clock_event_update(event);
5723 static int cpu_clock_event_add(struct perf_event *event, int flags)
5725 if (flags & PERF_EF_START)
5726 cpu_clock_event_start(event, flags);
5731 static void cpu_clock_event_del(struct perf_event *event, int flags)
5733 cpu_clock_event_stop(event, flags);
5736 static void cpu_clock_event_read(struct perf_event *event)
5738 cpu_clock_event_update(event);
5741 static int cpu_clock_event_init(struct perf_event *event)
5743 if (event->attr.type != PERF_TYPE_SOFTWARE)
5746 if (event->attr.config != PERF_COUNT_SW_CPU_CLOCK)
5749 perf_swevent_init_hrtimer(event);
5754 static struct pmu perf_cpu_clock = {
5755 .task_ctx_nr = perf_sw_context,
5757 .event_init = cpu_clock_event_init,
5758 .add = cpu_clock_event_add,
5759 .del = cpu_clock_event_del,
5760 .start = cpu_clock_event_start,
5761 .stop = cpu_clock_event_stop,
5762 .read = cpu_clock_event_read,
5766 * Software event: task time clock
5769 static void task_clock_event_update(struct perf_event *event, u64 now)
5774 prev = local64_xchg(&event->hw.prev_count, now);
5776 local64_add(delta, &event->count);
5779 static void task_clock_event_start(struct perf_event *event, int flags)
5781 local64_set(&event->hw.prev_count, event->ctx->time);
5782 perf_swevent_start_hrtimer(event);
5785 static void task_clock_event_stop(struct perf_event *event, int flags)
5787 perf_swevent_cancel_hrtimer(event);
5788 task_clock_event_update(event, event->ctx->time);
5791 static int task_clock_event_add(struct perf_event *event, int flags)
5793 if (flags & PERF_EF_START)
5794 task_clock_event_start(event, flags);
5799 static void task_clock_event_del(struct perf_event *event, int flags)
5801 task_clock_event_stop(event, PERF_EF_UPDATE);
5804 static void task_clock_event_read(struct perf_event *event)
5806 u64 now = perf_clock();
5807 u64 delta = now - event->ctx->timestamp;
5808 u64 time = event->ctx->time + delta;
5810 task_clock_event_update(event, time);
5813 static int task_clock_event_init(struct perf_event *event)
5815 if (event->attr.type != PERF_TYPE_SOFTWARE)
5818 if (event->attr.config != PERF_COUNT_SW_TASK_CLOCK)
5821 perf_swevent_init_hrtimer(event);
5826 static struct pmu perf_task_clock = {
5827 .task_ctx_nr = perf_sw_context,
5829 .event_init = task_clock_event_init,
5830 .add = task_clock_event_add,
5831 .del = task_clock_event_del,
5832 .start = task_clock_event_start,
5833 .stop = task_clock_event_stop,
5834 .read = task_clock_event_read,
5837 static void perf_pmu_nop_void(struct pmu *pmu)
5841 static int perf_pmu_nop_int(struct pmu *pmu)
5846 static void perf_pmu_start_txn(struct pmu *pmu)
5848 perf_pmu_disable(pmu);
5851 static int perf_pmu_commit_txn(struct pmu *pmu)
5853 perf_pmu_enable(pmu);
5857 static void perf_pmu_cancel_txn(struct pmu *pmu)
5859 perf_pmu_enable(pmu);
5863 * Ensures all contexts with the same task_ctx_nr have the same
5864 * pmu_cpu_context too.
5866 static void *find_pmu_context(int ctxn)
5873 list_for_each_entry(pmu, &pmus, entry) {
5874 if (pmu->task_ctx_nr == ctxn)
5875 return pmu->pmu_cpu_context;
5881 static void update_pmu_context(struct pmu *pmu, struct pmu *old_pmu)
5885 for_each_possible_cpu(cpu) {
5886 struct perf_cpu_context *cpuctx;
5888 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
5890 if (cpuctx->active_pmu == old_pmu)
5891 cpuctx->active_pmu = pmu;
5895 static void free_pmu_context(struct pmu *pmu)
5899 mutex_lock(&pmus_lock);
5901 * Like a real lame refcount.
5903 list_for_each_entry(i, &pmus, entry) {
5904 if (i->pmu_cpu_context == pmu->pmu_cpu_context) {
5905 update_pmu_context(i, pmu);
5910 free_percpu(pmu->pmu_cpu_context);
5912 mutex_unlock(&pmus_lock);
5914 static struct idr pmu_idr;
5917 type_show(struct device *dev, struct device_attribute *attr, char *page)
5919 struct pmu *pmu = dev_get_drvdata(dev);
5921 return snprintf(page, PAGE_SIZE-1, "%d\n", pmu->type);
5924 static struct device_attribute pmu_dev_attrs[] = {
5929 static int pmu_bus_running;
5930 static struct bus_type pmu_bus = {
5931 .name = "event_source",
5932 .dev_attrs = pmu_dev_attrs,
5935 static void pmu_dev_release(struct device *dev)
5940 static int pmu_dev_alloc(struct pmu *pmu)
5944 pmu->dev = kzalloc(sizeof(struct device), GFP_KERNEL);
5948 device_initialize(pmu->dev);
5949 ret = dev_set_name(pmu->dev, "%s", pmu->name);
5953 dev_set_drvdata(pmu->dev, pmu);
5954 pmu->dev->bus = &pmu_bus;
5955 pmu->dev->release = pmu_dev_release;
5956 ret = device_add(pmu->dev);
5964 put_device(pmu->dev);
5968 static struct lock_class_key cpuctx_mutex;
5969 static struct lock_class_key cpuctx_lock;
5971 int perf_pmu_register(struct pmu *pmu, char *name, int type)
5975 mutex_lock(&pmus_lock);
5977 pmu->pmu_disable_count = alloc_percpu(int);
5978 if (!pmu->pmu_disable_count)
5987 int err = idr_pre_get(&pmu_idr, GFP_KERNEL);
5991 err = idr_get_new_above(&pmu_idr, pmu, PERF_TYPE_MAX, &type);
5999 if (pmu_bus_running) {
6000 ret = pmu_dev_alloc(pmu);
6006 pmu->pmu_cpu_context = find_pmu_context(pmu->task_ctx_nr);
6007 if (pmu->pmu_cpu_context)
6008 goto got_cpu_context;
6010 pmu->pmu_cpu_context = alloc_percpu(struct perf_cpu_context);
6011 if (!pmu->pmu_cpu_context)
6014 for_each_possible_cpu(cpu) {
6015 struct perf_cpu_context *cpuctx;
6017 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
6018 __perf_event_init_context(&cpuctx->ctx);
6019 lockdep_set_class(&cpuctx->ctx.mutex, &cpuctx_mutex);
6020 lockdep_set_class(&cpuctx->ctx.lock, &cpuctx_lock);
6021 cpuctx->ctx.type = cpu_context;
6022 cpuctx->ctx.pmu = pmu;
6023 cpuctx->jiffies_interval = 1;
6024 INIT_LIST_HEAD(&cpuctx->rotation_list);
6025 cpuctx->active_pmu = pmu;
6029 if (!pmu->start_txn) {
6030 if (pmu->pmu_enable) {
6032 * If we have pmu_enable/pmu_disable calls, install
6033 * transaction stubs that use that to try and batch
6034 * hardware accesses.
6036 pmu->start_txn = perf_pmu_start_txn;
6037 pmu->commit_txn = perf_pmu_commit_txn;
6038 pmu->cancel_txn = perf_pmu_cancel_txn;
6040 pmu->start_txn = perf_pmu_nop_void;
6041 pmu->commit_txn = perf_pmu_nop_int;
6042 pmu->cancel_txn = perf_pmu_nop_void;
6046 if (!pmu->pmu_enable) {
6047 pmu->pmu_enable = perf_pmu_nop_void;
6048 pmu->pmu_disable = perf_pmu_nop_void;
6051 list_add_rcu(&pmu->entry, &pmus);
6054 mutex_unlock(&pmus_lock);
6059 device_del(pmu->dev);
6060 put_device(pmu->dev);
6063 if (pmu->type >= PERF_TYPE_MAX)
6064 idr_remove(&pmu_idr, pmu->type);
6067 free_percpu(pmu->pmu_disable_count);
6071 void perf_pmu_unregister(struct pmu *pmu)
6073 mutex_lock(&pmus_lock);
6074 list_del_rcu(&pmu->entry);
6075 mutex_unlock(&pmus_lock);
6078 * We dereference the pmu list under both SRCU and regular RCU, so
6079 * synchronize against both of those.
6081 synchronize_srcu(&pmus_srcu);
6084 free_percpu(pmu->pmu_disable_count);
6085 if (pmu->type >= PERF_TYPE_MAX)
6086 idr_remove(&pmu_idr, pmu->type);
6087 device_del(pmu->dev);
6088 put_device(pmu->dev);
6089 free_pmu_context(pmu);
6092 struct pmu *perf_init_event(struct perf_event *event)
6094 struct pmu *pmu = NULL;
6098 idx = srcu_read_lock(&pmus_srcu);
6101 pmu = idr_find(&pmu_idr, event->attr.type);
6104 ret = pmu->event_init(event);
6110 list_for_each_entry_rcu(pmu, &pmus, entry) {
6111 ret = pmu->event_init(event);
6115 if (ret != -ENOENT) {
6120 pmu = ERR_PTR(-ENOENT);
6122 srcu_read_unlock(&pmus_srcu, idx);
6128 * Allocate and initialize a event structure
6130 static struct perf_event *
6131 perf_event_alloc(struct perf_event_attr *attr, int cpu,
6132 struct task_struct *task,
6133 struct perf_event *group_leader,
6134 struct perf_event *parent_event,
6135 perf_overflow_handler_t overflow_handler)
6138 struct perf_event *event;
6139 struct hw_perf_event *hwc;
6142 if ((unsigned)cpu >= nr_cpu_ids) {
6143 if (!task || cpu != -1)
6144 return ERR_PTR(-EINVAL);
6147 event = kzalloc(sizeof(*event), GFP_KERNEL);
6149 return ERR_PTR(-ENOMEM);
6152 * Single events are their own group leaders, with an
6153 * empty sibling list:
6156 group_leader = event;
6158 mutex_init(&event->child_mutex);
6159 INIT_LIST_HEAD(&event->child_list);
6161 INIT_LIST_HEAD(&event->group_entry);
6162 INIT_LIST_HEAD(&event->event_entry);
6163 INIT_LIST_HEAD(&event->sibling_list);
6164 init_waitqueue_head(&event->waitq);
6165 init_irq_work(&event->pending, perf_pending_event);
6167 mutex_init(&event->mmap_mutex);
6170 event->attr = *attr;
6171 event->group_leader = group_leader;
6175 event->parent = parent_event;
6177 event->ns = get_pid_ns(current->nsproxy->pid_ns);
6178 event->id = atomic64_inc_return(&perf_event_id);
6180 event->state = PERF_EVENT_STATE_INACTIVE;
6183 event->attach_state = PERF_ATTACH_TASK;
6184 #ifdef CONFIG_HAVE_HW_BREAKPOINT
6186 * hw_breakpoint is a bit difficult here..
6188 if (attr->type == PERF_TYPE_BREAKPOINT)
6189 event->hw.bp_target = task;
6193 if (!overflow_handler && parent_event)
6194 overflow_handler = parent_event->overflow_handler;
6196 event->overflow_handler = overflow_handler;
6199 event->state = PERF_EVENT_STATE_OFF;
6204 hwc->sample_period = attr->sample_period;
6205 if (attr->freq && attr->sample_freq)
6206 hwc->sample_period = 1;
6207 hwc->last_period = hwc->sample_period;
6209 local64_set(&hwc->period_left, hwc->sample_period);
6212 * we currently do not support PERF_FORMAT_GROUP on inherited events
6214 if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP))
6217 pmu = perf_init_event(event);
6223 else if (IS_ERR(pmu))
6228 put_pid_ns(event->ns);
6230 return ERR_PTR(err);
6235 if (!event->parent) {
6236 if (event->attach_state & PERF_ATTACH_TASK)
6237 jump_label_inc(&perf_sched_events);
6238 if (event->attr.mmap || event->attr.mmap_data)
6239 atomic_inc(&nr_mmap_events);
6240 if (event->attr.comm)
6241 atomic_inc(&nr_comm_events);
6242 if (event->attr.task)
6243 atomic_inc(&nr_task_events);
6244 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) {
6245 err = get_callchain_buffers();
6248 return ERR_PTR(err);
6256 static int perf_copy_attr(struct perf_event_attr __user *uattr,
6257 struct perf_event_attr *attr)
6262 if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0))
6266 * zero the full structure, so that a short copy will be nice.
6268 memset(attr, 0, sizeof(*attr));
6270 ret = get_user(size, &uattr->size);
6274 if (size > PAGE_SIZE) /* silly large */
6277 if (!size) /* abi compat */
6278 size = PERF_ATTR_SIZE_VER0;
6280 if (size < PERF_ATTR_SIZE_VER0)
6284 * If we're handed a bigger struct than we know of,
6285 * ensure all the unknown bits are 0 - i.e. new
6286 * user-space does not rely on any kernel feature
6287 * extensions we dont know about yet.
6289 if (size > sizeof(*attr)) {
6290 unsigned char __user *addr;
6291 unsigned char __user *end;
6294 addr = (void __user *)uattr + sizeof(*attr);
6295 end = (void __user *)uattr + size;
6297 for (; addr < end; addr++) {
6298 ret = get_user(val, addr);
6304 size = sizeof(*attr);
6307 ret = copy_from_user(attr, uattr, size);
6312 * If the type exists, the corresponding creation will verify
6315 if (attr->type >= PERF_TYPE_MAX)
6318 if (attr->__reserved_1)
6321 if (attr->sample_type & ~(PERF_SAMPLE_MAX-1))
6324 if (attr->read_format & ~(PERF_FORMAT_MAX-1))
6331 put_user(sizeof(*attr), &uattr->size);
6337 perf_event_set_output(struct perf_event *event, struct perf_event *output_event)
6339 struct perf_buffer *buffer = NULL, *old_buffer = NULL;
6345 /* don't allow circular references */
6346 if (event == output_event)
6350 * Don't allow cross-cpu buffers
6352 if (output_event->cpu != event->cpu)
6356 * If its not a per-cpu buffer, it must be the same task.
6358 if (output_event->cpu == -1 && output_event->ctx != event->ctx)
6362 mutex_lock(&event->mmap_mutex);
6363 /* Can't redirect output if we've got an active mmap() */
6364 if (atomic_read(&event->mmap_count))
6368 /* get the buffer we want to redirect to */
6369 buffer = perf_buffer_get(output_event);
6374 old_buffer = event->buffer;
6375 rcu_assign_pointer(event->buffer, buffer);
6378 mutex_unlock(&event->mmap_mutex);
6381 perf_buffer_put(old_buffer);
6387 * sys_perf_event_open - open a performance event, associate it to a task/cpu
6389 * @attr_uptr: event_id type attributes for monitoring/sampling
6392 * @group_fd: group leader event fd
6394 SYSCALL_DEFINE5(perf_event_open,
6395 struct perf_event_attr __user *, attr_uptr,
6396 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
6398 struct perf_event *group_leader = NULL, *output_event = NULL;
6399 struct perf_event *event, *sibling;
6400 struct perf_event_attr attr;
6401 struct perf_event_context *ctx;
6402 struct file *event_file = NULL;
6403 struct file *group_file = NULL;
6404 struct task_struct *task = NULL;
6408 int fput_needed = 0;
6411 /* for future expandability... */
6412 if (flags & ~PERF_FLAG_ALL)
6415 err = perf_copy_attr(attr_uptr, &attr);
6419 if (!attr.exclude_kernel) {
6420 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
6425 if (attr.sample_freq > sysctl_perf_event_sample_rate)
6430 * In cgroup mode, the pid argument is used to pass the fd
6431 * opened to the cgroup directory in cgroupfs. The cpu argument
6432 * designates the cpu on which to monitor threads from that
6435 if ((flags & PERF_FLAG_PID_CGROUP) && (pid == -1 || cpu == -1))
6438 event_fd = get_unused_fd_flags(O_RDWR);
6442 if (group_fd != -1) {
6443 group_leader = perf_fget_light(group_fd, &fput_needed);
6444 if (IS_ERR(group_leader)) {
6445 err = PTR_ERR(group_leader);
6448 group_file = group_leader->filp;
6449 if (flags & PERF_FLAG_FD_OUTPUT)
6450 output_event = group_leader;
6451 if (flags & PERF_FLAG_FD_NO_GROUP)
6452 group_leader = NULL;
6455 if (pid != -1 && !(flags & PERF_FLAG_PID_CGROUP)) {
6456 task = find_lively_task_by_vpid(pid);
6458 err = PTR_ERR(task);
6463 event = perf_event_alloc(&attr, cpu, task, group_leader, NULL, NULL);
6464 if (IS_ERR(event)) {
6465 err = PTR_ERR(event);
6469 if (flags & PERF_FLAG_PID_CGROUP) {
6470 err = perf_cgroup_connect(pid, event, &attr, group_leader);
6475 * - that has cgroup constraint on event->cpu
6476 * - that may need work on context switch
6478 atomic_inc(&per_cpu(perf_cgroup_events, event->cpu));
6479 jump_label_inc(&perf_sched_events);
6483 * Special case software events and allow them to be part of
6484 * any hardware group.
6489 (is_software_event(event) != is_software_event(group_leader))) {
6490 if (is_software_event(event)) {
6492 * If event and group_leader are not both a software
6493 * event, and event is, then group leader is not.
6495 * Allow the addition of software events to !software
6496 * groups, this is safe because software events never
6499 pmu = group_leader->pmu;
6500 } else if (is_software_event(group_leader) &&
6501 (group_leader->group_flags & PERF_GROUP_SOFTWARE)) {
6503 * In case the group is a pure software group, and we
6504 * try to add a hardware event, move the whole group to
6505 * the hardware context.
6512 * Get the target context (task or percpu):
6514 ctx = find_get_context(pmu, task, cpu);
6521 put_task_struct(task);
6526 * Look up the group leader (we will attach this event to it):
6532 * Do not allow a recursive hierarchy (this new sibling
6533 * becoming part of another group-sibling):
6535 if (group_leader->group_leader != group_leader)
6538 * Do not allow to attach to a group in a different
6539 * task or CPU context:
6542 if (group_leader->ctx->type != ctx->type)
6545 if (group_leader->ctx != ctx)
6550 * Only a group leader can be exclusive or pinned
6552 if (attr.exclusive || attr.pinned)
6557 err = perf_event_set_output(event, output_event);
6562 event_file = anon_inode_getfile("[perf_event]", &perf_fops, event, O_RDWR);
6563 if (IS_ERR(event_file)) {
6564 err = PTR_ERR(event_file);
6569 struct perf_event_context *gctx = group_leader->ctx;
6571 mutex_lock(&gctx->mutex);
6572 perf_remove_from_context(group_leader);
6573 list_for_each_entry(sibling, &group_leader->sibling_list,
6575 perf_remove_from_context(sibling);
6578 mutex_unlock(&gctx->mutex);
6582 event->filp = event_file;
6583 WARN_ON_ONCE(ctx->parent_ctx);
6584 mutex_lock(&ctx->mutex);
6587 perf_install_in_context(ctx, group_leader, cpu);
6589 list_for_each_entry(sibling, &group_leader->sibling_list,
6591 perf_install_in_context(ctx, sibling, cpu);
6596 perf_install_in_context(ctx, event, cpu);
6598 perf_unpin_context(ctx);
6599 mutex_unlock(&ctx->mutex);
6601 event->owner = current;
6603 mutex_lock(¤t->perf_event_mutex);
6604 list_add_tail(&event->owner_entry, ¤t->perf_event_list);
6605 mutex_unlock(¤t->perf_event_mutex);
6608 * Precalculate sample_data sizes
6610 perf_event__header_size(event);
6611 perf_event__id_header_size(event);
6614 * Drop the reference on the group_event after placing the
6615 * new event on the sibling_list. This ensures destruction
6616 * of the group leader will find the pointer to itself in
6617 * perf_group_detach().
6619 fput_light(group_file, fput_needed);
6620 fd_install(event_fd, event_file);
6624 perf_unpin_context(ctx);
6630 put_task_struct(task);
6632 fput_light(group_file, fput_needed);
6634 put_unused_fd(event_fd);
6639 * perf_event_create_kernel_counter
6641 * @attr: attributes of the counter to create
6642 * @cpu: cpu in which the counter is bound
6643 * @task: task to profile (NULL for percpu)
6646 perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu,
6647 struct task_struct *task,
6648 perf_overflow_handler_t overflow_handler)
6650 struct perf_event_context *ctx;
6651 struct perf_event *event;
6655 * Get the target context (task or percpu):
6658 event = perf_event_alloc(attr, cpu, task, NULL, NULL, overflow_handler);
6659 if (IS_ERR(event)) {
6660 err = PTR_ERR(event);
6664 ctx = find_get_context(event->pmu, task, cpu);
6671 WARN_ON_ONCE(ctx->parent_ctx);
6672 mutex_lock(&ctx->mutex);
6673 perf_install_in_context(ctx, event, cpu);
6675 perf_unpin_context(ctx);
6676 mutex_unlock(&ctx->mutex);
6683 return ERR_PTR(err);
6685 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter);
6687 static void sync_child_event(struct perf_event *child_event,
6688 struct task_struct *child)
6690 struct perf_event *parent_event = child_event->parent;
6693 if (child_event->attr.inherit_stat)
6694 perf_event_read_event(child_event, child);
6696 child_val = perf_event_count(child_event);
6699 * Add back the child's count to the parent's count:
6701 atomic64_add(child_val, &parent_event->child_count);
6702 atomic64_add(child_event->total_time_enabled,
6703 &parent_event->child_total_time_enabled);
6704 atomic64_add(child_event->total_time_running,
6705 &parent_event->child_total_time_running);
6708 * Remove this event from the parent's list
6710 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
6711 mutex_lock(&parent_event->child_mutex);
6712 list_del_init(&child_event->child_list);
6713 mutex_unlock(&parent_event->child_mutex);
6716 * Release the parent event, if this was the last
6719 fput(parent_event->filp);
6723 __perf_event_exit_task(struct perf_event *child_event,
6724 struct perf_event_context *child_ctx,
6725 struct task_struct *child)
6727 if (child_event->parent) {
6728 raw_spin_lock_irq(&child_ctx->lock);
6729 perf_group_detach(child_event);
6730 raw_spin_unlock_irq(&child_ctx->lock);
6733 perf_remove_from_context(child_event);
6736 * It can happen that the parent exits first, and has events
6737 * that are still around due to the child reference. These
6738 * events need to be zapped.
6740 if (child_event->parent) {
6741 sync_child_event(child_event, child);
6742 free_event(child_event);
6746 static void perf_event_exit_task_context(struct task_struct *child, int ctxn)
6748 struct perf_event *child_event, *tmp;
6749 struct perf_event_context *child_ctx;
6750 unsigned long flags;
6752 if (likely(!child->perf_event_ctxp[ctxn])) {
6753 perf_event_task(child, NULL, 0);
6757 local_irq_save(flags);
6759 * We can't reschedule here because interrupts are disabled,
6760 * and either child is current or it is a task that can't be
6761 * scheduled, so we are now safe from rescheduling changing
6764 child_ctx = rcu_dereference_raw(child->perf_event_ctxp[ctxn]);
6767 * Take the context lock here so that if find_get_context is
6768 * reading child->perf_event_ctxp, we wait until it has
6769 * incremented the context's refcount before we do put_ctx below.
6771 raw_spin_lock(&child_ctx->lock);
6772 task_ctx_sched_out(child_ctx);
6773 child->perf_event_ctxp[ctxn] = NULL;
6775 * If this context is a clone; unclone it so it can't get
6776 * swapped to another process while we're removing all
6777 * the events from it.
6779 unclone_ctx(child_ctx);
6780 update_context_time(child_ctx);
6781 raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
6784 * Report the task dead after unscheduling the events so that we
6785 * won't get any samples after PERF_RECORD_EXIT. We can however still
6786 * get a few PERF_RECORD_READ events.
6788 perf_event_task(child, child_ctx, 0);
6791 * We can recurse on the same lock type through:
6793 * __perf_event_exit_task()
6794 * sync_child_event()
6795 * fput(parent_event->filp)
6797 * mutex_lock(&ctx->mutex)
6799 * But since its the parent context it won't be the same instance.
6801 mutex_lock(&child_ctx->mutex);
6804 list_for_each_entry_safe(child_event, tmp, &child_ctx->pinned_groups,
6806 __perf_event_exit_task(child_event, child_ctx, child);
6808 list_for_each_entry_safe(child_event, tmp, &child_ctx->flexible_groups,
6810 __perf_event_exit_task(child_event, child_ctx, child);
6813 * If the last event was a group event, it will have appended all
6814 * its siblings to the list, but we obtained 'tmp' before that which
6815 * will still point to the list head terminating the iteration.
6817 if (!list_empty(&child_ctx->pinned_groups) ||
6818 !list_empty(&child_ctx->flexible_groups))
6821 mutex_unlock(&child_ctx->mutex);
6827 * When a child task exits, feed back event values to parent events.
6829 void perf_event_exit_task(struct task_struct *child)
6831 struct perf_event *event, *tmp;
6834 mutex_lock(&child->perf_event_mutex);
6835 list_for_each_entry_safe(event, tmp, &child->perf_event_list,
6837 list_del_init(&event->owner_entry);
6840 * Ensure the list deletion is visible before we clear
6841 * the owner, closes a race against perf_release() where
6842 * we need to serialize on the owner->perf_event_mutex.
6845 event->owner = NULL;
6847 mutex_unlock(&child->perf_event_mutex);
6849 for_each_task_context_nr(ctxn)
6850 perf_event_exit_task_context(child, ctxn);
6853 static void perf_free_event(struct perf_event *event,
6854 struct perf_event_context *ctx)
6856 struct perf_event *parent = event->parent;
6858 if (WARN_ON_ONCE(!parent))
6861 mutex_lock(&parent->child_mutex);
6862 list_del_init(&event->child_list);
6863 mutex_unlock(&parent->child_mutex);
6867 perf_group_detach(event);
6868 list_del_event(event, ctx);
6873 * free an unexposed, unused context as created by inheritance by
6874 * perf_event_init_task below, used by fork() in case of fail.
6876 void perf_event_free_task(struct task_struct *task)
6878 struct perf_event_context *ctx;
6879 struct perf_event *event, *tmp;
6882 for_each_task_context_nr(ctxn) {
6883 ctx = task->perf_event_ctxp[ctxn];
6887 mutex_lock(&ctx->mutex);
6889 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups,
6891 perf_free_event(event, ctx);
6893 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups,
6895 perf_free_event(event, ctx);
6897 if (!list_empty(&ctx->pinned_groups) ||
6898 !list_empty(&ctx->flexible_groups))
6901 mutex_unlock(&ctx->mutex);
6907 void perf_event_delayed_put(struct task_struct *task)
6911 for_each_task_context_nr(ctxn)
6912 WARN_ON_ONCE(task->perf_event_ctxp[ctxn]);
6916 * inherit a event from parent task to child task:
6918 static struct perf_event *
6919 inherit_event(struct perf_event *parent_event,
6920 struct task_struct *parent,
6921 struct perf_event_context *parent_ctx,
6922 struct task_struct *child,
6923 struct perf_event *group_leader,
6924 struct perf_event_context *child_ctx)
6926 struct perf_event *child_event;
6927 unsigned long flags;
6930 * Instead of creating recursive hierarchies of events,
6931 * we link inherited events back to the original parent,
6932 * which has a filp for sure, which we use as the reference
6935 if (parent_event->parent)
6936 parent_event = parent_event->parent;
6938 child_event = perf_event_alloc(&parent_event->attr,
6941 group_leader, parent_event,
6943 if (IS_ERR(child_event))
6948 * Make the child state follow the state of the parent event,
6949 * not its attr.disabled bit. We hold the parent's mutex,
6950 * so we won't race with perf_event_{en, dis}able_family.
6952 if (parent_event->state >= PERF_EVENT_STATE_INACTIVE)
6953 child_event->state = PERF_EVENT_STATE_INACTIVE;
6955 child_event->state = PERF_EVENT_STATE_OFF;
6957 if (parent_event->attr.freq) {
6958 u64 sample_period = parent_event->hw.sample_period;
6959 struct hw_perf_event *hwc = &child_event->hw;
6961 hwc->sample_period = sample_period;
6962 hwc->last_period = sample_period;
6964 local64_set(&hwc->period_left, sample_period);
6967 child_event->ctx = child_ctx;
6968 child_event->overflow_handler = parent_event->overflow_handler;
6971 * Precalculate sample_data sizes
6973 perf_event__header_size(child_event);
6974 perf_event__id_header_size(child_event);
6977 * Link it up in the child's context:
6979 raw_spin_lock_irqsave(&child_ctx->lock, flags);
6980 add_event_to_ctx(child_event, child_ctx);
6981 raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
6984 * Get a reference to the parent filp - we will fput it
6985 * when the child event exits. This is safe to do because
6986 * we are in the parent and we know that the filp still
6987 * exists and has a nonzero count:
6989 atomic_long_inc(&parent_event->filp->f_count);
6992 * Link this into the parent event's child list
6994 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
6995 mutex_lock(&parent_event->child_mutex);
6996 list_add_tail(&child_event->child_list, &parent_event->child_list);
6997 mutex_unlock(&parent_event->child_mutex);
7002 static int inherit_group(struct perf_event *parent_event,
7003 struct task_struct *parent,
7004 struct perf_event_context *parent_ctx,
7005 struct task_struct *child,
7006 struct perf_event_context *child_ctx)
7008 struct perf_event *leader;
7009 struct perf_event *sub;
7010 struct perf_event *child_ctr;
7012 leader = inherit_event(parent_event, parent, parent_ctx,
7013 child, NULL, child_ctx);
7015 return PTR_ERR(leader);
7016 list_for_each_entry(sub, &parent_event->sibling_list, group_entry) {
7017 child_ctr = inherit_event(sub, parent, parent_ctx,
7018 child, leader, child_ctx);
7019 if (IS_ERR(child_ctr))
7020 return PTR_ERR(child_ctr);
7026 inherit_task_group(struct perf_event *event, struct task_struct *parent,
7027 struct perf_event_context *parent_ctx,
7028 struct task_struct *child, int ctxn,
7032 struct perf_event_context *child_ctx;
7034 if (!event->attr.inherit) {
7039 child_ctx = child->perf_event_ctxp[ctxn];
7042 * This is executed from the parent task context, so
7043 * inherit events that have been marked for cloning.
7044 * First allocate and initialize a context for the
7048 child_ctx = alloc_perf_context(event->pmu, child);
7052 child->perf_event_ctxp[ctxn] = child_ctx;
7055 ret = inherit_group(event, parent, parent_ctx,
7065 * Initialize the perf_event context in task_struct
7067 int perf_event_init_context(struct task_struct *child, int ctxn)
7069 struct perf_event_context *child_ctx, *parent_ctx;
7070 struct perf_event_context *cloned_ctx;
7071 struct perf_event *event;
7072 struct task_struct *parent = current;
7073 int inherited_all = 1;
7074 unsigned long flags;
7077 if (likely(!parent->perf_event_ctxp[ctxn]))
7081 * If the parent's context is a clone, pin it so it won't get
7084 parent_ctx = perf_pin_task_context(parent, ctxn);
7087 * No need to check if parent_ctx != NULL here; since we saw
7088 * it non-NULL earlier, the only reason for it to become NULL
7089 * is if we exit, and since we're currently in the middle of
7090 * a fork we can't be exiting at the same time.
7094 * Lock the parent list. No need to lock the child - not PID
7095 * hashed yet and not running, so nobody can access it.
7097 mutex_lock(&parent_ctx->mutex);
7100 * We dont have to disable NMIs - we are only looking at
7101 * the list, not manipulating it:
7103 list_for_each_entry(event, &parent_ctx->pinned_groups, group_entry) {
7104 ret = inherit_task_group(event, parent, parent_ctx,
7105 child, ctxn, &inherited_all);
7111 * We can't hold ctx->lock when iterating the ->flexible_group list due
7112 * to allocations, but we need to prevent rotation because
7113 * rotate_ctx() will change the list from interrupt context.
7115 raw_spin_lock_irqsave(&parent_ctx->lock, flags);
7116 parent_ctx->rotate_disable = 1;
7117 raw_spin_unlock_irqrestore(&parent_ctx->lock, flags);
7119 list_for_each_entry(event, &parent_ctx->flexible_groups, group_entry) {
7120 ret = inherit_task_group(event, parent, parent_ctx,
7121 child, ctxn, &inherited_all);
7126 raw_spin_lock_irqsave(&parent_ctx->lock, flags);
7127 parent_ctx->rotate_disable = 0;
7129 child_ctx = child->perf_event_ctxp[ctxn];
7131 if (child_ctx && inherited_all) {
7133 * Mark the child context as a clone of the parent
7134 * context, or of whatever the parent is a clone of.
7136 * Note that if the parent is a clone, the holding of
7137 * parent_ctx->lock avoids it from being uncloned.
7139 cloned_ctx = parent_ctx->parent_ctx;
7141 child_ctx->parent_ctx = cloned_ctx;
7142 child_ctx->parent_gen = parent_ctx->parent_gen;
7144 child_ctx->parent_ctx = parent_ctx;
7145 child_ctx->parent_gen = parent_ctx->generation;
7147 get_ctx(child_ctx->parent_ctx);
7150 raw_spin_unlock_irqrestore(&parent_ctx->lock, flags);
7151 mutex_unlock(&parent_ctx->mutex);
7153 perf_unpin_context(parent_ctx);
7154 put_ctx(parent_ctx);
7160 * Initialize the perf_event context in task_struct
7162 int perf_event_init_task(struct task_struct *child)
7166 memset(child->perf_event_ctxp, 0, sizeof(child->perf_event_ctxp));
7167 mutex_init(&child->perf_event_mutex);
7168 INIT_LIST_HEAD(&child->perf_event_list);
7170 for_each_task_context_nr(ctxn) {
7171 ret = perf_event_init_context(child, ctxn);
7179 static void __init perf_event_init_all_cpus(void)
7181 struct swevent_htable *swhash;
7184 for_each_possible_cpu(cpu) {
7185 swhash = &per_cpu(swevent_htable, cpu);
7186 mutex_init(&swhash->hlist_mutex);
7187 INIT_LIST_HEAD(&per_cpu(rotation_list, cpu));
7191 static void __cpuinit perf_event_init_cpu(int cpu)
7193 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
7195 mutex_lock(&swhash->hlist_mutex);
7196 if (swhash->hlist_refcount > 0) {
7197 struct swevent_hlist *hlist;
7199 hlist = kzalloc_node(sizeof(*hlist), GFP_KERNEL, cpu_to_node(cpu));
7201 rcu_assign_pointer(swhash->swevent_hlist, hlist);
7203 mutex_unlock(&swhash->hlist_mutex);
7206 #if defined CONFIG_HOTPLUG_CPU || defined CONFIG_KEXEC
7207 static void perf_pmu_rotate_stop(struct pmu *pmu)
7209 struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
7211 WARN_ON(!irqs_disabled());
7213 list_del_init(&cpuctx->rotation_list);
7216 static void __perf_event_exit_context(void *__info)
7218 struct perf_event_context *ctx = __info;
7219 struct perf_event *event, *tmp;
7221 perf_pmu_rotate_stop(ctx->pmu);
7223 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry)
7224 __perf_remove_from_context(event);
7225 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups, group_entry)
7226 __perf_remove_from_context(event);
7229 static void perf_event_exit_cpu_context(int cpu)
7231 struct perf_event_context *ctx;
7235 idx = srcu_read_lock(&pmus_srcu);
7236 list_for_each_entry_rcu(pmu, &pmus, entry) {
7237 ctx = &per_cpu_ptr(pmu->pmu_cpu_context, cpu)->ctx;
7239 mutex_lock(&ctx->mutex);
7240 smp_call_function_single(cpu, __perf_event_exit_context, ctx, 1);
7241 mutex_unlock(&ctx->mutex);
7243 srcu_read_unlock(&pmus_srcu, idx);
7246 static void perf_event_exit_cpu(int cpu)
7248 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
7250 mutex_lock(&swhash->hlist_mutex);
7251 swevent_hlist_release(swhash);
7252 mutex_unlock(&swhash->hlist_mutex);
7254 perf_event_exit_cpu_context(cpu);
7257 static inline void perf_event_exit_cpu(int cpu) { }
7261 perf_reboot(struct notifier_block *notifier, unsigned long val, void *v)
7265 for_each_online_cpu(cpu)
7266 perf_event_exit_cpu(cpu);
7272 * Run the perf reboot notifier at the very last possible moment so that
7273 * the generic watchdog code runs as long as possible.
7275 static struct notifier_block perf_reboot_notifier = {
7276 .notifier_call = perf_reboot,
7277 .priority = INT_MIN,
7280 static int __cpuinit
7281 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
7283 unsigned int cpu = (long)hcpu;
7285 switch (action & ~CPU_TASKS_FROZEN) {
7287 case CPU_UP_PREPARE:
7288 case CPU_DOWN_FAILED:
7289 perf_event_init_cpu(cpu);
7292 case CPU_UP_CANCELED:
7293 case CPU_DOWN_PREPARE:
7294 perf_event_exit_cpu(cpu);
7304 void __init perf_event_init(void)
7310 perf_event_init_all_cpus();
7311 init_srcu_struct(&pmus_srcu);
7312 perf_pmu_register(&perf_swevent, "software", PERF_TYPE_SOFTWARE);
7313 perf_pmu_register(&perf_cpu_clock, NULL, -1);
7314 perf_pmu_register(&perf_task_clock, NULL, -1);
7316 perf_cpu_notifier(perf_cpu_notify);
7317 register_reboot_notifier(&perf_reboot_notifier);
7319 ret = init_hw_breakpoint();
7320 WARN(ret, "hw_breakpoint initialization failed with: %d", ret);
7323 static int __init perf_event_sysfs_init(void)
7328 mutex_lock(&pmus_lock);
7330 ret = bus_register(&pmu_bus);
7334 list_for_each_entry(pmu, &pmus, entry) {
7335 if (!pmu->name || pmu->type < 0)
7338 ret = pmu_dev_alloc(pmu);
7339 WARN(ret, "Failed to register pmu: %s, reason %d\n", pmu->name, ret);
7341 pmu_bus_running = 1;
7345 mutex_unlock(&pmus_lock);
7349 device_initcall(perf_event_sysfs_init);
7351 #ifdef CONFIG_CGROUP_PERF
7352 static struct cgroup_subsys_state *perf_cgroup_create(
7353 struct cgroup_subsys *ss, struct cgroup *cont)
7355 struct perf_cgroup *jc;
7357 jc = kzalloc(sizeof(*jc), GFP_KERNEL);
7359 return ERR_PTR(-ENOMEM);
7361 jc->info = alloc_percpu(struct perf_cgroup_info);
7364 return ERR_PTR(-ENOMEM);
7370 static void perf_cgroup_destroy(struct cgroup_subsys *ss,
7371 struct cgroup *cont)
7373 struct perf_cgroup *jc;
7374 jc = container_of(cgroup_subsys_state(cont, perf_subsys_id),
7375 struct perf_cgroup, css);
7376 free_percpu(jc->info);
7380 static int __perf_cgroup_move(void *info)
7382 struct task_struct *task = info;
7383 perf_cgroup_switch(task, PERF_CGROUP_SWOUT | PERF_CGROUP_SWIN);
7387 static void perf_cgroup_move(struct task_struct *task)
7389 task_function_call(task, __perf_cgroup_move, task);
7392 static void perf_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
7393 struct cgroup *old_cgrp, struct task_struct *task,
7396 perf_cgroup_move(task);
7398 struct task_struct *c;
7400 list_for_each_entry_rcu(c, &task->thread_group, thread_group) {
7401 perf_cgroup_move(c);
7407 static void perf_cgroup_exit(struct cgroup_subsys *ss, struct cgroup *cgrp,
7408 struct cgroup *old_cgrp, struct task_struct *task)
7411 * cgroup_exit() is called in the copy_process() failure path.
7412 * Ignore this case since the task hasn't ran yet, this avoids
7413 * trying to poke a half freed task state from generic code.
7415 if (!(task->flags & PF_EXITING))
7418 perf_cgroup_move(task);
7421 struct cgroup_subsys perf_subsys = {
7422 .name = "perf_event",
7423 .subsys_id = perf_subsys_id,
7424 .create = perf_cgroup_create,
7425 .destroy = perf_cgroup_destroy,
7426 .exit = perf_cgroup_exit,
7427 .attach = perf_cgroup_attach,
7429 #endif /* CONFIG_CGROUP_PERF */