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 if (!ctx->nr_events && cpuctx->task_ctx == ctx) {
1125 cpuctx->task_ctx = NULL;
1127 raw_spin_unlock(&ctx->lock);
1134 * Remove the event from a task's (or a CPU's) list of events.
1136 * CPU events are removed with a smp call. For task events we only
1137 * call when the task is on a CPU.
1139 * If event->ctx is a cloned context, callers must make sure that
1140 * every task struct that event->ctx->task could possibly point to
1141 * remains valid. This is OK when called from perf_release since
1142 * that only calls us on the top-level context, which can't be a clone.
1143 * When called from perf_event_exit_task, it's OK because the
1144 * context has been detached from its task.
1146 static void perf_remove_from_context(struct perf_event *event)
1148 struct perf_event_context *ctx = event->ctx;
1149 struct task_struct *task = ctx->task;
1151 lockdep_assert_held(&ctx->mutex);
1155 * Per cpu events are removed via an smp call and
1156 * the removal is always successful.
1158 cpu_function_call(event->cpu, __perf_remove_from_context, event);
1163 if (!task_function_call(task, __perf_remove_from_context, event))
1166 raw_spin_lock_irq(&ctx->lock);
1168 * If we failed to find a running task, but find the context active now
1169 * that we've acquired the ctx->lock, retry.
1171 if (ctx->is_active) {
1172 raw_spin_unlock_irq(&ctx->lock);
1177 * Since the task isn't running, its safe to remove the event, us
1178 * holding the ctx->lock ensures the task won't get scheduled in.
1180 list_del_event(event, ctx);
1181 raw_spin_unlock_irq(&ctx->lock);
1185 * Cross CPU call to disable a performance event
1187 static int __perf_event_disable(void *info)
1189 struct perf_event *event = info;
1190 struct perf_event_context *ctx = event->ctx;
1191 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1194 * If this is a per-task event, need to check whether this
1195 * event's task is the current task on this cpu.
1197 * Can trigger due to concurrent perf_event_context_sched_out()
1198 * flipping contexts around.
1200 if (ctx->task && cpuctx->task_ctx != ctx)
1203 raw_spin_lock(&ctx->lock);
1206 * If the event is on, turn it off.
1207 * If it is in error state, leave it in error state.
1209 if (event->state >= PERF_EVENT_STATE_INACTIVE) {
1210 update_context_time(ctx);
1211 update_cgrp_time_from_event(event);
1212 update_group_times(event);
1213 if (event == event->group_leader)
1214 group_sched_out(event, cpuctx, ctx);
1216 event_sched_out(event, cpuctx, ctx);
1217 event->state = PERF_EVENT_STATE_OFF;
1220 raw_spin_unlock(&ctx->lock);
1228 * If event->ctx is a cloned context, callers must make sure that
1229 * every task struct that event->ctx->task could possibly point to
1230 * remains valid. This condition is satisifed when called through
1231 * perf_event_for_each_child or perf_event_for_each because they
1232 * hold the top-level event's child_mutex, so any descendant that
1233 * goes to exit will block in sync_child_event.
1234 * When called from perf_pending_event it's OK because event->ctx
1235 * is the current context on this CPU and preemption is disabled,
1236 * hence we can't get into perf_event_task_sched_out for this context.
1238 void perf_event_disable(struct perf_event *event)
1240 struct perf_event_context *ctx = event->ctx;
1241 struct task_struct *task = ctx->task;
1245 * Disable the event on the cpu that it's on
1247 cpu_function_call(event->cpu, __perf_event_disable, event);
1252 if (!task_function_call(task, __perf_event_disable, event))
1255 raw_spin_lock_irq(&ctx->lock);
1257 * If the event is still active, we need to retry the cross-call.
1259 if (event->state == PERF_EVENT_STATE_ACTIVE) {
1260 raw_spin_unlock_irq(&ctx->lock);
1262 * Reload the task pointer, it might have been changed by
1263 * a concurrent perf_event_context_sched_out().
1270 * Since we have the lock this context can't be scheduled
1271 * in, so we can change the state safely.
1273 if (event->state == PERF_EVENT_STATE_INACTIVE) {
1274 update_group_times(event);
1275 event->state = PERF_EVENT_STATE_OFF;
1277 raw_spin_unlock_irq(&ctx->lock);
1280 static void perf_set_shadow_time(struct perf_event *event,
1281 struct perf_event_context *ctx,
1285 * use the correct time source for the time snapshot
1287 * We could get by without this by leveraging the
1288 * fact that to get to this function, the caller
1289 * has most likely already called update_context_time()
1290 * and update_cgrp_time_xx() and thus both timestamp
1291 * are identical (or very close). Given that tstamp is,
1292 * already adjusted for cgroup, we could say that:
1293 * tstamp - ctx->timestamp
1295 * tstamp - cgrp->timestamp.
1297 * Then, in perf_output_read(), the calculation would
1298 * work with no changes because:
1299 * - event is guaranteed scheduled in
1300 * - no scheduled out in between
1301 * - thus the timestamp would be the same
1303 * But this is a bit hairy.
1305 * So instead, we have an explicit cgroup call to remain
1306 * within the time time source all along. We believe it
1307 * is cleaner and simpler to understand.
1309 if (is_cgroup_event(event))
1310 perf_cgroup_set_shadow_time(event, tstamp);
1312 event->shadow_ctx_time = tstamp - ctx->timestamp;
1315 #define MAX_INTERRUPTS (~0ULL)
1317 static void perf_log_throttle(struct perf_event *event, int enable);
1320 event_sched_in(struct perf_event *event,
1321 struct perf_cpu_context *cpuctx,
1322 struct perf_event_context *ctx)
1324 u64 tstamp = perf_event_time(event);
1326 if (event->state <= PERF_EVENT_STATE_OFF)
1329 event->state = PERF_EVENT_STATE_ACTIVE;
1330 event->oncpu = smp_processor_id();
1333 * Unthrottle events, since we scheduled we might have missed several
1334 * ticks already, also for a heavily scheduling task there is little
1335 * guarantee it'll get a tick in a timely manner.
1337 if (unlikely(event->hw.interrupts == MAX_INTERRUPTS)) {
1338 perf_log_throttle(event, 1);
1339 event->hw.interrupts = 0;
1343 * The new state must be visible before we turn it on in the hardware:
1347 if (event->pmu->add(event, PERF_EF_START)) {
1348 event->state = PERF_EVENT_STATE_INACTIVE;
1353 event->tstamp_running += tstamp - event->tstamp_stopped;
1355 perf_set_shadow_time(event, ctx, tstamp);
1357 if (!is_software_event(event))
1358 cpuctx->active_oncpu++;
1361 if (event->attr.exclusive)
1362 cpuctx->exclusive = 1;
1368 group_sched_in(struct perf_event *group_event,
1369 struct perf_cpu_context *cpuctx,
1370 struct perf_event_context *ctx)
1372 struct perf_event *event, *partial_group = NULL;
1373 struct pmu *pmu = group_event->pmu;
1374 u64 now = ctx->time;
1375 bool simulate = false;
1377 if (group_event->state == PERF_EVENT_STATE_OFF)
1380 pmu->start_txn(pmu);
1382 if (event_sched_in(group_event, cpuctx, ctx)) {
1383 pmu->cancel_txn(pmu);
1388 * Schedule in siblings as one group (if any):
1390 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
1391 if (event_sched_in(event, cpuctx, ctx)) {
1392 partial_group = event;
1397 if (!pmu->commit_txn(pmu))
1402 * Groups can be scheduled in as one unit only, so undo any
1403 * partial group before returning:
1404 * The events up to the failed event are scheduled out normally,
1405 * tstamp_stopped will be updated.
1407 * The failed events and the remaining siblings need to have
1408 * their timings updated as if they had gone thru event_sched_in()
1409 * and event_sched_out(). This is required to get consistent timings
1410 * across the group. This also takes care of the case where the group
1411 * could never be scheduled by ensuring tstamp_stopped is set to mark
1412 * the time the event was actually stopped, such that time delta
1413 * calculation in update_event_times() is correct.
1415 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
1416 if (event == partial_group)
1420 event->tstamp_running += now - event->tstamp_stopped;
1421 event->tstamp_stopped = now;
1423 event_sched_out(event, cpuctx, ctx);
1426 event_sched_out(group_event, cpuctx, ctx);
1428 pmu->cancel_txn(pmu);
1434 * Work out whether we can put this event group on the CPU now.
1436 static int group_can_go_on(struct perf_event *event,
1437 struct perf_cpu_context *cpuctx,
1441 * Groups consisting entirely of software events can always go on.
1443 if (event->group_flags & PERF_GROUP_SOFTWARE)
1446 * If an exclusive group is already on, no other hardware
1449 if (cpuctx->exclusive)
1452 * If this group is exclusive and there are already
1453 * events on the CPU, it can't go on.
1455 if (event->attr.exclusive && cpuctx->active_oncpu)
1458 * Otherwise, try to add it if all previous groups were able
1464 static void add_event_to_ctx(struct perf_event *event,
1465 struct perf_event_context *ctx)
1467 u64 tstamp = perf_event_time(event);
1469 list_add_event(event, ctx);
1470 perf_group_attach(event);
1471 event->tstamp_enabled = tstamp;
1472 event->tstamp_running = tstamp;
1473 event->tstamp_stopped = tstamp;
1476 static void task_ctx_sched_out(struct perf_event_context *ctx);
1478 ctx_sched_in(struct perf_event_context *ctx,
1479 struct perf_cpu_context *cpuctx,
1480 enum event_type_t event_type,
1481 struct task_struct *task);
1483 static void perf_event_sched_in(struct perf_cpu_context *cpuctx,
1484 struct perf_event_context *ctx,
1485 struct task_struct *task)
1487 cpu_ctx_sched_in(cpuctx, EVENT_PINNED, task);
1489 ctx_sched_in(ctx, cpuctx, EVENT_PINNED, task);
1490 cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE, task);
1492 ctx_sched_in(ctx, cpuctx, EVENT_FLEXIBLE, task);
1496 * Cross CPU call to install and enable a performance event
1498 * Must be called with ctx->mutex held
1500 static int __perf_install_in_context(void *info)
1502 struct perf_event *event = info;
1503 struct perf_event_context *ctx = event->ctx;
1504 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1505 struct perf_event_context *task_ctx = cpuctx->task_ctx;
1506 struct task_struct *task = current;
1508 perf_ctx_lock(cpuctx, task_ctx);
1509 perf_pmu_disable(cpuctx->ctx.pmu);
1512 * If there was an active task_ctx schedule it out.
1515 task_ctx_sched_out(task_ctx);
1518 * If the context we're installing events in is not the
1519 * active task_ctx, flip them.
1521 if (ctx->task && task_ctx != ctx) {
1523 raw_spin_unlock(&task_ctx->lock);
1524 raw_spin_lock(&ctx->lock);
1529 cpuctx->task_ctx = task_ctx;
1530 task = task_ctx->task;
1533 cpu_ctx_sched_out(cpuctx, EVENT_ALL);
1535 update_context_time(ctx);
1537 * update cgrp time only if current cgrp
1538 * matches event->cgrp. Must be done before
1539 * calling add_event_to_ctx()
1541 update_cgrp_time_from_event(event);
1543 add_event_to_ctx(event, ctx);
1546 * Schedule everything back in
1548 perf_event_sched_in(cpuctx, task_ctx, task);
1550 perf_pmu_enable(cpuctx->ctx.pmu);
1551 perf_ctx_unlock(cpuctx, task_ctx);
1557 * Attach a performance event to a context
1559 * First we add the event to the list with the hardware enable bit
1560 * in event->hw_config cleared.
1562 * If the event is attached to a task which is on a CPU we use a smp
1563 * call to enable it in the task context. The task might have been
1564 * scheduled away, but we check this in the smp call again.
1567 perf_install_in_context(struct perf_event_context *ctx,
1568 struct perf_event *event,
1571 struct task_struct *task = ctx->task;
1573 lockdep_assert_held(&ctx->mutex);
1579 * Per cpu events are installed via an smp call and
1580 * the install is always successful.
1582 cpu_function_call(cpu, __perf_install_in_context, event);
1587 if (!task_function_call(task, __perf_install_in_context, event))
1590 raw_spin_lock_irq(&ctx->lock);
1592 * If we failed to find a running task, but find the context active now
1593 * that we've acquired the ctx->lock, retry.
1595 if (ctx->is_active) {
1596 raw_spin_unlock_irq(&ctx->lock);
1601 * Since the task isn't running, its safe to add the event, us holding
1602 * the ctx->lock ensures the task won't get scheduled in.
1604 add_event_to_ctx(event, ctx);
1605 raw_spin_unlock_irq(&ctx->lock);
1609 * Put a event into inactive state and update time fields.
1610 * Enabling the leader of a group effectively enables all
1611 * the group members that aren't explicitly disabled, so we
1612 * have to update their ->tstamp_enabled also.
1613 * Note: this works for group members as well as group leaders
1614 * since the non-leader members' sibling_lists will be empty.
1616 static void __perf_event_mark_enabled(struct perf_event *event,
1617 struct perf_event_context *ctx)
1619 struct perf_event *sub;
1620 u64 tstamp = perf_event_time(event);
1622 event->state = PERF_EVENT_STATE_INACTIVE;
1623 event->tstamp_enabled = tstamp - event->total_time_enabled;
1624 list_for_each_entry(sub, &event->sibling_list, group_entry) {
1625 if (sub->state >= PERF_EVENT_STATE_INACTIVE)
1626 sub->tstamp_enabled = tstamp - sub->total_time_enabled;
1631 * Cross CPU call to enable a performance event
1633 static int __perf_event_enable(void *info)
1635 struct perf_event *event = info;
1636 struct perf_event_context *ctx = event->ctx;
1637 struct perf_event *leader = event->group_leader;
1638 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1641 if (WARN_ON_ONCE(!ctx->is_active))
1644 raw_spin_lock(&ctx->lock);
1645 update_context_time(ctx);
1647 if (event->state >= PERF_EVENT_STATE_INACTIVE)
1651 * set current task's cgroup time reference point
1653 perf_cgroup_set_timestamp(current, ctx);
1655 __perf_event_mark_enabled(event, ctx);
1657 if (!event_filter_match(event)) {
1658 if (is_cgroup_event(event))
1659 perf_cgroup_defer_enabled(event);
1664 * If the event is in a group and isn't the group leader,
1665 * then don't put it on unless the group is on.
1667 if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE)
1670 if (!group_can_go_on(event, cpuctx, 1)) {
1673 if (event == leader)
1674 err = group_sched_in(event, cpuctx, ctx);
1676 err = event_sched_in(event, cpuctx, ctx);
1681 * If this event can't go on and it's part of a
1682 * group, then the whole group has to come off.
1684 if (leader != event)
1685 group_sched_out(leader, cpuctx, ctx);
1686 if (leader->attr.pinned) {
1687 update_group_times(leader);
1688 leader->state = PERF_EVENT_STATE_ERROR;
1693 raw_spin_unlock(&ctx->lock);
1701 * If event->ctx is a cloned context, callers must make sure that
1702 * every task struct that event->ctx->task could possibly point to
1703 * remains valid. This condition is satisfied when called through
1704 * perf_event_for_each_child or perf_event_for_each as described
1705 * for perf_event_disable.
1707 void perf_event_enable(struct perf_event *event)
1709 struct perf_event_context *ctx = event->ctx;
1710 struct task_struct *task = ctx->task;
1714 * Enable the event on the cpu that it's on
1716 cpu_function_call(event->cpu, __perf_event_enable, event);
1720 raw_spin_lock_irq(&ctx->lock);
1721 if (event->state >= PERF_EVENT_STATE_INACTIVE)
1725 * If the event is in error state, clear that first.
1726 * That way, if we see the event in error state below, we
1727 * know that it has gone back into error state, as distinct
1728 * from the task having been scheduled away before the
1729 * cross-call arrived.
1731 if (event->state == PERF_EVENT_STATE_ERROR)
1732 event->state = PERF_EVENT_STATE_OFF;
1735 if (!ctx->is_active) {
1736 __perf_event_mark_enabled(event, ctx);
1740 raw_spin_unlock_irq(&ctx->lock);
1742 if (!task_function_call(task, __perf_event_enable, event))
1745 raw_spin_lock_irq(&ctx->lock);
1748 * If the context is active and the event is still off,
1749 * we need to retry the cross-call.
1751 if (ctx->is_active && event->state == PERF_EVENT_STATE_OFF) {
1753 * task could have been flipped by a concurrent
1754 * perf_event_context_sched_out()
1761 raw_spin_unlock_irq(&ctx->lock);
1764 static int perf_event_refresh(struct perf_event *event, int refresh)
1767 * not supported on inherited events
1769 if (event->attr.inherit || !is_sampling_event(event))
1772 atomic_add(refresh, &event->event_limit);
1773 perf_event_enable(event);
1778 static void ctx_sched_out(struct perf_event_context *ctx,
1779 struct perf_cpu_context *cpuctx,
1780 enum event_type_t event_type)
1782 struct perf_event *event;
1783 int is_active = ctx->is_active;
1785 ctx->is_active &= ~event_type;
1786 if (likely(!ctx->nr_events))
1789 update_context_time(ctx);
1790 update_cgrp_time_from_cpuctx(cpuctx);
1791 if (!ctx->nr_active)
1794 perf_pmu_disable(ctx->pmu);
1795 if ((is_active & EVENT_PINNED) && (event_type & EVENT_PINNED)) {
1796 list_for_each_entry(event, &ctx->pinned_groups, group_entry)
1797 group_sched_out(event, cpuctx, ctx);
1800 if ((is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE)) {
1801 list_for_each_entry(event, &ctx->flexible_groups, group_entry)
1802 group_sched_out(event, cpuctx, ctx);
1804 perf_pmu_enable(ctx->pmu);
1808 * Test whether two contexts are equivalent, i.e. whether they
1809 * have both been cloned from the same version of the same context
1810 * and they both have the same number of enabled events.
1811 * If the number of enabled events is the same, then the set
1812 * of enabled events should be the same, because these are both
1813 * inherited contexts, therefore we can't access individual events
1814 * in them directly with an fd; we can only enable/disable all
1815 * events via prctl, or enable/disable all events in a family
1816 * via ioctl, which will have the same effect on both contexts.
1818 static int context_equiv(struct perf_event_context *ctx1,
1819 struct perf_event_context *ctx2)
1821 return ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx
1822 && ctx1->parent_gen == ctx2->parent_gen
1823 && !ctx1->pin_count && !ctx2->pin_count;
1826 static void __perf_event_sync_stat(struct perf_event *event,
1827 struct perf_event *next_event)
1831 if (!event->attr.inherit_stat)
1835 * Update the event value, we cannot use perf_event_read()
1836 * because we're in the middle of a context switch and have IRQs
1837 * disabled, which upsets smp_call_function_single(), however
1838 * we know the event must be on the current CPU, therefore we
1839 * don't need to use it.
1841 switch (event->state) {
1842 case PERF_EVENT_STATE_ACTIVE:
1843 event->pmu->read(event);
1846 case PERF_EVENT_STATE_INACTIVE:
1847 update_event_times(event);
1855 * In order to keep per-task stats reliable we need to flip the event
1856 * values when we flip the contexts.
1858 value = local64_read(&next_event->count);
1859 value = local64_xchg(&event->count, value);
1860 local64_set(&next_event->count, value);
1862 swap(event->total_time_enabled, next_event->total_time_enabled);
1863 swap(event->total_time_running, next_event->total_time_running);
1866 * Since we swizzled the values, update the user visible data too.
1868 perf_event_update_userpage(event);
1869 perf_event_update_userpage(next_event);
1872 #define list_next_entry(pos, member) \
1873 list_entry(pos->member.next, typeof(*pos), member)
1875 static void perf_event_sync_stat(struct perf_event_context *ctx,
1876 struct perf_event_context *next_ctx)
1878 struct perf_event *event, *next_event;
1883 update_context_time(ctx);
1885 event = list_first_entry(&ctx->event_list,
1886 struct perf_event, event_entry);
1888 next_event = list_first_entry(&next_ctx->event_list,
1889 struct perf_event, event_entry);
1891 while (&event->event_entry != &ctx->event_list &&
1892 &next_event->event_entry != &next_ctx->event_list) {
1894 __perf_event_sync_stat(event, next_event);
1896 event = list_next_entry(event, event_entry);
1897 next_event = list_next_entry(next_event, event_entry);
1901 static void perf_event_context_sched_out(struct task_struct *task, int ctxn,
1902 struct task_struct *next)
1904 struct perf_event_context *ctx = task->perf_event_ctxp[ctxn];
1905 struct perf_event_context *next_ctx;
1906 struct perf_event_context *parent;
1907 struct perf_cpu_context *cpuctx;
1913 cpuctx = __get_cpu_context(ctx);
1914 if (!cpuctx->task_ctx)
1918 parent = rcu_dereference(ctx->parent_ctx);
1919 next_ctx = next->perf_event_ctxp[ctxn];
1920 if (parent && next_ctx &&
1921 rcu_dereference(next_ctx->parent_ctx) == parent) {
1923 * Looks like the two contexts are clones, so we might be
1924 * able to optimize the context switch. We lock both
1925 * contexts and check that they are clones under the
1926 * lock (including re-checking that neither has been
1927 * uncloned in the meantime). It doesn't matter which
1928 * order we take the locks because no other cpu could
1929 * be trying to lock both of these tasks.
1931 raw_spin_lock(&ctx->lock);
1932 raw_spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING);
1933 if (context_equiv(ctx, next_ctx)) {
1935 * XXX do we need a memory barrier of sorts
1936 * wrt to rcu_dereference() of perf_event_ctxp
1938 task->perf_event_ctxp[ctxn] = next_ctx;
1939 next->perf_event_ctxp[ctxn] = ctx;
1941 next_ctx->task = task;
1944 perf_event_sync_stat(ctx, next_ctx);
1946 raw_spin_unlock(&next_ctx->lock);
1947 raw_spin_unlock(&ctx->lock);
1952 raw_spin_lock(&ctx->lock);
1953 ctx_sched_out(ctx, cpuctx, EVENT_ALL);
1954 cpuctx->task_ctx = NULL;
1955 raw_spin_unlock(&ctx->lock);
1959 #define for_each_task_context_nr(ctxn) \
1960 for ((ctxn) = 0; (ctxn) < perf_nr_task_contexts; (ctxn)++)
1963 * Called from scheduler to remove the events of the current task,
1964 * with interrupts disabled.
1966 * We stop each event and update the event value in event->count.
1968 * This does not protect us against NMI, but disable()
1969 * sets the disabled bit in the control field of event _before_
1970 * accessing the event control register. If a NMI hits, then it will
1971 * not restart the event.
1973 void __perf_event_task_sched_out(struct task_struct *task,
1974 struct task_struct *next)
1978 for_each_task_context_nr(ctxn)
1979 perf_event_context_sched_out(task, ctxn, next);
1982 * if cgroup events exist on this CPU, then we need
1983 * to check if we have to switch out PMU state.
1984 * cgroup event are system-wide mode only
1986 if (atomic_read(&__get_cpu_var(perf_cgroup_events)))
1987 perf_cgroup_sched_out(task);
1990 static void task_ctx_sched_out(struct perf_event_context *ctx)
1992 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1994 if (!cpuctx->task_ctx)
1997 if (WARN_ON_ONCE(ctx != cpuctx->task_ctx))
2000 ctx_sched_out(ctx, cpuctx, EVENT_ALL);
2001 cpuctx->task_ctx = NULL;
2005 * Called with IRQs disabled
2007 static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
2008 enum event_type_t event_type)
2010 ctx_sched_out(&cpuctx->ctx, cpuctx, event_type);
2014 ctx_pinned_sched_in(struct perf_event_context *ctx,
2015 struct perf_cpu_context *cpuctx)
2017 struct perf_event *event;
2019 list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
2020 if (event->state <= PERF_EVENT_STATE_OFF)
2022 if (!event_filter_match(event))
2025 /* may need to reset tstamp_enabled */
2026 if (is_cgroup_event(event))
2027 perf_cgroup_mark_enabled(event, ctx);
2029 if (group_can_go_on(event, cpuctx, 1))
2030 group_sched_in(event, cpuctx, ctx);
2033 * If this pinned group hasn't been scheduled,
2034 * put it in error state.
2036 if (event->state == PERF_EVENT_STATE_INACTIVE) {
2037 update_group_times(event);
2038 event->state = PERF_EVENT_STATE_ERROR;
2044 ctx_flexible_sched_in(struct perf_event_context *ctx,
2045 struct perf_cpu_context *cpuctx)
2047 struct perf_event *event;
2050 list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
2051 /* Ignore events in OFF or ERROR state */
2052 if (event->state <= PERF_EVENT_STATE_OFF)
2055 * Listen to the 'cpu' scheduling filter constraint
2058 if (!event_filter_match(event))
2061 /* may need to reset tstamp_enabled */
2062 if (is_cgroup_event(event))
2063 perf_cgroup_mark_enabled(event, ctx);
2065 if (group_can_go_on(event, cpuctx, can_add_hw)) {
2066 if (group_sched_in(event, cpuctx, ctx))
2073 ctx_sched_in(struct perf_event_context *ctx,
2074 struct perf_cpu_context *cpuctx,
2075 enum event_type_t event_type,
2076 struct task_struct *task)
2079 int is_active = ctx->is_active;
2081 ctx->is_active |= event_type;
2082 if (likely(!ctx->nr_events))
2086 ctx->timestamp = now;
2087 perf_cgroup_set_timestamp(task, ctx);
2089 * First go through the list and put on any pinned groups
2090 * in order to give them the best chance of going on.
2092 if (!(is_active & EVENT_PINNED) && (event_type & EVENT_PINNED))
2093 ctx_pinned_sched_in(ctx, cpuctx);
2095 /* Then walk through the lower prio flexible groups */
2096 if (!(is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE))
2097 ctx_flexible_sched_in(ctx, cpuctx);
2100 static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
2101 enum event_type_t event_type,
2102 struct task_struct *task)
2104 struct perf_event_context *ctx = &cpuctx->ctx;
2106 ctx_sched_in(ctx, cpuctx, event_type, task);
2109 static void perf_event_context_sched_in(struct perf_event_context *ctx,
2110 struct task_struct *task)
2112 struct perf_cpu_context *cpuctx;
2114 cpuctx = __get_cpu_context(ctx);
2115 if (cpuctx->task_ctx == ctx)
2118 perf_ctx_lock(cpuctx, ctx);
2119 perf_pmu_disable(ctx->pmu);
2121 * We want to keep the following priority order:
2122 * cpu pinned (that don't need to move), task pinned,
2123 * cpu flexible, task flexible.
2125 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
2127 perf_event_sched_in(cpuctx, ctx, task);
2129 cpuctx->task_ctx = ctx;
2131 perf_pmu_enable(ctx->pmu);
2132 perf_ctx_unlock(cpuctx, ctx);
2135 * Since these rotations are per-cpu, we need to ensure the
2136 * cpu-context we got scheduled on is actually rotating.
2138 perf_pmu_rotate_start(ctx->pmu);
2142 * Called from scheduler to add the events of the current task
2143 * with interrupts disabled.
2145 * We restore the event value and then enable it.
2147 * This does not protect us against NMI, but enable()
2148 * sets the enabled bit in the control field of event _before_
2149 * accessing the event control register. If a NMI hits, then it will
2150 * keep the event running.
2152 void __perf_event_task_sched_in(struct task_struct *task)
2154 struct perf_event_context *ctx;
2157 for_each_task_context_nr(ctxn) {
2158 ctx = task->perf_event_ctxp[ctxn];
2162 perf_event_context_sched_in(ctx, task);
2165 * if cgroup events exist on this CPU, then we need
2166 * to check if we have to switch in PMU state.
2167 * cgroup event are system-wide mode only
2169 if (atomic_read(&__get_cpu_var(perf_cgroup_events)))
2170 perf_cgroup_sched_in(task);
2173 static u64 perf_calculate_period(struct perf_event *event, u64 nsec, u64 count)
2175 u64 frequency = event->attr.sample_freq;
2176 u64 sec = NSEC_PER_SEC;
2177 u64 divisor, dividend;
2179 int count_fls, nsec_fls, frequency_fls, sec_fls;
2181 count_fls = fls64(count);
2182 nsec_fls = fls64(nsec);
2183 frequency_fls = fls64(frequency);
2187 * We got @count in @nsec, with a target of sample_freq HZ
2188 * the target period becomes:
2191 * period = -------------------
2192 * @nsec * sample_freq
2197 * Reduce accuracy by one bit such that @a and @b converge
2198 * to a similar magnitude.
2200 #define REDUCE_FLS(a, b) \
2202 if (a##_fls > b##_fls) { \
2212 * Reduce accuracy until either term fits in a u64, then proceed with
2213 * the other, so that finally we can do a u64/u64 division.
2215 while (count_fls + sec_fls > 64 && nsec_fls + frequency_fls > 64) {
2216 REDUCE_FLS(nsec, frequency);
2217 REDUCE_FLS(sec, count);
2220 if (count_fls + sec_fls > 64) {
2221 divisor = nsec * frequency;
2223 while (count_fls + sec_fls > 64) {
2224 REDUCE_FLS(count, sec);
2228 dividend = count * sec;
2230 dividend = count * sec;
2232 while (nsec_fls + frequency_fls > 64) {
2233 REDUCE_FLS(nsec, frequency);
2237 divisor = nsec * frequency;
2243 return div64_u64(dividend, divisor);
2246 static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count)
2248 struct hw_perf_event *hwc = &event->hw;
2249 s64 period, sample_period;
2252 period = perf_calculate_period(event, nsec, count);
2254 delta = (s64)(period - hwc->sample_period);
2255 delta = (delta + 7) / 8; /* low pass filter */
2257 sample_period = hwc->sample_period + delta;
2262 hwc->sample_period = sample_period;
2264 if (local64_read(&hwc->period_left) > 8*sample_period) {
2265 event->pmu->stop(event, PERF_EF_UPDATE);
2266 local64_set(&hwc->period_left, 0);
2267 event->pmu->start(event, PERF_EF_RELOAD);
2271 static void perf_ctx_adjust_freq(struct perf_event_context *ctx, u64 period)
2273 struct perf_event *event;
2274 struct hw_perf_event *hwc;
2275 u64 interrupts, now;
2278 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
2279 if (event->state != PERF_EVENT_STATE_ACTIVE)
2282 if (!event_filter_match(event))
2287 interrupts = hwc->interrupts;
2288 hwc->interrupts = 0;
2291 * unthrottle events on the tick
2293 if (interrupts == MAX_INTERRUPTS) {
2294 perf_log_throttle(event, 1);
2295 event->pmu->start(event, 0);
2298 if (!event->attr.freq || !event->attr.sample_freq)
2301 event->pmu->read(event);
2302 now = local64_read(&event->count);
2303 delta = now - hwc->freq_count_stamp;
2304 hwc->freq_count_stamp = now;
2307 perf_adjust_period(event, period, delta);
2312 * Round-robin a context's events:
2314 static void rotate_ctx(struct perf_event_context *ctx)
2317 * Rotate the first entry last of non-pinned groups. Rotation might be
2318 * disabled by the inheritance code.
2320 if (!ctx->rotate_disable)
2321 list_rotate_left(&ctx->flexible_groups);
2325 * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
2326 * because they're strictly cpu affine and rotate_start is called with IRQs
2327 * disabled, while rotate_context is called from IRQ context.
2329 static void perf_rotate_context(struct perf_cpu_context *cpuctx)
2331 u64 interval = (u64)cpuctx->jiffies_interval * TICK_NSEC;
2332 struct perf_event_context *ctx = NULL;
2333 int rotate = 0, remove = 1;
2335 if (cpuctx->ctx.nr_events) {
2337 if (cpuctx->ctx.nr_events != cpuctx->ctx.nr_active)
2341 ctx = cpuctx->task_ctx;
2342 if (ctx && ctx->nr_events) {
2344 if (ctx->nr_events != ctx->nr_active)
2348 perf_ctx_lock(cpuctx, cpuctx->task_ctx);
2349 perf_pmu_disable(cpuctx->ctx.pmu);
2350 perf_ctx_adjust_freq(&cpuctx->ctx, interval);
2352 perf_ctx_adjust_freq(ctx, interval);
2357 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
2359 ctx_sched_out(ctx, cpuctx, EVENT_FLEXIBLE);
2361 rotate_ctx(&cpuctx->ctx);
2365 perf_event_sched_in(cpuctx, ctx, current);
2369 list_del_init(&cpuctx->rotation_list);
2371 perf_pmu_enable(cpuctx->ctx.pmu);
2372 perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
2375 void perf_event_task_tick(void)
2377 struct list_head *head = &__get_cpu_var(rotation_list);
2378 struct perf_cpu_context *cpuctx, *tmp;
2380 WARN_ON(!irqs_disabled());
2382 list_for_each_entry_safe(cpuctx, tmp, head, rotation_list) {
2383 if (cpuctx->jiffies_interval == 1 ||
2384 !(jiffies % cpuctx->jiffies_interval))
2385 perf_rotate_context(cpuctx);
2389 static int event_enable_on_exec(struct perf_event *event,
2390 struct perf_event_context *ctx)
2392 if (!event->attr.enable_on_exec)
2395 event->attr.enable_on_exec = 0;
2396 if (event->state >= PERF_EVENT_STATE_INACTIVE)
2399 __perf_event_mark_enabled(event, ctx);
2405 * Enable all of a task's events that have been marked enable-on-exec.
2406 * This expects task == current.
2408 static void perf_event_enable_on_exec(struct perf_event_context *ctx)
2410 struct perf_event *event;
2411 unsigned long flags;
2415 local_irq_save(flags);
2416 if (!ctx || !ctx->nr_events)
2420 * We must ctxsw out cgroup events to avoid conflict
2421 * when invoking perf_task_event_sched_in() later on
2422 * in this function. Otherwise we end up trying to
2423 * ctxswin cgroup events which are already scheduled
2426 perf_cgroup_sched_out(current);
2428 raw_spin_lock(&ctx->lock);
2429 task_ctx_sched_out(ctx);
2431 list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
2432 ret = event_enable_on_exec(event, ctx);
2437 list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
2438 ret = event_enable_on_exec(event, ctx);
2444 * Unclone this context if we enabled any event.
2449 raw_spin_unlock(&ctx->lock);
2452 * Also calls ctxswin for cgroup events, if any:
2454 perf_event_context_sched_in(ctx, ctx->task);
2456 local_irq_restore(flags);
2460 * Cross CPU call to read the hardware event
2462 static void __perf_event_read(void *info)
2464 struct perf_event *event = info;
2465 struct perf_event_context *ctx = event->ctx;
2466 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
2469 * If this is a task context, we need to check whether it is
2470 * the current task context of this cpu. If not it has been
2471 * scheduled out before the smp call arrived. In that case
2472 * event->count would have been updated to a recent sample
2473 * when the event was scheduled out.
2475 if (ctx->task && cpuctx->task_ctx != ctx)
2478 raw_spin_lock(&ctx->lock);
2479 if (ctx->is_active) {
2480 update_context_time(ctx);
2481 update_cgrp_time_from_event(event);
2483 update_event_times(event);
2484 if (event->state == PERF_EVENT_STATE_ACTIVE)
2485 event->pmu->read(event);
2486 raw_spin_unlock(&ctx->lock);
2489 static inline u64 perf_event_count(struct perf_event *event)
2491 return local64_read(&event->count) + atomic64_read(&event->child_count);
2494 static u64 perf_event_read(struct perf_event *event)
2497 * If event is enabled and currently active on a CPU, update the
2498 * value in the event structure:
2500 if (event->state == PERF_EVENT_STATE_ACTIVE) {
2501 smp_call_function_single(event->oncpu,
2502 __perf_event_read, event, 1);
2503 } else if (event->state == PERF_EVENT_STATE_INACTIVE) {
2504 struct perf_event_context *ctx = event->ctx;
2505 unsigned long flags;
2507 raw_spin_lock_irqsave(&ctx->lock, flags);
2509 * may read while context is not active
2510 * (e.g., thread is blocked), in that case
2511 * we cannot update context time
2513 if (ctx->is_active) {
2514 update_context_time(ctx);
2515 update_cgrp_time_from_event(event);
2517 update_event_times(event);
2518 raw_spin_unlock_irqrestore(&ctx->lock, flags);
2521 return perf_event_count(event);
2528 struct callchain_cpus_entries {
2529 struct rcu_head rcu_head;
2530 struct perf_callchain_entry *cpu_entries[0];
2533 static DEFINE_PER_CPU(int, callchain_recursion[PERF_NR_CONTEXTS]);
2534 static atomic_t nr_callchain_events;
2535 static DEFINE_MUTEX(callchain_mutex);
2536 struct callchain_cpus_entries *callchain_cpus_entries;
2539 __weak void perf_callchain_kernel(struct perf_callchain_entry *entry,
2540 struct pt_regs *regs)
2544 __weak void perf_callchain_user(struct perf_callchain_entry *entry,
2545 struct pt_regs *regs)
2549 static void release_callchain_buffers_rcu(struct rcu_head *head)
2551 struct callchain_cpus_entries *entries;
2554 entries = container_of(head, struct callchain_cpus_entries, rcu_head);
2556 for_each_possible_cpu(cpu)
2557 kfree(entries->cpu_entries[cpu]);
2562 static void release_callchain_buffers(void)
2564 struct callchain_cpus_entries *entries;
2566 entries = callchain_cpus_entries;
2567 rcu_assign_pointer(callchain_cpus_entries, NULL);
2568 call_rcu(&entries->rcu_head, release_callchain_buffers_rcu);
2571 static int alloc_callchain_buffers(void)
2575 struct callchain_cpus_entries *entries;
2578 * We can't use the percpu allocation API for data that can be
2579 * accessed from NMI. Use a temporary manual per cpu allocation
2580 * until that gets sorted out.
2582 size = offsetof(struct callchain_cpus_entries, cpu_entries[nr_cpu_ids]);
2584 entries = kzalloc(size, GFP_KERNEL);
2588 size = sizeof(struct perf_callchain_entry) * PERF_NR_CONTEXTS;
2590 for_each_possible_cpu(cpu) {
2591 entries->cpu_entries[cpu] = kmalloc_node(size, GFP_KERNEL,
2593 if (!entries->cpu_entries[cpu])
2597 rcu_assign_pointer(callchain_cpus_entries, entries);
2602 for_each_possible_cpu(cpu)
2603 kfree(entries->cpu_entries[cpu]);
2609 static int get_callchain_buffers(void)
2614 mutex_lock(&callchain_mutex);
2616 count = atomic_inc_return(&nr_callchain_events);
2617 if (WARN_ON_ONCE(count < 1)) {
2623 /* If the allocation failed, give up */
2624 if (!callchain_cpus_entries)
2629 err = alloc_callchain_buffers();
2631 release_callchain_buffers();
2633 mutex_unlock(&callchain_mutex);
2638 static void put_callchain_buffers(void)
2640 if (atomic_dec_and_mutex_lock(&nr_callchain_events, &callchain_mutex)) {
2641 release_callchain_buffers();
2642 mutex_unlock(&callchain_mutex);
2646 static int get_recursion_context(int *recursion)
2654 else if (in_softirq())
2659 if (recursion[rctx])
2668 static inline void put_recursion_context(int *recursion, int rctx)
2674 static struct perf_callchain_entry *get_callchain_entry(int *rctx)
2677 struct callchain_cpus_entries *entries;
2679 *rctx = get_recursion_context(__get_cpu_var(callchain_recursion));
2683 entries = rcu_dereference(callchain_cpus_entries);
2687 cpu = smp_processor_id();
2689 return &entries->cpu_entries[cpu][*rctx];
2693 put_callchain_entry(int rctx)
2695 put_recursion_context(__get_cpu_var(callchain_recursion), rctx);
2698 static struct perf_callchain_entry *perf_callchain(struct pt_regs *regs)
2701 struct perf_callchain_entry *entry;
2704 entry = get_callchain_entry(&rctx);
2713 if (!user_mode(regs)) {
2714 perf_callchain_store(entry, PERF_CONTEXT_KERNEL);
2715 perf_callchain_kernel(entry, regs);
2717 regs = task_pt_regs(current);
2723 perf_callchain_store(entry, PERF_CONTEXT_USER);
2724 perf_callchain_user(entry, regs);
2728 put_callchain_entry(rctx);
2734 * Initialize the perf_event context in a task_struct:
2736 static void __perf_event_init_context(struct perf_event_context *ctx)
2738 raw_spin_lock_init(&ctx->lock);
2739 mutex_init(&ctx->mutex);
2740 INIT_LIST_HEAD(&ctx->pinned_groups);
2741 INIT_LIST_HEAD(&ctx->flexible_groups);
2742 INIT_LIST_HEAD(&ctx->event_list);
2743 atomic_set(&ctx->refcount, 1);
2746 static struct perf_event_context *
2747 alloc_perf_context(struct pmu *pmu, struct task_struct *task)
2749 struct perf_event_context *ctx;
2751 ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL);
2755 __perf_event_init_context(ctx);
2758 get_task_struct(task);
2765 static struct task_struct *
2766 find_lively_task_by_vpid(pid_t vpid)
2768 struct task_struct *task;
2775 task = find_task_by_vpid(vpid);
2777 get_task_struct(task);
2781 return ERR_PTR(-ESRCH);
2783 /* Reuse ptrace permission checks for now. */
2785 if (!ptrace_may_access(task, PTRACE_MODE_READ))
2790 put_task_struct(task);
2791 return ERR_PTR(err);
2796 * Returns a matching context with refcount and pincount.
2798 static struct perf_event_context *
2799 find_get_context(struct pmu *pmu, struct task_struct *task, int cpu)
2801 struct perf_event_context *ctx;
2802 struct perf_cpu_context *cpuctx;
2803 unsigned long flags;
2807 /* Must be root to operate on a CPU event: */
2808 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
2809 return ERR_PTR(-EACCES);
2812 * We could be clever and allow to attach a event to an
2813 * offline CPU and activate it when the CPU comes up, but
2816 if (!cpu_online(cpu))
2817 return ERR_PTR(-ENODEV);
2819 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
2828 ctxn = pmu->task_ctx_nr;
2833 ctx = perf_lock_task_context(task, ctxn, &flags);
2837 raw_spin_unlock_irqrestore(&ctx->lock, flags);
2839 ctx = alloc_perf_context(pmu, task);
2845 mutex_lock(&task->perf_event_mutex);
2847 * If it has already passed perf_event_exit_task().
2848 * we must see PF_EXITING, it takes this mutex too.
2850 if (task->flags & PF_EXITING)
2852 else if (task->perf_event_ctxp[ctxn])
2857 rcu_assign_pointer(task->perf_event_ctxp[ctxn], ctx);
2859 mutex_unlock(&task->perf_event_mutex);
2861 if (unlikely(err)) {
2873 return ERR_PTR(err);
2876 static void perf_event_free_filter(struct perf_event *event);
2878 static void free_event_rcu(struct rcu_head *head)
2880 struct perf_event *event;
2882 event = container_of(head, struct perf_event, rcu_head);
2884 put_pid_ns(event->ns);
2885 perf_event_free_filter(event);
2889 static void perf_buffer_put(struct perf_buffer *buffer);
2891 static void free_event(struct perf_event *event)
2893 irq_work_sync(&event->pending);
2895 if (!event->parent) {
2896 if (event->attach_state & PERF_ATTACH_TASK)
2897 jump_label_dec(&perf_sched_events);
2898 if (event->attr.mmap || event->attr.mmap_data)
2899 atomic_dec(&nr_mmap_events);
2900 if (event->attr.comm)
2901 atomic_dec(&nr_comm_events);
2902 if (event->attr.task)
2903 atomic_dec(&nr_task_events);
2904 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN)
2905 put_callchain_buffers();
2906 if (is_cgroup_event(event)) {
2907 atomic_dec(&per_cpu(perf_cgroup_events, event->cpu));
2908 jump_label_dec(&perf_sched_events);
2912 if (event->buffer) {
2913 perf_buffer_put(event->buffer);
2914 event->buffer = NULL;
2917 if (is_cgroup_event(event))
2918 perf_detach_cgroup(event);
2921 event->destroy(event);
2924 put_ctx(event->ctx);
2926 call_rcu(&event->rcu_head, free_event_rcu);
2929 int perf_event_release_kernel(struct perf_event *event)
2931 struct perf_event_context *ctx = event->ctx;
2933 WARN_ON_ONCE(ctx->parent_ctx);
2935 * There are two ways this annotation is useful:
2937 * 1) there is a lock recursion from perf_event_exit_task
2938 * see the comment there.
2940 * 2) there is a lock-inversion with mmap_sem through
2941 * perf_event_read_group(), which takes faults while
2942 * holding ctx->mutex, however this is called after
2943 * the last filedesc died, so there is no possibility
2944 * to trigger the AB-BA case.
2946 mutex_lock_nested(&ctx->mutex, SINGLE_DEPTH_NESTING);
2947 raw_spin_lock_irq(&ctx->lock);
2948 perf_group_detach(event);
2949 raw_spin_unlock_irq(&ctx->lock);
2950 perf_remove_from_context(event);
2951 mutex_unlock(&ctx->mutex);
2957 EXPORT_SYMBOL_GPL(perf_event_release_kernel);
2960 * Called when the last reference to the file is gone.
2962 static int perf_release(struct inode *inode, struct file *file)
2964 struct perf_event *event = file->private_data;
2965 struct task_struct *owner;
2967 file->private_data = NULL;
2970 owner = ACCESS_ONCE(event->owner);
2972 * Matches the smp_wmb() in perf_event_exit_task(). If we observe
2973 * !owner it means the list deletion is complete and we can indeed
2974 * free this event, otherwise we need to serialize on
2975 * owner->perf_event_mutex.
2977 smp_read_barrier_depends();
2980 * Since delayed_put_task_struct() also drops the last
2981 * task reference we can safely take a new reference
2982 * while holding the rcu_read_lock().
2984 get_task_struct(owner);
2989 mutex_lock(&owner->perf_event_mutex);
2991 * We have to re-check the event->owner field, if it is cleared
2992 * we raced with perf_event_exit_task(), acquiring the mutex
2993 * ensured they're done, and we can proceed with freeing the
2997 list_del_init(&event->owner_entry);
2998 mutex_unlock(&owner->perf_event_mutex);
2999 put_task_struct(owner);
3002 return perf_event_release_kernel(event);
3005 u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running)
3007 struct perf_event *child;
3013 mutex_lock(&event->child_mutex);
3014 total += perf_event_read(event);
3015 *enabled += event->total_time_enabled +
3016 atomic64_read(&event->child_total_time_enabled);
3017 *running += event->total_time_running +
3018 atomic64_read(&event->child_total_time_running);
3020 list_for_each_entry(child, &event->child_list, child_list) {
3021 total += perf_event_read(child);
3022 *enabled += child->total_time_enabled;
3023 *running += child->total_time_running;
3025 mutex_unlock(&event->child_mutex);
3029 EXPORT_SYMBOL_GPL(perf_event_read_value);
3031 static int perf_event_read_group(struct perf_event *event,
3032 u64 read_format, char __user *buf)
3034 struct perf_event *leader = event->group_leader, *sub;
3035 int n = 0, size = 0, ret = -EFAULT;
3036 struct perf_event_context *ctx = leader->ctx;
3038 u64 count, enabled, running;
3040 mutex_lock(&ctx->mutex);
3041 count = perf_event_read_value(leader, &enabled, &running);
3043 values[n++] = 1 + leader->nr_siblings;
3044 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3045 values[n++] = enabled;
3046 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3047 values[n++] = running;
3048 values[n++] = count;
3049 if (read_format & PERF_FORMAT_ID)
3050 values[n++] = primary_event_id(leader);
3052 size = n * sizeof(u64);
3054 if (copy_to_user(buf, values, size))
3059 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
3062 values[n++] = perf_event_read_value(sub, &enabled, &running);
3063 if (read_format & PERF_FORMAT_ID)
3064 values[n++] = primary_event_id(sub);
3066 size = n * sizeof(u64);
3068 if (copy_to_user(buf + ret, values, size)) {
3076 mutex_unlock(&ctx->mutex);
3081 static int perf_event_read_one(struct perf_event *event,
3082 u64 read_format, char __user *buf)
3084 u64 enabled, running;
3088 values[n++] = perf_event_read_value(event, &enabled, &running);
3089 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3090 values[n++] = enabled;
3091 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3092 values[n++] = running;
3093 if (read_format & PERF_FORMAT_ID)
3094 values[n++] = primary_event_id(event);
3096 if (copy_to_user(buf, values, n * sizeof(u64)))
3099 return n * sizeof(u64);
3103 * Read the performance event - simple non blocking version for now
3106 perf_read_hw(struct perf_event *event, char __user *buf, size_t count)
3108 u64 read_format = event->attr.read_format;
3112 * Return end-of-file for a read on a event that is in
3113 * error state (i.e. because it was pinned but it couldn't be
3114 * scheduled on to the CPU at some point).
3116 if (event->state == PERF_EVENT_STATE_ERROR)
3119 if (count < event->read_size)
3122 WARN_ON_ONCE(event->ctx->parent_ctx);
3123 if (read_format & PERF_FORMAT_GROUP)
3124 ret = perf_event_read_group(event, read_format, buf);
3126 ret = perf_event_read_one(event, read_format, buf);
3132 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
3134 struct perf_event *event = file->private_data;
3136 return perf_read_hw(event, buf, count);
3139 static unsigned int perf_poll(struct file *file, poll_table *wait)
3141 struct perf_event *event = file->private_data;
3142 struct perf_buffer *buffer;
3143 unsigned int events = POLL_HUP;
3146 buffer = rcu_dereference(event->buffer);
3148 events = atomic_xchg(&buffer->poll, 0);
3151 poll_wait(file, &event->waitq, wait);
3156 static void perf_event_reset(struct perf_event *event)
3158 (void)perf_event_read(event);
3159 local64_set(&event->count, 0);
3160 perf_event_update_userpage(event);
3164 * Holding the top-level event's child_mutex means that any
3165 * descendant process that has inherited this event will block
3166 * in sync_child_event if it goes to exit, thus satisfying the
3167 * task existence requirements of perf_event_enable/disable.
3169 static void perf_event_for_each_child(struct perf_event *event,
3170 void (*func)(struct perf_event *))
3172 struct perf_event *child;
3174 WARN_ON_ONCE(event->ctx->parent_ctx);
3175 mutex_lock(&event->child_mutex);
3177 list_for_each_entry(child, &event->child_list, child_list)
3179 mutex_unlock(&event->child_mutex);
3182 static void perf_event_for_each(struct perf_event *event,
3183 void (*func)(struct perf_event *))
3185 struct perf_event_context *ctx = event->ctx;
3186 struct perf_event *sibling;
3188 WARN_ON_ONCE(ctx->parent_ctx);
3189 mutex_lock(&ctx->mutex);
3190 event = event->group_leader;
3192 perf_event_for_each_child(event, func);
3194 list_for_each_entry(sibling, &event->sibling_list, group_entry)
3195 perf_event_for_each_child(event, func);
3196 mutex_unlock(&ctx->mutex);
3199 static int perf_event_period(struct perf_event *event, u64 __user *arg)
3201 struct perf_event_context *ctx = event->ctx;
3205 if (!is_sampling_event(event))
3208 if (copy_from_user(&value, arg, sizeof(value)))
3214 raw_spin_lock_irq(&ctx->lock);
3215 if (event->attr.freq) {
3216 if (value > sysctl_perf_event_sample_rate) {
3221 event->attr.sample_freq = value;
3223 event->attr.sample_period = value;
3224 event->hw.sample_period = value;
3227 raw_spin_unlock_irq(&ctx->lock);
3232 static const struct file_operations perf_fops;
3234 static struct perf_event *perf_fget_light(int fd, int *fput_needed)
3238 file = fget_light(fd, fput_needed);
3240 return ERR_PTR(-EBADF);
3242 if (file->f_op != &perf_fops) {
3243 fput_light(file, *fput_needed);
3245 return ERR_PTR(-EBADF);
3248 return file->private_data;
3251 static int perf_event_set_output(struct perf_event *event,
3252 struct perf_event *output_event);
3253 static int perf_event_set_filter(struct perf_event *event, void __user *arg);
3255 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
3257 struct perf_event *event = file->private_data;
3258 void (*func)(struct perf_event *);
3262 case PERF_EVENT_IOC_ENABLE:
3263 func = perf_event_enable;
3265 case PERF_EVENT_IOC_DISABLE:
3266 func = perf_event_disable;
3268 case PERF_EVENT_IOC_RESET:
3269 func = perf_event_reset;
3272 case PERF_EVENT_IOC_REFRESH:
3273 return perf_event_refresh(event, arg);
3275 case PERF_EVENT_IOC_PERIOD:
3276 return perf_event_period(event, (u64 __user *)arg);
3278 case PERF_EVENT_IOC_SET_OUTPUT:
3280 struct perf_event *output_event = NULL;
3281 int fput_needed = 0;
3285 output_event = perf_fget_light(arg, &fput_needed);
3286 if (IS_ERR(output_event))
3287 return PTR_ERR(output_event);
3290 ret = perf_event_set_output(event, output_event);
3292 fput_light(output_event->filp, fput_needed);
3297 case PERF_EVENT_IOC_SET_FILTER:
3298 return perf_event_set_filter(event, (void __user *)arg);
3304 if (flags & PERF_IOC_FLAG_GROUP)
3305 perf_event_for_each(event, func);
3307 perf_event_for_each_child(event, func);
3312 int perf_event_task_enable(void)
3314 struct perf_event *event;
3316 mutex_lock(¤t->perf_event_mutex);
3317 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
3318 perf_event_for_each_child(event, perf_event_enable);
3319 mutex_unlock(¤t->perf_event_mutex);
3324 int perf_event_task_disable(void)
3326 struct perf_event *event;
3328 mutex_lock(¤t->perf_event_mutex);
3329 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
3330 perf_event_for_each_child(event, perf_event_disable);
3331 mutex_unlock(¤t->perf_event_mutex);
3336 #ifndef PERF_EVENT_INDEX_OFFSET
3337 # define PERF_EVENT_INDEX_OFFSET 0
3340 static int perf_event_index(struct perf_event *event)
3342 if (event->hw.state & PERF_HES_STOPPED)
3345 if (event->state != PERF_EVENT_STATE_ACTIVE)
3348 return event->hw.idx + 1 - PERF_EVENT_INDEX_OFFSET;
3352 * Callers need to ensure there can be no nesting of this function, otherwise
3353 * the seqlock logic goes bad. We can not serialize this because the arch
3354 * code calls this from NMI context.
3356 void perf_event_update_userpage(struct perf_event *event)
3358 struct perf_event_mmap_page *userpg;
3359 struct perf_buffer *buffer;
3362 buffer = rcu_dereference(event->buffer);
3366 userpg = buffer->user_page;
3369 * Disable preemption so as to not let the corresponding user-space
3370 * spin too long if we get preempted.
3375 userpg->index = perf_event_index(event);
3376 userpg->offset = perf_event_count(event);
3377 if (event->state == PERF_EVENT_STATE_ACTIVE)
3378 userpg->offset -= local64_read(&event->hw.prev_count);
3380 userpg->time_enabled = event->total_time_enabled +
3381 atomic64_read(&event->child_total_time_enabled);
3383 userpg->time_running = event->total_time_running +
3384 atomic64_read(&event->child_total_time_running);
3393 static unsigned long perf_data_size(struct perf_buffer *buffer);
3396 perf_buffer_init(struct perf_buffer *buffer, long watermark, int flags)
3398 long max_size = perf_data_size(buffer);
3401 buffer->watermark = min(max_size, watermark);
3403 if (!buffer->watermark)
3404 buffer->watermark = max_size / 2;
3406 if (flags & PERF_BUFFER_WRITABLE)
3407 buffer->writable = 1;
3409 atomic_set(&buffer->refcount, 1);
3412 #ifndef CONFIG_PERF_USE_VMALLOC
3415 * Back perf_mmap() with regular GFP_KERNEL-0 pages.
3418 static struct page *
3419 perf_mmap_to_page(struct perf_buffer *buffer, unsigned long pgoff)
3421 if (pgoff > buffer->nr_pages)
3425 return virt_to_page(buffer->user_page);
3427 return virt_to_page(buffer->data_pages[pgoff - 1]);
3430 static void *perf_mmap_alloc_page(int cpu)
3435 node = (cpu == -1) ? cpu : cpu_to_node(cpu);
3436 page = alloc_pages_node(node, GFP_KERNEL | __GFP_ZERO, 0);
3440 return page_address(page);
3443 static struct perf_buffer *
3444 perf_buffer_alloc(int nr_pages, long watermark, int cpu, int flags)
3446 struct perf_buffer *buffer;
3450 size = sizeof(struct perf_buffer);
3451 size += nr_pages * sizeof(void *);
3453 buffer = kzalloc(size, GFP_KERNEL);
3457 buffer->user_page = perf_mmap_alloc_page(cpu);
3458 if (!buffer->user_page)
3459 goto fail_user_page;
3461 for (i = 0; i < nr_pages; i++) {
3462 buffer->data_pages[i] = perf_mmap_alloc_page(cpu);
3463 if (!buffer->data_pages[i])
3464 goto fail_data_pages;
3467 buffer->nr_pages = nr_pages;
3469 perf_buffer_init(buffer, watermark, flags);
3474 for (i--; i >= 0; i--)
3475 free_page((unsigned long)buffer->data_pages[i]);
3477 free_page((unsigned long)buffer->user_page);
3486 static void perf_mmap_free_page(unsigned long addr)
3488 struct page *page = virt_to_page((void *)addr);
3490 page->mapping = NULL;
3494 static void perf_buffer_free(struct perf_buffer *buffer)
3498 perf_mmap_free_page((unsigned long)buffer->user_page);
3499 for (i = 0; i < buffer->nr_pages; i++)
3500 perf_mmap_free_page((unsigned long)buffer->data_pages[i]);
3504 static inline int page_order(struct perf_buffer *buffer)
3512 * Back perf_mmap() with vmalloc memory.
3514 * Required for architectures that have d-cache aliasing issues.
3517 static inline int page_order(struct perf_buffer *buffer)
3519 return buffer->page_order;
3522 static struct page *
3523 perf_mmap_to_page(struct perf_buffer *buffer, unsigned long pgoff)
3525 if (pgoff > (1UL << page_order(buffer)))
3528 return vmalloc_to_page((void *)buffer->user_page + pgoff * PAGE_SIZE);
3531 static void perf_mmap_unmark_page(void *addr)
3533 struct page *page = vmalloc_to_page(addr);
3535 page->mapping = NULL;
3538 static void perf_buffer_free_work(struct work_struct *work)
3540 struct perf_buffer *buffer;
3544 buffer = container_of(work, struct perf_buffer, work);
3545 nr = 1 << page_order(buffer);
3547 base = buffer->user_page;
3548 for (i = 0; i < nr + 1; i++)
3549 perf_mmap_unmark_page(base + (i * PAGE_SIZE));
3555 static void perf_buffer_free(struct perf_buffer *buffer)
3557 schedule_work(&buffer->work);
3560 static struct perf_buffer *
3561 perf_buffer_alloc(int nr_pages, long watermark, int cpu, int flags)
3563 struct perf_buffer *buffer;
3567 size = sizeof(struct perf_buffer);
3568 size += sizeof(void *);
3570 buffer = kzalloc(size, GFP_KERNEL);
3574 INIT_WORK(&buffer->work, perf_buffer_free_work);
3576 all_buf = vmalloc_user((nr_pages + 1) * PAGE_SIZE);
3580 buffer->user_page = all_buf;
3581 buffer->data_pages[0] = all_buf + PAGE_SIZE;
3582 buffer->page_order = ilog2(nr_pages);
3583 buffer->nr_pages = 1;
3585 perf_buffer_init(buffer, watermark, flags);
3598 static unsigned long perf_data_size(struct perf_buffer *buffer)
3600 return buffer->nr_pages << (PAGE_SHIFT + page_order(buffer));
3603 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
3605 struct perf_event *event = vma->vm_file->private_data;
3606 struct perf_buffer *buffer;
3607 int ret = VM_FAULT_SIGBUS;
3609 if (vmf->flags & FAULT_FLAG_MKWRITE) {
3610 if (vmf->pgoff == 0)
3616 buffer = rcu_dereference(event->buffer);
3620 if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE))
3623 vmf->page = perf_mmap_to_page(buffer, vmf->pgoff);
3627 get_page(vmf->page);
3628 vmf->page->mapping = vma->vm_file->f_mapping;
3629 vmf->page->index = vmf->pgoff;
3638 static void perf_buffer_free_rcu(struct rcu_head *rcu_head)
3640 struct perf_buffer *buffer;
3642 buffer = container_of(rcu_head, struct perf_buffer, rcu_head);
3643 perf_buffer_free(buffer);
3646 static struct perf_buffer *perf_buffer_get(struct perf_event *event)
3648 struct perf_buffer *buffer;
3651 buffer = rcu_dereference(event->buffer);
3653 if (!atomic_inc_not_zero(&buffer->refcount))
3661 static void perf_buffer_put(struct perf_buffer *buffer)
3663 if (!atomic_dec_and_test(&buffer->refcount))
3666 call_rcu(&buffer->rcu_head, perf_buffer_free_rcu);
3669 static void perf_mmap_open(struct vm_area_struct *vma)
3671 struct perf_event *event = vma->vm_file->private_data;
3673 atomic_inc(&event->mmap_count);
3676 static void perf_mmap_close(struct vm_area_struct *vma)
3678 struct perf_event *event = vma->vm_file->private_data;
3680 if (atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex)) {
3681 unsigned long size = perf_data_size(event->buffer);
3682 struct user_struct *user = event->mmap_user;
3683 struct perf_buffer *buffer = event->buffer;
3685 atomic_long_sub((size >> PAGE_SHIFT) + 1, &user->locked_vm);
3686 vma->vm_mm->locked_vm -= event->mmap_locked;
3687 rcu_assign_pointer(event->buffer, NULL);
3688 mutex_unlock(&event->mmap_mutex);
3690 perf_buffer_put(buffer);
3695 static const struct vm_operations_struct perf_mmap_vmops = {
3696 .open = perf_mmap_open,
3697 .close = perf_mmap_close,
3698 .fault = perf_mmap_fault,
3699 .page_mkwrite = perf_mmap_fault,
3702 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
3704 struct perf_event *event = file->private_data;
3705 unsigned long user_locked, user_lock_limit;
3706 struct user_struct *user = current_user();
3707 unsigned long locked, lock_limit;
3708 struct perf_buffer *buffer;
3709 unsigned long vma_size;
3710 unsigned long nr_pages;
3711 long user_extra, extra;
3712 int ret = 0, flags = 0;
3715 * Don't allow mmap() of inherited per-task counters. This would
3716 * create a performance issue due to all children writing to the
3719 if (event->cpu == -1 && event->attr.inherit)
3722 if (!(vma->vm_flags & VM_SHARED))
3725 vma_size = vma->vm_end - vma->vm_start;
3726 nr_pages = (vma_size / PAGE_SIZE) - 1;
3729 * If we have buffer pages ensure they're a power-of-two number, so we
3730 * can do bitmasks instead of modulo.
3732 if (nr_pages != 0 && !is_power_of_2(nr_pages))
3735 if (vma_size != PAGE_SIZE * (1 + nr_pages))
3738 if (vma->vm_pgoff != 0)
3741 WARN_ON_ONCE(event->ctx->parent_ctx);
3742 mutex_lock(&event->mmap_mutex);
3743 if (event->buffer) {
3744 if (event->buffer->nr_pages == nr_pages)
3745 atomic_inc(&event->buffer->refcount);
3751 user_extra = nr_pages + 1;
3752 user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10);
3755 * Increase the limit linearly with more CPUs:
3757 user_lock_limit *= num_online_cpus();
3759 user_locked = atomic_long_read(&user->locked_vm) + user_extra;
3762 if (user_locked > user_lock_limit)
3763 extra = user_locked - user_lock_limit;
3765 lock_limit = rlimit(RLIMIT_MEMLOCK);
3766 lock_limit >>= PAGE_SHIFT;
3767 locked = vma->vm_mm->locked_vm + extra;
3769 if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() &&
3770 !capable(CAP_IPC_LOCK)) {
3775 WARN_ON(event->buffer);
3777 if (vma->vm_flags & VM_WRITE)
3778 flags |= PERF_BUFFER_WRITABLE;
3780 buffer = perf_buffer_alloc(nr_pages, event->attr.wakeup_watermark,
3786 rcu_assign_pointer(event->buffer, buffer);
3788 atomic_long_add(user_extra, &user->locked_vm);
3789 event->mmap_locked = extra;
3790 event->mmap_user = get_current_user();
3791 vma->vm_mm->locked_vm += event->mmap_locked;
3795 atomic_inc(&event->mmap_count);
3796 mutex_unlock(&event->mmap_mutex);
3798 vma->vm_flags |= VM_RESERVED;
3799 vma->vm_ops = &perf_mmap_vmops;
3804 static int perf_fasync(int fd, struct file *filp, int on)
3806 struct inode *inode = filp->f_path.dentry->d_inode;
3807 struct perf_event *event = filp->private_data;
3810 mutex_lock(&inode->i_mutex);
3811 retval = fasync_helper(fd, filp, on, &event->fasync);
3812 mutex_unlock(&inode->i_mutex);
3820 static const struct file_operations perf_fops = {
3821 .llseek = no_llseek,
3822 .release = perf_release,
3825 .unlocked_ioctl = perf_ioctl,
3826 .compat_ioctl = perf_ioctl,
3828 .fasync = perf_fasync,
3834 * If there's data, ensure we set the poll() state and publish everything
3835 * to user-space before waking everybody up.
3838 void perf_event_wakeup(struct perf_event *event)
3840 wake_up_all(&event->waitq);
3842 if (event->pending_kill) {
3843 kill_fasync(&event->fasync, SIGIO, event->pending_kill);
3844 event->pending_kill = 0;
3848 static void perf_pending_event(struct irq_work *entry)
3850 struct perf_event *event = container_of(entry,
3851 struct perf_event, pending);
3853 if (event->pending_disable) {
3854 event->pending_disable = 0;
3855 __perf_event_disable(event);
3858 if (event->pending_wakeup) {
3859 event->pending_wakeup = 0;
3860 perf_event_wakeup(event);
3865 * We assume there is only KVM supporting the callbacks.
3866 * Later on, we might change it to a list if there is
3867 * another virtualization implementation supporting the callbacks.
3869 struct perf_guest_info_callbacks *perf_guest_cbs;
3871 int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
3873 perf_guest_cbs = cbs;
3876 EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks);
3878 int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
3880 perf_guest_cbs = NULL;
3883 EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks);
3888 static bool perf_output_space(struct perf_buffer *buffer, unsigned long tail,
3889 unsigned long offset, unsigned long head)
3893 if (!buffer->writable)
3896 mask = perf_data_size(buffer) - 1;
3898 offset = (offset - tail) & mask;
3899 head = (head - tail) & mask;
3901 if ((int)(head - offset) < 0)
3907 static void perf_output_wakeup(struct perf_output_handle *handle)
3909 atomic_set(&handle->buffer->poll, POLL_IN);
3912 handle->event->pending_wakeup = 1;
3913 irq_work_queue(&handle->event->pending);
3915 perf_event_wakeup(handle->event);
3919 * We need to ensure a later event_id doesn't publish a head when a former
3920 * event isn't done writing. However since we need to deal with NMIs we
3921 * cannot fully serialize things.
3923 * We only publish the head (and generate a wakeup) when the outer-most
3926 static void perf_output_get_handle(struct perf_output_handle *handle)
3928 struct perf_buffer *buffer = handle->buffer;
3931 local_inc(&buffer->nest);
3932 handle->wakeup = local_read(&buffer->wakeup);
3935 static void perf_output_put_handle(struct perf_output_handle *handle)
3937 struct perf_buffer *buffer = handle->buffer;
3941 head = local_read(&buffer->head);
3944 * IRQ/NMI can happen here, which means we can miss a head update.
3947 if (!local_dec_and_test(&buffer->nest))
3951 * Publish the known good head. Rely on the full barrier implied
3952 * by atomic_dec_and_test() order the buffer->head read and this
3955 buffer->user_page->data_head = head;
3958 * Now check if we missed an update, rely on the (compiler)
3959 * barrier in atomic_dec_and_test() to re-read buffer->head.
3961 if (unlikely(head != local_read(&buffer->head))) {
3962 local_inc(&buffer->nest);
3966 if (handle->wakeup != local_read(&buffer->wakeup))
3967 perf_output_wakeup(handle);
3973 __always_inline void perf_output_copy(struct perf_output_handle *handle,
3974 const void *buf, unsigned int len)
3977 unsigned long size = min_t(unsigned long, handle->size, len);
3979 memcpy(handle->addr, buf, size);
3982 handle->addr += size;
3984 handle->size -= size;
3985 if (!handle->size) {
3986 struct perf_buffer *buffer = handle->buffer;
3989 handle->page &= buffer->nr_pages - 1;
3990 handle->addr = buffer->data_pages[handle->page];
3991 handle->size = PAGE_SIZE << page_order(buffer);
3996 static void __perf_event_header__init_id(struct perf_event_header *header,
3997 struct perf_sample_data *data,
3998 struct perf_event *event)
4000 u64 sample_type = event->attr.sample_type;
4002 data->type = sample_type;
4003 header->size += event->id_header_size;
4005 if (sample_type & PERF_SAMPLE_TID) {
4006 /* namespace issues */
4007 data->tid_entry.pid = perf_event_pid(event, current);
4008 data->tid_entry.tid = perf_event_tid(event, current);
4011 if (sample_type & PERF_SAMPLE_TIME)
4012 data->time = perf_clock();
4014 if (sample_type & PERF_SAMPLE_ID)
4015 data->id = primary_event_id(event);
4017 if (sample_type & PERF_SAMPLE_STREAM_ID)
4018 data->stream_id = event->id;
4020 if (sample_type & PERF_SAMPLE_CPU) {
4021 data->cpu_entry.cpu = raw_smp_processor_id();
4022 data->cpu_entry.reserved = 0;
4026 static void perf_event_header__init_id(struct perf_event_header *header,
4027 struct perf_sample_data *data,
4028 struct perf_event *event)
4030 if (event->attr.sample_id_all)
4031 __perf_event_header__init_id(header, data, event);
4034 static void __perf_event__output_id_sample(struct perf_output_handle *handle,
4035 struct perf_sample_data *data)
4037 u64 sample_type = data->type;
4039 if (sample_type & PERF_SAMPLE_TID)
4040 perf_output_put(handle, data->tid_entry);
4042 if (sample_type & PERF_SAMPLE_TIME)
4043 perf_output_put(handle, data->time);
4045 if (sample_type & PERF_SAMPLE_ID)
4046 perf_output_put(handle, data->id);
4048 if (sample_type & PERF_SAMPLE_STREAM_ID)
4049 perf_output_put(handle, data->stream_id);
4051 if (sample_type & PERF_SAMPLE_CPU)
4052 perf_output_put(handle, data->cpu_entry);
4055 static void perf_event__output_id_sample(struct perf_event *event,
4056 struct perf_output_handle *handle,
4057 struct perf_sample_data *sample)
4059 if (event->attr.sample_id_all)
4060 __perf_event__output_id_sample(handle, sample);
4063 int perf_output_begin(struct perf_output_handle *handle,
4064 struct perf_event *event, unsigned int size,
4065 int nmi, int sample)
4067 struct perf_buffer *buffer;
4068 unsigned long tail, offset, head;
4070 struct perf_sample_data sample_data;
4072 struct perf_event_header header;
4079 * For inherited events we send all the output towards the parent.
4082 event = event->parent;
4084 buffer = rcu_dereference(event->buffer);
4088 handle->buffer = buffer;
4089 handle->event = event;
4091 handle->sample = sample;
4093 if (!buffer->nr_pages)
4096 have_lost = local_read(&buffer->lost);
4098 lost_event.header.size = sizeof(lost_event);
4099 perf_event_header__init_id(&lost_event.header, &sample_data,
4101 size += lost_event.header.size;
4104 perf_output_get_handle(handle);
4108 * Userspace could choose to issue a mb() before updating the
4109 * tail pointer. So that all reads will be completed before the
4112 tail = ACCESS_ONCE(buffer->user_page->data_tail);
4114 offset = head = local_read(&buffer->head);
4116 if (unlikely(!perf_output_space(buffer, tail, offset, head)))
4118 } while (local_cmpxchg(&buffer->head, offset, head) != offset);
4120 if (head - local_read(&buffer->wakeup) > buffer->watermark)
4121 local_add(buffer->watermark, &buffer->wakeup);
4123 handle->page = offset >> (PAGE_SHIFT + page_order(buffer));
4124 handle->page &= buffer->nr_pages - 1;
4125 handle->size = offset & ((PAGE_SIZE << page_order(buffer)) - 1);
4126 handle->addr = buffer->data_pages[handle->page];
4127 handle->addr += handle->size;
4128 handle->size = (PAGE_SIZE << page_order(buffer)) - handle->size;
4131 lost_event.header.type = PERF_RECORD_LOST;
4132 lost_event.header.misc = 0;
4133 lost_event.id = event->id;
4134 lost_event.lost = local_xchg(&buffer->lost, 0);
4136 perf_output_put(handle, lost_event);
4137 perf_event__output_id_sample(event, handle, &sample_data);
4143 local_inc(&buffer->lost);
4144 perf_output_put_handle(handle);
4151 void perf_output_end(struct perf_output_handle *handle)
4153 struct perf_event *event = handle->event;
4154 struct perf_buffer *buffer = handle->buffer;
4156 int wakeup_events = event->attr.wakeup_events;
4158 if (handle->sample && wakeup_events) {
4159 int events = local_inc_return(&buffer->events);
4160 if (events >= wakeup_events) {
4161 local_sub(wakeup_events, &buffer->events);
4162 local_inc(&buffer->wakeup);
4166 perf_output_put_handle(handle);
4170 static void perf_output_read_one(struct perf_output_handle *handle,
4171 struct perf_event *event,
4172 u64 enabled, u64 running)
4174 u64 read_format = event->attr.read_format;
4178 values[n++] = perf_event_count(event);
4179 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
4180 values[n++] = enabled +
4181 atomic64_read(&event->child_total_time_enabled);
4183 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
4184 values[n++] = running +
4185 atomic64_read(&event->child_total_time_running);
4187 if (read_format & PERF_FORMAT_ID)
4188 values[n++] = primary_event_id(event);
4190 perf_output_copy(handle, values, n * sizeof(u64));
4194 * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
4196 static void perf_output_read_group(struct perf_output_handle *handle,
4197 struct perf_event *event,
4198 u64 enabled, u64 running)
4200 struct perf_event *leader = event->group_leader, *sub;
4201 u64 read_format = event->attr.read_format;
4205 values[n++] = 1 + leader->nr_siblings;
4207 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
4208 values[n++] = enabled;
4210 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
4211 values[n++] = running;
4213 if (leader != event)
4214 leader->pmu->read(leader);
4216 values[n++] = perf_event_count(leader);
4217 if (read_format & PERF_FORMAT_ID)
4218 values[n++] = primary_event_id(leader);
4220 perf_output_copy(handle, values, n * sizeof(u64));
4222 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
4226 sub->pmu->read(sub);
4228 values[n++] = perf_event_count(sub);
4229 if (read_format & PERF_FORMAT_ID)
4230 values[n++] = primary_event_id(sub);
4232 perf_output_copy(handle, values, n * sizeof(u64));
4236 #define PERF_FORMAT_TOTAL_TIMES (PERF_FORMAT_TOTAL_TIME_ENABLED|\
4237 PERF_FORMAT_TOTAL_TIME_RUNNING)
4239 static void perf_output_read(struct perf_output_handle *handle,
4240 struct perf_event *event)
4242 u64 enabled = 0, running = 0, now, ctx_time;
4243 u64 read_format = event->attr.read_format;
4246 * compute total_time_enabled, total_time_running
4247 * based on snapshot values taken when the event
4248 * was last scheduled in.
4250 * we cannot simply called update_context_time()
4251 * because of locking issue as we are called in
4254 if (read_format & PERF_FORMAT_TOTAL_TIMES) {
4256 ctx_time = event->shadow_ctx_time + now;
4257 enabled = ctx_time - event->tstamp_enabled;
4258 running = ctx_time - event->tstamp_running;
4261 if (event->attr.read_format & PERF_FORMAT_GROUP)
4262 perf_output_read_group(handle, event, enabled, running);
4264 perf_output_read_one(handle, event, enabled, running);
4267 void perf_output_sample(struct perf_output_handle *handle,
4268 struct perf_event_header *header,
4269 struct perf_sample_data *data,
4270 struct perf_event *event)
4272 u64 sample_type = data->type;
4274 perf_output_put(handle, *header);
4276 if (sample_type & PERF_SAMPLE_IP)
4277 perf_output_put(handle, data->ip);
4279 if (sample_type & PERF_SAMPLE_TID)
4280 perf_output_put(handle, data->tid_entry);
4282 if (sample_type & PERF_SAMPLE_TIME)
4283 perf_output_put(handle, data->time);
4285 if (sample_type & PERF_SAMPLE_ADDR)
4286 perf_output_put(handle, data->addr);
4288 if (sample_type & PERF_SAMPLE_ID)
4289 perf_output_put(handle, data->id);
4291 if (sample_type & PERF_SAMPLE_STREAM_ID)
4292 perf_output_put(handle, data->stream_id);
4294 if (sample_type & PERF_SAMPLE_CPU)
4295 perf_output_put(handle, data->cpu_entry);
4297 if (sample_type & PERF_SAMPLE_PERIOD)
4298 perf_output_put(handle, data->period);
4300 if (sample_type & PERF_SAMPLE_READ)
4301 perf_output_read(handle, event);
4303 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
4304 if (data->callchain) {
4307 if (data->callchain)
4308 size += data->callchain->nr;
4310 size *= sizeof(u64);
4312 perf_output_copy(handle, data->callchain, size);
4315 perf_output_put(handle, nr);
4319 if (sample_type & PERF_SAMPLE_RAW) {
4321 perf_output_put(handle, data->raw->size);
4322 perf_output_copy(handle, data->raw->data,
4329 .size = sizeof(u32),
4332 perf_output_put(handle, raw);
4337 void perf_prepare_sample(struct perf_event_header *header,
4338 struct perf_sample_data *data,
4339 struct perf_event *event,
4340 struct pt_regs *regs)
4342 u64 sample_type = event->attr.sample_type;
4344 header->type = PERF_RECORD_SAMPLE;
4345 header->size = sizeof(*header) + event->header_size;
4348 header->misc |= perf_misc_flags(regs);
4350 __perf_event_header__init_id(header, data, event);
4352 if (sample_type & PERF_SAMPLE_IP)
4353 data->ip = perf_instruction_pointer(regs);
4355 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
4358 data->callchain = perf_callchain(regs);
4360 if (data->callchain)
4361 size += data->callchain->nr;
4363 header->size += size * sizeof(u64);
4366 if (sample_type & PERF_SAMPLE_RAW) {
4367 int size = sizeof(u32);
4370 size += data->raw->size;
4372 size += sizeof(u32);
4374 WARN_ON_ONCE(size & (sizeof(u64)-1));
4375 header->size += size;
4379 static void perf_event_output(struct perf_event *event, int nmi,
4380 struct perf_sample_data *data,
4381 struct pt_regs *regs)
4383 struct perf_output_handle handle;
4384 struct perf_event_header header;
4386 /* protect the callchain buffers */
4389 perf_prepare_sample(&header, data, event, regs);
4391 if (perf_output_begin(&handle, event, header.size, nmi, 1))
4394 perf_output_sample(&handle, &header, data, event);
4396 perf_output_end(&handle);
4406 struct perf_read_event {
4407 struct perf_event_header header;
4414 perf_event_read_event(struct perf_event *event,
4415 struct task_struct *task)
4417 struct perf_output_handle handle;
4418 struct perf_sample_data sample;
4419 struct perf_read_event read_event = {
4421 .type = PERF_RECORD_READ,
4423 .size = sizeof(read_event) + event->read_size,
4425 .pid = perf_event_pid(event, task),
4426 .tid = perf_event_tid(event, task),
4430 perf_event_header__init_id(&read_event.header, &sample, event);
4431 ret = perf_output_begin(&handle, event, read_event.header.size, 0, 0);
4435 perf_output_put(&handle, read_event);
4436 perf_output_read(&handle, event);
4437 perf_event__output_id_sample(event, &handle, &sample);
4439 perf_output_end(&handle);
4443 * task tracking -- fork/exit
4445 * enabled by: attr.comm | attr.mmap | attr.mmap_data | attr.task
4448 struct perf_task_event {
4449 struct task_struct *task;
4450 struct perf_event_context *task_ctx;
4453 struct perf_event_header header;
4463 static void perf_event_task_output(struct perf_event *event,
4464 struct perf_task_event *task_event)
4466 struct perf_output_handle handle;
4467 struct perf_sample_data sample;
4468 struct task_struct *task = task_event->task;
4469 int ret, size = task_event->event_id.header.size;
4471 perf_event_header__init_id(&task_event->event_id.header, &sample, event);
4473 ret = perf_output_begin(&handle, event,
4474 task_event->event_id.header.size, 0, 0);
4478 task_event->event_id.pid = perf_event_pid(event, task);
4479 task_event->event_id.ppid = perf_event_pid(event, current);
4481 task_event->event_id.tid = perf_event_tid(event, task);
4482 task_event->event_id.ptid = perf_event_tid(event, current);
4484 perf_output_put(&handle, task_event->event_id);
4486 perf_event__output_id_sample(event, &handle, &sample);
4488 perf_output_end(&handle);
4490 task_event->event_id.header.size = size;
4493 static int perf_event_task_match(struct perf_event *event)
4495 if (event->state < PERF_EVENT_STATE_INACTIVE)
4498 if (!event_filter_match(event))
4501 if (event->attr.comm || event->attr.mmap ||
4502 event->attr.mmap_data || event->attr.task)
4508 static void perf_event_task_ctx(struct perf_event_context *ctx,
4509 struct perf_task_event *task_event)
4511 struct perf_event *event;
4513 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
4514 if (perf_event_task_match(event))
4515 perf_event_task_output(event, task_event);
4519 static void perf_event_task_event(struct perf_task_event *task_event)
4521 struct perf_cpu_context *cpuctx;
4522 struct perf_event_context *ctx;
4527 list_for_each_entry_rcu(pmu, &pmus, entry) {
4528 cpuctx = get_cpu_ptr(pmu->pmu_cpu_context);
4529 if (cpuctx->active_pmu != pmu)
4531 perf_event_task_ctx(&cpuctx->ctx, task_event);
4533 ctx = task_event->task_ctx;
4535 ctxn = pmu->task_ctx_nr;
4538 ctx = rcu_dereference(current->perf_event_ctxp[ctxn]);
4541 perf_event_task_ctx(ctx, task_event);
4543 put_cpu_ptr(pmu->pmu_cpu_context);
4548 static void perf_event_task(struct task_struct *task,
4549 struct perf_event_context *task_ctx,
4552 struct perf_task_event task_event;
4554 if (!atomic_read(&nr_comm_events) &&
4555 !atomic_read(&nr_mmap_events) &&
4556 !atomic_read(&nr_task_events))
4559 task_event = (struct perf_task_event){
4561 .task_ctx = task_ctx,
4564 .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT,
4566 .size = sizeof(task_event.event_id),
4572 .time = perf_clock(),
4576 perf_event_task_event(&task_event);
4579 void perf_event_fork(struct task_struct *task)
4581 perf_event_task(task, NULL, 1);
4588 struct perf_comm_event {
4589 struct task_struct *task;
4594 struct perf_event_header header;
4601 static void perf_event_comm_output(struct perf_event *event,
4602 struct perf_comm_event *comm_event)
4604 struct perf_output_handle handle;
4605 struct perf_sample_data sample;
4606 int size = comm_event->event_id.header.size;
4609 perf_event_header__init_id(&comm_event->event_id.header, &sample, event);
4610 ret = perf_output_begin(&handle, event,
4611 comm_event->event_id.header.size, 0, 0);
4616 comm_event->event_id.pid = perf_event_pid(event, comm_event->task);
4617 comm_event->event_id.tid = perf_event_tid(event, comm_event->task);
4619 perf_output_put(&handle, comm_event->event_id);
4620 perf_output_copy(&handle, comm_event->comm,
4621 comm_event->comm_size);
4623 perf_event__output_id_sample(event, &handle, &sample);
4625 perf_output_end(&handle);
4627 comm_event->event_id.header.size = size;
4630 static int perf_event_comm_match(struct perf_event *event)
4632 if (event->state < PERF_EVENT_STATE_INACTIVE)
4635 if (!event_filter_match(event))
4638 if (event->attr.comm)
4644 static void perf_event_comm_ctx(struct perf_event_context *ctx,
4645 struct perf_comm_event *comm_event)
4647 struct perf_event *event;
4649 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
4650 if (perf_event_comm_match(event))
4651 perf_event_comm_output(event, comm_event);
4655 static void perf_event_comm_event(struct perf_comm_event *comm_event)
4657 struct perf_cpu_context *cpuctx;
4658 struct perf_event_context *ctx;
4659 char comm[TASK_COMM_LEN];
4664 memset(comm, 0, sizeof(comm));
4665 strlcpy(comm, comm_event->task->comm, sizeof(comm));
4666 size = ALIGN(strlen(comm)+1, sizeof(u64));
4668 comm_event->comm = comm;
4669 comm_event->comm_size = size;
4671 comm_event->event_id.header.size = sizeof(comm_event->event_id) + size;
4673 list_for_each_entry_rcu(pmu, &pmus, entry) {
4674 cpuctx = get_cpu_ptr(pmu->pmu_cpu_context);
4675 if (cpuctx->active_pmu != pmu)
4677 perf_event_comm_ctx(&cpuctx->ctx, comm_event);
4679 ctxn = pmu->task_ctx_nr;
4683 ctx = rcu_dereference(current->perf_event_ctxp[ctxn]);
4685 perf_event_comm_ctx(ctx, comm_event);
4687 put_cpu_ptr(pmu->pmu_cpu_context);
4692 void perf_event_comm(struct task_struct *task)
4694 struct perf_comm_event comm_event;
4695 struct perf_event_context *ctx;
4698 for_each_task_context_nr(ctxn) {
4699 ctx = task->perf_event_ctxp[ctxn];
4703 perf_event_enable_on_exec(ctx);
4706 if (!atomic_read(&nr_comm_events))
4709 comm_event = (struct perf_comm_event){
4715 .type = PERF_RECORD_COMM,
4724 perf_event_comm_event(&comm_event);
4731 struct perf_mmap_event {
4732 struct vm_area_struct *vma;
4734 const char *file_name;
4738 struct perf_event_header header;
4748 static void perf_event_mmap_output(struct perf_event *event,
4749 struct perf_mmap_event *mmap_event)
4751 struct perf_output_handle handle;
4752 struct perf_sample_data sample;
4753 int size = mmap_event->event_id.header.size;
4756 perf_event_header__init_id(&mmap_event->event_id.header, &sample, event);
4757 ret = perf_output_begin(&handle, event,
4758 mmap_event->event_id.header.size, 0, 0);
4762 mmap_event->event_id.pid = perf_event_pid(event, current);
4763 mmap_event->event_id.tid = perf_event_tid(event, current);
4765 perf_output_put(&handle, mmap_event->event_id);
4766 perf_output_copy(&handle, mmap_event->file_name,
4767 mmap_event->file_size);
4769 perf_event__output_id_sample(event, &handle, &sample);
4771 perf_output_end(&handle);
4773 mmap_event->event_id.header.size = size;
4776 static int perf_event_mmap_match(struct perf_event *event,
4777 struct perf_mmap_event *mmap_event,
4780 if (event->state < PERF_EVENT_STATE_INACTIVE)
4783 if (!event_filter_match(event))
4786 if ((!executable && event->attr.mmap_data) ||
4787 (executable && event->attr.mmap))
4793 static void perf_event_mmap_ctx(struct perf_event_context *ctx,
4794 struct perf_mmap_event *mmap_event,
4797 struct perf_event *event;
4799 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
4800 if (perf_event_mmap_match(event, mmap_event, executable))
4801 perf_event_mmap_output(event, mmap_event);
4805 static void perf_event_mmap_event(struct perf_mmap_event *mmap_event)
4807 struct perf_cpu_context *cpuctx;
4808 struct perf_event_context *ctx;
4809 struct vm_area_struct *vma = mmap_event->vma;
4810 struct file *file = vma->vm_file;
4818 memset(tmp, 0, sizeof(tmp));
4822 * d_path works from the end of the buffer backwards, so we
4823 * need to add enough zero bytes after the string to handle
4824 * the 64bit alignment we do later.
4826 buf = kzalloc(PATH_MAX + sizeof(u64), GFP_KERNEL);
4828 name = strncpy(tmp, "//enomem", sizeof(tmp));
4831 name = d_path(&file->f_path, buf, PATH_MAX);
4833 name = strncpy(tmp, "//toolong", sizeof(tmp));
4837 if (arch_vma_name(mmap_event->vma)) {
4838 name = strncpy(tmp, arch_vma_name(mmap_event->vma),
4844 name = strncpy(tmp, "[vdso]", sizeof(tmp));
4846 } else if (vma->vm_start <= vma->vm_mm->start_brk &&
4847 vma->vm_end >= vma->vm_mm->brk) {
4848 name = strncpy(tmp, "[heap]", sizeof(tmp));
4850 } else if (vma->vm_start <= vma->vm_mm->start_stack &&
4851 vma->vm_end >= vma->vm_mm->start_stack) {
4852 name = strncpy(tmp, "[stack]", sizeof(tmp));
4856 name = strncpy(tmp, "//anon", sizeof(tmp));
4861 size = ALIGN(strlen(name)+1, sizeof(u64));
4863 mmap_event->file_name = name;
4864 mmap_event->file_size = size;
4866 mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size;
4869 list_for_each_entry_rcu(pmu, &pmus, entry) {
4870 cpuctx = get_cpu_ptr(pmu->pmu_cpu_context);
4871 if (cpuctx->active_pmu != pmu)
4873 perf_event_mmap_ctx(&cpuctx->ctx, mmap_event,
4874 vma->vm_flags & VM_EXEC);
4876 ctxn = pmu->task_ctx_nr;
4880 ctx = rcu_dereference(current->perf_event_ctxp[ctxn]);
4882 perf_event_mmap_ctx(ctx, mmap_event,
4883 vma->vm_flags & VM_EXEC);
4886 put_cpu_ptr(pmu->pmu_cpu_context);
4893 void perf_event_mmap(struct vm_area_struct *vma)
4895 struct perf_mmap_event mmap_event;
4897 if (!atomic_read(&nr_mmap_events))
4900 mmap_event = (struct perf_mmap_event){
4906 .type = PERF_RECORD_MMAP,
4907 .misc = PERF_RECORD_MISC_USER,
4912 .start = vma->vm_start,
4913 .len = vma->vm_end - vma->vm_start,
4914 .pgoff = (u64)vma->vm_pgoff << PAGE_SHIFT,
4918 perf_event_mmap_event(&mmap_event);
4922 * IRQ throttle logging
4925 static void perf_log_throttle(struct perf_event *event, int enable)
4927 struct perf_output_handle handle;
4928 struct perf_sample_data sample;
4932 struct perf_event_header header;
4936 } throttle_event = {
4938 .type = PERF_RECORD_THROTTLE,
4940 .size = sizeof(throttle_event),
4942 .time = perf_clock(),
4943 .id = primary_event_id(event),
4944 .stream_id = event->id,
4948 throttle_event.header.type = PERF_RECORD_UNTHROTTLE;
4950 perf_event_header__init_id(&throttle_event.header, &sample, event);
4952 ret = perf_output_begin(&handle, event,
4953 throttle_event.header.size, 1, 0);
4957 perf_output_put(&handle, throttle_event);
4958 perf_event__output_id_sample(event, &handle, &sample);
4959 perf_output_end(&handle);
4963 * Generic event overflow handling, sampling.
4966 static int __perf_event_overflow(struct perf_event *event, int nmi,
4967 int throttle, struct perf_sample_data *data,
4968 struct pt_regs *regs)
4970 int events = atomic_read(&event->event_limit);
4971 struct hw_perf_event *hwc = &event->hw;
4975 * Non-sampling counters might still use the PMI to fold short
4976 * hardware counters, ignore those.
4978 if (unlikely(!is_sampling_event(event)))
4981 if (unlikely(hwc->interrupts >= max_samples_per_tick)) {
4983 hwc->interrupts = MAX_INTERRUPTS;
4984 perf_log_throttle(event, 0);
4990 if (event->attr.freq) {
4991 u64 now = perf_clock();
4992 s64 delta = now - hwc->freq_time_stamp;
4994 hwc->freq_time_stamp = now;
4996 if (delta > 0 && delta < 2*TICK_NSEC)
4997 perf_adjust_period(event, delta, hwc->last_period);
5001 * XXX event_limit might not quite work as expected on inherited
5005 event->pending_kill = POLL_IN;
5006 if (events && atomic_dec_and_test(&event->event_limit)) {
5008 event->pending_kill = POLL_HUP;
5010 event->pending_disable = 1;
5011 irq_work_queue(&event->pending);
5013 perf_event_disable(event);
5016 if (event->overflow_handler)
5017 event->overflow_handler(event, nmi, data, regs);
5019 perf_event_output(event, nmi, data, regs);
5021 if (event->fasync && event->pending_kill) {
5023 event->pending_wakeup = 1;
5024 irq_work_queue(&event->pending);
5026 perf_event_wakeup(event);
5032 int perf_event_overflow(struct perf_event *event, int nmi,
5033 struct perf_sample_data *data,
5034 struct pt_regs *regs)
5036 return __perf_event_overflow(event, nmi, 1, data, regs);
5040 * Generic software event infrastructure
5043 struct swevent_htable {
5044 struct swevent_hlist *swevent_hlist;
5045 struct mutex hlist_mutex;
5048 /* Recursion avoidance in each contexts */
5049 int recursion[PERF_NR_CONTEXTS];
5052 static DEFINE_PER_CPU(struct swevent_htable, swevent_htable);
5055 * We directly increment event->count and keep a second value in
5056 * event->hw.period_left to count intervals. This period event
5057 * is kept in the range [-sample_period, 0] so that we can use the
5061 static u64 perf_swevent_set_period(struct perf_event *event)
5063 struct hw_perf_event *hwc = &event->hw;
5064 u64 period = hwc->last_period;
5068 hwc->last_period = hwc->sample_period;
5071 old = val = local64_read(&hwc->period_left);
5075 nr = div64_u64(period + val, period);
5076 offset = nr * period;
5078 if (local64_cmpxchg(&hwc->period_left, old, val) != old)
5084 static void perf_swevent_overflow(struct perf_event *event, u64 overflow,
5085 int nmi, struct perf_sample_data *data,
5086 struct pt_regs *regs)
5088 struct hw_perf_event *hwc = &event->hw;
5091 data->period = event->hw.last_period;
5093 overflow = perf_swevent_set_period(event);
5095 if (hwc->interrupts == MAX_INTERRUPTS)
5098 for (; overflow; overflow--) {
5099 if (__perf_event_overflow(event, nmi, throttle,
5102 * We inhibit the overflow from happening when
5103 * hwc->interrupts == MAX_INTERRUPTS.
5111 static void perf_swevent_event(struct perf_event *event, u64 nr,
5112 int nmi, struct perf_sample_data *data,
5113 struct pt_regs *regs)
5115 struct hw_perf_event *hwc = &event->hw;
5117 local64_add(nr, &event->count);
5122 if (!is_sampling_event(event))
5125 if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq)
5126 return perf_swevent_overflow(event, 1, nmi, data, regs);
5128 if (local64_add_negative(nr, &hwc->period_left))
5131 perf_swevent_overflow(event, 0, nmi, data, regs);
5134 static int perf_exclude_event(struct perf_event *event,
5135 struct pt_regs *regs)
5137 if (event->hw.state & PERF_HES_STOPPED)
5141 if (event->attr.exclude_user && user_mode(regs))
5144 if (event->attr.exclude_kernel && !user_mode(regs))
5151 static int perf_swevent_match(struct perf_event *event,
5152 enum perf_type_id type,
5154 struct perf_sample_data *data,
5155 struct pt_regs *regs)
5157 if (event->attr.type != type)
5160 if (event->attr.config != event_id)
5163 if (perf_exclude_event(event, regs))
5169 static inline u64 swevent_hash(u64 type, u32 event_id)
5171 u64 val = event_id | (type << 32);
5173 return hash_64(val, SWEVENT_HLIST_BITS);
5176 static inline struct hlist_head *
5177 __find_swevent_head(struct swevent_hlist *hlist, u64 type, u32 event_id)
5179 u64 hash = swevent_hash(type, event_id);
5181 return &hlist->heads[hash];
5184 /* For the read side: events when they trigger */
5185 static inline struct hlist_head *
5186 find_swevent_head_rcu(struct swevent_htable *swhash, u64 type, u32 event_id)
5188 struct swevent_hlist *hlist;
5190 hlist = rcu_dereference(swhash->swevent_hlist);
5194 return __find_swevent_head(hlist, type, event_id);
5197 /* For the event head insertion and removal in the hlist */
5198 static inline struct hlist_head *
5199 find_swevent_head(struct swevent_htable *swhash, struct perf_event *event)
5201 struct swevent_hlist *hlist;
5202 u32 event_id = event->attr.config;
5203 u64 type = event->attr.type;
5206 * Event scheduling is always serialized against hlist allocation
5207 * and release. Which makes the protected version suitable here.
5208 * The context lock guarantees that.
5210 hlist = rcu_dereference_protected(swhash->swevent_hlist,
5211 lockdep_is_held(&event->ctx->lock));
5215 return __find_swevent_head(hlist, type, event_id);
5218 static void do_perf_sw_event(enum perf_type_id type, u32 event_id,
5220 struct perf_sample_data *data,
5221 struct pt_regs *regs)
5223 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
5224 struct perf_event *event;
5225 struct hlist_node *node;
5226 struct hlist_head *head;
5229 head = find_swevent_head_rcu(swhash, type, event_id);
5233 hlist_for_each_entry_rcu(event, node, head, hlist_entry) {
5234 if (perf_swevent_match(event, type, event_id, data, regs))
5235 perf_swevent_event(event, nr, nmi, data, regs);
5241 int perf_swevent_get_recursion_context(void)
5243 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
5245 return get_recursion_context(swhash->recursion);
5247 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context);
5249 inline void perf_swevent_put_recursion_context(int rctx)
5251 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
5253 put_recursion_context(swhash->recursion, rctx);
5256 void __perf_sw_event(u32 event_id, u64 nr, int nmi,
5257 struct pt_regs *regs, u64 addr)
5259 struct perf_sample_data data;
5262 preempt_disable_notrace();
5263 rctx = perf_swevent_get_recursion_context();
5267 perf_sample_data_init(&data, addr);
5269 do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, nmi, &data, regs);
5271 perf_swevent_put_recursion_context(rctx);
5272 preempt_enable_notrace();
5275 static void perf_swevent_read(struct perf_event *event)
5279 static int perf_swevent_add(struct perf_event *event, int flags)
5281 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
5282 struct hw_perf_event *hwc = &event->hw;
5283 struct hlist_head *head;
5285 if (is_sampling_event(event)) {
5286 hwc->last_period = hwc->sample_period;
5287 perf_swevent_set_period(event);
5290 hwc->state = !(flags & PERF_EF_START);
5292 head = find_swevent_head(swhash, event);
5293 if (WARN_ON_ONCE(!head))
5296 hlist_add_head_rcu(&event->hlist_entry, head);
5301 static void perf_swevent_del(struct perf_event *event, int flags)
5303 hlist_del_rcu(&event->hlist_entry);
5306 static void perf_swevent_start(struct perf_event *event, int flags)
5308 event->hw.state = 0;
5311 static void perf_swevent_stop(struct perf_event *event, int flags)
5313 event->hw.state = PERF_HES_STOPPED;
5316 /* Deref the hlist from the update side */
5317 static inline struct swevent_hlist *
5318 swevent_hlist_deref(struct swevent_htable *swhash)
5320 return rcu_dereference_protected(swhash->swevent_hlist,
5321 lockdep_is_held(&swhash->hlist_mutex));
5324 static void swevent_hlist_release(struct swevent_htable *swhash)
5326 struct swevent_hlist *hlist = swevent_hlist_deref(swhash);
5331 rcu_assign_pointer(swhash->swevent_hlist, NULL);
5332 kfree_rcu(hlist, rcu_head);
5335 static void swevent_hlist_put_cpu(struct perf_event *event, int cpu)
5337 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
5339 mutex_lock(&swhash->hlist_mutex);
5341 if (!--swhash->hlist_refcount)
5342 swevent_hlist_release(swhash);
5344 mutex_unlock(&swhash->hlist_mutex);
5347 static void swevent_hlist_put(struct perf_event *event)
5351 if (event->cpu != -1) {
5352 swevent_hlist_put_cpu(event, event->cpu);
5356 for_each_possible_cpu(cpu)
5357 swevent_hlist_put_cpu(event, cpu);
5360 static int swevent_hlist_get_cpu(struct perf_event *event, int cpu)
5362 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
5365 mutex_lock(&swhash->hlist_mutex);
5367 if (!swevent_hlist_deref(swhash) && cpu_online(cpu)) {
5368 struct swevent_hlist *hlist;
5370 hlist = kzalloc(sizeof(*hlist), GFP_KERNEL);
5375 rcu_assign_pointer(swhash->swevent_hlist, hlist);
5377 swhash->hlist_refcount++;
5379 mutex_unlock(&swhash->hlist_mutex);
5384 static int swevent_hlist_get(struct perf_event *event)
5387 int cpu, failed_cpu;
5389 if (event->cpu != -1)
5390 return swevent_hlist_get_cpu(event, event->cpu);
5393 for_each_possible_cpu(cpu) {
5394 err = swevent_hlist_get_cpu(event, cpu);
5404 for_each_possible_cpu(cpu) {
5405 if (cpu == failed_cpu)
5407 swevent_hlist_put_cpu(event, cpu);
5414 struct jump_label_key perf_swevent_enabled[PERF_COUNT_SW_MAX];
5416 static void sw_perf_event_destroy(struct perf_event *event)
5418 u64 event_id = event->attr.config;
5420 WARN_ON(event->parent);
5422 jump_label_dec(&perf_swevent_enabled[event_id]);
5423 swevent_hlist_put(event);
5426 static int perf_swevent_init(struct perf_event *event)
5428 int event_id = event->attr.config;
5430 if (event->attr.type != PERF_TYPE_SOFTWARE)
5434 case PERF_COUNT_SW_CPU_CLOCK:
5435 case PERF_COUNT_SW_TASK_CLOCK:
5442 if (event_id >= PERF_COUNT_SW_MAX)
5445 if (!event->parent) {
5448 err = swevent_hlist_get(event);
5452 jump_label_inc(&perf_swevent_enabled[event_id]);
5453 event->destroy = sw_perf_event_destroy;
5459 static struct pmu perf_swevent = {
5460 .task_ctx_nr = perf_sw_context,
5462 .event_init = perf_swevent_init,
5463 .add = perf_swevent_add,
5464 .del = perf_swevent_del,
5465 .start = perf_swevent_start,
5466 .stop = perf_swevent_stop,
5467 .read = perf_swevent_read,
5470 #ifdef CONFIG_EVENT_TRACING
5472 static int perf_tp_filter_match(struct perf_event *event,
5473 struct perf_sample_data *data)
5475 void *record = data->raw->data;
5477 if (likely(!event->filter) || filter_match_preds(event->filter, record))
5482 static int perf_tp_event_match(struct perf_event *event,
5483 struct perf_sample_data *data,
5484 struct pt_regs *regs)
5486 if (event->hw.state & PERF_HES_STOPPED)
5489 * All tracepoints are from kernel-space.
5491 if (event->attr.exclude_kernel)
5494 if (!perf_tp_filter_match(event, data))
5500 void perf_tp_event(u64 addr, u64 count, void *record, int entry_size,
5501 struct pt_regs *regs, struct hlist_head *head, int rctx)
5503 struct perf_sample_data data;
5504 struct perf_event *event;
5505 struct hlist_node *node;
5507 struct perf_raw_record raw = {
5512 perf_sample_data_init(&data, addr);
5515 hlist_for_each_entry_rcu(event, node, head, hlist_entry) {
5516 if (perf_tp_event_match(event, &data, regs))
5517 perf_swevent_event(event, count, 1, &data, regs);
5520 perf_swevent_put_recursion_context(rctx);
5522 EXPORT_SYMBOL_GPL(perf_tp_event);
5524 static void tp_perf_event_destroy(struct perf_event *event)
5526 perf_trace_destroy(event);
5529 static int perf_tp_event_init(struct perf_event *event)
5533 if (event->attr.type != PERF_TYPE_TRACEPOINT)
5536 err = perf_trace_init(event);
5540 event->destroy = tp_perf_event_destroy;
5545 static struct pmu perf_tracepoint = {
5546 .task_ctx_nr = perf_sw_context,
5548 .event_init = perf_tp_event_init,
5549 .add = perf_trace_add,
5550 .del = perf_trace_del,
5551 .start = perf_swevent_start,
5552 .stop = perf_swevent_stop,
5553 .read = perf_swevent_read,
5556 static inline void perf_tp_register(void)
5558 perf_pmu_register(&perf_tracepoint, "tracepoint", PERF_TYPE_TRACEPOINT);
5561 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
5566 if (event->attr.type != PERF_TYPE_TRACEPOINT)
5569 filter_str = strndup_user(arg, PAGE_SIZE);
5570 if (IS_ERR(filter_str))
5571 return PTR_ERR(filter_str);
5573 ret = ftrace_profile_set_filter(event, event->attr.config, filter_str);
5579 static void perf_event_free_filter(struct perf_event *event)
5581 ftrace_profile_free_filter(event);
5586 static inline void perf_tp_register(void)
5590 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
5595 static void perf_event_free_filter(struct perf_event *event)
5599 #endif /* CONFIG_EVENT_TRACING */
5601 #ifdef CONFIG_HAVE_HW_BREAKPOINT
5602 void perf_bp_event(struct perf_event *bp, void *data)
5604 struct perf_sample_data sample;
5605 struct pt_regs *regs = data;
5607 perf_sample_data_init(&sample, bp->attr.bp_addr);
5609 if (!bp->hw.state && !perf_exclude_event(bp, regs))
5610 perf_swevent_event(bp, 1, 1, &sample, regs);
5615 * hrtimer based swevent callback
5618 static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer)
5620 enum hrtimer_restart ret = HRTIMER_RESTART;
5621 struct perf_sample_data data;
5622 struct pt_regs *regs;
5623 struct perf_event *event;
5626 event = container_of(hrtimer, struct perf_event, hw.hrtimer);
5628 if (event->state != PERF_EVENT_STATE_ACTIVE)
5629 return HRTIMER_NORESTART;
5631 event->pmu->read(event);
5633 perf_sample_data_init(&data, 0);
5634 data.period = event->hw.last_period;
5635 regs = get_irq_regs();
5637 if (regs && !perf_exclude_event(event, regs)) {
5638 if (!(event->attr.exclude_idle && current->pid == 0))
5639 if (perf_event_overflow(event, 0, &data, regs))
5640 ret = HRTIMER_NORESTART;
5643 period = max_t(u64, 10000, event->hw.sample_period);
5644 hrtimer_forward_now(hrtimer, ns_to_ktime(period));
5649 static void perf_swevent_start_hrtimer(struct perf_event *event)
5651 struct hw_perf_event *hwc = &event->hw;
5654 if (!is_sampling_event(event))
5657 period = local64_read(&hwc->period_left);
5662 local64_set(&hwc->period_left, 0);
5664 period = max_t(u64, 10000, hwc->sample_period);
5666 __hrtimer_start_range_ns(&hwc->hrtimer,
5667 ns_to_ktime(period), 0,
5668 HRTIMER_MODE_REL_PINNED, 0);
5671 static void perf_swevent_cancel_hrtimer(struct perf_event *event)
5673 struct hw_perf_event *hwc = &event->hw;
5675 if (is_sampling_event(event)) {
5676 ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer);
5677 local64_set(&hwc->period_left, ktime_to_ns(remaining));
5679 hrtimer_cancel(&hwc->hrtimer);
5683 static void perf_swevent_init_hrtimer(struct perf_event *event)
5685 struct hw_perf_event *hwc = &event->hw;
5687 if (!is_sampling_event(event))
5690 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
5691 hwc->hrtimer.function = perf_swevent_hrtimer;
5694 * Since hrtimers have a fixed rate, we can do a static freq->period
5695 * mapping and avoid the whole period adjust feedback stuff.
5697 if (event->attr.freq) {
5698 long freq = event->attr.sample_freq;
5700 event->attr.sample_period = NSEC_PER_SEC / freq;
5701 hwc->sample_period = event->attr.sample_period;
5702 local64_set(&hwc->period_left, hwc->sample_period);
5703 event->attr.freq = 0;
5708 * Software event: cpu wall time clock
5711 static void cpu_clock_event_update(struct perf_event *event)
5716 now = local_clock();
5717 prev = local64_xchg(&event->hw.prev_count, now);
5718 local64_add(now - prev, &event->count);
5721 static void cpu_clock_event_start(struct perf_event *event, int flags)
5723 local64_set(&event->hw.prev_count, local_clock());
5724 perf_swevent_start_hrtimer(event);
5727 static void cpu_clock_event_stop(struct perf_event *event, int flags)
5729 perf_swevent_cancel_hrtimer(event);
5730 cpu_clock_event_update(event);
5733 static int cpu_clock_event_add(struct perf_event *event, int flags)
5735 if (flags & PERF_EF_START)
5736 cpu_clock_event_start(event, flags);
5741 static void cpu_clock_event_del(struct perf_event *event, int flags)
5743 cpu_clock_event_stop(event, flags);
5746 static void cpu_clock_event_read(struct perf_event *event)
5748 cpu_clock_event_update(event);
5751 static int cpu_clock_event_init(struct perf_event *event)
5753 if (event->attr.type != PERF_TYPE_SOFTWARE)
5756 if (event->attr.config != PERF_COUNT_SW_CPU_CLOCK)
5759 perf_swevent_init_hrtimer(event);
5764 static struct pmu perf_cpu_clock = {
5765 .task_ctx_nr = perf_sw_context,
5767 .event_init = cpu_clock_event_init,
5768 .add = cpu_clock_event_add,
5769 .del = cpu_clock_event_del,
5770 .start = cpu_clock_event_start,
5771 .stop = cpu_clock_event_stop,
5772 .read = cpu_clock_event_read,
5776 * Software event: task time clock
5779 static void task_clock_event_update(struct perf_event *event, u64 now)
5784 prev = local64_xchg(&event->hw.prev_count, now);
5786 local64_add(delta, &event->count);
5789 static void task_clock_event_start(struct perf_event *event, int flags)
5791 local64_set(&event->hw.prev_count, event->ctx->time);
5792 perf_swevent_start_hrtimer(event);
5795 static void task_clock_event_stop(struct perf_event *event, int flags)
5797 perf_swevent_cancel_hrtimer(event);
5798 task_clock_event_update(event, event->ctx->time);
5801 static int task_clock_event_add(struct perf_event *event, int flags)
5803 if (flags & PERF_EF_START)
5804 task_clock_event_start(event, flags);
5809 static void task_clock_event_del(struct perf_event *event, int flags)
5811 task_clock_event_stop(event, PERF_EF_UPDATE);
5814 static void task_clock_event_read(struct perf_event *event)
5816 u64 now = perf_clock();
5817 u64 delta = now - event->ctx->timestamp;
5818 u64 time = event->ctx->time + delta;
5820 task_clock_event_update(event, time);
5823 static int task_clock_event_init(struct perf_event *event)
5825 if (event->attr.type != PERF_TYPE_SOFTWARE)
5828 if (event->attr.config != PERF_COUNT_SW_TASK_CLOCK)
5831 perf_swevent_init_hrtimer(event);
5836 static struct pmu perf_task_clock = {
5837 .task_ctx_nr = perf_sw_context,
5839 .event_init = task_clock_event_init,
5840 .add = task_clock_event_add,
5841 .del = task_clock_event_del,
5842 .start = task_clock_event_start,
5843 .stop = task_clock_event_stop,
5844 .read = task_clock_event_read,
5847 static void perf_pmu_nop_void(struct pmu *pmu)
5851 static int perf_pmu_nop_int(struct pmu *pmu)
5856 static void perf_pmu_start_txn(struct pmu *pmu)
5858 perf_pmu_disable(pmu);
5861 static int perf_pmu_commit_txn(struct pmu *pmu)
5863 perf_pmu_enable(pmu);
5867 static void perf_pmu_cancel_txn(struct pmu *pmu)
5869 perf_pmu_enable(pmu);
5873 * Ensures all contexts with the same task_ctx_nr have the same
5874 * pmu_cpu_context too.
5876 static void *find_pmu_context(int ctxn)
5883 list_for_each_entry(pmu, &pmus, entry) {
5884 if (pmu->task_ctx_nr == ctxn)
5885 return pmu->pmu_cpu_context;
5891 static void update_pmu_context(struct pmu *pmu, struct pmu *old_pmu)
5895 for_each_possible_cpu(cpu) {
5896 struct perf_cpu_context *cpuctx;
5898 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
5900 if (cpuctx->active_pmu == old_pmu)
5901 cpuctx->active_pmu = pmu;
5905 static void free_pmu_context(struct pmu *pmu)
5909 mutex_lock(&pmus_lock);
5911 * Like a real lame refcount.
5913 list_for_each_entry(i, &pmus, entry) {
5914 if (i->pmu_cpu_context == pmu->pmu_cpu_context) {
5915 update_pmu_context(i, pmu);
5920 free_percpu(pmu->pmu_cpu_context);
5922 mutex_unlock(&pmus_lock);
5924 static struct idr pmu_idr;
5927 type_show(struct device *dev, struct device_attribute *attr, char *page)
5929 struct pmu *pmu = dev_get_drvdata(dev);
5931 return snprintf(page, PAGE_SIZE-1, "%d\n", pmu->type);
5934 static struct device_attribute pmu_dev_attrs[] = {
5939 static int pmu_bus_running;
5940 static struct bus_type pmu_bus = {
5941 .name = "event_source",
5942 .dev_attrs = pmu_dev_attrs,
5945 static void pmu_dev_release(struct device *dev)
5950 static int pmu_dev_alloc(struct pmu *pmu)
5954 pmu->dev = kzalloc(sizeof(struct device), GFP_KERNEL);
5958 device_initialize(pmu->dev);
5959 ret = dev_set_name(pmu->dev, "%s", pmu->name);
5963 dev_set_drvdata(pmu->dev, pmu);
5964 pmu->dev->bus = &pmu_bus;
5965 pmu->dev->release = pmu_dev_release;
5966 ret = device_add(pmu->dev);
5974 put_device(pmu->dev);
5978 static struct lock_class_key cpuctx_mutex;
5979 static struct lock_class_key cpuctx_lock;
5981 int perf_pmu_register(struct pmu *pmu, char *name, int type)
5985 mutex_lock(&pmus_lock);
5987 pmu->pmu_disable_count = alloc_percpu(int);
5988 if (!pmu->pmu_disable_count)
5997 int err = idr_pre_get(&pmu_idr, GFP_KERNEL);
6001 err = idr_get_new_above(&pmu_idr, pmu, PERF_TYPE_MAX, &type);
6009 if (pmu_bus_running) {
6010 ret = pmu_dev_alloc(pmu);
6016 pmu->pmu_cpu_context = find_pmu_context(pmu->task_ctx_nr);
6017 if (pmu->pmu_cpu_context)
6018 goto got_cpu_context;
6020 pmu->pmu_cpu_context = alloc_percpu(struct perf_cpu_context);
6021 if (!pmu->pmu_cpu_context)
6024 for_each_possible_cpu(cpu) {
6025 struct perf_cpu_context *cpuctx;
6027 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
6028 __perf_event_init_context(&cpuctx->ctx);
6029 lockdep_set_class(&cpuctx->ctx.mutex, &cpuctx_mutex);
6030 lockdep_set_class(&cpuctx->ctx.lock, &cpuctx_lock);
6031 cpuctx->ctx.type = cpu_context;
6032 cpuctx->ctx.pmu = pmu;
6033 cpuctx->jiffies_interval = 1;
6034 INIT_LIST_HEAD(&cpuctx->rotation_list);
6035 cpuctx->active_pmu = pmu;
6039 if (!pmu->start_txn) {
6040 if (pmu->pmu_enable) {
6042 * If we have pmu_enable/pmu_disable calls, install
6043 * transaction stubs that use that to try and batch
6044 * hardware accesses.
6046 pmu->start_txn = perf_pmu_start_txn;
6047 pmu->commit_txn = perf_pmu_commit_txn;
6048 pmu->cancel_txn = perf_pmu_cancel_txn;
6050 pmu->start_txn = perf_pmu_nop_void;
6051 pmu->commit_txn = perf_pmu_nop_int;
6052 pmu->cancel_txn = perf_pmu_nop_void;
6056 if (!pmu->pmu_enable) {
6057 pmu->pmu_enable = perf_pmu_nop_void;
6058 pmu->pmu_disable = perf_pmu_nop_void;
6061 list_add_rcu(&pmu->entry, &pmus);
6064 mutex_unlock(&pmus_lock);
6069 device_del(pmu->dev);
6070 put_device(pmu->dev);
6073 if (pmu->type >= PERF_TYPE_MAX)
6074 idr_remove(&pmu_idr, pmu->type);
6077 free_percpu(pmu->pmu_disable_count);
6081 void perf_pmu_unregister(struct pmu *pmu)
6083 mutex_lock(&pmus_lock);
6084 list_del_rcu(&pmu->entry);
6085 mutex_unlock(&pmus_lock);
6088 * We dereference the pmu list under both SRCU and regular RCU, so
6089 * synchronize against both of those.
6091 synchronize_srcu(&pmus_srcu);
6094 free_percpu(pmu->pmu_disable_count);
6095 if (pmu->type >= PERF_TYPE_MAX)
6096 idr_remove(&pmu_idr, pmu->type);
6097 device_del(pmu->dev);
6098 put_device(pmu->dev);
6099 free_pmu_context(pmu);
6102 struct pmu *perf_init_event(struct perf_event *event)
6104 struct pmu *pmu = NULL;
6108 idx = srcu_read_lock(&pmus_srcu);
6111 pmu = idr_find(&pmu_idr, event->attr.type);
6114 ret = pmu->event_init(event);
6120 list_for_each_entry_rcu(pmu, &pmus, entry) {
6121 ret = pmu->event_init(event);
6125 if (ret != -ENOENT) {
6130 pmu = ERR_PTR(-ENOENT);
6132 srcu_read_unlock(&pmus_srcu, idx);
6138 * Allocate and initialize a event structure
6140 static struct perf_event *
6141 perf_event_alloc(struct perf_event_attr *attr, int cpu,
6142 struct task_struct *task,
6143 struct perf_event *group_leader,
6144 struct perf_event *parent_event,
6145 perf_overflow_handler_t overflow_handler)
6148 struct perf_event *event;
6149 struct hw_perf_event *hwc;
6152 if ((unsigned)cpu >= nr_cpu_ids) {
6153 if (!task || cpu != -1)
6154 return ERR_PTR(-EINVAL);
6157 event = kzalloc(sizeof(*event), GFP_KERNEL);
6159 return ERR_PTR(-ENOMEM);
6162 * Single events are their own group leaders, with an
6163 * empty sibling list:
6166 group_leader = event;
6168 mutex_init(&event->child_mutex);
6169 INIT_LIST_HEAD(&event->child_list);
6171 INIT_LIST_HEAD(&event->group_entry);
6172 INIT_LIST_HEAD(&event->event_entry);
6173 INIT_LIST_HEAD(&event->sibling_list);
6174 init_waitqueue_head(&event->waitq);
6175 init_irq_work(&event->pending, perf_pending_event);
6177 mutex_init(&event->mmap_mutex);
6180 event->attr = *attr;
6181 event->group_leader = group_leader;
6185 event->parent = parent_event;
6187 event->ns = get_pid_ns(current->nsproxy->pid_ns);
6188 event->id = atomic64_inc_return(&perf_event_id);
6190 event->state = PERF_EVENT_STATE_INACTIVE;
6193 event->attach_state = PERF_ATTACH_TASK;
6194 #ifdef CONFIG_HAVE_HW_BREAKPOINT
6196 * hw_breakpoint is a bit difficult here..
6198 if (attr->type == PERF_TYPE_BREAKPOINT)
6199 event->hw.bp_target = task;
6203 if (!overflow_handler && parent_event)
6204 overflow_handler = parent_event->overflow_handler;
6206 event->overflow_handler = overflow_handler;
6209 event->state = PERF_EVENT_STATE_OFF;
6214 hwc->sample_period = attr->sample_period;
6215 if (attr->freq && attr->sample_freq)
6216 hwc->sample_period = 1;
6217 hwc->last_period = hwc->sample_period;
6219 local64_set(&hwc->period_left, hwc->sample_period);
6222 * we currently do not support PERF_FORMAT_GROUP on inherited events
6224 if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP))
6227 pmu = perf_init_event(event);
6233 else if (IS_ERR(pmu))
6238 put_pid_ns(event->ns);
6240 return ERR_PTR(err);
6245 if (!event->parent) {
6246 if (event->attach_state & PERF_ATTACH_TASK)
6247 jump_label_inc(&perf_sched_events);
6248 if (event->attr.mmap || event->attr.mmap_data)
6249 atomic_inc(&nr_mmap_events);
6250 if (event->attr.comm)
6251 atomic_inc(&nr_comm_events);
6252 if (event->attr.task)
6253 atomic_inc(&nr_task_events);
6254 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) {
6255 err = get_callchain_buffers();
6258 return ERR_PTR(err);
6266 static int perf_copy_attr(struct perf_event_attr __user *uattr,
6267 struct perf_event_attr *attr)
6272 if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0))
6276 * zero the full structure, so that a short copy will be nice.
6278 memset(attr, 0, sizeof(*attr));
6280 ret = get_user(size, &uattr->size);
6284 if (size > PAGE_SIZE) /* silly large */
6287 if (!size) /* abi compat */
6288 size = PERF_ATTR_SIZE_VER0;
6290 if (size < PERF_ATTR_SIZE_VER0)
6294 * If we're handed a bigger struct than we know of,
6295 * ensure all the unknown bits are 0 - i.e. new
6296 * user-space does not rely on any kernel feature
6297 * extensions we dont know about yet.
6299 if (size > sizeof(*attr)) {
6300 unsigned char __user *addr;
6301 unsigned char __user *end;
6304 addr = (void __user *)uattr + sizeof(*attr);
6305 end = (void __user *)uattr + size;
6307 for (; addr < end; addr++) {
6308 ret = get_user(val, addr);
6314 size = sizeof(*attr);
6317 ret = copy_from_user(attr, uattr, size);
6322 * If the type exists, the corresponding creation will verify
6325 if (attr->type >= PERF_TYPE_MAX)
6328 if (attr->__reserved_1)
6331 if (attr->sample_type & ~(PERF_SAMPLE_MAX-1))
6334 if (attr->read_format & ~(PERF_FORMAT_MAX-1))
6341 put_user(sizeof(*attr), &uattr->size);
6347 perf_event_set_output(struct perf_event *event, struct perf_event *output_event)
6349 struct perf_buffer *buffer = NULL, *old_buffer = NULL;
6355 /* don't allow circular references */
6356 if (event == output_event)
6360 * Don't allow cross-cpu buffers
6362 if (output_event->cpu != event->cpu)
6366 * If its not a per-cpu buffer, it must be the same task.
6368 if (output_event->cpu == -1 && output_event->ctx != event->ctx)
6372 mutex_lock(&event->mmap_mutex);
6373 /* Can't redirect output if we've got an active mmap() */
6374 if (atomic_read(&event->mmap_count))
6378 /* get the buffer we want to redirect to */
6379 buffer = perf_buffer_get(output_event);
6384 old_buffer = event->buffer;
6385 rcu_assign_pointer(event->buffer, buffer);
6388 mutex_unlock(&event->mmap_mutex);
6391 perf_buffer_put(old_buffer);
6397 * sys_perf_event_open - open a performance event, associate it to a task/cpu
6399 * @attr_uptr: event_id type attributes for monitoring/sampling
6402 * @group_fd: group leader event fd
6404 SYSCALL_DEFINE5(perf_event_open,
6405 struct perf_event_attr __user *, attr_uptr,
6406 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
6408 struct perf_event *group_leader = NULL, *output_event = NULL;
6409 struct perf_event *event, *sibling;
6410 struct perf_event_attr attr;
6411 struct perf_event_context *ctx;
6412 struct file *event_file = NULL;
6413 struct file *group_file = NULL;
6414 struct task_struct *task = NULL;
6418 int fput_needed = 0;
6421 /* for future expandability... */
6422 if (flags & ~PERF_FLAG_ALL)
6425 err = perf_copy_attr(attr_uptr, &attr);
6429 if (!attr.exclude_kernel) {
6430 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
6435 if (attr.sample_freq > sysctl_perf_event_sample_rate)
6440 * In cgroup mode, the pid argument is used to pass the fd
6441 * opened to the cgroup directory in cgroupfs. The cpu argument
6442 * designates the cpu on which to monitor threads from that
6445 if ((flags & PERF_FLAG_PID_CGROUP) && (pid == -1 || cpu == -1))
6448 event_fd = get_unused_fd_flags(O_RDWR);
6452 if (group_fd != -1) {
6453 group_leader = perf_fget_light(group_fd, &fput_needed);
6454 if (IS_ERR(group_leader)) {
6455 err = PTR_ERR(group_leader);
6458 group_file = group_leader->filp;
6459 if (flags & PERF_FLAG_FD_OUTPUT)
6460 output_event = group_leader;
6461 if (flags & PERF_FLAG_FD_NO_GROUP)
6462 group_leader = NULL;
6465 if (pid != -1 && !(flags & PERF_FLAG_PID_CGROUP)) {
6466 task = find_lively_task_by_vpid(pid);
6468 err = PTR_ERR(task);
6473 event = perf_event_alloc(&attr, cpu, task, group_leader, NULL, NULL);
6474 if (IS_ERR(event)) {
6475 err = PTR_ERR(event);
6479 if (flags & PERF_FLAG_PID_CGROUP) {
6480 err = perf_cgroup_connect(pid, event, &attr, group_leader);
6485 * - that has cgroup constraint on event->cpu
6486 * - that may need work on context switch
6488 atomic_inc(&per_cpu(perf_cgroup_events, event->cpu));
6489 jump_label_inc(&perf_sched_events);
6493 * Special case software events and allow them to be part of
6494 * any hardware group.
6499 (is_software_event(event) != is_software_event(group_leader))) {
6500 if (is_software_event(event)) {
6502 * If event and group_leader are not both a software
6503 * event, and event is, then group leader is not.
6505 * Allow the addition of software events to !software
6506 * groups, this is safe because software events never
6509 pmu = group_leader->pmu;
6510 } else if (is_software_event(group_leader) &&
6511 (group_leader->group_flags & PERF_GROUP_SOFTWARE)) {
6513 * In case the group is a pure software group, and we
6514 * try to add a hardware event, move the whole group to
6515 * the hardware context.
6522 * Get the target context (task or percpu):
6524 ctx = find_get_context(pmu, task, cpu);
6531 put_task_struct(task);
6536 * Look up the group leader (we will attach this event to it):
6542 * Do not allow a recursive hierarchy (this new sibling
6543 * becoming part of another group-sibling):
6545 if (group_leader->group_leader != group_leader)
6548 * Do not allow to attach to a group in a different
6549 * task or CPU context:
6552 if (group_leader->ctx->type != ctx->type)
6555 if (group_leader->ctx != ctx)
6560 * Only a group leader can be exclusive or pinned
6562 if (attr.exclusive || attr.pinned)
6567 err = perf_event_set_output(event, output_event);
6572 event_file = anon_inode_getfile("[perf_event]", &perf_fops, event, O_RDWR);
6573 if (IS_ERR(event_file)) {
6574 err = PTR_ERR(event_file);
6579 struct perf_event_context *gctx = group_leader->ctx;
6581 mutex_lock(&gctx->mutex);
6582 perf_remove_from_context(group_leader);
6583 list_for_each_entry(sibling, &group_leader->sibling_list,
6585 perf_remove_from_context(sibling);
6588 mutex_unlock(&gctx->mutex);
6592 event->filp = event_file;
6593 WARN_ON_ONCE(ctx->parent_ctx);
6594 mutex_lock(&ctx->mutex);
6597 perf_install_in_context(ctx, group_leader, cpu);
6599 list_for_each_entry(sibling, &group_leader->sibling_list,
6601 perf_install_in_context(ctx, sibling, cpu);
6606 perf_install_in_context(ctx, event, cpu);
6608 perf_unpin_context(ctx);
6609 mutex_unlock(&ctx->mutex);
6611 event->owner = current;
6613 mutex_lock(¤t->perf_event_mutex);
6614 list_add_tail(&event->owner_entry, ¤t->perf_event_list);
6615 mutex_unlock(¤t->perf_event_mutex);
6618 * Precalculate sample_data sizes
6620 perf_event__header_size(event);
6621 perf_event__id_header_size(event);
6624 * Drop the reference on the group_event after placing the
6625 * new event on the sibling_list. This ensures destruction
6626 * of the group leader will find the pointer to itself in
6627 * perf_group_detach().
6629 fput_light(group_file, fput_needed);
6630 fd_install(event_fd, event_file);
6634 perf_unpin_context(ctx);
6640 put_task_struct(task);
6642 fput_light(group_file, fput_needed);
6644 put_unused_fd(event_fd);
6649 * perf_event_create_kernel_counter
6651 * @attr: attributes of the counter to create
6652 * @cpu: cpu in which the counter is bound
6653 * @task: task to profile (NULL for percpu)
6656 perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu,
6657 struct task_struct *task,
6658 perf_overflow_handler_t overflow_handler)
6660 struct perf_event_context *ctx;
6661 struct perf_event *event;
6665 * Get the target context (task or percpu):
6668 event = perf_event_alloc(attr, cpu, task, NULL, NULL, overflow_handler);
6669 if (IS_ERR(event)) {
6670 err = PTR_ERR(event);
6674 ctx = find_get_context(event->pmu, task, cpu);
6681 WARN_ON_ONCE(ctx->parent_ctx);
6682 mutex_lock(&ctx->mutex);
6683 perf_install_in_context(ctx, event, cpu);
6685 perf_unpin_context(ctx);
6686 mutex_unlock(&ctx->mutex);
6693 return ERR_PTR(err);
6695 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter);
6697 static void sync_child_event(struct perf_event *child_event,
6698 struct task_struct *child)
6700 struct perf_event *parent_event = child_event->parent;
6703 if (child_event->attr.inherit_stat)
6704 perf_event_read_event(child_event, child);
6706 child_val = perf_event_count(child_event);
6709 * Add back the child's count to the parent's count:
6711 atomic64_add(child_val, &parent_event->child_count);
6712 atomic64_add(child_event->total_time_enabled,
6713 &parent_event->child_total_time_enabled);
6714 atomic64_add(child_event->total_time_running,
6715 &parent_event->child_total_time_running);
6718 * Remove this event from the parent's list
6720 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
6721 mutex_lock(&parent_event->child_mutex);
6722 list_del_init(&child_event->child_list);
6723 mutex_unlock(&parent_event->child_mutex);
6726 * Release the parent event, if this was the last
6729 fput(parent_event->filp);
6733 __perf_event_exit_task(struct perf_event *child_event,
6734 struct perf_event_context *child_ctx,
6735 struct task_struct *child)
6737 if (child_event->parent) {
6738 raw_spin_lock_irq(&child_ctx->lock);
6739 perf_group_detach(child_event);
6740 raw_spin_unlock_irq(&child_ctx->lock);
6743 perf_remove_from_context(child_event);
6746 * It can happen that the parent exits first, and has events
6747 * that are still around due to the child reference. These
6748 * events need to be zapped.
6750 if (child_event->parent) {
6751 sync_child_event(child_event, child);
6752 free_event(child_event);
6756 static void perf_event_exit_task_context(struct task_struct *child, int ctxn)
6758 struct perf_event *child_event, *tmp;
6759 struct perf_event_context *child_ctx;
6760 unsigned long flags;
6762 if (likely(!child->perf_event_ctxp[ctxn])) {
6763 perf_event_task(child, NULL, 0);
6767 local_irq_save(flags);
6769 * We can't reschedule here because interrupts are disabled,
6770 * and either child is current or it is a task that can't be
6771 * scheduled, so we are now safe from rescheduling changing
6774 child_ctx = rcu_dereference_raw(child->perf_event_ctxp[ctxn]);
6777 * Take the context lock here so that if find_get_context is
6778 * reading child->perf_event_ctxp, we wait until it has
6779 * incremented the context's refcount before we do put_ctx below.
6781 raw_spin_lock(&child_ctx->lock);
6782 task_ctx_sched_out(child_ctx);
6783 child->perf_event_ctxp[ctxn] = NULL;
6785 * If this context is a clone; unclone it so it can't get
6786 * swapped to another process while we're removing all
6787 * the events from it.
6789 unclone_ctx(child_ctx);
6790 update_context_time(child_ctx);
6791 raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
6794 * Report the task dead after unscheduling the events so that we
6795 * won't get any samples after PERF_RECORD_EXIT. We can however still
6796 * get a few PERF_RECORD_READ events.
6798 perf_event_task(child, child_ctx, 0);
6801 * We can recurse on the same lock type through:
6803 * __perf_event_exit_task()
6804 * sync_child_event()
6805 * fput(parent_event->filp)
6807 * mutex_lock(&ctx->mutex)
6809 * But since its the parent context it won't be the same instance.
6811 mutex_lock(&child_ctx->mutex);
6814 list_for_each_entry_safe(child_event, tmp, &child_ctx->pinned_groups,
6816 __perf_event_exit_task(child_event, child_ctx, child);
6818 list_for_each_entry_safe(child_event, tmp, &child_ctx->flexible_groups,
6820 __perf_event_exit_task(child_event, child_ctx, child);
6823 * If the last event was a group event, it will have appended all
6824 * its siblings to the list, but we obtained 'tmp' before that which
6825 * will still point to the list head terminating the iteration.
6827 if (!list_empty(&child_ctx->pinned_groups) ||
6828 !list_empty(&child_ctx->flexible_groups))
6831 mutex_unlock(&child_ctx->mutex);
6837 * When a child task exits, feed back event values to parent events.
6839 void perf_event_exit_task(struct task_struct *child)
6841 struct perf_event *event, *tmp;
6844 mutex_lock(&child->perf_event_mutex);
6845 list_for_each_entry_safe(event, tmp, &child->perf_event_list,
6847 list_del_init(&event->owner_entry);
6850 * Ensure the list deletion is visible before we clear
6851 * the owner, closes a race against perf_release() where
6852 * we need to serialize on the owner->perf_event_mutex.
6855 event->owner = NULL;
6857 mutex_unlock(&child->perf_event_mutex);
6859 for_each_task_context_nr(ctxn)
6860 perf_event_exit_task_context(child, ctxn);
6863 static void perf_free_event(struct perf_event *event,
6864 struct perf_event_context *ctx)
6866 struct perf_event *parent = event->parent;
6868 if (WARN_ON_ONCE(!parent))
6871 mutex_lock(&parent->child_mutex);
6872 list_del_init(&event->child_list);
6873 mutex_unlock(&parent->child_mutex);
6877 perf_group_detach(event);
6878 list_del_event(event, ctx);
6883 * free an unexposed, unused context as created by inheritance by
6884 * perf_event_init_task below, used by fork() in case of fail.
6886 void perf_event_free_task(struct task_struct *task)
6888 struct perf_event_context *ctx;
6889 struct perf_event *event, *tmp;
6892 for_each_task_context_nr(ctxn) {
6893 ctx = task->perf_event_ctxp[ctxn];
6897 mutex_lock(&ctx->mutex);
6899 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups,
6901 perf_free_event(event, ctx);
6903 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups,
6905 perf_free_event(event, ctx);
6907 if (!list_empty(&ctx->pinned_groups) ||
6908 !list_empty(&ctx->flexible_groups))
6911 mutex_unlock(&ctx->mutex);
6917 void perf_event_delayed_put(struct task_struct *task)
6921 for_each_task_context_nr(ctxn)
6922 WARN_ON_ONCE(task->perf_event_ctxp[ctxn]);
6926 * inherit a event from parent task to child task:
6928 static struct perf_event *
6929 inherit_event(struct perf_event *parent_event,
6930 struct task_struct *parent,
6931 struct perf_event_context *parent_ctx,
6932 struct task_struct *child,
6933 struct perf_event *group_leader,
6934 struct perf_event_context *child_ctx)
6936 struct perf_event *child_event;
6937 unsigned long flags;
6940 * Instead of creating recursive hierarchies of events,
6941 * we link inherited events back to the original parent,
6942 * which has a filp for sure, which we use as the reference
6945 if (parent_event->parent)
6946 parent_event = parent_event->parent;
6948 child_event = perf_event_alloc(&parent_event->attr,
6951 group_leader, parent_event,
6953 if (IS_ERR(child_event))
6958 * Make the child state follow the state of the parent event,
6959 * not its attr.disabled bit. We hold the parent's mutex,
6960 * so we won't race with perf_event_{en, dis}able_family.
6962 if (parent_event->state >= PERF_EVENT_STATE_INACTIVE)
6963 child_event->state = PERF_EVENT_STATE_INACTIVE;
6965 child_event->state = PERF_EVENT_STATE_OFF;
6967 if (parent_event->attr.freq) {
6968 u64 sample_period = parent_event->hw.sample_period;
6969 struct hw_perf_event *hwc = &child_event->hw;
6971 hwc->sample_period = sample_period;
6972 hwc->last_period = sample_period;
6974 local64_set(&hwc->period_left, sample_period);
6977 child_event->ctx = child_ctx;
6978 child_event->overflow_handler = parent_event->overflow_handler;
6981 * Precalculate sample_data sizes
6983 perf_event__header_size(child_event);
6984 perf_event__id_header_size(child_event);
6987 * Link it up in the child's context:
6989 raw_spin_lock_irqsave(&child_ctx->lock, flags);
6990 add_event_to_ctx(child_event, child_ctx);
6991 raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
6994 * Get a reference to the parent filp - we will fput it
6995 * when the child event exits. This is safe to do because
6996 * we are in the parent and we know that the filp still
6997 * exists and has a nonzero count:
6999 atomic_long_inc(&parent_event->filp->f_count);
7002 * Link this into the parent event's child list
7004 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
7005 mutex_lock(&parent_event->child_mutex);
7006 list_add_tail(&child_event->child_list, &parent_event->child_list);
7007 mutex_unlock(&parent_event->child_mutex);
7012 static int inherit_group(struct perf_event *parent_event,
7013 struct task_struct *parent,
7014 struct perf_event_context *parent_ctx,
7015 struct task_struct *child,
7016 struct perf_event_context *child_ctx)
7018 struct perf_event *leader;
7019 struct perf_event *sub;
7020 struct perf_event *child_ctr;
7022 leader = inherit_event(parent_event, parent, parent_ctx,
7023 child, NULL, child_ctx);
7025 return PTR_ERR(leader);
7026 list_for_each_entry(sub, &parent_event->sibling_list, group_entry) {
7027 child_ctr = inherit_event(sub, parent, parent_ctx,
7028 child, leader, child_ctx);
7029 if (IS_ERR(child_ctr))
7030 return PTR_ERR(child_ctr);
7036 inherit_task_group(struct perf_event *event, struct task_struct *parent,
7037 struct perf_event_context *parent_ctx,
7038 struct task_struct *child, int ctxn,
7042 struct perf_event_context *child_ctx;
7044 if (!event->attr.inherit) {
7049 child_ctx = child->perf_event_ctxp[ctxn];
7052 * This is executed from the parent task context, so
7053 * inherit events that have been marked for cloning.
7054 * First allocate and initialize a context for the
7058 child_ctx = alloc_perf_context(event->pmu, child);
7062 child->perf_event_ctxp[ctxn] = child_ctx;
7065 ret = inherit_group(event, parent, parent_ctx,
7075 * Initialize the perf_event context in task_struct
7077 int perf_event_init_context(struct task_struct *child, int ctxn)
7079 struct perf_event_context *child_ctx, *parent_ctx;
7080 struct perf_event_context *cloned_ctx;
7081 struct perf_event *event;
7082 struct task_struct *parent = current;
7083 int inherited_all = 1;
7084 unsigned long flags;
7087 if (likely(!parent->perf_event_ctxp[ctxn]))
7091 * If the parent's context is a clone, pin it so it won't get
7094 parent_ctx = perf_pin_task_context(parent, ctxn);
7097 * No need to check if parent_ctx != NULL here; since we saw
7098 * it non-NULL earlier, the only reason for it to become NULL
7099 * is if we exit, and since we're currently in the middle of
7100 * a fork we can't be exiting at the same time.
7104 * Lock the parent list. No need to lock the child - not PID
7105 * hashed yet and not running, so nobody can access it.
7107 mutex_lock(&parent_ctx->mutex);
7110 * We dont have to disable NMIs - we are only looking at
7111 * the list, not manipulating it:
7113 list_for_each_entry(event, &parent_ctx->pinned_groups, group_entry) {
7114 ret = inherit_task_group(event, parent, parent_ctx,
7115 child, ctxn, &inherited_all);
7121 * We can't hold ctx->lock when iterating the ->flexible_group list due
7122 * to allocations, but we need to prevent rotation because
7123 * rotate_ctx() will change the list from interrupt context.
7125 raw_spin_lock_irqsave(&parent_ctx->lock, flags);
7126 parent_ctx->rotate_disable = 1;
7127 raw_spin_unlock_irqrestore(&parent_ctx->lock, flags);
7129 list_for_each_entry(event, &parent_ctx->flexible_groups, group_entry) {
7130 ret = inherit_task_group(event, parent, parent_ctx,
7131 child, ctxn, &inherited_all);
7136 raw_spin_lock_irqsave(&parent_ctx->lock, flags);
7137 parent_ctx->rotate_disable = 0;
7139 child_ctx = child->perf_event_ctxp[ctxn];
7141 if (child_ctx && inherited_all) {
7143 * Mark the child context as a clone of the parent
7144 * context, or of whatever the parent is a clone of.
7146 * Note that if the parent is a clone, the holding of
7147 * parent_ctx->lock avoids it from being uncloned.
7149 cloned_ctx = parent_ctx->parent_ctx;
7151 child_ctx->parent_ctx = cloned_ctx;
7152 child_ctx->parent_gen = parent_ctx->parent_gen;
7154 child_ctx->parent_ctx = parent_ctx;
7155 child_ctx->parent_gen = parent_ctx->generation;
7157 get_ctx(child_ctx->parent_ctx);
7160 raw_spin_unlock_irqrestore(&parent_ctx->lock, flags);
7161 mutex_unlock(&parent_ctx->mutex);
7163 perf_unpin_context(parent_ctx);
7164 put_ctx(parent_ctx);
7170 * Initialize the perf_event context in task_struct
7172 int perf_event_init_task(struct task_struct *child)
7176 memset(child->perf_event_ctxp, 0, sizeof(child->perf_event_ctxp));
7177 mutex_init(&child->perf_event_mutex);
7178 INIT_LIST_HEAD(&child->perf_event_list);
7180 for_each_task_context_nr(ctxn) {
7181 ret = perf_event_init_context(child, ctxn);
7189 static void __init perf_event_init_all_cpus(void)
7191 struct swevent_htable *swhash;
7194 for_each_possible_cpu(cpu) {
7195 swhash = &per_cpu(swevent_htable, cpu);
7196 mutex_init(&swhash->hlist_mutex);
7197 INIT_LIST_HEAD(&per_cpu(rotation_list, cpu));
7201 static void __cpuinit perf_event_init_cpu(int cpu)
7203 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
7205 mutex_lock(&swhash->hlist_mutex);
7206 if (swhash->hlist_refcount > 0) {
7207 struct swevent_hlist *hlist;
7209 hlist = kzalloc_node(sizeof(*hlist), GFP_KERNEL, cpu_to_node(cpu));
7211 rcu_assign_pointer(swhash->swevent_hlist, hlist);
7213 mutex_unlock(&swhash->hlist_mutex);
7216 #if defined CONFIG_HOTPLUG_CPU || defined CONFIG_KEXEC
7217 static void perf_pmu_rotate_stop(struct pmu *pmu)
7219 struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
7221 WARN_ON(!irqs_disabled());
7223 list_del_init(&cpuctx->rotation_list);
7226 static void __perf_event_exit_context(void *__info)
7228 struct perf_event_context *ctx = __info;
7229 struct perf_event *event, *tmp;
7231 perf_pmu_rotate_stop(ctx->pmu);
7233 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry)
7234 __perf_remove_from_context(event);
7235 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups, group_entry)
7236 __perf_remove_from_context(event);
7239 static void perf_event_exit_cpu_context(int cpu)
7241 struct perf_event_context *ctx;
7245 idx = srcu_read_lock(&pmus_srcu);
7246 list_for_each_entry_rcu(pmu, &pmus, entry) {
7247 ctx = &per_cpu_ptr(pmu->pmu_cpu_context, cpu)->ctx;
7249 mutex_lock(&ctx->mutex);
7250 smp_call_function_single(cpu, __perf_event_exit_context, ctx, 1);
7251 mutex_unlock(&ctx->mutex);
7253 srcu_read_unlock(&pmus_srcu, idx);
7256 static void perf_event_exit_cpu(int cpu)
7258 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
7260 mutex_lock(&swhash->hlist_mutex);
7261 swevent_hlist_release(swhash);
7262 mutex_unlock(&swhash->hlist_mutex);
7264 perf_event_exit_cpu_context(cpu);
7267 static inline void perf_event_exit_cpu(int cpu) { }
7271 perf_reboot(struct notifier_block *notifier, unsigned long val, void *v)
7275 for_each_online_cpu(cpu)
7276 perf_event_exit_cpu(cpu);
7282 * Run the perf reboot notifier at the very last possible moment so that
7283 * the generic watchdog code runs as long as possible.
7285 static struct notifier_block perf_reboot_notifier = {
7286 .notifier_call = perf_reboot,
7287 .priority = INT_MIN,
7290 static int __cpuinit
7291 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
7293 unsigned int cpu = (long)hcpu;
7295 switch (action & ~CPU_TASKS_FROZEN) {
7297 case CPU_UP_PREPARE:
7298 case CPU_DOWN_FAILED:
7299 perf_event_init_cpu(cpu);
7302 case CPU_UP_CANCELED:
7303 case CPU_DOWN_PREPARE:
7304 perf_event_exit_cpu(cpu);
7314 void __init perf_event_init(void)
7320 perf_event_init_all_cpus();
7321 init_srcu_struct(&pmus_srcu);
7322 perf_pmu_register(&perf_swevent, "software", PERF_TYPE_SOFTWARE);
7323 perf_pmu_register(&perf_cpu_clock, NULL, -1);
7324 perf_pmu_register(&perf_task_clock, NULL, -1);
7326 perf_cpu_notifier(perf_cpu_notify);
7327 register_reboot_notifier(&perf_reboot_notifier);
7329 ret = init_hw_breakpoint();
7330 WARN(ret, "hw_breakpoint initialization failed with: %d", ret);
7333 static int __init perf_event_sysfs_init(void)
7338 mutex_lock(&pmus_lock);
7340 ret = bus_register(&pmu_bus);
7344 list_for_each_entry(pmu, &pmus, entry) {
7345 if (!pmu->name || pmu->type < 0)
7348 ret = pmu_dev_alloc(pmu);
7349 WARN(ret, "Failed to register pmu: %s, reason %d\n", pmu->name, ret);
7351 pmu_bus_running = 1;
7355 mutex_unlock(&pmus_lock);
7359 device_initcall(perf_event_sysfs_init);
7361 #ifdef CONFIG_CGROUP_PERF
7362 static struct cgroup_subsys_state *perf_cgroup_create(
7363 struct cgroup_subsys *ss, struct cgroup *cont)
7365 struct perf_cgroup *jc;
7367 jc = kzalloc(sizeof(*jc), GFP_KERNEL);
7369 return ERR_PTR(-ENOMEM);
7371 jc->info = alloc_percpu(struct perf_cgroup_info);
7374 return ERR_PTR(-ENOMEM);
7380 static void perf_cgroup_destroy(struct cgroup_subsys *ss,
7381 struct cgroup *cont)
7383 struct perf_cgroup *jc;
7384 jc = container_of(cgroup_subsys_state(cont, perf_subsys_id),
7385 struct perf_cgroup, css);
7386 free_percpu(jc->info);
7390 static int __perf_cgroup_move(void *info)
7392 struct task_struct *task = info;
7393 perf_cgroup_switch(task, PERF_CGROUP_SWOUT | PERF_CGROUP_SWIN);
7398 perf_cgroup_attach_task(struct cgroup *cgrp, struct task_struct *task)
7400 task_function_call(task, __perf_cgroup_move, task);
7403 static void perf_cgroup_exit(struct cgroup_subsys *ss, struct cgroup *cgrp,
7404 struct cgroup *old_cgrp, struct task_struct *task)
7407 * cgroup_exit() is called in the copy_process() failure path.
7408 * Ignore this case since the task hasn't ran yet, this avoids
7409 * trying to poke a half freed task state from generic code.
7411 if (!(task->flags & PF_EXITING))
7414 perf_cgroup_attach_task(cgrp, task);
7417 struct cgroup_subsys perf_subsys = {
7418 .name = "perf_event",
7419 .subsys_id = perf_subsys_id,
7420 .create = perf_cgroup_create,
7421 .destroy = perf_cgroup_destroy,
7422 .exit = perf_cgroup_exit,
7423 .attach_task = perf_cgroup_attach_task,
7425 #endif /* CONFIG_CGROUP_PERF */