2 * Performance events core code:
4 * Copyright (C) 2008 Thomas Gleixner <tglx@linutronix.de>
5 * Copyright (C) 2008-2011 Red Hat, Inc., Ingo Molnar
6 * Copyright (C) 2008-2011 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
7 * Copyright © 2009 Paul Mackerras, IBM Corp. <paulus@au1.ibm.com>
9 * For licensing details see kernel-base/COPYING
14 #include <linux/cpu.h>
15 #include <linux/smp.h>
16 #include <linux/idr.h>
17 #include <linux/file.h>
18 #include <linux/poll.h>
19 #include <linux/slab.h>
20 #include <linux/hash.h>
21 #include <linux/sysfs.h>
22 #include <linux/dcache.h>
23 #include <linux/percpu.h>
24 #include <linux/ptrace.h>
25 #include <linux/reboot.h>
26 #include <linux/vmstat.h>
27 #include <linux/device.h>
28 #include <linux/vmalloc.h>
29 #include <linux/hardirq.h>
30 #include <linux/rculist.h>
31 #include <linux/uaccess.h>
32 #include <linux/syscalls.h>
33 #include <linux/anon_inodes.h>
34 #include <linux/kernel_stat.h>
35 #include <linux/perf_event.h>
36 #include <linux/ftrace_event.h>
37 #include <linux/hw_breakpoint.h>
39 #include <asm/irq_regs.h>
41 struct remote_function_call {
42 struct task_struct *p;
43 int (*func)(void *info);
48 static void remote_function(void *data)
50 struct remote_function_call *tfc = data;
51 struct task_struct *p = tfc->p;
55 if (task_cpu(p) != smp_processor_id() || !task_curr(p))
59 tfc->ret = tfc->func(tfc->info);
63 * task_function_call - call a function on the cpu on which a task runs
64 * @p: the task to evaluate
65 * @func: the function to be called
66 * @info: the function call argument
68 * Calls the function @func when the task is currently running. This might
69 * be on the current CPU, which just calls the function directly
71 * returns: @func return value, or
72 * -ESRCH - when the process isn't running
73 * -EAGAIN - when the process moved away
76 task_function_call(struct task_struct *p, int (*func) (void *info), void *info)
78 struct remote_function_call data = {
82 .ret = -ESRCH, /* No such (running) process */
86 smp_call_function_single(task_cpu(p), remote_function, &data, 1);
92 * cpu_function_call - call a function on the cpu
93 * @func: the function to be called
94 * @info: the function call argument
96 * Calls the function @func on the remote cpu.
98 * returns: @func return value or -ENXIO when the cpu is offline
100 static int cpu_function_call(int cpu, int (*func) (void *info), void *info)
102 struct remote_function_call data = {
106 .ret = -ENXIO, /* No such CPU */
109 smp_call_function_single(cpu, remote_function, &data, 1);
114 #define PERF_FLAG_ALL (PERF_FLAG_FD_NO_GROUP |\
115 PERF_FLAG_FD_OUTPUT |\
116 PERF_FLAG_PID_CGROUP)
119 EVENT_FLEXIBLE = 0x1,
121 EVENT_ALL = EVENT_FLEXIBLE | EVENT_PINNED,
125 * perf_sched_events : >0 events exist
126 * perf_cgroup_events: >0 per-cpu cgroup events exist on this cpu
128 struct jump_label_key perf_sched_events __read_mostly;
129 static DEFINE_PER_CPU(atomic_t, perf_cgroup_events);
131 static atomic_t nr_mmap_events __read_mostly;
132 static atomic_t nr_comm_events __read_mostly;
133 static atomic_t nr_task_events __read_mostly;
135 static LIST_HEAD(pmus);
136 static DEFINE_MUTEX(pmus_lock);
137 static struct srcu_struct pmus_srcu;
140 * perf event paranoia level:
141 * -1 - not paranoid at all
142 * 0 - disallow raw tracepoint access for unpriv
143 * 1 - disallow cpu events for unpriv
144 * 2 - disallow kernel profiling for unpriv
146 int sysctl_perf_event_paranoid __read_mostly = 1;
148 /* Minimum for 512 kiB + 1 user control page */
149 int sysctl_perf_event_mlock __read_mostly = 512 + (PAGE_SIZE / 1024); /* 'free' kiB per user */
152 * max perf event sample rate
154 #define DEFAULT_MAX_SAMPLE_RATE 100000
155 int sysctl_perf_event_sample_rate __read_mostly = DEFAULT_MAX_SAMPLE_RATE;
156 static int max_samples_per_tick __read_mostly =
157 DIV_ROUND_UP(DEFAULT_MAX_SAMPLE_RATE, HZ);
159 int perf_proc_update_handler(struct ctl_table *table, int write,
160 void __user *buffer, size_t *lenp,
163 int ret = proc_dointvec(table, write, buffer, lenp, ppos);
168 max_samples_per_tick = DIV_ROUND_UP(sysctl_perf_event_sample_rate, HZ);
173 static atomic64_t perf_event_id;
175 static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
176 enum event_type_t event_type);
178 static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
179 enum event_type_t event_type,
180 struct task_struct *task);
182 static void update_context_time(struct perf_event_context *ctx);
183 static u64 perf_event_time(struct perf_event *event);
185 void __weak perf_event_print_debug(void) { }
187 extern __weak const char *perf_pmu_name(void)
192 static inline u64 perf_clock(void)
194 return local_clock();
197 static inline struct perf_cpu_context *
198 __get_cpu_context(struct perf_event_context *ctx)
200 return this_cpu_ptr(ctx->pmu->pmu_cpu_context);
203 static void perf_ctx_lock(struct perf_cpu_context *cpuctx,
204 struct perf_event_context *ctx)
206 raw_spin_lock(&cpuctx->ctx.lock);
208 raw_spin_lock(&ctx->lock);
211 static void perf_ctx_unlock(struct perf_cpu_context *cpuctx,
212 struct perf_event_context *ctx)
215 raw_spin_unlock(&ctx->lock);
216 raw_spin_unlock(&cpuctx->ctx.lock);
219 #ifdef CONFIG_CGROUP_PERF
222 * Must ensure cgroup is pinned (css_get) before calling
223 * this function. In other words, we cannot call this function
224 * if there is no cgroup event for the current CPU context.
226 static inline struct perf_cgroup *
227 perf_cgroup_from_task(struct task_struct *task)
229 return container_of(task_subsys_state(task, perf_subsys_id),
230 struct perf_cgroup, css);
234 perf_cgroup_match(struct perf_event *event)
236 struct perf_event_context *ctx = event->ctx;
237 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
239 return !event->cgrp || event->cgrp == cpuctx->cgrp;
242 static inline void perf_get_cgroup(struct perf_event *event)
244 css_get(&event->cgrp->css);
247 static inline void perf_put_cgroup(struct perf_event *event)
249 css_put(&event->cgrp->css);
252 static inline void perf_detach_cgroup(struct perf_event *event)
254 perf_put_cgroup(event);
258 static inline int is_cgroup_event(struct perf_event *event)
260 return event->cgrp != NULL;
263 static inline u64 perf_cgroup_event_time(struct perf_event *event)
265 struct perf_cgroup_info *t;
267 t = per_cpu_ptr(event->cgrp->info, event->cpu);
271 static inline void __update_cgrp_time(struct perf_cgroup *cgrp)
273 struct perf_cgroup_info *info;
278 info = this_cpu_ptr(cgrp->info);
280 info->time += now - info->timestamp;
281 info->timestamp = now;
284 static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx)
286 struct perf_cgroup *cgrp_out = cpuctx->cgrp;
288 __update_cgrp_time(cgrp_out);
291 static inline void update_cgrp_time_from_event(struct perf_event *event)
293 struct perf_cgroup *cgrp;
296 * ensure we access cgroup data only when needed and
297 * when we know the cgroup is pinned (css_get)
299 if (!is_cgroup_event(event))
302 cgrp = perf_cgroup_from_task(current);
304 * Do not update time when cgroup is not active
306 if (cgrp == event->cgrp)
307 __update_cgrp_time(event->cgrp);
311 perf_cgroup_set_timestamp(struct task_struct *task,
312 struct perf_event_context *ctx)
314 struct perf_cgroup *cgrp;
315 struct perf_cgroup_info *info;
318 * ctx->lock held by caller
319 * ensure we do not access cgroup data
320 * unless we have the cgroup pinned (css_get)
322 if (!task || !ctx->nr_cgroups)
325 cgrp = perf_cgroup_from_task(task);
326 info = this_cpu_ptr(cgrp->info);
327 info->timestamp = ctx->timestamp;
330 #define PERF_CGROUP_SWOUT 0x1 /* cgroup switch out every event */
331 #define PERF_CGROUP_SWIN 0x2 /* cgroup switch in events based on task */
334 * reschedule events based on the cgroup constraint of task.
336 * mode SWOUT : schedule out everything
337 * mode SWIN : schedule in based on cgroup for next
339 void perf_cgroup_switch(struct task_struct *task, int mode)
341 struct perf_cpu_context *cpuctx;
346 * disable interrupts to avoid geting nr_cgroup
347 * changes via __perf_event_disable(). Also
350 local_irq_save(flags);
353 * we reschedule only in the presence of cgroup
354 * constrained events.
358 list_for_each_entry_rcu(pmu, &pmus, entry) {
359 cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
362 * perf_cgroup_events says at least one
363 * context on this CPU has cgroup events.
365 * ctx->nr_cgroups reports the number of cgroup
366 * events for a context.
368 if (cpuctx->ctx.nr_cgroups > 0) {
369 perf_ctx_lock(cpuctx, cpuctx->task_ctx);
370 perf_pmu_disable(cpuctx->ctx.pmu);
372 if (mode & PERF_CGROUP_SWOUT) {
373 cpu_ctx_sched_out(cpuctx, EVENT_ALL);
375 * must not be done before ctxswout due
376 * to event_filter_match() in event_sched_out()
381 if (mode & PERF_CGROUP_SWIN) {
382 WARN_ON_ONCE(cpuctx->cgrp);
383 /* set cgrp before ctxsw in to
384 * allow event_filter_match() to not
385 * have to pass task around
387 cpuctx->cgrp = perf_cgroup_from_task(task);
388 cpu_ctx_sched_in(cpuctx, EVENT_ALL, task);
390 perf_pmu_enable(cpuctx->ctx.pmu);
391 perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
397 local_irq_restore(flags);
400 static inline void perf_cgroup_sched_out(struct task_struct *task)
402 perf_cgroup_switch(task, PERF_CGROUP_SWOUT);
405 static inline void perf_cgroup_sched_in(struct task_struct *task)
407 perf_cgroup_switch(task, PERF_CGROUP_SWIN);
410 static inline int perf_cgroup_connect(int fd, struct perf_event *event,
411 struct perf_event_attr *attr,
412 struct perf_event *group_leader)
414 struct perf_cgroup *cgrp;
415 struct cgroup_subsys_state *css;
417 int ret = 0, fput_needed;
419 file = fget_light(fd, &fput_needed);
423 css = cgroup_css_from_dir(file, perf_subsys_id);
429 cgrp = container_of(css, struct perf_cgroup, css);
432 /* must be done before we fput() the file */
433 perf_get_cgroup(event);
436 * all events in a group must monitor
437 * the same cgroup because a task belongs
438 * to only one perf cgroup at a time
440 if (group_leader && group_leader->cgrp != cgrp) {
441 perf_detach_cgroup(event);
445 fput_light(file, fput_needed);
450 perf_cgroup_set_shadow_time(struct perf_event *event, u64 now)
452 struct perf_cgroup_info *t;
453 t = per_cpu_ptr(event->cgrp->info, event->cpu);
454 event->shadow_ctx_time = now - t->timestamp;
458 perf_cgroup_defer_enabled(struct perf_event *event)
461 * when the current task's perf cgroup does not match
462 * the event's, we need to remember to call the
463 * perf_mark_enable() function the first time a task with
464 * a matching perf cgroup is scheduled in.
466 if (is_cgroup_event(event) && !perf_cgroup_match(event))
467 event->cgrp_defer_enabled = 1;
471 perf_cgroup_mark_enabled(struct perf_event *event,
472 struct perf_event_context *ctx)
474 struct perf_event *sub;
475 u64 tstamp = perf_event_time(event);
477 if (!event->cgrp_defer_enabled)
480 event->cgrp_defer_enabled = 0;
482 event->tstamp_enabled = tstamp - event->total_time_enabled;
483 list_for_each_entry(sub, &event->sibling_list, group_entry) {
484 if (sub->state >= PERF_EVENT_STATE_INACTIVE) {
485 sub->tstamp_enabled = tstamp - sub->total_time_enabled;
486 sub->cgrp_defer_enabled = 0;
490 #else /* !CONFIG_CGROUP_PERF */
493 perf_cgroup_match(struct perf_event *event)
498 static inline void perf_detach_cgroup(struct perf_event *event)
501 static inline int is_cgroup_event(struct perf_event *event)
506 static inline u64 perf_cgroup_event_cgrp_time(struct perf_event *event)
511 static inline void update_cgrp_time_from_event(struct perf_event *event)
515 static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx)
519 static inline void perf_cgroup_sched_out(struct task_struct *task)
523 static inline void perf_cgroup_sched_in(struct task_struct *task)
527 static inline int perf_cgroup_connect(pid_t pid, struct perf_event *event,
528 struct perf_event_attr *attr,
529 struct perf_event *group_leader)
535 perf_cgroup_set_timestamp(struct task_struct *task,
536 struct perf_event_context *ctx)
541 perf_cgroup_switch(struct task_struct *task, struct task_struct *next)
546 perf_cgroup_set_shadow_time(struct perf_event *event, u64 now)
550 static inline u64 perf_cgroup_event_time(struct perf_event *event)
556 perf_cgroup_defer_enabled(struct perf_event *event)
561 perf_cgroup_mark_enabled(struct perf_event *event,
562 struct perf_event_context *ctx)
567 void perf_pmu_disable(struct pmu *pmu)
569 int *count = this_cpu_ptr(pmu->pmu_disable_count);
571 pmu->pmu_disable(pmu);
574 void perf_pmu_enable(struct pmu *pmu)
576 int *count = this_cpu_ptr(pmu->pmu_disable_count);
578 pmu->pmu_enable(pmu);
581 static DEFINE_PER_CPU(struct list_head, rotation_list);
584 * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
585 * because they're strictly cpu affine and rotate_start is called with IRQs
586 * disabled, while rotate_context is called from IRQ context.
588 static void perf_pmu_rotate_start(struct pmu *pmu)
590 struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
591 struct list_head *head = &__get_cpu_var(rotation_list);
593 WARN_ON(!irqs_disabled());
595 if (list_empty(&cpuctx->rotation_list))
596 list_add(&cpuctx->rotation_list, head);
599 static void get_ctx(struct perf_event_context *ctx)
601 WARN_ON(!atomic_inc_not_zero(&ctx->refcount));
604 static void put_ctx(struct perf_event_context *ctx)
606 if (atomic_dec_and_test(&ctx->refcount)) {
608 put_ctx(ctx->parent_ctx);
610 put_task_struct(ctx->task);
611 kfree_rcu(ctx, rcu_head);
615 static void unclone_ctx(struct perf_event_context *ctx)
617 if (ctx->parent_ctx) {
618 put_ctx(ctx->parent_ctx);
619 ctx->parent_ctx = NULL;
623 static u32 perf_event_pid(struct perf_event *event, struct task_struct *p)
626 * only top level events have the pid namespace they were created in
629 event = event->parent;
631 return task_tgid_nr_ns(p, event->ns);
634 static u32 perf_event_tid(struct perf_event *event, struct task_struct *p)
637 * only top level events have the pid namespace they were created in
640 event = event->parent;
642 return task_pid_nr_ns(p, event->ns);
646 * If we inherit events we want to return the parent event id
649 static u64 primary_event_id(struct perf_event *event)
654 id = event->parent->id;
660 * Get the perf_event_context for a task and lock it.
661 * This has to cope with with the fact that until it is locked,
662 * the context could get moved to another task.
664 static struct perf_event_context *
665 perf_lock_task_context(struct task_struct *task, int ctxn, unsigned long *flags)
667 struct perf_event_context *ctx;
671 ctx = rcu_dereference(task->perf_event_ctxp[ctxn]);
674 * If this context is a clone of another, it might
675 * get swapped for another underneath us by
676 * perf_event_task_sched_out, though the
677 * rcu_read_lock() protects us from any context
678 * getting freed. Lock the context and check if it
679 * got swapped before we could get the lock, and retry
680 * if so. If we locked the right context, then it
681 * can't get swapped on us any more.
683 raw_spin_lock_irqsave(&ctx->lock, *flags);
684 if (ctx != rcu_dereference(task->perf_event_ctxp[ctxn])) {
685 raw_spin_unlock_irqrestore(&ctx->lock, *flags);
689 if (!atomic_inc_not_zero(&ctx->refcount)) {
690 raw_spin_unlock_irqrestore(&ctx->lock, *flags);
699 * Get the context for a task and increment its pin_count so it
700 * can't get swapped to another task. This also increments its
701 * reference count so that the context can't get freed.
703 static struct perf_event_context *
704 perf_pin_task_context(struct task_struct *task, int ctxn)
706 struct perf_event_context *ctx;
709 ctx = perf_lock_task_context(task, ctxn, &flags);
712 raw_spin_unlock_irqrestore(&ctx->lock, flags);
717 static void perf_unpin_context(struct perf_event_context *ctx)
721 raw_spin_lock_irqsave(&ctx->lock, flags);
723 raw_spin_unlock_irqrestore(&ctx->lock, flags);
727 * Update the record of the current time in a context.
729 static void update_context_time(struct perf_event_context *ctx)
731 u64 now = perf_clock();
733 ctx->time += now - ctx->timestamp;
734 ctx->timestamp = now;
737 static u64 perf_event_time(struct perf_event *event)
739 struct perf_event_context *ctx = event->ctx;
741 if (is_cgroup_event(event))
742 return perf_cgroup_event_time(event);
744 return ctx ? ctx->time : 0;
748 * Update the total_time_enabled and total_time_running fields for a event.
750 static void update_event_times(struct perf_event *event)
752 struct perf_event_context *ctx = event->ctx;
755 if (event->state < PERF_EVENT_STATE_INACTIVE ||
756 event->group_leader->state < PERF_EVENT_STATE_INACTIVE)
759 * in cgroup mode, time_enabled represents
760 * the time the event was enabled AND active
761 * tasks were in the monitored cgroup. This is
762 * independent of the activity of the context as
763 * there may be a mix of cgroup and non-cgroup events.
765 * That is why we treat cgroup events differently
768 if (is_cgroup_event(event))
769 run_end = perf_event_time(event);
770 else if (ctx->is_active)
773 run_end = event->tstamp_stopped;
775 event->total_time_enabled = run_end - event->tstamp_enabled;
777 if (event->state == PERF_EVENT_STATE_INACTIVE)
778 run_end = event->tstamp_stopped;
780 run_end = perf_event_time(event);
782 event->total_time_running = run_end - event->tstamp_running;
787 * Update total_time_enabled and total_time_running for all events in a group.
789 static void update_group_times(struct perf_event *leader)
791 struct perf_event *event;
793 update_event_times(leader);
794 list_for_each_entry(event, &leader->sibling_list, group_entry)
795 update_event_times(event);
798 static struct list_head *
799 ctx_group_list(struct perf_event *event, struct perf_event_context *ctx)
801 if (event->attr.pinned)
802 return &ctx->pinned_groups;
804 return &ctx->flexible_groups;
808 * Add a event from the lists for its context.
809 * Must be called with ctx->mutex and ctx->lock held.
812 list_add_event(struct perf_event *event, struct perf_event_context *ctx)
814 WARN_ON_ONCE(event->attach_state & PERF_ATTACH_CONTEXT);
815 event->attach_state |= PERF_ATTACH_CONTEXT;
818 * If we're a stand alone event or group leader, we go to the context
819 * list, group events are kept attached to the group so that
820 * perf_group_detach can, at all times, locate all siblings.
822 if (event->group_leader == event) {
823 struct list_head *list;
825 if (is_software_event(event))
826 event->group_flags |= PERF_GROUP_SOFTWARE;
828 list = ctx_group_list(event, ctx);
829 list_add_tail(&event->group_entry, list);
832 if (is_cgroup_event(event))
835 list_add_rcu(&event->event_entry, &ctx->event_list);
837 perf_pmu_rotate_start(ctx->pmu);
839 if (event->attr.inherit_stat)
844 * Called at perf_event creation and when events are attached/detached from a
847 static void perf_event__read_size(struct perf_event *event)
849 int entry = sizeof(u64); /* value */
853 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
856 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
859 if (event->attr.read_format & PERF_FORMAT_ID)
860 entry += sizeof(u64);
862 if (event->attr.read_format & PERF_FORMAT_GROUP) {
863 nr += event->group_leader->nr_siblings;
868 event->read_size = size;
871 static void perf_event__header_size(struct perf_event *event)
873 struct perf_sample_data *data;
874 u64 sample_type = event->attr.sample_type;
877 perf_event__read_size(event);
879 if (sample_type & PERF_SAMPLE_IP)
880 size += sizeof(data->ip);
882 if (sample_type & PERF_SAMPLE_ADDR)
883 size += sizeof(data->addr);
885 if (sample_type & PERF_SAMPLE_PERIOD)
886 size += sizeof(data->period);
888 if (sample_type & PERF_SAMPLE_READ)
889 size += event->read_size;
891 event->header_size = size;
894 static void perf_event__id_header_size(struct perf_event *event)
896 struct perf_sample_data *data;
897 u64 sample_type = event->attr.sample_type;
900 if (sample_type & PERF_SAMPLE_TID)
901 size += sizeof(data->tid_entry);
903 if (sample_type & PERF_SAMPLE_TIME)
904 size += sizeof(data->time);
906 if (sample_type & PERF_SAMPLE_ID)
907 size += sizeof(data->id);
909 if (sample_type & PERF_SAMPLE_STREAM_ID)
910 size += sizeof(data->stream_id);
912 if (sample_type & PERF_SAMPLE_CPU)
913 size += sizeof(data->cpu_entry);
915 event->id_header_size = size;
918 static void perf_group_attach(struct perf_event *event)
920 struct perf_event *group_leader = event->group_leader, *pos;
923 * We can have double attach due to group movement in perf_event_open.
925 if (event->attach_state & PERF_ATTACH_GROUP)
928 event->attach_state |= PERF_ATTACH_GROUP;
930 if (group_leader == event)
933 if (group_leader->group_flags & PERF_GROUP_SOFTWARE &&
934 !is_software_event(event))
935 group_leader->group_flags &= ~PERF_GROUP_SOFTWARE;
937 list_add_tail(&event->group_entry, &group_leader->sibling_list);
938 group_leader->nr_siblings++;
940 perf_event__header_size(group_leader);
942 list_for_each_entry(pos, &group_leader->sibling_list, group_entry)
943 perf_event__header_size(pos);
947 * Remove a event from the lists for its context.
948 * Must be called with ctx->mutex and ctx->lock held.
951 list_del_event(struct perf_event *event, struct perf_event_context *ctx)
953 struct perf_cpu_context *cpuctx;
955 * We can have double detach due to exit/hot-unplug + close.
957 if (!(event->attach_state & PERF_ATTACH_CONTEXT))
960 event->attach_state &= ~PERF_ATTACH_CONTEXT;
962 if (is_cgroup_event(event)) {
964 cpuctx = __get_cpu_context(ctx);
966 * if there are no more cgroup events
967 * then cler cgrp to avoid stale pointer
968 * in update_cgrp_time_from_cpuctx()
970 if (!ctx->nr_cgroups)
975 if (event->attr.inherit_stat)
978 list_del_rcu(&event->event_entry);
980 if (event->group_leader == event)
981 list_del_init(&event->group_entry);
983 update_group_times(event);
986 * If event was in error state, then keep it
987 * that way, otherwise bogus counts will be
988 * returned on read(). The only way to get out
989 * of error state is by explicit re-enabling
992 if (event->state > PERF_EVENT_STATE_OFF)
993 event->state = PERF_EVENT_STATE_OFF;
996 static void perf_group_detach(struct perf_event *event)
998 struct perf_event *sibling, *tmp;
999 struct list_head *list = NULL;
1002 * We can have double detach due to exit/hot-unplug + close.
1004 if (!(event->attach_state & PERF_ATTACH_GROUP))
1007 event->attach_state &= ~PERF_ATTACH_GROUP;
1010 * If this is a sibling, remove it from its group.
1012 if (event->group_leader != event) {
1013 list_del_init(&event->group_entry);
1014 event->group_leader->nr_siblings--;
1018 if (!list_empty(&event->group_entry))
1019 list = &event->group_entry;
1022 * If this was a group event with sibling events then
1023 * upgrade the siblings to singleton events by adding them
1024 * to whatever list we are on.
1026 list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) {
1028 list_move_tail(&sibling->group_entry, list);
1029 sibling->group_leader = sibling;
1031 /* Inherit group flags from the previous leader */
1032 sibling->group_flags = event->group_flags;
1036 perf_event__header_size(event->group_leader);
1038 list_for_each_entry(tmp, &event->group_leader->sibling_list, group_entry)
1039 perf_event__header_size(tmp);
1043 event_filter_match(struct perf_event *event)
1045 return (event->cpu == -1 || event->cpu == smp_processor_id())
1046 && perf_cgroup_match(event);
1050 event_sched_out(struct perf_event *event,
1051 struct perf_cpu_context *cpuctx,
1052 struct perf_event_context *ctx)
1054 u64 tstamp = perf_event_time(event);
1057 * An event which could not be activated because of
1058 * filter mismatch still needs to have its timings
1059 * maintained, otherwise bogus information is return
1060 * via read() for time_enabled, time_running:
1062 if (event->state == PERF_EVENT_STATE_INACTIVE
1063 && !event_filter_match(event)) {
1064 delta = tstamp - event->tstamp_stopped;
1065 event->tstamp_running += delta;
1066 event->tstamp_stopped = tstamp;
1069 if (event->state != PERF_EVENT_STATE_ACTIVE)
1072 event->state = PERF_EVENT_STATE_INACTIVE;
1073 if (event->pending_disable) {
1074 event->pending_disable = 0;
1075 event->state = PERF_EVENT_STATE_OFF;
1077 event->tstamp_stopped = tstamp;
1078 event->pmu->del(event, 0);
1081 if (!is_software_event(event))
1082 cpuctx->active_oncpu--;
1084 if (event->attr.exclusive || !cpuctx->active_oncpu)
1085 cpuctx->exclusive = 0;
1089 group_sched_out(struct perf_event *group_event,
1090 struct perf_cpu_context *cpuctx,
1091 struct perf_event_context *ctx)
1093 struct perf_event *event;
1094 int state = group_event->state;
1096 event_sched_out(group_event, cpuctx, ctx);
1099 * Schedule out siblings (if any):
1101 list_for_each_entry(event, &group_event->sibling_list, group_entry)
1102 event_sched_out(event, cpuctx, ctx);
1104 if (state == PERF_EVENT_STATE_ACTIVE && group_event->attr.exclusive)
1105 cpuctx->exclusive = 0;
1109 * Cross CPU call to remove a performance event
1111 * We disable the event on the hardware level first. After that we
1112 * remove it from the context list.
1114 static int __perf_remove_from_context(void *info)
1116 struct perf_event *event = info;
1117 struct perf_event_context *ctx = event->ctx;
1118 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1120 raw_spin_lock(&ctx->lock);
1121 event_sched_out(event, cpuctx, ctx);
1122 list_del_event(event, ctx);
1123 raw_spin_unlock(&ctx->lock);
1130 * Remove the event from a task's (or a CPU's) list of events.
1132 * CPU events are removed with a smp call. For task events we only
1133 * call when the task is on a CPU.
1135 * If event->ctx is a cloned context, callers must make sure that
1136 * every task struct that event->ctx->task could possibly point to
1137 * remains valid. This is OK when called from perf_release since
1138 * that only calls us on the top-level context, which can't be a clone.
1139 * When called from perf_event_exit_task, it's OK because the
1140 * context has been detached from its task.
1142 static void perf_remove_from_context(struct perf_event *event)
1144 struct perf_event_context *ctx = event->ctx;
1145 struct task_struct *task = ctx->task;
1147 lockdep_assert_held(&ctx->mutex);
1151 * Per cpu events are removed via an smp call and
1152 * the removal is always successful.
1154 cpu_function_call(event->cpu, __perf_remove_from_context, event);
1159 if (!task_function_call(task, __perf_remove_from_context, event))
1162 raw_spin_lock_irq(&ctx->lock);
1164 * If we failed to find a running task, but find the context active now
1165 * that we've acquired the ctx->lock, retry.
1167 if (ctx->is_active) {
1168 raw_spin_unlock_irq(&ctx->lock);
1173 * Since the task isn't running, its safe to remove the event, us
1174 * holding the ctx->lock ensures the task won't get scheduled in.
1176 list_del_event(event, ctx);
1177 raw_spin_unlock_irq(&ctx->lock);
1181 * Cross CPU call to disable a performance event
1183 static int __perf_event_disable(void *info)
1185 struct perf_event *event = info;
1186 struct perf_event_context *ctx = event->ctx;
1187 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1190 * If this is a per-task event, need to check whether this
1191 * event's task is the current task on this cpu.
1193 * Can trigger due to concurrent perf_event_context_sched_out()
1194 * flipping contexts around.
1196 if (ctx->task && cpuctx->task_ctx != ctx)
1199 raw_spin_lock(&ctx->lock);
1202 * If the event is on, turn it off.
1203 * If it is in error state, leave it in error state.
1205 if (event->state >= PERF_EVENT_STATE_INACTIVE) {
1206 update_context_time(ctx);
1207 update_cgrp_time_from_event(event);
1208 update_group_times(event);
1209 if (event == event->group_leader)
1210 group_sched_out(event, cpuctx, ctx);
1212 event_sched_out(event, cpuctx, ctx);
1213 event->state = PERF_EVENT_STATE_OFF;
1216 raw_spin_unlock(&ctx->lock);
1224 * If event->ctx is a cloned context, callers must make sure that
1225 * every task struct that event->ctx->task could possibly point to
1226 * remains valid. This condition is satisifed when called through
1227 * perf_event_for_each_child or perf_event_for_each because they
1228 * hold the top-level event's child_mutex, so any descendant that
1229 * goes to exit will block in sync_child_event.
1230 * When called from perf_pending_event it's OK because event->ctx
1231 * is the current context on this CPU and preemption is disabled,
1232 * hence we can't get into perf_event_task_sched_out for this context.
1234 void perf_event_disable(struct perf_event *event)
1236 struct perf_event_context *ctx = event->ctx;
1237 struct task_struct *task = ctx->task;
1241 * Disable the event on the cpu that it's on
1243 cpu_function_call(event->cpu, __perf_event_disable, event);
1248 if (!task_function_call(task, __perf_event_disable, event))
1251 raw_spin_lock_irq(&ctx->lock);
1253 * If the event is still active, we need to retry the cross-call.
1255 if (event->state == PERF_EVENT_STATE_ACTIVE) {
1256 raw_spin_unlock_irq(&ctx->lock);
1258 * Reload the task pointer, it might have been changed by
1259 * a concurrent perf_event_context_sched_out().
1266 * Since we have the lock this context can't be scheduled
1267 * in, so we can change the state safely.
1269 if (event->state == PERF_EVENT_STATE_INACTIVE) {
1270 update_group_times(event);
1271 event->state = PERF_EVENT_STATE_OFF;
1273 raw_spin_unlock_irq(&ctx->lock);
1276 static void perf_set_shadow_time(struct perf_event *event,
1277 struct perf_event_context *ctx,
1281 * use the correct time source for the time snapshot
1283 * We could get by without this by leveraging the
1284 * fact that to get to this function, the caller
1285 * has most likely already called update_context_time()
1286 * and update_cgrp_time_xx() and thus both timestamp
1287 * are identical (or very close). Given that tstamp is,
1288 * already adjusted for cgroup, we could say that:
1289 * tstamp - ctx->timestamp
1291 * tstamp - cgrp->timestamp.
1293 * Then, in perf_output_read(), the calculation would
1294 * work with no changes because:
1295 * - event is guaranteed scheduled in
1296 * - no scheduled out in between
1297 * - thus the timestamp would be the same
1299 * But this is a bit hairy.
1301 * So instead, we have an explicit cgroup call to remain
1302 * within the time time source all along. We believe it
1303 * is cleaner and simpler to understand.
1305 if (is_cgroup_event(event))
1306 perf_cgroup_set_shadow_time(event, tstamp);
1308 event->shadow_ctx_time = tstamp - ctx->timestamp;
1311 #define MAX_INTERRUPTS (~0ULL)
1313 static void perf_log_throttle(struct perf_event *event, int enable);
1316 event_sched_in(struct perf_event *event,
1317 struct perf_cpu_context *cpuctx,
1318 struct perf_event_context *ctx)
1320 u64 tstamp = perf_event_time(event);
1322 if (event->state <= PERF_EVENT_STATE_OFF)
1325 event->state = PERF_EVENT_STATE_ACTIVE;
1326 event->oncpu = smp_processor_id();
1329 * Unthrottle events, since we scheduled we might have missed several
1330 * ticks already, also for a heavily scheduling task there is little
1331 * guarantee it'll get a tick in a timely manner.
1333 if (unlikely(event->hw.interrupts == MAX_INTERRUPTS)) {
1334 perf_log_throttle(event, 1);
1335 event->hw.interrupts = 0;
1339 * The new state must be visible before we turn it on in the hardware:
1343 if (event->pmu->add(event, PERF_EF_START)) {
1344 event->state = PERF_EVENT_STATE_INACTIVE;
1349 event->tstamp_running += tstamp - event->tstamp_stopped;
1351 perf_set_shadow_time(event, ctx, tstamp);
1353 if (!is_software_event(event))
1354 cpuctx->active_oncpu++;
1357 if (event->attr.exclusive)
1358 cpuctx->exclusive = 1;
1364 group_sched_in(struct perf_event *group_event,
1365 struct perf_cpu_context *cpuctx,
1366 struct perf_event_context *ctx)
1368 struct perf_event *event, *partial_group = NULL;
1369 struct pmu *pmu = group_event->pmu;
1370 u64 now = ctx->time;
1371 bool simulate = false;
1373 if (group_event->state == PERF_EVENT_STATE_OFF)
1376 pmu->start_txn(pmu);
1378 if (event_sched_in(group_event, cpuctx, ctx)) {
1379 pmu->cancel_txn(pmu);
1384 * Schedule in siblings as one group (if any):
1386 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
1387 if (event_sched_in(event, cpuctx, ctx)) {
1388 partial_group = event;
1393 if (!pmu->commit_txn(pmu))
1398 * Groups can be scheduled in as one unit only, so undo any
1399 * partial group before returning:
1400 * The events up to the failed event are scheduled out normally,
1401 * tstamp_stopped will be updated.
1403 * The failed events and the remaining siblings need to have
1404 * their timings updated as if they had gone thru event_sched_in()
1405 * and event_sched_out(). This is required to get consistent timings
1406 * across the group. This also takes care of the case where the group
1407 * could never be scheduled by ensuring tstamp_stopped is set to mark
1408 * the time the event was actually stopped, such that time delta
1409 * calculation in update_event_times() is correct.
1411 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
1412 if (event == partial_group)
1416 event->tstamp_running += now - event->tstamp_stopped;
1417 event->tstamp_stopped = now;
1419 event_sched_out(event, cpuctx, ctx);
1422 event_sched_out(group_event, cpuctx, ctx);
1424 pmu->cancel_txn(pmu);
1430 * Work out whether we can put this event group on the CPU now.
1432 static int group_can_go_on(struct perf_event *event,
1433 struct perf_cpu_context *cpuctx,
1437 * Groups consisting entirely of software events can always go on.
1439 if (event->group_flags & PERF_GROUP_SOFTWARE)
1442 * If an exclusive group is already on, no other hardware
1445 if (cpuctx->exclusive)
1448 * If this group is exclusive and there are already
1449 * events on the CPU, it can't go on.
1451 if (event->attr.exclusive && cpuctx->active_oncpu)
1454 * Otherwise, try to add it if all previous groups were able
1460 static void add_event_to_ctx(struct perf_event *event,
1461 struct perf_event_context *ctx)
1463 u64 tstamp = perf_event_time(event);
1465 list_add_event(event, ctx);
1466 perf_group_attach(event);
1467 event->tstamp_enabled = tstamp;
1468 event->tstamp_running = tstamp;
1469 event->tstamp_stopped = tstamp;
1472 static void perf_event_context_sched_in(struct perf_event_context *ctx,
1473 struct task_struct *tsk);
1476 * Cross CPU call to install and enable a performance event
1478 * Must be called with ctx->mutex held
1480 static int __perf_install_in_context(void *info)
1482 struct perf_event *event = info;
1483 struct perf_event_context *ctx = event->ctx;
1484 struct perf_event *leader = event->group_leader;
1485 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1489 * In case we're installing a new context to an already running task,
1490 * could also happen before perf_event_task_sched_in() on architectures
1491 * which do context switches with IRQs enabled.
1493 if (ctx->task && !cpuctx->task_ctx)
1494 perf_event_context_sched_in(ctx, ctx->task);
1496 raw_spin_lock(&ctx->lock);
1498 update_context_time(ctx);
1500 * update cgrp time only if current cgrp
1501 * matches event->cgrp. Must be done before
1502 * calling add_event_to_ctx()
1504 update_cgrp_time_from_event(event);
1506 add_event_to_ctx(event, ctx);
1508 if (!event_filter_match(event))
1512 * Don't put the event on if it is disabled or if
1513 * it is in a group and the group isn't on.
1515 if (event->state != PERF_EVENT_STATE_INACTIVE ||
1516 (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE))
1520 * An exclusive event can't go on if there are already active
1521 * hardware events, and no hardware event can go on if there
1522 * is already an exclusive event on.
1524 if (!group_can_go_on(event, cpuctx, 1))
1527 err = event_sched_in(event, cpuctx, ctx);
1531 * This event couldn't go on. If it is in a group
1532 * then we have to pull the whole group off.
1533 * If the event group is pinned then put it in error state.
1535 if (leader != event)
1536 group_sched_out(leader, cpuctx, ctx);
1537 if (leader->attr.pinned) {
1538 update_group_times(leader);
1539 leader->state = PERF_EVENT_STATE_ERROR;
1544 raw_spin_unlock(&ctx->lock);
1550 * Attach a performance event to a context
1552 * First we add the event to the list with the hardware enable bit
1553 * in event->hw_config cleared.
1555 * If the event is attached to a task which is on a CPU we use a smp
1556 * call to enable it in the task context. The task might have been
1557 * scheduled away, but we check this in the smp call again.
1560 perf_install_in_context(struct perf_event_context *ctx,
1561 struct perf_event *event,
1564 struct task_struct *task = ctx->task;
1566 lockdep_assert_held(&ctx->mutex);
1572 * Per cpu events are installed via an smp call and
1573 * the install is always successful.
1575 cpu_function_call(cpu, __perf_install_in_context, event);
1580 if (!task_function_call(task, __perf_install_in_context, event))
1583 raw_spin_lock_irq(&ctx->lock);
1585 * If we failed to find a running task, but find the context active now
1586 * that we've acquired the ctx->lock, retry.
1588 if (ctx->is_active) {
1589 raw_spin_unlock_irq(&ctx->lock);
1594 * Since the task isn't running, its safe to add the event, us holding
1595 * the ctx->lock ensures the task won't get scheduled in.
1597 add_event_to_ctx(event, ctx);
1598 raw_spin_unlock_irq(&ctx->lock);
1602 * Put a event into inactive state and update time fields.
1603 * Enabling the leader of a group effectively enables all
1604 * the group members that aren't explicitly disabled, so we
1605 * have to update their ->tstamp_enabled also.
1606 * Note: this works for group members as well as group leaders
1607 * since the non-leader members' sibling_lists will be empty.
1609 static void __perf_event_mark_enabled(struct perf_event *event,
1610 struct perf_event_context *ctx)
1612 struct perf_event *sub;
1613 u64 tstamp = perf_event_time(event);
1615 event->state = PERF_EVENT_STATE_INACTIVE;
1616 event->tstamp_enabled = tstamp - event->total_time_enabled;
1617 list_for_each_entry(sub, &event->sibling_list, group_entry) {
1618 if (sub->state >= PERF_EVENT_STATE_INACTIVE)
1619 sub->tstamp_enabled = tstamp - sub->total_time_enabled;
1624 * Cross CPU call to enable a performance event
1626 static int __perf_event_enable(void *info)
1628 struct perf_event *event = info;
1629 struct perf_event_context *ctx = event->ctx;
1630 struct perf_event *leader = event->group_leader;
1631 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1634 if (WARN_ON_ONCE(!ctx->is_active))
1637 raw_spin_lock(&ctx->lock);
1638 update_context_time(ctx);
1640 if (event->state >= PERF_EVENT_STATE_INACTIVE)
1644 * set current task's cgroup time reference point
1646 perf_cgroup_set_timestamp(current, ctx);
1648 __perf_event_mark_enabled(event, ctx);
1650 if (!event_filter_match(event)) {
1651 if (is_cgroup_event(event))
1652 perf_cgroup_defer_enabled(event);
1657 * If the event is in a group and isn't the group leader,
1658 * then don't put it on unless the group is on.
1660 if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE)
1663 if (!group_can_go_on(event, cpuctx, 1)) {
1666 if (event == leader)
1667 err = group_sched_in(event, cpuctx, ctx);
1669 err = event_sched_in(event, cpuctx, ctx);
1674 * If this event can't go on and it's part of a
1675 * group, then the whole group has to come off.
1677 if (leader != event)
1678 group_sched_out(leader, cpuctx, ctx);
1679 if (leader->attr.pinned) {
1680 update_group_times(leader);
1681 leader->state = PERF_EVENT_STATE_ERROR;
1686 raw_spin_unlock(&ctx->lock);
1694 * If event->ctx is a cloned context, callers must make sure that
1695 * every task struct that event->ctx->task could possibly point to
1696 * remains valid. This condition is satisfied when called through
1697 * perf_event_for_each_child or perf_event_for_each as described
1698 * for perf_event_disable.
1700 void perf_event_enable(struct perf_event *event)
1702 struct perf_event_context *ctx = event->ctx;
1703 struct task_struct *task = ctx->task;
1707 * Enable the event on the cpu that it's on
1709 cpu_function_call(event->cpu, __perf_event_enable, event);
1713 raw_spin_lock_irq(&ctx->lock);
1714 if (event->state >= PERF_EVENT_STATE_INACTIVE)
1718 * If the event is in error state, clear that first.
1719 * That way, if we see the event in error state below, we
1720 * know that it has gone back into error state, as distinct
1721 * from the task having been scheduled away before the
1722 * cross-call arrived.
1724 if (event->state == PERF_EVENT_STATE_ERROR)
1725 event->state = PERF_EVENT_STATE_OFF;
1728 if (!ctx->is_active) {
1729 __perf_event_mark_enabled(event, ctx);
1733 raw_spin_unlock_irq(&ctx->lock);
1735 if (!task_function_call(task, __perf_event_enable, event))
1738 raw_spin_lock_irq(&ctx->lock);
1741 * If the context is active and the event is still off,
1742 * we need to retry the cross-call.
1744 if (ctx->is_active && event->state == PERF_EVENT_STATE_OFF) {
1746 * task could have been flipped by a concurrent
1747 * perf_event_context_sched_out()
1754 raw_spin_unlock_irq(&ctx->lock);
1757 static int perf_event_refresh(struct perf_event *event, int refresh)
1760 * not supported on inherited events
1762 if (event->attr.inherit || !is_sampling_event(event))
1765 atomic_add(refresh, &event->event_limit);
1766 perf_event_enable(event);
1771 static void ctx_sched_out(struct perf_event_context *ctx,
1772 struct perf_cpu_context *cpuctx,
1773 enum event_type_t event_type)
1775 struct perf_event *event;
1778 if (likely(!ctx->nr_events))
1781 update_context_time(ctx);
1782 update_cgrp_time_from_cpuctx(cpuctx);
1783 if (!ctx->nr_active)
1786 perf_pmu_disable(ctx->pmu);
1787 if (event_type & EVENT_PINNED) {
1788 list_for_each_entry(event, &ctx->pinned_groups, group_entry)
1789 group_sched_out(event, cpuctx, ctx);
1792 if (event_type & EVENT_FLEXIBLE) {
1793 list_for_each_entry(event, &ctx->flexible_groups, group_entry)
1794 group_sched_out(event, cpuctx, ctx);
1796 perf_pmu_enable(ctx->pmu);
1800 * Test whether two contexts are equivalent, i.e. whether they
1801 * have both been cloned from the same version of the same context
1802 * and they both have the same number of enabled events.
1803 * If the number of enabled events is the same, then the set
1804 * of enabled events should be the same, because these are both
1805 * inherited contexts, therefore we can't access individual events
1806 * in them directly with an fd; we can only enable/disable all
1807 * events via prctl, or enable/disable all events in a family
1808 * via ioctl, which will have the same effect on both contexts.
1810 static int context_equiv(struct perf_event_context *ctx1,
1811 struct perf_event_context *ctx2)
1813 return ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx
1814 && ctx1->parent_gen == ctx2->parent_gen
1815 && !ctx1->pin_count && !ctx2->pin_count;
1818 static void __perf_event_sync_stat(struct perf_event *event,
1819 struct perf_event *next_event)
1823 if (!event->attr.inherit_stat)
1827 * Update the event value, we cannot use perf_event_read()
1828 * because we're in the middle of a context switch and have IRQs
1829 * disabled, which upsets smp_call_function_single(), however
1830 * we know the event must be on the current CPU, therefore we
1831 * don't need to use it.
1833 switch (event->state) {
1834 case PERF_EVENT_STATE_ACTIVE:
1835 event->pmu->read(event);
1838 case PERF_EVENT_STATE_INACTIVE:
1839 update_event_times(event);
1847 * In order to keep per-task stats reliable we need to flip the event
1848 * values when we flip the contexts.
1850 value = local64_read(&next_event->count);
1851 value = local64_xchg(&event->count, value);
1852 local64_set(&next_event->count, value);
1854 swap(event->total_time_enabled, next_event->total_time_enabled);
1855 swap(event->total_time_running, next_event->total_time_running);
1858 * Since we swizzled the values, update the user visible data too.
1860 perf_event_update_userpage(event);
1861 perf_event_update_userpage(next_event);
1864 #define list_next_entry(pos, member) \
1865 list_entry(pos->member.next, typeof(*pos), member)
1867 static void perf_event_sync_stat(struct perf_event_context *ctx,
1868 struct perf_event_context *next_ctx)
1870 struct perf_event *event, *next_event;
1875 update_context_time(ctx);
1877 event = list_first_entry(&ctx->event_list,
1878 struct perf_event, event_entry);
1880 next_event = list_first_entry(&next_ctx->event_list,
1881 struct perf_event, event_entry);
1883 while (&event->event_entry != &ctx->event_list &&
1884 &next_event->event_entry != &next_ctx->event_list) {
1886 __perf_event_sync_stat(event, next_event);
1888 event = list_next_entry(event, event_entry);
1889 next_event = list_next_entry(next_event, event_entry);
1893 static void perf_event_context_sched_out(struct task_struct *task, int ctxn,
1894 struct task_struct *next)
1896 struct perf_event_context *ctx = task->perf_event_ctxp[ctxn];
1897 struct perf_event_context *next_ctx;
1898 struct perf_event_context *parent;
1899 struct perf_cpu_context *cpuctx;
1905 cpuctx = __get_cpu_context(ctx);
1906 if (!cpuctx->task_ctx)
1910 parent = rcu_dereference(ctx->parent_ctx);
1911 next_ctx = next->perf_event_ctxp[ctxn];
1912 if (parent && next_ctx &&
1913 rcu_dereference(next_ctx->parent_ctx) == parent) {
1915 * Looks like the two contexts are clones, so we might be
1916 * able to optimize the context switch. We lock both
1917 * contexts and check that they are clones under the
1918 * lock (including re-checking that neither has been
1919 * uncloned in the meantime). It doesn't matter which
1920 * order we take the locks because no other cpu could
1921 * be trying to lock both of these tasks.
1923 raw_spin_lock(&ctx->lock);
1924 raw_spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING);
1925 if (context_equiv(ctx, next_ctx)) {
1927 * XXX do we need a memory barrier of sorts
1928 * wrt to rcu_dereference() of perf_event_ctxp
1930 task->perf_event_ctxp[ctxn] = next_ctx;
1931 next->perf_event_ctxp[ctxn] = ctx;
1933 next_ctx->task = task;
1936 perf_event_sync_stat(ctx, next_ctx);
1938 raw_spin_unlock(&next_ctx->lock);
1939 raw_spin_unlock(&ctx->lock);
1944 raw_spin_lock(&ctx->lock);
1945 ctx_sched_out(ctx, cpuctx, EVENT_ALL);
1946 cpuctx->task_ctx = NULL;
1947 raw_spin_unlock(&ctx->lock);
1951 #define for_each_task_context_nr(ctxn) \
1952 for ((ctxn) = 0; (ctxn) < perf_nr_task_contexts; (ctxn)++)
1955 * Called from scheduler to remove the events of the current task,
1956 * with interrupts disabled.
1958 * We stop each event and update the event value in event->count.
1960 * This does not protect us against NMI, but disable()
1961 * sets the disabled bit in the control field of event _before_
1962 * accessing the event control register. If a NMI hits, then it will
1963 * not restart the event.
1965 void __perf_event_task_sched_out(struct task_struct *task,
1966 struct task_struct *next)
1970 for_each_task_context_nr(ctxn)
1971 perf_event_context_sched_out(task, ctxn, next);
1974 * if cgroup events exist on this CPU, then we need
1975 * to check if we have to switch out PMU state.
1976 * cgroup event are system-wide mode only
1978 if (atomic_read(&__get_cpu_var(perf_cgroup_events)))
1979 perf_cgroup_sched_out(task);
1982 static void task_ctx_sched_out(struct perf_event_context *ctx,
1983 enum event_type_t event_type)
1985 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1987 if (!cpuctx->task_ctx)
1990 if (WARN_ON_ONCE(ctx != cpuctx->task_ctx))
1993 ctx_sched_out(ctx, cpuctx, event_type);
1994 cpuctx->task_ctx = NULL;
1998 * Called with IRQs disabled
2000 static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
2001 enum event_type_t event_type)
2003 ctx_sched_out(&cpuctx->ctx, cpuctx, event_type);
2007 ctx_pinned_sched_in(struct perf_event_context *ctx,
2008 struct perf_cpu_context *cpuctx)
2010 struct perf_event *event;
2012 list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
2013 if (event->state <= PERF_EVENT_STATE_OFF)
2015 if (!event_filter_match(event))
2018 /* may need to reset tstamp_enabled */
2019 if (is_cgroup_event(event))
2020 perf_cgroup_mark_enabled(event, ctx);
2022 if (group_can_go_on(event, cpuctx, 1))
2023 group_sched_in(event, cpuctx, ctx);
2026 * If this pinned group hasn't been scheduled,
2027 * put it in error state.
2029 if (event->state == PERF_EVENT_STATE_INACTIVE) {
2030 update_group_times(event);
2031 event->state = PERF_EVENT_STATE_ERROR;
2037 ctx_flexible_sched_in(struct perf_event_context *ctx,
2038 struct perf_cpu_context *cpuctx)
2040 struct perf_event *event;
2043 list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
2044 /* Ignore events in OFF or ERROR state */
2045 if (event->state <= PERF_EVENT_STATE_OFF)
2048 * Listen to the 'cpu' scheduling filter constraint
2051 if (!event_filter_match(event))
2054 /* may need to reset tstamp_enabled */
2055 if (is_cgroup_event(event))
2056 perf_cgroup_mark_enabled(event, ctx);
2058 if (group_can_go_on(event, cpuctx, can_add_hw)) {
2059 if (group_sched_in(event, cpuctx, ctx))
2066 ctx_sched_in(struct perf_event_context *ctx,
2067 struct perf_cpu_context *cpuctx,
2068 enum event_type_t event_type,
2069 struct task_struct *task)
2074 if (likely(!ctx->nr_events))
2078 ctx->timestamp = now;
2079 perf_cgroup_set_timestamp(task, ctx);
2081 * First go through the list and put on any pinned groups
2082 * in order to give them the best chance of going on.
2084 if (event_type & EVENT_PINNED)
2085 ctx_pinned_sched_in(ctx, cpuctx);
2087 /* Then walk through the lower prio flexible groups */
2088 if (event_type & EVENT_FLEXIBLE)
2089 ctx_flexible_sched_in(ctx, cpuctx);
2092 static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
2093 enum event_type_t event_type,
2094 struct task_struct *task)
2096 struct perf_event_context *ctx = &cpuctx->ctx;
2098 ctx_sched_in(ctx, cpuctx, event_type, task);
2101 static void task_ctx_sched_in(struct perf_event_context *ctx,
2102 enum event_type_t event_type)
2104 struct perf_cpu_context *cpuctx;
2106 cpuctx = __get_cpu_context(ctx);
2107 if (cpuctx->task_ctx == ctx)
2110 ctx_sched_in(ctx, cpuctx, event_type, NULL);
2111 cpuctx->task_ctx = ctx;
2114 static void perf_event_context_sched_in(struct perf_event_context *ctx,
2115 struct task_struct *task)
2117 struct perf_cpu_context *cpuctx;
2119 cpuctx = __get_cpu_context(ctx);
2120 if (cpuctx->task_ctx == ctx)
2123 perf_ctx_lock(cpuctx, ctx);
2124 perf_pmu_disable(ctx->pmu);
2126 * We want to keep the following priority order:
2127 * cpu pinned (that don't need to move), task pinned,
2128 * cpu flexible, task flexible.
2130 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
2132 ctx_sched_in(ctx, cpuctx, EVENT_PINNED, task);
2133 cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE, task);
2134 ctx_sched_in(ctx, cpuctx, EVENT_FLEXIBLE, task);
2136 cpuctx->task_ctx = ctx;
2138 perf_pmu_enable(ctx->pmu);
2139 perf_ctx_unlock(cpuctx, ctx);
2142 * Since these rotations are per-cpu, we need to ensure the
2143 * cpu-context we got scheduled on is actually rotating.
2145 perf_pmu_rotate_start(ctx->pmu);
2149 * Called from scheduler to add the events of the current task
2150 * with interrupts disabled.
2152 * We restore the event value and then enable it.
2154 * This does not protect us against NMI, but enable()
2155 * sets the enabled bit in the control field of event _before_
2156 * accessing the event control register. If a NMI hits, then it will
2157 * keep the event running.
2159 void __perf_event_task_sched_in(struct task_struct *task)
2161 struct perf_event_context *ctx;
2164 for_each_task_context_nr(ctxn) {
2165 ctx = task->perf_event_ctxp[ctxn];
2169 perf_event_context_sched_in(ctx, task);
2172 * if cgroup events exist on this CPU, then we need
2173 * to check if we have to switch in PMU state.
2174 * cgroup event are system-wide mode only
2176 if (atomic_read(&__get_cpu_var(perf_cgroup_events)))
2177 perf_cgroup_sched_in(task);
2180 static u64 perf_calculate_period(struct perf_event *event, u64 nsec, u64 count)
2182 u64 frequency = event->attr.sample_freq;
2183 u64 sec = NSEC_PER_SEC;
2184 u64 divisor, dividend;
2186 int count_fls, nsec_fls, frequency_fls, sec_fls;
2188 count_fls = fls64(count);
2189 nsec_fls = fls64(nsec);
2190 frequency_fls = fls64(frequency);
2194 * We got @count in @nsec, with a target of sample_freq HZ
2195 * the target period becomes:
2198 * period = -------------------
2199 * @nsec * sample_freq
2204 * Reduce accuracy by one bit such that @a and @b converge
2205 * to a similar magnitude.
2207 #define REDUCE_FLS(a, b) \
2209 if (a##_fls > b##_fls) { \
2219 * Reduce accuracy until either term fits in a u64, then proceed with
2220 * the other, so that finally we can do a u64/u64 division.
2222 while (count_fls + sec_fls > 64 && nsec_fls + frequency_fls > 64) {
2223 REDUCE_FLS(nsec, frequency);
2224 REDUCE_FLS(sec, count);
2227 if (count_fls + sec_fls > 64) {
2228 divisor = nsec * frequency;
2230 while (count_fls + sec_fls > 64) {
2231 REDUCE_FLS(count, sec);
2235 dividend = count * sec;
2237 dividend = count * sec;
2239 while (nsec_fls + frequency_fls > 64) {
2240 REDUCE_FLS(nsec, frequency);
2244 divisor = nsec * frequency;
2250 return div64_u64(dividend, divisor);
2253 static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count)
2255 struct hw_perf_event *hwc = &event->hw;
2256 s64 period, sample_period;
2259 period = perf_calculate_period(event, nsec, count);
2261 delta = (s64)(period - hwc->sample_period);
2262 delta = (delta + 7) / 8; /* low pass filter */
2264 sample_period = hwc->sample_period + delta;
2269 hwc->sample_period = sample_period;
2271 if (local64_read(&hwc->period_left) > 8*sample_period) {
2272 event->pmu->stop(event, PERF_EF_UPDATE);
2273 local64_set(&hwc->period_left, 0);
2274 event->pmu->start(event, PERF_EF_RELOAD);
2278 static void perf_ctx_adjust_freq(struct perf_event_context *ctx, u64 period)
2280 struct perf_event *event;
2281 struct hw_perf_event *hwc;
2282 u64 interrupts, now;
2285 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
2286 if (event->state != PERF_EVENT_STATE_ACTIVE)
2289 if (!event_filter_match(event))
2294 interrupts = hwc->interrupts;
2295 hwc->interrupts = 0;
2298 * unthrottle events on the tick
2300 if (interrupts == MAX_INTERRUPTS) {
2301 perf_log_throttle(event, 1);
2302 event->pmu->start(event, 0);
2305 if (!event->attr.freq || !event->attr.sample_freq)
2308 event->pmu->read(event);
2309 now = local64_read(&event->count);
2310 delta = now - hwc->freq_count_stamp;
2311 hwc->freq_count_stamp = now;
2314 perf_adjust_period(event, period, delta);
2319 * Round-robin a context's events:
2321 static void rotate_ctx(struct perf_event_context *ctx)
2324 * Rotate the first entry last of non-pinned groups. Rotation might be
2325 * disabled by the inheritance code.
2327 if (!ctx->rotate_disable)
2328 list_rotate_left(&ctx->flexible_groups);
2332 * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
2333 * because they're strictly cpu affine and rotate_start is called with IRQs
2334 * disabled, while rotate_context is called from IRQ context.
2336 static void perf_rotate_context(struct perf_cpu_context *cpuctx)
2338 u64 interval = (u64)cpuctx->jiffies_interval * TICK_NSEC;
2339 struct perf_event_context *ctx = NULL;
2340 int rotate = 0, remove = 1;
2342 if (cpuctx->ctx.nr_events) {
2344 if (cpuctx->ctx.nr_events != cpuctx->ctx.nr_active)
2348 ctx = cpuctx->task_ctx;
2349 if (ctx && ctx->nr_events) {
2351 if (ctx->nr_events != ctx->nr_active)
2355 perf_ctx_lock(cpuctx, cpuctx->task_ctx);
2356 perf_pmu_disable(cpuctx->ctx.pmu);
2357 perf_ctx_adjust_freq(&cpuctx->ctx, interval);
2359 perf_ctx_adjust_freq(ctx, interval);
2364 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
2366 task_ctx_sched_out(ctx, EVENT_FLEXIBLE);
2368 rotate_ctx(&cpuctx->ctx);
2372 cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE, current);
2374 task_ctx_sched_in(ctx, EVENT_FLEXIBLE);
2378 list_del_init(&cpuctx->rotation_list);
2380 perf_pmu_enable(cpuctx->ctx.pmu);
2381 perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
2384 void perf_event_task_tick(void)
2386 struct list_head *head = &__get_cpu_var(rotation_list);
2387 struct perf_cpu_context *cpuctx, *tmp;
2389 WARN_ON(!irqs_disabled());
2391 list_for_each_entry_safe(cpuctx, tmp, head, rotation_list) {
2392 if (cpuctx->jiffies_interval == 1 ||
2393 !(jiffies % cpuctx->jiffies_interval))
2394 perf_rotate_context(cpuctx);
2398 static int event_enable_on_exec(struct perf_event *event,
2399 struct perf_event_context *ctx)
2401 if (!event->attr.enable_on_exec)
2404 event->attr.enable_on_exec = 0;
2405 if (event->state >= PERF_EVENT_STATE_INACTIVE)
2408 __perf_event_mark_enabled(event, ctx);
2414 * Enable all of a task's events that have been marked enable-on-exec.
2415 * This expects task == current.
2417 static void perf_event_enable_on_exec(struct perf_event_context *ctx)
2419 struct perf_event *event;
2420 unsigned long flags;
2424 local_irq_save(flags);
2425 if (!ctx || !ctx->nr_events)
2429 * We must ctxsw out cgroup events to avoid conflict
2430 * when invoking perf_task_event_sched_in() later on
2431 * in this function. Otherwise we end up trying to
2432 * ctxswin cgroup events which are already scheduled
2435 perf_cgroup_sched_out(current);
2437 raw_spin_lock(&ctx->lock);
2438 task_ctx_sched_out(ctx, EVENT_ALL);
2440 list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
2441 ret = event_enable_on_exec(event, ctx);
2446 list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
2447 ret = event_enable_on_exec(event, ctx);
2453 * Unclone this context if we enabled any event.
2458 raw_spin_unlock(&ctx->lock);
2461 * Also calls ctxswin for cgroup events, if any:
2463 perf_event_context_sched_in(ctx, ctx->task);
2465 local_irq_restore(flags);
2469 * Cross CPU call to read the hardware event
2471 static void __perf_event_read(void *info)
2473 struct perf_event *event = info;
2474 struct perf_event_context *ctx = event->ctx;
2475 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
2478 * If this is a task context, we need to check whether it is
2479 * the current task context of this cpu. If not it has been
2480 * scheduled out before the smp call arrived. In that case
2481 * event->count would have been updated to a recent sample
2482 * when the event was scheduled out.
2484 if (ctx->task && cpuctx->task_ctx != ctx)
2487 raw_spin_lock(&ctx->lock);
2488 if (ctx->is_active) {
2489 update_context_time(ctx);
2490 update_cgrp_time_from_event(event);
2492 update_event_times(event);
2493 if (event->state == PERF_EVENT_STATE_ACTIVE)
2494 event->pmu->read(event);
2495 raw_spin_unlock(&ctx->lock);
2498 static inline u64 perf_event_count(struct perf_event *event)
2500 return local64_read(&event->count) + atomic64_read(&event->child_count);
2503 static u64 perf_event_read(struct perf_event *event)
2506 * If event is enabled and currently active on a CPU, update the
2507 * value in the event structure:
2509 if (event->state == PERF_EVENT_STATE_ACTIVE) {
2510 smp_call_function_single(event->oncpu,
2511 __perf_event_read, event, 1);
2512 } else if (event->state == PERF_EVENT_STATE_INACTIVE) {
2513 struct perf_event_context *ctx = event->ctx;
2514 unsigned long flags;
2516 raw_spin_lock_irqsave(&ctx->lock, flags);
2518 * may read while context is not active
2519 * (e.g., thread is blocked), in that case
2520 * we cannot update context time
2522 if (ctx->is_active) {
2523 update_context_time(ctx);
2524 update_cgrp_time_from_event(event);
2526 update_event_times(event);
2527 raw_spin_unlock_irqrestore(&ctx->lock, flags);
2530 return perf_event_count(event);
2537 struct callchain_cpus_entries {
2538 struct rcu_head rcu_head;
2539 struct perf_callchain_entry *cpu_entries[0];
2542 static DEFINE_PER_CPU(int, callchain_recursion[PERF_NR_CONTEXTS]);
2543 static atomic_t nr_callchain_events;
2544 static DEFINE_MUTEX(callchain_mutex);
2545 struct callchain_cpus_entries *callchain_cpus_entries;
2548 __weak void perf_callchain_kernel(struct perf_callchain_entry *entry,
2549 struct pt_regs *regs)
2553 __weak void perf_callchain_user(struct perf_callchain_entry *entry,
2554 struct pt_regs *regs)
2558 static void release_callchain_buffers_rcu(struct rcu_head *head)
2560 struct callchain_cpus_entries *entries;
2563 entries = container_of(head, struct callchain_cpus_entries, rcu_head);
2565 for_each_possible_cpu(cpu)
2566 kfree(entries->cpu_entries[cpu]);
2571 static void release_callchain_buffers(void)
2573 struct callchain_cpus_entries *entries;
2575 entries = callchain_cpus_entries;
2576 rcu_assign_pointer(callchain_cpus_entries, NULL);
2577 call_rcu(&entries->rcu_head, release_callchain_buffers_rcu);
2580 static int alloc_callchain_buffers(void)
2584 struct callchain_cpus_entries *entries;
2587 * We can't use the percpu allocation API for data that can be
2588 * accessed from NMI. Use a temporary manual per cpu allocation
2589 * until that gets sorted out.
2591 size = offsetof(struct callchain_cpus_entries, cpu_entries[nr_cpu_ids]);
2593 entries = kzalloc(size, GFP_KERNEL);
2597 size = sizeof(struct perf_callchain_entry) * PERF_NR_CONTEXTS;
2599 for_each_possible_cpu(cpu) {
2600 entries->cpu_entries[cpu] = kmalloc_node(size, GFP_KERNEL,
2602 if (!entries->cpu_entries[cpu])
2606 rcu_assign_pointer(callchain_cpus_entries, entries);
2611 for_each_possible_cpu(cpu)
2612 kfree(entries->cpu_entries[cpu]);
2618 static int get_callchain_buffers(void)
2623 mutex_lock(&callchain_mutex);
2625 count = atomic_inc_return(&nr_callchain_events);
2626 if (WARN_ON_ONCE(count < 1)) {
2632 /* If the allocation failed, give up */
2633 if (!callchain_cpus_entries)
2638 err = alloc_callchain_buffers();
2640 release_callchain_buffers();
2642 mutex_unlock(&callchain_mutex);
2647 static void put_callchain_buffers(void)
2649 if (atomic_dec_and_mutex_lock(&nr_callchain_events, &callchain_mutex)) {
2650 release_callchain_buffers();
2651 mutex_unlock(&callchain_mutex);
2655 static int get_recursion_context(int *recursion)
2663 else if (in_softirq())
2668 if (recursion[rctx])
2677 static inline void put_recursion_context(int *recursion, int rctx)
2683 static struct perf_callchain_entry *get_callchain_entry(int *rctx)
2686 struct callchain_cpus_entries *entries;
2688 *rctx = get_recursion_context(__get_cpu_var(callchain_recursion));
2692 entries = rcu_dereference(callchain_cpus_entries);
2696 cpu = smp_processor_id();
2698 return &entries->cpu_entries[cpu][*rctx];
2702 put_callchain_entry(int rctx)
2704 put_recursion_context(__get_cpu_var(callchain_recursion), rctx);
2707 static struct perf_callchain_entry *perf_callchain(struct pt_regs *regs)
2710 struct perf_callchain_entry *entry;
2713 entry = get_callchain_entry(&rctx);
2722 if (!user_mode(regs)) {
2723 perf_callchain_store(entry, PERF_CONTEXT_KERNEL);
2724 perf_callchain_kernel(entry, regs);
2726 regs = task_pt_regs(current);
2732 perf_callchain_store(entry, PERF_CONTEXT_USER);
2733 perf_callchain_user(entry, regs);
2737 put_callchain_entry(rctx);
2743 * Initialize the perf_event context in a task_struct:
2745 static void __perf_event_init_context(struct perf_event_context *ctx)
2747 raw_spin_lock_init(&ctx->lock);
2748 mutex_init(&ctx->mutex);
2749 INIT_LIST_HEAD(&ctx->pinned_groups);
2750 INIT_LIST_HEAD(&ctx->flexible_groups);
2751 INIT_LIST_HEAD(&ctx->event_list);
2752 atomic_set(&ctx->refcount, 1);
2755 static struct perf_event_context *
2756 alloc_perf_context(struct pmu *pmu, struct task_struct *task)
2758 struct perf_event_context *ctx;
2760 ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL);
2764 __perf_event_init_context(ctx);
2767 get_task_struct(task);
2774 static struct task_struct *
2775 find_lively_task_by_vpid(pid_t vpid)
2777 struct task_struct *task;
2784 task = find_task_by_vpid(vpid);
2786 get_task_struct(task);
2790 return ERR_PTR(-ESRCH);
2792 /* Reuse ptrace permission checks for now. */
2794 if (!ptrace_may_access(task, PTRACE_MODE_READ))
2799 put_task_struct(task);
2800 return ERR_PTR(err);
2805 * Returns a matching context with refcount and pincount.
2807 static struct perf_event_context *
2808 find_get_context(struct pmu *pmu, struct task_struct *task, int cpu)
2810 struct perf_event_context *ctx;
2811 struct perf_cpu_context *cpuctx;
2812 unsigned long flags;
2816 /* Must be root to operate on a CPU event: */
2817 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
2818 return ERR_PTR(-EACCES);
2821 * We could be clever and allow to attach a event to an
2822 * offline CPU and activate it when the CPU comes up, but
2825 if (!cpu_online(cpu))
2826 return ERR_PTR(-ENODEV);
2828 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
2837 ctxn = pmu->task_ctx_nr;
2842 ctx = perf_lock_task_context(task, ctxn, &flags);
2846 raw_spin_unlock_irqrestore(&ctx->lock, flags);
2848 ctx = alloc_perf_context(pmu, task);
2854 mutex_lock(&task->perf_event_mutex);
2856 * If it has already passed perf_event_exit_task().
2857 * we must see PF_EXITING, it takes this mutex too.
2859 if (task->flags & PF_EXITING)
2861 else if (task->perf_event_ctxp[ctxn])
2866 rcu_assign_pointer(task->perf_event_ctxp[ctxn], ctx);
2868 mutex_unlock(&task->perf_event_mutex);
2870 if (unlikely(err)) {
2882 return ERR_PTR(err);
2885 static void perf_event_free_filter(struct perf_event *event);
2887 static void free_event_rcu(struct rcu_head *head)
2889 struct perf_event *event;
2891 event = container_of(head, struct perf_event, rcu_head);
2893 put_pid_ns(event->ns);
2894 perf_event_free_filter(event);
2898 static void perf_buffer_put(struct perf_buffer *buffer);
2900 static void free_event(struct perf_event *event)
2902 irq_work_sync(&event->pending);
2904 if (!event->parent) {
2905 if (event->attach_state & PERF_ATTACH_TASK)
2906 jump_label_dec(&perf_sched_events);
2907 if (event->attr.mmap || event->attr.mmap_data)
2908 atomic_dec(&nr_mmap_events);
2909 if (event->attr.comm)
2910 atomic_dec(&nr_comm_events);
2911 if (event->attr.task)
2912 atomic_dec(&nr_task_events);
2913 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN)
2914 put_callchain_buffers();
2915 if (is_cgroup_event(event)) {
2916 atomic_dec(&per_cpu(perf_cgroup_events, event->cpu));
2917 jump_label_dec(&perf_sched_events);
2921 if (event->buffer) {
2922 perf_buffer_put(event->buffer);
2923 event->buffer = NULL;
2926 if (is_cgroup_event(event))
2927 perf_detach_cgroup(event);
2930 event->destroy(event);
2933 put_ctx(event->ctx);
2935 call_rcu(&event->rcu_head, free_event_rcu);
2938 int perf_event_release_kernel(struct perf_event *event)
2940 struct perf_event_context *ctx = event->ctx;
2943 * Remove from the PMU, can't get re-enabled since we got
2944 * here because the last ref went.
2946 perf_event_disable(event);
2948 WARN_ON_ONCE(ctx->parent_ctx);
2950 * There are two ways this annotation is useful:
2952 * 1) there is a lock recursion from perf_event_exit_task
2953 * see the comment there.
2955 * 2) there is a lock-inversion with mmap_sem through
2956 * perf_event_read_group(), which takes faults while
2957 * holding ctx->mutex, however this is called after
2958 * the last filedesc died, so there is no possibility
2959 * to trigger the AB-BA case.
2961 mutex_lock_nested(&ctx->mutex, SINGLE_DEPTH_NESTING);
2962 raw_spin_lock_irq(&ctx->lock);
2963 perf_group_detach(event);
2964 list_del_event(event, ctx);
2965 raw_spin_unlock_irq(&ctx->lock);
2966 mutex_unlock(&ctx->mutex);
2972 EXPORT_SYMBOL_GPL(perf_event_release_kernel);
2975 * Called when the last reference to the file is gone.
2977 static int perf_release(struct inode *inode, struct file *file)
2979 struct perf_event *event = file->private_data;
2980 struct task_struct *owner;
2982 file->private_data = NULL;
2985 owner = ACCESS_ONCE(event->owner);
2987 * Matches the smp_wmb() in perf_event_exit_task(). If we observe
2988 * !owner it means the list deletion is complete and we can indeed
2989 * free this event, otherwise we need to serialize on
2990 * owner->perf_event_mutex.
2992 smp_read_barrier_depends();
2995 * Since delayed_put_task_struct() also drops the last
2996 * task reference we can safely take a new reference
2997 * while holding the rcu_read_lock().
2999 get_task_struct(owner);
3004 mutex_lock(&owner->perf_event_mutex);
3006 * We have to re-check the event->owner field, if it is cleared
3007 * we raced with perf_event_exit_task(), acquiring the mutex
3008 * ensured they're done, and we can proceed with freeing the
3012 list_del_init(&event->owner_entry);
3013 mutex_unlock(&owner->perf_event_mutex);
3014 put_task_struct(owner);
3017 return perf_event_release_kernel(event);
3020 u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running)
3022 struct perf_event *child;
3028 mutex_lock(&event->child_mutex);
3029 total += perf_event_read(event);
3030 *enabled += event->total_time_enabled +
3031 atomic64_read(&event->child_total_time_enabled);
3032 *running += event->total_time_running +
3033 atomic64_read(&event->child_total_time_running);
3035 list_for_each_entry(child, &event->child_list, child_list) {
3036 total += perf_event_read(child);
3037 *enabled += child->total_time_enabled;
3038 *running += child->total_time_running;
3040 mutex_unlock(&event->child_mutex);
3044 EXPORT_SYMBOL_GPL(perf_event_read_value);
3046 static int perf_event_read_group(struct perf_event *event,
3047 u64 read_format, char __user *buf)
3049 struct perf_event *leader = event->group_leader, *sub;
3050 int n = 0, size = 0, ret = -EFAULT;
3051 struct perf_event_context *ctx = leader->ctx;
3053 u64 count, enabled, running;
3055 mutex_lock(&ctx->mutex);
3056 count = perf_event_read_value(leader, &enabled, &running);
3058 values[n++] = 1 + leader->nr_siblings;
3059 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3060 values[n++] = enabled;
3061 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3062 values[n++] = running;
3063 values[n++] = count;
3064 if (read_format & PERF_FORMAT_ID)
3065 values[n++] = primary_event_id(leader);
3067 size = n * sizeof(u64);
3069 if (copy_to_user(buf, values, size))
3074 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
3077 values[n++] = perf_event_read_value(sub, &enabled, &running);
3078 if (read_format & PERF_FORMAT_ID)
3079 values[n++] = primary_event_id(sub);
3081 size = n * sizeof(u64);
3083 if (copy_to_user(buf + ret, values, size)) {
3091 mutex_unlock(&ctx->mutex);
3096 static int perf_event_read_one(struct perf_event *event,
3097 u64 read_format, char __user *buf)
3099 u64 enabled, running;
3103 values[n++] = perf_event_read_value(event, &enabled, &running);
3104 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3105 values[n++] = enabled;
3106 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3107 values[n++] = running;
3108 if (read_format & PERF_FORMAT_ID)
3109 values[n++] = primary_event_id(event);
3111 if (copy_to_user(buf, values, n * sizeof(u64)))
3114 return n * sizeof(u64);
3118 * Read the performance event - simple non blocking version for now
3121 perf_read_hw(struct perf_event *event, char __user *buf, size_t count)
3123 u64 read_format = event->attr.read_format;
3127 * Return end-of-file for a read on a event that is in
3128 * error state (i.e. because it was pinned but it couldn't be
3129 * scheduled on to the CPU at some point).
3131 if (event->state == PERF_EVENT_STATE_ERROR)
3134 if (count < event->read_size)
3137 WARN_ON_ONCE(event->ctx->parent_ctx);
3138 if (read_format & PERF_FORMAT_GROUP)
3139 ret = perf_event_read_group(event, read_format, buf);
3141 ret = perf_event_read_one(event, read_format, buf);
3147 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
3149 struct perf_event *event = file->private_data;
3151 return perf_read_hw(event, buf, count);
3154 static unsigned int perf_poll(struct file *file, poll_table *wait)
3156 struct perf_event *event = file->private_data;
3157 struct perf_buffer *buffer;
3158 unsigned int events = POLL_HUP;
3161 buffer = rcu_dereference(event->buffer);
3163 events = atomic_xchg(&buffer->poll, 0);
3166 poll_wait(file, &event->waitq, wait);
3171 static void perf_event_reset(struct perf_event *event)
3173 (void)perf_event_read(event);
3174 local64_set(&event->count, 0);
3175 perf_event_update_userpage(event);
3179 * Holding the top-level event's child_mutex means that any
3180 * descendant process that has inherited this event will block
3181 * in sync_child_event if it goes to exit, thus satisfying the
3182 * task existence requirements of perf_event_enable/disable.
3184 static void perf_event_for_each_child(struct perf_event *event,
3185 void (*func)(struct perf_event *))
3187 struct perf_event *child;
3189 WARN_ON_ONCE(event->ctx->parent_ctx);
3190 mutex_lock(&event->child_mutex);
3192 list_for_each_entry(child, &event->child_list, child_list)
3194 mutex_unlock(&event->child_mutex);
3197 static void perf_event_for_each(struct perf_event *event,
3198 void (*func)(struct perf_event *))
3200 struct perf_event_context *ctx = event->ctx;
3201 struct perf_event *sibling;
3203 WARN_ON_ONCE(ctx->parent_ctx);
3204 mutex_lock(&ctx->mutex);
3205 event = event->group_leader;
3207 perf_event_for_each_child(event, func);
3209 list_for_each_entry(sibling, &event->sibling_list, group_entry)
3210 perf_event_for_each_child(event, func);
3211 mutex_unlock(&ctx->mutex);
3214 static int perf_event_period(struct perf_event *event, u64 __user *arg)
3216 struct perf_event_context *ctx = event->ctx;
3220 if (!is_sampling_event(event))
3223 if (copy_from_user(&value, arg, sizeof(value)))
3229 raw_spin_lock_irq(&ctx->lock);
3230 if (event->attr.freq) {
3231 if (value > sysctl_perf_event_sample_rate) {
3236 event->attr.sample_freq = value;
3238 event->attr.sample_period = value;
3239 event->hw.sample_period = value;
3242 raw_spin_unlock_irq(&ctx->lock);
3247 static const struct file_operations perf_fops;
3249 static struct perf_event *perf_fget_light(int fd, int *fput_needed)
3253 file = fget_light(fd, fput_needed);
3255 return ERR_PTR(-EBADF);
3257 if (file->f_op != &perf_fops) {
3258 fput_light(file, *fput_needed);
3260 return ERR_PTR(-EBADF);
3263 return file->private_data;
3266 static int perf_event_set_output(struct perf_event *event,
3267 struct perf_event *output_event);
3268 static int perf_event_set_filter(struct perf_event *event, void __user *arg);
3270 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
3272 struct perf_event *event = file->private_data;
3273 void (*func)(struct perf_event *);
3277 case PERF_EVENT_IOC_ENABLE:
3278 func = perf_event_enable;
3280 case PERF_EVENT_IOC_DISABLE:
3281 func = perf_event_disable;
3283 case PERF_EVENT_IOC_RESET:
3284 func = perf_event_reset;
3287 case PERF_EVENT_IOC_REFRESH:
3288 return perf_event_refresh(event, arg);
3290 case PERF_EVENT_IOC_PERIOD:
3291 return perf_event_period(event, (u64 __user *)arg);
3293 case PERF_EVENT_IOC_SET_OUTPUT:
3295 struct perf_event *output_event = NULL;
3296 int fput_needed = 0;
3300 output_event = perf_fget_light(arg, &fput_needed);
3301 if (IS_ERR(output_event))
3302 return PTR_ERR(output_event);
3305 ret = perf_event_set_output(event, output_event);
3307 fput_light(output_event->filp, fput_needed);
3312 case PERF_EVENT_IOC_SET_FILTER:
3313 return perf_event_set_filter(event, (void __user *)arg);
3319 if (flags & PERF_IOC_FLAG_GROUP)
3320 perf_event_for_each(event, func);
3322 perf_event_for_each_child(event, func);
3327 int perf_event_task_enable(void)
3329 struct perf_event *event;
3331 mutex_lock(¤t->perf_event_mutex);
3332 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
3333 perf_event_for_each_child(event, perf_event_enable);
3334 mutex_unlock(¤t->perf_event_mutex);
3339 int perf_event_task_disable(void)
3341 struct perf_event *event;
3343 mutex_lock(¤t->perf_event_mutex);
3344 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
3345 perf_event_for_each_child(event, perf_event_disable);
3346 mutex_unlock(¤t->perf_event_mutex);
3351 #ifndef PERF_EVENT_INDEX_OFFSET
3352 # define PERF_EVENT_INDEX_OFFSET 0
3355 static int perf_event_index(struct perf_event *event)
3357 if (event->hw.state & PERF_HES_STOPPED)
3360 if (event->state != PERF_EVENT_STATE_ACTIVE)
3363 return event->hw.idx + 1 - PERF_EVENT_INDEX_OFFSET;
3367 * Callers need to ensure there can be no nesting of this function, otherwise
3368 * the seqlock logic goes bad. We can not serialize this because the arch
3369 * code calls this from NMI context.
3371 void perf_event_update_userpage(struct perf_event *event)
3373 struct perf_event_mmap_page *userpg;
3374 struct perf_buffer *buffer;
3377 buffer = rcu_dereference(event->buffer);
3381 userpg = buffer->user_page;
3384 * Disable preemption so as to not let the corresponding user-space
3385 * spin too long if we get preempted.
3390 userpg->index = perf_event_index(event);
3391 userpg->offset = perf_event_count(event);
3392 if (event->state == PERF_EVENT_STATE_ACTIVE)
3393 userpg->offset -= local64_read(&event->hw.prev_count);
3395 userpg->time_enabled = event->total_time_enabled +
3396 atomic64_read(&event->child_total_time_enabled);
3398 userpg->time_running = event->total_time_running +
3399 atomic64_read(&event->child_total_time_running);
3408 static unsigned long perf_data_size(struct perf_buffer *buffer);
3411 perf_buffer_init(struct perf_buffer *buffer, long watermark, int flags)
3413 long max_size = perf_data_size(buffer);
3416 buffer->watermark = min(max_size, watermark);
3418 if (!buffer->watermark)
3419 buffer->watermark = max_size / 2;
3421 if (flags & PERF_BUFFER_WRITABLE)
3422 buffer->writable = 1;
3424 atomic_set(&buffer->refcount, 1);
3427 #ifndef CONFIG_PERF_USE_VMALLOC
3430 * Back perf_mmap() with regular GFP_KERNEL-0 pages.
3433 static struct page *
3434 perf_mmap_to_page(struct perf_buffer *buffer, unsigned long pgoff)
3436 if (pgoff > buffer->nr_pages)
3440 return virt_to_page(buffer->user_page);
3442 return virt_to_page(buffer->data_pages[pgoff - 1]);
3445 static void *perf_mmap_alloc_page(int cpu)
3450 node = (cpu == -1) ? cpu : cpu_to_node(cpu);
3451 page = alloc_pages_node(node, GFP_KERNEL | __GFP_ZERO, 0);
3455 return page_address(page);
3458 static struct perf_buffer *
3459 perf_buffer_alloc(int nr_pages, long watermark, int cpu, int flags)
3461 struct perf_buffer *buffer;
3465 size = sizeof(struct perf_buffer);
3466 size += nr_pages * sizeof(void *);
3468 buffer = kzalloc(size, GFP_KERNEL);
3472 buffer->user_page = perf_mmap_alloc_page(cpu);
3473 if (!buffer->user_page)
3474 goto fail_user_page;
3476 for (i = 0; i < nr_pages; i++) {
3477 buffer->data_pages[i] = perf_mmap_alloc_page(cpu);
3478 if (!buffer->data_pages[i])
3479 goto fail_data_pages;
3482 buffer->nr_pages = nr_pages;
3484 perf_buffer_init(buffer, watermark, flags);
3489 for (i--; i >= 0; i--)
3490 free_page((unsigned long)buffer->data_pages[i]);
3492 free_page((unsigned long)buffer->user_page);
3501 static void perf_mmap_free_page(unsigned long addr)
3503 struct page *page = virt_to_page((void *)addr);
3505 page->mapping = NULL;
3509 static void perf_buffer_free(struct perf_buffer *buffer)
3513 perf_mmap_free_page((unsigned long)buffer->user_page);
3514 for (i = 0; i < buffer->nr_pages; i++)
3515 perf_mmap_free_page((unsigned long)buffer->data_pages[i]);
3519 static inline int page_order(struct perf_buffer *buffer)
3527 * Back perf_mmap() with vmalloc memory.
3529 * Required for architectures that have d-cache aliasing issues.
3532 static inline int page_order(struct perf_buffer *buffer)
3534 return buffer->page_order;
3537 static struct page *
3538 perf_mmap_to_page(struct perf_buffer *buffer, unsigned long pgoff)
3540 if (pgoff > (1UL << page_order(buffer)))
3543 return vmalloc_to_page((void *)buffer->user_page + pgoff * PAGE_SIZE);
3546 static void perf_mmap_unmark_page(void *addr)
3548 struct page *page = vmalloc_to_page(addr);
3550 page->mapping = NULL;
3553 static void perf_buffer_free_work(struct work_struct *work)
3555 struct perf_buffer *buffer;
3559 buffer = container_of(work, struct perf_buffer, work);
3560 nr = 1 << page_order(buffer);
3562 base = buffer->user_page;
3563 for (i = 0; i < nr + 1; i++)
3564 perf_mmap_unmark_page(base + (i * PAGE_SIZE));
3570 static void perf_buffer_free(struct perf_buffer *buffer)
3572 schedule_work(&buffer->work);
3575 static struct perf_buffer *
3576 perf_buffer_alloc(int nr_pages, long watermark, int cpu, int flags)
3578 struct perf_buffer *buffer;
3582 size = sizeof(struct perf_buffer);
3583 size += sizeof(void *);
3585 buffer = kzalloc(size, GFP_KERNEL);
3589 INIT_WORK(&buffer->work, perf_buffer_free_work);
3591 all_buf = vmalloc_user((nr_pages + 1) * PAGE_SIZE);
3595 buffer->user_page = all_buf;
3596 buffer->data_pages[0] = all_buf + PAGE_SIZE;
3597 buffer->page_order = ilog2(nr_pages);
3598 buffer->nr_pages = 1;
3600 perf_buffer_init(buffer, watermark, flags);
3613 static unsigned long perf_data_size(struct perf_buffer *buffer)
3615 return buffer->nr_pages << (PAGE_SHIFT + page_order(buffer));
3618 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
3620 struct perf_event *event = vma->vm_file->private_data;
3621 struct perf_buffer *buffer;
3622 int ret = VM_FAULT_SIGBUS;
3624 if (vmf->flags & FAULT_FLAG_MKWRITE) {
3625 if (vmf->pgoff == 0)
3631 buffer = rcu_dereference(event->buffer);
3635 if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE))
3638 vmf->page = perf_mmap_to_page(buffer, vmf->pgoff);
3642 get_page(vmf->page);
3643 vmf->page->mapping = vma->vm_file->f_mapping;
3644 vmf->page->index = vmf->pgoff;
3653 static void perf_buffer_free_rcu(struct rcu_head *rcu_head)
3655 struct perf_buffer *buffer;
3657 buffer = container_of(rcu_head, struct perf_buffer, rcu_head);
3658 perf_buffer_free(buffer);
3661 static struct perf_buffer *perf_buffer_get(struct perf_event *event)
3663 struct perf_buffer *buffer;
3666 buffer = rcu_dereference(event->buffer);
3668 if (!atomic_inc_not_zero(&buffer->refcount))
3676 static void perf_buffer_put(struct perf_buffer *buffer)
3678 if (!atomic_dec_and_test(&buffer->refcount))
3681 call_rcu(&buffer->rcu_head, perf_buffer_free_rcu);
3684 static void perf_mmap_open(struct vm_area_struct *vma)
3686 struct perf_event *event = vma->vm_file->private_data;
3688 atomic_inc(&event->mmap_count);
3691 static void perf_mmap_close(struct vm_area_struct *vma)
3693 struct perf_event *event = vma->vm_file->private_data;
3695 if (atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex)) {
3696 unsigned long size = perf_data_size(event->buffer);
3697 struct user_struct *user = event->mmap_user;
3698 struct perf_buffer *buffer = event->buffer;
3700 atomic_long_sub((size >> PAGE_SHIFT) + 1, &user->locked_vm);
3701 vma->vm_mm->locked_vm -= event->mmap_locked;
3702 rcu_assign_pointer(event->buffer, NULL);
3703 mutex_unlock(&event->mmap_mutex);
3705 perf_buffer_put(buffer);
3710 static const struct vm_operations_struct perf_mmap_vmops = {
3711 .open = perf_mmap_open,
3712 .close = perf_mmap_close,
3713 .fault = perf_mmap_fault,
3714 .page_mkwrite = perf_mmap_fault,
3717 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
3719 struct perf_event *event = file->private_data;
3720 unsigned long user_locked, user_lock_limit;
3721 struct user_struct *user = current_user();
3722 unsigned long locked, lock_limit;
3723 struct perf_buffer *buffer;
3724 unsigned long vma_size;
3725 unsigned long nr_pages;
3726 long user_extra, extra;
3727 int ret = 0, flags = 0;
3730 * Don't allow mmap() of inherited per-task counters. This would
3731 * create a performance issue due to all children writing to the
3734 if (event->cpu == -1 && event->attr.inherit)
3737 if (!(vma->vm_flags & VM_SHARED))
3740 vma_size = vma->vm_end - vma->vm_start;
3741 nr_pages = (vma_size / PAGE_SIZE) - 1;
3744 * If we have buffer pages ensure they're a power-of-two number, so we
3745 * can do bitmasks instead of modulo.
3747 if (nr_pages != 0 && !is_power_of_2(nr_pages))
3750 if (vma_size != PAGE_SIZE * (1 + nr_pages))
3753 if (vma->vm_pgoff != 0)
3756 WARN_ON_ONCE(event->ctx->parent_ctx);
3757 mutex_lock(&event->mmap_mutex);
3758 if (event->buffer) {
3759 if (event->buffer->nr_pages == nr_pages)
3760 atomic_inc(&event->buffer->refcount);
3766 user_extra = nr_pages + 1;
3767 user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10);
3770 * Increase the limit linearly with more CPUs:
3772 user_lock_limit *= num_online_cpus();
3774 user_locked = atomic_long_read(&user->locked_vm) + user_extra;
3777 if (user_locked > user_lock_limit)
3778 extra = user_locked - user_lock_limit;
3780 lock_limit = rlimit(RLIMIT_MEMLOCK);
3781 lock_limit >>= PAGE_SHIFT;
3782 locked = vma->vm_mm->locked_vm + extra;
3784 if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() &&
3785 !capable(CAP_IPC_LOCK)) {
3790 WARN_ON(event->buffer);
3792 if (vma->vm_flags & VM_WRITE)
3793 flags |= PERF_BUFFER_WRITABLE;
3795 buffer = perf_buffer_alloc(nr_pages, event->attr.wakeup_watermark,
3801 rcu_assign_pointer(event->buffer, buffer);
3803 atomic_long_add(user_extra, &user->locked_vm);
3804 event->mmap_locked = extra;
3805 event->mmap_user = get_current_user();
3806 vma->vm_mm->locked_vm += event->mmap_locked;
3810 atomic_inc(&event->mmap_count);
3811 mutex_unlock(&event->mmap_mutex);
3813 vma->vm_flags |= VM_RESERVED;
3814 vma->vm_ops = &perf_mmap_vmops;
3819 static int perf_fasync(int fd, struct file *filp, int on)
3821 struct inode *inode = filp->f_path.dentry->d_inode;
3822 struct perf_event *event = filp->private_data;
3825 mutex_lock(&inode->i_mutex);
3826 retval = fasync_helper(fd, filp, on, &event->fasync);
3827 mutex_unlock(&inode->i_mutex);
3835 static const struct file_operations perf_fops = {
3836 .llseek = no_llseek,
3837 .release = perf_release,
3840 .unlocked_ioctl = perf_ioctl,
3841 .compat_ioctl = perf_ioctl,
3843 .fasync = perf_fasync,
3849 * If there's data, ensure we set the poll() state and publish everything
3850 * to user-space before waking everybody up.
3853 void perf_event_wakeup(struct perf_event *event)
3855 wake_up_all(&event->waitq);
3857 if (event->pending_kill) {
3858 kill_fasync(&event->fasync, SIGIO, event->pending_kill);
3859 event->pending_kill = 0;
3863 static void perf_pending_event(struct irq_work *entry)
3865 struct perf_event *event = container_of(entry,
3866 struct perf_event, pending);
3868 if (event->pending_disable) {
3869 event->pending_disable = 0;
3870 __perf_event_disable(event);
3873 if (event->pending_wakeup) {
3874 event->pending_wakeup = 0;
3875 perf_event_wakeup(event);
3880 * We assume there is only KVM supporting the callbacks.
3881 * Later on, we might change it to a list if there is
3882 * another virtualization implementation supporting the callbacks.
3884 struct perf_guest_info_callbacks *perf_guest_cbs;
3886 int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
3888 perf_guest_cbs = cbs;
3891 EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks);
3893 int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
3895 perf_guest_cbs = NULL;
3898 EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks);
3903 static bool perf_output_space(struct perf_buffer *buffer, unsigned long tail,
3904 unsigned long offset, unsigned long head)
3908 if (!buffer->writable)
3911 mask = perf_data_size(buffer) - 1;
3913 offset = (offset - tail) & mask;
3914 head = (head - tail) & mask;
3916 if ((int)(head - offset) < 0)
3922 static void perf_output_wakeup(struct perf_output_handle *handle)
3924 atomic_set(&handle->buffer->poll, POLL_IN);
3927 handle->event->pending_wakeup = 1;
3928 irq_work_queue(&handle->event->pending);
3930 perf_event_wakeup(handle->event);
3934 * We need to ensure a later event_id doesn't publish a head when a former
3935 * event isn't done writing. However since we need to deal with NMIs we
3936 * cannot fully serialize things.
3938 * We only publish the head (and generate a wakeup) when the outer-most
3941 static void perf_output_get_handle(struct perf_output_handle *handle)
3943 struct perf_buffer *buffer = handle->buffer;
3946 local_inc(&buffer->nest);
3947 handle->wakeup = local_read(&buffer->wakeup);
3950 static void perf_output_put_handle(struct perf_output_handle *handle)
3952 struct perf_buffer *buffer = handle->buffer;
3956 head = local_read(&buffer->head);
3959 * IRQ/NMI can happen here, which means we can miss a head update.
3962 if (!local_dec_and_test(&buffer->nest))
3966 * Publish the known good head. Rely on the full barrier implied
3967 * by atomic_dec_and_test() order the buffer->head read and this
3970 buffer->user_page->data_head = head;
3973 * Now check if we missed an update, rely on the (compiler)
3974 * barrier in atomic_dec_and_test() to re-read buffer->head.
3976 if (unlikely(head != local_read(&buffer->head))) {
3977 local_inc(&buffer->nest);
3981 if (handle->wakeup != local_read(&buffer->wakeup))
3982 perf_output_wakeup(handle);
3988 __always_inline void perf_output_copy(struct perf_output_handle *handle,
3989 const void *buf, unsigned int len)
3992 unsigned long size = min_t(unsigned long, handle->size, len);
3994 memcpy(handle->addr, buf, size);
3997 handle->addr += size;
3999 handle->size -= size;
4000 if (!handle->size) {
4001 struct perf_buffer *buffer = handle->buffer;
4004 handle->page &= buffer->nr_pages - 1;
4005 handle->addr = buffer->data_pages[handle->page];
4006 handle->size = PAGE_SIZE << page_order(buffer);
4011 static void __perf_event_header__init_id(struct perf_event_header *header,
4012 struct perf_sample_data *data,
4013 struct perf_event *event)
4015 u64 sample_type = event->attr.sample_type;
4017 data->type = sample_type;
4018 header->size += event->id_header_size;
4020 if (sample_type & PERF_SAMPLE_TID) {
4021 /* namespace issues */
4022 data->tid_entry.pid = perf_event_pid(event, current);
4023 data->tid_entry.tid = perf_event_tid(event, current);
4026 if (sample_type & PERF_SAMPLE_TIME)
4027 data->time = perf_clock();
4029 if (sample_type & PERF_SAMPLE_ID)
4030 data->id = primary_event_id(event);
4032 if (sample_type & PERF_SAMPLE_STREAM_ID)
4033 data->stream_id = event->id;
4035 if (sample_type & PERF_SAMPLE_CPU) {
4036 data->cpu_entry.cpu = raw_smp_processor_id();
4037 data->cpu_entry.reserved = 0;
4041 static void perf_event_header__init_id(struct perf_event_header *header,
4042 struct perf_sample_data *data,
4043 struct perf_event *event)
4045 if (event->attr.sample_id_all)
4046 __perf_event_header__init_id(header, data, event);
4049 static void __perf_event__output_id_sample(struct perf_output_handle *handle,
4050 struct perf_sample_data *data)
4052 u64 sample_type = data->type;
4054 if (sample_type & PERF_SAMPLE_TID)
4055 perf_output_put(handle, data->tid_entry);
4057 if (sample_type & PERF_SAMPLE_TIME)
4058 perf_output_put(handle, data->time);
4060 if (sample_type & PERF_SAMPLE_ID)
4061 perf_output_put(handle, data->id);
4063 if (sample_type & PERF_SAMPLE_STREAM_ID)
4064 perf_output_put(handle, data->stream_id);
4066 if (sample_type & PERF_SAMPLE_CPU)
4067 perf_output_put(handle, data->cpu_entry);
4070 static void perf_event__output_id_sample(struct perf_event *event,
4071 struct perf_output_handle *handle,
4072 struct perf_sample_data *sample)
4074 if (event->attr.sample_id_all)
4075 __perf_event__output_id_sample(handle, sample);
4078 int perf_output_begin(struct perf_output_handle *handle,
4079 struct perf_event *event, unsigned int size,
4080 int nmi, int sample)
4082 struct perf_buffer *buffer;
4083 unsigned long tail, offset, head;
4085 struct perf_sample_data sample_data;
4087 struct perf_event_header header;
4094 * For inherited events we send all the output towards the parent.
4097 event = event->parent;
4099 buffer = rcu_dereference(event->buffer);
4103 handle->buffer = buffer;
4104 handle->event = event;
4106 handle->sample = sample;
4108 if (!buffer->nr_pages)
4111 have_lost = local_read(&buffer->lost);
4113 lost_event.header.size = sizeof(lost_event);
4114 perf_event_header__init_id(&lost_event.header, &sample_data,
4116 size += lost_event.header.size;
4119 perf_output_get_handle(handle);
4123 * Userspace could choose to issue a mb() before updating the
4124 * tail pointer. So that all reads will be completed before the
4127 tail = ACCESS_ONCE(buffer->user_page->data_tail);
4129 offset = head = local_read(&buffer->head);
4131 if (unlikely(!perf_output_space(buffer, tail, offset, head)))
4133 } while (local_cmpxchg(&buffer->head, offset, head) != offset);
4135 if (head - local_read(&buffer->wakeup) > buffer->watermark)
4136 local_add(buffer->watermark, &buffer->wakeup);
4138 handle->page = offset >> (PAGE_SHIFT + page_order(buffer));
4139 handle->page &= buffer->nr_pages - 1;
4140 handle->size = offset & ((PAGE_SIZE << page_order(buffer)) - 1);
4141 handle->addr = buffer->data_pages[handle->page];
4142 handle->addr += handle->size;
4143 handle->size = (PAGE_SIZE << page_order(buffer)) - handle->size;
4146 lost_event.header.type = PERF_RECORD_LOST;
4147 lost_event.header.misc = 0;
4148 lost_event.id = event->id;
4149 lost_event.lost = local_xchg(&buffer->lost, 0);
4151 perf_output_put(handle, lost_event);
4152 perf_event__output_id_sample(event, handle, &sample_data);
4158 local_inc(&buffer->lost);
4159 perf_output_put_handle(handle);
4166 void perf_output_end(struct perf_output_handle *handle)
4168 struct perf_event *event = handle->event;
4169 struct perf_buffer *buffer = handle->buffer;
4171 int wakeup_events = event->attr.wakeup_events;
4173 if (handle->sample && wakeup_events) {
4174 int events = local_inc_return(&buffer->events);
4175 if (events >= wakeup_events) {
4176 local_sub(wakeup_events, &buffer->events);
4177 local_inc(&buffer->wakeup);
4181 perf_output_put_handle(handle);
4185 static void perf_output_read_one(struct perf_output_handle *handle,
4186 struct perf_event *event,
4187 u64 enabled, u64 running)
4189 u64 read_format = event->attr.read_format;
4193 values[n++] = perf_event_count(event);
4194 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
4195 values[n++] = enabled +
4196 atomic64_read(&event->child_total_time_enabled);
4198 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
4199 values[n++] = running +
4200 atomic64_read(&event->child_total_time_running);
4202 if (read_format & PERF_FORMAT_ID)
4203 values[n++] = primary_event_id(event);
4205 perf_output_copy(handle, values, n * sizeof(u64));
4209 * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
4211 static void perf_output_read_group(struct perf_output_handle *handle,
4212 struct perf_event *event,
4213 u64 enabled, u64 running)
4215 struct perf_event *leader = event->group_leader, *sub;
4216 u64 read_format = event->attr.read_format;
4220 values[n++] = 1 + leader->nr_siblings;
4222 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
4223 values[n++] = enabled;
4225 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
4226 values[n++] = running;
4228 if (leader != event)
4229 leader->pmu->read(leader);
4231 values[n++] = perf_event_count(leader);
4232 if (read_format & PERF_FORMAT_ID)
4233 values[n++] = primary_event_id(leader);
4235 perf_output_copy(handle, values, n * sizeof(u64));
4237 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
4241 sub->pmu->read(sub);
4243 values[n++] = perf_event_count(sub);
4244 if (read_format & PERF_FORMAT_ID)
4245 values[n++] = primary_event_id(sub);
4247 perf_output_copy(handle, values, n * sizeof(u64));
4251 #define PERF_FORMAT_TOTAL_TIMES (PERF_FORMAT_TOTAL_TIME_ENABLED|\
4252 PERF_FORMAT_TOTAL_TIME_RUNNING)
4254 static void perf_output_read(struct perf_output_handle *handle,
4255 struct perf_event *event)
4257 u64 enabled = 0, running = 0, now, ctx_time;
4258 u64 read_format = event->attr.read_format;
4261 * compute total_time_enabled, total_time_running
4262 * based on snapshot values taken when the event
4263 * was last scheduled in.
4265 * we cannot simply called update_context_time()
4266 * because of locking issue as we are called in
4269 if (read_format & PERF_FORMAT_TOTAL_TIMES) {
4271 ctx_time = event->shadow_ctx_time + now;
4272 enabled = ctx_time - event->tstamp_enabled;
4273 running = ctx_time - event->tstamp_running;
4276 if (event->attr.read_format & PERF_FORMAT_GROUP)
4277 perf_output_read_group(handle, event, enabled, running);
4279 perf_output_read_one(handle, event, enabled, running);
4282 void perf_output_sample(struct perf_output_handle *handle,
4283 struct perf_event_header *header,
4284 struct perf_sample_data *data,
4285 struct perf_event *event)
4287 u64 sample_type = data->type;
4289 perf_output_put(handle, *header);
4291 if (sample_type & PERF_SAMPLE_IP)
4292 perf_output_put(handle, data->ip);
4294 if (sample_type & PERF_SAMPLE_TID)
4295 perf_output_put(handle, data->tid_entry);
4297 if (sample_type & PERF_SAMPLE_TIME)
4298 perf_output_put(handle, data->time);
4300 if (sample_type & PERF_SAMPLE_ADDR)
4301 perf_output_put(handle, data->addr);
4303 if (sample_type & PERF_SAMPLE_ID)
4304 perf_output_put(handle, data->id);
4306 if (sample_type & PERF_SAMPLE_STREAM_ID)
4307 perf_output_put(handle, data->stream_id);
4309 if (sample_type & PERF_SAMPLE_CPU)
4310 perf_output_put(handle, data->cpu_entry);
4312 if (sample_type & PERF_SAMPLE_PERIOD)
4313 perf_output_put(handle, data->period);
4315 if (sample_type & PERF_SAMPLE_READ)
4316 perf_output_read(handle, event);
4318 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
4319 if (data->callchain) {
4322 if (data->callchain)
4323 size += data->callchain->nr;
4325 size *= sizeof(u64);
4327 perf_output_copy(handle, data->callchain, size);
4330 perf_output_put(handle, nr);
4334 if (sample_type & PERF_SAMPLE_RAW) {
4336 perf_output_put(handle, data->raw->size);
4337 perf_output_copy(handle, data->raw->data,
4344 .size = sizeof(u32),
4347 perf_output_put(handle, raw);
4352 void perf_prepare_sample(struct perf_event_header *header,
4353 struct perf_sample_data *data,
4354 struct perf_event *event,
4355 struct pt_regs *regs)
4357 u64 sample_type = event->attr.sample_type;
4359 header->type = PERF_RECORD_SAMPLE;
4360 header->size = sizeof(*header) + event->header_size;
4363 header->misc |= perf_misc_flags(regs);
4365 __perf_event_header__init_id(header, data, event);
4367 if (sample_type & PERF_SAMPLE_IP)
4368 data->ip = perf_instruction_pointer(regs);
4370 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
4373 data->callchain = perf_callchain(regs);
4375 if (data->callchain)
4376 size += data->callchain->nr;
4378 header->size += size * sizeof(u64);
4381 if (sample_type & PERF_SAMPLE_RAW) {
4382 int size = sizeof(u32);
4385 size += data->raw->size;
4387 size += sizeof(u32);
4389 WARN_ON_ONCE(size & (sizeof(u64)-1));
4390 header->size += size;
4394 static void perf_event_output(struct perf_event *event, int nmi,
4395 struct perf_sample_data *data,
4396 struct pt_regs *regs)
4398 struct perf_output_handle handle;
4399 struct perf_event_header header;
4401 /* protect the callchain buffers */
4404 perf_prepare_sample(&header, data, event, regs);
4406 if (perf_output_begin(&handle, event, header.size, nmi, 1))
4409 perf_output_sample(&handle, &header, data, event);
4411 perf_output_end(&handle);
4421 struct perf_read_event {
4422 struct perf_event_header header;
4429 perf_event_read_event(struct perf_event *event,
4430 struct task_struct *task)
4432 struct perf_output_handle handle;
4433 struct perf_sample_data sample;
4434 struct perf_read_event read_event = {
4436 .type = PERF_RECORD_READ,
4438 .size = sizeof(read_event) + event->read_size,
4440 .pid = perf_event_pid(event, task),
4441 .tid = perf_event_tid(event, task),
4445 perf_event_header__init_id(&read_event.header, &sample, event);
4446 ret = perf_output_begin(&handle, event, read_event.header.size, 0, 0);
4450 perf_output_put(&handle, read_event);
4451 perf_output_read(&handle, event);
4452 perf_event__output_id_sample(event, &handle, &sample);
4454 perf_output_end(&handle);
4458 * task tracking -- fork/exit
4460 * enabled by: attr.comm | attr.mmap | attr.mmap_data | attr.task
4463 struct perf_task_event {
4464 struct task_struct *task;
4465 struct perf_event_context *task_ctx;
4468 struct perf_event_header header;
4478 static void perf_event_task_output(struct perf_event *event,
4479 struct perf_task_event *task_event)
4481 struct perf_output_handle handle;
4482 struct perf_sample_data sample;
4483 struct task_struct *task = task_event->task;
4484 int ret, size = task_event->event_id.header.size;
4486 perf_event_header__init_id(&task_event->event_id.header, &sample, event);
4488 ret = perf_output_begin(&handle, event,
4489 task_event->event_id.header.size, 0, 0);
4493 task_event->event_id.pid = perf_event_pid(event, task);
4494 task_event->event_id.ppid = perf_event_pid(event, current);
4496 task_event->event_id.tid = perf_event_tid(event, task);
4497 task_event->event_id.ptid = perf_event_tid(event, current);
4499 perf_output_put(&handle, task_event->event_id);
4501 perf_event__output_id_sample(event, &handle, &sample);
4503 perf_output_end(&handle);
4505 task_event->event_id.header.size = size;
4508 static int perf_event_task_match(struct perf_event *event)
4510 if (event->state < PERF_EVENT_STATE_INACTIVE)
4513 if (!event_filter_match(event))
4516 if (event->attr.comm || event->attr.mmap ||
4517 event->attr.mmap_data || event->attr.task)
4523 static void perf_event_task_ctx(struct perf_event_context *ctx,
4524 struct perf_task_event *task_event)
4526 struct perf_event *event;
4528 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
4529 if (perf_event_task_match(event))
4530 perf_event_task_output(event, task_event);
4534 static void perf_event_task_event(struct perf_task_event *task_event)
4536 struct perf_cpu_context *cpuctx;
4537 struct perf_event_context *ctx;
4542 list_for_each_entry_rcu(pmu, &pmus, entry) {
4543 cpuctx = get_cpu_ptr(pmu->pmu_cpu_context);
4544 if (cpuctx->active_pmu != pmu)
4546 perf_event_task_ctx(&cpuctx->ctx, task_event);
4548 ctx = task_event->task_ctx;
4550 ctxn = pmu->task_ctx_nr;
4553 ctx = rcu_dereference(current->perf_event_ctxp[ctxn]);
4556 perf_event_task_ctx(ctx, task_event);
4558 put_cpu_ptr(pmu->pmu_cpu_context);
4563 static void perf_event_task(struct task_struct *task,
4564 struct perf_event_context *task_ctx,
4567 struct perf_task_event task_event;
4569 if (!atomic_read(&nr_comm_events) &&
4570 !atomic_read(&nr_mmap_events) &&
4571 !atomic_read(&nr_task_events))
4574 task_event = (struct perf_task_event){
4576 .task_ctx = task_ctx,
4579 .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT,
4581 .size = sizeof(task_event.event_id),
4587 .time = perf_clock(),
4591 perf_event_task_event(&task_event);
4594 void perf_event_fork(struct task_struct *task)
4596 perf_event_task(task, NULL, 1);
4603 struct perf_comm_event {
4604 struct task_struct *task;
4609 struct perf_event_header header;
4616 static void perf_event_comm_output(struct perf_event *event,
4617 struct perf_comm_event *comm_event)
4619 struct perf_output_handle handle;
4620 struct perf_sample_data sample;
4621 int size = comm_event->event_id.header.size;
4624 perf_event_header__init_id(&comm_event->event_id.header, &sample, event);
4625 ret = perf_output_begin(&handle, event,
4626 comm_event->event_id.header.size, 0, 0);
4631 comm_event->event_id.pid = perf_event_pid(event, comm_event->task);
4632 comm_event->event_id.tid = perf_event_tid(event, comm_event->task);
4634 perf_output_put(&handle, comm_event->event_id);
4635 perf_output_copy(&handle, comm_event->comm,
4636 comm_event->comm_size);
4638 perf_event__output_id_sample(event, &handle, &sample);
4640 perf_output_end(&handle);
4642 comm_event->event_id.header.size = size;
4645 static int perf_event_comm_match(struct perf_event *event)
4647 if (event->state < PERF_EVENT_STATE_INACTIVE)
4650 if (!event_filter_match(event))
4653 if (event->attr.comm)
4659 static void perf_event_comm_ctx(struct perf_event_context *ctx,
4660 struct perf_comm_event *comm_event)
4662 struct perf_event *event;
4664 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
4665 if (perf_event_comm_match(event))
4666 perf_event_comm_output(event, comm_event);
4670 static void perf_event_comm_event(struct perf_comm_event *comm_event)
4672 struct perf_cpu_context *cpuctx;
4673 struct perf_event_context *ctx;
4674 char comm[TASK_COMM_LEN];
4679 memset(comm, 0, sizeof(comm));
4680 strlcpy(comm, comm_event->task->comm, sizeof(comm));
4681 size = ALIGN(strlen(comm)+1, sizeof(u64));
4683 comm_event->comm = comm;
4684 comm_event->comm_size = size;
4686 comm_event->event_id.header.size = sizeof(comm_event->event_id) + size;
4688 list_for_each_entry_rcu(pmu, &pmus, entry) {
4689 cpuctx = get_cpu_ptr(pmu->pmu_cpu_context);
4690 if (cpuctx->active_pmu != pmu)
4692 perf_event_comm_ctx(&cpuctx->ctx, comm_event);
4694 ctxn = pmu->task_ctx_nr;
4698 ctx = rcu_dereference(current->perf_event_ctxp[ctxn]);
4700 perf_event_comm_ctx(ctx, comm_event);
4702 put_cpu_ptr(pmu->pmu_cpu_context);
4707 void perf_event_comm(struct task_struct *task)
4709 struct perf_comm_event comm_event;
4710 struct perf_event_context *ctx;
4713 for_each_task_context_nr(ctxn) {
4714 ctx = task->perf_event_ctxp[ctxn];
4718 perf_event_enable_on_exec(ctx);
4721 if (!atomic_read(&nr_comm_events))
4724 comm_event = (struct perf_comm_event){
4730 .type = PERF_RECORD_COMM,
4739 perf_event_comm_event(&comm_event);
4746 struct perf_mmap_event {
4747 struct vm_area_struct *vma;
4749 const char *file_name;
4753 struct perf_event_header header;
4763 static void perf_event_mmap_output(struct perf_event *event,
4764 struct perf_mmap_event *mmap_event)
4766 struct perf_output_handle handle;
4767 struct perf_sample_data sample;
4768 int size = mmap_event->event_id.header.size;
4771 perf_event_header__init_id(&mmap_event->event_id.header, &sample, event);
4772 ret = perf_output_begin(&handle, event,
4773 mmap_event->event_id.header.size, 0, 0);
4777 mmap_event->event_id.pid = perf_event_pid(event, current);
4778 mmap_event->event_id.tid = perf_event_tid(event, current);
4780 perf_output_put(&handle, mmap_event->event_id);
4781 perf_output_copy(&handle, mmap_event->file_name,
4782 mmap_event->file_size);
4784 perf_event__output_id_sample(event, &handle, &sample);
4786 perf_output_end(&handle);
4788 mmap_event->event_id.header.size = size;
4791 static int perf_event_mmap_match(struct perf_event *event,
4792 struct perf_mmap_event *mmap_event,
4795 if (event->state < PERF_EVENT_STATE_INACTIVE)
4798 if (!event_filter_match(event))
4801 if ((!executable && event->attr.mmap_data) ||
4802 (executable && event->attr.mmap))
4808 static void perf_event_mmap_ctx(struct perf_event_context *ctx,
4809 struct perf_mmap_event *mmap_event,
4812 struct perf_event *event;
4814 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
4815 if (perf_event_mmap_match(event, mmap_event, executable))
4816 perf_event_mmap_output(event, mmap_event);
4820 static void perf_event_mmap_event(struct perf_mmap_event *mmap_event)
4822 struct perf_cpu_context *cpuctx;
4823 struct perf_event_context *ctx;
4824 struct vm_area_struct *vma = mmap_event->vma;
4825 struct file *file = vma->vm_file;
4833 memset(tmp, 0, sizeof(tmp));
4837 * d_path works from the end of the buffer backwards, so we
4838 * need to add enough zero bytes after the string to handle
4839 * the 64bit alignment we do later.
4841 buf = kzalloc(PATH_MAX + sizeof(u64), GFP_KERNEL);
4843 name = strncpy(tmp, "//enomem", sizeof(tmp));
4846 name = d_path(&file->f_path, buf, PATH_MAX);
4848 name = strncpy(tmp, "//toolong", sizeof(tmp));
4852 if (arch_vma_name(mmap_event->vma)) {
4853 name = strncpy(tmp, arch_vma_name(mmap_event->vma),
4859 name = strncpy(tmp, "[vdso]", sizeof(tmp));
4861 } else if (vma->vm_start <= vma->vm_mm->start_brk &&
4862 vma->vm_end >= vma->vm_mm->brk) {
4863 name = strncpy(tmp, "[heap]", sizeof(tmp));
4865 } else if (vma->vm_start <= vma->vm_mm->start_stack &&
4866 vma->vm_end >= vma->vm_mm->start_stack) {
4867 name = strncpy(tmp, "[stack]", sizeof(tmp));
4871 name = strncpy(tmp, "//anon", sizeof(tmp));
4876 size = ALIGN(strlen(name)+1, sizeof(u64));
4878 mmap_event->file_name = name;
4879 mmap_event->file_size = size;
4881 mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size;
4884 list_for_each_entry_rcu(pmu, &pmus, entry) {
4885 cpuctx = get_cpu_ptr(pmu->pmu_cpu_context);
4886 if (cpuctx->active_pmu != pmu)
4888 perf_event_mmap_ctx(&cpuctx->ctx, mmap_event,
4889 vma->vm_flags & VM_EXEC);
4891 ctxn = pmu->task_ctx_nr;
4895 ctx = rcu_dereference(current->perf_event_ctxp[ctxn]);
4897 perf_event_mmap_ctx(ctx, mmap_event,
4898 vma->vm_flags & VM_EXEC);
4901 put_cpu_ptr(pmu->pmu_cpu_context);
4908 void perf_event_mmap(struct vm_area_struct *vma)
4910 struct perf_mmap_event mmap_event;
4912 if (!atomic_read(&nr_mmap_events))
4915 mmap_event = (struct perf_mmap_event){
4921 .type = PERF_RECORD_MMAP,
4922 .misc = PERF_RECORD_MISC_USER,
4927 .start = vma->vm_start,
4928 .len = vma->vm_end - vma->vm_start,
4929 .pgoff = (u64)vma->vm_pgoff << PAGE_SHIFT,
4933 perf_event_mmap_event(&mmap_event);
4937 * IRQ throttle logging
4940 static void perf_log_throttle(struct perf_event *event, int enable)
4942 struct perf_output_handle handle;
4943 struct perf_sample_data sample;
4947 struct perf_event_header header;
4951 } throttle_event = {
4953 .type = PERF_RECORD_THROTTLE,
4955 .size = sizeof(throttle_event),
4957 .time = perf_clock(),
4958 .id = primary_event_id(event),
4959 .stream_id = event->id,
4963 throttle_event.header.type = PERF_RECORD_UNTHROTTLE;
4965 perf_event_header__init_id(&throttle_event.header, &sample, event);
4967 ret = perf_output_begin(&handle, event,
4968 throttle_event.header.size, 1, 0);
4972 perf_output_put(&handle, throttle_event);
4973 perf_event__output_id_sample(event, &handle, &sample);
4974 perf_output_end(&handle);
4978 * Generic event overflow handling, sampling.
4981 static int __perf_event_overflow(struct perf_event *event, int nmi,
4982 int throttle, struct perf_sample_data *data,
4983 struct pt_regs *regs)
4985 int events = atomic_read(&event->event_limit);
4986 struct hw_perf_event *hwc = &event->hw;
4990 * Non-sampling counters might still use the PMI to fold short
4991 * hardware counters, ignore those.
4993 if (unlikely(!is_sampling_event(event)))
4996 if (unlikely(hwc->interrupts >= max_samples_per_tick)) {
4998 hwc->interrupts = MAX_INTERRUPTS;
4999 perf_log_throttle(event, 0);
5005 if (event->attr.freq) {
5006 u64 now = perf_clock();
5007 s64 delta = now - hwc->freq_time_stamp;
5009 hwc->freq_time_stamp = now;
5011 if (delta > 0 && delta < 2*TICK_NSEC)
5012 perf_adjust_period(event, delta, hwc->last_period);
5016 * XXX event_limit might not quite work as expected on inherited
5020 event->pending_kill = POLL_IN;
5021 if (events && atomic_dec_and_test(&event->event_limit)) {
5023 event->pending_kill = POLL_HUP;
5025 event->pending_disable = 1;
5026 irq_work_queue(&event->pending);
5028 perf_event_disable(event);
5031 if (event->overflow_handler)
5032 event->overflow_handler(event, nmi, data, regs);
5034 perf_event_output(event, nmi, data, regs);
5036 if (event->fasync && event->pending_kill) {
5038 event->pending_wakeup = 1;
5039 irq_work_queue(&event->pending);
5041 perf_event_wakeup(event);
5047 int perf_event_overflow(struct perf_event *event, int nmi,
5048 struct perf_sample_data *data,
5049 struct pt_regs *regs)
5051 return __perf_event_overflow(event, nmi, 1, data, regs);
5055 * Generic software event infrastructure
5058 struct swevent_htable {
5059 struct swevent_hlist *swevent_hlist;
5060 struct mutex hlist_mutex;
5063 /* Recursion avoidance in each contexts */
5064 int recursion[PERF_NR_CONTEXTS];
5067 static DEFINE_PER_CPU(struct swevent_htable, swevent_htable);
5070 * We directly increment event->count and keep a second value in
5071 * event->hw.period_left to count intervals. This period event
5072 * is kept in the range [-sample_period, 0] so that we can use the
5076 static u64 perf_swevent_set_period(struct perf_event *event)
5078 struct hw_perf_event *hwc = &event->hw;
5079 u64 period = hwc->last_period;
5083 hwc->last_period = hwc->sample_period;
5086 old = val = local64_read(&hwc->period_left);
5090 nr = div64_u64(period + val, period);
5091 offset = nr * period;
5093 if (local64_cmpxchg(&hwc->period_left, old, val) != old)
5099 static void perf_swevent_overflow(struct perf_event *event, u64 overflow,
5100 int nmi, struct perf_sample_data *data,
5101 struct pt_regs *regs)
5103 struct hw_perf_event *hwc = &event->hw;
5106 data->period = event->hw.last_period;
5108 overflow = perf_swevent_set_period(event);
5110 if (hwc->interrupts == MAX_INTERRUPTS)
5113 for (; overflow; overflow--) {
5114 if (__perf_event_overflow(event, nmi, throttle,
5117 * We inhibit the overflow from happening when
5118 * hwc->interrupts == MAX_INTERRUPTS.
5126 static void perf_swevent_event(struct perf_event *event, u64 nr,
5127 int nmi, struct perf_sample_data *data,
5128 struct pt_regs *regs)
5130 struct hw_perf_event *hwc = &event->hw;
5132 local64_add(nr, &event->count);
5137 if (!is_sampling_event(event))
5140 if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq)
5141 return perf_swevent_overflow(event, 1, nmi, data, regs);
5143 if (local64_add_negative(nr, &hwc->period_left))
5146 perf_swevent_overflow(event, 0, nmi, data, regs);
5149 static int perf_exclude_event(struct perf_event *event,
5150 struct pt_regs *regs)
5152 if (event->hw.state & PERF_HES_STOPPED)
5156 if (event->attr.exclude_user && user_mode(regs))
5159 if (event->attr.exclude_kernel && !user_mode(regs))
5166 static int perf_swevent_match(struct perf_event *event,
5167 enum perf_type_id type,
5169 struct perf_sample_data *data,
5170 struct pt_regs *regs)
5172 if (event->attr.type != type)
5175 if (event->attr.config != event_id)
5178 if (perf_exclude_event(event, regs))
5184 static inline u64 swevent_hash(u64 type, u32 event_id)
5186 u64 val = event_id | (type << 32);
5188 return hash_64(val, SWEVENT_HLIST_BITS);
5191 static inline struct hlist_head *
5192 __find_swevent_head(struct swevent_hlist *hlist, u64 type, u32 event_id)
5194 u64 hash = swevent_hash(type, event_id);
5196 return &hlist->heads[hash];
5199 /* For the read side: events when they trigger */
5200 static inline struct hlist_head *
5201 find_swevent_head_rcu(struct swevent_htable *swhash, u64 type, u32 event_id)
5203 struct swevent_hlist *hlist;
5205 hlist = rcu_dereference(swhash->swevent_hlist);
5209 return __find_swevent_head(hlist, type, event_id);
5212 /* For the event head insertion and removal in the hlist */
5213 static inline struct hlist_head *
5214 find_swevent_head(struct swevent_htable *swhash, struct perf_event *event)
5216 struct swevent_hlist *hlist;
5217 u32 event_id = event->attr.config;
5218 u64 type = event->attr.type;
5221 * Event scheduling is always serialized against hlist allocation
5222 * and release. Which makes the protected version suitable here.
5223 * The context lock guarantees that.
5225 hlist = rcu_dereference_protected(swhash->swevent_hlist,
5226 lockdep_is_held(&event->ctx->lock));
5230 return __find_swevent_head(hlist, type, event_id);
5233 static void do_perf_sw_event(enum perf_type_id type, u32 event_id,
5235 struct perf_sample_data *data,
5236 struct pt_regs *regs)
5238 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
5239 struct perf_event *event;
5240 struct hlist_node *node;
5241 struct hlist_head *head;
5244 head = find_swevent_head_rcu(swhash, type, event_id);
5248 hlist_for_each_entry_rcu(event, node, head, hlist_entry) {
5249 if (perf_swevent_match(event, type, event_id, data, regs))
5250 perf_swevent_event(event, nr, nmi, data, regs);
5256 int perf_swevent_get_recursion_context(void)
5258 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
5260 return get_recursion_context(swhash->recursion);
5262 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context);
5264 inline void perf_swevent_put_recursion_context(int rctx)
5266 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
5268 put_recursion_context(swhash->recursion, rctx);
5271 void __perf_sw_event(u32 event_id, u64 nr, int nmi,
5272 struct pt_regs *regs, u64 addr)
5274 struct perf_sample_data data;
5277 preempt_disable_notrace();
5278 rctx = perf_swevent_get_recursion_context();
5282 perf_sample_data_init(&data, addr);
5284 do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, nmi, &data, regs);
5286 perf_swevent_put_recursion_context(rctx);
5287 preempt_enable_notrace();
5290 static void perf_swevent_read(struct perf_event *event)
5294 static int perf_swevent_add(struct perf_event *event, int flags)
5296 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
5297 struct hw_perf_event *hwc = &event->hw;
5298 struct hlist_head *head;
5300 if (is_sampling_event(event)) {
5301 hwc->last_period = hwc->sample_period;
5302 perf_swevent_set_period(event);
5305 hwc->state = !(flags & PERF_EF_START);
5307 head = find_swevent_head(swhash, event);
5308 if (WARN_ON_ONCE(!head))
5311 hlist_add_head_rcu(&event->hlist_entry, head);
5316 static void perf_swevent_del(struct perf_event *event, int flags)
5318 hlist_del_rcu(&event->hlist_entry);
5321 static void perf_swevent_start(struct perf_event *event, int flags)
5323 event->hw.state = 0;
5326 static void perf_swevent_stop(struct perf_event *event, int flags)
5328 event->hw.state = PERF_HES_STOPPED;
5331 /* Deref the hlist from the update side */
5332 static inline struct swevent_hlist *
5333 swevent_hlist_deref(struct swevent_htable *swhash)
5335 return rcu_dereference_protected(swhash->swevent_hlist,
5336 lockdep_is_held(&swhash->hlist_mutex));
5339 static void swevent_hlist_release(struct swevent_htable *swhash)
5341 struct swevent_hlist *hlist = swevent_hlist_deref(swhash);
5346 rcu_assign_pointer(swhash->swevent_hlist, NULL);
5347 kfree_rcu(hlist, rcu_head);
5350 static void swevent_hlist_put_cpu(struct perf_event *event, int cpu)
5352 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
5354 mutex_lock(&swhash->hlist_mutex);
5356 if (!--swhash->hlist_refcount)
5357 swevent_hlist_release(swhash);
5359 mutex_unlock(&swhash->hlist_mutex);
5362 static void swevent_hlist_put(struct perf_event *event)
5366 if (event->cpu != -1) {
5367 swevent_hlist_put_cpu(event, event->cpu);
5371 for_each_possible_cpu(cpu)
5372 swevent_hlist_put_cpu(event, cpu);
5375 static int swevent_hlist_get_cpu(struct perf_event *event, int cpu)
5377 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
5380 mutex_lock(&swhash->hlist_mutex);
5382 if (!swevent_hlist_deref(swhash) && cpu_online(cpu)) {
5383 struct swevent_hlist *hlist;
5385 hlist = kzalloc(sizeof(*hlist), GFP_KERNEL);
5390 rcu_assign_pointer(swhash->swevent_hlist, hlist);
5392 swhash->hlist_refcount++;
5394 mutex_unlock(&swhash->hlist_mutex);
5399 static int swevent_hlist_get(struct perf_event *event)
5402 int cpu, failed_cpu;
5404 if (event->cpu != -1)
5405 return swevent_hlist_get_cpu(event, event->cpu);
5408 for_each_possible_cpu(cpu) {
5409 err = swevent_hlist_get_cpu(event, cpu);
5419 for_each_possible_cpu(cpu) {
5420 if (cpu == failed_cpu)
5422 swevent_hlist_put_cpu(event, cpu);
5429 struct jump_label_key perf_swevent_enabled[PERF_COUNT_SW_MAX];
5431 static void sw_perf_event_destroy(struct perf_event *event)
5433 u64 event_id = event->attr.config;
5435 WARN_ON(event->parent);
5437 jump_label_dec(&perf_swevent_enabled[event_id]);
5438 swevent_hlist_put(event);
5441 static int perf_swevent_init(struct perf_event *event)
5443 int event_id = event->attr.config;
5445 if (event->attr.type != PERF_TYPE_SOFTWARE)
5449 case PERF_COUNT_SW_CPU_CLOCK:
5450 case PERF_COUNT_SW_TASK_CLOCK:
5457 if (event_id >= PERF_COUNT_SW_MAX)
5460 if (!event->parent) {
5463 err = swevent_hlist_get(event);
5467 jump_label_inc(&perf_swevent_enabled[event_id]);
5468 event->destroy = sw_perf_event_destroy;
5474 static struct pmu perf_swevent = {
5475 .task_ctx_nr = perf_sw_context,
5477 .event_init = perf_swevent_init,
5478 .add = perf_swevent_add,
5479 .del = perf_swevent_del,
5480 .start = perf_swevent_start,
5481 .stop = perf_swevent_stop,
5482 .read = perf_swevent_read,
5485 #ifdef CONFIG_EVENT_TRACING
5487 static int perf_tp_filter_match(struct perf_event *event,
5488 struct perf_sample_data *data)
5490 void *record = data->raw->data;
5492 if (likely(!event->filter) || filter_match_preds(event->filter, record))
5497 static int perf_tp_event_match(struct perf_event *event,
5498 struct perf_sample_data *data,
5499 struct pt_regs *regs)
5501 if (event->hw.state & PERF_HES_STOPPED)
5504 * All tracepoints are from kernel-space.
5506 if (event->attr.exclude_kernel)
5509 if (!perf_tp_filter_match(event, data))
5515 void perf_tp_event(u64 addr, u64 count, void *record, int entry_size,
5516 struct pt_regs *regs, struct hlist_head *head, int rctx)
5518 struct perf_sample_data data;
5519 struct perf_event *event;
5520 struct hlist_node *node;
5522 struct perf_raw_record raw = {
5527 perf_sample_data_init(&data, addr);
5530 hlist_for_each_entry_rcu(event, node, head, hlist_entry) {
5531 if (perf_tp_event_match(event, &data, regs))
5532 perf_swevent_event(event, count, 1, &data, regs);
5535 perf_swevent_put_recursion_context(rctx);
5537 EXPORT_SYMBOL_GPL(perf_tp_event);
5539 static void tp_perf_event_destroy(struct perf_event *event)
5541 perf_trace_destroy(event);
5544 static int perf_tp_event_init(struct perf_event *event)
5548 if (event->attr.type != PERF_TYPE_TRACEPOINT)
5551 err = perf_trace_init(event);
5555 event->destroy = tp_perf_event_destroy;
5560 static struct pmu perf_tracepoint = {
5561 .task_ctx_nr = perf_sw_context,
5563 .event_init = perf_tp_event_init,
5564 .add = perf_trace_add,
5565 .del = perf_trace_del,
5566 .start = perf_swevent_start,
5567 .stop = perf_swevent_stop,
5568 .read = perf_swevent_read,
5571 static inline void perf_tp_register(void)
5573 perf_pmu_register(&perf_tracepoint, "tracepoint", PERF_TYPE_TRACEPOINT);
5576 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
5581 if (event->attr.type != PERF_TYPE_TRACEPOINT)
5584 filter_str = strndup_user(arg, PAGE_SIZE);
5585 if (IS_ERR(filter_str))
5586 return PTR_ERR(filter_str);
5588 ret = ftrace_profile_set_filter(event, event->attr.config, filter_str);
5594 static void perf_event_free_filter(struct perf_event *event)
5596 ftrace_profile_free_filter(event);
5601 static inline void perf_tp_register(void)
5605 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
5610 static void perf_event_free_filter(struct perf_event *event)
5614 #endif /* CONFIG_EVENT_TRACING */
5616 #ifdef CONFIG_HAVE_HW_BREAKPOINT
5617 void perf_bp_event(struct perf_event *bp, void *data)
5619 struct perf_sample_data sample;
5620 struct pt_regs *regs = data;
5622 perf_sample_data_init(&sample, bp->attr.bp_addr);
5624 if (!bp->hw.state && !perf_exclude_event(bp, regs))
5625 perf_swevent_event(bp, 1, 1, &sample, regs);
5630 * hrtimer based swevent callback
5633 static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer)
5635 enum hrtimer_restart ret = HRTIMER_RESTART;
5636 struct perf_sample_data data;
5637 struct pt_regs *regs;
5638 struct perf_event *event;
5641 event = container_of(hrtimer, struct perf_event, hw.hrtimer);
5643 if (event->state != PERF_EVENT_STATE_ACTIVE)
5644 return HRTIMER_NORESTART;
5646 event->pmu->read(event);
5648 perf_sample_data_init(&data, 0);
5649 data.period = event->hw.last_period;
5650 regs = get_irq_regs();
5652 if (regs && !perf_exclude_event(event, regs)) {
5653 if (!(event->attr.exclude_idle && current->pid == 0))
5654 if (perf_event_overflow(event, 0, &data, regs))
5655 ret = HRTIMER_NORESTART;
5658 period = max_t(u64, 10000, event->hw.sample_period);
5659 hrtimer_forward_now(hrtimer, ns_to_ktime(period));
5664 static void perf_swevent_start_hrtimer(struct perf_event *event)
5666 struct hw_perf_event *hwc = &event->hw;
5669 if (!is_sampling_event(event))
5672 period = local64_read(&hwc->period_left);
5677 local64_set(&hwc->period_left, 0);
5679 period = max_t(u64, 10000, hwc->sample_period);
5681 __hrtimer_start_range_ns(&hwc->hrtimer,
5682 ns_to_ktime(period), 0,
5683 HRTIMER_MODE_REL_PINNED, 0);
5686 static void perf_swevent_cancel_hrtimer(struct perf_event *event)
5688 struct hw_perf_event *hwc = &event->hw;
5690 if (is_sampling_event(event)) {
5691 ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer);
5692 local64_set(&hwc->period_left, ktime_to_ns(remaining));
5694 hrtimer_cancel(&hwc->hrtimer);
5698 static void perf_swevent_init_hrtimer(struct perf_event *event)
5700 struct hw_perf_event *hwc = &event->hw;
5702 if (!is_sampling_event(event))
5705 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
5706 hwc->hrtimer.function = perf_swevent_hrtimer;
5709 * Since hrtimers have a fixed rate, we can do a static freq->period
5710 * mapping and avoid the whole period adjust feedback stuff.
5712 if (event->attr.freq) {
5713 long freq = event->attr.sample_freq;
5715 event->attr.sample_period = NSEC_PER_SEC / freq;
5716 hwc->sample_period = event->attr.sample_period;
5717 local64_set(&hwc->period_left, hwc->sample_period);
5718 event->attr.freq = 0;
5723 * Software event: cpu wall time clock
5726 static void cpu_clock_event_update(struct perf_event *event)
5731 now = local_clock();
5732 prev = local64_xchg(&event->hw.prev_count, now);
5733 local64_add(now - prev, &event->count);
5736 static void cpu_clock_event_start(struct perf_event *event, int flags)
5738 local64_set(&event->hw.prev_count, local_clock());
5739 perf_swevent_start_hrtimer(event);
5742 static void cpu_clock_event_stop(struct perf_event *event, int flags)
5744 perf_swevent_cancel_hrtimer(event);
5745 cpu_clock_event_update(event);
5748 static int cpu_clock_event_add(struct perf_event *event, int flags)
5750 if (flags & PERF_EF_START)
5751 cpu_clock_event_start(event, flags);
5756 static void cpu_clock_event_del(struct perf_event *event, int flags)
5758 cpu_clock_event_stop(event, flags);
5761 static void cpu_clock_event_read(struct perf_event *event)
5763 cpu_clock_event_update(event);
5766 static int cpu_clock_event_init(struct perf_event *event)
5768 if (event->attr.type != PERF_TYPE_SOFTWARE)
5771 if (event->attr.config != PERF_COUNT_SW_CPU_CLOCK)
5774 perf_swevent_init_hrtimer(event);
5779 static struct pmu perf_cpu_clock = {
5780 .task_ctx_nr = perf_sw_context,
5782 .event_init = cpu_clock_event_init,
5783 .add = cpu_clock_event_add,
5784 .del = cpu_clock_event_del,
5785 .start = cpu_clock_event_start,
5786 .stop = cpu_clock_event_stop,
5787 .read = cpu_clock_event_read,
5791 * Software event: task time clock
5794 static void task_clock_event_update(struct perf_event *event, u64 now)
5799 prev = local64_xchg(&event->hw.prev_count, now);
5801 local64_add(delta, &event->count);
5804 static void task_clock_event_start(struct perf_event *event, int flags)
5806 local64_set(&event->hw.prev_count, event->ctx->time);
5807 perf_swevent_start_hrtimer(event);
5810 static void task_clock_event_stop(struct perf_event *event, int flags)
5812 perf_swevent_cancel_hrtimer(event);
5813 task_clock_event_update(event, event->ctx->time);
5816 static int task_clock_event_add(struct perf_event *event, int flags)
5818 if (flags & PERF_EF_START)
5819 task_clock_event_start(event, flags);
5824 static void task_clock_event_del(struct perf_event *event, int flags)
5826 task_clock_event_stop(event, PERF_EF_UPDATE);
5829 static void task_clock_event_read(struct perf_event *event)
5831 u64 now = perf_clock();
5832 u64 delta = now - event->ctx->timestamp;
5833 u64 time = event->ctx->time + delta;
5835 task_clock_event_update(event, time);
5838 static int task_clock_event_init(struct perf_event *event)
5840 if (event->attr.type != PERF_TYPE_SOFTWARE)
5843 if (event->attr.config != PERF_COUNT_SW_TASK_CLOCK)
5846 perf_swevent_init_hrtimer(event);
5851 static struct pmu perf_task_clock = {
5852 .task_ctx_nr = perf_sw_context,
5854 .event_init = task_clock_event_init,
5855 .add = task_clock_event_add,
5856 .del = task_clock_event_del,
5857 .start = task_clock_event_start,
5858 .stop = task_clock_event_stop,
5859 .read = task_clock_event_read,
5862 static void perf_pmu_nop_void(struct pmu *pmu)
5866 static int perf_pmu_nop_int(struct pmu *pmu)
5871 static void perf_pmu_start_txn(struct pmu *pmu)
5873 perf_pmu_disable(pmu);
5876 static int perf_pmu_commit_txn(struct pmu *pmu)
5878 perf_pmu_enable(pmu);
5882 static void perf_pmu_cancel_txn(struct pmu *pmu)
5884 perf_pmu_enable(pmu);
5888 * Ensures all contexts with the same task_ctx_nr have the same
5889 * pmu_cpu_context too.
5891 static void *find_pmu_context(int ctxn)
5898 list_for_each_entry(pmu, &pmus, entry) {
5899 if (pmu->task_ctx_nr == ctxn)
5900 return pmu->pmu_cpu_context;
5906 static void update_pmu_context(struct pmu *pmu, struct pmu *old_pmu)
5910 for_each_possible_cpu(cpu) {
5911 struct perf_cpu_context *cpuctx;
5913 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
5915 if (cpuctx->active_pmu == old_pmu)
5916 cpuctx->active_pmu = pmu;
5920 static void free_pmu_context(struct pmu *pmu)
5924 mutex_lock(&pmus_lock);
5926 * Like a real lame refcount.
5928 list_for_each_entry(i, &pmus, entry) {
5929 if (i->pmu_cpu_context == pmu->pmu_cpu_context) {
5930 update_pmu_context(i, pmu);
5935 free_percpu(pmu->pmu_cpu_context);
5937 mutex_unlock(&pmus_lock);
5939 static struct idr pmu_idr;
5942 type_show(struct device *dev, struct device_attribute *attr, char *page)
5944 struct pmu *pmu = dev_get_drvdata(dev);
5946 return snprintf(page, PAGE_SIZE-1, "%d\n", pmu->type);
5949 static struct device_attribute pmu_dev_attrs[] = {
5954 static int pmu_bus_running;
5955 static struct bus_type pmu_bus = {
5956 .name = "event_source",
5957 .dev_attrs = pmu_dev_attrs,
5960 static void pmu_dev_release(struct device *dev)
5965 static int pmu_dev_alloc(struct pmu *pmu)
5969 pmu->dev = kzalloc(sizeof(struct device), GFP_KERNEL);
5973 device_initialize(pmu->dev);
5974 ret = dev_set_name(pmu->dev, "%s", pmu->name);
5978 dev_set_drvdata(pmu->dev, pmu);
5979 pmu->dev->bus = &pmu_bus;
5980 pmu->dev->release = pmu_dev_release;
5981 ret = device_add(pmu->dev);
5989 put_device(pmu->dev);
5993 static struct lock_class_key cpuctx_mutex;
5994 static struct lock_class_key cpuctx_lock;
5996 int perf_pmu_register(struct pmu *pmu, char *name, int type)
6000 mutex_lock(&pmus_lock);
6002 pmu->pmu_disable_count = alloc_percpu(int);
6003 if (!pmu->pmu_disable_count)
6012 int err = idr_pre_get(&pmu_idr, GFP_KERNEL);
6016 err = idr_get_new_above(&pmu_idr, pmu, PERF_TYPE_MAX, &type);
6024 if (pmu_bus_running) {
6025 ret = pmu_dev_alloc(pmu);
6031 pmu->pmu_cpu_context = find_pmu_context(pmu->task_ctx_nr);
6032 if (pmu->pmu_cpu_context)
6033 goto got_cpu_context;
6035 pmu->pmu_cpu_context = alloc_percpu(struct perf_cpu_context);
6036 if (!pmu->pmu_cpu_context)
6039 for_each_possible_cpu(cpu) {
6040 struct perf_cpu_context *cpuctx;
6042 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
6043 __perf_event_init_context(&cpuctx->ctx);
6044 lockdep_set_class(&cpuctx->ctx.mutex, &cpuctx_mutex);
6045 lockdep_set_class(&cpuctx->ctx.lock, &cpuctx_lock);
6046 cpuctx->ctx.type = cpu_context;
6047 cpuctx->ctx.pmu = pmu;
6048 cpuctx->jiffies_interval = 1;
6049 INIT_LIST_HEAD(&cpuctx->rotation_list);
6050 cpuctx->active_pmu = pmu;
6054 if (!pmu->start_txn) {
6055 if (pmu->pmu_enable) {
6057 * If we have pmu_enable/pmu_disable calls, install
6058 * transaction stubs that use that to try and batch
6059 * hardware accesses.
6061 pmu->start_txn = perf_pmu_start_txn;
6062 pmu->commit_txn = perf_pmu_commit_txn;
6063 pmu->cancel_txn = perf_pmu_cancel_txn;
6065 pmu->start_txn = perf_pmu_nop_void;
6066 pmu->commit_txn = perf_pmu_nop_int;
6067 pmu->cancel_txn = perf_pmu_nop_void;
6071 if (!pmu->pmu_enable) {
6072 pmu->pmu_enable = perf_pmu_nop_void;
6073 pmu->pmu_disable = perf_pmu_nop_void;
6076 list_add_rcu(&pmu->entry, &pmus);
6079 mutex_unlock(&pmus_lock);
6084 device_del(pmu->dev);
6085 put_device(pmu->dev);
6088 if (pmu->type >= PERF_TYPE_MAX)
6089 idr_remove(&pmu_idr, pmu->type);
6092 free_percpu(pmu->pmu_disable_count);
6096 void perf_pmu_unregister(struct pmu *pmu)
6098 mutex_lock(&pmus_lock);
6099 list_del_rcu(&pmu->entry);
6100 mutex_unlock(&pmus_lock);
6103 * We dereference the pmu list under both SRCU and regular RCU, so
6104 * synchronize against both of those.
6106 synchronize_srcu(&pmus_srcu);
6109 free_percpu(pmu->pmu_disable_count);
6110 if (pmu->type >= PERF_TYPE_MAX)
6111 idr_remove(&pmu_idr, pmu->type);
6112 device_del(pmu->dev);
6113 put_device(pmu->dev);
6114 free_pmu_context(pmu);
6117 struct pmu *perf_init_event(struct perf_event *event)
6119 struct pmu *pmu = NULL;
6123 idx = srcu_read_lock(&pmus_srcu);
6126 pmu = idr_find(&pmu_idr, event->attr.type);
6129 ret = pmu->event_init(event);
6135 list_for_each_entry_rcu(pmu, &pmus, entry) {
6136 ret = pmu->event_init(event);
6140 if (ret != -ENOENT) {
6145 pmu = ERR_PTR(-ENOENT);
6147 srcu_read_unlock(&pmus_srcu, idx);
6153 * Allocate and initialize a event structure
6155 static struct perf_event *
6156 perf_event_alloc(struct perf_event_attr *attr, int cpu,
6157 struct task_struct *task,
6158 struct perf_event *group_leader,
6159 struct perf_event *parent_event,
6160 perf_overflow_handler_t overflow_handler)
6163 struct perf_event *event;
6164 struct hw_perf_event *hwc;
6167 if ((unsigned)cpu >= nr_cpu_ids) {
6168 if (!task || cpu != -1)
6169 return ERR_PTR(-EINVAL);
6172 event = kzalloc(sizeof(*event), GFP_KERNEL);
6174 return ERR_PTR(-ENOMEM);
6177 * Single events are their own group leaders, with an
6178 * empty sibling list:
6181 group_leader = event;
6183 mutex_init(&event->child_mutex);
6184 INIT_LIST_HEAD(&event->child_list);
6186 INIT_LIST_HEAD(&event->group_entry);
6187 INIT_LIST_HEAD(&event->event_entry);
6188 INIT_LIST_HEAD(&event->sibling_list);
6189 init_waitqueue_head(&event->waitq);
6190 init_irq_work(&event->pending, perf_pending_event);
6192 mutex_init(&event->mmap_mutex);
6195 event->attr = *attr;
6196 event->group_leader = group_leader;
6200 event->parent = parent_event;
6202 event->ns = get_pid_ns(current->nsproxy->pid_ns);
6203 event->id = atomic64_inc_return(&perf_event_id);
6205 event->state = PERF_EVENT_STATE_INACTIVE;
6208 event->attach_state = PERF_ATTACH_TASK;
6209 #ifdef CONFIG_HAVE_HW_BREAKPOINT
6211 * hw_breakpoint is a bit difficult here..
6213 if (attr->type == PERF_TYPE_BREAKPOINT)
6214 event->hw.bp_target = task;
6218 if (!overflow_handler && parent_event)
6219 overflow_handler = parent_event->overflow_handler;
6221 event->overflow_handler = overflow_handler;
6224 event->state = PERF_EVENT_STATE_OFF;
6229 hwc->sample_period = attr->sample_period;
6230 if (attr->freq && attr->sample_freq)
6231 hwc->sample_period = 1;
6232 hwc->last_period = hwc->sample_period;
6234 local64_set(&hwc->period_left, hwc->sample_period);
6237 * we currently do not support PERF_FORMAT_GROUP on inherited events
6239 if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP))
6242 pmu = perf_init_event(event);
6248 else if (IS_ERR(pmu))
6253 put_pid_ns(event->ns);
6255 return ERR_PTR(err);
6260 if (!event->parent) {
6261 if (event->attach_state & PERF_ATTACH_TASK)
6262 jump_label_inc(&perf_sched_events);
6263 if (event->attr.mmap || event->attr.mmap_data)
6264 atomic_inc(&nr_mmap_events);
6265 if (event->attr.comm)
6266 atomic_inc(&nr_comm_events);
6267 if (event->attr.task)
6268 atomic_inc(&nr_task_events);
6269 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) {
6270 err = get_callchain_buffers();
6273 return ERR_PTR(err);
6281 static int perf_copy_attr(struct perf_event_attr __user *uattr,
6282 struct perf_event_attr *attr)
6287 if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0))
6291 * zero the full structure, so that a short copy will be nice.
6293 memset(attr, 0, sizeof(*attr));
6295 ret = get_user(size, &uattr->size);
6299 if (size > PAGE_SIZE) /* silly large */
6302 if (!size) /* abi compat */
6303 size = PERF_ATTR_SIZE_VER0;
6305 if (size < PERF_ATTR_SIZE_VER0)
6309 * If we're handed a bigger struct than we know of,
6310 * ensure all the unknown bits are 0 - i.e. new
6311 * user-space does not rely on any kernel feature
6312 * extensions we dont know about yet.
6314 if (size > sizeof(*attr)) {
6315 unsigned char __user *addr;
6316 unsigned char __user *end;
6319 addr = (void __user *)uattr + sizeof(*attr);
6320 end = (void __user *)uattr + size;
6322 for (; addr < end; addr++) {
6323 ret = get_user(val, addr);
6329 size = sizeof(*attr);
6332 ret = copy_from_user(attr, uattr, size);
6337 * If the type exists, the corresponding creation will verify
6340 if (attr->type >= PERF_TYPE_MAX)
6343 if (attr->__reserved_1)
6346 if (attr->sample_type & ~(PERF_SAMPLE_MAX-1))
6349 if (attr->read_format & ~(PERF_FORMAT_MAX-1))
6356 put_user(sizeof(*attr), &uattr->size);
6362 perf_event_set_output(struct perf_event *event, struct perf_event *output_event)
6364 struct perf_buffer *buffer = NULL, *old_buffer = NULL;
6370 /* don't allow circular references */
6371 if (event == output_event)
6375 * Don't allow cross-cpu buffers
6377 if (output_event->cpu != event->cpu)
6381 * If its not a per-cpu buffer, it must be the same task.
6383 if (output_event->cpu == -1 && output_event->ctx != event->ctx)
6387 mutex_lock(&event->mmap_mutex);
6388 /* Can't redirect output if we've got an active mmap() */
6389 if (atomic_read(&event->mmap_count))
6393 /* get the buffer we want to redirect to */
6394 buffer = perf_buffer_get(output_event);
6399 old_buffer = event->buffer;
6400 rcu_assign_pointer(event->buffer, buffer);
6403 mutex_unlock(&event->mmap_mutex);
6406 perf_buffer_put(old_buffer);
6412 * sys_perf_event_open - open a performance event, associate it to a task/cpu
6414 * @attr_uptr: event_id type attributes for monitoring/sampling
6417 * @group_fd: group leader event fd
6419 SYSCALL_DEFINE5(perf_event_open,
6420 struct perf_event_attr __user *, attr_uptr,
6421 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
6423 struct perf_event *group_leader = NULL, *output_event = NULL;
6424 struct perf_event *event, *sibling;
6425 struct perf_event_attr attr;
6426 struct perf_event_context *ctx;
6427 struct file *event_file = NULL;
6428 struct file *group_file = NULL;
6429 struct task_struct *task = NULL;
6433 int fput_needed = 0;
6436 /* for future expandability... */
6437 if (flags & ~PERF_FLAG_ALL)
6440 err = perf_copy_attr(attr_uptr, &attr);
6444 if (!attr.exclude_kernel) {
6445 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
6450 if (attr.sample_freq > sysctl_perf_event_sample_rate)
6455 * In cgroup mode, the pid argument is used to pass the fd
6456 * opened to the cgroup directory in cgroupfs. The cpu argument
6457 * designates the cpu on which to monitor threads from that
6460 if ((flags & PERF_FLAG_PID_CGROUP) && (pid == -1 || cpu == -1))
6463 event_fd = get_unused_fd_flags(O_RDWR);
6467 if (group_fd != -1) {
6468 group_leader = perf_fget_light(group_fd, &fput_needed);
6469 if (IS_ERR(group_leader)) {
6470 err = PTR_ERR(group_leader);
6473 group_file = group_leader->filp;
6474 if (flags & PERF_FLAG_FD_OUTPUT)
6475 output_event = group_leader;
6476 if (flags & PERF_FLAG_FD_NO_GROUP)
6477 group_leader = NULL;
6480 if (pid != -1 && !(flags & PERF_FLAG_PID_CGROUP)) {
6481 task = find_lively_task_by_vpid(pid);
6483 err = PTR_ERR(task);
6488 event = perf_event_alloc(&attr, cpu, task, group_leader, NULL, NULL);
6489 if (IS_ERR(event)) {
6490 err = PTR_ERR(event);
6494 if (flags & PERF_FLAG_PID_CGROUP) {
6495 err = perf_cgroup_connect(pid, event, &attr, group_leader);
6500 * - that has cgroup constraint on event->cpu
6501 * - that may need work on context switch
6503 atomic_inc(&per_cpu(perf_cgroup_events, event->cpu));
6504 jump_label_inc(&perf_sched_events);
6508 * Special case software events and allow them to be part of
6509 * any hardware group.
6514 (is_software_event(event) != is_software_event(group_leader))) {
6515 if (is_software_event(event)) {
6517 * If event and group_leader are not both a software
6518 * event, and event is, then group leader is not.
6520 * Allow the addition of software events to !software
6521 * groups, this is safe because software events never
6524 pmu = group_leader->pmu;
6525 } else if (is_software_event(group_leader) &&
6526 (group_leader->group_flags & PERF_GROUP_SOFTWARE)) {
6528 * In case the group is a pure software group, and we
6529 * try to add a hardware event, move the whole group to
6530 * the hardware context.
6537 * Get the target context (task or percpu):
6539 ctx = find_get_context(pmu, task, cpu);
6546 put_task_struct(task);
6551 * Look up the group leader (we will attach this event to it):
6557 * Do not allow a recursive hierarchy (this new sibling
6558 * becoming part of another group-sibling):
6560 if (group_leader->group_leader != group_leader)
6563 * Do not allow to attach to a group in a different
6564 * task or CPU context:
6567 if (group_leader->ctx->type != ctx->type)
6570 if (group_leader->ctx != ctx)
6575 * Only a group leader can be exclusive or pinned
6577 if (attr.exclusive || attr.pinned)
6582 err = perf_event_set_output(event, output_event);
6587 event_file = anon_inode_getfile("[perf_event]", &perf_fops, event, O_RDWR);
6588 if (IS_ERR(event_file)) {
6589 err = PTR_ERR(event_file);
6594 struct perf_event_context *gctx = group_leader->ctx;
6596 mutex_lock(&gctx->mutex);
6597 perf_remove_from_context(group_leader);
6598 list_for_each_entry(sibling, &group_leader->sibling_list,
6600 perf_remove_from_context(sibling);
6603 mutex_unlock(&gctx->mutex);
6607 event->filp = event_file;
6608 WARN_ON_ONCE(ctx->parent_ctx);
6609 mutex_lock(&ctx->mutex);
6612 perf_install_in_context(ctx, group_leader, cpu);
6614 list_for_each_entry(sibling, &group_leader->sibling_list,
6616 perf_install_in_context(ctx, sibling, cpu);
6621 perf_install_in_context(ctx, event, cpu);
6623 perf_unpin_context(ctx);
6624 mutex_unlock(&ctx->mutex);
6626 event->owner = current;
6628 mutex_lock(¤t->perf_event_mutex);
6629 list_add_tail(&event->owner_entry, ¤t->perf_event_list);
6630 mutex_unlock(¤t->perf_event_mutex);
6633 * Precalculate sample_data sizes
6635 perf_event__header_size(event);
6636 perf_event__id_header_size(event);
6639 * Drop the reference on the group_event after placing the
6640 * new event on the sibling_list. This ensures destruction
6641 * of the group leader will find the pointer to itself in
6642 * perf_group_detach().
6644 fput_light(group_file, fput_needed);
6645 fd_install(event_fd, event_file);
6649 perf_unpin_context(ctx);
6655 put_task_struct(task);
6657 fput_light(group_file, fput_needed);
6659 put_unused_fd(event_fd);
6664 * perf_event_create_kernel_counter
6666 * @attr: attributes of the counter to create
6667 * @cpu: cpu in which the counter is bound
6668 * @task: task to profile (NULL for percpu)
6671 perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu,
6672 struct task_struct *task,
6673 perf_overflow_handler_t overflow_handler)
6675 struct perf_event_context *ctx;
6676 struct perf_event *event;
6680 * Get the target context (task or percpu):
6683 event = perf_event_alloc(attr, cpu, task, NULL, NULL, overflow_handler);
6684 if (IS_ERR(event)) {
6685 err = PTR_ERR(event);
6689 ctx = find_get_context(event->pmu, task, cpu);
6696 WARN_ON_ONCE(ctx->parent_ctx);
6697 mutex_lock(&ctx->mutex);
6698 perf_install_in_context(ctx, event, cpu);
6700 perf_unpin_context(ctx);
6701 mutex_unlock(&ctx->mutex);
6708 return ERR_PTR(err);
6710 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter);
6712 static void sync_child_event(struct perf_event *child_event,
6713 struct task_struct *child)
6715 struct perf_event *parent_event = child_event->parent;
6718 if (child_event->attr.inherit_stat)
6719 perf_event_read_event(child_event, child);
6721 child_val = perf_event_count(child_event);
6724 * Add back the child's count to the parent's count:
6726 atomic64_add(child_val, &parent_event->child_count);
6727 atomic64_add(child_event->total_time_enabled,
6728 &parent_event->child_total_time_enabled);
6729 atomic64_add(child_event->total_time_running,
6730 &parent_event->child_total_time_running);
6733 * Remove this event from the parent's list
6735 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
6736 mutex_lock(&parent_event->child_mutex);
6737 list_del_init(&child_event->child_list);
6738 mutex_unlock(&parent_event->child_mutex);
6741 * Release the parent event, if this was the last
6744 fput(parent_event->filp);
6748 __perf_event_exit_task(struct perf_event *child_event,
6749 struct perf_event_context *child_ctx,
6750 struct task_struct *child)
6752 if (child_event->parent) {
6753 raw_spin_lock_irq(&child_ctx->lock);
6754 perf_group_detach(child_event);
6755 raw_spin_unlock_irq(&child_ctx->lock);
6758 perf_remove_from_context(child_event);
6761 * It can happen that the parent exits first, and has events
6762 * that are still around due to the child reference. These
6763 * events need to be zapped.
6765 if (child_event->parent) {
6766 sync_child_event(child_event, child);
6767 free_event(child_event);
6771 static void perf_event_exit_task_context(struct task_struct *child, int ctxn)
6773 struct perf_event *child_event, *tmp;
6774 struct perf_event_context *child_ctx;
6775 unsigned long flags;
6777 if (likely(!child->perf_event_ctxp[ctxn])) {
6778 perf_event_task(child, NULL, 0);
6782 local_irq_save(flags);
6784 * We can't reschedule here because interrupts are disabled,
6785 * and either child is current or it is a task that can't be
6786 * scheduled, so we are now safe from rescheduling changing
6789 child_ctx = rcu_dereference_raw(child->perf_event_ctxp[ctxn]);
6792 * Take the context lock here so that if find_get_context is
6793 * reading child->perf_event_ctxp, we wait until it has
6794 * incremented the context's refcount before we do put_ctx below.
6796 raw_spin_lock(&child_ctx->lock);
6797 task_ctx_sched_out(child_ctx, EVENT_ALL);
6798 child->perf_event_ctxp[ctxn] = NULL;
6800 * If this context is a clone; unclone it so it can't get
6801 * swapped to another process while we're removing all
6802 * the events from it.
6804 unclone_ctx(child_ctx);
6805 update_context_time(child_ctx);
6806 raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
6809 * Report the task dead after unscheduling the events so that we
6810 * won't get any samples after PERF_RECORD_EXIT. We can however still
6811 * get a few PERF_RECORD_READ events.
6813 perf_event_task(child, child_ctx, 0);
6816 * We can recurse on the same lock type through:
6818 * __perf_event_exit_task()
6819 * sync_child_event()
6820 * fput(parent_event->filp)
6822 * mutex_lock(&ctx->mutex)
6824 * But since its the parent context it won't be the same instance.
6826 mutex_lock(&child_ctx->mutex);
6829 list_for_each_entry_safe(child_event, tmp, &child_ctx->pinned_groups,
6831 __perf_event_exit_task(child_event, child_ctx, child);
6833 list_for_each_entry_safe(child_event, tmp, &child_ctx->flexible_groups,
6835 __perf_event_exit_task(child_event, child_ctx, child);
6838 * If the last event was a group event, it will have appended all
6839 * its siblings to the list, but we obtained 'tmp' before that which
6840 * will still point to the list head terminating the iteration.
6842 if (!list_empty(&child_ctx->pinned_groups) ||
6843 !list_empty(&child_ctx->flexible_groups))
6846 mutex_unlock(&child_ctx->mutex);
6852 * When a child task exits, feed back event values to parent events.
6854 void perf_event_exit_task(struct task_struct *child)
6856 struct perf_event *event, *tmp;
6859 mutex_lock(&child->perf_event_mutex);
6860 list_for_each_entry_safe(event, tmp, &child->perf_event_list,
6862 list_del_init(&event->owner_entry);
6865 * Ensure the list deletion is visible before we clear
6866 * the owner, closes a race against perf_release() where
6867 * we need to serialize on the owner->perf_event_mutex.
6870 event->owner = NULL;
6872 mutex_unlock(&child->perf_event_mutex);
6874 for_each_task_context_nr(ctxn)
6875 perf_event_exit_task_context(child, ctxn);
6878 static void perf_free_event(struct perf_event *event,
6879 struct perf_event_context *ctx)
6881 struct perf_event *parent = event->parent;
6883 if (WARN_ON_ONCE(!parent))
6886 mutex_lock(&parent->child_mutex);
6887 list_del_init(&event->child_list);
6888 mutex_unlock(&parent->child_mutex);
6892 perf_group_detach(event);
6893 list_del_event(event, ctx);
6898 * free an unexposed, unused context as created by inheritance by
6899 * perf_event_init_task below, used by fork() in case of fail.
6901 void perf_event_free_task(struct task_struct *task)
6903 struct perf_event_context *ctx;
6904 struct perf_event *event, *tmp;
6907 for_each_task_context_nr(ctxn) {
6908 ctx = task->perf_event_ctxp[ctxn];
6912 mutex_lock(&ctx->mutex);
6914 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups,
6916 perf_free_event(event, ctx);
6918 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups,
6920 perf_free_event(event, ctx);
6922 if (!list_empty(&ctx->pinned_groups) ||
6923 !list_empty(&ctx->flexible_groups))
6926 mutex_unlock(&ctx->mutex);
6932 void perf_event_delayed_put(struct task_struct *task)
6936 for_each_task_context_nr(ctxn)
6937 WARN_ON_ONCE(task->perf_event_ctxp[ctxn]);
6941 * inherit a event from parent task to child task:
6943 static struct perf_event *
6944 inherit_event(struct perf_event *parent_event,
6945 struct task_struct *parent,
6946 struct perf_event_context *parent_ctx,
6947 struct task_struct *child,
6948 struct perf_event *group_leader,
6949 struct perf_event_context *child_ctx)
6951 struct perf_event *child_event;
6952 unsigned long flags;
6955 * Instead of creating recursive hierarchies of events,
6956 * we link inherited events back to the original parent,
6957 * which has a filp for sure, which we use as the reference
6960 if (parent_event->parent)
6961 parent_event = parent_event->parent;
6963 child_event = perf_event_alloc(&parent_event->attr,
6966 group_leader, parent_event,
6968 if (IS_ERR(child_event))
6973 * Make the child state follow the state of the parent event,
6974 * not its attr.disabled bit. We hold the parent's mutex,
6975 * so we won't race with perf_event_{en, dis}able_family.
6977 if (parent_event->state >= PERF_EVENT_STATE_INACTIVE)
6978 child_event->state = PERF_EVENT_STATE_INACTIVE;
6980 child_event->state = PERF_EVENT_STATE_OFF;
6982 if (parent_event->attr.freq) {
6983 u64 sample_period = parent_event->hw.sample_period;
6984 struct hw_perf_event *hwc = &child_event->hw;
6986 hwc->sample_period = sample_period;
6987 hwc->last_period = sample_period;
6989 local64_set(&hwc->period_left, sample_period);
6992 child_event->ctx = child_ctx;
6993 child_event->overflow_handler = parent_event->overflow_handler;
6996 * Precalculate sample_data sizes
6998 perf_event__header_size(child_event);
6999 perf_event__id_header_size(child_event);
7002 * Link it up in the child's context:
7004 raw_spin_lock_irqsave(&child_ctx->lock, flags);
7005 add_event_to_ctx(child_event, child_ctx);
7006 raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
7009 * Get a reference to the parent filp - we will fput it
7010 * when the child event exits. This is safe to do because
7011 * we are in the parent and we know that the filp still
7012 * exists and has a nonzero count:
7014 atomic_long_inc(&parent_event->filp->f_count);
7017 * Link this into the parent event's child list
7019 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
7020 mutex_lock(&parent_event->child_mutex);
7021 list_add_tail(&child_event->child_list, &parent_event->child_list);
7022 mutex_unlock(&parent_event->child_mutex);
7027 static int inherit_group(struct perf_event *parent_event,
7028 struct task_struct *parent,
7029 struct perf_event_context *parent_ctx,
7030 struct task_struct *child,
7031 struct perf_event_context *child_ctx)
7033 struct perf_event *leader;
7034 struct perf_event *sub;
7035 struct perf_event *child_ctr;
7037 leader = inherit_event(parent_event, parent, parent_ctx,
7038 child, NULL, child_ctx);
7040 return PTR_ERR(leader);
7041 list_for_each_entry(sub, &parent_event->sibling_list, group_entry) {
7042 child_ctr = inherit_event(sub, parent, parent_ctx,
7043 child, leader, child_ctx);
7044 if (IS_ERR(child_ctr))
7045 return PTR_ERR(child_ctr);
7051 inherit_task_group(struct perf_event *event, struct task_struct *parent,
7052 struct perf_event_context *parent_ctx,
7053 struct task_struct *child, int ctxn,
7057 struct perf_event_context *child_ctx;
7059 if (!event->attr.inherit) {
7064 child_ctx = child->perf_event_ctxp[ctxn];
7067 * This is executed from the parent task context, so
7068 * inherit events that have been marked for cloning.
7069 * First allocate and initialize a context for the
7073 child_ctx = alloc_perf_context(event->pmu, child);
7077 child->perf_event_ctxp[ctxn] = child_ctx;
7080 ret = inherit_group(event, parent, parent_ctx,
7090 * Initialize the perf_event context in task_struct
7092 int perf_event_init_context(struct task_struct *child, int ctxn)
7094 struct perf_event_context *child_ctx, *parent_ctx;
7095 struct perf_event_context *cloned_ctx;
7096 struct perf_event *event;
7097 struct task_struct *parent = current;
7098 int inherited_all = 1;
7099 unsigned long flags;
7102 if (likely(!parent->perf_event_ctxp[ctxn]))
7106 * If the parent's context is a clone, pin it so it won't get
7109 parent_ctx = perf_pin_task_context(parent, ctxn);
7112 * No need to check if parent_ctx != NULL here; since we saw
7113 * it non-NULL earlier, the only reason for it to become NULL
7114 * is if we exit, and since we're currently in the middle of
7115 * a fork we can't be exiting at the same time.
7119 * Lock the parent list. No need to lock the child - not PID
7120 * hashed yet and not running, so nobody can access it.
7122 mutex_lock(&parent_ctx->mutex);
7125 * We dont have to disable NMIs - we are only looking at
7126 * the list, not manipulating it:
7128 list_for_each_entry(event, &parent_ctx->pinned_groups, group_entry) {
7129 ret = inherit_task_group(event, parent, parent_ctx,
7130 child, ctxn, &inherited_all);
7136 * We can't hold ctx->lock when iterating the ->flexible_group list due
7137 * to allocations, but we need to prevent rotation because
7138 * rotate_ctx() will change the list from interrupt context.
7140 raw_spin_lock_irqsave(&parent_ctx->lock, flags);
7141 parent_ctx->rotate_disable = 1;
7142 raw_spin_unlock_irqrestore(&parent_ctx->lock, flags);
7144 list_for_each_entry(event, &parent_ctx->flexible_groups, group_entry) {
7145 ret = inherit_task_group(event, parent, parent_ctx,
7146 child, ctxn, &inherited_all);
7151 raw_spin_lock_irqsave(&parent_ctx->lock, flags);
7152 parent_ctx->rotate_disable = 0;
7154 child_ctx = child->perf_event_ctxp[ctxn];
7156 if (child_ctx && inherited_all) {
7158 * Mark the child context as a clone of the parent
7159 * context, or of whatever the parent is a clone of.
7161 * Note that if the parent is a clone, the holding of
7162 * parent_ctx->lock avoids it from being uncloned.
7164 cloned_ctx = parent_ctx->parent_ctx;
7166 child_ctx->parent_ctx = cloned_ctx;
7167 child_ctx->parent_gen = parent_ctx->parent_gen;
7169 child_ctx->parent_ctx = parent_ctx;
7170 child_ctx->parent_gen = parent_ctx->generation;
7172 get_ctx(child_ctx->parent_ctx);
7175 raw_spin_unlock_irqrestore(&parent_ctx->lock, flags);
7176 mutex_unlock(&parent_ctx->mutex);
7178 perf_unpin_context(parent_ctx);
7179 put_ctx(parent_ctx);
7185 * Initialize the perf_event context in task_struct
7187 int perf_event_init_task(struct task_struct *child)
7191 memset(child->perf_event_ctxp, 0, sizeof(child->perf_event_ctxp));
7192 mutex_init(&child->perf_event_mutex);
7193 INIT_LIST_HEAD(&child->perf_event_list);
7195 for_each_task_context_nr(ctxn) {
7196 ret = perf_event_init_context(child, ctxn);
7204 static void __init perf_event_init_all_cpus(void)
7206 struct swevent_htable *swhash;
7209 for_each_possible_cpu(cpu) {
7210 swhash = &per_cpu(swevent_htable, cpu);
7211 mutex_init(&swhash->hlist_mutex);
7212 INIT_LIST_HEAD(&per_cpu(rotation_list, cpu));
7216 static void __cpuinit perf_event_init_cpu(int cpu)
7218 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
7220 mutex_lock(&swhash->hlist_mutex);
7221 if (swhash->hlist_refcount > 0) {
7222 struct swevent_hlist *hlist;
7224 hlist = kzalloc_node(sizeof(*hlist), GFP_KERNEL, cpu_to_node(cpu));
7226 rcu_assign_pointer(swhash->swevent_hlist, hlist);
7228 mutex_unlock(&swhash->hlist_mutex);
7231 #if defined CONFIG_HOTPLUG_CPU || defined CONFIG_KEXEC
7232 static void perf_pmu_rotate_stop(struct pmu *pmu)
7234 struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
7236 WARN_ON(!irqs_disabled());
7238 list_del_init(&cpuctx->rotation_list);
7241 static void __perf_event_exit_context(void *__info)
7243 struct perf_event_context *ctx = __info;
7244 struct perf_event *event, *tmp;
7246 perf_pmu_rotate_stop(ctx->pmu);
7248 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry)
7249 __perf_remove_from_context(event);
7250 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups, group_entry)
7251 __perf_remove_from_context(event);
7254 static void perf_event_exit_cpu_context(int cpu)
7256 struct perf_event_context *ctx;
7260 idx = srcu_read_lock(&pmus_srcu);
7261 list_for_each_entry_rcu(pmu, &pmus, entry) {
7262 ctx = &per_cpu_ptr(pmu->pmu_cpu_context, cpu)->ctx;
7264 mutex_lock(&ctx->mutex);
7265 smp_call_function_single(cpu, __perf_event_exit_context, ctx, 1);
7266 mutex_unlock(&ctx->mutex);
7268 srcu_read_unlock(&pmus_srcu, idx);
7271 static void perf_event_exit_cpu(int cpu)
7273 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
7275 mutex_lock(&swhash->hlist_mutex);
7276 swevent_hlist_release(swhash);
7277 mutex_unlock(&swhash->hlist_mutex);
7279 perf_event_exit_cpu_context(cpu);
7282 static inline void perf_event_exit_cpu(int cpu) { }
7286 perf_reboot(struct notifier_block *notifier, unsigned long val, void *v)
7290 for_each_online_cpu(cpu)
7291 perf_event_exit_cpu(cpu);
7297 * Run the perf reboot notifier at the very last possible moment so that
7298 * the generic watchdog code runs as long as possible.
7300 static struct notifier_block perf_reboot_notifier = {
7301 .notifier_call = perf_reboot,
7302 .priority = INT_MIN,
7305 static int __cpuinit
7306 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
7308 unsigned int cpu = (long)hcpu;
7310 switch (action & ~CPU_TASKS_FROZEN) {
7312 case CPU_UP_PREPARE:
7313 case CPU_DOWN_FAILED:
7314 perf_event_init_cpu(cpu);
7317 case CPU_UP_CANCELED:
7318 case CPU_DOWN_PREPARE:
7319 perf_event_exit_cpu(cpu);
7329 void __init perf_event_init(void)
7335 perf_event_init_all_cpus();
7336 init_srcu_struct(&pmus_srcu);
7337 perf_pmu_register(&perf_swevent, "software", PERF_TYPE_SOFTWARE);
7338 perf_pmu_register(&perf_cpu_clock, NULL, -1);
7339 perf_pmu_register(&perf_task_clock, NULL, -1);
7341 perf_cpu_notifier(perf_cpu_notify);
7342 register_reboot_notifier(&perf_reboot_notifier);
7344 ret = init_hw_breakpoint();
7345 WARN(ret, "hw_breakpoint initialization failed with: %d", ret);
7348 static int __init perf_event_sysfs_init(void)
7353 mutex_lock(&pmus_lock);
7355 ret = bus_register(&pmu_bus);
7359 list_for_each_entry(pmu, &pmus, entry) {
7360 if (!pmu->name || pmu->type < 0)
7363 ret = pmu_dev_alloc(pmu);
7364 WARN(ret, "Failed to register pmu: %s, reason %d\n", pmu->name, ret);
7366 pmu_bus_running = 1;
7370 mutex_unlock(&pmus_lock);
7374 device_initcall(perf_event_sysfs_init);
7376 #ifdef CONFIG_CGROUP_PERF
7377 static struct cgroup_subsys_state *perf_cgroup_create(
7378 struct cgroup_subsys *ss, struct cgroup *cont)
7380 struct perf_cgroup *jc;
7382 jc = kzalloc(sizeof(*jc), GFP_KERNEL);
7384 return ERR_PTR(-ENOMEM);
7386 jc->info = alloc_percpu(struct perf_cgroup_info);
7389 return ERR_PTR(-ENOMEM);
7395 static void perf_cgroup_destroy(struct cgroup_subsys *ss,
7396 struct cgroup *cont)
7398 struct perf_cgroup *jc;
7399 jc = container_of(cgroup_subsys_state(cont, perf_subsys_id),
7400 struct perf_cgroup, css);
7401 free_percpu(jc->info);
7405 static int __perf_cgroup_move(void *info)
7407 struct task_struct *task = info;
7408 perf_cgroup_switch(task, PERF_CGROUP_SWOUT | PERF_CGROUP_SWIN);
7412 static void perf_cgroup_move(struct task_struct *task)
7414 task_function_call(task, __perf_cgroup_move, task);
7417 static void perf_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
7418 struct cgroup *old_cgrp, struct task_struct *task,
7421 perf_cgroup_move(task);
7423 struct task_struct *c;
7425 list_for_each_entry_rcu(c, &task->thread_group, thread_group) {
7426 perf_cgroup_move(c);
7432 static void perf_cgroup_exit(struct cgroup_subsys *ss, struct cgroup *cgrp,
7433 struct cgroup *old_cgrp, struct task_struct *task)
7436 * cgroup_exit() is called in the copy_process() failure path.
7437 * Ignore this case since the task hasn't ran yet, this avoids
7438 * trying to poke a half freed task state from generic code.
7440 if (!(task->flags & PF_EXITING))
7443 perf_cgroup_move(task);
7446 struct cgroup_subsys perf_subsys = {
7447 .name = "perf_event",
7448 .subsys_id = perf_subsys_id,
7449 .create = perf_cgroup_create,
7450 .destroy = perf_cgroup_destroy,
7451 .exit = perf_cgroup_exit,
7452 .attach = perf_cgroup_attach,
7454 #endif /* CONFIG_CGROUP_PERF */