2 * Performance events core code:
4 * Copyright (C) 2008 Thomas Gleixner <tglx@linutronix.de>
5 * Copyright (C) 2008-2009 Red Hat, Inc., Ingo Molnar
6 * Copyright (C) 2008-2009 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
7 * Copyright © 2009 Paul Mackerras, IBM Corp. <paulus@au1.ibm.com>
9 * For licensing details see kernel-base/COPYING
14 #include <linux/cpu.h>
15 #include <linux/smp.h>
16 #include <linux/idr.h>
17 #include <linux/file.h>
18 #include <linux/poll.h>
19 #include <linux/slab.h>
20 #include <linux/hash.h>
21 #include <linux/sysfs.h>
22 #include <linux/dcache.h>
23 #include <linux/percpu.h>
24 #include <linux/ptrace.h>
25 #include <linux/reboot.h>
26 #include <linux/vmstat.h>
27 #include <linux/device.h>
28 #include <linux/vmalloc.h>
29 #include <linux/hardirq.h>
30 #include <linux/rculist.h>
31 #include <linux/uaccess.h>
32 #include <linux/syscalls.h>
33 #include <linux/anon_inodes.h>
34 #include <linux/kernel_stat.h>
35 #include <linux/perf_event.h>
36 #include <linux/ftrace_event.h>
37 #include <linux/hw_breakpoint.h>
39 #include <asm/irq_regs.h>
41 struct remote_function_call {
42 struct task_struct *p;
43 int (*func)(void *info);
48 static void remote_function(void *data)
50 struct remote_function_call *tfc = data;
51 struct task_struct *p = tfc->p;
55 if (task_cpu(p) != smp_processor_id() || !task_curr(p))
59 tfc->ret = tfc->func(tfc->info);
63 * task_function_call - call a function on the cpu on which a task runs
64 * @p: the task to evaluate
65 * @func: the function to be called
66 * @info: the function call argument
68 * Calls the function @func when the task is currently running. This might
69 * be on the current CPU, which just calls the function directly
71 * returns: @func return value, or
72 * -ESRCH - when the process isn't running
73 * -EAGAIN - when the process moved away
76 task_function_call(struct task_struct *p, int (*func) (void *info), void *info)
78 struct remote_function_call data = {
82 .ret = -ESRCH, /* No such (running) process */
86 smp_call_function_single(task_cpu(p), remote_function, &data, 1);
92 * cpu_function_call - call a function on the cpu
93 * @func: the function to be called
94 * @info: the function call argument
96 * Calls the function @func on the remote cpu.
98 * returns: @func return value or -ENXIO when the cpu is offline
100 static int cpu_function_call(int cpu, int (*func) (void *info), void *info)
102 struct remote_function_call data = {
106 .ret = -ENXIO, /* No such CPU */
109 smp_call_function_single(cpu, remote_function, &data, 1);
114 #define PERF_FLAG_ALL (PERF_FLAG_FD_NO_GROUP |\
115 PERF_FLAG_FD_OUTPUT |\
116 PERF_FLAG_PID_CGROUP)
119 EVENT_FLEXIBLE = 0x1,
121 EVENT_ALL = EVENT_FLEXIBLE | EVENT_PINNED,
125 * perf_sched_events : >0 events exist
126 * perf_cgroup_events: >0 per-cpu cgroup events exist on this cpu
128 atomic_t perf_sched_events __read_mostly;
129 static DEFINE_PER_CPU(atomic_t, perf_cgroup_events);
131 static atomic_t nr_mmap_events __read_mostly;
132 static atomic_t nr_comm_events __read_mostly;
133 static atomic_t nr_task_events __read_mostly;
135 static LIST_HEAD(pmus);
136 static DEFINE_MUTEX(pmus_lock);
137 static struct srcu_struct pmus_srcu;
140 * perf event paranoia level:
141 * -1 - not paranoid at all
142 * 0 - disallow raw tracepoint access for unpriv
143 * 1 - disallow cpu events for unpriv
144 * 2 - disallow kernel profiling for unpriv
146 int sysctl_perf_event_paranoid __read_mostly = 1;
148 /* 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 #ifdef CONFIG_CGROUP_PERF
206 * Must ensure cgroup is pinned (css_get) before calling
207 * this function. In other words, we cannot call this function
208 * if there is no cgroup event for the current CPU context.
210 static inline struct perf_cgroup *
211 perf_cgroup_from_task(struct task_struct *task)
213 return container_of(task_subsys_state(task, perf_subsys_id),
214 struct perf_cgroup, css);
218 perf_cgroup_match(struct perf_event *event)
220 struct perf_event_context *ctx = event->ctx;
221 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
223 return !event->cgrp || event->cgrp == cpuctx->cgrp;
226 static inline void perf_get_cgroup(struct perf_event *event)
228 css_get(&event->cgrp->css);
231 static inline void perf_put_cgroup(struct perf_event *event)
233 css_put(&event->cgrp->css);
236 static inline void perf_detach_cgroup(struct perf_event *event)
238 perf_put_cgroup(event);
242 static inline int is_cgroup_event(struct perf_event *event)
244 return event->cgrp != NULL;
247 static inline u64 perf_cgroup_event_time(struct perf_event *event)
249 struct perf_cgroup_info *t;
251 t = per_cpu_ptr(event->cgrp->info, event->cpu);
255 static inline void __update_cgrp_time(struct perf_cgroup *cgrp)
257 struct perf_cgroup_info *info;
262 info = this_cpu_ptr(cgrp->info);
264 info->time += now - info->timestamp;
265 info->timestamp = now;
268 static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx)
270 struct perf_cgroup *cgrp_out = cpuctx->cgrp;
272 __update_cgrp_time(cgrp_out);
275 static inline void update_cgrp_time_from_event(struct perf_event *event)
277 struct perf_cgroup *cgrp;
280 * ensure we access cgroup data only when needed and
281 * when we know the cgroup is pinned (css_get)
283 if (!is_cgroup_event(event))
286 cgrp = perf_cgroup_from_task(current);
288 * Do not update time when cgroup is not active
290 if (cgrp == event->cgrp)
291 __update_cgrp_time(event->cgrp);
295 perf_cgroup_set_timestamp(struct task_struct *task,
296 struct perf_event_context *ctx)
298 struct perf_cgroup *cgrp;
299 struct perf_cgroup_info *info;
302 * ctx->lock held by caller
303 * ensure we do not access cgroup data
304 * unless we have the cgroup pinned (css_get)
306 if (!task || !ctx->nr_cgroups)
309 cgrp = perf_cgroup_from_task(task);
310 info = this_cpu_ptr(cgrp->info);
311 info->timestamp = ctx->timestamp;
314 #define PERF_CGROUP_SWOUT 0x1 /* cgroup switch out every event */
315 #define PERF_CGROUP_SWIN 0x2 /* cgroup switch in events based on task */
318 * reschedule events based on the cgroup constraint of task.
320 * mode SWOUT : schedule out everything
321 * mode SWIN : schedule in based on cgroup for next
323 void perf_cgroup_switch(struct task_struct *task, int mode)
325 struct perf_cpu_context *cpuctx;
330 * disable interrupts to avoid geting nr_cgroup
331 * changes via __perf_event_disable(). Also
334 local_irq_save(flags);
337 * we reschedule only in the presence of cgroup
338 * constrained events.
342 list_for_each_entry_rcu(pmu, &pmus, entry) {
344 cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
346 perf_pmu_disable(cpuctx->ctx.pmu);
349 * perf_cgroup_events says at least one
350 * context on this CPU has cgroup events.
352 * ctx->nr_cgroups reports the number of cgroup
353 * events for a context.
355 if (cpuctx->ctx.nr_cgroups > 0) {
357 if (mode & PERF_CGROUP_SWOUT) {
358 cpu_ctx_sched_out(cpuctx, EVENT_ALL);
360 * must not be done before ctxswout due
361 * to event_filter_match() in event_sched_out()
366 if (mode & PERF_CGROUP_SWIN) {
367 /* set cgrp before ctxsw in to
368 * allow event_filter_match() to not
369 * have to pass task around
371 cpuctx->cgrp = perf_cgroup_from_task(task);
372 cpu_ctx_sched_in(cpuctx, EVENT_ALL, task);
376 perf_pmu_enable(cpuctx->ctx.pmu);
381 local_irq_restore(flags);
384 static inline void perf_cgroup_sched_out(struct task_struct *task)
386 perf_cgroup_switch(task, PERF_CGROUP_SWOUT);
389 static inline void perf_cgroup_sched_in(struct task_struct *task)
391 perf_cgroup_switch(task, PERF_CGROUP_SWIN);
394 static inline int perf_cgroup_connect(int fd, struct perf_event *event,
395 struct perf_event_attr *attr,
396 struct perf_event *group_leader)
398 struct perf_cgroup *cgrp;
399 struct cgroup_subsys_state *css;
401 int ret = 0, fput_needed;
403 file = fget_light(fd, &fput_needed);
407 css = cgroup_css_from_dir(file, perf_subsys_id);
413 cgrp = container_of(css, struct perf_cgroup, css);
416 /* must be done before we fput() the file */
417 perf_get_cgroup(event);
420 * all events in a group must monitor
421 * the same cgroup because a task belongs
422 * to only one perf cgroup at a time
424 if (group_leader && group_leader->cgrp != cgrp) {
425 perf_detach_cgroup(event);
429 fput_light(file, fput_needed);
434 perf_cgroup_set_shadow_time(struct perf_event *event, u64 now)
436 struct perf_cgroup_info *t;
437 t = per_cpu_ptr(event->cgrp->info, event->cpu);
438 event->shadow_ctx_time = now - t->timestamp;
442 perf_cgroup_defer_enabled(struct perf_event *event)
445 * when the current task's perf cgroup does not match
446 * the event's, we need to remember to call the
447 * perf_mark_enable() function the first time a task with
448 * a matching perf cgroup is scheduled in.
450 if (is_cgroup_event(event) && !perf_cgroup_match(event))
451 event->cgrp_defer_enabled = 1;
455 perf_cgroup_mark_enabled(struct perf_event *event,
456 struct perf_event_context *ctx)
458 struct perf_event *sub;
459 u64 tstamp = perf_event_time(event);
461 if (!event->cgrp_defer_enabled)
464 event->cgrp_defer_enabled = 0;
466 event->tstamp_enabled = tstamp - event->total_time_enabled;
467 list_for_each_entry(sub, &event->sibling_list, group_entry) {
468 if (sub->state >= PERF_EVENT_STATE_INACTIVE) {
469 sub->tstamp_enabled = tstamp - sub->total_time_enabled;
470 sub->cgrp_defer_enabled = 0;
474 #else /* !CONFIG_CGROUP_PERF */
477 perf_cgroup_match(struct perf_event *event)
482 static inline void perf_detach_cgroup(struct perf_event *event)
485 static inline int is_cgroup_event(struct perf_event *event)
490 static inline u64 perf_cgroup_event_cgrp_time(struct perf_event *event)
495 static inline void update_cgrp_time_from_event(struct perf_event *event)
499 static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx)
503 static inline void perf_cgroup_sched_out(struct task_struct *task)
507 static inline void perf_cgroup_sched_in(struct task_struct *task)
511 static inline int perf_cgroup_connect(pid_t pid, struct perf_event *event,
512 struct perf_event_attr *attr,
513 struct perf_event *group_leader)
519 perf_cgroup_set_timestamp(struct task_struct *task,
520 struct perf_event_context *ctx)
525 perf_cgroup_switch(struct task_struct *task, struct task_struct *next)
530 perf_cgroup_set_shadow_time(struct perf_event *event, u64 now)
534 static inline u64 perf_cgroup_event_time(struct perf_event *event)
540 perf_cgroup_defer_enabled(struct perf_event *event)
545 perf_cgroup_mark_enabled(struct perf_event *event,
546 struct perf_event_context *ctx)
551 void perf_pmu_disable(struct pmu *pmu)
553 int *count = this_cpu_ptr(pmu->pmu_disable_count);
555 pmu->pmu_disable(pmu);
558 void perf_pmu_enable(struct pmu *pmu)
560 int *count = this_cpu_ptr(pmu->pmu_disable_count);
562 pmu->pmu_enable(pmu);
565 static DEFINE_PER_CPU(struct list_head, rotation_list);
568 * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
569 * because they're strictly cpu affine and rotate_start is called with IRQs
570 * disabled, while rotate_context is called from IRQ context.
572 static void perf_pmu_rotate_start(struct pmu *pmu)
574 struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
575 struct list_head *head = &__get_cpu_var(rotation_list);
577 WARN_ON(!irqs_disabled());
579 if (list_empty(&cpuctx->rotation_list))
580 list_add(&cpuctx->rotation_list, head);
583 static void get_ctx(struct perf_event_context *ctx)
585 WARN_ON(!atomic_inc_not_zero(&ctx->refcount));
588 static void free_ctx(struct rcu_head *head)
590 struct perf_event_context *ctx;
592 ctx = container_of(head, struct perf_event_context, rcu_head);
596 static void put_ctx(struct perf_event_context *ctx)
598 if (atomic_dec_and_test(&ctx->refcount)) {
600 put_ctx(ctx->parent_ctx);
602 put_task_struct(ctx->task);
603 call_rcu(&ctx->rcu_head, free_ctx);
607 static void unclone_ctx(struct perf_event_context *ctx)
609 if (ctx->parent_ctx) {
610 put_ctx(ctx->parent_ctx);
611 ctx->parent_ctx = NULL;
615 static u32 perf_event_pid(struct perf_event *event, struct task_struct *p)
618 * only top level events have the pid namespace they were created in
621 event = event->parent;
623 return task_tgid_nr_ns(p, event->ns);
626 static u32 perf_event_tid(struct perf_event *event, struct task_struct *p)
629 * only top level events have the pid namespace they were created in
632 event = event->parent;
634 return task_pid_nr_ns(p, event->ns);
638 * If we inherit events we want to return the parent event id
641 static u64 primary_event_id(struct perf_event *event)
646 id = event->parent->id;
652 * Get the perf_event_context for a task and lock it.
653 * This has to cope with with the fact that until it is locked,
654 * the context could get moved to another task.
656 static struct perf_event_context *
657 perf_lock_task_context(struct task_struct *task, int ctxn, unsigned long *flags)
659 struct perf_event_context *ctx;
663 ctx = rcu_dereference(task->perf_event_ctxp[ctxn]);
666 * If this context is a clone of another, it might
667 * get swapped for another underneath us by
668 * perf_event_task_sched_out, though the
669 * rcu_read_lock() protects us from any context
670 * getting freed. Lock the context and check if it
671 * got swapped before we could get the lock, and retry
672 * if so. If we locked the right context, then it
673 * can't get swapped on us any more.
675 raw_spin_lock_irqsave(&ctx->lock, *flags);
676 if (ctx != rcu_dereference(task->perf_event_ctxp[ctxn])) {
677 raw_spin_unlock_irqrestore(&ctx->lock, *flags);
681 if (!atomic_inc_not_zero(&ctx->refcount)) {
682 raw_spin_unlock_irqrestore(&ctx->lock, *flags);
691 * Get the context for a task and increment its pin_count so it
692 * can't get swapped to another task. This also increments its
693 * reference count so that the context can't get freed.
695 static struct perf_event_context *
696 perf_pin_task_context(struct task_struct *task, int ctxn)
698 struct perf_event_context *ctx;
701 ctx = perf_lock_task_context(task, ctxn, &flags);
704 raw_spin_unlock_irqrestore(&ctx->lock, flags);
709 static void perf_unpin_context(struct perf_event_context *ctx)
713 raw_spin_lock_irqsave(&ctx->lock, flags);
715 raw_spin_unlock_irqrestore(&ctx->lock, flags);
719 * Update the record of the current time in a context.
721 static void update_context_time(struct perf_event_context *ctx)
723 u64 now = perf_clock();
725 ctx->time += now - ctx->timestamp;
726 ctx->timestamp = now;
729 static u64 perf_event_time(struct perf_event *event)
731 struct perf_event_context *ctx = event->ctx;
733 if (is_cgroup_event(event))
734 return perf_cgroup_event_time(event);
736 return ctx ? ctx->time : 0;
740 * Update the total_time_enabled and total_time_running fields for a event.
742 static void update_event_times(struct perf_event *event)
744 struct perf_event_context *ctx = event->ctx;
747 if (event->state < PERF_EVENT_STATE_INACTIVE ||
748 event->group_leader->state < PERF_EVENT_STATE_INACTIVE)
751 * in cgroup mode, time_enabled represents
752 * the time the event was enabled AND active
753 * tasks were in the monitored cgroup. This is
754 * independent of the activity of the context as
755 * there may be a mix of cgroup and non-cgroup events.
757 * That is why we treat cgroup events differently
760 if (is_cgroup_event(event))
761 run_end = perf_event_time(event);
762 else if (ctx->is_active)
765 run_end = event->tstamp_stopped;
767 event->total_time_enabled = run_end - event->tstamp_enabled;
769 if (event->state == PERF_EVENT_STATE_INACTIVE)
770 run_end = event->tstamp_stopped;
772 run_end = perf_event_time(event);
774 event->total_time_running = run_end - event->tstamp_running;
779 * Update total_time_enabled and total_time_running for all events in a group.
781 static void update_group_times(struct perf_event *leader)
783 struct perf_event *event;
785 update_event_times(leader);
786 list_for_each_entry(event, &leader->sibling_list, group_entry)
787 update_event_times(event);
790 static struct list_head *
791 ctx_group_list(struct perf_event *event, struct perf_event_context *ctx)
793 if (event->attr.pinned)
794 return &ctx->pinned_groups;
796 return &ctx->flexible_groups;
800 * Add a event from the lists for its context.
801 * Must be called with ctx->mutex and ctx->lock held.
804 list_add_event(struct perf_event *event, struct perf_event_context *ctx)
806 WARN_ON_ONCE(event->attach_state & PERF_ATTACH_CONTEXT);
807 event->attach_state |= PERF_ATTACH_CONTEXT;
810 * If we're a stand alone event or group leader, we go to the context
811 * list, group events are kept attached to the group so that
812 * perf_group_detach can, at all times, locate all siblings.
814 if (event->group_leader == event) {
815 struct list_head *list;
817 if (is_software_event(event))
818 event->group_flags |= PERF_GROUP_SOFTWARE;
820 list = ctx_group_list(event, ctx);
821 list_add_tail(&event->group_entry, list);
824 if (is_cgroup_event(event))
827 list_add_rcu(&event->event_entry, &ctx->event_list);
829 perf_pmu_rotate_start(ctx->pmu);
831 if (event->attr.inherit_stat)
836 * Called at perf_event creation and when events are attached/detached from a
839 static void perf_event__read_size(struct perf_event *event)
841 int entry = sizeof(u64); /* value */
845 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
848 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
851 if (event->attr.read_format & PERF_FORMAT_ID)
852 entry += sizeof(u64);
854 if (event->attr.read_format & PERF_FORMAT_GROUP) {
855 nr += event->group_leader->nr_siblings;
860 event->read_size = size;
863 static void perf_event__header_size(struct perf_event *event)
865 struct perf_sample_data *data;
866 u64 sample_type = event->attr.sample_type;
869 perf_event__read_size(event);
871 if (sample_type & PERF_SAMPLE_IP)
872 size += sizeof(data->ip);
874 if (sample_type & PERF_SAMPLE_ADDR)
875 size += sizeof(data->addr);
877 if (sample_type & PERF_SAMPLE_PERIOD)
878 size += sizeof(data->period);
880 if (sample_type & PERF_SAMPLE_READ)
881 size += event->read_size;
883 event->header_size = size;
886 static void perf_event__id_header_size(struct perf_event *event)
888 struct perf_sample_data *data;
889 u64 sample_type = event->attr.sample_type;
892 if (sample_type & PERF_SAMPLE_TID)
893 size += sizeof(data->tid_entry);
895 if (sample_type & PERF_SAMPLE_TIME)
896 size += sizeof(data->time);
898 if (sample_type & PERF_SAMPLE_ID)
899 size += sizeof(data->id);
901 if (sample_type & PERF_SAMPLE_STREAM_ID)
902 size += sizeof(data->stream_id);
904 if (sample_type & PERF_SAMPLE_CPU)
905 size += sizeof(data->cpu_entry);
907 event->id_header_size = size;
910 static void perf_group_attach(struct perf_event *event)
912 struct perf_event *group_leader = event->group_leader, *pos;
915 * We can have double attach due to group movement in perf_event_open.
917 if (event->attach_state & PERF_ATTACH_GROUP)
920 event->attach_state |= PERF_ATTACH_GROUP;
922 if (group_leader == event)
925 if (group_leader->group_flags & PERF_GROUP_SOFTWARE &&
926 !is_software_event(event))
927 group_leader->group_flags &= ~PERF_GROUP_SOFTWARE;
929 list_add_tail(&event->group_entry, &group_leader->sibling_list);
930 group_leader->nr_siblings++;
932 perf_event__header_size(group_leader);
934 list_for_each_entry(pos, &group_leader->sibling_list, group_entry)
935 perf_event__header_size(pos);
939 * Remove a event from the lists for its context.
940 * Must be called with ctx->mutex and ctx->lock held.
943 list_del_event(struct perf_event *event, struct perf_event_context *ctx)
945 struct perf_cpu_context *cpuctx;
947 * We can have double detach due to exit/hot-unplug + close.
949 if (!(event->attach_state & PERF_ATTACH_CONTEXT))
952 event->attach_state &= ~PERF_ATTACH_CONTEXT;
954 if (is_cgroup_event(event)) {
956 cpuctx = __get_cpu_context(ctx);
958 * if there are no more cgroup events
959 * then cler cgrp to avoid stale pointer
960 * in update_cgrp_time_from_cpuctx()
962 if (!ctx->nr_cgroups)
967 if (event->attr.inherit_stat)
970 list_del_rcu(&event->event_entry);
972 if (event->group_leader == event)
973 list_del_init(&event->group_entry);
975 update_group_times(event);
978 * If event was in error state, then keep it
979 * that way, otherwise bogus counts will be
980 * returned on read(). The only way to get out
981 * of error state is by explicit re-enabling
984 if (event->state > PERF_EVENT_STATE_OFF)
985 event->state = PERF_EVENT_STATE_OFF;
988 static void perf_group_detach(struct perf_event *event)
990 struct perf_event *sibling, *tmp;
991 struct list_head *list = NULL;
994 * We can have double detach due to exit/hot-unplug + close.
996 if (!(event->attach_state & PERF_ATTACH_GROUP))
999 event->attach_state &= ~PERF_ATTACH_GROUP;
1002 * If this is a sibling, remove it from its group.
1004 if (event->group_leader != event) {
1005 list_del_init(&event->group_entry);
1006 event->group_leader->nr_siblings--;
1010 if (!list_empty(&event->group_entry))
1011 list = &event->group_entry;
1014 * If this was a group event with sibling events then
1015 * upgrade the siblings to singleton events by adding them
1016 * to whatever list we are on.
1018 list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) {
1020 list_move_tail(&sibling->group_entry, list);
1021 sibling->group_leader = sibling;
1023 /* Inherit group flags from the previous leader */
1024 sibling->group_flags = event->group_flags;
1028 perf_event__header_size(event->group_leader);
1030 list_for_each_entry(tmp, &event->group_leader->sibling_list, group_entry)
1031 perf_event__header_size(tmp);
1035 event_filter_match(struct perf_event *event)
1037 return (event->cpu == -1 || event->cpu == smp_processor_id())
1038 && perf_cgroup_match(event);
1042 event_sched_out(struct perf_event *event,
1043 struct perf_cpu_context *cpuctx,
1044 struct perf_event_context *ctx)
1046 u64 tstamp = perf_event_time(event);
1049 * An event which could not be activated because of
1050 * filter mismatch still needs to have its timings
1051 * maintained, otherwise bogus information is return
1052 * via read() for time_enabled, time_running:
1054 if (event->state == PERF_EVENT_STATE_INACTIVE
1055 && !event_filter_match(event)) {
1056 delta = tstamp - event->tstamp_stopped;
1057 event->tstamp_running += delta;
1058 event->tstamp_stopped = tstamp;
1061 if (event->state != PERF_EVENT_STATE_ACTIVE)
1064 event->state = PERF_EVENT_STATE_INACTIVE;
1065 if (event->pending_disable) {
1066 event->pending_disable = 0;
1067 event->state = PERF_EVENT_STATE_OFF;
1069 event->tstamp_stopped = tstamp;
1070 event->pmu->del(event, 0);
1073 if (!is_software_event(event))
1074 cpuctx->active_oncpu--;
1076 if (event->attr.exclusive || !cpuctx->active_oncpu)
1077 cpuctx->exclusive = 0;
1081 group_sched_out(struct perf_event *group_event,
1082 struct perf_cpu_context *cpuctx,
1083 struct perf_event_context *ctx)
1085 struct perf_event *event;
1086 int state = group_event->state;
1088 event_sched_out(group_event, cpuctx, ctx);
1091 * Schedule out siblings (if any):
1093 list_for_each_entry(event, &group_event->sibling_list, group_entry)
1094 event_sched_out(event, cpuctx, ctx);
1096 if (state == PERF_EVENT_STATE_ACTIVE && group_event->attr.exclusive)
1097 cpuctx->exclusive = 0;
1101 * Cross CPU call to remove a performance event
1103 * We disable the event on the hardware level first. After that we
1104 * remove it from the context list.
1106 static int __perf_remove_from_context(void *info)
1108 struct perf_event *event = info;
1109 struct perf_event_context *ctx = event->ctx;
1110 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1112 raw_spin_lock(&ctx->lock);
1113 event_sched_out(event, cpuctx, ctx);
1114 list_del_event(event, ctx);
1115 raw_spin_unlock(&ctx->lock);
1122 * Remove the event from a task's (or a CPU's) list of events.
1124 * CPU events are removed with a smp call. For task events we only
1125 * call when the task is on a CPU.
1127 * If event->ctx is a cloned context, callers must make sure that
1128 * every task struct that event->ctx->task could possibly point to
1129 * remains valid. This is OK when called from perf_release since
1130 * that only calls us on the top-level context, which can't be a clone.
1131 * When called from perf_event_exit_task, it's OK because the
1132 * context has been detached from its task.
1134 static void perf_remove_from_context(struct perf_event *event)
1136 struct perf_event_context *ctx = event->ctx;
1137 struct task_struct *task = ctx->task;
1139 lockdep_assert_held(&ctx->mutex);
1143 * Per cpu events are removed via an smp call and
1144 * the removal is always successful.
1146 cpu_function_call(event->cpu, __perf_remove_from_context, event);
1151 if (!task_function_call(task, __perf_remove_from_context, event))
1154 raw_spin_lock_irq(&ctx->lock);
1156 * If we failed to find a running task, but find the context active now
1157 * that we've acquired the ctx->lock, retry.
1159 if (ctx->is_active) {
1160 raw_spin_unlock_irq(&ctx->lock);
1165 * Since the task isn't running, its safe to remove the event, us
1166 * holding the ctx->lock ensures the task won't get scheduled in.
1168 list_del_event(event, ctx);
1169 raw_spin_unlock_irq(&ctx->lock);
1173 * Cross CPU call to disable a performance event
1175 static int __perf_event_disable(void *info)
1177 struct perf_event *event = info;
1178 struct perf_event_context *ctx = event->ctx;
1179 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1182 * If this is a per-task event, need to check whether this
1183 * event's task is the current task on this cpu.
1185 * Can trigger due to concurrent perf_event_context_sched_out()
1186 * flipping contexts around.
1188 if (ctx->task && cpuctx->task_ctx != ctx)
1191 raw_spin_lock(&ctx->lock);
1194 * If the event is on, turn it off.
1195 * If it is in error state, leave it in error state.
1197 if (event->state >= PERF_EVENT_STATE_INACTIVE) {
1198 update_context_time(ctx);
1199 update_cgrp_time_from_event(event);
1200 update_group_times(event);
1201 if (event == event->group_leader)
1202 group_sched_out(event, cpuctx, ctx);
1204 event_sched_out(event, cpuctx, ctx);
1205 event->state = PERF_EVENT_STATE_OFF;
1208 raw_spin_unlock(&ctx->lock);
1216 * If event->ctx is a cloned context, callers must make sure that
1217 * every task struct that event->ctx->task could possibly point to
1218 * remains valid. This condition is satisifed when called through
1219 * perf_event_for_each_child or perf_event_for_each because they
1220 * hold the top-level event's child_mutex, so any descendant that
1221 * goes to exit will block in sync_child_event.
1222 * When called from perf_pending_event it's OK because event->ctx
1223 * is the current context on this CPU and preemption is disabled,
1224 * hence we can't get into perf_event_task_sched_out for this context.
1226 void perf_event_disable(struct perf_event *event)
1228 struct perf_event_context *ctx = event->ctx;
1229 struct task_struct *task = ctx->task;
1233 * Disable the event on the cpu that it's on
1235 cpu_function_call(event->cpu, __perf_event_disable, event);
1240 if (!task_function_call(task, __perf_event_disable, event))
1243 raw_spin_lock_irq(&ctx->lock);
1245 * If the event is still active, we need to retry the cross-call.
1247 if (event->state == PERF_EVENT_STATE_ACTIVE) {
1248 raw_spin_unlock_irq(&ctx->lock);
1250 * Reload the task pointer, it might have been changed by
1251 * a concurrent perf_event_context_sched_out().
1258 * Since we have the lock this context can't be scheduled
1259 * in, so we can change the state safely.
1261 if (event->state == PERF_EVENT_STATE_INACTIVE) {
1262 update_group_times(event);
1263 event->state = PERF_EVENT_STATE_OFF;
1265 raw_spin_unlock_irq(&ctx->lock);
1268 static void perf_set_shadow_time(struct perf_event *event,
1269 struct perf_event_context *ctx,
1273 * use the correct time source for the time snapshot
1275 * We could get by without this by leveraging the
1276 * fact that to get to this function, the caller
1277 * has most likely already called update_context_time()
1278 * and update_cgrp_time_xx() and thus both timestamp
1279 * are identical (or very close). Given that tstamp is,
1280 * already adjusted for cgroup, we could say that:
1281 * tstamp - ctx->timestamp
1283 * tstamp - cgrp->timestamp.
1285 * Then, in perf_output_read(), the calculation would
1286 * work with no changes because:
1287 * - event is guaranteed scheduled in
1288 * - no scheduled out in between
1289 * - thus the timestamp would be the same
1291 * But this is a bit hairy.
1293 * So instead, we have an explicit cgroup call to remain
1294 * within the time time source all along. We believe it
1295 * is cleaner and simpler to understand.
1297 if (is_cgroup_event(event))
1298 perf_cgroup_set_shadow_time(event, tstamp);
1300 event->shadow_ctx_time = tstamp - ctx->timestamp;
1303 #define MAX_INTERRUPTS (~0ULL)
1305 static void perf_log_throttle(struct perf_event *event, int enable);
1308 event_sched_in(struct perf_event *event,
1309 struct perf_cpu_context *cpuctx,
1310 struct perf_event_context *ctx)
1312 u64 tstamp = perf_event_time(event);
1314 if (event->state <= PERF_EVENT_STATE_OFF)
1317 event->state = PERF_EVENT_STATE_ACTIVE;
1318 event->oncpu = smp_processor_id();
1321 * Unthrottle events, since we scheduled we might have missed several
1322 * ticks already, also for a heavily scheduling task there is little
1323 * guarantee it'll get a tick in a timely manner.
1325 if (unlikely(event->hw.interrupts == MAX_INTERRUPTS)) {
1326 perf_log_throttle(event, 1);
1327 event->hw.interrupts = 0;
1331 * The new state must be visible before we turn it on in the hardware:
1335 if (event->pmu->add(event, PERF_EF_START)) {
1336 event->state = PERF_EVENT_STATE_INACTIVE;
1341 event->tstamp_running += tstamp - event->tstamp_stopped;
1343 perf_set_shadow_time(event, ctx, tstamp);
1345 if (!is_software_event(event))
1346 cpuctx->active_oncpu++;
1349 if (event->attr.exclusive)
1350 cpuctx->exclusive = 1;
1356 group_sched_in(struct perf_event *group_event,
1357 struct perf_cpu_context *cpuctx,
1358 struct perf_event_context *ctx)
1360 struct perf_event *event, *partial_group = NULL;
1361 struct pmu *pmu = group_event->pmu;
1362 u64 now = ctx->time;
1363 bool simulate = false;
1365 if (group_event->state == PERF_EVENT_STATE_OFF)
1368 pmu->start_txn(pmu);
1370 if (event_sched_in(group_event, cpuctx, ctx)) {
1371 pmu->cancel_txn(pmu);
1376 * Schedule in siblings as one group (if any):
1378 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
1379 if (event_sched_in(event, cpuctx, ctx)) {
1380 partial_group = event;
1385 if (!pmu->commit_txn(pmu))
1390 * Groups can be scheduled in as one unit only, so undo any
1391 * partial group before returning:
1392 * The events up to the failed event are scheduled out normally,
1393 * tstamp_stopped will be updated.
1395 * The failed events and the remaining siblings need to have
1396 * their timings updated as if they had gone thru event_sched_in()
1397 * and event_sched_out(). This is required to get consistent timings
1398 * across the group. This also takes care of the case where the group
1399 * could never be scheduled by ensuring tstamp_stopped is set to mark
1400 * the time the event was actually stopped, such that time delta
1401 * calculation in update_event_times() is correct.
1403 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
1404 if (event == partial_group)
1408 event->tstamp_running += now - event->tstamp_stopped;
1409 event->tstamp_stopped = now;
1411 event_sched_out(event, cpuctx, ctx);
1414 event_sched_out(group_event, cpuctx, ctx);
1416 pmu->cancel_txn(pmu);
1422 * Work out whether we can put this event group on the CPU now.
1424 static int group_can_go_on(struct perf_event *event,
1425 struct perf_cpu_context *cpuctx,
1429 * Groups consisting entirely of software events can always go on.
1431 if (event->group_flags & PERF_GROUP_SOFTWARE)
1434 * If an exclusive group is already on, no other hardware
1437 if (cpuctx->exclusive)
1440 * If this group is exclusive and there are already
1441 * events on the CPU, it can't go on.
1443 if (event->attr.exclusive && cpuctx->active_oncpu)
1446 * Otherwise, try to add it if all previous groups were able
1452 static void add_event_to_ctx(struct perf_event *event,
1453 struct perf_event_context *ctx)
1455 u64 tstamp = perf_event_time(event);
1457 list_add_event(event, ctx);
1458 perf_group_attach(event);
1459 event->tstamp_enabled = tstamp;
1460 event->tstamp_running = tstamp;
1461 event->tstamp_stopped = tstamp;
1464 static void perf_event_context_sched_in(struct perf_event_context *ctx,
1465 struct task_struct *tsk);
1468 * Cross CPU call to install and enable a performance event
1470 * Must be called with ctx->mutex held
1472 static int __perf_install_in_context(void *info)
1474 struct perf_event *event = info;
1475 struct perf_event_context *ctx = event->ctx;
1476 struct perf_event *leader = event->group_leader;
1477 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1481 * In case we're installing a new context to an already running task,
1482 * could also happen before perf_event_task_sched_in() on architectures
1483 * which do context switches with IRQs enabled.
1485 if (ctx->task && !cpuctx->task_ctx)
1486 perf_event_context_sched_in(ctx, ctx->task);
1488 raw_spin_lock(&ctx->lock);
1490 update_context_time(ctx);
1492 * update cgrp time only if current cgrp
1493 * matches event->cgrp. Must be done before
1494 * calling add_event_to_ctx()
1496 update_cgrp_time_from_event(event);
1498 add_event_to_ctx(event, ctx);
1500 if (!event_filter_match(event))
1504 * Don't put the event on if it is disabled or if
1505 * it is in a group and the group isn't on.
1507 if (event->state != PERF_EVENT_STATE_INACTIVE ||
1508 (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE))
1512 * An exclusive event can't go on if there are already active
1513 * hardware events, and no hardware event can go on if there
1514 * is already an exclusive event on.
1516 if (!group_can_go_on(event, cpuctx, 1))
1519 err = event_sched_in(event, cpuctx, ctx);
1523 * This event couldn't go on. If it is in a group
1524 * then we have to pull the whole group off.
1525 * If the event group is pinned then put it in error state.
1527 if (leader != event)
1528 group_sched_out(leader, cpuctx, ctx);
1529 if (leader->attr.pinned) {
1530 update_group_times(leader);
1531 leader->state = PERF_EVENT_STATE_ERROR;
1536 raw_spin_unlock(&ctx->lock);
1542 * Attach a performance event to a context
1544 * First we add the event to the list with the hardware enable bit
1545 * in event->hw_config cleared.
1547 * If the event is attached to a task which is on a CPU we use a smp
1548 * call to enable it in the task context. The task might have been
1549 * scheduled away, but we check this in the smp call again.
1552 perf_install_in_context(struct perf_event_context *ctx,
1553 struct perf_event *event,
1556 struct task_struct *task = ctx->task;
1558 lockdep_assert_held(&ctx->mutex);
1564 * Per cpu events are installed via an smp call and
1565 * the install is always successful.
1567 cpu_function_call(cpu, __perf_install_in_context, event);
1572 if (!task_function_call(task, __perf_install_in_context, event))
1575 raw_spin_lock_irq(&ctx->lock);
1577 * If we failed to find a running task, but find the context active now
1578 * that we've acquired the ctx->lock, retry.
1580 if (ctx->is_active) {
1581 raw_spin_unlock_irq(&ctx->lock);
1586 * Since the task isn't running, its safe to add the event, us holding
1587 * the ctx->lock ensures the task won't get scheduled in.
1589 add_event_to_ctx(event, ctx);
1590 raw_spin_unlock_irq(&ctx->lock);
1594 * Put a event into inactive state and update time fields.
1595 * Enabling the leader of a group effectively enables all
1596 * the group members that aren't explicitly disabled, so we
1597 * have to update their ->tstamp_enabled also.
1598 * Note: this works for group members as well as group leaders
1599 * since the non-leader members' sibling_lists will be empty.
1601 static void __perf_event_mark_enabled(struct perf_event *event,
1602 struct perf_event_context *ctx)
1604 struct perf_event *sub;
1605 u64 tstamp = perf_event_time(event);
1607 event->state = PERF_EVENT_STATE_INACTIVE;
1608 event->tstamp_enabled = tstamp - event->total_time_enabled;
1609 list_for_each_entry(sub, &event->sibling_list, group_entry) {
1610 if (sub->state >= PERF_EVENT_STATE_INACTIVE)
1611 sub->tstamp_enabled = tstamp - sub->total_time_enabled;
1616 * Cross CPU call to enable a performance event
1618 static int __perf_event_enable(void *info)
1620 struct perf_event *event = info;
1621 struct perf_event_context *ctx = event->ctx;
1622 struct perf_event *leader = event->group_leader;
1623 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1626 if (WARN_ON_ONCE(!ctx->is_active))
1629 raw_spin_lock(&ctx->lock);
1630 update_context_time(ctx);
1632 if (event->state >= PERF_EVENT_STATE_INACTIVE)
1636 * set current task's cgroup time reference point
1638 perf_cgroup_set_timestamp(current, ctx);
1640 __perf_event_mark_enabled(event, ctx);
1642 if (!event_filter_match(event)) {
1643 if (is_cgroup_event(event))
1644 perf_cgroup_defer_enabled(event);
1649 * If the event is in a group and isn't the group leader,
1650 * then don't put it on unless the group is on.
1652 if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE)
1655 if (!group_can_go_on(event, cpuctx, 1)) {
1658 if (event == leader)
1659 err = group_sched_in(event, cpuctx, ctx);
1661 err = event_sched_in(event, cpuctx, ctx);
1666 * If this event can't go on and it's part of a
1667 * group, then the whole group has to come off.
1669 if (leader != event)
1670 group_sched_out(leader, cpuctx, ctx);
1671 if (leader->attr.pinned) {
1672 update_group_times(leader);
1673 leader->state = PERF_EVENT_STATE_ERROR;
1678 raw_spin_unlock(&ctx->lock);
1686 * If event->ctx is a cloned context, callers must make sure that
1687 * every task struct that event->ctx->task could possibly point to
1688 * remains valid. This condition is satisfied when called through
1689 * perf_event_for_each_child or perf_event_for_each as described
1690 * for perf_event_disable.
1692 void perf_event_enable(struct perf_event *event)
1694 struct perf_event_context *ctx = event->ctx;
1695 struct task_struct *task = ctx->task;
1699 * Enable the event on the cpu that it's on
1701 cpu_function_call(event->cpu, __perf_event_enable, event);
1705 raw_spin_lock_irq(&ctx->lock);
1706 if (event->state >= PERF_EVENT_STATE_INACTIVE)
1710 * If the event is in error state, clear that first.
1711 * That way, if we see the event in error state below, we
1712 * know that it has gone back into error state, as distinct
1713 * from the task having been scheduled away before the
1714 * cross-call arrived.
1716 if (event->state == PERF_EVENT_STATE_ERROR)
1717 event->state = PERF_EVENT_STATE_OFF;
1720 if (!ctx->is_active) {
1721 __perf_event_mark_enabled(event, ctx);
1725 raw_spin_unlock_irq(&ctx->lock);
1727 if (!task_function_call(task, __perf_event_enable, event))
1730 raw_spin_lock_irq(&ctx->lock);
1733 * If the context is active and the event is still off,
1734 * we need to retry the cross-call.
1736 if (ctx->is_active && event->state == PERF_EVENT_STATE_OFF) {
1738 * task could have been flipped by a concurrent
1739 * perf_event_context_sched_out()
1746 raw_spin_unlock_irq(&ctx->lock);
1749 static int perf_event_refresh(struct perf_event *event, int refresh)
1752 * not supported on inherited events
1754 if (event->attr.inherit || !is_sampling_event(event))
1757 atomic_add(refresh, &event->event_limit);
1758 perf_event_enable(event);
1763 static void ctx_sched_out(struct perf_event_context *ctx,
1764 struct perf_cpu_context *cpuctx,
1765 enum event_type_t event_type)
1767 struct perf_event *event;
1769 raw_spin_lock(&ctx->lock);
1770 perf_pmu_disable(ctx->pmu);
1772 if (likely(!ctx->nr_events))
1774 update_context_time(ctx);
1775 update_cgrp_time_from_cpuctx(cpuctx);
1777 if (!ctx->nr_active)
1780 if (event_type & EVENT_PINNED) {
1781 list_for_each_entry(event, &ctx->pinned_groups, group_entry)
1782 group_sched_out(event, cpuctx, ctx);
1785 if (event_type & EVENT_FLEXIBLE) {
1786 list_for_each_entry(event, &ctx->flexible_groups, group_entry)
1787 group_sched_out(event, cpuctx, ctx);
1790 perf_pmu_enable(ctx->pmu);
1791 raw_spin_unlock(&ctx->lock);
1795 * Test whether two contexts are equivalent, i.e. whether they
1796 * have both been cloned from the same version of the same context
1797 * and they both have the same number of enabled events.
1798 * If the number of enabled events is the same, then the set
1799 * of enabled events should be the same, because these are both
1800 * inherited contexts, therefore we can't access individual events
1801 * in them directly with an fd; we can only enable/disable all
1802 * events via prctl, or enable/disable all events in a family
1803 * via ioctl, which will have the same effect on both contexts.
1805 static int context_equiv(struct perf_event_context *ctx1,
1806 struct perf_event_context *ctx2)
1808 return ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx
1809 && ctx1->parent_gen == ctx2->parent_gen
1810 && !ctx1->pin_count && !ctx2->pin_count;
1813 static void __perf_event_sync_stat(struct perf_event *event,
1814 struct perf_event *next_event)
1818 if (!event->attr.inherit_stat)
1822 * Update the event value, we cannot use perf_event_read()
1823 * because we're in the middle of a context switch and have IRQs
1824 * disabled, which upsets smp_call_function_single(), however
1825 * we know the event must be on the current CPU, therefore we
1826 * don't need to use it.
1828 switch (event->state) {
1829 case PERF_EVENT_STATE_ACTIVE:
1830 event->pmu->read(event);
1833 case PERF_EVENT_STATE_INACTIVE:
1834 update_event_times(event);
1842 * In order to keep per-task stats reliable we need to flip the event
1843 * values when we flip the contexts.
1845 value = local64_read(&next_event->count);
1846 value = local64_xchg(&event->count, value);
1847 local64_set(&next_event->count, value);
1849 swap(event->total_time_enabled, next_event->total_time_enabled);
1850 swap(event->total_time_running, next_event->total_time_running);
1853 * Since we swizzled the values, update the user visible data too.
1855 perf_event_update_userpage(event);
1856 perf_event_update_userpage(next_event);
1859 #define list_next_entry(pos, member) \
1860 list_entry(pos->member.next, typeof(*pos), member)
1862 static void perf_event_sync_stat(struct perf_event_context *ctx,
1863 struct perf_event_context *next_ctx)
1865 struct perf_event *event, *next_event;
1870 update_context_time(ctx);
1872 event = list_first_entry(&ctx->event_list,
1873 struct perf_event, event_entry);
1875 next_event = list_first_entry(&next_ctx->event_list,
1876 struct perf_event, event_entry);
1878 while (&event->event_entry != &ctx->event_list &&
1879 &next_event->event_entry != &next_ctx->event_list) {
1881 __perf_event_sync_stat(event, next_event);
1883 event = list_next_entry(event, event_entry);
1884 next_event = list_next_entry(next_event, event_entry);
1888 static void perf_event_context_sched_out(struct task_struct *task, int ctxn,
1889 struct task_struct *next)
1891 struct perf_event_context *ctx = task->perf_event_ctxp[ctxn];
1892 struct perf_event_context *next_ctx;
1893 struct perf_event_context *parent;
1894 struct perf_cpu_context *cpuctx;
1900 cpuctx = __get_cpu_context(ctx);
1901 if (!cpuctx->task_ctx)
1905 parent = rcu_dereference(ctx->parent_ctx);
1906 next_ctx = next->perf_event_ctxp[ctxn];
1907 if (parent && next_ctx &&
1908 rcu_dereference(next_ctx->parent_ctx) == parent) {
1910 * Looks like the two contexts are clones, so we might be
1911 * able to optimize the context switch. We lock both
1912 * contexts and check that they are clones under the
1913 * lock (including re-checking that neither has been
1914 * uncloned in the meantime). It doesn't matter which
1915 * order we take the locks because no other cpu could
1916 * be trying to lock both of these tasks.
1918 raw_spin_lock(&ctx->lock);
1919 raw_spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING);
1920 if (context_equiv(ctx, next_ctx)) {
1922 * XXX do we need a memory barrier of sorts
1923 * wrt to rcu_dereference() of perf_event_ctxp
1925 task->perf_event_ctxp[ctxn] = next_ctx;
1926 next->perf_event_ctxp[ctxn] = ctx;
1928 next_ctx->task = task;
1931 perf_event_sync_stat(ctx, next_ctx);
1933 raw_spin_unlock(&next_ctx->lock);
1934 raw_spin_unlock(&ctx->lock);
1939 ctx_sched_out(ctx, cpuctx, EVENT_ALL);
1940 cpuctx->task_ctx = NULL;
1944 #define for_each_task_context_nr(ctxn) \
1945 for ((ctxn) = 0; (ctxn) < perf_nr_task_contexts; (ctxn)++)
1948 * Called from scheduler to remove the events of the current task,
1949 * with interrupts disabled.
1951 * We stop each event and update the event value in event->count.
1953 * This does not protect us against NMI, but disable()
1954 * sets the disabled bit in the control field of event _before_
1955 * accessing the event control register. If a NMI hits, then it will
1956 * not restart the event.
1958 void __perf_event_task_sched_out(struct task_struct *task,
1959 struct task_struct *next)
1963 for_each_task_context_nr(ctxn)
1964 perf_event_context_sched_out(task, ctxn, next);
1967 * if cgroup events exist on this CPU, then we need
1968 * to check if we have to switch out PMU state.
1969 * cgroup event are system-wide mode only
1971 if (atomic_read(&__get_cpu_var(perf_cgroup_events)))
1972 perf_cgroup_sched_out(task);
1975 static void task_ctx_sched_out(struct perf_event_context *ctx,
1976 enum event_type_t event_type)
1978 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1980 if (!cpuctx->task_ctx)
1983 if (WARN_ON_ONCE(ctx != cpuctx->task_ctx))
1986 ctx_sched_out(ctx, cpuctx, event_type);
1987 cpuctx->task_ctx = NULL;
1991 * Called with IRQs disabled
1993 static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
1994 enum event_type_t event_type)
1996 ctx_sched_out(&cpuctx->ctx, cpuctx, event_type);
2000 ctx_pinned_sched_in(struct perf_event_context *ctx,
2001 struct perf_cpu_context *cpuctx)
2003 struct perf_event *event;
2005 list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
2006 if (event->state <= PERF_EVENT_STATE_OFF)
2008 if (!event_filter_match(event))
2011 /* may need to reset tstamp_enabled */
2012 if (is_cgroup_event(event))
2013 perf_cgroup_mark_enabled(event, ctx);
2015 if (group_can_go_on(event, cpuctx, 1))
2016 group_sched_in(event, cpuctx, ctx);
2019 * If this pinned group hasn't been scheduled,
2020 * put it in error state.
2022 if (event->state == PERF_EVENT_STATE_INACTIVE) {
2023 update_group_times(event);
2024 event->state = PERF_EVENT_STATE_ERROR;
2030 ctx_flexible_sched_in(struct perf_event_context *ctx,
2031 struct perf_cpu_context *cpuctx)
2033 struct perf_event *event;
2036 list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
2037 /* Ignore events in OFF or ERROR state */
2038 if (event->state <= PERF_EVENT_STATE_OFF)
2041 * Listen to the 'cpu' scheduling filter constraint
2044 if (!event_filter_match(event))
2047 /* may need to reset tstamp_enabled */
2048 if (is_cgroup_event(event))
2049 perf_cgroup_mark_enabled(event, ctx);
2051 if (group_can_go_on(event, cpuctx, can_add_hw)) {
2052 if (group_sched_in(event, cpuctx, ctx))
2059 ctx_sched_in(struct perf_event_context *ctx,
2060 struct perf_cpu_context *cpuctx,
2061 enum event_type_t event_type,
2062 struct task_struct *task)
2066 raw_spin_lock(&ctx->lock);
2068 if (likely(!ctx->nr_events))
2072 ctx->timestamp = now;
2073 perf_cgroup_set_timestamp(task, ctx);
2075 * First go through the list and put on any pinned groups
2076 * in order to give them the best chance of going on.
2078 if (event_type & EVENT_PINNED)
2079 ctx_pinned_sched_in(ctx, cpuctx);
2081 /* Then walk through the lower prio flexible groups */
2082 if (event_type & EVENT_FLEXIBLE)
2083 ctx_flexible_sched_in(ctx, cpuctx);
2086 raw_spin_unlock(&ctx->lock);
2089 static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
2090 enum event_type_t event_type,
2091 struct task_struct *task)
2093 struct perf_event_context *ctx = &cpuctx->ctx;
2095 ctx_sched_in(ctx, cpuctx, event_type, task);
2098 static void task_ctx_sched_in(struct perf_event_context *ctx,
2099 enum event_type_t event_type)
2101 struct perf_cpu_context *cpuctx;
2103 cpuctx = __get_cpu_context(ctx);
2104 if (cpuctx->task_ctx == ctx)
2107 ctx_sched_in(ctx, cpuctx, event_type, NULL);
2108 cpuctx->task_ctx = ctx;
2111 static void perf_event_context_sched_in(struct perf_event_context *ctx,
2112 struct task_struct *task)
2114 struct perf_cpu_context *cpuctx;
2116 cpuctx = __get_cpu_context(ctx);
2117 if (cpuctx->task_ctx == ctx)
2120 perf_pmu_disable(ctx->pmu);
2122 * We want to keep the following priority order:
2123 * cpu pinned (that don't need to move), task pinned,
2124 * cpu flexible, task flexible.
2126 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
2128 ctx_sched_in(ctx, cpuctx, EVENT_PINNED, task);
2129 cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE, task);
2130 ctx_sched_in(ctx, cpuctx, EVENT_FLEXIBLE, task);
2132 cpuctx->task_ctx = ctx;
2135 * Since these rotations are per-cpu, we need to ensure the
2136 * cpu-context we got scheduled on is actually rotating.
2138 perf_pmu_rotate_start(ctx->pmu);
2139 perf_pmu_enable(ctx->pmu);
2143 * Called from scheduler to add the events of the current task
2144 * with interrupts disabled.
2146 * We restore the event value and then enable it.
2148 * This does not protect us against NMI, but enable()
2149 * sets the enabled bit in the control field of event _before_
2150 * accessing the event control register. If a NMI hits, then it will
2151 * keep the event running.
2153 void __perf_event_task_sched_in(struct task_struct *task)
2155 struct perf_event_context *ctx;
2158 for_each_task_context_nr(ctxn) {
2159 ctx = task->perf_event_ctxp[ctxn];
2163 perf_event_context_sched_in(ctx, task);
2166 * if cgroup events exist on this CPU, then we need
2167 * to check if we have to switch in PMU state.
2168 * cgroup event are system-wide mode only
2170 if (atomic_read(&__get_cpu_var(perf_cgroup_events)))
2171 perf_cgroup_sched_in(task);
2174 static u64 perf_calculate_period(struct perf_event *event, u64 nsec, u64 count)
2176 u64 frequency = event->attr.sample_freq;
2177 u64 sec = NSEC_PER_SEC;
2178 u64 divisor, dividend;
2180 int count_fls, nsec_fls, frequency_fls, sec_fls;
2182 count_fls = fls64(count);
2183 nsec_fls = fls64(nsec);
2184 frequency_fls = fls64(frequency);
2188 * We got @count in @nsec, with a target of sample_freq HZ
2189 * the target period becomes:
2192 * period = -------------------
2193 * @nsec * sample_freq
2198 * Reduce accuracy by one bit such that @a and @b converge
2199 * to a similar magnitude.
2201 #define REDUCE_FLS(a, b) \
2203 if (a##_fls > b##_fls) { \
2213 * Reduce accuracy until either term fits in a u64, then proceed with
2214 * the other, so that finally we can do a u64/u64 division.
2216 while (count_fls + sec_fls > 64 && nsec_fls + frequency_fls > 64) {
2217 REDUCE_FLS(nsec, frequency);
2218 REDUCE_FLS(sec, count);
2221 if (count_fls + sec_fls > 64) {
2222 divisor = nsec * frequency;
2224 while (count_fls + sec_fls > 64) {
2225 REDUCE_FLS(count, sec);
2229 dividend = count * sec;
2231 dividend = count * sec;
2233 while (nsec_fls + frequency_fls > 64) {
2234 REDUCE_FLS(nsec, frequency);
2238 divisor = nsec * frequency;
2244 return div64_u64(dividend, divisor);
2247 static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count)
2249 struct hw_perf_event *hwc = &event->hw;
2250 s64 period, sample_period;
2253 period = perf_calculate_period(event, nsec, count);
2255 delta = (s64)(period - hwc->sample_period);
2256 delta = (delta + 7) / 8; /* low pass filter */
2258 sample_period = hwc->sample_period + delta;
2263 hwc->sample_period = sample_period;
2265 if (local64_read(&hwc->period_left) > 8*sample_period) {
2266 event->pmu->stop(event, PERF_EF_UPDATE);
2267 local64_set(&hwc->period_left, 0);
2268 event->pmu->start(event, PERF_EF_RELOAD);
2272 static void perf_ctx_adjust_freq(struct perf_event_context *ctx, u64 period)
2274 struct perf_event *event;
2275 struct hw_perf_event *hwc;
2276 u64 interrupts, now;
2279 raw_spin_lock(&ctx->lock);
2280 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
2281 if (event->state != PERF_EVENT_STATE_ACTIVE)
2284 if (!event_filter_match(event))
2289 interrupts = hwc->interrupts;
2290 hwc->interrupts = 0;
2293 * unthrottle events on the tick
2295 if (interrupts == MAX_INTERRUPTS) {
2296 perf_log_throttle(event, 1);
2297 event->pmu->start(event, 0);
2300 if (!event->attr.freq || !event->attr.sample_freq)
2303 event->pmu->read(event);
2304 now = local64_read(&event->count);
2305 delta = now - hwc->freq_count_stamp;
2306 hwc->freq_count_stamp = now;
2309 perf_adjust_period(event, period, delta);
2311 raw_spin_unlock(&ctx->lock);
2315 * Round-robin a context's events:
2317 static void rotate_ctx(struct perf_event_context *ctx)
2319 raw_spin_lock(&ctx->lock);
2322 * Rotate the first entry last of non-pinned groups. Rotation might be
2323 * disabled by the inheritance code.
2325 if (!ctx->rotate_disable)
2326 list_rotate_left(&ctx->flexible_groups);
2328 raw_spin_unlock(&ctx->lock);
2332 * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
2333 * because they're strictly cpu affine and rotate_start is called with IRQs
2334 * disabled, while rotate_context is called from IRQ context.
2336 static void perf_rotate_context(struct perf_cpu_context *cpuctx)
2338 u64 interval = (u64)cpuctx->jiffies_interval * TICK_NSEC;
2339 struct perf_event_context *ctx = NULL;
2340 int rotate = 0, remove = 1;
2342 if (cpuctx->ctx.nr_events) {
2344 if (cpuctx->ctx.nr_events != cpuctx->ctx.nr_active)
2348 ctx = cpuctx->task_ctx;
2349 if (ctx && ctx->nr_events) {
2351 if (ctx->nr_events != ctx->nr_active)
2355 perf_pmu_disable(cpuctx->ctx.pmu);
2356 perf_ctx_adjust_freq(&cpuctx->ctx, interval);
2358 perf_ctx_adjust_freq(ctx, interval);
2363 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
2365 task_ctx_sched_out(ctx, EVENT_FLEXIBLE);
2367 rotate_ctx(&cpuctx->ctx);
2371 cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE, current);
2373 task_ctx_sched_in(ctx, EVENT_FLEXIBLE);
2377 list_del_init(&cpuctx->rotation_list);
2379 perf_pmu_enable(cpuctx->ctx.pmu);
2382 void perf_event_task_tick(void)
2384 struct list_head *head = &__get_cpu_var(rotation_list);
2385 struct perf_cpu_context *cpuctx, *tmp;
2387 WARN_ON(!irqs_disabled());
2389 list_for_each_entry_safe(cpuctx, tmp, head, rotation_list) {
2390 if (cpuctx->jiffies_interval == 1 ||
2391 !(jiffies % cpuctx->jiffies_interval))
2392 perf_rotate_context(cpuctx);
2396 static int event_enable_on_exec(struct perf_event *event,
2397 struct perf_event_context *ctx)
2399 if (!event->attr.enable_on_exec)
2402 event->attr.enable_on_exec = 0;
2403 if (event->state >= PERF_EVENT_STATE_INACTIVE)
2406 __perf_event_mark_enabled(event, ctx);
2412 * Enable all of a task's events that have been marked enable-on-exec.
2413 * This expects task == current.
2415 static void perf_event_enable_on_exec(struct perf_event_context *ctx)
2417 struct perf_event *event;
2418 unsigned long flags;
2422 local_irq_save(flags);
2423 if (!ctx || !ctx->nr_events)
2426 task_ctx_sched_out(ctx, EVENT_ALL);
2428 raw_spin_lock(&ctx->lock);
2430 list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
2431 ret = event_enable_on_exec(event, ctx);
2436 list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
2437 ret = event_enable_on_exec(event, ctx);
2443 * Unclone this context if we enabled any event.
2448 raw_spin_unlock(&ctx->lock);
2450 perf_event_context_sched_in(ctx, ctx->task);
2452 local_irq_restore(flags);
2456 * Cross CPU call to read the hardware event
2458 static void __perf_event_read(void *info)
2460 struct perf_event *event = info;
2461 struct perf_event_context *ctx = event->ctx;
2462 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
2465 * If this is a task context, we need to check whether it is
2466 * the current task context of this cpu. If not it has been
2467 * scheduled out before the smp call arrived. In that case
2468 * event->count would have been updated to a recent sample
2469 * when the event was scheduled out.
2471 if (ctx->task && cpuctx->task_ctx != ctx)
2474 raw_spin_lock(&ctx->lock);
2475 if (ctx->is_active) {
2476 update_context_time(ctx);
2477 update_cgrp_time_from_event(event);
2479 update_event_times(event);
2480 if (event->state == PERF_EVENT_STATE_ACTIVE)
2481 event->pmu->read(event);
2482 raw_spin_unlock(&ctx->lock);
2485 static inline u64 perf_event_count(struct perf_event *event)
2487 return local64_read(&event->count) + atomic64_read(&event->child_count);
2490 static u64 perf_event_read(struct perf_event *event)
2493 * If event is enabled and currently active on a CPU, update the
2494 * value in the event structure:
2496 if (event->state == PERF_EVENT_STATE_ACTIVE) {
2497 smp_call_function_single(event->oncpu,
2498 __perf_event_read, event, 1);
2499 } else if (event->state == PERF_EVENT_STATE_INACTIVE) {
2500 struct perf_event_context *ctx = event->ctx;
2501 unsigned long flags;
2503 raw_spin_lock_irqsave(&ctx->lock, flags);
2505 * may read while context is not active
2506 * (e.g., thread is blocked), in that case
2507 * we cannot update context time
2509 if (ctx->is_active) {
2510 update_context_time(ctx);
2511 update_cgrp_time_from_event(event);
2513 update_event_times(event);
2514 raw_spin_unlock_irqrestore(&ctx->lock, flags);
2517 return perf_event_count(event);
2524 struct callchain_cpus_entries {
2525 struct rcu_head rcu_head;
2526 struct perf_callchain_entry *cpu_entries[0];
2529 static DEFINE_PER_CPU(int, callchain_recursion[PERF_NR_CONTEXTS]);
2530 static atomic_t nr_callchain_events;
2531 static DEFINE_MUTEX(callchain_mutex);
2532 struct callchain_cpus_entries *callchain_cpus_entries;
2535 __weak void perf_callchain_kernel(struct perf_callchain_entry *entry,
2536 struct pt_regs *regs)
2540 __weak void perf_callchain_user(struct perf_callchain_entry *entry,
2541 struct pt_regs *regs)
2545 static void release_callchain_buffers_rcu(struct rcu_head *head)
2547 struct callchain_cpus_entries *entries;
2550 entries = container_of(head, struct callchain_cpus_entries, rcu_head);
2552 for_each_possible_cpu(cpu)
2553 kfree(entries->cpu_entries[cpu]);
2558 static void release_callchain_buffers(void)
2560 struct callchain_cpus_entries *entries;
2562 entries = callchain_cpus_entries;
2563 rcu_assign_pointer(callchain_cpus_entries, NULL);
2564 call_rcu(&entries->rcu_head, release_callchain_buffers_rcu);
2567 static int alloc_callchain_buffers(void)
2571 struct callchain_cpus_entries *entries;
2574 * We can't use the percpu allocation API for data that can be
2575 * accessed from NMI. Use a temporary manual per cpu allocation
2576 * until that gets sorted out.
2578 size = offsetof(struct callchain_cpus_entries, cpu_entries[nr_cpu_ids]);
2580 entries = kzalloc(size, GFP_KERNEL);
2584 size = sizeof(struct perf_callchain_entry) * PERF_NR_CONTEXTS;
2586 for_each_possible_cpu(cpu) {
2587 entries->cpu_entries[cpu] = kmalloc_node(size, GFP_KERNEL,
2589 if (!entries->cpu_entries[cpu])
2593 rcu_assign_pointer(callchain_cpus_entries, entries);
2598 for_each_possible_cpu(cpu)
2599 kfree(entries->cpu_entries[cpu]);
2605 static int get_callchain_buffers(void)
2610 mutex_lock(&callchain_mutex);
2612 count = atomic_inc_return(&nr_callchain_events);
2613 if (WARN_ON_ONCE(count < 1)) {
2619 /* If the allocation failed, give up */
2620 if (!callchain_cpus_entries)
2625 err = alloc_callchain_buffers();
2627 release_callchain_buffers();
2629 mutex_unlock(&callchain_mutex);
2634 static void put_callchain_buffers(void)
2636 if (atomic_dec_and_mutex_lock(&nr_callchain_events, &callchain_mutex)) {
2637 release_callchain_buffers();
2638 mutex_unlock(&callchain_mutex);
2642 static int get_recursion_context(int *recursion)
2650 else if (in_softirq())
2655 if (recursion[rctx])
2664 static inline void put_recursion_context(int *recursion, int rctx)
2670 static struct perf_callchain_entry *get_callchain_entry(int *rctx)
2673 struct callchain_cpus_entries *entries;
2675 *rctx = get_recursion_context(__get_cpu_var(callchain_recursion));
2679 entries = rcu_dereference(callchain_cpus_entries);
2683 cpu = smp_processor_id();
2685 return &entries->cpu_entries[cpu][*rctx];
2689 put_callchain_entry(int rctx)
2691 put_recursion_context(__get_cpu_var(callchain_recursion), rctx);
2694 static struct perf_callchain_entry *perf_callchain(struct pt_regs *regs)
2697 struct perf_callchain_entry *entry;
2700 entry = get_callchain_entry(&rctx);
2709 if (!user_mode(regs)) {
2710 perf_callchain_store(entry, PERF_CONTEXT_KERNEL);
2711 perf_callchain_kernel(entry, regs);
2713 regs = task_pt_regs(current);
2719 perf_callchain_store(entry, PERF_CONTEXT_USER);
2720 perf_callchain_user(entry, regs);
2724 put_callchain_entry(rctx);
2730 * Initialize the perf_event context in a task_struct:
2732 static void __perf_event_init_context(struct perf_event_context *ctx)
2734 raw_spin_lock_init(&ctx->lock);
2735 mutex_init(&ctx->mutex);
2736 INIT_LIST_HEAD(&ctx->pinned_groups);
2737 INIT_LIST_HEAD(&ctx->flexible_groups);
2738 INIT_LIST_HEAD(&ctx->event_list);
2739 atomic_set(&ctx->refcount, 1);
2742 static struct perf_event_context *
2743 alloc_perf_context(struct pmu *pmu, struct task_struct *task)
2745 struct perf_event_context *ctx;
2747 ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL);
2751 __perf_event_init_context(ctx);
2754 get_task_struct(task);
2761 static struct task_struct *
2762 find_lively_task_by_vpid(pid_t vpid)
2764 struct task_struct *task;
2771 task = find_task_by_vpid(vpid);
2773 get_task_struct(task);
2777 return ERR_PTR(-ESRCH);
2779 /* Reuse ptrace permission checks for now. */
2781 if (!ptrace_may_access(task, PTRACE_MODE_READ))
2786 put_task_struct(task);
2787 return ERR_PTR(err);
2792 * Returns a matching context with refcount and pincount.
2794 static struct perf_event_context *
2795 find_get_context(struct pmu *pmu, struct task_struct *task, int cpu)
2797 struct perf_event_context *ctx;
2798 struct perf_cpu_context *cpuctx;
2799 unsigned long flags;
2803 /* Must be root to operate on a CPU event: */
2804 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
2805 return ERR_PTR(-EACCES);
2808 * We could be clever and allow to attach a event to an
2809 * offline CPU and activate it when the CPU comes up, but
2812 if (!cpu_online(cpu))
2813 return ERR_PTR(-ENODEV);
2815 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
2824 ctxn = pmu->task_ctx_nr;
2829 ctx = perf_lock_task_context(task, ctxn, &flags);
2833 raw_spin_unlock_irqrestore(&ctx->lock, flags);
2837 ctx = alloc_perf_context(pmu, task);
2845 mutex_lock(&task->perf_event_mutex);
2847 * If it has already passed perf_event_exit_task().
2848 * we must see PF_EXITING, it takes this mutex too.
2850 if (task->flags & PF_EXITING)
2852 else if (task->perf_event_ctxp[ctxn])
2856 rcu_assign_pointer(task->perf_event_ctxp[ctxn], ctx);
2858 mutex_unlock(&task->perf_event_mutex);
2860 if (unlikely(err)) {
2861 put_task_struct(task);
2873 return ERR_PTR(err);
2876 static void perf_event_free_filter(struct perf_event *event);
2878 static void free_event_rcu(struct rcu_head *head)
2880 struct perf_event *event;
2882 event = container_of(head, struct perf_event, rcu_head);
2884 put_pid_ns(event->ns);
2885 perf_event_free_filter(event);
2889 static void perf_buffer_put(struct perf_buffer *buffer);
2891 static void free_event(struct perf_event *event)
2893 irq_work_sync(&event->pending);
2895 if (!event->parent) {
2896 if (event->attach_state & PERF_ATTACH_TASK)
2897 jump_label_dec(&perf_sched_events);
2898 if (event->attr.mmap || event->attr.mmap_data)
2899 atomic_dec(&nr_mmap_events);
2900 if (event->attr.comm)
2901 atomic_dec(&nr_comm_events);
2902 if (event->attr.task)
2903 atomic_dec(&nr_task_events);
2904 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN)
2905 put_callchain_buffers();
2906 if (is_cgroup_event(event)) {
2907 atomic_dec(&per_cpu(perf_cgroup_events, event->cpu));
2908 jump_label_dec(&perf_sched_events);
2912 if (event->buffer) {
2913 perf_buffer_put(event->buffer);
2914 event->buffer = NULL;
2917 if (is_cgroup_event(event))
2918 perf_detach_cgroup(event);
2921 event->destroy(event);
2924 put_ctx(event->ctx);
2926 call_rcu(&event->rcu_head, free_event_rcu);
2929 int perf_event_release_kernel(struct perf_event *event)
2931 struct perf_event_context *ctx = event->ctx;
2934 * Remove from the PMU, can't get re-enabled since we got
2935 * here because the last ref went.
2937 perf_event_disable(event);
2939 WARN_ON_ONCE(ctx->parent_ctx);
2941 * There are two ways this annotation is useful:
2943 * 1) there is a lock recursion from perf_event_exit_task
2944 * see the comment there.
2946 * 2) there is a lock-inversion with mmap_sem through
2947 * perf_event_read_group(), which takes faults while
2948 * holding ctx->mutex, however this is called after
2949 * the last filedesc died, so there is no possibility
2950 * to trigger the AB-BA case.
2952 mutex_lock_nested(&ctx->mutex, SINGLE_DEPTH_NESTING);
2953 raw_spin_lock_irq(&ctx->lock);
2954 perf_group_detach(event);
2955 list_del_event(event, ctx);
2956 raw_spin_unlock_irq(&ctx->lock);
2957 mutex_unlock(&ctx->mutex);
2963 EXPORT_SYMBOL_GPL(perf_event_release_kernel);
2966 * Called when the last reference to the file is gone.
2968 static int perf_release(struct inode *inode, struct file *file)
2970 struct perf_event *event = file->private_data;
2971 struct task_struct *owner;
2973 file->private_data = NULL;
2976 owner = ACCESS_ONCE(event->owner);
2978 * Matches the smp_wmb() in perf_event_exit_task(). If we observe
2979 * !owner it means the list deletion is complete and we can indeed
2980 * free this event, otherwise we need to serialize on
2981 * owner->perf_event_mutex.
2983 smp_read_barrier_depends();
2986 * Since delayed_put_task_struct() also drops the last
2987 * task reference we can safely take a new reference
2988 * while holding the rcu_read_lock().
2990 get_task_struct(owner);
2995 mutex_lock(&owner->perf_event_mutex);
2997 * We have to re-check the event->owner field, if it is cleared
2998 * we raced with perf_event_exit_task(), acquiring the mutex
2999 * ensured they're done, and we can proceed with freeing the
3003 list_del_init(&event->owner_entry);
3004 mutex_unlock(&owner->perf_event_mutex);
3005 put_task_struct(owner);
3008 return perf_event_release_kernel(event);
3011 u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running)
3013 struct perf_event *child;
3019 mutex_lock(&event->child_mutex);
3020 total += perf_event_read(event);
3021 *enabled += event->total_time_enabled +
3022 atomic64_read(&event->child_total_time_enabled);
3023 *running += event->total_time_running +
3024 atomic64_read(&event->child_total_time_running);
3026 list_for_each_entry(child, &event->child_list, child_list) {
3027 total += perf_event_read(child);
3028 *enabled += child->total_time_enabled;
3029 *running += child->total_time_running;
3031 mutex_unlock(&event->child_mutex);
3035 EXPORT_SYMBOL_GPL(perf_event_read_value);
3037 static int perf_event_read_group(struct perf_event *event,
3038 u64 read_format, char __user *buf)
3040 struct perf_event *leader = event->group_leader, *sub;
3041 int n = 0, size = 0, ret = -EFAULT;
3042 struct perf_event_context *ctx = leader->ctx;
3044 u64 count, enabled, running;
3046 mutex_lock(&ctx->mutex);
3047 count = perf_event_read_value(leader, &enabled, &running);
3049 values[n++] = 1 + leader->nr_siblings;
3050 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3051 values[n++] = enabled;
3052 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3053 values[n++] = running;
3054 values[n++] = count;
3055 if (read_format & PERF_FORMAT_ID)
3056 values[n++] = primary_event_id(leader);
3058 size = n * sizeof(u64);
3060 if (copy_to_user(buf, values, size))
3065 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
3068 values[n++] = perf_event_read_value(sub, &enabled, &running);
3069 if (read_format & PERF_FORMAT_ID)
3070 values[n++] = primary_event_id(sub);
3072 size = n * sizeof(u64);
3074 if (copy_to_user(buf + ret, values, size)) {
3082 mutex_unlock(&ctx->mutex);
3087 static int perf_event_read_one(struct perf_event *event,
3088 u64 read_format, char __user *buf)
3090 u64 enabled, running;
3094 values[n++] = perf_event_read_value(event, &enabled, &running);
3095 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3096 values[n++] = enabled;
3097 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3098 values[n++] = running;
3099 if (read_format & PERF_FORMAT_ID)
3100 values[n++] = primary_event_id(event);
3102 if (copy_to_user(buf, values, n * sizeof(u64)))
3105 return n * sizeof(u64);
3109 * Read the performance event - simple non blocking version for now
3112 perf_read_hw(struct perf_event *event, char __user *buf, size_t count)
3114 u64 read_format = event->attr.read_format;
3118 * Return end-of-file for a read on a event that is in
3119 * error state (i.e. because it was pinned but it couldn't be
3120 * scheduled on to the CPU at some point).
3122 if (event->state == PERF_EVENT_STATE_ERROR)
3125 if (count < event->read_size)
3128 WARN_ON_ONCE(event->ctx->parent_ctx);
3129 if (read_format & PERF_FORMAT_GROUP)
3130 ret = perf_event_read_group(event, read_format, buf);
3132 ret = perf_event_read_one(event, read_format, buf);
3138 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
3140 struct perf_event *event = file->private_data;
3142 return perf_read_hw(event, buf, count);
3145 static unsigned int perf_poll(struct file *file, poll_table *wait)
3147 struct perf_event *event = file->private_data;
3148 struct perf_buffer *buffer;
3149 unsigned int events = POLL_HUP;
3152 buffer = rcu_dereference(event->buffer);
3154 events = atomic_xchg(&buffer->poll, 0);
3157 poll_wait(file, &event->waitq, wait);
3162 static void perf_event_reset(struct perf_event *event)
3164 (void)perf_event_read(event);
3165 local64_set(&event->count, 0);
3166 perf_event_update_userpage(event);
3170 * Holding the top-level event's child_mutex means that any
3171 * descendant process that has inherited this event will block
3172 * in sync_child_event if it goes to exit, thus satisfying the
3173 * task existence requirements of perf_event_enable/disable.
3175 static void perf_event_for_each_child(struct perf_event *event,
3176 void (*func)(struct perf_event *))
3178 struct perf_event *child;
3180 WARN_ON_ONCE(event->ctx->parent_ctx);
3181 mutex_lock(&event->child_mutex);
3183 list_for_each_entry(child, &event->child_list, child_list)
3185 mutex_unlock(&event->child_mutex);
3188 static void perf_event_for_each(struct perf_event *event,
3189 void (*func)(struct perf_event *))
3191 struct perf_event_context *ctx = event->ctx;
3192 struct perf_event *sibling;
3194 WARN_ON_ONCE(ctx->parent_ctx);
3195 mutex_lock(&ctx->mutex);
3196 event = event->group_leader;
3198 perf_event_for_each_child(event, func);
3200 list_for_each_entry(sibling, &event->sibling_list, group_entry)
3201 perf_event_for_each_child(event, func);
3202 mutex_unlock(&ctx->mutex);
3205 static int perf_event_period(struct perf_event *event, u64 __user *arg)
3207 struct perf_event_context *ctx = event->ctx;
3211 if (!is_sampling_event(event))
3214 if (copy_from_user(&value, arg, sizeof(value)))
3220 raw_spin_lock_irq(&ctx->lock);
3221 if (event->attr.freq) {
3222 if (value > sysctl_perf_event_sample_rate) {
3227 event->attr.sample_freq = value;
3229 event->attr.sample_period = value;
3230 event->hw.sample_period = value;
3233 raw_spin_unlock_irq(&ctx->lock);
3238 static const struct file_operations perf_fops;
3240 static struct perf_event *perf_fget_light(int fd, int *fput_needed)
3244 file = fget_light(fd, fput_needed);
3246 return ERR_PTR(-EBADF);
3248 if (file->f_op != &perf_fops) {
3249 fput_light(file, *fput_needed);
3251 return ERR_PTR(-EBADF);
3254 return file->private_data;
3257 static int perf_event_set_output(struct perf_event *event,
3258 struct perf_event *output_event);
3259 static int perf_event_set_filter(struct perf_event *event, void __user *arg);
3261 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
3263 struct perf_event *event = file->private_data;
3264 void (*func)(struct perf_event *);
3268 case PERF_EVENT_IOC_ENABLE:
3269 func = perf_event_enable;
3271 case PERF_EVENT_IOC_DISABLE:
3272 func = perf_event_disable;
3274 case PERF_EVENT_IOC_RESET:
3275 func = perf_event_reset;
3278 case PERF_EVENT_IOC_REFRESH:
3279 return perf_event_refresh(event, arg);
3281 case PERF_EVENT_IOC_PERIOD:
3282 return perf_event_period(event, (u64 __user *)arg);
3284 case PERF_EVENT_IOC_SET_OUTPUT:
3286 struct perf_event *output_event = NULL;
3287 int fput_needed = 0;
3291 output_event = perf_fget_light(arg, &fput_needed);
3292 if (IS_ERR(output_event))
3293 return PTR_ERR(output_event);
3296 ret = perf_event_set_output(event, output_event);
3298 fput_light(output_event->filp, fput_needed);
3303 case PERF_EVENT_IOC_SET_FILTER:
3304 return perf_event_set_filter(event, (void __user *)arg);
3310 if (flags & PERF_IOC_FLAG_GROUP)
3311 perf_event_for_each(event, func);
3313 perf_event_for_each_child(event, func);
3318 int perf_event_task_enable(void)
3320 struct perf_event *event;
3322 mutex_lock(¤t->perf_event_mutex);
3323 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
3324 perf_event_for_each_child(event, perf_event_enable);
3325 mutex_unlock(¤t->perf_event_mutex);
3330 int perf_event_task_disable(void)
3332 struct perf_event *event;
3334 mutex_lock(¤t->perf_event_mutex);
3335 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
3336 perf_event_for_each_child(event, perf_event_disable);
3337 mutex_unlock(¤t->perf_event_mutex);
3342 #ifndef PERF_EVENT_INDEX_OFFSET
3343 # define PERF_EVENT_INDEX_OFFSET 0
3346 static int perf_event_index(struct perf_event *event)
3348 if (event->hw.state & PERF_HES_STOPPED)
3351 if (event->state != PERF_EVENT_STATE_ACTIVE)
3354 return event->hw.idx + 1 - PERF_EVENT_INDEX_OFFSET;
3358 * Callers need to ensure there can be no nesting of this function, otherwise
3359 * the seqlock logic goes bad. We can not serialize this because the arch
3360 * code calls this from NMI context.
3362 void perf_event_update_userpage(struct perf_event *event)
3364 struct perf_event_mmap_page *userpg;
3365 struct perf_buffer *buffer;
3368 buffer = rcu_dereference(event->buffer);
3372 userpg = buffer->user_page;
3375 * Disable preemption so as to not let the corresponding user-space
3376 * spin too long if we get preempted.
3381 userpg->index = perf_event_index(event);
3382 userpg->offset = perf_event_count(event);
3383 if (event->state == PERF_EVENT_STATE_ACTIVE)
3384 userpg->offset -= local64_read(&event->hw.prev_count);
3386 userpg->time_enabled = event->total_time_enabled +
3387 atomic64_read(&event->child_total_time_enabled);
3389 userpg->time_running = event->total_time_running +
3390 atomic64_read(&event->child_total_time_running);
3399 static unsigned long perf_data_size(struct perf_buffer *buffer);
3402 perf_buffer_init(struct perf_buffer *buffer, long watermark, int flags)
3404 long max_size = perf_data_size(buffer);
3407 buffer->watermark = min(max_size, watermark);
3409 if (!buffer->watermark)
3410 buffer->watermark = max_size / 2;
3412 if (flags & PERF_BUFFER_WRITABLE)
3413 buffer->writable = 1;
3415 atomic_set(&buffer->refcount, 1);
3418 #ifndef CONFIG_PERF_USE_VMALLOC
3421 * Back perf_mmap() with regular GFP_KERNEL-0 pages.
3424 static struct page *
3425 perf_mmap_to_page(struct perf_buffer *buffer, unsigned long pgoff)
3427 if (pgoff > buffer->nr_pages)
3431 return virt_to_page(buffer->user_page);
3433 return virt_to_page(buffer->data_pages[pgoff - 1]);
3436 static void *perf_mmap_alloc_page(int cpu)
3441 node = (cpu == -1) ? cpu : cpu_to_node(cpu);
3442 page = alloc_pages_node(node, GFP_KERNEL | __GFP_ZERO, 0);
3446 return page_address(page);
3449 static struct perf_buffer *
3450 perf_buffer_alloc(int nr_pages, long watermark, int cpu, int flags)
3452 struct perf_buffer *buffer;
3456 size = sizeof(struct perf_buffer);
3457 size += nr_pages * sizeof(void *);
3459 buffer = kzalloc(size, GFP_KERNEL);
3463 buffer->user_page = perf_mmap_alloc_page(cpu);
3464 if (!buffer->user_page)
3465 goto fail_user_page;
3467 for (i = 0; i < nr_pages; i++) {
3468 buffer->data_pages[i] = perf_mmap_alloc_page(cpu);
3469 if (!buffer->data_pages[i])
3470 goto fail_data_pages;
3473 buffer->nr_pages = nr_pages;
3475 perf_buffer_init(buffer, watermark, flags);
3480 for (i--; i >= 0; i--)
3481 free_page((unsigned long)buffer->data_pages[i]);
3483 free_page((unsigned long)buffer->user_page);
3492 static void perf_mmap_free_page(unsigned long addr)
3494 struct page *page = virt_to_page((void *)addr);
3496 page->mapping = NULL;
3500 static void perf_buffer_free(struct perf_buffer *buffer)
3504 perf_mmap_free_page((unsigned long)buffer->user_page);
3505 for (i = 0; i < buffer->nr_pages; i++)
3506 perf_mmap_free_page((unsigned long)buffer->data_pages[i]);
3510 static inline int page_order(struct perf_buffer *buffer)
3518 * Back perf_mmap() with vmalloc memory.
3520 * Required for architectures that have d-cache aliasing issues.
3523 static inline int page_order(struct perf_buffer *buffer)
3525 return buffer->page_order;
3528 static struct page *
3529 perf_mmap_to_page(struct perf_buffer *buffer, unsigned long pgoff)
3531 if (pgoff > (1UL << page_order(buffer)))
3534 return vmalloc_to_page((void *)buffer->user_page + pgoff * PAGE_SIZE);
3537 static void perf_mmap_unmark_page(void *addr)
3539 struct page *page = vmalloc_to_page(addr);
3541 page->mapping = NULL;
3544 static void perf_buffer_free_work(struct work_struct *work)
3546 struct perf_buffer *buffer;
3550 buffer = container_of(work, struct perf_buffer, work);
3551 nr = 1 << page_order(buffer);
3553 base = buffer->user_page;
3554 for (i = 0; i < nr + 1; i++)
3555 perf_mmap_unmark_page(base + (i * PAGE_SIZE));
3561 static void perf_buffer_free(struct perf_buffer *buffer)
3563 schedule_work(&buffer->work);
3566 static struct perf_buffer *
3567 perf_buffer_alloc(int nr_pages, long watermark, int cpu, int flags)
3569 struct perf_buffer *buffer;
3573 size = sizeof(struct perf_buffer);
3574 size += sizeof(void *);
3576 buffer = kzalloc(size, GFP_KERNEL);
3580 INIT_WORK(&buffer->work, perf_buffer_free_work);
3582 all_buf = vmalloc_user((nr_pages + 1) * PAGE_SIZE);
3586 buffer->user_page = all_buf;
3587 buffer->data_pages[0] = all_buf + PAGE_SIZE;
3588 buffer->page_order = ilog2(nr_pages);
3589 buffer->nr_pages = 1;
3591 perf_buffer_init(buffer, watermark, flags);
3604 static unsigned long perf_data_size(struct perf_buffer *buffer)
3606 return buffer->nr_pages << (PAGE_SHIFT + page_order(buffer));
3609 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
3611 struct perf_event *event = vma->vm_file->private_data;
3612 struct perf_buffer *buffer;
3613 int ret = VM_FAULT_SIGBUS;
3615 if (vmf->flags & FAULT_FLAG_MKWRITE) {
3616 if (vmf->pgoff == 0)
3622 buffer = rcu_dereference(event->buffer);
3626 if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE))
3629 vmf->page = perf_mmap_to_page(buffer, vmf->pgoff);
3633 get_page(vmf->page);
3634 vmf->page->mapping = vma->vm_file->f_mapping;
3635 vmf->page->index = vmf->pgoff;
3644 static void perf_buffer_free_rcu(struct rcu_head *rcu_head)
3646 struct perf_buffer *buffer;
3648 buffer = container_of(rcu_head, struct perf_buffer, rcu_head);
3649 perf_buffer_free(buffer);
3652 static struct perf_buffer *perf_buffer_get(struct perf_event *event)
3654 struct perf_buffer *buffer;
3657 buffer = rcu_dereference(event->buffer);
3659 if (!atomic_inc_not_zero(&buffer->refcount))
3667 static void perf_buffer_put(struct perf_buffer *buffer)
3669 if (!atomic_dec_and_test(&buffer->refcount))
3672 call_rcu(&buffer->rcu_head, perf_buffer_free_rcu);
3675 static void perf_mmap_open(struct vm_area_struct *vma)
3677 struct perf_event *event = vma->vm_file->private_data;
3679 atomic_inc(&event->mmap_count);
3682 static void perf_mmap_close(struct vm_area_struct *vma)
3684 struct perf_event *event = vma->vm_file->private_data;
3686 if (atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex)) {
3687 unsigned long size = perf_data_size(event->buffer);
3688 struct user_struct *user = event->mmap_user;
3689 struct perf_buffer *buffer = event->buffer;
3691 atomic_long_sub((size >> PAGE_SHIFT) + 1, &user->locked_vm);
3692 vma->vm_mm->locked_vm -= event->mmap_locked;
3693 rcu_assign_pointer(event->buffer, NULL);
3694 mutex_unlock(&event->mmap_mutex);
3696 perf_buffer_put(buffer);
3701 static const struct vm_operations_struct perf_mmap_vmops = {
3702 .open = perf_mmap_open,
3703 .close = perf_mmap_close,
3704 .fault = perf_mmap_fault,
3705 .page_mkwrite = perf_mmap_fault,
3708 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
3710 struct perf_event *event = file->private_data;
3711 unsigned long user_locked, user_lock_limit;
3712 struct user_struct *user = current_user();
3713 unsigned long locked, lock_limit;
3714 struct perf_buffer *buffer;
3715 unsigned long vma_size;
3716 unsigned long nr_pages;
3717 long user_extra, extra;
3718 int ret = 0, flags = 0;
3721 * Don't allow mmap() of inherited per-task counters. This would
3722 * create a performance issue due to all children writing to the
3725 if (event->cpu == -1 && event->attr.inherit)
3728 if (!(vma->vm_flags & VM_SHARED))
3731 vma_size = vma->vm_end - vma->vm_start;
3732 nr_pages = (vma_size / PAGE_SIZE) - 1;
3735 * If we have buffer pages ensure they're a power-of-two number, so we
3736 * can do bitmasks instead of modulo.
3738 if (nr_pages != 0 && !is_power_of_2(nr_pages))
3741 if (vma_size != PAGE_SIZE * (1 + nr_pages))
3744 if (vma->vm_pgoff != 0)
3747 WARN_ON_ONCE(event->ctx->parent_ctx);
3748 mutex_lock(&event->mmap_mutex);
3749 if (event->buffer) {
3750 if (event->buffer->nr_pages == nr_pages)
3751 atomic_inc(&event->buffer->refcount);
3757 user_extra = nr_pages + 1;
3758 user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10);
3761 * Increase the limit linearly with more CPUs:
3763 user_lock_limit *= num_online_cpus();
3765 user_locked = atomic_long_read(&user->locked_vm) + user_extra;
3768 if (user_locked > user_lock_limit)
3769 extra = user_locked - user_lock_limit;
3771 lock_limit = rlimit(RLIMIT_MEMLOCK);
3772 lock_limit >>= PAGE_SHIFT;
3773 locked = vma->vm_mm->locked_vm + extra;
3775 if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() &&
3776 !capable(CAP_IPC_LOCK)) {
3781 WARN_ON(event->buffer);
3783 if (vma->vm_flags & VM_WRITE)
3784 flags |= PERF_BUFFER_WRITABLE;
3786 buffer = perf_buffer_alloc(nr_pages, event->attr.wakeup_watermark,
3792 rcu_assign_pointer(event->buffer, buffer);
3794 atomic_long_add(user_extra, &user->locked_vm);
3795 event->mmap_locked = extra;
3796 event->mmap_user = get_current_user();
3797 vma->vm_mm->locked_vm += event->mmap_locked;
3801 atomic_inc(&event->mmap_count);
3802 mutex_unlock(&event->mmap_mutex);
3804 vma->vm_flags |= VM_RESERVED;
3805 vma->vm_ops = &perf_mmap_vmops;
3810 static int perf_fasync(int fd, struct file *filp, int on)
3812 struct inode *inode = filp->f_path.dentry->d_inode;
3813 struct perf_event *event = filp->private_data;
3816 mutex_lock(&inode->i_mutex);
3817 retval = fasync_helper(fd, filp, on, &event->fasync);
3818 mutex_unlock(&inode->i_mutex);
3826 static const struct file_operations perf_fops = {
3827 .llseek = no_llseek,
3828 .release = perf_release,
3831 .unlocked_ioctl = perf_ioctl,
3832 .compat_ioctl = perf_ioctl,
3834 .fasync = perf_fasync,
3840 * If there's data, ensure we set the poll() state and publish everything
3841 * to user-space before waking everybody up.
3844 void perf_event_wakeup(struct perf_event *event)
3846 wake_up_all(&event->waitq);
3848 if (event->pending_kill) {
3849 kill_fasync(&event->fasync, SIGIO, event->pending_kill);
3850 event->pending_kill = 0;
3854 static void perf_pending_event(struct irq_work *entry)
3856 struct perf_event *event = container_of(entry,
3857 struct perf_event, pending);
3859 if (event->pending_disable) {
3860 event->pending_disable = 0;
3861 __perf_event_disable(event);
3864 if (event->pending_wakeup) {
3865 event->pending_wakeup = 0;
3866 perf_event_wakeup(event);
3871 * We assume there is only KVM supporting the callbacks.
3872 * Later on, we might change it to a list if there is
3873 * another virtualization implementation supporting the callbacks.
3875 struct perf_guest_info_callbacks *perf_guest_cbs;
3877 int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
3879 perf_guest_cbs = cbs;
3882 EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks);
3884 int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
3886 perf_guest_cbs = NULL;
3889 EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks);
3894 static bool perf_output_space(struct perf_buffer *buffer, unsigned long tail,
3895 unsigned long offset, unsigned long head)
3899 if (!buffer->writable)
3902 mask = perf_data_size(buffer) - 1;
3904 offset = (offset - tail) & mask;
3905 head = (head - tail) & mask;
3907 if ((int)(head - offset) < 0)
3913 static void perf_output_wakeup(struct perf_output_handle *handle)
3915 atomic_set(&handle->buffer->poll, POLL_IN);
3918 handle->event->pending_wakeup = 1;
3919 irq_work_queue(&handle->event->pending);
3921 perf_event_wakeup(handle->event);
3925 * We need to ensure a later event_id doesn't publish a head when a former
3926 * event isn't done writing. However since we need to deal with NMIs we
3927 * cannot fully serialize things.
3929 * We only publish the head (and generate a wakeup) when the outer-most
3932 static void perf_output_get_handle(struct perf_output_handle *handle)
3934 struct perf_buffer *buffer = handle->buffer;
3937 local_inc(&buffer->nest);
3938 handle->wakeup = local_read(&buffer->wakeup);
3941 static void perf_output_put_handle(struct perf_output_handle *handle)
3943 struct perf_buffer *buffer = handle->buffer;
3947 head = local_read(&buffer->head);
3950 * IRQ/NMI can happen here, which means we can miss a head update.
3953 if (!local_dec_and_test(&buffer->nest))
3957 * Publish the known good head. Rely on the full barrier implied
3958 * by atomic_dec_and_test() order the buffer->head read and this
3961 buffer->user_page->data_head = head;
3964 * Now check if we missed an update, rely on the (compiler)
3965 * barrier in atomic_dec_and_test() to re-read buffer->head.
3967 if (unlikely(head != local_read(&buffer->head))) {
3968 local_inc(&buffer->nest);
3972 if (handle->wakeup != local_read(&buffer->wakeup))
3973 perf_output_wakeup(handle);
3979 __always_inline void perf_output_copy(struct perf_output_handle *handle,
3980 const void *buf, unsigned int len)
3983 unsigned long size = min_t(unsigned long, handle->size, len);
3985 memcpy(handle->addr, buf, size);
3988 handle->addr += size;
3990 handle->size -= size;
3991 if (!handle->size) {
3992 struct perf_buffer *buffer = handle->buffer;
3995 handle->page &= buffer->nr_pages - 1;
3996 handle->addr = buffer->data_pages[handle->page];
3997 handle->size = PAGE_SIZE << page_order(buffer);
4002 static void __perf_event_header__init_id(struct perf_event_header *header,
4003 struct perf_sample_data *data,
4004 struct perf_event *event)
4006 u64 sample_type = event->attr.sample_type;
4008 data->type = sample_type;
4009 header->size += event->id_header_size;
4011 if (sample_type & PERF_SAMPLE_TID) {
4012 /* namespace issues */
4013 data->tid_entry.pid = perf_event_pid(event, current);
4014 data->tid_entry.tid = perf_event_tid(event, current);
4017 if (sample_type & PERF_SAMPLE_TIME)
4018 data->time = perf_clock();
4020 if (sample_type & PERF_SAMPLE_ID)
4021 data->id = primary_event_id(event);
4023 if (sample_type & PERF_SAMPLE_STREAM_ID)
4024 data->stream_id = event->id;
4026 if (sample_type & PERF_SAMPLE_CPU) {
4027 data->cpu_entry.cpu = raw_smp_processor_id();
4028 data->cpu_entry.reserved = 0;
4032 static void perf_event_header__init_id(struct perf_event_header *header,
4033 struct perf_sample_data *data,
4034 struct perf_event *event)
4036 if (event->attr.sample_id_all)
4037 __perf_event_header__init_id(header, data, event);
4040 static void __perf_event__output_id_sample(struct perf_output_handle *handle,
4041 struct perf_sample_data *data)
4043 u64 sample_type = data->type;
4045 if (sample_type & PERF_SAMPLE_TID)
4046 perf_output_put(handle, data->tid_entry);
4048 if (sample_type & PERF_SAMPLE_TIME)
4049 perf_output_put(handle, data->time);
4051 if (sample_type & PERF_SAMPLE_ID)
4052 perf_output_put(handle, data->id);
4054 if (sample_type & PERF_SAMPLE_STREAM_ID)
4055 perf_output_put(handle, data->stream_id);
4057 if (sample_type & PERF_SAMPLE_CPU)
4058 perf_output_put(handle, data->cpu_entry);
4061 static void perf_event__output_id_sample(struct perf_event *event,
4062 struct perf_output_handle *handle,
4063 struct perf_sample_data *sample)
4065 if (event->attr.sample_id_all)
4066 __perf_event__output_id_sample(handle, sample);
4069 int perf_output_begin(struct perf_output_handle *handle,
4070 struct perf_event *event, unsigned int size,
4071 int nmi, int sample)
4073 struct perf_buffer *buffer;
4074 unsigned long tail, offset, head;
4076 struct perf_sample_data sample_data;
4078 struct perf_event_header header;
4085 * For inherited events we send all the output towards the parent.
4088 event = event->parent;
4090 buffer = rcu_dereference(event->buffer);
4094 handle->buffer = buffer;
4095 handle->event = event;
4097 handle->sample = sample;
4099 if (!buffer->nr_pages)
4102 have_lost = local_read(&buffer->lost);
4104 lost_event.header.size = sizeof(lost_event);
4105 perf_event_header__init_id(&lost_event.header, &sample_data,
4107 size += lost_event.header.size;
4110 perf_output_get_handle(handle);
4114 * Userspace could choose to issue a mb() before updating the
4115 * tail pointer. So that all reads will be completed before the
4118 tail = ACCESS_ONCE(buffer->user_page->data_tail);
4120 offset = head = local_read(&buffer->head);
4122 if (unlikely(!perf_output_space(buffer, tail, offset, head)))
4124 } while (local_cmpxchg(&buffer->head, offset, head) != offset);
4126 if (head - local_read(&buffer->wakeup) > buffer->watermark)
4127 local_add(buffer->watermark, &buffer->wakeup);
4129 handle->page = offset >> (PAGE_SHIFT + page_order(buffer));
4130 handle->page &= buffer->nr_pages - 1;
4131 handle->size = offset & ((PAGE_SIZE << page_order(buffer)) - 1);
4132 handle->addr = buffer->data_pages[handle->page];
4133 handle->addr += handle->size;
4134 handle->size = (PAGE_SIZE << page_order(buffer)) - handle->size;
4137 lost_event.header.type = PERF_RECORD_LOST;
4138 lost_event.header.misc = 0;
4139 lost_event.id = event->id;
4140 lost_event.lost = local_xchg(&buffer->lost, 0);
4142 perf_output_put(handle, lost_event);
4143 perf_event__output_id_sample(event, handle, &sample_data);
4149 local_inc(&buffer->lost);
4150 perf_output_put_handle(handle);
4157 void perf_output_end(struct perf_output_handle *handle)
4159 struct perf_event *event = handle->event;
4160 struct perf_buffer *buffer = handle->buffer;
4162 int wakeup_events = event->attr.wakeup_events;
4164 if (handle->sample && wakeup_events) {
4165 int events = local_inc_return(&buffer->events);
4166 if (events >= wakeup_events) {
4167 local_sub(wakeup_events, &buffer->events);
4168 local_inc(&buffer->wakeup);
4172 perf_output_put_handle(handle);
4176 static void perf_output_read_one(struct perf_output_handle *handle,
4177 struct perf_event *event,
4178 u64 enabled, u64 running)
4180 u64 read_format = event->attr.read_format;
4184 values[n++] = perf_event_count(event);
4185 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
4186 values[n++] = enabled +
4187 atomic64_read(&event->child_total_time_enabled);
4189 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
4190 values[n++] = running +
4191 atomic64_read(&event->child_total_time_running);
4193 if (read_format & PERF_FORMAT_ID)
4194 values[n++] = primary_event_id(event);
4196 perf_output_copy(handle, values, n * sizeof(u64));
4200 * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
4202 static void perf_output_read_group(struct perf_output_handle *handle,
4203 struct perf_event *event,
4204 u64 enabled, u64 running)
4206 struct perf_event *leader = event->group_leader, *sub;
4207 u64 read_format = event->attr.read_format;
4211 values[n++] = 1 + leader->nr_siblings;
4213 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
4214 values[n++] = enabled;
4216 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
4217 values[n++] = running;
4219 if (leader != event)
4220 leader->pmu->read(leader);
4222 values[n++] = perf_event_count(leader);
4223 if (read_format & PERF_FORMAT_ID)
4224 values[n++] = primary_event_id(leader);
4226 perf_output_copy(handle, values, n * sizeof(u64));
4228 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
4232 sub->pmu->read(sub);
4234 values[n++] = perf_event_count(sub);
4235 if (read_format & PERF_FORMAT_ID)
4236 values[n++] = primary_event_id(sub);
4238 perf_output_copy(handle, values, n * sizeof(u64));
4242 #define PERF_FORMAT_TOTAL_TIMES (PERF_FORMAT_TOTAL_TIME_ENABLED|\
4243 PERF_FORMAT_TOTAL_TIME_RUNNING)
4245 static void perf_output_read(struct perf_output_handle *handle,
4246 struct perf_event *event)
4248 u64 enabled = 0, running = 0, now, ctx_time;
4249 u64 read_format = event->attr.read_format;
4252 * compute total_time_enabled, total_time_running
4253 * based on snapshot values taken when the event
4254 * was last scheduled in.
4256 * we cannot simply called update_context_time()
4257 * because of locking issue as we are called in
4260 if (read_format & PERF_FORMAT_TOTAL_TIMES) {
4262 ctx_time = event->shadow_ctx_time + now;
4263 enabled = ctx_time - event->tstamp_enabled;
4264 running = ctx_time - event->tstamp_running;
4267 if (event->attr.read_format & PERF_FORMAT_GROUP)
4268 perf_output_read_group(handle, event, enabled, running);
4270 perf_output_read_one(handle, event, enabled, running);
4273 void perf_output_sample(struct perf_output_handle *handle,
4274 struct perf_event_header *header,
4275 struct perf_sample_data *data,
4276 struct perf_event *event)
4278 u64 sample_type = data->type;
4280 perf_output_put(handle, *header);
4282 if (sample_type & PERF_SAMPLE_IP)
4283 perf_output_put(handle, data->ip);
4285 if (sample_type & PERF_SAMPLE_TID)
4286 perf_output_put(handle, data->tid_entry);
4288 if (sample_type & PERF_SAMPLE_TIME)
4289 perf_output_put(handle, data->time);
4291 if (sample_type & PERF_SAMPLE_ADDR)
4292 perf_output_put(handle, data->addr);
4294 if (sample_type & PERF_SAMPLE_ID)
4295 perf_output_put(handle, data->id);
4297 if (sample_type & PERF_SAMPLE_STREAM_ID)
4298 perf_output_put(handle, data->stream_id);
4300 if (sample_type & PERF_SAMPLE_CPU)
4301 perf_output_put(handle, data->cpu_entry);
4303 if (sample_type & PERF_SAMPLE_PERIOD)
4304 perf_output_put(handle, data->period);
4306 if (sample_type & PERF_SAMPLE_READ)
4307 perf_output_read(handle, event);
4309 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
4310 if (data->callchain) {
4313 if (data->callchain)
4314 size += data->callchain->nr;
4316 size *= sizeof(u64);
4318 perf_output_copy(handle, data->callchain, size);
4321 perf_output_put(handle, nr);
4325 if (sample_type & PERF_SAMPLE_RAW) {
4327 perf_output_put(handle, data->raw->size);
4328 perf_output_copy(handle, data->raw->data,
4335 .size = sizeof(u32),
4338 perf_output_put(handle, raw);
4343 void perf_prepare_sample(struct perf_event_header *header,
4344 struct perf_sample_data *data,
4345 struct perf_event *event,
4346 struct pt_regs *regs)
4348 u64 sample_type = event->attr.sample_type;
4350 header->type = PERF_RECORD_SAMPLE;
4351 header->size = sizeof(*header) + event->header_size;
4354 header->misc |= perf_misc_flags(regs);
4356 __perf_event_header__init_id(header, data, event);
4358 if (sample_type & PERF_SAMPLE_IP)
4359 data->ip = perf_instruction_pointer(regs);
4361 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
4364 data->callchain = perf_callchain(regs);
4366 if (data->callchain)
4367 size += data->callchain->nr;
4369 header->size += size * sizeof(u64);
4372 if (sample_type & PERF_SAMPLE_RAW) {
4373 int size = sizeof(u32);
4376 size += data->raw->size;
4378 size += sizeof(u32);
4380 WARN_ON_ONCE(size & (sizeof(u64)-1));
4381 header->size += size;
4385 static void perf_event_output(struct perf_event *event, int nmi,
4386 struct perf_sample_data *data,
4387 struct pt_regs *regs)
4389 struct perf_output_handle handle;
4390 struct perf_event_header header;
4392 /* protect the callchain buffers */
4395 perf_prepare_sample(&header, data, event, regs);
4397 if (perf_output_begin(&handle, event, header.size, nmi, 1))
4400 perf_output_sample(&handle, &header, data, event);
4402 perf_output_end(&handle);
4412 struct perf_read_event {
4413 struct perf_event_header header;
4420 perf_event_read_event(struct perf_event *event,
4421 struct task_struct *task)
4423 struct perf_output_handle handle;
4424 struct perf_sample_data sample;
4425 struct perf_read_event read_event = {
4427 .type = PERF_RECORD_READ,
4429 .size = sizeof(read_event) + event->read_size,
4431 .pid = perf_event_pid(event, task),
4432 .tid = perf_event_tid(event, task),
4436 perf_event_header__init_id(&read_event.header, &sample, event);
4437 ret = perf_output_begin(&handle, event, read_event.header.size, 0, 0);
4441 perf_output_put(&handle, read_event);
4442 perf_output_read(&handle, event);
4443 perf_event__output_id_sample(event, &handle, &sample);
4445 perf_output_end(&handle);
4449 * task tracking -- fork/exit
4451 * enabled by: attr.comm | attr.mmap | attr.mmap_data | attr.task
4454 struct perf_task_event {
4455 struct task_struct *task;
4456 struct perf_event_context *task_ctx;
4459 struct perf_event_header header;
4469 static void perf_event_task_output(struct perf_event *event,
4470 struct perf_task_event *task_event)
4472 struct perf_output_handle handle;
4473 struct perf_sample_data sample;
4474 struct task_struct *task = task_event->task;
4475 int ret, size = task_event->event_id.header.size;
4477 perf_event_header__init_id(&task_event->event_id.header, &sample, event);
4479 ret = perf_output_begin(&handle, event,
4480 task_event->event_id.header.size, 0, 0);
4484 task_event->event_id.pid = perf_event_pid(event, task);
4485 task_event->event_id.ppid = perf_event_pid(event, current);
4487 task_event->event_id.tid = perf_event_tid(event, task);
4488 task_event->event_id.ptid = perf_event_tid(event, current);
4490 perf_output_put(&handle, task_event->event_id);
4492 perf_event__output_id_sample(event, &handle, &sample);
4494 perf_output_end(&handle);
4496 task_event->event_id.header.size = size;
4499 static int perf_event_task_match(struct perf_event *event)
4501 if (event->state < PERF_EVENT_STATE_INACTIVE)
4504 if (!event_filter_match(event))
4507 if (event->attr.comm || event->attr.mmap ||
4508 event->attr.mmap_data || event->attr.task)
4514 static void perf_event_task_ctx(struct perf_event_context *ctx,
4515 struct perf_task_event *task_event)
4517 struct perf_event *event;
4519 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
4520 if (perf_event_task_match(event))
4521 perf_event_task_output(event, task_event);
4525 static void perf_event_task_event(struct perf_task_event *task_event)
4527 struct perf_cpu_context *cpuctx;
4528 struct perf_event_context *ctx;
4533 list_for_each_entry_rcu(pmu, &pmus, entry) {
4534 cpuctx = get_cpu_ptr(pmu->pmu_cpu_context);
4535 if (cpuctx->active_pmu != pmu)
4537 perf_event_task_ctx(&cpuctx->ctx, task_event);
4539 ctx = task_event->task_ctx;
4541 ctxn = pmu->task_ctx_nr;
4544 ctx = rcu_dereference(current->perf_event_ctxp[ctxn]);
4547 perf_event_task_ctx(ctx, task_event);
4549 put_cpu_ptr(pmu->pmu_cpu_context);
4554 static void perf_event_task(struct task_struct *task,
4555 struct perf_event_context *task_ctx,
4558 struct perf_task_event task_event;
4560 if (!atomic_read(&nr_comm_events) &&
4561 !atomic_read(&nr_mmap_events) &&
4562 !atomic_read(&nr_task_events))
4565 task_event = (struct perf_task_event){
4567 .task_ctx = task_ctx,
4570 .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT,
4572 .size = sizeof(task_event.event_id),
4578 .time = perf_clock(),
4582 perf_event_task_event(&task_event);
4585 void perf_event_fork(struct task_struct *task)
4587 perf_event_task(task, NULL, 1);
4594 struct perf_comm_event {
4595 struct task_struct *task;
4600 struct perf_event_header header;
4607 static void perf_event_comm_output(struct perf_event *event,
4608 struct perf_comm_event *comm_event)
4610 struct perf_output_handle handle;
4611 struct perf_sample_data sample;
4612 int size = comm_event->event_id.header.size;
4615 perf_event_header__init_id(&comm_event->event_id.header, &sample, event);
4616 ret = perf_output_begin(&handle, event,
4617 comm_event->event_id.header.size, 0, 0);
4622 comm_event->event_id.pid = perf_event_pid(event, comm_event->task);
4623 comm_event->event_id.tid = perf_event_tid(event, comm_event->task);
4625 perf_output_put(&handle, comm_event->event_id);
4626 perf_output_copy(&handle, comm_event->comm,
4627 comm_event->comm_size);
4629 perf_event__output_id_sample(event, &handle, &sample);
4631 perf_output_end(&handle);
4633 comm_event->event_id.header.size = size;
4636 static int perf_event_comm_match(struct perf_event *event)
4638 if (event->state < PERF_EVENT_STATE_INACTIVE)
4641 if (!event_filter_match(event))
4644 if (event->attr.comm)
4650 static void perf_event_comm_ctx(struct perf_event_context *ctx,
4651 struct perf_comm_event *comm_event)
4653 struct perf_event *event;
4655 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
4656 if (perf_event_comm_match(event))
4657 perf_event_comm_output(event, comm_event);
4661 static void perf_event_comm_event(struct perf_comm_event *comm_event)
4663 struct perf_cpu_context *cpuctx;
4664 struct perf_event_context *ctx;
4665 char comm[TASK_COMM_LEN];
4670 memset(comm, 0, sizeof(comm));
4671 strlcpy(comm, comm_event->task->comm, sizeof(comm));
4672 size = ALIGN(strlen(comm)+1, sizeof(u64));
4674 comm_event->comm = comm;
4675 comm_event->comm_size = size;
4677 comm_event->event_id.header.size = sizeof(comm_event->event_id) + size;
4679 list_for_each_entry_rcu(pmu, &pmus, entry) {
4680 cpuctx = get_cpu_ptr(pmu->pmu_cpu_context);
4681 if (cpuctx->active_pmu != pmu)
4683 perf_event_comm_ctx(&cpuctx->ctx, comm_event);
4685 ctxn = pmu->task_ctx_nr;
4689 ctx = rcu_dereference(current->perf_event_ctxp[ctxn]);
4691 perf_event_comm_ctx(ctx, comm_event);
4693 put_cpu_ptr(pmu->pmu_cpu_context);
4698 void perf_event_comm(struct task_struct *task)
4700 struct perf_comm_event comm_event;
4701 struct perf_event_context *ctx;
4704 for_each_task_context_nr(ctxn) {
4705 ctx = task->perf_event_ctxp[ctxn];
4709 perf_event_enable_on_exec(ctx);
4712 if (!atomic_read(&nr_comm_events))
4715 comm_event = (struct perf_comm_event){
4721 .type = PERF_RECORD_COMM,
4730 perf_event_comm_event(&comm_event);
4737 struct perf_mmap_event {
4738 struct vm_area_struct *vma;
4740 const char *file_name;
4744 struct perf_event_header header;
4754 static void perf_event_mmap_output(struct perf_event *event,
4755 struct perf_mmap_event *mmap_event)
4757 struct perf_output_handle handle;
4758 struct perf_sample_data sample;
4759 int size = mmap_event->event_id.header.size;
4762 perf_event_header__init_id(&mmap_event->event_id.header, &sample, event);
4763 ret = perf_output_begin(&handle, event,
4764 mmap_event->event_id.header.size, 0, 0);
4768 mmap_event->event_id.pid = perf_event_pid(event, current);
4769 mmap_event->event_id.tid = perf_event_tid(event, current);
4771 perf_output_put(&handle, mmap_event->event_id);
4772 perf_output_copy(&handle, mmap_event->file_name,
4773 mmap_event->file_size);
4775 perf_event__output_id_sample(event, &handle, &sample);
4777 perf_output_end(&handle);
4779 mmap_event->event_id.header.size = size;
4782 static int perf_event_mmap_match(struct perf_event *event,
4783 struct perf_mmap_event *mmap_event,
4786 if (event->state < PERF_EVENT_STATE_INACTIVE)
4789 if (!event_filter_match(event))
4792 if ((!executable && event->attr.mmap_data) ||
4793 (executable && event->attr.mmap))
4799 static void perf_event_mmap_ctx(struct perf_event_context *ctx,
4800 struct perf_mmap_event *mmap_event,
4803 struct perf_event *event;
4805 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
4806 if (perf_event_mmap_match(event, mmap_event, executable))
4807 perf_event_mmap_output(event, mmap_event);
4811 static void perf_event_mmap_event(struct perf_mmap_event *mmap_event)
4813 struct perf_cpu_context *cpuctx;
4814 struct perf_event_context *ctx;
4815 struct vm_area_struct *vma = mmap_event->vma;
4816 struct file *file = vma->vm_file;
4824 memset(tmp, 0, sizeof(tmp));
4828 * d_path works from the end of the buffer backwards, so we
4829 * need to add enough zero bytes after the string to handle
4830 * the 64bit alignment we do later.
4832 buf = kzalloc(PATH_MAX + sizeof(u64), GFP_KERNEL);
4834 name = strncpy(tmp, "//enomem", sizeof(tmp));
4837 name = d_path(&file->f_path, buf, PATH_MAX);
4839 name = strncpy(tmp, "//toolong", sizeof(tmp));
4843 if (arch_vma_name(mmap_event->vma)) {
4844 name = strncpy(tmp, arch_vma_name(mmap_event->vma),
4850 name = strncpy(tmp, "[vdso]", sizeof(tmp));
4852 } else if (vma->vm_start <= vma->vm_mm->start_brk &&
4853 vma->vm_end >= vma->vm_mm->brk) {
4854 name = strncpy(tmp, "[heap]", sizeof(tmp));
4856 } else if (vma->vm_start <= vma->vm_mm->start_stack &&
4857 vma->vm_end >= vma->vm_mm->start_stack) {
4858 name = strncpy(tmp, "[stack]", sizeof(tmp));
4862 name = strncpy(tmp, "//anon", sizeof(tmp));
4867 size = ALIGN(strlen(name)+1, sizeof(u64));
4869 mmap_event->file_name = name;
4870 mmap_event->file_size = size;
4872 mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size;
4875 list_for_each_entry_rcu(pmu, &pmus, entry) {
4876 cpuctx = get_cpu_ptr(pmu->pmu_cpu_context);
4877 if (cpuctx->active_pmu != pmu)
4879 perf_event_mmap_ctx(&cpuctx->ctx, mmap_event,
4880 vma->vm_flags & VM_EXEC);
4882 ctxn = pmu->task_ctx_nr;
4886 ctx = rcu_dereference(current->perf_event_ctxp[ctxn]);
4888 perf_event_mmap_ctx(ctx, mmap_event,
4889 vma->vm_flags & VM_EXEC);
4892 put_cpu_ptr(pmu->pmu_cpu_context);
4899 void perf_event_mmap(struct vm_area_struct *vma)
4901 struct perf_mmap_event mmap_event;
4903 if (!atomic_read(&nr_mmap_events))
4906 mmap_event = (struct perf_mmap_event){
4912 .type = PERF_RECORD_MMAP,
4913 .misc = PERF_RECORD_MISC_USER,
4918 .start = vma->vm_start,
4919 .len = vma->vm_end - vma->vm_start,
4920 .pgoff = (u64)vma->vm_pgoff << PAGE_SHIFT,
4924 perf_event_mmap_event(&mmap_event);
4928 * IRQ throttle logging
4931 static void perf_log_throttle(struct perf_event *event, int enable)
4933 struct perf_output_handle handle;
4934 struct perf_sample_data sample;
4938 struct perf_event_header header;
4942 } throttle_event = {
4944 .type = PERF_RECORD_THROTTLE,
4946 .size = sizeof(throttle_event),
4948 .time = perf_clock(),
4949 .id = primary_event_id(event),
4950 .stream_id = event->id,
4954 throttle_event.header.type = PERF_RECORD_UNTHROTTLE;
4956 perf_event_header__init_id(&throttle_event.header, &sample, event);
4958 ret = perf_output_begin(&handle, event,
4959 throttle_event.header.size, 1, 0);
4963 perf_output_put(&handle, throttle_event);
4964 perf_event__output_id_sample(event, &handle, &sample);
4965 perf_output_end(&handle);
4969 * Generic event overflow handling, sampling.
4972 static int __perf_event_overflow(struct perf_event *event, int nmi,
4973 int throttle, struct perf_sample_data *data,
4974 struct pt_regs *regs)
4976 int events = atomic_read(&event->event_limit);
4977 struct hw_perf_event *hwc = &event->hw;
4981 * Non-sampling counters might still use the PMI to fold short
4982 * hardware counters, ignore those.
4984 if (unlikely(!is_sampling_event(event)))
4987 if (unlikely(hwc->interrupts >= max_samples_per_tick)) {
4989 hwc->interrupts = MAX_INTERRUPTS;
4990 perf_log_throttle(event, 0);
4996 if (event->attr.freq) {
4997 u64 now = perf_clock();
4998 s64 delta = now - hwc->freq_time_stamp;
5000 hwc->freq_time_stamp = now;
5002 if (delta > 0 && delta < 2*TICK_NSEC)
5003 perf_adjust_period(event, delta, hwc->last_period);
5007 * XXX event_limit might not quite work as expected on inherited
5011 event->pending_kill = POLL_IN;
5012 if (events && atomic_dec_and_test(&event->event_limit)) {
5014 event->pending_kill = POLL_HUP;
5016 event->pending_disable = 1;
5017 irq_work_queue(&event->pending);
5019 perf_event_disable(event);
5022 if (event->overflow_handler)
5023 event->overflow_handler(event, nmi, data, regs);
5025 perf_event_output(event, nmi, data, regs);
5030 int perf_event_overflow(struct perf_event *event, int nmi,
5031 struct perf_sample_data *data,
5032 struct pt_regs *regs)
5034 return __perf_event_overflow(event, nmi, 1, data, regs);
5038 * Generic software event infrastructure
5041 struct swevent_htable {
5042 struct swevent_hlist *swevent_hlist;
5043 struct mutex hlist_mutex;
5046 /* Recursion avoidance in each contexts */
5047 int recursion[PERF_NR_CONTEXTS];
5050 static DEFINE_PER_CPU(struct swevent_htable, swevent_htable);
5053 * We directly increment event->count and keep a second value in
5054 * event->hw.period_left to count intervals. This period event
5055 * is kept in the range [-sample_period, 0] so that we can use the
5059 static u64 perf_swevent_set_period(struct perf_event *event)
5061 struct hw_perf_event *hwc = &event->hw;
5062 u64 period = hwc->last_period;
5066 hwc->last_period = hwc->sample_period;
5069 old = val = local64_read(&hwc->period_left);
5073 nr = div64_u64(period + val, period);
5074 offset = nr * period;
5076 if (local64_cmpxchg(&hwc->period_left, old, val) != old)
5082 static void perf_swevent_overflow(struct perf_event *event, u64 overflow,
5083 int nmi, struct perf_sample_data *data,
5084 struct pt_regs *regs)
5086 struct hw_perf_event *hwc = &event->hw;
5089 data->period = event->hw.last_period;
5091 overflow = perf_swevent_set_period(event);
5093 if (hwc->interrupts == MAX_INTERRUPTS)
5096 for (; overflow; overflow--) {
5097 if (__perf_event_overflow(event, nmi, throttle,
5100 * We inhibit the overflow from happening when
5101 * hwc->interrupts == MAX_INTERRUPTS.
5109 static void perf_swevent_event(struct perf_event *event, u64 nr,
5110 int nmi, struct perf_sample_data *data,
5111 struct pt_regs *regs)
5113 struct hw_perf_event *hwc = &event->hw;
5115 local64_add(nr, &event->count);
5120 if (!is_sampling_event(event))
5123 if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq)
5124 return perf_swevent_overflow(event, 1, nmi, data, regs);
5126 if (local64_add_negative(nr, &hwc->period_left))
5129 perf_swevent_overflow(event, 0, nmi, data, regs);
5132 static int perf_exclude_event(struct perf_event *event,
5133 struct pt_regs *regs)
5135 if (event->hw.state & PERF_HES_STOPPED)
5139 if (event->attr.exclude_user && user_mode(regs))
5142 if (event->attr.exclude_kernel && !user_mode(regs))
5149 static int perf_swevent_match(struct perf_event *event,
5150 enum perf_type_id type,
5152 struct perf_sample_data *data,
5153 struct pt_regs *regs)
5155 if (event->attr.type != type)
5158 if (event->attr.config != event_id)
5161 if (perf_exclude_event(event, regs))
5167 static inline u64 swevent_hash(u64 type, u32 event_id)
5169 u64 val = event_id | (type << 32);
5171 return hash_64(val, SWEVENT_HLIST_BITS);
5174 static inline struct hlist_head *
5175 __find_swevent_head(struct swevent_hlist *hlist, u64 type, u32 event_id)
5177 u64 hash = swevent_hash(type, event_id);
5179 return &hlist->heads[hash];
5182 /* For the read side: events when they trigger */
5183 static inline struct hlist_head *
5184 find_swevent_head_rcu(struct swevent_htable *swhash, u64 type, u32 event_id)
5186 struct swevent_hlist *hlist;
5188 hlist = rcu_dereference(swhash->swevent_hlist);
5192 return __find_swevent_head(hlist, type, event_id);
5195 /* For the event head insertion and removal in the hlist */
5196 static inline struct hlist_head *
5197 find_swevent_head(struct swevent_htable *swhash, struct perf_event *event)
5199 struct swevent_hlist *hlist;
5200 u32 event_id = event->attr.config;
5201 u64 type = event->attr.type;
5204 * Event scheduling is always serialized against hlist allocation
5205 * and release. Which makes the protected version suitable here.
5206 * The context lock guarantees that.
5208 hlist = rcu_dereference_protected(swhash->swevent_hlist,
5209 lockdep_is_held(&event->ctx->lock));
5213 return __find_swevent_head(hlist, type, event_id);
5216 static void do_perf_sw_event(enum perf_type_id type, u32 event_id,
5218 struct perf_sample_data *data,
5219 struct pt_regs *regs)
5221 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
5222 struct perf_event *event;
5223 struct hlist_node *node;
5224 struct hlist_head *head;
5227 head = find_swevent_head_rcu(swhash, type, event_id);
5231 hlist_for_each_entry_rcu(event, node, head, hlist_entry) {
5232 if (perf_swevent_match(event, type, event_id, data, regs))
5233 perf_swevent_event(event, nr, nmi, data, regs);
5239 int perf_swevent_get_recursion_context(void)
5241 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
5243 return get_recursion_context(swhash->recursion);
5245 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context);
5247 inline void perf_swevent_put_recursion_context(int rctx)
5249 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
5251 put_recursion_context(swhash->recursion, rctx);
5254 void __perf_sw_event(u32 event_id, u64 nr, int nmi,
5255 struct pt_regs *regs, u64 addr)
5257 struct perf_sample_data data;
5260 preempt_disable_notrace();
5261 rctx = perf_swevent_get_recursion_context();
5265 perf_sample_data_init(&data, addr);
5267 do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, nmi, &data, regs);
5269 perf_swevent_put_recursion_context(rctx);
5270 preempt_enable_notrace();
5273 static void perf_swevent_read(struct perf_event *event)
5277 static int perf_swevent_add(struct perf_event *event, int flags)
5279 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
5280 struct hw_perf_event *hwc = &event->hw;
5281 struct hlist_head *head;
5283 if (is_sampling_event(event)) {
5284 hwc->last_period = hwc->sample_period;
5285 perf_swevent_set_period(event);
5288 hwc->state = !(flags & PERF_EF_START);
5290 head = find_swevent_head(swhash, event);
5291 if (WARN_ON_ONCE(!head))
5294 hlist_add_head_rcu(&event->hlist_entry, head);
5299 static void perf_swevent_del(struct perf_event *event, int flags)
5301 hlist_del_rcu(&event->hlist_entry);
5304 static void perf_swevent_start(struct perf_event *event, int flags)
5306 event->hw.state = 0;
5309 static void perf_swevent_stop(struct perf_event *event, int flags)
5311 event->hw.state = PERF_HES_STOPPED;
5314 /* Deref the hlist from the update side */
5315 static inline struct swevent_hlist *
5316 swevent_hlist_deref(struct swevent_htable *swhash)
5318 return rcu_dereference_protected(swhash->swevent_hlist,
5319 lockdep_is_held(&swhash->hlist_mutex));
5322 static void swevent_hlist_release_rcu(struct rcu_head *rcu_head)
5324 struct swevent_hlist *hlist;
5326 hlist = container_of(rcu_head, struct swevent_hlist, rcu_head);
5330 static void swevent_hlist_release(struct swevent_htable *swhash)
5332 struct swevent_hlist *hlist = swevent_hlist_deref(swhash);
5337 rcu_assign_pointer(swhash->swevent_hlist, NULL);
5338 call_rcu(&hlist->rcu_head, swevent_hlist_release_rcu);
5341 static void swevent_hlist_put_cpu(struct perf_event *event, int cpu)
5343 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
5345 mutex_lock(&swhash->hlist_mutex);
5347 if (!--swhash->hlist_refcount)
5348 swevent_hlist_release(swhash);
5350 mutex_unlock(&swhash->hlist_mutex);
5353 static void swevent_hlist_put(struct perf_event *event)
5357 if (event->cpu != -1) {
5358 swevent_hlist_put_cpu(event, event->cpu);
5362 for_each_possible_cpu(cpu)
5363 swevent_hlist_put_cpu(event, cpu);
5366 static int swevent_hlist_get_cpu(struct perf_event *event, int cpu)
5368 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
5371 mutex_lock(&swhash->hlist_mutex);
5373 if (!swevent_hlist_deref(swhash) && cpu_online(cpu)) {
5374 struct swevent_hlist *hlist;
5376 hlist = kzalloc(sizeof(*hlist), GFP_KERNEL);
5381 rcu_assign_pointer(swhash->swevent_hlist, hlist);
5383 swhash->hlist_refcount++;
5385 mutex_unlock(&swhash->hlist_mutex);
5390 static int swevent_hlist_get(struct perf_event *event)
5393 int cpu, failed_cpu;
5395 if (event->cpu != -1)
5396 return swevent_hlist_get_cpu(event, event->cpu);
5399 for_each_possible_cpu(cpu) {
5400 err = swevent_hlist_get_cpu(event, cpu);
5410 for_each_possible_cpu(cpu) {
5411 if (cpu == failed_cpu)
5413 swevent_hlist_put_cpu(event, cpu);
5420 atomic_t perf_swevent_enabled[PERF_COUNT_SW_MAX];
5422 static void sw_perf_event_destroy(struct perf_event *event)
5424 u64 event_id = event->attr.config;
5426 WARN_ON(event->parent);
5428 jump_label_dec(&perf_swevent_enabled[event_id]);
5429 swevent_hlist_put(event);
5432 static int perf_swevent_init(struct perf_event *event)
5434 int event_id = event->attr.config;
5436 if (event->attr.type != PERF_TYPE_SOFTWARE)
5440 case PERF_COUNT_SW_CPU_CLOCK:
5441 case PERF_COUNT_SW_TASK_CLOCK:
5448 if (event_id >= PERF_COUNT_SW_MAX)
5451 if (!event->parent) {
5454 err = swevent_hlist_get(event);
5458 jump_label_inc(&perf_swevent_enabled[event_id]);
5459 event->destroy = sw_perf_event_destroy;
5465 static struct pmu perf_swevent = {
5466 .task_ctx_nr = perf_sw_context,
5468 .event_init = perf_swevent_init,
5469 .add = perf_swevent_add,
5470 .del = perf_swevent_del,
5471 .start = perf_swevent_start,
5472 .stop = perf_swevent_stop,
5473 .read = perf_swevent_read,
5476 #ifdef CONFIG_EVENT_TRACING
5478 static int perf_tp_filter_match(struct perf_event *event,
5479 struct perf_sample_data *data)
5481 void *record = data->raw->data;
5483 if (likely(!event->filter) || filter_match_preds(event->filter, record))
5488 static int perf_tp_event_match(struct perf_event *event,
5489 struct perf_sample_data *data,
5490 struct pt_regs *regs)
5492 if (event->hw.state & PERF_HES_STOPPED)
5495 * All tracepoints are from kernel-space.
5497 if (event->attr.exclude_kernel)
5500 if (!perf_tp_filter_match(event, data))
5506 void perf_tp_event(u64 addr, u64 count, void *record, int entry_size,
5507 struct pt_regs *regs, struct hlist_head *head, int rctx)
5509 struct perf_sample_data data;
5510 struct perf_event *event;
5511 struct hlist_node *node;
5513 struct perf_raw_record raw = {
5518 perf_sample_data_init(&data, addr);
5521 hlist_for_each_entry_rcu(event, node, head, hlist_entry) {
5522 if (perf_tp_event_match(event, &data, regs))
5523 perf_swevent_event(event, count, 1, &data, regs);
5526 perf_swevent_put_recursion_context(rctx);
5528 EXPORT_SYMBOL_GPL(perf_tp_event);
5530 static void tp_perf_event_destroy(struct perf_event *event)
5532 perf_trace_destroy(event);
5535 static int perf_tp_event_init(struct perf_event *event)
5539 if (event->attr.type != PERF_TYPE_TRACEPOINT)
5542 err = perf_trace_init(event);
5546 event->destroy = tp_perf_event_destroy;
5551 static struct pmu perf_tracepoint = {
5552 .task_ctx_nr = perf_sw_context,
5554 .event_init = perf_tp_event_init,
5555 .add = perf_trace_add,
5556 .del = perf_trace_del,
5557 .start = perf_swevent_start,
5558 .stop = perf_swevent_stop,
5559 .read = perf_swevent_read,
5562 static inline void perf_tp_register(void)
5564 perf_pmu_register(&perf_tracepoint, "tracepoint", PERF_TYPE_TRACEPOINT);
5567 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
5572 if (event->attr.type != PERF_TYPE_TRACEPOINT)
5575 filter_str = strndup_user(arg, PAGE_SIZE);
5576 if (IS_ERR(filter_str))
5577 return PTR_ERR(filter_str);
5579 ret = ftrace_profile_set_filter(event, event->attr.config, filter_str);
5585 static void perf_event_free_filter(struct perf_event *event)
5587 ftrace_profile_free_filter(event);
5592 static inline void perf_tp_register(void)
5596 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
5601 static void perf_event_free_filter(struct perf_event *event)
5605 #endif /* CONFIG_EVENT_TRACING */
5607 #ifdef CONFIG_HAVE_HW_BREAKPOINT
5608 void perf_bp_event(struct perf_event *bp, void *data)
5610 struct perf_sample_data sample;
5611 struct pt_regs *regs = data;
5613 perf_sample_data_init(&sample, bp->attr.bp_addr);
5615 if (!bp->hw.state && !perf_exclude_event(bp, regs))
5616 perf_swevent_event(bp, 1, 1, &sample, regs);
5621 * hrtimer based swevent callback
5624 static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer)
5626 enum hrtimer_restart ret = HRTIMER_RESTART;
5627 struct perf_sample_data data;
5628 struct pt_regs *regs;
5629 struct perf_event *event;
5632 event = container_of(hrtimer, struct perf_event, hw.hrtimer);
5634 if (event->state != PERF_EVENT_STATE_ACTIVE)
5635 return HRTIMER_NORESTART;
5637 event->pmu->read(event);
5639 perf_sample_data_init(&data, 0);
5640 data.period = event->hw.last_period;
5641 regs = get_irq_regs();
5643 if (regs && !perf_exclude_event(event, regs)) {
5644 if (!(event->attr.exclude_idle && current->pid == 0))
5645 if (perf_event_overflow(event, 0, &data, regs))
5646 ret = HRTIMER_NORESTART;
5649 period = max_t(u64, 10000, event->hw.sample_period);
5650 hrtimer_forward_now(hrtimer, ns_to_ktime(period));
5655 static void perf_swevent_start_hrtimer(struct perf_event *event)
5657 struct hw_perf_event *hwc = &event->hw;
5660 if (!is_sampling_event(event))
5663 period = local64_read(&hwc->period_left);
5668 local64_set(&hwc->period_left, 0);
5670 period = max_t(u64, 10000, hwc->sample_period);
5672 __hrtimer_start_range_ns(&hwc->hrtimer,
5673 ns_to_ktime(period), 0,
5674 HRTIMER_MODE_REL_PINNED, 0);
5677 static void perf_swevent_cancel_hrtimer(struct perf_event *event)
5679 struct hw_perf_event *hwc = &event->hw;
5681 if (is_sampling_event(event)) {
5682 ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer);
5683 local64_set(&hwc->period_left, ktime_to_ns(remaining));
5685 hrtimer_cancel(&hwc->hrtimer);
5689 static void perf_swevent_init_hrtimer(struct perf_event *event)
5691 struct hw_perf_event *hwc = &event->hw;
5693 if (!is_sampling_event(event))
5696 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
5697 hwc->hrtimer.function = perf_swevent_hrtimer;
5700 * Since hrtimers have a fixed rate, we can do a static freq->period
5701 * mapping and avoid the whole period adjust feedback stuff.
5703 if (event->attr.freq) {
5704 long freq = event->attr.sample_freq;
5706 event->attr.sample_period = NSEC_PER_SEC / freq;
5707 hwc->sample_period = event->attr.sample_period;
5708 local64_set(&hwc->period_left, hwc->sample_period);
5709 event->attr.freq = 0;
5714 * Software event: cpu wall time clock
5717 static void cpu_clock_event_update(struct perf_event *event)
5722 now = local_clock();
5723 prev = local64_xchg(&event->hw.prev_count, now);
5724 local64_add(now - prev, &event->count);
5727 static void cpu_clock_event_start(struct perf_event *event, int flags)
5729 local64_set(&event->hw.prev_count, local_clock());
5730 perf_swevent_start_hrtimer(event);
5733 static void cpu_clock_event_stop(struct perf_event *event, int flags)
5735 perf_swevent_cancel_hrtimer(event);
5736 cpu_clock_event_update(event);
5739 static int cpu_clock_event_add(struct perf_event *event, int flags)
5741 if (flags & PERF_EF_START)
5742 cpu_clock_event_start(event, flags);
5747 static void cpu_clock_event_del(struct perf_event *event, int flags)
5749 cpu_clock_event_stop(event, flags);
5752 static void cpu_clock_event_read(struct perf_event *event)
5754 cpu_clock_event_update(event);
5757 static int cpu_clock_event_init(struct perf_event *event)
5759 if (event->attr.type != PERF_TYPE_SOFTWARE)
5762 if (event->attr.config != PERF_COUNT_SW_CPU_CLOCK)
5765 perf_swevent_init_hrtimer(event);
5770 static struct pmu perf_cpu_clock = {
5771 .task_ctx_nr = perf_sw_context,
5773 .event_init = cpu_clock_event_init,
5774 .add = cpu_clock_event_add,
5775 .del = cpu_clock_event_del,
5776 .start = cpu_clock_event_start,
5777 .stop = cpu_clock_event_stop,
5778 .read = cpu_clock_event_read,
5782 * Software event: task time clock
5785 static void task_clock_event_update(struct perf_event *event, u64 now)
5790 prev = local64_xchg(&event->hw.prev_count, now);
5792 local64_add(delta, &event->count);
5795 static void task_clock_event_start(struct perf_event *event, int flags)
5797 local64_set(&event->hw.prev_count, event->ctx->time);
5798 perf_swevent_start_hrtimer(event);
5801 static void task_clock_event_stop(struct perf_event *event, int flags)
5803 perf_swevent_cancel_hrtimer(event);
5804 task_clock_event_update(event, event->ctx->time);
5807 static int task_clock_event_add(struct perf_event *event, int flags)
5809 if (flags & PERF_EF_START)
5810 task_clock_event_start(event, flags);
5815 static void task_clock_event_del(struct perf_event *event, int flags)
5817 task_clock_event_stop(event, PERF_EF_UPDATE);
5820 static void task_clock_event_read(struct perf_event *event)
5822 u64 now = perf_clock();
5823 u64 delta = now - event->ctx->timestamp;
5824 u64 time = event->ctx->time + delta;
5826 task_clock_event_update(event, time);
5829 static int task_clock_event_init(struct perf_event *event)
5831 if (event->attr.type != PERF_TYPE_SOFTWARE)
5834 if (event->attr.config != PERF_COUNT_SW_TASK_CLOCK)
5837 perf_swevent_init_hrtimer(event);
5842 static struct pmu perf_task_clock = {
5843 .task_ctx_nr = perf_sw_context,
5845 .event_init = task_clock_event_init,
5846 .add = task_clock_event_add,
5847 .del = task_clock_event_del,
5848 .start = task_clock_event_start,
5849 .stop = task_clock_event_stop,
5850 .read = task_clock_event_read,
5853 static void perf_pmu_nop_void(struct pmu *pmu)
5857 static int perf_pmu_nop_int(struct pmu *pmu)
5862 static void perf_pmu_start_txn(struct pmu *pmu)
5864 perf_pmu_disable(pmu);
5867 static int perf_pmu_commit_txn(struct pmu *pmu)
5869 perf_pmu_enable(pmu);
5873 static void perf_pmu_cancel_txn(struct pmu *pmu)
5875 perf_pmu_enable(pmu);
5879 * Ensures all contexts with the same task_ctx_nr have the same
5880 * pmu_cpu_context too.
5882 static void *find_pmu_context(int ctxn)
5889 list_for_each_entry(pmu, &pmus, entry) {
5890 if (pmu->task_ctx_nr == ctxn)
5891 return pmu->pmu_cpu_context;
5897 static void update_pmu_context(struct pmu *pmu, struct pmu *old_pmu)
5901 for_each_possible_cpu(cpu) {
5902 struct perf_cpu_context *cpuctx;
5904 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
5906 if (cpuctx->active_pmu == old_pmu)
5907 cpuctx->active_pmu = pmu;
5911 static void free_pmu_context(struct pmu *pmu)
5915 mutex_lock(&pmus_lock);
5917 * Like a real lame refcount.
5919 list_for_each_entry(i, &pmus, entry) {
5920 if (i->pmu_cpu_context == pmu->pmu_cpu_context) {
5921 update_pmu_context(i, pmu);
5926 free_percpu(pmu->pmu_cpu_context);
5928 mutex_unlock(&pmus_lock);
5930 static struct idr pmu_idr;
5933 type_show(struct device *dev, struct device_attribute *attr, char *page)
5935 struct pmu *pmu = dev_get_drvdata(dev);
5937 return snprintf(page, PAGE_SIZE-1, "%d\n", pmu->type);
5940 static struct device_attribute pmu_dev_attrs[] = {
5945 static int pmu_bus_running;
5946 static struct bus_type pmu_bus = {
5947 .name = "event_source",
5948 .dev_attrs = pmu_dev_attrs,
5951 static void pmu_dev_release(struct device *dev)
5956 static int pmu_dev_alloc(struct pmu *pmu)
5960 pmu->dev = kzalloc(sizeof(struct device), GFP_KERNEL);
5964 device_initialize(pmu->dev);
5965 ret = dev_set_name(pmu->dev, "%s", pmu->name);
5969 dev_set_drvdata(pmu->dev, pmu);
5970 pmu->dev->bus = &pmu_bus;
5971 pmu->dev->release = pmu_dev_release;
5972 ret = device_add(pmu->dev);
5980 put_device(pmu->dev);
5984 static struct lock_class_key cpuctx_mutex;
5986 int perf_pmu_register(struct pmu *pmu, char *name, int type)
5990 mutex_lock(&pmus_lock);
5992 pmu->pmu_disable_count = alloc_percpu(int);
5993 if (!pmu->pmu_disable_count)
6002 int err = idr_pre_get(&pmu_idr, GFP_KERNEL);
6006 err = idr_get_new_above(&pmu_idr, pmu, PERF_TYPE_MAX, &type);
6014 if (pmu_bus_running) {
6015 ret = pmu_dev_alloc(pmu);
6021 pmu->pmu_cpu_context = find_pmu_context(pmu->task_ctx_nr);
6022 if (pmu->pmu_cpu_context)
6023 goto got_cpu_context;
6025 pmu->pmu_cpu_context = alloc_percpu(struct perf_cpu_context);
6026 if (!pmu->pmu_cpu_context)
6029 for_each_possible_cpu(cpu) {
6030 struct perf_cpu_context *cpuctx;
6032 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
6033 __perf_event_init_context(&cpuctx->ctx);
6034 lockdep_set_class(&cpuctx->ctx.mutex, &cpuctx_mutex);
6035 cpuctx->ctx.type = cpu_context;
6036 cpuctx->ctx.pmu = pmu;
6037 cpuctx->jiffies_interval = 1;
6038 INIT_LIST_HEAD(&cpuctx->rotation_list);
6039 cpuctx->active_pmu = pmu;
6043 if (!pmu->start_txn) {
6044 if (pmu->pmu_enable) {
6046 * If we have pmu_enable/pmu_disable calls, install
6047 * transaction stubs that use that to try and batch
6048 * hardware accesses.
6050 pmu->start_txn = perf_pmu_start_txn;
6051 pmu->commit_txn = perf_pmu_commit_txn;
6052 pmu->cancel_txn = perf_pmu_cancel_txn;
6054 pmu->start_txn = perf_pmu_nop_void;
6055 pmu->commit_txn = perf_pmu_nop_int;
6056 pmu->cancel_txn = perf_pmu_nop_void;
6060 if (!pmu->pmu_enable) {
6061 pmu->pmu_enable = perf_pmu_nop_void;
6062 pmu->pmu_disable = perf_pmu_nop_void;
6065 list_add_rcu(&pmu->entry, &pmus);
6068 mutex_unlock(&pmus_lock);
6073 device_del(pmu->dev);
6074 put_device(pmu->dev);
6077 if (pmu->type >= PERF_TYPE_MAX)
6078 idr_remove(&pmu_idr, pmu->type);
6081 free_percpu(pmu->pmu_disable_count);
6085 void perf_pmu_unregister(struct pmu *pmu)
6087 mutex_lock(&pmus_lock);
6088 list_del_rcu(&pmu->entry);
6089 mutex_unlock(&pmus_lock);
6092 * We dereference the pmu list under both SRCU and regular RCU, so
6093 * synchronize against both of those.
6095 synchronize_srcu(&pmus_srcu);
6098 free_percpu(pmu->pmu_disable_count);
6099 if (pmu->type >= PERF_TYPE_MAX)
6100 idr_remove(&pmu_idr, pmu->type);
6101 device_del(pmu->dev);
6102 put_device(pmu->dev);
6103 free_pmu_context(pmu);
6106 struct pmu *perf_init_event(struct perf_event *event)
6108 struct pmu *pmu = NULL;
6112 idx = srcu_read_lock(&pmus_srcu);
6115 pmu = idr_find(&pmu_idr, event->attr.type);
6118 ret = pmu->event_init(event);
6124 list_for_each_entry_rcu(pmu, &pmus, entry) {
6125 ret = pmu->event_init(event);
6129 if (ret != -ENOENT) {
6134 pmu = ERR_PTR(-ENOENT);
6136 srcu_read_unlock(&pmus_srcu, idx);
6142 * Allocate and initialize a event structure
6144 static struct perf_event *
6145 perf_event_alloc(struct perf_event_attr *attr, int cpu,
6146 struct task_struct *task,
6147 struct perf_event *group_leader,
6148 struct perf_event *parent_event,
6149 perf_overflow_handler_t overflow_handler)
6152 struct perf_event *event;
6153 struct hw_perf_event *hwc;
6156 if ((unsigned)cpu >= nr_cpu_ids) {
6157 if (!task || cpu != -1)
6158 return ERR_PTR(-EINVAL);
6161 event = kzalloc(sizeof(*event), GFP_KERNEL);
6163 return ERR_PTR(-ENOMEM);
6166 * Single events are their own group leaders, with an
6167 * empty sibling list:
6170 group_leader = event;
6172 mutex_init(&event->child_mutex);
6173 INIT_LIST_HEAD(&event->child_list);
6175 INIT_LIST_HEAD(&event->group_entry);
6176 INIT_LIST_HEAD(&event->event_entry);
6177 INIT_LIST_HEAD(&event->sibling_list);
6178 init_waitqueue_head(&event->waitq);
6179 init_irq_work(&event->pending, perf_pending_event);
6181 mutex_init(&event->mmap_mutex);
6184 event->attr = *attr;
6185 event->group_leader = group_leader;
6189 event->parent = parent_event;
6191 event->ns = get_pid_ns(current->nsproxy->pid_ns);
6192 event->id = atomic64_inc_return(&perf_event_id);
6194 event->state = PERF_EVENT_STATE_INACTIVE;
6197 event->attach_state = PERF_ATTACH_TASK;
6198 #ifdef CONFIG_HAVE_HW_BREAKPOINT
6200 * hw_breakpoint is a bit difficult here..
6202 if (attr->type == PERF_TYPE_BREAKPOINT)
6203 event->hw.bp_target = task;
6207 if (!overflow_handler && parent_event)
6208 overflow_handler = parent_event->overflow_handler;
6210 event->overflow_handler = overflow_handler;
6213 event->state = PERF_EVENT_STATE_OFF;
6218 hwc->sample_period = attr->sample_period;
6219 if (attr->freq && attr->sample_freq)
6220 hwc->sample_period = 1;
6221 hwc->last_period = hwc->sample_period;
6223 local64_set(&hwc->period_left, hwc->sample_period);
6226 * we currently do not support PERF_FORMAT_GROUP on inherited events
6228 if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP))
6231 pmu = perf_init_event(event);
6237 else if (IS_ERR(pmu))
6242 put_pid_ns(event->ns);
6244 return ERR_PTR(err);
6249 if (!event->parent) {
6250 if (event->attach_state & PERF_ATTACH_TASK)
6251 jump_label_inc(&perf_sched_events);
6252 if (event->attr.mmap || event->attr.mmap_data)
6253 atomic_inc(&nr_mmap_events);
6254 if (event->attr.comm)
6255 atomic_inc(&nr_comm_events);
6256 if (event->attr.task)
6257 atomic_inc(&nr_task_events);
6258 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) {
6259 err = get_callchain_buffers();
6262 return ERR_PTR(err);
6270 static int perf_copy_attr(struct perf_event_attr __user *uattr,
6271 struct perf_event_attr *attr)
6276 if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0))
6280 * zero the full structure, so that a short copy will be nice.
6282 memset(attr, 0, sizeof(*attr));
6284 ret = get_user(size, &uattr->size);
6288 if (size > PAGE_SIZE) /* silly large */
6291 if (!size) /* abi compat */
6292 size = PERF_ATTR_SIZE_VER0;
6294 if (size < PERF_ATTR_SIZE_VER0)
6298 * If we're handed a bigger struct than we know of,
6299 * ensure all the unknown bits are 0 - i.e. new
6300 * user-space does not rely on any kernel feature
6301 * extensions we dont know about yet.
6303 if (size > sizeof(*attr)) {
6304 unsigned char __user *addr;
6305 unsigned char __user *end;
6308 addr = (void __user *)uattr + sizeof(*attr);
6309 end = (void __user *)uattr + size;
6311 for (; addr < end; addr++) {
6312 ret = get_user(val, addr);
6318 size = sizeof(*attr);
6321 ret = copy_from_user(attr, uattr, size);
6326 * If the type exists, the corresponding creation will verify
6329 if (attr->type >= PERF_TYPE_MAX)
6332 if (attr->__reserved_1)
6335 if (attr->sample_type & ~(PERF_SAMPLE_MAX-1))
6338 if (attr->read_format & ~(PERF_FORMAT_MAX-1))
6345 put_user(sizeof(*attr), &uattr->size);
6351 perf_event_set_output(struct perf_event *event, struct perf_event *output_event)
6353 struct perf_buffer *buffer = NULL, *old_buffer = NULL;
6359 /* don't allow circular references */
6360 if (event == output_event)
6364 * Don't allow cross-cpu buffers
6366 if (output_event->cpu != event->cpu)
6370 * If its not a per-cpu buffer, it must be the same task.
6372 if (output_event->cpu == -1 && output_event->ctx != event->ctx)
6376 mutex_lock(&event->mmap_mutex);
6377 /* Can't redirect output if we've got an active mmap() */
6378 if (atomic_read(&event->mmap_count))
6382 /* get the buffer we want to redirect to */
6383 buffer = perf_buffer_get(output_event);
6388 old_buffer = event->buffer;
6389 rcu_assign_pointer(event->buffer, buffer);
6392 mutex_unlock(&event->mmap_mutex);
6395 perf_buffer_put(old_buffer);
6401 * sys_perf_event_open - open a performance event, associate it to a task/cpu
6403 * @attr_uptr: event_id type attributes for monitoring/sampling
6406 * @group_fd: group leader event fd
6408 SYSCALL_DEFINE5(perf_event_open,
6409 struct perf_event_attr __user *, attr_uptr,
6410 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
6412 struct perf_event *group_leader = NULL, *output_event = NULL;
6413 struct perf_event *event, *sibling;
6414 struct perf_event_attr attr;
6415 struct perf_event_context *ctx;
6416 struct file *event_file = NULL;
6417 struct file *group_file = NULL;
6418 struct task_struct *task = NULL;
6422 int fput_needed = 0;
6425 /* for future expandability... */
6426 if (flags & ~PERF_FLAG_ALL)
6429 err = perf_copy_attr(attr_uptr, &attr);
6433 if (!attr.exclude_kernel) {
6434 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
6439 if (attr.sample_freq > sysctl_perf_event_sample_rate)
6444 * In cgroup mode, the pid argument is used to pass the fd
6445 * opened to the cgroup directory in cgroupfs. The cpu argument
6446 * designates the cpu on which to monitor threads from that
6449 if ((flags & PERF_FLAG_PID_CGROUP) && (pid == -1 || cpu == -1))
6452 event_fd = get_unused_fd_flags(O_RDWR);
6456 if (group_fd != -1) {
6457 group_leader = perf_fget_light(group_fd, &fput_needed);
6458 if (IS_ERR(group_leader)) {
6459 err = PTR_ERR(group_leader);
6462 group_file = group_leader->filp;
6463 if (flags & PERF_FLAG_FD_OUTPUT)
6464 output_event = group_leader;
6465 if (flags & PERF_FLAG_FD_NO_GROUP)
6466 group_leader = NULL;
6469 if (pid != -1 && !(flags & PERF_FLAG_PID_CGROUP)) {
6470 task = find_lively_task_by_vpid(pid);
6472 err = PTR_ERR(task);
6477 event = perf_event_alloc(&attr, cpu, task, group_leader, NULL, NULL);
6478 if (IS_ERR(event)) {
6479 err = PTR_ERR(event);
6483 if (flags & PERF_FLAG_PID_CGROUP) {
6484 err = perf_cgroup_connect(pid, event, &attr, group_leader);
6489 * - that has cgroup constraint on event->cpu
6490 * - that may need work on context switch
6492 atomic_inc(&per_cpu(perf_cgroup_events, event->cpu));
6493 jump_label_inc(&perf_sched_events);
6497 * Special case software events and allow them to be part of
6498 * any hardware group.
6503 (is_software_event(event) != is_software_event(group_leader))) {
6504 if (is_software_event(event)) {
6506 * If event and group_leader are not both a software
6507 * event, and event is, then group leader is not.
6509 * Allow the addition of software events to !software
6510 * groups, this is safe because software events never
6513 pmu = group_leader->pmu;
6514 } else if (is_software_event(group_leader) &&
6515 (group_leader->group_flags & PERF_GROUP_SOFTWARE)) {
6517 * In case the group is a pure software group, and we
6518 * try to add a hardware event, move the whole group to
6519 * the hardware context.
6526 * Get the target context (task or percpu):
6528 ctx = find_get_context(pmu, task, cpu);
6535 put_task_struct(task);
6540 * Look up the group leader (we will attach this event to it):
6546 * Do not allow a recursive hierarchy (this new sibling
6547 * becoming part of another group-sibling):
6549 if (group_leader->group_leader != group_leader)
6552 * Do not allow to attach to a group in a different
6553 * task or CPU context:
6556 if (group_leader->ctx->type != ctx->type)
6559 if (group_leader->ctx != ctx)
6564 * Only a group leader can be exclusive or pinned
6566 if (attr.exclusive || attr.pinned)
6571 err = perf_event_set_output(event, output_event);
6576 event_file = anon_inode_getfile("[perf_event]", &perf_fops, event, O_RDWR);
6577 if (IS_ERR(event_file)) {
6578 err = PTR_ERR(event_file);
6583 struct perf_event_context *gctx = group_leader->ctx;
6585 mutex_lock(&gctx->mutex);
6586 perf_remove_from_context(group_leader);
6587 list_for_each_entry(sibling, &group_leader->sibling_list,
6589 perf_remove_from_context(sibling);
6592 mutex_unlock(&gctx->mutex);
6596 event->filp = event_file;
6597 WARN_ON_ONCE(ctx->parent_ctx);
6598 mutex_lock(&ctx->mutex);
6601 perf_install_in_context(ctx, group_leader, cpu);
6603 list_for_each_entry(sibling, &group_leader->sibling_list,
6605 perf_install_in_context(ctx, sibling, cpu);
6610 perf_install_in_context(ctx, event, cpu);
6612 perf_unpin_context(ctx);
6613 mutex_unlock(&ctx->mutex);
6615 event->owner = current;
6617 mutex_lock(¤t->perf_event_mutex);
6618 list_add_tail(&event->owner_entry, ¤t->perf_event_list);
6619 mutex_unlock(¤t->perf_event_mutex);
6622 * Precalculate sample_data sizes
6624 perf_event__header_size(event);
6625 perf_event__id_header_size(event);
6628 * Drop the reference on the group_event after placing the
6629 * new event on the sibling_list. This ensures destruction
6630 * of the group leader will find the pointer to itself in
6631 * perf_group_detach().
6633 fput_light(group_file, fput_needed);
6634 fd_install(event_fd, event_file);
6638 perf_unpin_context(ctx);
6644 put_task_struct(task);
6646 fput_light(group_file, fput_needed);
6648 put_unused_fd(event_fd);
6653 * perf_event_create_kernel_counter
6655 * @attr: attributes of the counter to create
6656 * @cpu: cpu in which the counter is bound
6657 * @task: task to profile (NULL for percpu)
6660 perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu,
6661 struct task_struct *task,
6662 perf_overflow_handler_t overflow_handler)
6664 struct perf_event_context *ctx;
6665 struct perf_event *event;
6669 * Get the target context (task or percpu):
6672 event = perf_event_alloc(attr, cpu, task, NULL, NULL, overflow_handler);
6673 if (IS_ERR(event)) {
6674 err = PTR_ERR(event);
6678 ctx = find_get_context(event->pmu, task, cpu);
6685 WARN_ON_ONCE(ctx->parent_ctx);
6686 mutex_lock(&ctx->mutex);
6687 perf_install_in_context(ctx, event, cpu);
6689 perf_unpin_context(ctx);
6690 mutex_unlock(&ctx->mutex);
6697 return ERR_PTR(err);
6699 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter);
6701 static void sync_child_event(struct perf_event *child_event,
6702 struct task_struct *child)
6704 struct perf_event *parent_event = child_event->parent;
6707 if (child_event->attr.inherit_stat)
6708 perf_event_read_event(child_event, child);
6710 child_val = perf_event_count(child_event);
6713 * Add back the child's count to the parent's count:
6715 atomic64_add(child_val, &parent_event->child_count);
6716 atomic64_add(child_event->total_time_enabled,
6717 &parent_event->child_total_time_enabled);
6718 atomic64_add(child_event->total_time_running,
6719 &parent_event->child_total_time_running);
6722 * Remove this event from the parent's list
6724 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
6725 mutex_lock(&parent_event->child_mutex);
6726 list_del_init(&child_event->child_list);
6727 mutex_unlock(&parent_event->child_mutex);
6730 * Release the parent event, if this was the last
6733 fput(parent_event->filp);
6737 __perf_event_exit_task(struct perf_event *child_event,
6738 struct perf_event_context *child_ctx,
6739 struct task_struct *child)
6741 if (child_event->parent) {
6742 raw_spin_lock_irq(&child_ctx->lock);
6743 perf_group_detach(child_event);
6744 raw_spin_unlock_irq(&child_ctx->lock);
6747 perf_remove_from_context(child_event);
6750 * It can happen that the parent exits first, and has events
6751 * that are still around due to the child reference. These
6752 * events need to be zapped.
6754 if (child_event->parent) {
6755 sync_child_event(child_event, child);
6756 free_event(child_event);
6760 static void perf_event_exit_task_context(struct task_struct *child, int ctxn)
6762 struct perf_event *child_event, *tmp;
6763 struct perf_event_context *child_ctx;
6764 unsigned long flags;
6766 if (likely(!child->perf_event_ctxp[ctxn])) {
6767 perf_event_task(child, NULL, 0);
6771 local_irq_save(flags);
6773 * We can't reschedule here because interrupts are disabled,
6774 * and either child is current or it is a task that can't be
6775 * scheduled, so we are now safe from rescheduling changing
6778 child_ctx = rcu_dereference_raw(child->perf_event_ctxp[ctxn]);
6779 task_ctx_sched_out(child_ctx, EVENT_ALL);
6782 * Take the context lock here so that if find_get_context is
6783 * reading child->perf_event_ctxp, we wait until it has
6784 * incremented the context's refcount before we do put_ctx below.
6786 raw_spin_lock(&child_ctx->lock);
6787 child->perf_event_ctxp[ctxn] = NULL;
6789 * If this context is a clone; unclone it so it can't get
6790 * swapped to another process while we're removing all
6791 * the events from it.
6793 unclone_ctx(child_ctx);
6794 update_context_time(child_ctx);
6795 raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
6798 * Report the task dead after unscheduling the events so that we
6799 * won't get any samples after PERF_RECORD_EXIT. We can however still
6800 * get a few PERF_RECORD_READ events.
6802 perf_event_task(child, child_ctx, 0);
6805 * We can recurse on the same lock type through:
6807 * __perf_event_exit_task()
6808 * sync_child_event()
6809 * fput(parent_event->filp)
6811 * mutex_lock(&ctx->mutex)
6813 * But since its the parent context it won't be the same instance.
6815 mutex_lock(&child_ctx->mutex);
6818 list_for_each_entry_safe(child_event, tmp, &child_ctx->pinned_groups,
6820 __perf_event_exit_task(child_event, child_ctx, child);
6822 list_for_each_entry_safe(child_event, tmp, &child_ctx->flexible_groups,
6824 __perf_event_exit_task(child_event, child_ctx, child);
6827 * If the last event was a group event, it will have appended all
6828 * its siblings to the list, but we obtained 'tmp' before that which
6829 * will still point to the list head terminating the iteration.
6831 if (!list_empty(&child_ctx->pinned_groups) ||
6832 !list_empty(&child_ctx->flexible_groups))
6835 mutex_unlock(&child_ctx->mutex);
6841 * When a child task exits, feed back event values to parent events.
6843 void perf_event_exit_task(struct task_struct *child)
6845 struct perf_event *event, *tmp;
6848 mutex_lock(&child->perf_event_mutex);
6849 list_for_each_entry_safe(event, tmp, &child->perf_event_list,
6851 list_del_init(&event->owner_entry);
6854 * Ensure the list deletion is visible before we clear
6855 * the owner, closes a race against perf_release() where
6856 * we need to serialize on the owner->perf_event_mutex.
6859 event->owner = NULL;
6861 mutex_unlock(&child->perf_event_mutex);
6863 for_each_task_context_nr(ctxn)
6864 perf_event_exit_task_context(child, ctxn);
6867 static void perf_free_event(struct perf_event *event,
6868 struct perf_event_context *ctx)
6870 struct perf_event *parent = event->parent;
6872 if (WARN_ON_ONCE(!parent))
6875 mutex_lock(&parent->child_mutex);
6876 list_del_init(&event->child_list);
6877 mutex_unlock(&parent->child_mutex);
6881 perf_group_detach(event);
6882 list_del_event(event, ctx);
6887 * free an unexposed, unused context as created by inheritance by
6888 * perf_event_init_task below, used by fork() in case of fail.
6890 void perf_event_free_task(struct task_struct *task)
6892 struct perf_event_context *ctx;
6893 struct perf_event *event, *tmp;
6896 for_each_task_context_nr(ctxn) {
6897 ctx = task->perf_event_ctxp[ctxn];
6901 mutex_lock(&ctx->mutex);
6903 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups,
6905 perf_free_event(event, ctx);
6907 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups,
6909 perf_free_event(event, ctx);
6911 if (!list_empty(&ctx->pinned_groups) ||
6912 !list_empty(&ctx->flexible_groups))
6915 mutex_unlock(&ctx->mutex);
6921 void perf_event_delayed_put(struct task_struct *task)
6925 for_each_task_context_nr(ctxn)
6926 WARN_ON_ONCE(task->perf_event_ctxp[ctxn]);
6930 * inherit a event from parent task to child task:
6932 static struct perf_event *
6933 inherit_event(struct perf_event *parent_event,
6934 struct task_struct *parent,
6935 struct perf_event_context *parent_ctx,
6936 struct task_struct *child,
6937 struct perf_event *group_leader,
6938 struct perf_event_context *child_ctx)
6940 struct perf_event *child_event;
6941 unsigned long flags;
6944 * Instead of creating recursive hierarchies of events,
6945 * we link inherited events back to the original parent,
6946 * which has a filp for sure, which we use as the reference
6949 if (parent_event->parent)
6950 parent_event = parent_event->parent;
6952 child_event = perf_event_alloc(&parent_event->attr,
6955 group_leader, parent_event,
6957 if (IS_ERR(child_event))
6962 * Make the child state follow the state of the parent event,
6963 * not its attr.disabled bit. We hold the parent's mutex,
6964 * so we won't race with perf_event_{en, dis}able_family.
6966 if (parent_event->state >= PERF_EVENT_STATE_INACTIVE)
6967 child_event->state = PERF_EVENT_STATE_INACTIVE;
6969 child_event->state = PERF_EVENT_STATE_OFF;
6971 if (parent_event->attr.freq) {
6972 u64 sample_period = parent_event->hw.sample_period;
6973 struct hw_perf_event *hwc = &child_event->hw;
6975 hwc->sample_period = sample_period;
6976 hwc->last_period = sample_period;
6978 local64_set(&hwc->period_left, sample_period);
6981 child_event->ctx = child_ctx;
6982 child_event->overflow_handler = parent_event->overflow_handler;
6985 * Precalculate sample_data sizes
6987 perf_event__header_size(child_event);
6988 perf_event__id_header_size(child_event);
6991 * Link it up in the child's context:
6993 raw_spin_lock_irqsave(&child_ctx->lock, flags);
6994 add_event_to_ctx(child_event, child_ctx);
6995 raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
6998 * Get a reference to the parent filp - we will fput it
6999 * when the child event exits. This is safe to do because
7000 * we are in the parent and we know that the filp still
7001 * exists and has a nonzero count:
7003 atomic_long_inc(&parent_event->filp->f_count);
7006 * Link this into the parent event's child list
7008 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
7009 mutex_lock(&parent_event->child_mutex);
7010 list_add_tail(&child_event->child_list, &parent_event->child_list);
7011 mutex_unlock(&parent_event->child_mutex);
7016 static int inherit_group(struct perf_event *parent_event,
7017 struct task_struct *parent,
7018 struct perf_event_context *parent_ctx,
7019 struct task_struct *child,
7020 struct perf_event_context *child_ctx)
7022 struct perf_event *leader;
7023 struct perf_event *sub;
7024 struct perf_event *child_ctr;
7026 leader = inherit_event(parent_event, parent, parent_ctx,
7027 child, NULL, child_ctx);
7029 return PTR_ERR(leader);
7030 list_for_each_entry(sub, &parent_event->sibling_list, group_entry) {
7031 child_ctr = inherit_event(sub, parent, parent_ctx,
7032 child, leader, child_ctx);
7033 if (IS_ERR(child_ctr))
7034 return PTR_ERR(child_ctr);
7040 inherit_task_group(struct perf_event *event, struct task_struct *parent,
7041 struct perf_event_context *parent_ctx,
7042 struct task_struct *child, int ctxn,
7046 struct perf_event_context *child_ctx;
7048 if (!event->attr.inherit) {
7053 child_ctx = child->perf_event_ctxp[ctxn];
7056 * This is executed from the parent task context, so
7057 * inherit events that have been marked for cloning.
7058 * First allocate and initialize a context for the
7062 child_ctx = alloc_perf_context(event->pmu, child);
7066 child->perf_event_ctxp[ctxn] = child_ctx;
7069 ret = inherit_group(event, parent, parent_ctx,
7079 * Initialize the perf_event context in task_struct
7081 int perf_event_init_context(struct task_struct *child, int ctxn)
7083 struct perf_event_context *child_ctx, *parent_ctx;
7084 struct perf_event_context *cloned_ctx;
7085 struct perf_event *event;
7086 struct task_struct *parent = current;
7087 int inherited_all = 1;
7088 unsigned long flags;
7091 if (likely(!parent->perf_event_ctxp[ctxn]))
7095 * If the parent's context is a clone, pin it so it won't get
7098 parent_ctx = perf_pin_task_context(parent, ctxn);
7101 * No need to check if parent_ctx != NULL here; since we saw
7102 * it non-NULL earlier, the only reason for it to become NULL
7103 * is if we exit, and since we're currently in the middle of
7104 * a fork we can't be exiting at the same time.
7108 * Lock the parent list. No need to lock the child - not PID
7109 * hashed yet and not running, so nobody can access it.
7111 mutex_lock(&parent_ctx->mutex);
7114 * We dont have to disable NMIs - we are only looking at
7115 * the list, not manipulating it:
7117 list_for_each_entry(event, &parent_ctx->pinned_groups, group_entry) {
7118 ret = inherit_task_group(event, parent, parent_ctx,
7119 child, ctxn, &inherited_all);
7125 * We can't hold ctx->lock when iterating the ->flexible_group list due
7126 * to allocations, but we need to prevent rotation because
7127 * rotate_ctx() will change the list from interrupt context.
7129 raw_spin_lock_irqsave(&parent_ctx->lock, flags);
7130 parent_ctx->rotate_disable = 1;
7131 raw_spin_unlock_irqrestore(&parent_ctx->lock, flags);
7133 list_for_each_entry(event, &parent_ctx->flexible_groups, group_entry) {
7134 ret = inherit_task_group(event, parent, parent_ctx,
7135 child, ctxn, &inherited_all);
7140 raw_spin_lock_irqsave(&parent_ctx->lock, flags);
7141 parent_ctx->rotate_disable = 0;
7143 child_ctx = child->perf_event_ctxp[ctxn];
7145 if (child_ctx && inherited_all) {
7147 * Mark the child context as a clone of the parent
7148 * context, or of whatever the parent is a clone of.
7150 * Note that if the parent is a clone, the holding of
7151 * parent_ctx->lock avoids it from being uncloned.
7153 cloned_ctx = parent_ctx->parent_ctx;
7155 child_ctx->parent_ctx = cloned_ctx;
7156 child_ctx->parent_gen = parent_ctx->parent_gen;
7158 child_ctx->parent_ctx = parent_ctx;
7159 child_ctx->parent_gen = parent_ctx->generation;
7161 get_ctx(child_ctx->parent_ctx);
7164 raw_spin_unlock_irqrestore(&parent_ctx->lock, flags);
7165 mutex_unlock(&parent_ctx->mutex);
7167 perf_unpin_context(parent_ctx);
7168 put_ctx(parent_ctx);
7174 * Initialize the perf_event context in task_struct
7176 int perf_event_init_task(struct task_struct *child)
7180 memset(child->perf_event_ctxp, 0, sizeof(child->perf_event_ctxp));
7181 mutex_init(&child->perf_event_mutex);
7182 INIT_LIST_HEAD(&child->perf_event_list);
7184 for_each_task_context_nr(ctxn) {
7185 ret = perf_event_init_context(child, ctxn);
7193 static void __init perf_event_init_all_cpus(void)
7195 struct swevent_htable *swhash;
7198 for_each_possible_cpu(cpu) {
7199 swhash = &per_cpu(swevent_htable, cpu);
7200 mutex_init(&swhash->hlist_mutex);
7201 INIT_LIST_HEAD(&per_cpu(rotation_list, cpu));
7205 static void __cpuinit perf_event_init_cpu(int cpu)
7207 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
7209 mutex_lock(&swhash->hlist_mutex);
7210 if (swhash->hlist_refcount > 0) {
7211 struct swevent_hlist *hlist;
7213 hlist = kzalloc_node(sizeof(*hlist), GFP_KERNEL, cpu_to_node(cpu));
7215 rcu_assign_pointer(swhash->swevent_hlist, hlist);
7217 mutex_unlock(&swhash->hlist_mutex);
7220 #if defined CONFIG_HOTPLUG_CPU || defined CONFIG_KEXEC
7221 static void perf_pmu_rotate_stop(struct pmu *pmu)
7223 struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
7225 WARN_ON(!irqs_disabled());
7227 list_del_init(&cpuctx->rotation_list);
7230 static void __perf_event_exit_context(void *__info)
7232 struct perf_event_context *ctx = __info;
7233 struct perf_event *event, *tmp;
7235 perf_pmu_rotate_stop(ctx->pmu);
7237 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry)
7238 __perf_remove_from_context(event);
7239 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups, group_entry)
7240 __perf_remove_from_context(event);
7243 static void perf_event_exit_cpu_context(int cpu)
7245 struct perf_event_context *ctx;
7249 idx = srcu_read_lock(&pmus_srcu);
7250 list_for_each_entry_rcu(pmu, &pmus, entry) {
7251 ctx = &per_cpu_ptr(pmu->pmu_cpu_context, cpu)->ctx;
7253 mutex_lock(&ctx->mutex);
7254 smp_call_function_single(cpu, __perf_event_exit_context, ctx, 1);
7255 mutex_unlock(&ctx->mutex);
7257 srcu_read_unlock(&pmus_srcu, idx);
7260 static void perf_event_exit_cpu(int cpu)
7262 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
7264 mutex_lock(&swhash->hlist_mutex);
7265 swevent_hlist_release(swhash);
7266 mutex_unlock(&swhash->hlist_mutex);
7268 perf_event_exit_cpu_context(cpu);
7271 static inline void perf_event_exit_cpu(int cpu) { }
7275 perf_reboot(struct notifier_block *notifier, unsigned long val, void *v)
7279 for_each_online_cpu(cpu)
7280 perf_event_exit_cpu(cpu);
7286 * Run the perf reboot notifier at the very last possible moment so that
7287 * the generic watchdog code runs as long as possible.
7289 static struct notifier_block perf_reboot_notifier = {
7290 .notifier_call = perf_reboot,
7291 .priority = INT_MIN,
7294 static int __cpuinit
7295 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
7297 unsigned int cpu = (long)hcpu;
7299 switch (action & ~CPU_TASKS_FROZEN) {
7301 case CPU_UP_PREPARE:
7302 case CPU_DOWN_FAILED:
7303 perf_event_init_cpu(cpu);
7306 case CPU_UP_CANCELED:
7307 case CPU_DOWN_PREPARE:
7308 perf_event_exit_cpu(cpu);
7318 void __init perf_event_init(void)
7324 perf_event_init_all_cpus();
7325 init_srcu_struct(&pmus_srcu);
7326 perf_pmu_register(&perf_swevent, "software", PERF_TYPE_SOFTWARE);
7327 perf_pmu_register(&perf_cpu_clock, NULL, -1);
7328 perf_pmu_register(&perf_task_clock, NULL, -1);
7330 perf_cpu_notifier(perf_cpu_notify);
7331 register_reboot_notifier(&perf_reboot_notifier);
7333 ret = init_hw_breakpoint();
7334 WARN(ret, "hw_breakpoint initialization failed with: %d", ret);
7337 static int __init perf_event_sysfs_init(void)
7342 mutex_lock(&pmus_lock);
7344 ret = bus_register(&pmu_bus);
7348 list_for_each_entry(pmu, &pmus, entry) {
7349 if (!pmu->name || pmu->type < 0)
7352 ret = pmu_dev_alloc(pmu);
7353 WARN(ret, "Failed to register pmu: %s, reason %d\n", pmu->name, ret);
7355 pmu_bus_running = 1;
7359 mutex_unlock(&pmus_lock);
7363 device_initcall(perf_event_sysfs_init);
7365 #ifdef CONFIG_CGROUP_PERF
7366 static struct cgroup_subsys_state *perf_cgroup_create(
7367 struct cgroup_subsys *ss, struct cgroup *cont)
7369 struct perf_cgroup *jc;
7371 jc = kzalloc(sizeof(*jc), GFP_KERNEL);
7373 return ERR_PTR(-ENOMEM);
7375 jc->info = alloc_percpu(struct perf_cgroup_info);
7378 return ERR_PTR(-ENOMEM);
7384 static void perf_cgroup_destroy(struct cgroup_subsys *ss,
7385 struct cgroup *cont)
7387 struct perf_cgroup *jc;
7388 jc = container_of(cgroup_subsys_state(cont, perf_subsys_id),
7389 struct perf_cgroup, css);
7390 free_percpu(jc->info);
7394 static int __perf_cgroup_move(void *info)
7396 struct task_struct *task = info;
7397 perf_cgroup_switch(task, PERF_CGROUP_SWOUT | PERF_CGROUP_SWIN);
7401 static void perf_cgroup_move(struct task_struct *task)
7403 task_function_call(task, __perf_cgroup_move, task);
7406 static void perf_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
7407 struct cgroup *old_cgrp, struct task_struct *task,
7410 perf_cgroup_move(task);
7412 struct task_struct *c;
7414 list_for_each_entry_rcu(c, &task->thread_group, thread_group) {
7415 perf_cgroup_move(c);
7421 static void perf_cgroup_exit(struct cgroup_subsys *ss, struct cgroup *cgrp,
7422 struct cgroup *old_cgrp, struct task_struct *task)
7425 * cgroup_exit() is called in the copy_process() failure path.
7426 * Ignore this case since the task hasn't ran yet, this avoids
7427 * trying to poke a half freed task state from generic code.
7429 if (!(task->flags & PF_EXITING))
7432 perf_cgroup_move(task);
7435 struct cgroup_subsys perf_subsys = {
7436 .name = "perf_event",
7437 .subsys_id = perf_subsys_id,
7438 .create = perf_cgroup_create,
7439 .destroy = perf_cgroup_destroy,
7440 .exit = perf_cgroup_exit,
7441 .attach = perf_cgroup_attach,
7443 #endif /* CONFIG_CGROUP_PERF */