2 * Performance events core code:
4 * Copyright (C) 2008 Thomas Gleixner <tglx@linutronix.de>
5 * Copyright (C) 2008-2011 Red Hat, Inc., Ingo Molnar
6 * Copyright (C) 2008-2011 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
7 * Copyright © 2009 Paul Mackerras, IBM Corp. <paulus@au1.ibm.com>
9 * For licensing details see kernel-base/COPYING
14 #include <linux/cpu.h>
15 #include <linux/smp.h>
16 #include <linux/idr.h>
17 #include <linux/file.h>
18 #include <linux/poll.h>
19 #include <linux/slab.h>
20 #include <linux/hash.h>
21 #include <linux/sysfs.h>
22 #include <linux/dcache.h>
23 #include <linux/percpu.h>
24 #include <linux/ptrace.h>
25 #include <linux/reboot.h>
26 #include <linux/vmstat.h>
27 #include <linux/device.h>
28 #include <linux/vmalloc.h>
29 #include <linux/hardirq.h>
30 #include <linux/rculist.h>
31 #include <linux/uaccess.h>
32 #include <linux/syscalls.h>
33 #include <linux/anon_inodes.h>
34 #include <linux/kernel_stat.h>
35 #include <linux/perf_event.h>
36 #include <linux/ftrace_event.h>
37 #include <linux/hw_breakpoint.h>
39 #include <asm/irq_regs.h>
41 struct remote_function_call {
42 struct task_struct *p;
43 int (*func)(void *info);
48 static void remote_function(void *data)
50 struct remote_function_call *tfc = data;
51 struct task_struct *p = tfc->p;
55 if (task_cpu(p) != smp_processor_id() || !task_curr(p))
59 tfc->ret = tfc->func(tfc->info);
63 * task_function_call - call a function on the cpu on which a task runs
64 * @p: the task to evaluate
65 * @func: the function to be called
66 * @info: the function call argument
68 * Calls the function @func when the task is currently running. This might
69 * be on the current CPU, which just calls the function directly
71 * returns: @func return value, or
72 * -ESRCH - when the process isn't running
73 * -EAGAIN - when the process moved away
76 task_function_call(struct task_struct *p, int (*func) (void *info), void *info)
78 struct remote_function_call data = {
82 .ret = -ESRCH, /* No such (running) process */
86 smp_call_function_single(task_cpu(p), remote_function, &data, 1);
92 * cpu_function_call - call a function on the cpu
93 * @func: the function to be called
94 * @info: the function call argument
96 * Calls the function @func on the remote cpu.
98 * returns: @func return value or -ENXIO when the cpu is offline
100 static int cpu_function_call(int cpu, int (*func) (void *info), void *info)
102 struct remote_function_call data = {
106 .ret = -ENXIO, /* No such CPU */
109 smp_call_function_single(cpu, remote_function, &data, 1);
114 #define PERF_FLAG_ALL (PERF_FLAG_FD_NO_GROUP |\
115 PERF_FLAG_FD_OUTPUT |\
116 PERF_FLAG_PID_CGROUP)
119 EVENT_FLEXIBLE = 0x1,
121 EVENT_ALL = EVENT_FLEXIBLE | EVENT_PINNED,
125 * perf_sched_events : >0 events exist
126 * perf_cgroup_events: >0 per-cpu cgroup events exist on this cpu
128 struct jump_label_key perf_sched_events __read_mostly;
129 static DEFINE_PER_CPU(atomic_t, perf_cgroup_events);
131 static atomic_t nr_mmap_events __read_mostly;
132 static atomic_t nr_comm_events __read_mostly;
133 static atomic_t nr_task_events __read_mostly;
135 static LIST_HEAD(pmus);
136 static DEFINE_MUTEX(pmus_lock);
137 static struct srcu_struct pmus_srcu;
140 * perf event paranoia level:
141 * -1 - not paranoid at all
142 * 0 - disallow raw tracepoint access for unpriv
143 * 1 - disallow cpu events for unpriv
144 * 2 - disallow kernel profiling for unpriv
146 int sysctl_perf_event_paranoid __read_mostly = 1;
148 /* Minimum for 512 kiB + 1 user control page */
149 int sysctl_perf_event_mlock __read_mostly = 512 + (PAGE_SIZE / 1024); /* 'free' kiB per user */
152 * max perf event sample rate
154 #define DEFAULT_MAX_SAMPLE_RATE 100000
155 int sysctl_perf_event_sample_rate __read_mostly = DEFAULT_MAX_SAMPLE_RATE;
156 static int max_samples_per_tick __read_mostly =
157 DIV_ROUND_UP(DEFAULT_MAX_SAMPLE_RATE, HZ);
159 int perf_proc_update_handler(struct ctl_table *table, int write,
160 void __user *buffer, size_t *lenp,
163 int ret = proc_dointvec(table, write, buffer, lenp, ppos);
168 max_samples_per_tick = DIV_ROUND_UP(sysctl_perf_event_sample_rate, HZ);
173 static atomic64_t perf_event_id;
175 static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
176 enum event_type_t event_type);
178 static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
179 enum event_type_t event_type,
180 struct task_struct *task);
182 static void update_context_time(struct perf_event_context *ctx);
183 static u64 perf_event_time(struct perf_event *event);
185 void __weak perf_event_print_debug(void) { }
187 extern __weak const char *perf_pmu_name(void)
192 static inline u64 perf_clock(void)
194 return local_clock();
197 static inline struct perf_cpu_context *
198 __get_cpu_context(struct perf_event_context *ctx)
200 return this_cpu_ptr(ctx->pmu->pmu_cpu_context);
203 #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 WARN_ON_ONCE(cpuctx->cgrp);
368 /* set cgrp before ctxsw in to
369 * allow event_filter_match() to not
370 * have to pass task around
372 cpuctx->cgrp = perf_cgroup_from_task(task);
373 cpu_ctx_sched_in(cpuctx, EVENT_ALL, task);
377 perf_pmu_enable(cpuctx->ctx.pmu);
382 local_irq_restore(flags);
385 static inline void perf_cgroup_sched_out(struct task_struct *task)
387 perf_cgroup_switch(task, PERF_CGROUP_SWOUT);
390 static inline void perf_cgroup_sched_in(struct task_struct *task)
392 perf_cgroup_switch(task, PERF_CGROUP_SWIN);
395 static inline int perf_cgroup_connect(int fd, struct perf_event *event,
396 struct perf_event_attr *attr,
397 struct perf_event *group_leader)
399 struct perf_cgroup *cgrp;
400 struct cgroup_subsys_state *css;
402 int ret = 0, fput_needed;
404 file = fget_light(fd, &fput_needed);
408 css = cgroup_css_from_dir(file, perf_subsys_id);
414 cgrp = container_of(css, struct perf_cgroup, css);
417 /* must be done before we fput() the file */
418 perf_get_cgroup(event);
421 * all events in a group must monitor
422 * the same cgroup because a task belongs
423 * to only one perf cgroup at a time
425 if (group_leader && group_leader->cgrp != cgrp) {
426 perf_detach_cgroup(event);
430 fput_light(file, fput_needed);
435 perf_cgroup_set_shadow_time(struct perf_event *event, u64 now)
437 struct perf_cgroup_info *t;
438 t = per_cpu_ptr(event->cgrp->info, event->cpu);
439 event->shadow_ctx_time = now - t->timestamp;
443 perf_cgroup_defer_enabled(struct perf_event *event)
446 * when the current task's perf cgroup does not match
447 * the event's, we need to remember to call the
448 * perf_mark_enable() function the first time a task with
449 * a matching perf cgroup is scheduled in.
451 if (is_cgroup_event(event) && !perf_cgroup_match(event))
452 event->cgrp_defer_enabled = 1;
456 perf_cgroup_mark_enabled(struct perf_event *event,
457 struct perf_event_context *ctx)
459 struct perf_event *sub;
460 u64 tstamp = perf_event_time(event);
462 if (!event->cgrp_defer_enabled)
465 event->cgrp_defer_enabled = 0;
467 event->tstamp_enabled = tstamp - event->total_time_enabled;
468 list_for_each_entry(sub, &event->sibling_list, group_entry) {
469 if (sub->state >= PERF_EVENT_STATE_INACTIVE) {
470 sub->tstamp_enabled = tstamp - sub->total_time_enabled;
471 sub->cgrp_defer_enabled = 0;
475 #else /* !CONFIG_CGROUP_PERF */
478 perf_cgroup_match(struct perf_event *event)
483 static inline void perf_detach_cgroup(struct perf_event *event)
486 static inline int is_cgroup_event(struct perf_event *event)
491 static inline u64 perf_cgroup_event_cgrp_time(struct perf_event *event)
496 static inline void update_cgrp_time_from_event(struct perf_event *event)
500 static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx)
504 static inline void perf_cgroup_sched_out(struct task_struct *task)
508 static inline void perf_cgroup_sched_in(struct task_struct *task)
512 static inline int perf_cgroup_connect(pid_t pid, struct perf_event *event,
513 struct perf_event_attr *attr,
514 struct perf_event *group_leader)
520 perf_cgroup_set_timestamp(struct task_struct *task,
521 struct perf_event_context *ctx)
526 perf_cgroup_switch(struct task_struct *task, struct task_struct *next)
531 perf_cgroup_set_shadow_time(struct perf_event *event, u64 now)
535 static inline u64 perf_cgroup_event_time(struct perf_event *event)
541 perf_cgroup_defer_enabled(struct perf_event *event)
546 perf_cgroup_mark_enabled(struct perf_event *event,
547 struct perf_event_context *ctx)
552 void perf_pmu_disable(struct pmu *pmu)
554 int *count = this_cpu_ptr(pmu->pmu_disable_count);
556 pmu->pmu_disable(pmu);
559 void perf_pmu_enable(struct pmu *pmu)
561 int *count = this_cpu_ptr(pmu->pmu_disable_count);
563 pmu->pmu_enable(pmu);
566 static DEFINE_PER_CPU(struct list_head, rotation_list);
569 * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
570 * because they're strictly cpu affine and rotate_start is called with IRQs
571 * disabled, while rotate_context is called from IRQ context.
573 static void perf_pmu_rotate_start(struct pmu *pmu)
575 struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
576 struct list_head *head = &__get_cpu_var(rotation_list);
578 WARN_ON(!irqs_disabled());
580 if (list_empty(&cpuctx->rotation_list))
581 list_add(&cpuctx->rotation_list, head);
584 static void get_ctx(struct perf_event_context *ctx)
586 WARN_ON(!atomic_inc_not_zero(&ctx->refcount));
589 static void put_ctx(struct perf_event_context *ctx)
591 if (atomic_dec_and_test(&ctx->refcount)) {
593 put_ctx(ctx->parent_ctx);
595 put_task_struct(ctx->task);
596 kfree_rcu(ctx, rcu_head);
600 static void unclone_ctx(struct perf_event_context *ctx)
602 if (ctx->parent_ctx) {
603 put_ctx(ctx->parent_ctx);
604 ctx->parent_ctx = NULL;
608 static u32 perf_event_pid(struct perf_event *event, struct task_struct *p)
611 * only top level events have the pid namespace they were created in
614 event = event->parent;
616 return task_tgid_nr_ns(p, event->ns);
619 static u32 perf_event_tid(struct perf_event *event, struct task_struct *p)
622 * only top level events have the pid namespace they were created in
625 event = event->parent;
627 return task_pid_nr_ns(p, event->ns);
631 * If we inherit events we want to return the parent event id
634 static u64 primary_event_id(struct perf_event *event)
639 id = event->parent->id;
645 * Get the perf_event_context for a task and lock it.
646 * This has to cope with with the fact that until it is locked,
647 * the context could get moved to another task.
649 static struct perf_event_context *
650 perf_lock_task_context(struct task_struct *task, int ctxn, unsigned long *flags)
652 struct perf_event_context *ctx;
656 ctx = rcu_dereference(task->perf_event_ctxp[ctxn]);
659 * If this context is a clone of another, it might
660 * get swapped for another underneath us by
661 * perf_event_task_sched_out, though the
662 * rcu_read_lock() protects us from any context
663 * getting freed. Lock the context and check if it
664 * got swapped before we could get the lock, and retry
665 * if so. If we locked the right context, then it
666 * can't get swapped on us any more.
668 raw_spin_lock_irqsave(&ctx->lock, *flags);
669 if (ctx != rcu_dereference(task->perf_event_ctxp[ctxn])) {
670 raw_spin_unlock_irqrestore(&ctx->lock, *flags);
674 if (!atomic_inc_not_zero(&ctx->refcount)) {
675 raw_spin_unlock_irqrestore(&ctx->lock, *flags);
684 * Get the context for a task and increment its pin_count so it
685 * can't get swapped to another task. This also increments its
686 * reference count so that the context can't get freed.
688 static struct perf_event_context *
689 perf_pin_task_context(struct task_struct *task, int ctxn)
691 struct perf_event_context *ctx;
694 ctx = perf_lock_task_context(task, ctxn, &flags);
697 raw_spin_unlock_irqrestore(&ctx->lock, flags);
702 static void perf_unpin_context(struct perf_event_context *ctx)
706 raw_spin_lock_irqsave(&ctx->lock, flags);
708 raw_spin_unlock_irqrestore(&ctx->lock, flags);
712 * Update the record of the current time in a context.
714 static void update_context_time(struct perf_event_context *ctx)
716 u64 now = perf_clock();
718 ctx->time += now - ctx->timestamp;
719 ctx->timestamp = now;
722 static u64 perf_event_time(struct perf_event *event)
724 struct perf_event_context *ctx = event->ctx;
726 if (is_cgroup_event(event))
727 return perf_cgroup_event_time(event);
729 return ctx ? ctx->time : 0;
733 * Update the total_time_enabled and total_time_running fields for a event.
735 static void update_event_times(struct perf_event *event)
737 struct perf_event_context *ctx = event->ctx;
740 if (event->state < PERF_EVENT_STATE_INACTIVE ||
741 event->group_leader->state < PERF_EVENT_STATE_INACTIVE)
744 * in cgroup mode, time_enabled represents
745 * the time the event was enabled AND active
746 * tasks were in the monitored cgroup. This is
747 * independent of the activity of the context as
748 * there may be a mix of cgroup and non-cgroup events.
750 * That is why we treat cgroup events differently
753 if (is_cgroup_event(event))
754 run_end = perf_event_time(event);
755 else if (ctx->is_active)
758 run_end = event->tstamp_stopped;
760 event->total_time_enabled = run_end - event->tstamp_enabled;
762 if (event->state == PERF_EVENT_STATE_INACTIVE)
763 run_end = event->tstamp_stopped;
765 run_end = perf_event_time(event);
767 event->total_time_running = run_end - event->tstamp_running;
772 * Update total_time_enabled and total_time_running for all events in a group.
774 static void update_group_times(struct perf_event *leader)
776 struct perf_event *event;
778 update_event_times(leader);
779 list_for_each_entry(event, &leader->sibling_list, group_entry)
780 update_event_times(event);
783 static struct list_head *
784 ctx_group_list(struct perf_event *event, struct perf_event_context *ctx)
786 if (event->attr.pinned)
787 return &ctx->pinned_groups;
789 return &ctx->flexible_groups;
793 * Add a event from the lists for its context.
794 * Must be called with ctx->mutex and ctx->lock held.
797 list_add_event(struct perf_event *event, struct perf_event_context *ctx)
799 WARN_ON_ONCE(event->attach_state & PERF_ATTACH_CONTEXT);
800 event->attach_state |= PERF_ATTACH_CONTEXT;
803 * If we're a stand alone event or group leader, we go to the context
804 * list, group events are kept attached to the group so that
805 * perf_group_detach can, at all times, locate all siblings.
807 if (event->group_leader == event) {
808 struct list_head *list;
810 if (is_software_event(event))
811 event->group_flags |= PERF_GROUP_SOFTWARE;
813 list = ctx_group_list(event, ctx);
814 list_add_tail(&event->group_entry, list);
817 if (is_cgroup_event(event))
820 list_add_rcu(&event->event_entry, &ctx->event_list);
822 perf_pmu_rotate_start(ctx->pmu);
824 if (event->attr.inherit_stat)
829 * Called at perf_event creation and when events are attached/detached from a
832 static void perf_event__read_size(struct perf_event *event)
834 int entry = sizeof(u64); /* value */
838 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
841 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
844 if (event->attr.read_format & PERF_FORMAT_ID)
845 entry += sizeof(u64);
847 if (event->attr.read_format & PERF_FORMAT_GROUP) {
848 nr += event->group_leader->nr_siblings;
853 event->read_size = size;
856 static void perf_event__header_size(struct perf_event *event)
858 struct perf_sample_data *data;
859 u64 sample_type = event->attr.sample_type;
862 perf_event__read_size(event);
864 if (sample_type & PERF_SAMPLE_IP)
865 size += sizeof(data->ip);
867 if (sample_type & PERF_SAMPLE_ADDR)
868 size += sizeof(data->addr);
870 if (sample_type & PERF_SAMPLE_PERIOD)
871 size += sizeof(data->period);
873 if (sample_type & PERF_SAMPLE_READ)
874 size += event->read_size;
876 event->header_size = size;
879 static void perf_event__id_header_size(struct perf_event *event)
881 struct perf_sample_data *data;
882 u64 sample_type = event->attr.sample_type;
885 if (sample_type & PERF_SAMPLE_TID)
886 size += sizeof(data->tid_entry);
888 if (sample_type & PERF_SAMPLE_TIME)
889 size += sizeof(data->time);
891 if (sample_type & PERF_SAMPLE_ID)
892 size += sizeof(data->id);
894 if (sample_type & PERF_SAMPLE_STREAM_ID)
895 size += sizeof(data->stream_id);
897 if (sample_type & PERF_SAMPLE_CPU)
898 size += sizeof(data->cpu_entry);
900 event->id_header_size = size;
903 static void perf_group_attach(struct perf_event *event)
905 struct perf_event *group_leader = event->group_leader, *pos;
908 * We can have double attach due to group movement in perf_event_open.
910 if (event->attach_state & PERF_ATTACH_GROUP)
913 event->attach_state |= PERF_ATTACH_GROUP;
915 if (group_leader == event)
918 if (group_leader->group_flags & PERF_GROUP_SOFTWARE &&
919 !is_software_event(event))
920 group_leader->group_flags &= ~PERF_GROUP_SOFTWARE;
922 list_add_tail(&event->group_entry, &group_leader->sibling_list);
923 group_leader->nr_siblings++;
925 perf_event__header_size(group_leader);
927 list_for_each_entry(pos, &group_leader->sibling_list, group_entry)
928 perf_event__header_size(pos);
932 * Remove a event from the lists for its context.
933 * Must be called with ctx->mutex and ctx->lock held.
936 list_del_event(struct perf_event *event, struct perf_event_context *ctx)
938 struct perf_cpu_context *cpuctx;
940 * We can have double detach due to exit/hot-unplug + close.
942 if (!(event->attach_state & PERF_ATTACH_CONTEXT))
945 event->attach_state &= ~PERF_ATTACH_CONTEXT;
947 if (is_cgroup_event(event)) {
949 cpuctx = __get_cpu_context(ctx);
951 * if there are no more cgroup events
952 * then cler cgrp to avoid stale pointer
953 * in update_cgrp_time_from_cpuctx()
955 if (!ctx->nr_cgroups)
960 if (event->attr.inherit_stat)
963 list_del_rcu(&event->event_entry);
965 if (event->group_leader == event)
966 list_del_init(&event->group_entry);
968 update_group_times(event);
971 * If event was in error state, then keep it
972 * that way, otherwise bogus counts will be
973 * returned on read(). The only way to get out
974 * of error state is by explicit re-enabling
977 if (event->state > PERF_EVENT_STATE_OFF)
978 event->state = PERF_EVENT_STATE_OFF;
981 static void perf_group_detach(struct perf_event *event)
983 struct perf_event *sibling, *tmp;
984 struct list_head *list = NULL;
987 * We can have double detach due to exit/hot-unplug + close.
989 if (!(event->attach_state & PERF_ATTACH_GROUP))
992 event->attach_state &= ~PERF_ATTACH_GROUP;
995 * If this is a sibling, remove it from its group.
997 if (event->group_leader != event) {
998 list_del_init(&event->group_entry);
999 event->group_leader->nr_siblings--;
1003 if (!list_empty(&event->group_entry))
1004 list = &event->group_entry;
1007 * If this was a group event with sibling events then
1008 * upgrade the siblings to singleton events by adding them
1009 * to whatever list we are on.
1011 list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) {
1013 list_move_tail(&sibling->group_entry, list);
1014 sibling->group_leader = sibling;
1016 /* Inherit group flags from the previous leader */
1017 sibling->group_flags = event->group_flags;
1021 perf_event__header_size(event->group_leader);
1023 list_for_each_entry(tmp, &event->group_leader->sibling_list, group_entry)
1024 perf_event__header_size(tmp);
1028 event_filter_match(struct perf_event *event)
1030 return (event->cpu == -1 || event->cpu == smp_processor_id())
1031 && perf_cgroup_match(event);
1035 event_sched_out(struct perf_event *event,
1036 struct perf_cpu_context *cpuctx,
1037 struct perf_event_context *ctx)
1039 u64 tstamp = perf_event_time(event);
1042 * An event which could not be activated because of
1043 * filter mismatch still needs to have its timings
1044 * maintained, otherwise bogus information is return
1045 * via read() for time_enabled, time_running:
1047 if (event->state == PERF_EVENT_STATE_INACTIVE
1048 && !event_filter_match(event)) {
1049 delta = tstamp - event->tstamp_stopped;
1050 event->tstamp_running += delta;
1051 event->tstamp_stopped = tstamp;
1054 if (event->state != PERF_EVENT_STATE_ACTIVE)
1057 event->state = PERF_EVENT_STATE_INACTIVE;
1058 if (event->pending_disable) {
1059 event->pending_disable = 0;
1060 event->state = PERF_EVENT_STATE_OFF;
1062 event->tstamp_stopped = tstamp;
1063 event->pmu->del(event, 0);
1066 if (!is_software_event(event))
1067 cpuctx->active_oncpu--;
1069 if (event->attr.exclusive || !cpuctx->active_oncpu)
1070 cpuctx->exclusive = 0;
1074 group_sched_out(struct perf_event *group_event,
1075 struct perf_cpu_context *cpuctx,
1076 struct perf_event_context *ctx)
1078 struct perf_event *event;
1079 int state = group_event->state;
1081 event_sched_out(group_event, cpuctx, ctx);
1084 * Schedule out siblings (if any):
1086 list_for_each_entry(event, &group_event->sibling_list, group_entry)
1087 event_sched_out(event, cpuctx, ctx);
1089 if (state == PERF_EVENT_STATE_ACTIVE && group_event->attr.exclusive)
1090 cpuctx->exclusive = 0;
1094 * Cross CPU call to remove a performance event
1096 * We disable the event on the hardware level first. After that we
1097 * remove it from the context list.
1099 static int __perf_remove_from_context(void *info)
1101 struct perf_event *event = info;
1102 struct perf_event_context *ctx = event->ctx;
1103 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1105 raw_spin_lock(&ctx->lock);
1106 event_sched_out(event, cpuctx, ctx);
1107 list_del_event(event, ctx);
1108 raw_spin_unlock(&ctx->lock);
1115 * Remove the event from a task's (or a CPU's) list of events.
1117 * CPU events are removed with a smp call. For task events we only
1118 * call when the task is on a CPU.
1120 * If event->ctx is a cloned context, callers must make sure that
1121 * every task struct that event->ctx->task could possibly point to
1122 * remains valid. This is OK when called from perf_release since
1123 * that only calls us on the top-level context, which can't be a clone.
1124 * When called from perf_event_exit_task, it's OK because the
1125 * context has been detached from its task.
1127 static void perf_remove_from_context(struct perf_event *event)
1129 struct perf_event_context *ctx = event->ctx;
1130 struct task_struct *task = ctx->task;
1132 lockdep_assert_held(&ctx->mutex);
1136 * Per cpu events are removed via an smp call and
1137 * the removal is always successful.
1139 cpu_function_call(event->cpu, __perf_remove_from_context, event);
1144 if (!task_function_call(task, __perf_remove_from_context, event))
1147 raw_spin_lock_irq(&ctx->lock);
1149 * If we failed to find a running task, but find the context active now
1150 * that we've acquired the ctx->lock, retry.
1152 if (ctx->is_active) {
1153 raw_spin_unlock_irq(&ctx->lock);
1158 * Since the task isn't running, its safe to remove the event, us
1159 * holding the ctx->lock ensures the task won't get scheduled in.
1161 list_del_event(event, ctx);
1162 raw_spin_unlock_irq(&ctx->lock);
1166 * Cross CPU call to disable a performance event
1168 static int __perf_event_disable(void *info)
1170 struct perf_event *event = info;
1171 struct perf_event_context *ctx = event->ctx;
1172 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1175 * If this is a per-task event, need to check whether this
1176 * event's task is the current task on this cpu.
1178 * Can trigger due to concurrent perf_event_context_sched_out()
1179 * flipping contexts around.
1181 if (ctx->task && cpuctx->task_ctx != ctx)
1184 raw_spin_lock(&ctx->lock);
1187 * If the event is on, turn it off.
1188 * If it is in error state, leave it in error state.
1190 if (event->state >= PERF_EVENT_STATE_INACTIVE) {
1191 update_context_time(ctx);
1192 update_cgrp_time_from_event(event);
1193 update_group_times(event);
1194 if (event == event->group_leader)
1195 group_sched_out(event, cpuctx, ctx);
1197 event_sched_out(event, cpuctx, ctx);
1198 event->state = PERF_EVENT_STATE_OFF;
1201 raw_spin_unlock(&ctx->lock);
1209 * If event->ctx is a cloned context, callers must make sure that
1210 * every task struct that event->ctx->task could possibly point to
1211 * remains valid. This condition is satisifed when called through
1212 * perf_event_for_each_child or perf_event_for_each because they
1213 * hold the top-level event's child_mutex, so any descendant that
1214 * goes to exit will block in sync_child_event.
1215 * When called from perf_pending_event it's OK because event->ctx
1216 * is the current context on this CPU and preemption is disabled,
1217 * hence we can't get into perf_event_task_sched_out for this context.
1219 void perf_event_disable(struct perf_event *event)
1221 struct perf_event_context *ctx = event->ctx;
1222 struct task_struct *task = ctx->task;
1226 * Disable the event on the cpu that it's on
1228 cpu_function_call(event->cpu, __perf_event_disable, event);
1233 if (!task_function_call(task, __perf_event_disable, event))
1236 raw_spin_lock_irq(&ctx->lock);
1238 * If the event is still active, we need to retry the cross-call.
1240 if (event->state == PERF_EVENT_STATE_ACTIVE) {
1241 raw_spin_unlock_irq(&ctx->lock);
1243 * Reload the task pointer, it might have been changed by
1244 * a concurrent perf_event_context_sched_out().
1251 * Since we have the lock this context can't be scheduled
1252 * in, so we can change the state safely.
1254 if (event->state == PERF_EVENT_STATE_INACTIVE) {
1255 update_group_times(event);
1256 event->state = PERF_EVENT_STATE_OFF;
1258 raw_spin_unlock_irq(&ctx->lock);
1261 static void perf_set_shadow_time(struct perf_event *event,
1262 struct perf_event_context *ctx,
1266 * use the correct time source for the time snapshot
1268 * We could get by without this by leveraging the
1269 * fact that to get to this function, the caller
1270 * has most likely already called update_context_time()
1271 * and update_cgrp_time_xx() and thus both timestamp
1272 * are identical (or very close). Given that tstamp is,
1273 * already adjusted for cgroup, we could say that:
1274 * tstamp - ctx->timestamp
1276 * tstamp - cgrp->timestamp.
1278 * Then, in perf_output_read(), the calculation would
1279 * work with no changes because:
1280 * - event is guaranteed scheduled in
1281 * - no scheduled out in between
1282 * - thus the timestamp would be the same
1284 * But this is a bit hairy.
1286 * So instead, we have an explicit cgroup call to remain
1287 * within the time time source all along. We believe it
1288 * is cleaner and simpler to understand.
1290 if (is_cgroup_event(event))
1291 perf_cgroup_set_shadow_time(event, tstamp);
1293 event->shadow_ctx_time = tstamp - ctx->timestamp;
1296 #define MAX_INTERRUPTS (~0ULL)
1298 static void perf_log_throttle(struct perf_event *event, int enable);
1301 event_sched_in(struct perf_event *event,
1302 struct perf_cpu_context *cpuctx,
1303 struct perf_event_context *ctx)
1305 u64 tstamp = perf_event_time(event);
1307 if (event->state <= PERF_EVENT_STATE_OFF)
1310 event->state = PERF_EVENT_STATE_ACTIVE;
1311 event->oncpu = smp_processor_id();
1314 * Unthrottle events, since we scheduled we might have missed several
1315 * ticks already, also for a heavily scheduling task there is little
1316 * guarantee it'll get a tick in a timely manner.
1318 if (unlikely(event->hw.interrupts == MAX_INTERRUPTS)) {
1319 perf_log_throttle(event, 1);
1320 event->hw.interrupts = 0;
1324 * The new state must be visible before we turn it on in the hardware:
1328 if (event->pmu->add(event, PERF_EF_START)) {
1329 event->state = PERF_EVENT_STATE_INACTIVE;
1334 event->tstamp_running += tstamp - event->tstamp_stopped;
1336 perf_set_shadow_time(event, ctx, tstamp);
1338 if (!is_software_event(event))
1339 cpuctx->active_oncpu++;
1342 if (event->attr.exclusive)
1343 cpuctx->exclusive = 1;
1349 group_sched_in(struct perf_event *group_event,
1350 struct perf_cpu_context *cpuctx,
1351 struct perf_event_context *ctx)
1353 struct perf_event *event, *partial_group = NULL;
1354 struct pmu *pmu = group_event->pmu;
1355 u64 now = ctx->time;
1356 bool simulate = false;
1358 if (group_event->state == PERF_EVENT_STATE_OFF)
1361 pmu->start_txn(pmu);
1363 if (event_sched_in(group_event, cpuctx, ctx)) {
1364 pmu->cancel_txn(pmu);
1369 * Schedule in siblings as one group (if any):
1371 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
1372 if (event_sched_in(event, cpuctx, ctx)) {
1373 partial_group = event;
1378 if (!pmu->commit_txn(pmu))
1383 * Groups can be scheduled in as one unit only, so undo any
1384 * partial group before returning:
1385 * The events up to the failed event are scheduled out normally,
1386 * tstamp_stopped will be updated.
1388 * The failed events and the remaining siblings need to have
1389 * their timings updated as if they had gone thru event_sched_in()
1390 * and event_sched_out(). This is required to get consistent timings
1391 * across the group. This also takes care of the case where the group
1392 * could never be scheduled by ensuring tstamp_stopped is set to mark
1393 * the time the event was actually stopped, such that time delta
1394 * calculation in update_event_times() is correct.
1396 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
1397 if (event == partial_group)
1401 event->tstamp_running += now - event->tstamp_stopped;
1402 event->tstamp_stopped = now;
1404 event_sched_out(event, cpuctx, ctx);
1407 event_sched_out(group_event, cpuctx, ctx);
1409 pmu->cancel_txn(pmu);
1415 * Work out whether we can put this event group on the CPU now.
1417 static int group_can_go_on(struct perf_event *event,
1418 struct perf_cpu_context *cpuctx,
1422 * Groups consisting entirely of software events can always go on.
1424 if (event->group_flags & PERF_GROUP_SOFTWARE)
1427 * If an exclusive group is already on, no other hardware
1430 if (cpuctx->exclusive)
1433 * If this group is exclusive and there are already
1434 * events on the CPU, it can't go on.
1436 if (event->attr.exclusive && cpuctx->active_oncpu)
1439 * Otherwise, try to add it if all previous groups were able
1445 static void add_event_to_ctx(struct perf_event *event,
1446 struct perf_event_context *ctx)
1448 u64 tstamp = perf_event_time(event);
1450 list_add_event(event, ctx);
1451 perf_group_attach(event);
1452 event->tstamp_enabled = tstamp;
1453 event->tstamp_running = tstamp;
1454 event->tstamp_stopped = tstamp;
1457 static void perf_event_context_sched_in(struct perf_event_context *ctx,
1458 struct task_struct *tsk);
1461 * Cross CPU call to install and enable a performance event
1463 * Must be called with ctx->mutex held
1465 static int __perf_install_in_context(void *info)
1467 struct perf_event *event = info;
1468 struct perf_event_context *ctx = event->ctx;
1469 struct perf_event *leader = event->group_leader;
1470 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1474 * In case we're installing a new context to an already running task,
1475 * could also happen before perf_event_task_sched_in() on architectures
1476 * which do context switches with IRQs enabled.
1478 if (ctx->task && !cpuctx->task_ctx)
1479 perf_event_context_sched_in(ctx, ctx->task);
1481 raw_spin_lock(&ctx->lock);
1483 update_context_time(ctx);
1485 * update cgrp time only if current cgrp
1486 * matches event->cgrp. Must be done before
1487 * calling add_event_to_ctx()
1489 update_cgrp_time_from_event(event);
1491 add_event_to_ctx(event, ctx);
1493 if (!event_filter_match(event))
1497 * Don't put the event on if it is disabled or if
1498 * it is in a group and the group isn't on.
1500 if (event->state != PERF_EVENT_STATE_INACTIVE ||
1501 (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE))
1505 * An exclusive event can't go on if there are already active
1506 * hardware events, and no hardware event can go on if there
1507 * is already an exclusive event on.
1509 if (!group_can_go_on(event, cpuctx, 1))
1512 err = event_sched_in(event, cpuctx, ctx);
1516 * This event couldn't go on. If it is in a group
1517 * then we have to pull the whole group off.
1518 * If the event group is pinned then put it in error state.
1520 if (leader != event)
1521 group_sched_out(leader, cpuctx, ctx);
1522 if (leader->attr.pinned) {
1523 update_group_times(leader);
1524 leader->state = PERF_EVENT_STATE_ERROR;
1529 raw_spin_unlock(&ctx->lock);
1535 * Attach a performance event to a context
1537 * First we add the event to the list with the hardware enable bit
1538 * in event->hw_config cleared.
1540 * If the event is attached to a task which is on a CPU we use a smp
1541 * call to enable it in the task context. The task might have been
1542 * scheduled away, but we check this in the smp call again.
1545 perf_install_in_context(struct perf_event_context *ctx,
1546 struct perf_event *event,
1549 struct task_struct *task = ctx->task;
1551 lockdep_assert_held(&ctx->mutex);
1557 * Per cpu events are installed via an smp call and
1558 * the install is always successful.
1560 cpu_function_call(cpu, __perf_install_in_context, event);
1565 if (!task_function_call(task, __perf_install_in_context, event))
1568 raw_spin_lock_irq(&ctx->lock);
1570 * If we failed to find a running task, but find the context active now
1571 * that we've acquired the ctx->lock, retry.
1573 if (ctx->is_active) {
1574 raw_spin_unlock_irq(&ctx->lock);
1579 * Since the task isn't running, its safe to add the event, us holding
1580 * the ctx->lock ensures the task won't get scheduled in.
1582 add_event_to_ctx(event, ctx);
1583 raw_spin_unlock_irq(&ctx->lock);
1587 * Put a event into inactive state and update time fields.
1588 * Enabling the leader of a group effectively enables all
1589 * the group members that aren't explicitly disabled, so we
1590 * have to update their ->tstamp_enabled also.
1591 * Note: this works for group members as well as group leaders
1592 * since the non-leader members' sibling_lists will be empty.
1594 static void __perf_event_mark_enabled(struct perf_event *event,
1595 struct perf_event_context *ctx)
1597 struct perf_event *sub;
1598 u64 tstamp = perf_event_time(event);
1600 event->state = PERF_EVENT_STATE_INACTIVE;
1601 event->tstamp_enabled = tstamp - event->total_time_enabled;
1602 list_for_each_entry(sub, &event->sibling_list, group_entry) {
1603 if (sub->state >= PERF_EVENT_STATE_INACTIVE)
1604 sub->tstamp_enabled = tstamp - sub->total_time_enabled;
1609 * Cross CPU call to enable a performance event
1611 static int __perf_event_enable(void *info)
1613 struct perf_event *event = info;
1614 struct perf_event_context *ctx = event->ctx;
1615 struct perf_event *leader = event->group_leader;
1616 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1619 if (WARN_ON_ONCE(!ctx->is_active))
1622 raw_spin_lock(&ctx->lock);
1623 update_context_time(ctx);
1625 if (event->state >= PERF_EVENT_STATE_INACTIVE)
1629 * set current task's cgroup time reference point
1631 perf_cgroup_set_timestamp(current, ctx);
1633 __perf_event_mark_enabled(event, ctx);
1635 if (!event_filter_match(event)) {
1636 if (is_cgroup_event(event))
1637 perf_cgroup_defer_enabled(event);
1642 * If the event is in a group and isn't the group leader,
1643 * then don't put it on unless the group is on.
1645 if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE)
1648 if (!group_can_go_on(event, cpuctx, 1)) {
1651 if (event == leader)
1652 err = group_sched_in(event, cpuctx, ctx);
1654 err = event_sched_in(event, cpuctx, ctx);
1659 * If this event can't go on and it's part of a
1660 * group, then the whole group has to come off.
1662 if (leader != event)
1663 group_sched_out(leader, cpuctx, ctx);
1664 if (leader->attr.pinned) {
1665 update_group_times(leader);
1666 leader->state = PERF_EVENT_STATE_ERROR;
1671 raw_spin_unlock(&ctx->lock);
1679 * If event->ctx is a cloned context, callers must make sure that
1680 * every task struct that event->ctx->task could possibly point to
1681 * remains valid. This condition is satisfied when called through
1682 * perf_event_for_each_child or perf_event_for_each as described
1683 * for perf_event_disable.
1685 void perf_event_enable(struct perf_event *event)
1687 struct perf_event_context *ctx = event->ctx;
1688 struct task_struct *task = ctx->task;
1692 * Enable the event on the cpu that it's on
1694 cpu_function_call(event->cpu, __perf_event_enable, event);
1698 raw_spin_lock_irq(&ctx->lock);
1699 if (event->state >= PERF_EVENT_STATE_INACTIVE)
1703 * If the event is in error state, clear that first.
1704 * That way, if we see the event in error state below, we
1705 * know that it has gone back into error state, as distinct
1706 * from the task having been scheduled away before the
1707 * cross-call arrived.
1709 if (event->state == PERF_EVENT_STATE_ERROR)
1710 event->state = PERF_EVENT_STATE_OFF;
1713 if (!ctx->is_active) {
1714 __perf_event_mark_enabled(event, ctx);
1718 raw_spin_unlock_irq(&ctx->lock);
1720 if (!task_function_call(task, __perf_event_enable, event))
1723 raw_spin_lock_irq(&ctx->lock);
1726 * If the context is active and the event is still off,
1727 * we need to retry the cross-call.
1729 if (ctx->is_active && event->state == PERF_EVENT_STATE_OFF) {
1731 * task could have been flipped by a concurrent
1732 * perf_event_context_sched_out()
1739 raw_spin_unlock_irq(&ctx->lock);
1742 static int perf_event_refresh(struct perf_event *event, int refresh)
1745 * not supported on inherited events
1747 if (event->attr.inherit || !is_sampling_event(event))
1750 atomic_add(refresh, &event->event_limit);
1751 perf_event_enable(event);
1756 static void ctx_sched_out(struct perf_event_context *ctx,
1757 struct perf_cpu_context *cpuctx,
1758 enum event_type_t event_type)
1760 struct perf_event *event;
1762 raw_spin_lock(&ctx->lock);
1763 perf_pmu_disable(ctx->pmu);
1765 if (likely(!ctx->nr_events))
1767 update_context_time(ctx);
1768 update_cgrp_time_from_cpuctx(cpuctx);
1770 if (!ctx->nr_active)
1773 if (event_type & EVENT_PINNED) {
1774 list_for_each_entry(event, &ctx->pinned_groups, group_entry)
1775 group_sched_out(event, cpuctx, ctx);
1778 if (event_type & EVENT_FLEXIBLE) {
1779 list_for_each_entry(event, &ctx->flexible_groups, group_entry)
1780 group_sched_out(event, cpuctx, ctx);
1783 perf_pmu_enable(ctx->pmu);
1784 raw_spin_unlock(&ctx->lock);
1788 * Test whether two contexts are equivalent, i.e. whether they
1789 * have both been cloned from the same version of the same context
1790 * and they both have the same number of enabled events.
1791 * If the number of enabled events is the same, then the set
1792 * of enabled events should be the same, because these are both
1793 * inherited contexts, therefore we can't access individual events
1794 * in them directly with an fd; we can only enable/disable all
1795 * events via prctl, or enable/disable all events in a family
1796 * via ioctl, which will have the same effect on both contexts.
1798 static int context_equiv(struct perf_event_context *ctx1,
1799 struct perf_event_context *ctx2)
1801 return ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx
1802 && ctx1->parent_gen == ctx2->parent_gen
1803 && !ctx1->pin_count && !ctx2->pin_count;
1806 static void __perf_event_sync_stat(struct perf_event *event,
1807 struct perf_event *next_event)
1811 if (!event->attr.inherit_stat)
1815 * Update the event value, we cannot use perf_event_read()
1816 * because we're in the middle of a context switch and have IRQs
1817 * disabled, which upsets smp_call_function_single(), however
1818 * we know the event must be on the current CPU, therefore we
1819 * don't need to use it.
1821 switch (event->state) {
1822 case PERF_EVENT_STATE_ACTIVE:
1823 event->pmu->read(event);
1826 case PERF_EVENT_STATE_INACTIVE:
1827 update_event_times(event);
1835 * In order to keep per-task stats reliable we need to flip the event
1836 * values when we flip the contexts.
1838 value = local64_read(&next_event->count);
1839 value = local64_xchg(&event->count, value);
1840 local64_set(&next_event->count, value);
1842 swap(event->total_time_enabled, next_event->total_time_enabled);
1843 swap(event->total_time_running, next_event->total_time_running);
1846 * Since we swizzled the values, update the user visible data too.
1848 perf_event_update_userpage(event);
1849 perf_event_update_userpage(next_event);
1852 #define list_next_entry(pos, member) \
1853 list_entry(pos->member.next, typeof(*pos), member)
1855 static void perf_event_sync_stat(struct perf_event_context *ctx,
1856 struct perf_event_context *next_ctx)
1858 struct perf_event *event, *next_event;
1863 update_context_time(ctx);
1865 event = list_first_entry(&ctx->event_list,
1866 struct perf_event, event_entry);
1868 next_event = list_first_entry(&next_ctx->event_list,
1869 struct perf_event, event_entry);
1871 while (&event->event_entry != &ctx->event_list &&
1872 &next_event->event_entry != &next_ctx->event_list) {
1874 __perf_event_sync_stat(event, next_event);
1876 event = list_next_entry(event, event_entry);
1877 next_event = list_next_entry(next_event, event_entry);
1881 static void perf_event_context_sched_out(struct task_struct *task, int ctxn,
1882 struct task_struct *next)
1884 struct perf_event_context *ctx = task->perf_event_ctxp[ctxn];
1885 struct perf_event_context *next_ctx;
1886 struct perf_event_context *parent;
1887 struct perf_cpu_context *cpuctx;
1893 cpuctx = __get_cpu_context(ctx);
1894 if (!cpuctx->task_ctx)
1898 parent = rcu_dereference(ctx->parent_ctx);
1899 next_ctx = next->perf_event_ctxp[ctxn];
1900 if (parent && next_ctx &&
1901 rcu_dereference(next_ctx->parent_ctx) == parent) {
1903 * Looks like the two contexts are clones, so we might be
1904 * able to optimize the context switch. We lock both
1905 * contexts and check that they are clones under the
1906 * lock (including re-checking that neither has been
1907 * uncloned in the meantime). It doesn't matter which
1908 * order we take the locks because no other cpu could
1909 * be trying to lock both of these tasks.
1911 raw_spin_lock(&ctx->lock);
1912 raw_spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING);
1913 if (context_equiv(ctx, next_ctx)) {
1915 * XXX do we need a memory barrier of sorts
1916 * wrt to rcu_dereference() of perf_event_ctxp
1918 task->perf_event_ctxp[ctxn] = next_ctx;
1919 next->perf_event_ctxp[ctxn] = ctx;
1921 next_ctx->task = task;
1924 perf_event_sync_stat(ctx, next_ctx);
1926 raw_spin_unlock(&next_ctx->lock);
1927 raw_spin_unlock(&ctx->lock);
1932 ctx_sched_out(ctx, cpuctx, EVENT_ALL);
1933 cpuctx->task_ctx = NULL;
1937 #define for_each_task_context_nr(ctxn) \
1938 for ((ctxn) = 0; (ctxn) < perf_nr_task_contexts; (ctxn)++)
1941 * Called from scheduler to remove the events of the current task,
1942 * with interrupts disabled.
1944 * We stop each event and update the event value in event->count.
1946 * This does not protect us against NMI, but disable()
1947 * sets the disabled bit in the control field of event _before_
1948 * accessing the event control register. If a NMI hits, then it will
1949 * not restart the event.
1951 void __perf_event_task_sched_out(struct task_struct *task,
1952 struct task_struct *next)
1956 for_each_task_context_nr(ctxn)
1957 perf_event_context_sched_out(task, ctxn, next);
1960 * if cgroup events exist on this CPU, then we need
1961 * to check if we have to switch out PMU state.
1962 * cgroup event are system-wide mode only
1964 if (atomic_read(&__get_cpu_var(perf_cgroup_events)))
1965 perf_cgroup_sched_out(task);
1968 static void task_ctx_sched_out(struct perf_event_context *ctx,
1969 enum event_type_t event_type)
1971 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1973 if (!cpuctx->task_ctx)
1976 if (WARN_ON_ONCE(ctx != cpuctx->task_ctx))
1979 ctx_sched_out(ctx, cpuctx, event_type);
1980 cpuctx->task_ctx = NULL;
1984 * Called with IRQs disabled
1986 static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
1987 enum event_type_t event_type)
1989 ctx_sched_out(&cpuctx->ctx, cpuctx, event_type);
1993 ctx_pinned_sched_in(struct perf_event_context *ctx,
1994 struct perf_cpu_context *cpuctx)
1996 struct perf_event *event;
1998 list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
1999 if (event->state <= PERF_EVENT_STATE_OFF)
2001 if (!event_filter_match(event))
2004 /* may need to reset tstamp_enabled */
2005 if (is_cgroup_event(event))
2006 perf_cgroup_mark_enabled(event, ctx);
2008 if (group_can_go_on(event, cpuctx, 1))
2009 group_sched_in(event, cpuctx, ctx);
2012 * If this pinned group hasn't been scheduled,
2013 * put it in error state.
2015 if (event->state == PERF_EVENT_STATE_INACTIVE) {
2016 update_group_times(event);
2017 event->state = PERF_EVENT_STATE_ERROR;
2023 ctx_flexible_sched_in(struct perf_event_context *ctx,
2024 struct perf_cpu_context *cpuctx)
2026 struct perf_event *event;
2029 list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
2030 /* Ignore events in OFF or ERROR state */
2031 if (event->state <= PERF_EVENT_STATE_OFF)
2034 * Listen to the 'cpu' scheduling filter constraint
2037 if (!event_filter_match(event))
2040 /* may need to reset tstamp_enabled */
2041 if (is_cgroup_event(event))
2042 perf_cgroup_mark_enabled(event, ctx);
2044 if (group_can_go_on(event, cpuctx, can_add_hw)) {
2045 if (group_sched_in(event, cpuctx, ctx))
2052 ctx_sched_in(struct perf_event_context *ctx,
2053 struct perf_cpu_context *cpuctx,
2054 enum event_type_t event_type,
2055 struct task_struct *task)
2059 raw_spin_lock(&ctx->lock);
2061 if (likely(!ctx->nr_events))
2065 ctx->timestamp = now;
2066 perf_cgroup_set_timestamp(task, ctx);
2068 * First go through the list and put on any pinned groups
2069 * in order to give them the best chance of going on.
2071 if (event_type & EVENT_PINNED)
2072 ctx_pinned_sched_in(ctx, cpuctx);
2074 /* Then walk through the lower prio flexible groups */
2075 if (event_type & EVENT_FLEXIBLE)
2076 ctx_flexible_sched_in(ctx, cpuctx);
2079 raw_spin_unlock(&ctx->lock);
2082 static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
2083 enum event_type_t event_type,
2084 struct task_struct *task)
2086 struct perf_event_context *ctx = &cpuctx->ctx;
2088 ctx_sched_in(ctx, cpuctx, event_type, task);
2091 static void task_ctx_sched_in(struct perf_event_context *ctx,
2092 enum event_type_t event_type)
2094 struct perf_cpu_context *cpuctx;
2096 cpuctx = __get_cpu_context(ctx);
2097 if (cpuctx->task_ctx == ctx)
2100 ctx_sched_in(ctx, cpuctx, event_type, NULL);
2101 cpuctx->task_ctx = ctx;
2104 static void perf_event_context_sched_in(struct perf_event_context *ctx,
2105 struct task_struct *task)
2107 struct perf_cpu_context *cpuctx;
2109 cpuctx = __get_cpu_context(ctx);
2110 if (cpuctx->task_ctx == ctx)
2113 perf_pmu_disable(ctx->pmu);
2115 * We want to keep the following priority order:
2116 * cpu pinned (that don't need to move), task pinned,
2117 * cpu flexible, task flexible.
2119 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
2121 ctx_sched_in(ctx, cpuctx, EVENT_PINNED, task);
2122 cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE, task);
2123 ctx_sched_in(ctx, cpuctx, EVENT_FLEXIBLE, task);
2125 cpuctx->task_ctx = ctx;
2128 * Since these rotations are per-cpu, we need to ensure the
2129 * cpu-context we got scheduled on is actually rotating.
2131 perf_pmu_rotate_start(ctx->pmu);
2132 perf_pmu_enable(ctx->pmu);
2136 * Called from scheduler to add the events of the current task
2137 * with interrupts disabled.
2139 * We restore the event value and then enable it.
2141 * This does not protect us against NMI, but enable()
2142 * sets the enabled bit in the control field of event _before_
2143 * accessing the event control register. If a NMI hits, then it will
2144 * keep the event running.
2146 void __perf_event_task_sched_in(struct task_struct *task)
2148 struct perf_event_context *ctx;
2151 for_each_task_context_nr(ctxn) {
2152 ctx = task->perf_event_ctxp[ctxn];
2156 perf_event_context_sched_in(ctx, task);
2159 * if cgroup events exist on this CPU, then we need
2160 * to check if we have to switch in PMU state.
2161 * cgroup event are system-wide mode only
2163 if (atomic_read(&__get_cpu_var(perf_cgroup_events)))
2164 perf_cgroup_sched_in(task);
2167 static u64 perf_calculate_period(struct perf_event *event, u64 nsec, u64 count)
2169 u64 frequency = event->attr.sample_freq;
2170 u64 sec = NSEC_PER_SEC;
2171 u64 divisor, dividend;
2173 int count_fls, nsec_fls, frequency_fls, sec_fls;
2175 count_fls = fls64(count);
2176 nsec_fls = fls64(nsec);
2177 frequency_fls = fls64(frequency);
2181 * We got @count in @nsec, with a target of sample_freq HZ
2182 * the target period becomes:
2185 * period = -------------------
2186 * @nsec * sample_freq
2191 * Reduce accuracy by one bit such that @a and @b converge
2192 * to a similar magnitude.
2194 #define REDUCE_FLS(a, b) \
2196 if (a##_fls > b##_fls) { \
2206 * Reduce accuracy until either term fits in a u64, then proceed with
2207 * the other, so that finally we can do a u64/u64 division.
2209 while (count_fls + sec_fls > 64 && nsec_fls + frequency_fls > 64) {
2210 REDUCE_FLS(nsec, frequency);
2211 REDUCE_FLS(sec, count);
2214 if (count_fls + sec_fls > 64) {
2215 divisor = nsec * frequency;
2217 while (count_fls + sec_fls > 64) {
2218 REDUCE_FLS(count, sec);
2222 dividend = count * sec;
2224 dividend = count * sec;
2226 while (nsec_fls + frequency_fls > 64) {
2227 REDUCE_FLS(nsec, frequency);
2231 divisor = nsec * frequency;
2237 return div64_u64(dividend, divisor);
2240 static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count)
2242 struct hw_perf_event *hwc = &event->hw;
2243 s64 period, sample_period;
2246 period = perf_calculate_period(event, nsec, count);
2248 delta = (s64)(period - hwc->sample_period);
2249 delta = (delta + 7) / 8; /* low pass filter */
2251 sample_period = hwc->sample_period + delta;
2256 hwc->sample_period = sample_period;
2258 if (local64_read(&hwc->period_left) > 8*sample_period) {
2259 event->pmu->stop(event, PERF_EF_UPDATE);
2260 local64_set(&hwc->period_left, 0);
2261 event->pmu->start(event, PERF_EF_RELOAD);
2265 static void perf_ctx_adjust_freq(struct perf_event_context *ctx, u64 period)
2267 struct perf_event *event;
2268 struct hw_perf_event *hwc;
2269 u64 interrupts, now;
2272 raw_spin_lock(&ctx->lock);
2273 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
2274 if (event->state != PERF_EVENT_STATE_ACTIVE)
2277 if (!event_filter_match(event))
2282 interrupts = hwc->interrupts;
2283 hwc->interrupts = 0;
2286 * unthrottle events on the tick
2288 if (interrupts == MAX_INTERRUPTS) {
2289 perf_log_throttle(event, 1);
2290 event->pmu->start(event, 0);
2293 if (!event->attr.freq || !event->attr.sample_freq)
2296 event->pmu->read(event);
2297 now = local64_read(&event->count);
2298 delta = now - hwc->freq_count_stamp;
2299 hwc->freq_count_stamp = now;
2302 perf_adjust_period(event, period, delta);
2304 raw_spin_unlock(&ctx->lock);
2308 * Round-robin a context's events:
2310 static void rotate_ctx(struct perf_event_context *ctx)
2312 raw_spin_lock(&ctx->lock);
2315 * Rotate the first entry last of non-pinned groups. Rotation might be
2316 * disabled by the inheritance code.
2318 if (!ctx->rotate_disable)
2319 list_rotate_left(&ctx->flexible_groups);
2321 raw_spin_unlock(&ctx->lock);
2325 * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
2326 * because they're strictly cpu affine and rotate_start is called with IRQs
2327 * disabled, while rotate_context is called from IRQ context.
2329 static void perf_rotate_context(struct perf_cpu_context *cpuctx)
2331 u64 interval = (u64)cpuctx->jiffies_interval * TICK_NSEC;
2332 struct perf_event_context *ctx = NULL;
2333 int rotate = 0, remove = 1;
2335 if (cpuctx->ctx.nr_events) {
2337 if (cpuctx->ctx.nr_events != cpuctx->ctx.nr_active)
2341 ctx = cpuctx->task_ctx;
2342 if (ctx && ctx->nr_events) {
2344 if (ctx->nr_events != ctx->nr_active)
2348 perf_pmu_disable(cpuctx->ctx.pmu);
2349 perf_ctx_adjust_freq(&cpuctx->ctx, interval);
2351 perf_ctx_adjust_freq(ctx, interval);
2356 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
2358 task_ctx_sched_out(ctx, EVENT_FLEXIBLE);
2360 rotate_ctx(&cpuctx->ctx);
2364 cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE, current);
2366 task_ctx_sched_in(ctx, EVENT_FLEXIBLE);
2370 list_del_init(&cpuctx->rotation_list);
2372 perf_pmu_enable(cpuctx->ctx.pmu);
2375 void perf_event_task_tick(void)
2377 struct list_head *head = &__get_cpu_var(rotation_list);
2378 struct perf_cpu_context *cpuctx, *tmp;
2380 WARN_ON(!irqs_disabled());
2382 list_for_each_entry_safe(cpuctx, tmp, head, rotation_list) {
2383 if (cpuctx->jiffies_interval == 1 ||
2384 !(jiffies % cpuctx->jiffies_interval))
2385 perf_rotate_context(cpuctx);
2389 static int event_enable_on_exec(struct perf_event *event,
2390 struct perf_event_context *ctx)
2392 if (!event->attr.enable_on_exec)
2395 event->attr.enable_on_exec = 0;
2396 if (event->state >= PERF_EVENT_STATE_INACTIVE)
2399 __perf_event_mark_enabled(event, ctx);
2405 * Enable all of a task's events that have been marked enable-on-exec.
2406 * This expects task == current.
2408 static void perf_event_enable_on_exec(struct perf_event_context *ctx)
2410 struct perf_event *event;
2411 unsigned long flags;
2415 local_irq_save(flags);
2416 if (!ctx || !ctx->nr_events)
2420 * We must ctxsw out cgroup events to avoid conflict
2421 * when invoking perf_task_event_sched_in() later on
2422 * in this function. Otherwise we end up trying to
2423 * ctxswin cgroup events which are already scheduled
2426 perf_cgroup_sched_out(current);
2427 task_ctx_sched_out(ctx, EVENT_ALL);
2429 raw_spin_lock(&ctx->lock);
2431 list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
2432 ret = event_enable_on_exec(event, ctx);
2437 list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
2438 ret = event_enable_on_exec(event, ctx);
2444 * Unclone this context if we enabled any event.
2449 raw_spin_unlock(&ctx->lock);
2452 * Also calls ctxswin for cgroup events, if any:
2454 perf_event_context_sched_in(ctx, ctx->task);
2456 local_irq_restore(flags);
2460 * Cross CPU call to read the hardware event
2462 static void __perf_event_read(void *info)
2464 struct perf_event *event = info;
2465 struct perf_event_context *ctx = event->ctx;
2466 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
2469 * If this is a task context, we need to check whether it is
2470 * the current task context of this cpu. If not it has been
2471 * scheduled out before the smp call arrived. In that case
2472 * event->count would have been updated to a recent sample
2473 * when the event was scheduled out.
2475 if (ctx->task && cpuctx->task_ctx != ctx)
2478 raw_spin_lock(&ctx->lock);
2479 if (ctx->is_active) {
2480 update_context_time(ctx);
2481 update_cgrp_time_from_event(event);
2483 update_event_times(event);
2484 if (event->state == PERF_EVENT_STATE_ACTIVE)
2485 event->pmu->read(event);
2486 raw_spin_unlock(&ctx->lock);
2489 static inline u64 perf_event_count(struct perf_event *event)
2491 return local64_read(&event->count) + atomic64_read(&event->child_count);
2494 static u64 perf_event_read(struct perf_event *event)
2497 * If event is enabled and currently active on a CPU, update the
2498 * value in the event structure:
2500 if (event->state == PERF_EVENT_STATE_ACTIVE) {
2501 smp_call_function_single(event->oncpu,
2502 __perf_event_read, event, 1);
2503 } else if (event->state == PERF_EVENT_STATE_INACTIVE) {
2504 struct perf_event_context *ctx = event->ctx;
2505 unsigned long flags;
2507 raw_spin_lock_irqsave(&ctx->lock, flags);
2509 * may read while context is not active
2510 * (e.g., thread is blocked), in that case
2511 * we cannot update context time
2513 if (ctx->is_active) {
2514 update_context_time(ctx);
2515 update_cgrp_time_from_event(event);
2517 update_event_times(event);
2518 raw_spin_unlock_irqrestore(&ctx->lock, flags);
2521 return perf_event_count(event);
2528 struct callchain_cpus_entries {
2529 struct rcu_head rcu_head;
2530 struct perf_callchain_entry *cpu_entries[0];
2533 static DEFINE_PER_CPU(int, callchain_recursion[PERF_NR_CONTEXTS]);
2534 static atomic_t nr_callchain_events;
2535 static DEFINE_MUTEX(callchain_mutex);
2536 struct callchain_cpus_entries *callchain_cpus_entries;
2539 __weak void perf_callchain_kernel(struct perf_callchain_entry *entry,
2540 struct pt_regs *regs)
2544 __weak void perf_callchain_user(struct perf_callchain_entry *entry,
2545 struct pt_regs *regs)
2549 static void release_callchain_buffers_rcu(struct rcu_head *head)
2551 struct callchain_cpus_entries *entries;
2554 entries = container_of(head, struct callchain_cpus_entries, rcu_head);
2556 for_each_possible_cpu(cpu)
2557 kfree(entries->cpu_entries[cpu]);
2562 static void release_callchain_buffers(void)
2564 struct callchain_cpus_entries *entries;
2566 entries = callchain_cpus_entries;
2567 rcu_assign_pointer(callchain_cpus_entries, NULL);
2568 call_rcu(&entries->rcu_head, release_callchain_buffers_rcu);
2571 static int alloc_callchain_buffers(void)
2575 struct callchain_cpus_entries *entries;
2578 * We can't use the percpu allocation API for data that can be
2579 * accessed from NMI. Use a temporary manual per cpu allocation
2580 * until that gets sorted out.
2582 size = offsetof(struct callchain_cpus_entries, cpu_entries[nr_cpu_ids]);
2584 entries = kzalloc(size, GFP_KERNEL);
2588 size = sizeof(struct perf_callchain_entry) * PERF_NR_CONTEXTS;
2590 for_each_possible_cpu(cpu) {
2591 entries->cpu_entries[cpu] = kmalloc_node(size, GFP_KERNEL,
2593 if (!entries->cpu_entries[cpu])
2597 rcu_assign_pointer(callchain_cpus_entries, entries);
2602 for_each_possible_cpu(cpu)
2603 kfree(entries->cpu_entries[cpu]);
2609 static int get_callchain_buffers(void)
2614 mutex_lock(&callchain_mutex);
2616 count = atomic_inc_return(&nr_callchain_events);
2617 if (WARN_ON_ONCE(count < 1)) {
2623 /* If the allocation failed, give up */
2624 if (!callchain_cpus_entries)
2629 err = alloc_callchain_buffers();
2631 release_callchain_buffers();
2633 mutex_unlock(&callchain_mutex);
2638 static void put_callchain_buffers(void)
2640 if (atomic_dec_and_mutex_lock(&nr_callchain_events, &callchain_mutex)) {
2641 release_callchain_buffers();
2642 mutex_unlock(&callchain_mutex);
2646 static int get_recursion_context(int *recursion)
2654 else if (in_softirq())
2659 if (recursion[rctx])
2668 static inline void put_recursion_context(int *recursion, int rctx)
2674 static struct perf_callchain_entry *get_callchain_entry(int *rctx)
2677 struct callchain_cpus_entries *entries;
2679 *rctx = get_recursion_context(__get_cpu_var(callchain_recursion));
2683 entries = rcu_dereference(callchain_cpus_entries);
2687 cpu = smp_processor_id();
2689 return &entries->cpu_entries[cpu][*rctx];
2693 put_callchain_entry(int rctx)
2695 put_recursion_context(__get_cpu_var(callchain_recursion), rctx);
2698 static struct perf_callchain_entry *perf_callchain(struct pt_regs *regs)
2701 struct perf_callchain_entry *entry;
2704 entry = get_callchain_entry(&rctx);
2713 if (!user_mode(regs)) {
2714 perf_callchain_store(entry, PERF_CONTEXT_KERNEL);
2715 perf_callchain_kernel(entry, regs);
2717 regs = task_pt_regs(current);
2723 perf_callchain_store(entry, PERF_CONTEXT_USER);
2724 perf_callchain_user(entry, regs);
2728 put_callchain_entry(rctx);
2734 * Initialize the perf_event context in a task_struct:
2736 static void __perf_event_init_context(struct perf_event_context *ctx)
2738 raw_spin_lock_init(&ctx->lock);
2739 mutex_init(&ctx->mutex);
2740 INIT_LIST_HEAD(&ctx->pinned_groups);
2741 INIT_LIST_HEAD(&ctx->flexible_groups);
2742 INIT_LIST_HEAD(&ctx->event_list);
2743 atomic_set(&ctx->refcount, 1);
2746 static struct perf_event_context *
2747 alloc_perf_context(struct pmu *pmu, struct task_struct *task)
2749 struct perf_event_context *ctx;
2751 ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL);
2755 __perf_event_init_context(ctx);
2758 get_task_struct(task);
2765 static struct task_struct *
2766 find_lively_task_by_vpid(pid_t vpid)
2768 struct task_struct *task;
2775 task = find_task_by_vpid(vpid);
2777 get_task_struct(task);
2781 return ERR_PTR(-ESRCH);
2783 /* Reuse ptrace permission checks for now. */
2785 if (!ptrace_may_access(task, PTRACE_MODE_READ))
2790 put_task_struct(task);
2791 return ERR_PTR(err);
2796 * Returns a matching context with refcount and pincount.
2798 static struct perf_event_context *
2799 find_get_context(struct pmu *pmu, struct task_struct *task, int cpu)
2801 struct perf_event_context *ctx;
2802 struct perf_cpu_context *cpuctx;
2803 unsigned long flags;
2807 /* Must be root to operate on a CPU event: */
2808 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
2809 return ERR_PTR(-EACCES);
2812 * We could be clever and allow to attach a event to an
2813 * offline CPU and activate it when the CPU comes up, but
2816 if (!cpu_online(cpu))
2817 return ERR_PTR(-ENODEV);
2819 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
2828 ctxn = pmu->task_ctx_nr;
2833 ctx = perf_lock_task_context(task, ctxn, &flags);
2837 raw_spin_unlock_irqrestore(&ctx->lock, flags);
2841 ctx = alloc_perf_context(pmu, task);
2849 mutex_lock(&task->perf_event_mutex);
2851 * If it has already passed perf_event_exit_task().
2852 * we must see PF_EXITING, it takes this mutex too.
2854 if (task->flags & PF_EXITING)
2856 else if (task->perf_event_ctxp[ctxn])
2860 rcu_assign_pointer(task->perf_event_ctxp[ctxn], ctx);
2862 mutex_unlock(&task->perf_event_mutex);
2864 if (unlikely(err)) {
2865 put_task_struct(task);
2877 return ERR_PTR(err);
2880 static void perf_event_free_filter(struct perf_event *event);
2882 static void free_event_rcu(struct rcu_head *head)
2884 struct perf_event *event;
2886 event = container_of(head, struct perf_event, rcu_head);
2888 put_pid_ns(event->ns);
2889 perf_event_free_filter(event);
2893 static void perf_buffer_put(struct perf_buffer *buffer);
2895 static void free_event(struct perf_event *event)
2897 irq_work_sync(&event->pending);
2899 if (!event->parent) {
2900 if (event->attach_state & PERF_ATTACH_TASK)
2901 jump_label_dec(&perf_sched_events);
2902 if (event->attr.mmap || event->attr.mmap_data)
2903 atomic_dec(&nr_mmap_events);
2904 if (event->attr.comm)
2905 atomic_dec(&nr_comm_events);
2906 if (event->attr.task)
2907 atomic_dec(&nr_task_events);
2908 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN)
2909 put_callchain_buffers();
2910 if (is_cgroup_event(event)) {
2911 atomic_dec(&per_cpu(perf_cgroup_events, event->cpu));
2912 jump_label_dec(&perf_sched_events);
2916 if (event->buffer) {
2917 perf_buffer_put(event->buffer);
2918 event->buffer = NULL;
2921 if (is_cgroup_event(event))
2922 perf_detach_cgroup(event);
2925 event->destroy(event);
2928 put_ctx(event->ctx);
2930 call_rcu(&event->rcu_head, free_event_rcu);
2933 int perf_event_release_kernel(struct perf_event *event)
2935 struct perf_event_context *ctx = event->ctx;
2938 * Remove from the PMU, can't get re-enabled since we got
2939 * here because the last ref went.
2941 perf_event_disable(event);
2943 WARN_ON_ONCE(ctx->parent_ctx);
2945 * There are two ways this annotation is useful:
2947 * 1) there is a lock recursion from perf_event_exit_task
2948 * see the comment there.
2950 * 2) there is a lock-inversion with mmap_sem through
2951 * perf_event_read_group(), which takes faults while
2952 * holding ctx->mutex, however this is called after
2953 * the last filedesc died, so there is no possibility
2954 * to trigger the AB-BA case.
2956 mutex_lock_nested(&ctx->mutex, SINGLE_DEPTH_NESTING);
2957 raw_spin_lock_irq(&ctx->lock);
2958 perf_group_detach(event);
2959 list_del_event(event, ctx);
2960 raw_spin_unlock_irq(&ctx->lock);
2961 mutex_unlock(&ctx->mutex);
2967 EXPORT_SYMBOL_GPL(perf_event_release_kernel);
2970 * Called when the last reference to the file is gone.
2972 static int perf_release(struct inode *inode, struct file *file)
2974 struct perf_event *event = file->private_data;
2975 struct task_struct *owner;
2977 file->private_data = NULL;
2980 owner = ACCESS_ONCE(event->owner);
2982 * Matches the smp_wmb() in perf_event_exit_task(). If we observe
2983 * !owner it means the list deletion is complete and we can indeed
2984 * free this event, otherwise we need to serialize on
2985 * owner->perf_event_mutex.
2987 smp_read_barrier_depends();
2990 * Since delayed_put_task_struct() also drops the last
2991 * task reference we can safely take a new reference
2992 * while holding the rcu_read_lock().
2994 get_task_struct(owner);
2999 mutex_lock(&owner->perf_event_mutex);
3001 * We have to re-check the event->owner field, if it is cleared
3002 * we raced with perf_event_exit_task(), acquiring the mutex
3003 * ensured they're done, and we can proceed with freeing the
3007 list_del_init(&event->owner_entry);
3008 mutex_unlock(&owner->perf_event_mutex);
3009 put_task_struct(owner);
3012 return perf_event_release_kernel(event);
3015 u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running)
3017 struct perf_event *child;
3023 mutex_lock(&event->child_mutex);
3024 total += perf_event_read(event);
3025 *enabled += event->total_time_enabled +
3026 atomic64_read(&event->child_total_time_enabled);
3027 *running += event->total_time_running +
3028 atomic64_read(&event->child_total_time_running);
3030 list_for_each_entry(child, &event->child_list, child_list) {
3031 total += perf_event_read(child);
3032 *enabled += child->total_time_enabled;
3033 *running += child->total_time_running;
3035 mutex_unlock(&event->child_mutex);
3039 EXPORT_SYMBOL_GPL(perf_event_read_value);
3041 static int perf_event_read_group(struct perf_event *event,
3042 u64 read_format, char __user *buf)
3044 struct perf_event *leader = event->group_leader, *sub;
3045 int n = 0, size = 0, ret = -EFAULT;
3046 struct perf_event_context *ctx = leader->ctx;
3048 u64 count, enabled, running;
3050 mutex_lock(&ctx->mutex);
3051 count = perf_event_read_value(leader, &enabled, &running);
3053 values[n++] = 1 + leader->nr_siblings;
3054 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3055 values[n++] = enabled;
3056 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3057 values[n++] = running;
3058 values[n++] = count;
3059 if (read_format & PERF_FORMAT_ID)
3060 values[n++] = primary_event_id(leader);
3062 size = n * sizeof(u64);
3064 if (copy_to_user(buf, values, size))
3069 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
3072 values[n++] = perf_event_read_value(sub, &enabled, &running);
3073 if (read_format & PERF_FORMAT_ID)
3074 values[n++] = primary_event_id(sub);
3076 size = n * sizeof(u64);
3078 if (copy_to_user(buf + ret, values, size)) {
3086 mutex_unlock(&ctx->mutex);
3091 static int perf_event_read_one(struct perf_event *event,
3092 u64 read_format, char __user *buf)
3094 u64 enabled, running;
3098 values[n++] = perf_event_read_value(event, &enabled, &running);
3099 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3100 values[n++] = enabled;
3101 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3102 values[n++] = running;
3103 if (read_format & PERF_FORMAT_ID)
3104 values[n++] = primary_event_id(event);
3106 if (copy_to_user(buf, values, n * sizeof(u64)))
3109 return n * sizeof(u64);
3113 * Read the performance event - simple non blocking version for now
3116 perf_read_hw(struct perf_event *event, char __user *buf, size_t count)
3118 u64 read_format = event->attr.read_format;
3122 * Return end-of-file for a read on a event that is in
3123 * error state (i.e. because it was pinned but it couldn't be
3124 * scheduled on to the CPU at some point).
3126 if (event->state == PERF_EVENT_STATE_ERROR)
3129 if (count < event->read_size)
3132 WARN_ON_ONCE(event->ctx->parent_ctx);
3133 if (read_format & PERF_FORMAT_GROUP)
3134 ret = perf_event_read_group(event, read_format, buf);
3136 ret = perf_event_read_one(event, read_format, buf);
3142 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
3144 struct perf_event *event = file->private_data;
3146 return perf_read_hw(event, buf, count);
3149 static unsigned int perf_poll(struct file *file, poll_table *wait)
3151 struct perf_event *event = file->private_data;
3152 struct perf_buffer *buffer;
3153 unsigned int events = POLL_HUP;
3156 buffer = rcu_dereference(event->buffer);
3158 events = atomic_xchg(&buffer->poll, 0);
3161 poll_wait(file, &event->waitq, wait);
3166 static void perf_event_reset(struct perf_event *event)
3168 (void)perf_event_read(event);
3169 local64_set(&event->count, 0);
3170 perf_event_update_userpage(event);
3174 * Holding the top-level event's child_mutex means that any
3175 * descendant process that has inherited this event will block
3176 * in sync_child_event if it goes to exit, thus satisfying the
3177 * task existence requirements of perf_event_enable/disable.
3179 static void perf_event_for_each_child(struct perf_event *event,
3180 void (*func)(struct perf_event *))
3182 struct perf_event *child;
3184 WARN_ON_ONCE(event->ctx->parent_ctx);
3185 mutex_lock(&event->child_mutex);
3187 list_for_each_entry(child, &event->child_list, child_list)
3189 mutex_unlock(&event->child_mutex);
3192 static void perf_event_for_each(struct perf_event *event,
3193 void (*func)(struct perf_event *))
3195 struct perf_event_context *ctx = event->ctx;
3196 struct perf_event *sibling;
3198 WARN_ON_ONCE(ctx->parent_ctx);
3199 mutex_lock(&ctx->mutex);
3200 event = event->group_leader;
3202 perf_event_for_each_child(event, func);
3204 list_for_each_entry(sibling, &event->sibling_list, group_entry)
3205 perf_event_for_each_child(event, func);
3206 mutex_unlock(&ctx->mutex);
3209 static int perf_event_period(struct perf_event *event, u64 __user *arg)
3211 struct perf_event_context *ctx = event->ctx;
3215 if (!is_sampling_event(event))
3218 if (copy_from_user(&value, arg, sizeof(value)))
3224 raw_spin_lock_irq(&ctx->lock);
3225 if (event->attr.freq) {
3226 if (value > sysctl_perf_event_sample_rate) {
3231 event->attr.sample_freq = value;
3233 event->attr.sample_period = value;
3234 event->hw.sample_period = value;
3237 raw_spin_unlock_irq(&ctx->lock);
3242 static const struct file_operations perf_fops;
3244 static struct perf_event *perf_fget_light(int fd, int *fput_needed)
3248 file = fget_light(fd, fput_needed);
3250 return ERR_PTR(-EBADF);
3252 if (file->f_op != &perf_fops) {
3253 fput_light(file, *fput_needed);
3255 return ERR_PTR(-EBADF);
3258 return file->private_data;
3261 static int perf_event_set_output(struct perf_event *event,
3262 struct perf_event *output_event);
3263 static int perf_event_set_filter(struct perf_event *event, void __user *arg);
3265 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
3267 struct perf_event *event = file->private_data;
3268 void (*func)(struct perf_event *);
3272 case PERF_EVENT_IOC_ENABLE:
3273 func = perf_event_enable;
3275 case PERF_EVENT_IOC_DISABLE:
3276 func = perf_event_disable;
3278 case PERF_EVENT_IOC_RESET:
3279 func = perf_event_reset;
3282 case PERF_EVENT_IOC_REFRESH:
3283 return perf_event_refresh(event, arg);
3285 case PERF_EVENT_IOC_PERIOD:
3286 return perf_event_period(event, (u64 __user *)arg);
3288 case PERF_EVENT_IOC_SET_OUTPUT:
3290 struct perf_event *output_event = NULL;
3291 int fput_needed = 0;
3295 output_event = perf_fget_light(arg, &fput_needed);
3296 if (IS_ERR(output_event))
3297 return PTR_ERR(output_event);
3300 ret = perf_event_set_output(event, output_event);
3302 fput_light(output_event->filp, fput_needed);
3307 case PERF_EVENT_IOC_SET_FILTER:
3308 return perf_event_set_filter(event, (void __user *)arg);
3314 if (flags & PERF_IOC_FLAG_GROUP)
3315 perf_event_for_each(event, func);
3317 perf_event_for_each_child(event, func);
3322 int perf_event_task_enable(void)
3324 struct perf_event *event;
3326 mutex_lock(¤t->perf_event_mutex);
3327 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
3328 perf_event_for_each_child(event, perf_event_enable);
3329 mutex_unlock(¤t->perf_event_mutex);
3334 int perf_event_task_disable(void)
3336 struct perf_event *event;
3338 mutex_lock(¤t->perf_event_mutex);
3339 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
3340 perf_event_for_each_child(event, perf_event_disable);
3341 mutex_unlock(¤t->perf_event_mutex);
3346 #ifndef PERF_EVENT_INDEX_OFFSET
3347 # define PERF_EVENT_INDEX_OFFSET 0
3350 static int perf_event_index(struct perf_event *event)
3352 if (event->hw.state & PERF_HES_STOPPED)
3355 if (event->state != PERF_EVENT_STATE_ACTIVE)
3358 return event->hw.idx + 1 - PERF_EVENT_INDEX_OFFSET;
3362 * Callers need to ensure there can be no nesting of this function, otherwise
3363 * the seqlock logic goes bad. We can not serialize this because the arch
3364 * code calls this from NMI context.
3366 void perf_event_update_userpage(struct perf_event *event)
3368 struct perf_event_mmap_page *userpg;
3369 struct perf_buffer *buffer;
3372 buffer = rcu_dereference(event->buffer);
3376 userpg = buffer->user_page;
3379 * Disable preemption so as to not let the corresponding user-space
3380 * spin too long if we get preempted.
3385 userpg->index = perf_event_index(event);
3386 userpg->offset = perf_event_count(event);
3387 if (event->state == PERF_EVENT_STATE_ACTIVE)
3388 userpg->offset -= local64_read(&event->hw.prev_count);
3390 userpg->time_enabled = event->total_time_enabled +
3391 atomic64_read(&event->child_total_time_enabled);
3393 userpg->time_running = event->total_time_running +
3394 atomic64_read(&event->child_total_time_running);
3403 static unsigned long perf_data_size(struct perf_buffer *buffer);
3406 perf_buffer_init(struct perf_buffer *buffer, long watermark, int flags)
3408 long max_size = perf_data_size(buffer);
3411 buffer->watermark = min(max_size, watermark);
3413 if (!buffer->watermark)
3414 buffer->watermark = max_size / 2;
3416 if (flags & PERF_BUFFER_WRITABLE)
3417 buffer->writable = 1;
3419 atomic_set(&buffer->refcount, 1);
3422 #ifndef CONFIG_PERF_USE_VMALLOC
3425 * Back perf_mmap() with regular GFP_KERNEL-0 pages.
3428 static struct page *
3429 perf_mmap_to_page(struct perf_buffer *buffer, unsigned long pgoff)
3431 if (pgoff > buffer->nr_pages)
3435 return virt_to_page(buffer->user_page);
3437 return virt_to_page(buffer->data_pages[pgoff - 1]);
3440 static void *perf_mmap_alloc_page(int cpu)
3445 node = (cpu == -1) ? cpu : cpu_to_node(cpu);
3446 page = alloc_pages_node(node, GFP_KERNEL | __GFP_ZERO, 0);
3450 return page_address(page);
3453 static struct perf_buffer *
3454 perf_buffer_alloc(int nr_pages, long watermark, int cpu, int flags)
3456 struct perf_buffer *buffer;
3460 size = sizeof(struct perf_buffer);
3461 size += nr_pages * sizeof(void *);
3463 buffer = kzalloc(size, GFP_KERNEL);
3467 buffer->user_page = perf_mmap_alloc_page(cpu);
3468 if (!buffer->user_page)
3469 goto fail_user_page;
3471 for (i = 0; i < nr_pages; i++) {
3472 buffer->data_pages[i] = perf_mmap_alloc_page(cpu);
3473 if (!buffer->data_pages[i])
3474 goto fail_data_pages;
3477 buffer->nr_pages = nr_pages;
3479 perf_buffer_init(buffer, watermark, flags);
3484 for (i--; i >= 0; i--)
3485 free_page((unsigned long)buffer->data_pages[i]);
3487 free_page((unsigned long)buffer->user_page);
3496 static void perf_mmap_free_page(unsigned long addr)
3498 struct page *page = virt_to_page((void *)addr);
3500 page->mapping = NULL;
3504 static void perf_buffer_free(struct perf_buffer *buffer)
3508 perf_mmap_free_page((unsigned long)buffer->user_page);
3509 for (i = 0; i < buffer->nr_pages; i++)
3510 perf_mmap_free_page((unsigned long)buffer->data_pages[i]);
3514 static inline int page_order(struct perf_buffer *buffer)
3522 * Back perf_mmap() with vmalloc memory.
3524 * Required for architectures that have d-cache aliasing issues.
3527 static inline int page_order(struct perf_buffer *buffer)
3529 return buffer->page_order;
3532 static struct page *
3533 perf_mmap_to_page(struct perf_buffer *buffer, unsigned long pgoff)
3535 if (pgoff > (1UL << page_order(buffer)))
3538 return vmalloc_to_page((void *)buffer->user_page + pgoff * PAGE_SIZE);
3541 static void perf_mmap_unmark_page(void *addr)
3543 struct page *page = vmalloc_to_page(addr);
3545 page->mapping = NULL;
3548 static void perf_buffer_free_work(struct work_struct *work)
3550 struct perf_buffer *buffer;
3554 buffer = container_of(work, struct perf_buffer, work);
3555 nr = 1 << page_order(buffer);
3557 base = buffer->user_page;
3558 for (i = 0; i < nr + 1; i++)
3559 perf_mmap_unmark_page(base + (i * PAGE_SIZE));
3565 static void perf_buffer_free(struct perf_buffer *buffer)
3567 schedule_work(&buffer->work);
3570 static struct perf_buffer *
3571 perf_buffer_alloc(int nr_pages, long watermark, int cpu, int flags)
3573 struct perf_buffer *buffer;
3577 size = sizeof(struct perf_buffer);
3578 size += sizeof(void *);
3580 buffer = kzalloc(size, GFP_KERNEL);
3584 INIT_WORK(&buffer->work, perf_buffer_free_work);
3586 all_buf = vmalloc_user((nr_pages + 1) * PAGE_SIZE);
3590 buffer->user_page = all_buf;
3591 buffer->data_pages[0] = all_buf + PAGE_SIZE;
3592 buffer->page_order = ilog2(nr_pages);
3593 buffer->nr_pages = 1;
3595 perf_buffer_init(buffer, watermark, flags);
3608 static unsigned long perf_data_size(struct perf_buffer *buffer)
3610 return buffer->nr_pages << (PAGE_SHIFT + page_order(buffer));
3613 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
3615 struct perf_event *event = vma->vm_file->private_data;
3616 struct perf_buffer *buffer;
3617 int ret = VM_FAULT_SIGBUS;
3619 if (vmf->flags & FAULT_FLAG_MKWRITE) {
3620 if (vmf->pgoff == 0)
3626 buffer = rcu_dereference(event->buffer);
3630 if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE))
3633 vmf->page = perf_mmap_to_page(buffer, vmf->pgoff);
3637 get_page(vmf->page);
3638 vmf->page->mapping = vma->vm_file->f_mapping;
3639 vmf->page->index = vmf->pgoff;
3648 static void perf_buffer_free_rcu(struct rcu_head *rcu_head)
3650 struct perf_buffer *buffer;
3652 buffer = container_of(rcu_head, struct perf_buffer, rcu_head);
3653 perf_buffer_free(buffer);
3656 static struct perf_buffer *perf_buffer_get(struct perf_event *event)
3658 struct perf_buffer *buffer;
3661 buffer = rcu_dereference(event->buffer);
3663 if (!atomic_inc_not_zero(&buffer->refcount))
3671 static void perf_buffer_put(struct perf_buffer *buffer)
3673 if (!atomic_dec_and_test(&buffer->refcount))
3676 call_rcu(&buffer->rcu_head, perf_buffer_free_rcu);
3679 static void perf_mmap_open(struct vm_area_struct *vma)
3681 struct perf_event *event = vma->vm_file->private_data;
3683 atomic_inc(&event->mmap_count);
3686 static void perf_mmap_close(struct vm_area_struct *vma)
3688 struct perf_event *event = vma->vm_file->private_data;
3690 if (atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex)) {
3691 unsigned long size = perf_data_size(event->buffer);
3692 struct user_struct *user = event->mmap_user;
3693 struct perf_buffer *buffer = event->buffer;
3695 atomic_long_sub((size >> PAGE_SHIFT) + 1, &user->locked_vm);
3696 vma->vm_mm->locked_vm -= event->mmap_locked;
3697 rcu_assign_pointer(event->buffer, NULL);
3698 mutex_unlock(&event->mmap_mutex);
3700 perf_buffer_put(buffer);
3705 static const struct vm_operations_struct perf_mmap_vmops = {
3706 .open = perf_mmap_open,
3707 .close = perf_mmap_close,
3708 .fault = perf_mmap_fault,
3709 .page_mkwrite = perf_mmap_fault,
3712 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
3714 struct perf_event *event = file->private_data;
3715 unsigned long user_locked, user_lock_limit;
3716 struct user_struct *user = current_user();
3717 unsigned long locked, lock_limit;
3718 struct perf_buffer *buffer;
3719 unsigned long vma_size;
3720 unsigned long nr_pages;
3721 long user_extra, extra;
3722 int ret = 0, flags = 0;
3725 * Don't allow mmap() of inherited per-task counters. This would
3726 * create a performance issue due to all children writing to the
3729 if (event->cpu == -1 && event->attr.inherit)
3732 if (!(vma->vm_flags & VM_SHARED))
3735 vma_size = vma->vm_end - vma->vm_start;
3736 nr_pages = (vma_size / PAGE_SIZE) - 1;
3739 * If we have buffer pages ensure they're a power-of-two number, so we
3740 * can do bitmasks instead of modulo.
3742 if (nr_pages != 0 && !is_power_of_2(nr_pages))
3745 if (vma_size != PAGE_SIZE * (1 + nr_pages))
3748 if (vma->vm_pgoff != 0)
3751 WARN_ON_ONCE(event->ctx->parent_ctx);
3752 mutex_lock(&event->mmap_mutex);
3753 if (event->buffer) {
3754 if (event->buffer->nr_pages == nr_pages)
3755 atomic_inc(&event->buffer->refcount);
3761 user_extra = nr_pages + 1;
3762 user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10);
3765 * Increase the limit linearly with more CPUs:
3767 user_lock_limit *= num_online_cpus();
3769 user_locked = atomic_long_read(&user->locked_vm) + user_extra;
3772 if (user_locked > user_lock_limit)
3773 extra = user_locked - user_lock_limit;
3775 lock_limit = rlimit(RLIMIT_MEMLOCK);
3776 lock_limit >>= PAGE_SHIFT;
3777 locked = vma->vm_mm->locked_vm + extra;
3779 if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() &&
3780 !capable(CAP_IPC_LOCK)) {
3785 WARN_ON(event->buffer);
3787 if (vma->vm_flags & VM_WRITE)
3788 flags |= PERF_BUFFER_WRITABLE;
3790 buffer = perf_buffer_alloc(nr_pages, event->attr.wakeup_watermark,
3796 rcu_assign_pointer(event->buffer, buffer);
3798 atomic_long_add(user_extra, &user->locked_vm);
3799 event->mmap_locked = extra;
3800 event->mmap_user = get_current_user();
3801 vma->vm_mm->locked_vm += event->mmap_locked;
3805 atomic_inc(&event->mmap_count);
3806 mutex_unlock(&event->mmap_mutex);
3808 vma->vm_flags |= VM_RESERVED;
3809 vma->vm_ops = &perf_mmap_vmops;
3814 static int perf_fasync(int fd, struct file *filp, int on)
3816 struct inode *inode = filp->f_path.dentry->d_inode;
3817 struct perf_event *event = filp->private_data;
3820 mutex_lock(&inode->i_mutex);
3821 retval = fasync_helper(fd, filp, on, &event->fasync);
3822 mutex_unlock(&inode->i_mutex);
3830 static const struct file_operations perf_fops = {
3831 .llseek = no_llseek,
3832 .release = perf_release,
3835 .unlocked_ioctl = perf_ioctl,
3836 .compat_ioctl = perf_ioctl,
3838 .fasync = perf_fasync,
3844 * If there's data, ensure we set the poll() state and publish everything
3845 * to user-space before waking everybody up.
3848 void perf_event_wakeup(struct perf_event *event)
3850 wake_up_all(&event->waitq);
3852 if (event->pending_kill) {
3853 kill_fasync(&event->fasync, SIGIO, event->pending_kill);
3854 event->pending_kill = 0;
3858 static void perf_pending_event(struct irq_work *entry)
3860 struct perf_event *event = container_of(entry,
3861 struct perf_event, pending);
3863 if (event->pending_disable) {
3864 event->pending_disable = 0;
3865 __perf_event_disable(event);
3868 if (event->pending_wakeup) {
3869 event->pending_wakeup = 0;
3870 perf_event_wakeup(event);
3875 * We assume there is only KVM supporting the callbacks.
3876 * Later on, we might change it to a list if there is
3877 * another virtualization implementation supporting the callbacks.
3879 struct perf_guest_info_callbacks *perf_guest_cbs;
3881 int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
3883 perf_guest_cbs = cbs;
3886 EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks);
3888 int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
3890 perf_guest_cbs = NULL;
3893 EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks);
3898 static bool perf_output_space(struct perf_buffer *buffer, unsigned long tail,
3899 unsigned long offset, unsigned long head)
3903 if (!buffer->writable)
3906 mask = perf_data_size(buffer) - 1;
3908 offset = (offset - tail) & mask;
3909 head = (head - tail) & mask;
3911 if ((int)(head - offset) < 0)
3917 static void perf_output_wakeup(struct perf_output_handle *handle)
3919 atomic_set(&handle->buffer->poll, POLL_IN);
3922 handle->event->pending_wakeup = 1;
3923 irq_work_queue(&handle->event->pending);
3925 perf_event_wakeup(handle->event);
3929 * We need to ensure a later event_id doesn't publish a head when a former
3930 * event isn't done writing. However since we need to deal with NMIs we
3931 * cannot fully serialize things.
3933 * We only publish the head (and generate a wakeup) when the outer-most
3936 static void perf_output_get_handle(struct perf_output_handle *handle)
3938 struct perf_buffer *buffer = handle->buffer;
3941 local_inc(&buffer->nest);
3942 handle->wakeup = local_read(&buffer->wakeup);
3945 static void perf_output_put_handle(struct perf_output_handle *handle)
3947 struct perf_buffer *buffer = handle->buffer;
3951 head = local_read(&buffer->head);
3954 * IRQ/NMI can happen here, which means we can miss a head update.
3957 if (!local_dec_and_test(&buffer->nest))
3961 * Publish the known good head. Rely on the full barrier implied
3962 * by atomic_dec_and_test() order the buffer->head read and this
3965 buffer->user_page->data_head = head;
3968 * Now check if we missed an update, rely on the (compiler)
3969 * barrier in atomic_dec_and_test() to re-read buffer->head.
3971 if (unlikely(head != local_read(&buffer->head))) {
3972 local_inc(&buffer->nest);
3976 if (handle->wakeup != local_read(&buffer->wakeup))
3977 perf_output_wakeup(handle);
3983 __always_inline void perf_output_copy(struct perf_output_handle *handle,
3984 const void *buf, unsigned int len)
3987 unsigned long size = min_t(unsigned long, handle->size, len);
3989 memcpy(handle->addr, buf, size);
3992 handle->addr += size;
3994 handle->size -= size;
3995 if (!handle->size) {
3996 struct perf_buffer *buffer = handle->buffer;
3999 handle->page &= buffer->nr_pages - 1;
4000 handle->addr = buffer->data_pages[handle->page];
4001 handle->size = PAGE_SIZE << page_order(buffer);
4006 static void __perf_event_header__init_id(struct perf_event_header *header,
4007 struct perf_sample_data *data,
4008 struct perf_event *event)
4010 u64 sample_type = event->attr.sample_type;
4012 data->type = sample_type;
4013 header->size += event->id_header_size;
4015 if (sample_type & PERF_SAMPLE_TID) {
4016 /* namespace issues */
4017 data->tid_entry.pid = perf_event_pid(event, current);
4018 data->tid_entry.tid = perf_event_tid(event, current);
4021 if (sample_type & PERF_SAMPLE_TIME)
4022 data->time = perf_clock();
4024 if (sample_type & PERF_SAMPLE_ID)
4025 data->id = primary_event_id(event);
4027 if (sample_type & PERF_SAMPLE_STREAM_ID)
4028 data->stream_id = event->id;
4030 if (sample_type & PERF_SAMPLE_CPU) {
4031 data->cpu_entry.cpu = raw_smp_processor_id();
4032 data->cpu_entry.reserved = 0;
4036 static void perf_event_header__init_id(struct perf_event_header *header,
4037 struct perf_sample_data *data,
4038 struct perf_event *event)
4040 if (event->attr.sample_id_all)
4041 __perf_event_header__init_id(header, data, event);
4044 static void __perf_event__output_id_sample(struct perf_output_handle *handle,
4045 struct perf_sample_data *data)
4047 u64 sample_type = data->type;
4049 if (sample_type & PERF_SAMPLE_TID)
4050 perf_output_put(handle, data->tid_entry);
4052 if (sample_type & PERF_SAMPLE_TIME)
4053 perf_output_put(handle, data->time);
4055 if (sample_type & PERF_SAMPLE_ID)
4056 perf_output_put(handle, data->id);
4058 if (sample_type & PERF_SAMPLE_STREAM_ID)
4059 perf_output_put(handle, data->stream_id);
4061 if (sample_type & PERF_SAMPLE_CPU)
4062 perf_output_put(handle, data->cpu_entry);
4065 static void perf_event__output_id_sample(struct perf_event *event,
4066 struct perf_output_handle *handle,
4067 struct perf_sample_data *sample)
4069 if (event->attr.sample_id_all)
4070 __perf_event__output_id_sample(handle, sample);
4073 int perf_output_begin(struct perf_output_handle *handle,
4074 struct perf_event *event, unsigned int size,
4075 int nmi, int sample)
4077 struct perf_buffer *buffer;
4078 unsigned long tail, offset, head;
4080 struct perf_sample_data sample_data;
4082 struct perf_event_header header;
4089 * For inherited events we send all the output towards the parent.
4092 event = event->parent;
4094 buffer = rcu_dereference(event->buffer);
4098 handle->buffer = buffer;
4099 handle->event = event;
4101 handle->sample = sample;
4103 if (!buffer->nr_pages)
4106 have_lost = local_read(&buffer->lost);
4108 lost_event.header.size = sizeof(lost_event);
4109 perf_event_header__init_id(&lost_event.header, &sample_data,
4111 size += lost_event.header.size;
4114 perf_output_get_handle(handle);
4118 * Userspace could choose to issue a mb() before updating the
4119 * tail pointer. So that all reads will be completed before the
4122 tail = ACCESS_ONCE(buffer->user_page->data_tail);
4124 offset = head = local_read(&buffer->head);
4126 if (unlikely(!perf_output_space(buffer, tail, offset, head)))
4128 } while (local_cmpxchg(&buffer->head, offset, head) != offset);
4130 if (head - local_read(&buffer->wakeup) > buffer->watermark)
4131 local_add(buffer->watermark, &buffer->wakeup);
4133 handle->page = offset >> (PAGE_SHIFT + page_order(buffer));
4134 handle->page &= buffer->nr_pages - 1;
4135 handle->size = offset & ((PAGE_SIZE << page_order(buffer)) - 1);
4136 handle->addr = buffer->data_pages[handle->page];
4137 handle->addr += handle->size;
4138 handle->size = (PAGE_SIZE << page_order(buffer)) - handle->size;
4141 lost_event.header.type = PERF_RECORD_LOST;
4142 lost_event.header.misc = 0;
4143 lost_event.id = event->id;
4144 lost_event.lost = local_xchg(&buffer->lost, 0);
4146 perf_output_put(handle, lost_event);
4147 perf_event__output_id_sample(event, handle, &sample_data);
4153 local_inc(&buffer->lost);
4154 perf_output_put_handle(handle);
4161 void perf_output_end(struct perf_output_handle *handle)
4163 struct perf_event *event = handle->event;
4164 struct perf_buffer *buffer = handle->buffer;
4166 int wakeup_events = event->attr.wakeup_events;
4168 if (handle->sample && wakeup_events) {
4169 int events = local_inc_return(&buffer->events);
4170 if (events >= wakeup_events) {
4171 local_sub(wakeup_events, &buffer->events);
4172 local_inc(&buffer->wakeup);
4176 perf_output_put_handle(handle);
4180 static void perf_output_read_one(struct perf_output_handle *handle,
4181 struct perf_event *event,
4182 u64 enabled, u64 running)
4184 u64 read_format = event->attr.read_format;
4188 values[n++] = perf_event_count(event);
4189 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
4190 values[n++] = enabled +
4191 atomic64_read(&event->child_total_time_enabled);
4193 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
4194 values[n++] = running +
4195 atomic64_read(&event->child_total_time_running);
4197 if (read_format & PERF_FORMAT_ID)
4198 values[n++] = primary_event_id(event);
4200 perf_output_copy(handle, values, n * sizeof(u64));
4204 * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
4206 static void perf_output_read_group(struct perf_output_handle *handle,
4207 struct perf_event *event,
4208 u64 enabled, u64 running)
4210 struct perf_event *leader = event->group_leader, *sub;
4211 u64 read_format = event->attr.read_format;
4215 values[n++] = 1 + leader->nr_siblings;
4217 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
4218 values[n++] = enabled;
4220 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
4221 values[n++] = running;
4223 if (leader != event)
4224 leader->pmu->read(leader);
4226 values[n++] = perf_event_count(leader);
4227 if (read_format & PERF_FORMAT_ID)
4228 values[n++] = primary_event_id(leader);
4230 perf_output_copy(handle, values, n * sizeof(u64));
4232 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
4236 sub->pmu->read(sub);
4238 values[n++] = perf_event_count(sub);
4239 if (read_format & PERF_FORMAT_ID)
4240 values[n++] = primary_event_id(sub);
4242 perf_output_copy(handle, values, n * sizeof(u64));
4246 #define PERF_FORMAT_TOTAL_TIMES (PERF_FORMAT_TOTAL_TIME_ENABLED|\
4247 PERF_FORMAT_TOTAL_TIME_RUNNING)
4249 static void perf_output_read(struct perf_output_handle *handle,
4250 struct perf_event *event)
4252 u64 enabled = 0, running = 0, now, ctx_time;
4253 u64 read_format = event->attr.read_format;
4256 * compute total_time_enabled, total_time_running
4257 * based on snapshot values taken when the event
4258 * was last scheduled in.
4260 * we cannot simply called update_context_time()
4261 * because of locking issue as we are called in
4264 if (read_format & PERF_FORMAT_TOTAL_TIMES) {
4266 ctx_time = event->shadow_ctx_time + now;
4267 enabled = ctx_time - event->tstamp_enabled;
4268 running = ctx_time - event->tstamp_running;
4271 if (event->attr.read_format & PERF_FORMAT_GROUP)
4272 perf_output_read_group(handle, event, enabled, running);
4274 perf_output_read_one(handle, event, enabled, running);
4277 void perf_output_sample(struct perf_output_handle *handle,
4278 struct perf_event_header *header,
4279 struct perf_sample_data *data,
4280 struct perf_event *event)
4282 u64 sample_type = data->type;
4284 perf_output_put(handle, *header);
4286 if (sample_type & PERF_SAMPLE_IP)
4287 perf_output_put(handle, data->ip);
4289 if (sample_type & PERF_SAMPLE_TID)
4290 perf_output_put(handle, data->tid_entry);
4292 if (sample_type & PERF_SAMPLE_TIME)
4293 perf_output_put(handle, data->time);
4295 if (sample_type & PERF_SAMPLE_ADDR)
4296 perf_output_put(handle, data->addr);
4298 if (sample_type & PERF_SAMPLE_ID)
4299 perf_output_put(handle, data->id);
4301 if (sample_type & PERF_SAMPLE_STREAM_ID)
4302 perf_output_put(handle, data->stream_id);
4304 if (sample_type & PERF_SAMPLE_CPU)
4305 perf_output_put(handle, data->cpu_entry);
4307 if (sample_type & PERF_SAMPLE_PERIOD)
4308 perf_output_put(handle, data->period);
4310 if (sample_type & PERF_SAMPLE_READ)
4311 perf_output_read(handle, event);
4313 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
4314 if (data->callchain) {
4317 if (data->callchain)
4318 size += data->callchain->nr;
4320 size *= sizeof(u64);
4322 perf_output_copy(handle, data->callchain, size);
4325 perf_output_put(handle, nr);
4329 if (sample_type & PERF_SAMPLE_RAW) {
4331 perf_output_put(handle, data->raw->size);
4332 perf_output_copy(handle, data->raw->data,
4339 .size = sizeof(u32),
4342 perf_output_put(handle, raw);
4347 void perf_prepare_sample(struct perf_event_header *header,
4348 struct perf_sample_data *data,
4349 struct perf_event *event,
4350 struct pt_regs *regs)
4352 u64 sample_type = event->attr.sample_type;
4354 header->type = PERF_RECORD_SAMPLE;
4355 header->size = sizeof(*header) + event->header_size;
4358 header->misc |= perf_misc_flags(regs);
4360 __perf_event_header__init_id(header, data, event);
4362 if (sample_type & PERF_SAMPLE_IP)
4363 data->ip = perf_instruction_pointer(regs);
4365 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
4368 data->callchain = perf_callchain(regs);
4370 if (data->callchain)
4371 size += data->callchain->nr;
4373 header->size += size * sizeof(u64);
4376 if (sample_type & PERF_SAMPLE_RAW) {
4377 int size = sizeof(u32);
4380 size += data->raw->size;
4382 size += sizeof(u32);
4384 WARN_ON_ONCE(size & (sizeof(u64)-1));
4385 header->size += size;
4389 static void perf_event_output(struct perf_event *event, int nmi,
4390 struct perf_sample_data *data,
4391 struct pt_regs *regs)
4393 struct perf_output_handle handle;
4394 struct perf_event_header header;
4396 /* protect the callchain buffers */
4399 perf_prepare_sample(&header, data, event, regs);
4401 if (perf_output_begin(&handle, event, header.size, nmi, 1))
4404 perf_output_sample(&handle, &header, data, event);
4406 perf_output_end(&handle);
4416 struct perf_read_event {
4417 struct perf_event_header header;
4424 perf_event_read_event(struct perf_event *event,
4425 struct task_struct *task)
4427 struct perf_output_handle handle;
4428 struct perf_sample_data sample;
4429 struct perf_read_event read_event = {
4431 .type = PERF_RECORD_READ,
4433 .size = sizeof(read_event) + event->read_size,
4435 .pid = perf_event_pid(event, task),
4436 .tid = perf_event_tid(event, task),
4440 perf_event_header__init_id(&read_event.header, &sample, event);
4441 ret = perf_output_begin(&handle, event, read_event.header.size, 0, 0);
4445 perf_output_put(&handle, read_event);
4446 perf_output_read(&handle, event);
4447 perf_event__output_id_sample(event, &handle, &sample);
4449 perf_output_end(&handle);
4453 * task tracking -- fork/exit
4455 * enabled by: attr.comm | attr.mmap | attr.mmap_data | attr.task
4458 struct perf_task_event {
4459 struct task_struct *task;
4460 struct perf_event_context *task_ctx;
4463 struct perf_event_header header;
4473 static void perf_event_task_output(struct perf_event *event,
4474 struct perf_task_event *task_event)
4476 struct perf_output_handle handle;
4477 struct perf_sample_data sample;
4478 struct task_struct *task = task_event->task;
4479 int ret, size = task_event->event_id.header.size;
4481 perf_event_header__init_id(&task_event->event_id.header, &sample, event);
4483 ret = perf_output_begin(&handle, event,
4484 task_event->event_id.header.size, 0, 0);
4488 task_event->event_id.pid = perf_event_pid(event, task);
4489 task_event->event_id.ppid = perf_event_pid(event, current);
4491 task_event->event_id.tid = perf_event_tid(event, task);
4492 task_event->event_id.ptid = perf_event_tid(event, current);
4494 perf_output_put(&handle, task_event->event_id);
4496 perf_event__output_id_sample(event, &handle, &sample);
4498 perf_output_end(&handle);
4500 task_event->event_id.header.size = size;
4503 static int perf_event_task_match(struct perf_event *event)
4505 if (event->state < PERF_EVENT_STATE_INACTIVE)
4508 if (!event_filter_match(event))
4511 if (event->attr.comm || event->attr.mmap ||
4512 event->attr.mmap_data || event->attr.task)
4518 static void perf_event_task_ctx(struct perf_event_context *ctx,
4519 struct perf_task_event *task_event)
4521 struct perf_event *event;
4523 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
4524 if (perf_event_task_match(event))
4525 perf_event_task_output(event, task_event);
4529 static void perf_event_task_event(struct perf_task_event *task_event)
4531 struct perf_cpu_context *cpuctx;
4532 struct perf_event_context *ctx;
4537 list_for_each_entry_rcu(pmu, &pmus, entry) {
4538 cpuctx = get_cpu_ptr(pmu->pmu_cpu_context);
4539 if (cpuctx->active_pmu != pmu)
4541 perf_event_task_ctx(&cpuctx->ctx, task_event);
4543 ctx = task_event->task_ctx;
4545 ctxn = pmu->task_ctx_nr;
4548 ctx = rcu_dereference(current->perf_event_ctxp[ctxn]);
4551 perf_event_task_ctx(ctx, task_event);
4553 put_cpu_ptr(pmu->pmu_cpu_context);
4558 static void perf_event_task(struct task_struct *task,
4559 struct perf_event_context *task_ctx,
4562 struct perf_task_event task_event;
4564 if (!atomic_read(&nr_comm_events) &&
4565 !atomic_read(&nr_mmap_events) &&
4566 !atomic_read(&nr_task_events))
4569 task_event = (struct perf_task_event){
4571 .task_ctx = task_ctx,
4574 .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT,
4576 .size = sizeof(task_event.event_id),
4582 .time = perf_clock(),
4586 perf_event_task_event(&task_event);
4589 void perf_event_fork(struct task_struct *task)
4591 perf_event_task(task, NULL, 1);
4598 struct perf_comm_event {
4599 struct task_struct *task;
4604 struct perf_event_header header;
4611 static void perf_event_comm_output(struct perf_event *event,
4612 struct perf_comm_event *comm_event)
4614 struct perf_output_handle handle;
4615 struct perf_sample_data sample;
4616 int size = comm_event->event_id.header.size;
4619 perf_event_header__init_id(&comm_event->event_id.header, &sample, event);
4620 ret = perf_output_begin(&handle, event,
4621 comm_event->event_id.header.size, 0, 0);
4626 comm_event->event_id.pid = perf_event_pid(event, comm_event->task);
4627 comm_event->event_id.tid = perf_event_tid(event, comm_event->task);
4629 perf_output_put(&handle, comm_event->event_id);
4630 perf_output_copy(&handle, comm_event->comm,
4631 comm_event->comm_size);
4633 perf_event__output_id_sample(event, &handle, &sample);
4635 perf_output_end(&handle);
4637 comm_event->event_id.header.size = size;
4640 static int perf_event_comm_match(struct perf_event *event)
4642 if (event->state < PERF_EVENT_STATE_INACTIVE)
4645 if (!event_filter_match(event))
4648 if (event->attr.comm)
4654 static void perf_event_comm_ctx(struct perf_event_context *ctx,
4655 struct perf_comm_event *comm_event)
4657 struct perf_event *event;
4659 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
4660 if (perf_event_comm_match(event))
4661 perf_event_comm_output(event, comm_event);
4665 static void perf_event_comm_event(struct perf_comm_event *comm_event)
4667 struct perf_cpu_context *cpuctx;
4668 struct perf_event_context *ctx;
4669 char comm[TASK_COMM_LEN];
4674 memset(comm, 0, sizeof(comm));
4675 strlcpy(comm, comm_event->task->comm, sizeof(comm));
4676 size = ALIGN(strlen(comm)+1, sizeof(u64));
4678 comm_event->comm = comm;
4679 comm_event->comm_size = size;
4681 comm_event->event_id.header.size = sizeof(comm_event->event_id) + size;
4683 list_for_each_entry_rcu(pmu, &pmus, entry) {
4684 cpuctx = get_cpu_ptr(pmu->pmu_cpu_context);
4685 if (cpuctx->active_pmu != pmu)
4687 perf_event_comm_ctx(&cpuctx->ctx, comm_event);
4689 ctxn = pmu->task_ctx_nr;
4693 ctx = rcu_dereference(current->perf_event_ctxp[ctxn]);
4695 perf_event_comm_ctx(ctx, comm_event);
4697 put_cpu_ptr(pmu->pmu_cpu_context);
4702 void perf_event_comm(struct task_struct *task)
4704 struct perf_comm_event comm_event;
4705 struct perf_event_context *ctx;
4708 for_each_task_context_nr(ctxn) {
4709 ctx = task->perf_event_ctxp[ctxn];
4713 perf_event_enable_on_exec(ctx);
4716 if (!atomic_read(&nr_comm_events))
4719 comm_event = (struct perf_comm_event){
4725 .type = PERF_RECORD_COMM,
4734 perf_event_comm_event(&comm_event);
4741 struct perf_mmap_event {
4742 struct vm_area_struct *vma;
4744 const char *file_name;
4748 struct perf_event_header header;
4758 static void perf_event_mmap_output(struct perf_event *event,
4759 struct perf_mmap_event *mmap_event)
4761 struct perf_output_handle handle;
4762 struct perf_sample_data sample;
4763 int size = mmap_event->event_id.header.size;
4766 perf_event_header__init_id(&mmap_event->event_id.header, &sample, event);
4767 ret = perf_output_begin(&handle, event,
4768 mmap_event->event_id.header.size, 0, 0);
4772 mmap_event->event_id.pid = perf_event_pid(event, current);
4773 mmap_event->event_id.tid = perf_event_tid(event, current);
4775 perf_output_put(&handle, mmap_event->event_id);
4776 perf_output_copy(&handle, mmap_event->file_name,
4777 mmap_event->file_size);
4779 perf_event__output_id_sample(event, &handle, &sample);
4781 perf_output_end(&handle);
4783 mmap_event->event_id.header.size = size;
4786 static int perf_event_mmap_match(struct perf_event *event,
4787 struct perf_mmap_event *mmap_event,
4790 if (event->state < PERF_EVENT_STATE_INACTIVE)
4793 if (!event_filter_match(event))
4796 if ((!executable && event->attr.mmap_data) ||
4797 (executable && event->attr.mmap))
4803 static void perf_event_mmap_ctx(struct perf_event_context *ctx,
4804 struct perf_mmap_event *mmap_event,
4807 struct perf_event *event;
4809 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
4810 if (perf_event_mmap_match(event, mmap_event, executable))
4811 perf_event_mmap_output(event, mmap_event);
4815 static void perf_event_mmap_event(struct perf_mmap_event *mmap_event)
4817 struct perf_cpu_context *cpuctx;
4818 struct perf_event_context *ctx;
4819 struct vm_area_struct *vma = mmap_event->vma;
4820 struct file *file = vma->vm_file;
4828 memset(tmp, 0, sizeof(tmp));
4832 * d_path works from the end of the buffer backwards, so we
4833 * need to add enough zero bytes after the string to handle
4834 * the 64bit alignment we do later.
4836 buf = kzalloc(PATH_MAX + sizeof(u64), GFP_KERNEL);
4838 name = strncpy(tmp, "//enomem", sizeof(tmp));
4841 name = d_path(&file->f_path, buf, PATH_MAX);
4843 name = strncpy(tmp, "//toolong", sizeof(tmp));
4847 if (arch_vma_name(mmap_event->vma)) {
4848 name = strncpy(tmp, arch_vma_name(mmap_event->vma),
4854 name = strncpy(tmp, "[vdso]", sizeof(tmp));
4856 } else if (vma->vm_start <= vma->vm_mm->start_brk &&
4857 vma->vm_end >= vma->vm_mm->brk) {
4858 name = strncpy(tmp, "[heap]", sizeof(tmp));
4860 } else if (vma->vm_start <= vma->vm_mm->start_stack &&
4861 vma->vm_end >= vma->vm_mm->start_stack) {
4862 name = strncpy(tmp, "[stack]", sizeof(tmp));
4866 name = strncpy(tmp, "//anon", sizeof(tmp));
4871 size = ALIGN(strlen(name)+1, sizeof(u64));
4873 mmap_event->file_name = name;
4874 mmap_event->file_size = size;
4876 mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size;
4879 list_for_each_entry_rcu(pmu, &pmus, entry) {
4880 cpuctx = get_cpu_ptr(pmu->pmu_cpu_context);
4881 if (cpuctx->active_pmu != pmu)
4883 perf_event_mmap_ctx(&cpuctx->ctx, mmap_event,
4884 vma->vm_flags & VM_EXEC);
4886 ctxn = pmu->task_ctx_nr;
4890 ctx = rcu_dereference(current->perf_event_ctxp[ctxn]);
4892 perf_event_mmap_ctx(ctx, mmap_event,
4893 vma->vm_flags & VM_EXEC);
4896 put_cpu_ptr(pmu->pmu_cpu_context);
4903 void perf_event_mmap(struct vm_area_struct *vma)
4905 struct perf_mmap_event mmap_event;
4907 if (!atomic_read(&nr_mmap_events))
4910 mmap_event = (struct perf_mmap_event){
4916 .type = PERF_RECORD_MMAP,
4917 .misc = PERF_RECORD_MISC_USER,
4922 .start = vma->vm_start,
4923 .len = vma->vm_end - vma->vm_start,
4924 .pgoff = (u64)vma->vm_pgoff << PAGE_SHIFT,
4928 perf_event_mmap_event(&mmap_event);
4932 * IRQ throttle logging
4935 static void perf_log_throttle(struct perf_event *event, int enable)
4937 struct perf_output_handle handle;
4938 struct perf_sample_data sample;
4942 struct perf_event_header header;
4946 } throttle_event = {
4948 .type = PERF_RECORD_THROTTLE,
4950 .size = sizeof(throttle_event),
4952 .time = perf_clock(),
4953 .id = primary_event_id(event),
4954 .stream_id = event->id,
4958 throttle_event.header.type = PERF_RECORD_UNTHROTTLE;
4960 perf_event_header__init_id(&throttle_event.header, &sample, event);
4962 ret = perf_output_begin(&handle, event,
4963 throttle_event.header.size, 1, 0);
4967 perf_output_put(&handle, throttle_event);
4968 perf_event__output_id_sample(event, &handle, &sample);
4969 perf_output_end(&handle);
4973 * Generic event overflow handling, sampling.
4976 static int __perf_event_overflow(struct perf_event *event, int nmi,
4977 int throttle, struct perf_sample_data *data,
4978 struct pt_regs *regs)
4980 int events = atomic_read(&event->event_limit);
4981 struct hw_perf_event *hwc = &event->hw;
4985 * Non-sampling counters might still use the PMI to fold short
4986 * hardware counters, ignore those.
4988 if (unlikely(!is_sampling_event(event)))
4991 if (unlikely(hwc->interrupts >= max_samples_per_tick)) {
4993 hwc->interrupts = MAX_INTERRUPTS;
4994 perf_log_throttle(event, 0);
5000 if (event->attr.freq) {
5001 u64 now = perf_clock();
5002 s64 delta = now - hwc->freq_time_stamp;
5004 hwc->freq_time_stamp = now;
5006 if (delta > 0 && delta < 2*TICK_NSEC)
5007 perf_adjust_period(event, delta, hwc->last_period);
5011 * XXX event_limit might not quite work as expected on inherited
5015 event->pending_kill = POLL_IN;
5016 if (events && atomic_dec_and_test(&event->event_limit)) {
5018 event->pending_kill = POLL_HUP;
5020 event->pending_disable = 1;
5021 irq_work_queue(&event->pending);
5023 perf_event_disable(event);
5026 if (event->overflow_handler)
5027 event->overflow_handler(event, nmi, data, regs);
5029 perf_event_output(event, nmi, data, regs);
5031 if (event->fasync && event->pending_kill) {
5033 event->pending_wakeup = 1;
5034 irq_work_queue(&event->pending);
5036 perf_event_wakeup(event);
5042 int perf_event_overflow(struct perf_event *event, int nmi,
5043 struct perf_sample_data *data,
5044 struct pt_regs *regs)
5046 return __perf_event_overflow(event, nmi, 1, data, regs);
5050 * Generic software event infrastructure
5053 struct swevent_htable {
5054 struct swevent_hlist *swevent_hlist;
5055 struct mutex hlist_mutex;
5058 /* Recursion avoidance in each contexts */
5059 int recursion[PERF_NR_CONTEXTS];
5062 static DEFINE_PER_CPU(struct swevent_htable, swevent_htable);
5065 * We directly increment event->count and keep a second value in
5066 * event->hw.period_left to count intervals. This period event
5067 * is kept in the range [-sample_period, 0] so that we can use the
5071 static u64 perf_swevent_set_period(struct perf_event *event)
5073 struct hw_perf_event *hwc = &event->hw;
5074 u64 period = hwc->last_period;
5078 hwc->last_period = hwc->sample_period;
5081 old = val = local64_read(&hwc->period_left);
5085 nr = div64_u64(period + val, period);
5086 offset = nr * period;
5088 if (local64_cmpxchg(&hwc->period_left, old, val) != old)
5094 static void perf_swevent_overflow(struct perf_event *event, u64 overflow,
5095 int nmi, struct perf_sample_data *data,
5096 struct pt_regs *regs)
5098 struct hw_perf_event *hwc = &event->hw;
5101 data->period = event->hw.last_period;
5103 overflow = perf_swevent_set_period(event);
5105 if (hwc->interrupts == MAX_INTERRUPTS)
5108 for (; overflow; overflow--) {
5109 if (__perf_event_overflow(event, nmi, throttle,
5112 * We inhibit the overflow from happening when
5113 * hwc->interrupts == MAX_INTERRUPTS.
5121 static void perf_swevent_event(struct perf_event *event, u64 nr,
5122 int nmi, struct perf_sample_data *data,
5123 struct pt_regs *regs)
5125 struct hw_perf_event *hwc = &event->hw;
5127 local64_add(nr, &event->count);
5132 if (!is_sampling_event(event))
5135 if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq)
5136 return perf_swevent_overflow(event, 1, nmi, data, regs);
5138 if (local64_add_negative(nr, &hwc->period_left))
5141 perf_swevent_overflow(event, 0, nmi, data, regs);
5144 static int perf_exclude_event(struct perf_event *event,
5145 struct pt_regs *regs)
5147 if (event->hw.state & PERF_HES_STOPPED)
5151 if (event->attr.exclude_user && user_mode(regs))
5154 if (event->attr.exclude_kernel && !user_mode(regs))
5161 static int perf_swevent_match(struct perf_event *event,
5162 enum perf_type_id type,
5164 struct perf_sample_data *data,
5165 struct pt_regs *regs)
5167 if (event->attr.type != type)
5170 if (event->attr.config != event_id)
5173 if (perf_exclude_event(event, regs))
5179 static inline u64 swevent_hash(u64 type, u32 event_id)
5181 u64 val = event_id | (type << 32);
5183 return hash_64(val, SWEVENT_HLIST_BITS);
5186 static inline struct hlist_head *
5187 __find_swevent_head(struct swevent_hlist *hlist, u64 type, u32 event_id)
5189 u64 hash = swevent_hash(type, event_id);
5191 return &hlist->heads[hash];
5194 /* For the read side: events when they trigger */
5195 static inline struct hlist_head *
5196 find_swevent_head_rcu(struct swevent_htable *swhash, u64 type, u32 event_id)
5198 struct swevent_hlist *hlist;
5200 hlist = rcu_dereference(swhash->swevent_hlist);
5204 return __find_swevent_head(hlist, type, event_id);
5207 /* For the event head insertion and removal in the hlist */
5208 static inline struct hlist_head *
5209 find_swevent_head(struct swevent_htable *swhash, struct perf_event *event)
5211 struct swevent_hlist *hlist;
5212 u32 event_id = event->attr.config;
5213 u64 type = event->attr.type;
5216 * Event scheduling is always serialized against hlist allocation
5217 * and release. Which makes the protected version suitable here.
5218 * The context lock guarantees that.
5220 hlist = rcu_dereference_protected(swhash->swevent_hlist,
5221 lockdep_is_held(&event->ctx->lock));
5225 return __find_swevent_head(hlist, type, event_id);
5228 static void do_perf_sw_event(enum perf_type_id type, u32 event_id,
5230 struct perf_sample_data *data,
5231 struct pt_regs *regs)
5233 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
5234 struct perf_event *event;
5235 struct hlist_node *node;
5236 struct hlist_head *head;
5239 head = find_swevent_head_rcu(swhash, type, event_id);
5243 hlist_for_each_entry_rcu(event, node, head, hlist_entry) {
5244 if (perf_swevent_match(event, type, event_id, data, regs))
5245 perf_swevent_event(event, nr, nmi, data, regs);
5251 int perf_swevent_get_recursion_context(void)
5253 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
5255 return get_recursion_context(swhash->recursion);
5257 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context);
5259 inline void perf_swevent_put_recursion_context(int rctx)
5261 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
5263 put_recursion_context(swhash->recursion, rctx);
5266 void __perf_sw_event(u32 event_id, u64 nr, int nmi,
5267 struct pt_regs *regs, u64 addr)
5269 struct perf_sample_data data;
5272 preempt_disable_notrace();
5273 rctx = perf_swevent_get_recursion_context();
5277 perf_sample_data_init(&data, addr);
5279 do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, nmi, &data, regs);
5281 perf_swevent_put_recursion_context(rctx);
5282 preempt_enable_notrace();
5285 static void perf_swevent_read(struct perf_event *event)
5289 static int perf_swevent_add(struct perf_event *event, int flags)
5291 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
5292 struct hw_perf_event *hwc = &event->hw;
5293 struct hlist_head *head;
5295 if (is_sampling_event(event)) {
5296 hwc->last_period = hwc->sample_period;
5297 perf_swevent_set_period(event);
5300 hwc->state = !(flags & PERF_EF_START);
5302 head = find_swevent_head(swhash, event);
5303 if (WARN_ON_ONCE(!head))
5306 hlist_add_head_rcu(&event->hlist_entry, head);
5311 static void perf_swevent_del(struct perf_event *event, int flags)
5313 hlist_del_rcu(&event->hlist_entry);
5316 static void perf_swevent_start(struct perf_event *event, int flags)
5318 event->hw.state = 0;
5321 static void perf_swevent_stop(struct perf_event *event, int flags)
5323 event->hw.state = PERF_HES_STOPPED;
5326 /* Deref the hlist from the update side */
5327 static inline struct swevent_hlist *
5328 swevent_hlist_deref(struct swevent_htable *swhash)
5330 return rcu_dereference_protected(swhash->swevent_hlist,
5331 lockdep_is_held(&swhash->hlist_mutex));
5334 static void swevent_hlist_release(struct swevent_htable *swhash)
5336 struct swevent_hlist *hlist = swevent_hlist_deref(swhash);
5341 rcu_assign_pointer(swhash->swevent_hlist, NULL);
5342 kfree_rcu(hlist, rcu_head);
5345 static void swevent_hlist_put_cpu(struct perf_event *event, int cpu)
5347 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
5349 mutex_lock(&swhash->hlist_mutex);
5351 if (!--swhash->hlist_refcount)
5352 swevent_hlist_release(swhash);
5354 mutex_unlock(&swhash->hlist_mutex);
5357 static void swevent_hlist_put(struct perf_event *event)
5361 if (event->cpu != -1) {
5362 swevent_hlist_put_cpu(event, event->cpu);
5366 for_each_possible_cpu(cpu)
5367 swevent_hlist_put_cpu(event, cpu);
5370 static int swevent_hlist_get_cpu(struct perf_event *event, int cpu)
5372 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
5375 mutex_lock(&swhash->hlist_mutex);
5377 if (!swevent_hlist_deref(swhash) && cpu_online(cpu)) {
5378 struct swevent_hlist *hlist;
5380 hlist = kzalloc(sizeof(*hlist), GFP_KERNEL);
5385 rcu_assign_pointer(swhash->swevent_hlist, hlist);
5387 swhash->hlist_refcount++;
5389 mutex_unlock(&swhash->hlist_mutex);
5394 static int swevent_hlist_get(struct perf_event *event)
5397 int cpu, failed_cpu;
5399 if (event->cpu != -1)
5400 return swevent_hlist_get_cpu(event, event->cpu);
5403 for_each_possible_cpu(cpu) {
5404 err = swevent_hlist_get_cpu(event, cpu);
5414 for_each_possible_cpu(cpu) {
5415 if (cpu == failed_cpu)
5417 swevent_hlist_put_cpu(event, cpu);
5424 struct jump_label_key perf_swevent_enabled[PERF_COUNT_SW_MAX];
5426 static void sw_perf_event_destroy(struct perf_event *event)
5428 u64 event_id = event->attr.config;
5430 WARN_ON(event->parent);
5432 jump_label_dec(&perf_swevent_enabled[event_id]);
5433 swevent_hlist_put(event);
5436 static int perf_swevent_init(struct perf_event *event)
5438 int event_id = event->attr.config;
5440 if (event->attr.type != PERF_TYPE_SOFTWARE)
5444 case PERF_COUNT_SW_CPU_CLOCK:
5445 case PERF_COUNT_SW_TASK_CLOCK:
5452 if (event_id >= PERF_COUNT_SW_MAX)
5455 if (!event->parent) {
5458 err = swevent_hlist_get(event);
5462 jump_label_inc(&perf_swevent_enabled[event_id]);
5463 event->destroy = sw_perf_event_destroy;
5469 static struct pmu perf_swevent = {
5470 .task_ctx_nr = perf_sw_context,
5472 .event_init = perf_swevent_init,
5473 .add = perf_swevent_add,
5474 .del = perf_swevent_del,
5475 .start = perf_swevent_start,
5476 .stop = perf_swevent_stop,
5477 .read = perf_swevent_read,
5480 #ifdef CONFIG_EVENT_TRACING
5482 static int perf_tp_filter_match(struct perf_event *event,
5483 struct perf_sample_data *data)
5485 void *record = data->raw->data;
5487 if (likely(!event->filter) || filter_match_preds(event->filter, record))
5492 static int perf_tp_event_match(struct perf_event *event,
5493 struct perf_sample_data *data,
5494 struct pt_regs *regs)
5496 if (event->hw.state & PERF_HES_STOPPED)
5499 * All tracepoints are from kernel-space.
5501 if (event->attr.exclude_kernel)
5504 if (!perf_tp_filter_match(event, data))
5510 void perf_tp_event(u64 addr, u64 count, void *record, int entry_size,
5511 struct pt_regs *regs, struct hlist_head *head, int rctx)
5513 struct perf_sample_data data;
5514 struct perf_event *event;
5515 struct hlist_node *node;
5517 struct perf_raw_record raw = {
5522 perf_sample_data_init(&data, addr);
5525 hlist_for_each_entry_rcu(event, node, head, hlist_entry) {
5526 if (perf_tp_event_match(event, &data, regs))
5527 perf_swevent_event(event, count, 1, &data, regs);
5530 perf_swevent_put_recursion_context(rctx);
5532 EXPORT_SYMBOL_GPL(perf_tp_event);
5534 static void tp_perf_event_destroy(struct perf_event *event)
5536 perf_trace_destroy(event);
5539 static int perf_tp_event_init(struct perf_event *event)
5543 if (event->attr.type != PERF_TYPE_TRACEPOINT)
5546 err = perf_trace_init(event);
5550 event->destroy = tp_perf_event_destroy;
5555 static struct pmu perf_tracepoint = {
5556 .task_ctx_nr = perf_sw_context,
5558 .event_init = perf_tp_event_init,
5559 .add = perf_trace_add,
5560 .del = perf_trace_del,
5561 .start = perf_swevent_start,
5562 .stop = perf_swevent_stop,
5563 .read = perf_swevent_read,
5566 static inline void perf_tp_register(void)
5568 perf_pmu_register(&perf_tracepoint, "tracepoint", PERF_TYPE_TRACEPOINT);
5571 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
5576 if (event->attr.type != PERF_TYPE_TRACEPOINT)
5579 filter_str = strndup_user(arg, PAGE_SIZE);
5580 if (IS_ERR(filter_str))
5581 return PTR_ERR(filter_str);
5583 ret = ftrace_profile_set_filter(event, event->attr.config, filter_str);
5589 static void perf_event_free_filter(struct perf_event *event)
5591 ftrace_profile_free_filter(event);
5596 static inline void perf_tp_register(void)
5600 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
5605 static void perf_event_free_filter(struct perf_event *event)
5609 #endif /* CONFIG_EVENT_TRACING */
5611 #ifdef CONFIG_HAVE_HW_BREAKPOINT
5612 void perf_bp_event(struct perf_event *bp, void *data)
5614 struct perf_sample_data sample;
5615 struct pt_regs *regs = data;
5617 perf_sample_data_init(&sample, bp->attr.bp_addr);
5619 if (!bp->hw.state && !perf_exclude_event(bp, regs))
5620 perf_swevent_event(bp, 1, 1, &sample, regs);
5625 * hrtimer based swevent callback
5628 static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer)
5630 enum hrtimer_restart ret = HRTIMER_RESTART;
5631 struct perf_sample_data data;
5632 struct pt_regs *regs;
5633 struct perf_event *event;
5636 event = container_of(hrtimer, struct perf_event, hw.hrtimer);
5638 if (event->state != PERF_EVENT_STATE_ACTIVE)
5639 return HRTIMER_NORESTART;
5641 event->pmu->read(event);
5643 perf_sample_data_init(&data, 0);
5644 data.period = event->hw.last_period;
5645 regs = get_irq_regs();
5647 if (regs && !perf_exclude_event(event, regs)) {
5648 if (!(event->attr.exclude_idle && current->pid == 0))
5649 if (perf_event_overflow(event, 0, &data, regs))
5650 ret = HRTIMER_NORESTART;
5653 period = max_t(u64, 10000, event->hw.sample_period);
5654 hrtimer_forward_now(hrtimer, ns_to_ktime(period));
5659 static void perf_swevent_start_hrtimer(struct perf_event *event)
5661 struct hw_perf_event *hwc = &event->hw;
5664 if (!is_sampling_event(event))
5667 period = local64_read(&hwc->period_left);
5672 local64_set(&hwc->period_left, 0);
5674 period = max_t(u64, 10000, hwc->sample_period);
5676 __hrtimer_start_range_ns(&hwc->hrtimer,
5677 ns_to_ktime(period), 0,
5678 HRTIMER_MODE_REL_PINNED, 0);
5681 static void perf_swevent_cancel_hrtimer(struct perf_event *event)
5683 struct hw_perf_event *hwc = &event->hw;
5685 if (is_sampling_event(event)) {
5686 ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer);
5687 local64_set(&hwc->period_left, ktime_to_ns(remaining));
5689 hrtimer_cancel(&hwc->hrtimer);
5693 static void perf_swevent_init_hrtimer(struct perf_event *event)
5695 struct hw_perf_event *hwc = &event->hw;
5697 if (!is_sampling_event(event))
5700 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
5701 hwc->hrtimer.function = perf_swevent_hrtimer;
5704 * Since hrtimers have a fixed rate, we can do a static freq->period
5705 * mapping and avoid the whole period adjust feedback stuff.
5707 if (event->attr.freq) {
5708 long freq = event->attr.sample_freq;
5710 event->attr.sample_period = NSEC_PER_SEC / freq;
5711 hwc->sample_period = event->attr.sample_period;
5712 local64_set(&hwc->period_left, hwc->sample_period);
5713 event->attr.freq = 0;
5718 * Software event: cpu wall time clock
5721 static void cpu_clock_event_update(struct perf_event *event)
5726 now = local_clock();
5727 prev = local64_xchg(&event->hw.prev_count, now);
5728 local64_add(now - prev, &event->count);
5731 static void cpu_clock_event_start(struct perf_event *event, int flags)
5733 local64_set(&event->hw.prev_count, local_clock());
5734 perf_swevent_start_hrtimer(event);
5737 static void cpu_clock_event_stop(struct perf_event *event, int flags)
5739 perf_swevent_cancel_hrtimer(event);
5740 cpu_clock_event_update(event);
5743 static int cpu_clock_event_add(struct perf_event *event, int flags)
5745 if (flags & PERF_EF_START)
5746 cpu_clock_event_start(event, flags);
5751 static void cpu_clock_event_del(struct perf_event *event, int flags)
5753 cpu_clock_event_stop(event, flags);
5756 static void cpu_clock_event_read(struct perf_event *event)
5758 cpu_clock_event_update(event);
5761 static int cpu_clock_event_init(struct perf_event *event)
5763 if (event->attr.type != PERF_TYPE_SOFTWARE)
5766 if (event->attr.config != PERF_COUNT_SW_CPU_CLOCK)
5769 perf_swevent_init_hrtimer(event);
5774 static struct pmu perf_cpu_clock = {
5775 .task_ctx_nr = perf_sw_context,
5777 .event_init = cpu_clock_event_init,
5778 .add = cpu_clock_event_add,
5779 .del = cpu_clock_event_del,
5780 .start = cpu_clock_event_start,
5781 .stop = cpu_clock_event_stop,
5782 .read = cpu_clock_event_read,
5786 * Software event: task time clock
5789 static void task_clock_event_update(struct perf_event *event, u64 now)
5794 prev = local64_xchg(&event->hw.prev_count, now);
5796 local64_add(delta, &event->count);
5799 static void task_clock_event_start(struct perf_event *event, int flags)
5801 local64_set(&event->hw.prev_count, event->ctx->time);
5802 perf_swevent_start_hrtimer(event);
5805 static void task_clock_event_stop(struct perf_event *event, int flags)
5807 perf_swevent_cancel_hrtimer(event);
5808 task_clock_event_update(event, event->ctx->time);
5811 static int task_clock_event_add(struct perf_event *event, int flags)
5813 if (flags & PERF_EF_START)
5814 task_clock_event_start(event, flags);
5819 static void task_clock_event_del(struct perf_event *event, int flags)
5821 task_clock_event_stop(event, PERF_EF_UPDATE);
5824 static void task_clock_event_read(struct perf_event *event)
5826 u64 now = perf_clock();
5827 u64 delta = now - event->ctx->timestamp;
5828 u64 time = event->ctx->time + delta;
5830 task_clock_event_update(event, time);
5833 static int task_clock_event_init(struct perf_event *event)
5835 if (event->attr.type != PERF_TYPE_SOFTWARE)
5838 if (event->attr.config != PERF_COUNT_SW_TASK_CLOCK)
5841 perf_swevent_init_hrtimer(event);
5846 static struct pmu perf_task_clock = {
5847 .task_ctx_nr = perf_sw_context,
5849 .event_init = task_clock_event_init,
5850 .add = task_clock_event_add,
5851 .del = task_clock_event_del,
5852 .start = task_clock_event_start,
5853 .stop = task_clock_event_stop,
5854 .read = task_clock_event_read,
5857 static void perf_pmu_nop_void(struct pmu *pmu)
5861 static int perf_pmu_nop_int(struct pmu *pmu)
5866 static void perf_pmu_start_txn(struct pmu *pmu)
5868 perf_pmu_disable(pmu);
5871 static int perf_pmu_commit_txn(struct pmu *pmu)
5873 perf_pmu_enable(pmu);
5877 static void perf_pmu_cancel_txn(struct pmu *pmu)
5879 perf_pmu_enable(pmu);
5883 * Ensures all contexts with the same task_ctx_nr have the same
5884 * pmu_cpu_context too.
5886 static void *find_pmu_context(int ctxn)
5893 list_for_each_entry(pmu, &pmus, entry) {
5894 if (pmu->task_ctx_nr == ctxn)
5895 return pmu->pmu_cpu_context;
5901 static void update_pmu_context(struct pmu *pmu, struct pmu *old_pmu)
5905 for_each_possible_cpu(cpu) {
5906 struct perf_cpu_context *cpuctx;
5908 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
5910 if (cpuctx->active_pmu == old_pmu)
5911 cpuctx->active_pmu = pmu;
5915 static void free_pmu_context(struct pmu *pmu)
5919 mutex_lock(&pmus_lock);
5921 * Like a real lame refcount.
5923 list_for_each_entry(i, &pmus, entry) {
5924 if (i->pmu_cpu_context == pmu->pmu_cpu_context) {
5925 update_pmu_context(i, pmu);
5930 free_percpu(pmu->pmu_cpu_context);
5932 mutex_unlock(&pmus_lock);
5934 static struct idr pmu_idr;
5937 type_show(struct device *dev, struct device_attribute *attr, char *page)
5939 struct pmu *pmu = dev_get_drvdata(dev);
5941 return snprintf(page, PAGE_SIZE-1, "%d\n", pmu->type);
5944 static struct device_attribute pmu_dev_attrs[] = {
5949 static int pmu_bus_running;
5950 static struct bus_type pmu_bus = {
5951 .name = "event_source",
5952 .dev_attrs = pmu_dev_attrs,
5955 static void pmu_dev_release(struct device *dev)
5960 static int pmu_dev_alloc(struct pmu *pmu)
5964 pmu->dev = kzalloc(sizeof(struct device), GFP_KERNEL);
5968 device_initialize(pmu->dev);
5969 ret = dev_set_name(pmu->dev, "%s", pmu->name);
5973 dev_set_drvdata(pmu->dev, pmu);
5974 pmu->dev->bus = &pmu_bus;
5975 pmu->dev->release = pmu_dev_release;
5976 ret = device_add(pmu->dev);
5984 put_device(pmu->dev);
5988 static struct lock_class_key cpuctx_mutex;
5990 int perf_pmu_register(struct pmu *pmu, char *name, int type)
5994 mutex_lock(&pmus_lock);
5996 pmu->pmu_disable_count = alloc_percpu(int);
5997 if (!pmu->pmu_disable_count)
6006 int err = idr_pre_get(&pmu_idr, GFP_KERNEL);
6010 err = idr_get_new_above(&pmu_idr, pmu, PERF_TYPE_MAX, &type);
6018 if (pmu_bus_running) {
6019 ret = pmu_dev_alloc(pmu);
6025 pmu->pmu_cpu_context = find_pmu_context(pmu->task_ctx_nr);
6026 if (pmu->pmu_cpu_context)
6027 goto got_cpu_context;
6029 pmu->pmu_cpu_context = alloc_percpu(struct perf_cpu_context);
6030 if (!pmu->pmu_cpu_context)
6033 for_each_possible_cpu(cpu) {
6034 struct perf_cpu_context *cpuctx;
6036 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
6037 __perf_event_init_context(&cpuctx->ctx);
6038 lockdep_set_class(&cpuctx->ctx.mutex, &cpuctx_mutex);
6039 cpuctx->ctx.type = cpu_context;
6040 cpuctx->ctx.pmu = pmu;
6041 cpuctx->jiffies_interval = 1;
6042 INIT_LIST_HEAD(&cpuctx->rotation_list);
6043 cpuctx->active_pmu = pmu;
6047 if (!pmu->start_txn) {
6048 if (pmu->pmu_enable) {
6050 * If we have pmu_enable/pmu_disable calls, install
6051 * transaction stubs that use that to try and batch
6052 * hardware accesses.
6054 pmu->start_txn = perf_pmu_start_txn;
6055 pmu->commit_txn = perf_pmu_commit_txn;
6056 pmu->cancel_txn = perf_pmu_cancel_txn;
6058 pmu->start_txn = perf_pmu_nop_void;
6059 pmu->commit_txn = perf_pmu_nop_int;
6060 pmu->cancel_txn = perf_pmu_nop_void;
6064 if (!pmu->pmu_enable) {
6065 pmu->pmu_enable = perf_pmu_nop_void;
6066 pmu->pmu_disable = perf_pmu_nop_void;
6069 list_add_rcu(&pmu->entry, &pmus);
6072 mutex_unlock(&pmus_lock);
6077 device_del(pmu->dev);
6078 put_device(pmu->dev);
6081 if (pmu->type >= PERF_TYPE_MAX)
6082 idr_remove(&pmu_idr, pmu->type);
6085 free_percpu(pmu->pmu_disable_count);
6089 void perf_pmu_unregister(struct pmu *pmu)
6091 mutex_lock(&pmus_lock);
6092 list_del_rcu(&pmu->entry);
6093 mutex_unlock(&pmus_lock);
6096 * We dereference the pmu list under both SRCU and regular RCU, so
6097 * synchronize against both of those.
6099 synchronize_srcu(&pmus_srcu);
6102 free_percpu(pmu->pmu_disable_count);
6103 if (pmu->type >= PERF_TYPE_MAX)
6104 idr_remove(&pmu_idr, pmu->type);
6105 device_del(pmu->dev);
6106 put_device(pmu->dev);
6107 free_pmu_context(pmu);
6110 struct pmu *perf_init_event(struct perf_event *event)
6112 struct pmu *pmu = NULL;
6116 idx = srcu_read_lock(&pmus_srcu);
6119 pmu = idr_find(&pmu_idr, event->attr.type);
6122 ret = pmu->event_init(event);
6128 list_for_each_entry_rcu(pmu, &pmus, entry) {
6129 ret = pmu->event_init(event);
6133 if (ret != -ENOENT) {
6138 pmu = ERR_PTR(-ENOENT);
6140 srcu_read_unlock(&pmus_srcu, idx);
6146 * Allocate and initialize a event structure
6148 static struct perf_event *
6149 perf_event_alloc(struct perf_event_attr *attr, int cpu,
6150 struct task_struct *task,
6151 struct perf_event *group_leader,
6152 struct perf_event *parent_event,
6153 perf_overflow_handler_t overflow_handler)
6156 struct perf_event *event;
6157 struct hw_perf_event *hwc;
6160 if ((unsigned)cpu >= nr_cpu_ids) {
6161 if (!task || cpu != -1)
6162 return ERR_PTR(-EINVAL);
6165 event = kzalloc(sizeof(*event), GFP_KERNEL);
6167 return ERR_PTR(-ENOMEM);
6170 * Single events are their own group leaders, with an
6171 * empty sibling list:
6174 group_leader = event;
6176 mutex_init(&event->child_mutex);
6177 INIT_LIST_HEAD(&event->child_list);
6179 INIT_LIST_HEAD(&event->group_entry);
6180 INIT_LIST_HEAD(&event->event_entry);
6181 INIT_LIST_HEAD(&event->sibling_list);
6182 init_waitqueue_head(&event->waitq);
6183 init_irq_work(&event->pending, perf_pending_event);
6185 mutex_init(&event->mmap_mutex);
6188 event->attr = *attr;
6189 event->group_leader = group_leader;
6193 event->parent = parent_event;
6195 event->ns = get_pid_ns(current->nsproxy->pid_ns);
6196 event->id = atomic64_inc_return(&perf_event_id);
6198 event->state = PERF_EVENT_STATE_INACTIVE;
6201 event->attach_state = PERF_ATTACH_TASK;
6202 #ifdef CONFIG_HAVE_HW_BREAKPOINT
6204 * hw_breakpoint is a bit difficult here..
6206 if (attr->type == PERF_TYPE_BREAKPOINT)
6207 event->hw.bp_target = task;
6211 if (!overflow_handler && parent_event)
6212 overflow_handler = parent_event->overflow_handler;
6214 event->overflow_handler = overflow_handler;
6217 event->state = PERF_EVENT_STATE_OFF;
6222 hwc->sample_period = attr->sample_period;
6223 if (attr->freq && attr->sample_freq)
6224 hwc->sample_period = 1;
6225 hwc->last_period = hwc->sample_period;
6227 local64_set(&hwc->period_left, hwc->sample_period);
6230 * we currently do not support PERF_FORMAT_GROUP on inherited events
6232 if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP))
6235 pmu = perf_init_event(event);
6241 else if (IS_ERR(pmu))
6246 put_pid_ns(event->ns);
6248 return ERR_PTR(err);
6253 if (!event->parent) {
6254 if (event->attach_state & PERF_ATTACH_TASK)
6255 jump_label_inc(&perf_sched_events);
6256 if (event->attr.mmap || event->attr.mmap_data)
6257 atomic_inc(&nr_mmap_events);
6258 if (event->attr.comm)
6259 atomic_inc(&nr_comm_events);
6260 if (event->attr.task)
6261 atomic_inc(&nr_task_events);
6262 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) {
6263 err = get_callchain_buffers();
6266 return ERR_PTR(err);
6274 static int perf_copy_attr(struct perf_event_attr __user *uattr,
6275 struct perf_event_attr *attr)
6280 if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0))
6284 * zero the full structure, so that a short copy will be nice.
6286 memset(attr, 0, sizeof(*attr));
6288 ret = get_user(size, &uattr->size);
6292 if (size > PAGE_SIZE) /* silly large */
6295 if (!size) /* abi compat */
6296 size = PERF_ATTR_SIZE_VER0;
6298 if (size < PERF_ATTR_SIZE_VER0)
6302 * If we're handed a bigger struct than we know of,
6303 * ensure all the unknown bits are 0 - i.e. new
6304 * user-space does not rely on any kernel feature
6305 * extensions we dont know about yet.
6307 if (size > sizeof(*attr)) {
6308 unsigned char __user *addr;
6309 unsigned char __user *end;
6312 addr = (void __user *)uattr + sizeof(*attr);
6313 end = (void __user *)uattr + size;
6315 for (; addr < end; addr++) {
6316 ret = get_user(val, addr);
6322 size = sizeof(*attr);
6325 ret = copy_from_user(attr, uattr, size);
6330 * If the type exists, the corresponding creation will verify
6333 if (attr->type >= PERF_TYPE_MAX)
6336 if (attr->__reserved_1)
6339 if (attr->sample_type & ~(PERF_SAMPLE_MAX-1))
6342 if (attr->read_format & ~(PERF_FORMAT_MAX-1))
6349 put_user(sizeof(*attr), &uattr->size);
6355 perf_event_set_output(struct perf_event *event, struct perf_event *output_event)
6357 struct perf_buffer *buffer = NULL, *old_buffer = NULL;
6363 /* don't allow circular references */
6364 if (event == output_event)
6368 * Don't allow cross-cpu buffers
6370 if (output_event->cpu != event->cpu)
6374 * If its not a per-cpu buffer, it must be the same task.
6376 if (output_event->cpu == -1 && output_event->ctx != event->ctx)
6380 mutex_lock(&event->mmap_mutex);
6381 /* Can't redirect output if we've got an active mmap() */
6382 if (atomic_read(&event->mmap_count))
6386 /* get the buffer we want to redirect to */
6387 buffer = perf_buffer_get(output_event);
6392 old_buffer = event->buffer;
6393 rcu_assign_pointer(event->buffer, buffer);
6396 mutex_unlock(&event->mmap_mutex);
6399 perf_buffer_put(old_buffer);
6405 * sys_perf_event_open - open a performance event, associate it to a task/cpu
6407 * @attr_uptr: event_id type attributes for monitoring/sampling
6410 * @group_fd: group leader event fd
6412 SYSCALL_DEFINE5(perf_event_open,
6413 struct perf_event_attr __user *, attr_uptr,
6414 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
6416 struct perf_event *group_leader = NULL, *output_event = NULL;
6417 struct perf_event *event, *sibling;
6418 struct perf_event_attr attr;
6419 struct perf_event_context *ctx;
6420 struct file *event_file = NULL;
6421 struct file *group_file = NULL;
6422 struct task_struct *task = NULL;
6426 int fput_needed = 0;
6429 /* for future expandability... */
6430 if (flags & ~PERF_FLAG_ALL)
6433 err = perf_copy_attr(attr_uptr, &attr);
6437 if (!attr.exclude_kernel) {
6438 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
6443 if (attr.sample_freq > sysctl_perf_event_sample_rate)
6448 * In cgroup mode, the pid argument is used to pass the fd
6449 * opened to the cgroup directory in cgroupfs. The cpu argument
6450 * designates the cpu on which to monitor threads from that
6453 if ((flags & PERF_FLAG_PID_CGROUP) && (pid == -1 || cpu == -1))
6456 event_fd = get_unused_fd_flags(O_RDWR);
6460 if (group_fd != -1) {
6461 group_leader = perf_fget_light(group_fd, &fput_needed);
6462 if (IS_ERR(group_leader)) {
6463 err = PTR_ERR(group_leader);
6466 group_file = group_leader->filp;
6467 if (flags & PERF_FLAG_FD_OUTPUT)
6468 output_event = group_leader;
6469 if (flags & PERF_FLAG_FD_NO_GROUP)
6470 group_leader = NULL;
6473 if (pid != -1 && !(flags & PERF_FLAG_PID_CGROUP)) {
6474 task = find_lively_task_by_vpid(pid);
6476 err = PTR_ERR(task);
6481 event = perf_event_alloc(&attr, cpu, task, group_leader, NULL, NULL);
6482 if (IS_ERR(event)) {
6483 err = PTR_ERR(event);
6487 if (flags & PERF_FLAG_PID_CGROUP) {
6488 err = perf_cgroup_connect(pid, event, &attr, group_leader);
6493 * - that has cgroup constraint on event->cpu
6494 * - that may need work on context switch
6496 atomic_inc(&per_cpu(perf_cgroup_events, event->cpu));
6497 jump_label_inc(&perf_sched_events);
6501 * Special case software events and allow them to be part of
6502 * any hardware group.
6507 (is_software_event(event) != is_software_event(group_leader))) {
6508 if (is_software_event(event)) {
6510 * If event and group_leader are not both a software
6511 * event, and event is, then group leader is not.
6513 * Allow the addition of software events to !software
6514 * groups, this is safe because software events never
6517 pmu = group_leader->pmu;
6518 } else if (is_software_event(group_leader) &&
6519 (group_leader->group_flags & PERF_GROUP_SOFTWARE)) {
6521 * In case the group is a pure software group, and we
6522 * try to add a hardware event, move the whole group to
6523 * the hardware context.
6530 * Get the target context (task or percpu):
6532 ctx = find_get_context(pmu, task, cpu);
6539 put_task_struct(task);
6544 * Look up the group leader (we will attach this event to it):
6550 * Do not allow a recursive hierarchy (this new sibling
6551 * becoming part of another group-sibling):
6553 if (group_leader->group_leader != group_leader)
6556 * Do not allow to attach to a group in a different
6557 * task or CPU context:
6560 if (group_leader->ctx->type != ctx->type)
6563 if (group_leader->ctx != ctx)
6568 * Only a group leader can be exclusive or pinned
6570 if (attr.exclusive || attr.pinned)
6575 err = perf_event_set_output(event, output_event);
6580 event_file = anon_inode_getfile("[perf_event]", &perf_fops, event, O_RDWR);
6581 if (IS_ERR(event_file)) {
6582 err = PTR_ERR(event_file);
6587 struct perf_event_context *gctx = group_leader->ctx;
6589 mutex_lock(&gctx->mutex);
6590 perf_remove_from_context(group_leader);
6591 list_for_each_entry(sibling, &group_leader->sibling_list,
6593 perf_remove_from_context(sibling);
6596 mutex_unlock(&gctx->mutex);
6600 event->filp = event_file;
6601 WARN_ON_ONCE(ctx->parent_ctx);
6602 mutex_lock(&ctx->mutex);
6605 perf_install_in_context(ctx, group_leader, cpu);
6607 list_for_each_entry(sibling, &group_leader->sibling_list,
6609 perf_install_in_context(ctx, sibling, cpu);
6614 perf_install_in_context(ctx, event, cpu);
6616 perf_unpin_context(ctx);
6617 mutex_unlock(&ctx->mutex);
6619 event->owner = current;
6621 mutex_lock(¤t->perf_event_mutex);
6622 list_add_tail(&event->owner_entry, ¤t->perf_event_list);
6623 mutex_unlock(¤t->perf_event_mutex);
6626 * Precalculate sample_data sizes
6628 perf_event__header_size(event);
6629 perf_event__id_header_size(event);
6632 * Drop the reference on the group_event after placing the
6633 * new event on the sibling_list. This ensures destruction
6634 * of the group leader will find the pointer to itself in
6635 * perf_group_detach().
6637 fput_light(group_file, fput_needed);
6638 fd_install(event_fd, event_file);
6642 perf_unpin_context(ctx);
6648 put_task_struct(task);
6650 fput_light(group_file, fput_needed);
6652 put_unused_fd(event_fd);
6657 * perf_event_create_kernel_counter
6659 * @attr: attributes of the counter to create
6660 * @cpu: cpu in which the counter is bound
6661 * @task: task to profile (NULL for percpu)
6664 perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu,
6665 struct task_struct *task,
6666 perf_overflow_handler_t overflow_handler)
6668 struct perf_event_context *ctx;
6669 struct perf_event *event;
6673 * Get the target context (task or percpu):
6676 event = perf_event_alloc(attr, cpu, task, NULL, NULL, overflow_handler);
6677 if (IS_ERR(event)) {
6678 err = PTR_ERR(event);
6682 ctx = find_get_context(event->pmu, task, cpu);
6689 WARN_ON_ONCE(ctx->parent_ctx);
6690 mutex_lock(&ctx->mutex);
6691 perf_install_in_context(ctx, event, cpu);
6693 perf_unpin_context(ctx);
6694 mutex_unlock(&ctx->mutex);
6701 return ERR_PTR(err);
6703 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter);
6705 static void sync_child_event(struct perf_event *child_event,
6706 struct task_struct *child)
6708 struct perf_event *parent_event = child_event->parent;
6711 if (child_event->attr.inherit_stat)
6712 perf_event_read_event(child_event, child);
6714 child_val = perf_event_count(child_event);
6717 * Add back the child's count to the parent's count:
6719 atomic64_add(child_val, &parent_event->child_count);
6720 atomic64_add(child_event->total_time_enabled,
6721 &parent_event->child_total_time_enabled);
6722 atomic64_add(child_event->total_time_running,
6723 &parent_event->child_total_time_running);
6726 * Remove this event from the parent's list
6728 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
6729 mutex_lock(&parent_event->child_mutex);
6730 list_del_init(&child_event->child_list);
6731 mutex_unlock(&parent_event->child_mutex);
6734 * Release the parent event, if this was the last
6737 fput(parent_event->filp);
6741 __perf_event_exit_task(struct perf_event *child_event,
6742 struct perf_event_context *child_ctx,
6743 struct task_struct *child)
6745 if (child_event->parent) {
6746 raw_spin_lock_irq(&child_ctx->lock);
6747 perf_group_detach(child_event);
6748 raw_spin_unlock_irq(&child_ctx->lock);
6751 perf_remove_from_context(child_event);
6754 * It can happen that the parent exits first, and has events
6755 * that are still around due to the child reference. These
6756 * events need to be zapped.
6758 if (child_event->parent) {
6759 sync_child_event(child_event, child);
6760 free_event(child_event);
6764 static void perf_event_exit_task_context(struct task_struct *child, int ctxn)
6766 struct perf_event *child_event, *tmp;
6767 struct perf_event_context *child_ctx;
6768 unsigned long flags;
6770 if (likely(!child->perf_event_ctxp[ctxn])) {
6771 perf_event_task(child, NULL, 0);
6775 local_irq_save(flags);
6777 * We can't reschedule here because interrupts are disabled,
6778 * and either child is current or it is a task that can't be
6779 * scheduled, so we are now safe from rescheduling changing
6782 child_ctx = rcu_dereference_raw(child->perf_event_ctxp[ctxn]);
6783 task_ctx_sched_out(child_ctx, EVENT_ALL);
6786 * Take the context lock here so that if find_get_context is
6787 * reading child->perf_event_ctxp, we wait until it has
6788 * incremented the context's refcount before we do put_ctx below.
6790 raw_spin_lock(&child_ctx->lock);
6791 child->perf_event_ctxp[ctxn] = NULL;
6793 * If this context is a clone; unclone it so it can't get
6794 * swapped to another process while we're removing all
6795 * the events from it.
6797 unclone_ctx(child_ctx);
6798 update_context_time(child_ctx);
6799 raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
6802 * Report the task dead after unscheduling the events so that we
6803 * won't get any samples after PERF_RECORD_EXIT. We can however still
6804 * get a few PERF_RECORD_READ events.
6806 perf_event_task(child, child_ctx, 0);
6809 * We can recurse on the same lock type through:
6811 * __perf_event_exit_task()
6812 * sync_child_event()
6813 * fput(parent_event->filp)
6815 * mutex_lock(&ctx->mutex)
6817 * But since its the parent context it won't be the same instance.
6819 mutex_lock(&child_ctx->mutex);
6822 list_for_each_entry_safe(child_event, tmp, &child_ctx->pinned_groups,
6824 __perf_event_exit_task(child_event, child_ctx, child);
6826 list_for_each_entry_safe(child_event, tmp, &child_ctx->flexible_groups,
6828 __perf_event_exit_task(child_event, child_ctx, child);
6831 * If the last event was a group event, it will have appended all
6832 * its siblings to the list, but we obtained 'tmp' before that which
6833 * will still point to the list head terminating the iteration.
6835 if (!list_empty(&child_ctx->pinned_groups) ||
6836 !list_empty(&child_ctx->flexible_groups))
6839 mutex_unlock(&child_ctx->mutex);
6845 * When a child task exits, feed back event values to parent events.
6847 void perf_event_exit_task(struct task_struct *child)
6849 struct perf_event *event, *tmp;
6852 mutex_lock(&child->perf_event_mutex);
6853 list_for_each_entry_safe(event, tmp, &child->perf_event_list,
6855 list_del_init(&event->owner_entry);
6858 * Ensure the list deletion is visible before we clear
6859 * the owner, closes a race against perf_release() where
6860 * we need to serialize on the owner->perf_event_mutex.
6863 event->owner = NULL;
6865 mutex_unlock(&child->perf_event_mutex);
6867 for_each_task_context_nr(ctxn)
6868 perf_event_exit_task_context(child, ctxn);
6871 static void perf_free_event(struct perf_event *event,
6872 struct perf_event_context *ctx)
6874 struct perf_event *parent = event->parent;
6876 if (WARN_ON_ONCE(!parent))
6879 mutex_lock(&parent->child_mutex);
6880 list_del_init(&event->child_list);
6881 mutex_unlock(&parent->child_mutex);
6885 perf_group_detach(event);
6886 list_del_event(event, ctx);
6891 * free an unexposed, unused context as created by inheritance by
6892 * perf_event_init_task below, used by fork() in case of fail.
6894 void perf_event_free_task(struct task_struct *task)
6896 struct perf_event_context *ctx;
6897 struct perf_event *event, *tmp;
6900 for_each_task_context_nr(ctxn) {
6901 ctx = task->perf_event_ctxp[ctxn];
6905 mutex_lock(&ctx->mutex);
6907 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups,
6909 perf_free_event(event, ctx);
6911 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups,
6913 perf_free_event(event, ctx);
6915 if (!list_empty(&ctx->pinned_groups) ||
6916 !list_empty(&ctx->flexible_groups))
6919 mutex_unlock(&ctx->mutex);
6925 void perf_event_delayed_put(struct task_struct *task)
6929 for_each_task_context_nr(ctxn)
6930 WARN_ON_ONCE(task->perf_event_ctxp[ctxn]);
6934 * inherit a event from parent task to child task:
6936 static struct perf_event *
6937 inherit_event(struct perf_event *parent_event,
6938 struct task_struct *parent,
6939 struct perf_event_context *parent_ctx,
6940 struct task_struct *child,
6941 struct perf_event *group_leader,
6942 struct perf_event_context *child_ctx)
6944 struct perf_event *child_event;
6945 unsigned long flags;
6948 * Instead of creating recursive hierarchies of events,
6949 * we link inherited events back to the original parent,
6950 * which has a filp for sure, which we use as the reference
6953 if (parent_event->parent)
6954 parent_event = parent_event->parent;
6956 child_event = perf_event_alloc(&parent_event->attr,
6959 group_leader, parent_event,
6961 if (IS_ERR(child_event))
6966 * Make the child state follow the state of the parent event,
6967 * not its attr.disabled bit. We hold the parent's mutex,
6968 * so we won't race with perf_event_{en, dis}able_family.
6970 if (parent_event->state >= PERF_EVENT_STATE_INACTIVE)
6971 child_event->state = PERF_EVENT_STATE_INACTIVE;
6973 child_event->state = PERF_EVENT_STATE_OFF;
6975 if (parent_event->attr.freq) {
6976 u64 sample_period = parent_event->hw.sample_period;
6977 struct hw_perf_event *hwc = &child_event->hw;
6979 hwc->sample_period = sample_period;
6980 hwc->last_period = sample_period;
6982 local64_set(&hwc->period_left, sample_period);
6985 child_event->ctx = child_ctx;
6986 child_event->overflow_handler = parent_event->overflow_handler;
6989 * Precalculate sample_data sizes
6991 perf_event__header_size(child_event);
6992 perf_event__id_header_size(child_event);
6995 * Link it up in the child's context:
6997 raw_spin_lock_irqsave(&child_ctx->lock, flags);
6998 add_event_to_ctx(child_event, child_ctx);
6999 raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
7002 * Get a reference to the parent filp - we will fput it
7003 * when the child event exits. This is safe to do because
7004 * we are in the parent and we know that the filp still
7005 * exists and has a nonzero count:
7007 atomic_long_inc(&parent_event->filp->f_count);
7010 * Link this into the parent event's child list
7012 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
7013 mutex_lock(&parent_event->child_mutex);
7014 list_add_tail(&child_event->child_list, &parent_event->child_list);
7015 mutex_unlock(&parent_event->child_mutex);
7020 static int inherit_group(struct perf_event *parent_event,
7021 struct task_struct *parent,
7022 struct perf_event_context *parent_ctx,
7023 struct task_struct *child,
7024 struct perf_event_context *child_ctx)
7026 struct perf_event *leader;
7027 struct perf_event *sub;
7028 struct perf_event *child_ctr;
7030 leader = inherit_event(parent_event, parent, parent_ctx,
7031 child, NULL, child_ctx);
7033 return PTR_ERR(leader);
7034 list_for_each_entry(sub, &parent_event->sibling_list, group_entry) {
7035 child_ctr = inherit_event(sub, parent, parent_ctx,
7036 child, leader, child_ctx);
7037 if (IS_ERR(child_ctr))
7038 return PTR_ERR(child_ctr);
7044 inherit_task_group(struct perf_event *event, struct task_struct *parent,
7045 struct perf_event_context *parent_ctx,
7046 struct task_struct *child, int ctxn,
7050 struct perf_event_context *child_ctx;
7052 if (!event->attr.inherit) {
7057 child_ctx = child->perf_event_ctxp[ctxn];
7060 * This is executed from the parent task context, so
7061 * inherit events that have been marked for cloning.
7062 * First allocate and initialize a context for the
7066 child_ctx = alloc_perf_context(event->pmu, child);
7070 child->perf_event_ctxp[ctxn] = child_ctx;
7073 ret = inherit_group(event, parent, parent_ctx,
7083 * Initialize the perf_event context in task_struct
7085 int perf_event_init_context(struct task_struct *child, int ctxn)
7087 struct perf_event_context *child_ctx, *parent_ctx;
7088 struct perf_event_context *cloned_ctx;
7089 struct perf_event *event;
7090 struct task_struct *parent = current;
7091 int inherited_all = 1;
7092 unsigned long flags;
7095 if (likely(!parent->perf_event_ctxp[ctxn]))
7099 * If the parent's context is a clone, pin it so it won't get
7102 parent_ctx = perf_pin_task_context(parent, ctxn);
7105 * No need to check if parent_ctx != NULL here; since we saw
7106 * it non-NULL earlier, the only reason for it to become NULL
7107 * is if we exit, and since we're currently in the middle of
7108 * a fork we can't be exiting at the same time.
7112 * Lock the parent list. No need to lock the child - not PID
7113 * hashed yet and not running, so nobody can access it.
7115 mutex_lock(&parent_ctx->mutex);
7118 * We dont have to disable NMIs - we are only looking at
7119 * the list, not manipulating it:
7121 list_for_each_entry(event, &parent_ctx->pinned_groups, group_entry) {
7122 ret = inherit_task_group(event, parent, parent_ctx,
7123 child, ctxn, &inherited_all);
7129 * We can't hold ctx->lock when iterating the ->flexible_group list due
7130 * to allocations, but we need to prevent rotation because
7131 * rotate_ctx() will change the list from interrupt context.
7133 raw_spin_lock_irqsave(&parent_ctx->lock, flags);
7134 parent_ctx->rotate_disable = 1;
7135 raw_spin_unlock_irqrestore(&parent_ctx->lock, flags);
7137 list_for_each_entry(event, &parent_ctx->flexible_groups, group_entry) {
7138 ret = inherit_task_group(event, parent, parent_ctx,
7139 child, ctxn, &inherited_all);
7144 raw_spin_lock_irqsave(&parent_ctx->lock, flags);
7145 parent_ctx->rotate_disable = 0;
7147 child_ctx = child->perf_event_ctxp[ctxn];
7149 if (child_ctx && inherited_all) {
7151 * Mark the child context as a clone of the parent
7152 * context, or of whatever the parent is a clone of.
7154 * Note that if the parent is a clone, the holding of
7155 * parent_ctx->lock avoids it from being uncloned.
7157 cloned_ctx = parent_ctx->parent_ctx;
7159 child_ctx->parent_ctx = cloned_ctx;
7160 child_ctx->parent_gen = parent_ctx->parent_gen;
7162 child_ctx->parent_ctx = parent_ctx;
7163 child_ctx->parent_gen = parent_ctx->generation;
7165 get_ctx(child_ctx->parent_ctx);
7168 raw_spin_unlock_irqrestore(&parent_ctx->lock, flags);
7169 mutex_unlock(&parent_ctx->mutex);
7171 perf_unpin_context(parent_ctx);
7172 put_ctx(parent_ctx);
7178 * Initialize the perf_event context in task_struct
7180 int perf_event_init_task(struct task_struct *child)
7184 memset(child->perf_event_ctxp, 0, sizeof(child->perf_event_ctxp));
7185 mutex_init(&child->perf_event_mutex);
7186 INIT_LIST_HEAD(&child->perf_event_list);
7188 for_each_task_context_nr(ctxn) {
7189 ret = perf_event_init_context(child, ctxn);
7197 static void __init perf_event_init_all_cpus(void)
7199 struct swevent_htable *swhash;
7202 for_each_possible_cpu(cpu) {
7203 swhash = &per_cpu(swevent_htable, cpu);
7204 mutex_init(&swhash->hlist_mutex);
7205 INIT_LIST_HEAD(&per_cpu(rotation_list, cpu));
7209 static void __cpuinit perf_event_init_cpu(int cpu)
7211 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
7213 mutex_lock(&swhash->hlist_mutex);
7214 if (swhash->hlist_refcount > 0) {
7215 struct swevent_hlist *hlist;
7217 hlist = kzalloc_node(sizeof(*hlist), GFP_KERNEL, cpu_to_node(cpu));
7219 rcu_assign_pointer(swhash->swevent_hlist, hlist);
7221 mutex_unlock(&swhash->hlist_mutex);
7224 #if defined CONFIG_HOTPLUG_CPU || defined CONFIG_KEXEC
7225 static void perf_pmu_rotate_stop(struct pmu *pmu)
7227 struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
7229 WARN_ON(!irqs_disabled());
7231 list_del_init(&cpuctx->rotation_list);
7234 static void __perf_event_exit_context(void *__info)
7236 struct perf_event_context *ctx = __info;
7237 struct perf_event *event, *tmp;
7239 perf_pmu_rotate_stop(ctx->pmu);
7241 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry)
7242 __perf_remove_from_context(event);
7243 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups, group_entry)
7244 __perf_remove_from_context(event);
7247 static void perf_event_exit_cpu_context(int cpu)
7249 struct perf_event_context *ctx;
7253 idx = srcu_read_lock(&pmus_srcu);
7254 list_for_each_entry_rcu(pmu, &pmus, entry) {
7255 ctx = &per_cpu_ptr(pmu->pmu_cpu_context, cpu)->ctx;
7257 mutex_lock(&ctx->mutex);
7258 smp_call_function_single(cpu, __perf_event_exit_context, ctx, 1);
7259 mutex_unlock(&ctx->mutex);
7261 srcu_read_unlock(&pmus_srcu, idx);
7264 static void perf_event_exit_cpu(int cpu)
7266 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
7268 mutex_lock(&swhash->hlist_mutex);
7269 swevent_hlist_release(swhash);
7270 mutex_unlock(&swhash->hlist_mutex);
7272 perf_event_exit_cpu_context(cpu);
7275 static inline void perf_event_exit_cpu(int cpu) { }
7279 perf_reboot(struct notifier_block *notifier, unsigned long val, void *v)
7283 for_each_online_cpu(cpu)
7284 perf_event_exit_cpu(cpu);
7290 * Run the perf reboot notifier at the very last possible moment so that
7291 * the generic watchdog code runs as long as possible.
7293 static struct notifier_block perf_reboot_notifier = {
7294 .notifier_call = perf_reboot,
7295 .priority = INT_MIN,
7298 static int __cpuinit
7299 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
7301 unsigned int cpu = (long)hcpu;
7303 switch (action & ~CPU_TASKS_FROZEN) {
7305 case CPU_UP_PREPARE:
7306 case CPU_DOWN_FAILED:
7307 perf_event_init_cpu(cpu);
7310 case CPU_UP_CANCELED:
7311 case CPU_DOWN_PREPARE:
7312 perf_event_exit_cpu(cpu);
7322 void __init perf_event_init(void)
7328 perf_event_init_all_cpus();
7329 init_srcu_struct(&pmus_srcu);
7330 perf_pmu_register(&perf_swevent, "software", PERF_TYPE_SOFTWARE);
7331 perf_pmu_register(&perf_cpu_clock, NULL, -1);
7332 perf_pmu_register(&perf_task_clock, NULL, -1);
7334 perf_cpu_notifier(perf_cpu_notify);
7335 register_reboot_notifier(&perf_reboot_notifier);
7337 ret = init_hw_breakpoint();
7338 WARN(ret, "hw_breakpoint initialization failed with: %d", ret);
7341 static int __init perf_event_sysfs_init(void)
7346 mutex_lock(&pmus_lock);
7348 ret = bus_register(&pmu_bus);
7352 list_for_each_entry(pmu, &pmus, entry) {
7353 if (!pmu->name || pmu->type < 0)
7356 ret = pmu_dev_alloc(pmu);
7357 WARN(ret, "Failed to register pmu: %s, reason %d\n", pmu->name, ret);
7359 pmu_bus_running = 1;
7363 mutex_unlock(&pmus_lock);
7367 device_initcall(perf_event_sysfs_init);
7369 #ifdef CONFIG_CGROUP_PERF
7370 static struct cgroup_subsys_state *perf_cgroup_create(
7371 struct cgroup_subsys *ss, struct cgroup *cont)
7373 struct perf_cgroup *jc;
7375 jc = kzalloc(sizeof(*jc), GFP_KERNEL);
7377 return ERR_PTR(-ENOMEM);
7379 jc->info = alloc_percpu(struct perf_cgroup_info);
7382 return ERR_PTR(-ENOMEM);
7388 static void perf_cgroup_destroy(struct cgroup_subsys *ss,
7389 struct cgroup *cont)
7391 struct perf_cgroup *jc;
7392 jc = container_of(cgroup_subsys_state(cont, perf_subsys_id),
7393 struct perf_cgroup, css);
7394 free_percpu(jc->info);
7398 static int __perf_cgroup_move(void *info)
7400 struct task_struct *task = info;
7401 perf_cgroup_switch(task, PERF_CGROUP_SWOUT | PERF_CGROUP_SWIN);
7406 perf_cgroup_attach_task(struct cgroup *cgrp, struct task_struct *task)
7408 task_function_call(task, __perf_cgroup_move, task);
7411 static void perf_cgroup_exit(struct cgroup_subsys *ss, struct cgroup *cgrp,
7412 struct cgroup *old_cgrp, struct task_struct *task)
7415 * cgroup_exit() is called in the copy_process() failure path.
7416 * Ignore this case since the task hasn't ran yet, this avoids
7417 * trying to poke a half freed task state from generic code.
7419 if (!(task->flags & PF_EXITING))
7422 perf_cgroup_attach_task(cgrp, task);
7425 struct cgroup_subsys perf_subsys = {
7426 .name = "perf_event",
7427 .subsys_id = perf_subsys_id,
7428 .create = perf_cgroup_create,
7429 .destroy = perf_cgroup_destroy,
7430 .exit = perf_cgroup_exit,
7431 .attach_task = perf_cgroup_attach_task,
7433 #endif /* CONFIG_CGROUP_PERF */