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/export.h>
29 #include <linux/vmalloc.h>
30 #include <linux/hardirq.h>
31 #include <linux/rculist.h>
32 #include <linux/uaccess.h>
33 #include <linux/syscalls.h>
34 #include <linux/anon_inodes.h>
35 #include <linux/kernel_stat.h>
36 #include <linux/perf_event.h>
37 #include <linux/ftrace_event.h>
38 #include <linux/hw_breakpoint.h>
42 #include <asm/irq_regs.h>
44 struct remote_function_call {
45 struct task_struct *p;
46 int (*func)(void *info);
51 static void remote_function(void *data)
53 struct remote_function_call *tfc = data;
54 struct task_struct *p = tfc->p;
58 if (task_cpu(p) != smp_processor_id() || !task_curr(p))
62 tfc->ret = tfc->func(tfc->info);
66 * task_function_call - call a function on the cpu on which a task runs
67 * @p: the task to evaluate
68 * @func: the function to be called
69 * @info: the function call argument
71 * Calls the function @func when the task is currently running. This might
72 * be on the current CPU, which just calls the function directly
74 * returns: @func return value, or
75 * -ESRCH - when the process isn't running
76 * -EAGAIN - when the process moved away
79 task_function_call(struct task_struct *p, int (*func) (void *info), void *info)
81 struct remote_function_call data = {
85 .ret = -ESRCH, /* No such (running) process */
89 smp_call_function_single(task_cpu(p), remote_function, &data, 1);
95 * cpu_function_call - call a function on the cpu
96 * @func: the function to be called
97 * @info: the function call argument
99 * Calls the function @func on the remote cpu.
101 * returns: @func return value or -ENXIO when the cpu is offline
103 static int cpu_function_call(int cpu, int (*func) (void *info), void *info)
105 struct remote_function_call data = {
109 .ret = -ENXIO, /* No such CPU */
112 smp_call_function_single(cpu, remote_function, &data, 1);
117 #define PERF_FLAG_ALL (PERF_FLAG_FD_NO_GROUP |\
118 PERF_FLAG_FD_OUTPUT |\
119 PERF_FLAG_PID_CGROUP)
122 * branch priv levels that need permission checks
124 #define PERF_SAMPLE_BRANCH_PERM_PLM \
125 (PERF_SAMPLE_BRANCH_KERNEL |\
126 PERF_SAMPLE_BRANCH_HV)
129 EVENT_FLEXIBLE = 0x1,
131 EVENT_ALL = EVENT_FLEXIBLE | EVENT_PINNED,
135 * perf_sched_events : >0 events exist
136 * perf_cgroup_events: >0 per-cpu cgroup events exist on this cpu
138 struct static_key_deferred perf_sched_events __read_mostly;
139 static DEFINE_PER_CPU(atomic_t, perf_cgroup_events);
140 static DEFINE_PER_CPU(atomic_t, perf_branch_stack_events);
142 static atomic_t nr_mmap_events __read_mostly;
143 static atomic_t nr_comm_events __read_mostly;
144 static atomic_t nr_task_events __read_mostly;
146 static LIST_HEAD(pmus);
147 static DEFINE_MUTEX(pmus_lock);
148 static struct srcu_struct pmus_srcu;
151 * perf event paranoia level:
152 * -1 - not paranoid at all
153 * 0 - disallow raw tracepoint access for unpriv
154 * 1 - disallow cpu events for unpriv
155 * 2 - disallow kernel profiling for unpriv
157 int sysctl_perf_event_paranoid __read_mostly = 1;
159 /* Minimum for 512 kiB + 1 user control page */
160 int sysctl_perf_event_mlock __read_mostly = 512 + (PAGE_SIZE / 1024); /* 'free' kiB per user */
163 * max perf event sample rate
165 #define DEFAULT_MAX_SAMPLE_RATE 100000
166 int sysctl_perf_event_sample_rate __read_mostly = DEFAULT_MAX_SAMPLE_RATE;
167 static int max_samples_per_tick __read_mostly =
168 DIV_ROUND_UP(DEFAULT_MAX_SAMPLE_RATE, HZ);
170 int perf_proc_update_handler(struct ctl_table *table, int write,
171 void __user *buffer, size_t *lenp,
174 int ret = proc_dointvec(table, write, buffer, lenp, ppos);
179 max_samples_per_tick = DIV_ROUND_UP(sysctl_perf_event_sample_rate, HZ);
184 static atomic64_t perf_event_id;
186 static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
187 enum event_type_t event_type);
189 static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
190 enum event_type_t event_type,
191 struct task_struct *task);
193 static void update_context_time(struct perf_event_context *ctx);
194 static u64 perf_event_time(struct perf_event *event);
196 static void ring_buffer_attach(struct perf_event *event,
197 struct ring_buffer *rb);
199 void __weak perf_event_print_debug(void) { }
201 extern __weak const char *perf_pmu_name(void)
206 static inline u64 perf_clock(void)
208 return local_clock();
211 static inline struct perf_cpu_context *
212 __get_cpu_context(struct perf_event_context *ctx)
214 return this_cpu_ptr(ctx->pmu->pmu_cpu_context);
217 static void perf_ctx_lock(struct perf_cpu_context *cpuctx,
218 struct perf_event_context *ctx)
220 raw_spin_lock(&cpuctx->ctx.lock);
222 raw_spin_lock(&ctx->lock);
225 static void perf_ctx_unlock(struct perf_cpu_context *cpuctx,
226 struct perf_event_context *ctx)
229 raw_spin_unlock(&ctx->lock);
230 raw_spin_unlock(&cpuctx->ctx.lock);
233 #ifdef CONFIG_CGROUP_PERF
236 * Must ensure cgroup is pinned (css_get) before calling
237 * this function. In other words, we cannot call this function
238 * if there is no cgroup event for the current CPU context.
240 static inline struct perf_cgroup *
241 perf_cgroup_from_task(struct task_struct *task)
243 return container_of(task_subsys_state(task, perf_subsys_id),
244 struct perf_cgroup, css);
248 perf_cgroup_match(struct perf_event *event)
250 struct perf_event_context *ctx = event->ctx;
251 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
253 return !event->cgrp || event->cgrp == cpuctx->cgrp;
256 static inline bool perf_tryget_cgroup(struct perf_event *event)
258 return css_tryget(&event->cgrp->css);
261 static inline void perf_put_cgroup(struct perf_event *event)
263 css_put(&event->cgrp->css);
266 static inline void perf_detach_cgroup(struct perf_event *event)
268 perf_put_cgroup(event);
272 static inline int is_cgroup_event(struct perf_event *event)
274 return event->cgrp != NULL;
277 static inline u64 perf_cgroup_event_time(struct perf_event *event)
279 struct perf_cgroup_info *t;
281 t = per_cpu_ptr(event->cgrp->info, event->cpu);
285 static inline void __update_cgrp_time(struct perf_cgroup *cgrp)
287 struct perf_cgroup_info *info;
292 info = this_cpu_ptr(cgrp->info);
294 info->time += now - info->timestamp;
295 info->timestamp = now;
298 static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx)
300 struct perf_cgroup *cgrp_out = cpuctx->cgrp;
302 __update_cgrp_time(cgrp_out);
305 static inline void update_cgrp_time_from_event(struct perf_event *event)
307 struct perf_cgroup *cgrp;
310 * ensure we access cgroup data only when needed and
311 * when we know the cgroup is pinned (css_get)
313 if (!is_cgroup_event(event))
316 cgrp = perf_cgroup_from_task(current);
318 * Do not update time when cgroup is not active
320 if (cgrp == event->cgrp)
321 __update_cgrp_time(event->cgrp);
325 perf_cgroup_set_timestamp(struct task_struct *task,
326 struct perf_event_context *ctx)
328 struct perf_cgroup *cgrp;
329 struct perf_cgroup_info *info;
332 * ctx->lock held by caller
333 * ensure we do not access cgroup data
334 * unless we have the cgroup pinned (css_get)
336 if (!task || !ctx->nr_cgroups)
339 cgrp = perf_cgroup_from_task(task);
340 info = this_cpu_ptr(cgrp->info);
341 info->timestamp = ctx->timestamp;
344 #define PERF_CGROUP_SWOUT 0x1 /* cgroup switch out every event */
345 #define PERF_CGROUP_SWIN 0x2 /* cgroup switch in events based on task */
348 * reschedule events based on the cgroup constraint of task.
350 * mode SWOUT : schedule out everything
351 * mode SWIN : schedule in based on cgroup for next
353 void perf_cgroup_switch(struct task_struct *task, int mode)
355 struct perf_cpu_context *cpuctx;
360 * disable interrupts to avoid geting nr_cgroup
361 * changes via __perf_event_disable(). Also
364 local_irq_save(flags);
367 * we reschedule only in the presence of cgroup
368 * constrained events.
372 list_for_each_entry_rcu(pmu, &pmus, entry) {
373 cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
376 * perf_cgroup_events says at least one
377 * context on this CPU has cgroup events.
379 * ctx->nr_cgroups reports the number of cgroup
380 * events for a context.
382 if (cpuctx->ctx.nr_cgroups > 0) {
383 perf_ctx_lock(cpuctx, cpuctx->task_ctx);
384 perf_pmu_disable(cpuctx->ctx.pmu);
386 if (mode & PERF_CGROUP_SWOUT) {
387 cpu_ctx_sched_out(cpuctx, EVENT_ALL);
389 * must not be done before ctxswout due
390 * to event_filter_match() in event_sched_out()
395 if (mode & PERF_CGROUP_SWIN) {
396 WARN_ON_ONCE(cpuctx->cgrp);
397 /* set cgrp before ctxsw in to
398 * allow event_filter_match() to not
399 * have to pass task around
401 cpuctx->cgrp = perf_cgroup_from_task(task);
402 cpu_ctx_sched_in(cpuctx, EVENT_ALL, task);
404 perf_pmu_enable(cpuctx->ctx.pmu);
405 perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
411 local_irq_restore(flags);
414 static inline void perf_cgroup_sched_out(struct task_struct *task,
415 struct task_struct *next)
417 struct perf_cgroup *cgrp1;
418 struct perf_cgroup *cgrp2 = NULL;
421 * we come here when we know perf_cgroup_events > 0
423 cgrp1 = perf_cgroup_from_task(task);
426 * next is NULL when called from perf_event_enable_on_exec()
427 * that will systematically cause a cgroup_switch()
430 cgrp2 = perf_cgroup_from_task(next);
433 * only schedule out current cgroup events if we know
434 * that we are switching to a different cgroup. Otherwise,
435 * do no touch the cgroup events.
438 perf_cgroup_switch(task, PERF_CGROUP_SWOUT);
441 static inline void perf_cgroup_sched_in(struct task_struct *prev,
442 struct task_struct *task)
444 struct perf_cgroup *cgrp1;
445 struct perf_cgroup *cgrp2 = NULL;
448 * we come here when we know perf_cgroup_events > 0
450 cgrp1 = perf_cgroup_from_task(task);
452 /* prev can never be NULL */
453 cgrp2 = perf_cgroup_from_task(prev);
456 * only need to schedule in cgroup events if we are changing
457 * cgroup during ctxsw. Cgroup events were not scheduled
458 * out of ctxsw out if that was not the case.
461 perf_cgroup_switch(task, PERF_CGROUP_SWIN);
464 static inline int perf_cgroup_connect(int fd, struct perf_event *event,
465 struct perf_event_attr *attr,
466 struct perf_event *group_leader)
468 struct perf_cgroup *cgrp;
469 struct cgroup_subsys_state *css;
471 int ret = 0, fput_needed;
473 file = fget_light(fd, &fput_needed);
477 css = cgroup_css_from_dir(file, perf_subsys_id);
483 cgrp = container_of(css, struct perf_cgroup, css);
486 /* must be done before we fput() the file */
487 if (!perf_tryget_cgroup(event)) {
494 * all events in a group must monitor
495 * the same cgroup because a task belongs
496 * to only one perf cgroup at a time
498 if (group_leader && group_leader->cgrp != cgrp) {
499 perf_detach_cgroup(event);
503 fput_light(file, fput_needed);
508 perf_cgroup_set_shadow_time(struct perf_event *event, u64 now)
510 struct perf_cgroup_info *t;
511 t = per_cpu_ptr(event->cgrp->info, event->cpu);
512 event->shadow_ctx_time = now - t->timestamp;
516 perf_cgroup_defer_enabled(struct perf_event *event)
519 * when the current task's perf cgroup does not match
520 * the event's, we need to remember to call the
521 * perf_mark_enable() function the first time a task with
522 * a matching perf cgroup is scheduled in.
524 if (is_cgroup_event(event) && !perf_cgroup_match(event))
525 event->cgrp_defer_enabled = 1;
529 perf_cgroup_mark_enabled(struct perf_event *event,
530 struct perf_event_context *ctx)
532 struct perf_event *sub;
533 u64 tstamp = perf_event_time(event);
535 if (!event->cgrp_defer_enabled)
538 event->cgrp_defer_enabled = 0;
540 event->tstamp_enabled = tstamp - event->total_time_enabled;
541 list_for_each_entry(sub, &event->sibling_list, group_entry) {
542 if (sub->state >= PERF_EVENT_STATE_INACTIVE) {
543 sub->tstamp_enabled = tstamp - sub->total_time_enabled;
544 sub->cgrp_defer_enabled = 0;
548 #else /* !CONFIG_CGROUP_PERF */
551 perf_cgroup_match(struct perf_event *event)
556 static inline void perf_detach_cgroup(struct perf_event *event)
559 static inline int is_cgroup_event(struct perf_event *event)
564 static inline u64 perf_cgroup_event_cgrp_time(struct perf_event *event)
569 static inline void update_cgrp_time_from_event(struct perf_event *event)
573 static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx)
577 static inline void perf_cgroup_sched_out(struct task_struct *task,
578 struct task_struct *next)
582 static inline void perf_cgroup_sched_in(struct task_struct *prev,
583 struct task_struct *task)
587 static inline int perf_cgroup_connect(pid_t pid, struct perf_event *event,
588 struct perf_event_attr *attr,
589 struct perf_event *group_leader)
595 perf_cgroup_set_timestamp(struct task_struct *task,
596 struct perf_event_context *ctx)
601 perf_cgroup_switch(struct task_struct *task, struct task_struct *next)
606 perf_cgroup_set_shadow_time(struct perf_event *event, u64 now)
610 static inline u64 perf_cgroup_event_time(struct perf_event *event)
616 perf_cgroup_defer_enabled(struct perf_event *event)
621 perf_cgroup_mark_enabled(struct perf_event *event,
622 struct perf_event_context *ctx)
627 void perf_pmu_disable(struct pmu *pmu)
629 int *count = this_cpu_ptr(pmu->pmu_disable_count);
631 pmu->pmu_disable(pmu);
634 void perf_pmu_enable(struct pmu *pmu)
636 int *count = this_cpu_ptr(pmu->pmu_disable_count);
638 pmu->pmu_enable(pmu);
641 static DEFINE_PER_CPU(struct list_head, rotation_list);
644 * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
645 * because they're strictly cpu affine and rotate_start is called with IRQs
646 * disabled, while rotate_context is called from IRQ context.
648 static void perf_pmu_rotate_start(struct pmu *pmu)
650 struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
651 struct list_head *head = &__get_cpu_var(rotation_list);
653 WARN_ON(!irqs_disabled());
655 if (list_empty(&cpuctx->rotation_list))
656 list_add(&cpuctx->rotation_list, head);
659 static void get_ctx(struct perf_event_context *ctx)
661 WARN_ON(!atomic_inc_not_zero(&ctx->refcount));
664 static void put_ctx(struct perf_event_context *ctx)
666 if (atomic_dec_and_test(&ctx->refcount)) {
668 put_ctx(ctx->parent_ctx);
670 put_task_struct(ctx->task);
671 kfree_rcu(ctx, rcu_head);
675 static void unclone_ctx(struct perf_event_context *ctx)
677 if (ctx->parent_ctx) {
678 put_ctx(ctx->parent_ctx);
679 ctx->parent_ctx = NULL;
683 static u32 perf_event_pid(struct perf_event *event, struct task_struct *p)
686 * only top level events have the pid namespace they were created in
689 event = event->parent;
691 return task_tgid_nr_ns(p, event->ns);
694 static u32 perf_event_tid(struct perf_event *event, struct task_struct *p)
697 * only top level events have the pid namespace they were created in
700 event = event->parent;
702 return task_pid_nr_ns(p, event->ns);
706 * If we inherit events we want to return the parent event id
709 static u64 primary_event_id(struct perf_event *event)
714 id = event->parent->id;
720 * Get the perf_event_context for a task and lock it.
721 * This has to cope with with the fact that until it is locked,
722 * the context could get moved to another task.
724 static struct perf_event_context *
725 perf_lock_task_context(struct task_struct *task, int ctxn, unsigned long *flags)
727 struct perf_event_context *ctx;
731 ctx = rcu_dereference(task->perf_event_ctxp[ctxn]);
734 * If this context is a clone of another, it might
735 * get swapped for another underneath us by
736 * perf_event_task_sched_out, though the
737 * rcu_read_lock() protects us from any context
738 * getting freed. Lock the context and check if it
739 * got swapped before we could get the lock, and retry
740 * if so. If we locked the right context, then it
741 * can't get swapped on us any more.
743 raw_spin_lock_irqsave(&ctx->lock, *flags);
744 if (ctx != rcu_dereference(task->perf_event_ctxp[ctxn])) {
745 raw_spin_unlock_irqrestore(&ctx->lock, *flags);
749 if (!atomic_inc_not_zero(&ctx->refcount)) {
750 raw_spin_unlock_irqrestore(&ctx->lock, *flags);
759 * Get the context for a task and increment its pin_count so it
760 * can't get swapped to another task. This also increments its
761 * reference count so that the context can't get freed.
763 static struct perf_event_context *
764 perf_pin_task_context(struct task_struct *task, int ctxn)
766 struct perf_event_context *ctx;
769 ctx = perf_lock_task_context(task, ctxn, &flags);
772 raw_spin_unlock_irqrestore(&ctx->lock, flags);
777 static void perf_unpin_context(struct perf_event_context *ctx)
781 raw_spin_lock_irqsave(&ctx->lock, flags);
783 raw_spin_unlock_irqrestore(&ctx->lock, flags);
787 * Update the record of the current time in a context.
789 static void update_context_time(struct perf_event_context *ctx)
791 u64 now = perf_clock();
793 ctx->time += now - ctx->timestamp;
794 ctx->timestamp = now;
797 static u64 perf_event_time(struct perf_event *event)
799 struct perf_event_context *ctx = event->ctx;
801 if (is_cgroup_event(event))
802 return perf_cgroup_event_time(event);
804 return ctx ? ctx->time : 0;
808 * Update the total_time_enabled and total_time_running fields for a event.
809 * The caller of this function needs to hold the ctx->lock.
811 static void update_event_times(struct perf_event *event)
813 struct perf_event_context *ctx = event->ctx;
816 if (event->state < PERF_EVENT_STATE_INACTIVE ||
817 event->group_leader->state < PERF_EVENT_STATE_INACTIVE)
820 * in cgroup mode, time_enabled represents
821 * the time the event was enabled AND active
822 * tasks were in the monitored cgroup. This is
823 * independent of the activity of the context as
824 * there may be a mix of cgroup and non-cgroup events.
826 * That is why we treat cgroup events differently
829 if (is_cgroup_event(event))
830 run_end = perf_cgroup_event_time(event);
831 else if (ctx->is_active)
834 run_end = event->tstamp_stopped;
836 event->total_time_enabled = run_end - event->tstamp_enabled;
838 if (event->state == PERF_EVENT_STATE_INACTIVE)
839 run_end = event->tstamp_stopped;
841 run_end = perf_event_time(event);
843 event->total_time_running = run_end - event->tstamp_running;
848 * Update total_time_enabled and total_time_running for all events in a group.
850 static void update_group_times(struct perf_event *leader)
852 struct perf_event *event;
854 update_event_times(leader);
855 list_for_each_entry(event, &leader->sibling_list, group_entry)
856 update_event_times(event);
859 static struct list_head *
860 ctx_group_list(struct perf_event *event, struct perf_event_context *ctx)
862 if (event->attr.pinned)
863 return &ctx->pinned_groups;
865 return &ctx->flexible_groups;
869 * Add a event from the lists for its context.
870 * Must be called with ctx->mutex and ctx->lock held.
873 list_add_event(struct perf_event *event, struct perf_event_context *ctx)
875 WARN_ON_ONCE(event->attach_state & PERF_ATTACH_CONTEXT);
876 event->attach_state |= PERF_ATTACH_CONTEXT;
879 * If we're a stand alone event or group leader, we go to the context
880 * list, group events are kept attached to the group so that
881 * perf_group_detach can, at all times, locate all siblings.
883 if (event->group_leader == event) {
884 struct list_head *list;
886 if (is_software_event(event))
887 event->group_flags |= PERF_GROUP_SOFTWARE;
889 list = ctx_group_list(event, ctx);
890 list_add_tail(&event->group_entry, list);
893 if (is_cgroup_event(event))
896 if (has_branch_stack(event))
897 ctx->nr_branch_stack++;
899 list_add_rcu(&event->event_entry, &ctx->event_list);
901 perf_pmu_rotate_start(ctx->pmu);
903 if (event->attr.inherit_stat)
908 * Called at perf_event creation and when events are attached/detached from a
911 static void perf_event__read_size(struct perf_event *event)
913 int entry = sizeof(u64); /* value */
917 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
920 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
923 if (event->attr.read_format & PERF_FORMAT_ID)
924 entry += sizeof(u64);
926 if (event->attr.read_format & PERF_FORMAT_GROUP) {
927 nr += event->group_leader->nr_siblings;
932 event->read_size = size;
935 static void perf_event__header_size(struct perf_event *event)
937 struct perf_sample_data *data;
938 u64 sample_type = event->attr.sample_type;
941 perf_event__read_size(event);
943 if (sample_type & PERF_SAMPLE_IP)
944 size += sizeof(data->ip);
946 if (sample_type & PERF_SAMPLE_ADDR)
947 size += sizeof(data->addr);
949 if (sample_type & PERF_SAMPLE_PERIOD)
950 size += sizeof(data->period);
952 if (sample_type & PERF_SAMPLE_READ)
953 size += event->read_size;
955 event->header_size = size;
958 static void perf_event__id_header_size(struct perf_event *event)
960 struct perf_sample_data *data;
961 u64 sample_type = event->attr.sample_type;
964 if (sample_type & PERF_SAMPLE_TID)
965 size += sizeof(data->tid_entry);
967 if (sample_type & PERF_SAMPLE_TIME)
968 size += sizeof(data->time);
970 if (sample_type & PERF_SAMPLE_ID)
971 size += sizeof(data->id);
973 if (sample_type & PERF_SAMPLE_STREAM_ID)
974 size += sizeof(data->stream_id);
976 if (sample_type & PERF_SAMPLE_CPU)
977 size += sizeof(data->cpu_entry);
979 event->id_header_size = size;
982 static void perf_group_attach(struct perf_event *event)
984 struct perf_event *group_leader = event->group_leader, *pos;
987 * We can have double attach due to group movement in perf_event_open.
989 if (event->attach_state & PERF_ATTACH_GROUP)
992 event->attach_state |= PERF_ATTACH_GROUP;
994 if (group_leader == event)
997 if (group_leader->group_flags & PERF_GROUP_SOFTWARE &&
998 !is_software_event(event))
999 group_leader->group_flags &= ~PERF_GROUP_SOFTWARE;
1001 list_add_tail(&event->group_entry, &group_leader->sibling_list);
1002 group_leader->nr_siblings++;
1004 perf_event__header_size(group_leader);
1006 list_for_each_entry(pos, &group_leader->sibling_list, group_entry)
1007 perf_event__header_size(pos);
1011 * Remove a event from the lists for its context.
1012 * Must be called with ctx->mutex and ctx->lock held.
1015 list_del_event(struct perf_event *event, struct perf_event_context *ctx)
1017 struct perf_cpu_context *cpuctx;
1019 * We can have double detach due to exit/hot-unplug + close.
1021 if (!(event->attach_state & PERF_ATTACH_CONTEXT))
1024 event->attach_state &= ~PERF_ATTACH_CONTEXT;
1026 if (is_cgroup_event(event)) {
1028 cpuctx = __get_cpu_context(ctx);
1030 * if there are no more cgroup events
1031 * then cler cgrp to avoid stale pointer
1032 * in update_cgrp_time_from_cpuctx()
1034 if (!ctx->nr_cgroups)
1035 cpuctx->cgrp = NULL;
1038 if (has_branch_stack(event))
1039 ctx->nr_branch_stack--;
1042 if (event->attr.inherit_stat)
1045 list_del_rcu(&event->event_entry);
1047 if (event->group_leader == event)
1048 list_del_init(&event->group_entry);
1050 update_group_times(event);
1053 * If event was in error state, then keep it
1054 * that way, otherwise bogus counts will be
1055 * returned on read(). The only way to get out
1056 * of error state is by explicit re-enabling
1059 if (event->state > PERF_EVENT_STATE_OFF)
1060 event->state = PERF_EVENT_STATE_OFF;
1063 static void perf_group_detach(struct perf_event *event)
1065 struct perf_event *sibling, *tmp;
1066 struct list_head *list = NULL;
1069 * We can have double detach due to exit/hot-unplug + close.
1071 if (!(event->attach_state & PERF_ATTACH_GROUP))
1074 event->attach_state &= ~PERF_ATTACH_GROUP;
1077 * If this is a sibling, remove it from its group.
1079 if (event->group_leader != event) {
1080 list_del_init(&event->group_entry);
1081 event->group_leader->nr_siblings--;
1085 if (!list_empty(&event->group_entry))
1086 list = &event->group_entry;
1089 * If this was a group event with sibling events then
1090 * upgrade the siblings to singleton events by adding them
1091 * to whatever list we are on.
1093 list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) {
1095 list_move_tail(&sibling->group_entry, list);
1096 sibling->group_leader = sibling;
1098 /* Inherit group flags from the previous leader */
1099 sibling->group_flags = event->group_flags;
1103 perf_event__header_size(event->group_leader);
1105 list_for_each_entry(tmp, &event->group_leader->sibling_list, group_entry)
1106 perf_event__header_size(tmp);
1110 event_filter_match(struct perf_event *event)
1112 return (event->cpu == -1 || event->cpu == smp_processor_id())
1113 && perf_cgroup_match(event);
1117 event_sched_out(struct perf_event *event,
1118 struct perf_cpu_context *cpuctx,
1119 struct perf_event_context *ctx)
1121 u64 tstamp = perf_event_time(event);
1124 * An event which could not be activated because of
1125 * filter mismatch still needs to have its timings
1126 * maintained, otherwise bogus information is return
1127 * via read() for time_enabled, time_running:
1129 if (event->state == PERF_EVENT_STATE_INACTIVE
1130 && !event_filter_match(event)) {
1131 delta = tstamp - event->tstamp_stopped;
1132 event->tstamp_running += delta;
1133 event->tstamp_stopped = tstamp;
1136 if (event->state != PERF_EVENT_STATE_ACTIVE)
1139 event->state = PERF_EVENT_STATE_INACTIVE;
1140 if (event->pending_disable) {
1141 event->pending_disable = 0;
1142 event->state = PERF_EVENT_STATE_OFF;
1144 event->tstamp_stopped = tstamp;
1145 event->pmu->del(event, 0);
1148 if (!is_software_event(event))
1149 cpuctx->active_oncpu--;
1151 if (event->attr.freq && event->attr.sample_freq)
1153 if (event->attr.exclusive || !cpuctx->active_oncpu)
1154 cpuctx->exclusive = 0;
1158 group_sched_out(struct perf_event *group_event,
1159 struct perf_cpu_context *cpuctx,
1160 struct perf_event_context *ctx)
1162 struct perf_event *event;
1163 int state = group_event->state;
1165 event_sched_out(group_event, cpuctx, ctx);
1168 * Schedule out siblings (if any):
1170 list_for_each_entry(event, &group_event->sibling_list, group_entry)
1171 event_sched_out(event, cpuctx, ctx);
1173 if (state == PERF_EVENT_STATE_ACTIVE && group_event->attr.exclusive)
1174 cpuctx->exclusive = 0;
1178 * Cross CPU call to remove a performance event
1180 * We disable the event on the hardware level first. After that we
1181 * remove it from the context list.
1183 static int __perf_remove_from_context(void *info)
1185 struct perf_event *event = info;
1186 struct perf_event_context *ctx = event->ctx;
1187 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1189 raw_spin_lock(&ctx->lock);
1190 event_sched_out(event, cpuctx, ctx);
1191 list_del_event(event, ctx);
1192 if (!ctx->nr_events && cpuctx->task_ctx == ctx) {
1194 cpuctx->task_ctx = NULL;
1196 raw_spin_unlock(&ctx->lock);
1203 * Remove the event from a task's (or a CPU's) list of events.
1205 * CPU events are removed with a smp call. For task events we only
1206 * call when the task is on a CPU.
1208 * If event->ctx is a cloned context, callers must make sure that
1209 * every task struct that event->ctx->task could possibly point to
1210 * remains valid. This is OK when called from perf_release since
1211 * that only calls us on the top-level context, which can't be a clone.
1212 * When called from perf_event_exit_task, it's OK because the
1213 * context has been detached from its task.
1215 static void perf_remove_from_context(struct perf_event *event)
1217 struct perf_event_context *ctx = event->ctx;
1218 struct task_struct *task = ctx->task;
1220 lockdep_assert_held(&ctx->mutex);
1224 * Per cpu events are removed via an smp call and
1225 * the removal is always successful.
1227 cpu_function_call(event->cpu, __perf_remove_from_context, event);
1232 if (!task_function_call(task, __perf_remove_from_context, event))
1235 raw_spin_lock_irq(&ctx->lock);
1237 * If we failed to find a running task, but find the context active now
1238 * that we've acquired the ctx->lock, retry.
1240 if (ctx->is_active) {
1241 raw_spin_unlock_irq(&ctx->lock);
1246 * Since the task isn't running, its safe to remove the event, us
1247 * holding the ctx->lock ensures the task won't get scheduled in.
1249 list_del_event(event, ctx);
1250 raw_spin_unlock_irq(&ctx->lock);
1254 * Cross CPU call to disable a performance event
1256 int __perf_event_disable(void *info)
1258 struct perf_event *event = info;
1259 struct perf_event_context *ctx = event->ctx;
1260 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1263 * If this is a per-task event, need to check whether this
1264 * event's task is the current task on this cpu.
1266 * Can trigger due to concurrent perf_event_context_sched_out()
1267 * flipping contexts around.
1269 if (ctx->task && cpuctx->task_ctx != ctx)
1272 raw_spin_lock(&ctx->lock);
1275 * If the event is on, turn it off.
1276 * If it is in error state, leave it in error state.
1278 if (event->state >= PERF_EVENT_STATE_INACTIVE) {
1279 update_context_time(ctx);
1280 update_cgrp_time_from_event(event);
1281 update_group_times(event);
1282 if (event == event->group_leader)
1283 group_sched_out(event, cpuctx, ctx);
1285 event_sched_out(event, cpuctx, ctx);
1286 event->state = PERF_EVENT_STATE_OFF;
1289 raw_spin_unlock(&ctx->lock);
1297 * If event->ctx is a cloned context, callers must make sure that
1298 * every task struct that event->ctx->task could possibly point to
1299 * remains valid. This condition is satisifed when called through
1300 * perf_event_for_each_child or perf_event_for_each because they
1301 * hold the top-level event's child_mutex, so any descendant that
1302 * goes to exit will block in sync_child_event.
1303 * When called from perf_pending_event it's OK because event->ctx
1304 * is the current context on this CPU and preemption is disabled,
1305 * hence we can't get into perf_event_task_sched_out for this context.
1307 void perf_event_disable(struct perf_event *event)
1309 struct perf_event_context *ctx = event->ctx;
1310 struct task_struct *task = ctx->task;
1314 * Disable the event on the cpu that it's on
1316 cpu_function_call(event->cpu, __perf_event_disable, event);
1321 if (!task_function_call(task, __perf_event_disable, event))
1324 raw_spin_lock_irq(&ctx->lock);
1326 * If the event is still active, we need to retry the cross-call.
1328 if (event->state == PERF_EVENT_STATE_ACTIVE) {
1329 raw_spin_unlock_irq(&ctx->lock);
1331 * Reload the task pointer, it might have been changed by
1332 * a concurrent perf_event_context_sched_out().
1339 * Since we have the lock this context can't be scheduled
1340 * in, so we can change the state safely.
1342 if (event->state == PERF_EVENT_STATE_INACTIVE) {
1343 update_group_times(event);
1344 event->state = PERF_EVENT_STATE_OFF;
1346 raw_spin_unlock_irq(&ctx->lock);
1348 EXPORT_SYMBOL_GPL(perf_event_disable);
1350 static void perf_set_shadow_time(struct perf_event *event,
1351 struct perf_event_context *ctx,
1355 * use the correct time source for the time snapshot
1357 * We could get by without this by leveraging the
1358 * fact that to get to this function, the caller
1359 * has most likely already called update_context_time()
1360 * and update_cgrp_time_xx() and thus both timestamp
1361 * are identical (or very close). Given that tstamp is,
1362 * already adjusted for cgroup, we could say that:
1363 * tstamp - ctx->timestamp
1365 * tstamp - cgrp->timestamp.
1367 * Then, in perf_output_read(), the calculation would
1368 * work with no changes because:
1369 * - event is guaranteed scheduled in
1370 * - no scheduled out in between
1371 * - thus the timestamp would be the same
1373 * But this is a bit hairy.
1375 * So instead, we have an explicit cgroup call to remain
1376 * within the time time source all along. We believe it
1377 * is cleaner and simpler to understand.
1379 if (is_cgroup_event(event))
1380 perf_cgroup_set_shadow_time(event, tstamp);
1382 event->shadow_ctx_time = tstamp - ctx->timestamp;
1385 #define MAX_INTERRUPTS (~0ULL)
1387 static void perf_log_throttle(struct perf_event *event, int enable);
1390 event_sched_in(struct perf_event *event,
1391 struct perf_cpu_context *cpuctx,
1392 struct perf_event_context *ctx)
1394 u64 tstamp = perf_event_time(event);
1396 if (event->state <= PERF_EVENT_STATE_OFF)
1399 event->state = PERF_EVENT_STATE_ACTIVE;
1400 event->oncpu = smp_processor_id();
1403 * Unthrottle events, since we scheduled we might have missed several
1404 * ticks already, also for a heavily scheduling task there is little
1405 * guarantee it'll get a tick in a timely manner.
1407 if (unlikely(event->hw.interrupts == MAX_INTERRUPTS)) {
1408 perf_log_throttle(event, 1);
1409 event->hw.interrupts = 0;
1413 * The new state must be visible before we turn it on in the hardware:
1417 if (event->pmu->add(event, PERF_EF_START)) {
1418 event->state = PERF_EVENT_STATE_INACTIVE;
1423 event->tstamp_running += tstamp - event->tstamp_stopped;
1425 perf_set_shadow_time(event, ctx, tstamp);
1427 if (!is_software_event(event))
1428 cpuctx->active_oncpu++;
1430 if (event->attr.freq && event->attr.sample_freq)
1433 if (event->attr.exclusive)
1434 cpuctx->exclusive = 1;
1440 group_sched_in(struct perf_event *group_event,
1441 struct perf_cpu_context *cpuctx,
1442 struct perf_event_context *ctx)
1444 struct perf_event *event, *partial_group = NULL;
1445 struct pmu *pmu = group_event->pmu;
1446 u64 now = ctx->time;
1447 bool simulate = false;
1449 if (group_event->state == PERF_EVENT_STATE_OFF)
1452 pmu->start_txn(pmu);
1454 if (event_sched_in(group_event, cpuctx, ctx)) {
1455 pmu->cancel_txn(pmu);
1460 * Schedule in siblings as one group (if any):
1462 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
1463 if (event_sched_in(event, cpuctx, ctx)) {
1464 partial_group = event;
1469 if (!pmu->commit_txn(pmu))
1474 * Groups can be scheduled in as one unit only, so undo any
1475 * partial group before returning:
1476 * The events up to the failed event are scheduled out normally,
1477 * tstamp_stopped will be updated.
1479 * The failed events and the remaining siblings need to have
1480 * their timings updated as if they had gone thru event_sched_in()
1481 * and event_sched_out(). This is required to get consistent timings
1482 * across the group. This also takes care of the case where the group
1483 * could never be scheduled by ensuring tstamp_stopped is set to mark
1484 * the time the event was actually stopped, such that time delta
1485 * calculation in update_event_times() is correct.
1487 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
1488 if (event == partial_group)
1492 event->tstamp_running += now - event->tstamp_stopped;
1493 event->tstamp_stopped = now;
1495 event_sched_out(event, cpuctx, ctx);
1498 event_sched_out(group_event, cpuctx, ctx);
1500 pmu->cancel_txn(pmu);
1506 * Work out whether we can put this event group on the CPU now.
1508 static int group_can_go_on(struct perf_event *event,
1509 struct perf_cpu_context *cpuctx,
1513 * Groups consisting entirely of software events can always go on.
1515 if (event->group_flags & PERF_GROUP_SOFTWARE)
1518 * If an exclusive group is already on, no other hardware
1521 if (cpuctx->exclusive)
1524 * If this group is exclusive and there are already
1525 * events on the CPU, it can't go on.
1527 if (event->attr.exclusive && cpuctx->active_oncpu)
1530 * Otherwise, try to add it if all previous groups were able
1536 static void add_event_to_ctx(struct perf_event *event,
1537 struct perf_event_context *ctx)
1539 u64 tstamp = perf_event_time(event);
1541 list_add_event(event, ctx);
1542 perf_group_attach(event);
1543 event->tstamp_enabled = tstamp;
1544 event->tstamp_running = tstamp;
1545 event->tstamp_stopped = tstamp;
1548 static void task_ctx_sched_out(struct perf_event_context *ctx);
1550 ctx_sched_in(struct perf_event_context *ctx,
1551 struct perf_cpu_context *cpuctx,
1552 enum event_type_t event_type,
1553 struct task_struct *task);
1555 static void perf_event_sched_in(struct perf_cpu_context *cpuctx,
1556 struct perf_event_context *ctx,
1557 struct task_struct *task)
1559 cpu_ctx_sched_in(cpuctx, EVENT_PINNED, task);
1561 ctx_sched_in(ctx, cpuctx, EVENT_PINNED, task);
1562 cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE, task);
1564 ctx_sched_in(ctx, cpuctx, EVENT_FLEXIBLE, task);
1568 * Cross CPU call to install and enable a performance event
1570 * Must be called with ctx->mutex held
1572 static int __perf_install_in_context(void *info)
1574 struct perf_event *event = info;
1575 struct perf_event_context *ctx = event->ctx;
1576 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1577 struct perf_event_context *task_ctx = cpuctx->task_ctx;
1578 struct task_struct *task = current;
1580 perf_ctx_lock(cpuctx, task_ctx);
1581 perf_pmu_disable(cpuctx->ctx.pmu);
1584 * If there was an active task_ctx schedule it out.
1587 task_ctx_sched_out(task_ctx);
1590 * If the context we're installing events in is not the
1591 * active task_ctx, flip them.
1593 if (ctx->task && task_ctx != ctx) {
1595 raw_spin_unlock(&task_ctx->lock);
1596 raw_spin_lock(&ctx->lock);
1601 cpuctx->task_ctx = task_ctx;
1602 task = task_ctx->task;
1605 cpu_ctx_sched_out(cpuctx, EVENT_ALL);
1607 update_context_time(ctx);
1609 * update cgrp time only if current cgrp
1610 * matches event->cgrp. Must be done before
1611 * calling add_event_to_ctx()
1613 update_cgrp_time_from_event(event);
1615 add_event_to_ctx(event, ctx);
1618 * Schedule everything back in
1620 perf_event_sched_in(cpuctx, task_ctx, task);
1622 perf_pmu_enable(cpuctx->ctx.pmu);
1623 perf_ctx_unlock(cpuctx, task_ctx);
1629 * Attach a performance event to a context
1631 * First we add the event to the list with the hardware enable bit
1632 * in event->hw_config cleared.
1634 * If the event is attached to a task which is on a CPU we use a smp
1635 * call to enable it in the task context. The task might have been
1636 * scheduled away, but we check this in the smp call again.
1639 perf_install_in_context(struct perf_event_context *ctx,
1640 struct perf_event *event,
1643 struct task_struct *task = ctx->task;
1645 lockdep_assert_held(&ctx->mutex);
1648 if (event->cpu != -1)
1653 * Per cpu events are installed via an smp call and
1654 * the install is always successful.
1656 cpu_function_call(cpu, __perf_install_in_context, event);
1661 if (!task_function_call(task, __perf_install_in_context, event))
1664 raw_spin_lock_irq(&ctx->lock);
1666 * If we failed to find a running task, but find the context active now
1667 * that we've acquired the ctx->lock, retry.
1669 if (ctx->is_active) {
1670 raw_spin_unlock_irq(&ctx->lock);
1675 * Since the task isn't running, its safe to add the event, us holding
1676 * the ctx->lock ensures the task won't get scheduled in.
1678 add_event_to_ctx(event, ctx);
1679 raw_spin_unlock_irq(&ctx->lock);
1683 * Put a event into inactive state and update time fields.
1684 * Enabling the leader of a group effectively enables all
1685 * the group members that aren't explicitly disabled, so we
1686 * have to update their ->tstamp_enabled also.
1687 * Note: this works for group members as well as group leaders
1688 * since the non-leader members' sibling_lists will be empty.
1690 static void __perf_event_mark_enabled(struct perf_event *event)
1692 struct perf_event *sub;
1693 u64 tstamp = perf_event_time(event);
1695 event->state = PERF_EVENT_STATE_INACTIVE;
1696 event->tstamp_enabled = tstamp - event->total_time_enabled;
1697 list_for_each_entry(sub, &event->sibling_list, group_entry) {
1698 if (sub->state >= PERF_EVENT_STATE_INACTIVE)
1699 sub->tstamp_enabled = tstamp - sub->total_time_enabled;
1704 * Cross CPU call to enable a performance event
1706 static int __perf_event_enable(void *info)
1708 struct perf_event *event = info;
1709 struct perf_event_context *ctx = event->ctx;
1710 struct perf_event *leader = event->group_leader;
1711 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1714 if (WARN_ON_ONCE(!ctx->is_active))
1717 raw_spin_lock(&ctx->lock);
1718 update_context_time(ctx);
1720 if (event->state >= PERF_EVENT_STATE_INACTIVE)
1724 * set current task's cgroup time reference point
1726 perf_cgroup_set_timestamp(current, ctx);
1728 __perf_event_mark_enabled(event);
1730 if (!event_filter_match(event)) {
1731 if (is_cgroup_event(event))
1732 perf_cgroup_defer_enabled(event);
1737 * If the event is in a group and isn't the group leader,
1738 * then don't put it on unless the group is on.
1740 if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE)
1743 if (!group_can_go_on(event, cpuctx, 1)) {
1746 if (event == leader)
1747 err = group_sched_in(event, cpuctx, ctx);
1749 err = event_sched_in(event, cpuctx, ctx);
1754 * If this event can't go on and it's part of a
1755 * group, then the whole group has to come off.
1757 if (leader != event)
1758 group_sched_out(leader, cpuctx, ctx);
1759 if (leader->attr.pinned) {
1760 update_group_times(leader);
1761 leader->state = PERF_EVENT_STATE_ERROR;
1766 raw_spin_unlock(&ctx->lock);
1774 * If event->ctx is a cloned context, callers must make sure that
1775 * every task struct that event->ctx->task could possibly point to
1776 * remains valid. This condition is satisfied when called through
1777 * perf_event_for_each_child or perf_event_for_each as described
1778 * for perf_event_disable.
1780 void perf_event_enable(struct perf_event *event)
1782 struct perf_event_context *ctx = event->ctx;
1783 struct task_struct *task = ctx->task;
1787 * Enable the event on the cpu that it's on
1789 cpu_function_call(event->cpu, __perf_event_enable, event);
1793 raw_spin_lock_irq(&ctx->lock);
1794 if (event->state >= PERF_EVENT_STATE_INACTIVE)
1798 * If the event is in error state, clear that first.
1799 * That way, if we see the event in error state below, we
1800 * know that it has gone back into error state, as distinct
1801 * from the task having been scheduled away before the
1802 * cross-call arrived.
1804 if (event->state == PERF_EVENT_STATE_ERROR)
1805 event->state = PERF_EVENT_STATE_OFF;
1808 if (!ctx->is_active) {
1809 __perf_event_mark_enabled(event);
1813 raw_spin_unlock_irq(&ctx->lock);
1815 if (!task_function_call(task, __perf_event_enable, event))
1818 raw_spin_lock_irq(&ctx->lock);
1821 * If the context is active and the event is still off,
1822 * we need to retry the cross-call.
1824 if (ctx->is_active && event->state == PERF_EVENT_STATE_OFF) {
1826 * task could have been flipped by a concurrent
1827 * perf_event_context_sched_out()
1834 raw_spin_unlock_irq(&ctx->lock);
1836 EXPORT_SYMBOL_GPL(perf_event_enable);
1838 int perf_event_refresh(struct perf_event *event, int refresh)
1841 * not supported on inherited events
1843 if (event->attr.inherit || !is_sampling_event(event))
1846 atomic_add(refresh, &event->event_limit);
1847 perf_event_enable(event);
1851 EXPORT_SYMBOL_GPL(perf_event_refresh);
1853 static void ctx_sched_out(struct perf_event_context *ctx,
1854 struct perf_cpu_context *cpuctx,
1855 enum event_type_t event_type)
1857 struct perf_event *event;
1858 int is_active = ctx->is_active;
1860 ctx->is_active &= ~event_type;
1861 if (likely(!ctx->nr_events))
1864 update_context_time(ctx);
1865 update_cgrp_time_from_cpuctx(cpuctx);
1866 if (!ctx->nr_active)
1869 perf_pmu_disable(ctx->pmu);
1870 if ((is_active & EVENT_PINNED) && (event_type & EVENT_PINNED)) {
1871 list_for_each_entry(event, &ctx->pinned_groups, group_entry)
1872 group_sched_out(event, cpuctx, ctx);
1875 if ((is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE)) {
1876 list_for_each_entry(event, &ctx->flexible_groups, group_entry)
1877 group_sched_out(event, cpuctx, ctx);
1879 perf_pmu_enable(ctx->pmu);
1883 * Test whether two contexts are equivalent, i.e. whether they
1884 * have both been cloned from the same version of the same context
1885 * and they both have the same number of enabled events.
1886 * If the number of enabled events is the same, then the set
1887 * of enabled events should be the same, because these are both
1888 * inherited contexts, therefore we can't access individual events
1889 * in them directly with an fd; we can only enable/disable all
1890 * events via prctl, or enable/disable all events in a family
1891 * via ioctl, which will have the same effect on both contexts.
1893 static int context_equiv(struct perf_event_context *ctx1,
1894 struct perf_event_context *ctx2)
1896 return ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx
1897 && ctx1->parent_gen == ctx2->parent_gen
1898 && !ctx1->pin_count && !ctx2->pin_count;
1901 static void __perf_event_sync_stat(struct perf_event *event,
1902 struct perf_event *next_event)
1906 if (!event->attr.inherit_stat)
1910 * Update the event value, we cannot use perf_event_read()
1911 * because we're in the middle of a context switch and have IRQs
1912 * disabled, which upsets smp_call_function_single(), however
1913 * we know the event must be on the current CPU, therefore we
1914 * don't need to use it.
1916 switch (event->state) {
1917 case PERF_EVENT_STATE_ACTIVE:
1918 event->pmu->read(event);
1921 case PERF_EVENT_STATE_INACTIVE:
1922 update_event_times(event);
1930 * In order to keep per-task stats reliable we need to flip the event
1931 * values when we flip the contexts.
1933 value = local64_read(&next_event->count);
1934 value = local64_xchg(&event->count, value);
1935 local64_set(&next_event->count, value);
1937 swap(event->total_time_enabled, next_event->total_time_enabled);
1938 swap(event->total_time_running, next_event->total_time_running);
1941 * Since we swizzled the values, update the user visible data too.
1943 perf_event_update_userpage(event);
1944 perf_event_update_userpage(next_event);
1947 #define list_next_entry(pos, member) \
1948 list_entry(pos->member.next, typeof(*pos), member)
1950 static void perf_event_sync_stat(struct perf_event_context *ctx,
1951 struct perf_event_context *next_ctx)
1953 struct perf_event *event, *next_event;
1958 update_context_time(ctx);
1960 event = list_first_entry(&ctx->event_list,
1961 struct perf_event, event_entry);
1963 next_event = list_first_entry(&next_ctx->event_list,
1964 struct perf_event, event_entry);
1966 while (&event->event_entry != &ctx->event_list &&
1967 &next_event->event_entry != &next_ctx->event_list) {
1969 __perf_event_sync_stat(event, next_event);
1971 event = list_next_entry(event, event_entry);
1972 next_event = list_next_entry(next_event, event_entry);
1976 static void perf_event_context_sched_out(struct task_struct *task, int ctxn,
1977 struct task_struct *next)
1979 struct perf_event_context *ctx = task->perf_event_ctxp[ctxn];
1980 struct perf_event_context *next_ctx;
1981 struct perf_event_context *parent;
1982 struct perf_cpu_context *cpuctx;
1988 cpuctx = __get_cpu_context(ctx);
1989 if (!cpuctx->task_ctx)
1993 parent = rcu_dereference(ctx->parent_ctx);
1994 next_ctx = next->perf_event_ctxp[ctxn];
1995 if (parent && next_ctx &&
1996 rcu_dereference(next_ctx->parent_ctx) == parent) {
1998 * Looks like the two contexts are clones, so we might be
1999 * able to optimize the context switch. We lock both
2000 * contexts and check that they are clones under the
2001 * lock (including re-checking that neither has been
2002 * uncloned in the meantime). It doesn't matter which
2003 * order we take the locks because no other cpu could
2004 * be trying to lock both of these tasks.
2006 raw_spin_lock(&ctx->lock);
2007 raw_spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING);
2008 if (context_equiv(ctx, next_ctx)) {
2010 * XXX do we need a memory barrier of sorts
2011 * wrt to rcu_dereference() of perf_event_ctxp
2013 task->perf_event_ctxp[ctxn] = next_ctx;
2014 next->perf_event_ctxp[ctxn] = ctx;
2016 next_ctx->task = task;
2019 perf_event_sync_stat(ctx, next_ctx);
2021 raw_spin_unlock(&next_ctx->lock);
2022 raw_spin_unlock(&ctx->lock);
2027 raw_spin_lock(&ctx->lock);
2028 ctx_sched_out(ctx, cpuctx, EVENT_ALL);
2029 cpuctx->task_ctx = NULL;
2030 raw_spin_unlock(&ctx->lock);
2034 #define for_each_task_context_nr(ctxn) \
2035 for ((ctxn) = 0; (ctxn) < perf_nr_task_contexts; (ctxn)++)
2038 * Called from scheduler to remove the events of the current task,
2039 * with interrupts disabled.
2041 * We stop each event and update the event value in event->count.
2043 * This does not protect us against NMI, but disable()
2044 * sets the disabled bit in the control field of event _before_
2045 * accessing the event control register. If a NMI hits, then it will
2046 * not restart the event.
2048 void __perf_event_task_sched_out(struct task_struct *task,
2049 struct task_struct *next)
2053 for_each_task_context_nr(ctxn)
2054 perf_event_context_sched_out(task, ctxn, next);
2057 * if cgroup events exist on this CPU, then we need
2058 * to check if we have to switch out PMU state.
2059 * cgroup event are system-wide mode only
2061 if (atomic_read(&__get_cpu_var(perf_cgroup_events)))
2062 perf_cgroup_sched_out(task, next);
2065 static void task_ctx_sched_out(struct perf_event_context *ctx)
2067 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
2069 if (!cpuctx->task_ctx)
2072 if (WARN_ON_ONCE(ctx != cpuctx->task_ctx))
2075 ctx_sched_out(ctx, cpuctx, EVENT_ALL);
2076 cpuctx->task_ctx = NULL;
2080 * Called with IRQs disabled
2082 static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
2083 enum event_type_t event_type)
2085 ctx_sched_out(&cpuctx->ctx, cpuctx, event_type);
2089 ctx_pinned_sched_in(struct perf_event_context *ctx,
2090 struct perf_cpu_context *cpuctx)
2092 struct perf_event *event;
2094 list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
2095 if (event->state <= PERF_EVENT_STATE_OFF)
2097 if (!event_filter_match(event))
2100 /* may need to reset tstamp_enabled */
2101 if (is_cgroup_event(event))
2102 perf_cgroup_mark_enabled(event, ctx);
2104 if (group_can_go_on(event, cpuctx, 1))
2105 group_sched_in(event, cpuctx, ctx);
2108 * If this pinned group hasn't been scheduled,
2109 * put it in error state.
2111 if (event->state == PERF_EVENT_STATE_INACTIVE) {
2112 update_group_times(event);
2113 event->state = PERF_EVENT_STATE_ERROR;
2119 ctx_flexible_sched_in(struct perf_event_context *ctx,
2120 struct perf_cpu_context *cpuctx)
2122 struct perf_event *event;
2125 list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
2126 /* Ignore events in OFF or ERROR state */
2127 if (event->state <= PERF_EVENT_STATE_OFF)
2130 * Listen to the 'cpu' scheduling filter constraint
2133 if (!event_filter_match(event))
2136 /* may need to reset tstamp_enabled */
2137 if (is_cgroup_event(event))
2138 perf_cgroup_mark_enabled(event, ctx);
2140 if (group_can_go_on(event, cpuctx, can_add_hw)) {
2141 if (group_sched_in(event, cpuctx, ctx))
2148 ctx_sched_in(struct perf_event_context *ctx,
2149 struct perf_cpu_context *cpuctx,
2150 enum event_type_t event_type,
2151 struct task_struct *task)
2154 int is_active = ctx->is_active;
2156 ctx->is_active |= event_type;
2157 if (likely(!ctx->nr_events))
2161 ctx->timestamp = now;
2162 perf_cgroup_set_timestamp(task, ctx);
2164 * First go through the list and put on any pinned groups
2165 * in order to give them the best chance of going on.
2167 if (!(is_active & EVENT_PINNED) && (event_type & EVENT_PINNED))
2168 ctx_pinned_sched_in(ctx, cpuctx);
2170 /* Then walk through the lower prio flexible groups */
2171 if (!(is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE))
2172 ctx_flexible_sched_in(ctx, cpuctx);
2175 static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
2176 enum event_type_t event_type,
2177 struct task_struct *task)
2179 struct perf_event_context *ctx = &cpuctx->ctx;
2181 ctx_sched_in(ctx, cpuctx, event_type, task);
2184 static void perf_event_context_sched_in(struct perf_event_context *ctx,
2185 struct task_struct *task)
2187 struct perf_cpu_context *cpuctx;
2189 cpuctx = __get_cpu_context(ctx);
2190 if (cpuctx->task_ctx == ctx)
2193 perf_ctx_lock(cpuctx, ctx);
2194 perf_pmu_disable(ctx->pmu);
2196 * We want to keep the following priority order:
2197 * cpu pinned (that don't need to move), task pinned,
2198 * cpu flexible, task flexible.
2200 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
2203 cpuctx->task_ctx = ctx;
2205 perf_event_sched_in(cpuctx, cpuctx->task_ctx, task);
2207 perf_pmu_enable(ctx->pmu);
2208 perf_ctx_unlock(cpuctx, ctx);
2211 * Since these rotations are per-cpu, we need to ensure the
2212 * cpu-context we got scheduled on is actually rotating.
2214 perf_pmu_rotate_start(ctx->pmu);
2218 * When sampling the branck stack in system-wide, it may be necessary
2219 * to flush the stack on context switch. This happens when the branch
2220 * stack does not tag its entries with the pid of the current task.
2221 * Otherwise it becomes impossible to associate a branch entry with a
2222 * task. This ambiguity is more likely to appear when the branch stack
2223 * supports priv level filtering and the user sets it to monitor only
2224 * at the user level (which could be a useful measurement in system-wide
2225 * mode). In that case, the risk is high of having a branch stack with
2226 * branch from multiple tasks. Flushing may mean dropping the existing
2227 * entries or stashing them somewhere in the PMU specific code layer.
2229 * This function provides the context switch callback to the lower code
2230 * layer. It is invoked ONLY when there is at least one system-wide context
2231 * with at least one active event using taken branch sampling.
2233 static void perf_branch_stack_sched_in(struct task_struct *prev,
2234 struct task_struct *task)
2236 struct perf_cpu_context *cpuctx;
2238 unsigned long flags;
2240 /* no need to flush branch stack if not changing task */
2244 local_irq_save(flags);
2248 list_for_each_entry_rcu(pmu, &pmus, entry) {
2249 cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
2252 * check if the context has at least one
2253 * event using PERF_SAMPLE_BRANCH_STACK
2255 if (cpuctx->ctx.nr_branch_stack > 0
2256 && pmu->flush_branch_stack) {
2258 pmu = cpuctx->ctx.pmu;
2260 perf_ctx_lock(cpuctx, cpuctx->task_ctx);
2262 perf_pmu_disable(pmu);
2264 pmu->flush_branch_stack();
2266 perf_pmu_enable(pmu);
2268 perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
2274 local_irq_restore(flags);
2278 * Called from scheduler to add the events of the current task
2279 * with interrupts disabled.
2281 * We restore the event value and then enable it.
2283 * This does not protect us against NMI, but enable()
2284 * sets the enabled bit in the control field of event _before_
2285 * accessing the event control register. If a NMI hits, then it will
2286 * keep the event running.
2288 void __perf_event_task_sched_in(struct task_struct *prev,
2289 struct task_struct *task)
2291 struct perf_event_context *ctx;
2294 for_each_task_context_nr(ctxn) {
2295 ctx = task->perf_event_ctxp[ctxn];
2299 perf_event_context_sched_in(ctx, task);
2302 * if cgroup events exist on this CPU, then we need
2303 * to check if we have to switch in PMU state.
2304 * cgroup event are system-wide mode only
2306 if (atomic_read(&__get_cpu_var(perf_cgroup_events)))
2307 perf_cgroup_sched_in(prev, task);
2309 /* check for system-wide branch_stack events */
2310 if (atomic_read(&__get_cpu_var(perf_branch_stack_events)))
2311 perf_branch_stack_sched_in(prev, task);
2314 static u64 perf_calculate_period(struct perf_event *event, u64 nsec, u64 count)
2316 u64 frequency = event->attr.sample_freq;
2317 u64 sec = NSEC_PER_SEC;
2318 u64 divisor, dividend;
2320 int count_fls, nsec_fls, frequency_fls, sec_fls;
2322 count_fls = fls64(count);
2323 nsec_fls = fls64(nsec);
2324 frequency_fls = fls64(frequency);
2328 * We got @count in @nsec, with a target of sample_freq HZ
2329 * the target period becomes:
2332 * period = -------------------
2333 * @nsec * sample_freq
2338 * Reduce accuracy by one bit such that @a and @b converge
2339 * to a similar magnitude.
2341 #define REDUCE_FLS(a, b) \
2343 if (a##_fls > b##_fls) { \
2353 * Reduce accuracy until either term fits in a u64, then proceed with
2354 * the other, so that finally we can do a u64/u64 division.
2356 while (count_fls + sec_fls > 64 && nsec_fls + frequency_fls > 64) {
2357 REDUCE_FLS(nsec, frequency);
2358 REDUCE_FLS(sec, count);
2361 if (count_fls + sec_fls > 64) {
2362 divisor = nsec * frequency;
2364 while (count_fls + sec_fls > 64) {
2365 REDUCE_FLS(count, sec);
2369 dividend = count * sec;
2371 dividend = count * sec;
2373 while (nsec_fls + frequency_fls > 64) {
2374 REDUCE_FLS(nsec, frequency);
2378 divisor = nsec * frequency;
2384 return div64_u64(dividend, divisor);
2387 static DEFINE_PER_CPU(int, perf_throttled_count);
2388 static DEFINE_PER_CPU(u64, perf_throttled_seq);
2390 static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count, bool disable)
2392 struct hw_perf_event *hwc = &event->hw;
2393 s64 period, sample_period;
2396 period = perf_calculate_period(event, nsec, count);
2398 delta = (s64)(period - hwc->sample_period);
2399 delta = (delta + 7) / 8; /* low pass filter */
2401 sample_period = hwc->sample_period + delta;
2406 hwc->sample_period = sample_period;
2408 if (local64_read(&hwc->period_left) > 8*sample_period) {
2410 event->pmu->stop(event, PERF_EF_UPDATE);
2412 local64_set(&hwc->period_left, 0);
2415 event->pmu->start(event, PERF_EF_RELOAD);
2420 * combine freq adjustment with unthrottling to avoid two passes over the
2421 * events. At the same time, make sure, having freq events does not change
2422 * the rate of unthrottling as that would introduce bias.
2424 static void perf_adjust_freq_unthr_context(struct perf_event_context *ctx,
2427 struct perf_event *event;
2428 struct hw_perf_event *hwc;
2429 u64 now, period = TICK_NSEC;
2433 * only need to iterate over all events iff:
2434 * - context have events in frequency mode (needs freq adjust)
2435 * - there are events to unthrottle on this cpu
2437 if (!(ctx->nr_freq || needs_unthr))
2440 raw_spin_lock(&ctx->lock);
2441 perf_pmu_disable(ctx->pmu);
2443 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
2444 if (event->state != PERF_EVENT_STATE_ACTIVE)
2447 if (!event_filter_match(event))
2452 if (needs_unthr && hwc->interrupts == MAX_INTERRUPTS) {
2453 hwc->interrupts = 0;
2454 perf_log_throttle(event, 1);
2455 event->pmu->start(event, 0);
2458 if (!event->attr.freq || !event->attr.sample_freq)
2462 * stop the event and update event->count
2464 event->pmu->stop(event, PERF_EF_UPDATE);
2466 now = local64_read(&event->count);
2467 delta = now - hwc->freq_count_stamp;
2468 hwc->freq_count_stamp = now;
2472 * reload only if value has changed
2473 * we have stopped the event so tell that
2474 * to perf_adjust_period() to avoid stopping it
2478 perf_adjust_period(event, period, delta, false);
2480 event->pmu->start(event, delta > 0 ? PERF_EF_RELOAD : 0);
2483 perf_pmu_enable(ctx->pmu);
2484 raw_spin_unlock(&ctx->lock);
2488 * Round-robin a context's events:
2490 static void rotate_ctx(struct perf_event_context *ctx)
2493 * Rotate the first entry last of non-pinned groups. Rotation might be
2494 * disabled by the inheritance code.
2496 if (!ctx->rotate_disable)
2497 list_rotate_left(&ctx->flexible_groups);
2501 * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
2502 * because they're strictly cpu affine and rotate_start is called with IRQs
2503 * disabled, while rotate_context is called from IRQ context.
2505 static void perf_rotate_context(struct perf_cpu_context *cpuctx)
2507 struct perf_event_context *ctx = NULL;
2508 int rotate = 0, remove = 1;
2510 if (cpuctx->ctx.nr_events) {
2512 if (cpuctx->ctx.nr_events != cpuctx->ctx.nr_active)
2516 ctx = cpuctx->task_ctx;
2517 if (ctx && ctx->nr_events) {
2519 if (ctx->nr_events != ctx->nr_active)
2526 perf_ctx_lock(cpuctx, cpuctx->task_ctx);
2527 perf_pmu_disable(cpuctx->ctx.pmu);
2529 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
2531 ctx_sched_out(ctx, cpuctx, EVENT_FLEXIBLE);
2533 rotate_ctx(&cpuctx->ctx);
2537 perf_event_sched_in(cpuctx, ctx, current);
2539 perf_pmu_enable(cpuctx->ctx.pmu);
2540 perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
2543 list_del_init(&cpuctx->rotation_list);
2546 void perf_event_task_tick(void)
2548 struct list_head *head = &__get_cpu_var(rotation_list);
2549 struct perf_cpu_context *cpuctx, *tmp;
2550 struct perf_event_context *ctx;
2553 WARN_ON(!irqs_disabled());
2555 __this_cpu_inc(perf_throttled_seq);
2556 throttled = __this_cpu_xchg(perf_throttled_count, 0);
2558 list_for_each_entry_safe(cpuctx, tmp, head, rotation_list) {
2560 perf_adjust_freq_unthr_context(ctx, throttled);
2562 ctx = cpuctx->task_ctx;
2564 perf_adjust_freq_unthr_context(ctx, throttled);
2566 if (cpuctx->jiffies_interval == 1 ||
2567 !(jiffies % cpuctx->jiffies_interval))
2568 perf_rotate_context(cpuctx);
2572 static int event_enable_on_exec(struct perf_event *event,
2573 struct perf_event_context *ctx)
2575 if (!event->attr.enable_on_exec)
2578 event->attr.enable_on_exec = 0;
2579 if (event->state >= PERF_EVENT_STATE_INACTIVE)
2582 __perf_event_mark_enabled(event);
2588 * Enable all of a task's events that have been marked enable-on-exec.
2589 * This expects task == current.
2591 static void perf_event_enable_on_exec(struct perf_event_context *ctx)
2593 struct perf_event *event;
2594 unsigned long flags;
2598 local_irq_save(flags);
2599 if (!ctx || !ctx->nr_events)
2603 * We must ctxsw out cgroup events to avoid conflict
2604 * when invoking perf_task_event_sched_in() later on
2605 * in this function. Otherwise we end up trying to
2606 * ctxswin cgroup events which are already scheduled
2609 perf_cgroup_sched_out(current, NULL);
2611 raw_spin_lock(&ctx->lock);
2612 task_ctx_sched_out(ctx);
2614 list_for_each_entry(event, &ctx->event_list, event_entry) {
2615 ret = event_enable_on_exec(event, ctx);
2621 * Unclone this context if we enabled any event.
2626 raw_spin_unlock(&ctx->lock);
2629 * Also calls ctxswin for cgroup events, if any:
2631 perf_event_context_sched_in(ctx, ctx->task);
2633 local_irq_restore(flags);
2637 * Cross CPU call to read the hardware event
2639 static void __perf_event_read(void *info)
2641 struct perf_event *event = info;
2642 struct perf_event_context *ctx = event->ctx;
2643 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
2646 * If this is a task context, we need to check whether it is
2647 * the current task context of this cpu. If not it has been
2648 * scheduled out before the smp call arrived. In that case
2649 * event->count would have been updated to a recent sample
2650 * when the event was scheduled out.
2652 if (ctx->task && cpuctx->task_ctx != ctx)
2655 raw_spin_lock(&ctx->lock);
2656 if (ctx->is_active) {
2657 update_context_time(ctx);
2658 update_cgrp_time_from_event(event);
2660 update_event_times(event);
2661 if (event->state == PERF_EVENT_STATE_ACTIVE)
2662 event->pmu->read(event);
2663 raw_spin_unlock(&ctx->lock);
2666 static inline u64 perf_event_count(struct perf_event *event)
2668 return local64_read(&event->count) + atomic64_read(&event->child_count);
2671 static u64 perf_event_read(struct perf_event *event)
2674 * If event is enabled and currently active on a CPU, update the
2675 * value in the event structure:
2677 if (event->state == PERF_EVENT_STATE_ACTIVE) {
2678 smp_call_function_single(event->oncpu,
2679 __perf_event_read, event, 1);
2680 } else if (event->state == PERF_EVENT_STATE_INACTIVE) {
2681 struct perf_event_context *ctx = event->ctx;
2682 unsigned long flags;
2684 raw_spin_lock_irqsave(&ctx->lock, flags);
2686 * may read while context is not active
2687 * (e.g., thread is blocked), in that case
2688 * we cannot update context time
2690 if (ctx->is_active) {
2691 update_context_time(ctx);
2692 update_cgrp_time_from_event(event);
2694 update_event_times(event);
2695 raw_spin_unlock_irqrestore(&ctx->lock, flags);
2698 return perf_event_count(event);
2702 * Initialize the perf_event context in a task_struct:
2704 static void __perf_event_init_context(struct perf_event_context *ctx)
2706 raw_spin_lock_init(&ctx->lock);
2707 mutex_init(&ctx->mutex);
2708 INIT_LIST_HEAD(&ctx->pinned_groups);
2709 INIT_LIST_HEAD(&ctx->flexible_groups);
2710 INIT_LIST_HEAD(&ctx->event_list);
2711 atomic_set(&ctx->refcount, 1);
2714 static struct perf_event_context *
2715 alloc_perf_context(struct pmu *pmu, struct task_struct *task)
2717 struct perf_event_context *ctx;
2719 ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL);
2723 __perf_event_init_context(ctx);
2726 get_task_struct(task);
2733 static struct task_struct *
2734 find_lively_task_by_vpid(pid_t vpid)
2736 struct task_struct *task;
2743 task = find_task_by_vpid(vpid);
2745 get_task_struct(task);
2749 return ERR_PTR(-ESRCH);
2751 /* Reuse ptrace permission checks for now. */
2753 if (!ptrace_may_access(task, PTRACE_MODE_READ))
2758 put_task_struct(task);
2759 return ERR_PTR(err);
2764 * Returns a matching context with refcount and pincount.
2766 static struct perf_event_context *
2767 find_get_context(struct pmu *pmu, struct task_struct *task, int cpu)
2769 struct perf_event_context *ctx;
2770 struct perf_cpu_context *cpuctx;
2771 unsigned long flags;
2775 /* Must be root to operate on a CPU event: */
2776 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
2777 return ERR_PTR(-EACCES);
2780 * We could be clever and allow to attach a event to an
2781 * offline CPU and activate it when the CPU comes up, but
2784 if (!cpu_online(cpu))
2785 return ERR_PTR(-ENODEV);
2787 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
2796 ctxn = pmu->task_ctx_nr;
2801 ctx = perf_lock_task_context(task, ctxn, &flags);
2805 raw_spin_unlock_irqrestore(&ctx->lock, flags);
2807 ctx = alloc_perf_context(pmu, task);
2813 mutex_lock(&task->perf_event_mutex);
2815 * If it has already passed perf_event_exit_task().
2816 * we must see PF_EXITING, it takes this mutex too.
2818 if (task->flags & PF_EXITING)
2820 else if (task->perf_event_ctxp[ctxn])
2825 rcu_assign_pointer(task->perf_event_ctxp[ctxn], ctx);
2827 mutex_unlock(&task->perf_event_mutex);
2829 if (unlikely(err)) {
2841 return ERR_PTR(err);
2844 static void perf_event_free_filter(struct perf_event *event);
2846 static void free_event_rcu(struct rcu_head *head)
2848 struct perf_event *event;
2850 event = container_of(head, struct perf_event, rcu_head);
2852 put_pid_ns(event->ns);
2853 perf_event_free_filter(event);
2857 static void ring_buffer_put(struct ring_buffer *rb);
2859 static void free_event(struct perf_event *event)
2861 irq_work_sync(&event->pending);
2863 if (!event->parent) {
2864 if (event->attach_state & PERF_ATTACH_TASK)
2865 static_key_slow_dec_deferred(&perf_sched_events);
2866 if (event->attr.mmap || event->attr.mmap_data)
2867 atomic_dec(&nr_mmap_events);
2868 if (event->attr.comm)
2869 atomic_dec(&nr_comm_events);
2870 if (event->attr.task)
2871 atomic_dec(&nr_task_events);
2872 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN)
2873 put_callchain_buffers();
2874 if (is_cgroup_event(event)) {
2875 atomic_dec(&per_cpu(perf_cgroup_events, event->cpu));
2876 static_key_slow_dec_deferred(&perf_sched_events);
2879 if (has_branch_stack(event)) {
2880 static_key_slow_dec_deferred(&perf_sched_events);
2881 /* is system-wide event */
2882 if (!(event->attach_state & PERF_ATTACH_TASK))
2883 atomic_dec(&per_cpu(perf_branch_stack_events,
2889 ring_buffer_put(event->rb);
2893 if (is_cgroup_event(event))
2894 perf_detach_cgroup(event);
2897 event->destroy(event);
2900 put_ctx(event->ctx);
2902 call_rcu(&event->rcu_head, free_event_rcu);
2905 int perf_event_release_kernel(struct perf_event *event)
2907 struct perf_event_context *ctx = event->ctx;
2909 WARN_ON_ONCE(ctx->parent_ctx);
2911 * There are two ways this annotation is useful:
2913 * 1) there is a lock recursion from perf_event_exit_task
2914 * see the comment there.
2916 * 2) there is a lock-inversion with mmap_sem through
2917 * perf_event_read_group(), which takes faults while
2918 * holding ctx->mutex, however this is called after
2919 * the last filedesc died, so there is no possibility
2920 * to trigger the AB-BA case.
2922 mutex_lock_nested(&ctx->mutex, SINGLE_DEPTH_NESTING);
2923 raw_spin_lock_irq(&ctx->lock);
2924 perf_group_detach(event);
2925 raw_spin_unlock_irq(&ctx->lock);
2926 perf_remove_from_context(event);
2927 mutex_unlock(&ctx->mutex);
2933 EXPORT_SYMBOL_GPL(perf_event_release_kernel);
2936 * Called when the last reference to the file is gone.
2938 static void put_event(struct perf_event *event)
2940 struct task_struct *owner;
2942 if (!atomic_long_dec_and_test(&event->refcount))
2946 owner = ACCESS_ONCE(event->owner);
2948 * Matches the smp_wmb() in perf_event_exit_task(). If we observe
2949 * !owner it means the list deletion is complete and we can indeed
2950 * free this event, otherwise we need to serialize on
2951 * owner->perf_event_mutex.
2953 smp_read_barrier_depends();
2956 * Since delayed_put_task_struct() also drops the last
2957 * task reference we can safely take a new reference
2958 * while holding the rcu_read_lock().
2960 get_task_struct(owner);
2965 mutex_lock(&owner->perf_event_mutex);
2967 * We have to re-check the event->owner field, if it is cleared
2968 * we raced with perf_event_exit_task(), acquiring the mutex
2969 * ensured they're done, and we can proceed with freeing the
2973 list_del_init(&event->owner_entry);
2974 mutex_unlock(&owner->perf_event_mutex);
2975 put_task_struct(owner);
2978 perf_event_release_kernel(event);
2981 static int perf_release(struct inode *inode, struct file *file)
2983 put_event(file->private_data);
2987 u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running)
2989 struct perf_event *child;
2995 mutex_lock(&event->child_mutex);
2996 total += perf_event_read(event);
2997 *enabled += event->total_time_enabled +
2998 atomic64_read(&event->child_total_time_enabled);
2999 *running += event->total_time_running +
3000 atomic64_read(&event->child_total_time_running);
3002 list_for_each_entry(child, &event->child_list, child_list) {
3003 total += perf_event_read(child);
3004 *enabled += child->total_time_enabled;
3005 *running += child->total_time_running;
3007 mutex_unlock(&event->child_mutex);
3011 EXPORT_SYMBOL_GPL(perf_event_read_value);
3013 static int perf_event_read_group(struct perf_event *event,
3014 u64 read_format, char __user *buf)
3016 struct perf_event *leader = event->group_leader, *sub;
3017 int n = 0, size = 0, ret = -EFAULT;
3018 struct perf_event_context *ctx = leader->ctx;
3020 u64 count, enabled, running;
3022 mutex_lock(&ctx->mutex);
3023 count = perf_event_read_value(leader, &enabled, &running);
3025 values[n++] = 1 + leader->nr_siblings;
3026 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3027 values[n++] = enabled;
3028 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3029 values[n++] = running;
3030 values[n++] = count;
3031 if (read_format & PERF_FORMAT_ID)
3032 values[n++] = primary_event_id(leader);
3034 size = n * sizeof(u64);
3036 if (copy_to_user(buf, values, size))
3041 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
3044 values[n++] = perf_event_read_value(sub, &enabled, &running);
3045 if (read_format & PERF_FORMAT_ID)
3046 values[n++] = primary_event_id(sub);
3048 size = n * sizeof(u64);
3050 if (copy_to_user(buf + ret, values, size)) {
3058 mutex_unlock(&ctx->mutex);
3063 static int perf_event_read_one(struct perf_event *event,
3064 u64 read_format, char __user *buf)
3066 u64 enabled, running;
3070 values[n++] = perf_event_read_value(event, &enabled, &running);
3071 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3072 values[n++] = enabled;
3073 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3074 values[n++] = running;
3075 if (read_format & PERF_FORMAT_ID)
3076 values[n++] = primary_event_id(event);
3078 if (copy_to_user(buf, values, n * sizeof(u64)))
3081 return n * sizeof(u64);
3085 * Read the performance event - simple non blocking version for now
3088 perf_read_hw(struct perf_event *event, char __user *buf, size_t count)
3090 u64 read_format = event->attr.read_format;
3094 * Return end-of-file for a read on a event that is in
3095 * error state (i.e. because it was pinned but it couldn't be
3096 * scheduled on to the CPU at some point).
3098 if (event->state == PERF_EVENT_STATE_ERROR)
3101 if (count < event->read_size)
3104 WARN_ON_ONCE(event->ctx->parent_ctx);
3105 if (read_format & PERF_FORMAT_GROUP)
3106 ret = perf_event_read_group(event, read_format, buf);
3108 ret = perf_event_read_one(event, read_format, buf);
3114 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
3116 struct perf_event *event = file->private_data;
3118 return perf_read_hw(event, buf, count);
3121 static unsigned int perf_poll(struct file *file, poll_table *wait)
3123 struct perf_event *event = file->private_data;
3124 struct ring_buffer *rb;
3125 unsigned int events = POLL_HUP;
3128 * Race between perf_event_set_output() and perf_poll(): perf_poll()
3129 * grabs the rb reference but perf_event_set_output() overrides it.
3130 * Here is the timeline for two threads T1, T2:
3131 * t0: T1, rb = rcu_dereference(event->rb)
3132 * t1: T2, old_rb = event->rb
3133 * t2: T2, event->rb = new rb
3134 * t3: T2, ring_buffer_detach(old_rb)
3135 * t4: T1, ring_buffer_attach(rb1)
3136 * t5: T1, poll_wait(event->waitq)
3138 * To avoid this problem, we grab mmap_mutex in perf_poll()
3139 * thereby ensuring that the assignment of the new ring buffer
3140 * and the detachment of the old buffer appear atomic to perf_poll()
3142 mutex_lock(&event->mmap_mutex);
3145 rb = rcu_dereference(event->rb);
3147 ring_buffer_attach(event, rb);
3148 events = atomic_xchg(&rb->poll, 0);
3152 mutex_unlock(&event->mmap_mutex);
3154 poll_wait(file, &event->waitq, wait);
3159 static void perf_event_reset(struct perf_event *event)
3161 (void)perf_event_read(event);
3162 local64_set(&event->count, 0);
3163 perf_event_update_userpage(event);
3167 * Holding the top-level event's child_mutex means that any
3168 * descendant process that has inherited this event will block
3169 * in sync_child_event if it goes to exit, thus satisfying the
3170 * task existence requirements of perf_event_enable/disable.
3172 static void perf_event_for_each_child(struct perf_event *event,
3173 void (*func)(struct perf_event *))
3175 struct perf_event *child;
3177 WARN_ON_ONCE(event->ctx->parent_ctx);
3178 mutex_lock(&event->child_mutex);
3180 list_for_each_entry(child, &event->child_list, child_list)
3182 mutex_unlock(&event->child_mutex);
3185 static void perf_event_for_each(struct perf_event *event,
3186 void (*func)(struct perf_event *))
3188 struct perf_event_context *ctx = event->ctx;
3189 struct perf_event *sibling;
3191 WARN_ON_ONCE(ctx->parent_ctx);
3192 mutex_lock(&ctx->mutex);
3193 event = event->group_leader;
3195 perf_event_for_each_child(event, func);
3196 list_for_each_entry(sibling, &event->sibling_list, group_entry)
3197 perf_event_for_each_child(sibling, func);
3198 mutex_unlock(&ctx->mutex);
3201 static int perf_event_period(struct perf_event *event, u64 __user *arg)
3203 struct perf_event_context *ctx = event->ctx;
3207 if (!is_sampling_event(event))
3210 if (copy_from_user(&value, arg, sizeof(value)))
3216 raw_spin_lock_irq(&ctx->lock);
3217 if (event->attr.freq) {
3218 if (value > sysctl_perf_event_sample_rate) {
3223 event->attr.sample_freq = value;
3225 event->attr.sample_period = value;
3226 event->hw.sample_period = value;
3229 raw_spin_unlock_irq(&ctx->lock);
3234 static const struct file_operations perf_fops;
3236 static struct file *perf_fget_light(int fd, int *fput_needed)
3240 file = fget_light(fd, fput_needed);
3242 return ERR_PTR(-EBADF);
3244 if (file->f_op != &perf_fops) {
3245 fput_light(file, *fput_needed);
3247 return ERR_PTR(-EBADF);
3253 static int perf_event_set_output(struct perf_event *event,
3254 struct perf_event *output_event);
3255 static int perf_event_set_filter(struct perf_event *event, void __user *arg);
3257 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
3259 struct perf_event *event = file->private_data;
3260 void (*func)(struct perf_event *);
3264 case PERF_EVENT_IOC_ENABLE:
3265 func = perf_event_enable;
3267 case PERF_EVENT_IOC_DISABLE:
3268 func = perf_event_disable;
3270 case PERF_EVENT_IOC_RESET:
3271 func = perf_event_reset;
3274 case PERF_EVENT_IOC_REFRESH:
3275 return perf_event_refresh(event, arg);
3277 case PERF_EVENT_IOC_PERIOD:
3278 return perf_event_period(event, (u64 __user *)arg);
3280 case PERF_EVENT_IOC_SET_OUTPUT:
3282 struct file *output_file = NULL;
3283 struct perf_event *output_event = NULL;
3284 int fput_needed = 0;
3288 output_file = perf_fget_light(arg, &fput_needed);
3289 if (IS_ERR(output_file))
3290 return PTR_ERR(output_file);
3291 output_event = output_file->private_data;
3294 ret = perf_event_set_output(event, output_event);
3296 fput_light(output_file, fput_needed);
3301 case PERF_EVENT_IOC_SET_FILTER:
3302 return perf_event_set_filter(event, (void __user *)arg);
3308 if (flags & PERF_IOC_FLAG_GROUP)
3309 perf_event_for_each(event, func);
3311 perf_event_for_each_child(event, func);
3316 int perf_event_task_enable(void)
3318 struct perf_event *event;
3320 mutex_lock(¤t->perf_event_mutex);
3321 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
3322 perf_event_for_each_child(event, perf_event_enable);
3323 mutex_unlock(¤t->perf_event_mutex);
3328 int perf_event_task_disable(void)
3330 struct perf_event *event;
3332 mutex_lock(¤t->perf_event_mutex);
3333 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
3334 perf_event_for_each_child(event, perf_event_disable);
3335 mutex_unlock(¤t->perf_event_mutex);
3340 static int perf_event_index(struct perf_event *event)
3342 if (event->hw.state & PERF_HES_STOPPED)
3345 if (event->state != PERF_EVENT_STATE_ACTIVE)
3348 return event->pmu->event_idx(event);
3351 static void calc_timer_values(struct perf_event *event,
3358 *now = perf_clock();
3359 ctx_time = event->shadow_ctx_time + *now;
3360 *enabled = ctx_time - event->tstamp_enabled;
3361 *running = ctx_time - event->tstamp_running;
3364 void __weak arch_perf_update_userpage(struct perf_event_mmap_page *userpg, u64 now)
3369 * Callers need to ensure there can be no nesting of this function, otherwise
3370 * the seqlock logic goes bad. We can not serialize this because the arch
3371 * code calls this from NMI context.
3373 void perf_event_update_userpage(struct perf_event *event)
3375 struct perf_event_mmap_page *userpg;
3376 struct ring_buffer *rb;
3377 u64 enabled, running, now;
3381 * compute total_time_enabled, total_time_running
3382 * based on snapshot values taken when the event
3383 * was last scheduled in.
3385 * we cannot simply called update_context_time()
3386 * because of locking issue as we can be called in
3389 calc_timer_values(event, &now, &enabled, &running);
3390 rb = rcu_dereference(event->rb);
3394 userpg = rb->user_page;
3397 * Disable preemption so as to not let the corresponding user-space
3398 * spin too long if we get preempted.
3403 userpg->index = perf_event_index(event);
3404 userpg->offset = perf_event_count(event);
3406 userpg->offset -= local64_read(&event->hw.prev_count);
3408 userpg->time_enabled = enabled +
3409 atomic64_read(&event->child_total_time_enabled);
3411 userpg->time_running = running +
3412 atomic64_read(&event->child_total_time_running);
3414 arch_perf_update_userpage(userpg, now);
3423 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
3425 struct perf_event *event = vma->vm_file->private_data;
3426 struct ring_buffer *rb;
3427 int ret = VM_FAULT_SIGBUS;
3429 if (vmf->flags & FAULT_FLAG_MKWRITE) {
3430 if (vmf->pgoff == 0)
3436 rb = rcu_dereference(event->rb);
3440 if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE))
3443 vmf->page = perf_mmap_to_page(rb, vmf->pgoff);
3447 get_page(vmf->page);
3448 vmf->page->mapping = vma->vm_file->f_mapping;
3449 vmf->page->index = vmf->pgoff;
3458 static void ring_buffer_attach(struct perf_event *event,
3459 struct ring_buffer *rb)
3461 unsigned long flags;
3463 if (!list_empty(&event->rb_entry))
3466 spin_lock_irqsave(&rb->event_lock, flags);
3467 if (!list_empty(&event->rb_entry))
3470 list_add(&event->rb_entry, &rb->event_list);
3472 spin_unlock_irqrestore(&rb->event_lock, flags);
3475 static void ring_buffer_detach(struct perf_event *event,
3476 struct ring_buffer *rb)
3478 unsigned long flags;
3480 if (list_empty(&event->rb_entry))
3483 spin_lock_irqsave(&rb->event_lock, flags);
3484 list_del_init(&event->rb_entry);
3485 wake_up_all(&event->waitq);
3486 spin_unlock_irqrestore(&rb->event_lock, flags);
3489 static void ring_buffer_wakeup(struct perf_event *event)
3491 struct ring_buffer *rb;
3494 rb = rcu_dereference(event->rb);
3498 list_for_each_entry_rcu(event, &rb->event_list, rb_entry)
3499 wake_up_all(&event->waitq);
3505 static void rb_free_rcu(struct rcu_head *rcu_head)
3507 struct ring_buffer *rb;
3509 rb = container_of(rcu_head, struct ring_buffer, rcu_head);
3513 static struct ring_buffer *ring_buffer_get(struct perf_event *event)
3515 struct ring_buffer *rb;
3518 rb = rcu_dereference(event->rb);
3520 if (!atomic_inc_not_zero(&rb->refcount))
3528 static void ring_buffer_put(struct ring_buffer *rb)
3530 struct perf_event *event, *n;
3531 unsigned long flags;
3533 if (!atomic_dec_and_test(&rb->refcount))
3536 spin_lock_irqsave(&rb->event_lock, flags);
3537 list_for_each_entry_safe(event, n, &rb->event_list, rb_entry) {
3538 list_del_init(&event->rb_entry);
3539 wake_up_all(&event->waitq);
3541 spin_unlock_irqrestore(&rb->event_lock, flags);
3543 call_rcu(&rb->rcu_head, rb_free_rcu);
3546 static void perf_mmap_open(struct vm_area_struct *vma)
3548 struct perf_event *event = vma->vm_file->private_data;
3550 atomic_inc(&event->mmap_count);
3553 static void perf_mmap_close(struct vm_area_struct *vma)
3555 struct perf_event *event = vma->vm_file->private_data;
3557 if (atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex)) {
3558 unsigned long size = perf_data_size(event->rb);
3559 struct user_struct *user = event->mmap_user;
3560 struct ring_buffer *rb = event->rb;
3562 atomic_long_sub((size >> PAGE_SHIFT) + 1, &user->locked_vm);
3563 vma->vm_mm->pinned_vm -= event->mmap_locked;
3564 rcu_assign_pointer(event->rb, NULL);
3565 ring_buffer_detach(event, rb);
3566 mutex_unlock(&event->mmap_mutex);
3568 ring_buffer_put(rb);
3573 static const struct vm_operations_struct perf_mmap_vmops = {
3574 .open = perf_mmap_open,
3575 .close = perf_mmap_close,
3576 .fault = perf_mmap_fault,
3577 .page_mkwrite = perf_mmap_fault,
3580 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
3582 struct perf_event *event = file->private_data;
3583 unsigned long user_locked, user_lock_limit;
3584 struct user_struct *user = current_user();
3585 unsigned long locked, lock_limit;
3586 struct ring_buffer *rb;
3587 unsigned long vma_size;
3588 unsigned long nr_pages;
3589 long user_extra, extra;
3590 int ret = 0, flags = 0;
3593 * Don't allow mmap() of inherited per-task counters. This would
3594 * create a performance issue due to all children writing to the
3597 if (event->cpu == -1 && event->attr.inherit)
3600 if (!(vma->vm_flags & VM_SHARED))
3603 vma_size = vma->vm_end - vma->vm_start;
3604 nr_pages = (vma_size / PAGE_SIZE) - 1;
3607 * If we have rb pages ensure they're a power-of-two number, so we
3608 * can do bitmasks instead of modulo.
3610 if (nr_pages != 0 && !is_power_of_2(nr_pages))
3613 if (vma_size != PAGE_SIZE * (1 + nr_pages))
3616 if (vma->vm_pgoff != 0)
3619 WARN_ON_ONCE(event->ctx->parent_ctx);
3620 mutex_lock(&event->mmap_mutex);
3622 if (event->rb->nr_pages == nr_pages)
3623 atomic_inc(&event->rb->refcount);
3629 user_extra = nr_pages + 1;
3630 user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10);
3633 * Increase the limit linearly with more CPUs:
3635 user_lock_limit *= num_online_cpus();
3637 user_locked = atomic_long_read(&user->locked_vm) + user_extra;
3640 if (user_locked > user_lock_limit)
3641 extra = user_locked - user_lock_limit;
3643 lock_limit = rlimit(RLIMIT_MEMLOCK);
3644 lock_limit >>= PAGE_SHIFT;
3645 locked = vma->vm_mm->pinned_vm + extra;
3647 if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() &&
3648 !capable(CAP_IPC_LOCK)) {
3655 if (vma->vm_flags & VM_WRITE)
3656 flags |= RING_BUFFER_WRITABLE;
3658 rb = rb_alloc(nr_pages,
3659 event->attr.watermark ? event->attr.wakeup_watermark : 0,
3666 rcu_assign_pointer(event->rb, rb);
3668 atomic_long_add(user_extra, &user->locked_vm);
3669 event->mmap_locked = extra;
3670 event->mmap_user = get_current_user();
3671 vma->vm_mm->pinned_vm += event->mmap_locked;
3673 perf_event_update_userpage(event);
3677 atomic_inc(&event->mmap_count);
3678 mutex_unlock(&event->mmap_mutex);
3680 vma->vm_flags |= VM_RESERVED;
3681 vma->vm_ops = &perf_mmap_vmops;
3686 static int perf_fasync(int fd, struct file *filp, int on)
3688 struct inode *inode = filp->f_path.dentry->d_inode;
3689 struct perf_event *event = filp->private_data;
3692 mutex_lock(&inode->i_mutex);
3693 retval = fasync_helper(fd, filp, on, &event->fasync);
3694 mutex_unlock(&inode->i_mutex);
3702 static const struct file_operations perf_fops = {
3703 .llseek = no_llseek,
3704 .release = perf_release,
3707 .unlocked_ioctl = perf_ioctl,
3708 .compat_ioctl = perf_ioctl,
3710 .fasync = perf_fasync,
3716 * If there's data, ensure we set the poll() state and publish everything
3717 * to user-space before waking everybody up.
3720 void perf_event_wakeup(struct perf_event *event)
3722 ring_buffer_wakeup(event);
3724 if (event->pending_kill) {
3725 kill_fasync(&event->fasync, SIGIO, event->pending_kill);
3726 event->pending_kill = 0;
3730 static void perf_pending_event(struct irq_work *entry)
3732 struct perf_event *event = container_of(entry,
3733 struct perf_event, pending);
3735 if (event->pending_disable) {
3736 event->pending_disable = 0;
3737 __perf_event_disable(event);
3740 if (event->pending_wakeup) {
3741 event->pending_wakeup = 0;
3742 perf_event_wakeup(event);
3747 * We assume there is only KVM supporting the callbacks.
3748 * Later on, we might change it to a list if there is
3749 * another virtualization implementation supporting the callbacks.
3751 struct perf_guest_info_callbacks *perf_guest_cbs;
3753 int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
3755 perf_guest_cbs = cbs;
3758 EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks);
3760 int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
3762 perf_guest_cbs = NULL;
3765 EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks);
3767 static void __perf_event_header__init_id(struct perf_event_header *header,
3768 struct perf_sample_data *data,
3769 struct perf_event *event)
3771 u64 sample_type = event->attr.sample_type;
3773 data->type = sample_type;
3774 header->size += event->id_header_size;
3776 if (sample_type & PERF_SAMPLE_TID) {
3777 /* namespace issues */
3778 data->tid_entry.pid = perf_event_pid(event, current);
3779 data->tid_entry.tid = perf_event_tid(event, current);
3782 if (sample_type & PERF_SAMPLE_TIME)
3783 data->time = perf_clock();
3785 if (sample_type & PERF_SAMPLE_ID)
3786 data->id = primary_event_id(event);
3788 if (sample_type & PERF_SAMPLE_STREAM_ID)
3789 data->stream_id = event->id;
3791 if (sample_type & PERF_SAMPLE_CPU) {
3792 data->cpu_entry.cpu = raw_smp_processor_id();
3793 data->cpu_entry.reserved = 0;
3797 void perf_event_header__init_id(struct perf_event_header *header,
3798 struct perf_sample_data *data,
3799 struct perf_event *event)
3801 if (event->attr.sample_id_all)
3802 __perf_event_header__init_id(header, data, event);
3805 static void __perf_event__output_id_sample(struct perf_output_handle *handle,
3806 struct perf_sample_data *data)
3808 u64 sample_type = data->type;
3810 if (sample_type & PERF_SAMPLE_TID)
3811 perf_output_put(handle, data->tid_entry);
3813 if (sample_type & PERF_SAMPLE_TIME)
3814 perf_output_put(handle, data->time);
3816 if (sample_type & PERF_SAMPLE_ID)
3817 perf_output_put(handle, data->id);
3819 if (sample_type & PERF_SAMPLE_STREAM_ID)
3820 perf_output_put(handle, data->stream_id);
3822 if (sample_type & PERF_SAMPLE_CPU)
3823 perf_output_put(handle, data->cpu_entry);
3826 void perf_event__output_id_sample(struct perf_event *event,
3827 struct perf_output_handle *handle,
3828 struct perf_sample_data *sample)
3830 if (event->attr.sample_id_all)
3831 __perf_event__output_id_sample(handle, sample);
3834 static void perf_output_read_one(struct perf_output_handle *handle,
3835 struct perf_event *event,
3836 u64 enabled, u64 running)
3838 u64 read_format = event->attr.read_format;
3842 values[n++] = perf_event_count(event);
3843 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
3844 values[n++] = enabled +
3845 atomic64_read(&event->child_total_time_enabled);
3847 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
3848 values[n++] = running +
3849 atomic64_read(&event->child_total_time_running);
3851 if (read_format & PERF_FORMAT_ID)
3852 values[n++] = primary_event_id(event);
3854 __output_copy(handle, values, n * sizeof(u64));
3858 * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
3860 static void perf_output_read_group(struct perf_output_handle *handle,
3861 struct perf_event *event,
3862 u64 enabled, u64 running)
3864 struct perf_event *leader = event->group_leader, *sub;
3865 u64 read_format = event->attr.read_format;
3869 values[n++] = 1 + leader->nr_siblings;
3871 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3872 values[n++] = enabled;
3874 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3875 values[n++] = running;
3877 if (leader != event)
3878 leader->pmu->read(leader);
3880 values[n++] = perf_event_count(leader);
3881 if (read_format & PERF_FORMAT_ID)
3882 values[n++] = primary_event_id(leader);
3884 __output_copy(handle, values, n * sizeof(u64));
3886 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
3890 sub->pmu->read(sub);
3892 values[n++] = perf_event_count(sub);
3893 if (read_format & PERF_FORMAT_ID)
3894 values[n++] = primary_event_id(sub);
3896 __output_copy(handle, values, n * sizeof(u64));
3900 #define PERF_FORMAT_TOTAL_TIMES (PERF_FORMAT_TOTAL_TIME_ENABLED|\
3901 PERF_FORMAT_TOTAL_TIME_RUNNING)
3903 static void perf_output_read(struct perf_output_handle *handle,
3904 struct perf_event *event)
3906 u64 enabled = 0, running = 0, now;
3907 u64 read_format = event->attr.read_format;
3910 * compute total_time_enabled, total_time_running
3911 * based on snapshot values taken when the event
3912 * was last scheduled in.
3914 * we cannot simply called update_context_time()
3915 * because of locking issue as we are called in
3918 if (read_format & PERF_FORMAT_TOTAL_TIMES)
3919 calc_timer_values(event, &now, &enabled, &running);
3921 if (event->attr.read_format & PERF_FORMAT_GROUP)
3922 perf_output_read_group(handle, event, enabled, running);
3924 perf_output_read_one(handle, event, enabled, running);
3927 void perf_output_sample(struct perf_output_handle *handle,
3928 struct perf_event_header *header,
3929 struct perf_sample_data *data,
3930 struct perf_event *event)
3932 u64 sample_type = data->type;
3934 perf_output_put(handle, *header);
3936 if (sample_type & PERF_SAMPLE_IP)
3937 perf_output_put(handle, data->ip);
3939 if (sample_type & PERF_SAMPLE_TID)
3940 perf_output_put(handle, data->tid_entry);
3942 if (sample_type & PERF_SAMPLE_TIME)
3943 perf_output_put(handle, data->time);
3945 if (sample_type & PERF_SAMPLE_ADDR)
3946 perf_output_put(handle, data->addr);
3948 if (sample_type & PERF_SAMPLE_ID)
3949 perf_output_put(handle, data->id);
3951 if (sample_type & PERF_SAMPLE_STREAM_ID)
3952 perf_output_put(handle, data->stream_id);
3954 if (sample_type & PERF_SAMPLE_CPU)
3955 perf_output_put(handle, data->cpu_entry);
3957 if (sample_type & PERF_SAMPLE_PERIOD)
3958 perf_output_put(handle, data->period);
3960 if (sample_type & PERF_SAMPLE_READ)
3961 perf_output_read(handle, event);
3963 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3964 if (data->callchain) {
3967 if (data->callchain)
3968 size += data->callchain->nr;
3970 size *= sizeof(u64);
3972 __output_copy(handle, data->callchain, size);
3975 perf_output_put(handle, nr);
3979 if (sample_type & PERF_SAMPLE_RAW) {
3981 perf_output_put(handle, data->raw->size);
3982 __output_copy(handle, data->raw->data,
3989 .size = sizeof(u32),
3992 perf_output_put(handle, raw);
3996 if (!event->attr.watermark) {
3997 int wakeup_events = event->attr.wakeup_events;
3999 if (wakeup_events) {
4000 struct ring_buffer *rb = handle->rb;
4001 int events = local_inc_return(&rb->events);
4003 if (events >= wakeup_events) {
4004 local_sub(wakeup_events, &rb->events);
4005 local_inc(&rb->wakeup);
4010 if (sample_type & PERF_SAMPLE_BRANCH_STACK) {
4011 if (data->br_stack) {
4014 size = data->br_stack->nr
4015 * sizeof(struct perf_branch_entry);
4017 perf_output_put(handle, data->br_stack->nr);
4018 perf_output_copy(handle, data->br_stack->entries, size);
4021 * we always store at least the value of nr
4024 perf_output_put(handle, nr);
4029 void perf_prepare_sample(struct perf_event_header *header,
4030 struct perf_sample_data *data,
4031 struct perf_event *event,
4032 struct pt_regs *regs)
4034 u64 sample_type = event->attr.sample_type;
4036 header->type = PERF_RECORD_SAMPLE;
4037 header->size = sizeof(*header) + event->header_size;
4040 header->misc |= perf_misc_flags(regs);
4042 __perf_event_header__init_id(header, data, event);
4044 if (sample_type & PERF_SAMPLE_IP)
4045 data->ip = perf_instruction_pointer(regs);
4047 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
4050 data->callchain = perf_callchain(event, regs);
4052 if (data->callchain)
4053 size += data->callchain->nr;
4055 header->size += size * sizeof(u64);
4058 if (sample_type & PERF_SAMPLE_RAW) {
4059 int size = sizeof(u32);
4062 size += data->raw->size;
4064 size += sizeof(u32);
4066 WARN_ON_ONCE(size & (sizeof(u64)-1));
4067 header->size += size;
4070 if (sample_type & PERF_SAMPLE_BRANCH_STACK) {
4071 int size = sizeof(u64); /* nr */
4072 if (data->br_stack) {
4073 size += data->br_stack->nr
4074 * sizeof(struct perf_branch_entry);
4076 header->size += size;
4080 static void perf_event_output(struct perf_event *event,
4081 struct perf_sample_data *data,
4082 struct pt_regs *regs)
4084 struct perf_output_handle handle;
4085 struct perf_event_header header;
4087 /* protect the callchain buffers */
4090 perf_prepare_sample(&header, data, event, regs);
4092 if (perf_output_begin(&handle, event, header.size))
4095 perf_output_sample(&handle, &header, data, event);
4097 perf_output_end(&handle);
4107 struct perf_read_event {
4108 struct perf_event_header header;
4115 perf_event_read_event(struct perf_event *event,
4116 struct task_struct *task)
4118 struct perf_output_handle handle;
4119 struct perf_sample_data sample;
4120 struct perf_read_event read_event = {
4122 .type = PERF_RECORD_READ,
4124 .size = sizeof(read_event) + event->read_size,
4126 .pid = perf_event_pid(event, task),
4127 .tid = perf_event_tid(event, task),
4131 perf_event_header__init_id(&read_event.header, &sample, event);
4132 ret = perf_output_begin(&handle, event, read_event.header.size);
4136 perf_output_put(&handle, read_event);
4137 perf_output_read(&handle, event);
4138 perf_event__output_id_sample(event, &handle, &sample);
4140 perf_output_end(&handle);
4144 * task tracking -- fork/exit
4146 * enabled by: attr.comm | attr.mmap | attr.mmap_data | attr.task
4149 struct perf_task_event {
4150 struct task_struct *task;
4151 struct perf_event_context *task_ctx;
4154 struct perf_event_header header;
4164 static void perf_event_task_output(struct perf_event *event,
4165 struct perf_task_event *task_event)
4167 struct perf_output_handle handle;
4168 struct perf_sample_data sample;
4169 struct task_struct *task = task_event->task;
4170 int ret, size = task_event->event_id.header.size;
4172 perf_event_header__init_id(&task_event->event_id.header, &sample, event);
4174 ret = perf_output_begin(&handle, event,
4175 task_event->event_id.header.size);
4179 task_event->event_id.pid = perf_event_pid(event, task);
4180 task_event->event_id.ppid = perf_event_pid(event, current);
4182 task_event->event_id.tid = perf_event_tid(event, task);
4183 task_event->event_id.ptid = perf_event_tid(event, current);
4185 perf_output_put(&handle, task_event->event_id);
4187 perf_event__output_id_sample(event, &handle, &sample);
4189 perf_output_end(&handle);
4191 task_event->event_id.header.size = size;
4194 static int perf_event_task_match(struct perf_event *event)
4196 if (event->state < PERF_EVENT_STATE_INACTIVE)
4199 if (!event_filter_match(event))
4202 if (event->attr.comm || event->attr.mmap ||
4203 event->attr.mmap_data || event->attr.task)
4209 static void perf_event_task_ctx(struct perf_event_context *ctx,
4210 struct perf_task_event *task_event)
4212 struct perf_event *event;
4214 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
4215 if (perf_event_task_match(event))
4216 perf_event_task_output(event, task_event);
4220 static void perf_event_task_event(struct perf_task_event *task_event)
4222 struct perf_cpu_context *cpuctx;
4223 struct perf_event_context *ctx;
4228 list_for_each_entry_rcu(pmu, &pmus, entry) {
4229 cpuctx = get_cpu_ptr(pmu->pmu_cpu_context);
4230 if (cpuctx->active_pmu != pmu)
4232 perf_event_task_ctx(&cpuctx->ctx, task_event);
4234 ctx = task_event->task_ctx;
4236 ctxn = pmu->task_ctx_nr;
4239 ctx = rcu_dereference(current->perf_event_ctxp[ctxn]);
4242 perf_event_task_ctx(ctx, task_event);
4244 put_cpu_ptr(pmu->pmu_cpu_context);
4249 static void perf_event_task(struct task_struct *task,
4250 struct perf_event_context *task_ctx,
4253 struct perf_task_event task_event;
4255 if (!atomic_read(&nr_comm_events) &&
4256 !atomic_read(&nr_mmap_events) &&
4257 !atomic_read(&nr_task_events))
4260 task_event = (struct perf_task_event){
4262 .task_ctx = task_ctx,
4265 .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT,
4267 .size = sizeof(task_event.event_id),
4273 .time = perf_clock(),
4277 perf_event_task_event(&task_event);
4280 void perf_event_fork(struct task_struct *task)
4282 perf_event_task(task, NULL, 1);
4289 struct perf_comm_event {
4290 struct task_struct *task;
4295 struct perf_event_header header;
4302 static void perf_event_comm_output(struct perf_event *event,
4303 struct perf_comm_event *comm_event)
4305 struct perf_output_handle handle;
4306 struct perf_sample_data sample;
4307 int size = comm_event->event_id.header.size;
4310 perf_event_header__init_id(&comm_event->event_id.header, &sample, event);
4311 ret = perf_output_begin(&handle, event,
4312 comm_event->event_id.header.size);
4317 comm_event->event_id.pid = perf_event_pid(event, comm_event->task);
4318 comm_event->event_id.tid = perf_event_tid(event, comm_event->task);
4320 perf_output_put(&handle, comm_event->event_id);
4321 __output_copy(&handle, comm_event->comm,
4322 comm_event->comm_size);
4324 perf_event__output_id_sample(event, &handle, &sample);
4326 perf_output_end(&handle);
4328 comm_event->event_id.header.size = size;
4331 static int perf_event_comm_match(struct perf_event *event)
4333 if (event->state < PERF_EVENT_STATE_INACTIVE)
4336 if (!event_filter_match(event))
4339 if (event->attr.comm)
4345 static void perf_event_comm_ctx(struct perf_event_context *ctx,
4346 struct perf_comm_event *comm_event)
4348 struct perf_event *event;
4350 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
4351 if (perf_event_comm_match(event))
4352 perf_event_comm_output(event, comm_event);
4356 static void perf_event_comm_event(struct perf_comm_event *comm_event)
4358 struct perf_cpu_context *cpuctx;
4359 struct perf_event_context *ctx;
4360 char comm[TASK_COMM_LEN];
4365 memset(comm, 0, sizeof(comm));
4366 strlcpy(comm, comm_event->task->comm, sizeof(comm));
4367 size = ALIGN(strlen(comm)+1, sizeof(u64));
4369 comm_event->comm = comm;
4370 comm_event->comm_size = size;
4372 comm_event->event_id.header.size = sizeof(comm_event->event_id) + size;
4374 list_for_each_entry_rcu(pmu, &pmus, entry) {
4375 cpuctx = get_cpu_ptr(pmu->pmu_cpu_context);
4376 if (cpuctx->active_pmu != pmu)
4378 perf_event_comm_ctx(&cpuctx->ctx, comm_event);
4380 ctxn = pmu->task_ctx_nr;
4384 ctx = rcu_dereference(current->perf_event_ctxp[ctxn]);
4386 perf_event_comm_ctx(ctx, comm_event);
4388 put_cpu_ptr(pmu->pmu_cpu_context);
4393 void perf_event_comm(struct task_struct *task)
4395 struct perf_comm_event comm_event;
4396 struct perf_event_context *ctx;
4399 for_each_task_context_nr(ctxn) {
4400 ctx = task->perf_event_ctxp[ctxn];
4404 perf_event_enable_on_exec(ctx);
4407 if (!atomic_read(&nr_comm_events))
4410 comm_event = (struct perf_comm_event){
4416 .type = PERF_RECORD_COMM,
4425 perf_event_comm_event(&comm_event);
4432 struct perf_mmap_event {
4433 struct vm_area_struct *vma;
4435 const char *file_name;
4439 struct perf_event_header header;
4449 static void perf_event_mmap_output(struct perf_event *event,
4450 struct perf_mmap_event *mmap_event)
4452 struct perf_output_handle handle;
4453 struct perf_sample_data sample;
4454 int size = mmap_event->event_id.header.size;
4457 perf_event_header__init_id(&mmap_event->event_id.header, &sample, event);
4458 ret = perf_output_begin(&handle, event,
4459 mmap_event->event_id.header.size);
4463 mmap_event->event_id.pid = perf_event_pid(event, current);
4464 mmap_event->event_id.tid = perf_event_tid(event, current);
4466 perf_output_put(&handle, mmap_event->event_id);
4467 __output_copy(&handle, mmap_event->file_name,
4468 mmap_event->file_size);
4470 perf_event__output_id_sample(event, &handle, &sample);
4472 perf_output_end(&handle);
4474 mmap_event->event_id.header.size = size;
4477 static int perf_event_mmap_match(struct perf_event *event,
4478 struct perf_mmap_event *mmap_event,
4481 if (event->state < PERF_EVENT_STATE_INACTIVE)
4484 if (!event_filter_match(event))
4487 if ((!executable && event->attr.mmap_data) ||
4488 (executable && event->attr.mmap))
4494 static void perf_event_mmap_ctx(struct perf_event_context *ctx,
4495 struct perf_mmap_event *mmap_event,
4498 struct perf_event *event;
4500 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
4501 if (perf_event_mmap_match(event, mmap_event, executable))
4502 perf_event_mmap_output(event, mmap_event);
4506 static void perf_event_mmap_event(struct perf_mmap_event *mmap_event)
4508 struct perf_cpu_context *cpuctx;
4509 struct perf_event_context *ctx;
4510 struct vm_area_struct *vma = mmap_event->vma;
4511 struct file *file = vma->vm_file;
4519 memset(tmp, 0, sizeof(tmp));
4523 * d_path works from the end of the rb backwards, so we
4524 * need to add enough zero bytes after the string to handle
4525 * the 64bit alignment we do later.
4527 buf = kzalloc(PATH_MAX + sizeof(u64), GFP_KERNEL);
4529 name = strncpy(tmp, "//enomem", sizeof(tmp));
4532 name = d_path(&file->f_path, buf, PATH_MAX);
4534 name = strncpy(tmp, "//toolong", sizeof(tmp));
4538 if (arch_vma_name(mmap_event->vma)) {
4539 name = strncpy(tmp, arch_vma_name(mmap_event->vma),
4545 name = strncpy(tmp, "[vdso]", sizeof(tmp));
4547 } else if (vma->vm_start <= vma->vm_mm->start_brk &&
4548 vma->vm_end >= vma->vm_mm->brk) {
4549 name = strncpy(tmp, "[heap]", sizeof(tmp));
4551 } else if (vma->vm_start <= vma->vm_mm->start_stack &&
4552 vma->vm_end >= vma->vm_mm->start_stack) {
4553 name = strncpy(tmp, "[stack]", sizeof(tmp));
4557 name = strncpy(tmp, "//anon", sizeof(tmp));
4562 size = ALIGN(strlen(name)+1, sizeof(u64));
4564 mmap_event->file_name = name;
4565 mmap_event->file_size = size;
4567 mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size;
4570 list_for_each_entry_rcu(pmu, &pmus, entry) {
4571 cpuctx = get_cpu_ptr(pmu->pmu_cpu_context);
4572 if (cpuctx->active_pmu != pmu)
4574 perf_event_mmap_ctx(&cpuctx->ctx, mmap_event,
4575 vma->vm_flags & VM_EXEC);
4577 ctxn = pmu->task_ctx_nr;
4581 ctx = rcu_dereference(current->perf_event_ctxp[ctxn]);
4583 perf_event_mmap_ctx(ctx, mmap_event,
4584 vma->vm_flags & VM_EXEC);
4587 put_cpu_ptr(pmu->pmu_cpu_context);
4594 void perf_event_mmap(struct vm_area_struct *vma)
4596 struct perf_mmap_event mmap_event;
4598 if (!atomic_read(&nr_mmap_events))
4601 mmap_event = (struct perf_mmap_event){
4607 .type = PERF_RECORD_MMAP,
4608 .misc = PERF_RECORD_MISC_USER,
4613 .start = vma->vm_start,
4614 .len = vma->vm_end - vma->vm_start,
4615 .pgoff = (u64)vma->vm_pgoff << PAGE_SHIFT,
4619 perf_event_mmap_event(&mmap_event);
4623 * IRQ throttle logging
4626 static void perf_log_throttle(struct perf_event *event, int enable)
4628 struct perf_output_handle handle;
4629 struct perf_sample_data sample;
4633 struct perf_event_header header;
4637 } throttle_event = {
4639 .type = PERF_RECORD_THROTTLE,
4641 .size = sizeof(throttle_event),
4643 .time = perf_clock(),
4644 .id = primary_event_id(event),
4645 .stream_id = event->id,
4649 throttle_event.header.type = PERF_RECORD_UNTHROTTLE;
4651 perf_event_header__init_id(&throttle_event.header, &sample, event);
4653 ret = perf_output_begin(&handle, event,
4654 throttle_event.header.size);
4658 perf_output_put(&handle, throttle_event);
4659 perf_event__output_id_sample(event, &handle, &sample);
4660 perf_output_end(&handle);
4664 * Generic event overflow handling, sampling.
4667 static int __perf_event_overflow(struct perf_event *event,
4668 int throttle, struct perf_sample_data *data,
4669 struct pt_regs *regs)
4671 int events = atomic_read(&event->event_limit);
4672 struct hw_perf_event *hwc = &event->hw;
4677 * Non-sampling counters might still use the PMI to fold short
4678 * hardware counters, ignore those.
4680 if (unlikely(!is_sampling_event(event)))
4683 seq = __this_cpu_read(perf_throttled_seq);
4684 if (seq != hwc->interrupts_seq) {
4685 hwc->interrupts_seq = seq;
4686 hwc->interrupts = 1;
4689 if (unlikely(throttle
4690 && hwc->interrupts >= max_samples_per_tick)) {
4691 __this_cpu_inc(perf_throttled_count);
4692 hwc->interrupts = MAX_INTERRUPTS;
4693 perf_log_throttle(event, 0);
4698 if (event->attr.freq) {
4699 u64 now = perf_clock();
4700 s64 delta = now - hwc->freq_time_stamp;
4702 hwc->freq_time_stamp = now;
4704 if (delta > 0 && delta < 2*TICK_NSEC)
4705 perf_adjust_period(event, delta, hwc->last_period, true);
4709 * XXX event_limit might not quite work as expected on inherited
4713 event->pending_kill = POLL_IN;
4714 if (events && atomic_dec_and_test(&event->event_limit)) {
4716 event->pending_kill = POLL_HUP;
4717 event->pending_disable = 1;
4718 irq_work_queue(&event->pending);
4721 if (event->overflow_handler)
4722 event->overflow_handler(event, data, regs);
4724 perf_event_output(event, data, regs);
4726 if (event->fasync && event->pending_kill) {
4727 event->pending_wakeup = 1;
4728 irq_work_queue(&event->pending);
4734 int perf_event_overflow(struct perf_event *event,
4735 struct perf_sample_data *data,
4736 struct pt_regs *regs)
4738 return __perf_event_overflow(event, 1, data, regs);
4742 * Generic software event infrastructure
4745 struct swevent_htable {
4746 struct swevent_hlist *swevent_hlist;
4747 struct mutex hlist_mutex;
4750 /* Recursion avoidance in each contexts */
4751 int recursion[PERF_NR_CONTEXTS];
4754 static DEFINE_PER_CPU(struct swevent_htable, swevent_htable);
4757 * We directly increment event->count and keep a second value in
4758 * event->hw.period_left to count intervals. This period event
4759 * is kept in the range [-sample_period, 0] so that we can use the
4763 static u64 perf_swevent_set_period(struct perf_event *event)
4765 struct hw_perf_event *hwc = &event->hw;
4766 u64 period = hwc->last_period;
4770 hwc->last_period = hwc->sample_period;
4773 old = val = local64_read(&hwc->period_left);
4777 nr = div64_u64(period + val, period);
4778 offset = nr * period;
4780 if (local64_cmpxchg(&hwc->period_left, old, val) != old)
4786 static void perf_swevent_overflow(struct perf_event *event, u64 overflow,
4787 struct perf_sample_data *data,
4788 struct pt_regs *regs)
4790 struct hw_perf_event *hwc = &event->hw;
4794 overflow = perf_swevent_set_period(event);
4796 if (hwc->interrupts == MAX_INTERRUPTS)
4799 for (; overflow; overflow--) {
4800 if (__perf_event_overflow(event, throttle,
4803 * We inhibit the overflow from happening when
4804 * hwc->interrupts == MAX_INTERRUPTS.
4812 static void perf_swevent_event(struct perf_event *event, u64 nr,
4813 struct perf_sample_data *data,
4814 struct pt_regs *regs)
4816 struct hw_perf_event *hwc = &event->hw;
4818 local64_add(nr, &event->count);
4823 if (!is_sampling_event(event))
4826 if ((event->attr.sample_type & PERF_SAMPLE_PERIOD) && !event->attr.freq) {
4828 return perf_swevent_overflow(event, 1, data, regs);
4830 data->period = event->hw.last_period;
4832 if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq)
4833 return perf_swevent_overflow(event, 1, data, regs);
4835 if (local64_add_negative(nr, &hwc->period_left))
4838 perf_swevent_overflow(event, 0, data, regs);
4841 static int perf_exclude_event(struct perf_event *event,
4842 struct pt_regs *regs)
4844 if (event->hw.state & PERF_HES_STOPPED)
4848 if (event->attr.exclude_user && user_mode(regs))
4851 if (event->attr.exclude_kernel && !user_mode(regs))
4858 static int perf_swevent_match(struct perf_event *event,
4859 enum perf_type_id type,
4861 struct perf_sample_data *data,
4862 struct pt_regs *regs)
4864 if (event->attr.type != type)
4867 if (event->attr.config != event_id)
4870 if (perf_exclude_event(event, regs))
4876 static inline u64 swevent_hash(u64 type, u32 event_id)
4878 u64 val = event_id | (type << 32);
4880 return hash_64(val, SWEVENT_HLIST_BITS);
4883 static inline struct hlist_head *
4884 __find_swevent_head(struct swevent_hlist *hlist, u64 type, u32 event_id)
4886 u64 hash = swevent_hash(type, event_id);
4888 return &hlist->heads[hash];
4891 /* For the read side: events when they trigger */
4892 static inline struct hlist_head *
4893 find_swevent_head_rcu(struct swevent_htable *swhash, u64 type, u32 event_id)
4895 struct swevent_hlist *hlist;
4897 hlist = rcu_dereference(swhash->swevent_hlist);
4901 return __find_swevent_head(hlist, type, event_id);
4904 /* For the event head insertion and removal in the hlist */
4905 static inline struct hlist_head *
4906 find_swevent_head(struct swevent_htable *swhash, struct perf_event *event)
4908 struct swevent_hlist *hlist;
4909 u32 event_id = event->attr.config;
4910 u64 type = event->attr.type;
4913 * Event scheduling is always serialized against hlist allocation
4914 * and release. Which makes the protected version suitable here.
4915 * The context lock guarantees that.
4917 hlist = rcu_dereference_protected(swhash->swevent_hlist,
4918 lockdep_is_held(&event->ctx->lock));
4922 return __find_swevent_head(hlist, type, event_id);
4925 static void do_perf_sw_event(enum perf_type_id type, u32 event_id,
4927 struct perf_sample_data *data,
4928 struct pt_regs *regs)
4930 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
4931 struct perf_event *event;
4932 struct hlist_node *node;
4933 struct hlist_head *head;
4936 head = find_swevent_head_rcu(swhash, type, event_id);
4940 hlist_for_each_entry_rcu(event, node, head, hlist_entry) {
4941 if (perf_swevent_match(event, type, event_id, data, regs))
4942 perf_swevent_event(event, nr, data, regs);
4948 int perf_swevent_get_recursion_context(void)
4950 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
4952 return get_recursion_context(swhash->recursion);
4954 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context);
4956 inline void perf_swevent_put_recursion_context(int rctx)
4958 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
4960 put_recursion_context(swhash->recursion, rctx);
4963 void __perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)
4965 struct perf_sample_data data;
4968 preempt_disable_notrace();
4969 rctx = perf_swevent_get_recursion_context();
4973 perf_sample_data_init(&data, addr, 0);
4975 do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, &data, regs);
4977 perf_swevent_put_recursion_context(rctx);
4978 preempt_enable_notrace();
4981 static void perf_swevent_read(struct perf_event *event)
4985 static int perf_swevent_add(struct perf_event *event, int flags)
4987 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
4988 struct hw_perf_event *hwc = &event->hw;
4989 struct hlist_head *head;
4991 if (is_sampling_event(event)) {
4992 hwc->last_period = hwc->sample_period;
4993 perf_swevent_set_period(event);
4996 hwc->state = !(flags & PERF_EF_START);
4998 head = find_swevent_head(swhash, event);
4999 if (WARN_ON_ONCE(!head))
5002 hlist_add_head_rcu(&event->hlist_entry, head);
5007 static void perf_swevent_del(struct perf_event *event, int flags)
5009 hlist_del_rcu(&event->hlist_entry);
5012 static void perf_swevent_start(struct perf_event *event, int flags)
5014 event->hw.state = 0;
5017 static void perf_swevent_stop(struct perf_event *event, int flags)
5019 event->hw.state = PERF_HES_STOPPED;
5022 /* Deref the hlist from the update side */
5023 static inline struct swevent_hlist *
5024 swevent_hlist_deref(struct swevent_htable *swhash)
5026 return rcu_dereference_protected(swhash->swevent_hlist,
5027 lockdep_is_held(&swhash->hlist_mutex));
5030 static void swevent_hlist_release(struct swevent_htable *swhash)
5032 struct swevent_hlist *hlist = swevent_hlist_deref(swhash);
5037 rcu_assign_pointer(swhash->swevent_hlist, NULL);
5038 kfree_rcu(hlist, rcu_head);
5041 static void swevent_hlist_put_cpu(struct perf_event *event, int cpu)
5043 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
5045 mutex_lock(&swhash->hlist_mutex);
5047 if (!--swhash->hlist_refcount)
5048 swevent_hlist_release(swhash);
5050 mutex_unlock(&swhash->hlist_mutex);
5053 static void swevent_hlist_put(struct perf_event *event)
5057 if (event->cpu != -1) {
5058 swevent_hlist_put_cpu(event, event->cpu);
5062 for_each_possible_cpu(cpu)
5063 swevent_hlist_put_cpu(event, cpu);
5066 static int swevent_hlist_get_cpu(struct perf_event *event, int cpu)
5068 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
5071 mutex_lock(&swhash->hlist_mutex);
5073 if (!swevent_hlist_deref(swhash) && cpu_online(cpu)) {
5074 struct swevent_hlist *hlist;
5076 hlist = kzalloc(sizeof(*hlist), GFP_KERNEL);
5081 rcu_assign_pointer(swhash->swevent_hlist, hlist);
5083 swhash->hlist_refcount++;
5085 mutex_unlock(&swhash->hlist_mutex);
5090 static int swevent_hlist_get(struct perf_event *event)
5093 int cpu, failed_cpu;
5095 if (event->cpu != -1)
5096 return swevent_hlist_get_cpu(event, event->cpu);
5099 for_each_possible_cpu(cpu) {
5100 err = swevent_hlist_get_cpu(event, cpu);
5110 for_each_possible_cpu(cpu) {
5111 if (cpu == failed_cpu)
5113 swevent_hlist_put_cpu(event, cpu);
5120 struct static_key perf_swevent_enabled[PERF_COUNT_SW_MAX];
5122 static void sw_perf_event_destroy(struct perf_event *event)
5124 u64 event_id = event->attr.config;
5126 WARN_ON(event->parent);
5128 static_key_slow_dec(&perf_swevent_enabled[event_id]);
5129 swevent_hlist_put(event);
5132 static int perf_swevent_init(struct perf_event *event)
5134 int event_id = event->attr.config;
5136 if (event->attr.type != PERF_TYPE_SOFTWARE)
5140 * no branch sampling for software events
5142 if (has_branch_stack(event))
5146 case PERF_COUNT_SW_CPU_CLOCK:
5147 case PERF_COUNT_SW_TASK_CLOCK:
5154 if (event_id >= PERF_COUNT_SW_MAX)
5157 if (!event->parent) {
5160 err = swevent_hlist_get(event);
5164 static_key_slow_inc(&perf_swevent_enabled[event_id]);
5165 event->destroy = sw_perf_event_destroy;
5171 static int perf_swevent_event_idx(struct perf_event *event)
5176 static struct pmu perf_swevent = {
5177 .task_ctx_nr = perf_sw_context,
5179 .event_init = perf_swevent_init,
5180 .add = perf_swevent_add,
5181 .del = perf_swevent_del,
5182 .start = perf_swevent_start,
5183 .stop = perf_swevent_stop,
5184 .read = perf_swevent_read,
5186 .event_idx = perf_swevent_event_idx,
5189 #ifdef CONFIG_EVENT_TRACING
5191 static int perf_tp_filter_match(struct perf_event *event,
5192 struct perf_sample_data *data)
5194 void *record = data->raw->data;
5196 if (likely(!event->filter) || filter_match_preds(event->filter, record))
5201 static int perf_tp_event_match(struct perf_event *event,
5202 struct perf_sample_data *data,
5203 struct pt_regs *regs)
5205 if (event->hw.state & PERF_HES_STOPPED)
5208 * All tracepoints are from kernel-space.
5210 if (event->attr.exclude_kernel)
5213 if (!perf_tp_filter_match(event, data))
5219 void perf_tp_event(u64 addr, u64 count, void *record, int entry_size,
5220 struct pt_regs *regs, struct hlist_head *head, int rctx,
5221 struct task_struct *task)
5223 struct perf_sample_data data;
5224 struct perf_event *event;
5225 struct hlist_node *node;
5227 struct perf_raw_record raw = {
5232 perf_sample_data_init(&data, addr, 0);
5235 hlist_for_each_entry_rcu(event, node, head, hlist_entry) {
5236 if (perf_tp_event_match(event, &data, regs))
5237 perf_swevent_event(event, count, &data, regs);
5241 * If we got specified a target task, also iterate its context and
5242 * deliver this event there too.
5244 if (task && task != current) {
5245 struct perf_event_context *ctx;
5246 struct trace_entry *entry = record;
5249 ctx = rcu_dereference(task->perf_event_ctxp[perf_sw_context]);
5253 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
5254 if (event->attr.type != PERF_TYPE_TRACEPOINT)
5256 if (event->attr.config != entry->type)
5258 if (perf_tp_event_match(event, &data, regs))
5259 perf_swevent_event(event, count, &data, regs);
5265 perf_swevent_put_recursion_context(rctx);
5267 EXPORT_SYMBOL_GPL(perf_tp_event);
5269 static void tp_perf_event_destroy(struct perf_event *event)
5271 perf_trace_destroy(event);
5274 static int perf_tp_event_init(struct perf_event *event)
5278 if (event->attr.type != PERF_TYPE_TRACEPOINT)
5282 * no branch sampling for tracepoint events
5284 if (has_branch_stack(event))
5287 err = perf_trace_init(event);
5291 event->destroy = tp_perf_event_destroy;
5296 static struct pmu perf_tracepoint = {
5297 .task_ctx_nr = perf_sw_context,
5299 .event_init = perf_tp_event_init,
5300 .add = perf_trace_add,
5301 .del = perf_trace_del,
5302 .start = perf_swevent_start,
5303 .stop = perf_swevent_stop,
5304 .read = perf_swevent_read,
5306 .event_idx = perf_swevent_event_idx,
5309 static inline void perf_tp_register(void)
5311 perf_pmu_register(&perf_tracepoint, "tracepoint", PERF_TYPE_TRACEPOINT);
5314 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
5319 if (event->attr.type != PERF_TYPE_TRACEPOINT)
5322 filter_str = strndup_user(arg, PAGE_SIZE);
5323 if (IS_ERR(filter_str))
5324 return PTR_ERR(filter_str);
5326 ret = ftrace_profile_set_filter(event, event->attr.config, filter_str);
5332 static void perf_event_free_filter(struct perf_event *event)
5334 ftrace_profile_free_filter(event);
5339 static inline void perf_tp_register(void)
5343 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
5348 static void perf_event_free_filter(struct perf_event *event)
5352 #endif /* CONFIG_EVENT_TRACING */
5354 #ifdef CONFIG_HAVE_HW_BREAKPOINT
5355 void perf_bp_event(struct perf_event *bp, void *data)
5357 struct perf_sample_data sample;
5358 struct pt_regs *regs = data;
5360 perf_sample_data_init(&sample, bp->attr.bp_addr, 0);
5362 if (!bp->hw.state && !perf_exclude_event(bp, regs))
5363 perf_swevent_event(bp, 1, &sample, regs);
5368 * hrtimer based swevent callback
5371 static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer)
5373 enum hrtimer_restart ret = HRTIMER_RESTART;
5374 struct perf_sample_data data;
5375 struct pt_regs *regs;
5376 struct perf_event *event;
5379 event = container_of(hrtimer, struct perf_event, hw.hrtimer);
5381 if (event->state != PERF_EVENT_STATE_ACTIVE)
5382 return HRTIMER_NORESTART;
5384 event->pmu->read(event);
5386 perf_sample_data_init(&data, 0, event->hw.last_period);
5387 regs = get_irq_regs();
5389 if (regs && !perf_exclude_event(event, regs)) {
5390 if (!(event->attr.exclude_idle && is_idle_task(current)))
5391 if (__perf_event_overflow(event, 1, &data, regs))
5392 ret = HRTIMER_NORESTART;
5395 period = max_t(u64, 10000, event->hw.sample_period);
5396 hrtimer_forward_now(hrtimer, ns_to_ktime(period));
5401 static void perf_swevent_start_hrtimer(struct perf_event *event)
5403 struct hw_perf_event *hwc = &event->hw;
5406 if (!is_sampling_event(event))
5409 period = local64_read(&hwc->period_left);
5414 local64_set(&hwc->period_left, 0);
5416 period = max_t(u64, 10000, hwc->sample_period);
5418 __hrtimer_start_range_ns(&hwc->hrtimer,
5419 ns_to_ktime(period), 0,
5420 HRTIMER_MODE_REL_PINNED, 0);
5423 static void perf_swevent_cancel_hrtimer(struct perf_event *event)
5425 struct hw_perf_event *hwc = &event->hw;
5427 if (is_sampling_event(event)) {
5428 ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer);
5429 local64_set(&hwc->period_left, ktime_to_ns(remaining));
5431 hrtimer_cancel(&hwc->hrtimer);
5435 static void perf_swevent_init_hrtimer(struct perf_event *event)
5437 struct hw_perf_event *hwc = &event->hw;
5439 if (!is_sampling_event(event))
5442 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
5443 hwc->hrtimer.function = perf_swevent_hrtimer;
5446 * Since hrtimers have a fixed rate, we can do a static freq->period
5447 * mapping and avoid the whole period adjust feedback stuff.
5449 if (event->attr.freq) {
5450 long freq = event->attr.sample_freq;
5452 event->attr.sample_period = NSEC_PER_SEC / freq;
5453 hwc->sample_period = event->attr.sample_period;
5454 local64_set(&hwc->period_left, hwc->sample_period);
5455 event->attr.freq = 0;
5460 * Software event: cpu wall time clock
5463 static void cpu_clock_event_update(struct perf_event *event)
5468 now = local_clock();
5469 prev = local64_xchg(&event->hw.prev_count, now);
5470 local64_add(now - prev, &event->count);
5473 static void cpu_clock_event_start(struct perf_event *event, int flags)
5475 local64_set(&event->hw.prev_count, local_clock());
5476 perf_swevent_start_hrtimer(event);
5479 static void cpu_clock_event_stop(struct perf_event *event, int flags)
5481 perf_swevent_cancel_hrtimer(event);
5482 cpu_clock_event_update(event);
5485 static int cpu_clock_event_add(struct perf_event *event, int flags)
5487 if (flags & PERF_EF_START)
5488 cpu_clock_event_start(event, flags);
5493 static void cpu_clock_event_del(struct perf_event *event, int flags)
5495 cpu_clock_event_stop(event, flags);
5498 static void cpu_clock_event_read(struct perf_event *event)
5500 cpu_clock_event_update(event);
5503 static int cpu_clock_event_init(struct perf_event *event)
5505 if (event->attr.type != PERF_TYPE_SOFTWARE)
5508 if (event->attr.config != PERF_COUNT_SW_CPU_CLOCK)
5512 * no branch sampling for software events
5514 if (has_branch_stack(event))
5517 perf_swevent_init_hrtimer(event);
5522 static struct pmu perf_cpu_clock = {
5523 .task_ctx_nr = perf_sw_context,
5525 .event_init = cpu_clock_event_init,
5526 .add = cpu_clock_event_add,
5527 .del = cpu_clock_event_del,
5528 .start = cpu_clock_event_start,
5529 .stop = cpu_clock_event_stop,
5530 .read = cpu_clock_event_read,
5532 .event_idx = perf_swevent_event_idx,
5536 * Software event: task time clock
5539 static void task_clock_event_update(struct perf_event *event, u64 now)
5544 prev = local64_xchg(&event->hw.prev_count, now);
5546 local64_add(delta, &event->count);
5549 static void task_clock_event_start(struct perf_event *event, int flags)
5551 local64_set(&event->hw.prev_count, event->ctx->time);
5552 perf_swevent_start_hrtimer(event);
5555 static void task_clock_event_stop(struct perf_event *event, int flags)
5557 perf_swevent_cancel_hrtimer(event);
5558 task_clock_event_update(event, event->ctx->time);
5561 static int task_clock_event_add(struct perf_event *event, int flags)
5563 if (flags & PERF_EF_START)
5564 task_clock_event_start(event, flags);
5569 static void task_clock_event_del(struct perf_event *event, int flags)
5571 task_clock_event_stop(event, PERF_EF_UPDATE);
5574 static void task_clock_event_read(struct perf_event *event)
5576 u64 now = perf_clock();
5577 u64 delta = now - event->ctx->timestamp;
5578 u64 time = event->ctx->time + delta;
5580 task_clock_event_update(event, time);
5583 static int task_clock_event_init(struct perf_event *event)
5585 if (event->attr.type != PERF_TYPE_SOFTWARE)
5588 if (event->attr.config != PERF_COUNT_SW_TASK_CLOCK)
5592 * no branch sampling for software events
5594 if (has_branch_stack(event))
5597 perf_swevent_init_hrtimer(event);
5602 static struct pmu perf_task_clock = {
5603 .task_ctx_nr = perf_sw_context,
5605 .event_init = task_clock_event_init,
5606 .add = task_clock_event_add,
5607 .del = task_clock_event_del,
5608 .start = task_clock_event_start,
5609 .stop = task_clock_event_stop,
5610 .read = task_clock_event_read,
5612 .event_idx = perf_swevent_event_idx,
5615 static void perf_pmu_nop_void(struct pmu *pmu)
5619 static int perf_pmu_nop_int(struct pmu *pmu)
5624 static void perf_pmu_start_txn(struct pmu *pmu)
5626 perf_pmu_disable(pmu);
5629 static int perf_pmu_commit_txn(struct pmu *pmu)
5631 perf_pmu_enable(pmu);
5635 static void perf_pmu_cancel_txn(struct pmu *pmu)
5637 perf_pmu_enable(pmu);
5640 static int perf_event_idx_default(struct perf_event *event)
5642 return event->hw.idx + 1;
5646 * Ensures all contexts with the same task_ctx_nr have the same
5647 * pmu_cpu_context too.
5649 static void *find_pmu_context(int ctxn)
5656 list_for_each_entry(pmu, &pmus, entry) {
5657 if (pmu->task_ctx_nr == ctxn)
5658 return pmu->pmu_cpu_context;
5664 static void update_pmu_context(struct pmu *pmu, struct pmu *old_pmu)
5668 for_each_possible_cpu(cpu) {
5669 struct perf_cpu_context *cpuctx;
5671 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
5673 if (cpuctx->active_pmu == old_pmu)
5674 cpuctx->active_pmu = pmu;
5678 static void free_pmu_context(struct pmu *pmu)
5682 mutex_lock(&pmus_lock);
5684 * Like a real lame refcount.
5686 list_for_each_entry(i, &pmus, entry) {
5687 if (i->pmu_cpu_context == pmu->pmu_cpu_context) {
5688 update_pmu_context(i, pmu);
5693 free_percpu(pmu->pmu_cpu_context);
5695 mutex_unlock(&pmus_lock);
5697 static struct idr pmu_idr;
5700 type_show(struct device *dev, struct device_attribute *attr, char *page)
5702 struct pmu *pmu = dev_get_drvdata(dev);
5704 return snprintf(page, PAGE_SIZE-1, "%d\n", pmu->type);
5707 static struct device_attribute pmu_dev_attrs[] = {
5712 static int pmu_bus_running;
5713 static struct bus_type pmu_bus = {
5714 .name = "event_source",
5715 .dev_attrs = pmu_dev_attrs,
5718 static void pmu_dev_release(struct device *dev)
5723 static int pmu_dev_alloc(struct pmu *pmu)
5727 pmu->dev = kzalloc(sizeof(struct device), GFP_KERNEL);
5731 pmu->dev->groups = pmu->attr_groups;
5732 device_initialize(pmu->dev);
5733 ret = dev_set_name(pmu->dev, "%s", pmu->name);
5737 dev_set_drvdata(pmu->dev, pmu);
5738 pmu->dev->bus = &pmu_bus;
5739 pmu->dev->release = pmu_dev_release;
5740 ret = device_add(pmu->dev);
5748 put_device(pmu->dev);
5752 static struct lock_class_key cpuctx_mutex;
5753 static struct lock_class_key cpuctx_lock;
5755 int perf_pmu_register(struct pmu *pmu, char *name, int type)
5759 mutex_lock(&pmus_lock);
5761 pmu->pmu_disable_count = alloc_percpu(int);
5762 if (!pmu->pmu_disable_count)
5771 int err = idr_pre_get(&pmu_idr, GFP_KERNEL);
5775 err = idr_get_new_above(&pmu_idr, pmu, PERF_TYPE_MAX, &type);
5783 if (pmu_bus_running) {
5784 ret = pmu_dev_alloc(pmu);
5790 pmu->pmu_cpu_context = find_pmu_context(pmu->task_ctx_nr);
5791 if (pmu->pmu_cpu_context)
5792 goto got_cpu_context;
5794 pmu->pmu_cpu_context = alloc_percpu(struct perf_cpu_context);
5795 if (!pmu->pmu_cpu_context)
5798 for_each_possible_cpu(cpu) {
5799 struct perf_cpu_context *cpuctx;
5801 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
5802 __perf_event_init_context(&cpuctx->ctx);
5803 lockdep_set_class(&cpuctx->ctx.mutex, &cpuctx_mutex);
5804 lockdep_set_class(&cpuctx->ctx.lock, &cpuctx_lock);
5805 cpuctx->ctx.type = cpu_context;
5806 cpuctx->ctx.pmu = pmu;
5807 cpuctx->jiffies_interval = 1;
5808 INIT_LIST_HEAD(&cpuctx->rotation_list);
5809 cpuctx->active_pmu = pmu;
5813 if (!pmu->start_txn) {
5814 if (pmu->pmu_enable) {
5816 * If we have pmu_enable/pmu_disable calls, install
5817 * transaction stubs that use that to try and batch
5818 * hardware accesses.
5820 pmu->start_txn = perf_pmu_start_txn;
5821 pmu->commit_txn = perf_pmu_commit_txn;
5822 pmu->cancel_txn = perf_pmu_cancel_txn;
5824 pmu->start_txn = perf_pmu_nop_void;
5825 pmu->commit_txn = perf_pmu_nop_int;
5826 pmu->cancel_txn = perf_pmu_nop_void;
5830 if (!pmu->pmu_enable) {
5831 pmu->pmu_enable = perf_pmu_nop_void;
5832 pmu->pmu_disable = perf_pmu_nop_void;
5835 if (!pmu->event_idx)
5836 pmu->event_idx = perf_event_idx_default;
5838 list_add_rcu(&pmu->entry, &pmus);
5841 mutex_unlock(&pmus_lock);
5846 device_del(pmu->dev);
5847 put_device(pmu->dev);
5850 if (pmu->type >= PERF_TYPE_MAX)
5851 idr_remove(&pmu_idr, pmu->type);
5854 free_percpu(pmu->pmu_disable_count);
5858 void perf_pmu_unregister(struct pmu *pmu)
5860 mutex_lock(&pmus_lock);
5861 list_del_rcu(&pmu->entry);
5862 mutex_unlock(&pmus_lock);
5865 * We dereference the pmu list under both SRCU and regular RCU, so
5866 * synchronize against both of those.
5868 synchronize_srcu(&pmus_srcu);
5871 free_percpu(pmu->pmu_disable_count);
5872 if (pmu->type >= PERF_TYPE_MAX)
5873 idr_remove(&pmu_idr, pmu->type);
5874 device_del(pmu->dev);
5875 put_device(pmu->dev);
5876 free_pmu_context(pmu);
5879 struct pmu *perf_init_event(struct perf_event *event)
5881 struct pmu *pmu = NULL;
5885 idx = srcu_read_lock(&pmus_srcu);
5888 pmu = idr_find(&pmu_idr, event->attr.type);
5892 ret = pmu->event_init(event);
5898 list_for_each_entry_rcu(pmu, &pmus, entry) {
5900 ret = pmu->event_init(event);
5904 if (ret != -ENOENT) {
5909 pmu = ERR_PTR(-ENOENT);
5911 srcu_read_unlock(&pmus_srcu, idx);
5917 * Allocate and initialize a event structure
5919 static struct perf_event *
5920 perf_event_alloc(struct perf_event_attr *attr, int cpu,
5921 struct task_struct *task,
5922 struct perf_event *group_leader,
5923 struct perf_event *parent_event,
5924 perf_overflow_handler_t overflow_handler,
5928 struct perf_event *event;
5929 struct hw_perf_event *hwc;
5932 if ((unsigned)cpu >= nr_cpu_ids) {
5933 if (!task || cpu != -1)
5934 return ERR_PTR(-EINVAL);
5937 event = kzalloc(sizeof(*event), GFP_KERNEL);
5939 return ERR_PTR(-ENOMEM);
5942 * Single events are their own group leaders, with an
5943 * empty sibling list:
5946 group_leader = event;
5948 mutex_init(&event->child_mutex);
5949 INIT_LIST_HEAD(&event->child_list);
5951 INIT_LIST_HEAD(&event->group_entry);
5952 INIT_LIST_HEAD(&event->event_entry);
5953 INIT_LIST_HEAD(&event->sibling_list);
5954 INIT_LIST_HEAD(&event->rb_entry);
5956 init_waitqueue_head(&event->waitq);
5957 init_irq_work(&event->pending, perf_pending_event);
5959 mutex_init(&event->mmap_mutex);
5961 atomic_long_set(&event->refcount, 1);
5963 event->attr = *attr;
5964 event->group_leader = group_leader;
5968 event->parent = parent_event;
5970 event->ns = get_pid_ns(current->nsproxy->pid_ns);
5971 event->id = atomic64_inc_return(&perf_event_id);
5973 event->state = PERF_EVENT_STATE_INACTIVE;
5976 event->attach_state = PERF_ATTACH_TASK;
5977 #ifdef CONFIG_HAVE_HW_BREAKPOINT
5979 * hw_breakpoint is a bit difficult here..
5981 if (attr->type == PERF_TYPE_BREAKPOINT)
5982 event->hw.bp_target = task;
5986 if (!overflow_handler && parent_event) {
5987 overflow_handler = parent_event->overflow_handler;
5988 context = parent_event->overflow_handler_context;
5991 event->overflow_handler = overflow_handler;
5992 event->overflow_handler_context = context;
5995 event->state = PERF_EVENT_STATE_OFF;
6000 hwc->sample_period = attr->sample_period;
6001 if (attr->freq && attr->sample_freq)
6002 hwc->sample_period = 1;
6003 hwc->last_period = hwc->sample_period;
6005 local64_set(&hwc->period_left, hwc->sample_period);
6008 * we currently do not support PERF_FORMAT_GROUP on inherited events
6010 if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP))
6013 pmu = perf_init_event(event);
6019 else if (IS_ERR(pmu))
6024 put_pid_ns(event->ns);
6026 return ERR_PTR(err);
6029 if (!event->parent) {
6030 if (event->attach_state & PERF_ATTACH_TASK)
6031 static_key_slow_inc(&perf_sched_events.key);
6032 if (event->attr.mmap || event->attr.mmap_data)
6033 atomic_inc(&nr_mmap_events);
6034 if (event->attr.comm)
6035 atomic_inc(&nr_comm_events);
6036 if (event->attr.task)
6037 atomic_inc(&nr_task_events);
6038 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) {
6039 err = get_callchain_buffers();
6042 return ERR_PTR(err);
6045 if (has_branch_stack(event)) {
6046 static_key_slow_inc(&perf_sched_events.key);
6047 if (!(event->attach_state & PERF_ATTACH_TASK))
6048 atomic_inc(&per_cpu(perf_branch_stack_events,
6056 static int perf_copy_attr(struct perf_event_attr __user *uattr,
6057 struct perf_event_attr *attr)
6062 if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0))
6066 * zero the full structure, so that a short copy will be nice.
6068 memset(attr, 0, sizeof(*attr));
6070 ret = get_user(size, &uattr->size);
6074 if (size > PAGE_SIZE) /* silly large */
6077 if (!size) /* abi compat */
6078 size = PERF_ATTR_SIZE_VER0;
6080 if (size < PERF_ATTR_SIZE_VER0)
6084 * If we're handed a bigger struct than we know of,
6085 * ensure all the unknown bits are 0 - i.e. new
6086 * user-space does not rely on any kernel feature
6087 * extensions we dont know about yet.
6089 if (size > sizeof(*attr)) {
6090 unsigned char __user *addr;
6091 unsigned char __user *end;
6094 addr = (void __user *)uattr + sizeof(*attr);
6095 end = (void __user *)uattr + size;
6097 for (; addr < end; addr++) {
6098 ret = get_user(val, addr);
6104 size = sizeof(*attr);
6107 ret = copy_from_user(attr, uattr, size);
6111 if (attr->__reserved_1)
6114 if (attr->sample_type & ~(PERF_SAMPLE_MAX-1))
6117 if (attr->read_format & ~(PERF_FORMAT_MAX-1))
6120 if (attr->sample_type & PERF_SAMPLE_BRANCH_STACK) {
6121 u64 mask = attr->branch_sample_type;
6123 /* only using defined bits */
6124 if (mask & ~(PERF_SAMPLE_BRANCH_MAX-1))
6127 /* at least one branch bit must be set */
6128 if (!(mask & ~PERF_SAMPLE_BRANCH_PLM_ALL))
6131 /* kernel level capture: check permissions */
6132 if ((mask & PERF_SAMPLE_BRANCH_PERM_PLM)
6133 && perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
6136 /* propagate priv level, when not set for branch */
6137 if (!(mask & PERF_SAMPLE_BRANCH_PLM_ALL)) {
6139 /* exclude_kernel checked on syscall entry */
6140 if (!attr->exclude_kernel)
6141 mask |= PERF_SAMPLE_BRANCH_KERNEL;
6143 if (!attr->exclude_user)
6144 mask |= PERF_SAMPLE_BRANCH_USER;
6146 if (!attr->exclude_hv)
6147 mask |= PERF_SAMPLE_BRANCH_HV;
6149 * adjust user setting (for HW filter setup)
6151 attr->branch_sample_type = mask;
6158 put_user(sizeof(*attr), &uattr->size);
6164 perf_event_set_output(struct perf_event *event, struct perf_event *output_event)
6166 struct ring_buffer *rb = NULL, *old_rb = NULL;
6172 /* don't allow circular references */
6173 if (event == output_event)
6177 * Don't allow cross-cpu buffers
6179 if (output_event->cpu != event->cpu)
6183 * If its not a per-cpu rb, it must be the same task.
6185 if (output_event->cpu == -1 && output_event->ctx != event->ctx)
6189 mutex_lock(&event->mmap_mutex);
6190 /* Can't redirect output if we've got an active mmap() */
6191 if (atomic_read(&event->mmap_count))
6195 /* get the rb we want to redirect to */
6196 rb = ring_buffer_get(output_event);
6202 rcu_assign_pointer(event->rb, rb);
6204 ring_buffer_detach(event, old_rb);
6207 mutex_unlock(&event->mmap_mutex);
6210 ring_buffer_put(old_rb);
6216 * sys_perf_event_open - open a performance event, associate it to a task/cpu
6218 * @attr_uptr: event_id type attributes for monitoring/sampling
6221 * @group_fd: group leader event fd
6223 SYSCALL_DEFINE5(perf_event_open,
6224 struct perf_event_attr __user *, attr_uptr,
6225 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
6227 struct perf_event *group_leader = NULL, *output_event = NULL;
6228 struct perf_event *event, *sibling;
6229 struct perf_event_attr attr;
6230 struct perf_event_context *ctx;
6231 struct file *event_file = NULL;
6232 struct file *group_file = NULL;
6233 struct task_struct *task = NULL;
6237 int fput_needed = 0;
6240 /* for future expandability... */
6241 if (flags & ~PERF_FLAG_ALL)
6244 err = perf_copy_attr(attr_uptr, &attr);
6248 if (!attr.exclude_kernel) {
6249 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
6254 if (attr.sample_freq > sysctl_perf_event_sample_rate)
6259 * In cgroup mode, the pid argument is used to pass the fd
6260 * opened to the cgroup directory in cgroupfs. The cpu argument
6261 * designates the cpu on which to monitor threads from that
6264 if ((flags & PERF_FLAG_PID_CGROUP) && (pid == -1 || cpu == -1))
6267 event_fd = get_unused_fd_flags(O_RDWR);
6271 if (group_fd != -1) {
6272 group_file = perf_fget_light(group_fd, &fput_needed);
6273 if (IS_ERR(group_file)) {
6274 err = PTR_ERR(group_file);
6277 group_leader = group_file->private_data;
6278 if (flags & PERF_FLAG_FD_OUTPUT)
6279 output_event = group_leader;
6280 if (flags & PERF_FLAG_FD_NO_GROUP)
6281 group_leader = NULL;
6284 if (pid != -1 && !(flags & PERF_FLAG_PID_CGROUP)) {
6285 task = find_lively_task_by_vpid(pid);
6287 err = PTR_ERR(task);
6294 event = perf_event_alloc(&attr, cpu, task, group_leader, NULL,
6296 if (IS_ERR(event)) {
6297 err = PTR_ERR(event);
6301 if (flags & PERF_FLAG_PID_CGROUP) {
6302 err = perf_cgroup_connect(pid, event, &attr, group_leader);
6307 * - that has cgroup constraint on event->cpu
6308 * - that may need work on context switch
6310 atomic_inc(&per_cpu(perf_cgroup_events, event->cpu));
6311 static_key_slow_inc(&perf_sched_events.key);
6315 * Special case software events and allow them to be part of
6316 * any hardware group.
6321 (is_software_event(event) != is_software_event(group_leader))) {
6322 if (is_software_event(event)) {
6324 * If event and group_leader are not both a software
6325 * event, and event is, then group leader is not.
6327 * Allow the addition of software events to !software
6328 * groups, this is safe because software events never
6331 pmu = group_leader->pmu;
6332 } else if (is_software_event(group_leader) &&
6333 (group_leader->group_flags & PERF_GROUP_SOFTWARE)) {
6335 * In case the group is a pure software group, and we
6336 * try to add a hardware event, move the whole group to
6337 * the hardware context.
6344 * Get the target context (task or percpu):
6346 ctx = find_get_context(pmu, task, event->cpu);
6353 put_task_struct(task);
6358 * Look up the group leader (we will attach this event to it):
6364 * Do not allow a recursive hierarchy (this new sibling
6365 * becoming part of another group-sibling):
6367 if (group_leader->group_leader != group_leader)
6370 * Do not allow to attach to a group in a different
6371 * task or CPU context:
6374 if (group_leader->ctx->type != ctx->type)
6377 if (group_leader->ctx != ctx)
6382 * Only a group leader can be exclusive or pinned
6384 if (attr.exclusive || attr.pinned)
6389 err = perf_event_set_output(event, output_event);
6394 event_file = anon_inode_getfile("[perf_event]", &perf_fops, event, O_RDWR);
6395 if (IS_ERR(event_file)) {
6396 err = PTR_ERR(event_file);
6401 struct perf_event_context *gctx = group_leader->ctx;
6403 mutex_lock(&gctx->mutex);
6404 perf_remove_from_context(group_leader);
6405 list_for_each_entry(sibling, &group_leader->sibling_list,
6407 perf_remove_from_context(sibling);
6410 mutex_unlock(&gctx->mutex);
6414 WARN_ON_ONCE(ctx->parent_ctx);
6415 mutex_lock(&ctx->mutex);
6419 perf_install_in_context(ctx, group_leader, event->cpu);
6421 list_for_each_entry(sibling, &group_leader->sibling_list,
6423 perf_install_in_context(ctx, sibling, event->cpu);
6428 perf_install_in_context(ctx, event, event->cpu);
6430 perf_unpin_context(ctx);
6431 mutex_unlock(&ctx->mutex);
6435 event->owner = current;
6437 mutex_lock(¤t->perf_event_mutex);
6438 list_add_tail(&event->owner_entry, ¤t->perf_event_list);
6439 mutex_unlock(¤t->perf_event_mutex);
6442 * Precalculate sample_data sizes
6444 perf_event__header_size(event);
6445 perf_event__id_header_size(event);
6448 * Drop the reference on the group_event after placing the
6449 * new event on the sibling_list. This ensures destruction
6450 * of the group leader will find the pointer to itself in
6451 * perf_group_detach().
6453 fput_light(group_file, fput_needed);
6454 fd_install(event_fd, event_file);
6458 perf_unpin_context(ctx);
6465 put_task_struct(task);
6467 fput_light(group_file, fput_needed);
6469 put_unused_fd(event_fd);
6474 * perf_event_create_kernel_counter
6476 * @attr: attributes of the counter to create
6477 * @cpu: cpu in which the counter is bound
6478 * @task: task to profile (NULL for percpu)
6481 perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu,
6482 struct task_struct *task,
6483 perf_overflow_handler_t overflow_handler,
6486 struct perf_event_context *ctx;
6487 struct perf_event *event;
6491 * Get the target context (task or percpu):
6494 event = perf_event_alloc(attr, cpu, task, NULL, NULL,
6495 overflow_handler, context);
6496 if (IS_ERR(event)) {
6497 err = PTR_ERR(event);
6501 ctx = find_get_context(event->pmu, task, cpu);
6507 WARN_ON_ONCE(ctx->parent_ctx);
6508 mutex_lock(&ctx->mutex);
6509 perf_install_in_context(ctx, event, cpu);
6511 perf_unpin_context(ctx);
6512 mutex_unlock(&ctx->mutex);
6519 return ERR_PTR(err);
6521 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter);
6523 void perf_pmu_migrate_context(struct pmu *pmu, int src_cpu, int dst_cpu)
6525 struct perf_event_context *src_ctx;
6526 struct perf_event_context *dst_ctx;
6527 struct perf_event *event, *tmp;
6530 src_ctx = &per_cpu_ptr(pmu->pmu_cpu_context, src_cpu)->ctx;
6531 dst_ctx = &per_cpu_ptr(pmu->pmu_cpu_context, dst_cpu)->ctx;
6533 mutex_lock(&src_ctx->mutex);
6534 list_for_each_entry_safe(event, tmp, &src_ctx->event_list,
6536 perf_remove_from_context(event);
6538 list_add(&event->event_entry, &events);
6540 mutex_unlock(&src_ctx->mutex);
6544 mutex_lock(&dst_ctx->mutex);
6545 list_for_each_entry_safe(event, tmp, &events, event_entry) {
6546 list_del(&event->event_entry);
6547 if (event->state >= PERF_EVENT_STATE_OFF)
6548 event->state = PERF_EVENT_STATE_INACTIVE;
6549 perf_install_in_context(dst_ctx, event, dst_cpu);
6552 mutex_unlock(&dst_ctx->mutex);
6554 EXPORT_SYMBOL_GPL(perf_pmu_migrate_context);
6556 static void sync_child_event(struct perf_event *child_event,
6557 struct task_struct *child)
6559 struct perf_event *parent_event = child_event->parent;
6562 if (child_event->attr.inherit_stat)
6563 perf_event_read_event(child_event, child);
6565 child_val = perf_event_count(child_event);
6568 * Add back the child's count to the parent's count:
6570 atomic64_add(child_val, &parent_event->child_count);
6571 atomic64_add(child_event->total_time_enabled,
6572 &parent_event->child_total_time_enabled);
6573 atomic64_add(child_event->total_time_running,
6574 &parent_event->child_total_time_running);
6577 * Remove this event from the parent's list
6579 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
6580 mutex_lock(&parent_event->child_mutex);
6581 list_del_init(&child_event->child_list);
6582 mutex_unlock(&parent_event->child_mutex);
6585 * Release the parent event, if this was the last
6588 put_event(parent_event);
6592 __perf_event_exit_task(struct perf_event *child_event,
6593 struct perf_event_context *child_ctx,
6594 struct task_struct *child)
6596 if (child_event->parent) {
6597 raw_spin_lock_irq(&child_ctx->lock);
6598 perf_group_detach(child_event);
6599 raw_spin_unlock_irq(&child_ctx->lock);
6602 perf_remove_from_context(child_event);
6605 * It can happen that the parent exits first, and has events
6606 * that are still around due to the child reference. These
6607 * events need to be zapped.
6609 if (child_event->parent) {
6610 sync_child_event(child_event, child);
6611 free_event(child_event);
6615 static void perf_event_exit_task_context(struct task_struct *child, int ctxn)
6617 struct perf_event *child_event, *tmp;
6618 struct perf_event_context *child_ctx;
6619 unsigned long flags;
6621 if (likely(!child->perf_event_ctxp[ctxn])) {
6622 perf_event_task(child, NULL, 0);
6626 local_irq_save(flags);
6628 * We can't reschedule here because interrupts are disabled,
6629 * and either child is current or it is a task that can't be
6630 * scheduled, so we are now safe from rescheduling changing
6633 child_ctx = rcu_dereference_raw(child->perf_event_ctxp[ctxn]);
6636 * Take the context lock here so that if find_get_context is
6637 * reading child->perf_event_ctxp, we wait until it has
6638 * incremented the context's refcount before we do put_ctx below.
6640 raw_spin_lock(&child_ctx->lock);
6641 task_ctx_sched_out(child_ctx);
6642 child->perf_event_ctxp[ctxn] = NULL;
6644 * If this context is a clone; unclone it so it can't get
6645 * swapped to another process while we're removing all
6646 * the events from it.
6648 unclone_ctx(child_ctx);
6649 update_context_time(child_ctx);
6650 raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
6653 * Report the task dead after unscheduling the events so that we
6654 * won't get any samples after PERF_RECORD_EXIT. We can however still
6655 * get a few PERF_RECORD_READ events.
6657 perf_event_task(child, child_ctx, 0);
6660 * We can recurse on the same lock type through:
6662 * __perf_event_exit_task()
6663 * sync_child_event()
6665 * mutex_lock(&ctx->mutex)
6667 * But since its the parent context it won't be the same instance.
6669 mutex_lock(&child_ctx->mutex);
6672 list_for_each_entry_safe(child_event, tmp, &child_ctx->pinned_groups,
6674 __perf_event_exit_task(child_event, child_ctx, child);
6676 list_for_each_entry_safe(child_event, tmp, &child_ctx->flexible_groups,
6678 __perf_event_exit_task(child_event, child_ctx, child);
6681 * If the last event was a group event, it will have appended all
6682 * its siblings to the list, but we obtained 'tmp' before that which
6683 * will still point to the list head terminating the iteration.
6685 if (!list_empty(&child_ctx->pinned_groups) ||
6686 !list_empty(&child_ctx->flexible_groups))
6689 mutex_unlock(&child_ctx->mutex);
6695 * When a child task exits, feed back event values to parent events.
6697 void perf_event_exit_task(struct task_struct *child)
6699 struct perf_event *event, *tmp;
6702 mutex_lock(&child->perf_event_mutex);
6703 list_for_each_entry_safe(event, tmp, &child->perf_event_list,
6705 list_del_init(&event->owner_entry);
6708 * Ensure the list deletion is visible before we clear
6709 * the owner, closes a race against perf_release() where
6710 * we need to serialize on the owner->perf_event_mutex.
6713 event->owner = NULL;
6715 mutex_unlock(&child->perf_event_mutex);
6717 for_each_task_context_nr(ctxn)
6718 perf_event_exit_task_context(child, ctxn);
6721 static void perf_free_event(struct perf_event *event,
6722 struct perf_event_context *ctx)
6724 struct perf_event *parent = event->parent;
6726 if (WARN_ON_ONCE(!parent))
6729 mutex_lock(&parent->child_mutex);
6730 list_del_init(&event->child_list);
6731 mutex_unlock(&parent->child_mutex);
6735 perf_group_detach(event);
6736 list_del_event(event, ctx);
6741 * free an unexposed, unused context as created by inheritance by
6742 * perf_event_init_task below, used by fork() in case of fail.
6744 void perf_event_free_task(struct task_struct *task)
6746 struct perf_event_context *ctx;
6747 struct perf_event *event, *tmp;
6750 for_each_task_context_nr(ctxn) {
6751 ctx = task->perf_event_ctxp[ctxn];
6755 mutex_lock(&ctx->mutex);
6757 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups,
6759 perf_free_event(event, ctx);
6761 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups,
6763 perf_free_event(event, ctx);
6765 if (!list_empty(&ctx->pinned_groups) ||
6766 !list_empty(&ctx->flexible_groups))
6769 mutex_unlock(&ctx->mutex);
6775 void perf_event_delayed_put(struct task_struct *task)
6779 for_each_task_context_nr(ctxn)
6780 WARN_ON_ONCE(task->perf_event_ctxp[ctxn]);
6784 * inherit a event from parent task to child task:
6786 static struct perf_event *
6787 inherit_event(struct perf_event *parent_event,
6788 struct task_struct *parent,
6789 struct perf_event_context *parent_ctx,
6790 struct task_struct *child,
6791 struct perf_event *group_leader,
6792 struct perf_event_context *child_ctx)
6794 struct perf_event *child_event;
6795 unsigned long flags;
6798 * Instead of creating recursive hierarchies of events,
6799 * we link inherited events back to the original parent,
6800 * which has a filp for sure, which we use as the reference
6803 if (parent_event->parent)
6804 parent_event = parent_event->parent;
6806 child_event = perf_event_alloc(&parent_event->attr,
6809 group_leader, parent_event,
6811 if (IS_ERR(child_event))
6814 if (!atomic_long_inc_not_zero(&parent_event->refcount)) {
6815 free_event(child_event);
6822 * Make the child state follow the state of the parent event,
6823 * not its attr.disabled bit. We hold the parent's mutex,
6824 * so we won't race with perf_event_{en, dis}able_family.
6826 if (parent_event->state >= PERF_EVENT_STATE_INACTIVE)
6827 child_event->state = PERF_EVENT_STATE_INACTIVE;
6829 child_event->state = PERF_EVENT_STATE_OFF;
6831 if (parent_event->attr.freq) {
6832 u64 sample_period = parent_event->hw.sample_period;
6833 struct hw_perf_event *hwc = &child_event->hw;
6835 hwc->sample_period = sample_period;
6836 hwc->last_period = sample_period;
6838 local64_set(&hwc->period_left, sample_period);
6841 child_event->ctx = child_ctx;
6842 child_event->overflow_handler = parent_event->overflow_handler;
6843 child_event->overflow_handler_context
6844 = parent_event->overflow_handler_context;
6847 * Precalculate sample_data sizes
6849 perf_event__header_size(child_event);
6850 perf_event__id_header_size(child_event);
6853 * Link it up in the child's context:
6855 raw_spin_lock_irqsave(&child_ctx->lock, flags);
6856 add_event_to_ctx(child_event, child_ctx);
6857 raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
6860 * Link this into the parent event's child list
6862 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
6863 mutex_lock(&parent_event->child_mutex);
6864 list_add_tail(&child_event->child_list, &parent_event->child_list);
6865 mutex_unlock(&parent_event->child_mutex);
6870 static int inherit_group(struct perf_event *parent_event,
6871 struct task_struct *parent,
6872 struct perf_event_context *parent_ctx,
6873 struct task_struct *child,
6874 struct perf_event_context *child_ctx)
6876 struct perf_event *leader;
6877 struct perf_event *sub;
6878 struct perf_event *child_ctr;
6880 leader = inherit_event(parent_event, parent, parent_ctx,
6881 child, NULL, child_ctx);
6883 return PTR_ERR(leader);
6884 list_for_each_entry(sub, &parent_event->sibling_list, group_entry) {
6885 child_ctr = inherit_event(sub, parent, parent_ctx,
6886 child, leader, child_ctx);
6887 if (IS_ERR(child_ctr))
6888 return PTR_ERR(child_ctr);
6894 inherit_task_group(struct perf_event *event, struct task_struct *parent,
6895 struct perf_event_context *parent_ctx,
6896 struct task_struct *child, int ctxn,
6900 struct perf_event_context *child_ctx;
6902 if (!event->attr.inherit) {
6907 child_ctx = child->perf_event_ctxp[ctxn];
6910 * This is executed from the parent task context, so
6911 * inherit events that have been marked for cloning.
6912 * First allocate and initialize a context for the
6916 child_ctx = alloc_perf_context(event->pmu, child);
6920 child->perf_event_ctxp[ctxn] = child_ctx;
6923 ret = inherit_group(event, parent, parent_ctx,
6933 * Initialize the perf_event context in task_struct
6935 int perf_event_init_context(struct task_struct *child, int ctxn)
6937 struct perf_event_context *child_ctx, *parent_ctx;
6938 struct perf_event_context *cloned_ctx;
6939 struct perf_event *event;
6940 struct task_struct *parent = current;
6941 int inherited_all = 1;
6942 unsigned long flags;
6945 if (likely(!parent->perf_event_ctxp[ctxn]))
6949 * If the parent's context is a clone, pin it so it won't get
6952 parent_ctx = perf_pin_task_context(parent, ctxn);
6955 * No need to check if parent_ctx != NULL here; since we saw
6956 * it non-NULL earlier, the only reason for it to become NULL
6957 * is if we exit, and since we're currently in the middle of
6958 * a fork we can't be exiting at the same time.
6962 * Lock the parent list. No need to lock the child - not PID
6963 * hashed yet and not running, so nobody can access it.
6965 mutex_lock(&parent_ctx->mutex);
6968 * We dont have to disable NMIs - we are only looking at
6969 * the list, not manipulating it:
6971 list_for_each_entry(event, &parent_ctx->pinned_groups, group_entry) {
6972 ret = inherit_task_group(event, parent, parent_ctx,
6973 child, ctxn, &inherited_all);
6979 * We can't hold ctx->lock when iterating the ->flexible_group list due
6980 * to allocations, but we need to prevent rotation because
6981 * rotate_ctx() will change the list from interrupt context.
6983 raw_spin_lock_irqsave(&parent_ctx->lock, flags);
6984 parent_ctx->rotate_disable = 1;
6985 raw_spin_unlock_irqrestore(&parent_ctx->lock, flags);
6987 list_for_each_entry(event, &parent_ctx->flexible_groups, group_entry) {
6988 ret = inherit_task_group(event, parent, parent_ctx,
6989 child, ctxn, &inherited_all);
6994 raw_spin_lock_irqsave(&parent_ctx->lock, flags);
6995 parent_ctx->rotate_disable = 0;
6997 child_ctx = child->perf_event_ctxp[ctxn];
6999 if (child_ctx && inherited_all) {
7001 * Mark the child context as a clone of the parent
7002 * context, or of whatever the parent is a clone of.
7004 * Note that if the parent is a clone, the holding of
7005 * parent_ctx->lock avoids it from being uncloned.
7007 cloned_ctx = parent_ctx->parent_ctx;
7009 child_ctx->parent_ctx = cloned_ctx;
7010 child_ctx->parent_gen = parent_ctx->parent_gen;
7012 child_ctx->parent_ctx = parent_ctx;
7013 child_ctx->parent_gen = parent_ctx->generation;
7015 get_ctx(child_ctx->parent_ctx);
7018 raw_spin_unlock_irqrestore(&parent_ctx->lock, flags);
7019 mutex_unlock(&parent_ctx->mutex);
7021 perf_unpin_context(parent_ctx);
7022 put_ctx(parent_ctx);
7028 * Initialize the perf_event context in task_struct
7030 int perf_event_init_task(struct task_struct *child)
7034 memset(child->perf_event_ctxp, 0, sizeof(child->perf_event_ctxp));
7035 mutex_init(&child->perf_event_mutex);
7036 INIT_LIST_HEAD(&child->perf_event_list);
7038 for_each_task_context_nr(ctxn) {
7039 ret = perf_event_init_context(child, ctxn);
7047 static void __init perf_event_init_all_cpus(void)
7049 struct swevent_htable *swhash;
7052 for_each_possible_cpu(cpu) {
7053 swhash = &per_cpu(swevent_htable, cpu);
7054 mutex_init(&swhash->hlist_mutex);
7055 INIT_LIST_HEAD(&per_cpu(rotation_list, cpu));
7059 static void __cpuinit perf_event_init_cpu(int cpu)
7061 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
7063 mutex_lock(&swhash->hlist_mutex);
7064 if (swhash->hlist_refcount > 0) {
7065 struct swevent_hlist *hlist;
7067 hlist = kzalloc_node(sizeof(*hlist), GFP_KERNEL, cpu_to_node(cpu));
7069 rcu_assign_pointer(swhash->swevent_hlist, hlist);
7071 mutex_unlock(&swhash->hlist_mutex);
7074 #if defined CONFIG_HOTPLUG_CPU || defined CONFIG_KEXEC
7075 static void perf_pmu_rotate_stop(struct pmu *pmu)
7077 struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
7079 WARN_ON(!irqs_disabled());
7081 list_del_init(&cpuctx->rotation_list);
7084 static void __perf_event_exit_context(void *__info)
7086 struct perf_event_context *ctx = __info;
7087 struct perf_event *event, *tmp;
7089 perf_pmu_rotate_stop(ctx->pmu);
7091 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry)
7092 __perf_remove_from_context(event);
7093 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups, group_entry)
7094 __perf_remove_from_context(event);
7097 static void perf_event_exit_cpu_context(int cpu)
7099 struct perf_event_context *ctx;
7103 idx = srcu_read_lock(&pmus_srcu);
7104 list_for_each_entry_rcu(pmu, &pmus, entry) {
7105 ctx = &per_cpu_ptr(pmu->pmu_cpu_context, cpu)->ctx;
7107 mutex_lock(&ctx->mutex);
7108 smp_call_function_single(cpu, __perf_event_exit_context, ctx, 1);
7109 mutex_unlock(&ctx->mutex);
7111 srcu_read_unlock(&pmus_srcu, idx);
7114 static void perf_event_exit_cpu(int cpu)
7116 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
7118 mutex_lock(&swhash->hlist_mutex);
7119 swevent_hlist_release(swhash);
7120 mutex_unlock(&swhash->hlist_mutex);
7122 perf_event_exit_cpu_context(cpu);
7125 static inline void perf_event_exit_cpu(int cpu) { }
7129 perf_reboot(struct notifier_block *notifier, unsigned long val, void *v)
7133 for_each_online_cpu(cpu)
7134 perf_event_exit_cpu(cpu);
7140 * Run the perf reboot notifier at the very last possible moment so that
7141 * the generic watchdog code runs as long as possible.
7143 static struct notifier_block perf_reboot_notifier = {
7144 .notifier_call = perf_reboot,
7145 .priority = INT_MIN,
7148 static int __cpuinit
7149 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
7151 unsigned int cpu = (long)hcpu;
7153 switch (action & ~CPU_TASKS_FROZEN) {
7155 case CPU_UP_PREPARE:
7156 case CPU_DOWN_FAILED:
7157 perf_event_init_cpu(cpu);
7160 case CPU_UP_CANCELED:
7161 case CPU_DOWN_PREPARE:
7162 perf_event_exit_cpu(cpu);
7172 void __init perf_event_init(void)
7178 perf_event_init_all_cpus();
7179 init_srcu_struct(&pmus_srcu);
7180 perf_pmu_register(&perf_swevent, "software", PERF_TYPE_SOFTWARE);
7181 perf_pmu_register(&perf_cpu_clock, NULL, -1);
7182 perf_pmu_register(&perf_task_clock, NULL, -1);
7184 perf_cpu_notifier(perf_cpu_notify);
7185 register_reboot_notifier(&perf_reboot_notifier);
7187 ret = init_hw_breakpoint();
7188 WARN(ret, "hw_breakpoint initialization failed with: %d", ret);
7190 /* do not patch jump label more than once per second */
7191 jump_label_rate_limit(&perf_sched_events, HZ);
7194 * Build time assertion that we keep the data_head at the intended
7195 * location. IOW, validation we got the __reserved[] size right.
7197 BUILD_BUG_ON((offsetof(struct perf_event_mmap_page, data_head))
7201 static int __init perf_event_sysfs_init(void)
7206 mutex_lock(&pmus_lock);
7208 ret = bus_register(&pmu_bus);
7212 list_for_each_entry(pmu, &pmus, entry) {
7213 if (!pmu->name || pmu->type < 0)
7216 ret = pmu_dev_alloc(pmu);
7217 WARN(ret, "Failed to register pmu: %s, reason %d\n", pmu->name, ret);
7219 pmu_bus_running = 1;
7223 mutex_unlock(&pmus_lock);
7227 device_initcall(perf_event_sysfs_init);
7229 #ifdef CONFIG_CGROUP_PERF
7230 static struct cgroup_subsys_state *perf_cgroup_create(struct cgroup *cont)
7232 struct perf_cgroup *jc;
7234 jc = kzalloc(sizeof(*jc), GFP_KERNEL);
7236 return ERR_PTR(-ENOMEM);
7238 jc->info = alloc_percpu(struct perf_cgroup_info);
7241 return ERR_PTR(-ENOMEM);
7247 static void perf_cgroup_destroy(struct cgroup *cont)
7249 struct perf_cgroup *jc;
7250 jc = container_of(cgroup_subsys_state(cont, perf_subsys_id),
7251 struct perf_cgroup, css);
7252 free_percpu(jc->info);
7256 static int __perf_cgroup_move(void *info)
7258 struct task_struct *task = info;
7259 perf_cgroup_switch(task, PERF_CGROUP_SWOUT | PERF_CGROUP_SWIN);
7263 static void perf_cgroup_attach(struct cgroup *cgrp, struct cgroup_taskset *tset)
7265 struct task_struct *task;
7267 cgroup_taskset_for_each(task, cgrp, tset)
7268 task_function_call(task, __perf_cgroup_move, task);
7271 static void perf_cgroup_exit(struct cgroup *cgrp, struct cgroup *old_cgrp,
7272 struct task_struct *task)
7275 * cgroup_exit() is called in the copy_process() failure path.
7276 * Ignore this case since the task hasn't ran yet, this avoids
7277 * trying to poke a half freed task state from generic code.
7279 if (!(task->flags & PF_EXITING))
7282 task_function_call(task, __perf_cgroup_move, task);
7285 struct cgroup_subsys perf_subsys = {
7286 .name = "perf_event",
7287 .subsys_id = perf_subsys_id,
7288 .create = perf_cgroup_create,
7289 .destroy = perf_cgroup_destroy,
7290 .exit = perf_cgroup_exit,
7291 .attach = perf_cgroup_attach,
7293 #endif /* CONFIG_CGROUP_PERF */