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/tick.h>
22 #include <linux/sysfs.h>
23 #include <linux/dcache.h>
24 #include <linux/percpu.h>
25 #include <linux/ptrace.h>
26 #include <linux/reboot.h>
27 #include <linux/vmstat.h>
28 #include <linux/device.h>
29 #include <linux/export.h>
30 #include <linux/vmalloc.h>
31 #include <linux/hardirq.h>
32 #include <linux/rculist.h>
33 #include <linux/uaccess.h>
34 #include <linux/syscalls.h>
35 #include <linux/anon_inodes.h>
36 #include <linux/kernel_stat.h>
37 #include <linux/perf_event.h>
38 #include <linux/ftrace_event.h>
39 #include <linux/hw_breakpoint.h>
40 #include <linux/mm_types.h>
41 #include <linux/cgroup.h>
45 #include <asm/irq_regs.h>
47 struct remote_function_call {
48 struct task_struct *p;
49 int (*func)(void *info);
54 static void remote_function(void *data)
56 struct remote_function_call *tfc = data;
57 struct task_struct *p = tfc->p;
61 if (task_cpu(p) != smp_processor_id() || !task_curr(p))
65 tfc->ret = tfc->func(tfc->info);
69 * task_function_call - call a function on the cpu on which a task runs
70 * @p: the task to evaluate
71 * @func: the function to be called
72 * @info: the function call argument
74 * Calls the function @func when the task is currently running. This might
75 * be on the current CPU, which just calls the function directly
77 * returns: @func return value, or
78 * -ESRCH - when the process isn't running
79 * -EAGAIN - when the process moved away
82 task_function_call(struct task_struct *p, int (*func) (void *info), void *info)
84 struct remote_function_call data = {
88 .ret = -ESRCH, /* No such (running) process */
92 smp_call_function_single(task_cpu(p), remote_function, &data, 1);
98 * cpu_function_call - call a function on the cpu
99 * @func: the function to be called
100 * @info: the function call argument
102 * Calls the function @func on the remote cpu.
104 * returns: @func return value or -ENXIO when the cpu is offline
106 static int cpu_function_call(int cpu, int (*func) (void *info), void *info)
108 struct remote_function_call data = {
112 .ret = -ENXIO, /* No such CPU */
115 smp_call_function_single(cpu, remote_function, &data, 1);
120 #define PERF_FLAG_ALL (PERF_FLAG_FD_NO_GROUP |\
121 PERF_FLAG_FD_OUTPUT |\
122 PERF_FLAG_PID_CGROUP)
125 * branch priv levels that need permission checks
127 #define PERF_SAMPLE_BRANCH_PERM_PLM \
128 (PERF_SAMPLE_BRANCH_KERNEL |\
129 PERF_SAMPLE_BRANCH_HV)
132 EVENT_FLEXIBLE = 0x1,
134 EVENT_ALL = EVENT_FLEXIBLE | EVENT_PINNED,
138 * perf_sched_events : >0 events exist
139 * perf_cgroup_events: >0 per-cpu cgroup events exist on this cpu
141 struct static_key_deferred perf_sched_events __read_mostly;
142 static DEFINE_PER_CPU(atomic_t, perf_cgroup_events);
143 static DEFINE_PER_CPU(atomic_t, perf_branch_stack_events);
145 static atomic_t nr_mmap_events __read_mostly;
146 static atomic_t nr_comm_events __read_mostly;
147 static atomic_t nr_task_events __read_mostly;
149 static LIST_HEAD(pmus);
150 static DEFINE_MUTEX(pmus_lock);
151 static struct srcu_struct pmus_srcu;
154 * perf event paranoia level:
155 * -1 - not paranoid at all
156 * 0 - disallow raw tracepoint access for unpriv
157 * 1 - disallow cpu events for unpriv
158 * 2 - disallow kernel profiling for unpriv
160 int sysctl_perf_event_paranoid __read_mostly = 1;
162 /* Minimum for 512 kiB + 1 user control page */
163 int sysctl_perf_event_mlock __read_mostly = 512 + (PAGE_SIZE / 1024); /* 'free' kiB per user */
166 * max perf event sample rate
168 #define DEFAULT_MAX_SAMPLE_RATE 100000
169 int sysctl_perf_event_sample_rate __read_mostly = DEFAULT_MAX_SAMPLE_RATE;
170 static int max_samples_per_tick __read_mostly =
171 DIV_ROUND_UP(DEFAULT_MAX_SAMPLE_RATE, HZ);
173 int perf_proc_update_handler(struct ctl_table *table, int write,
174 void __user *buffer, size_t *lenp,
177 int ret = proc_dointvec(table, write, buffer, lenp, ppos);
182 max_samples_per_tick = DIV_ROUND_UP(sysctl_perf_event_sample_rate, HZ);
187 static atomic64_t perf_event_id;
189 static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
190 enum event_type_t event_type);
192 static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
193 enum event_type_t event_type,
194 struct task_struct *task);
196 static void update_context_time(struct perf_event_context *ctx);
197 static u64 perf_event_time(struct perf_event *event);
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 * perf_cgroup_info keeps track of time_enabled for a cgroup.
237 * This is a per-cpu dynamically allocated data structure.
239 struct perf_cgroup_info {
245 struct cgroup_subsys_state css;
246 struct perf_cgroup_info __percpu *info;
250 * Must ensure cgroup is pinned (css_get) before calling
251 * this function. In other words, we cannot call this function
252 * if there is no cgroup event for the current CPU context.
254 static inline struct perf_cgroup *
255 perf_cgroup_from_task(struct task_struct *task)
257 return container_of(task_subsys_state(task, perf_subsys_id),
258 struct perf_cgroup, css);
262 perf_cgroup_match(struct perf_event *event)
264 struct perf_event_context *ctx = event->ctx;
265 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
267 /* @event doesn't care about cgroup */
271 /* wants specific cgroup scope but @cpuctx isn't associated with any */
276 * Cgroup scoping is recursive. An event enabled for a cgroup is
277 * also enabled for all its descendant cgroups. If @cpuctx's
278 * cgroup is a descendant of @event's (the test covers identity
279 * case), it's a match.
281 return cgroup_is_descendant(cpuctx->cgrp->css.cgroup,
282 event->cgrp->css.cgroup);
285 static inline bool perf_tryget_cgroup(struct perf_event *event)
287 return css_tryget(&event->cgrp->css);
290 static inline void perf_put_cgroup(struct perf_event *event)
292 css_put(&event->cgrp->css);
295 static inline void perf_detach_cgroup(struct perf_event *event)
297 perf_put_cgroup(event);
301 static inline int is_cgroup_event(struct perf_event *event)
303 return event->cgrp != NULL;
306 static inline u64 perf_cgroup_event_time(struct perf_event *event)
308 struct perf_cgroup_info *t;
310 t = per_cpu_ptr(event->cgrp->info, event->cpu);
314 static inline void __update_cgrp_time(struct perf_cgroup *cgrp)
316 struct perf_cgroup_info *info;
321 info = this_cpu_ptr(cgrp->info);
323 info->time += now - info->timestamp;
324 info->timestamp = now;
327 static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx)
329 struct perf_cgroup *cgrp_out = cpuctx->cgrp;
331 __update_cgrp_time(cgrp_out);
334 static inline void update_cgrp_time_from_event(struct perf_event *event)
336 struct perf_cgroup *cgrp;
339 * ensure we access cgroup data only when needed and
340 * when we know the cgroup is pinned (css_get)
342 if (!is_cgroup_event(event))
345 cgrp = perf_cgroup_from_task(current);
347 * Do not update time when cgroup is not active
349 if (cgrp == event->cgrp)
350 __update_cgrp_time(event->cgrp);
354 perf_cgroup_set_timestamp(struct task_struct *task,
355 struct perf_event_context *ctx)
357 struct perf_cgroup *cgrp;
358 struct perf_cgroup_info *info;
361 * ctx->lock held by caller
362 * ensure we do not access cgroup data
363 * unless we have the cgroup pinned (css_get)
365 if (!task || !ctx->nr_cgroups)
368 cgrp = perf_cgroup_from_task(task);
369 info = this_cpu_ptr(cgrp->info);
370 info->timestamp = ctx->timestamp;
373 #define PERF_CGROUP_SWOUT 0x1 /* cgroup switch out every event */
374 #define PERF_CGROUP_SWIN 0x2 /* cgroup switch in events based on task */
377 * reschedule events based on the cgroup constraint of task.
379 * mode SWOUT : schedule out everything
380 * mode SWIN : schedule in based on cgroup for next
382 void perf_cgroup_switch(struct task_struct *task, int mode)
384 struct perf_cpu_context *cpuctx;
389 * disable interrupts to avoid geting nr_cgroup
390 * changes via __perf_event_disable(). Also
393 local_irq_save(flags);
396 * we reschedule only in the presence of cgroup
397 * constrained events.
401 list_for_each_entry_rcu(pmu, &pmus, entry) {
402 cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
403 if (cpuctx->unique_pmu != pmu)
404 continue; /* ensure we process each cpuctx once */
407 * perf_cgroup_events says at least one
408 * context on this CPU has cgroup events.
410 * ctx->nr_cgroups reports the number of cgroup
411 * events for a context.
413 if (cpuctx->ctx.nr_cgroups > 0) {
414 perf_ctx_lock(cpuctx, cpuctx->task_ctx);
415 perf_pmu_disable(cpuctx->ctx.pmu);
417 if (mode & PERF_CGROUP_SWOUT) {
418 cpu_ctx_sched_out(cpuctx, EVENT_ALL);
420 * must not be done before ctxswout due
421 * to event_filter_match() in event_sched_out()
426 if (mode & PERF_CGROUP_SWIN) {
427 WARN_ON_ONCE(cpuctx->cgrp);
429 * set cgrp before ctxsw in to allow
430 * event_filter_match() to not have to pass
433 cpuctx->cgrp = perf_cgroup_from_task(task);
434 cpu_ctx_sched_in(cpuctx, EVENT_ALL, task);
436 perf_pmu_enable(cpuctx->ctx.pmu);
437 perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
443 local_irq_restore(flags);
446 static inline void perf_cgroup_sched_out(struct task_struct *task,
447 struct task_struct *next)
449 struct perf_cgroup *cgrp1;
450 struct perf_cgroup *cgrp2 = NULL;
453 * we come here when we know perf_cgroup_events > 0
455 cgrp1 = perf_cgroup_from_task(task);
458 * next is NULL when called from perf_event_enable_on_exec()
459 * that will systematically cause a cgroup_switch()
462 cgrp2 = perf_cgroup_from_task(next);
465 * only schedule out current cgroup events if we know
466 * that we are switching to a different cgroup. Otherwise,
467 * do no touch the cgroup events.
470 perf_cgroup_switch(task, PERF_CGROUP_SWOUT);
473 static inline void perf_cgroup_sched_in(struct task_struct *prev,
474 struct task_struct *task)
476 struct perf_cgroup *cgrp1;
477 struct perf_cgroup *cgrp2 = NULL;
480 * we come here when we know perf_cgroup_events > 0
482 cgrp1 = perf_cgroup_from_task(task);
484 /* prev can never be NULL */
485 cgrp2 = perf_cgroup_from_task(prev);
488 * only need to schedule in cgroup events if we are changing
489 * cgroup during ctxsw. Cgroup events were not scheduled
490 * out of ctxsw out if that was not the case.
493 perf_cgroup_switch(task, PERF_CGROUP_SWIN);
496 static inline int perf_cgroup_connect(int fd, struct perf_event *event,
497 struct perf_event_attr *attr,
498 struct perf_event *group_leader)
500 struct perf_cgroup *cgrp;
501 struct cgroup_subsys_state *css;
502 struct fd f = fdget(fd);
508 css = cgroup_css_from_dir(f.file, perf_subsys_id);
514 cgrp = container_of(css, struct perf_cgroup, css);
517 /* must be done before we fput() the file */
518 if (!perf_tryget_cgroup(event)) {
525 * all events in a group must monitor
526 * the same cgroup because a task belongs
527 * to only one perf cgroup at a time
529 if (group_leader && group_leader->cgrp != cgrp) {
530 perf_detach_cgroup(event);
539 perf_cgroup_set_shadow_time(struct perf_event *event, u64 now)
541 struct perf_cgroup_info *t;
542 t = per_cpu_ptr(event->cgrp->info, event->cpu);
543 event->shadow_ctx_time = now - t->timestamp;
547 perf_cgroup_defer_enabled(struct perf_event *event)
550 * when the current task's perf cgroup does not match
551 * the event's, we need to remember to call the
552 * perf_mark_enable() function the first time a task with
553 * a matching perf cgroup is scheduled in.
555 if (is_cgroup_event(event) && !perf_cgroup_match(event))
556 event->cgrp_defer_enabled = 1;
560 perf_cgroup_mark_enabled(struct perf_event *event,
561 struct perf_event_context *ctx)
563 struct perf_event *sub;
564 u64 tstamp = perf_event_time(event);
566 if (!event->cgrp_defer_enabled)
569 event->cgrp_defer_enabled = 0;
571 event->tstamp_enabled = tstamp - event->total_time_enabled;
572 list_for_each_entry(sub, &event->sibling_list, group_entry) {
573 if (sub->state >= PERF_EVENT_STATE_INACTIVE) {
574 sub->tstamp_enabled = tstamp - sub->total_time_enabled;
575 sub->cgrp_defer_enabled = 0;
579 #else /* !CONFIG_CGROUP_PERF */
582 perf_cgroup_match(struct perf_event *event)
587 static inline void perf_detach_cgroup(struct perf_event *event)
590 static inline int is_cgroup_event(struct perf_event *event)
595 static inline u64 perf_cgroup_event_cgrp_time(struct perf_event *event)
600 static inline void update_cgrp_time_from_event(struct perf_event *event)
604 static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx)
608 static inline void perf_cgroup_sched_out(struct task_struct *task,
609 struct task_struct *next)
613 static inline void perf_cgroup_sched_in(struct task_struct *prev,
614 struct task_struct *task)
618 static inline int perf_cgroup_connect(pid_t pid, struct perf_event *event,
619 struct perf_event_attr *attr,
620 struct perf_event *group_leader)
626 perf_cgroup_set_timestamp(struct task_struct *task,
627 struct perf_event_context *ctx)
632 perf_cgroup_switch(struct task_struct *task, struct task_struct *next)
637 perf_cgroup_set_shadow_time(struct perf_event *event, u64 now)
641 static inline u64 perf_cgroup_event_time(struct perf_event *event)
647 perf_cgroup_defer_enabled(struct perf_event *event)
652 perf_cgroup_mark_enabled(struct perf_event *event,
653 struct perf_event_context *ctx)
658 void perf_pmu_disable(struct pmu *pmu)
660 int *count = this_cpu_ptr(pmu->pmu_disable_count);
662 pmu->pmu_disable(pmu);
665 void perf_pmu_enable(struct pmu *pmu)
667 int *count = this_cpu_ptr(pmu->pmu_disable_count);
669 pmu->pmu_enable(pmu);
672 static DEFINE_PER_CPU(struct list_head, rotation_list);
675 * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
676 * because they're strictly cpu affine and rotate_start is called with IRQs
677 * disabled, while rotate_context is called from IRQ context.
679 static void perf_pmu_rotate_start(struct pmu *pmu)
681 struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
682 struct list_head *head = &__get_cpu_var(rotation_list);
684 WARN_ON(!irqs_disabled());
686 if (list_empty(&cpuctx->rotation_list)) {
687 int was_empty = list_empty(head);
688 list_add(&cpuctx->rotation_list, head);
690 tick_nohz_full_kick();
694 static void get_ctx(struct perf_event_context *ctx)
696 WARN_ON(!atomic_inc_not_zero(&ctx->refcount));
699 static void put_ctx(struct perf_event_context *ctx)
701 if (atomic_dec_and_test(&ctx->refcount)) {
703 put_ctx(ctx->parent_ctx);
705 put_task_struct(ctx->task);
706 kfree_rcu(ctx, rcu_head);
710 static void unclone_ctx(struct perf_event_context *ctx)
712 if (ctx->parent_ctx) {
713 put_ctx(ctx->parent_ctx);
714 ctx->parent_ctx = NULL;
718 static u32 perf_event_pid(struct perf_event *event, struct task_struct *p)
721 * only top level events have the pid namespace they were created in
724 event = event->parent;
726 return task_tgid_nr_ns(p, event->ns);
729 static u32 perf_event_tid(struct perf_event *event, struct task_struct *p)
732 * only top level events have the pid namespace they were created in
735 event = event->parent;
737 return task_pid_nr_ns(p, event->ns);
741 * If we inherit events we want to return the parent event id
744 static u64 primary_event_id(struct perf_event *event)
749 id = event->parent->id;
755 * Get the perf_event_context for a task and lock it.
756 * This has to cope with with the fact that until it is locked,
757 * the context could get moved to another task.
759 static struct perf_event_context *
760 perf_lock_task_context(struct task_struct *task, int ctxn, unsigned long *flags)
762 struct perf_event_context *ctx;
766 * One of the few rules of preemptible RCU is that one cannot do
767 * rcu_read_unlock() while holding a scheduler (or nested) lock when
768 * part of the read side critical section was preemptible -- see
769 * rcu_read_unlock_special().
771 * Since ctx->lock nests under rq->lock we must ensure the entire read
772 * side critical section is non-preemptible.
776 ctx = rcu_dereference(task->perf_event_ctxp[ctxn]);
779 * If this context is a clone of another, it might
780 * get swapped for another underneath us by
781 * perf_event_task_sched_out, though the
782 * rcu_read_lock() protects us from any context
783 * getting freed. Lock the context and check if it
784 * got swapped before we could get the lock, and retry
785 * if so. If we locked the right context, then it
786 * can't get swapped on us any more.
788 raw_spin_lock_irqsave(&ctx->lock, *flags);
789 if (ctx != rcu_dereference(task->perf_event_ctxp[ctxn])) {
790 raw_spin_unlock_irqrestore(&ctx->lock, *flags);
796 if (!atomic_inc_not_zero(&ctx->refcount)) {
797 raw_spin_unlock_irqrestore(&ctx->lock, *flags);
807 * Get the context for a task and increment its pin_count so it
808 * can't get swapped to another task. This also increments its
809 * reference count so that the context can't get freed.
811 static struct perf_event_context *
812 perf_pin_task_context(struct task_struct *task, int ctxn)
814 struct perf_event_context *ctx;
817 ctx = perf_lock_task_context(task, ctxn, &flags);
820 raw_spin_unlock_irqrestore(&ctx->lock, flags);
825 static void perf_unpin_context(struct perf_event_context *ctx)
829 raw_spin_lock_irqsave(&ctx->lock, flags);
831 raw_spin_unlock_irqrestore(&ctx->lock, flags);
835 * Update the record of the current time in a context.
837 static void update_context_time(struct perf_event_context *ctx)
839 u64 now = perf_clock();
841 ctx->time += now - ctx->timestamp;
842 ctx->timestamp = now;
845 static u64 perf_event_time(struct perf_event *event)
847 struct perf_event_context *ctx = event->ctx;
849 if (is_cgroup_event(event))
850 return perf_cgroup_event_time(event);
852 return ctx ? ctx->time : 0;
856 * Update the total_time_enabled and total_time_running fields for a event.
857 * The caller of this function needs to hold the ctx->lock.
859 static void update_event_times(struct perf_event *event)
861 struct perf_event_context *ctx = event->ctx;
864 if (event->state < PERF_EVENT_STATE_INACTIVE ||
865 event->group_leader->state < PERF_EVENT_STATE_INACTIVE)
868 * in cgroup mode, time_enabled represents
869 * the time the event was enabled AND active
870 * tasks were in the monitored cgroup. This is
871 * independent of the activity of the context as
872 * there may be a mix of cgroup and non-cgroup events.
874 * That is why we treat cgroup events differently
877 if (is_cgroup_event(event))
878 run_end = perf_cgroup_event_time(event);
879 else if (ctx->is_active)
882 run_end = event->tstamp_stopped;
884 event->total_time_enabled = run_end - event->tstamp_enabled;
886 if (event->state == PERF_EVENT_STATE_INACTIVE)
887 run_end = event->tstamp_stopped;
889 run_end = perf_event_time(event);
891 event->total_time_running = run_end - event->tstamp_running;
896 * Update total_time_enabled and total_time_running for all events in a group.
898 static void update_group_times(struct perf_event *leader)
900 struct perf_event *event;
902 update_event_times(leader);
903 list_for_each_entry(event, &leader->sibling_list, group_entry)
904 update_event_times(event);
907 static struct list_head *
908 ctx_group_list(struct perf_event *event, struct perf_event_context *ctx)
910 if (event->attr.pinned)
911 return &ctx->pinned_groups;
913 return &ctx->flexible_groups;
917 * Add a event from the lists for its context.
918 * Must be called with ctx->mutex and ctx->lock held.
921 list_add_event(struct perf_event *event, struct perf_event_context *ctx)
923 WARN_ON_ONCE(event->attach_state & PERF_ATTACH_CONTEXT);
924 event->attach_state |= PERF_ATTACH_CONTEXT;
927 * If we're a stand alone event or group leader, we go to the context
928 * list, group events are kept attached to the group so that
929 * perf_group_detach can, at all times, locate all siblings.
931 if (event->group_leader == event) {
932 struct list_head *list;
934 if (is_software_event(event))
935 event->group_flags |= PERF_GROUP_SOFTWARE;
937 list = ctx_group_list(event, ctx);
938 list_add_tail(&event->group_entry, list);
941 if (is_cgroup_event(event))
944 if (has_branch_stack(event))
945 ctx->nr_branch_stack++;
947 list_add_rcu(&event->event_entry, &ctx->event_list);
949 perf_pmu_rotate_start(ctx->pmu);
951 if (event->attr.inherit_stat)
956 * Initialize event state based on the perf_event_attr::disabled.
958 static inline void perf_event__state_init(struct perf_event *event)
960 event->state = event->attr.disabled ? PERF_EVENT_STATE_OFF :
961 PERF_EVENT_STATE_INACTIVE;
965 * Called at perf_event creation and when events are attached/detached from a
968 static void perf_event__read_size(struct perf_event *event)
970 int entry = sizeof(u64); /* value */
974 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
977 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
980 if (event->attr.read_format & PERF_FORMAT_ID)
981 entry += sizeof(u64);
983 if (event->attr.read_format & PERF_FORMAT_GROUP) {
984 nr += event->group_leader->nr_siblings;
989 event->read_size = size;
992 static void perf_event__header_size(struct perf_event *event)
994 struct perf_sample_data *data;
995 u64 sample_type = event->attr.sample_type;
998 perf_event__read_size(event);
1000 if (sample_type & PERF_SAMPLE_IP)
1001 size += sizeof(data->ip);
1003 if (sample_type & PERF_SAMPLE_ADDR)
1004 size += sizeof(data->addr);
1006 if (sample_type & PERF_SAMPLE_PERIOD)
1007 size += sizeof(data->period);
1009 if (sample_type & PERF_SAMPLE_WEIGHT)
1010 size += sizeof(data->weight);
1012 if (sample_type & PERF_SAMPLE_READ)
1013 size += event->read_size;
1015 if (sample_type & PERF_SAMPLE_DATA_SRC)
1016 size += sizeof(data->data_src.val);
1018 event->header_size = size;
1021 static void perf_event__id_header_size(struct perf_event *event)
1023 struct perf_sample_data *data;
1024 u64 sample_type = event->attr.sample_type;
1027 if (sample_type & PERF_SAMPLE_TID)
1028 size += sizeof(data->tid_entry);
1030 if (sample_type & PERF_SAMPLE_TIME)
1031 size += sizeof(data->time);
1033 if (sample_type & PERF_SAMPLE_ID)
1034 size += sizeof(data->id);
1036 if (sample_type & PERF_SAMPLE_STREAM_ID)
1037 size += sizeof(data->stream_id);
1039 if (sample_type & PERF_SAMPLE_CPU)
1040 size += sizeof(data->cpu_entry);
1042 event->id_header_size = size;
1045 static void perf_group_attach(struct perf_event *event)
1047 struct perf_event *group_leader = event->group_leader, *pos;
1050 * We can have double attach due to group movement in perf_event_open.
1052 if (event->attach_state & PERF_ATTACH_GROUP)
1055 event->attach_state |= PERF_ATTACH_GROUP;
1057 if (group_leader == event)
1060 if (group_leader->group_flags & PERF_GROUP_SOFTWARE &&
1061 !is_software_event(event))
1062 group_leader->group_flags &= ~PERF_GROUP_SOFTWARE;
1064 list_add_tail(&event->group_entry, &group_leader->sibling_list);
1065 group_leader->nr_siblings++;
1067 perf_event__header_size(group_leader);
1069 list_for_each_entry(pos, &group_leader->sibling_list, group_entry)
1070 perf_event__header_size(pos);
1074 * Remove a event from the lists for its context.
1075 * Must be called with ctx->mutex and ctx->lock held.
1078 list_del_event(struct perf_event *event, struct perf_event_context *ctx)
1080 struct perf_cpu_context *cpuctx;
1082 * We can have double detach due to exit/hot-unplug + close.
1084 if (!(event->attach_state & PERF_ATTACH_CONTEXT))
1087 event->attach_state &= ~PERF_ATTACH_CONTEXT;
1089 if (is_cgroup_event(event)) {
1091 cpuctx = __get_cpu_context(ctx);
1093 * if there are no more cgroup events
1094 * then cler cgrp to avoid stale pointer
1095 * in update_cgrp_time_from_cpuctx()
1097 if (!ctx->nr_cgroups)
1098 cpuctx->cgrp = NULL;
1101 if (has_branch_stack(event))
1102 ctx->nr_branch_stack--;
1105 if (event->attr.inherit_stat)
1108 list_del_rcu(&event->event_entry);
1110 if (event->group_leader == event)
1111 list_del_init(&event->group_entry);
1113 update_group_times(event);
1116 * If event was in error state, then keep it
1117 * that way, otherwise bogus counts will be
1118 * returned on read(). The only way to get out
1119 * of error state is by explicit re-enabling
1122 if (event->state > PERF_EVENT_STATE_OFF)
1123 event->state = PERF_EVENT_STATE_OFF;
1126 static void perf_group_detach(struct perf_event *event)
1128 struct perf_event *sibling, *tmp;
1129 struct list_head *list = NULL;
1132 * We can have double detach due to exit/hot-unplug + close.
1134 if (!(event->attach_state & PERF_ATTACH_GROUP))
1137 event->attach_state &= ~PERF_ATTACH_GROUP;
1140 * If this is a sibling, remove it from its group.
1142 if (event->group_leader != event) {
1143 list_del_init(&event->group_entry);
1144 event->group_leader->nr_siblings--;
1148 if (!list_empty(&event->group_entry))
1149 list = &event->group_entry;
1152 * If this was a group event with sibling events then
1153 * upgrade the siblings to singleton events by adding them
1154 * to whatever list we are on.
1156 list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) {
1158 list_move_tail(&sibling->group_entry, list);
1159 sibling->group_leader = sibling;
1161 /* Inherit group flags from the previous leader */
1162 sibling->group_flags = event->group_flags;
1166 perf_event__header_size(event->group_leader);
1168 list_for_each_entry(tmp, &event->group_leader->sibling_list, group_entry)
1169 perf_event__header_size(tmp);
1173 event_filter_match(struct perf_event *event)
1175 return (event->cpu == -1 || event->cpu == smp_processor_id())
1176 && perf_cgroup_match(event);
1180 event_sched_out(struct perf_event *event,
1181 struct perf_cpu_context *cpuctx,
1182 struct perf_event_context *ctx)
1184 u64 tstamp = perf_event_time(event);
1187 * An event which could not be activated because of
1188 * filter mismatch still needs to have its timings
1189 * maintained, otherwise bogus information is return
1190 * via read() for time_enabled, time_running:
1192 if (event->state == PERF_EVENT_STATE_INACTIVE
1193 && !event_filter_match(event)) {
1194 delta = tstamp - event->tstamp_stopped;
1195 event->tstamp_running += delta;
1196 event->tstamp_stopped = tstamp;
1199 if (event->state != PERF_EVENT_STATE_ACTIVE)
1202 event->state = PERF_EVENT_STATE_INACTIVE;
1203 if (event->pending_disable) {
1204 event->pending_disable = 0;
1205 event->state = PERF_EVENT_STATE_OFF;
1207 event->tstamp_stopped = tstamp;
1208 event->pmu->del(event, 0);
1211 if (!is_software_event(event))
1212 cpuctx->active_oncpu--;
1214 if (event->attr.freq && event->attr.sample_freq)
1216 if (event->attr.exclusive || !cpuctx->active_oncpu)
1217 cpuctx->exclusive = 0;
1221 group_sched_out(struct perf_event *group_event,
1222 struct perf_cpu_context *cpuctx,
1223 struct perf_event_context *ctx)
1225 struct perf_event *event;
1226 int state = group_event->state;
1228 event_sched_out(group_event, cpuctx, ctx);
1231 * Schedule out siblings (if any):
1233 list_for_each_entry(event, &group_event->sibling_list, group_entry)
1234 event_sched_out(event, cpuctx, ctx);
1236 if (state == PERF_EVENT_STATE_ACTIVE && group_event->attr.exclusive)
1237 cpuctx->exclusive = 0;
1241 * Cross CPU call to remove a performance event
1243 * We disable the event on the hardware level first. After that we
1244 * remove it from the context list.
1246 static int __perf_remove_from_context(void *info)
1248 struct perf_event *event = info;
1249 struct perf_event_context *ctx = event->ctx;
1250 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1252 raw_spin_lock(&ctx->lock);
1253 event_sched_out(event, cpuctx, ctx);
1254 list_del_event(event, ctx);
1255 if (!ctx->nr_events && cpuctx->task_ctx == ctx) {
1257 cpuctx->task_ctx = NULL;
1259 raw_spin_unlock(&ctx->lock);
1266 * Remove the event from a task's (or a CPU's) list of events.
1268 * CPU events are removed with a smp call. For task events we only
1269 * call when the task is on a CPU.
1271 * If event->ctx is a cloned context, callers must make sure that
1272 * every task struct that event->ctx->task could possibly point to
1273 * remains valid. This is OK when called from perf_release since
1274 * that only calls us on the top-level context, which can't be a clone.
1275 * When called from perf_event_exit_task, it's OK because the
1276 * context has been detached from its task.
1278 static void perf_remove_from_context(struct perf_event *event)
1280 struct perf_event_context *ctx = event->ctx;
1281 struct task_struct *task = ctx->task;
1283 lockdep_assert_held(&ctx->mutex);
1287 * Per cpu events are removed via an smp call and
1288 * the removal is always successful.
1290 cpu_function_call(event->cpu, __perf_remove_from_context, event);
1295 if (!task_function_call(task, __perf_remove_from_context, event))
1298 raw_spin_lock_irq(&ctx->lock);
1300 * If we failed to find a running task, but find the context active now
1301 * that we've acquired the ctx->lock, retry.
1303 if (ctx->is_active) {
1304 raw_spin_unlock_irq(&ctx->lock);
1309 * Since the task isn't running, its safe to remove the event, us
1310 * holding the ctx->lock ensures the task won't get scheduled in.
1312 list_del_event(event, ctx);
1313 raw_spin_unlock_irq(&ctx->lock);
1317 * Cross CPU call to disable a performance event
1319 int __perf_event_disable(void *info)
1321 struct perf_event *event = info;
1322 struct perf_event_context *ctx = event->ctx;
1323 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1326 * If this is a per-task event, need to check whether this
1327 * event's task is the current task on this cpu.
1329 * Can trigger due to concurrent perf_event_context_sched_out()
1330 * flipping contexts around.
1332 if (ctx->task && cpuctx->task_ctx != ctx)
1335 raw_spin_lock(&ctx->lock);
1338 * If the event is on, turn it off.
1339 * If it is in error state, leave it in error state.
1341 if (event->state >= PERF_EVENT_STATE_INACTIVE) {
1342 update_context_time(ctx);
1343 update_cgrp_time_from_event(event);
1344 update_group_times(event);
1345 if (event == event->group_leader)
1346 group_sched_out(event, cpuctx, ctx);
1348 event_sched_out(event, cpuctx, ctx);
1349 event->state = PERF_EVENT_STATE_OFF;
1352 raw_spin_unlock(&ctx->lock);
1360 * If event->ctx is a cloned context, callers must make sure that
1361 * every task struct that event->ctx->task could possibly point to
1362 * remains valid. This condition is satisifed when called through
1363 * perf_event_for_each_child or perf_event_for_each because they
1364 * hold the top-level event's child_mutex, so any descendant that
1365 * goes to exit will block in sync_child_event.
1366 * When called from perf_pending_event it's OK because event->ctx
1367 * is the current context on this CPU and preemption is disabled,
1368 * hence we can't get into perf_event_task_sched_out for this context.
1370 void perf_event_disable(struct perf_event *event)
1372 struct perf_event_context *ctx = event->ctx;
1373 struct task_struct *task = ctx->task;
1377 * Disable the event on the cpu that it's on
1379 cpu_function_call(event->cpu, __perf_event_disable, event);
1384 if (!task_function_call(task, __perf_event_disable, event))
1387 raw_spin_lock_irq(&ctx->lock);
1389 * If the event is still active, we need to retry the cross-call.
1391 if (event->state == PERF_EVENT_STATE_ACTIVE) {
1392 raw_spin_unlock_irq(&ctx->lock);
1394 * Reload the task pointer, it might have been changed by
1395 * a concurrent perf_event_context_sched_out().
1402 * Since we have the lock this context can't be scheduled
1403 * in, so we can change the state safely.
1405 if (event->state == PERF_EVENT_STATE_INACTIVE) {
1406 update_group_times(event);
1407 event->state = PERF_EVENT_STATE_OFF;
1409 raw_spin_unlock_irq(&ctx->lock);
1411 EXPORT_SYMBOL_GPL(perf_event_disable);
1413 static void perf_set_shadow_time(struct perf_event *event,
1414 struct perf_event_context *ctx,
1418 * use the correct time source for the time snapshot
1420 * We could get by without this by leveraging the
1421 * fact that to get to this function, the caller
1422 * has most likely already called update_context_time()
1423 * and update_cgrp_time_xx() and thus both timestamp
1424 * are identical (or very close). Given that tstamp is,
1425 * already adjusted for cgroup, we could say that:
1426 * tstamp - ctx->timestamp
1428 * tstamp - cgrp->timestamp.
1430 * Then, in perf_output_read(), the calculation would
1431 * work with no changes because:
1432 * - event is guaranteed scheduled in
1433 * - no scheduled out in between
1434 * - thus the timestamp would be the same
1436 * But this is a bit hairy.
1438 * So instead, we have an explicit cgroup call to remain
1439 * within the time time source all along. We believe it
1440 * is cleaner and simpler to understand.
1442 if (is_cgroup_event(event))
1443 perf_cgroup_set_shadow_time(event, tstamp);
1445 event->shadow_ctx_time = tstamp - ctx->timestamp;
1448 #define MAX_INTERRUPTS (~0ULL)
1450 static void perf_log_throttle(struct perf_event *event, int enable);
1453 event_sched_in(struct perf_event *event,
1454 struct perf_cpu_context *cpuctx,
1455 struct perf_event_context *ctx)
1457 u64 tstamp = perf_event_time(event);
1459 if (event->state <= PERF_EVENT_STATE_OFF)
1462 event->state = PERF_EVENT_STATE_ACTIVE;
1463 event->oncpu = smp_processor_id();
1466 * Unthrottle events, since we scheduled we might have missed several
1467 * ticks already, also for a heavily scheduling task there is little
1468 * guarantee it'll get a tick in a timely manner.
1470 if (unlikely(event->hw.interrupts == MAX_INTERRUPTS)) {
1471 perf_log_throttle(event, 1);
1472 event->hw.interrupts = 0;
1476 * The new state must be visible before we turn it on in the hardware:
1480 if (event->pmu->add(event, PERF_EF_START)) {
1481 event->state = PERF_EVENT_STATE_INACTIVE;
1486 event->tstamp_running += tstamp - event->tstamp_stopped;
1488 perf_set_shadow_time(event, ctx, tstamp);
1490 if (!is_software_event(event))
1491 cpuctx->active_oncpu++;
1493 if (event->attr.freq && event->attr.sample_freq)
1496 if (event->attr.exclusive)
1497 cpuctx->exclusive = 1;
1503 group_sched_in(struct perf_event *group_event,
1504 struct perf_cpu_context *cpuctx,
1505 struct perf_event_context *ctx)
1507 struct perf_event *event, *partial_group = NULL;
1508 struct pmu *pmu = group_event->pmu;
1509 u64 now = ctx->time;
1510 bool simulate = false;
1512 if (group_event->state == PERF_EVENT_STATE_OFF)
1515 pmu->start_txn(pmu);
1517 if (event_sched_in(group_event, cpuctx, ctx)) {
1518 pmu->cancel_txn(pmu);
1523 * Schedule in siblings as one group (if any):
1525 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
1526 if (event_sched_in(event, cpuctx, ctx)) {
1527 partial_group = event;
1532 if (!pmu->commit_txn(pmu))
1537 * Groups can be scheduled in as one unit only, so undo any
1538 * partial group before returning:
1539 * The events up to the failed event are scheduled out normally,
1540 * tstamp_stopped will be updated.
1542 * The failed events and the remaining siblings need to have
1543 * their timings updated as if they had gone thru event_sched_in()
1544 * and event_sched_out(). This is required to get consistent timings
1545 * across the group. This also takes care of the case where the group
1546 * could never be scheduled by ensuring tstamp_stopped is set to mark
1547 * the time the event was actually stopped, such that time delta
1548 * calculation in update_event_times() is correct.
1550 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
1551 if (event == partial_group)
1555 event->tstamp_running += now - event->tstamp_stopped;
1556 event->tstamp_stopped = now;
1558 event_sched_out(event, cpuctx, ctx);
1561 event_sched_out(group_event, cpuctx, ctx);
1563 pmu->cancel_txn(pmu);
1569 * Work out whether we can put this event group on the CPU now.
1571 static int group_can_go_on(struct perf_event *event,
1572 struct perf_cpu_context *cpuctx,
1576 * Groups consisting entirely of software events can always go on.
1578 if (event->group_flags & PERF_GROUP_SOFTWARE)
1581 * If an exclusive group is already on, no other hardware
1584 if (cpuctx->exclusive)
1587 * If this group is exclusive and there are already
1588 * events on the CPU, it can't go on.
1590 if (event->attr.exclusive && cpuctx->active_oncpu)
1593 * Otherwise, try to add it if all previous groups were able
1599 static void add_event_to_ctx(struct perf_event *event,
1600 struct perf_event_context *ctx)
1602 u64 tstamp = perf_event_time(event);
1604 list_add_event(event, ctx);
1605 perf_group_attach(event);
1606 event->tstamp_enabled = tstamp;
1607 event->tstamp_running = tstamp;
1608 event->tstamp_stopped = tstamp;
1611 static void task_ctx_sched_out(struct perf_event_context *ctx);
1613 ctx_sched_in(struct perf_event_context *ctx,
1614 struct perf_cpu_context *cpuctx,
1615 enum event_type_t event_type,
1616 struct task_struct *task);
1618 static void perf_event_sched_in(struct perf_cpu_context *cpuctx,
1619 struct perf_event_context *ctx,
1620 struct task_struct *task)
1622 cpu_ctx_sched_in(cpuctx, EVENT_PINNED, task);
1624 ctx_sched_in(ctx, cpuctx, EVENT_PINNED, task);
1625 cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE, task);
1627 ctx_sched_in(ctx, cpuctx, EVENT_FLEXIBLE, task);
1631 * Cross CPU call to install and enable a performance event
1633 * Must be called with ctx->mutex held
1635 static int __perf_install_in_context(void *info)
1637 struct perf_event *event = info;
1638 struct perf_event_context *ctx = event->ctx;
1639 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1640 struct perf_event_context *task_ctx = cpuctx->task_ctx;
1641 struct task_struct *task = current;
1643 perf_ctx_lock(cpuctx, task_ctx);
1644 perf_pmu_disable(cpuctx->ctx.pmu);
1647 * If there was an active task_ctx schedule it out.
1650 task_ctx_sched_out(task_ctx);
1653 * If the context we're installing events in is not the
1654 * active task_ctx, flip them.
1656 if (ctx->task && task_ctx != ctx) {
1658 raw_spin_unlock(&task_ctx->lock);
1659 raw_spin_lock(&ctx->lock);
1664 cpuctx->task_ctx = task_ctx;
1665 task = task_ctx->task;
1668 cpu_ctx_sched_out(cpuctx, EVENT_ALL);
1670 update_context_time(ctx);
1672 * update cgrp time only if current cgrp
1673 * matches event->cgrp. Must be done before
1674 * calling add_event_to_ctx()
1676 update_cgrp_time_from_event(event);
1678 add_event_to_ctx(event, ctx);
1681 * Schedule everything back in
1683 perf_event_sched_in(cpuctx, task_ctx, task);
1685 perf_pmu_enable(cpuctx->ctx.pmu);
1686 perf_ctx_unlock(cpuctx, task_ctx);
1692 * Attach a performance event to a context
1694 * First we add the event to the list with the hardware enable bit
1695 * in event->hw_config cleared.
1697 * If the event is attached to a task which is on a CPU we use a smp
1698 * call to enable it in the task context. The task might have been
1699 * scheduled away, but we check this in the smp call again.
1702 perf_install_in_context(struct perf_event_context *ctx,
1703 struct perf_event *event,
1706 struct task_struct *task = ctx->task;
1708 lockdep_assert_held(&ctx->mutex);
1711 if (event->cpu != -1)
1716 * Per cpu events are installed via an smp call and
1717 * the install is always successful.
1719 cpu_function_call(cpu, __perf_install_in_context, event);
1724 if (!task_function_call(task, __perf_install_in_context, event))
1727 raw_spin_lock_irq(&ctx->lock);
1729 * If we failed to find a running task, but find the context active now
1730 * that we've acquired the ctx->lock, retry.
1732 if (ctx->is_active) {
1733 raw_spin_unlock_irq(&ctx->lock);
1738 * Since the task isn't running, its safe to add the event, us holding
1739 * the ctx->lock ensures the task won't get scheduled in.
1741 add_event_to_ctx(event, ctx);
1742 raw_spin_unlock_irq(&ctx->lock);
1746 * Put a event into inactive state and update time fields.
1747 * Enabling the leader of a group effectively enables all
1748 * the group members that aren't explicitly disabled, so we
1749 * have to update their ->tstamp_enabled also.
1750 * Note: this works for group members as well as group leaders
1751 * since the non-leader members' sibling_lists will be empty.
1753 static void __perf_event_mark_enabled(struct perf_event *event)
1755 struct perf_event *sub;
1756 u64 tstamp = perf_event_time(event);
1758 event->state = PERF_EVENT_STATE_INACTIVE;
1759 event->tstamp_enabled = tstamp - event->total_time_enabled;
1760 list_for_each_entry(sub, &event->sibling_list, group_entry) {
1761 if (sub->state >= PERF_EVENT_STATE_INACTIVE)
1762 sub->tstamp_enabled = tstamp - sub->total_time_enabled;
1767 * Cross CPU call to enable a performance event
1769 static int __perf_event_enable(void *info)
1771 struct perf_event *event = info;
1772 struct perf_event_context *ctx = event->ctx;
1773 struct perf_event *leader = event->group_leader;
1774 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1778 * There's a time window between 'ctx->is_active' check
1779 * in perf_event_enable function and this place having:
1781 * - ctx->lock unlocked
1783 * where the task could be killed and 'ctx' deactivated
1784 * by perf_event_exit_task.
1786 if (!ctx->is_active)
1789 raw_spin_lock(&ctx->lock);
1790 update_context_time(ctx);
1792 if (event->state >= PERF_EVENT_STATE_INACTIVE)
1796 * set current task's cgroup time reference point
1798 perf_cgroup_set_timestamp(current, ctx);
1800 __perf_event_mark_enabled(event);
1802 if (!event_filter_match(event)) {
1803 if (is_cgroup_event(event))
1804 perf_cgroup_defer_enabled(event);
1809 * If the event is in a group and isn't the group leader,
1810 * then don't put it on unless the group is on.
1812 if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE)
1815 if (!group_can_go_on(event, cpuctx, 1)) {
1818 if (event == leader)
1819 err = group_sched_in(event, cpuctx, ctx);
1821 err = event_sched_in(event, cpuctx, ctx);
1826 * If this event can't go on and it's part of a
1827 * group, then the whole group has to come off.
1829 if (leader != event)
1830 group_sched_out(leader, cpuctx, ctx);
1831 if (leader->attr.pinned) {
1832 update_group_times(leader);
1833 leader->state = PERF_EVENT_STATE_ERROR;
1838 raw_spin_unlock(&ctx->lock);
1846 * If event->ctx is a cloned context, callers must make sure that
1847 * every task struct that event->ctx->task could possibly point to
1848 * remains valid. This condition is satisfied when called through
1849 * perf_event_for_each_child or perf_event_for_each as described
1850 * for perf_event_disable.
1852 void perf_event_enable(struct perf_event *event)
1854 struct perf_event_context *ctx = event->ctx;
1855 struct task_struct *task = ctx->task;
1859 * Enable the event on the cpu that it's on
1861 cpu_function_call(event->cpu, __perf_event_enable, event);
1865 raw_spin_lock_irq(&ctx->lock);
1866 if (event->state >= PERF_EVENT_STATE_INACTIVE)
1870 * If the event is in error state, clear that first.
1871 * That way, if we see the event in error state below, we
1872 * know that it has gone back into error state, as distinct
1873 * from the task having been scheduled away before the
1874 * cross-call arrived.
1876 if (event->state == PERF_EVENT_STATE_ERROR)
1877 event->state = PERF_EVENT_STATE_OFF;
1880 if (!ctx->is_active) {
1881 __perf_event_mark_enabled(event);
1885 raw_spin_unlock_irq(&ctx->lock);
1887 if (!task_function_call(task, __perf_event_enable, event))
1890 raw_spin_lock_irq(&ctx->lock);
1893 * If the context is active and the event is still off,
1894 * we need to retry the cross-call.
1896 if (ctx->is_active && event->state == PERF_EVENT_STATE_OFF) {
1898 * task could have been flipped by a concurrent
1899 * perf_event_context_sched_out()
1906 raw_spin_unlock_irq(&ctx->lock);
1908 EXPORT_SYMBOL_GPL(perf_event_enable);
1910 int perf_event_refresh(struct perf_event *event, int refresh)
1913 * not supported on inherited events
1915 if (event->attr.inherit || !is_sampling_event(event))
1918 atomic_add(refresh, &event->event_limit);
1919 perf_event_enable(event);
1923 EXPORT_SYMBOL_GPL(perf_event_refresh);
1925 static void ctx_sched_out(struct perf_event_context *ctx,
1926 struct perf_cpu_context *cpuctx,
1927 enum event_type_t event_type)
1929 struct perf_event *event;
1930 int is_active = ctx->is_active;
1932 ctx->is_active &= ~event_type;
1933 if (likely(!ctx->nr_events))
1936 update_context_time(ctx);
1937 update_cgrp_time_from_cpuctx(cpuctx);
1938 if (!ctx->nr_active)
1941 perf_pmu_disable(ctx->pmu);
1942 if ((is_active & EVENT_PINNED) && (event_type & EVENT_PINNED)) {
1943 list_for_each_entry(event, &ctx->pinned_groups, group_entry)
1944 group_sched_out(event, cpuctx, ctx);
1947 if ((is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE)) {
1948 list_for_each_entry(event, &ctx->flexible_groups, group_entry)
1949 group_sched_out(event, cpuctx, ctx);
1951 perf_pmu_enable(ctx->pmu);
1955 * Test whether two contexts are equivalent, i.e. whether they
1956 * have both been cloned from the same version of the same context
1957 * and they both have the same number of enabled events.
1958 * If the number of enabled events is the same, then the set
1959 * of enabled events should be the same, because these are both
1960 * inherited contexts, therefore we can't access individual events
1961 * in them directly with an fd; we can only enable/disable all
1962 * events via prctl, or enable/disable all events in a family
1963 * via ioctl, which will have the same effect on both contexts.
1965 static int context_equiv(struct perf_event_context *ctx1,
1966 struct perf_event_context *ctx2)
1968 return ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx
1969 && ctx1->parent_gen == ctx2->parent_gen
1970 && !ctx1->pin_count && !ctx2->pin_count;
1973 static void __perf_event_sync_stat(struct perf_event *event,
1974 struct perf_event *next_event)
1978 if (!event->attr.inherit_stat)
1982 * Update the event value, we cannot use perf_event_read()
1983 * because we're in the middle of a context switch and have IRQs
1984 * disabled, which upsets smp_call_function_single(), however
1985 * we know the event must be on the current CPU, therefore we
1986 * don't need to use it.
1988 switch (event->state) {
1989 case PERF_EVENT_STATE_ACTIVE:
1990 event->pmu->read(event);
1993 case PERF_EVENT_STATE_INACTIVE:
1994 update_event_times(event);
2002 * In order to keep per-task stats reliable we need to flip the event
2003 * values when we flip the contexts.
2005 value = local64_read(&next_event->count);
2006 value = local64_xchg(&event->count, value);
2007 local64_set(&next_event->count, value);
2009 swap(event->total_time_enabled, next_event->total_time_enabled);
2010 swap(event->total_time_running, next_event->total_time_running);
2013 * Since we swizzled the values, update the user visible data too.
2015 perf_event_update_userpage(event);
2016 perf_event_update_userpage(next_event);
2019 static void perf_event_sync_stat(struct perf_event_context *ctx,
2020 struct perf_event_context *next_ctx)
2022 struct perf_event *event, *next_event;
2027 update_context_time(ctx);
2029 event = list_first_entry(&ctx->event_list,
2030 struct perf_event, event_entry);
2032 next_event = list_first_entry(&next_ctx->event_list,
2033 struct perf_event, event_entry);
2035 while (&event->event_entry != &ctx->event_list &&
2036 &next_event->event_entry != &next_ctx->event_list) {
2038 __perf_event_sync_stat(event, next_event);
2040 event = list_next_entry(event, event_entry);
2041 next_event = list_next_entry(next_event, event_entry);
2045 static void perf_event_context_sched_out(struct task_struct *task, int ctxn,
2046 struct task_struct *next)
2048 struct perf_event_context *ctx = task->perf_event_ctxp[ctxn];
2049 struct perf_event_context *next_ctx;
2050 struct perf_event_context *parent;
2051 struct perf_cpu_context *cpuctx;
2057 cpuctx = __get_cpu_context(ctx);
2058 if (!cpuctx->task_ctx)
2062 parent = rcu_dereference(ctx->parent_ctx);
2063 next_ctx = next->perf_event_ctxp[ctxn];
2064 if (parent && next_ctx &&
2065 rcu_dereference(next_ctx->parent_ctx) == parent) {
2067 * Looks like the two contexts are clones, so we might be
2068 * able to optimize the context switch. We lock both
2069 * contexts and check that they are clones under the
2070 * lock (including re-checking that neither has been
2071 * uncloned in the meantime). It doesn't matter which
2072 * order we take the locks because no other cpu could
2073 * be trying to lock both of these tasks.
2075 raw_spin_lock(&ctx->lock);
2076 raw_spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING);
2077 if (context_equiv(ctx, next_ctx)) {
2079 * XXX do we need a memory barrier of sorts
2080 * wrt to rcu_dereference() of perf_event_ctxp
2082 task->perf_event_ctxp[ctxn] = next_ctx;
2083 next->perf_event_ctxp[ctxn] = ctx;
2085 next_ctx->task = task;
2088 perf_event_sync_stat(ctx, next_ctx);
2090 raw_spin_unlock(&next_ctx->lock);
2091 raw_spin_unlock(&ctx->lock);
2096 raw_spin_lock(&ctx->lock);
2097 ctx_sched_out(ctx, cpuctx, EVENT_ALL);
2098 cpuctx->task_ctx = NULL;
2099 raw_spin_unlock(&ctx->lock);
2103 #define for_each_task_context_nr(ctxn) \
2104 for ((ctxn) = 0; (ctxn) < perf_nr_task_contexts; (ctxn)++)
2107 * Called from scheduler to remove the events of the current task,
2108 * with interrupts disabled.
2110 * We stop each event and update the event value in event->count.
2112 * This does not protect us against NMI, but disable()
2113 * sets the disabled bit in the control field of event _before_
2114 * accessing the event control register. If a NMI hits, then it will
2115 * not restart the event.
2117 void __perf_event_task_sched_out(struct task_struct *task,
2118 struct task_struct *next)
2122 for_each_task_context_nr(ctxn)
2123 perf_event_context_sched_out(task, ctxn, next);
2126 * if cgroup events exist on this CPU, then we need
2127 * to check if we have to switch out PMU state.
2128 * cgroup event are system-wide mode only
2130 if (atomic_read(&__get_cpu_var(perf_cgroup_events)))
2131 perf_cgroup_sched_out(task, next);
2134 static void task_ctx_sched_out(struct perf_event_context *ctx)
2136 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
2138 if (!cpuctx->task_ctx)
2141 if (WARN_ON_ONCE(ctx != cpuctx->task_ctx))
2144 ctx_sched_out(ctx, cpuctx, EVENT_ALL);
2145 cpuctx->task_ctx = NULL;
2149 * Called with IRQs disabled
2151 static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
2152 enum event_type_t event_type)
2154 ctx_sched_out(&cpuctx->ctx, cpuctx, event_type);
2158 ctx_pinned_sched_in(struct perf_event_context *ctx,
2159 struct perf_cpu_context *cpuctx)
2161 struct perf_event *event;
2163 list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
2164 if (event->state <= PERF_EVENT_STATE_OFF)
2166 if (!event_filter_match(event))
2169 /* may need to reset tstamp_enabled */
2170 if (is_cgroup_event(event))
2171 perf_cgroup_mark_enabled(event, ctx);
2173 if (group_can_go_on(event, cpuctx, 1))
2174 group_sched_in(event, cpuctx, ctx);
2177 * If this pinned group hasn't been scheduled,
2178 * put it in error state.
2180 if (event->state == PERF_EVENT_STATE_INACTIVE) {
2181 update_group_times(event);
2182 event->state = PERF_EVENT_STATE_ERROR;
2188 ctx_flexible_sched_in(struct perf_event_context *ctx,
2189 struct perf_cpu_context *cpuctx)
2191 struct perf_event *event;
2194 list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
2195 /* Ignore events in OFF or ERROR state */
2196 if (event->state <= PERF_EVENT_STATE_OFF)
2199 * Listen to the 'cpu' scheduling filter constraint
2202 if (!event_filter_match(event))
2205 /* may need to reset tstamp_enabled */
2206 if (is_cgroup_event(event))
2207 perf_cgroup_mark_enabled(event, ctx);
2209 if (group_can_go_on(event, cpuctx, can_add_hw)) {
2210 if (group_sched_in(event, cpuctx, ctx))
2217 ctx_sched_in(struct perf_event_context *ctx,
2218 struct perf_cpu_context *cpuctx,
2219 enum event_type_t event_type,
2220 struct task_struct *task)
2223 int is_active = ctx->is_active;
2225 ctx->is_active |= event_type;
2226 if (likely(!ctx->nr_events))
2230 ctx->timestamp = now;
2231 perf_cgroup_set_timestamp(task, ctx);
2233 * First go through the list and put on any pinned groups
2234 * in order to give them the best chance of going on.
2236 if (!(is_active & EVENT_PINNED) && (event_type & EVENT_PINNED))
2237 ctx_pinned_sched_in(ctx, cpuctx);
2239 /* Then walk through the lower prio flexible groups */
2240 if (!(is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE))
2241 ctx_flexible_sched_in(ctx, cpuctx);
2244 static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
2245 enum event_type_t event_type,
2246 struct task_struct *task)
2248 struct perf_event_context *ctx = &cpuctx->ctx;
2250 ctx_sched_in(ctx, cpuctx, event_type, task);
2253 static void perf_event_context_sched_in(struct perf_event_context *ctx,
2254 struct task_struct *task)
2256 struct perf_cpu_context *cpuctx;
2258 cpuctx = __get_cpu_context(ctx);
2259 if (cpuctx->task_ctx == ctx)
2262 perf_ctx_lock(cpuctx, ctx);
2263 perf_pmu_disable(ctx->pmu);
2265 * We want to keep the following priority order:
2266 * cpu pinned (that don't need to move), task pinned,
2267 * cpu flexible, task flexible.
2269 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
2272 cpuctx->task_ctx = ctx;
2274 perf_event_sched_in(cpuctx, cpuctx->task_ctx, task);
2276 perf_pmu_enable(ctx->pmu);
2277 perf_ctx_unlock(cpuctx, ctx);
2280 * Since these rotations are per-cpu, we need to ensure the
2281 * cpu-context we got scheduled on is actually rotating.
2283 perf_pmu_rotate_start(ctx->pmu);
2287 * When sampling the branck stack in system-wide, it may be necessary
2288 * to flush the stack on context switch. This happens when the branch
2289 * stack does not tag its entries with the pid of the current task.
2290 * Otherwise it becomes impossible to associate a branch entry with a
2291 * task. This ambiguity is more likely to appear when the branch stack
2292 * supports priv level filtering and the user sets it to monitor only
2293 * at the user level (which could be a useful measurement in system-wide
2294 * mode). In that case, the risk is high of having a branch stack with
2295 * branch from multiple tasks. Flushing may mean dropping the existing
2296 * entries or stashing them somewhere in the PMU specific code layer.
2298 * This function provides the context switch callback to the lower code
2299 * layer. It is invoked ONLY when there is at least one system-wide context
2300 * with at least one active event using taken branch sampling.
2302 static void perf_branch_stack_sched_in(struct task_struct *prev,
2303 struct task_struct *task)
2305 struct perf_cpu_context *cpuctx;
2307 unsigned long flags;
2309 /* no need to flush branch stack if not changing task */
2313 local_irq_save(flags);
2317 list_for_each_entry_rcu(pmu, &pmus, entry) {
2318 cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
2321 * check if the context has at least one
2322 * event using PERF_SAMPLE_BRANCH_STACK
2324 if (cpuctx->ctx.nr_branch_stack > 0
2325 && pmu->flush_branch_stack) {
2327 pmu = cpuctx->ctx.pmu;
2329 perf_ctx_lock(cpuctx, cpuctx->task_ctx);
2331 perf_pmu_disable(pmu);
2333 pmu->flush_branch_stack();
2335 perf_pmu_enable(pmu);
2337 perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
2343 local_irq_restore(flags);
2347 * Called from scheduler to add the events of the current task
2348 * with interrupts disabled.
2350 * We restore the event value and then enable it.
2352 * This does not protect us against NMI, but enable()
2353 * sets the enabled bit in the control field of event _before_
2354 * accessing the event control register. If a NMI hits, then it will
2355 * keep the event running.
2357 void __perf_event_task_sched_in(struct task_struct *prev,
2358 struct task_struct *task)
2360 struct perf_event_context *ctx;
2363 for_each_task_context_nr(ctxn) {
2364 ctx = task->perf_event_ctxp[ctxn];
2368 perf_event_context_sched_in(ctx, task);
2371 * if cgroup events exist on this CPU, then we need
2372 * to check if we have to switch in PMU state.
2373 * cgroup event are system-wide mode only
2375 if (atomic_read(&__get_cpu_var(perf_cgroup_events)))
2376 perf_cgroup_sched_in(prev, task);
2378 /* check for system-wide branch_stack events */
2379 if (atomic_read(&__get_cpu_var(perf_branch_stack_events)))
2380 perf_branch_stack_sched_in(prev, task);
2383 static u64 perf_calculate_period(struct perf_event *event, u64 nsec, u64 count)
2385 u64 frequency = event->attr.sample_freq;
2386 u64 sec = NSEC_PER_SEC;
2387 u64 divisor, dividend;
2389 int count_fls, nsec_fls, frequency_fls, sec_fls;
2391 count_fls = fls64(count);
2392 nsec_fls = fls64(nsec);
2393 frequency_fls = fls64(frequency);
2397 * We got @count in @nsec, with a target of sample_freq HZ
2398 * the target period becomes:
2401 * period = -------------------
2402 * @nsec * sample_freq
2407 * Reduce accuracy by one bit such that @a and @b converge
2408 * to a similar magnitude.
2410 #define REDUCE_FLS(a, b) \
2412 if (a##_fls > b##_fls) { \
2422 * Reduce accuracy until either term fits in a u64, then proceed with
2423 * the other, so that finally we can do a u64/u64 division.
2425 while (count_fls + sec_fls > 64 && nsec_fls + frequency_fls > 64) {
2426 REDUCE_FLS(nsec, frequency);
2427 REDUCE_FLS(sec, count);
2430 if (count_fls + sec_fls > 64) {
2431 divisor = nsec * frequency;
2433 while (count_fls + sec_fls > 64) {
2434 REDUCE_FLS(count, sec);
2438 dividend = count * sec;
2440 dividend = count * sec;
2442 while (nsec_fls + frequency_fls > 64) {
2443 REDUCE_FLS(nsec, frequency);
2447 divisor = nsec * frequency;
2453 return div64_u64(dividend, divisor);
2456 static DEFINE_PER_CPU(int, perf_throttled_count);
2457 static DEFINE_PER_CPU(u64, perf_throttled_seq);
2459 static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count, bool disable)
2461 struct hw_perf_event *hwc = &event->hw;
2462 s64 period, sample_period;
2465 period = perf_calculate_period(event, nsec, count);
2467 delta = (s64)(period - hwc->sample_period);
2468 delta = (delta + 7) / 8; /* low pass filter */
2470 sample_period = hwc->sample_period + delta;
2475 hwc->sample_period = sample_period;
2477 if (local64_read(&hwc->period_left) > 8*sample_period) {
2479 event->pmu->stop(event, PERF_EF_UPDATE);
2481 local64_set(&hwc->period_left, 0);
2484 event->pmu->start(event, PERF_EF_RELOAD);
2489 * combine freq adjustment with unthrottling to avoid two passes over the
2490 * events. At the same time, make sure, having freq events does not change
2491 * the rate of unthrottling as that would introduce bias.
2493 static void perf_adjust_freq_unthr_context(struct perf_event_context *ctx,
2496 struct perf_event *event;
2497 struct hw_perf_event *hwc;
2498 u64 now, period = TICK_NSEC;
2502 * only need to iterate over all events iff:
2503 * - context have events in frequency mode (needs freq adjust)
2504 * - there are events to unthrottle on this cpu
2506 if (!(ctx->nr_freq || needs_unthr))
2509 raw_spin_lock(&ctx->lock);
2510 perf_pmu_disable(ctx->pmu);
2512 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
2513 if (event->state != PERF_EVENT_STATE_ACTIVE)
2516 if (!event_filter_match(event))
2521 if (needs_unthr && hwc->interrupts == MAX_INTERRUPTS) {
2522 hwc->interrupts = 0;
2523 perf_log_throttle(event, 1);
2524 event->pmu->start(event, 0);
2527 if (!event->attr.freq || !event->attr.sample_freq)
2531 * stop the event and update event->count
2533 event->pmu->stop(event, PERF_EF_UPDATE);
2535 now = local64_read(&event->count);
2536 delta = now - hwc->freq_count_stamp;
2537 hwc->freq_count_stamp = now;
2541 * reload only if value has changed
2542 * we have stopped the event so tell that
2543 * to perf_adjust_period() to avoid stopping it
2547 perf_adjust_period(event, period, delta, false);
2549 event->pmu->start(event, delta > 0 ? PERF_EF_RELOAD : 0);
2552 perf_pmu_enable(ctx->pmu);
2553 raw_spin_unlock(&ctx->lock);
2557 * Round-robin a context's events:
2559 static void rotate_ctx(struct perf_event_context *ctx)
2562 * Rotate the first entry last of non-pinned groups. Rotation might be
2563 * disabled by the inheritance code.
2565 if (!ctx->rotate_disable)
2566 list_rotate_left(&ctx->flexible_groups);
2570 * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
2571 * because they're strictly cpu affine and rotate_start is called with IRQs
2572 * disabled, while rotate_context is called from IRQ context.
2574 static void perf_rotate_context(struct perf_cpu_context *cpuctx)
2576 struct perf_event_context *ctx = NULL;
2577 int rotate = 0, remove = 1;
2579 if (cpuctx->ctx.nr_events) {
2581 if (cpuctx->ctx.nr_events != cpuctx->ctx.nr_active)
2585 ctx = cpuctx->task_ctx;
2586 if (ctx && ctx->nr_events) {
2588 if (ctx->nr_events != ctx->nr_active)
2595 perf_ctx_lock(cpuctx, cpuctx->task_ctx);
2596 perf_pmu_disable(cpuctx->ctx.pmu);
2598 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
2600 ctx_sched_out(ctx, cpuctx, EVENT_FLEXIBLE);
2602 rotate_ctx(&cpuctx->ctx);
2606 perf_event_sched_in(cpuctx, ctx, current);
2608 perf_pmu_enable(cpuctx->ctx.pmu);
2609 perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
2612 list_del_init(&cpuctx->rotation_list);
2615 #ifdef CONFIG_NO_HZ_FULL
2616 bool perf_event_can_stop_tick(void)
2618 if (list_empty(&__get_cpu_var(rotation_list)))
2625 void perf_event_task_tick(void)
2627 struct list_head *head = &__get_cpu_var(rotation_list);
2628 struct perf_cpu_context *cpuctx, *tmp;
2629 struct perf_event_context *ctx;
2632 WARN_ON(!irqs_disabled());
2634 __this_cpu_inc(perf_throttled_seq);
2635 throttled = __this_cpu_xchg(perf_throttled_count, 0);
2637 list_for_each_entry_safe(cpuctx, tmp, head, rotation_list) {
2639 perf_adjust_freq_unthr_context(ctx, throttled);
2641 ctx = cpuctx->task_ctx;
2643 perf_adjust_freq_unthr_context(ctx, throttled);
2645 if (cpuctx->jiffies_interval == 1 ||
2646 !(jiffies % cpuctx->jiffies_interval))
2647 perf_rotate_context(cpuctx);
2651 static int event_enable_on_exec(struct perf_event *event,
2652 struct perf_event_context *ctx)
2654 if (!event->attr.enable_on_exec)
2657 event->attr.enable_on_exec = 0;
2658 if (event->state >= PERF_EVENT_STATE_INACTIVE)
2661 __perf_event_mark_enabled(event);
2667 * Enable all of a task's events that have been marked enable-on-exec.
2668 * This expects task == current.
2670 static void perf_event_enable_on_exec(struct perf_event_context *ctx)
2672 struct perf_event *event;
2673 unsigned long flags;
2677 local_irq_save(flags);
2678 if (!ctx || !ctx->nr_events)
2682 * We must ctxsw out cgroup events to avoid conflict
2683 * when invoking perf_task_event_sched_in() later on
2684 * in this function. Otherwise we end up trying to
2685 * ctxswin cgroup events which are already scheduled
2688 perf_cgroup_sched_out(current, NULL);
2690 raw_spin_lock(&ctx->lock);
2691 task_ctx_sched_out(ctx);
2693 list_for_each_entry(event, &ctx->event_list, event_entry) {
2694 ret = event_enable_on_exec(event, ctx);
2700 * Unclone this context if we enabled any event.
2705 raw_spin_unlock(&ctx->lock);
2708 * Also calls ctxswin for cgroup events, if any:
2710 perf_event_context_sched_in(ctx, ctx->task);
2712 local_irq_restore(flags);
2716 * Cross CPU call to read the hardware event
2718 static void __perf_event_read(void *info)
2720 struct perf_event *event = info;
2721 struct perf_event_context *ctx = event->ctx;
2722 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
2725 * If this is a task context, we need to check whether it is
2726 * the current task context of this cpu. If not it has been
2727 * scheduled out before the smp call arrived. In that case
2728 * event->count would have been updated to a recent sample
2729 * when the event was scheduled out.
2731 if (ctx->task && cpuctx->task_ctx != ctx)
2734 raw_spin_lock(&ctx->lock);
2735 if (ctx->is_active) {
2736 update_context_time(ctx);
2737 update_cgrp_time_from_event(event);
2739 update_event_times(event);
2740 if (event->state == PERF_EVENT_STATE_ACTIVE)
2741 event->pmu->read(event);
2742 raw_spin_unlock(&ctx->lock);
2745 static inline u64 perf_event_count(struct perf_event *event)
2747 return local64_read(&event->count) + atomic64_read(&event->child_count);
2750 static u64 perf_event_read(struct perf_event *event)
2753 * If event is enabled and currently active on a CPU, update the
2754 * value in the event structure:
2756 if (event->state == PERF_EVENT_STATE_ACTIVE) {
2757 smp_call_function_single(event->oncpu,
2758 __perf_event_read, event, 1);
2759 } else if (event->state == PERF_EVENT_STATE_INACTIVE) {
2760 struct perf_event_context *ctx = event->ctx;
2761 unsigned long flags;
2763 raw_spin_lock_irqsave(&ctx->lock, flags);
2765 * may read while context is not active
2766 * (e.g., thread is blocked), in that case
2767 * we cannot update context time
2769 if (ctx->is_active) {
2770 update_context_time(ctx);
2771 update_cgrp_time_from_event(event);
2773 update_event_times(event);
2774 raw_spin_unlock_irqrestore(&ctx->lock, flags);
2777 return perf_event_count(event);
2781 * Initialize the perf_event context in a task_struct:
2783 static void __perf_event_init_context(struct perf_event_context *ctx)
2785 raw_spin_lock_init(&ctx->lock);
2786 mutex_init(&ctx->mutex);
2787 INIT_LIST_HEAD(&ctx->pinned_groups);
2788 INIT_LIST_HEAD(&ctx->flexible_groups);
2789 INIT_LIST_HEAD(&ctx->event_list);
2790 atomic_set(&ctx->refcount, 1);
2793 static struct perf_event_context *
2794 alloc_perf_context(struct pmu *pmu, struct task_struct *task)
2796 struct perf_event_context *ctx;
2798 ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL);
2802 __perf_event_init_context(ctx);
2805 get_task_struct(task);
2812 static struct task_struct *
2813 find_lively_task_by_vpid(pid_t vpid)
2815 struct task_struct *task;
2822 task = find_task_by_vpid(vpid);
2824 get_task_struct(task);
2828 return ERR_PTR(-ESRCH);
2830 /* Reuse ptrace permission checks for now. */
2832 if (!ptrace_may_access(task, PTRACE_MODE_READ))
2837 put_task_struct(task);
2838 return ERR_PTR(err);
2843 * Returns a matching context with refcount and pincount.
2845 static struct perf_event_context *
2846 find_get_context(struct pmu *pmu, struct task_struct *task, int cpu)
2848 struct perf_event_context *ctx;
2849 struct perf_cpu_context *cpuctx;
2850 unsigned long flags;
2854 /* Must be root to operate on a CPU event: */
2855 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
2856 return ERR_PTR(-EACCES);
2859 * We could be clever and allow to attach a event to an
2860 * offline CPU and activate it when the CPU comes up, but
2863 if (!cpu_online(cpu))
2864 return ERR_PTR(-ENODEV);
2866 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
2875 ctxn = pmu->task_ctx_nr;
2880 ctx = perf_lock_task_context(task, ctxn, &flags);
2884 raw_spin_unlock_irqrestore(&ctx->lock, flags);
2886 ctx = alloc_perf_context(pmu, task);
2892 mutex_lock(&task->perf_event_mutex);
2894 * If it has already passed perf_event_exit_task().
2895 * we must see PF_EXITING, it takes this mutex too.
2897 if (task->flags & PF_EXITING)
2899 else if (task->perf_event_ctxp[ctxn])
2904 rcu_assign_pointer(task->perf_event_ctxp[ctxn], ctx);
2906 mutex_unlock(&task->perf_event_mutex);
2908 if (unlikely(err)) {
2920 return ERR_PTR(err);
2923 static void perf_event_free_filter(struct perf_event *event);
2925 static void free_event_rcu(struct rcu_head *head)
2927 struct perf_event *event;
2929 event = container_of(head, struct perf_event, rcu_head);
2931 put_pid_ns(event->ns);
2932 perf_event_free_filter(event);
2936 static void ring_buffer_put(struct ring_buffer *rb);
2937 static void ring_buffer_detach(struct perf_event *event, struct ring_buffer *rb);
2939 static void free_event(struct perf_event *event)
2941 irq_work_sync(&event->pending);
2943 if (!event->parent) {
2944 if (event->attach_state & PERF_ATTACH_TASK)
2945 static_key_slow_dec_deferred(&perf_sched_events);
2946 if (event->attr.mmap || event->attr.mmap_data)
2947 atomic_dec(&nr_mmap_events);
2948 if (event->attr.comm)
2949 atomic_dec(&nr_comm_events);
2950 if (event->attr.task)
2951 atomic_dec(&nr_task_events);
2952 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN)
2953 put_callchain_buffers();
2954 if (is_cgroup_event(event)) {
2955 atomic_dec(&per_cpu(perf_cgroup_events, event->cpu));
2956 static_key_slow_dec_deferred(&perf_sched_events);
2959 if (has_branch_stack(event)) {
2960 static_key_slow_dec_deferred(&perf_sched_events);
2961 /* is system-wide event */
2962 if (!(event->attach_state & PERF_ATTACH_TASK)) {
2963 atomic_dec(&per_cpu(perf_branch_stack_events,
2970 struct ring_buffer *rb;
2973 * Can happen when we close an event with re-directed output.
2975 * Since we have a 0 refcount, perf_mmap_close() will skip
2976 * over us; possibly making our ring_buffer_put() the last.
2978 mutex_lock(&event->mmap_mutex);
2981 rcu_assign_pointer(event->rb, NULL);
2982 ring_buffer_detach(event, rb);
2983 ring_buffer_put(rb); /* could be last */
2985 mutex_unlock(&event->mmap_mutex);
2988 if (is_cgroup_event(event))
2989 perf_detach_cgroup(event);
2992 event->destroy(event);
2995 put_ctx(event->ctx);
2997 call_rcu(&event->rcu_head, free_event_rcu);
3000 int perf_event_release_kernel(struct perf_event *event)
3002 struct perf_event_context *ctx = event->ctx;
3004 WARN_ON_ONCE(ctx->parent_ctx);
3006 * There are two ways this annotation is useful:
3008 * 1) there is a lock recursion from perf_event_exit_task
3009 * see the comment there.
3011 * 2) there is a lock-inversion with mmap_sem through
3012 * perf_event_read_group(), which takes faults while
3013 * holding ctx->mutex, however this is called after
3014 * the last filedesc died, so there is no possibility
3015 * to trigger the AB-BA case.
3017 mutex_lock_nested(&ctx->mutex, SINGLE_DEPTH_NESTING);
3018 raw_spin_lock_irq(&ctx->lock);
3019 perf_group_detach(event);
3020 raw_spin_unlock_irq(&ctx->lock);
3021 perf_remove_from_context(event);
3022 mutex_unlock(&ctx->mutex);
3028 EXPORT_SYMBOL_GPL(perf_event_release_kernel);
3031 * Called when the last reference to the file is gone.
3033 static void put_event(struct perf_event *event)
3035 struct task_struct *owner;
3037 if (!atomic_long_dec_and_test(&event->refcount))
3041 owner = ACCESS_ONCE(event->owner);
3043 * Matches the smp_wmb() in perf_event_exit_task(). If we observe
3044 * !owner it means the list deletion is complete and we can indeed
3045 * free this event, otherwise we need to serialize on
3046 * owner->perf_event_mutex.
3048 smp_read_barrier_depends();
3051 * Since delayed_put_task_struct() also drops the last
3052 * task reference we can safely take a new reference
3053 * while holding the rcu_read_lock().
3055 get_task_struct(owner);
3060 mutex_lock(&owner->perf_event_mutex);
3062 * We have to re-check the event->owner field, if it is cleared
3063 * we raced with perf_event_exit_task(), acquiring the mutex
3064 * ensured they're done, and we can proceed with freeing the
3068 list_del_init(&event->owner_entry);
3069 mutex_unlock(&owner->perf_event_mutex);
3070 put_task_struct(owner);
3073 perf_event_release_kernel(event);
3076 static int perf_release(struct inode *inode, struct file *file)
3078 put_event(file->private_data);
3082 u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running)
3084 struct perf_event *child;
3090 mutex_lock(&event->child_mutex);
3091 total += perf_event_read(event);
3092 *enabled += event->total_time_enabled +
3093 atomic64_read(&event->child_total_time_enabled);
3094 *running += event->total_time_running +
3095 atomic64_read(&event->child_total_time_running);
3097 list_for_each_entry(child, &event->child_list, child_list) {
3098 total += perf_event_read(child);
3099 *enabled += child->total_time_enabled;
3100 *running += child->total_time_running;
3102 mutex_unlock(&event->child_mutex);
3106 EXPORT_SYMBOL_GPL(perf_event_read_value);
3108 static int perf_event_read_group(struct perf_event *event,
3109 u64 read_format, char __user *buf)
3111 struct perf_event *leader = event->group_leader, *sub;
3112 int n = 0, size = 0, ret = -EFAULT;
3113 struct perf_event_context *ctx = leader->ctx;
3115 u64 count, enabled, running;
3117 mutex_lock(&ctx->mutex);
3118 count = perf_event_read_value(leader, &enabled, &running);
3120 values[n++] = 1 + leader->nr_siblings;
3121 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3122 values[n++] = enabled;
3123 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3124 values[n++] = running;
3125 values[n++] = count;
3126 if (read_format & PERF_FORMAT_ID)
3127 values[n++] = primary_event_id(leader);
3129 size = n * sizeof(u64);
3131 if (copy_to_user(buf, values, size))
3136 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
3139 values[n++] = perf_event_read_value(sub, &enabled, &running);
3140 if (read_format & PERF_FORMAT_ID)
3141 values[n++] = primary_event_id(sub);
3143 size = n * sizeof(u64);
3145 if (copy_to_user(buf + ret, values, size)) {
3153 mutex_unlock(&ctx->mutex);
3158 static int perf_event_read_one(struct perf_event *event,
3159 u64 read_format, char __user *buf)
3161 u64 enabled, running;
3165 values[n++] = perf_event_read_value(event, &enabled, &running);
3166 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3167 values[n++] = enabled;
3168 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3169 values[n++] = running;
3170 if (read_format & PERF_FORMAT_ID)
3171 values[n++] = primary_event_id(event);
3173 if (copy_to_user(buf, values, n * sizeof(u64)))
3176 return n * sizeof(u64);
3180 * Read the performance event - simple non blocking version for now
3183 perf_read_hw(struct perf_event *event, char __user *buf, size_t count)
3185 u64 read_format = event->attr.read_format;
3189 * Return end-of-file for a read on a event that is in
3190 * error state (i.e. because it was pinned but it couldn't be
3191 * scheduled on to the CPU at some point).
3193 if (event->state == PERF_EVENT_STATE_ERROR)
3196 if (count < event->read_size)
3199 WARN_ON_ONCE(event->ctx->parent_ctx);
3200 if (read_format & PERF_FORMAT_GROUP)
3201 ret = perf_event_read_group(event, read_format, buf);
3203 ret = perf_event_read_one(event, read_format, buf);
3209 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
3211 struct perf_event *event = file->private_data;
3213 return perf_read_hw(event, buf, count);
3216 static unsigned int perf_poll(struct file *file, poll_table *wait)
3218 struct perf_event *event = file->private_data;
3219 struct ring_buffer *rb;
3220 unsigned int events = POLL_HUP;
3223 * Pin the event->rb by taking event->mmap_mutex; otherwise
3224 * perf_event_set_output() can swizzle our rb and make us miss wakeups.
3226 mutex_lock(&event->mmap_mutex);
3229 events = atomic_xchg(&rb->poll, 0);
3230 mutex_unlock(&event->mmap_mutex);
3232 poll_wait(file, &event->waitq, wait);
3237 static void perf_event_reset(struct perf_event *event)
3239 (void)perf_event_read(event);
3240 local64_set(&event->count, 0);
3241 perf_event_update_userpage(event);
3245 * Holding the top-level event's child_mutex means that any
3246 * descendant process that has inherited this event will block
3247 * in sync_child_event if it goes to exit, thus satisfying the
3248 * task existence requirements of perf_event_enable/disable.
3250 static void perf_event_for_each_child(struct perf_event *event,
3251 void (*func)(struct perf_event *))
3253 struct perf_event *child;
3255 WARN_ON_ONCE(event->ctx->parent_ctx);
3256 mutex_lock(&event->child_mutex);
3258 list_for_each_entry(child, &event->child_list, child_list)
3260 mutex_unlock(&event->child_mutex);
3263 static void perf_event_for_each(struct perf_event *event,
3264 void (*func)(struct perf_event *))
3266 struct perf_event_context *ctx = event->ctx;
3267 struct perf_event *sibling;
3269 WARN_ON_ONCE(ctx->parent_ctx);
3270 mutex_lock(&ctx->mutex);
3271 event = event->group_leader;
3273 perf_event_for_each_child(event, func);
3274 list_for_each_entry(sibling, &event->sibling_list, group_entry)
3275 perf_event_for_each_child(sibling, func);
3276 mutex_unlock(&ctx->mutex);
3279 static int perf_event_period(struct perf_event *event, u64 __user *arg)
3281 struct perf_event_context *ctx = event->ctx;
3285 if (!is_sampling_event(event))
3288 if (copy_from_user(&value, arg, sizeof(value)))
3294 raw_spin_lock_irq(&ctx->lock);
3295 if (event->attr.freq) {
3296 if (value > sysctl_perf_event_sample_rate) {
3301 event->attr.sample_freq = value;
3303 event->attr.sample_period = value;
3304 event->hw.sample_period = value;
3307 raw_spin_unlock_irq(&ctx->lock);
3312 static const struct file_operations perf_fops;
3314 static inline int perf_fget_light(int fd, struct fd *p)
3316 struct fd f = fdget(fd);
3320 if (f.file->f_op != &perf_fops) {
3328 static int perf_event_set_output(struct perf_event *event,
3329 struct perf_event *output_event);
3330 static int perf_event_set_filter(struct perf_event *event, void __user *arg);
3332 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
3334 struct perf_event *event = file->private_data;
3335 void (*func)(struct perf_event *);
3339 case PERF_EVENT_IOC_ENABLE:
3340 func = perf_event_enable;
3342 case PERF_EVENT_IOC_DISABLE:
3343 func = perf_event_disable;
3345 case PERF_EVENT_IOC_RESET:
3346 func = perf_event_reset;
3349 case PERF_EVENT_IOC_REFRESH:
3350 return perf_event_refresh(event, arg);
3352 case PERF_EVENT_IOC_PERIOD:
3353 return perf_event_period(event, (u64 __user *)arg);
3355 case PERF_EVENT_IOC_SET_OUTPUT:
3359 struct perf_event *output_event;
3361 ret = perf_fget_light(arg, &output);
3364 output_event = output.file->private_data;
3365 ret = perf_event_set_output(event, output_event);
3368 ret = perf_event_set_output(event, NULL);
3373 case PERF_EVENT_IOC_SET_FILTER:
3374 return perf_event_set_filter(event, (void __user *)arg);
3380 if (flags & PERF_IOC_FLAG_GROUP)
3381 perf_event_for_each(event, func);
3383 perf_event_for_each_child(event, func);
3388 int perf_event_task_enable(void)
3390 struct perf_event *event;
3392 mutex_lock(¤t->perf_event_mutex);
3393 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
3394 perf_event_for_each_child(event, perf_event_enable);
3395 mutex_unlock(¤t->perf_event_mutex);
3400 int perf_event_task_disable(void)
3402 struct perf_event *event;
3404 mutex_lock(¤t->perf_event_mutex);
3405 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
3406 perf_event_for_each_child(event, perf_event_disable);
3407 mutex_unlock(¤t->perf_event_mutex);
3412 static int perf_event_index(struct perf_event *event)
3414 if (event->hw.state & PERF_HES_STOPPED)
3417 if (event->state != PERF_EVENT_STATE_ACTIVE)
3420 return event->pmu->event_idx(event);
3423 static void calc_timer_values(struct perf_event *event,
3430 *now = perf_clock();
3431 ctx_time = event->shadow_ctx_time + *now;
3432 *enabled = ctx_time - event->tstamp_enabled;
3433 *running = ctx_time - event->tstamp_running;
3436 void __weak arch_perf_update_userpage(struct perf_event_mmap_page *userpg, u64 now)
3441 * Callers need to ensure there can be no nesting of this function, otherwise
3442 * the seqlock logic goes bad. We can not serialize this because the arch
3443 * code calls this from NMI context.
3445 void perf_event_update_userpage(struct perf_event *event)
3447 struct perf_event_mmap_page *userpg;
3448 struct ring_buffer *rb;
3449 u64 enabled, running, now;
3453 * compute total_time_enabled, total_time_running
3454 * based on snapshot values taken when the event
3455 * was last scheduled in.
3457 * we cannot simply called update_context_time()
3458 * because of locking issue as we can be called in
3461 calc_timer_values(event, &now, &enabled, &running);
3462 rb = rcu_dereference(event->rb);
3466 userpg = rb->user_page;
3469 * Disable preemption so as to not let the corresponding user-space
3470 * spin too long if we get preempted.
3475 userpg->index = perf_event_index(event);
3476 userpg->offset = perf_event_count(event);
3478 userpg->offset -= local64_read(&event->hw.prev_count);
3480 userpg->time_enabled = enabled +
3481 atomic64_read(&event->child_total_time_enabled);
3483 userpg->time_running = running +
3484 atomic64_read(&event->child_total_time_running);
3486 arch_perf_update_userpage(userpg, now);
3495 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
3497 struct perf_event *event = vma->vm_file->private_data;
3498 struct ring_buffer *rb;
3499 int ret = VM_FAULT_SIGBUS;
3501 if (vmf->flags & FAULT_FLAG_MKWRITE) {
3502 if (vmf->pgoff == 0)
3508 rb = rcu_dereference(event->rb);
3512 if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE))
3515 vmf->page = perf_mmap_to_page(rb, vmf->pgoff);
3519 get_page(vmf->page);
3520 vmf->page->mapping = vma->vm_file->f_mapping;
3521 vmf->page->index = vmf->pgoff;
3530 static void ring_buffer_attach(struct perf_event *event,
3531 struct ring_buffer *rb)
3533 unsigned long flags;
3535 if (!list_empty(&event->rb_entry))
3538 spin_lock_irqsave(&rb->event_lock, flags);
3539 if (list_empty(&event->rb_entry))
3540 list_add(&event->rb_entry, &rb->event_list);
3541 spin_unlock_irqrestore(&rb->event_lock, flags);
3544 static void ring_buffer_detach(struct perf_event *event, struct ring_buffer *rb)
3546 unsigned long flags;
3548 if (list_empty(&event->rb_entry))
3551 spin_lock_irqsave(&rb->event_lock, flags);
3552 list_del_init(&event->rb_entry);
3553 wake_up_all(&event->waitq);
3554 spin_unlock_irqrestore(&rb->event_lock, flags);
3557 static void ring_buffer_wakeup(struct perf_event *event)
3559 struct ring_buffer *rb;
3562 rb = rcu_dereference(event->rb);
3564 list_for_each_entry_rcu(event, &rb->event_list, rb_entry)
3565 wake_up_all(&event->waitq);
3570 static void rb_free_rcu(struct rcu_head *rcu_head)
3572 struct ring_buffer *rb;
3574 rb = container_of(rcu_head, struct ring_buffer, rcu_head);
3578 static struct ring_buffer *ring_buffer_get(struct perf_event *event)
3580 struct ring_buffer *rb;
3583 rb = rcu_dereference(event->rb);
3585 if (!atomic_inc_not_zero(&rb->refcount))
3593 static void ring_buffer_put(struct ring_buffer *rb)
3595 if (!atomic_dec_and_test(&rb->refcount))
3598 WARN_ON_ONCE(!list_empty(&rb->event_list));
3600 call_rcu(&rb->rcu_head, rb_free_rcu);
3603 static void perf_mmap_open(struct vm_area_struct *vma)
3605 struct perf_event *event = vma->vm_file->private_data;
3607 atomic_inc(&event->mmap_count);
3608 atomic_inc(&event->rb->mmap_count);
3612 * A buffer can be mmap()ed multiple times; either directly through the same
3613 * event, or through other events by use of perf_event_set_output().
3615 * In order to undo the VM accounting done by perf_mmap() we need to destroy
3616 * the buffer here, where we still have a VM context. This means we need
3617 * to detach all events redirecting to us.
3619 static void perf_mmap_close(struct vm_area_struct *vma)
3621 struct perf_event *event = vma->vm_file->private_data;
3623 struct ring_buffer *rb = event->rb;
3624 struct user_struct *mmap_user = rb->mmap_user;
3625 int mmap_locked = rb->mmap_locked;
3626 unsigned long size = perf_data_size(rb);
3628 atomic_dec(&rb->mmap_count);
3630 if (!atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex))
3633 /* Detach current event from the buffer. */
3634 rcu_assign_pointer(event->rb, NULL);
3635 ring_buffer_detach(event, rb);
3636 mutex_unlock(&event->mmap_mutex);
3638 /* If there's still other mmap()s of this buffer, we're done. */
3639 if (atomic_read(&rb->mmap_count)) {
3640 ring_buffer_put(rb); /* can't be last */
3645 * No other mmap()s, detach from all other events that might redirect
3646 * into the now unreachable buffer. Somewhat complicated by the
3647 * fact that rb::event_lock otherwise nests inside mmap_mutex.
3651 list_for_each_entry_rcu(event, &rb->event_list, rb_entry) {
3652 if (!atomic_long_inc_not_zero(&event->refcount)) {
3654 * This event is en-route to free_event() which will
3655 * detach it and remove it from the list.
3661 mutex_lock(&event->mmap_mutex);
3663 * Check we didn't race with perf_event_set_output() which can
3664 * swizzle the rb from under us while we were waiting to
3665 * acquire mmap_mutex.
3667 * If we find a different rb; ignore this event, a next
3668 * iteration will no longer find it on the list. We have to
3669 * still restart the iteration to make sure we're not now
3670 * iterating the wrong list.
3672 if (event->rb == rb) {
3673 rcu_assign_pointer(event->rb, NULL);
3674 ring_buffer_detach(event, rb);
3675 ring_buffer_put(rb); /* can't be last, we still have one */
3677 mutex_unlock(&event->mmap_mutex);
3681 * Restart the iteration; either we're on the wrong list or
3682 * destroyed its integrity by doing a deletion.
3689 * It could be there's still a few 0-ref events on the list; they'll
3690 * get cleaned up by free_event() -- they'll also still have their
3691 * ref on the rb and will free it whenever they are done with it.
3693 * Aside from that, this buffer is 'fully' detached and unmapped,
3694 * undo the VM accounting.
3697 atomic_long_sub((size >> PAGE_SHIFT) + 1, &mmap_user->locked_vm);
3698 vma->vm_mm->pinned_vm -= mmap_locked;
3699 free_uid(mmap_user);
3701 ring_buffer_put(rb); /* could be last */
3704 static const struct vm_operations_struct perf_mmap_vmops = {
3705 .open = perf_mmap_open,
3706 .close = perf_mmap_close,
3707 .fault = perf_mmap_fault,
3708 .page_mkwrite = perf_mmap_fault,
3711 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
3713 struct perf_event *event = file->private_data;
3714 unsigned long user_locked, user_lock_limit;
3715 struct user_struct *user = current_user();
3716 unsigned long locked, lock_limit;
3717 struct ring_buffer *rb;
3718 unsigned long vma_size;
3719 unsigned long nr_pages;
3720 long user_extra, extra;
3721 int ret = 0, flags = 0;
3724 * Don't allow mmap() of inherited per-task counters. This would
3725 * create a performance issue due to all children writing to the
3728 if (event->cpu == -1 && event->attr.inherit)
3731 if (!(vma->vm_flags & VM_SHARED))
3734 vma_size = vma->vm_end - vma->vm_start;
3735 nr_pages = (vma_size / PAGE_SIZE) - 1;
3738 * If we have rb pages ensure they're a power-of-two number, so we
3739 * can do bitmasks instead of modulo.
3741 if (nr_pages != 0 && !is_power_of_2(nr_pages))
3744 if (vma_size != PAGE_SIZE * (1 + nr_pages))
3747 if (vma->vm_pgoff != 0)
3750 WARN_ON_ONCE(event->ctx->parent_ctx);
3752 mutex_lock(&event->mmap_mutex);
3754 if (event->rb->nr_pages != nr_pages) {
3759 if (!atomic_inc_not_zero(&event->rb->mmap_count)) {
3761 * Raced against perf_mmap_close() through
3762 * perf_event_set_output(). Try again, hope for better
3765 mutex_unlock(&event->mmap_mutex);
3772 user_extra = nr_pages + 1;
3773 user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10);
3776 * Increase the limit linearly with more CPUs:
3778 user_lock_limit *= num_online_cpus();
3780 user_locked = atomic_long_read(&user->locked_vm) + user_extra;
3783 if (user_locked > user_lock_limit)
3784 extra = user_locked - user_lock_limit;
3786 lock_limit = rlimit(RLIMIT_MEMLOCK);
3787 lock_limit >>= PAGE_SHIFT;
3788 locked = vma->vm_mm->pinned_vm + extra;
3790 if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() &&
3791 !capable(CAP_IPC_LOCK)) {
3798 if (vma->vm_flags & VM_WRITE)
3799 flags |= RING_BUFFER_WRITABLE;
3801 rb = rb_alloc(nr_pages,
3802 event->attr.watermark ? event->attr.wakeup_watermark : 0,
3810 atomic_set(&rb->mmap_count, 1);
3811 rb->mmap_locked = extra;
3812 rb->mmap_user = get_current_user();
3814 atomic_long_add(user_extra, &user->locked_vm);
3815 vma->vm_mm->pinned_vm += extra;
3817 ring_buffer_attach(event, rb);
3818 rcu_assign_pointer(event->rb, rb);
3820 perf_event_update_userpage(event);
3824 atomic_inc(&event->mmap_count);
3825 mutex_unlock(&event->mmap_mutex);
3828 * Since pinned accounting is per vm we cannot allow fork() to copy our
3831 vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND | VM_DONTDUMP;
3832 vma->vm_ops = &perf_mmap_vmops;
3837 static int perf_fasync(int fd, struct file *filp, int on)
3839 struct inode *inode = file_inode(filp);
3840 struct perf_event *event = filp->private_data;
3843 mutex_lock(&inode->i_mutex);
3844 retval = fasync_helper(fd, filp, on, &event->fasync);
3845 mutex_unlock(&inode->i_mutex);
3853 static const struct file_operations perf_fops = {
3854 .llseek = no_llseek,
3855 .release = perf_release,
3858 .unlocked_ioctl = perf_ioctl,
3859 .compat_ioctl = perf_ioctl,
3861 .fasync = perf_fasync,
3867 * If there's data, ensure we set the poll() state and publish everything
3868 * to user-space before waking everybody up.
3871 void perf_event_wakeup(struct perf_event *event)
3873 ring_buffer_wakeup(event);
3875 if (event->pending_kill) {
3876 kill_fasync(&event->fasync, SIGIO, event->pending_kill);
3877 event->pending_kill = 0;
3881 static void perf_pending_event(struct irq_work *entry)
3883 struct perf_event *event = container_of(entry,
3884 struct perf_event, pending);
3886 if (event->pending_disable) {
3887 event->pending_disable = 0;
3888 __perf_event_disable(event);
3891 if (event->pending_wakeup) {
3892 event->pending_wakeup = 0;
3893 perf_event_wakeup(event);
3898 * We assume there is only KVM supporting the callbacks.
3899 * Later on, we might change it to a list if there is
3900 * another virtualization implementation supporting the callbacks.
3902 struct perf_guest_info_callbacks *perf_guest_cbs;
3904 int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
3906 perf_guest_cbs = cbs;
3909 EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks);
3911 int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
3913 perf_guest_cbs = NULL;
3916 EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks);
3919 perf_output_sample_regs(struct perf_output_handle *handle,
3920 struct pt_regs *regs, u64 mask)
3924 for_each_set_bit(bit, (const unsigned long *) &mask,
3925 sizeof(mask) * BITS_PER_BYTE) {
3928 val = perf_reg_value(regs, bit);
3929 perf_output_put(handle, val);
3933 static void perf_sample_regs_user(struct perf_regs_user *regs_user,
3934 struct pt_regs *regs)
3936 if (!user_mode(regs)) {
3938 regs = task_pt_regs(current);
3944 regs_user->regs = regs;
3945 regs_user->abi = perf_reg_abi(current);
3950 * Get remaining task size from user stack pointer.
3952 * It'd be better to take stack vma map and limit this more
3953 * precisly, but there's no way to get it safely under interrupt,
3954 * so using TASK_SIZE as limit.
3956 static u64 perf_ustack_task_size(struct pt_regs *regs)
3958 unsigned long addr = perf_user_stack_pointer(regs);
3960 if (!addr || addr >= TASK_SIZE)
3963 return TASK_SIZE - addr;
3967 perf_sample_ustack_size(u16 stack_size, u16 header_size,
3968 struct pt_regs *regs)
3972 /* No regs, no stack pointer, no dump. */
3977 * Check if we fit in with the requested stack size into the:
3979 * If we don't, we limit the size to the TASK_SIZE.
3981 * - remaining sample size
3982 * If we don't, we customize the stack size to
3983 * fit in to the remaining sample size.
3986 task_size = min((u64) USHRT_MAX, perf_ustack_task_size(regs));
3987 stack_size = min(stack_size, (u16) task_size);
3989 /* Current header size plus static size and dynamic size. */
3990 header_size += 2 * sizeof(u64);
3992 /* Do we fit in with the current stack dump size? */
3993 if ((u16) (header_size + stack_size) < header_size) {
3995 * If we overflow the maximum size for the sample,
3996 * we customize the stack dump size to fit in.
3998 stack_size = USHRT_MAX - header_size - sizeof(u64);
3999 stack_size = round_up(stack_size, sizeof(u64));
4006 perf_output_sample_ustack(struct perf_output_handle *handle, u64 dump_size,
4007 struct pt_regs *regs)
4009 /* Case of a kernel thread, nothing to dump */
4012 perf_output_put(handle, size);
4021 * - the size requested by user or the best one we can fit
4022 * in to the sample max size
4024 * - user stack dump data
4026 * - the actual dumped size
4030 perf_output_put(handle, dump_size);
4033 sp = perf_user_stack_pointer(regs);
4034 rem = __output_copy_user(handle, (void *) sp, dump_size);
4035 dyn_size = dump_size - rem;
4037 perf_output_skip(handle, rem);
4040 perf_output_put(handle, dyn_size);
4044 static void __perf_event_header__init_id(struct perf_event_header *header,
4045 struct perf_sample_data *data,
4046 struct perf_event *event)
4048 u64 sample_type = event->attr.sample_type;
4050 data->type = sample_type;
4051 header->size += event->id_header_size;
4053 if (sample_type & PERF_SAMPLE_TID) {
4054 /* namespace issues */
4055 data->tid_entry.pid = perf_event_pid(event, current);
4056 data->tid_entry.tid = perf_event_tid(event, current);
4059 if (sample_type & PERF_SAMPLE_TIME)
4060 data->time = perf_clock();
4062 if (sample_type & PERF_SAMPLE_ID)
4063 data->id = primary_event_id(event);
4065 if (sample_type & PERF_SAMPLE_STREAM_ID)
4066 data->stream_id = event->id;
4068 if (sample_type & PERF_SAMPLE_CPU) {
4069 data->cpu_entry.cpu = raw_smp_processor_id();
4070 data->cpu_entry.reserved = 0;
4074 void perf_event_header__init_id(struct perf_event_header *header,
4075 struct perf_sample_data *data,
4076 struct perf_event *event)
4078 if (event->attr.sample_id_all)
4079 __perf_event_header__init_id(header, data, event);
4082 static void __perf_event__output_id_sample(struct perf_output_handle *handle,
4083 struct perf_sample_data *data)
4085 u64 sample_type = data->type;
4087 if (sample_type & PERF_SAMPLE_TID)
4088 perf_output_put(handle, data->tid_entry);
4090 if (sample_type & PERF_SAMPLE_TIME)
4091 perf_output_put(handle, data->time);
4093 if (sample_type & PERF_SAMPLE_ID)
4094 perf_output_put(handle, data->id);
4096 if (sample_type & PERF_SAMPLE_STREAM_ID)
4097 perf_output_put(handle, data->stream_id);
4099 if (sample_type & PERF_SAMPLE_CPU)
4100 perf_output_put(handle, data->cpu_entry);
4103 void perf_event__output_id_sample(struct perf_event *event,
4104 struct perf_output_handle *handle,
4105 struct perf_sample_data *sample)
4107 if (event->attr.sample_id_all)
4108 __perf_event__output_id_sample(handle, sample);
4111 static void perf_output_read_one(struct perf_output_handle *handle,
4112 struct perf_event *event,
4113 u64 enabled, u64 running)
4115 u64 read_format = event->attr.read_format;
4119 values[n++] = perf_event_count(event);
4120 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
4121 values[n++] = enabled +
4122 atomic64_read(&event->child_total_time_enabled);
4124 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
4125 values[n++] = running +
4126 atomic64_read(&event->child_total_time_running);
4128 if (read_format & PERF_FORMAT_ID)
4129 values[n++] = primary_event_id(event);
4131 __output_copy(handle, values, n * sizeof(u64));
4135 * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
4137 static void perf_output_read_group(struct perf_output_handle *handle,
4138 struct perf_event *event,
4139 u64 enabled, u64 running)
4141 struct perf_event *leader = event->group_leader, *sub;
4142 u64 read_format = event->attr.read_format;
4146 values[n++] = 1 + leader->nr_siblings;
4148 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
4149 values[n++] = enabled;
4151 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
4152 values[n++] = running;
4154 if (leader != event)
4155 leader->pmu->read(leader);
4157 values[n++] = perf_event_count(leader);
4158 if (read_format & PERF_FORMAT_ID)
4159 values[n++] = primary_event_id(leader);
4161 __output_copy(handle, values, n * sizeof(u64));
4163 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
4167 sub->pmu->read(sub);
4169 values[n++] = perf_event_count(sub);
4170 if (read_format & PERF_FORMAT_ID)
4171 values[n++] = primary_event_id(sub);
4173 __output_copy(handle, values, n * sizeof(u64));
4177 #define PERF_FORMAT_TOTAL_TIMES (PERF_FORMAT_TOTAL_TIME_ENABLED|\
4178 PERF_FORMAT_TOTAL_TIME_RUNNING)
4180 static void perf_output_read(struct perf_output_handle *handle,
4181 struct perf_event *event)
4183 u64 enabled = 0, running = 0, now;
4184 u64 read_format = event->attr.read_format;
4187 * compute total_time_enabled, total_time_running
4188 * based on snapshot values taken when the event
4189 * was last scheduled in.
4191 * we cannot simply called update_context_time()
4192 * because of locking issue as we are called in
4195 if (read_format & PERF_FORMAT_TOTAL_TIMES)
4196 calc_timer_values(event, &now, &enabled, &running);
4198 if (event->attr.read_format & PERF_FORMAT_GROUP)
4199 perf_output_read_group(handle, event, enabled, running);
4201 perf_output_read_one(handle, event, enabled, running);
4204 void perf_output_sample(struct perf_output_handle *handle,
4205 struct perf_event_header *header,
4206 struct perf_sample_data *data,
4207 struct perf_event *event)
4209 u64 sample_type = data->type;
4211 perf_output_put(handle, *header);
4213 if (sample_type & PERF_SAMPLE_IP)
4214 perf_output_put(handle, data->ip);
4216 if (sample_type & PERF_SAMPLE_TID)
4217 perf_output_put(handle, data->tid_entry);
4219 if (sample_type & PERF_SAMPLE_TIME)
4220 perf_output_put(handle, data->time);
4222 if (sample_type & PERF_SAMPLE_ADDR)
4223 perf_output_put(handle, data->addr);
4225 if (sample_type & PERF_SAMPLE_ID)
4226 perf_output_put(handle, data->id);
4228 if (sample_type & PERF_SAMPLE_STREAM_ID)
4229 perf_output_put(handle, data->stream_id);
4231 if (sample_type & PERF_SAMPLE_CPU)
4232 perf_output_put(handle, data->cpu_entry);
4234 if (sample_type & PERF_SAMPLE_PERIOD)
4235 perf_output_put(handle, data->period);
4237 if (sample_type & PERF_SAMPLE_READ)
4238 perf_output_read(handle, event);
4240 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
4241 if (data->callchain) {
4244 if (data->callchain)
4245 size += data->callchain->nr;
4247 size *= sizeof(u64);
4249 __output_copy(handle, data->callchain, size);
4252 perf_output_put(handle, nr);
4256 if (sample_type & PERF_SAMPLE_RAW) {
4258 perf_output_put(handle, data->raw->size);
4259 __output_copy(handle, data->raw->data,
4266 .size = sizeof(u32),
4269 perf_output_put(handle, raw);
4273 if (!event->attr.watermark) {
4274 int wakeup_events = event->attr.wakeup_events;
4276 if (wakeup_events) {
4277 struct ring_buffer *rb = handle->rb;
4278 int events = local_inc_return(&rb->events);
4280 if (events >= wakeup_events) {
4281 local_sub(wakeup_events, &rb->events);
4282 local_inc(&rb->wakeup);
4287 if (sample_type & PERF_SAMPLE_BRANCH_STACK) {
4288 if (data->br_stack) {
4291 size = data->br_stack->nr
4292 * sizeof(struct perf_branch_entry);
4294 perf_output_put(handle, data->br_stack->nr);
4295 perf_output_copy(handle, data->br_stack->entries, size);
4298 * we always store at least the value of nr
4301 perf_output_put(handle, nr);
4305 if (sample_type & PERF_SAMPLE_REGS_USER) {
4306 u64 abi = data->regs_user.abi;
4309 * If there are no regs to dump, notice it through
4310 * first u64 being zero (PERF_SAMPLE_REGS_ABI_NONE).
4312 perf_output_put(handle, abi);
4315 u64 mask = event->attr.sample_regs_user;
4316 perf_output_sample_regs(handle,
4317 data->regs_user.regs,
4322 if (sample_type & PERF_SAMPLE_STACK_USER)
4323 perf_output_sample_ustack(handle,
4324 data->stack_user_size,
4325 data->regs_user.regs);
4327 if (sample_type & PERF_SAMPLE_WEIGHT)
4328 perf_output_put(handle, data->weight);
4330 if (sample_type & PERF_SAMPLE_DATA_SRC)
4331 perf_output_put(handle, data->data_src.val);
4334 void perf_prepare_sample(struct perf_event_header *header,
4335 struct perf_sample_data *data,
4336 struct perf_event *event,
4337 struct pt_regs *regs)
4339 u64 sample_type = event->attr.sample_type;
4341 header->type = PERF_RECORD_SAMPLE;
4342 header->size = sizeof(*header) + event->header_size;
4345 header->misc |= perf_misc_flags(regs);
4347 __perf_event_header__init_id(header, data, event);
4349 if (sample_type & PERF_SAMPLE_IP)
4350 data->ip = perf_instruction_pointer(regs);
4352 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
4355 data->callchain = perf_callchain(event, regs);
4357 if (data->callchain)
4358 size += data->callchain->nr;
4360 header->size += size * sizeof(u64);
4363 if (sample_type & PERF_SAMPLE_RAW) {
4364 int size = sizeof(u32);
4367 size += data->raw->size;
4369 size += sizeof(u32);
4371 WARN_ON_ONCE(size & (sizeof(u64)-1));
4372 header->size += size;
4375 if (sample_type & PERF_SAMPLE_BRANCH_STACK) {
4376 int size = sizeof(u64); /* nr */
4377 if (data->br_stack) {
4378 size += data->br_stack->nr
4379 * sizeof(struct perf_branch_entry);
4381 header->size += size;
4384 if (sample_type & PERF_SAMPLE_REGS_USER) {
4385 /* regs dump ABI info */
4386 int size = sizeof(u64);
4388 perf_sample_regs_user(&data->regs_user, regs);
4390 if (data->regs_user.regs) {
4391 u64 mask = event->attr.sample_regs_user;
4392 size += hweight64(mask) * sizeof(u64);
4395 header->size += size;
4398 if (sample_type & PERF_SAMPLE_STACK_USER) {
4400 * Either we need PERF_SAMPLE_STACK_USER bit to be allways
4401 * processed as the last one or have additional check added
4402 * in case new sample type is added, because we could eat
4403 * up the rest of the sample size.
4405 struct perf_regs_user *uregs = &data->regs_user;
4406 u16 stack_size = event->attr.sample_stack_user;
4407 u16 size = sizeof(u64);
4410 perf_sample_regs_user(uregs, regs);
4412 stack_size = perf_sample_ustack_size(stack_size, header->size,
4416 * If there is something to dump, add space for the dump
4417 * itself and for the field that tells the dynamic size,
4418 * which is how many have been actually dumped.
4421 size += sizeof(u64) + stack_size;
4423 data->stack_user_size = stack_size;
4424 header->size += size;
4428 static void perf_event_output(struct perf_event *event,
4429 struct perf_sample_data *data,
4430 struct pt_regs *regs)
4432 struct perf_output_handle handle;
4433 struct perf_event_header header;
4435 /* protect the callchain buffers */
4438 perf_prepare_sample(&header, data, event, regs);
4440 if (perf_output_begin(&handle, event, header.size))
4443 perf_output_sample(&handle, &header, data, event);
4445 perf_output_end(&handle);
4455 struct perf_read_event {
4456 struct perf_event_header header;
4463 perf_event_read_event(struct perf_event *event,
4464 struct task_struct *task)
4466 struct perf_output_handle handle;
4467 struct perf_sample_data sample;
4468 struct perf_read_event read_event = {
4470 .type = PERF_RECORD_READ,
4472 .size = sizeof(read_event) + event->read_size,
4474 .pid = perf_event_pid(event, task),
4475 .tid = perf_event_tid(event, task),
4479 perf_event_header__init_id(&read_event.header, &sample, event);
4480 ret = perf_output_begin(&handle, event, read_event.header.size);
4484 perf_output_put(&handle, read_event);
4485 perf_output_read(&handle, event);
4486 perf_event__output_id_sample(event, &handle, &sample);
4488 perf_output_end(&handle);
4491 typedef int (perf_event_aux_match_cb)(struct perf_event *event, void *data);
4492 typedef void (perf_event_aux_output_cb)(struct perf_event *event, void *data);
4495 perf_event_aux_ctx(struct perf_event_context *ctx,
4496 perf_event_aux_match_cb match,
4497 perf_event_aux_output_cb output,
4500 struct perf_event *event;
4502 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
4503 if (event->state < PERF_EVENT_STATE_INACTIVE)
4505 if (!event_filter_match(event))
4507 if (match(event, data))
4508 output(event, data);
4513 perf_event_aux(perf_event_aux_match_cb match,
4514 perf_event_aux_output_cb output,
4516 struct perf_event_context *task_ctx)
4518 struct perf_cpu_context *cpuctx;
4519 struct perf_event_context *ctx;
4524 list_for_each_entry_rcu(pmu, &pmus, entry) {
4525 cpuctx = get_cpu_ptr(pmu->pmu_cpu_context);
4526 if (cpuctx->unique_pmu != pmu)
4528 perf_event_aux_ctx(&cpuctx->ctx, match, output, data);
4531 ctxn = pmu->task_ctx_nr;
4534 ctx = rcu_dereference(current->perf_event_ctxp[ctxn]);
4536 perf_event_aux_ctx(ctx, match, output, data);
4538 put_cpu_ptr(pmu->pmu_cpu_context);
4543 perf_event_aux_ctx(task_ctx, match, output, data);
4550 * task tracking -- fork/exit
4552 * enabled by: attr.comm | attr.mmap | attr.mmap_data | attr.task
4555 struct perf_task_event {
4556 struct task_struct *task;
4557 struct perf_event_context *task_ctx;
4560 struct perf_event_header header;
4570 static void perf_event_task_output(struct perf_event *event,
4573 struct perf_task_event *task_event = data;
4574 struct perf_output_handle handle;
4575 struct perf_sample_data sample;
4576 struct task_struct *task = task_event->task;
4577 int ret, size = task_event->event_id.header.size;
4579 perf_event_header__init_id(&task_event->event_id.header, &sample, event);
4581 ret = perf_output_begin(&handle, event,
4582 task_event->event_id.header.size);
4586 task_event->event_id.pid = perf_event_pid(event, task);
4587 task_event->event_id.ppid = perf_event_pid(event, current);
4589 task_event->event_id.tid = perf_event_tid(event, task);
4590 task_event->event_id.ptid = perf_event_tid(event, current);
4592 perf_output_put(&handle, task_event->event_id);
4594 perf_event__output_id_sample(event, &handle, &sample);
4596 perf_output_end(&handle);
4598 task_event->event_id.header.size = size;
4601 static int perf_event_task_match(struct perf_event *event,
4602 void *data __maybe_unused)
4604 return event->attr.comm || event->attr.mmap ||
4605 event->attr.mmap_data || event->attr.task;
4608 static void perf_event_task(struct task_struct *task,
4609 struct perf_event_context *task_ctx,
4612 struct perf_task_event task_event;
4614 if (!atomic_read(&nr_comm_events) &&
4615 !atomic_read(&nr_mmap_events) &&
4616 !atomic_read(&nr_task_events))
4619 task_event = (struct perf_task_event){
4621 .task_ctx = task_ctx,
4624 .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT,
4626 .size = sizeof(task_event.event_id),
4632 .time = perf_clock(),
4636 perf_event_aux(perf_event_task_match,
4637 perf_event_task_output,
4642 void perf_event_fork(struct task_struct *task)
4644 perf_event_task(task, NULL, 1);
4651 struct perf_comm_event {
4652 struct task_struct *task;
4657 struct perf_event_header header;
4664 static void perf_event_comm_output(struct perf_event *event,
4667 struct perf_comm_event *comm_event = data;
4668 struct perf_output_handle handle;
4669 struct perf_sample_data sample;
4670 int size = comm_event->event_id.header.size;
4673 perf_event_header__init_id(&comm_event->event_id.header, &sample, event);
4674 ret = perf_output_begin(&handle, event,
4675 comm_event->event_id.header.size);
4680 comm_event->event_id.pid = perf_event_pid(event, comm_event->task);
4681 comm_event->event_id.tid = perf_event_tid(event, comm_event->task);
4683 perf_output_put(&handle, comm_event->event_id);
4684 __output_copy(&handle, comm_event->comm,
4685 comm_event->comm_size);
4687 perf_event__output_id_sample(event, &handle, &sample);
4689 perf_output_end(&handle);
4691 comm_event->event_id.header.size = size;
4694 static int perf_event_comm_match(struct perf_event *event,
4695 void *data __maybe_unused)
4697 return event->attr.comm;
4700 static void perf_event_comm_event(struct perf_comm_event *comm_event)
4702 char comm[TASK_COMM_LEN];
4705 memset(comm, 0, sizeof(comm));
4706 strlcpy(comm, comm_event->task->comm, sizeof(comm));
4707 size = ALIGN(strlen(comm)+1, sizeof(u64));
4709 comm_event->comm = comm;
4710 comm_event->comm_size = size;
4712 comm_event->event_id.header.size = sizeof(comm_event->event_id) + size;
4714 perf_event_aux(perf_event_comm_match,
4715 perf_event_comm_output,
4720 void perf_event_comm(struct task_struct *task)
4722 struct perf_comm_event comm_event;
4723 struct perf_event_context *ctx;
4727 for_each_task_context_nr(ctxn) {
4728 ctx = task->perf_event_ctxp[ctxn];
4732 perf_event_enable_on_exec(ctx);
4736 if (!atomic_read(&nr_comm_events))
4739 comm_event = (struct perf_comm_event){
4745 .type = PERF_RECORD_COMM,
4754 perf_event_comm_event(&comm_event);
4761 struct perf_mmap_event {
4762 struct vm_area_struct *vma;
4764 const char *file_name;
4768 struct perf_event_header header;
4778 static void perf_event_mmap_output(struct perf_event *event,
4781 struct perf_mmap_event *mmap_event = data;
4782 struct perf_output_handle handle;
4783 struct perf_sample_data sample;
4784 int size = mmap_event->event_id.header.size;
4787 perf_event_header__init_id(&mmap_event->event_id.header, &sample, event);
4788 ret = perf_output_begin(&handle, event,
4789 mmap_event->event_id.header.size);
4793 mmap_event->event_id.pid = perf_event_pid(event, current);
4794 mmap_event->event_id.tid = perf_event_tid(event, current);
4796 perf_output_put(&handle, mmap_event->event_id);
4797 __output_copy(&handle, mmap_event->file_name,
4798 mmap_event->file_size);
4800 perf_event__output_id_sample(event, &handle, &sample);
4802 perf_output_end(&handle);
4804 mmap_event->event_id.header.size = size;
4807 static int perf_event_mmap_match(struct perf_event *event,
4810 struct perf_mmap_event *mmap_event = data;
4811 struct vm_area_struct *vma = mmap_event->vma;
4812 int executable = vma->vm_flags & VM_EXEC;
4814 return (!executable && event->attr.mmap_data) ||
4815 (executable && event->attr.mmap);
4818 static void perf_event_mmap_event(struct perf_mmap_event *mmap_event)
4820 struct vm_area_struct *vma = mmap_event->vma;
4821 struct file *file = vma->vm_file;
4827 memset(tmp, 0, sizeof(tmp));
4831 * d_path works from the end of the rb backwards, so we
4832 * need to add enough zero bytes after the string to handle
4833 * the 64bit alignment we do later.
4835 buf = kzalloc(PATH_MAX + sizeof(u64), GFP_KERNEL);
4837 name = strncpy(tmp, "//enomem", sizeof(tmp));
4840 name = d_path(&file->f_path, buf, PATH_MAX);
4842 name = strncpy(tmp, "//toolong", sizeof(tmp));
4846 if (arch_vma_name(mmap_event->vma)) {
4847 name = strncpy(tmp, arch_vma_name(mmap_event->vma),
4849 tmp[sizeof(tmp) - 1] = '\0';
4854 name = strncpy(tmp, "[vdso]", sizeof(tmp));
4856 } else if (vma->vm_start <= vma->vm_mm->start_brk &&
4857 vma->vm_end >= vma->vm_mm->brk) {
4858 name = strncpy(tmp, "[heap]", sizeof(tmp));
4860 } else if (vma->vm_start <= vma->vm_mm->start_stack &&
4861 vma->vm_end >= vma->vm_mm->start_stack) {
4862 name = strncpy(tmp, "[stack]", sizeof(tmp));
4866 name = strncpy(tmp, "//anon", sizeof(tmp));
4871 size = ALIGN(strlen(name)+1, sizeof(u64));
4873 mmap_event->file_name = name;
4874 mmap_event->file_size = size;
4876 if (!(vma->vm_flags & VM_EXEC))
4877 mmap_event->event_id.header.misc |= PERF_RECORD_MISC_MMAP_DATA;
4879 mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size;
4881 perf_event_aux(perf_event_mmap_match,
4882 perf_event_mmap_output,
4889 void perf_event_mmap(struct vm_area_struct *vma)
4891 struct perf_mmap_event mmap_event;
4893 if (!atomic_read(&nr_mmap_events))
4896 mmap_event = (struct perf_mmap_event){
4902 .type = PERF_RECORD_MMAP,
4903 .misc = PERF_RECORD_MISC_USER,
4908 .start = vma->vm_start,
4909 .len = vma->vm_end - vma->vm_start,
4910 .pgoff = (u64)vma->vm_pgoff << PAGE_SHIFT,
4914 perf_event_mmap_event(&mmap_event);
4918 * IRQ throttle logging
4921 static void perf_log_throttle(struct perf_event *event, int enable)
4923 struct perf_output_handle handle;
4924 struct perf_sample_data sample;
4928 struct perf_event_header header;
4932 } throttle_event = {
4934 .type = PERF_RECORD_THROTTLE,
4936 .size = sizeof(throttle_event),
4938 .time = perf_clock(),
4939 .id = primary_event_id(event),
4940 .stream_id = event->id,
4944 throttle_event.header.type = PERF_RECORD_UNTHROTTLE;
4946 perf_event_header__init_id(&throttle_event.header, &sample, event);
4948 ret = perf_output_begin(&handle, event,
4949 throttle_event.header.size);
4953 perf_output_put(&handle, throttle_event);
4954 perf_event__output_id_sample(event, &handle, &sample);
4955 perf_output_end(&handle);
4959 * Generic event overflow handling, sampling.
4962 static int __perf_event_overflow(struct perf_event *event,
4963 int throttle, struct perf_sample_data *data,
4964 struct pt_regs *regs)
4966 int events = atomic_read(&event->event_limit);
4967 struct hw_perf_event *hwc = &event->hw;
4972 * Non-sampling counters might still use the PMI to fold short
4973 * hardware counters, ignore those.
4975 if (unlikely(!is_sampling_event(event)))
4978 seq = __this_cpu_read(perf_throttled_seq);
4979 if (seq != hwc->interrupts_seq) {
4980 hwc->interrupts_seq = seq;
4981 hwc->interrupts = 1;
4984 if (unlikely(throttle
4985 && hwc->interrupts >= max_samples_per_tick)) {
4986 __this_cpu_inc(perf_throttled_count);
4987 hwc->interrupts = MAX_INTERRUPTS;
4988 perf_log_throttle(event, 0);
4993 if (event->attr.freq) {
4994 u64 now = perf_clock();
4995 s64 delta = now - hwc->freq_time_stamp;
4997 hwc->freq_time_stamp = now;
4999 if (delta > 0 && delta < 2*TICK_NSEC)
5000 perf_adjust_period(event, delta, hwc->last_period, true);
5004 * XXX event_limit might not quite work as expected on inherited
5008 event->pending_kill = POLL_IN;
5009 if (events && atomic_dec_and_test(&event->event_limit)) {
5011 event->pending_kill = POLL_HUP;
5012 event->pending_disable = 1;
5013 irq_work_queue(&event->pending);
5016 if (event->overflow_handler)
5017 event->overflow_handler(event, data, regs);
5019 perf_event_output(event, data, regs);
5021 if (event->fasync && event->pending_kill) {
5022 event->pending_wakeup = 1;
5023 irq_work_queue(&event->pending);
5029 int perf_event_overflow(struct perf_event *event,
5030 struct perf_sample_data *data,
5031 struct pt_regs *regs)
5033 return __perf_event_overflow(event, 1, data, regs);
5037 * Generic software event infrastructure
5040 struct swevent_htable {
5041 struct swevent_hlist *swevent_hlist;
5042 struct mutex hlist_mutex;
5045 /* Recursion avoidance in each contexts */
5046 int recursion[PERF_NR_CONTEXTS];
5048 /* Keeps track of cpu being initialized/exited */
5052 static DEFINE_PER_CPU(struct swevent_htable, swevent_htable);
5055 * We directly increment event->count and keep a second value in
5056 * event->hw.period_left to count intervals. This period event
5057 * is kept in the range [-sample_period, 0] so that we can use the
5061 static u64 perf_swevent_set_period(struct perf_event *event)
5063 struct hw_perf_event *hwc = &event->hw;
5064 u64 period = hwc->last_period;
5068 hwc->last_period = hwc->sample_period;
5071 old = val = local64_read(&hwc->period_left);
5075 nr = div64_u64(period + val, period);
5076 offset = nr * period;
5078 if (local64_cmpxchg(&hwc->period_left, old, val) != old)
5084 static void perf_swevent_overflow(struct perf_event *event, u64 overflow,
5085 struct perf_sample_data *data,
5086 struct pt_regs *regs)
5088 struct hw_perf_event *hwc = &event->hw;
5092 overflow = perf_swevent_set_period(event);
5094 if (hwc->interrupts == MAX_INTERRUPTS)
5097 for (; overflow; overflow--) {
5098 if (__perf_event_overflow(event, throttle,
5101 * We inhibit the overflow from happening when
5102 * hwc->interrupts == MAX_INTERRUPTS.
5110 static void perf_swevent_event(struct perf_event *event, u64 nr,
5111 struct perf_sample_data *data,
5112 struct pt_regs *regs)
5114 struct hw_perf_event *hwc = &event->hw;
5116 local64_add(nr, &event->count);
5121 if (!is_sampling_event(event))
5124 if ((event->attr.sample_type & PERF_SAMPLE_PERIOD) && !event->attr.freq) {
5126 return perf_swevent_overflow(event, 1, data, regs);
5128 data->period = event->hw.last_period;
5130 if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq)
5131 return perf_swevent_overflow(event, 1, data, regs);
5133 if (local64_add_negative(nr, &hwc->period_left))
5136 perf_swevent_overflow(event, 0, data, regs);
5139 static int perf_exclude_event(struct perf_event *event,
5140 struct pt_regs *regs)
5142 if (event->hw.state & PERF_HES_STOPPED)
5146 if (event->attr.exclude_user && user_mode(regs))
5149 if (event->attr.exclude_kernel && !user_mode(regs))
5156 static int perf_swevent_match(struct perf_event *event,
5157 enum perf_type_id type,
5159 struct perf_sample_data *data,
5160 struct pt_regs *regs)
5162 if (event->attr.type != type)
5165 if (event->attr.config != event_id)
5168 if (perf_exclude_event(event, regs))
5174 static inline u64 swevent_hash(u64 type, u32 event_id)
5176 u64 val = event_id | (type << 32);
5178 return hash_64(val, SWEVENT_HLIST_BITS);
5181 static inline struct hlist_head *
5182 __find_swevent_head(struct swevent_hlist *hlist, u64 type, u32 event_id)
5184 u64 hash = swevent_hash(type, event_id);
5186 return &hlist->heads[hash];
5189 /* For the read side: events when they trigger */
5190 static inline struct hlist_head *
5191 find_swevent_head_rcu(struct swevent_htable *swhash, u64 type, u32 event_id)
5193 struct swevent_hlist *hlist;
5195 hlist = rcu_dereference(swhash->swevent_hlist);
5199 return __find_swevent_head(hlist, type, event_id);
5202 /* For the event head insertion and removal in the hlist */
5203 static inline struct hlist_head *
5204 find_swevent_head(struct swevent_htable *swhash, struct perf_event *event)
5206 struct swevent_hlist *hlist;
5207 u32 event_id = event->attr.config;
5208 u64 type = event->attr.type;
5211 * Event scheduling is always serialized against hlist allocation
5212 * and release. Which makes the protected version suitable here.
5213 * The context lock guarantees that.
5215 hlist = rcu_dereference_protected(swhash->swevent_hlist,
5216 lockdep_is_held(&event->ctx->lock));
5220 return __find_swevent_head(hlist, type, event_id);
5223 static void do_perf_sw_event(enum perf_type_id type, u32 event_id,
5225 struct perf_sample_data *data,
5226 struct pt_regs *regs)
5228 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
5229 struct perf_event *event;
5230 struct hlist_head *head;
5233 head = find_swevent_head_rcu(swhash, type, event_id);
5237 hlist_for_each_entry_rcu(event, head, hlist_entry) {
5238 if (perf_swevent_match(event, type, event_id, data, regs))
5239 perf_swevent_event(event, nr, data, regs);
5245 int perf_swevent_get_recursion_context(void)
5247 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
5249 return get_recursion_context(swhash->recursion);
5251 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context);
5253 inline void perf_swevent_put_recursion_context(int rctx)
5255 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
5257 put_recursion_context(swhash->recursion, rctx);
5260 void __perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)
5262 struct perf_sample_data data;
5265 preempt_disable_notrace();
5266 rctx = perf_swevent_get_recursion_context();
5270 perf_sample_data_init(&data, addr, 0);
5272 do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, &data, regs);
5274 perf_swevent_put_recursion_context(rctx);
5275 preempt_enable_notrace();
5278 static void perf_swevent_read(struct perf_event *event)
5282 static int perf_swevent_add(struct perf_event *event, int flags)
5284 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
5285 struct hw_perf_event *hwc = &event->hw;
5286 struct hlist_head *head;
5288 if (is_sampling_event(event)) {
5289 hwc->last_period = hwc->sample_period;
5290 perf_swevent_set_period(event);
5293 hwc->state = !(flags & PERF_EF_START);
5295 head = find_swevent_head(swhash, event);
5298 * We can race with cpu hotplug code. Do not
5299 * WARN if the cpu just got unplugged.
5301 WARN_ON_ONCE(swhash->online);
5305 hlist_add_head_rcu(&event->hlist_entry, head);
5310 static void perf_swevent_del(struct perf_event *event, int flags)
5312 hlist_del_rcu(&event->hlist_entry);
5315 static void perf_swevent_start(struct perf_event *event, int flags)
5317 event->hw.state = 0;
5320 static void perf_swevent_stop(struct perf_event *event, int flags)
5322 event->hw.state = PERF_HES_STOPPED;
5325 /* Deref the hlist from the update side */
5326 static inline struct swevent_hlist *
5327 swevent_hlist_deref(struct swevent_htable *swhash)
5329 return rcu_dereference_protected(swhash->swevent_hlist,
5330 lockdep_is_held(&swhash->hlist_mutex));
5333 static void swevent_hlist_release(struct swevent_htable *swhash)
5335 struct swevent_hlist *hlist = swevent_hlist_deref(swhash);
5340 rcu_assign_pointer(swhash->swevent_hlist, NULL);
5341 kfree_rcu(hlist, rcu_head);
5344 static void swevent_hlist_put_cpu(struct perf_event *event, int cpu)
5346 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
5348 mutex_lock(&swhash->hlist_mutex);
5350 if (!--swhash->hlist_refcount)
5351 swevent_hlist_release(swhash);
5353 mutex_unlock(&swhash->hlist_mutex);
5356 static void swevent_hlist_put(struct perf_event *event)
5360 if (event->cpu != -1) {
5361 swevent_hlist_put_cpu(event, event->cpu);
5365 for_each_possible_cpu(cpu)
5366 swevent_hlist_put_cpu(event, cpu);
5369 static int swevent_hlist_get_cpu(struct perf_event *event, int cpu)
5371 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
5374 mutex_lock(&swhash->hlist_mutex);
5376 if (!swevent_hlist_deref(swhash) && cpu_online(cpu)) {
5377 struct swevent_hlist *hlist;
5379 hlist = kzalloc(sizeof(*hlist), GFP_KERNEL);
5384 rcu_assign_pointer(swhash->swevent_hlist, hlist);
5386 swhash->hlist_refcount++;
5388 mutex_unlock(&swhash->hlist_mutex);
5393 static int swevent_hlist_get(struct perf_event *event)
5396 int cpu, failed_cpu;
5398 if (event->cpu != -1)
5399 return swevent_hlist_get_cpu(event, event->cpu);
5402 for_each_possible_cpu(cpu) {
5403 err = swevent_hlist_get_cpu(event, cpu);
5413 for_each_possible_cpu(cpu) {
5414 if (cpu == failed_cpu)
5416 swevent_hlist_put_cpu(event, cpu);
5423 struct static_key perf_swevent_enabled[PERF_COUNT_SW_MAX];
5425 static void sw_perf_event_destroy(struct perf_event *event)
5427 u64 event_id = event->attr.config;
5429 WARN_ON(event->parent);
5431 static_key_slow_dec(&perf_swevent_enabled[event_id]);
5432 swevent_hlist_put(event);
5435 static int perf_swevent_init(struct perf_event *event)
5437 u64 event_id = event->attr.config;
5439 if (event->attr.type != PERF_TYPE_SOFTWARE)
5443 * no branch sampling for software events
5445 if (has_branch_stack(event))
5449 case PERF_COUNT_SW_CPU_CLOCK:
5450 case PERF_COUNT_SW_TASK_CLOCK:
5457 if (event_id >= PERF_COUNT_SW_MAX)
5460 if (!event->parent) {
5463 err = swevent_hlist_get(event);
5467 static_key_slow_inc(&perf_swevent_enabled[event_id]);
5468 event->destroy = sw_perf_event_destroy;
5474 static int perf_swevent_event_idx(struct perf_event *event)
5479 static struct pmu perf_swevent = {
5480 .task_ctx_nr = perf_sw_context,
5482 .event_init = perf_swevent_init,
5483 .add = perf_swevent_add,
5484 .del = perf_swevent_del,
5485 .start = perf_swevent_start,
5486 .stop = perf_swevent_stop,
5487 .read = perf_swevent_read,
5489 .event_idx = perf_swevent_event_idx,
5492 #ifdef CONFIG_EVENT_TRACING
5494 static int perf_tp_filter_match(struct perf_event *event,
5495 struct perf_sample_data *data)
5497 void *record = data->raw->data;
5499 if (likely(!event->filter) || filter_match_preds(event->filter, record))
5504 static int perf_tp_event_match(struct perf_event *event,
5505 struct perf_sample_data *data,
5506 struct pt_regs *regs)
5508 if (event->hw.state & PERF_HES_STOPPED)
5511 * All tracepoints are from kernel-space.
5513 if (event->attr.exclude_kernel)
5516 if (!perf_tp_filter_match(event, data))
5522 void perf_tp_event(u64 addr, u64 count, void *record, int entry_size,
5523 struct pt_regs *regs, struct hlist_head *head, int rctx,
5524 struct task_struct *task)
5526 struct perf_sample_data data;
5527 struct perf_event *event;
5529 struct perf_raw_record raw = {
5534 perf_sample_data_init(&data, addr, 0);
5537 hlist_for_each_entry_rcu(event, head, hlist_entry) {
5538 if (perf_tp_event_match(event, &data, regs))
5539 perf_swevent_event(event, count, &data, regs);
5543 * If we got specified a target task, also iterate its context and
5544 * deliver this event there too.
5546 if (task && task != current) {
5547 struct perf_event_context *ctx;
5548 struct trace_entry *entry = record;
5551 ctx = rcu_dereference(task->perf_event_ctxp[perf_sw_context]);
5555 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
5556 if (event->attr.type != PERF_TYPE_TRACEPOINT)
5558 if (event->attr.config != entry->type)
5560 if (perf_tp_event_match(event, &data, regs))
5561 perf_swevent_event(event, count, &data, regs);
5567 perf_swevent_put_recursion_context(rctx);
5569 EXPORT_SYMBOL_GPL(perf_tp_event);
5571 static void tp_perf_event_destroy(struct perf_event *event)
5573 perf_trace_destroy(event);
5576 static int perf_tp_event_init(struct perf_event *event)
5580 if (event->attr.type != PERF_TYPE_TRACEPOINT)
5584 * no branch sampling for tracepoint events
5586 if (has_branch_stack(event))
5589 err = perf_trace_init(event);
5593 event->destroy = tp_perf_event_destroy;
5598 static struct pmu perf_tracepoint = {
5599 .task_ctx_nr = perf_sw_context,
5601 .event_init = perf_tp_event_init,
5602 .add = perf_trace_add,
5603 .del = perf_trace_del,
5604 .start = perf_swevent_start,
5605 .stop = perf_swevent_stop,
5606 .read = perf_swevent_read,
5608 .event_idx = perf_swevent_event_idx,
5611 static inline void perf_tp_register(void)
5613 perf_pmu_register(&perf_tracepoint, "tracepoint", PERF_TYPE_TRACEPOINT);
5616 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
5621 if (event->attr.type != PERF_TYPE_TRACEPOINT)
5624 filter_str = strndup_user(arg, PAGE_SIZE);
5625 if (IS_ERR(filter_str))
5626 return PTR_ERR(filter_str);
5628 ret = ftrace_profile_set_filter(event, event->attr.config, filter_str);
5634 static void perf_event_free_filter(struct perf_event *event)
5636 ftrace_profile_free_filter(event);
5641 static inline void perf_tp_register(void)
5645 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
5650 static void perf_event_free_filter(struct perf_event *event)
5654 #endif /* CONFIG_EVENT_TRACING */
5656 #ifdef CONFIG_HAVE_HW_BREAKPOINT
5657 void perf_bp_event(struct perf_event *bp, void *data)
5659 struct perf_sample_data sample;
5660 struct pt_regs *regs = data;
5662 perf_sample_data_init(&sample, bp->attr.bp_addr, 0);
5664 if (!bp->hw.state && !perf_exclude_event(bp, regs))
5665 perf_swevent_event(bp, 1, &sample, regs);
5670 * hrtimer based swevent callback
5673 static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer)
5675 enum hrtimer_restart ret = HRTIMER_RESTART;
5676 struct perf_sample_data data;
5677 struct pt_regs *regs;
5678 struct perf_event *event;
5681 event = container_of(hrtimer, struct perf_event, hw.hrtimer);
5683 if (event->state != PERF_EVENT_STATE_ACTIVE)
5684 return HRTIMER_NORESTART;
5686 event->pmu->read(event);
5688 perf_sample_data_init(&data, 0, event->hw.last_period);
5689 regs = get_irq_regs();
5691 if (regs && !perf_exclude_event(event, regs)) {
5692 if (!(event->attr.exclude_idle && is_idle_task(current)))
5693 if (__perf_event_overflow(event, 1, &data, regs))
5694 ret = HRTIMER_NORESTART;
5697 period = max_t(u64, 10000, event->hw.sample_period);
5698 hrtimer_forward_now(hrtimer, ns_to_ktime(period));
5703 static void perf_swevent_start_hrtimer(struct perf_event *event)
5705 struct hw_perf_event *hwc = &event->hw;
5708 if (!is_sampling_event(event))
5711 period = local64_read(&hwc->period_left);
5716 local64_set(&hwc->period_left, 0);
5718 period = max_t(u64, 10000, hwc->sample_period);
5720 __hrtimer_start_range_ns(&hwc->hrtimer,
5721 ns_to_ktime(period), 0,
5722 HRTIMER_MODE_REL_PINNED, 0);
5725 static void perf_swevent_cancel_hrtimer(struct perf_event *event)
5727 struct hw_perf_event *hwc = &event->hw;
5729 if (is_sampling_event(event)) {
5730 ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer);
5731 local64_set(&hwc->period_left, ktime_to_ns(remaining));
5733 hrtimer_cancel(&hwc->hrtimer);
5737 static void perf_swevent_init_hrtimer(struct perf_event *event)
5739 struct hw_perf_event *hwc = &event->hw;
5741 if (!is_sampling_event(event))
5744 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
5745 hwc->hrtimer.function = perf_swevent_hrtimer;
5748 * Since hrtimers have a fixed rate, we can do a static freq->period
5749 * mapping and avoid the whole period adjust feedback stuff.
5751 if (event->attr.freq) {
5752 long freq = event->attr.sample_freq;
5754 event->attr.sample_period = NSEC_PER_SEC / freq;
5755 hwc->sample_period = event->attr.sample_period;
5756 local64_set(&hwc->period_left, hwc->sample_period);
5757 hwc->last_period = hwc->sample_period;
5758 event->attr.freq = 0;
5763 * Software event: cpu wall time clock
5766 static void cpu_clock_event_update(struct perf_event *event)
5771 now = local_clock();
5772 prev = local64_xchg(&event->hw.prev_count, now);
5773 local64_add(now - prev, &event->count);
5776 static void cpu_clock_event_start(struct perf_event *event, int flags)
5778 local64_set(&event->hw.prev_count, local_clock());
5779 perf_swevent_start_hrtimer(event);
5782 static void cpu_clock_event_stop(struct perf_event *event, int flags)
5784 perf_swevent_cancel_hrtimer(event);
5785 cpu_clock_event_update(event);
5788 static int cpu_clock_event_add(struct perf_event *event, int flags)
5790 if (flags & PERF_EF_START)
5791 cpu_clock_event_start(event, flags);
5796 static void cpu_clock_event_del(struct perf_event *event, int flags)
5798 cpu_clock_event_stop(event, flags);
5801 static void cpu_clock_event_read(struct perf_event *event)
5803 cpu_clock_event_update(event);
5806 static int cpu_clock_event_init(struct perf_event *event)
5808 if (event->attr.type != PERF_TYPE_SOFTWARE)
5811 if (event->attr.config != PERF_COUNT_SW_CPU_CLOCK)
5815 * no branch sampling for software events
5817 if (has_branch_stack(event))
5820 perf_swevent_init_hrtimer(event);
5825 static struct pmu perf_cpu_clock = {
5826 .task_ctx_nr = perf_sw_context,
5828 .event_init = cpu_clock_event_init,
5829 .add = cpu_clock_event_add,
5830 .del = cpu_clock_event_del,
5831 .start = cpu_clock_event_start,
5832 .stop = cpu_clock_event_stop,
5833 .read = cpu_clock_event_read,
5835 .event_idx = perf_swevent_event_idx,
5839 * Software event: task time clock
5842 static void task_clock_event_update(struct perf_event *event, u64 now)
5847 prev = local64_xchg(&event->hw.prev_count, now);
5849 local64_add(delta, &event->count);
5852 static void task_clock_event_start(struct perf_event *event, int flags)
5854 local64_set(&event->hw.prev_count, event->ctx->time);
5855 perf_swevent_start_hrtimer(event);
5858 static void task_clock_event_stop(struct perf_event *event, int flags)
5860 perf_swevent_cancel_hrtimer(event);
5861 task_clock_event_update(event, event->ctx->time);
5864 static int task_clock_event_add(struct perf_event *event, int flags)
5866 if (flags & PERF_EF_START)
5867 task_clock_event_start(event, flags);
5872 static void task_clock_event_del(struct perf_event *event, int flags)
5874 task_clock_event_stop(event, PERF_EF_UPDATE);
5877 static void task_clock_event_read(struct perf_event *event)
5879 u64 now = perf_clock();
5880 u64 delta = now - event->ctx->timestamp;
5881 u64 time = event->ctx->time + delta;
5883 task_clock_event_update(event, time);
5886 static int task_clock_event_init(struct perf_event *event)
5888 if (event->attr.type != PERF_TYPE_SOFTWARE)
5891 if (event->attr.config != PERF_COUNT_SW_TASK_CLOCK)
5895 * no branch sampling for software events
5897 if (has_branch_stack(event))
5900 perf_swevent_init_hrtimer(event);
5905 static struct pmu perf_task_clock = {
5906 .task_ctx_nr = perf_sw_context,
5908 .event_init = task_clock_event_init,
5909 .add = task_clock_event_add,
5910 .del = task_clock_event_del,
5911 .start = task_clock_event_start,
5912 .stop = task_clock_event_stop,
5913 .read = task_clock_event_read,
5915 .event_idx = perf_swevent_event_idx,
5918 static void perf_pmu_nop_void(struct pmu *pmu)
5922 static int perf_pmu_nop_int(struct pmu *pmu)
5927 static void perf_pmu_start_txn(struct pmu *pmu)
5929 perf_pmu_disable(pmu);
5932 static int perf_pmu_commit_txn(struct pmu *pmu)
5934 perf_pmu_enable(pmu);
5938 static void perf_pmu_cancel_txn(struct pmu *pmu)
5940 perf_pmu_enable(pmu);
5943 static int perf_event_idx_default(struct perf_event *event)
5945 return event->hw.idx + 1;
5949 * Ensures all contexts with the same task_ctx_nr have the same
5950 * pmu_cpu_context too.
5952 static void *find_pmu_context(int ctxn)
5959 list_for_each_entry(pmu, &pmus, entry) {
5960 if (pmu->task_ctx_nr == ctxn)
5961 return pmu->pmu_cpu_context;
5967 static void update_pmu_context(struct pmu *pmu, struct pmu *old_pmu)
5971 for_each_possible_cpu(cpu) {
5972 struct perf_cpu_context *cpuctx;
5974 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
5976 if (cpuctx->unique_pmu == old_pmu)
5977 cpuctx->unique_pmu = pmu;
5981 static void free_pmu_context(struct pmu *pmu)
5985 mutex_lock(&pmus_lock);
5987 * Like a real lame refcount.
5989 list_for_each_entry(i, &pmus, entry) {
5990 if (i->pmu_cpu_context == pmu->pmu_cpu_context) {
5991 update_pmu_context(i, pmu);
5996 free_percpu(pmu->pmu_cpu_context);
5998 mutex_unlock(&pmus_lock);
6000 static struct idr pmu_idr;
6003 type_show(struct device *dev, struct device_attribute *attr, char *page)
6005 struct pmu *pmu = dev_get_drvdata(dev);
6007 return snprintf(page, PAGE_SIZE-1, "%d\n", pmu->type);
6010 static struct device_attribute pmu_dev_attrs[] = {
6015 static int pmu_bus_running;
6016 static struct bus_type pmu_bus = {
6017 .name = "event_source",
6018 .dev_attrs = pmu_dev_attrs,
6021 static void pmu_dev_release(struct device *dev)
6026 static int pmu_dev_alloc(struct pmu *pmu)
6030 pmu->dev = kzalloc(sizeof(struct device), GFP_KERNEL);
6034 pmu->dev->groups = pmu->attr_groups;
6035 device_initialize(pmu->dev);
6036 ret = dev_set_name(pmu->dev, "%s", pmu->name);
6040 dev_set_drvdata(pmu->dev, pmu);
6041 pmu->dev->bus = &pmu_bus;
6042 pmu->dev->release = pmu_dev_release;
6043 ret = device_add(pmu->dev);
6051 put_device(pmu->dev);
6055 static struct lock_class_key cpuctx_mutex;
6056 static struct lock_class_key cpuctx_lock;
6058 int perf_pmu_register(struct pmu *pmu, char *name, int type)
6062 mutex_lock(&pmus_lock);
6064 pmu->pmu_disable_count = alloc_percpu(int);
6065 if (!pmu->pmu_disable_count)
6074 type = idr_alloc(&pmu_idr, pmu, PERF_TYPE_MAX, 0, GFP_KERNEL);
6082 if (pmu_bus_running) {
6083 ret = pmu_dev_alloc(pmu);
6089 pmu->pmu_cpu_context = find_pmu_context(pmu->task_ctx_nr);
6090 if (pmu->pmu_cpu_context)
6091 goto got_cpu_context;
6094 pmu->pmu_cpu_context = alloc_percpu(struct perf_cpu_context);
6095 if (!pmu->pmu_cpu_context)
6098 for_each_possible_cpu(cpu) {
6099 struct perf_cpu_context *cpuctx;
6101 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
6102 __perf_event_init_context(&cpuctx->ctx);
6103 lockdep_set_class(&cpuctx->ctx.mutex, &cpuctx_mutex);
6104 lockdep_set_class(&cpuctx->ctx.lock, &cpuctx_lock);
6105 cpuctx->ctx.type = cpu_context;
6106 cpuctx->ctx.pmu = pmu;
6107 cpuctx->jiffies_interval = 1;
6108 INIT_LIST_HEAD(&cpuctx->rotation_list);
6109 cpuctx->unique_pmu = pmu;
6113 if (!pmu->start_txn) {
6114 if (pmu->pmu_enable) {
6116 * If we have pmu_enable/pmu_disable calls, install
6117 * transaction stubs that use that to try and batch
6118 * hardware accesses.
6120 pmu->start_txn = perf_pmu_start_txn;
6121 pmu->commit_txn = perf_pmu_commit_txn;
6122 pmu->cancel_txn = perf_pmu_cancel_txn;
6124 pmu->start_txn = perf_pmu_nop_void;
6125 pmu->commit_txn = perf_pmu_nop_int;
6126 pmu->cancel_txn = perf_pmu_nop_void;
6130 if (!pmu->pmu_enable) {
6131 pmu->pmu_enable = perf_pmu_nop_void;
6132 pmu->pmu_disable = perf_pmu_nop_void;
6135 if (!pmu->event_idx)
6136 pmu->event_idx = perf_event_idx_default;
6138 list_add_rcu(&pmu->entry, &pmus);
6141 mutex_unlock(&pmus_lock);
6146 device_del(pmu->dev);
6147 put_device(pmu->dev);
6150 if (pmu->type >= PERF_TYPE_MAX)
6151 idr_remove(&pmu_idr, pmu->type);
6154 free_percpu(pmu->pmu_disable_count);
6158 void perf_pmu_unregister(struct pmu *pmu)
6160 mutex_lock(&pmus_lock);
6161 list_del_rcu(&pmu->entry);
6162 mutex_unlock(&pmus_lock);
6165 * We dereference the pmu list under both SRCU and regular RCU, so
6166 * synchronize against both of those.
6168 synchronize_srcu(&pmus_srcu);
6171 free_percpu(pmu->pmu_disable_count);
6172 if (pmu->type >= PERF_TYPE_MAX)
6173 idr_remove(&pmu_idr, pmu->type);
6174 device_del(pmu->dev);
6175 put_device(pmu->dev);
6176 free_pmu_context(pmu);
6179 struct pmu *perf_init_event(struct perf_event *event)
6181 struct pmu *pmu = NULL;
6185 idx = srcu_read_lock(&pmus_srcu);
6188 pmu = idr_find(&pmu_idr, event->attr.type);
6192 ret = pmu->event_init(event);
6198 list_for_each_entry_rcu(pmu, &pmus, entry) {
6200 ret = pmu->event_init(event);
6204 if (ret != -ENOENT) {
6209 pmu = ERR_PTR(-ENOENT);
6211 srcu_read_unlock(&pmus_srcu, idx);
6217 * Allocate and initialize a event structure
6219 static struct perf_event *
6220 perf_event_alloc(struct perf_event_attr *attr, int cpu,
6221 struct task_struct *task,
6222 struct perf_event *group_leader,
6223 struct perf_event *parent_event,
6224 perf_overflow_handler_t overflow_handler,
6228 struct perf_event *event;
6229 struct hw_perf_event *hwc;
6232 if ((unsigned)cpu >= nr_cpu_ids) {
6233 if (!task || cpu != -1)
6234 return ERR_PTR(-EINVAL);
6237 event = kzalloc(sizeof(*event), GFP_KERNEL);
6239 return ERR_PTR(-ENOMEM);
6242 * Single events are their own group leaders, with an
6243 * empty sibling list:
6246 group_leader = event;
6248 mutex_init(&event->child_mutex);
6249 INIT_LIST_HEAD(&event->child_list);
6251 INIT_LIST_HEAD(&event->group_entry);
6252 INIT_LIST_HEAD(&event->event_entry);
6253 INIT_LIST_HEAD(&event->sibling_list);
6254 INIT_LIST_HEAD(&event->rb_entry);
6256 init_waitqueue_head(&event->waitq);
6257 init_irq_work(&event->pending, perf_pending_event);
6259 mutex_init(&event->mmap_mutex);
6261 atomic_long_set(&event->refcount, 1);
6263 event->attr = *attr;
6264 event->group_leader = group_leader;
6268 event->parent = parent_event;
6270 event->ns = get_pid_ns(task_active_pid_ns(current));
6271 event->id = atomic64_inc_return(&perf_event_id);
6273 event->state = PERF_EVENT_STATE_INACTIVE;
6276 event->attach_state = PERF_ATTACH_TASK;
6278 if (attr->type == PERF_TYPE_TRACEPOINT)
6279 event->hw.tp_target = task;
6280 #ifdef CONFIG_HAVE_HW_BREAKPOINT
6282 * hw_breakpoint is a bit difficult here..
6284 else if (attr->type == PERF_TYPE_BREAKPOINT)
6285 event->hw.bp_target = task;
6289 if (!overflow_handler && parent_event) {
6290 overflow_handler = parent_event->overflow_handler;
6291 context = parent_event->overflow_handler_context;
6294 event->overflow_handler = overflow_handler;
6295 event->overflow_handler_context = context;
6297 perf_event__state_init(event);
6302 hwc->sample_period = attr->sample_period;
6303 if (attr->freq && attr->sample_freq)
6304 hwc->sample_period = 1;
6305 hwc->last_period = hwc->sample_period;
6307 local64_set(&hwc->period_left, hwc->sample_period);
6310 * we currently do not support PERF_FORMAT_GROUP on inherited events
6312 if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP))
6315 pmu = perf_init_event(event);
6321 else if (IS_ERR(pmu))
6326 put_pid_ns(event->ns);
6328 return ERR_PTR(err);
6331 if (!event->parent) {
6332 if (event->attach_state & PERF_ATTACH_TASK)
6333 static_key_slow_inc(&perf_sched_events.key);
6334 if (event->attr.mmap || event->attr.mmap_data)
6335 atomic_inc(&nr_mmap_events);
6336 if (event->attr.comm)
6337 atomic_inc(&nr_comm_events);
6338 if (event->attr.task)
6339 atomic_inc(&nr_task_events);
6340 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) {
6341 err = get_callchain_buffers();
6344 return ERR_PTR(err);
6347 if (has_branch_stack(event)) {
6348 static_key_slow_inc(&perf_sched_events.key);
6349 if (!(event->attach_state & PERF_ATTACH_TASK))
6350 atomic_inc(&per_cpu(perf_branch_stack_events,
6358 static int perf_copy_attr(struct perf_event_attr __user *uattr,
6359 struct perf_event_attr *attr)
6364 if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0))
6368 * zero the full structure, so that a short copy will be nice.
6370 memset(attr, 0, sizeof(*attr));
6372 ret = get_user(size, &uattr->size);
6376 if (size > PAGE_SIZE) /* silly large */
6379 if (!size) /* abi compat */
6380 size = PERF_ATTR_SIZE_VER0;
6382 if (size < PERF_ATTR_SIZE_VER0)
6386 * If we're handed a bigger struct than we know of,
6387 * ensure all the unknown bits are 0 - i.e. new
6388 * user-space does not rely on any kernel feature
6389 * extensions we dont know about yet.
6391 if (size > sizeof(*attr)) {
6392 unsigned char __user *addr;
6393 unsigned char __user *end;
6396 addr = (void __user *)uattr + sizeof(*attr);
6397 end = (void __user *)uattr + size;
6399 for (; addr < end; addr++) {
6400 ret = get_user(val, addr);
6406 size = sizeof(*attr);
6409 ret = copy_from_user(attr, uattr, size);
6413 if (attr->__reserved_1)
6416 if (attr->sample_type & ~(PERF_SAMPLE_MAX-1))
6419 if (attr->read_format & ~(PERF_FORMAT_MAX-1))
6422 if (attr->sample_type & PERF_SAMPLE_BRANCH_STACK) {
6423 u64 mask = attr->branch_sample_type;
6425 /* only using defined bits */
6426 if (mask & ~(PERF_SAMPLE_BRANCH_MAX-1))
6429 /* at least one branch bit must be set */
6430 if (!(mask & ~PERF_SAMPLE_BRANCH_PLM_ALL))
6433 /* kernel level capture: check permissions */
6434 if ((mask & PERF_SAMPLE_BRANCH_PERM_PLM)
6435 && perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
6438 /* propagate priv level, when not set for branch */
6439 if (!(mask & PERF_SAMPLE_BRANCH_PLM_ALL)) {
6441 /* exclude_kernel checked on syscall entry */
6442 if (!attr->exclude_kernel)
6443 mask |= PERF_SAMPLE_BRANCH_KERNEL;
6445 if (!attr->exclude_user)
6446 mask |= PERF_SAMPLE_BRANCH_USER;
6448 if (!attr->exclude_hv)
6449 mask |= PERF_SAMPLE_BRANCH_HV;
6451 * adjust user setting (for HW filter setup)
6453 attr->branch_sample_type = mask;
6457 if (attr->sample_type & PERF_SAMPLE_REGS_USER) {
6458 ret = perf_reg_validate(attr->sample_regs_user);
6463 if (attr->sample_type & PERF_SAMPLE_STACK_USER) {
6464 if (!arch_perf_have_user_stack_dump())
6468 * We have __u32 type for the size, but so far
6469 * we can only use __u16 as maximum due to the
6470 * __u16 sample size limit.
6472 if (attr->sample_stack_user >= USHRT_MAX)
6474 else if (!IS_ALIGNED(attr->sample_stack_user, sizeof(u64)))
6482 put_user(sizeof(*attr), &uattr->size);
6488 perf_event_set_output(struct perf_event *event, struct perf_event *output_event)
6490 struct ring_buffer *rb = NULL, *old_rb = NULL;
6496 /* don't allow circular references */
6497 if (event == output_event)
6501 * Don't allow cross-cpu buffers
6503 if (output_event->cpu != event->cpu)
6507 * If its not a per-cpu rb, it must be the same task.
6509 if (output_event->cpu == -1 && output_event->ctx != event->ctx)
6513 mutex_lock(&event->mmap_mutex);
6514 /* Can't redirect output if we've got an active mmap() */
6515 if (atomic_read(&event->mmap_count))
6521 /* get the rb we want to redirect to */
6522 rb = ring_buffer_get(output_event);
6528 ring_buffer_detach(event, old_rb);
6531 ring_buffer_attach(event, rb);
6533 rcu_assign_pointer(event->rb, rb);
6536 ring_buffer_put(old_rb);
6538 * Since we detached before setting the new rb, so that we
6539 * could attach the new rb, we could have missed a wakeup.
6542 wake_up_all(&event->waitq);
6547 mutex_unlock(&event->mmap_mutex);
6554 * sys_perf_event_open - open a performance event, associate it to a task/cpu
6556 * @attr_uptr: event_id type attributes for monitoring/sampling
6559 * @group_fd: group leader event fd
6561 SYSCALL_DEFINE5(perf_event_open,
6562 struct perf_event_attr __user *, attr_uptr,
6563 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
6565 struct perf_event *group_leader = NULL, *output_event = NULL;
6566 struct perf_event *event, *sibling;
6567 struct perf_event_attr attr;
6568 struct perf_event_context *ctx;
6569 struct file *event_file = NULL;
6570 struct fd group = {NULL, 0};
6571 struct task_struct *task = NULL;
6577 /* for future expandability... */
6578 if (flags & ~PERF_FLAG_ALL)
6581 err = perf_copy_attr(attr_uptr, &attr);
6585 if (!attr.exclude_kernel) {
6586 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
6591 if (attr.sample_freq > sysctl_perf_event_sample_rate)
6594 if (attr.sample_period & (1ULL << 63))
6599 * In cgroup mode, the pid argument is used to pass the fd
6600 * opened to the cgroup directory in cgroupfs. The cpu argument
6601 * designates the cpu on which to monitor threads from that
6604 if ((flags & PERF_FLAG_PID_CGROUP) && (pid == -1 || cpu == -1))
6607 event_fd = get_unused_fd();
6611 if (group_fd != -1) {
6612 err = perf_fget_light(group_fd, &group);
6615 group_leader = group.file->private_data;
6616 if (flags & PERF_FLAG_FD_OUTPUT)
6617 output_event = group_leader;
6618 if (flags & PERF_FLAG_FD_NO_GROUP)
6619 group_leader = NULL;
6622 if (pid != -1 && !(flags & PERF_FLAG_PID_CGROUP)) {
6623 task = find_lively_task_by_vpid(pid);
6625 err = PTR_ERR(task);
6632 event = perf_event_alloc(&attr, cpu, task, group_leader, NULL,
6634 if (IS_ERR(event)) {
6635 err = PTR_ERR(event);
6639 if (flags & PERF_FLAG_PID_CGROUP) {
6640 err = perf_cgroup_connect(pid, event, &attr, group_leader);
6645 * - that has cgroup constraint on event->cpu
6646 * - that may need work on context switch
6648 atomic_inc(&per_cpu(perf_cgroup_events, event->cpu));
6649 static_key_slow_inc(&perf_sched_events.key);
6653 * Special case software events and allow them to be part of
6654 * any hardware group.
6659 (is_software_event(event) != is_software_event(group_leader))) {
6660 if (is_software_event(event)) {
6662 * If event and group_leader are not both a software
6663 * event, and event is, then group leader is not.
6665 * Allow the addition of software events to !software
6666 * groups, this is safe because software events never
6669 pmu = group_leader->pmu;
6670 } else if (is_software_event(group_leader) &&
6671 (group_leader->group_flags & PERF_GROUP_SOFTWARE)) {
6673 * In case the group is a pure software group, and we
6674 * try to add a hardware event, move the whole group to
6675 * the hardware context.
6682 * Get the target context (task or percpu):
6684 ctx = find_get_context(pmu, task, event->cpu);
6691 put_task_struct(task);
6696 * Look up the group leader (we will attach this event to it):
6702 * Do not allow a recursive hierarchy (this new sibling
6703 * becoming part of another group-sibling):
6705 if (group_leader->group_leader != group_leader)
6708 * Do not allow to attach to a group in a different
6709 * task or CPU context:
6712 if (group_leader->ctx->type != ctx->type)
6715 if (group_leader->ctx != ctx)
6720 * Only a group leader can be exclusive or pinned
6722 if (attr.exclusive || attr.pinned)
6727 err = perf_event_set_output(event, output_event);
6732 event_file = anon_inode_getfile("[perf_event]", &perf_fops, event, O_RDWR);
6733 if (IS_ERR(event_file)) {
6734 err = PTR_ERR(event_file);
6739 struct perf_event_context *gctx = group_leader->ctx;
6741 mutex_lock(&gctx->mutex);
6742 perf_remove_from_context(group_leader);
6745 * Removing from the context ends up with disabled
6746 * event. What we want here is event in the initial
6747 * startup state, ready to be add into new context.
6749 perf_event__state_init(group_leader);
6750 list_for_each_entry(sibling, &group_leader->sibling_list,
6752 perf_remove_from_context(sibling);
6753 perf_event__state_init(sibling);
6756 mutex_unlock(&gctx->mutex);
6760 WARN_ON_ONCE(ctx->parent_ctx);
6761 mutex_lock(&ctx->mutex);
6765 perf_install_in_context(ctx, group_leader, event->cpu);
6767 list_for_each_entry(sibling, &group_leader->sibling_list,
6769 perf_install_in_context(ctx, sibling, event->cpu);
6774 perf_install_in_context(ctx, event, event->cpu);
6776 perf_unpin_context(ctx);
6777 mutex_unlock(&ctx->mutex);
6781 event->owner = current;
6783 mutex_lock(¤t->perf_event_mutex);
6784 list_add_tail(&event->owner_entry, ¤t->perf_event_list);
6785 mutex_unlock(¤t->perf_event_mutex);
6788 * Precalculate sample_data sizes
6790 perf_event__header_size(event);
6791 perf_event__id_header_size(event);
6794 * Drop the reference on the group_event after placing the
6795 * new event on the sibling_list. This ensures destruction
6796 * of the group leader will find the pointer to itself in
6797 * perf_group_detach().
6800 fd_install(event_fd, event_file);
6804 perf_unpin_context(ctx);
6811 put_task_struct(task);
6815 put_unused_fd(event_fd);
6820 * perf_event_create_kernel_counter
6822 * @attr: attributes of the counter to create
6823 * @cpu: cpu in which the counter is bound
6824 * @task: task to profile (NULL for percpu)
6827 perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu,
6828 struct task_struct *task,
6829 perf_overflow_handler_t overflow_handler,
6832 struct perf_event_context *ctx;
6833 struct perf_event *event;
6837 * Get the target context (task or percpu):
6840 event = perf_event_alloc(attr, cpu, task, NULL, NULL,
6841 overflow_handler, context);
6842 if (IS_ERR(event)) {
6843 err = PTR_ERR(event);
6847 ctx = find_get_context(event->pmu, task, cpu);
6853 WARN_ON_ONCE(ctx->parent_ctx);
6854 mutex_lock(&ctx->mutex);
6855 perf_install_in_context(ctx, event, cpu);
6857 perf_unpin_context(ctx);
6858 mutex_unlock(&ctx->mutex);
6865 return ERR_PTR(err);
6867 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter);
6869 void perf_pmu_migrate_context(struct pmu *pmu, int src_cpu, int dst_cpu)
6871 struct perf_event_context *src_ctx;
6872 struct perf_event_context *dst_ctx;
6873 struct perf_event *event, *tmp;
6876 src_ctx = &per_cpu_ptr(pmu->pmu_cpu_context, src_cpu)->ctx;
6877 dst_ctx = &per_cpu_ptr(pmu->pmu_cpu_context, dst_cpu)->ctx;
6879 mutex_lock(&src_ctx->mutex);
6880 list_for_each_entry_safe(event, tmp, &src_ctx->event_list,
6882 perf_remove_from_context(event);
6884 list_add(&event->event_entry, &events);
6886 mutex_unlock(&src_ctx->mutex);
6890 mutex_lock(&dst_ctx->mutex);
6891 list_for_each_entry_safe(event, tmp, &events, event_entry) {
6892 list_del(&event->event_entry);
6893 if (event->state >= PERF_EVENT_STATE_OFF)
6894 event->state = PERF_EVENT_STATE_INACTIVE;
6895 perf_install_in_context(dst_ctx, event, dst_cpu);
6898 mutex_unlock(&dst_ctx->mutex);
6900 EXPORT_SYMBOL_GPL(perf_pmu_migrate_context);
6902 static void sync_child_event(struct perf_event *child_event,
6903 struct task_struct *child)
6905 struct perf_event *parent_event = child_event->parent;
6908 if (child_event->attr.inherit_stat)
6909 perf_event_read_event(child_event, child);
6911 child_val = perf_event_count(child_event);
6914 * Add back the child's count to the parent's count:
6916 atomic64_add(child_val, &parent_event->child_count);
6917 atomic64_add(child_event->total_time_enabled,
6918 &parent_event->child_total_time_enabled);
6919 atomic64_add(child_event->total_time_running,
6920 &parent_event->child_total_time_running);
6923 * Remove this event from the parent's list
6925 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
6926 mutex_lock(&parent_event->child_mutex);
6927 list_del_init(&child_event->child_list);
6928 mutex_unlock(&parent_event->child_mutex);
6931 * Release the parent event, if this was the last
6934 put_event(parent_event);
6938 __perf_event_exit_task(struct perf_event *child_event,
6939 struct perf_event_context *child_ctx,
6940 struct task_struct *child)
6942 if (child_event->parent) {
6943 raw_spin_lock_irq(&child_ctx->lock);
6944 perf_group_detach(child_event);
6945 raw_spin_unlock_irq(&child_ctx->lock);
6948 perf_remove_from_context(child_event);
6951 * It can happen that the parent exits first, and has events
6952 * that are still around due to the child reference. These
6953 * events need to be zapped.
6955 if (child_event->parent) {
6956 sync_child_event(child_event, child);
6957 free_event(child_event);
6961 static void perf_event_exit_task_context(struct task_struct *child, int ctxn)
6963 struct perf_event *child_event, *tmp;
6964 struct perf_event_context *child_ctx;
6965 unsigned long flags;
6967 if (likely(!child->perf_event_ctxp[ctxn])) {
6968 perf_event_task(child, NULL, 0);
6972 local_irq_save(flags);
6974 * We can't reschedule here because interrupts are disabled,
6975 * and either child is current or it is a task that can't be
6976 * scheduled, so we are now safe from rescheduling changing
6979 child_ctx = rcu_dereference_raw(child->perf_event_ctxp[ctxn]);
6982 * Take the context lock here so that if find_get_context is
6983 * reading child->perf_event_ctxp, we wait until it has
6984 * incremented the context's refcount before we do put_ctx below.
6986 raw_spin_lock(&child_ctx->lock);
6987 task_ctx_sched_out(child_ctx);
6988 child->perf_event_ctxp[ctxn] = NULL;
6990 * If this context is a clone; unclone it so it can't get
6991 * swapped to another process while we're removing all
6992 * the events from it.
6994 unclone_ctx(child_ctx);
6995 update_context_time(child_ctx);
6996 raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
6999 * Report the task dead after unscheduling the events so that we
7000 * won't get any samples after PERF_RECORD_EXIT. We can however still
7001 * get a few PERF_RECORD_READ events.
7003 perf_event_task(child, child_ctx, 0);
7006 * We can recurse on the same lock type through:
7008 * __perf_event_exit_task()
7009 * sync_child_event()
7011 * mutex_lock(&ctx->mutex)
7013 * But since its the parent context it won't be the same instance.
7015 mutex_lock(&child_ctx->mutex);
7018 list_for_each_entry_safe(child_event, tmp, &child_ctx->pinned_groups,
7020 __perf_event_exit_task(child_event, child_ctx, child);
7022 list_for_each_entry_safe(child_event, tmp, &child_ctx->flexible_groups,
7024 __perf_event_exit_task(child_event, child_ctx, child);
7027 * If the last event was a group event, it will have appended all
7028 * its siblings to the list, but we obtained 'tmp' before that which
7029 * will still point to the list head terminating the iteration.
7031 if (!list_empty(&child_ctx->pinned_groups) ||
7032 !list_empty(&child_ctx->flexible_groups))
7035 mutex_unlock(&child_ctx->mutex);
7041 * When a child task exits, feed back event values to parent events.
7043 void perf_event_exit_task(struct task_struct *child)
7045 struct perf_event *event, *tmp;
7048 mutex_lock(&child->perf_event_mutex);
7049 list_for_each_entry_safe(event, tmp, &child->perf_event_list,
7051 list_del_init(&event->owner_entry);
7054 * Ensure the list deletion is visible before we clear
7055 * the owner, closes a race against perf_release() where
7056 * we need to serialize on the owner->perf_event_mutex.
7059 event->owner = NULL;
7061 mutex_unlock(&child->perf_event_mutex);
7063 for_each_task_context_nr(ctxn)
7064 perf_event_exit_task_context(child, ctxn);
7067 static void perf_free_event(struct perf_event *event,
7068 struct perf_event_context *ctx)
7070 struct perf_event *parent = event->parent;
7072 if (WARN_ON_ONCE(!parent))
7075 mutex_lock(&parent->child_mutex);
7076 list_del_init(&event->child_list);
7077 mutex_unlock(&parent->child_mutex);
7081 perf_group_detach(event);
7082 list_del_event(event, ctx);
7087 * free an unexposed, unused context as created by inheritance by
7088 * perf_event_init_task below, used by fork() in case of fail.
7090 void perf_event_free_task(struct task_struct *task)
7092 struct perf_event_context *ctx;
7093 struct perf_event *event, *tmp;
7096 for_each_task_context_nr(ctxn) {
7097 ctx = task->perf_event_ctxp[ctxn];
7101 mutex_lock(&ctx->mutex);
7103 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups,
7105 perf_free_event(event, ctx);
7107 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups,
7109 perf_free_event(event, ctx);
7111 if (!list_empty(&ctx->pinned_groups) ||
7112 !list_empty(&ctx->flexible_groups))
7115 mutex_unlock(&ctx->mutex);
7121 void perf_event_delayed_put(struct task_struct *task)
7125 for_each_task_context_nr(ctxn)
7126 WARN_ON_ONCE(task->perf_event_ctxp[ctxn]);
7130 * inherit a event from parent task to child task:
7132 static struct perf_event *
7133 inherit_event(struct perf_event *parent_event,
7134 struct task_struct *parent,
7135 struct perf_event_context *parent_ctx,
7136 struct task_struct *child,
7137 struct perf_event *group_leader,
7138 struct perf_event_context *child_ctx)
7140 struct perf_event *child_event;
7141 unsigned long flags;
7144 * Instead of creating recursive hierarchies of events,
7145 * we link inherited events back to the original parent,
7146 * which has a filp for sure, which we use as the reference
7149 if (parent_event->parent)
7150 parent_event = parent_event->parent;
7152 child_event = perf_event_alloc(&parent_event->attr,
7155 group_leader, parent_event,
7157 if (IS_ERR(child_event))
7160 if (!atomic_long_inc_not_zero(&parent_event->refcount)) {
7161 free_event(child_event);
7168 * Make the child state follow the state of the parent event,
7169 * not its attr.disabled bit. We hold the parent's mutex,
7170 * so we won't race with perf_event_{en, dis}able_family.
7172 if (parent_event->state >= PERF_EVENT_STATE_INACTIVE)
7173 child_event->state = PERF_EVENT_STATE_INACTIVE;
7175 child_event->state = PERF_EVENT_STATE_OFF;
7177 if (parent_event->attr.freq) {
7178 u64 sample_period = parent_event->hw.sample_period;
7179 struct hw_perf_event *hwc = &child_event->hw;
7181 hwc->sample_period = sample_period;
7182 hwc->last_period = sample_period;
7184 local64_set(&hwc->period_left, sample_period);
7187 child_event->ctx = child_ctx;
7188 child_event->overflow_handler = parent_event->overflow_handler;
7189 child_event->overflow_handler_context
7190 = parent_event->overflow_handler_context;
7193 * Precalculate sample_data sizes
7195 perf_event__header_size(child_event);
7196 perf_event__id_header_size(child_event);
7199 * Link it up in the child's context:
7201 raw_spin_lock_irqsave(&child_ctx->lock, flags);
7202 add_event_to_ctx(child_event, child_ctx);
7203 raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
7206 * Link this into the parent event's child list
7208 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
7209 mutex_lock(&parent_event->child_mutex);
7210 list_add_tail(&child_event->child_list, &parent_event->child_list);
7211 mutex_unlock(&parent_event->child_mutex);
7216 static int inherit_group(struct perf_event *parent_event,
7217 struct task_struct *parent,
7218 struct perf_event_context *parent_ctx,
7219 struct task_struct *child,
7220 struct perf_event_context *child_ctx)
7222 struct perf_event *leader;
7223 struct perf_event *sub;
7224 struct perf_event *child_ctr;
7226 leader = inherit_event(parent_event, parent, parent_ctx,
7227 child, NULL, child_ctx);
7229 return PTR_ERR(leader);
7230 list_for_each_entry(sub, &parent_event->sibling_list, group_entry) {
7231 child_ctr = inherit_event(sub, parent, parent_ctx,
7232 child, leader, child_ctx);
7233 if (IS_ERR(child_ctr))
7234 return PTR_ERR(child_ctr);
7240 inherit_task_group(struct perf_event *event, struct task_struct *parent,
7241 struct perf_event_context *parent_ctx,
7242 struct task_struct *child, int ctxn,
7246 struct perf_event_context *child_ctx;
7248 if (!event->attr.inherit) {
7253 child_ctx = child->perf_event_ctxp[ctxn];
7256 * This is executed from the parent task context, so
7257 * inherit events that have been marked for cloning.
7258 * First allocate and initialize a context for the
7262 child_ctx = alloc_perf_context(parent_ctx->pmu, child);
7266 child->perf_event_ctxp[ctxn] = child_ctx;
7269 ret = inherit_group(event, parent, parent_ctx,
7279 * Initialize the perf_event context in task_struct
7281 int perf_event_init_context(struct task_struct *child, int ctxn)
7283 struct perf_event_context *child_ctx, *parent_ctx;
7284 struct perf_event_context *cloned_ctx;
7285 struct perf_event *event;
7286 struct task_struct *parent = current;
7287 int inherited_all = 1;
7288 unsigned long flags;
7291 if (likely(!parent->perf_event_ctxp[ctxn]))
7295 * If the parent's context is a clone, pin it so it won't get
7298 parent_ctx = perf_pin_task_context(parent, ctxn);
7301 * No need to check if parent_ctx != NULL here; since we saw
7302 * it non-NULL earlier, the only reason for it to become NULL
7303 * is if we exit, and since we're currently in the middle of
7304 * a fork we can't be exiting at the same time.
7308 * Lock the parent list. No need to lock the child - not PID
7309 * hashed yet and not running, so nobody can access it.
7311 mutex_lock(&parent_ctx->mutex);
7314 * We dont have to disable NMIs - we are only looking at
7315 * the list, not manipulating it:
7317 list_for_each_entry(event, &parent_ctx->pinned_groups, group_entry) {
7318 ret = inherit_task_group(event, parent, parent_ctx,
7319 child, ctxn, &inherited_all);
7325 * We can't hold ctx->lock when iterating the ->flexible_group list due
7326 * to allocations, but we need to prevent rotation because
7327 * rotate_ctx() will change the list from interrupt context.
7329 raw_spin_lock_irqsave(&parent_ctx->lock, flags);
7330 parent_ctx->rotate_disable = 1;
7331 raw_spin_unlock_irqrestore(&parent_ctx->lock, flags);
7333 list_for_each_entry(event, &parent_ctx->flexible_groups, group_entry) {
7334 ret = inherit_task_group(event, parent, parent_ctx,
7335 child, ctxn, &inherited_all);
7340 raw_spin_lock_irqsave(&parent_ctx->lock, flags);
7341 parent_ctx->rotate_disable = 0;
7343 child_ctx = child->perf_event_ctxp[ctxn];
7345 if (child_ctx && inherited_all) {
7347 * Mark the child context as a clone of the parent
7348 * context, or of whatever the parent is a clone of.
7350 * Note that if the parent is a clone, the holding of
7351 * parent_ctx->lock avoids it from being uncloned.
7353 cloned_ctx = parent_ctx->parent_ctx;
7355 child_ctx->parent_ctx = cloned_ctx;
7356 child_ctx->parent_gen = parent_ctx->parent_gen;
7358 child_ctx->parent_ctx = parent_ctx;
7359 child_ctx->parent_gen = parent_ctx->generation;
7361 get_ctx(child_ctx->parent_ctx);
7364 raw_spin_unlock_irqrestore(&parent_ctx->lock, flags);
7365 mutex_unlock(&parent_ctx->mutex);
7367 perf_unpin_context(parent_ctx);
7368 put_ctx(parent_ctx);
7374 * Initialize the perf_event context in task_struct
7376 int perf_event_init_task(struct task_struct *child)
7380 memset(child->perf_event_ctxp, 0, sizeof(child->perf_event_ctxp));
7381 mutex_init(&child->perf_event_mutex);
7382 INIT_LIST_HEAD(&child->perf_event_list);
7384 for_each_task_context_nr(ctxn) {
7385 ret = perf_event_init_context(child, ctxn);
7393 static void __init perf_event_init_all_cpus(void)
7395 struct swevent_htable *swhash;
7398 for_each_possible_cpu(cpu) {
7399 swhash = &per_cpu(swevent_htable, cpu);
7400 mutex_init(&swhash->hlist_mutex);
7401 INIT_LIST_HEAD(&per_cpu(rotation_list, cpu));
7405 static void __cpuinit perf_event_init_cpu(int cpu)
7407 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
7409 mutex_lock(&swhash->hlist_mutex);
7410 swhash->online = true;
7411 if (swhash->hlist_refcount > 0) {
7412 struct swevent_hlist *hlist;
7414 hlist = kzalloc_node(sizeof(*hlist), GFP_KERNEL, cpu_to_node(cpu));
7416 rcu_assign_pointer(swhash->swevent_hlist, hlist);
7418 mutex_unlock(&swhash->hlist_mutex);
7421 #if defined CONFIG_HOTPLUG_CPU || defined CONFIG_KEXEC
7422 static void perf_pmu_rotate_stop(struct pmu *pmu)
7424 struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
7426 WARN_ON(!irqs_disabled());
7428 list_del_init(&cpuctx->rotation_list);
7431 static void __perf_event_exit_context(void *__info)
7433 struct perf_event_context *ctx = __info;
7434 struct perf_event *event;
7436 perf_pmu_rotate_stop(ctx->pmu);
7439 list_for_each_entry_rcu(event, &ctx->event_list, event_entry)
7440 __perf_remove_from_context(event);
7444 static void perf_event_exit_cpu_context(int cpu)
7446 struct perf_event_context *ctx;
7450 idx = srcu_read_lock(&pmus_srcu);
7451 list_for_each_entry_rcu(pmu, &pmus, entry) {
7452 ctx = &per_cpu_ptr(pmu->pmu_cpu_context, cpu)->ctx;
7454 mutex_lock(&ctx->mutex);
7455 smp_call_function_single(cpu, __perf_event_exit_context, ctx, 1);
7456 mutex_unlock(&ctx->mutex);
7458 srcu_read_unlock(&pmus_srcu, idx);
7461 static void perf_event_exit_cpu(int cpu)
7463 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
7465 perf_event_exit_cpu_context(cpu);
7467 mutex_lock(&swhash->hlist_mutex);
7468 swhash->online = false;
7469 swevent_hlist_release(swhash);
7470 mutex_unlock(&swhash->hlist_mutex);
7473 static inline void perf_event_exit_cpu(int cpu) { }
7477 perf_reboot(struct notifier_block *notifier, unsigned long val, void *v)
7481 for_each_online_cpu(cpu)
7482 perf_event_exit_cpu(cpu);
7488 * Run the perf reboot notifier at the very last possible moment so that
7489 * the generic watchdog code runs as long as possible.
7491 static struct notifier_block perf_reboot_notifier = {
7492 .notifier_call = perf_reboot,
7493 .priority = INT_MIN,
7496 static int __cpuinit
7497 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
7499 unsigned int cpu = (long)hcpu;
7501 switch (action & ~CPU_TASKS_FROZEN) {
7503 case CPU_UP_PREPARE:
7504 case CPU_DOWN_FAILED:
7505 perf_event_init_cpu(cpu);
7508 case CPU_UP_CANCELED:
7509 case CPU_DOWN_PREPARE:
7510 perf_event_exit_cpu(cpu);
7520 void __init perf_event_init(void)
7526 perf_event_init_all_cpus();
7527 init_srcu_struct(&pmus_srcu);
7528 perf_pmu_register(&perf_swevent, "software", PERF_TYPE_SOFTWARE);
7529 perf_pmu_register(&perf_cpu_clock, NULL, -1);
7530 perf_pmu_register(&perf_task_clock, NULL, -1);
7532 perf_cpu_notifier(perf_cpu_notify);
7533 register_reboot_notifier(&perf_reboot_notifier);
7535 ret = init_hw_breakpoint();
7536 WARN(ret, "hw_breakpoint initialization failed with: %d", ret);
7538 /* do not patch jump label more than once per second */
7539 jump_label_rate_limit(&perf_sched_events, HZ);
7542 * Build time assertion that we keep the data_head at the intended
7543 * location. IOW, validation we got the __reserved[] size right.
7545 BUILD_BUG_ON((offsetof(struct perf_event_mmap_page, data_head))
7549 static int __init perf_event_sysfs_init(void)
7554 mutex_lock(&pmus_lock);
7556 ret = bus_register(&pmu_bus);
7560 list_for_each_entry(pmu, &pmus, entry) {
7561 if (!pmu->name || pmu->type < 0)
7564 ret = pmu_dev_alloc(pmu);
7565 WARN(ret, "Failed to register pmu: %s, reason %d\n", pmu->name, ret);
7567 pmu_bus_running = 1;
7571 mutex_unlock(&pmus_lock);
7575 device_initcall(perf_event_sysfs_init);
7577 #ifdef CONFIG_CGROUP_PERF
7578 static struct cgroup_subsys_state *perf_cgroup_css_alloc(struct cgroup *cont)
7580 struct perf_cgroup *jc;
7582 jc = kzalloc(sizeof(*jc), GFP_KERNEL);
7584 return ERR_PTR(-ENOMEM);
7586 jc->info = alloc_percpu(struct perf_cgroup_info);
7589 return ERR_PTR(-ENOMEM);
7595 static void perf_cgroup_css_free(struct cgroup *cont)
7597 struct perf_cgroup *jc;
7598 jc = container_of(cgroup_subsys_state(cont, perf_subsys_id),
7599 struct perf_cgroup, css);
7600 free_percpu(jc->info);
7604 static int __perf_cgroup_move(void *info)
7606 struct task_struct *task = info;
7607 perf_cgroup_switch(task, PERF_CGROUP_SWOUT | PERF_CGROUP_SWIN);
7611 static void perf_cgroup_attach(struct cgroup *cgrp, struct cgroup_taskset *tset)
7613 struct task_struct *task;
7615 cgroup_taskset_for_each(task, cgrp, tset)
7616 task_function_call(task, __perf_cgroup_move, task);
7619 static void perf_cgroup_exit(struct cgroup *cgrp, struct cgroup *old_cgrp,
7620 struct task_struct *task)
7623 * cgroup_exit() is called in the copy_process() failure path.
7624 * Ignore this case since the task hasn't ran yet, this avoids
7625 * trying to poke a half freed task state from generic code.
7627 if (!(task->flags & PF_EXITING))
7630 task_function_call(task, __perf_cgroup_move, task);
7633 struct cgroup_subsys perf_subsys = {
7634 .name = "perf_event",
7635 .subsys_id = perf_subsys_id,
7636 .css_alloc = perf_cgroup_css_alloc,
7637 .css_free = perf_cgroup_css_free,
7638 .exit = perf_cgroup_exit,
7639 .attach = perf_cgroup_attach,
7641 #endif /* CONFIG_CGROUP_PERF */