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 #define list_next_entry(pos, member) \
2020 list_entry(pos->member.next, typeof(*pos), member)
2022 static void perf_event_sync_stat(struct perf_event_context *ctx,
2023 struct perf_event_context *next_ctx)
2025 struct perf_event *event, *next_event;
2030 update_context_time(ctx);
2032 event = list_first_entry(&ctx->event_list,
2033 struct perf_event, event_entry);
2035 next_event = list_first_entry(&next_ctx->event_list,
2036 struct perf_event, event_entry);
2038 while (&event->event_entry != &ctx->event_list &&
2039 &next_event->event_entry != &next_ctx->event_list) {
2041 __perf_event_sync_stat(event, next_event);
2043 event = list_next_entry(event, event_entry);
2044 next_event = list_next_entry(next_event, event_entry);
2048 static void perf_event_context_sched_out(struct task_struct *task, int ctxn,
2049 struct task_struct *next)
2051 struct perf_event_context *ctx = task->perf_event_ctxp[ctxn];
2052 struct perf_event_context *next_ctx;
2053 struct perf_event_context *parent;
2054 struct perf_cpu_context *cpuctx;
2060 cpuctx = __get_cpu_context(ctx);
2061 if (!cpuctx->task_ctx)
2065 parent = rcu_dereference(ctx->parent_ctx);
2066 next_ctx = next->perf_event_ctxp[ctxn];
2067 if (parent && next_ctx &&
2068 rcu_dereference(next_ctx->parent_ctx) == parent) {
2070 * Looks like the two contexts are clones, so we might be
2071 * able to optimize the context switch. We lock both
2072 * contexts and check that they are clones under the
2073 * lock (including re-checking that neither has been
2074 * uncloned in the meantime). It doesn't matter which
2075 * order we take the locks because no other cpu could
2076 * be trying to lock both of these tasks.
2078 raw_spin_lock(&ctx->lock);
2079 raw_spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING);
2080 if (context_equiv(ctx, next_ctx)) {
2082 * XXX do we need a memory barrier of sorts
2083 * wrt to rcu_dereference() of perf_event_ctxp
2085 task->perf_event_ctxp[ctxn] = next_ctx;
2086 next->perf_event_ctxp[ctxn] = ctx;
2088 next_ctx->task = task;
2091 perf_event_sync_stat(ctx, next_ctx);
2093 raw_spin_unlock(&next_ctx->lock);
2094 raw_spin_unlock(&ctx->lock);
2099 raw_spin_lock(&ctx->lock);
2100 ctx_sched_out(ctx, cpuctx, EVENT_ALL);
2101 cpuctx->task_ctx = NULL;
2102 raw_spin_unlock(&ctx->lock);
2106 #define for_each_task_context_nr(ctxn) \
2107 for ((ctxn) = 0; (ctxn) < perf_nr_task_contexts; (ctxn)++)
2110 * Called from scheduler to remove the events of the current task,
2111 * with interrupts disabled.
2113 * We stop each event and update the event value in event->count.
2115 * This does not protect us against NMI, but disable()
2116 * sets the disabled bit in the control field of event _before_
2117 * accessing the event control register. If a NMI hits, then it will
2118 * not restart the event.
2120 void __perf_event_task_sched_out(struct task_struct *task,
2121 struct task_struct *next)
2125 for_each_task_context_nr(ctxn)
2126 perf_event_context_sched_out(task, ctxn, next);
2129 * if cgroup events exist on this CPU, then we need
2130 * to check if we have to switch out PMU state.
2131 * cgroup event are system-wide mode only
2133 if (atomic_read(&__get_cpu_var(perf_cgroup_events)))
2134 perf_cgroup_sched_out(task, next);
2137 static void task_ctx_sched_out(struct perf_event_context *ctx)
2139 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
2141 if (!cpuctx->task_ctx)
2144 if (WARN_ON_ONCE(ctx != cpuctx->task_ctx))
2147 ctx_sched_out(ctx, cpuctx, EVENT_ALL);
2148 cpuctx->task_ctx = NULL;
2152 * Called with IRQs disabled
2154 static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
2155 enum event_type_t event_type)
2157 ctx_sched_out(&cpuctx->ctx, cpuctx, event_type);
2161 ctx_pinned_sched_in(struct perf_event_context *ctx,
2162 struct perf_cpu_context *cpuctx)
2164 struct perf_event *event;
2166 list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
2167 if (event->state <= PERF_EVENT_STATE_OFF)
2169 if (!event_filter_match(event))
2172 /* may need to reset tstamp_enabled */
2173 if (is_cgroup_event(event))
2174 perf_cgroup_mark_enabled(event, ctx);
2176 if (group_can_go_on(event, cpuctx, 1))
2177 group_sched_in(event, cpuctx, ctx);
2180 * If this pinned group hasn't been scheduled,
2181 * put it in error state.
2183 if (event->state == PERF_EVENT_STATE_INACTIVE) {
2184 update_group_times(event);
2185 event->state = PERF_EVENT_STATE_ERROR;
2191 ctx_flexible_sched_in(struct perf_event_context *ctx,
2192 struct perf_cpu_context *cpuctx)
2194 struct perf_event *event;
2197 list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
2198 /* Ignore events in OFF or ERROR state */
2199 if (event->state <= PERF_EVENT_STATE_OFF)
2202 * Listen to the 'cpu' scheduling filter constraint
2205 if (!event_filter_match(event))
2208 /* may need to reset tstamp_enabled */
2209 if (is_cgroup_event(event))
2210 perf_cgroup_mark_enabled(event, ctx);
2212 if (group_can_go_on(event, cpuctx, can_add_hw)) {
2213 if (group_sched_in(event, cpuctx, ctx))
2220 ctx_sched_in(struct perf_event_context *ctx,
2221 struct perf_cpu_context *cpuctx,
2222 enum event_type_t event_type,
2223 struct task_struct *task)
2226 int is_active = ctx->is_active;
2228 ctx->is_active |= event_type;
2229 if (likely(!ctx->nr_events))
2233 ctx->timestamp = now;
2234 perf_cgroup_set_timestamp(task, ctx);
2236 * First go through the list and put on any pinned groups
2237 * in order to give them the best chance of going on.
2239 if (!(is_active & EVENT_PINNED) && (event_type & EVENT_PINNED))
2240 ctx_pinned_sched_in(ctx, cpuctx);
2242 /* Then walk through the lower prio flexible groups */
2243 if (!(is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE))
2244 ctx_flexible_sched_in(ctx, cpuctx);
2247 static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
2248 enum event_type_t event_type,
2249 struct task_struct *task)
2251 struct perf_event_context *ctx = &cpuctx->ctx;
2253 ctx_sched_in(ctx, cpuctx, event_type, task);
2256 static void perf_event_context_sched_in(struct perf_event_context *ctx,
2257 struct task_struct *task)
2259 struct perf_cpu_context *cpuctx;
2261 cpuctx = __get_cpu_context(ctx);
2262 if (cpuctx->task_ctx == ctx)
2265 perf_ctx_lock(cpuctx, ctx);
2266 perf_pmu_disable(ctx->pmu);
2268 * We want to keep the following priority order:
2269 * cpu pinned (that don't need to move), task pinned,
2270 * cpu flexible, task flexible.
2272 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
2275 cpuctx->task_ctx = ctx;
2277 perf_event_sched_in(cpuctx, cpuctx->task_ctx, task);
2279 perf_pmu_enable(ctx->pmu);
2280 perf_ctx_unlock(cpuctx, ctx);
2283 * Since these rotations are per-cpu, we need to ensure the
2284 * cpu-context we got scheduled on is actually rotating.
2286 perf_pmu_rotate_start(ctx->pmu);
2290 * When sampling the branck stack in system-wide, it may be necessary
2291 * to flush the stack on context switch. This happens when the branch
2292 * stack does not tag its entries with the pid of the current task.
2293 * Otherwise it becomes impossible to associate a branch entry with a
2294 * task. This ambiguity is more likely to appear when the branch stack
2295 * supports priv level filtering and the user sets it to monitor only
2296 * at the user level (which could be a useful measurement in system-wide
2297 * mode). In that case, the risk is high of having a branch stack with
2298 * branch from multiple tasks. Flushing may mean dropping the existing
2299 * entries or stashing them somewhere in the PMU specific code layer.
2301 * This function provides the context switch callback to the lower code
2302 * layer. It is invoked ONLY when there is at least one system-wide context
2303 * with at least one active event using taken branch sampling.
2305 static void perf_branch_stack_sched_in(struct task_struct *prev,
2306 struct task_struct *task)
2308 struct perf_cpu_context *cpuctx;
2310 unsigned long flags;
2312 /* no need to flush branch stack if not changing task */
2316 local_irq_save(flags);
2320 list_for_each_entry_rcu(pmu, &pmus, entry) {
2321 cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
2324 * check if the context has at least one
2325 * event using PERF_SAMPLE_BRANCH_STACK
2327 if (cpuctx->ctx.nr_branch_stack > 0
2328 && pmu->flush_branch_stack) {
2330 pmu = cpuctx->ctx.pmu;
2332 perf_ctx_lock(cpuctx, cpuctx->task_ctx);
2334 perf_pmu_disable(pmu);
2336 pmu->flush_branch_stack();
2338 perf_pmu_enable(pmu);
2340 perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
2346 local_irq_restore(flags);
2350 * Called from scheduler to add the events of the current task
2351 * with interrupts disabled.
2353 * We restore the event value and then enable it.
2355 * This does not protect us against NMI, but enable()
2356 * sets the enabled bit in the control field of event _before_
2357 * accessing the event control register. If a NMI hits, then it will
2358 * keep the event running.
2360 void __perf_event_task_sched_in(struct task_struct *prev,
2361 struct task_struct *task)
2363 struct perf_event_context *ctx;
2366 for_each_task_context_nr(ctxn) {
2367 ctx = task->perf_event_ctxp[ctxn];
2371 perf_event_context_sched_in(ctx, task);
2374 * if cgroup events exist on this CPU, then we need
2375 * to check if we have to switch in PMU state.
2376 * cgroup event are system-wide mode only
2378 if (atomic_read(&__get_cpu_var(perf_cgroup_events)))
2379 perf_cgroup_sched_in(prev, task);
2381 /* check for system-wide branch_stack events */
2382 if (atomic_read(&__get_cpu_var(perf_branch_stack_events)))
2383 perf_branch_stack_sched_in(prev, task);
2386 static u64 perf_calculate_period(struct perf_event *event, u64 nsec, u64 count)
2388 u64 frequency = event->attr.sample_freq;
2389 u64 sec = NSEC_PER_SEC;
2390 u64 divisor, dividend;
2392 int count_fls, nsec_fls, frequency_fls, sec_fls;
2394 count_fls = fls64(count);
2395 nsec_fls = fls64(nsec);
2396 frequency_fls = fls64(frequency);
2400 * We got @count in @nsec, with a target of sample_freq HZ
2401 * the target period becomes:
2404 * period = -------------------
2405 * @nsec * sample_freq
2410 * Reduce accuracy by one bit such that @a and @b converge
2411 * to a similar magnitude.
2413 #define REDUCE_FLS(a, b) \
2415 if (a##_fls > b##_fls) { \
2425 * Reduce accuracy until either term fits in a u64, then proceed with
2426 * the other, so that finally we can do a u64/u64 division.
2428 while (count_fls + sec_fls > 64 && nsec_fls + frequency_fls > 64) {
2429 REDUCE_FLS(nsec, frequency);
2430 REDUCE_FLS(sec, count);
2433 if (count_fls + sec_fls > 64) {
2434 divisor = nsec * frequency;
2436 while (count_fls + sec_fls > 64) {
2437 REDUCE_FLS(count, sec);
2441 dividend = count * sec;
2443 dividend = count * sec;
2445 while (nsec_fls + frequency_fls > 64) {
2446 REDUCE_FLS(nsec, frequency);
2450 divisor = nsec * frequency;
2456 return div64_u64(dividend, divisor);
2459 static DEFINE_PER_CPU(int, perf_throttled_count);
2460 static DEFINE_PER_CPU(u64, perf_throttled_seq);
2462 static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count, bool disable)
2464 struct hw_perf_event *hwc = &event->hw;
2465 s64 period, sample_period;
2468 period = perf_calculate_period(event, nsec, count);
2470 delta = (s64)(period - hwc->sample_period);
2471 delta = (delta + 7) / 8; /* low pass filter */
2473 sample_period = hwc->sample_period + delta;
2478 hwc->sample_period = sample_period;
2480 if (local64_read(&hwc->period_left) > 8*sample_period) {
2482 event->pmu->stop(event, PERF_EF_UPDATE);
2484 local64_set(&hwc->period_left, 0);
2487 event->pmu->start(event, PERF_EF_RELOAD);
2492 * combine freq adjustment with unthrottling to avoid two passes over the
2493 * events. At the same time, make sure, having freq events does not change
2494 * the rate of unthrottling as that would introduce bias.
2496 static void perf_adjust_freq_unthr_context(struct perf_event_context *ctx,
2499 struct perf_event *event;
2500 struct hw_perf_event *hwc;
2501 u64 now, period = TICK_NSEC;
2505 * only need to iterate over all events iff:
2506 * - context have events in frequency mode (needs freq adjust)
2507 * - there are events to unthrottle on this cpu
2509 if (!(ctx->nr_freq || needs_unthr))
2512 raw_spin_lock(&ctx->lock);
2513 perf_pmu_disable(ctx->pmu);
2515 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
2516 if (event->state != PERF_EVENT_STATE_ACTIVE)
2519 if (!event_filter_match(event))
2524 if (needs_unthr && hwc->interrupts == MAX_INTERRUPTS) {
2525 hwc->interrupts = 0;
2526 perf_log_throttle(event, 1);
2527 event->pmu->start(event, 0);
2530 if (!event->attr.freq || !event->attr.sample_freq)
2534 * stop the event and update event->count
2536 event->pmu->stop(event, PERF_EF_UPDATE);
2538 now = local64_read(&event->count);
2539 delta = now - hwc->freq_count_stamp;
2540 hwc->freq_count_stamp = now;
2544 * reload only if value has changed
2545 * we have stopped the event so tell that
2546 * to perf_adjust_period() to avoid stopping it
2550 perf_adjust_period(event, period, delta, false);
2552 event->pmu->start(event, delta > 0 ? PERF_EF_RELOAD : 0);
2555 perf_pmu_enable(ctx->pmu);
2556 raw_spin_unlock(&ctx->lock);
2560 * Round-robin a context's events:
2562 static void rotate_ctx(struct perf_event_context *ctx)
2565 * Rotate the first entry last of non-pinned groups. Rotation might be
2566 * disabled by the inheritance code.
2568 if (!ctx->rotate_disable)
2569 list_rotate_left(&ctx->flexible_groups);
2573 * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
2574 * because they're strictly cpu affine and rotate_start is called with IRQs
2575 * disabled, while rotate_context is called from IRQ context.
2577 static void perf_rotate_context(struct perf_cpu_context *cpuctx)
2579 struct perf_event_context *ctx = NULL;
2580 int rotate = 0, remove = 1;
2582 if (cpuctx->ctx.nr_events) {
2584 if (cpuctx->ctx.nr_events != cpuctx->ctx.nr_active)
2588 ctx = cpuctx->task_ctx;
2589 if (ctx && ctx->nr_events) {
2591 if (ctx->nr_events != ctx->nr_active)
2598 perf_ctx_lock(cpuctx, cpuctx->task_ctx);
2599 perf_pmu_disable(cpuctx->ctx.pmu);
2601 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
2603 ctx_sched_out(ctx, cpuctx, EVENT_FLEXIBLE);
2605 rotate_ctx(&cpuctx->ctx);
2609 perf_event_sched_in(cpuctx, ctx, current);
2611 perf_pmu_enable(cpuctx->ctx.pmu);
2612 perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
2615 list_del_init(&cpuctx->rotation_list);
2618 #ifdef CONFIG_NO_HZ_FULL
2619 bool perf_event_can_stop_tick(void)
2621 if (list_empty(&__get_cpu_var(rotation_list)))
2628 void perf_event_task_tick(void)
2630 struct list_head *head = &__get_cpu_var(rotation_list);
2631 struct perf_cpu_context *cpuctx, *tmp;
2632 struct perf_event_context *ctx;
2635 WARN_ON(!irqs_disabled());
2637 __this_cpu_inc(perf_throttled_seq);
2638 throttled = __this_cpu_xchg(perf_throttled_count, 0);
2640 list_for_each_entry_safe(cpuctx, tmp, head, rotation_list) {
2642 perf_adjust_freq_unthr_context(ctx, throttled);
2644 ctx = cpuctx->task_ctx;
2646 perf_adjust_freq_unthr_context(ctx, throttled);
2648 if (cpuctx->jiffies_interval == 1 ||
2649 !(jiffies % cpuctx->jiffies_interval))
2650 perf_rotate_context(cpuctx);
2654 static int event_enable_on_exec(struct perf_event *event,
2655 struct perf_event_context *ctx)
2657 if (!event->attr.enable_on_exec)
2660 event->attr.enable_on_exec = 0;
2661 if (event->state >= PERF_EVENT_STATE_INACTIVE)
2664 __perf_event_mark_enabled(event);
2670 * Enable all of a task's events that have been marked enable-on-exec.
2671 * This expects task == current.
2673 static void perf_event_enable_on_exec(struct perf_event_context *ctx)
2675 struct perf_event *event;
2676 unsigned long flags;
2680 local_irq_save(flags);
2681 if (!ctx || !ctx->nr_events)
2685 * We must ctxsw out cgroup events to avoid conflict
2686 * when invoking perf_task_event_sched_in() later on
2687 * in this function. Otherwise we end up trying to
2688 * ctxswin cgroup events which are already scheduled
2691 perf_cgroup_sched_out(current, NULL);
2693 raw_spin_lock(&ctx->lock);
2694 task_ctx_sched_out(ctx);
2696 list_for_each_entry(event, &ctx->event_list, event_entry) {
2697 ret = event_enable_on_exec(event, ctx);
2703 * Unclone this context if we enabled any event.
2708 raw_spin_unlock(&ctx->lock);
2711 * Also calls ctxswin for cgroup events, if any:
2713 perf_event_context_sched_in(ctx, ctx->task);
2715 local_irq_restore(flags);
2719 * Cross CPU call to read the hardware event
2721 static void __perf_event_read(void *info)
2723 struct perf_event *event = info;
2724 struct perf_event_context *ctx = event->ctx;
2725 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
2728 * If this is a task context, we need to check whether it is
2729 * the current task context of this cpu. If not it has been
2730 * scheduled out before the smp call arrived. In that case
2731 * event->count would have been updated to a recent sample
2732 * when the event was scheduled out.
2734 if (ctx->task && cpuctx->task_ctx != ctx)
2737 raw_spin_lock(&ctx->lock);
2738 if (ctx->is_active) {
2739 update_context_time(ctx);
2740 update_cgrp_time_from_event(event);
2742 update_event_times(event);
2743 if (event->state == PERF_EVENT_STATE_ACTIVE)
2744 event->pmu->read(event);
2745 raw_spin_unlock(&ctx->lock);
2748 static inline u64 perf_event_count(struct perf_event *event)
2750 return local64_read(&event->count) + atomic64_read(&event->child_count);
2753 static u64 perf_event_read(struct perf_event *event)
2756 * If event is enabled and currently active on a CPU, update the
2757 * value in the event structure:
2759 if (event->state == PERF_EVENT_STATE_ACTIVE) {
2760 smp_call_function_single(event->oncpu,
2761 __perf_event_read, event, 1);
2762 } else if (event->state == PERF_EVENT_STATE_INACTIVE) {
2763 struct perf_event_context *ctx = event->ctx;
2764 unsigned long flags;
2766 raw_spin_lock_irqsave(&ctx->lock, flags);
2768 * may read while context is not active
2769 * (e.g., thread is blocked), in that case
2770 * we cannot update context time
2772 if (ctx->is_active) {
2773 update_context_time(ctx);
2774 update_cgrp_time_from_event(event);
2776 update_event_times(event);
2777 raw_spin_unlock_irqrestore(&ctx->lock, flags);
2780 return perf_event_count(event);
2784 * Initialize the perf_event context in a task_struct:
2786 static void __perf_event_init_context(struct perf_event_context *ctx)
2788 raw_spin_lock_init(&ctx->lock);
2789 mutex_init(&ctx->mutex);
2790 INIT_LIST_HEAD(&ctx->pinned_groups);
2791 INIT_LIST_HEAD(&ctx->flexible_groups);
2792 INIT_LIST_HEAD(&ctx->event_list);
2793 atomic_set(&ctx->refcount, 1);
2796 static struct perf_event_context *
2797 alloc_perf_context(struct pmu *pmu, struct task_struct *task)
2799 struct perf_event_context *ctx;
2801 ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL);
2805 __perf_event_init_context(ctx);
2808 get_task_struct(task);
2815 static struct task_struct *
2816 find_lively_task_by_vpid(pid_t vpid)
2818 struct task_struct *task;
2825 task = find_task_by_vpid(vpid);
2827 get_task_struct(task);
2831 return ERR_PTR(-ESRCH);
2833 /* Reuse ptrace permission checks for now. */
2835 if (!ptrace_may_access(task, PTRACE_MODE_READ))
2840 put_task_struct(task);
2841 return ERR_PTR(err);
2846 * Returns a matching context with refcount and pincount.
2848 static struct perf_event_context *
2849 find_get_context(struct pmu *pmu, struct task_struct *task, int cpu)
2851 struct perf_event_context *ctx;
2852 struct perf_cpu_context *cpuctx;
2853 unsigned long flags;
2857 /* Must be root to operate on a CPU event: */
2858 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
2859 return ERR_PTR(-EACCES);
2862 * We could be clever and allow to attach a event to an
2863 * offline CPU and activate it when the CPU comes up, but
2866 if (!cpu_online(cpu))
2867 return ERR_PTR(-ENODEV);
2869 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
2878 ctxn = pmu->task_ctx_nr;
2883 ctx = perf_lock_task_context(task, ctxn, &flags);
2887 raw_spin_unlock_irqrestore(&ctx->lock, flags);
2889 ctx = alloc_perf_context(pmu, task);
2895 mutex_lock(&task->perf_event_mutex);
2897 * If it has already passed perf_event_exit_task().
2898 * we must see PF_EXITING, it takes this mutex too.
2900 if (task->flags & PF_EXITING)
2902 else if (task->perf_event_ctxp[ctxn])
2907 rcu_assign_pointer(task->perf_event_ctxp[ctxn], ctx);
2909 mutex_unlock(&task->perf_event_mutex);
2911 if (unlikely(err)) {
2923 return ERR_PTR(err);
2926 static void perf_event_free_filter(struct perf_event *event);
2928 static void free_event_rcu(struct rcu_head *head)
2930 struct perf_event *event;
2932 event = container_of(head, struct perf_event, rcu_head);
2934 put_pid_ns(event->ns);
2935 perf_event_free_filter(event);
2939 static void ring_buffer_put(struct ring_buffer *rb);
2940 static void ring_buffer_detach(struct perf_event *event, struct ring_buffer *rb);
2942 static void free_event(struct perf_event *event)
2944 irq_work_sync(&event->pending);
2946 if (!event->parent) {
2947 if (event->attach_state & PERF_ATTACH_TASK)
2948 static_key_slow_dec_deferred(&perf_sched_events);
2949 if (event->attr.mmap || event->attr.mmap_data)
2950 atomic_dec(&nr_mmap_events);
2951 if (event->attr.comm)
2952 atomic_dec(&nr_comm_events);
2953 if (event->attr.task)
2954 atomic_dec(&nr_task_events);
2955 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN)
2956 put_callchain_buffers();
2957 if (is_cgroup_event(event)) {
2958 atomic_dec(&per_cpu(perf_cgroup_events, event->cpu));
2959 static_key_slow_dec_deferred(&perf_sched_events);
2962 if (has_branch_stack(event)) {
2963 static_key_slow_dec_deferred(&perf_sched_events);
2964 /* is system-wide event */
2965 if (!(event->attach_state & PERF_ATTACH_TASK)) {
2966 atomic_dec(&per_cpu(perf_branch_stack_events,
2973 struct ring_buffer *rb;
2976 * Can happen when we close an event with re-directed output.
2978 * Since we have a 0 refcount, perf_mmap_close() will skip
2979 * over us; possibly making our ring_buffer_put() the last.
2981 mutex_lock(&event->mmap_mutex);
2984 rcu_assign_pointer(event->rb, NULL);
2985 ring_buffer_detach(event, rb);
2986 ring_buffer_put(rb); /* could be last */
2988 mutex_unlock(&event->mmap_mutex);
2991 if (is_cgroup_event(event))
2992 perf_detach_cgroup(event);
2995 event->destroy(event);
2998 put_ctx(event->ctx);
3000 call_rcu(&event->rcu_head, free_event_rcu);
3003 int perf_event_release_kernel(struct perf_event *event)
3005 struct perf_event_context *ctx = event->ctx;
3007 WARN_ON_ONCE(ctx->parent_ctx);
3009 * There are two ways this annotation is useful:
3011 * 1) there is a lock recursion from perf_event_exit_task
3012 * see the comment there.
3014 * 2) there is a lock-inversion with mmap_sem through
3015 * perf_event_read_group(), which takes faults while
3016 * holding ctx->mutex, however this is called after
3017 * the last filedesc died, so there is no possibility
3018 * to trigger the AB-BA case.
3020 mutex_lock_nested(&ctx->mutex, SINGLE_DEPTH_NESTING);
3021 raw_spin_lock_irq(&ctx->lock);
3022 perf_group_detach(event);
3023 raw_spin_unlock_irq(&ctx->lock);
3024 perf_remove_from_context(event);
3025 mutex_unlock(&ctx->mutex);
3031 EXPORT_SYMBOL_GPL(perf_event_release_kernel);
3034 * Called when the last reference to the file is gone.
3036 static void put_event(struct perf_event *event)
3038 struct task_struct *owner;
3040 if (!atomic_long_dec_and_test(&event->refcount))
3044 owner = ACCESS_ONCE(event->owner);
3046 * Matches the smp_wmb() in perf_event_exit_task(). If we observe
3047 * !owner it means the list deletion is complete and we can indeed
3048 * free this event, otherwise we need to serialize on
3049 * owner->perf_event_mutex.
3051 smp_read_barrier_depends();
3054 * Since delayed_put_task_struct() also drops the last
3055 * task reference we can safely take a new reference
3056 * while holding the rcu_read_lock().
3058 get_task_struct(owner);
3063 mutex_lock(&owner->perf_event_mutex);
3065 * We have to re-check the event->owner field, if it is cleared
3066 * we raced with perf_event_exit_task(), acquiring the mutex
3067 * ensured they're done, and we can proceed with freeing the
3071 list_del_init(&event->owner_entry);
3072 mutex_unlock(&owner->perf_event_mutex);
3073 put_task_struct(owner);
3076 perf_event_release_kernel(event);
3079 static int perf_release(struct inode *inode, struct file *file)
3081 put_event(file->private_data);
3085 u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running)
3087 struct perf_event *child;
3093 mutex_lock(&event->child_mutex);
3094 total += perf_event_read(event);
3095 *enabled += event->total_time_enabled +
3096 atomic64_read(&event->child_total_time_enabled);
3097 *running += event->total_time_running +
3098 atomic64_read(&event->child_total_time_running);
3100 list_for_each_entry(child, &event->child_list, child_list) {
3101 total += perf_event_read(child);
3102 *enabled += child->total_time_enabled;
3103 *running += child->total_time_running;
3105 mutex_unlock(&event->child_mutex);
3109 EXPORT_SYMBOL_GPL(perf_event_read_value);
3111 static int perf_event_read_group(struct perf_event *event,
3112 u64 read_format, char __user *buf)
3114 struct perf_event *leader = event->group_leader, *sub;
3115 int n = 0, size = 0, ret = -EFAULT;
3116 struct perf_event_context *ctx = leader->ctx;
3118 u64 count, enabled, running;
3120 mutex_lock(&ctx->mutex);
3121 count = perf_event_read_value(leader, &enabled, &running);
3123 values[n++] = 1 + leader->nr_siblings;
3124 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3125 values[n++] = enabled;
3126 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3127 values[n++] = running;
3128 values[n++] = count;
3129 if (read_format & PERF_FORMAT_ID)
3130 values[n++] = primary_event_id(leader);
3132 size = n * sizeof(u64);
3134 if (copy_to_user(buf, values, size))
3139 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
3142 values[n++] = perf_event_read_value(sub, &enabled, &running);
3143 if (read_format & PERF_FORMAT_ID)
3144 values[n++] = primary_event_id(sub);
3146 size = n * sizeof(u64);
3148 if (copy_to_user(buf + ret, values, size)) {
3156 mutex_unlock(&ctx->mutex);
3161 static int perf_event_read_one(struct perf_event *event,
3162 u64 read_format, char __user *buf)
3164 u64 enabled, running;
3168 values[n++] = perf_event_read_value(event, &enabled, &running);
3169 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3170 values[n++] = enabled;
3171 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3172 values[n++] = running;
3173 if (read_format & PERF_FORMAT_ID)
3174 values[n++] = primary_event_id(event);
3176 if (copy_to_user(buf, values, n * sizeof(u64)))
3179 return n * sizeof(u64);
3183 * Read the performance event - simple non blocking version for now
3186 perf_read_hw(struct perf_event *event, char __user *buf, size_t count)
3188 u64 read_format = event->attr.read_format;
3192 * Return end-of-file for a read on a event that is in
3193 * error state (i.e. because it was pinned but it couldn't be
3194 * scheduled on to the CPU at some point).
3196 if (event->state == PERF_EVENT_STATE_ERROR)
3199 if (count < event->read_size)
3202 WARN_ON_ONCE(event->ctx->parent_ctx);
3203 if (read_format & PERF_FORMAT_GROUP)
3204 ret = perf_event_read_group(event, read_format, buf);
3206 ret = perf_event_read_one(event, read_format, buf);
3212 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
3214 struct perf_event *event = file->private_data;
3216 return perf_read_hw(event, buf, count);
3219 static unsigned int perf_poll(struct file *file, poll_table *wait)
3221 struct perf_event *event = file->private_data;
3222 struct ring_buffer *rb;
3223 unsigned int events = POLL_HUP;
3226 * Pin the event->rb by taking event->mmap_mutex; otherwise
3227 * perf_event_set_output() can swizzle our rb and make us miss wakeups.
3229 mutex_lock(&event->mmap_mutex);
3232 events = atomic_xchg(&rb->poll, 0);
3233 mutex_unlock(&event->mmap_mutex);
3235 poll_wait(file, &event->waitq, wait);
3240 static void perf_event_reset(struct perf_event *event)
3242 (void)perf_event_read(event);
3243 local64_set(&event->count, 0);
3244 perf_event_update_userpage(event);
3248 * Holding the top-level event's child_mutex means that any
3249 * descendant process that has inherited this event will block
3250 * in sync_child_event if it goes to exit, thus satisfying the
3251 * task existence requirements of perf_event_enable/disable.
3253 static void perf_event_for_each_child(struct perf_event *event,
3254 void (*func)(struct perf_event *))
3256 struct perf_event *child;
3258 WARN_ON_ONCE(event->ctx->parent_ctx);
3259 mutex_lock(&event->child_mutex);
3261 list_for_each_entry(child, &event->child_list, child_list)
3263 mutex_unlock(&event->child_mutex);
3266 static void perf_event_for_each(struct perf_event *event,
3267 void (*func)(struct perf_event *))
3269 struct perf_event_context *ctx = event->ctx;
3270 struct perf_event *sibling;
3272 WARN_ON_ONCE(ctx->parent_ctx);
3273 mutex_lock(&ctx->mutex);
3274 event = event->group_leader;
3276 perf_event_for_each_child(event, func);
3277 list_for_each_entry(sibling, &event->sibling_list, group_entry)
3278 perf_event_for_each_child(sibling, func);
3279 mutex_unlock(&ctx->mutex);
3282 static int perf_event_period(struct perf_event *event, u64 __user *arg)
3284 struct perf_event_context *ctx = event->ctx;
3288 if (!is_sampling_event(event))
3291 if (copy_from_user(&value, arg, sizeof(value)))
3297 raw_spin_lock_irq(&ctx->lock);
3298 if (event->attr.freq) {
3299 if (value > sysctl_perf_event_sample_rate) {
3304 event->attr.sample_freq = value;
3306 event->attr.sample_period = value;
3307 event->hw.sample_period = value;
3310 raw_spin_unlock_irq(&ctx->lock);
3315 static const struct file_operations perf_fops;
3317 static inline int perf_fget_light(int fd, struct fd *p)
3319 struct fd f = fdget(fd);
3323 if (f.file->f_op != &perf_fops) {
3331 static int perf_event_set_output(struct perf_event *event,
3332 struct perf_event *output_event);
3333 static int perf_event_set_filter(struct perf_event *event, void __user *arg);
3335 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
3337 struct perf_event *event = file->private_data;
3338 void (*func)(struct perf_event *);
3342 case PERF_EVENT_IOC_ENABLE:
3343 func = perf_event_enable;
3345 case PERF_EVENT_IOC_DISABLE:
3346 func = perf_event_disable;
3348 case PERF_EVENT_IOC_RESET:
3349 func = perf_event_reset;
3352 case PERF_EVENT_IOC_REFRESH:
3353 return perf_event_refresh(event, arg);
3355 case PERF_EVENT_IOC_PERIOD:
3356 return perf_event_period(event, (u64 __user *)arg);
3358 case PERF_EVENT_IOC_SET_OUTPUT:
3362 struct perf_event *output_event;
3364 ret = perf_fget_light(arg, &output);
3367 output_event = output.file->private_data;
3368 ret = perf_event_set_output(event, output_event);
3371 ret = perf_event_set_output(event, NULL);
3376 case PERF_EVENT_IOC_SET_FILTER:
3377 return perf_event_set_filter(event, (void __user *)arg);
3383 if (flags & PERF_IOC_FLAG_GROUP)
3384 perf_event_for_each(event, func);
3386 perf_event_for_each_child(event, func);
3391 int perf_event_task_enable(void)
3393 struct perf_event *event;
3395 mutex_lock(¤t->perf_event_mutex);
3396 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
3397 perf_event_for_each_child(event, perf_event_enable);
3398 mutex_unlock(¤t->perf_event_mutex);
3403 int perf_event_task_disable(void)
3405 struct perf_event *event;
3407 mutex_lock(¤t->perf_event_mutex);
3408 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
3409 perf_event_for_each_child(event, perf_event_disable);
3410 mutex_unlock(¤t->perf_event_mutex);
3415 static int perf_event_index(struct perf_event *event)
3417 if (event->hw.state & PERF_HES_STOPPED)
3420 if (event->state != PERF_EVENT_STATE_ACTIVE)
3423 return event->pmu->event_idx(event);
3426 static void calc_timer_values(struct perf_event *event,
3433 *now = perf_clock();
3434 ctx_time = event->shadow_ctx_time + *now;
3435 *enabled = ctx_time - event->tstamp_enabled;
3436 *running = ctx_time - event->tstamp_running;
3439 void __weak arch_perf_update_userpage(struct perf_event_mmap_page *userpg, u64 now)
3444 * Callers need to ensure there can be no nesting of this function, otherwise
3445 * the seqlock logic goes bad. We can not serialize this because the arch
3446 * code calls this from NMI context.
3448 void perf_event_update_userpage(struct perf_event *event)
3450 struct perf_event_mmap_page *userpg;
3451 struct ring_buffer *rb;
3452 u64 enabled, running, now;
3456 * compute total_time_enabled, total_time_running
3457 * based on snapshot values taken when the event
3458 * was last scheduled in.
3460 * we cannot simply called update_context_time()
3461 * because of locking issue as we can be called in
3464 calc_timer_values(event, &now, &enabled, &running);
3465 rb = rcu_dereference(event->rb);
3469 userpg = rb->user_page;
3472 * Disable preemption so as to not let the corresponding user-space
3473 * spin too long if we get preempted.
3478 userpg->index = perf_event_index(event);
3479 userpg->offset = perf_event_count(event);
3481 userpg->offset -= local64_read(&event->hw.prev_count);
3483 userpg->time_enabled = enabled +
3484 atomic64_read(&event->child_total_time_enabled);
3486 userpg->time_running = running +
3487 atomic64_read(&event->child_total_time_running);
3489 arch_perf_update_userpage(userpg, now);
3498 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
3500 struct perf_event *event = vma->vm_file->private_data;
3501 struct ring_buffer *rb;
3502 int ret = VM_FAULT_SIGBUS;
3504 if (vmf->flags & FAULT_FLAG_MKWRITE) {
3505 if (vmf->pgoff == 0)
3511 rb = rcu_dereference(event->rb);
3515 if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE))
3518 vmf->page = perf_mmap_to_page(rb, vmf->pgoff);
3522 get_page(vmf->page);
3523 vmf->page->mapping = vma->vm_file->f_mapping;
3524 vmf->page->index = vmf->pgoff;
3533 static void ring_buffer_attach(struct perf_event *event,
3534 struct ring_buffer *rb)
3536 unsigned long flags;
3538 if (!list_empty(&event->rb_entry))
3541 spin_lock_irqsave(&rb->event_lock, flags);
3542 if (list_empty(&event->rb_entry))
3543 list_add(&event->rb_entry, &rb->event_list);
3544 spin_unlock_irqrestore(&rb->event_lock, flags);
3547 static void ring_buffer_detach(struct perf_event *event, struct ring_buffer *rb)
3549 unsigned long flags;
3551 if (list_empty(&event->rb_entry))
3554 spin_lock_irqsave(&rb->event_lock, flags);
3555 list_del_init(&event->rb_entry);
3556 wake_up_all(&event->waitq);
3557 spin_unlock_irqrestore(&rb->event_lock, flags);
3560 static void ring_buffer_wakeup(struct perf_event *event)
3562 struct ring_buffer *rb;
3565 rb = rcu_dereference(event->rb);
3567 list_for_each_entry_rcu(event, &rb->event_list, rb_entry)
3568 wake_up_all(&event->waitq);
3573 static void rb_free_rcu(struct rcu_head *rcu_head)
3575 struct ring_buffer *rb;
3577 rb = container_of(rcu_head, struct ring_buffer, rcu_head);
3581 static struct ring_buffer *ring_buffer_get(struct perf_event *event)
3583 struct ring_buffer *rb;
3586 rb = rcu_dereference(event->rb);
3588 if (!atomic_inc_not_zero(&rb->refcount))
3596 static void ring_buffer_put(struct ring_buffer *rb)
3598 if (!atomic_dec_and_test(&rb->refcount))
3601 WARN_ON_ONCE(!list_empty(&rb->event_list));
3603 call_rcu(&rb->rcu_head, rb_free_rcu);
3606 static void perf_mmap_open(struct vm_area_struct *vma)
3608 struct perf_event *event = vma->vm_file->private_data;
3610 atomic_inc(&event->mmap_count);
3611 atomic_inc(&event->rb->mmap_count);
3615 * A buffer can be mmap()ed multiple times; either directly through the same
3616 * event, or through other events by use of perf_event_set_output().
3618 * In order to undo the VM accounting done by perf_mmap() we need to destroy
3619 * the buffer here, where we still have a VM context. This means we need
3620 * to detach all events redirecting to us.
3622 static void perf_mmap_close(struct vm_area_struct *vma)
3624 struct perf_event *event = vma->vm_file->private_data;
3626 struct ring_buffer *rb = event->rb;
3627 struct user_struct *mmap_user = rb->mmap_user;
3628 int mmap_locked = rb->mmap_locked;
3629 unsigned long size = perf_data_size(rb);
3631 atomic_dec(&rb->mmap_count);
3633 if (!atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex))
3636 /* Detach current event from the buffer. */
3637 rcu_assign_pointer(event->rb, NULL);
3638 ring_buffer_detach(event, rb);
3639 mutex_unlock(&event->mmap_mutex);
3641 /* If there's still other mmap()s of this buffer, we're done. */
3642 if (atomic_read(&rb->mmap_count)) {
3643 ring_buffer_put(rb); /* can't be last */
3648 * No other mmap()s, detach from all other events that might redirect
3649 * into the now unreachable buffer. Somewhat complicated by the
3650 * fact that rb::event_lock otherwise nests inside mmap_mutex.
3654 list_for_each_entry_rcu(event, &rb->event_list, rb_entry) {
3655 if (!atomic_long_inc_not_zero(&event->refcount)) {
3657 * This event is en-route to free_event() which will
3658 * detach it and remove it from the list.
3664 mutex_lock(&event->mmap_mutex);
3666 * Check we didn't race with perf_event_set_output() which can
3667 * swizzle the rb from under us while we were waiting to
3668 * acquire mmap_mutex.
3670 * If we find a different rb; ignore this event, a next
3671 * iteration will no longer find it on the list. We have to
3672 * still restart the iteration to make sure we're not now
3673 * iterating the wrong list.
3675 if (event->rb == rb) {
3676 rcu_assign_pointer(event->rb, NULL);
3677 ring_buffer_detach(event, rb);
3678 ring_buffer_put(rb); /* can't be last, we still have one */
3680 mutex_unlock(&event->mmap_mutex);
3684 * Restart the iteration; either we're on the wrong list or
3685 * destroyed its integrity by doing a deletion.
3692 * It could be there's still a few 0-ref events on the list; they'll
3693 * get cleaned up by free_event() -- they'll also still have their
3694 * ref on the rb and will free it whenever they are done with it.
3696 * Aside from that, this buffer is 'fully' detached and unmapped,
3697 * undo the VM accounting.
3700 atomic_long_sub((size >> PAGE_SHIFT) + 1, &mmap_user->locked_vm);
3701 vma->vm_mm->pinned_vm -= mmap_locked;
3702 free_uid(mmap_user);
3704 ring_buffer_put(rb); /* could be last */
3707 static const struct vm_operations_struct perf_mmap_vmops = {
3708 .open = perf_mmap_open,
3709 .close = perf_mmap_close,
3710 .fault = perf_mmap_fault,
3711 .page_mkwrite = perf_mmap_fault,
3714 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
3716 struct perf_event *event = file->private_data;
3717 unsigned long user_locked, user_lock_limit;
3718 struct user_struct *user = current_user();
3719 unsigned long locked, lock_limit;
3720 struct ring_buffer *rb;
3721 unsigned long vma_size;
3722 unsigned long nr_pages;
3723 long user_extra, extra;
3724 int ret = 0, flags = 0;
3727 * Don't allow mmap() of inherited per-task counters. This would
3728 * create a performance issue due to all children writing to the
3731 if (event->cpu == -1 && event->attr.inherit)
3734 if (!(vma->vm_flags & VM_SHARED))
3737 vma_size = vma->vm_end - vma->vm_start;
3738 nr_pages = (vma_size / PAGE_SIZE) - 1;
3741 * If we have rb pages ensure they're a power-of-two number, so we
3742 * can do bitmasks instead of modulo.
3744 if (nr_pages != 0 && !is_power_of_2(nr_pages))
3747 if (vma_size != PAGE_SIZE * (1 + nr_pages))
3750 if (vma->vm_pgoff != 0)
3753 WARN_ON_ONCE(event->ctx->parent_ctx);
3755 mutex_lock(&event->mmap_mutex);
3757 if (event->rb->nr_pages != nr_pages) {
3762 if (!atomic_inc_not_zero(&event->rb->mmap_count)) {
3764 * Raced against perf_mmap_close() through
3765 * perf_event_set_output(). Try again, hope for better
3768 mutex_unlock(&event->mmap_mutex);
3775 user_extra = nr_pages + 1;
3776 user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10);
3779 * Increase the limit linearly with more CPUs:
3781 user_lock_limit *= num_online_cpus();
3783 user_locked = atomic_long_read(&user->locked_vm) + user_extra;
3786 if (user_locked > user_lock_limit)
3787 extra = user_locked - user_lock_limit;
3789 lock_limit = rlimit(RLIMIT_MEMLOCK);
3790 lock_limit >>= PAGE_SHIFT;
3791 locked = vma->vm_mm->pinned_vm + extra;
3793 if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() &&
3794 !capable(CAP_IPC_LOCK)) {
3801 if (vma->vm_flags & VM_WRITE)
3802 flags |= RING_BUFFER_WRITABLE;
3804 rb = rb_alloc(nr_pages,
3805 event->attr.watermark ? event->attr.wakeup_watermark : 0,
3813 atomic_set(&rb->mmap_count, 1);
3814 rb->mmap_locked = extra;
3815 rb->mmap_user = get_current_user();
3817 atomic_long_add(user_extra, &user->locked_vm);
3818 vma->vm_mm->pinned_vm += extra;
3820 ring_buffer_attach(event, rb);
3821 rcu_assign_pointer(event->rb, rb);
3823 perf_event_update_userpage(event);
3827 atomic_inc(&event->mmap_count);
3828 mutex_unlock(&event->mmap_mutex);
3831 * Since pinned accounting is per vm we cannot allow fork() to copy our
3834 vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND | VM_DONTDUMP;
3835 vma->vm_ops = &perf_mmap_vmops;
3840 static int perf_fasync(int fd, struct file *filp, int on)
3842 struct inode *inode = file_inode(filp);
3843 struct perf_event *event = filp->private_data;
3846 mutex_lock(&inode->i_mutex);
3847 retval = fasync_helper(fd, filp, on, &event->fasync);
3848 mutex_unlock(&inode->i_mutex);
3856 static const struct file_operations perf_fops = {
3857 .llseek = no_llseek,
3858 .release = perf_release,
3861 .unlocked_ioctl = perf_ioctl,
3862 .compat_ioctl = perf_ioctl,
3864 .fasync = perf_fasync,
3870 * If there's data, ensure we set the poll() state and publish everything
3871 * to user-space before waking everybody up.
3874 void perf_event_wakeup(struct perf_event *event)
3876 ring_buffer_wakeup(event);
3878 if (event->pending_kill) {
3879 kill_fasync(&event->fasync, SIGIO, event->pending_kill);
3880 event->pending_kill = 0;
3884 static void perf_pending_event(struct irq_work *entry)
3886 struct perf_event *event = container_of(entry,
3887 struct perf_event, pending);
3889 if (event->pending_disable) {
3890 event->pending_disable = 0;
3891 __perf_event_disable(event);
3894 if (event->pending_wakeup) {
3895 event->pending_wakeup = 0;
3896 perf_event_wakeup(event);
3901 * We assume there is only KVM supporting the callbacks.
3902 * Later on, we might change it to a list if there is
3903 * another virtualization implementation supporting the callbacks.
3905 struct perf_guest_info_callbacks *perf_guest_cbs;
3907 int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
3909 perf_guest_cbs = cbs;
3912 EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks);
3914 int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
3916 perf_guest_cbs = NULL;
3919 EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks);
3922 perf_output_sample_regs(struct perf_output_handle *handle,
3923 struct pt_regs *regs, u64 mask)
3927 for_each_set_bit(bit, (const unsigned long *) &mask,
3928 sizeof(mask) * BITS_PER_BYTE) {
3931 val = perf_reg_value(regs, bit);
3932 perf_output_put(handle, val);
3936 static void perf_sample_regs_user(struct perf_regs_user *regs_user,
3937 struct pt_regs *regs)
3939 if (!user_mode(regs)) {
3941 regs = task_pt_regs(current);
3947 regs_user->regs = regs;
3948 regs_user->abi = perf_reg_abi(current);
3953 * Get remaining task size from user stack pointer.
3955 * It'd be better to take stack vma map and limit this more
3956 * precisly, but there's no way to get it safely under interrupt,
3957 * so using TASK_SIZE as limit.
3959 static u64 perf_ustack_task_size(struct pt_regs *regs)
3961 unsigned long addr = perf_user_stack_pointer(regs);
3963 if (!addr || addr >= TASK_SIZE)
3966 return TASK_SIZE - addr;
3970 perf_sample_ustack_size(u16 stack_size, u16 header_size,
3971 struct pt_regs *regs)
3975 /* No regs, no stack pointer, no dump. */
3980 * Check if we fit in with the requested stack size into the:
3982 * If we don't, we limit the size to the TASK_SIZE.
3984 * - remaining sample size
3985 * If we don't, we customize the stack size to
3986 * fit in to the remaining sample size.
3989 task_size = min((u64) USHRT_MAX, perf_ustack_task_size(regs));
3990 stack_size = min(stack_size, (u16) task_size);
3992 /* Current header size plus static size and dynamic size. */
3993 header_size += 2 * sizeof(u64);
3995 /* Do we fit in with the current stack dump size? */
3996 if ((u16) (header_size + stack_size) < header_size) {
3998 * If we overflow the maximum size for the sample,
3999 * we customize the stack dump size to fit in.
4001 stack_size = USHRT_MAX - header_size - sizeof(u64);
4002 stack_size = round_up(stack_size, sizeof(u64));
4009 perf_output_sample_ustack(struct perf_output_handle *handle, u64 dump_size,
4010 struct pt_regs *regs)
4012 /* Case of a kernel thread, nothing to dump */
4015 perf_output_put(handle, size);
4024 * - the size requested by user or the best one we can fit
4025 * in to the sample max size
4027 * - user stack dump data
4029 * - the actual dumped size
4033 perf_output_put(handle, dump_size);
4036 sp = perf_user_stack_pointer(regs);
4037 rem = __output_copy_user(handle, (void *) sp, dump_size);
4038 dyn_size = dump_size - rem;
4040 perf_output_skip(handle, rem);
4043 perf_output_put(handle, dyn_size);
4047 static void __perf_event_header__init_id(struct perf_event_header *header,
4048 struct perf_sample_data *data,
4049 struct perf_event *event)
4051 u64 sample_type = event->attr.sample_type;
4053 data->type = sample_type;
4054 header->size += event->id_header_size;
4056 if (sample_type & PERF_SAMPLE_TID) {
4057 /* namespace issues */
4058 data->tid_entry.pid = perf_event_pid(event, current);
4059 data->tid_entry.tid = perf_event_tid(event, current);
4062 if (sample_type & PERF_SAMPLE_TIME)
4063 data->time = perf_clock();
4065 if (sample_type & PERF_SAMPLE_ID)
4066 data->id = primary_event_id(event);
4068 if (sample_type & PERF_SAMPLE_STREAM_ID)
4069 data->stream_id = event->id;
4071 if (sample_type & PERF_SAMPLE_CPU) {
4072 data->cpu_entry.cpu = raw_smp_processor_id();
4073 data->cpu_entry.reserved = 0;
4077 void perf_event_header__init_id(struct perf_event_header *header,
4078 struct perf_sample_data *data,
4079 struct perf_event *event)
4081 if (event->attr.sample_id_all)
4082 __perf_event_header__init_id(header, data, event);
4085 static void __perf_event__output_id_sample(struct perf_output_handle *handle,
4086 struct perf_sample_data *data)
4088 u64 sample_type = data->type;
4090 if (sample_type & PERF_SAMPLE_TID)
4091 perf_output_put(handle, data->tid_entry);
4093 if (sample_type & PERF_SAMPLE_TIME)
4094 perf_output_put(handle, data->time);
4096 if (sample_type & PERF_SAMPLE_ID)
4097 perf_output_put(handle, data->id);
4099 if (sample_type & PERF_SAMPLE_STREAM_ID)
4100 perf_output_put(handle, data->stream_id);
4102 if (sample_type & PERF_SAMPLE_CPU)
4103 perf_output_put(handle, data->cpu_entry);
4106 void perf_event__output_id_sample(struct perf_event *event,
4107 struct perf_output_handle *handle,
4108 struct perf_sample_data *sample)
4110 if (event->attr.sample_id_all)
4111 __perf_event__output_id_sample(handle, sample);
4114 static void perf_output_read_one(struct perf_output_handle *handle,
4115 struct perf_event *event,
4116 u64 enabled, u64 running)
4118 u64 read_format = event->attr.read_format;
4122 values[n++] = perf_event_count(event);
4123 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
4124 values[n++] = enabled +
4125 atomic64_read(&event->child_total_time_enabled);
4127 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
4128 values[n++] = running +
4129 atomic64_read(&event->child_total_time_running);
4131 if (read_format & PERF_FORMAT_ID)
4132 values[n++] = primary_event_id(event);
4134 __output_copy(handle, values, n * sizeof(u64));
4138 * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
4140 static void perf_output_read_group(struct perf_output_handle *handle,
4141 struct perf_event *event,
4142 u64 enabled, u64 running)
4144 struct perf_event *leader = event->group_leader, *sub;
4145 u64 read_format = event->attr.read_format;
4149 values[n++] = 1 + leader->nr_siblings;
4151 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
4152 values[n++] = enabled;
4154 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
4155 values[n++] = running;
4157 if (leader != event)
4158 leader->pmu->read(leader);
4160 values[n++] = perf_event_count(leader);
4161 if (read_format & PERF_FORMAT_ID)
4162 values[n++] = primary_event_id(leader);
4164 __output_copy(handle, values, n * sizeof(u64));
4166 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
4170 sub->pmu->read(sub);
4172 values[n++] = perf_event_count(sub);
4173 if (read_format & PERF_FORMAT_ID)
4174 values[n++] = primary_event_id(sub);
4176 __output_copy(handle, values, n * sizeof(u64));
4180 #define PERF_FORMAT_TOTAL_TIMES (PERF_FORMAT_TOTAL_TIME_ENABLED|\
4181 PERF_FORMAT_TOTAL_TIME_RUNNING)
4183 static void perf_output_read(struct perf_output_handle *handle,
4184 struct perf_event *event)
4186 u64 enabled = 0, running = 0, now;
4187 u64 read_format = event->attr.read_format;
4190 * compute total_time_enabled, total_time_running
4191 * based on snapshot values taken when the event
4192 * was last scheduled in.
4194 * we cannot simply called update_context_time()
4195 * because of locking issue as we are called in
4198 if (read_format & PERF_FORMAT_TOTAL_TIMES)
4199 calc_timer_values(event, &now, &enabled, &running);
4201 if (event->attr.read_format & PERF_FORMAT_GROUP)
4202 perf_output_read_group(handle, event, enabled, running);
4204 perf_output_read_one(handle, event, enabled, running);
4207 void perf_output_sample(struct perf_output_handle *handle,
4208 struct perf_event_header *header,
4209 struct perf_sample_data *data,
4210 struct perf_event *event)
4212 u64 sample_type = data->type;
4214 perf_output_put(handle, *header);
4216 if (sample_type & PERF_SAMPLE_IP)
4217 perf_output_put(handle, data->ip);
4219 if (sample_type & PERF_SAMPLE_TID)
4220 perf_output_put(handle, data->tid_entry);
4222 if (sample_type & PERF_SAMPLE_TIME)
4223 perf_output_put(handle, data->time);
4225 if (sample_type & PERF_SAMPLE_ADDR)
4226 perf_output_put(handle, data->addr);
4228 if (sample_type & PERF_SAMPLE_ID)
4229 perf_output_put(handle, data->id);
4231 if (sample_type & PERF_SAMPLE_STREAM_ID)
4232 perf_output_put(handle, data->stream_id);
4234 if (sample_type & PERF_SAMPLE_CPU)
4235 perf_output_put(handle, data->cpu_entry);
4237 if (sample_type & PERF_SAMPLE_PERIOD)
4238 perf_output_put(handle, data->period);
4240 if (sample_type & PERF_SAMPLE_READ)
4241 perf_output_read(handle, event);
4243 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
4244 if (data->callchain) {
4247 if (data->callchain)
4248 size += data->callchain->nr;
4250 size *= sizeof(u64);
4252 __output_copy(handle, data->callchain, size);
4255 perf_output_put(handle, nr);
4259 if (sample_type & PERF_SAMPLE_RAW) {
4261 perf_output_put(handle, data->raw->size);
4262 __output_copy(handle, data->raw->data,
4269 .size = sizeof(u32),
4272 perf_output_put(handle, raw);
4276 if (!event->attr.watermark) {
4277 int wakeup_events = event->attr.wakeup_events;
4279 if (wakeup_events) {
4280 struct ring_buffer *rb = handle->rb;
4281 int events = local_inc_return(&rb->events);
4283 if (events >= wakeup_events) {
4284 local_sub(wakeup_events, &rb->events);
4285 local_inc(&rb->wakeup);
4290 if (sample_type & PERF_SAMPLE_BRANCH_STACK) {
4291 if (data->br_stack) {
4294 size = data->br_stack->nr
4295 * sizeof(struct perf_branch_entry);
4297 perf_output_put(handle, data->br_stack->nr);
4298 perf_output_copy(handle, data->br_stack->entries, size);
4301 * we always store at least the value of nr
4304 perf_output_put(handle, nr);
4308 if (sample_type & PERF_SAMPLE_REGS_USER) {
4309 u64 abi = data->regs_user.abi;
4312 * If there are no regs to dump, notice it through
4313 * first u64 being zero (PERF_SAMPLE_REGS_ABI_NONE).
4315 perf_output_put(handle, abi);
4318 u64 mask = event->attr.sample_regs_user;
4319 perf_output_sample_regs(handle,
4320 data->regs_user.regs,
4325 if (sample_type & PERF_SAMPLE_STACK_USER)
4326 perf_output_sample_ustack(handle,
4327 data->stack_user_size,
4328 data->regs_user.regs);
4330 if (sample_type & PERF_SAMPLE_WEIGHT)
4331 perf_output_put(handle, data->weight);
4333 if (sample_type & PERF_SAMPLE_DATA_SRC)
4334 perf_output_put(handle, data->data_src.val);
4337 void perf_prepare_sample(struct perf_event_header *header,
4338 struct perf_sample_data *data,
4339 struct perf_event *event,
4340 struct pt_regs *regs)
4342 u64 sample_type = event->attr.sample_type;
4344 header->type = PERF_RECORD_SAMPLE;
4345 header->size = sizeof(*header) + event->header_size;
4348 header->misc |= perf_misc_flags(regs);
4350 __perf_event_header__init_id(header, data, event);
4352 if (sample_type & PERF_SAMPLE_IP)
4353 data->ip = perf_instruction_pointer(regs);
4355 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
4358 data->callchain = perf_callchain(event, regs);
4360 if (data->callchain)
4361 size += data->callchain->nr;
4363 header->size += size * sizeof(u64);
4366 if (sample_type & PERF_SAMPLE_RAW) {
4367 int size = sizeof(u32);
4370 size += data->raw->size;
4372 size += sizeof(u32);
4374 WARN_ON_ONCE(size & (sizeof(u64)-1));
4375 header->size += size;
4378 if (sample_type & PERF_SAMPLE_BRANCH_STACK) {
4379 int size = sizeof(u64); /* nr */
4380 if (data->br_stack) {
4381 size += data->br_stack->nr
4382 * sizeof(struct perf_branch_entry);
4384 header->size += size;
4387 if (sample_type & PERF_SAMPLE_REGS_USER) {
4388 /* regs dump ABI info */
4389 int size = sizeof(u64);
4391 perf_sample_regs_user(&data->regs_user, regs);
4393 if (data->regs_user.regs) {
4394 u64 mask = event->attr.sample_regs_user;
4395 size += hweight64(mask) * sizeof(u64);
4398 header->size += size;
4401 if (sample_type & PERF_SAMPLE_STACK_USER) {
4403 * Either we need PERF_SAMPLE_STACK_USER bit to be allways
4404 * processed as the last one or have additional check added
4405 * in case new sample type is added, because we could eat
4406 * up the rest of the sample size.
4408 struct perf_regs_user *uregs = &data->regs_user;
4409 u16 stack_size = event->attr.sample_stack_user;
4410 u16 size = sizeof(u64);
4413 perf_sample_regs_user(uregs, regs);
4415 stack_size = perf_sample_ustack_size(stack_size, header->size,
4419 * If there is something to dump, add space for the dump
4420 * itself and for the field that tells the dynamic size,
4421 * which is how many have been actually dumped.
4424 size += sizeof(u64) + stack_size;
4426 data->stack_user_size = stack_size;
4427 header->size += size;
4431 static void perf_event_output(struct perf_event *event,
4432 struct perf_sample_data *data,
4433 struct pt_regs *regs)
4435 struct perf_output_handle handle;
4436 struct perf_event_header header;
4438 /* protect the callchain buffers */
4441 perf_prepare_sample(&header, data, event, regs);
4443 if (perf_output_begin(&handle, event, header.size))
4446 perf_output_sample(&handle, &header, data, event);
4448 perf_output_end(&handle);
4458 struct perf_read_event {
4459 struct perf_event_header header;
4466 perf_event_read_event(struct perf_event *event,
4467 struct task_struct *task)
4469 struct perf_output_handle handle;
4470 struct perf_sample_data sample;
4471 struct perf_read_event read_event = {
4473 .type = PERF_RECORD_READ,
4475 .size = sizeof(read_event) + event->read_size,
4477 .pid = perf_event_pid(event, task),
4478 .tid = perf_event_tid(event, task),
4482 perf_event_header__init_id(&read_event.header, &sample, event);
4483 ret = perf_output_begin(&handle, event, read_event.header.size);
4487 perf_output_put(&handle, read_event);
4488 perf_output_read(&handle, event);
4489 perf_event__output_id_sample(event, &handle, &sample);
4491 perf_output_end(&handle);
4494 typedef int (perf_event_aux_match_cb)(struct perf_event *event, void *data);
4495 typedef void (perf_event_aux_output_cb)(struct perf_event *event, void *data);
4498 perf_event_aux_ctx(struct perf_event_context *ctx,
4499 perf_event_aux_match_cb match,
4500 perf_event_aux_output_cb output,
4503 struct perf_event *event;
4505 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
4506 if (event->state < PERF_EVENT_STATE_INACTIVE)
4508 if (!event_filter_match(event))
4510 if (match(event, data))
4511 output(event, data);
4516 perf_event_aux(perf_event_aux_match_cb match,
4517 perf_event_aux_output_cb output,
4519 struct perf_event_context *task_ctx)
4521 struct perf_cpu_context *cpuctx;
4522 struct perf_event_context *ctx;
4527 list_for_each_entry_rcu(pmu, &pmus, entry) {
4528 cpuctx = get_cpu_ptr(pmu->pmu_cpu_context);
4529 if (cpuctx->unique_pmu != pmu)
4531 perf_event_aux_ctx(&cpuctx->ctx, match, output, data);
4534 ctxn = pmu->task_ctx_nr;
4537 ctx = rcu_dereference(current->perf_event_ctxp[ctxn]);
4539 perf_event_aux_ctx(ctx, match, output, data);
4541 put_cpu_ptr(pmu->pmu_cpu_context);
4546 perf_event_aux_ctx(task_ctx, match, output, data);
4553 * task tracking -- fork/exit
4555 * enabled by: attr.comm | attr.mmap | attr.mmap_data | attr.task
4558 struct perf_task_event {
4559 struct task_struct *task;
4560 struct perf_event_context *task_ctx;
4563 struct perf_event_header header;
4573 static void perf_event_task_output(struct perf_event *event,
4576 struct perf_task_event *task_event = data;
4577 struct perf_output_handle handle;
4578 struct perf_sample_data sample;
4579 struct task_struct *task = task_event->task;
4580 int ret, size = task_event->event_id.header.size;
4582 perf_event_header__init_id(&task_event->event_id.header, &sample, event);
4584 ret = perf_output_begin(&handle, event,
4585 task_event->event_id.header.size);
4589 task_event->event_id.pid = perf_event_pid(event, task);
4590 task_event->event_id.ppid = perf_event_pid(event, current);
4592 task_event->event_id.tid = perf_event_tid(event, task);
4593 task_event->event_id.ptid = perf_event_tid(event, current);
4595 perf_output_put(&handle, task_event->event_id);
4597 perf_event__output_id_sample(event, &handle, &sample);
4599 perf_output_end(&handle);
4601 task_event->event_id.header.size = size;
4604 static int perf_event_task_match(struct perf_event *event,
4605 void *data __maybe_unused)
4607 return event->attr.comm || event->attr.mmap ||
4608 event->attr.mmap_data || event->attr.task;
4611 static void perf_event_task(struct task_struct *task,
4612 struct perf_event_context *task_ctx,
4615 struct perf_task_event task_event;
4617 if (!atomic_read(&nr_comm_events) &&
4618 !atomic_read(&nr_mmap_events) &&
4619 !atomic_read(&nr_task_events))
4622 task_event = (struct perf_task_event){
4624 .task_ctx = task_ctx,
4627 .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT,
4629 .size = sizeof(task_event.event_id),
4635 .time = perf_clock(),
4639 perf_event_aux(perf_event_task_match,
4640 perf_event_task_output,
4645 void perf_event_fork(struct task_struct *task)
4647 perf_event_task(task, NULL, 1);
4654 struct perf_comm_event {
4655 struct task_struct *task;
4660 struct perf_event_header header;
4667 static void perf_event_comm_output(struct perf_event *event,
4670 struct perf_comm_event *comm_event = data;
4671 struct perf_output_handle handle;
4672 struct perf_sample_data sample;
4673 int size = comm_event->event_id.header.size;
4676 perf_event_header__init_id(&comm_event->event_id.header, &sample, event);
4677 ret = perf_output_begin(&handle, event,
4678 comm_event->event_id.header.size);
4683 comm_event->event_id.pid = perf_event_pid(event, comm_event->task);
4684 comm_event->event_id.tid = perf_event_tid(event, comm_event->task);
4686 perf_output_put(&handle, comm_event->event_id);
4687 __output_copy(&handle, comm_event->comm,
4688 comm_event->comm_size);
4690 perf_event__output_id_sample(event, &handle, &sample);
4692 perf_output_end(&handle);
4694 comm_event->event_id.header.size = size;
4697 static int perf_event_comm_match(struct perf_event *event,
4698 void *data __maybe_unused)
4700 return event->attr.comm;
4703 static void perf_event_comm_event(struct perf_comm_event *comm_event)
4705 char comm[TASK_COMM_LEN];
4708 memset(comm, 0, sizeof(comm));
4709 strlcpy(comm, comm_event->task->comm, sizeof(comm));
4710 size = ALIGN(strlen(comm)+1, sizeof(u64));
4712 comm_event->comm = comm;
4713 comm_event->comm_size = size;
4715 comm_event->event_id.header.size = sizeof(comm_event->event_id) + size;
4717 perf_event_aux(perf_event_comm_match,
4718 perf_event_comm_output,
4723 void perf_event_comm(struct task_struct *task)
4725 struct perf_comm_event comm_event;
4726 struct perf_event_context *ctx;
4730 for_each_task_context_nr(ctxn) {
4731 ctx = task->perf_event_ctxp[ctxn];
4735 perf_event_enable_on_exec(ctx);
4739 if (!atomic_read(&nr_comm_events))
4742 comm_event = (struct perf_comm_event){
4748 .type = PERF_RECORD_COMM,
4757 perf_event_comm_event(&comm_event);
4764 struct perf_mmap_event {
4765 struct vm_area_struct *vma;
4767 const char *file_name;
4771 struct perf_event_header header;
4781 static void perf_event_mmap_output(struct perf_event *event,
4784 struct perf_mmap_event *mmap_event = data;
4785 struct perf_output_handle handle;
4786 struct perf_sample_data sample;
4787 int size = mmap_event->event_id.header.size;
4790 perf_event_header__init_id(&mmap_event->event_id.header, &sample, event);
4791 ret = perf_output_begin(&handle, event,
4792 mmap_event->event_id.header.size);
4796 mmap_event->event_id.pid = perf_event_pid(event, current);
4797 mmap_event->event_id.tid = perf_event_tid(event, current);
4799 perf_output_put(&handle, mmap_event->event_id);
4800 __output_copy(&handle, mmap_event->file_name,
4801 mmap_event->file_size);
4803 perf_event__output_id_sample(event, &handle, &sample);
4805 perf_output_end(&handle);
4807 mmap_event->event_id.header.size = size;
4810 static int perf_event_mmap_match(struct perf_event *event,
4813 struct perf_mmap_event *mmap_event = data;
4814 struct vm_area_struct *vma = mmap_event->vma;
4815 int executable = vma->vm_flags & VM_EXEC;
4817 return (!executable && event->attr.mmap_data) ||
4818 (executable && event->attr.mmap);
4821 static void perf_event_mmap_event(struct perf_mmap_event *mmap_event)
4823 struct vm_area_struct *vma = mmap_event->vma;
4824 struct file *file = vma->vm_file;
4830 memset(tmp, 0, sizeof(tmp));
4834 * d_path works from the end of the rb backwards, so we
4835 * need to add enough zero bytes after the string to handle
4836 * the 64bit alignment we do later.
4838 buf = kzalloc(PATH_MAX + sizeof(u64), GFP_KERNEL);
4840 name = strncpy(tmp, "//enomem", sizeof(tmp));
4843 name = d_path(&file->f_path, buf, PATH_MAX);
4845 name = strncpy(tmp, "//toolong", sizeof(tmp));
4849 if (arch_vma_name(mmap_event->vma)) {
4850 name = strncpy(tmp, arch_vma_name(mmap_event->vma),
4852 tmp[sizeof(tmp) - 1] = '\0';
4857 name = strncpy(tmp, "[vdso]", sizeof(tmp));
4859 } else if (vma->vm_start <= vma->vm_mm->start_brk &&
4860 vma->vm_end >= vma->vm_mm->brk) {
4861 name = strncpy(tmp, "[heap]", sizeof(tmp));
4863 } else if (vma->vm_start <= vma->vm_mm->start_stack &&
4864 vma->vm_end >= vma->vm_mm->start_stack) {
4865 name = strncpy(tmp, "[stack]", sizeof(tmp));
4869 name = strncpy(tmp, "//anon", sizeof(tmp));
4874 size = ALIGN(strlen(name)+1, sizeof(u64));
4876 mmap_event->file_name = name;
4877 mmap_event->file_size = size;
4879 if (!(vma->vm_flags & VM_EXEC))
4880 mmap_event->event_id.header.misc |= PERF_RECORD_MISC_MMAP_DATA;
4882 mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size;
4884 perf_event_aux(perf_event_mmap_match,
4885 perf_event_mmap_output,
4892 void perf_event_mmap(struct vm_area_struct *vma)
4894 struct perf_mmap_event mmap_event;
4896 if (!atomic_read(&nr_mmap_events))
4899 mmap_event = (struct perf_mmap_event){
4905 .type = PERF_RECORD_MMAP,
4906 .misc = PERF_RECORD_MISC_USER,
4911 .start = vma->vm_start,
4912 .len = vma->vm_end - vma->vm_start,
4913 .pgoff = (u64)vma->vm_pgoff << PAGE_SHIFT,
4917 perf_event_mmap_event(&mmap_event);
4921 * IRQ throttle logging
4924 static void perf_log_throttle(struct perf_event *event, int enable)
4926 struct perf_output_handle handle;
4927 struct perf_sample_data sample;
4931 struct perf_event_header header;
4935 } throttle_event = {
4937 .type = PERF_RECORD_THROTTLE,
4939 .size = sizeof(throttle_event),
4941 .time = perf_clock(),
4942 .id = primary_event_id(event),
4943 .stream_id = event->id,
4947 throttle_event.header.type = PERF_RECORD_UNTHROTTLE;
4949 perf_event_header__init_id(&throttle_event.header, &sample, event);
4951 ret = perf_output_begin(&handle, event,
4952 throttle_event.header.size);
4956 perf_output_put(&handle, throttle_event);
4957 perf_event__output_id_sample(event, &handle, &sample);
4958 perf_output_end(&handle);
4962 * Generic event overflow handling, sampling.
4965 static int __perf_event_overflow(struct perf_event *event,
4966 int throttle, struct perf_sample_data *data,
4967 struct pt_regs *regs)
4969 int events = atomic_read(&event->event_limit);
4970 struct hw_perf_event *hwc = &event->hw;
4975 * Non-sampling counters might still use the PMI to fold short
4976 * hardware counters, ignore those.
4978 if (unlikely(!is_sampling_event(event)))
4981 seq = __this_cpu_read(perf_throttled_seq);
4982 if (seq != hwc->interrupts_seq) {
4983 hwc->interrupts_seq = seq;
4984 hwc->interrupts = 1;
4987 if (unlikely(throttle
4988 && hwc->interrupts >= max_samples_per_tick)) {
4989 __this_cpu_inc(perf_throttled_count);
4990 hwc->interrupts = MAX_INTERRUPTS;
4991 perf_log_throttle(event, 0);
4996 if (event->attr.freq) {
4997 u64 now = perf_clock();
4998 s64 delta = now - hwc->freq_time_stamp;
5000 hwc->freq_time_stamp = now;
5002 if (delta > 0 && delta < 2*TICK_NSEC)
5003 perf_adjust_period(event, delta, hwc->last_period, true);
5007 * XXX event_limit might not quite work as expected on inherited
5011 event->pending_kill = POLL_IN;
5012 if (events && atomic_dec_and_test(&event->event_limit)) {
5014 event->pending_kill = POLL_HUP;
5015 event->pending_disable = 1;
5016 irq_work_queue(&event->pending);
5019 if (event->overflow_handler)
5020 event->overflow_handler(event, data, regs);
5022 perf_event_output(event, data, regs);
5024 if (event->fasync && event->pending_kill) {
5025 event->pending_wakeup = 1;
5026 irq_work_queue(&event->pending);
5032 int perf_event_overflow(struct perf_event *event,
5033 struct perf_sample_data *data,
5034 struct pt_regs *regs)
5036 return __perf_event_overflow(event, 1, data, regs);
5040 * Generic software event infrastructure
5043 struct swevent_htable {
5044 struct swevent_hlist *swevent_hlist;
5045 struct mutex hlist_mutex;
5048 /* Recursion avoidance in each contexts */
5049 int recursion[PERF_NR_CONTEXTS];
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);
5296 if (WARN_ON_ONCE(!head))
5299 hlist_add_head_rcu(&event->hlist_entry, head);
5304 static void perf_swevent_del(struct perf_event *event, int flags)
5306 hlist_del_rcu(&event->hlist_entry);
5309 static void perf_swevent_start(struct perf_event *event, int flags)
5311 event->hw.state = 0;
5314 static void perf_swevent_stop(struct perf_event *event, int flags)
5316 event->hw.state = PERF_HES_STOPPED;
5319 /* Deref the hlist from the update side */
5320 static inline struct swevent_hlist *
5321 swevent_hlist_deref(struct swevent_htable *swhash)
5323 return rcu_dereference_protected(swhash->swevent_hlist,
5324 lockdep_is_held(&swhash->hlist_mutex));
5327 static void swevent_hlist_release(struct swevent_htable *swhash)
5329 struct swevent_hlist *hlist = swevent_hlist_deref(swhash);
5334 rcu_assign_pointer(swhash->swevent_hlist, NULL);
5335 kfree_rcu(hlist, rcu_head);
5338 static void swevent_hlist_put_cpu(struct perf_event *event, int cpu)
5340 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
5342 mutex_lock(&swhash->hlist_mutex);
5344 if (!--swhash->hlist_refcount)
5345 swevent_hlist_release(swhash);
5347 mutex_unlock(&swhash->hlist_mutex);
5350 static void swevent_hlist_put(struct perf_event *event)
5354 if (event->cpu != -1) {
5355 swevent_hlist_put_cpu(event, event->cpu);
5359 for_each_possible_cpu(cpu)
5360 swevent_hlist_put_cpu(event, cpu);
5363 static int swevent_hlist_get_cpu(struct perf_event *event, int cpu)
5365 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
5368 mutex_lock(&swhash->hlist_mutex);
5370 if (!swevent_hlist_deref(swhash) && cpu_online(cpu)) {
5371 struct swevent_hlist *hlist;
5373 hlist = kzalloc(sizeof(*hlist), GFP_KERNEL);
5378 rcu_assign_pointer(swhash->swevent_hlist, hlist);
5380 swhash->hlist_refcount++;
5382 mutex_unlock(&swhash->hlist_mutex);
5387 static int swevent_hlist_get(struct perf_event *event)
5390 int cpu, failed_cpu;
5392 if (event->cpu != -1)
5393 return swevent_hlist_get_cpu(event, event->cpu);
5396 for_each_possible_cpu(cpu) {
5397 err = swevent_hlist_get_cpu(event, cpu);
5407 for_each_possible_cpu(cpu) {
5408 if (cpu == failed_cpu)
5410 swevent_hlist_put_cpu(event, cpu);
5417 struct static_key perf_swevent_enabled[PERF_COUNT_SW_MAX];
5419 static void sw_perf_event_destroy(struct perf_event *event)
5421 u64 event_id = event->attr.config;
5423 WARN_ON(event->parent);
5425 static_key_slow_dec(&perf_swevent_enabled[event_id]);
5426 swevent_hlist_put(event);
5429 static int perf_swevent_init(struct perf_event *event)
5431 u64 event_id = event->attr.config;
5433 if (event->attr.type != PERF_TYPE_SOFTWARE)
5437 * no branch sampling for software events
5439 if (has_branch_stack(event))
5443 case PERF_COUNT_SW_CPU_CLOCK:
5444 case PERF_COUNT_SW_TASK_CLOCK:
5451 if (event_id >= PERF_COUNT_SW_MAX)
5454 if (!event->parent) {
5457 err = swevent_hlist_get(event);
5461 static_key_slow_inc(&perf_swevent_enabled[event_id]);
5462 event->destroy = sw_perf_event_destroy;
5468 static int perf_swevent_event_idx(struct perf_event *event)
5473 static struct pmu perf_swevent = {
5474 .task_ctx_nr = perf_sw_context,
5476 .event_init = perf_swevent_init,
5477 .add = perf_swevent_add,
5478 .del = perf_swevent_del,
5479 .start = perf_swevent_start,
5480 .stop = perf_swevent_stop,
5481 .read = perf_swevent_read,
5483 .event_idx = perf_swevent_event_idx,
5486 #ifdef CONFIG_EVENT_TRACING
5488 static int perf_tp_filter_match(struct perf_event *event,
5489 struct perf_sample_data *data)
5491 void *record = data->raw->data;
5493 if (likely(!event->filter) || filter_match_preds(event->filter, record))
5498 static int perf_tp_event_match(struct perf_event *event,
5499 struct perf_sample_data *data,
5500 struct pt_regs *regs)
5502 if (event->hw.state & PERF_HES_STOPPED)
5505 * All tracepoints are from kernel-space.
5507 if (event->attr.exclude_kernel)
5510 if (!perf_tp_filter_match(event, data))
5516 void perf_tp_event(u64 addr, u64 count, void *record, int entry_size,
5517 struct pt_regs *regs, struct hlist_head *head, int rctx,
5518 struct task_struct *task)
5520 struct perf_sample_data data;
5521 struct perf_event *event;
5523 struct perf_raw_record raw = {
5528 perf_sample_data_init(&data, addr, 0);
5531 hlist_for_each_entry_rcu(event, head, hlist_entry) {
5532 if (perf_tp_event_match(event, &data, regs))
5533 perf_swevent_event(event, count, &data, regs);
5537 * If we got specified a target task, also iterate its context and
5538 * deliver this event there too.
5540 if (task && task != current) {
5541 struct perf_event_context *ctx;
5542 struct trace_entry *entry = record;
5545 ctx = rcu_dereference(task->perf_event_ctxp[perf_sw_context]);
5549 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
5550 if (event->attr.type != PERF_TYPE_TRACEPOINT)
5552 if (event->attr.config != entry->type)
5554 if (perf_tp_event_match(event, &data, regs))
5555 perf_swevent_event(event, count, &data, regs);
5561 perf_swevent_put_recursion_context(rctx);
5563 EXPORT_SYMBOL_GPL(perf_tp_event);
5565 static void tp_perf_event_destroy(struct perf_event *event)
5567 perf_trace_destroy(event);
5570 static int perf_tp_event_init(struct perf_event *event)
5574 if (event->attr.type != PERF_TYPE_TRACEPOINT)
5578 * no branch sampling for tracepoint events
5580 if (has_branch_stack(event))
5583 err = perf_trace_init(event);
5587 event->destroy = tp_perf_event_destroy;
5592 static struct pmu perf_tracepoint = {
5593 .task_ctx_nr = perf_sw_context,
5595 .event_init = perf_tp_event_init,
5596 .add = perf_trace_add,
5597 .del = perf_trace_del,
5598 .start = perf_swevent_start,
5599 .stop = perf_swevent_stop,
5600 .read = perf_swevent_read,
5602 .event_idx = perf_swevent_event_idx,
5605 static inline void perf_tp_register(void)
5607 perf_pmu_register(&perf_tracepoint, "tracepoint", PERF_TYPE_TRACEPOINT);
5610 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
5615 if (event->attr.type != PERF_TYPE_TRACEPOINT)
5618 filter_str = strndup_user(arg, PAGE_SIZE);
5619 if (IS_ERR(filter_str))
5620 return PTR_ERR(filter_str);
5622 ret = ftrace_profile_set_filter(event, event->attr.config, filter_str);
5628 static void perf_event_free_filter(struct perf_event *event)
5630 ftrace_profile_free_filter(event);
5635 static inline void perf_tp_register(void)
5639 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
5644 static void perf_event_free_filter(struct perf_event *event)
5648 #endif /* CONFIG_EVENT_TRACING */
5650 #ifdef CONFIG_HAVE_HW_BREAKPOINT
5651 void perf_bp_event(struct perf_event *bp, void *data)
5653 struct perf_sample_data sample;
5654 struct pt_regs *regs = data;
5656 perf_sample_data_init(&sample, bp->attr.bp_addr, 0);
5658 if (!bp->hw.state && !perf_exclude_event(bp, regs))
5659 perf_swevent_event(bp, 1, &sample, regs);
5664 * hrtimer based swevent callback
5667 static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer)
5669 enum hrtimer_restart ret = HRTIMER_RESTART;
5670 struct perf_sample_data data;
5671 struct pt_regs *regs;
5672 struct perf_event *event;
5675 event = container_of(hrtimer, struct perf_event, hw.hrtimer);
5677 if (event->state != PERF_EVENT_STATE_ACTIVE)
5678 return HRTIMER_NORESTART;
5680 event->pmu->read(event);
5682 perf_sample_data_init(&data, 0, event->hw.last_period);
5683 regs = get_irq_regs();
5685 if (regs && !perf_exclude_event(event, regs)) {
5686 if (!(event->attr.exclude_idle && is_idle_task(current)))
5687 if (__perf_event_overflow(event, 1, &data, regs))
5688 ret = HRTIMER_NORESTART;
5691 period = max_t(u64, 10000, event->hw.sample_period);
5692 hrtimer_forward_now(hrtimer, ns_to_ktime(period));
5697 static void perf_swevent_start_hrtimer(struct perf_event *event)
5699 struct hw_perf_event *hwc = &event->hw;
5702 if (!is_sampling_event(event))
5705 period = local64_read(&hwc->period_left);
5710 local64_set(&hwc->period_left, 0);
5712 period = max_t(u64, 10000, hwc->sample_period);
5714 __hrtimer_start_range_ns(&hwc->hrtimer,
5715 ns_to_ktime(period), 0,
5716 HRTIMER_MODE_REL_PINNED, 0);
5719 static void perf_swevent_cancel_hrtimer(struct perf_event *event)
5721 struct hw_perf_event *hwc = &event->hw;
5723 if (is_sampling_event(event)) {
5724 ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer);
5725 local64_set(&hwc->period_left, ktime_to_ns(remaining));
5727 hrtimer_cancel(&hwc->hrtimer);
5731 static void perf_swevent_init_hrtimer(struct perf_event *event)
5733 struct hw_perf_event *hwc = &event->hw;
5735 if (!is_sampling_event(event))
5738 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
5739 hwc->hrtimer.function = perf_swevent_hrtimer;
5742 * Since hrtimers have a fixed rate, we can do a static freq->period
5743 * mapping and avoid the whole period adjust feedback stuff.
5745 if (event->attr.freq) {
5746 long freq = event->attr.sample_freq;
5748 event->attr.sample_period = NSEC_PER_SEC / freq;
5749 hwc->sample_period = event->attr.sample_period;
5750 local64_set(&hwc->period_left, hwc->sample_period);
5751 hwc->last_period = hwc->sample_period;
5752 event->attr.freq = 0;
5757 * Software event: cpu wall time clock
5760 static void cpu_clock_event_update(struct perf_event *event)
5765 now = local_clock();
5766 prev = local64_xchg(&event->hw.prev_count, now);
5767 local64_add(now - prev, &event->count);
5770 static void cpu_clock_event_start(struct perf_event *event, int flags)
5772 local64_set(&event->hw.prev_count, local_clock());
5773 perf_swevent_start_hrtimer(event);
5776 static void cpu_clock_event_stop(struct perf_event *event, int flags)
5778 perf_swevent_cancel_hrtimer(event);
5779 cpu_clock_event_update(event);
5782 static int cpu_clock_event_add(struct perf_event *event, int flags)
5784 if (flags & PERF_EF_START)
5785 cpu_clock_event_start(event, flags);
5790 static void cpu_clock_event_del(struct perf_event *event, int flags)
5792 cpu_clock_event_stop(event, flags);
5795 static void cpu_clock_event_read(struct perf_event *event)
5797 cpu_clock_event_update(event);
5800 static int cpu_clock_event_init(struct perf_event *event)
5802 if (event->attr.type != PERF_TYPE_SOFTWARE)
5805 if (event->attr.config != PERF_COUNT_SW_CPU_CLOCK)
5809 * no branch sampling for software events
5811 if (has_branch_stack(event))
5814 perf_swevent_init_hrtimer(event);
5819 static struct pmu perf_cpu_clock = {
5820 .task_ctx_nr = perf_sw_context,
5822 .event_init = cpu_clock_event_init,
5823 .add = cpu_clock_event_add,
5824 .del = cpu_clock_event_del,
5825 .start = cpu_clock_event_start,
5826 .stop = cpu_clock_event_stop,
5827 .read = cpu_clock_event_read,
5829 .event_idx = perf_swevent_event_idx,
5833 * Software event: task time clock
5836 static void task_clock_event_update(struct perf_event *event, u64 now)
5841 prev = local64_xchg(&event->hw.prev_count, now);
5843 local64_add(delta, &event->count);
5846 static void task_clock_event_start(struct perf_event *event, int flags)
5848 local64_set(&event->hw.prev_count, event->ctx->time);
5849 perf_swevent_start_hrtimer(event);
5852 static void task_clock_event_stop(struct perf_event *event, int flags)
5854 perf_swevent_cancel_hrtimer(event);
5855 task_clock_event_update(event, event->ctx->time);
5858 static int task_clock_event_add(struct perf_event *event, int flags)
5860 if (flags & PERF_EF_START)
5861 task_clock_event_start(event, flags);
5866 static void task_clock_event_del(struct perf_event *event, int flags)
5868 task_clock_event_stop(event, PERF_EF_UPDATE);
5871 static void task_clock_event_read(struct perf_event *event)
5873 u64 now = perf_clock();
5874 u64 delta = now - event->ctx->timestamp;
5875 u64 time = event->ctx->time + delta;
5877 task_clock_event_update(event, time);
5880 static int task_clock_event_init(struct perf_event *event)
5882 if (event->attr.type != PERF_TYPE_SOFTWARE)
5885 if (event->attr.config != PERF_COUNT_SW_TASK_CLOCK)
5889 * no branch sampling for software events
5891 if (has_branch_stack(event))
5894 perf_swevent_init_hrtimer(event);
5899 static struct pmu perf_task_clock = {
5900 .task_ctx_nr = perf_sw_context,
5902 .event_init = task_clock_event_init,
5903 .add = task_clock_event_add,
5904 .del = task_clock_event_del,
5905 .start = task_clock_event_start,
5906 .stop = task_clock_event_stop,
5907 .read = task_clock_event_read,
5909 .event_idx = perf_swevent_event_idx,
5912 static void perf_pmu_nop_void(struct pmu *pmu)
5916 static int perf_pmu_nop_int(struct pmu *pmu)
5921 static void perf_pmu_start_txn(struct pmu *pmu)
5923 perf_pmu_disable(pmu);
5926 static int perf_pmu_commit_txn(struct pmu *pmu)
5928 perf_pmu_enable(pmu);
5932 static void perf_pmu_cancel_txn(struct pmu *pmu)
5934 perf_pmu_enable(pmu);
5937 static int perf_event_idx_default(struct perf_event *event)
5939 return event->hw.idx + 1;
5943 * Ensures all contexts with the same task_ctx_nr have the same
5944 * pmu_cpu_context too.
5946 static void *find_pmu_context(int ctxn)
5953 list_for_each_entry(pmu, &pmus, entry) {
5954 if (pmu->task_ctx_nr == ctxn)
5955 return pmu->pmu_cpu_context;
5961 static void update_pmu_context(struct pmu *pmu, struct pmu *old_pmu)
5965 for_each_possible_cpu(cpu) {
5966 struct perf_cpu_context *cpuctx;
5968 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
5970 if (cpuctx->unique_pmu == old_pmu)
5971 cpuctx->unique_pmu = pmu;
5975 static void free_pmu_context(struct pmu *pmu)
5979 mutex_lock(&pmus_lock);
5981 * Like a real lame refcount.
5983 list_for_each_entry(i, &pmus, entry) {
5984 if (i->pmu_cpu_context == pmu->pmu_cpu_context) {
5985 update_pmu_context(i, pmu);
5990 free_percpu(pmu->pmu_cpu_context);
5992 mutex_unlock(&pmus_lock);
5994 static struct idr pmu_idr;
5997 type_show(struct device *dev, struct device_attribute *attr, char *page)
5999 struct pmu *pmu = dev_get_drvdata(dev);
6001 return snprintf(page, PAGE_SIZE-1, "%d\n", pmu->type);
6004 static struct device_attribute pmu_dev_attrs[] = {
6009 static int pmu_bus_running;
6010 static struct bus_type pmu_bus = {
6011 .name = "event_source",
6012 .dev_attrs = pmu_dev_attrs,
6015 static void pmu_dev_release(struct device *dev)
6020 static int pmu_dev_alloc(struct pmu *pmu)
6024 pmu->dev = kzalloc(sizeof(struct device), GFP_KERNEL);
6028 pmu->dev->groups = pmu->attr_groups;
6029 device_initialize(pmu->dev);
6030 ret = dev_set_name(pmu->dev, "%s", pmu->name);
6034 dev_set_drvdata(pmu->dev, pmu);
6035 pmu->dev->bus = &pmu_bus;
6036 pmu->dev->release = pmu_dev_release;
6037 ret = device_add(pmu->dev);
6045 put_device(pmu->dev);
6049 static struct lock_class_key cpuctx_mutex;
6050 static struct lock_class_key cpuctx_lock;
6052 int perf_pmu_register(struct pmu *pmu, char *name, int type)
6056 mutex_lock(&pmus_lock);
6058 pmu->pmu_disable_count = alloc_percpu(int);
6059 if (!pmu->pmu_disable_count)
6068 type = idr_alloc(&pmu_idr, pmu, PERF_TYPE_MAX, 0, GFP_KERNEL);
6076 if (pmu_bus_running) {
6077 ret = pmu_dev_alloc(pmu);
6083 pmu->pmu_cpu_context = find_pmu_context(pmu->task_ctx_nr);
6084 if (pmu->pmu_cpu_context)
6085 goto got_cpu_context;
6088 pmu->pmu_cpu_context = alloc_percpu(struct perf_cpu_context);
6089 if (!pmu->pmu_cpu_context)
6092 for_each_possible_cpu(cpu) {
6093 struct perf_cpu_context *cpuctx;
6095 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
6096 __perf_event_init_context(&cpuctx->ctx);
6097 lockdep_set_class(&cpuctx->ctx.mutex, &cpuctx_mutex);
6098 lockdep_set_class(&cpuctx->ctx.lock, &cpuctx_lock);
6099 cpuctx->ctx.type = cpu_context;
6100 cpuctx->ctx.pmu = pmu;
6101 cpuctx->jiffies_interval = 1;
6102 INIT_LIST_HEAD(&cpuctx->rotation_list);
6103 cpuctx->unique_pmu = pmu;
6107 if (!pmu->start_txn) {
6108 if (pmu->pmu_enable) {
6110 * If we have pmu_enable/pmu_disable calls, install
6111 * transaction stubs that use that to try and batch
6112 * hardware accesses.
6114 pmu->start_txn = perf_pmu_start_txn;
6115 pmu->commit_txn = perf_pmu_commit_txn;
6116 pmu->cancel_txn = perf_pmu_cancel_txn;
6118 pmu->start_txn = perf_pmu_nop_void;
6119 pmu->commit_txn = perf_pmu_nop_int;
6120 pmu->cancel_txn = perf_pmu_nop_void;
6124 if (!pmu->pmu_enable) {
6125 pmu->pmu_enable = perf_pmu_nop_void;
6126 pmu->pmu_disable = perf_pmu_nop_void;
6129 if (!pmu->event_idx)
6130 pmu->event_idx = perf_event_idx_default;
6132 list_add_rcu(&pmu->entry, &pmus);
6135 mutex_unlock(&pmus_lock);
6140 device_del(pmu->dev);
6141 put_device(pmu->dev);
6144 if (pmu->type >= PERF_TYPE_MAX)
6145 idr_remove(&pmu_idr, pmu->type);
6148 free_percpu(pmu->pmu_disable_count);
6152 void perf_pmu_unregister(struct pmu *pmu)
6154 mutex_lock(&pmus_lock);
6155 list_del_rcu(&pmu->entry);
6156 mutex_unlock(&pmus_lock);
6159 * We dereference the pmu list under both SRCU and regular RCU, so
6160 * synchronize against both of those.
6162 synchronize_srcu(&pmus_srcu);
6165 free_percpu(pmu->pmu_disable_count);
6166 if (pmu->type >= PERF_TYPE_MAX)
6167 idr_remove(&pmu_idr, pmu->type);
6168 device_del(pmu->dev);
6169 put_device(pmu->dev);
6170 free_pmu_context(pmu);
6173 struct pmu *perf_init_event(struct perf_event *event)
6175 struct pmu *pmu = NULL;
6179 idx = srcu_read_lock(&pmus_srcu);
6182 pmu = idr_find(&pmu_idr, event->attr.type);
6186 ret = pmu->event_init(event);
6192 list_for_each_entry_rcu(pmu, &pmus, entry) {
6194 ret = pmu->event_init(event);
6198 if (ret != -ENOENT) {
6203 pmu = ERR_PTR(-ENOENT);
6205 srcu_read_unlock(&pmus_srcu, idx);
6211 * Allocate and initialize a event structure
6213 static struct perf_event *
6214 perf_event_alloc(struct perf_event_attr *attr, int cpu,
6215 struct task_struct *task,
6216 struct perf_event *group_leader,
6217 struct perf_event *parent_event,
6218 perf_overflow_handler_t overflow_handler,
6222 struct perf_event *event;
6223 struct hw_perf_event *hwc;
6226 if ((unsigned)cpu >= nr_cpu_ids) {
6227 if (!task || cpu != -1)
6228 return ERR_PTR(-EINVAL);
6231 event = kzalloc(sizeof(*event), GFP_KERNEL);
6233 return ERR_PTR(-ENOMEM);
6236 * Single events are their own group leaders, with an
6237 * empty sibling list:
6240 group_leader = event;
6242 mutex_init(&event->child_mutex);
6243 INIT_LIST_HEAD(&event->child_list);
6245 INIT_LIST_HEAD(&event->group_entry);
6246 INIT_LIST_HEAD(&event->event_entry);
6247 INIT_LIST_HEAD(&event->sibling_list);
6248 INIT_LIST_HEAD(&event->rb_entry);
6250 init_waitqueue_head(&event->waitq);
6251 init_irq_work(&event->pending, perf_pending_event);
6253 mutex_init(&event->mmap_mutex);
6255 atomic_long_set(&event->refcount, 1);
6257 event->attr = *attr;
6258 event->group_leader = group_leader;
6262 event->parent = parent_event;
6264 event->ns = get_pid_ns(task_active_pid_ns(current));
6265 event->id = atomic64_inc_return(&perf_event_id);
6267 event->state = PERF_EVENT_STATE_INACTIVE;
6270 event->attach_state = PERF_ATTACH_TASK;
6272 if (attr->type == PERF_TYPE_TRACEPOINT)
6273 event->hw.tp_target = task;
6274 #ifdef CONFIG_HAVE_HW_BREAKPOINT
6276 * hw_breakpoint is a bit difficult here..
6278 else if (attr->type == PERF_TYPE_BREAKPOINT)
6279 event->hw.bp_target = task;
6283 if (!overflow_handler && parent_event) {
6284 overflow_handler = parent_event->overflow_handler;
6285 context = parent_event->overflow_handler_context;
6288 event->overflow_handler = overflow_handler;
6289 event->overflow_handler_context = context;
6291 perf_event__state_init(event);
6296 hwc->sample_period = attr->sample_period;
6297 if (attr->freq && attr->sample_freq)
6298 hwc->sample_period = 1;
6299 hwc->last_period = hwc->sample_period;
6301 local64_set(&hwc->period_left, hwc->sample_period);
6304 * we currently do not support PERF_FORMAT_GROUP on inherited events
6306 if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP))
6309 pmu = perf_init_event(event);
6315 else if (IS_ERR(pmu))
6320 put_pid_ns(event->ns);
6322 return ERR_PTR(err);
6325 if (!event->parent) {
6326 if (event->attach_state & PERF_ATTACH_TASK)
6327 static_key_slow_inc(&perf_sched_events.key);
6328 if (event->attr.mmap || event->attr.mmap_data)
6329 atomic_inc(&nr_mmap_events);
6330 if (event->attr.comm)
6331 atomic_inc(&nr_comm_events);
6332 if (event->attr.task)
6333 atomic_inc(&nr_task_events);
6334 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) {
6335 err = get_callchain_buffers();
6338 return ERR_PTR(err);
6341 if (has_branch_stack(event)) {
6342 static_key_slow_inc(&perf_sched_events.key);
6343 if (!(event->attach_state & PERF_ATTACH_TASK))
6344 atomic_inc(&per_cpu(perf_branch_stack_events,
6352 static int perf_copy_attr(struct perf_event_attr __user *uattr,
6353 struct perf_event_attr *attr)
6358 if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0))
6362 * zero the full structure, so that a short copy will be nice.
6364 memset(attr, 0, sizeof(*attr));
6366 ret = get_user(size, &uattr->size);
6370 if (size > PAGE_SIZE) /* silly large */
6373 if (!size) /* abi compat */
6374 size = PERF_ATTR_SIZE_VER0;
6376 if (size < PERF_ATTR_SIZE_VER0)
6380 * If we're handed a bigger struct than we know of,
6381 * ensure all the unknown bits are 0 - i.e. new
6382 * user-space does not rely on any kernel feature
6383 * extensions we dont know about yet.
6385 if (size > sizeof(*attr)) {
6386 unsigned char __user *addr;
6387 unsigned char __user *end;
6390 addr = (void __user *)uattr + sizeof(*attr);
6391 end = (void __user *)uattr + size;
6393 for (; addr < end; addr++) {
6394 ret = get_user(val, addr);
6400 size = sizeof(*attr);
6403 ret = copy_from_user(attr, uattr, size);
6407 if (attr->__reserved_1)
6410 if (attr->sample_type & ~(PERF_SAMPLE_MAX-1))
6413 if (attr->read_format & ~(PERF_FORMAT_MAX-1))
6416 if (attr->sample_type & PERF_SAMPLE_BRANCH_STACK) {
6417 u64 mask = attr->branch_sample_type;
6419 /* only using defined bits */
6420 if (mask & ~(PERF_SAMPLE_BRANCH_MAX-1))
6423 /* at least one branch bit must be set */
6424 if (!(mask & ~PERF_SAMPLE_BRANCH_PLM_ALL))
6427 /* kernel level capture: check permissions */
6428 if ((mask & PERF_SAMPLE_BRANCH_PERM_PLM)
6429 && perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
6432 /* propagate priv level, when not set for branch */
6433 if (!(mask & PERF_SAMPLE_BRANCH_PLM_ALL)) {
6435 /* exclude_kernel checked on syscall entry */
6436 if (!attr->exclude_kernel)
6437 mask |= PERF_SAMPLE_BRANCH_KERNEL;
6439 if (!attr->exclude_user)
6440 mask |= PERF_SAMPLE_BRANCH_USER;
6442 if (!attr->exclude_hv)
6443 mask |= PERF_SAMPLE_BRANCH_HV;
6445 * adjust user setting (for HW filter setup)
6447 attr->branch_sample_type = mask;
6451 if (attr->sample_type & PERF_SAMPLE_REGS_USER) {
6452 ret = perf_reg_validate(attr->sample_regs_user);
6457 if (attr->sample_type & PERF_SAMPLE_STACK_USER) {
6458 if (!arch_perf_have_user_stack_dump())
6462 * We have __u32 type for the size, but so far
6463 * we can only use __u16 as maximum due to the
6464 * __u16 sample size limit.
6466 if (attr->sample_stack_user >= USHRT_MAX)
6468 else if (!IS_ALIGNED(attr->sample_stack_user, sizeof(u64)))
6476 put_user(sizeof(*attr), &uattr->size);
6482 perf_event_set_output(struct perf_event *event, struct perf_event *output_event)
6484 struct ring_buffer *rb = NULL, *old_rb = NULL;
6490 /* don't allow circular references */
6491 if (event == output_event)
6495 * Don't allow cross-cpu buffers
6497 if (output_event->cpu != event->cpu)
6501 * If its not a per-cpu rb, it must be the same task.
6503 if (output_event->cpu == -1 && output_event->ctx != event->ctx)
6507 mutex_lock(&event->mmap_mutex);
6508 /* Can't redirect output if we've got an active mmap() */
6509 if (atomic_read(&event->mmap_count))
6515 /* get the rb we want to redirect to */
6516 rb = ring_buffer_get(output_event);
6522 ring_buffer_detach(event, old_rb);
6525 ring_buffer_attach(event, rb);
6527 rcu_assign_pointer(event->rb, rb);
6530 ring_buffer_put(old_rb);
6532 * Since we detached before setting the new rb, so that we
6533 * could attach the new rb, we could have missed a wakeup.
6536 wake_up_all(&event->waitq);
6541 mutex_unlock(&event->mmap_mutex);
6548 * sys_perf_event_open - open a performance event, associate it to a task/cpu
6550 * @attr_uptr: event_id type attributes for monitoring/sampling
6553 * @group_fd: group leader event fd
6555 SYSCALL_DEFINE5(perf_event_open,
6556 struct perf_event_attr __user *, attr_uptr,
6557 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
6559 struct perf_event *group_leader = NULL, *output_event = NULL;
6560 struct perf_event *event, *sibling;
6561 struct perf_event_attr attr;
6562 struct perf_event_context *ctx;
6563 struct file *event_file = NULL;
6564 struct fd group = {NULL, 0};
6565 struct task_struct *task = NULL;
6571 /* for future expandability... */
6572 if (flags & ~PERF_FLAG_ALL)
6575 err = perf_copy_attr(attr_uptr, &attr);
6579 if (!attr.exclude_kernel) {
6580 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
6585 if (attr.sample_freq > sysctl_perf_event_sample_rate)
6590 * In cgroup mode, the pid argument is used to pass the fd
6591 * opened to the cgroup directory in cgroupfs. The cpu argument
6592 * designates the cpu on which to monitor threads from that
6595 if ((flags & PERF_FLAG_PID_CGROUP) && (pid == -1 || cpu == -1))
6598 event_fd = get_unused_fd();
6602 if (group_fd != -1) {
6603 err = perf_fget_light(group_fd, &group);
6606 group_leader = group.file->private_data;
6607 if (flags & PERF_FLAG_FD_OUTPUT)
6608 output_event = group_leader;
6609 if (flags & PERF_FLAG_FD_NO_GROUP)
6610 group_leader = NULL;
6613 if (pid != -1 && !(flags & PERF_FLAG_PID_CGROUP)) {
6614 task = find_lively_task_by_vpid(pid);
6616 err = PTR_ERR(task);
6623 event = perf_event_alloc(&attr, cpu, task, group_leader, NULL,
6625 if (IS_ERR(event)) {
6626 err = PTR_ERR(event);
6630 if (flags & PERF_FLAG_PID_CGROUP) {
6631 err = perf_cgroup_connect(pid, event, &attr, group_leader);
6636 * - that has cgroup constraint on event->cpu
6637 * - that may need work on context switch
6639 atomic_inc(&per_cpu(perf_cgroup_events, event->cpu));
6640 static_key_slow_inc(&perf_sched_events.key);
6644 * Special case software events and allow them to be part of
6645 * any hardware group.
6650 (is_software_event(event) != is_software_event(group_leader))) {
6651 if (is_software_event(event)) {
6653 * If event and group_leader are not both a software
6654 * event, and event is, then group leader is not.
6656 * Allow the addition of software events to !software
6657 * groups, this is safe because software events never
6660 pmu = group_leader->pmu;
6661 } else if (is_software_event(group_leader) &&
6662 (group_leader->group_flags & PERF_GROUP_SOFTWARE)) {
6664 * In case the group is a pure software group, and we
6665 * try to add a hardware event, move the whole group to
6666 * the hardware context.
6673 * Get the target context (task or percpu):
6675 ctx = find_get_context(pmu, task, event->cpu);
6682 put_task_struct(task);
6687 * Look up the group leader (we will attach this event to it):
6693 * Do not allow a recursive hierarchy (this new sibling
6694 * becoming part of another group-sibling):
6696 if (group_leader->group_leader != group_leader)
6699 * Do not allow to attach to a group in a different
6700 * task or CPU context:
6703 if (group_leader->ctx->type != ctx->type)
6706 if (group_leader->ctx != ctx)
6711 * Only a group leader can be exclusive or pinned
6713 if (attr.exclusive || attr.pinned)
6718 err = perf_event_set_output(event, output_event);
6723 event_file = anon_inode_getfile("[perf_event]", &perf_fops, event, O_RDWR);
6724 if (IS_ERR(event_file)) {
6725 err = PTR_ERR(event_file);
6730 struct perf_event_context *gctx = group_leader->ctx;
6732 mutex_lock(&gctx->mutex);
6733 perf_remove_from_context(group_leader);
6736 * Removing from the context ends up with disabled
6737 * event. What we want here is event in the initial
6738 * startup state, ready to be add into new context.
6740 perf_event__state_init(group_leader);
6741 list_for_each_entry(sibling, &group_leader->sibling_list,
6743 perf_remove_from_context(sibling);
6744 perf_event__state_init(sibling);
6747 mutex_unlock(&gctx->mutex);
6751 WARN_ON_ONCE(ctx->parent_ctx);
6752 mutex_lock(&ctx->mutex);
6756 perf_install_in_context(ctx, group_leader, event->cpu);
6758 list_for_each_entry(sibling, &group_leader->sibling_list,
6760 perf_install_in_context(ctx, sibling, event->cpu);
6765 perf_install_in_context(ctx, event, event->cpu);
6767 perf_unpin_context(ctx);
6768 mutex_unlock(&ctx->mutex);
6772 event->owner = current;
6774 mutex_lock(¤t->perf_event_mutex);
6775 list_add_tail(&event->owner_entry, ¤t->perf_event_list);
6776 mutex_unlock(¤t->perf_event_mutex);
6779 * Precalculate sample_data sizes
6781 perf_event__header_size(event);
6782 perf_event__id_header_size(event);
6785 * Drop the reference on the group_event after placing the
6786 * new event on the sibling_list. This ensures destruction
6787 * of the group leader will find the pointer to itself in
6788 * perf_group_detach().
6791 fd_install(event_fd, event_file);
6795 perf_unpin_context(ctx);
6802 put_task_struct(task);
6806 put_unused_fd(event_fd);
6811 * perf_event_create_kernel_counter
6813 * @attr: attributes of the counter to create
6814 * @cpu: cpu in which the counter is bound
6815 * @task: task to profile (NULL for percpu)
6818 perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu,
6819 struct task_struct *task,
6820 perf_overflow_handler_t overflow_handler,
6823 struct perf_event_context *ctx;
6824 struct perf_event *event;
6828 * Get the target context (task or percpu):
6831 event = perf_event_alloc(attr, cpu, task, NULL, NULL,
6832 overflow_handler, context);
6833 if (IS_ERR(event)) {
6834 err = PTR_ERR(event);
6838 ctx = find_get_context(event->pmu, task, cpu);
6844 WARN_ON_ONCE(ctx->parent_ctx);
6845 mutex_lock(&ctx->mutex);
6846 perf_install_in_context(ctx, event, cpu);
6848 perf_unpin_context(ctx);
6849 mutex_unlock(&ctx->mutex);
6856 return ERR_PTR(err);
6858 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter);
6860 void perf_pmu_migrate_context(struct pmu *pmu, int src_cpu, int dst_cpu)
6862 struct perf_event_context *src_ctx;
6863 struct perf_event_context *dst_ctx;
6864 struct perf_event *event, *tmp;
6867 src_ctx = &per_cpu_ptr(pmu->pmu_cpu_context, src_cpu)->ctx;
6868 dst_ctx = &per_cpu_ptr(pmu->pmu_cpu_context, dst_cpu)->ctx;
6870 mutex_lock(&src_ctx->mutex);
6871 list_for_each_entry_safe(event, tmp, &src_ctx->event_list,
6873 perf_remove_from_context(event);
6875 list_add(&event->event_entry, &events);
6877 mutex_unlock(&src_ctx->mutex);
6881 mutex_lock(&dst_ctx->mutex);
6882 list_for_each_entry_safe(event, tmp, &events, event_entry) {
6883 list_del(&event->event_entry);
6884 if (event->state >= PERF_EVENT_STATE_OFF)
6885 event->state = PERF_EVENT_STATE_INACTIVE;
6886 perf_install_in_context(dst_ctx, event, dst_cpu);
6889 mutex_unlock(&dst_ctx->mutex);
6891 EXPORT_SYMBOL_GPL(perf_pmu_migrate_context);
6893 static void sync_child_event(struct perf_event *child_event,
6894 struct task_struct *child)
6896 struct perf_event *parent_event = child_event->parent;
6899 if (child_event->attr.inherit_stat)
6900 perf_event_read_event(child_event, child);
6902 child_val = perf_event_count(child_event);
6905 * Add back the child's count to the parent's count:
6907 atomic64_add(child_val, &parent_event->child_count);
6908 atomic64_add(child_event->total_time_enabled,
6909 &parent_event->child_total_time_enabled);
6910 atomic64_add(child_event->total_time_running,
6911 &parent_event->child_total_time_running);
6914 * Remove this event from the parent's list
6916 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
6917 mutex_lock(&parent_event->child_mutex);
6918 list_del_init(&child_event->child_list);
6919 mutex_unlock(&parent_event->child_mutex);
6922 * Release the parent event, if this was the last
6925 put_event(parent_event);
6929 __perf_event_exit_task(struct perf_event *child_event,
6930 struct perf_event_context *child_ctx,
6931 struct task_struct *child)
6933 if (child_event->parent) {
6934 raw_spin_lock_irq(&child_ctx->lock);
6935 perf_group_detach(child_event);
6936 raw_spin_unlock_irq(&child_ctx->lock);
6939 perf_remove_from_context(child_event);
6942 * It can happen that the parent exits first, and has events
6943 * that are still around due to the child reference. These
6944 * events need to be zapped.
6946 if (child_event->parent) {
6947 sync_child_event(child_event, child);
6948 free_event(child_event);
6952 static void perf_event_exit_task_context(struct task_struct *child, int ctxn)
6954 struct perf_event *child_event, *tmp;
6955 struct perf_event_context *child_ctx;
6956 unsigned long flags;
6958 if (likely(!child->perf_event_ctxp[ctxn])) {
6959 perf_event_task(child, NULL, 0);
6963 local_irq_save(flags);
6965 * We can't reschedule here because interrupts are disabled,
6966 * and either child is current or it is a task that can't be
6967 * scheduled, so we are now safe from rescheduling changing
6970 child_ctx = rcu_dereference_raw(child->perf_event_ctxp[ctxn]);
6973 * Take the context lock here so that if find_get_context is
6974 * reading child->perf_event_ctxp, we wait until it has
6975 * incremented the context's refcount before we do put_ctx below.
6977 raw_spin_lock(&child_ctx->lock);
6978 task_ctx_sched_out(child_ctx);
6979 child->perf_event_ctxp[ctxn] = NULL;
6981 * If this context is a clone; unclone it so it can't get
6982 * swapped to another process while we're removing all
6983 * the events from it.
6985 unclone_ctx(child_ctx);
6986 update_context_time(child_ctx);
6987 raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
6990 * Report the task dead after unscheduling the events so that we
6991 * won't get any samples after PERF_RECORD_EXIT. We can however still
6992 * get a few PERF_RECORD_READ events.
6994 perf_event_task(child, child_ctx, 0);
6997 * We can recurse on the same lock type through:
6999 * __perf_event_exit_task()
7000 * sync_child_event()
7002 * mutex_lock(&ctx->mutex)
7004 * But since its the parent context it won't be the same instance.
7006 mutex_lock(&child_ctx->mutex);
7009 list_for_each_entry_safe(child_event, tmp, &child_ctx->pinned_groups,
7011 __perf_event_exit_task(child_event, child_ctx, child);
7013 list_for_each_entry_safe(child_event, tmp, &child_ctx->flexible_groups,
7015 __perf_event_exit_task(child_event, child_ctx, child);
7018 * If the last event was a group event, it will have appended all
7019 * its siblings to the list, but we obtained 'tmp' before that which
7020 * will still point to the list head terminating the iteration.
7022 if (!list_empty(&child_ctx->pinned_groups) ||
7023 !list_empty(&child_ctx->flexible_groups))
7026 mutex_unlock(&child_ctx->mutex);
7032 * When a child task exits, feed back event values to parent events.
7034 void perf_event_exit_task(struct task_struct *child)
7036 struct perf_event *event, *tmp;
7039 mutex_lock(&child->perf_event_mutex);
7040 list_for_each_entry_safe(event, tmp, &child->perf_event_list,
7042 list_del_init(&event->owner_entry);
7045 * Ensure the list deletion is visible before we clear
7046 * the owner, closes a race against perf_release() where
7047 * we need to serialize on the owner->perf_event_mutex.
7050 event->owner = NULL;
7052 mutex_unlock(&child->perf_event_mutex);
7054 for_each_task_context_nr(ctxn)
7055 perf_event_exit_task_context(child, ctxn);
7058 static void perf_free_event(struct perf_event *event,
7059 struct perf_event_context *ctx)
7061 struct perf_event *parent = event->parent;
7063 if (WARN_ON_ONCE(!parent))
7066 mutex_lock(&parent->child_mutex);
7067 list_del_init(&event->child_list);
7068 mutex_unlock(&parent->child_mutex);
7072 perf_group_detach(event);
7073 list_del_event(event, ctx);
7078 * free an unexposed, unused context as created by inheritance by
7079 * perf_event_init_task below, used by fork() in case of fail.
7081 void perf_event_free_task(struct task_struct *task)
7083 struct perf_event_context *ctx;
7084 struct perf_event *event, *tmp;
7087 for_each_task_context_nr(ctxn) {
7088 ctx = task->perf_event_ctxp[ctxn];
7092 mutex_lock(&ctx->mutex);
7094 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups,
7096 perf_free_event(event, ctx);
7098 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups,
7100 perf_free_event(event, ctx);
7102 if (!list_empty(&ctx->pinned_groups) ||
7103 !list_empty(&ctx->flexible_groups))
7106 mutex_unlock(&ctx->mutex);
7112 void perf_event_delayed_put(struct task_struct *task)
7116 for_each_task_context_nr(ctxn)
7117 WARN_ON_ONCE(task->perf_event_ctxp[ctxn]);
7121 * inherit a event from parent task to child task:
7123 static struct perf_event *
7124 inherit_event(struct perf_event *parent_event,
7125 struct task_struct *parent,
7126 struct perf_event_context *parent_ctx,
7127 struct task_struct *child,
7128 struct perf_event *group_leader,
7129 struct perf_event_context *child_ctx)
7131 struct perf_event *child_event;
7132 unsigned long flags;
7135 * Instead of creating recursive hierarchies of events,
7136 * we link inherited events back to the original parent,
7137 * which has a filp for sure, which we use as the reference
7140 if (parent_event->parent)
7141 parent_event = parent_event->parent;
7143 child_event = perf_event_alloc(&parent_event->attr,
7146 group_leader, parent_event,
7148 if (IS_ERR(child_event))
7151 if (!atomic_long_inc_not_zero(&parent_event->refcount)) {
7152 free_event(child_event);
7159 * Make the child state follow the state of the parent event,
7160 * not its attr.disabled bit. We hold the parent's mutex,
7161 * so we won't race with perf_event_{en, dis}able_family.
7163 if (parent_event->state >= PERF_EVENT_STATE_INACTIVE)
7164 child_event->state = PERF_EVENT_STATE_INACTIVE;
7166 child_event->state = PERF_EVENT_STATE_OFF;
7168 if (parent_event->attr.freq) {
7169 u64 sample_period = parent_event->hw.sample_period;
7170 struct hw_perf_event *hwc = &child_event->hw;
7172 hwc->sample_period = sample_period;
7173 hwc->last_period = sample_period;
7175 local64_set(&hwc->period_left, sample_period);
7178 child_event->ctx = child_ctx;
7179 child_event->overflow_handler = parent_event->overflow_handler;
7180 child_event->overflow_handler_context
7181 = parent_event->overflow_handler_context;
7184 * Precalculate sample_data sizes
7186 perf_event__header_size(child_event);
7187 perf_event__id_header_size(child_event);
7190 * Link it up in the child's context:
7192 raw_spin_lock_irqsave(&child_ctx->lock, flags);
7193 add_event_to_ctx(child_event, child_ctx);
7194 raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
7197 * Link this into the parent event's child list
7199 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
7200 mutex_lock(&parent_event->child_mutex);
7201 list_add_tail(&child_event->child_list, &parent_event->child_list);
7202 mutex_unlock(&parent_event->child_mutex);
7207 static int inherit_group(struct perf_event *parent_event,
7208 struct task_struct *parent,
7209 struct perf_event_context *parent_ctx,
7210 struct task_struct *child,
7211 struct perf_event_context *child_ctx)
7213 struct perf_event *leader;
7214 struct perf_event *sub;
7215 struct perf_event *child_ctr;
7217 leader = inherit_event(parent_event, parent, parent_ctx,
7218 child, NULL, child_ctx);
7220 return PTR_ERR(leader);
7221 list_for_each_entry(sub, &parent_event->sibling_list, group_entry) {
7222 child_ctr = inherit_event(sub, parent, parent_ctx,
7223 child, leader, child_ctx);
7224 if (IS_ERR(child_ctr))
7225 return PTR_ERR(child_ctr);
7231 inherit_task_group(struct perf_event *event, struct task_struct *parent,
7232 struct perf_event_context *parent_ctx,
7233 struct task_struct *child, int ctxn,
7237 struct perf_event_context *child_ctx;
7239 if (!event->attr.inherit) {
7244 child_ctx = child->perf_event_ctxp[ctxn];
7247 * This is executed from the parent task context, so
7248 * inherit events that have been marked for cloning.
7249 * First allocate and initialize a context for the
7253 child_ctx = alloc_perf_context(parent_ctx->pmu, child);
7257 child->perf_event_ctxp[ctxn] = child_ctx;
7260 ret = inherit_group(event, parent, parent_ctx,
7270 * Initialize the perf_event context in task_struct
7272 int perf_event_init_context(struct task_struct *child, int ctxn)
7274 struct perf_event_context *child_ctx, *parent_ctx;
7275 struct perf_event_context *cloned_ctx;
7276 struct perf_event *event;
7277 struct task_struct *parent = current;
7278 int inherited_all = 1;
7279 unsigned long flags;
7282 if (likely(!parent->perf_event_ctxp[ctxn]))
7286 * If the parent's context is a clone, pin it so it won't get
7289 parent_ctx = perf_pin_task_context(parent, ctxn);
7292 * No need to check if parent_ctx != NULL here; since we saw
7293 * it non-NULL earlier, the only reason for it to become NULL
7294 * is if we exit, and since we're currently in the middle of
7295 * a fork we can't be exiting at the same time.
7299 * Lock the parent list. No need to lock the child - not PID
7300 * hashed yet and not running, so nobody can access it.
7302 mutex_lock(&parent_ctx->mutex);
7305 * We dont have to disable NMIs - we are only looking at
7306 * the list, not manipulating it:
7308 list_for_each_entry(event, &parent_ctx->pinned_groups, group_entry) {
7309 ret = inherit_task_group(event, parent, parent_ctx,
7310 child, ctxn, &inherited_all);
7316 * We can't hold ctx->lock when iterating the ->flexible_group list due
7317 * to allocations, but we need to prevent rotation because
7318 * rotate_ctx() will change the list from interrupt context.
7320 raw_spin_lock_irqsave(&parent_ctx->lock, flags);
7321 parent_ctx->rotate_disable = 1;
7322 raw_spin_unlock_irqrestore(&parent_ctx->lock, flags);
7324 list_for_each_entry(event, &parent_ctx->flexible_groups, group_entry) {
7325 ret = inherit_task_group(event, parent, parent_ctx,
7326 child, ctxn, &inherited_all);
7331 raw_spin_lock_irqsave(&parent_ctx->lock, flags);
7332 parent_ctx->rotate_disable = 0;
7334 child_ctx = child->perf_event_ctxp[ctxn];
7336 if (child_ctx && inherited_all) {
7338 * Mark the child context as a clone of the parent
7339 * context, or of whatever the parent is a clone of.
7341 * Note that if the parent is a clone, the holding of
7342 * parent_ctx->lock avoids it from being uncloned.
7344 cloned_ctx = parent_ctx->parent_ctx;
7346 child_ctx->parent_ctx = cloned_ctx;
7347 child_ctx->parent_gen = parent_ctx->parent_gen;
7349 child_ctx->parent_ctx = parent_ctx;
7350 child_ctx->parent_gen = parent_ctx->generation;
7352 get_ctx(child_ctx->parent_ctx);
7355 raw_spin_unlock_irqrestore(&parent_ctx->lock, flags);
7356 mutex_unlock(&parent_ctx->mutex);
7358 perf_unpin_context(parent_ctx);
7359 put_ctx(parent_ctx);
7365 * Initialize the perf_event context in task_struct
7367 int perf_event_init_task(struct task_struct *child)
7371 memset(child->perf_event_ctxp, 0, sizeof(child->perf_event_ctxp));
7372 mutex_init(&child->perf_event_mutex);
7373 INIT_LIST_HEAD(&child->perf_event_list);
7375 for_each_task_context_nr(ctxn) {
7376 ret = perf_event_init_context(child, ctxn);
7384 static void __init perf_event_init_all_cpus(void)
7386 struct swevent_htable *swhash;
7389 for_each_possible_cpu(cpu) {
7390 swhash = &per_cpu(swevent_htable, cpu);
7391 mutex_init(&swhash->hlist_mutex);
7392 INIT_LIST_HEAD(&per_cpu(rotation_list, cpu));
7396 static void __cpuinit perf_event_init_cpu(int cpu)
7398 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
7400 mutex_lock(&swhash->hlist_mutex);
7401 if (swhash->hlist_refcount > 0) {
7402 struct swevent_hlist *hlist;
7404 hlist = kzalloc_node(sizeof(*hlist), GFP_KERNEL, cpu_to_node(cpu));
7406 rcu_assign_pointer(swhash->swevent_hlist, hlist);
7408 mutex_unlock(&swhash->hlist_mutex);
7411 #if defined CONFIG_HOTPLUG_CPU || defined CONFIG_KEXEC
7412 static void perf_pmu_rotate_stop(struct pmu *pmu)
7414 struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
7416 WARN_ON(!irqs_disabled());
7418 list_del_init(&cpuctx->rotation_list);
7421 static void __perf_event_exit_context(void *__info)
7423 struct perf_event_context *ctx = __info;
7424 struct perf_event *event, *tmp;
7426 perf_pmu_rotate_stop(ctx->pmu);
7428 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry)
7429 __perf_remove_from_context(event);
7430 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups, group_entry)
7431 __perf_remove_from_context(event);
7434 static void perf_event_exit_cpu_context(int cpu)
7436 struct perf_event_context *ctx;
7440 idx = srcu_read_lock(&pmus_srcu);
7441 list_for_each_entry_rcu(pmu, &pmus, entry) {
7442 ctx = &per_cpu_ptr(pmu->pmu_cpu_context, cpu)->ctx;
7444 mutex_lock(&ctx->mutex);
7445 smp_call_function_single(cpu, __perf_event_exit_context, ctx, 1);
7446 mutex_unlock(&ctx->mutex);
7448 srcu_read_unlock(&pmus_srcu, idx);
7451 static void perf_event_exit_cpu(int cpu)
7453 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
7455 mutex_lock(&swhash->hlist_mutex);
7456 swevent_hlist_release(swhash);
7457 mutex_unlock(&swhash->hlist_mutex);
7459 perf_event_exit_cpu_context(cpu);
7462 static inline void perf_event_exit_cpu(int cpu) { }
7466 perf_reboot(struct notifier_block *notifier, unsigned long val, void *v)
7470 for_each_online_cpu(cpu)
7471 perf_event_exit_cpu(cpu);
7477 * Run the perf reboot notifier at the very last possible moment so that
7478 * the generic watchdog code runs as long as possible.
7480 static struct notifier_block perf_reboot_notifier = {
7481 .notifier_call = perf_reboot,
7482 .priority = INT_MIN,
7485 static int __cpuinit
7486 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
7488 unsigned int cpu = (long)hcpu;
7490 switch (action & ~CPU_TASKS_FROZEN) {
7492 case CPU_UP_PREPARE:
7493 case CPU_DOWN_FAILED:
7494 perf_event_init_cpu(cpu);
7497 case CPU_UP_CANCELED:
7498 case CPU_DOWN_PREPARE:
7499 perf_event_exit_cpu(cpu);
7509 void __init perf_event_init(void)
7515 perf_event_init_all_cpus();
7516 init_srcu_struct(&pmus_srcu);
7517 perf_pmu_register(&perf_swevent, "software", PERF_TYPE_SOFTWARE);
7518 perf_pmu_register(&perf_cpu_clock, NULL, -1);
7519 perf_pmu_register(&perf_task_clock, NULL, -1);
7521 perf_cpu_notifier(perf_cpu_notify);
7522 register_reboot_notifier(&perf_reboot_notifier);
7524 ret = init_hw_breakpoint();
7525 WARN(ret, "hw_breakpoint initialization failed with: %d", ret);
7527 /* do not patch jump label more than once per second */
7528 jump_label_rate_limit(&perf_sched_events, HZ);
7531 * Build time assertion that we keep the data_head at the intended
7532 * location. IOW, validation we got the __reserved[] size right.
7534 BUILD_BUG_ON((offsetof(struct perf_event_mmap_page, data_head))
7538 static int __init perf_event_sysfs_init(void)
7543 mutex_lock(&pmus_lock);
7545 ret = bus_register(&pmu_bus);
7549 list_for_each_entry(pmu, &pmus, entry) {
7550 if (!pmu->name || pmu->type < 0)
7553 ret = pmu_dev_alloc(pmu);
7554 WARN(ret, "Failed to register pmu: %s, reason %d\n", pmu->name, ret);
7556 pmu_bus_running = 1;
7560 mutex_unlock(&pmus_lock);
7564 device_initcall(perf_event_sysfs_init);
7566 #ifdef CONFIG_CGROUP_PERF
7567 static struct cgroup_subsys_state *perf_cgroup_css_alloc(struct cgroup *cont)
7569 struct perf_cgroup *jc;
7571 jc = kzalloc(sizeof(*jc), GFP_KERNEL);
7573 return ERR_PTR(-ENOMEM);
7575 jc->info = alloc_percpu(struct perf_cgroup_info);
7578 return ERR_PTR(-ENOMEM);
7584 static void perf_cgroup_css_free(struct cgroup *cont)
7586 struct perf_cgroup *jc;
7587 jc = container_of(cgroup_subsys_state(cont, perf_subsys_id),
7588 struct perf_cgroup, css);
7589 free_percpu(jc->info);
7593 static int __perf_cgroup_move(void *info)
7595 struct task_struct *task = info;
7596 perf_cgroup_switch(task, PERF_CGROUP_SWOUT | PERF_CGROUP_SWIN);
7600 static void perf_cgroup_attach(struct cgroup *cgrp, struct cgroup_taskset *tset)
7602 struct task_struct *task;
7604 cgroup_taskset_for_each(task, cgrp, tset)
7605 task_function_call(task, __perf_cgroup_move, task);
7608 static void perf_cgroup_exit(struct cgroup *cgrp, struct cgroup *old_cgrp,
7609 struct task_struct *task)
7612 * cgroup_exit() is called in the copy_process() failure path.
7613 * Ignore this case since the task hasn't ran yet, this avoids
7614 * trying to poke a half freed task state from generic code.
7616 if (!(task->flags & PF_EXITING))
7619 task_function_call(task, __perf_cgroup_move, task);
7622 struct cgroup_subsys perf_subsys = {
7623 .name = "perf_event",
7624 .subsys_id = perf_subsys_id,
7625 .css_alloc = perf_cgroup_css_alloc,
7626 .css_free = perf_cgroup_css_free,
7627 .exit = perf_cgroup_exit,
7628 .attach = perf_cgroup_attach,
7630 #endif /* CONFIG_CGROUP_PERF */