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 ctx = rcu_dereference(task->perf_event_ctxp[ctxn]);
769 * If this context is a clone of another, it might
770 * get swapped for another underneath us by
771 * perf_event_task_sched_out, though the
772 * rcu_read_lock() protects us from any context
773 * getting freed. Lock the context and check if it
774 * got swapped before we could get the lock, and retry
775 * if so. If we locked the right context, then it
776 * can't get swapped on us any more.
778 raw_spin_lock_irqsave(&ctx->lock, *flags);
779 if (ctx != rcu_dereference(task->perf_event_ctxp[ctxn])) {
780 raw_spin_unlock_irqrestore(&ctx->lock, *flags);
784 if (!atomic_inc_not_zero(&ctx->refcount)) {
785 raw_spin_unlock_irqrestore(&ctx->lock, *flags);
794 * Get the context for a task and increment its pin_count so it
795 * can't get swapped to another task. This also increments its
796 * reference count so that the context can't get freed.
798 static struct perf_event_context *
799 perf_pin_task_context(struct task_struct *task, int ctxn)
801 struct perf_event_context *ctx;
804 ctx = perf_lock_task_context(task, ctxn, &flags);
807 raw_spin_unlock_irqrestore(&ctx->lock, flags);
812 static void perf_unpin_context(struct perf_event_context *ctx)
816 raw_spin_lock_irqsave(&ctx->lock, flags);
818 raw_spin_unlock_irqrestore(&ctx->lock, flags);
822 * Update the record of the current time in a context.
824 static void update_context_time(struct perf_event_context *ctx)
826 u64 now = perf_clock();
828 ctx->time += now - ctx->timestamp;
829 ctx->timestamp = now;
832 static u64 perf_event_time(struct perf_event *event)
834 struct perf_event_context *ctx = event->ctx;
836 if (is_cgroup_event(event))
837 return perf_cgroup_event_time(event);
839 return ctx ? ctx->time : 0;
843 * Update the total_time_enabled and total_time_running fields for a event.
844 * The caller of this function needs to hold the ctx->lock.
846 static void update_event_times(struct perf_event *event)
848 struct perf_event_context *ctx = event->ctx;
851 if (event->state < PERF_EVENT_STATE_INACTIVE ||
852 event->group_leader->state < PERF_EVENT_STATE_INACTIVE)
855 * in cgroup mode, time_enabled represents
856 * the time the event was enabled AND active
857 * tasks were in the monitored cgroup. This is
858 * independent of the activity of the context as
859 * there may be a mix of cgroup and non-cgroup events.
861 * That is why we treat cgroup events differently
864 if (is_cgroup_event(event))
865 run_end = perf_cgroup_event_time(event);
866 else if (ctx->is_active)
869 run_end = event->tstamp_stopped;
871 event->total_time_enabled = run_end - event->tstamp_enabled;
873 if (event->state == PERF_EVENT_STATE_INACTIVE)
874 run_end = event->tstamp_stopped;
876 run_end = perf_event_time(event);
878 event->total_time_running = run_end - event->tstamp_running;
883 * Update total_time_enabled and total_time_running for all events in a group.
885 static void update_group_times(struct perf_event *leader)
887 struct perf_event *event;
889 update_event_times(leader);
890 list_for_each_entry(event, &leader->sibling_list, group_entry)
891 update_event_times(event);
894 static struct list_head *
895 ctx_group_list(struct perf_event *event, struct perf_event_context *ctx)
897 if (event->attr.pinned)
898 return &ctx->pinned_groups;
900 return &ctx->flexible_groups;
904 * Add a event from the lists for its context.
905 * Must be called with ctx->mutex and ctx->lock held.
908 list_add_event(struct perf_event *event, struct perf_event_context *ctx)
910 WARN_ON_ONCE(event->attach_state & PERF_ATTACH_CONTEXT);
911 event->attach_state |= PERF_ATTACH_CONTEXT;
914 * If we're a stand alone event or group leader, we go to the context
915 * list, group events are kept attached to the group so that
916 * perf_group_detach can, at all times, locate all siblings.
918 if (event->group_leader == event) {
919 struct list_head *list;
921 if (is_software_event(event))
922 event->group_flags |= PERF_GROUP_SOFTWARE;
924 list = ctx_group_list(event, ctx);
925 list_add_tail(&event->group_entry, list);
928 if (is_cgroup_event(event))
931 if (has_branch_stack(event))
932 ctx->nr_branch_stack++;
934 list_add_rcu(&event->event_entry, &ctx->event_list);
936 perf_pmu_rotate_start(ctx->pmu);
938 if (event->attr.inherit_stat)
943 * Initialize event state based on the perf_event_attr::disabled.
945 static inline void perf_event__state_init(struct perf_event *event)
947 event->state = event->attr.disabled ? PERF_EVENT_STATE_OFF :
948 PERF_EVENT_STATE_INACTIVE;
952 * Called at perf_event creation and when events are attached/detached from a
955 static void perf_event__read_size(struct perf_event *event)
957 int entry = sizeof(u64); /* value */
961 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
964 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
967 if (event->attr.read_format & PERF_FORMAT_ID)
968 entry += sizeof(u64);
970 if (event->attr.read_format & PERF_FORMAT_GROUP) {
971 nr += event->group_leader->nr_siblings;
976 event->read_size = size;
979 static void perf_event__header_size(struct perf_event *event)
981 struct perf_sample_data *data;
982 u64 sample_type = event->attr.sample_type;
985 perf_event__read_size(event);
987 if (sample_type & PERF_SAMPLE_IP)
988 size += sizeof(data->ip);
990 if (sample_type & PERF_SAMPLE_ADDR)
991 size += sizeof(data->addr);
993 if (sample_type & PERF_SAMPLE_PERIOD)
994 size += sizeof(data->period);
996 if (sample_type & PERF_SAMPLE_WEIGHT)
997 size += sizeof(data->weight);
999 if (sample_type & PERF_SAMPLE_READ)
1000 size += event->read_size;
1002 if (sample_type & PERF_SAMPLE_DATA_SRC)
1003 size += sizeof(data->data_src.val);
1005 event->header_size = size;
1008 static void perf_event__id_header_size(struct perf_event *event)
1010 struct perf_sample_data *data;
1011 u64 sample_type = event->attr.sample_type;
1014 if (sample_type & PERF_SAMPLE_TID)
1015 size += sizeof(data->tid_entry);
1017 if (sample_type & PERF_SAMPLE_TIME)
1018 size += sizeof(data->time);
1020 if (sample_type & PERF_SAMPLE_ID)
1021 size += sizeof(data->id);
1023 if (sample_type & PERF_SAMPLE_STREAM_ID)
1024 size += sizeof(data->stream_id);
1026 if (sample_type & PERF_SAMPLE_CPU)
1027 size += sizeof(data->cpu_entry);
1029 event->id_header_size = size;
1032 static void perf_group_attach(struct perf_event *event)
1034 struct perf_event *group_leader = event->group_leader, *pos;
1037 * We can have double attach due to group movement in perf_event_open.
1039 if (event->attach_state & PERF_ATTACH_GROUP)
1042 event->attach_state |= PERF_ATTACH_GROUP;
1044 if (group_leader == event)
1047 if (group_leader->group_flags & PERF_GROUP_SOFTWARE &&
1048 !is_software_event(event))
1049 group_leader->group_flags &= ~PERF_GROUP_SOFTWARE;
1051 list_add_tail(&event->group_entry, &group_leader->sibling_list);
1052 group_leader->nr_siblings++;
1054 perf_event__header_size(group_leader);
1056 list_for_each_entry(pos, &group_leader->sibling_list, group_entry)
1057 perf_event__header_size(pos);
1061 * Remove a event from the lists for its context.
1062 * Must be called with ctx->mutex and ctx->lock held.
1065 list_del_event(struct perf_event *event, struct perf_event_context *ctx)
1067 struct perf_cpu_context *cpuctx;
1069 * We can have double detach due to exit/hot-unplug + close.
1071 if (!(event->attach_state & PERF_ATTACH_CONTEXT))
1074 event->attach_state &= ~PERF_ATTACH_CONTEXT;
1076 if (is_cgroup_event(event)) {
1078 cpuctx = __get_cpu_context(ctx);
1080 * if there are no more cgroup events
1081 * then cler cgrp to avoid stale pointer
1082 * in update_cgrp_time_from_cpuctx()
1084 if (!ctx->nr_cgroups)
1085 cpuctx->cgrp = NULL;
1088 if (has_branch_stack(event))
1089 ctx->nr_branch_stack--;
1092 if (event->attr.inherit_stat)
1095 list_del_rcu(&event->event_entry);
1097 if (event->group_leader == event)
1098 list_del_init(&event->group_entry);
1100 update_group_times(event);
1103 * If event was in error state, then keep it
1104 * that way, otherwise bogus counts will be
1105 * returned on read(). The only way to get out
1106 * of error state is by explicit re-enabling
1109 if (event->state > PERF_EVENT_STATE_OFF)
1110 event->state = PERF_EVENT_STATE_OFF;
1113 static void perf_group_detach(struct perf_event *event)
1115 struct perf_event *sibling, *tmp;
1116 struct list_head *list = NULL;
1119 * We can have double detach due to exit/hot-unplug + close.
1121 if (!(event->attach_state & PERF_ATTACH_GROUP))
1124 event->attach_state &= ~PERF_ATTACH_GROUP;
1127 * If this is a sibling, remove it from its group.
1129 if (event->group_leader != event) {
1130 list_del_init(&event->group_entry);
1131 event->group_leader->nr_siblings--;
1135 if (!list_empty(&event->group_entry))
1136 list = &event->group_entry;
1139 * If this was a group event with sibling events then
1140 * upgrade the siblings to singleton events by adding them
1141 * to whatever list we are on.
1143 list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) {
1145 list_move_tail(&sibling->group_entry, list);
1146 sibling->group_leader = sibling;
1148 /* Inherit group flags from the previous leader */
1149 sibling->group_flags = event->group_flags;
1153 perf_event__header_size(event->group_leader);
1155 list_for_each_entry(tmp, &event->group_leader->sibling_list, group_entry)
1156 perf_event__header_size(tmp);
1160 event_filter_match(struct perf_event *event)
1162 return (event->cpu == -1 || event->cpu == smp_processor_id())
1163 && perf_cgroup_match(event);
1167 event_sched_out(struct perf_event *event,
1168 struct perf_cpu_context *cpuctx,
1169 struct perf_event_context *ctx)
1171 u64 tstamp = perf_event_time(event);
1174 * An event which could not be activated because of
1175 * filter mismatch still needs to have its timings
1176 * maintained, otherwise bogus information is return
1177 * via read() for time_enabled, time_running:
1179 if (event->state == PERF_EVENT_STATE_INACTIVE
1180 && !event_filter_match(event)) {
1181 delta = tstamp - event->tstamp_stopped;
1182 event->tstamp_running += delta;
1183 event->tstamp_stopped = tstamp;
1186 if (event->state != PERF_EVENT_STATE_ACTIVE)
1189 event->state = PERF_EVENT_STATE_INACTIVE;
1190 if (event->pending_disable) {
1191 event->pending_disable = 0;
1192 event->state = PERF_EVENT_STATE_OFF;
1194 event->tstamp_stopped = tstamp;
1195 event->pmu->del(event, 0);
1198 if (!is_software_event(event))
1199 cpuctx->active_oncpu--;
1201 if (event->attr.freq && event->attr.sample_freq)
1203 if (event->attr.exclusive || !cpuctx->active_oncpu)
1204 cpuctx->exclusive = 0;
1208 group_sched_out(struct perf_event *group_event,
1209 struct perf_cpu_context *cpuctx,
1210 struct perf_event_context *ctx)
1212 struct perf_event *event;
1213 int state = group_event->state;
1215 event_sched_out(group_event, cpuctx, ctx);
1218 * Schedule out siblings (if any):
1220 list_for_each_entry(event, &group_event->sibling_list, group_entry)
1221 event_sched_out(event, cpuctx, ctx);
1223 if (state == PERF_EVENT_STATE_ACTIVE && group_event->attr.exclusive)
1224 cpuctx->exclusive = 0;
1228 * Cross CPU call to remove a performance event
1230 * We disable the event on the hardware level first. After that we
1231 * remove it from the context list.
1233 static int __perf_remove_from_context(void *info)
1235 struct perf_event *event = info;
1236 struct perf_event_context *ctx = event->ctx;
1237 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1239 raw_spin_lock(&ctx->lock);
1240 event_sched_out(event, cpuctx, ctx);
1241 list_del_event(event, ctx);
1242 if (!ctx->nr_events && cpuctx->task_ctx == ctx) {
1244 cpuctx->task_ctx = NULL;
1246 raw_spin_unlock(&ctx->lock);
1253 * Remove the event from a task's (or a CPU's) list of events.
1255 * CPU events are removed with a smp call. For task events we only
1256 * call when the task is on a CPU.
1258 * If event->ctx is a cloned context, callers must make sure that
1259 * every task struct that event->ctx->task could possibly point to
1260 * remains valid. This is OK when called from perf_release since
1261 * that only calls us on the top-level context, which can't be a clone.
1262 * When called from perf_event_exit_task, it's OK because the
1263 * context has been detached from its task.
1265 static void perf_remove_from_context(struct perf_event *event)
1267 struct perf_event_context *ctx = event->ctx;
1268 struct task_struct *task = ctx->task;
1270 lockdep_assert_held(&ctx->mutex);
1274 * Per cpu events are removed via an smp call and
1275 * the removal is always successful.
1277 cpu_function_call(event->cpu, __perf_remove_from_context, event);
1282 if (!task_function_call(task, __perf_remove_from_context, event))
1285 raw_spin_lock_irq(&ctx->lock);
1287 * If we failed to find a running task, but find the context active now
1288 * that we've acquired the ctx->lock, retry.
1290 if (ctx->is_active) {
1291 raw_spin_unlock_irq(&ctx->lock);
1296 * Since the task isn't running, its safe to remove the event, us
1297 * holding the ctx->lock ensures the task won't get scheduled in.
1299 list_del_event(event, ctx);
1300 raw_spin_unlock_irq(&ctx->lock);
1304 * Cross CPU call to disable a performance event
1306 int __perf_event_disable(void *info)
1308 struct perf_event *event = info;
1309 struct perf_event_context *ctx = event->ctx;
1310 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1313 * If this is a per-task event, need to check whether this
1314 * event's task is the current task on this cpu.
1316 * Can trigger due to concurrent perf_event_context_sched_out()
1317 * flipping contexts around.
1319 if (ctx->task && cpuctx->task_ctx != ctx)
1322 raw_spin_lock(&ctx->lock);
1325 * If the event is on, turn it off.
1326 * If it is in error state, leave it in error state.
1328 if (event->state >= PERF_EVENT_STATE_INACTIVE) {
1329 update_context_time(ctx);
1330 update_cgrp_time_from_event(event);
1331 update_group_times(event);
1332 if (event == event->group_leader)
1333 group_sched_out(event, cpuctx, ctx);
1335 event_sched_out(event, cpuctx, ctx);
1336 event->state = PERF_EVENT_STATE_OFF;
1339 raw_spin_unlock(&ctx->lock);
1347 * If event->ctx is a cloned context, callers must make sure that
1348 * every task struct that event->ctx->task could possibly point to
1349 * remains valid. This condition is satisifed when called through
1350 * perf_event_for_each_child or perf_event_for_each because they
1351 * hold the top-level event's child_mutex, so any descendant that
1352 * goes to exit will block in sync_child_event.
1353 * When called from perf_pending_event it's OK because event->ctx
1354 * is the current context on this CPU and preemption is disabled,
1355 * hence we can't get into perf_event_task_sched_out for this context.
1357 void perf_event_disable(struct perf_event *event)
1359 struct perf_event_context *ctx = event->ctx;
1360 struct task_struct *task = ctx->task;
1364 * Disable the event on the cpu that it's on
1366 cpu_function_call(event->cpu, __perf_event_disable, event);
1371 if (!task_function_call(task, __perf_event_disable, event))
1374 raw_spin_lock_irq(&ctx->lock);
1376 * If the event is still active, we need to retry the cross-call.
1378 if (event->state == PERF_EVENT_STATE_ACTIVE) {
1379 raw_spin_unlock_irq(&ctx->lock);
1381 * Reload the task pointer, it might have been changed by
1382 * a concurrent perf_event_context_sched_out().
1389 * Since we have the lock this context can't be scheduled
1390 * in, so we can change the state safely.
1392 if (event->state == PERF_EVENT_STATE_INACTIVE) {
1393 update_group_times(event);
1394 event->state = PERF_EVENT_STATE_OFF;
1396 raw_spin_unlock_irq(&ctx->lock);
1398 EXPORT_SYMBOL_GPL(perf_event_disable);
1400 static void perf_set_shadow_time(struct perf_event *event,
1401 struct perf_event_context *ctx,
1405 * use the correct time source for the time snapshot
1407 * We could get by without this by leveraging the
1408 * fact that to get to this function, the caller
1409 * has most likely already called update_context_time()
1410 * and update_cgrp_time_xx() and thus both timestamp
1411 * are identical (or very close). Given that tstamp is,
1412 * already adjusted for cgroup, we could say that:
1413 * tstamp - ctx->timestamp
1415 * tstamp - cgrp->timestamp.
1417 * Then, in perf_output_read(), the calculation would
1418 * work with no changes because:
1419 * - event is guaranteed scheduled in
1420 * - no scheduled out in between
1421 * - thus the timestamp would be the same
1423 * But this is a bit hairy.
1425 * So instead, we have an explicit cgroup call to remain
1426 * within the time time source all along. We believe it
1427 * is cleaner and simpler to understand.
1429 if (is_cgroup_event(event))
1430 perf_cgroup_set_shadow_time(event, tstamp);
1432 event->shadow_ctx_time = tstamp - ctx->timestamp;
1435 #define MAX_INTERRUPTS (~0ULL)
1437 static void perf_log_throttle(struct perf_event *event, int enable);
1440 event_sched_in(struct perf_event *event,
1441 struct perf_cpu_context *cpuctx,
1442 struct perf_event_context *ctx)
1444 u64 tstamp = perf_event_time(event);
1446 if (event->state <= PERF_EVENT_STATE_OFF)
1449 event->state = PERF_EVENT_STATE_ACTIVE;
1450 event->oncpu = smp_processor_id();
1453 * Unthrottle events, since we scheduled we might have missed several
1454 * ticks already, also for a heavily scheduling task there is little
1455 * guarantee it'll get a tick in a timely manner.
1457 if (unlikely(event->hw.interrupts == MAX_INTERRUPTS)) {
1458 perf_log_throttle(event, 1);
1459 event->hw.interrupts = 0;
1463 * The new state must be visible before we turn it on in the hardware:
1467 if (event->pmu->add(event, PERF_EF_START)) {
1468 event->state = PERF_EVENT_STATE_INACTIVE;
1473 event->tstamp_running += tstamp - event->tstamp_stopped;
1475 perf_set_shadow_time(event, ctx, tstamp);
1477 if (!is_software_event(event))
1478 cpuctx->active_oncpu++;
1480 if (event->attr.freq && event->attr.sample_freq)
1483 if (event->attr.exclusive)
1484 cpuctx->exclusive = 1;
1490 group_sched_in(struct perf_event *group_event,
1491 struct perf_cpu_context *cpuctx,
1492 struct perf_event_context *ctx)
1494 struct perf_event *event, *partial_group = NULL;
1495 struct pmu *pmu = group_event->pmu;
1496 u64 now = ctx->time;
1497 bool simulate = false;
1499 if (group_event->state == PERF_EVENT_STATE_OFF)
1502 pmu->start_txn(pmu);
1504 if (event_sched_in(group_event, cpuctx, ctx)) {
1505 pmu->cancel_txn(pmu);
1510 * Schedule in siblings as one group (if any):
1512 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
1513 if (event_sched_in(event, cpuctx, ctx)) {
1514 partial_group = event;
1519 if (!pmu->commit_txn(pmu))
1524 * Groups can be scheduled in as one unit only, so undo any
1525 * partial group before returning:
1526 * The events up to the failed event are scheduled out normally,
1527 * tstamp_stopped will be updated.
1529 * The failed events and the remaining siblings need to have
1530 * their timings updated as if they had gone thru event_sched_in()
1531 * and event_sched_out(). This is required to get consistent timings
1532 * across the group. This also takes care of the case where the group
1533 * could never be scheduled by ensuring tstamp_stopped is set to mark
1534 * the time the event was actually stopped, such that time delta
1535 * calculation in update_event_times() is correct.
1537 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
1538 if (event == partial_group)
1542 event->tstamp_running += now - event->tstamp_stopped;
1543 event->tstamp_stopped = now;
1545 event_sched_out(event, cpuctx, ctx);
1548 event_sched_out(group_event, cpuctx, ctx);
1550 pmu->cancel_txn(pmu);
1556 * Work out whether we can put this event group on the CPU now.
1558 static int group_can_go_on(struct perf_event *event,
1559 struct perf_cpu_context *cpuctx,
1563 * Groups consisting entirely of software events can always go on.
1565 if (event->group_flags & PERF_GROUP_SOFTWARE)
1568 * If an exclusive group is already on, no other hardware
1571 if (cpuctx->exclusive)
1574 * If this group is exclusive and there are already
1575 * events on the CPU, it can't go on.
1577 if (event->attr.exclusive && cpuctx->active_oncpu)
1580 * Otherwise, try to add it if all previous groups were able
1586 static void add_event_to_ctx(struct perf_event *event,
1587 struct perf_event_context *ctx)
1589 u64 tstamp = perf_event_time(event);
1591 list_add_event(event, ctx);
1592 perf_group_attach(event);
1593 event->tstamp_enabled = tstamp;
1594 event->tstamp_running = tstamp;
1595 event->tstamp_stopped = tstamp;
1598 static void task_ctx_sched_out(struct perf_event_context *ctx);
1600 ctx_sched_in(struct perf_event_context *ctx,
1601 struct perf_cpu_context *cpuctx,
1602 enum event_type_t event_type,
1603 struct task_struct *task);
1605 static void perf_event_sched_in(struct perf_cpu_context *cpuctx,
1606 struct perf_event_context *ctx,
1607 struct task_struct *task)
1609 cpu_ctx_sched_in(cpuctx, EVENT_PINNED, task);
1611 ctx_sched_in(ctx, cpuctx, EVENT_PINNED, task);
1612 cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE, task);
1614 ctx_sched_in(ctx, cpuctx, EVENT_FLEXIBLE, task);
1618 * Cross CPU call to install and enable a performance event
1620 * Must be called with ctx->mutex held
1622 static int __perf_install_in_context(void *info)
1624 struct perf_event *event = info;
1625 struct perf_event_context *ctx = event->ctx;
1626 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1627 struct perf_event_context *task_ctx = cpuctx->task_ctx;
1628 struct task_struct *task = current;
1630 perf_ctx_lock(cpuctx, task_ctx);
1631 perf_pmu_disable(cpuctx->ctx.pmu);
1634 * If there was an active task_ctx schedule it out.
1637 task_ctx_sched_out(task_ctx);
1640 * If the context we're installing events in is not the
1641 * active task_ctx, flip them.
1643 if (ctx->task && task_ctx != ctx) {
1645 raw_spin_unlock(&task_ctx->lock);
1646 raw_spin_lock(&ctx->lock);
1651 cpuctx->task_ctx = task_ctx;
1652 task = task_ctx->task;
1655 cpu_ctx_sched_out(cpuctx, EVENT_ALL);
1657 update_context_time(ctx);
1659 * update cgrp time only if current cgrp
1660 * matches event->cgrp. Must be done before
1661 * calling add_event_to_ctx()
1663 update_cgrp_time_from_event(event);
1665 add_event_to_ctx(event, ctx);
1668 * Schedule everything back in
1670 perf_event_sched_in(cpuctx, task_ctx, task);
1672 perf_pmu_enable(cpuctx->ctx.pmu);
1673 perf_ctx_unlock(cpuctx, task_ctx);
1679 * Attach a performance event to a context
1681 * First we add the event to the list with the hardware enable bit
1682 * in event->hw_config cleared.
1684 * If the event is attached to a task which is on a CPU we use a smp
1685 * call to enable it in the task context. The task might have been
1686 * scheduled away, but we check this in the smp call again.
1689 perf_install_in_context(struct perf_event_context *ctx,
1690 struct perf_event *event,
1693 struct task_struct *task = ctx->task;
1695 lockdep_assert_held(&ctx->mutex);
1698 if (event->cpu != -1)
1703 * Per cpu events are installed via an smp call and
1704 * the install is always successful.
1706 cpu_function_call(cpu, __perf_install_in_context, event);
1711 if (!task_function_call(task, __perf_install_in_context, event))
1714 raw_spin_lock_irq(&ctx->lock);
1716 * If we failed to find a running task, but find the context active now
1717 * that we've acquired the ctx->lock, retry.
1719 if (ctx->is_active) {
1720 raw_spin_unlock_irq(&ctx->lock);
1725 * Since the task isn't running, its safe to add the event, us holding
1726 * the ctx->lock ensures the task won't get scheduled in.
1728 add_event_to_ctx(event, ctx);
1729 raw_spin_unlock_irq(&ctx->lock);
1733 * Put a event into inactive state and update time fields.
1734 * Enabling the leader of a group effectively enables all
1735 * the group members that aren't explicitly disabled, so we
1736 * have to update their ->tstamp_enabled also.
1737 * Note: this works for group members as well as group leaders
1738 * since the non-leader members' sibling_lists will be empty.
1740 static void __perf_event_mark_enabled(struct perf_event *event)
1742 struct perf_event *sub;
1743 u64 tstamp = perf_event_time(event);
1745 event->state = PERF_EVENT_STATE_INACTIVE;
1746 event->tstamp_enabled = tstamp - event->total_time_enabled;
1747 list_for_each_entry(sub, &event->sibling_list, group_entry) {
1748 if (sub->state >= PERF_EVENT_STATE_INACTIVE)
1749 sub->tstamp_enabled = tstamp - sub->total_time_enabled;
1754 * Cross CPU call to enable a performance event
1756 static int __perf_event_enable(void *info)
1758 struct perf_event *event = info;
1759 struct perf_event_context *ctx = event->ctx;
1760 struct perf_event *leader = event->group_leader;
1761 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1765 * There's a time window between 'ctx->is_active' check
1766 * in perf_event_enable function and this place having:
1768 * - ctx->lock unlocked
1770 * where the task could be killed and 'ctx' deactivated
1771 * by perf_event_exit_task.
1773 if (!ctx->is_active)
1776 raw_spin_lock(&ctx->lock);
1777 update_context_time(ctx);
1779 if (event->state >= PERF_EVENT_STATE_INACTIVE)
1783 * set current task's cgroup time reference point
1785 perf_cgroup_set_timestamp(current, ctx);
1787 __perf_event_mark_enabled(event);
1789 if (!event_filter_match(event)) {
1790 if (is_cgroup_event(event))
1791 perf_cgroup_defer_enabled(event);
1796 * If the event is in a group and isn't the group leader,
1797 * then don't put it on unless the group is on.
1799 if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE)
1802 if (!group_can_go_on(event, cpuctx, 1)) {
1805 if (event == leader)
1806 err = group_sched_in(event, cpuctx, ctx);
1808 err = event_sched_in(event, cpuctx, ctx);
1813 * If this event can't go on and it's part of a
1814 * group, then the whole group has to come off.
1816 if (leader != event)
1817 group_sched_out(leader, cpuctx, ctx);
1818 if (leader->attr.pinned) {
1819 update_group_times(leader);
1820 leader->state = PERF_EVENT_STATE_ERROR;
1825 raw_spin_unlock(&ctx->lock);
1833 * If event->ctx is a cloned context, callers must make sure that
1834 * every task struct that event->ctx->task could possibly point to
1835 * remains valid. This condition is satisfied when called through
1836 * perf_event_for_each_child or perf_event_for_each as described
1837 * for perf_event_disable.
1839 void perf_event_enable(struct perf_event *event)
1841 struct perf_event_context *ctx = event->ctx;
1842 struct task_struct *task = ctx->task;
1846 * Enable the event on the cpu that it's on
1848 cpu_function_call(event->cpu, __perf_event_enable, event);
1852 raw_spin_lock_irq(&ctx->lock);
1853 if (event->state >= PERF_EVENT_STATE_INACTIVE)
1857 * If the event is in error state, clear that first.
1858 * That way, if we see the event in error state below, we
1859 * know that it has gone back into error state, as distinct
1860 * from the task having been scheduled away before the
1861 * cross-call arrived.
1863 if (event->state == PERF_EVENT_STATE_ERROR)
1864 event->state = PERF_EVENT_STATE_OFF;
1867 if (!ctx->is_active) {
1868 __perf_event_mark_enabled(event);
1872 raw_spin_unlock_irq(&ctx->lock);
1874 if (!task_function_call(task, __perf_event_enable, event))
1877 raw_spin_lock_irq(&ctx->lock);
1880 * If the context is active and the event is still off,
1881 * we need to retry the cross-call.
1883 if (ctx->is_active && event->state == PERF_EVENT_STATE_OFF) {
1885 * task could have been flipped by a concurrent
1886 * perf_event_context_sched_out()
1893 raw_spin_unlock_irq(&ctx->lock);
1895 EXPORT_SYMBOL_GPL(perf_event_enable);
1897 int perf_event_refresh(struct perf_event *event, int refresh)
1900 * not supported on inherited events
1902 if (event->attr.inherit || !is_sampling_event(event))
1905 atomic_add(refresh, &event->event_limit);
1906 perf_event_enable(event);
1910 EXPORT_SYMBOL_GPL(perf_event_refresh);
1912 static void ctx_sched_out(struct perf_event_context *ctx,
1913 struct perf_cpu_context *cpuctx,
1914 enum event_type_t event_type)
1916 struct perf_event *event;
1917 int is_active = ctx->is_active;
1919 ctx->is_active &= ~event_type;
1920 if (likely(!ctx->nr_events))
1923 update_context_time(ctx);
1924 update_cgrp_time_from_cpuctx(cpuctx);
1925 if (!ctx->nr_active)
1928 perf_pmu_disable(ctx->pmu);
1929 if ((is_active & EVENT_PINNED) && (event_type & EVENT_PINNED)) {
1930 list_for_each_entry(event, &ctx->pinned_groups, group_entry)
1931 group_sched_out(event, cpuctx, ctx);
1934 if ((is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE)) {
1935 list_for_each_entry(event, &ctx->flexible_groups, group_entry)
1936 group_sched_out(event, cpuctx, ctx);
1938 perf_pmu_enable(ctx->pmu);
1942 * Test whether two contexts are equivalent, i.e. whether they
1943 * have both been cloned from the same version of the same context
1944 * and they both have the same number of enabled events.
1945 * If the number of enabled events is the same, then the set
1946 * of enabled events should be the same, because these are both
1947 * inherited contexts, therefore we can't access individual events
1948 * in them directly with an fd; we can only enable/disable all
1949 * events via prctl, or enable/disable all events in a family
1950 * via ioctl, which will have the same effect on both contexts.
1952 static int context_equiv(struct perf_event_context *ctx1,
1953 struct perf_event_context *ctx2)
1955 return ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx
1956 && ctx1->parent_gen == ctx2->parent_gen
1957 && !ctx1->pin_count && !ctx2->pin_count;
1960 static void __perf_event_sync_stat(struct perf_event *event,
1961 struct perf_event *next_event)
1965 if (!event->attr.inherit_stat)
1969 * Update the event value, we cannot use perf_event_read()
1970 * because we're in the middle of a context switch and have IRQs
1971 * disabled, which upsets smp_call_function_single(), however
1972 * we know the event must be on the current CPU, therefore we
1973 * don't need to use it.
1975 switch (event->state) {
1976 case PERF_EVENT_STATE_ACTIVE:
1977 event->pmu->read(event);
1980 case PERF_EVENT_STATE_INACTIVE:
1981 update_event_times(event);
1989 * In order to keep per-task stats reliable we need to flip the event
1990 * values when we flip the contexts.
1992 value = local64_read(&next_event->count);
1993 value = local64_xchg(&event->count, value);
1994 local64_set(&next_event->count, value);
1996 swap(event->total_time_enabled, next_event->total_time_enabled);
1997 swap(event->total_time_running, next_event->total_time_running);
2000 * Since we swizzled the values, update the user visible data too.
2002 perf_event_update_userpage(event);
2003 perf_event_update_userpage(next_event);
2006 #define list_next_entry(pos, member) \
2007 list_entry(pos->member.next, typeof(*pos), member)
2009 static void perf_event_sync_stat(struct perf_event_context *ctx,
2010 struct perf_event_context *next_ctx)
2012 struct perf_event *event, *next_event;
2017 update_context_time(ctx);
2019 event = list_first_entry(&ctx->event_list,
2020 struct perf_event, event_entry);
2022 next_event = list_first_entry(&next_ctx->event_list,
2023 struct perf_event, event_entry);
2025 while (&event->event_entry != &ctx->event_list &&
2026 &next_event->event_entry != &next_ctx->event_list) {
2028 __perf_event_sync_stat(event, next_event);
2030 event = list_next_entry(event, event_entry);
2031 next_event = list_next_entry(next_event, event_entry);
2035 static void perf_event_context_sched_out(struct task_struct *task, int ctxn,
2036 struct task_struct *next)
2038 struct perf_event_context *ctx = task->perf_event_ctxp[ctxn];
2039 struct perf_event_context *next_ctx;
2040 struct perf_event_context *parent;
2041 struct perf_cpu_context *cpuctx;
2047 cpuctx = __get_cpu_context(ctx);
2048 if (!cpuctx->task_ctx)
2052 parent = rcu_dereference(ctx->parent_ctx);
2053 next_ctx = next->perf_event_ctxp[ctxn];
2054 if (parent && next_ctx &&
2055 rcu_dereference(next_ctx->parent_ctx) == parent) {
2057 * Looks like the two contexts are clones, so we might be
2058 * able to optimize the context switch. We lock both
2059 * contexts and check that they are clones under the
2060 * lock (including re-checking that neither has been
2061 * uncloned in the meantime). It doesn't matter which
2062 * order we take the locks because no other cpu could
2063 * be trying to lock both of these tasks.
2065 raw_spin_lock(&ctx->lock);
2066 raw_spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING);
2067 if (context_equiv(ctx, next_ctx)) {
2069 * XXX do we need a memory barrier of sorts
2070 * wrt to rcu_dereference() of perf_event_ctxp
2072 task->perf_event_ctxp[ctxn] = next_ctx;
2073 next->perf_event_ctxp[ctxn] = ctx;
2075 next_ctx->task = task;
2078 perf_event_sync_stat(ctx, next_ctx);
2080 raw_spin_unlock(&next_ctx->lock);
2081 raw_spin_unlock(&ctx->lock);
2086 raw_spin_lock(&ctx->lock);
2087 ctx_sched_out(ctx, cpuctx, EVENT_ALL);
2088 cpuctx->task_ctx = NULL;
2089 raw_spin_unlock(&ctx->lock);
2093 #define for_each_task_context_nr(ctxn) \
2094 for ((ctxn) = 0; (ctxn) < perf_nr_task_contexts; (ctxn)++)
2097 * Called from scheduler to remove the events of the current task,
2098 * with interrupts disabled.
2100 * We stop each event and update the event value in event->count.
2102 * This does not protect us against NMI, but disable()
2103 * sets the disabled bit in the control field of event _before_
2104 * accessing the event control register. If a NMI hits, then it will
2105 * not restart the event.
2107 void __perf_event_task_sched_out(struct task_struct *task,
2108 struct task_struct *next)
2112 for_each_task_context_nr(ctxn)
2113 perf_event_context_sched_out(task, ctxn, next);
2116 * if cgroup events exist on this CPU, then we need
2117 * to check if we have to switch out PMU state.
2118 * cgroup event are system-wide mode only
2120 if (atomic_read(&__get_cpu_var(perf_cgroup_events)))
2121 perf_cgroup_sched_out(task, next);
2124 static void task_ctx_sched_out(struct perf_event_context *ctx)
2126 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
2128 if (!cpuctx->task_ctx)
2131 if (WARN_ON_ONCE(ctx != cpuctx->task_ctx))
2134 ctx_sched_out(ctx, cpuctx, EVENT_ALL);
2135 cpuctx->task_ctx = NULL;
2139 * Called with IRQs disabled
2141 static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
2142 enum event_type_t event_type)
2144 ctx_sched_out(&cpuctx->ctx, cpuctx, event_type);
2148 ctx_pinned_sched_in(struct perf_event_context *ctx,
2149 struct perf_cpu_context *cpuctx)
2151 struct perf_event *event;
2153 list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
2154 if (event->state <= PERF_EVENT_STATE_OFF)
2156 if (!event_filter_match(event))
2159 /* may need to reset tstamp_enabled */
2160 if (is_cgroup_event(event))
2161 perf_cgroup_mark_enabled(event, ctx);
2163 if (group_can_go_on(event, cpuctx, 1))
2164 group_sched_in(event, cpuctx, ctx);
2167 * If this pinned group hasn't been scheduled,
2168 * put it in error state.
2170 if (event->state == PERF_EVENT_STATE_INACTIVE) {
2171 update_group_times(event);
2172 event->state = PERF_EVENT_STATE_ERROR;
2178 ctx_flexible_sched_in(struct perf_event_context *ctx,
2179 struct perf_cpu_context *cpuctx)
2181 struct perf_event *event;
2184 list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
2185 /* Ignore events in OFF or ERROR state */
2186 if (event->state <= PERF_EVENT_STATE_OFF)
2189 * Listen to the 'cpu' scheduling filter constraint
2192 if (!event_filter_match(event))
2195 /* may need to reset tstamp_enabled */
2196 if (is_cgroup_event(event))
2197 perf_cgroup_mark_enabled(event, ctx);
2199 if (group_can_go_on(event, cpuctx, can_add_hw)) {
2200 if (group_sched_in(event, cpuctx, ctx))
2207 ctx_sched_in(struct perf_event_context *ctx,
2208 struct perf_cpu_context *cpuctx,
2209 enum event_type_t event_type,
2210 struct task_struct *task)
2213 int is_active = ctx->is_active;
2215 ctx->is_active |= event_type;
2216 if (likely(!ctx->nr_events))
2220 ctx->timestamp = now;
2221 perf_cgroup_set_timestamp(task, ctx);
2223 * First go through the list and put on any pinned groups
2224 * in order to give them the best chance of going on.
2226 if (!(is_active & EVENT_PINNED) && (event_type & EVENT_PINNED))
2227 ctx_pinned_sched_in(ctx, cpuctx);
2229 /* Then walk through the lower prio flexible groups */
2230 if (!(is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE))
2231 ctx_flexible_sched_in(ctx, cpuctx);
2234 static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
2235 enum event_type_t event_type,
2236 struct task_struct *task)
2238 struct perf_event_context *ctx = &cpuctx->ctx;
2240 ctx_sched_in(ctx, cpuctx, event_type, task);
2243 static void perf_event_context_sched_in(struct perf_event_context *ctx,
2244 struct task_struct *task)
2246 struct perf_cpu_context *cpuctx;
2248 cpuctx = __get_cpu_context(ctx);
2249 if (cpuctx->task_ctx == ctx)
2252 perf_ctx_lock(cpuctx, ctx);
2253 perf_pmu_disable(ctx->pmu);
2255 * We want to keep the following priority order:
2256 * cpu pinned (that don't need to move), task pinned,
2257 * cpu flexible, task flexible.
2259 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
2262 cpuctx->task_ctx = ctx;
2264 perf_event_sched_in(cpuctx, cpuctx->task_ctx, task);
2266 perf_pmu_enable(ctx->pmu);
2267 perf_ctx_unlock(cpuctx, ctx);
2270 * Since these rotations are per-cpu, we need to ensure the
2271 * cpu-context we got scheduled on is actually rotating.
2273 perf_pmu_rotate_start(ctx->pmu);
2277 * When sampling the branck stack in system-wide, it may be necessary
2278 * to flush the stack on context switch. This happens when the branch
2279 * stack does not tag its entries with the pid of the current task.
2280 * Otherwise it becomes impossible to associate a branch entry with a
2281 * task. This ambiguity is more likely to appear when the branch stack
2282 * supports priv level filtering and the user sets it to monitor only
2283 * at the user level (which could be a useful measurement in system-wide
2284 * mode). In that case, the risk is high of having a branch stack with
2285 * branch from multiple tasks. Flushing may mean dropping the existing
2286 * entries or stashing them somewhere in the PMU specific code layer.
2288 * This function provides the context switch callback to the lower code
2289 * layer. It is invoked ONLY when there is at least one system-wide context
2290 * with at least one active event using taken branch sampling.
2292 static void perf_branch_stack_sched_in(struct task_struct *prev,
2293 struct task_struct *task)
2295 struct perf_cpu_context *cpuctx;
2297 unsigned long flags;
2299 /* no need to flush branch stack if not changing task */
2303 local_irq_save(flags);
2307 list_for_each_entry_rcu(pmu, &pmus, entry) {
2308 cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
2311 * check if the context has at least one
2312 * event using PERF_SAMPLE_BRANCH_STACK
2314 if (cpuctx->ctx.nr_branch_stack > 0
2315 && pmu->flush_branch_stack) {
2317 pmu = cpuctx->ctx.pmu;
2319 perf_ctx_lock(cpuctx, cpuctx->task_ctx);
2321 perf_pmu_disable(pmu);
2323 pmu->flush_branch_stack();
2325 perf_pmu_enable(pmu);
2327 perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
2333 local_irq_restore(flags);
2337 * Called from scheduler to add the events of the current task
2338 * with interrupts disabled.
2340 * We restore the event value and then enable it.
2342 * This does not protect us against NMI, but enable()
2343 * sets the enabled bit in the control field of event _before_
2344 * accessing the event control register. If a NMI hits, then it will
2345 * keep the event running.
2347 void __perf_event_task_sched_in(struct task_struct *prev,
2348 struct task_struct *task)
2350 struct perf_event_context *ctx;
2353 for_each_task_context_nr(ctxn) {
2354 ctx = task->perf_event_ctxp[ctxn];
2358 perf_event_context_sched_in(ctx, task);
2361 * if cgroup events exist on this CPU, then we need
2362 * to check if we have to switch in PMU state.
2363 * cgroup event are system-wide mode only
2365 if (atomic_read(&__get_cpu_var(perf_cgroup_events)))
2366 perf_cgroup_sched_in(prev, task);
2368 /* check for system-wide branch_stack events */
2369 if (atomic_read(&__get_cpu_var(perf_branch_stack_events)))
2370 perf_branch_stack_sched_in(prev, task);
2373 static u64 perf_calculate_period(struct perf_event *event, u64 nsec, u64 count)
2375 u64 frequency = event->attr.sample_freq;
2376 u64 sec = NSEC_PER_SEC;
2377 u64 divisor, dividend;
2379 int count_fls, nsec_fls, frequency_fls, sec_fls;
2381 count_fls = fls64(count);
2382 nsec_fls = fls64(nsec);
2383 frequency_fls = fls64(frequency);
2387 * We got @count in @nsec, with a target of sample_freq HZ
2388 * the target period becomes:
2391 * period = -------------------
2392 * @nsec * sample_freq
2397 * Reduce accuracy by one bit such that @a and @b converge
2398 * to a similar magnitude.
2400 #define REDUCE_FLS(a, b) \
2402 if (a##_fls > b##_fls) { \
2412 * Reduce accuracy until either term fits in a u64, then proceed with
2413 * the other, so that finally we can do a u64/u64 division.
2415 while (count_fls + sec_fls > 64 && nsec_fls + frequency_fls > 64) {
2416 REDUCE_FLS(nsec, frequency);
2417 REDUCE_FLS(sec, count);
2420 if (count_fls + sec_fls > 64) {
2421 divisor = nsec * frequency;
2423 while (count_fls + sec_fls > 64) {
2424 REDUCE_FLS(count, sec);
2428 dividend = count * sec;
2430 dividend = count * sec;
2432 while (nsec_fls + frequency_fls > 64) {
2433 REDUCE_FLS(nsec, frequency);
2437 divisor = nsec * frequency;
2443 return div64_u64(dividend, divisor);
2446 static DEFINE_PER_CPU(int, perf_throttled_count);
2447 static DEFINE_PER_CPU(u64, perf_throttled_seq);
2449 static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count, bool disable)
2451 struct hw_perf_event *hwc = &event->hw;
2452 s64 period, sample_period;
2455 period = perf_calculate_period(event, nsec, count);
2457 delta = (s64)(period - hwc->sample_period);
2458 delta = (delta + 7) / 8; /* low pass filter */
2460 sample_period = hwc->sample_period + delta;
2465 hwc->sample_period = sample_period;
2467 if (local64_read(&hwc->period_left) > 8*sample_period) {
2469 event->pmu->stop(event, PERF_EF_UPDATE);
2471 local64_set(&hwc->period_left, 0);
2474 event->pmu->start(event, PERF_EF_RELOAD);
2479 * combine freq adjustment with unthrottling to avoid two passes over the
2480 * events. At the same time, make sure, having freq events does not change
2481 * the rate of unthrottling as that would introduce bias.
2483 static void perf_adjust_freq_unthr_context(struct perf_event_context *ctx,
2486 struct perf_event *event;
2487 struct hw_perf_event *hwc;
2488 u64 now, period = TICK_NSEC;
2492 * only need to iterate over all events iff:
2493 * - context have events in frequency mode (needs freq adjust)
2494 * - there are events to unthrottle on this cpu
2496 if (!(ctx->nr_freq || needs_unthr))
2499 raw_spin_lock(&ctx->lock);
2500 perf_pmu_disable(ctx->pmu);
2502 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
2503 if (event->state != PERF_EVENT_STATE_ACTIVE)
2506 if (!event_filter_match(event))
2511 if (needs_unthr && hwc->interrupts == MAX_INTERRUPTS) {
2512 hwc->interrupts = 0;
2513 perf_log_throttle(event, 1);
2514 event->pmu->start(event, 0);
2517 if (!event->attr.freq || !event->attr.sample_freq)
2521 * stop the event and update event->count
2523 event->pmu->stop(event, PERF_EF_UPDATE);
2525 now = local64_read(&event->count);
2526 delta = now - hwc->freq_count_stamp;
2527 hwc->freq_count_stamp = now;
2531 * reload only if value has changed
2532 * we have stopped the event so tell that
2533 * to perf_adjust_period() to avoid stopping it
2537 perf_adjust_period(event, period, delta, false);
2539 event->pmu->start(event, delta > 0 ? PERF_EF_RELOAD : 0);
2542 perf_pmu_enable(ctx->pmu);
2543 raw_spin_unlock(&ctx->lock);
2547 * Round-robin a context's events:
2549 static void rotate_ctx(struct perf_event_context *ctx)
2552 * Rotate the first entry last of non-pinned groups. Rotation might be
2553 * disabled by the inheritance code.
2555 if (!ctx->rotate_disable)
2556 list_rotate_left(&ctx->flexible_groups);
2560 * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
2561 * because they're strictly cpu affine and rotate_start is called with IRQs
2562 * disabled, while rotate_context is called from IRQ context.
2564 static void perf_rotate_context(struct perf_cpu_context *cpuctx)
2566 struct perf_event_context *ctx = NULL;
2567 int rotate = 0, remove = 1;
2569 if (cpuctx->ctx.nr_events) {
2571 if (cpuctx->ctx.nr_events != cpuctx->ctx.nr_active)
2575 ctx = cpuctx->task_ctx;
2576 if (ctx && ctx->nr_events) {
2578 if (ctx->nr_events != ctx->nr_active)
2585 perf_ctx_lock(cpuctx, cpuctx->task_ctx);
2586 perf_pmu_disable(cpuctx->ctx.pmu);
2588 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
2590 ctx_sched_out(ctx, cpuctx, EVENT_FLEXIBLE);
2592 rotate_ctx(&cpuctx->ctx);
2596 perf_event_sched_in(cpuctx, ctx, current);
2598 perf_pmu_enable(cpuctx->ctx.pmu);
2599 perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
2602 list_del_init(&cpuctx->rotation_list);
2605 #ifdef CONFIG_NO_HZ_FULL
2606 bool perf_event_can_stop_tick(void)
2608 if (list_empty(&__get_cpu_var(rotation_list)))
2615 void perf_event_task_tick(void)
2617 struct list_head *head = &__get_cpu_var(rotation_list);
2618 struct perf_cpu_context *cpuctx, *tmp;
2619 struct perf_event_context *ctx;
2622 WARN_ON(!irqs_disabled());
2624 __this_cpu_inc(perf_throttled_seq);
2625 throttled = __this_cpu_xchg(perf_throttled_count, 0);
2627 list_for_each_entry_safe(cpuctx, tmp, head, rotation_list) {
2629 perf_adjust_freq_unthr_context(ctx, throttled);
2631 ctx = cpuctx->task_ctx;
2633 perf_adjust_freq_unthr_context(ctx, throttled);
2635 if (cpuctx->jiffies_interval == 1 ||
2636 !(jiffies % cpuctx->jiffies_interval))
2637 perf_rotate_context(cpuctx);
2641 static int event_enable_on_exec(struct perf_event *event,
2642 struct perf_event_context *ctx)
2644 if (!event->attr.enable_on_exec)
2647 event->attr.enable_on_exec = 0;
2648 if (event->state >= PERF_EVENT_STATE_INACTIVE)
2651 __perf_event_mark_enabled(event);
2657 * Enable all of a task's events that have been marked enable-on-exec.
2658 * This expects task == current.
2660 static void perf_event_enable_on_exec(struct perf_event_context *ctx)
2662 struct perf_event *event;
2663 unsigned long flags;
2667 local_irq_save(flags);
2668 if (!ctx || !ctx->nr_events)
2672 * We must ctxsw out cgroup events to avoid conflict
2673 * when invoking perf_task_event_sched_in() later on
2674 * in this function. Otherwise we end up trying to
2675 * ctxswin cgroup events which are already scheduled
2678 perf_cgroup_sched_out(current, NULL);
2680 raw_spin_lock(&ctx->lock);
2681 task_ctx_sched_out(ctx);
2683 list_for_each_entry(event, &ctx->event_list, event_entry) {
2684 ret = event_enable_on_exec(event, ctx);
2690 * Unclone this context if we enabled any event.
2695 raw_spin_unlock(&ctx->lock);
2698 * Also calls ctxswin for cgroup events, if any:
2700 perf_event_context_sched_in(ctx, ctx->task);
2702 local_irq_restore(flags);
2706 * Cross CPU call to read the hardware event
2708 static void __perf_event_read(void *info)
2710 struct perf_event *event = info;
2711 struct perf_event_context *ctx = event->ctx;
2712 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
2715 * If this is a task context, we need to check whether it is
2716 * the current task context of this cpu. If not it has been
2717 * scheduled out before the smp call arrived. In that case
2718 * event->count would have been updated to a recent sample
2719 * when the event was scheduled out.
2721 if (ctx->task && cpuctx->task_ctx != ctx)
2724 raw_spin_lock(&ctx->lock);
2725 if (ctx->is_active) {
2726 update_context_time(ctx);
2727 update_cgrp_time_from_event(event);
2729 update_event_times(event);
2730 if (event->state == PERF_EVENT_STATE_ACTIVE)
2731 event->pmu->read(event);
2732 raw_spin_unlock(&ctx->lock);
2735 static inline u64 perf_event_count(struct perf_event *event)
2737 return local64_read(&event->count) + atomic64_read(&event->child_count);
2740 static u64 perf_event_read(struct perf_event *event)
2743 * If event is enabled and currently active on a CPU, update the
2744 * value in the event structure:
2746 if (event->state == PERF_EVENT_STATE_ACTIVE) {
2747 smp_call_function_single(event->oncpu,
2748 __perf_event_read, event, 1);
2749 } else if (event->state == PERF_EVENT_STATE_INACTIVE) {
2750 struct perf_event_context *ctx = event->ctx;
2751 unsigned long flags;
2753 raw_spin_lock_irqsave(&ctx->lock, flags);
2755 * may read while context is not active
2756 * (e.g., thread is blocked), in that case
2757 * we cannot update context time
2759 if (ctx->is_active) {
2760 update_context_time(ctx);
2761 update_cgrp_time_from_event(event);
2763 update_event_times(event);
2764 raw_spin_unlock_irqrestore(&ctx->lock, flags);
2767 return perf_event_count(event);
2771 * Initialize the perf_event context in a task_struct:
2773 static void __perf_event_init_context(struct perf_event_context *ctx)
2775 raw_spin_lock_init(&ctx->lock);
2776 mutex_init(&ctx->mutex);
2777 INIT_LIST_HEAD(&ctx->pinned_groups);
2778 INIT_LIST_HEAD(&ctx->flexible_groups);
2779 INIT_LIST_HEAD(&ctx->event_list);
2780 atomic_set(&ctx->refcount, 1);
2783 static struct perf_event_context *
2784 alloc_perf_context(struct pmu *pmu, struct task_struct *task)
2786 struct perf_event_context *ctx;
2788 ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL);
2792 __perf_event_init_context(ctx);
2795 get_task_struct(task);
2802 static struct task_struct *
2803 find_lively_task_by_vpid(pid_t vpid)
2805 struct task_struct *task;
2812 task = find_task_by_vpid(vpid);
2814 get_task_struct(task);
2818 return ERR_PTR(-ESRCH);
2820 /* Reuse ptrace permission checks for now. */
2822 if (!ptrace_may_access(task, PTRACE_MODE_READ))
2827 put_task_struct(task);
2828 return ERR_PTR(err);
2833 * Returns a matching context with refcount and pincount.
2835 static struct perf_event_context *
2836 find_get_context(struct pmu *pmu, struct task_struct *task, int cpu)
2838 struct perf_event_context *ctx;
2839 struct perf_cpu_context *cpuctx;
2840 unsigned long flags;
2844 /* Must be root to operate on a CPU event: */
2845 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
2846 return ERR_PTR(-EACCES);
2849 * We could be clever and allow to attach a event to an
2850 * offline CPU and activate it when the CPU comes up, but
2853 if (!cpu_online(cpu))
2854 return ERR_PTR(-ENODEV);
2856 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
2865 ctxn = pmu->task_ctx_nr;
2870 ctx = perf_lock_task_context(task, ctxn, &flags);
2874 raw_spin_unlock_irqrestore(&ctx->lock, flags);
2876 ctx = alloc_perf_context(pmu, task);
2882 mutex_lock(&task->perf_event_mutex);
2884 * If it has already passed perf_event_exit_task().
2885 * we must see PF_EXITING, it takes this mutex too.
2887 if (task->flags & PF_EXITING)
2889 else if (task->perf_event_ctxp[ctxn])
2894 rcu_assign_pointer(task->perf_event_ctxp[ctxn], ctx);
2896 mutex_unlock(&task->perf_event_mutex);
2898 if (unlikely(err)) {
2910 return ERR_PTR(err);
2913 static void perf_event_free_filter(struct perf_event *event);
2915 static void free_event_rcu(struct rcu_head *head)
2917 struct perf_event *event;
2919 event = container_of(head, struct perf_event, rcu_head);
2921 put_pid_ns(event->ns);
2922 perf_event_free_filter(event);
2926 static void ring_buffer_put(struct ring_buffer *rb);
2927 static void ring_buffer_detach(struct perf_event *event, struct ring_buffer *rb);
2929 static void free_event(struct perf_event *event)
2931 irq_work_sync(&event->pending);
2933 if (!event->parent) {
2934 if (event->attach_state & PERF_ATTACH_TASK)
2935 static_key_slow_dec_deferred(&perf_sched_events);
2936 if (event->attr.mmap || event->attr.mmap_data)
2937 atomic_dec(&nr_mmap_events);
2938 if (event->attr.comm)
2939 atomic_dec(&nr_comm_events);
2940 if (event->attr.task)
2941 atomic_dec(&nr_task_events);
2942 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN)
2943 put_callchain_buffers();
2944 if (is_cgroup_event(event)) {
2945 atomic_dec(&per_cpu(perf_cgroup_events, event->cpu));
2946 static_key_slow_dec_deferred(&perf_sched_events);
2949 if (has_branch_stack(event)) {
2950 static_key_slow_dec_deferred(&perf_sched_events);
2951 /* is system-wide event */
2952 if (!(event->attach_state & PERF_ATTACH_TASK)) {
2953 atomic_dec(&per_cpu(perf_branch_stack_events,
2960 struct ring_buffer *rb;
2963 * Can happen when we close an event with re-directed output.
2965 * Since we have a 0 refcount, perf_mmap_close() will skip
2966 * over us; possibly making our ring_buffer_put() the last.
2968 mutex_lock(&event->mmap_mutex);
2971 rcu_assign_pointer(event->rb, NULL);
2972 ring_buffer_detach(event, rb);
2973 ring_buffer_put(rb); /* could be last */
2975 mutex_unlock(&event->mmap_mutex);
2978 if (is_cgroup_event(event))
2979 perf_detach_cgroup(event);
2982 event->destroy(event);
2985 put_ctx(event->ctx);
2987 call_rcu(&event->rcu_head, free_event_rcu);
2990 int perf_event_release_kernel(struct perf_event *event)
2992 struct perf_event_context *ctx = event->ctx;
2994 WARN_ON_ONCE(ctx->parent_ctx);
2996 * There are two ways this annotation is useful:
2998 * 1) there is a lock recursion from perf_event_exit_task
2999 * see the comment there.
3001 * 2) there is a lock-inversion with mmap_sem through
3002 * perf_event_read_group(), which takes faults while
3003 * holding ctx->mutex, however this is called after
3004 * the last filedesc died, so there is no possibility
3005 * to trigger the AB-BA case.
3007 mutex_lock_nested(&ctx->mutex, SINGLE_DEPTH_NESTING);
3008 raw_spin_lock_irq(&ctx->lock);
3009 perf_group_detach(event);
3010 raw_spin_unlock_irq(&ctx->lock);
3011 perf_remove_from_context(event);
3012 mutex_unlock(&ctx->mutex);
3018 EXPORT_SYMBOL_GPL(perf_event_release_kernel);
3021 * Called when the last reference to the file is gone.
3023 static void put_event(struct perf_event *event)
3025 struct task_struct *owner;
3027 if (!atomic_long_dec_and_test(&event->refcount))
3031 owner = ACCESS_ONCE(event->owner);
3033 * Matches the smp_wmb() in perf_event_exit_task(). If we observe
3034 * !owner it means the list deletion is complete and we can indeed
3035 * free this event, otherwise we need to serialize on
3036 * owner->perf_event_mutex.
3038 smp_read_barrier_depends();
3041 * Since delayed_put_task_struct() also drops the last
3042 * task reference we can safely take a new reference
3043 * while holding the rcu_read_lock().
3045 get_task_struct(owner);
3050 mutex_lock(&owner->perf_event_mutex);
3052 * We have to re-check the event->owner field, if it is cleared
3053 * we raced with perf_event_exit_task(), acquiring the mutex
3054 * ensured they're done, and we can proceed with freeing the
3058 list_del_init(&event->owner_entry);
3059 mutex_unlock(&owner->perf_event_mutex);
3060 put_task_struct(owner);
3063 perf_event_release_kernel(event);
3066 static int perf_release(struct inode *inode, struct file *file)
3068 put_event(file->private_data);
3072 u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running)
3074 struct perf_event *child;
3080 mutex_lock(&event->child_mutex);
3081 total += perf_event_read(event);
3082 *enabled += event->total_time_enabled +
3083 atomic64_read(&event->child_total_time_enabled);
3084 *running += event->total_time_running +
3085 atomic64_read(&event->child_total_time_running);
3087 list_for_each_entry(child, &event->child_list, child_list) {
3088 total += perf_event_read(child);
3089 *enabled += child->total_time_enabled;
3090 *running += child->total_time_running;
3092 mutex_unlock(&event->child_mutex);
3096 EXPORT_SYMBOL_GPL(perf_event_read_value);
3098 static int perf_event_read_group(struct perf_event *event,
3099 u64 read_format, char __user *buf)
3101 struct perf_event *leader = event->group_leader, *sub;
3102 int n = 0, size = 0, ret = -EFAULT;
3103 struct perf_event_context *ctx = leader->ctx;
3105 u64 count, enabled, running;
3107 mutex_lock(&ctx->mutex);
3108 count = perf_event_read_value(leader, &enabled, &running);
3110 values[n++] = 1 + leader->nr_siblings;
3111 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3112 values[n++] = enabled;
3113 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3114 values[n++] = running;
3115 values[n++] = count;
3116 if (read_format & PERF_FORMAT_ID)
3117 values[n++] = primary_event_id(leader);
3119 size = n * sizeof(u64);
3121 if (copy_to_user(buf, values, size))
3126 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
3129 values[n++] = perf_event_read_value(sub, &enabled, &running);
3130 if (read_format & PERF_FORMAT_ID)
3131 values[n++] = primary_event_id(sub);
3133 size = n * sizeof(u64);
3135 if (copy_to_user(buf + ret, values, size)) {
3143 mutex_unlock(&ctx->mutex);
3148 static int perf_event_read_one(struct perf_event *event,
3149 u64 read_format, char __user *buf)
3151 u64 enabled, running;
3155 values[n++] = perf_event_read_value(event, &enabled, &running);
3156 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3157 values[n++] = enabled;
3158 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3159 values[n++] = running;
3160 if (read_format & PERF_FORMAT_ID)
3161 values[n++] = primary_event_id(event);
3163 if (copy_to_user(buf, values, n * sizeof(u64)))
3166 return n * sizeof(u64);
3170 * Read the performance event - simple non blocking version for now
3173 perf_read_hw(struct perf_event *event, char __user *buf, size_t count)
3175 u64 read_format = event->attr.read_format;
3179 * Return end-of-file for a read on a event that is in
3180 * error state (i.e. because it was pinned but it couldn't be
3181 * scheduled on to the CPU at some point).
3183 if (event->state == PERF_EVENT_STATE_ERROR)
3186 if (count < event->read_size)
3189 WARN_ON_ONCE(event->ctx->parent_ctx);
3190 if (read_format & PERF_FORMAT_GROUP)
3191 ret = perf_event_read_group(event, read_format, buf);
3193 ret = perf_event_read_one(event, read_format, buf);
3199 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
3201 struct perf_event *event = file->private_data;
3203 return perf_read_hw(event, buf, count);
3206 static unsigned int perf_poll(struct file *file, poll_table *wait)
3208 struct perf_event *event = file->private_data;
3209 struct ring_buffer *rb;
3210 unsigned int events = POLL_HUP;
3213 * Pin the event->rb by taking event->mmap_mutex; otherwise
3214 * perf_event_set_output() can swizzle our rb and make us miss wakeups.
3216 mutex_lock(&event->mmap_mutex);
3219 events = atomic_xchg(&rb->poll, 0);
3220 mutex_unlock(&event->mmap_mutex);
3222 poll_wait(file, &event->waitq, wait);
3227 static void perf_event_reset(struct perf_event *event)
3229 (void)perf_event_read(event);
3230 local64_set(&event->count, 0);
3231 perf_event_update_userpage(event);
3235 * Holding the top-level event's child_mutex means that any
3236 * descendant process that has inherited this event will block
3237 * in sync_child_event if it goes to exit, thus satisfying the
3238 * task existence requirements of perf_event_enable/disable.
3240 static void perf_event_for_each_child(struct perf_event *event,
3241 void (*func)(struct perf_event *))
3243 struct perf_event *child;
3245 WARN_ON_ONCE(event->ctx->parent_ctx);
3246 mutex_lock(&event->child_mutex);
3248 list_for_each_entry(child, &event->child_list, child_list)
3250 mutex_unlock(&event->child_mutex);
3253 static void perf_event_for_each(struct perf_event *event,
3254 void (*func)(struct perf_event *))
3256 struct perf_event_context *ctx = event->ctx;
3257 struct perf_event *sibling;
3259 WARN_ON_ONCE(ctx->parent_ctx);
3260 mutex_lock(&ctx->mutex);
3261 event = event->group_leader;
3263 perf_event_for_each_child(event, func);
3264 list_for_each_entry(sibling, &event->sibling_list, group_entry)
3265 perf_event_for_each_child(sibling, func);
3266 mutex_unlock(&ctx->mutex);
3269 static int perf_event_period(struct perf_event *event, u64 __user *arg)
3271 struct perf_event_context *ctx = event->ctx;
3275 if (!is_sampling_event(event))
3278 if (copy_from_user(&value, arg, sizeof(value)))
3284 raw_spin_lock_irq(&ctx->lock);
3285 if (event->attr.freq) {
3286 if (value > sysctl_perf_event_sample_rate) {
3291 event->attr.sample_freq = value;
3293 event->attr.sample_period = value;
3294 event->hw.sample_period = value;
3297 raw_spin_unlock_irq(&ctx->lock);
3302 static const struct file_operations perf_fops;
3304 static inline int perf_fget_light(int fd, struct fd *p)
3306 struct fd f = fdget(fd);
3310 if (f.file->f_op != &perf_fops) {
3318 static int perf_event_set_output(struct perf_event *event,
3319 struct perf_event *output_event);
3320 static int perf_event_set_filter(struct perf_event *event, void __user *arg);
3322 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
3324 struct perf_event *event = file->private_data;
3325 void (*func)(struct perf_event *);
3329 case PERF_EVENT_IOC_ENABLE:
3330 func = perf_event_enable;
3332 case PERF_EVENT_IOC_DISABLE:
3333 func = perf_event_disable;
3335 case PERF_EVENT_IOC_RESET:
3336 func = perf_event_reset;
3339 case PERF_EVENT_IOC_REFRESH:
3340 return perf_event_refresh(event, arg);
3342 case PERF_EVENT_IOC_PERIOD:
3343 return perf_event_period(event, (u64 __user *)arg);
3345 case PERF_EVENT_IOC_SET_OUTPUT:
3349 struct perf_event *output_event;
3351 ret = perf_fget_light(arg, &output);
3354 output_event = output.file->private_data;
3355 ret = perf_event_set_output(event, output_event);
3358 ret = perf_event_set_output(event, NULL);
3363 case PERF_EVENT_IOC_SET_FILTER:
3364 return perf_event_set_filter(event, (void __user *)arg);
3370 if (flags & PERF_IOC_FLAG_GROUP)
3371 perf_event_for_each(event, func);
3373 perf_event_for_each_child(event, func);
3378 int perf_event_task_enable(void)
3380 struct perf_event *event;
3382 mutex_lock(¤t->perf_event_mutex);
3383 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
3384 perf_event_for_each_child(event, perf_event_enable);
3385 mutex_unlock(¤t->perf_event_mutex);
3390 int perf_event_task_disable(void)
3392 struct perf_event *event;
3394 mutex_lock(¤t->perf_event_mutex);
3395 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
3396 perf_event_for_each_child(event, perf_event_disable);
3397 mutex_unlock(¤t->perf_event_mutex);
3402 static int perf_event_index(struct perf_event *event)
3404 if (event->hw.state & PERF_HES_STOPPED)
3407 if (event->state != PERF_EVENT_STATE_ACTIVE)
3410 return event->pmu->event_idx(event);
3413 static void calc_timer_values(struct perf_event *event,
3420 *now = perf_clock();
3421 ctx_time = event->shadow_ctx_time + *now;
3422 *enabled = ctx_time - event->tstamp_enabled;
3423 *running = ctx_time - event->tstamp_running;
3426 void __weak arch_perf_update_userpage(struct perf_event_mmap_page *userpg, u64 now)
3431 * Callers need to ensure there can be no nesting of this function, otherwise
3432 * the seqlock logic goes bad. We can not serialize this because the arch
3433 * code calls this from NMI context.
3435 void perf_event_update_userpage(struct perf_event *event)
3437 struct perf_event_mmap_page *userpg;
3438 struct ring_buffer *rb;
3439 u64 enabled, running, now;
3443 * compute total_time_enabled, total_time_running
3444 * based on snapshot values taken when the event
3445 * was last scheduled in.
3447 * we cannot simply called update_context_time()
3448 * because of locking issue as we can be called in
3451 calc_timer_values(event, &now, &enabled, &running);
3452 rb = rcu_dereference(event->rb);
3456 userpg = rb->user_page;
3459 * Disable preemption so as to not let the corresponding user-space
3460 * spin too long if we get preempted.
3465 userpg->index = perf_event_index(event);
3466 userpg->offset = perf_event_count(event);
3468 userpg->offset -= local64_read(&event->hw.prev_count);
3470 userpg->time_enabled = enabled +
3471 atomic64_read(&event->child_total_time_enabled);
3473 userpg->time_running = running +
3474 atomic64_read(&event->child_total_time_running);
3476 arch_perf_update_userpage(userpg, now);
3485 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
3487 struct perf_event *event = vma->vm_file->private_data;
3488 struct ring_buffer *rb;
3489 int ret = VM_FAULT_SIGBUS;
3491 if (vmf->flags & FAULT_FLAG_MKWRITE) {
3492 if (vmf->pgoff == 0)
3498 rb = rcu_dereference(event->rb);
3502 if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE))
3505 vmf->page = perf_mmap_to_page(rb, vmf->pgoff);
3509 get_page(vmf->page);
3510 vmf->page->mapping = vma->vm_file->f_mapping;
3511 vmf->page->index = vmf->pgoff;
3520 static void ring_buffer_attach(struct perf_event *event,
3521 struct ring_buffer *rb)
3523 unsigned long flags;
3525 if (!list_empty(&event->rb_entry))
3528 spin_lock_irqsave(&rb->event_lock, flags);
3529 if (list_empty(&event->rb_entry))
3530 list_add(&event->rb_entry, &rb->event_list);
3531 spin_unlock_irqrestore(&rb->event_lock, flags);
3534 static void ring_buffer_detach(struct perf_event *event, struct ring_buffer *rb)
3536 unsigned long flags;
3538 if (list_empty(&event->rb_entry))
3541 spin_lock_irqsave(&rb->event_lock, flags);
3542 list_del_init(&event->rb_entry);
3543 wake_up_all(&event->waitq);
3544 spin_unlock_irqrestore(&rb->event_lock, flags);
3547 static void ring_buffer_wakeup(struct perf_event *event)
3549 struct ring_buffer *rb;
3552 rb = rcu_dereference(event->rb);
3554 list_for_each_entry_rcu(event, &rb->event_list, rb_entry)
3555 wake_up_all(&event->waitq);
3560 static void rb_free_rcu(struct rcu_head *rcu_head)
3562 struct ring_buffer *rb;
3564 rb = container_of(rcu_head, struct ring_buffer, rcu_head);
3568 static struct ring_buffer *ring_buffer_get(struct perf_event *event)
3570 struct ring_buffer *rb;
3573 rb = rcu_dereference(event->rb);
3575 if (!atomic_inc_not_zero(&rb->refcount))
3583 static void ring_buffer_put(struct ring_buffer *rb)
3585 if (!atomic_dec_and_test(&rb->refcount))
3588 WARN_ON_ONCE(!list_empty(&rb->event_list));
3590 call_rcu(&rb->rcu_head, rb_free_rcu);
3593 static void perf_mmap_open(struct vm_area_struct *vma)
3595 struct perf_event *event = vma->vm_file->private_data;
3597 atomic_inc(&event->mmap_count);
3598 atomic_inc(&event->rb->mmap_count);
3602 * A buffer can be mmap()ed multiple times; either directly through the same
3603 * event, or through other events by use of perf_event_set_output().
3605 * In order to undo the VM accounting done by perf_mmap() we need to destroy
3606 * the buffer here, where we still have a VM context. This means we need
3607 * to detach all events redirecting to us.
3609 static void perf_mmap_close(struct vm_area_struct *vma)
3611 struct perf_event *event = vma->vm_file->private_data;
3613 struct ring_buffer *rb = event->rb;
3614 struct user_struct *mmap_user = rb->mmap_user;
3615 int mmap_locked = rb->mmap_locked;
3616 unsigned long size = perf_data_size(rb);
3618 atomic_dec(&rb->mmap_count);
3620 if (!atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex))
3623 /* Detach current event from the buffer. */
3624 rcu_assign_pointer(event->rb, NULL);
3625 ring_buffer_detach(event, rb);
3626 mutex_unlock(&event->mmap_mutex);
3628 /* If there's still other mmap()s of this buffer, we're done. */
3629 if (atomic_read(&rb->mmap_count)) {
3630 ring_buffer_put(rb); /* can't be last */
3635 * No other mmap()s, detach from all other events that might redirect
3636 * into the now unreachable buffer. Somewhat complicated by the
3637 * fact that rb::event_lock otherwise nests inside mmap_mutex.
3641 list_for_each_entry_rcu(event, &rb->event_list, rb_entry) {
3642 if (!atomic_long_inc_not_zero(&event->refcount)) {
3644 * This event is en-route to free_event() which will
3645 * detach it and remove it from the list.
3651 mutex_lock(&event->mmap_mutex);
3653 * Check we didn't race with perf_event_set_output() which can
3654 * swizzle the rb from under us while we were waiting to
3655 * acquire mmap_mutex.
3657 * If we find a different rb; ignore this event, a next
3658 * iteration will no longer find it on the list. We have to
3659 * still restart the iteration to make sure we're not now
3660 * iterating the wrong list.
3662 if (event->rb == rb) {
3663 rcu_assign_pointer(event->rb, NULL);
3664 ring_buffer_detach(event, rb);
3665 ring_buffer_put(rb); /* can't be last, we still have one */
3667 mutex_unlock(&event->mmap_mutex);
3671 * Restart the iteration; either we're on the wrong list or
3672 * destroyed its integrity by doing a deletion.
3679 * It could be there's still a few 0-ref events on the list; they'll
3680 * get cleaned up by free_event() -- they'll also still have their
3681 * ref on the rb and will free it whenever they are done with it.
3683 * Aside from that, this buffer is 'fully' detached and unmapped,
3684 * undo the VM accounting.
3687 atomic_long_sub((size >> PAGE_SHIFT) + 1, &mmap_user->locked_vm);
3688 vma->vm_mm->pinned_vm -= mmap_locked;
3689 free_uid(mmap_user);
3691 ring_buffer_put(rb); /* could be last */
3694 static const struct vm_operations_struct perf_mmap_vmops = {
3695 .open = perf_mmap_open,
3696 .close = perf_mmap_close,
3697 .fault = perf_mmap_fault,
3698 .page_mkwrite = perf_mmap_fault,
3701 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
3703 struct perf_event *event = file->private_data;
3704 unsigned long user_locked, user_lock_limit;
3705 struct user_struct *user = current_user();
3706 unsigned long locked, lock_limit;
3707 struct ring_buffer *rb;
3708 unsigned long vma_size;
3709 unsigned long nr_pages;
3710 long user_extra, extra;
3711 int ret = 0, flags = 0;
3714 * Don't allow mmap() of inherited per-task counters. This would
3715 * create a performance issue due to all children writing to the
3718 if (event->cpu == -1 && event->attr.inherit)
3721 if (!(vma->vm_flags & VM_SHARED))
3724 vma_size = vma->vm_end - vma->vm_start;
3725 nr_pages = (vma_size / PAGE_SIZE) - 1;
3728 * If we have rb pages ensure they're a power-of-two number, so we
3729 * can do bitmasks instead of modulo.
3731 if (nr_pages != 0 && !is_power_of_2(nr_pages))
3734 if (vma_size != PAGE_SIZE * (1 + nr_pages))
3737 if (vma->vm_pgoff != 0)
3740 WARN_ON_ONCE(event->ctx->parent_ctx);
3742 mutex_lock(&event->mmap_mutex);
3744 if (event->rb->nr_pages != nr_pages) {
3749 if (!atomic_inc_not_zero(&event->rb->mmap_count)) {
3751 * Raced against perf_mmap_close() through
3752 * perf_event_set_output(). Try again, hope for better
3755 mutex_unlock(&event->mmap_mutex);
3762 user_extra = nr_pages + 1;
3763 user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10);
3766 * Increase the limit linearly with more CPUs:
3768 user_lock_limit *= num_online_cpus();
3770 user_locked = atomic_long_read(&user->locked_vm) + user_extra;
3773 if (user_locked > user_lock_limit)
3774 extra = user_locked - user_lock_limit;
3776 lock_limit = rlimit(RLIMIT_MEMLOCK);
3777 lock_limit >>= PAGE_SHIFT;
3778 locked = vma->vm_mm->pinned_vm + extra;
3780 if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() &&
3781 !capable(CAP_IPC_LOCK)) {
3788 if (vma->vm_flags & VM_WRITE)
3789 flags |= RING_BUFFER_WRITABLE;
3791 rb = rb_alloc(nr_pages,
3792 event->attr.watermark ? event->attr.wakeup_watermark : 0,
3800 atomic_set(&rb->mmap_count, 1);
3801 rb->mmap_locked = extra;
3802 rb->mmap_user = get_current_user();
3804 atomic_long_add(user_extra, &user->locked_vm);
3805 vma->vm_mm->pinned_vm += extra;
3807 ring_buffer_attach(event, rb);
3808 rcu_assign_pointer(event->rb, rb);
3810 perf_event_update_userpage(event);
3814 atomic_inc(&event->mmap_count);
3815 mutex_unlock(&event->mmap_mutex);
3818 * Since pinned accounting is per vm we cannot allow fork() to copy our
3821 vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND | VM_DONTDUMP;
3822 vma->vm_ops = &perf_mmap_vmops;
3827 static int perf_fasync(int fd, struct file *filp, int on)
3829 struct inode *inode = file_inode(filp);
3830 struct perf_event *event = filp->private_data;
3833 mutex_lock(&inode->i_mutex);
3834 retval = fasync_helper(fd, filp, on, &event->fasync);
3835 mutex_unlock(&inode->i_mutex);
3843 static const struct file_operations perf_fops = {
3844 .llseek = no_llseek,
3845 .release = perf_release,
3848 .unlocked_ioctl = perf_ioctl,
3849 .compat_ioctl = perf_ioctl,
3851 .fasync = perf_fasync,
3857 * If there's data, ensure we set the poll() state and publish everything
3858 * to user-space before waking everybody up.
3861 void perf_event_wakeup(struct perf_event *event)
3863 ring_buffer_wakeup(event);
3865 if (event->pending_kill) {
3866 kill_fasync(&event->fasync, SIGIO, event->pending_kill);
3867 event->pending_kill = 0;
3871 static void perf_pending_event(struct irq_work *entry)
3873 struct perf_event *event = container_of(entry,
3874 struct perf_event, pending);
3876 if (event->pending_disable) {
3877 event->pending_disable = 0;
3878 __perf_event_disable(event);
3881 if (event->pending_wakeup) {
3882 event->pending_wakeup = 0;
3883 perf_event_wakeup(event);
3888 * We assume there is only KVM supporting the callbacks.
3889 * Later on, we might change it to a list if there is
3890 * another virtualization implementation supporting the callbacks.
3892 struct perf_guest_info_callbacks *perf_guest_cbs;
3894 int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
3896 perf_guest_cbs = cbs;
3899 EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks);
3901 int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
3903 perf_guest_cbs = NULL;
3906 EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks);
3909 perf_output_sample_regs(struct perf_output_handle *handle,
3910 struct pt_regs *regs, u64 mask)
3914 for_each_set_bit(bit, (const unsigned long *) &mask,
3915 sizeof(mask) * BITS_PER_BYTE) {
3918 val = perf_reg_value(regs, bit);
3919 perf_output_put(handle, val);
3923 static void perf_sample_regs_user(struct perf_regs_user *regs_user,
3924 struct pt_regs *regs)
3926 if (!user_mode(regs)) {
3928 regs = task_pt_regs(current);
3934 regs_user->regs = regs;
3935 regs_user->abi = perf_reg_abi(current);
3940 * Get remaining task size from user stack pointer.
3942 * It'd be better to take stack vma map and limit this more
3943 * precisly, but there's no way to get it safely under interrupt,
3944 * so using TASK_SIZE as limit.
3946 static u64 perf_ustack_task_size(struct pt_regs *regs)
3948 unsigned long addr = perf_user_stack_pointer(regs);
3950 if (!addr || addr >= TASK_SIZE)
3953 return TASK_SIZE - addr;
3957 perf_sample_ustack_size(u16 stack_size, u16 header_size,
3958 struct pt_regs *regs)
3962 /* No regs, no stack pointer, no dump. */
3967 * Check if we fit in with the requested stack size into the:
3969 * If we don't, we limit the size to the TASK_SIZE.
3971 * - remaining sample size
3972 * If we don't, we customize the stack size to
3973 * fit in to the remaining sample size.
3976 task_size = min((u64) USHRT_MAX, perf_ustack_task_size(regs));
3977 stack_size = min(stack_size, (u16) task_size);
3979 /* Current header size plus static size and dynamic size. */
3980 header_size += 2 * sizeof(u64);
3982 /* Do we fit in with the current stack dump size? */
3983 if ((u16) (header_size + stack_size) < header_size) {
3985 * If we overflow the maximum size for the sample,
3986 * we customize the stack dump size to fit in.
3988 stack_size = USHRT_MAX - header_size - sizeof(u64);
3989 stack_size = round_up(stack_size, sizeof(u64));
3996 perf_output_sample_ustack(struct perf_output_handle *handle, u64 dump_size,
3997 struct pt_regs *regs)
3999 /* Case of a kernel thread, nothing to dump */
4002 perf_output_put(handle, size);
4011 * - the size requested by user or the best one we can fit
4012 * in to the sample max size
4014 * - user stack dump data
4016 * - the actual dumped size
4020 perf_output_put(handle, dump_size);
4023 sp = perf_user_stack_pointer(regs);
4024 rem = __output_copy_user(handle, (void *) sp, dump_size);
4025 dyn_size = dump_size - rem;
4027 perf_output_skip(handle, rem);
4030 perf_output_put(handle, dyn_size);
4034 static void __perf_event_header__init_id(struct perf_event_header *header,
4035 struct perf_sample_data *data,
4036 struct perf_event *event)
4038 u64 sample_type = event->attr.sample_type;
4040 data->type = sample_type;
4041 header->size += event->id_header_size;
4043 if (sample_type & PERF_SAMPLE_TID) {
4044 /* namespace issues */
4045 data->tid_entry.pid = perf_event_pid(event, current);
4046 data->tid_entry.tid = perf_event_tid(event, current);
4049 if (sample_type & PERF_SAMPLE_TIME)
4050 data->time = perf_clock();
4052 if (sample_type & PERF_SAMPLE_ID)
4053 data->id = primary_event_id(event);
4055 if (sample_type & PERF_SAMPLE_STREAM_ID)
4056 data->stream_id = event->id;
4058 if (sample_type & PERF_SAMPLE_CPU) {
4059 data->cpu_entry.cpu = raw_smp_processor_id();
4060 data->cpu_entry.reserved = 0;
4064 void perf_event_header__init_id(struct perf_event_header *header,
4065 struct perf_sample_data *data,
4066 struct perf_event *event)
4068 if (event->attr.sample_id_all)
4069 __perf_event_header__init_id(header, data, event);
4072 static void __perf_event__output_id_sample(struct perf_output_handle *handle,
4073 struct perf_sample_data *data)
4075 u64 sample_type = data->type;
4077 if (sample_type & PERF_SAMPLE_TID)
4078 perf_output_put(handle, data->tid_entry);
4080 if (sample_type & PERF_SAMPLE_TIME)
4081 perf_output_put(handle, data->time);
4083 if (sample_type & PERF_SAMPLE_ID)
4084 perf_output_put(handle, data->id);
4086 if (sample_type & PERF_SAMPLE_STREAM_ID)
4087 perf_output_put(handle, data->stream_id);
4089 if (sample_type & PERF_SAMPLE_CPU)
4090 perf_output_put(handle, data->cpu_entry);
4093 void perf_event__output_id_sample(struct perf_event *event,
4094 struct perf_output_handle *handle,
4095 struct perf_sample_data *sample)
4097 if (event->attr.sample_id_all)
4098 __perf_event__output_id_sample(handle, sample);
4101 static void perf_output_read_one(struct perf_output_handle *handle,
4102 struct perf_event *event,
4103 u64 enabled, u64 running)
4105 u64 read_format = event->attr.read_format;
4109 values[n++] = perf_event_count(event);
4110 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
4111 values[n++] = enabled +
4112 atomic64_read(&event->child_total_time_enabled);
4114 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
4115 values[n++] = running +
4116 atomic64_read(&event->child_total_time_running);
4118 if (read_format & PERF_FORMAT_ID)
4119 values[n++] = primary_event_id(event);
4121 __output_copy(handle, values, n * sizeof(u64));
4125 * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
4127 static void perf_output_read_group(struct perf_output_handle *handle,
4128 struct perf_event *event,
4129 u64 enabled, u64 running)
4131 struct perf_event *leader = event->group_leader, *sub;
4132 u64 read_format = event->attr.read_format;
4136 values[n++] = 1 + leader->nr_siblings;
4138 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
4139 values[n++] = enabled;
4141 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
4142 values[n++] = running;
4144 if (leader != event)
4145 leader->pmu->read(leader);
4147 values[n++] = perf_event_count(leader);
4148 if (read_format & PERF_FORMAT_ID)
4149 values[n++] = primary_event_id(leader);
4151 __output_copy(handle, values, n * sizeof(u64));
4153 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
4157 sub->pmu->read(sub);
4159 values[n++] = perf_event_count(sub);
4160 if (read_format & PERF_FORMAT_ID)
4161 values[n++] = primary_event_id(sub);
4163 __output_copy(handle, values, n * sizeof(u64));
4167 #define PERF_FORMAT_TOTAL_TIMES (PERF_FORMAT_TOTAL_TIME_ENABLED|\
4168 PERF_FORMAT_TOTAL_TIME_RUNNING)
4170 static void perf_output_read(struct perf_output_handle *handle,
4171 struct perf_event *event)
4173 u64 enabled = 0, running = 0, now;
4174 u64 read_format = event->attr.read_format;
4177 * compute total_time_enabled, total_time_running
4178 * based on snapshot values taken when the event
4179 * was last scheduled in.
4181 * we cannot simply called update_context_time()
4182 * because of locking issue as we are called in
4185 if (read_format & PERF_FORMAT_TOTAL_TIMES)
4186 calc_timer_values(event, &now, &enabled, &running);
4188 if (event->attr.read_format & PERF_FORMAT_GROUP)
4189 perf_output_read_group(handle, event, enabled, running);
4191 perf_output_read_one(handle, event, enabled, running);
4194 void perf_output_sample(struct perf_output_handle *handle,
4195 struct perf_event_header *header,
4196 struct perf_sample_data *data,
4197 struct perf_event *event)
4199 u64 sample_type = data->type;
4201 perf_output_put(handle, *header);
4203 if (sample_type & PERF_SAMPLE_IP)
4204 perf_output_put(handle, data->ip);
4206 if (sample_type & PERF_SAMPLE_TID)
4207 perf_output_put(handle, data->tid_entry);
4209 if (sample_type & PERF_SAMPLE_TIME)
4210 perf_output_put(handle, data->time);
4212 if (sample_type & PERF_SAMPLE_ADDR)
4213 perf_output_put(handle, data->addr);
4215 if (sample_type & PERF_SAMPLE_ID)
4216 perf_output_put(handle, data->id);
4218 if (sample_type & PERF_SAMPLE_STREAM_ID)
4219 perf_output_put(handle, data->stream_id);
4221 if (sample_type & PERF_SAMPLE_CPU)
4222 perf_output_put(handle, data->cpu_entry);
4224 if (sample_type & PERF_SAMPLE_PERIOD)
4225 perf_output_put(handle, data->period);
4227 if (sample_type & PERF_SAMPLE_READ)
4228 perf_output_read(handle, event);
4230 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
4231 if (data->callchain) {
4234 if (data->callchain)
4235 size += data->callchain->nr;
4237 size *= sizeof(u64);
4239 __output_copy(handle, data->callchain, size);
4242 perf_output_put(handle, nr);
4246 if (sample_type & PERF_SAMPLE_RAW) {
4248 perf_output_put(handle, data->raw->size);
4249 __output_copy(handle, data->raw->data,
4256 .size = sizeof(u32),
4259 perf_output_put(handle, raw);
4263 if (!event->attr.watermark) {
4264 int wakeup_events = event->attr.wakeup_events;
4266 if (wakeup_events) {
4267 struct ring_buffer *rb = handle->rb;
4268 int events = local_inc_return(&rb->events);
4270 if (events >= wakeup_events) {
4271 local_sub(wakeup_events, &rb->events);
4272 local_inc(&rb->wakeup);
4277 if (sample_type & PERF_SAMPLE_BRANCH_STACK) {
4278 if (data->br_stack) {
4281 size = data->br_stack->nr
4282 * sizeof(struct perf_branch_entry);
4284 perf_output_put(handle, data->br_stack->nr);
4285 perf_output_copy(handle, data->br_stack->entries, size);
4288 * we always store at least the value of nr
4291 perf_output_put(handle, nr);
4295 if (sample_type & PERF_SAMPLE_REGS_USER) {
4296 u64 abi = data->regs_user.abi;
4299 * If there are no regs to dump, notice it through
4300 * first u64 being zero (PERF_SAMPLE_REGS_ABI_NONE).
4302 perf_output_put(handle, abi);
4305 u64 mask = event->attr.sample_regs_user;
4306 perf_output_sample_regs(handle,
4307 data->regs_user.regs,
4312 if (sample_type & PERF_SAMPLE_STACK_USER)
4313 perf_output_sample_ustack(handle,
4314 data->stack_user_size,
4315 data->regs_user.regs);
4317 if (sample_type & PERF_SAMPLE_WEIGHT)
4318 perf_output_put(handle, data->weight);
4320 if (sample_type & PERF_SAMPLE_DATA_SRC)
4321 perf_output_put(handle, data->data_src.val);
4324 void perf_prepare_sample(struct perf_event_header *header,
4325 struct perf_sample_data *data,
4326 struct perf_event *event,
4327 struct pt_regs *regs)
4329 u64 sample_type = event->attr.sample_type;
4331 header->type = PERF_RECORD_SAMPLE;
4332 header->size = sizeof(*header) + event->header_size;
4335 header->misc |= perf_misc_flags(regs);
4337 __perf_event_header__init_id(header, data, event);
4339 if (sample_type & PERF_SAMPLE_IP)
4340 data->ip = perf_instruction_pointer(regs);
4342 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
4345 data->callchain = perf_callchain(event, regs);
4347 if (data->callchain)
4348 size += data->callchain->nr;
4350 header->size += size * sizeof(u64);
4353 if (sample_type & PERF_SAMPLE_RAW) {
4354 int size = sizeof(u32);
4357 size += data->raw->size;
4359 size += sizeof(u32);
4361 WARN_ON_ONCE(size & (sizeof(u64)-1));
4362 header->size += size;
4365 if (sample_type & PERF_SAMPLE_BRANCH_STACK) {
4366 int size = sizeof(u64); /* nr */
4367 if (data->br_stack) {
4368 size += data->br_stack->nr
4369 * sizeof(struct perf_branch_entry);
4371 header->size += size;
4374 if (sample_type & PERF_SAMPLE_REGS_USER) {
4375 /* regs dump ABI info */
4376 int size = sizeof(u64);
4378 perf_sample_regs_user(&data->regs_user, regs);
4380 if (data->regs_user.regs) {
4381 u64 mask = event->attr.sample_regs_user;
4382 size += hweight64(mask) * sizeof(u64);
4385 header->size += size;
4388 if (sample_type & PERF_SAMPLE_STACK_USER) {
4390 * Either we need PERF_SAMPLE_STACK_USER bit to be allways
4391 * processed as the last one or have additional check added
4392 * in case new sample type is added, because we could eat
4393 * up the rest of the sample size.
4395 struct perf_regs_user *uregs = &data->regs_user;
4396 u16 stack_size = event->attr.sample_stack_user;
4397 u16 size = sizeof(u64);
4400 perf_sample_regs_user(uregs, regs);
4402 stack_size = perf_sample_ustack_size(stack_size, header->size,
4406 * If there is something to dump, add space for the dump
4407 * itself and for the field that tells the dynamic size,
4408 * which is how many have been actually dumped.
4411 size += sizeof(u64) + stack_size;
4413 data->stack_user_size = stack_size;
4414 header->size += size;
4418 static void perf_event_output(struct perf_event *event,
4419 struct perf_sample_data *data,
4420 struct pt_regs *regs)
4422 struct perf_output_handle handle;
4423 struct perf_event_header header;
4425 /* protect the callchain buffers */
4428 perf_prepare_sample(&header, data, event, regs);
4430 if (perf_output_begin(&handle, event, header.size))
4433 perf_output_sample(&handle, &header, data, event);
4435 perf_output_end(&handle);
4445 struct perf_read_event {
4446 struct perf_event_header header;
4453 perf_event_read_event(struct perf_event *event,
4454 struct task_struct *task)
4456 struct perf_output_handle handle;
4457 struct perf_sample_data sample;
4458 struct perf_read_event read_event = {
4460 .type = PERF_RECORD_READ,
4462 .size = sizeof(read_event) + event->read_size,
4464 .pid = perf_event_pid(event, task),
4465 .tid = perf_event_tid(event, task),
4469 perf_event_header__init_id(&read_event.header, &sample, event);
4470 ret = perf_output_begin(&handle, event, read_event.header.size);
4474 perf_output_put(&handle, read_event);
4475 perf_output_read(&handle, event);
4476 perf_event__output_id_sample(event, &handle, &sample);
4478 perf_output_end(&handle);
4481 typedef int (perf_event_aux_match_cb)(struct perf_event *event, void *data);
4482 typedef void (perf_event_aux_output_cb)(struct perf_event *event, void *data);
4485 perf_event_aux_ctx(struct perf_event_context *ctx,
4486 perf_event_aux_match_cb match,
4487 perf_event_aux_output_cb output,
4490 struct perf_event *event;
4492 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
4493 if (event->state < PERF_EVENT_STATE_INACTIVE)
4495 if (!event_filter_match(event))
4497 if (match(event, data))
4498 output(event, data);
4503 perf_event_aux(perf_event_aux_match_cb match,
4504 perf_event_aux_output_cb output,
4506 struct perf_event_context *task_ctx)
4508 struct perf_cpu_context *cpuctx;
4509 struct perf_event_context *ctx;
4514 list_for_each_entry_rcu(pmu, &pmus, entry) {
4515 cpuctx = get_cpu_ptr(pmu->pmu_cpu_context);
4516 if (cpuctx->unique_pmu != pmu)
4518 perf_event_aux_ctx(&cpuctx->ctx, match, output, data);
4521 ctxn = pmu->task_ctx_nr;
4524 ctx = rcu_dereference(current->perf_event_ctxp[ctxn]);
4526 perf_event_aux_ctx(ctx, match, output, data);
4528 put_cpu_ptr(pmu->pmu_cpu_context);
4533 perf_event_aux_ctx(task_ctx, match, output, data);
4540 * task tracking -- fork/exit
4542 * enabled by: attr.comm | attr.mmap | attr.mmap_data | attr.task
4545 struct perf_task_event {
4546 struct task_struct *task;
4547 struct perf_event_context *task_ctx;
4550 struct perf_event_header header;
4560 static void perf_event_task_output(struct perf_event *event,
4563 struct perf_task_event *task_event = data;
4564 struct perf_output_handle handle;
4565 struct perf_sample_data sample;
4566 struct task_struct *task = task_event->task;
4567 int ret, size = task_event->event_id.header.size;
4569 perf_event_header__init_id(&task_event->event_id.header, &sample, event);
4571 ret = perf_output_begin(&handle, event,
4572 task_event->event_id.header.size);
4576 task_event->event_id.pid = perf_event_pid(event, task);
4577 task_event->event_id.ppid = perf_event_pid(event, current);
4579 task_event->event_id.tid = perf_event_tid(event, task);
4580 task_event->event_id.ptid = perf_event_tid(event, current);
4582 perf_output_put(&handle, task_event->event_id);
4584 perf_event__output_id_sample(event, &handle, &sample);
4586 perf_output_end(&handle);
4588 task_event->event_id.header.size = size;
4591 static int perf_event_task_match(struct perf_event *event,
4592 void *data __maybe_unused)
4594 return event->attr.comm || event->attr.mmap ||
4595 event->attr.mmap_data || event->attr.task;
4598 static void perf_event_task(struct task_struct *task,
4599 struct perf_event_context *task_ctx,
4602 struct perf_task_event task_event;
4604 if (!atomic_read(&nr_comm_events) &&
4605 !atomic_read(&nr_mmap_events) &&
4606 !atomic_read(&nr_task_events))
4609 task_event = (struct perf_task_event){
4611 .task_ctx = task_ctx,
4614 .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT,
4616 .size = sizeof(task_event.event_id),
4622 .time = perf_clock(),
4626 perf_event_aux(perf_event_task_match,
4627 perf_event_task_output,
4632 void perf_event_fork(struct task_struct *task)
4634 perf_event_task(task, NULL, 1);
4641 struct perf_comm_event {
4642 struct task_struct *task;
4647 struct perf_event_header header;
4654 static void perf_event_comm_output(struct perf_event *event,
4657 struct perf_comm_event *comm_event = data;
4658 struct perf_output_handle handle;
4659 struct perf_sample_data sample;
4660 int size = comm_event->event_id.header.size;
4663 perf_event_header__init_id(&comm_event->event_id.header, &sample, event);
4664 ret = perf_output_begin(&handle, event,
4665 comm_event->event_id.header.size);
4670 comm_event->event_id.pid = perf_event_pid(event, comm_event->task);
4671 comm_event->event_id.tid = perf_event_tid(event, comm_event->task);
4673 perf_output_put(&handle, comm_event->event_id);
4674 __output_copy(&handle, comm_event->comm,
4675 comm_event->comm_size);
4677 perf_event__output_id_sample(event, &handle, &sample);
4679 perf_output_end(&handle);
4681 comm_event->event_id.header.size = size;
4684 static int perf_event_comm_match(struct perf_event *event,
4685 void *data __maybe_unused)
4687 return event->attr.comm;
4690 static void perf_event_comm_event(struct perf_comm_event *comm_event)
4692 char comm[TASK_COMM_LEN];
4695 memset(comm, 0, sizeof(comm));
4696 strlcpy(comm, comm_event->task->comm, sizeof(comm));
4697 size = ALIGN(strlen(comm)+1, sizeof(u64));
4699 comm_event->comm = comm;
4700 comm_event->comm_size = size;
4702 comm_event->event_id.header.size = sizeof(comm_event->event_id) + size;
4704 perf_event_aux(perf_event_comm_match,
4705 perf_event_comm_output,
4710 void perf_event_comm(struct task_struct *task)
4712 struct perf_comm_event comm_event;
4713 struct perf_event_context *ctx;
4717 for_each_task_context_nr(ctxn) {
4718 ctx = task->perf_event_ctxp[ctxn];
4722 perf_event_enable_on_exec(ctx);
4726 if (!atomic_read(&nr_comm_events))
4729 comm_event = (struct perf_comm_event){
4735 .type = PERF_RECORD_COMM,
4744 perf_event_comm_event(&comm_event);
4751 struct perf_mmap_event {
4752 struct vm_area_struct *vma;
4754 const char *file_name;
4758 struct perf_event_header header;
4768 static void perf_event_mmap_output(struct perf_event *event,
4771 struct perf_mmap_event *mmap_event = data;
4772 struct perf_output_handle handle;
4773 struct perf_sample_data sample;
4774 int size = mmap_event->event_id.header.size;
4777 perf_event_header__init_id(&mmap_event->event_id.header, &sample, event);
4778 ret = perf_output_begin(&handle, event,
4779 mmap_event->event_id.header.size);
4783 mmap_event->event_id.pid = perf_event_pid(event, current);
4784 mmap_event->event_id.tid = perf_event_tid(event, current);
4786 perf_output_put(&handle, mmap_event->event_id);
4787 __output_copy(&handle, mmap_event->file_name,
4788 mmap_event->file_size);
4790 perf_event__output_id_sample(event, &handle, &sample);
4792 perf_output_end(&handle);
4794 mmap_event->event_id.header.size = size;
4797 static int perf_event_mmap_match(struct perf_event *event,
4800 struct perf_mmap_event *mmap_event = data;
4801 struct vm_area_struct *vma = mmap_event->vma;
4802 int executable = vma->vm_flags & VM_EXEC;
4804 return (!executable && event->attr.mmap_data) ||
4805 (executable && event->attr.mmap);
4808 static void perf_event_mmap_event(struct perf_mmap_event *mmap_event)
4810 struct vm_area_struct *vma = mmap_event->vma;
4811 struct file *file = vma->vm_file;
4817 memset(tmp, 0, sizeof(tmp));
4821 * d_path works from the end of the rb backwards, so we
4822 * need to add enough zero bytes after the string to handle
4823 * the 64bit alignment we do later.
4825 buf = kzalloc(PATH_MAX + sizeof(u64), GFP_KERNEL);
4827 name = strncpy(tmp, "//enomem", sizeof(tmp));
4830 name = d_path(&file->f_path, buf, PATH_MAX);
4832 name = strncpy(tmp, "//toolong", sizeof(tmp));
4836 if (arch_vma_name(mmap_event->vma)) {
4837 name = strncpy(tmp, arch_vma_name(mmap_event->vma),
4839 tmp[sizeof(tmp) - 1] = '\0';
4844 name = strncpy(tmp, "[vdso]", sizeof(tmp));
4846 } else if (vma->vm_start <= vma->vm_mm->start_brk &&
4847 vma->vm_end >= vma->vm_mm->brk) {
4848 name = strncpy(tmp, "[heap]", sizeof(tmp));
4850 } else if (vma->vm_start <= vma->vm_mm->start_stack &&
4851 vma->vm_end >= vma->vm_mm->start_stack) {
4852 name = strncpy(tmp, "[stack]", sizeof(tmp));
4856 name = strncpy(tmp, "//anon", sizeof(tmp));
4861 size = ALIGN(strlen(name)+1, sizeof(u64));
4863 mmap_event->file_name = name;
4864 mmap_event->file_size = size;
4866 if (!(vma->vm_flags & VM_EXEC))
4867 mmap_event->event_id.header.misc |= PERF_RECORD_MISC_MMAP_DATA;
4869 mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size;
4871 perf_event_aux(perf_event_mmap_match,
4872 perf_event_mmap_output,
4879 void perf_event_mmap(struct vm_area_struct *vma)
4881 struct perf_mmap_event mmap_event;
4883 if (!atomic_read(&nr_mmap_events))
4886 mmap_event = (struct perf_mmap_event){
4892 .type = PERF_RECORD_MMAP,
4893 .misc = PERF_RECORD_MISC_USER,
4898 .start = vma->vm_start,
4899 .len = vma->vm_end - vma->vm_start,
4900 .pgoff = (u64)vma->vm_pgoff << PAGE_SHIFT,
4904 perf_event_mmap_event(&mmap_event);
4908 * IRQ throttle logging
4911 static void perf_log_throttle(struct perf_event *event, int enable)
4913 struct perf_output_handle handle;
4914 struct perf_sample_data sample;
4918 struct perf_event_header header;
4922 } throttle_event = {
4924 .type = PERF_RECORD_THROTTLE,
4926 .size = sizeof(throttle_event),
4928 .time = perf_clock(),
4929 .id = primary_event_id(event),
4930 .stream_id = event->id,
4934 throttle_event.header.type = PERF_RECORD_UNTHROTTLE;
4936 perf_event_header__init_id(&throttle_event.header, &sample, event);
4938 ret = perf_output_begin(&handle, event,
4939 throttle_event.header.size);
4943 perf_output_put(&handle, throttle_event);
4944 perf_event__output_id_sample(event, &handle, &sample);
4945 perf_output_end(&handle);
4949 * Generic event overflow handling, sampling.
4952 static int __perf_event_overflow(struct perf_event *event,
4953 int throttle, struct perf_sample_data *data,
4954 struct pt_regs *regs)
4956 int events = atomic_read(&event->event_limit);
4957 struct hw_perf_event *hwc = &event->hw;
4962 * Non-sampling counters might still use the PMI to fold short
4963 * hardware counters, ignore those.
4965 if (unlikely(!is_sampling_event(event)))
4968 seq = __this_cpu_read(perf_throttled_seq);
4969 if (seq != hwc->interrupts_seq) {
4970 hwc->interrupts_seq = seq;
4971 hwc->interrupts = 1;
4974 if (unlikely(throttle
4975 && hwc->interrupts >= max_samples_per_tick)) {
4976 __this_cpu_inc(perf_throttled_count);
4977 hwc->interrupts = MAX_INTERRUPTS;
4978 perf_log_throttle(event, 0);
4983 if (event->attr.freq) {
4984 u64 now = perf_clock();
4985 s64 delta = now - hwc->freq_time_stamp;
4987 hwc->freq_time_stamp = now;
4989 if (delta > 0 && delta < 2*TICK_NSEC)
4990 perf_adjust_period(event, delta, hwc->last_period, true);
4994 * XXX event_limit might not quite work as expected on inherited
4998 event->pending_kill = POLL_IN;
4999 if (events && atomic_dec_and_test(&event->event_limit)) {
5001 event->pending_kill = POLL_HUP;
5002 event->pending_disable = 1;
5003 irq_work_queue(&event->pending);
5006 if (event->overflow_handler)
5007 event->overflow_handler(event, data, regs);
5009 perf_event_output(event, data, regs);
5011 if (event->fasync && event->pending_kill) {
5012 event->pending_wakeup = 1;
5013 irq_work_queue(&event->pending);
5019 int perf_event_overflow(struct perf_event *event,
5020 struct perf_sample_data *data,
5021 struct pt_regs *regs)
5023 return __perf_event_overflow(event, 1, data, regs);
5027 * Generic software event infrastructure
5030 struct swevent_htable {
5031 struct swevent_hlist *swevent_hlist;
5032 struct mutex hlist_mutex;
5035 /* Recursion avoidance in each contexts */
5036 int recursion[PERF_NR_CONTEXTS];
5039 static DEFINE_PER_CPU(struct swevent_htable, swevent_htable);
5042 * We directly increment event->count and keep a second value in
5043 * event->hw.period_left to count intervals. This period event
5044 * is kept in the range [-sample_period, 0] so that we can use the
5048 static u64 perf_swevent_set_period(struct perf_event *event)
5050 struct hw_perf_event *hwc = &event->hw;
5051 u64 period = hwc->last_period;
5055 hwc->last_period = hwc->sample_period;
5058 old = val = local64_read(&hwc->period_left);
5062 nr = div64_u64(period + val, period);
5063 offset = nr * period;
5065 if (local64_cmpxchg(&hwc->period_left, old, val) != old)
5071 static void perf_swevent_overflow(struct perf_event *event, u64 overflow,
5072 struct perf_sample_data *data,
5073 struct pt_regs *regs)
5075 struct hw_perf_event *hwc = &event->hw;
5079 overflow = perf_swevent_set_period(event);
5081 if (hwc->interrupts == MAX_INTERRUPTS)
5084 for (; overflow; overflow--) {
5085 if (__perf_event_overflow(event, throttle,
5088 * We inhibit the overflow from happening when
5089 * hwc->interrupts == MAX_INTERRUPTS.
5097 static void perf_swevent_event(struct perf_event *event, u64 nr,
5098 struct perf_sample_data *data,
5099 struct pt_regs *regs)
5101 struct hw_perf_event *hwc = &event->hw;
5103 local64_add(nr, &event->count);
5108 if (!is_sampling_event(event))
5111 if ((event->attr.sample_type & PERF_SAMPLE_PERIOD) && !event->attr.freq) {
5113 return perf_swevent_overflow(event, 1, data, regs);
5115 data->period = event->hw.last_period;
5117 if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq)
5118 return perf_swevent_overflow(event, 1, data, regs);
5120 if (local64_add_negative(nr, &hwc->period_left))
5123 perf_swevent_overflow(event, 0, data, regs);
5126 static int perf_exclude_event(struct perf_event *event,
5127 struct pt_regs *regs)
5129 if (event->hw.state & PERF_HES_STOPPED)
5133 if (event->attr.exclude_user && user_mode(regs))
5136 if (event->attr.exclude_kernel && !user_mode(regs))
5143 static int perf_swevent_match(struct perf_event *event,
5144 enum perf_type_id type,
5146 struct perf_sample_data *data,
5147 struct pt_regs *regs)
5149 if (event->attr.type != type)
5152 if (event->attr.config != event_id)
5155 if (perf_exclude_event(event, regs))
5161 static inline u64 swevent_hash(u64 type, u32 event_id)
5163 u64 val = event_id | (type << 32);
5165 return hash_64(val, SWEVENT_HLIST_BITS);
5168 static inline struct hlist_head *
5169 __find_swevent_head(struct swevent_hlist *hlist, u64 type, u32 event_id)
5171 u64 hash = swevent_hash(type, event_id);
5173 return &hlist->heads[hash];
5176 /* For the read side: events when they trigger */
5177 static inline struct hlist_head *
5178 find_swevent_head_rcu(struct swevent_htable *swhash, u64 type, u32 event_id)
5180 struct swevent_hlist *hlist;
5182 hlist = rcu_dereference(swhash->swevent_hlist);
5186 return __find_swevent_head(hlist, type, event_id);
5189 /* For the event head insertion and removal in the hlist */
5190 static inline struct hlist_head *
5191 find_swevent_head(struct swevent_htable *swhash, struct perf_event *event)
5193 struct swevent_hlist *hlist;
5194 u32 event_id = event->attr.config;
5195 u64 type = event->attr.type;
5198 * Event scheduling is always serialized against hlist allocation
5199 * and release. Which makes the protected version suitable here.
5200 * The context lock guarantees that.
5202 hlist = rcu_dereference_protected(swhash->swevent_hlist,
5203 lockdep_is_held(&event->ctx->lock));
5207 return __find_swevent_head(hlist, type, event_id);
5210 static void do_perf_sw_event(enum perf_type_id type, u32 event_id,
5212 struct perf_sample_data *data,
5213 struct pt_regs *regs)
5215 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
5216 struct perf_event *event;
5217 struct hlist_head *head;
5220 head = find_swevent_head_rcu(swhash, type, event_id);
5224 hlist_for_each_entry_rcu(event, head, hlist_entry) {
5225 if (perf_swevent_match(event, type, event_id, data, regs))
5226 perf_swevent_event(event, nr, data, regs);
5232 int perf_swevent_get_recursion_context(void)
5234 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
5236 return get_recursion_context(swhash->recursion);
5238 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context);
5240 inline void perf_swevent_put_recursion_context(int rctx)
5242 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
5244 put_recursion_context(swhash->recursion, rctx);
5247 void __perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)
5249 struct perf_sample_data data;
5252 preempt_disable_notrace();
5253 rctx = perf_swevent_get_recursion_context();
5257 perf_sample_data_init(&data, addr, 0);
5259 do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, &data, regs);
5261 perf_swevent_put_recursion_context(rctx);
5262 preempt_enable_notrace();
5265 static void perf_swevent_read(struct perf_event *event)
5269 static int perf_swevent_add(struct perf_event *event, int flags)
5271 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
5272 struct hw_perf_event *hwc = &event->hw;
5273 struct hlist_head *head;
5275 if (is_sampling_event(event)) {
5276 hwc->last_period = hwc->sample_period;
5277 perf_swevent_set_period(event);
5280 hwc->state = !(flags & PERF_EF_START);
5282 head = find_swevent_head(swhash, event);
5283 if (WARN_ON_ONCE(!head))
5286 hlist_add_head_rcu(&event->hlist_entry, head);
5291 static void perf_swevent_del(struct perf_event *event, int flags)
5293 hlist_del_rcu(&event->hlist_entry);
5296 static void perf_swevent_start(struct perf_event *event, int flags)
5298 event->hw.state = 0;
5301 static void perf_swevent_stop(struct perf_event *event, int flags)
5303 event->hw.state = PERF_HES_STOPPED;
5306 /* Deref the hlist from the update side */
5307 static inline struct swevent_hlist *
5308 swevent_hlist_deref(struct swevent_htable *swhash)
5310 return rcu_dereference_protected(swhash->swevent_hlist,
5311 lockdep_is_held(&swhash->hlist_mutex));
5314 static void swevent_hlist_release(struct swevent_htable *swhash)
5316 struct swevent_hlist *hlist = swevent_hlist_deref(swhash);
5321 rcu_assign_pointer(swhash->swevent_hlist, NULL);
5322 kfree_rcu(hlist, rcu_head);
5325 static void swevent_hlist_put_cpu(struct perf_event *event, int cpu)
5327 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
5329 mutex_lock(&swhash->hlist_mutex);
5331 if (!--swhash->hlist_refcount)
5332 swevent_hlist_release(swhash);
5334 mutex_unlock(&swhash->hlist_mutex);
5337 static void swevent_hlist_put(struct perf_event *event)
5341 if (event->cpu != -1) {
5342 swevent_hlist_put_cpu(event, event->cpu);
5346 for_each_possible_cpu(cpu)
5347 swevent_hlist_put_cpu(event, cpu);
5350 static int swevent_hlist_get_cpu(struct perf_event *event, int cpu)
5352 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
5355 mutex_lock(&swhash->hlist_mutex);
5357 if (!swevent_hlist_deref(swhash) && cpu_online(cpu)) {
5358 struct swevent_hlist *hlist;
5360 hlist = kzalloc(sizeof(*hlist), GFP_KERNEL);
5365 rcu_assign_pointer(swhash->swevent_hlist, hlist);
5367 swhash->hlist_refcount++;
5369 mutex_unlock(&swhash->hlist_mutex);
5374 static int swevent_hlist_get(struct perf_event *event)
5377 int cpu, failed_cpu;
5379 if (event->cpu != -1)
5380 return swevent_hlist_get_cpu(event, event->cpu);
5383 for_each_possible_cpu(cpu) {
5384 err = swevent_hlist_get_cpu(event, cpu);
5394 for_each_possible_cpu(cpu) {
5395 if (cpu == failed_cpu)
5397 swevent_hlist_put_cpu(event, cpu);
5404 struct static_key perf_swevent_enabled[PERF_COUNT_SW_MAX];
5406 static void sw_perf_event_destroy(struct perf_event *event)
5408 u64 event_id = event->attr.config;
5410 WARN_ON(event->parent);
5412 static_key_slow_dec(&perf_swevent_enabled[event_id]);
5413 swevent_hlist_put(event);
5416 static int perf_swevent_init(struct perf_event *event)
5418 u64 event_id = event->attr.config;
5420 if (event->attr.type != PERF_TYPE_SOFTWARE)
5424 * no branch sampling for software events
5426 if (has_branch_stack(event))
5430 case PERF_COUNT_SW_CPU_CLOCK:
5431 case PERF_COUNT_SW_TASK_CLOCK:
5438 if (event_id >= PERF_COUNT_SW_MAX)
5441 if (!event->parent) {
5444 err = swevent_hlist_get(event);
5448 static_key_slow_inc(&perf_swevent_enabled[event_id]);
5449 event->destroy = sw_perf_event_destroy;
5455 static int perf_swevent_event_idx(struct perf_event *event)
5460 static struct pmu perf_swevent = {
5461 .task_ctx_nr = perf_sw_context,
5463 .event_init = perf_swevent_init,
5464 .add = perf_swevent_add,
5465 .del = perf_swevent_del,
5466 .start = perf_swevent_start,
5467 .stop = perf_swevent_stop,
5468 .read = perf_swevent_read,
5470 .event_idx = perf_swevent_event_idx,
5473 #ifdef CONFIG_EVENT_TRACING
5475 static int perf_tp_filter_match(struct perf_event *event,
5476 struct perf_sample_data *data)
5478 void *record = data->raw->data;
5480 if (likely(!event->filter) || filter_match_preds(event->filter, record))
5485 static int perf_tp_event_match(struct perf_event *event,
5486 struct perf_sample_data *data,
5487 struct pt_regs *regs)
5489 if (event->hw.state & PERF_HES_STOPPED)
5492 * All tracepoints are from kernel-space.
5494 if (event->attr.exclude_kernel)
5497 if (!perf_tp_filter_match(event, data))
5503 void perf_tp_event(u64 addr, u64 count, void *record, int entry_size,
5504 struct pt_regs *regs, struct hlist_head *head, int rctx,
5505 struct task_struct *task)
5507 struct perf_sample_data data;
5508 struct perf_event *event;
5510 struct perf_raw_record raw = {
5515 perf_sample_data_init(&data, addr, 0);
5518 hlist_for_each_entry_rcu(event, head, hlist_entry) {
5519 if (perf_tp_event_match(event, &data, regs))
5520 perf_swevent_event(event, count, &data, regs);
5524 * If we got specified a target task, also iterate its context and
5525 * deliver this event there too.
5527 if (task && task != current) {
5528 struct perf_event_context *ctx;
5529 struct trace_entry *entry = record;
5532 ctx = rcu_dereference(task->perf_event_ctxp[perf_sw_context]);
5536 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
5537 if (event->attr.type != PERF_TYPE_TRACEPOINT)
5539 if (event->attr.config != entry->type)
5541 if (perf_tp_event_match(event, &data, regs))
5542 perf_swevent_event(event, count, &data, regs);
5548 perf_swevent_put_recursion_context(rctx);
5550 EXPORT_SYMBOL_GPL(perf_tp_event);
5552 static void tp_perf_event_destroy(struct perf_event *event)
5554 perf_trace_destroy(event);
5557 static int perf_tp_event_init(struct perf_event *event)
5561 if (event->attr.type != PERF_TYPE_TRACEPOINT)
5565 * no branch sampling for tracepoint events
5567 if (has_branch_stack(event))
5570 err = perf_trace_init(event);
5574 event->destroy = tp_perf_event_destroy;
5579 static struct pmu perf_tracepoint = {
5580 .task_ctx_nr = perf_sw_context,
5582 .event_init = perf_tp_event_init,
5583 .add = perf_trace_add,
5584 .del = perf_trace_del,
5585 .start = perf_swevent_start,
5586 .stop = perf_swevent_stop,
5587 .read = perf_swevent_read,
5589 .event_idx = perf_swevent_event_idx,
5592 static inline void perf_tp_register(void)
5594 perf_pmu_register(&perf_tracepoint, "tracepoint", PERF_TYPE_TRACEPOINT);
5597 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
5602 if (event->attr.type != PERF_TYPE_TRACEPOINT)
5605 filter_str = strndup_user(arg, PAGE_SIZE);
5606 if (IS_ERR(filter_str))
5607 return PTR_ERR(filter_str);
5609 ret = ftrace_profile_set_filter(event, event->attr.config, filter_str);
5615 static void perf_event_free_filter(struct perf_event *event)
5617 ftrace_profile_free_filter(event);
5622 static inline void perf_tp_register(void)
5626 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
5631 static void perf_event_free_filter(struct perf_event *event)
5635 #endif /* CONFIG_EVENT_TRACING */
5637 #ifdef CONFIG_HAVE_HW_BREAKPOINT
5638 void perf_bp_event(struct perf_event *bp, void *data)
5640 struct perf_sample_data sample;
5641 struct pt_regs *regs = data;
5643 perf_sample_data_init(&sample, bp->attr.bp_addr, 0);
5645 if (!bp->hw.state && !perf_exclude_event(bp, regs))
5646 perf_swevent_event(bp, 1, &sample, regs);
5651 * hrtimer based swevent callback
5654 static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer)
5656 enum hrtimer_restart ret = HRTIMER_RESTART;
5657 struct perf_sample_data data;
5658 struct pt_regs *regs;
5659 struct perf_event *event;
5662 event = container_of(hrtimer, struct perf_event, hw.hrtimer);
5664 if (event->state != PERF_EVENT_STATE_ACTIVE)
5665 return HRTIMER_NORESTART;
5667 event->pmu->read(event);
5669 perf_sample_data_init(&data, 0, event->hw.last_period);
5670 regs = get_irq_regs();
5672 if (regs && !perf_exclude_event(event, regs)) {
5673 if (!(event->attr.exclude_idle && is_idle_task(current)))
5674 if (__perf_event_overflow(event, 1, &data, regs))
5675 ret = HRTIMER_NORESTART;
5678 period = max_t(u64, 10000, event->hw.sample_period);
5679 hrtimer_forward_now(hrtimer, ns_to_ktime(period));
5684 static void perf_swevent_start_hrtimer(struct perf_event *event)
5686 struct hw_perf_event *hwc = &event->hw;
5689 if (!is_sampling_event(event))
5692 period = local64_read(&hwc->period_left);
5697 local64_set(&hwc->period_left, 0);
5699 period = max_t(u64, 10000, hwc->sample_period);
5701 __hrtimer_start_range_ns(&hwc->hrtimer,
5702 ns_to_ktime(period), 0,
5703 HRTIMER_MODE_REL_PINNED, 0);
5706 static void perf_swevent_cancel_hrtimer(struct perf_event *event)
5708 struct hw_perf_event *hwc = &event->hw;
5710 if (is_sampling_event(event)) {
5711 ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer);
5712 local64_set(&hwc->period_left, ktime_to_ns(remaining));
5714 hrtimer_cancel(&hwc->hrtimer);
5718 static void perf_swevent_init_hrtimer(struct perf_event *event)
5720 struct hw_perf_event *hwc = &event->hw;
5722 if (!is_sampling_event(event))
5725 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
5726 hwc->hrtimer.function = perf_swevent_hrtimer;
5729 * Since hrtimers have a fixed rate, we can do a static freq->period
5730 * mapping and avoid the whole period adjust feedback stuff.
5732 if (event->attr.freq) {
5733 long freq = event->attr.sample_freq;
5735 event->attr.sample_period = NSEC_PER_SEC / freq;
5736 hwc->sample_period = event->attr.sample_period;
5737 local64_set(&hwc->period_left, hwc->sample_period);
5738 hwc->last_period = hwc->sample_period;
5739 event->attr.freq = 0;
5744 * Software event: cpu wall time clock
5747 static void cpu_clock_event_update(struct perf_event *event)
5752 now = local_clock();
5753 prev = local64_xchg(&event->hw.prev_count, now);
5754 local64_add(now - prev, &event->count);
5757 static void cpu_clock_event_start(struct perf_event *event, int flags)
5759 local64_set(&event->hw.prev_count, local_clock());
5760 perf_swevent_start_hrtimer(event);
5763 static void cpu_clock_event_stop(struct perf_event *event, int flags)
5765 perf_swevent_cancel_hrtimer(event);
5766 cpu_clock_event_update(event);
5769 static int cpu_clock_event_add(struct perf_event *event, int flags)
5771 if (flags & PERF_EF_START)
5772 cpu_clock_event_start(event, flags);
5777 static void cpu_clock_event_del(struct perf_event *event, int flags)
5779 cpu_clock_event_stop(event, flags);
5782 static void cpu_clock_event_read(struct perf_event *event)
5784 cpu_clock_event_update(event);
5787 static int cpu_clock_event_init(struct perf_event *event)
5789 if (event->attr.type != PERF_TYPE_SOFTWARE)
5792 if (event->attr.config != PERF_COUNT_SW_CPU_CLOCK)
5796 * no branch sampling for software events
5798 if (has_branch_stack(event))
5801 perf_swevent_init_hrtimer(event);
5806 static struct pmu perf_cpu_clock = {
5807 .task_ctx_nr = perf_sw_context,
5809 .event_init = cpu_clock_event_init,
5810 .add = cpu_clock_event_add,
5811 .del = cpu_clock_event_del,
5812 .start = cpu_clock_event_start,
5813 .stop = cpu_clock_event_stop,
5814 .read = cpu_clock_event_read,
5816 .event_idx = perf_swevent_event_idx,
5820 * Software event: task time clock
5823 static void task_clock_event_update(struct perf_event *event, u64 now)
5828 prev = local64_xchg(&event->hw.prev_count, now);
5830 local64_add(delta, &event->count);
5833 static void task_clock_event_start(struct perf_event *event, int flags)
5835 local64_set(&event->hw.prev_count, event->ctx->time);
5836 perf_swevent_start_hrtimer(event);
5839 static void task_clock_event_stop(struct perf_event *event, int flags)
5841 perf_swevent_cancel_hrtimer(event);
5842 task_clock_event_update(event, event->ctx->time);
5845 static int task_clock_event_add(struct perf_event *event, int flags)
5847 if (flags & PERF_EF_START)
5848 task_clock_event_start(event, flags);
5853 static void task_clock_event_del(struct perf_event *event, int flags)
5855 task_clock_event_stop(event, PERF_EF_UPDATE);
5858 static void task_clock_event_read(struct perf_event *event)
5860 u64 now = perf_clock();
5861 u64 delta = now - event->ctx->timestamp;
5862 u64 time = event->ctx->time + delta;
5864 task_clock_event_update(event, time);
5867 static int task_clock_event_init(struct perf_event *event)
5869 if (event->attr.type != PERF_TYPE_SOFTWARE)
5872 if (event->attr.config != PERF_COUNT_SW_TASK_CLOCK)
5876 * no branch sampling for software events
5878 if (has_branch_stack(event))
5881 perf_swevent_init_hrtimer(event);
5886 static struct pmu perf_task_clock = {
5887 .task_ctx_nr = perf_sw_context,
5889 .event_init = task_clock_event_init,
5890 .add = task_clock_event_add,
5891 .del = task_clock_event_del,
5892 .start = task_clock_event_start,
5893 .stop = task_clock_event_stop,
5894 .read = task_clock_event_read,
5896 .event_idx = perf_swevent_event_idx,
5899 static void perf_pmu_nop_void(struct pmu *pmu)
5903 static int perf_pmu_nop_int(struct pmu *pmu)
5908 static void perf_pmu_start_txn(struct pmu *pmu)
5910 perf_pmu_disable(pmu);
5913 static int perf_pmu_commit_txn(struct pmu *pmu)
5915 perf_pmu_enable(pmu);
5919 static void perf_pmu_cancel_txn(struct pmu *pmu)
5921 perf_pmu_enable(pmu);
5924 static int perf_event_idx_default(struct perf_event *event)
5926 return event->hw.idx + 1;
5930 * Ensures all contexts with the same task_ctx_nr have the same
5931 * pmu_cpu_context too.
5933 static void *find_pmu_context(int ctxn)
5940 list_for_each_entry(pmu, &pmus, entry) {
5941 if (pmu->task_ctx_nr == ctxn)
5942 return pmu->pmu_cpu_context;
5948 static void update_pmu_context(struct pmu *pmu, struct pmu *old_pmu)
5952 for_each_possible_cpu(cpu) {
5953 struct perf_cpu_context *cpuctx;
5955 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
5957 if (cpuctx->unique_pmu == old_pmu)
5958 cpuctx->unique_pmu = pmu;
5962 static void free_pmu_context(struct pmu *pmu)
5966 mutex_lock(&pmus_lock);
5968 * Like a real lame refcount.
5970 list_for_each_entry(i, &pmus, entry) {
5971 if (i->pmu_cpu_context == pmu->pmu_cpu_context) {
5972 update_pmu_context(i, pmu);
5977 free_percpu(pmu->pmu_cpu_context);
5979 mutex_unlock(&pmus_lock);
5981 static struct idr pmu_idr;
5984 type_show(struct device *dev, struct device_attribute *attr, char *page)
5986 struct pmu *pmu = dev_get_drvdata(dev);
5988 return snprintf(page, PAGE_SIZE-1, "%d\n", pmu->type);
5991 static struct device_attribute pmu_dev_attrs[] = {
5996 static int pmu_bus_running;
5997 static struct bus_type pmu_bus = {
5998 .name = "event_source",
5999 .dev_attrs = pmu_dev_attrs,
6002 static void pmu_dev_release(struct device *dev)
6007 static int pmu_dev_alloc(struct pmu *pmu)
6011 pmu->dev = kzalloc(sizeof(struct device), GFP_KERNEL);
6015 pmu->dev->groups = pmu->attr_groups;
6016 device_initialize(pmu->dev);
6017 ret = dev_set_name(pmu->dev, "%s", pmu->name);
6021 dev_set_drvdata(pmu->dev, pmu);
6022 pmu->dev->bus = &pmu_bus;
6023 pmu->dev->release = pmu_dev_release;
6024 ret = device_add(pmu->dev);
6032 put_device(pmu->dev);
6036 static struct lock_class_key cpuctx_mutex;
6037 static struct lock_class_key cpuctx_lock;
6039 int perf_pmu_register(struct pmu *pmu, char *name, int type)
6043 mutex_lock(&pmus_lock);
6045 pmu->pmu_disable_count = alloc_percpu(int);
6046 if (!pmu->pmu_disable_count)
6055 type = idr_alloc(&pmu_idr, pmu, PERF_TYPE_MAX, 0, GFP_KERNEL);
6063 if (pmu_bus_running) {
6064 ret = pmu_dev_alloc(pmu);
6070 pmu->pmu_cpu_context = find_pmu_context(pmu->task_ctx_nr);
6071 if (pmu->pmu_cpu_context)
6072 goto got_cpu_context;
6075 pmu->pmu_cpu_context = alloc_percpu(struct perf_cpu_context);
6076 if (!pmu->pmu_cpu_context)
6079 for_each_possible_cpu(cpu) {
6080 struct perf_cpu_context *cpuctx;
6082 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
6083 __perf_event_init_context(&cpuctx->ctx);
6084 lockdep_set_class(&cpuctx->ctx.mutex, &cpuctx_mutex);
6085 lockdep_set_class(&cpuctx->ctx.lock, &cpuctx_lock);
6086 cpuctx->ctx.type = cpu_context;
6087 cpuctx->ctx.pmu = pmu;
6088 cpuctx->jiffies_interval = 1;
6089 INIT_LIST_HEAD(&cpuctx->rotation_list);
6090 cpuctx->unique_pmu = pmu;
6094 if (!pmu->start_txn) {
6095 if (pmu->pmu_enable) {
6097 * If we have pmu_enable/pmu_disable calls, install
6098 * transaction stubs that use that to try and batch
6099 * hardware accesses.
6101 pmu->start_txn = perf_pmu_start_txn;
6102 pmu->commit_txn = perf_pmu_commit_txn;
6103 pmu->cancel_txn = perf_pmu_cancel_txn;
6105 pmu->start_txn = perf_pmu_nop_void;
6106 pmu->commit_txn = perf_pmu_nop_int;
6107 pmu->cancel_txn = perf_pmu_nop_void;
6111 if (!pmu->pmu_enable) {
6112 pmu->pmu_enable = perf_pmu_nop_void;
6113 pmu->pmu_disable = perf_pmu_nop_void;
6116 if (!pmu->event_idx)
6117 pmu->event_idx = perf_event_idx_default;
6119 list_add_rcu(&pmu->entry, &pmus);
6122 mutex_unlock(&pmus_lock);
6127 device_del(pmu->dev);
6128 put_device(pmu->dev);
6131 if (pmu->type >= PERF_TYPE_MAX)
6132 idr_remove(&pmu_idr, pmu->type);
6135 free_percpu(pmu->pmu_disable_count);
6139 void perf_pmu_unregister(struct pmu *pmu)
6141 mutex_lock(&pmus_lock);
6142 list_del_rcu(&pmu->entry);
6143 mutex_unlock(&pmus_lock);
6146 * We dereference the pmu list under both SRCU and regular RCU, so
6147 * synchronize against both of those.
6149 synchronize_srcu(&pmus_srcu);
6152 free_percpu(pmu->pmu_disable_count);
6153 if (pmu->type >= PERF_TYPE_MAX)
6154 idr_remove(&pmu_idr, pmu->type);
6155 device_del(pmu->dev);
6156 put_device(pmu->dev);
6157 free_pmu_context(pmu);
6160 struct pmu *perf_init_event(struct perf_event *event)
6162 struct pmu *pmu = NULL;
6166 idx = srcu_read_lock(&pmus_srcu);
6169 pmu = idr_find(&pmu_idr, event->attr.type);
6173 ret = pmu->event_init(event);
6179 list_for_each_entry_rcu(pmu, &pmus, entry) {
6181 ret = pmu->event_init(event);
6185 if (ret != -ENOENT) {
6190 pmu = ERR_PTR(-ENOENT);
6192 srcu_read_unlock(&pmus_srcu, idx);
6198 * Allocate and initialize a event structure
6200 static struct perf_event *
6201 perf_event_alloc(struct perf_event_attr *attr, int cpu,
6202 struct task_struct *task,
6203 struct perf_event *group_leader,
6204 struct perf_event *parent_event,
6205 perf_overflow_handler_t overflow_handler,
6209 struct perf_event *event;
6210 struct hw_perf_event *hwc;
6213 if ((unsigned)cpu >= nr_cpu_ids) {
6214 if (!task || cpu != -1)
6215 return ERR_PTR(-EINVAL);
6218 event = kzalloc(sizeof(*event), GFP_KERNEL);
6220 return ERR_PTR(-ENOMEM);
6223 * Single events are their own group leaders, with an
6224 * empty sibling list:
6227 group_leader = event;
6229 mutex_init(&event->child_mutex);
6230 INIT_LIST_HEAD(&event->child_list);
6232 INIT_LIST_HEAD(&event->group_entry);
6233 INIT_LIST_HEAD(&event->event_entry);
6234 INIT_LIST_HEAD(&event->sibling_list);
6235 INIT_LIST_HEAD(&event->rb_entry);
6237 init_waitqueue_head(&event->waitq);
6238 init_irq_work(&event->pending, perf_pending_event);
6240 mutex_init(&event->mmap_mutex);
6242 atomic_long_set(&event->refcount, 1);
6244 event->attr = *attr;
6245 event->group_leader = group_leader;
6249 event->parent = parent_event;
6251 event->ns = get_pid_ns(task_active_pid_ns(current));
6252 event->id = atomic64_inc_return(&perf_event_id);
6254 event->state = PERF_EVENT_STATE_INACTIVE;
6257 event->attach_state = PERF_ATTACH_TASK;
6259 if (attr->type == PERF_TYPE_TRACEPOINT)
6260 event->hw.tp_target = task;
6261 #ifdef CONFIG_HAVE_HW_BREAKPOINT
6263 * hw_breakpoint is a bit difficult here..
6265 else if (attr->type == PERF_TYPE_BREAKPOINT)
6266 event->hw.bp_target = task;
6270 if (!overflow_handler && parent_event) {
6271 overflow_handler = parent_event->overflow_handler;
6272 context = parent_event->overflow_handler_context;
6275 event->overflow_handler = overflow_handler;
6276 event->overflow_handler_context = context;
6278 perf_event__state_init(event);
6283 hwc->sample_period = attr->sample_period;
6284 if (attr->freq && attr->sample_freq)
6285 hwc->sample_period = 1;
6286 hwc->last_period = hwc->sample_period;
6288 local64_set(&hwc->period_left, hwc->sample_period);
6291 * we currently do not support PERF_FORMAT_GROUP on inherited events
6293 if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP))
6296 pmu = perf_init_event(event);
6302 else if (IS_ERR(pmu))
6307 put_pid_ns(event->ns);
6309 return ERR_PTR(err);
6312 if (!event->parent) {
6313 if (event->attach_state & PERF_ATTACH_TASK)
6314 static_key_slow_inc(&perf_sched_events.key);
6315 if (event->attr.mmap || event->attr.mmap_data)
6316 atomic_inc(&nr_mmap_events);
6317 if (event->attr.comm)
6318 atomic_inc(&nr_comm_events);
6319 if (event->attr.task)
6320 atomic_inc(&nr_task_events);
6321 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) {
6322 err = get_callchain_buffers();
6325 return ERR_PTR(err);
6328 if (has_branch_stack(event)) {
6329 static_key_slow_inc(&perf_sched_events.key);
6330 if (!(event->attach_state & PERF_ATTACH_TASK))
6331 atomic_inc(&per_cpu(perf_branch_stack_events,
6339 static int perf_copy_attr(struct perf_event_attr __user *uattr,
6340 struct perf_event_attr *attr)
6345 if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0))
6349 * zero the full structure, so that a short copy will be nice.
6351 memset(attr, 0, sizeof(*attr));
6353 ret = get_user(size, &uattr->size);
6357 if (size > PAGE_SIZE) /* silly large */
6360 if (!size) /* abi compat */
6361 size = PERF_ATTR_SIZE_VER0;
6363 if (size < PERF_ATTR_SIZE_VER0)
6367 * If we're handed a bigger struct than we know of,
6368 * ensure all the unknown bits are 0 - i.e. new
6369 * user-space does not rely on any kernel feature
6370 * extensions we dont know about yet.
6372 if (size > sizeof(*attr)) {
6373 unsigned char __user *addr;
6374 unsigned char __user *end;
6377 addr = (void __user *)uattr + sizeof(*attr);
6378 end = (void __user *)uattr + size;
6380 for (; addr < end; addr++) {
6381 ret = get_user(val, addr);
6387 size = sizeof(*attr);
6390 ret = copy_from_user(attr, uattr, size);
6394 if (attr->__reserved_1)
6397 if (attr->sample_type & ~(PERF_SAMPLE_MAX-1))
6400 if (attr->read_format & ~(PERF_FORMAT_MAX-1))
6403 if (attr->sample_type & PERF_SAMPLE_BRANCH_STACK) {
6404 u64 mask = attr->branch_sample_type;
6406 /* only using defined bits */
6407 if (mask & ~(PERF_SAMPLE_BRANCH_MAX-1))
6410 /* at least one branch bit must be set */
6411 if (!(mask & ~PERF_SAMPLE_BRANCH_PLM_ALL))
6414 /* kernel level capture: check permissions */
6415 if ((mask & PERF_SAMPLE_BRANCH_PERM_PLM)
6416 && perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
6419 /* propagate priv level, when not set for branch */
6420 if (!(mask & PERF_SAMPLE_BRANCH_PLM_ALL)) {
6422 /* exclude_kernel checked on syscall entry */
6423 if (!attr->exclude_kernel)
6424 mask |= PERF_SAMPLE_BRANCH_KERNEL;
6426 if (!attr->exclude_user)
6427 mask |= PERF_SAMPLE_BRANCH_USER;
6429 if (!attr->exclude_hv)
6430 mask |= PERF_SAMPLE_BRANCH_HV;
6432 * adjust user setting (for HW filter setup)
6434 attr->branch_sample_type = mask;
6438 if (attr->sample_type & PERF_SAMPLE_REGS_USER) {
6439 ret = perf_reg_validate(attr->sample_regs_user);
6444 if (attr->sample_type & PERF_SAMPLE_STACK_USER) {
6445 if (!arch_perf_have_user_stack_dump())
6449 * We have __u32 type for the size, but so far
6450 * we can only use __u16 as maximum due to the
6451 * __u16 sample size limit.
6453 if (attr->sample_stack_user >= USHRT_MAX)
6455 else if (!IS_ALIGNED(attr->sample_stack_user, sizeof(u64)))
6463 put_user(sizeof(*attr), &uattr->size);
6469 perf_event_set_output(struct perf_event *event, struct perf_event *output_event)
6471 struct ring_buffer *rb = NULL, *old_rb = NULL;
6477 /* don't allow circular references */
6478 if (event == output_event)
6482 * Don't allow cross-cpu buffers
6484 if (output_event->cpu != event->cpu)
6488 * If its not a per-cpu rb, it must be the same task.
6490 if (output_event->cpu == -1 && output_event->ctx != event->ctx)
6494 mutex_lock(&event->mmap_mutex);
6495 /* Can't redirect output if we've got an active mmap() */
6496 if (atomic_read(&event->mmap_count))
6502 /* get the rb we want to redirect to */
6503 rb = ring_buffer_get(output_event);
6509 ring_buffer_detach(event, old_rb);
6512 ring_buffer_attach(event, rb);
6514 rcu_assign_pointer(event->rb, rb);
6517 ring_buffer_put(old_rb);
6519 * Since we detached before setting the new rb, so that we
6520 * could attach the new rb, we could have missed a wakeup.
6523 wake_up_all(&event->waitq);
6528 mutex_unlock(&event->mmap_mutex);
6535 * sys_perf_event_open - open a performance event, associate it to a task/cpu
6537 * @attr_uptr: event_id type attributes for monitoring/sampling
6540 * @group_fd: group leader event fd
6542 SYSCALL_DEFINE5(perf_event_open,
6543 struct perf_event_attr __user *, attr_uptr,
6544 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
6546 struct perf_event *group_leader = NULL, *output_event = NULL;
6547 struct perf_event *event, *sibling;
6548 struct perf_event_attr attr;
6549 struct perf_event_context *ctx;
6550 struct file *event_file = NULL;
6551 struct fd group = {NULL, 0};
6552 struct task_struct *task = NULL;
6558 /* for future expandability... */
6559 if (flags & ~PERF_FLAG_ALL)
6562 err = perf_copy_attr(attr_uptr, &attr);
6566 if (!attr.exclude_kernel) {
6567 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
6572 if (attr.sample_freq > sysctl_perf_event_sample_rate)
6577 * In cgroup mode, the pid argument is used to pass the fd
6578 * opened to the cgroup directory in cgroupfs. The cpu argument
6579 * designates the cpu on which to monitor threads from that
6582 if ((flags & PERF_FLAG_PID_CGROUP) && (pid == -1 || cpu == -1))
6585 event_fd = get_unused_fd();
6589 if (group_fd != -1) {
6590 err = perf_fget_light(group_fd, &group);
6593 group_leader = group.file->private_data;
6594 if (flags & PERF_FLAG_FD_OUTPUT)
6595 output_event = group_leader;
6596 if (flags & PERF_FLAG_FD_NO_GROUP)
6597 group_leader = NULL;
6600 if (pid != -1 && !(flags & PERF_FLAG_PID_CGROUP)) {
6601 task = find_lively_task_by_vpid(pid);
6603 err = PTR_ERR(task);
6610 event = perf_event_alloc(&attr, cpu, task, group_leader, NULL,
6612 if (IS_ERR(event)) {
6613 err = PTR_ERR(event);
6617 if (flags & PERF_FLAG_PID_CGROUP) {
6618 err = perf_cgroup_connect(pid, event, &attr, group_leader);
6623 * - that has cgroup constraint on event->cpu
6624 * - that may need work on context switch
6626 atomic_inc(&per_cpu(perf_cgroup_events, event->cpu));
6627 static_key_slow_inc(&perf_sched_events.key);
6631 * Special case software events and allow them to be part of
6632 * any hardware group.
6637 (is_software_event(event) != is_software_event(group_leader))) {
6638 if (is_software_event(event)) {
6640 * If event and group_leader are not both a software
6641 * event, and event is, then group leader is not.
6643 * Allow the addition of software events to !software
6644 * groups, this is safe because software events never
6647 pmu = group_leader->pmu;
6648 } else if (is_software_event(group_leader) &&
6649 (group_leader->group_flags & PERF_GROUP_SOFTWARE)) {
6651 * In case the group is a pure software group, and we
6652 * try to add a hardware event, move the whole group to
6653 * the hardware context.
6660 * Get the target context (task or percpu):
6662 ctx = find_get_context(pmu, task, event->cpu);
6669 put_task_struct(task);
6674 * Look up the group leader (we will attach this event to it):
6680 * Do not allow a recursive hierarchy (this new sibling
6681 * becoming part of another group-sibling):
6683 if (group_leader->group_leader != group_leader)
6686 * Do not allow to attach to a group in a different
6687 * task or CPU context:
6690 if (group_leader->ctx->type != ctx->type)
6693 if (group_leader->ctx != ctx)
6698 * Only a group leader can be exclusive or pinned
6700 if (attr.exclusive || attr.pinned)
6705 err = perf_event_set_output(event, output_event);
6710 event_file = anon_inode_getfile("[perf_event]", &perf_fops, event, O_RDWR);
6711 if (IS_ERR(event_file)) {
6712 err = PTR_ERR(event_file);
6717 struct perf_event_context *gctx = group_leader->ctx;
6719 mutex_lock(&gctx->mutex);
6720 perf_remove_from_context(group_leader);
6723 * Removing from the context ends up with disabled
6724 * event. What we want here is event in the initial
6725 * startup state, ready to be add into new context.
6727 perf_event__state_init(group_leader);
6728 list_for_each_entry(sibling, &group_leader->sibling_list,
6730 perf_remove_from_context(sibling);
6731 perf_event__state_init(sibling);
6734 mutex_unlock(&gctx->mutex);
6738 WARN_ON_ONCE(ctx->parent_ctx);
6739 mutex_lock(&ctx->mutex);
6743 perf_install_in_context(ctx, group_leader, event->cpu);
6745 list_for_each_entry(sibling, &group_leader->sibling_list,
6747 perf_install_in_context(ctx, sibling, event->cpu);
6752 perf_install_in_context(ctx, event, event->cpu);
6754 perf_unpin_context(ctx);
6755 mutex_unlock(&ctx->mutex);
6759 event->owner = current;
6761 mutex_lock(¤t->perf_event_mutex);
6762 list_add_tail(&event->owner_entry, ¤t->perf_event_list);
6763 mutex_unlock(¤t->perf_event_mutex);
6766 * Precalculate sample_data sizes
6768 perf_event__header_size(event);
6769 perf_event__id_header_size(event);
6772 * Drop the reference on the group_event after placing the
6773 * new event on the sibling_list. This ensures destruction
6774 * of the group leader will find the pointer to itself in
6775 * perf_group_detach().
6778 fd_install(event_fd, event_file);
6782 perf_unpin_context(ctx);
6789 put_task_struct(task);
6793 put_unused_fd(event_fd);
6798 * perf_event_create_kernel_counter
6800 * @attr: attributes of the counter to create
6801 * @cpu: cpu in which the counter is bound
6802 * @task: task to profile (NULL for percpu)
6805 perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu,
6806 struct task_struct *task,
6807 perf_overflow_handler_t overflow_handler,
6810 struct perf_event_context *ctx;
6811 struct perf_event *event;
6815 * Get the target context (task or percpu):
6818 event = perf_event_alloc(attr, cpu, task, NULL, NULL,
6819 overflow_handler, context);
6820 if (IS_ERR(event)) {
6821 err = PTR_ERR(event);
6825 ctx = find_get_context(event->pmu, task, cpu);
6831 WARN_ON_ONCE(ctx->parent_ctx);
6832 mutex_lock(&ctx->mutex);
6833 perf_install_in_context(ctx, event, cpu);
6835 perf_unpin_context(ctx);
6836 mutex_unlock(&ctx->mutex);
6843 return ERR_PTR(err);
6845 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter);
6847 void perf_pmu_migrate_context(struct pmu *pmu, int src_cpu, int dst_cpu)
6849 struct perf_event_context *src_ctx;
6850 struct perf_event_context *dst_ctx;
6851 struct perf_event *event, *tmp;
6854 src_ctx = &per_cpu_ptr(pmu->pmu_cpu_context, src_cpu)->ctx;
6855 dst_ctx = &per_cpu_ptr(pmu->pmu_cpu_context, dst_cpu)->ctx;
6857 mutex_lock(&src_ctx->mutex);
6858 list_for_each_entry_safe(event, tmp, &src_ctx->event_list,
6860 perf_remove_from_context(event);
6862 list_add(&event->event_entry, &events);
6864 mutex_unlock(&src_ctx->mutex);
6868 mutex_lock(&dst_ctx->mutex);
6869 list_for_each_entry_safe(event, tmp, &events, event_entry) {
6870 list_del(&event->event_entry);
6871 if (event->state >= PERF_EVENT_STATE_OFF)
6872 event->state = PERF_EVENT_STATE_INACTIVE;
6873 perf_install_in_context(dst_ctx, event, dst_cpu);
6876 mutex_unlock(&dst_ctx->mutex);
6878 EXPORT_SYMBOL_GPL(perf_pmu_migrate_context);
6880 static void sync_child_event(struct perf_event *child_event,
6881 struct task_struct *child)
6883 struct perf_event *parent_event = child_event->parent;
6886 if (child_event->attr.inherit_stat)
6887 perf_event_read_event(child_event, child);
6889 child_val = perf_event_count(child_event);
6892 * Add back the child's count to the parent's count:
6894 atomic64_add(child_val, &parent_event->child_count);
6895 atomic64_add(child_event->total_time_enabled,
6896 &parent_event->child_total_time_enabled);
6897 atomic64_add(child_event->total_time_running,
6898 &parent_event->child_total_time_running);
6901 * Remove this event from the parent's list
6903 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
6904 mutex_lock(&parent_event->child_mutex);
6905 list_del_init(&child_event->child_list);
6906 mutex_unlock(&parent_event->child_mutex);
6909 * Release the parent event, if this was the last
6912 put_event(parent_event);
6916 __perf_event_exit_task(struct perf_event *child_event,
6917 struct perf_event_context *child_ctx,
6918 struct task_struct *child)
6920 if (child_event->parent) {
6921 raw_spin_lock_irq(&child_ctx->lock);
6922 perf_group_detach(child_event);
6923 raw_spin_unlock_irq(&child_ctx->lock);
6926 perf_remove_from_context(child_event);
6929 * It can happen that the parent exits first, and has events
6930 * that are still around due to the child reference. These
6931 * events need to be zapped.
6933 if (child_event->parent) {
6934 sync_child_event(child_event, child);
6935 free_event(child_event);
6939 static void perf_event_exit_task_context(struct task_struct *child, int ctxn)
6941 struct perf_event *child_event, *tmp;
6942 struct perf_event_context *child_ctx;
6943 unsigned long flags;
6945 if (likely(!child->perf_event_ctxp[ctxn])) {
6946 perf_event_task(child, NULL, 0);
6950 local_irq_save(flags);
6952 * We can't reschedule here because interrupts are disabled,
6953 * and either child is current or it is a task that can't be
6954 * scheduled, so we are now safe from rescheduling changing
6957 child_ctx = rcu_dereference_raw(child->perf_event_ctxp[ctxn]);
6960 * Take the context lock here so that if find_get_context is
6961 * reading child->perf_event_ctxp, we wait until it has
6962 * incremented the context's refcount before we do put_ctx below.
6964 raw_spin_lock(&child_ctx->lock);
6965 task_ctx_sched_out(child_ctx);
6966 child->perf_event_ctxp[ctxn] = NULL;
6968 * If this context is a clone; unclone it so it can't get
6969 * swapped to another process while we're removing all
6970 * the events from it.
6972 unclone_ctx(child_ctx);
6973 update_context_time(child_ctx);
6974 raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
6977 * Report the task dead after unscheduling the events so that we
6978 * won't get any samples after PERF_RECORD_EXIT. We can however still
6979 * get a few PERF_RECORD_READ events.
6981 perf_event_task(child, child_ctx, 0);
6984 * We can recurse on the same lock type through:
6986 * __perf_event_exit_task()
6987 * sync_child_event()
6989 * mutex_lock(&ctx->mutex)
6991 * But since its the parent context it won't be the same instance.
6993 mutex_lock(&child_ctx->mutex);
6996 list_for_each_entry_safe(child_event, tmp, &child_ctx->pinned_groups,
6998 __perf_event_exit_task(child_event, child_ctx, child);
7000 list_for_each_entry_safe(child_event, tmp, &child_ctx->flexible_groups,
7002 __perf_event_exit_task(child_event, child_ctx, child);
7005 * If the last event was a group event, it will have appended all
7006 * its siblings to the list, but we obtained 'tmp' before that which
7007 * will still point to the list head terminating the iteration.
7009 if (!list_empty(&child_ctx->pinned_groups) ||
7010 !list_empty(&child_ctx->flexible_groups))
7013 mutex_unlock(&child_ctx->mutex);
7019 * When a child task exits, feed back event values to parent events.
7021 void perf_event_exit_task(struct task_struct *child)
7023 struct perf_event *event, *tmp;
7026 mutex_lock(&child->perf_event_mutex);
7027 list_for_each_entry_safe(event, tmp, &child->perf_event_list,
7029 list_del_init(&event->owner_entry);
7032 * Ensure the list deletion is visible before we clear
7033 * the owner, closes a race against perf_release() where
7034 * we need to serialize on the owner->perf_event_mutex.
7037 event->owner = NULL;
7039 mutex_unlock(&child->perf_event_mutex);
7041 for_each_task_context_nr(ctxn)
7042 perf_event_exit_task_context(child, ctxn);
7045 static void perf_free_event(struct perf_event *event,
7046 struct perf_event_context *ctx)
7048 struct perf_event *parent = event->parent;
7050 if (WARN_ON_ONCE(!parent))
7053 mutex_lock(&parent->child_mutex);
7054 list_del_init(&event->child_list);
7055 mutex_unlock(&parent->child_mutex);
7059 perf_group_detach(event);
7060 list_del_event(event, ctx);
7065 * free an unexposed, unused context as created by inheritance by
7066 * perf_event_init_task below, used by fork() in case of fail.
7068 void perf_event_free_task(struct task_struct *task)
7070 struct perf_event_context *ctx;
7071 struct perf_event *event, *tmp;
7074 for_each_task_context_nr(ctxn) {
7075 ctx = task->perf_event_ctxp[ctxn];
7079 mutex_lock(&ctx->mutex);
7081 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups,
7083 perf_free_event(event, ctx);
7085 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups,
7087 perf_free_event(event, ctx);
7089 if (!list_empty(&ctx->pinned_groups) ||
7090 !list_empty(&ctx->flexible_groups))
7093 mutex_unlock(&ctx->mutex);
7099 void perf_event_delayed_put(struct task_struct *task)
7103 for_each_task_context_nr(ctxn)
7104 WARN_ON_ONCE(task->perf_event_ctxp[ctxn]);
7108 * inherit a event from parent task to child task:
7110 static struct perf_event *
7111 inherit_event(struct perf_event *parent_event,
7112 struct task_struct *parent,
7113 struct perf_event_context *parent_ctx,
7114 struct task_struct *child,
7115 struct perf_event *group_leader,
7116 struct perf_event_context *child_ctx)
7118 struct perf_event *child_event;
7119 unsigned long flags;
7122 * Instead of creating recursive hierarchies of events,
7123 * we link inherited events back to the original parent,
7124 * which has a filp for sure, which we use as the reference
7127 if (parent_event->parent)
7128 parent_event = parent_event->parent;
7130 child_event = perf_event_alloc(&parent_event->attr,
7133 group_leader, parent_event,
7135 if (IS_ERR(child_event))
7138 if (!atomic_long_inc_not_zero(&parent_event->refcount)) {
7139 free_event(child_event);
7146 * Make the child state follow the state of the parent event,
7147 * not its attr.disabled bit. We hold the parent's mutex,
7148 * so we won't race with perf_event_{en, dis}able_family.
7150 if (parent_event->state >= PERF_EVENT_STATE_INACTIVE)
7151 child_event->state = PERF_EVENT_STATE_INACTIVE;
7153 child_event->state = PERF_EVENT_STATE_OFF;
7155 if (parent_event->attr.freq) {
7156 u64 sample_period = parent_event->hw.sample_period;
7157 struct hw_perf_event *hwc = &child_event->hw;
7159 hwc->sample_period = sample_period;
7160 hwc->last_period = sample_period;
7162 local64_set(&hwc->period_left, sample_period);
7165 child_event->ctx = child_ctx;
7166 child_event->overflow_handler = parent_event->overflow_handler;
7167 child_event->overflow_handler_context
7168 = parent_event->overflow_handler_context;
7171 * Precalculate sample_data sizes
7173 perf_event__header_size(child_event);
7174 perf_event__id_header_size(child_event);
7177 * Link it up in the child's context:
7179 raw_spin_lock_irqsave(&child_ctx->lock, flags);
7180 add_event_to_ctx(child_event, child_ctx);
7181 raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
7184 * Link this into the parent event's child list
7186 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
7187 mutex_lock(&parent_event->child_mutex);
7188 list_add_tail(&child_event->child_list, &parent_event->child_list);
7189 mutex_unlock(&parent_event->child_mutex);
7194 static int inherit_group(struct perf_event *parent_event,
7195 struct task_struct *parent,
7196 struct perf_event_context *parent_ctx,
7197 struct task_struct *child,
7198 struct perf_event_context *child_ctx)
7200 struct perf_event *leader;
7201 struct perf_event *sub;
7202 struct perf_event *child_ctr;
7204 leader = inherit_event(parent_event, parent, parent_ctx,
7205 child, NULL, child_ctx);
7207 return PTR_ERR(leader);
7208 list_for_each_entry(sub, &parent_event->sibling_list, group_entry) {
7209 child_ctr = inherit_event(sub, parent, parent_ctx,
7210 child, leader, child_ctx);
7211 if (IS_ERR(child_ctr))
7212 return PTR_ERR(child_ctr);
7218 inherit_task_group(struct perf_event *event, struct task_struct *parent,
7219 struct perf_event_context *parent_ctx,
7220 struct task_struct *child, int ctxn,
7224 struct perf_event_context *child_ctx;
7226 if (!event->attr.inherit) {
7231 child_ctx = child->perf_event_ctxp[ctxn];
7234 * This is executed from the parent task context, so
7235 * inherit events that have been marked for cloning.
7236 * First allocate and initialize a context for the
7240 child_ctx = alloc_perf_context(parent_ctx->pmu, child);
7244 child->perf_event_ctxp[ctxn] = child_ctx;
7247 ret = inherit_group(event, parent, parent_ctx,
7257 * Initialize the perf_event context in task_struct
7259 int perf_event_init_context(struct task_struct *child, int ctxn)
7261 struct perf_event_context *child_ctx, *parent_ctx;
7262 struct perf_event_context *cloned_ctx;
7263 struct perf_event *event;
7264 struct task_struct *parent = current;
7265 int inherited_all = 1;
7266 unsigned long flags;
7269 if (likely(!parent->perf_event_ctxp[ctxn]))
7273 * If the parent's context is a clone, pin it so it won't get
7276 parent_ctx = perf_pin_task_context(parent, ctxn);
7279 * No need to check if parent_ctx != NULL here; since we saw
7280 * it non-NULL earlier, the only reason for it to become NULL
7281 * is if we exit, and since we're currently in the middle of
7282 * a fork we can't be exiting at the same time.
7286 * Lock the parent list. No need to lock the child - not PID
7287 * hashed yet and not running, so nobody can access it.
7289 mutex_lock(&parent_ctx->mutex);
7292 * We dont have to disable NMIs - we are only looking at
7293 * the list, not manipulating it:
7295 list_for_each_entry(event, &parent_ctx->pinned_groups, group_entry) {
7296 ret = inherit_task_group(event, parent, parent_ctx,
7297 child, ctxn, &inherited_all);
7303 * We can't hold ctx->lock when iterating the ->flexible_group list due
7304 * to allocations, but we need to prevent rotation because
7305 * rotate_ctx() will change the list from interrupt context.
7307 raw_spin_lock_irqsave(&parent_ctx->lock, flags);
7308 parent_ctx->rotate_disable = 1;
7309 raw_spin_unlock_irqrestore(&parent_ctx->lock, flags);
7311 list_for_each_entry(event, &parent_ctx->flexible_groups, group_entry) {
7312 ret = inherit_task_group(event, parent, parent_ctx,
7313 child, ctxn, &inherited_all);
7318 raw_spin_lock_irqsave(&parent_ctx->lock, flags);
7319 parent_ctx->rotate_disable = 0;
7321 child_ctx = child->perf_event_ctxp[ctxn];
7323 if (child_ctx && inherited_all) {
7325 * Mark the child context as a clone of the parent
7326 * context, or of whatever the parent is a clone of.
7328 * Note that if the parent is a clone, the holding of
7329 * parent_ctx->lock avoids it from being uncloned.
7331 cloned_ctx = parent_ctx->parent_ctx;
7333 child_ctx->parent_ctx = cloned_ctx;
7334 child_ctx->parent_gen = parent_ctx->parent_gen;
7336 child_ctx->parent_ctx = parent_ctx;
7337 child_ctx->parent_gen = parent_ctx->generation;
7339 get_ctx(child_ctx->parent_ctx);
7342 raw_spin_unlock_irqrestore(&parent_ctx->lock, flags);
7343 mutex_unlock(&parent_ctx->mutex);
7345 perf_unpin_context(parent_ctx);
7346 put_ctx(parent_ctx);
7352 * Initialize the perf_event context in task_struct
7354 int perf_event_init_task(struct task_struct *child)
7358 memset(child->perf_event_ctxp, 0, sizeof(child->perf_event_ctxp));
7359 mutex_init(&child->perf_event_mutex);
7360 INIT_LIST_HEAD(&child->perf_event_list);
7362 for_each_task_context_nr(ctxn) {
7363 ret = perf_event_init_context(child, ctxn);
7371 static void __init perf_event_init_all_cpus(void)
7373 struct swevent_htable *swhash;
7376 for_each_possible_cpu(cpu) {
7377 swhash = &per_cpu(swevent_htable, cpu);
7378 mutex_init(&swhash->hlist_mutex);
7379 INIT_LIST_HEAD(&per_cpu(rotation_list, cpu));
7383 static void __cpuinit perf_event_init_cpu(int cpu)
7385 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
7387 mutex_lock(&swhash->hlist_mutex);
7388 if (swhash->hlist_refcount > 0) {
7389 struct swevent_hlist *hlist;
7391 hlist = kzalloc_node(sizeof(*hlist), GFP_KERNEL, cpu_to_node(cpu));
7393 rcu_assign_pointer(swhash->swevent_hlist, hlist);
7395 mutex_unlock(&swhash->hlist_mutex);
7398 #if defined CONFIG_HOTPLUG_CPU || defined CONFIG_KEXEC
7399 static void perf_pmu_rotate_stop(struct pmu *pmu)
7401 struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
7403 WARN_ON(!irqs_disabled());
7405 list_del_init(&cpuctx->rotation_list);
7408 static void __perf_event_exit_context(void *__info)
7410 struct perf_event_context *ctx = __info;
7411 struct perf_event *event, *tmp;
7413 perf_pmu_rotate_stop(ctx->pmu);
7415 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry)
7416 __perf_remove_from_context(event);
7417 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups, group_entry)
7418 __perf_remove_from_context(event);
7421 static void perf_event_exit_cpu_context(int cpu)
7423 struct perf_event_context *ctx;
7427 idx = srcu_read_lock(&pmus_srcu);
7428 list_for_each_entry_rcu(pmu, &pmus, entry) {
7429 ctx = &per_cpu_ptr(pmu->pmu_cpu_context, cpu)->ctx;
7431 mutex_lock(&ctx->mutex);
7432 smp_call_function_single(cpu, __perf_event_exit_context, ctx, 1);
7433 mutex_unlock(&ctx->mutex);
7435 srcu_read_unlock(&pmus_srcu, idx);
7438 static void perf_event_exit_cpu(int cpu)
7440 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
7442 mutex_lock(&swhash->hlist_mutex);
7443 swevent_hlist_release(swhash);
7444 mutex_unlock(&swhash->hlist_mutex);
7446 perf_event_exit_cpu_context(cpu);
7449 static inline void perf_event_exit_cpu(int cpu) { }
7453 perf_reboot(struct notifier_block *notifier, unsigned long val, void *v)
7457 for_each_online_cpu(cpu)
7458 perf_event_exit_cpu(cpu);
7464 * Run the perf reboot notifier at the very last possible moment so that
7465 * the generic watchdog code runs as long as possible.
7467 static struct notifier_block perf_reboot_notifier = {
7468 .notifier_call = perf_reboot,
7469 .priority = INT_MIN,
7472 static int __cpuinit
7473 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
7475 unsigned int cpu = (long)hcpu;
7477 switch (action & ~CPU_TASKS_FROZEN) {
7479 case CPU_UP_PREPARE:
7480 case CPU_DOWN_FAILED:
7481 perf_event_init_cpu(cpu);
7484 case CPU_UP_CANCELED:
7485 case CPU_DOWN_PREPARE:
7486 perf_event_exit_cpu(cpu);
7496 void __init perf_event_init(void)
7502 perf_event_init_all_cpus();
7503 init_srcu_struct(&pmus_srcu);
7504 perf_pmu_register(&perf_swevent, "software", PERF_TYPE_SOFTWARE);
7505 perf_pmu_register(&perf_cpu_clock, NULL, -1);
7506 perf_pmu_register(&perf_task_clock, NULL, -1);
7508 perf_cpu_notifier(perf_cpu_notify);
7509 register_reboot_notifier(&perf_reboot_notifier);
7511 ret = init_hw_breakpoint();
7512 WARN(ret, "hw_breakpoint initialization failed with: %d", ret);
7514 /* do not patch jump label more than once per second */
7515 jump_label_rate_limit(&perf_sched_events, HZ);
7518 * Build time assertion that we keep the data_head at the intended
7519 * location. IOW, validation we got the __reserved[] size right.
7521 BUILD_BUG_ON((offsetof(struct perf_event_mmap_page, data_head))
7525 static int __init perf_event_sysfs_init(void)
7530 mutex_lock(&pmus_lock);
7532 ret = bus_register(&pmu_bus);
7536 list_for_each_entry(pmu, &pmus, entry) {
7537 if (!pmu->name || pmu->type < 0)
7540 ret = pmu_dev_alloc(pmu);
7541 WARN(ret, "Failed to register pmu: %s, reason %d\n", pmu->name, ret);
7543 pmu_bus_running = 1;
7547 mutex_unlock(&pmus_lock);
7551 device_initcall(perf_event_sysfs_init);
7553 #ifdef CONFIG_CGROUP_PERF
7554 static struct cgroup_subsys_state *perf_cgroup_css_alloc(struct cgroup *cont)
7556 struct perf_cgroup *jc;
7558 jc = kzalloc(sizeof(*jc), GFP_KERNEL);
7560 return ERR_PTR(-ENOMEM);
7562 jc->info = alloc_percpu(struct perf_cgroup_info);
7565 return ERR_PTR(-ENOMEM);
7571 static void perf_cgroup_css_free(struct cgroup *cont)
7573 struct perf_cgroup *jc;
7574 jc = container_of(cgroup_subsys_state(cont, perf_subsys_id),
7575 struct perf_cgroup, css);
7576 free_percpu(jc->info);
7580 static int __perf_cgroup_move(void *info)
7582 struct task_struct *task = info;
7583 perf_cgroup_switch(task, PERF_CGROUP_SWOUT | PERF_CGROUP_SWIN);
7587 static void perf_cgroup_attach(struct cgroup *cgrp, struct cgroup_taskset *tset)
7589 struct task_struct *task;
7591 cgroup_taskset_for_each(task, cgrp, tset)
7592 task_function_call(task, __perf_cgroup_move, task);
7595 static void perf_cgroup_exit(struct cgroup *cgrp, struct cgroup *old_cgrp,
7596 struct task_struct *task)
7599 * cgroup_exit() is called in the copy_process() failure path.
7600 * Ignore this case since the task hasn't ran yet, this avoids
7601 * trying to poke a half freed task state from generic code.
7603 if (!(task->flags & PF_EXITING))
7606 task_function_call(task, __perf_cgroup_move, task);
7609 struct cgroup_subsys perf_subsys = {
7610 .name = "perf_event",
7611 .subsys_id = perf_subsys_id,
7612 .css_alloc = perf_cgroup_css_alloc,
7613 .css_free = perf_cgroup_css_free,
7614 .exit = perf_cgroup_exit,
7615 .attach = perf_cgroup_attach,
7617 #endif /* CONFIG_CGROUP_PERF */