2 * Performance events core code:
4 * Copyright (C) 2008 Thomas Gleixner <tglx@linutronix.de>
5 * Copyright (C) 2008-2011 Red Hat, Inc., Ingo Molnar
6 * Copyright (C) 2008-2011 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
7 * Copyright © 2009 Paul Mackerras, IBM Corp. <paulus@au1.ibm.com>
9 * For licensing details see kernel-base/COPYING
14 #include <linux/cpu.h>
15 #include <linux/smp.h>
16 #include <linux/idr.h>
17 #include <linux/file.h>
18 #include <linux/poll.h>
19 #include <linux/slab.h>
20 #include <linux/hash.h>
21 #include <linux/sysfs.h>
22 #include <linux/dcache.h>
23 #include <linux/percpu.h>
24 #include <linux/ptrace.h>
25 #include <linux/reboot.h>
26 #include <linux/vmstat.h>
27 #include <linux/device.h>
28 #include <linux/export.h>
29 #include <linux/vmalloc.h>
30 #include <linux/hardirq.h>
31 #include <linux/rculist.h>
32 #include <linux/uaccess.h>
33 #include <linux/syscalls.h>
34 #include <linux/anon_inodes.h>
35 #include <linux/kernel_stat.h>
36 #include <linux/perf_event.h>
37 #include <linux/ftrace_event.h>
38 #include <linux/hw_breakpoint.h>
39 #include <linux/mm_types.h>
43 #include <asm/irq_regs.h>
45 struct remote_function_call {
46 struct task_struct *p;
47 int (*func)(void *info);
52 static void remote_function(void *data)
54 struct remote_function_call *tfc = data;
55 struct task_struct *p = tfc->p;
59 if (task_cpu(p) != smp_processor_id() || !task_curr(p))
63 tfc->ret = tfc->func(tfc->info);
67 * task_function_call - call a function on the cpu on which a task runs
68 * @p: the task to evaluate
69 * @func: the function to be called
70 * @info: the function call argument
72 * Calls the function @func when the task is currently running. This might
73 * be on the current CPU, which just calls the function directly
75 * returns: @func return value, or
76 * -ESRCH - when the process isn't running
77 * -EAGAIN - when the process moved away
80 task_function_call(struct task_struct *p, int (*func) (void *info), void *info)
82 struct remote_function_call data = {
86 .ret = -ESRCH, /* No such (running) process */
90 smp_call_function_single(task_cpu(p), remote_function, &data, 1);
96 * cpu_function_call - call a function on the cpu
97 * @func: the function to be called
98 * @info: the function call argument
100 * Calls the function @func on the remote cpu.
102 * returns: @func return value or -ENXIO when the cpu is offline
104 static int cpu_function_call(int cpu, int (*func) (void *info), void *info)
106 struct remote_function_call data = {
110 .ret = -ENXIO, /* No such CPU */
113 smp_call_function_single(cpu, remote_function, &data, 1);
118 #define PERF_FLAG_ALL (PERF_FLAG_FD_NO_GROUP |\
119 PERF_FLAG_FD_OUTPUT |\
120 PERF_FLAG_PID_CGROUP)
123 * branch priv levels that need permission checks
125 #define PERF_SAMPLE_BRANCH_PERM_PLM \
126 (PERF_SAMPLE_BRANCH_KERNEL |\
127 PERF_SAMPLE_BRANCH_HV)
130 EVENT_FLEXIBLE = 0x1,
132 EVENT_ALL = EVENT_FLEXIBLE | EVENT_PINNED,
136 * perf_sched_events : >0 events exist
137 * perf_cgroup_events: >0 per-cpu cgroup events exist on this cpu
139 struct static_key_deferred perf_sched_events __read_mostly;
140 static DEFINE_PER_CPU(atomic_t, perf_cgroup_events);
141 static DEFINE_PER_CPU(atomic_t, perf_branch_stack_events);
143 static atomic_t nr_mmap_events __read_mostly;
144 static atomic_t nr_comm_events __read_mostly;
145 static atomic_t nr_task_events __read_mostly;
147 static LIST_HEAD(pmus);
148 static DEFINE_MUTEX(pmus_lock);
149 static struct srcu_struct pmus_srcu;
152 * perf event paranoia level:
153 * -1 - not paranoid at all
154 * 0 - disallow raw tracepoint access for unpriv
155 * 1 - disallow cpu events for unpriv
156 * 2 - disallow kernel profiling for unpriv
158 int sysctl_perf_event_paranoid __read_mostly = 1;
160 /* Minimum for 512 kiB + 1 user control page */
161 int sysctl_perf_event_mlock __read_mostly = 512 + (PAGE_SIZE / 1024); /* 'free' kiB per user */
164 * max perf event sample rate
166 #define DEFAULT_MAX_SAMPLE_RATE 100000
167 int sysctl_perf_event_sample_rate __read_mostly = DEFAULT_MAX_SAMPLE_RATE;
168 static int max_samples_per_tick __read_mostly =
169 DIV_ROUND_UP(DEFAULT_MAX_SAMPLE_RATE, HZ);
171 int perf_proc_update_handler(struct ctl_table *table, int write,
172 void __user *buffer, size_t *lenp,
175 int ret = proc_dointvec(table, write, buffer, lenp, ppos);
180 max_samples_per_tick = DIV_ROUND_UP(sysctl_perf_event_sample_rate, HZ);
185 static atomic64_t perf_event_id;
187 static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
188 enum event_type_t event_type);
190 static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
191 enum event_type_t event_type,
192 struct task_struct *task);
194 static void update_context_time(struct perf_event_context *ctx);
195 static u64 perf_event_time(struct perf_event *event);
197 static void ring_buffer_attach(struct perf_event *event,
198 struct ring_buffer *rb);
200 void __weak perf_event_print_debug(void) { }
202 extern __weak const char *perf_pmu_name(void)
207 static inline u64 perf_clock(void)
209 return local_clock();
212 static inline struct perf_cpu_context *
213 __get_cpu_context(struct perf_event_context *ctx)
215 return this_cpu_ptr(ctx->pmu->pmu_cpu_context);
218 static void perf_ctx_lock(struct perf_cpu_context *cpuctx,
219 struct perf_event_context *ctx)
221 raw_spin_lock(&cpuctx->ctx.lock);
223 raw_spin_lock(&ctx->lock);
226 static void perf_ctx_unlock(struct perf_cpu_context *cpuctx,
227 struct perf_event_context *ctx)
230 raw_spin_unlock(&ctx->lock);
231 raw_spin_unlock(&cpuctx->ctx.lock);
234 #ifdef CONFIG_CGROUP_PERF
237 * Must ensure cgroup is pinned (css_get) before calling
238 * this function. In other words, we cannot call this function
239 * if there is no cgroup event for the current CPU context.
241 static inline struct perf_cgroup *
242 perf_cgroup_from_task(struct task_struct *task)
244 return container_of(task_subsys_state(task, perf_subsys_id),
245 struct perf_cgroup, css);
249 perf_cgroup_match(struct perf_event *event)
251 struct perf_event_context *ctx = event->ctx;
252 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
254 return !event->cgrp || event->cgrp == cpuctx->cgrp;
257 static inline bool perf_tryget_cgroup(struct perf_event *event)
259 return css_tryget(&event->cgrp->css);
262 static inline void perf_put_cgroup(struct perf_event *event)
264 css_put(&event->cgrp->css);
267 static inline void perf_detach_cgroup(struct perf_event *event)
269 perf_put_cgroup(event);
273 static inline int is_cgroup_event(struct perf_event *event)
275 return event->cgrp != NULL;
278 static inline u64 perf_cgroup_event_time(struct perf_event *event)
280 struct perf_cgroup_info *t;
282 t = per_cpu_ptr(event->cgrp->info, event->cpu);
286 static inline void __update_cgrp_time(struct perf_cgroup *cgrp)
288 struct perf_cgroup_info *info;
293 info = this_cpu_ptr(cgrp->info);
295 info->time += now - info->timestamp;
296 info->timestamp = now;
299 static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx)
301 struct perf_cgroup *cgrp_out = cpuctx->cgrp;
303 __update_cgrp_time(cgrp_out);
306 static inline void update_cgrp_time_from_event(struct perf_event *event)
308 struct perf_cgroup *cgrp;
311 * ensure we access cgroup data only when needed and
312 * when we know the cgroup is pinned (css_get)
314 if (!is_cgroup_event(event))
317 cgrp = perf_cgroup_from_task(current);
319 * Do not update time when cgroup is not active
321 if (cgrp == event->cgrp)
322 __update_cgrp_time(event->cgrp);
326 perf_cgroup_set_timestamp(struct task_struct *task,
327 struct perf_event_context *ctx)
329 struct perf_cgroup *cgrp;
330 struct perf_cgroup_info *info;
333 * ctx->lock held by caller
334 * ensure we do not access cgroup data
335 * unless we have the cgroup pinned (css_get)
337 if (!task || !ctx->nr_cgroups)
340 cgrp = perf_cgroup_from_task(task);
341 info = this_cpu_ptr(cgrp->info);
342 info->timestamp = ctx->timestamp;
345 #define PERF_CGROUP_SWOUT 0x1 /* cgroup switch out every event */
346 #define PERF_CGROUP_SWIN 0x2 /* cgroup switch in events based on task */
349 * reschedule events based on the cgroup constraint of task.
351 * mode SWOUT : schedule out everything
352 * mode SWIN : schedule in based on cgroup for next
354 void perf_cgroup_switch(struct task_struct *task, int mode)
356 struct perf_cpu_context *cpuctx;
361 * disable interrupts to avoid geting nr_cgroup
362 * changes via __perf_event_disable(). Also
365 local_irq_save(flags);
368 * we reschedule only in the presence of cgroup
369 * constrained events.
373 list_for_each_entry_rcu(pmu, &pmus, entry) {
374 cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
377 * perf_cgroup_events says at least one
378 * context on this CPU has cgroup events.
380 * ctx->nr_cgroups reports the number of cgroup
381 * events for a context.
383 if (cpuctx->ctx.nr_cgroups > 0) {
384 perf_ctx_lock(cpuctx, cpuctx->task_ctx);
385 perf_pmu_disable(cpuctx->ctx.pmu);
387 if (mode & PERF_CGROUP_SWOUT) {
388 cpu_ctx_sched_out(cpuctx, EVENT_ALL);
390 * must not be done before ctxswout due
391 * to event_filter_match() in event_sched_out()
396 if (mode & PERF_CGROUP_SWIN) {
397 WARN_ON_ONCE(cpuctx->cgrp);
398 /* set cgrp before ctxsw in to
399 * allow event_filter_match() to not
400 * have to pass task around
402 cpuctx->cgrp = perf_cgroup_from_task(task);
403 cpu_ctx_sched_in(cpuctx, EVENT_ALL, task);
405 perf_pmu_enable(cpuctx->ctx.pmu);
406 perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
412 local_irq_restore(flags);
415 static inline void perf_cgroup_sched_out(struct task_struct *task,
416 struct task_struct *next)
418 struct perf_cgroup *cgrp1;
419 struct perf_cgroup *cgrp2 = NULL;
422 * we come here when we know perf_cgroup_events > 0
424 cgrp1 = perf_cgroup_from_task(task);
427 * next is NULL when called from perf_event_enable_on_exec()
428 * that will systematically cause a cgroup_switch()
431 cgrp2 = perf_cgroup_from_task(next);
434 * only schedule out current cgroup events if we know
435 * that we are switching to a different cgroup. Otherwise,
436 * do no touch the cgroup events.
439 perf_cgroup_switch(task, PERF_CGROUP_SWOUT);
442 static inline void perf_cgroup_sched_in(struct task_struct *prev,
443 struct task_struct *task)
445 struct perf_cgroup *cgrp1;
446 struct perf_cgroup *cgrp2 = NULL;
449 * we come here when we know perf_cgroup_events > 0
451 cgrp1 = perf_cgroup_from_task(task);
453 /* prev can never be NULL */
454 cgrp2 = perf_cgroup_from_task(prev);
457 * only need to schedule in cgroup events if we are changing
458 * cgroup during ctxsw. Cgroup events were not scheduled
459 * out of ctxsw out if that was not the case.
462 perf_cgroup_switch(task, PERF_CGROUP_SWIN);
465 static inline int perf_cgroup_connect(int fd, struct perf_event *event,
466 struct perf_event_attr *attr,
467 struct perf_event *group_leader)
469 struct perf_cgroup *cgrp;
470 struct cgroup_subsys_state *css;
472 int ret = 0, fput_needed;
474 file = fget_light(fd, &fput_needed);
478 css = cgroup_css_from_dir(file, perf_subsys_id);
484 cgrp = container_of(css, struct perf_cgroup, css);
487 /* must be done before we fput() the file */
488 if (!perf_tryget_cgroup(event)) {
495 * all events in a group must monitor
496 * the same cgroup because a task belongs
497 * to only one perf cgroup at a time
499 if (group_leader && group_leader->cgrp != cgrp) {
500 perf_detach_cgroup(event);
504 fput_light(file, fput_needed);
509 perf_cgroup_set_shadow_time(struct perf_event *event, u64 now)
511 struct perf_cgroup_info *t;
512 t = per_cpu_ptr(event->cgrp->info, event->cpu);
513 event->shadow_ctx_time = now - t->timestamp;
517 perf_cgroup_defer_enabled(struct perf_event *event)
520 * when the current task's perf cgroup does not match
521 * the event's, we need to remember to call the
522 * perf_mark_enable() function the first time a task with
523 * a matching perf cgroup is scheduled in.
525 if (is_cgroup_event(event) && !perf_cgroup_match(event))
526 event->cgrp_defer_enabled = 1;
530 perf_cgroup_mark_enabled(struct perf_event *event,
531 struct perf_event_context *ctx)
533 struct perf_event *sub;
534 u64 tstamp = perf_event_time(event);
536 if (!event->cgrp_defer_enabled)
539 event->cgrp_defer_enabled = 0;
541 event->tstamp_enabled = tstamp - event->total_time_enabled;
542 list_for_each_entry(sub, &event->sibling_list, group_entry) {
543 if (sub->state >= PERF_EVENT_STATE_INACTIVE) {
544 sub->tstamp_enabled = tstamp - sub->total_time_enabled;
545 sub->cgrp_defer_enabled = 0;
549 #else /* !CONFIG_CGROUP_PERF */
552 perf_cgroup_match(struct perf_event *event)
557 static inline void perf_detach_cgroup(struct perf_event *event)
560 static inline int is_cgroup_event(struct perf_event *event)
565 static inline u64 perf_cgroup_event_cgrp_time(struct perf_event *event)
570 static inline void update_cgrp_time_from_event(struct perf_event *event)
574 static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx)
578 static inline void perf_cgroup_sched_out(struct task_struct *task,
579 struct task_struct *next)
583 static inline void perf_cgroup_sched_in(struct task_struct *prev,
584 struct task_struct *task)
588 static inline int perf_cgroup_connect(pid_t pid, struct perf_event *event,
589 struct perf_event_attr *attr,
590 struct perf_event *group_leader)
596 perf_cgroup_set_timestamp(struct task_struct *task,
597 struct perf_event_context *ctx)
602 perf_cgroup_switch(struct task_struct *task, struct task_struct *next)
607 perf_cgroup_set_shadow_time(struct perf_event *event, u64 now)
611 static inline u64 perf_cgroup_event_time(struct perf_event *event)
617 perf_cgroup_defer_enabled(struct perf_event *event)
622 perf_cgroup_mark_enabled(struct perf_event *event,
623 struct perf_event_context *ctx)
628 void perf_pmu_disable(struct pmu *pmu)
630 int *count = this_cpu_ptr(pmu->pmu_disable_count);
632 pmu->pmu_disable(pmu);
635 void perf_pmu_enable(struct pmu *pmu)
637 int *count = this_cpu_ptr(pmu->pmu_disable_count);
639 pmu->pmu_enable(pmu);
642 static DEFINE_PER_CPU(struct list_head, rotation_list);
645 * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
646 * because they're strictly cpu affine and rotate_start is called with IRQs
647 * disabled, while rotate_context is called from IRQ context.
649 static void perf_pmu_rotate_start(struct pmu *pmu)
651 struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
652 struct list_head *head = &__get_cpu_var(rotation_list);
654 WARN_ON(!irqs_disabled());
656 if (list_empty(&cpuctx->rotation_list))
657 list_add(&cpuctx->rotation_list, head);
660 static void get_ctx(struct perf_event_context *ctx)
662 WARN_ON(!atomic_inc_not_zero(&ctx->refcount));
665 static void put_ctx(struct perf_event_context *ctx)
667 if (atomic_dec_and_test(&ctx->refcount)) {
669 put_ctx(ctx->parent_ctx);
671 put_task_struct(ctx->task);
672 kfree_rcu(ctx, rcu_head);
676 static void unclone_ctx(struct perf_event_context *ctx)
678 if (ctx->parent_ctx) {
679 put_ctx(ctx->parent_ctx);
680 ctx->parent_ctx = NULL;
684 static u32 perf_event_pid(struct perf_event *event, struct task_struct *p)
687 * only top level events have the pid namespace they were created in
690 event = event->parent;
692 return task_tgid_nr_ns(p, event->ns);
695 static u32 perf_event_tid(struct perf_event *event, struct task_struct *p)
698 * only top level events have the pid namespace they were created in
701 event = event->parent;
703 return task_pid_nr_ns(p, event->ns);
707 * If we inherit events we want to return the parent event id
710 static u64 primary_event_id(struct perf_event *event)
715 id = event->parent->id;
721 * Get the perf_event_context for a task and lock it.
722 * This has to cope with with the fact that until it is locked,
723 * the context could get moved to another task.
725 static struct perf_event_context *
726 perf_lock_task_context(struct task_struct *task, int ctxn, unsigned long *flags)
728 struct perf_event_context *ctx;
732 ctx = rcu_dereference(task->perf_event_ctxp[ctxn]);
735 * If this context is a clone of another, it might
736 * get swapped for another underneath us by
737 * perf_event_task_sched_out, though the
738 * rcu_read_lock() protects us from any context
739 * getting freed. Lock the context and check if it
740 * got swapped before we could get the lock, and retry
741 * if so. If we locked the right context, then it
742 * can't get swapped on us any more.
744 raw_spin_lock_irqsave(&ctx->lock, *flags);
745 if (ctx != rcu_dereference(task->perf_event_ctxp[ctxn])) {
746 raw_spin_unlock_irqrestore(&ctx->lock, *flags);
750 if (!atomic_inc_not_zero(&ctx->refcount)) {
751 raw_spin_unlock_irqrestore(&ctx->lock, *flags);
760 * Get the context for a task and increment its pin_count so it
761 * can't get swapped to another task. This also increments its
762 * reference count so that the context can't get freed.
764 static struct perf_event_context *
765 perf_pin_task_context(struct task_struct *task, int ctxn)
767 struct perf_event_context *ctx;
770 ctx = perf_lock_task_context(task, ctxn, &flags);
773 raw_spin_unlock_irqrestore(&ctx->lock, flags);
778 static void perf_unpin_context(struct perf_event_context *ctx)
782 raw_spin_lock_irqsave(&ctx->lock, flags);
784 raw_spin_unlock_irqrestore(&ctx->lock, flags);
788 * Update the record of the current time in a context.
790 static void update_context_time(struct perf_event_context *ctx)
792 u64 now = perf_clock();
794 ctx->time += now - ctx->timestamp;
795 ctx->timestamp = now;
798 static u64 perf_event_time(struct perf_event *event)
800 struct perf_event_context *ctx = event->ctx;
802 if (is_cgroup_event(event))
803 return perf_cgroup_event_time(event);
805 return ctx ? ctx->time : 0;
809 * Update the total_time_enabled and total_time_running fields for a event.
810 * The caller of this function needs to hold the ctx->lock.
812 static void update_event_times(struct perf_event *event)
814 struct perf_event_context *ctx = event->ctx;
817 if (event->state < PERF_EVENT_STATE_INACTIVE ||
818 event->group_leader->state < PERF_EVENT_STATE_INACTIVE)
821 * in cgroup mode, time_enabled represents
822 * the time the event was enabled AND active
823 * tasks were in the monitored cgroup. This is
824 * independent of the activity of the context as
825 * there may be a mix of cgroup and non-cgroup events.
827 * That is why we treat cgroup events differently
830 if (is_cgroup_event(event))
831 run_end = perf_cgroup_event_time(event);
832 else if (ctx->is_active)
835 run_end = event->tstamp_stopped;
837 event->total_time_enabled = run_end - event->tstamp_enabled;
839 if (event->state == PERF_EVENT_STATE_INACTIVE)
840 run_end = event->tstamp_stopped;
842 run_end = perf_event_time(event);
844 event->total_time_running = run_end - event->tstamp_running;
849 * Update total_time_enabled and total_time_running for all events in a group.
851 static void update_group_times(struct perf_event *leader)
853 struct perf_event *event;
855 update_event_times(leader);
856 list_for_each_entry(event, &leader->sibling_list, group_entry)
857 update_event_times(event);
860 static struct list_head *
861 ctx_group_list(struct perf_event *event, struct perf_event_context *ctx)
863 if (event->attr.pinned)
864 return &ctx->pinned_groups;
866 return &ctx->flexible_groups;
870 * Add a event from the lists for its context.
871 * Must be called with ctx->mutex and ctx->lock held.
874 list_add_event(struct perf_event *event, struct perf_event_context *ctx)
876 WARN_ON_ONCE(event->attach_state & PERF_ATTACH_CONTEXT);
877 event->attach_state |= PERF_ATTACH_CONTEXT;
880 * If we're a stand alone event or group leader, we go to the context
881 * list, group events are kept attached to the group so that
882 * perf_group_detach can, at all times, locate all siblings.
884 if (event->group_leader == event) {
885 struct list_head *list;
887 if (is_software_event(event))
888 event->group_flags |= PERF_GROUP_SOFTWARE;
890 list = ctx_group_list(event, ctx);
891 list_add_tail(&event->group_entry, list);
894 if (is_cgroup_event(event))
897 if (has_branch_stack(event))
898 ctx->nr_branch_stack++;
900 list_add_rcu(&event->event_entry, &ctx->event_list);
902 perf_pmu_rotate_start(ctx->pmu);
904 if (event->attr.inherit_stat)
909 * Called at perf_event creation and when events are attached/detached from a
912 static void perf_event__read_size(struct perf_event *event)
914 int entry = sizeof(u64); /* value */
918 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
921 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
924 if (event->attr.read_format & PERF_FORMAT_ID)
925 entry += sizeof(u64);
927 if (event->attr.read_format & PERF_FORMAT_GROUP) {
928 nr += event->group_leader->nr_siblings;
933 event->read_size = size;
936 static void perf_event__header_size(struct perf_event *event)
938 struct perf_sample_data *data;
939 u64 sample_type = event->attr.sample_type;
942 perf_event__read_size(event);
944 if (sample_type & PERF_SAMPLE_IP)
945 size += sizeof(data->ip);
947 if (sample_type & PERF_SAMPLE_ADDR)
948 size += sizeof(data->addr);
950 if (sample_type & PERF_SAMPLE_PERIOD)
951 size += sizeof(data->period);
953 if (sample_type & PERF_SAMPLE_READ)
954 size += event->read_size;
956 event->header_size = size;
959 static void perf_event__id_header_size(struct perf_event *event)
961 struct perf_sample_data *data;
962 u64 sample_type = event->attr.sample_type;
965 if (sample_type & PERF_SAMPLE_TID)
966 size += sizeof(data->tid_entry);
968 if (sample_type & PERF_SAMPLE_TIME)
969 size += sizeof(data->time);
971 if (sample_type & PERF_SAMPLE_ID)
972 size += sizeof(data->id);
974 if (sample_type & PERF_SAMPLE_STREAM_ID)
975 size += sizeof(data->stream_id);
977 if (sample_type & PERF_SAMPLE_CPU)
978 size += sizeof(data->cpu_entry);
980 event->id_header_size = size;
983 static void perf_group_attach(struct perf_event *event)
985 struct perf_event *group_leader = event->group_leader, *pos;
988 * We can have double attach due to group movement in perf_event_open.
990 if (event->attach_state & PERF_ATTACH_GROUP)
993 event->attach_state |= PERF_ATTACH_GROUP;
995 if (group_leader == event)
998 if (group_leader->group_flags & PERF_GROUP_SOFTWARE &&
999 !is_software_event(event))
1000 group_leader->group_flags &= ~PERF_GROUP_SOFTWARE;
1002 list_add_tail(&event->group_entry, &group_leader->sibling_list);
1003 group_leader->nr_siblings++;
1005 perf_event__header_size(group_leader);
1007 list_for_each_entry(pos, &group_leader->sibling_list, group_entry)
1008 perf_event__header_size(pos);
1012 * Remove a event from the lists for its context.
1013 * Must be called with ctx->mutex and ctx->lock held.
1016 list_del_event(struct perf_event *event, struct perf_event_context *ctx)
1018 struct perf_cpu_context *cpuctx;
1020 * We can have double detach due to exit/hot-unplug + close.
1022 if (!(event->attach_state & PERF_ATTACH_CONTEXT))
1025 event->attach_state &= ~PERF_ATTACH_CONTEXT;
1027 if (is_cgroup_event(event)) {
1029 cpuctx = __get_cpu_context(ctx);
1031 * if there are no more cgroup events
1032 * then cler cgrp to avoid stale pointer
1033 * in update_cgrp_time_from_cpuctx()
1035 if (!ctx->nr_cgroups)
1036 cpuctx->cgrp = NULL;
1039 if (has_branch_stack(event))
1040 ctx->nr_branch_stack--;
1043 if (event->attr.inherit_stat)
1046 list_del_rcu(&event->event_entry);
1048 if (event->group_leader == event)
1049 list_del_init(&event->group_entry);
1051 update_group_times(event);
1054 * If event was in error state, then keep it
1055 * that way, otherwise bogus counts will be
1056 * returned on read(). The only way to get out
1057 * of error state is by explicit re-enabling
1060 if (event->state > PERF_EVENT_STATE_OFF)
1061 event->state = PERF_EVENT_STATE_OFF;
1064 static void perf_group_detach(struct perf_event *event)
1066 struct perf_event *sibling, *tmp;
1067 struct list_head *list = NULL;
1070 * We can have double detach due to exit/hot-unplug + close.
1072 if (!(event->attach_state & PERF_ATTACH_GROUP))
1075 event->attach_state &= ~PERF_ATTACH_GROUP;
1078 * If this is a sibling, remove it from its group.
1080 if (event->group_leader != event) {
1081 list_del_init(&event->group_entry);
1082 event->group_leader->nr_siblings--;
1086 if (!list_empty(&event->group_entry))
1087 list = &event->group_entry;
1090 * If this was a group event with sibling events then
1091 * upgrade the siblings to singleton events by adding them
1092 * to whatever list we are on.
1094 list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) {
1096 list_move_tail(&sibling->group_entry, list);
1097 sibling->group_leader = sibling;
1099 /* Inherit group flags from the previous leader */
1100 sibling->group_flags = event->group_flags;
1104 perf_event__header_size(event->group_leader);
1106 list_for_each_entry(tmp, &event->group_leader->sibling_list, group_entry)
1107 perf_event__header_size(tmp);
1111 event_filter_match(struct perf_event *event)
1113 return (event->cpu == -1 || event->cpu == smp_processor_id())
1114 && perf_cgroup_match(event);
1118 event_sched_out(struct perf_event *event,
1119 struct perf_cpu_context *cpuctx,
1120 struct perf_event_context *ctx)
1122 u64 tstamp = perf_event_time(event);
1125 * An event which could not be activated because of
1126 * filter mismatch still needs to have its timings
1127 * maintained, otherwise bogus information is return
1128 * via read() for time_enabled, time_running:
1130 if (event->state == PERF_EVENT_STATE_INACTIVE
1131 && !event_filter_match(event)) {
1132 delta = tstamp - event->tstamp_stopped;
1133 event->tstamp_running += delta;
1134 event->tstamp_stopped = tstamp;
1137 if (event->state != PERF_EVENT_STATE_ACTIVE)
1140 event->state = PERF_EVENT_STATE_INACTIVE;
1141 if (event->pending_disable) {
1142 event->pending_disable = 0;
1143 event->state = PERF_EVENT_STATE_OFF;
1145 event->tstamp_stopped = tstamp;
1146 event->pmu->del(event, 0);
1149 if (!is_software_event(event))
1150 cpuctx->active_oncpu--;
1152 if (event->attr.freq && event->attr.sample_freq)
1154 if (event->attr.exclusive || !cpuctx->active_oncpu)
1155 cpuctx->exclusive = 0;
1159 group_sched_out(struct perf_event *group_event,
1160 struct perf_cpu_context *cpuctx,
1161 struct perf_event_context *ctx)
1163 struct perf_event *event;
1164 int state = group_event->state;
1166 event_sched_out(group_event, cpuctx, ctx);
1169 * Schedule out siblings (if any):
1171 list_for_each_entry(event, &group_event->sibling_list, group_entry)
1172 event_sched_out(event, cpuctx, ctx);
1174 if (state == PERF_EVENT_STATE_ACTIVE && group_event->attr.exclusive)
1175 cpuctx->exclusive = 0;
1179 * Cross CPU call to remove a performance event
1181 * We disable the event on the hardware level first. After that we
1182 * remove it from the context list.
1184 static int __perf_remove_from_context(void *info)
1186 struct perf_event *event = info;
1187 struct perf_event_context *ctx = event->ctx;
1188 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1190 raw_spin_lock(&ctx->lock);
1191 event_sched_out(event, cpuctx, ctx);
1192 list_del_event(event, ctx);
1193 if (!ctx->nr_events && cpuctx->task_ctx == ctx) {
1195 cpuctx->task_ctx = NULL;
1197 raw_spin_unlock(&ctx->lock);
1204 * Remove the event from a task's (or a CPU's) list of events.
1206 * CPU events are removed with a smp call. For task events we only
1207 * call when the task is on a CPU.
1209 * If event->ctx is a cloned context, callers must make sure that
1210 * every task struct that event->ctx->task could possibly point to
1211 * remains valid. This is OK when called from perf_release since
1212 * that only calls us on the top-level context, which can't be a clone.
1213 * When called from perf_event_exit_task, it's OK because the
1214 * context has been detached from its task.
1216 static void perf_remove_from_context(struct perf_event *event)
1218 struct perf_event_context *ctx = event->ctx;
1219 struct task_struct *task = ctx->task;
1221 lockdep_assert_held(&ctx->mutex);
1225 * Per cpu events are removed via an smp call and
1226 * the removal is always successful.
1228 cpu_function_call(event->cpu, __perf_remove_from_context, event);
1233 if (!task_function_call(task, __perf_remove_from_context, event))
1236 raw_spin_lock_irq(&ctx->lock);
1238 * If we failed to find a running task, but find the context active now
1239 * that we've acquired the ctx->lock, retry.
1241 if (ctx->is_active) {
1242 raw_spin_unlock_irq(&ctx->lock);
1247 * Since the task isn't running, its safe to remove the event, us
1248 * holding the ctx->lock ensures the task won't get scheduled in.
1250 list_del_event(event, ctx);
1251 raw_spin_unlock_irq(&ctx->lock);
1255 * Cross CPU call to disable a performance event
1257 static int __perf_event_disable(void *info)
1259 struct perf_event *event = info;
1260 struct perf_event_context *ctx = event->ctx;
1261 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1264 * If this is a per-task event, need to check whether this
1265 * event's task is the current task on this cpu.
1267 * Can trigger due to concurrent perf_event_context_sched_out()
1268 * flipping contexts around.
1270 if (ctx->task && cpuctx->task_ctx != ctx)
1273 raw_spin_lock(&ctx->lock);
1276 * If the event is on, turn it off.
1277 * If it is in error state, leave it in error state.
1279 if (event->state >= PERF_EVENT_STATE_INACTIVE) {
1280 update_context_time(ctx);
1281 update_cgrp_time_from_event(event);
1282 update_group_times(event);
1283 if (event == event->group_leader)
1284 group_sched_out(event, cpuctx, ctx);
1286 event_sched_out(event, cpuctx, ctx);
1287 event->state = PERF_EVENT_STATE_OFF;
1290 raw_spin_unlock(&ctx->lock);
1298 * If event->ctx is a cloned context, callers must make sure that
1299 * every task struct that event->ctx->task could possibly point to
1300 * remains valid. This condition is satisifed when called through
1301 * perf_event_for_each_child or perf_event_for_each because they
1302 * hold the top-level event's child_mutex, so any descendant that
1303 * goes to exit will block in sync_child_event.
1304 * When called from perf_pending_event it's OK because event->ctx
1305 * is the current context on this CPU and preemption is disabled,
1306 * hence we can't get into perf_event_task_sched_out for this context.
1308 void perf_event_disable(struct perf_event *event)
1310 struct perf_event_context *ctx = event->ctx;
1311 struct task_struct *task = ctx->task;
1315 * Disable the event on the cpu that it's on
1317 cpu_function_call(event->cpu, __perf_event_disable, event);
1322 if (!task_function_call(task, __perf_event_disable, event))
1325 raw_spin_lock_irq(&ctx->lock);
1327 * If the event is still active, we need to retry the cross-call.
1329 if (event->state == PERF_EVENT_STATE_ACTIVE) {
1330 raw_spin_unlock_irq(&ctx->lock);
1332 * Reload the task pointer, it might have been changed by
1333 * a concurrent perf_event_context_sched_out().
1340 * Since we have the lock this context can't be scheduled
1341 * in, so we can change the state safely.
1343 if (event->state == PERF_EVENT_STATE_INACTIVE) {
1344 update_group_times(event);
1345 event->state = PERF_EVENT_STATE_OFF;
1347 raw_spin_unlock_irq(&ctx->lock);
1349 EXPORT_SYMBOL_GPL(perf_event_disable);
1351 static void perf_set_shadow_time(struct perf_event *event,
1352 struct perf_event_context *ctx,
1356 * use the correct time source for the time snapshot
1358 * We could get by without this by leveraging the
1359 * fact that to get to this function, the caller
1360 * has most likely already called update_context_time()
1361 * and update_cgrp_time_xx() and thus both timestamp
1362 * are identical (or very close). Given that tstamp is,
1363 * already adjusted for cgroup, we could say that:
1364 * tstamp - ctx->timestamp
1366 * tstamp - cgrp->timestamp.
1368 * Then, in perf_output_read(), the calculation would
1369 * work with no changes because:
1370 * - event is guaranteed scheduled in
1371 * - no scheduled out in between
1372 * - thus the timestamp would be the same
1374 * But this is a bit hairy.
1376 * So instead, we have an explicit cgroup call to remain
1377 * within the time time source all along. We believe it
1378 * is cleaner and simpler to understand.
1380 if (is_cgroup_event(event))
1381 perf_cgroup_set_shadow_time(event, tstamp);
1383 event->shadow_ctx_time = tstamp - ctx->timestamp;
1386 #define MAX_INTERRUPTS (~0ULL)
1388 static void perf_log_throttle(struct perf_event *event, int enable);
1391 event_sched_in(struct perf_event *event,
1392 struct perf_cpu_context *cpuctx,
1393 struct perf_event_context *ctx)
1395 u64 tstamp = perf_event_time(event);
1397 if (event->state <= PERF_EVENT_STATE_OFF)
1400 event->state = PERF_EVENT_STATE_ACTIVE;
1401 event->oncpu = smp_processor_id();
1404 * Unthrottle events, since we scheduled we might have missed several
1405 * ticks already, also for a heavily scheduling task there is little
1406 * guarantee it'll get a tick in a timely manner.
1408 if (unlikely(event->hw.interrupts == MAX_INTERRUPTS)) {
1409 perf_log_throttle(event, 1);
1410 event->hw.interrupts = 0;
1414 * The new state must be visible before we turn it on in the hardware:
1418 if (event->pmu->add(event, PERF_EF_START)) {
1419 event->state = PERF_EVENT_STATE_INACTIVE;
1424 event->tstamp_running += tstamp - event->tstamp_stopped;
1426 perf_set_shadow_time(event, ctx, tstamp);
1428 if (!is_software_event(event))
1429 cpuctx->active_oncpu++;
1431 if (event->attr.freq && event->attr.sample_freq)
1434 if (event->attr.exclusive)
1435 cpuctx->exclusive = 1;
1441 group_sched_in(struct perf_event *group_event,
1442 struct perf_cpu_context *cpuctx,
1443 struct perf_event_context *ctx)
1445 struct perf_event *event, *partial_group = NULL;
1446 struct pmu *pmu = group_event->pmu;
1447 u64 now = ctx->time;
1448 bool simulate = false;
1450 if (group_event->state == PERF_EVENT_STATE_OFF)
1453 pmu->start_txn(pmu);
1455 if (event_sched_in(group_event, cpuctx, ctx)) {
1456 pmu->cancel_txn(pmu);
1461 * Schedule in siblings as one group (if any):
1463 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
1464 if (event_sched_in(event, cpuctx, ctx)) {
1465 partial_group = event;
1470 if (!pmu->commit_txn(pmu))
1475 * Groups can be scheduled in as one unit only, so undo any
1476 * partial group before returning:
1477 * The events up to the failed event are scheduled out normally,
1478 * tstamp_stopped will be updated.
1480 * The failed events and the remaining siblings need to have
1481 * their timings updated as if they had gone thru event_sched_in()
1482 * and event_sched_out(). This is required to get consistent timings
1483 * across the group. This also takes care of the case where the group
1484 * could never be scheduled by ensuring tstamp_stopped is set to mark
1485 * the time the event was actually stopped, such that time delta
1486 * calculation in update_event_times() is correct.
1488 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
1489 if (event == partial_group)
1493 event->tstamp_running += now - event->tstamp_stopped;
1494 event->tstamp_stopped = now;
1496 event_sched_out(event, cpuctx, ctx);
1499 event_sched_out(group_event, cpuctx, ctx);
1501 pmu->cancel_txn(pmu);
1507 * Work out whether we can put this event group on the CPU now.
1509 static int group_can_go_on(struct perf_event *event,
1510 struct perf_cpu_context *cpuctx,
1514 * Groups consisting entirely of software events can always go on.
1516 if (event->group_flags & PERF_GROUP_SOFTWARE)
1519 * If an exclusive group is already on, no other hardware
1522 if (cpuctx->exclusive)
1525 * If this group is exclusive and there are already
1526 * events on the CPU, it can't go on.
1528 if (event->attr.exclusive && cpuctx->active_oncpu)
1531 * Otherwise, try to add it if all previous groups were able
1537 static void add_event_to_ctx(struct perf_event *event,
1538 struct perf_event_context *ctx)
1540 u64 tstamp = perf_event_time(event);
1542 list_add_event(event, ctx);
1543 perf_group_attach(event);
1544 event->tstamp_enabled = tstamp;
1545 event->tstamp_running = tstamp;
1546 event->tstamp_stopped = tstamp;
1549 static void task_ctx_sched_out(struct perf_event_context *ctx);
1551 ctx_sched_in(struct perf_event_context *ctx,
1552 struct perf_cpu_context *cpuctx,
1553 enum event_type_t event_type,
1554 struct task_struct *task);
1556 static void perf_event_sched_in(struct perf_cpu_context *cpuctx,
1557 struct perf_event_context *ctx,
1558 struct task_struct *task)
1560 cpu_ctx_sched_in(cpuctx, EVENT_PINNED, task);
1562 ctx_sched_in(ctx, cpuctx, EVENT_PINNED, task);
1563 cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE, task);
1565 ctx_sched_in(ctx, cpuctx, EVENT_FLEXIBLE, task);
1569 * Cross CPU call to install and enable a performance event
1571 * Must be called with ctx->mutex held
1573 static int __perf_install_in_context(void *info)
1575 struct perf_event *event = info;
1576 struct perf_event_context *ctx = event->ctx;
1577 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1578 struct perf_event_context *task_ctx = cpuctx->task_ctx;
1579 struct task_struct *task = current;
1581 perf_ctx_lock(cpuctx, task_ctx);
1582 perf_pmu_disable(cpuctx->ctx.pmu);
1585 * If there was an active task_ctx schedule it out.
1588 task_ctx_sched_out(task_ctx);
1591 * If the context we're installing events in is not the
1592 * active task_ctx, flip them.
1594 if (ctx->task && task_ctx != ctx) {
1596 raw_spin_unlock(&task_ctx->lock);
1597 raw_spin_lock(&ctx->lock);
1602 cpuctx->task_ctx = task_ctx;
1603 task = task_ctx->task;
1606 cpu_ctx_sched_out(cpuctx, EVENT_ALL);
1608 update_context_time(ctx);
1610 * update cgrp time only if current cgrp
1611 * matches event->cgrp. Must be done before
1612 * calling add_event_to_ctx()
1614 update_cgrp_time_from_event(event);
1616 add_event_to_ctx(event, ctx);
1619 * Schedule everything back in
1621 perf_event_sched_in(cpuctx, task_ctx, task);
1623 perf_pmu_enable(cpuctx->ctx.pmu);
1624 perf_ctx_unlock(cpuctx, task_ctx);
1630 * Attach a performance event to a context
1632 * First we add the event to the list with the hardware enable bit
1633 * in event->hw_config cleared.
1635 * If the event is attached to a task which is on a CPU we use a smp
1636 * call to enable it in the task context. The task might have been
1637 * scheduled away, but we check this in the smp call again.
1640 perf_install_in_context(struct perf_event_context *ctx,
1641 struct perf_event *event,
1644 struct task_struct *task = ctx->task;
1646 lockdep_assert_held(&ctx->mutex);
1649 if (event->cpu != -1)
1654 * Per cpu events are installed via an smp call and
1655 * the install is always successful.
1657 cpu_function_call(cpu, __perf_install_in_context, event);
1662 if (!task_function_call(task, __perf_install_in_context, event))
1665 raw_spin_lock_irq(&ctx->lock);
1667 * If we failed to find a running task, but find the context active now
1668 * that we've acquired the ctx->lock, retry.
1670 if (ctx->is_active) {
1671 raw_spin_unlock_irq(&ctx->lock);
1676 * Since the task isn't running, its safe to add the event, us holding
1677 * the ctx->lock ensures the task won't get scheduled in.
1679 add_event_to_ctx(event, ctx);
1680 raw_spin_unlock_irq(&ctx->lock);
1684 * Put a event into inactive state and update time fields.
1685 * Enabling the leader of a group effectively enables all
1686 * the group members that aren't explicitly disabled, so we
1687 * have to update their ->tstamp_enabled also.
1688 * Note: this works for group members as well as group leaders
1689 * since the non-leader members' sibling_lists will be empty.
1691 static void __perf_event_mark_enabled(struct perf_event *event)
1693 struct perf_event *sub;
1694 u64 tstamp = perf_event_time(event);
1696 event->state = PERF_EVENT_STATE_INACTIVE;
1697 event->tstamp_enabled = tstamp - event->total_time_enabled;
1698 list_for_each_entry(sub, &event->sibling_list, group_entry) {
1699 if (sub->state >= PERF_EVENT_STATE_INACTIVE)
1700 sub->tstamp_enabled = tstamp - sub->total_time_enabled;
1705 * Cross CPU call to enable a performance event
1707 static int __perf_event_enable(void *info)
1709 struct perf_event *event = info;
1710 struct perf_event_context *ctx = event->ctx;
1711 struct perf_event *leader = event->group_leader;
1712 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1715 if (WARN_ON_ONCE(!ctx->is_active))
1718 raw_spin_lock(&ctx->lock);
1719 update_context_time(ctx);
1721 if (event->state >= PERF_EVENT_STATE_INACTIVE)
1725 * set current task's cgroup time reference point
1727 perf_cgroup_set_timestamp(current, ctx);
1729 __perf_event_mark_enabled(event);
1731 if (!event_filter_match(event)) {
1732 if (is_cgroup_event(event))
1733 perf_cgroup_defer_enabled(event);
1738 * If the event is in a group and isn't the group leader,
1739 * then don't put it on unless the group is on.
1741 if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE)
1744 if (!group_can_go_on(event, cpuctx, 1)) {
1747 if (event == leader)
1748 err = group_sched_in(event, cpuctx, ctx);
1750 err = event_sched_in(event, cpuctx, ctx);
1755 * If this event can't go on and it's part of a
1756 * group, then the whole group has to come off.
1758 if (leader != event)
1759 group_sched_out(leader, cpuctx, ctx);
1760 if (leader->attr.pinned) {
1761 update_group_times(leader);
1762 leader->state = PERF_EVENT_STATE_ERROR;
1767 raw_spin_unlock(&ctx->lock);
1775 * If event->ctx is a cloned context, callers must make sure that
1776 * every task struct that event->ctx->task could possibly point to
1777 * remains valid. This condition is satisfied when called through
1778 * perf_event_for_each_child or perf_event_for_each as described
1779 * for perf_event_disable.
1781 void perf_event_enable(struct perf_event *event)
1783 struct perf_event_context *ctx = event->ctx;
1784 struct task_struct *task = ctx->task;
1788 * Enable the event on the cpu that it's on
1790 cpu_function_call(event->cpu, __perf_event_enable, event);
1794 raw_spin_lock_irq(&ctx->lock);
1795 if (event->state >= PERF_EVENT_STATE_INACTIVE)
1799 * If the event is in error state, clear that first.
1800 * That way, if we see the event in error state below, we
1801 * know that it has gone back into error state, as distinct
1802 * from the task having been scheduled away before the
1803 * cross-call arrived.
1805 if (event->state == PERF_EVENT_STATE_ERROR)
1806 event->state = PERF_EVENT_STATE_OFF;
1809 if (!ctx->is_active) {
1810 __perf_event_mark_enabled(event);
1814 raw_spin_unlock_irq(&ctx->lock);
1816 if (!task_function_call(task, __perf_event_enable, event))
1819 raw_spin_lock_irq(&ctx->lock);
1822 * If the context is active and the event is still off,
1823 * we need to retry the cross-call.
1825 if (ctx->is_active && event->state == PERF_EVENT_STATE_OFF) {
1827 * task could have been flipped by a concurrent
1828 * perf_event_context_sched_out()
1835 raw_spin_unlock_irq(&ctx->lock);
1837 EXPORT_SYMBOL_GPL(perf_event_enable);
1839 int perf_event_refresh(struct perf_event *event, int refresh)
1842 * not supported on inherited events
1844 if (event->attr.inherit || !is_sampling_event(event))
1847 atomic_add(refresh, &event->event_limit);
1848 perf_event_enable(event);
1852 EXPORT_SYMBOL_GPL(perf_event_refresh);
1854 static void ctx_sched_out(struct perf_event_context *ctx,
1855 struct perf_cpu_context *cpuctx,
1856 enum event_type_t event_type)
1858 struct perf_event *event;
1859 int is_active = ctx->is_active;
1861 ctx->is_active &= ~event_type;
1862 if (likely(!ctx->nr_events))
1865 update_context_time(ctx);
1866 update_cgrp_time_from_cpuctx(cpuctx);
1867 if (!ctx->nr_active)
1870 perf_pmu_disable(ctx->pmu);
1871 if ((is_active & EVENT_PINNED) && (event_type & EVENT_PINNED)) {
1872 list_for_each_entry(event, &ctx->pinned_groups, group_entry)
1873 group_sched_out(event, cpuctx, ctx);
1876 if ((is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE)) {
1877 list_for_each_entry(event, &ctx->flexible_groups, group_entry)
1878 group_sched_out(event, cpuctx, ctx);
1880 perf_pmu_enable(ctx->pmu);
1884 * Test whether two contexts are equivalent, i.e. whether they
1885 * have both been cloned from the same version of the same context
1886 * and they both have the same number of enabled events.
1887 * If the number of enabled events is the same, then the set
1888 * of enabled events should be the same, because these are both
1889 * inherited contexts, therefore we can't access individual events
1890 * in them directly with an fd; we can only enable/disable all
1891 * events via prctl, or enable/disable all events in a family
1892 * via ioctl, which will have the same effect on both contexts.
1894 static int context_equiv(struct perf_event_context *ctx1,
1895 struct perf_event_context *ctx2)
1897 return ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx
1898 && ctx1->parent_gen == ctx2->parent_gen
1899 && !ctx1->pin_count && !ctx2->pin_count;
1902 static void __perf_event_sync_stat(struct perf_event *event,
1903 struct perf_event *next_event)
1907 if (!event->attr.inherit_stat)
1911 * Update the event value, we cannot use perf_event_read()
1912 * because we're in the middle of a context switch and have IRQs
1913 * disabled, which upsets smp_call_function_single(), however
1914 * we know the event must be on the current CPU, therefore we
1915 * don't need to use it.
1917 switch (event->state) {
1918 case PERF_EVENT_STATE_ACTIVE:
1919 event->pmu->read(event);
1922 case PERF_EVENT_STATE_INACTIVE:
1923 update_event_times(event);
1931 * In order to keep per-task stats reliable we need to flip the event
1932 * values when we flip the contexts.
1934 value = local64_read(&next_event->count);
1935 value = local64_xchg(&event->count, value);
1936 local64_set(&next_event->count, value);
1938 swap(event->total_time_enabled, next_event->total_time_enabled);
1939 swap(event->total_time_running, next_event->total_time_running);
1942 * Since we swizzled the values, update the user visible data too.
1944 perf_event_update_userpage(event);
1945 perf_event_update_userpage(next_event);
1948 #define list_next_entry(pos, member) \
1949 list_entry(pos->member.next, typeof(*pos), member)
1951 static void perf_event_sync_stat(struct perf_event_context *ctx,
1952 struct perf_event_context *next_ctx)
1954 struct perf_event *event, *next_event;
1959 update_context_time(ctx);
1961 event = list_first_entry(&ctx->event_list,
1962 struct perf_event, event_entry);
1964 next_event = list_first_entry(&next_ctx->event_list,
1965 struct perf_event, event_entry);
1967 while (&event->event_entry != &ctx->event_list &&
1968 &next_event->event_entry != &next_ctx->event_list) {
1970 __perf_event_sync_stat(event, next_event);
1972 event = list_next_entry(event, event_entry);
1973 next_event = list_next_entry(next_event, event_entry);
1977 static void perf_event_context_sched_out(struct task_struct *task, int ctxn,
1978 struct task_struct *next)
1980 struct perf_event_context *ctx = task->perf_event_ctxp[ctxn];
1981 struct perf_event_context *next_ctx;
1982 struct perf_event_context *parent;
1983 struct perf_cpu_context *cpuctx;
1989 cpuctx = __get_cpu_context(ctx);
1990 if (!cpuctx->task_ctx)
1994 parent = rcu_dereference(ctx->parent_ctx);
1995 next_ctx = next->perf_event_ctxp[ctxn];
1996 if (parent && next_ctx &&
1997 rcu_dereference(next_ctx->parent_ctx) == parent) {
1999 * Looks like the two contexts are clones, so we might be
2000 * able to optimize the context switch. We lock both
2001 * contexts and check that they are clones under the
2002 * lock (including re-checking that neither has been
2003 * uncloned in the meantime). It doesn't matter which
2004 * order we take the locks because no other cpu could
2005 * be trying to lock both of these tasks.
2007 raw_spin_lock(&ctx->lock);
2008 raw_spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING);
2009 if (context_equiv(ctx, next_ctx)) {
2011 * XXX do we need a memory barrier of sorts
2012 * wrt to rcu_dereference() of perf_event_ctxp
2014 task->perf_event_ctxp[ctxn] = next_ctx;
2015 next->perf_event_ctxp[ctxn] = ctx;
2017 next_ctx->task = task;
2020 perf_event_sync_stat(ctx, next_ctx);
2022 raw_spin_unlock(&next_ctx->lock);
2023 raw_spin_unlock(&ctx->lock);
2028 raw_spin_lock(&ctx->lock);
2029 ctx_sched_out(ctx, cpuctx, EVENT_ALL);
2030 cpuctx->task_ctx = NULL;
2031 raw_spin_unlock(&ctx->lock);
2035 #define for_each_task_context_nr(ctxn) \
2036 for ((ctxn) = 0; (ctxn) < perf_nr_task_contexts; (ctxn)++)
2039 * Called from scheduler to remove the events of the current task,
2040 * with interrupts disabled.
2042 * We stop each event and update the event value in event->count.
2044 * This does not protect us against NMI, but disable()
2045 * sets the disabled bit in the control field of event _before_
2046 * accessing the event control register. If a NMI hits, then it will
2047 * not restart the event.
2049 void __perf_event_task_sched_out(struct task_struct *task,
2050 struct task_struct *next)
2054 for_each_task_context_nr(ctxn)
2055 perf_event_context_sched_out(task, ctxn, next);
2058 * if cgroup events exist on this CPU, then we need
2059 * to check if we have to switch out PMU state.
2060 * cgroup event are system-wide mode only
2062 if (atomic_read(&__get_cpu_var(perf_cgroup_events)))
2063 perf_cgroup_sched_out(task, next);
2066 static void task_ctx_sched_out(struct perf_event_context *ctx)
2068 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
2070 if (!cpuctx->task_ctx)
2073 if (WARN_ON_ONCE(ctx != cpuctx->task_ctx))
2076 ctx_sched_out(ctx, cpuctx, EVENT_ALL);
2077 cpuctx->task_ctx = NULL;
2081 * Called with IRQs disabled
2083 static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
2084 enum event_type_t event_type)
2086 ctx_sched_out(&cpuctx->ctx, cpuctx, event_type);
2090 ctx_pinned_sched_in(struct perf_event_context *ctx,
2091 struct perf_cpu_context *cpuctx)
2093 struct perf_event *event;
2095 list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
2096 if (event->state <= PERF_EVENT_STATE_OFF)
2098 if (!event_filter_match(event))
2101 /* may need to reset tstamp_enabled */
2102 if (is_cgroup_event(event))
2103 perf_cgroup_mark_enabled(event, ctx);
2105 if (group_can_go_on(event, cpuctx, 1))
2106 group_sched_in(event, cpuctx, ctx);
2109 * If this pinned group hasn't been scheduled,
2110 * put it in error state.
2112 if (event->state == PERF_EVENT_STATE_INACTIVE) {
2113 update_group_times(event);
2114 event->state = PERF_EVENT_STATE_ERROR;
2120 ctx_flexible_sched_in(struct perf_event_context *ctx,
2121 struct perf_cpu_context *cpuctx)
2123 struct perf_event *event;
2126 list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
2127 /* Ignore events in OFF or ERROR state */
2128 if (event->state <= PERF_EVENT_STATE_OFF)
2131 * Listen to the 'cpu' scheduling filter constraint
2134 if (!event_filter_match(event))
2137 /* may need to reset tstamp_enabled */
2138 if (is_cgroup_event(event))
2139 perf_cgroup_mark_enabled(event, ctx);
2141 if (group_can_go_on(event, cpuctx, can_add_hw)) {
2142 if (group_sched_in(event, cpuctx, ctx))
2149 ctx_sched_in(struct perf_event_context *ctx,
2150 struct perf_cpu_context *cpuctx,
2151 enum event_type_t event_type,
2152 struct task_struct *task)
2155 int is_active = ctx->is_active;
2157 ctx->is_active |= event_type;
2158 if (likely(!ctx->nr_events))
2162 ctx->timestamp = now;
2163 perf_cgroup_set_timestamp(task, ctx);
2165 * First go through the list and put on any pinned groups
2166 * in order to give them the best chance of going on.
2168 if (!(is_active & EVENT_PINNED) && (event_type & EVENT_PINNED))
2169 ctx_pinned_sched_in(ctx, cpuctx);
2171 /* Then walk through the lower prio flexible groups */
2172 if (!(is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE))
2173 ctx_flexible_sched_in(ctx, cpuctx);
2176 static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
2177 enum event_type_t event_type,
2178 struct task_struct *task)
2180 struct perf_event_context *ctx = &cpuctx->ctx;
2182 ctx_sched_in(ctx, cpuctx, event_type, task);
2185 static void perf_event_context_sched_in(struct perf_event_context *ctx,
2186 struct task_struct *task)
2188 struct perf_cpu_context *cpuctx;
2190 cpuctx = __get_cpu_context(ctx);
2191 if (cpuctx->task_ctx == ctx)
2194 perf_ctx_lock(cpuctx, ctx);
2195 perf_pmu_disable(ctx->pmu);
2197 * We want to keep the following priority order:
2198 * cpu pinned (that don't need to move), task pinned,
2199 * cpu flexible, task flexible.
2201 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
2204 cpuctx->task_ctx = ctx;
2206 perf_event_sched_in(cpuctx, cpuctx->task_ctx, task);
2208 perf_pmu_enable(ctx->pmu);
2209 perf_ctx_unlock(cpuctx, ctx);
2212 * Since these rotations are per-cpu, we need to ensure the
2213 * cpu-context we got scheduled on is actually rotating.
2215 perf_pmu_rotate_start(ctx->pmu);
2219 * When sampling the branck stack in system-wide, it may be necessary
2220 * to flush the stack on context switch. This happens when the branch
2221 * stack does not tag its entries with the pid of the current task.
2222 * Otherwise it becomes impossible to associate a branch entry with a
2223 * task. This ambiguity is more likely to appear when the branch stack
2224 * supports priv level filtering and the user sets it to monitor only
2225 * at the user level (which could be a useful measurement in system-wide
2226 * mode). In that case, the risk is high of having a branch stack with
2227 * branch from multiple tasks. Flushing may mean dropping the existing
2228 * entries or stashing them somewhere in the PMU specific code layer.
2230 * This function provides the context switch callback to the lower code
2231 * layer. It is invoked ONLY when there is at least one system-wide context
2232 * with at least one active event using taken branch sampling.
2234 static void perf_branch_stack_sched_in(struct task_struct *prev,
2235 struct task_struct *task)
2237 struct perf_cpu_context *cpuctx;
2239 unsigned long flags;
2241 /* no need to flush branch stack if not changing task */
2245 local_irq_save(flags);
2249 list_for_each_entry_rcu(pmu, &pmus, entry) {
2250 cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
2253 * check if the context has at least one
2254 * event using PERF_SAMPLE_BRANCH_STACK
2256 if (cpuctx->ctx.nr_branch_stack > 0
2257 && pmu->flush_branch_stack) {
2259 pmu = cpuctx->ctx.pmu;
2261 perf_ctx_lock(cpuctx, cpuctx->task_ctx);
2263 perf_pmu_disable(pmu);
2265 pmu->flush_branch_stack();
2267 perf_pmu_enable(pmu);
2269 perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
2275 local_irq_restore(flags);
2279 * Called from scheduler to add the events of the current task
2280 * with interrupts disabled.
2282 * We restore the event value and then enable it.
2284 * This does not protect us against NMI, but enable()
2285 * sets the enabled bit in the control field of event _before_
2286 * accessing the event control register. If a NMI hits, then it will
2287 * keep the event running.
2289 void __perf_event_task_sched_in(struct task_struct *prev,
2290 struct task_struct *task)
2292 struct perf_event_context *ctx;
2295 for_each_task_context_nr(ctxn) {
2296 ctx = task->perf_event_ctxp[ctxn];
2300 perf_event_context_sched_in(ctx, task);
2303 * if cgroup events exist on this CPU, then we need
2304 * to check if we have to switch in PMU state.
2305 * cgroup event are system-wide mode only
2307 if (atomic_read(&__get_cpu_var(perf_cgroup_events)))
2308 perf_cgroup_sched_in(prev, task);
2310 /* check for system-wide branch_stack events */
2311 if (atomic_read(&__get_cpu_var(perf_branch_stack_events)))
2312 perf_branch_stack_sched_in(prev, task);
2315 static u64 perf_calculate_period(struct perf_event *event, u64 nsec, u64 count)
2317 u64 frequency = event->attr.sample_freq;
2318 u64 sec = NSEC_PER_SEC;
2319 u64 divisor, dividend;
2321 int count_fls, nsec_fls, frequency_fls, sec_fls;
2323 count_fls = fls64(count);
2324 nsec_fls = fls64(nsec);
2325 frequency_fls = fls64(frequency);
2329 * We got @count in @nsec, with a target of sample_freq HZ
2330 * the target period becomes:
2333 * period = -------------------
2334 * @nsec * sample_freq
2339 * Reduce accuracy by one bit such that @a and @b converge
2340 * to a similar magnitude.
2342 #define REDUCE_FLS(a, b) \
2344 if (a##_fls > b##_fls) { \
2354 * Reduce accuracy until either term fits in a u64, then proceed with
2355 * the other, so that finally we can do a u64/u64 division.
2357 while (count_fls + sec_fls > 64 && nsec_fls + frequency_fls > 64) {
2358 REDUCE_FLS(nsec, frequency);
2359 REDUCE_FLS(sec, count);
2362 if (count_fls + sec_fls > 64) {
2363 divisor = nsec * frequency;
2365 while (count_fls + sec_fls > 64) {
2366 REDUCE_FLS(count, sec);
2370 dividend = count * sec;
2372 dividend = count * sec;
2374 while (nsec_fls + frequency_fls > 64) {
2375 REDUCE_FLS(nsec, frequency);
2379 divisor = nsec * frequency;
2385 return div64_u64(dividend, divisor);
2388 static DEFINE_PER_CPU(int, perf_throttled_count);
2389 static DEFINE_PER_CPU(u64, perf_throttled_seq);
2391 static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count, bool disable)
2393 struct hw_perf_event *hwc = &event->hw;
2394 s64 period, sample_period;
2397 period = perf_calculate_period(event, nsec, count);
2399 delta = (s64)(period - hwc->sample_period);
2400 delta = (delta + 7) / 8; /* low pass filter */
2402 sample_period = hwc->sample_period + delta;
2407 hwc->sample_period = sample_period;
2409 if (local64_read(&hwc->period_left) > 8*sample_period) {
2411 event->pmu->stop(event, PERF_EF_UPDATE);
2413 local64_set(&hwc->period_left, 0);
2416 event->pmu->start(event, PERF_EF_RELOAD);
2421 * combine freq adjustment with unthrottling to avoid two passes over the
2422 * events. At the same time, make sure, having freq events does not change
2423 * the rate of unthrottling as that would introduce bias.
2425 static void perf_adjust_freq_unthr_context(struct perf_event_context *ctx,
2428 struct perf_event *event;
2429 struct hw_perf_event *hwc;
2430 u64 now, period = TICK_NSEC;
2434 * only need to iterate over all events iff:
2435 * - context have events in frequency mode (needs freq adjust)
2436 * - there are events to unthrottle on this cpu
2438 if (!(ctx->nr_freq || needs_unthr))
2441 raw_spin_lock(&ctx->lock);
2442 perf_pmu_disable(ctx->pmu);
2444 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
2445 if (event->state != PERF_EVENT_STATE_ACTIVE)
2448 if (!event_filter_match(event))
2453 if (needs_unthr && hwc->interrupts == MAX_INTERRUPTS) {
2454 hwc->interrupts = 0;
2455 perf_log_throttle(event, 1);
2456 event->pmu->start(event, 0);
2459 if (!event->attr.freq || !event->attr.sample_freq)
2463 * stop the event and update event->count
2465 event->pmu->stop(event, PERF_EF_UPDATE);
2467 now = local64_read(&event->count);
2468 delta = now - hwc->freq_count_stamp;
2469 hwc->freq_count_stamp = now;
2473 * reload only if value has changed
2474 * we have stopped the event so tell that
2475 * to perf_adjust_period() to avoid stopping it
2479 perf_adjust_period(event, period, delta, false);
2481 event->pmu->start(event, delta > 0 ? PERF_EF_RELOAD : 0);
2484 perf_pmu_enable(ctx->pmu);
2485 raw_spin_unlock(&ctx->lock);
2489 * Round-robin a context's events:
2491 static void rotate_ctx(struct perf_event_context *ctx)
2494 * Rotate the first entry last of non-pinned groups. Rotation might be
2495 * disabled by the inheritance code.
2497 if (!ctx->rotate_disable)
2498 list_rotate_left(&ctx->flexible_groups);
2502 * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
2503 * because they're strictly cpu affine and rotate_start is called with IRQs
2504 * disabled, while rotate_context is called from IRQ context.
2506 static void perf_rotate_context(struct perf_cpu_context *cpuctx)
2508 struct perf_event_context *ctx = NULL;
2509 int rotate = 0, remove = 1;
2511 if (cpuctx->ctx.nr_events) {
2513 if (cpuctx->ctx.nr_events != cpuctx->ctx.nr_active)
2517 ctx = cpuctx->task_ctx;
2518 if (ctx && ctx->nr_events) {
2520 if (ctx->nr_events != ctx->nr_active)
2527 perf_ctx_lock(cpuctx, cpuctx->task_ctx);
2528 perf_pmu_disable(cpuctx->ctx.pmu);
2530 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
2532 ctx_sched_out(ctx, cpuctx, EVENT_FLEXIBLE);
2534 rotate_ctx(&cpuctx->ctx);
2538 perf_event_sched_in(cpuctx, ctx, current);
2540 perf_pmu_enable(cpuctx->ctx.pmu);
2541 perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
2544 list_del_init(&cpuctx->rotation_list);
2547 void perf_event_task_tick(void)
2549 struct list_head *head = &__get_cpu_var(rotation_list);
2550 struct perf_cpu_context *cpuctx, *tmp;
2551 struct perf_event_context *ctx;
2554 WARN_ON(!irqs_disabled());
2556 __this_cpu_inc(perf_throttled_seq);
2557 throttled = __this_cpu_xchg(perf_throttled_count, 0);
2559 list_for_each_entry_safe(cpuctx, tmp, head, rotation_list) {
2561 perf_adjust_freq_unthr_context(ctx, throttled);
2563 ctx = cpuctx->task_ctx;
2565 perf_adjust_freq_unthr_context(ctx, throttled);
2567 if (cpuctx->jiffies_interval == 1 ||
2568 !(jiffies % cpuctx->jiffies_interval))
2569 perf_rotate_context(cpuctx);
2573 static int event_enable_on_exec(struct perf_event *event,
2574 struct perf_event_context *ctx)
2576 if (!event->attr.enable_on_exec)
2579 event->attr.enable_on_exec = 0;
2580 if (event->state >= PERF_EVENT_STATE_INACTIVE)
2583 __perf_event_mark_enabled(event);
2589 * Enable all of a task's events that have been marked enable-on-exec.
2590 * This expects task == current.
2592 static void perf_event_enable_on_exec(struct perf_event_context *ctx)
2594 struct perf_event *event;
2595 unsigned long flags;
2599 local_irq_save(flags);
2600 if (!ctx || !ctx->nr_events)
2604 * We must ctxsw out cgroup events to avoid conflict
2605 * when invoking perf_task_event_sched_in() later on
2606 * in this function. Otherwise we end up trying to
2607 * ctxswin cgroup events which are already scheduled
2610 perf_cgroup_sched_out(current, NULL);
2612 raw_spin_lock(&ctx->lock);
2613 task_ctx_sched_out(ctx);
2615 list_for_each_entry(event, &ctx->event_list, event_entry) {
2616 ret = event_enable_on_exec(event, ctx);
2622 * Unclone this context if we enabled any event.
2627 raw_spin_unlock(&ctx->lock);
2630 * Also calls ctxswin for cgroup events, if any:
2632 perf_event_context_sched_in(ctx, ctx->task);
2634 local_irq_restore(flags);
2638 * Cross CPU call to read the hardware event
2640 static void __perf_event_read(void *info)
2642 struct perf_event *event = info;
2643 struct perf_event_context *ctx = event->ctx;
2644 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
2647 * If this is a task context, we need to check whether it is
2648 * the current task context of this cpu. If not it has been
2649 * scheduled out before the smp call arrived. In that case
2650 * event->count would have been updated to a recent sample
2651 * when the event was scheduled out.
2653 if (ctx->task && cpuctx->task_ctx != ctx)
2656 raw_spin_lock(&ctx->lock);
2657 if (ctx->is_active) {
2658 update_context_time(ctx);
2659 update_cgrp_time_from_event(event);
2661 update_event_times(event);
2662 if (event->state == PERF_EVENT_STATE_ACTIVE)
2663 event->pmu->read(event);
2664 raw_spin_unlock(&ctx->lock);
2667 static inline u64 perf_event_count(struct perf_event *event)
2669 return local64_read(&event->count) + atomic64_read(&event->child_count);
2672 static u64 perf_event_read(struct perf_event *event)
2675 * If event is enabled and currently active on a CPU, update the
2676 * value in the event structure:
2678 if (event->state == PERF_EVENT_STATE_ACTIVE) {
2679 smp_call_function_single(event->oncpu,
2680 __perf_event_read, event, 1);
2681 } else if (event->state == PERF_EVENT_STATE_INACTIVE) {
2682 struct perf_event_context *ctx = event->ctx;
2683 unsigned long flags;
2685 raw_spin_lock_irqsave(&ctx->lock, flags);
2687 * may read while context is not active
2688 * (e.g., thread is blocked), in that case
2689 * we cannot update context time
2691 if (ctx->is_active) {
2692 update_context_time(ctx);
2693 update_cgrp_time_from_event(event);
2695 update_event_times(event);
2696 raw_spin_unlock_irqrestore(&ctx->lock, flags);
2699 return perf_event_count(event);
2703 * Initialize the perf_event context in a task_struct:
2705 static void __perf_event_init_context(struct perf_event_context *ctx)
2707 raw_spin_lock_init(&ctx->lock);
2708 mutex_init(&ctx->mutex);
2709 INIT_LIST_HEAD(&ctx->pinned_groups);
2710 INIT_LIST_HEAD(&ctx->flexible_groups);
2711 INIT_LIST_HEAD(&ctx->event_list);
2712 atomic_set(&ctx->refcount, 1);
2715 static struct perf_event_context *
2716 alloc_perf_context(struct pmu *pmu, struct task_struct *task)
2718 struct perf_event_context *ctx;
2720 ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL);
2724 __perf_event_init_context(ctx);
2727 get_task_struct(task);
2734 static struct task_struct *
2735 find_lively_task_by_vpid(pid_t vpid)
2737 struct task_struct *task;
2744 task = find_task_by_vpid(vpid);
2746 get_task_struct(task);
2750 return ERR_PTR(-ESRCH);
2752 /* Reuse ptrace permission checks for now. */
2754 if (!ptrace_may_access(task, PTRACE_MODE_READ))
2759 put_task_struct(task);
2760 return ERR_PTR(err);
2765 * Returns a matching context with refcount and pincount.
2767 static struct perf_event_context *
2768 find_get_context(struct pmu *pmu, struct task_struct *task, int cpu)
2770 struct perf_event_context *ctx;
2771 struct perf_cpu_context *cpuctx;
2772 unsigned long flags;
2776 /* Must be root to operate on a CPU event: */
2777 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
2778 return ERR_PTR(-EACCES);
2781 * We could be clever and allow to attach a event to an
2782 * offline CPU and activate it when the CPU comes up, but
2785 if (!cpu_online(cpu))
2786 return ERR_PTR(-ENODEV);
2788 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
2797 ctxn = pmu->task_ctx_nr;
2802 ctx = perf_lock_task_context(task, ctxn, &flags);
2806 raw_spin_unlock_irqrestore(&ctx->lock, flags);
2808 ctx = alloc_perf_context(pmu, task);
2814 mutex_lock(&task->perf_event_mutex);
2816 * If it has already passed perf_event_exit_task().
2817 * we must see PF_EXITING, it takes this mutex too.
2819 if (task->flags & PF_EXITING)
2821 else if (task->perf_event_ctxp[ctxn])
2826 rcu_assign_pointer(task->perf_event_ctxp[ctxn], ctx);
2828 mutex_unlock(&task->perf_event_mutex);
2830 if (unlikely(err)) {
2842 return ERR_PTR(err);
2845 static void perf_event_free_filter(struct perf_event *event);
2847 static void free_event_rcu(struct rcu_head *head)
2849 struct perf_event *event;
2851 event = container_of(head, struct perf_event, rcu_head);
2853 put_pid_ns(event->ns);
2854 perf_event_free_filter(event);
2858 static void ring_buffer_put(struct ring_buffer *rb);
2860 static void free_event(struct perf_event *event)
2862 irq_work_sync(&event->pending);
2864 if (!event->parent) {
2865 if (event->attach_state & PERF_ATTACH_TASK)
2866 static_key_slow_dec_deferred(&perf_sched_events);
2867 if (event->attr.mmap || event->attr.mmap_data)
2868 atomic_dec(&nr_mmap_events);
2869 if (event->attr.comm)
2870 atomic_dec(&nr_comm_events);
2871 if (event->attr.task)
2872 atomic_dec(&nr_task_events);
2873 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN)
2874 put_callchain_buffers();
2875 if (is_cgroup_event(event)) {
2876 atomic_dec(&per_cpu(perf_cgroup_events, event->cpu));
2877 static_key_slow_dec_deferred(&perf_sched_events);
2880 if (has_branch_stack(event)) {
2881 static_key_slow_dec_deferred(&perf_sched_events);
2882 /* is system-wide event */
2883 if (!(event->attach_state & PERF_ATTACH_TASK))
2884 atomic_dec(&per_cpu(perf_branch_stack_events,
2890 ring_buffer_put(event->rb);
2894 if (is_cgroup_event(event))
2895 perf_detach_cgroup(event);
2898 event->destroy(event);
2901 put_ctx(event->ctx);
2903 call_rcu(&event->rcu_head, free_event_rcu);
2906 int perf_event_release_kernel(struct perf_event *event)
2908 struct perf_event_context *ctx = event->ctx;
2910 WARN_ON_ONCE(ctx->parent_ctx);
2912 * There are two ways this annotation is useful:
2914 * 1) there is a lock recursion from perf_event_exit_task
2915 * see the comment there.
2917 * 2) there is a lock-inversion with mmap_sem through
2918 * perf_event_read_group(), which takes faults while
2919 * holding ctx->mutex, however this is called after
2920 * the last filedesc died, so there is no possibility
2921 * to trigger the AB-BA case.
2923 mutex_lock_nested(&ctx->mutex, SINGLE_DEPTH_NESTING);
2924 raw_spin_lock_irq(&ctx->lock);
2925 perf_group_detach(event);
2926 raw_spin_unlock_irq(&ctx->lock);
2927 perf_remove_from_context(event);
2928 mutex_unlock(&ctx->mutex);
2934 EXPORT_SYMBOL_GPL(perf_event_release_kernel);
2937 * Called when the last reference to the file is gone.
2939 static int perf_release(struct inode *inode, struct file *file)
2941 struct perf_event *event = file->private_data;
2942 struct task_struct *owner;
2944 file->private_data = NULL;
2947 owner = ACCESS_ONCE(event->owner);
2949 * Matches the smp_wmb() in perf_event_exit_task(). If we observe
2950 * !owner it means the list deletion is complete and we can indeed
2951 * free this event, otherwise we need to serialize on
2952 * owner->perf_event_mutex.
2954 smp_read_barrier_depends();
2957 * Since delayed_put_task_struct() also drops the last
2958 * task reference we can safely take a new reference
2959 * while holding the rcu_read_lock().
2961 get_task_struct(owner);
2966 mutex_lock(&owner->perf_event_mutex);
2968 * We have to re-check the event->owner field, if it is cleared
2969 * we raced with perf_event_exit_task(), acquiring the mutex
2970 * ensured they're done, and we can proceed with freeing the
2974 list_del_init(&event->owner_entry);
2975 mutex_unlock(&owner->perf_event_mutex);
2976 put_task_struct(owner);
2979 return perf_event_release_kernel(event);
2982 u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running)
2984 struct perf_event *child;
2990 mutex_lock(&event->child_mutex);
2991 total += perf_event_read(event);
2992 *enabled += event->total_time_enabled +
2993 atomic64_read(&event->child_total_time_enabled);
2994 *running += event->total_time_running +
2995 atomic64_read(&event->child_total_time_running);
2997 list_for_each_entry(child, &event->child_list, child_list) {
2998 total += perf_event_read(child);
2999 *enabled += child->total_time_enabled;
3000 *running += child->total_time_running;
3002 mutex_unlock(&event->child_mutex);
3006 EXPORT_SYMBOL_GPL(perf_event_read_value);
3008 static int perf_event_read_group(struct perf_event *event,
3009 u64 read_format, char __user *buf)
3011 struct perf_event *leader = event->group_leader, *sub;
3012 int n = 0, size = 0, ret = -EFAULT;
3013 struct perf_event_context *ctx = leader->ctx;
3015 u64 count, enabled, running;
3017 mutex_lock(&ctx->mutex);
3018 count = perf_event_read_value(leader, &enabled, &running);
3020 values[n++] = 1 + leader->nr_siblings;
3021 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3022 values[n++] = enabled;
3023 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3024 values[n++] = running;
3025 values[n++] = count;
3026 if (read_format & PERF_FORMAT_ID)
3027 values[n++] = primary_event_id(leader);
3029 size = n * sizeof(u64);
3031 if (copy_to_user(buf, values, size))
3036 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
3039 values[n++] = perf_event_read_value(sub, &enabled, &running);
3040 if (read_format & PERF_FORMAT_ID)
3041 values[n++] = primary_event_id(sub);
3043 size = n * sizeof(u64);
3045 if (copy_to_user(buf + ret, values, size)) {
3053 mutex_unlock(&ctx->mutex);
3058 static int perf_event_read_one(struct perf_event *event,
3059 u64 read_format, char __user *buf)
3061 u64 enabled, running;
3065 values[n++] = perf_event_read_value(event, &enabled, &running);
3066 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3067 values[n++] = enabled;
3068 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3069 values[n++] = running;
3070 if (read_format & PERF_FORMAT_ID)
3071 values[n++] = primary_event_id(event);
3073 if (copy_to_user(buf, values, n * sizeof(u64)))
3076 return n * sizeof(u64);
3080 * Read the performance event - simple non blocking version for now
3083 perf_read_hw(struct perf_event *event, char __user *buf, size_t count)
3085 u64 read_format = event->attr.read_format;
3089 * Return end-of-file for a read on a event that is in
3090 * error state (i.e. because it was pinned but it couldn't be
3091 * scheduled on to the CPU at some point).
3093 if (event->state == PERF_EVENT_STATE_ERROR)
3096 if (count < event->read_size)
3099 WARN_ON_ONCE(event->ctx->parent_ctx);
3100 if (read_format & PERF_FORMAT_GROUP)
3101 ret = perf_event_read_group(event, read_format, buf);
3103 ret = perf_event_read_one(event, read_format, buf);
3109 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
3111 struct perf_event *event = file->private_data;
3113 return perf_read_hw(event, buf, count);
3116 static unsigned int perf_poll(struct file *file, poll_table *wait)
3118 struct perf_event *event = file->private_data;
3119 struct ring_buffer *rb;
3120 unsigned int events = POLL_HUP;
3123 * Race between perf_event_set_output() and perf_poll(): perf_poll()
3124 * grabs the rb reference but perf_event_set_output() overrides it.
3125 * Here is the timeline for two threads T1, T2:
3126 * t0: T1, rb = rcu_dereference(event->rb)
3127 * t1: T2, old_rb = event->rb
3128 * t2: T2, event->rb = new rb
3129 * t3: T2, ring_buffer_detach(old_rb)
3130 * t4: T1, ring_buffer_attach(rb1)
3131 * t5: T1, poll_wait(event->waitq)
3133 * To avoid this problem, we grab mmap_mutex in perf_poll()
3134 * thereby ensuring that the assignment of the new ring buffer
3135 * and the detachment of the old buffer appear atomic to perf_poll()
3137 mutex_lock(&event->mmap_mutex);
3140 rb = rcu_dereference(event->rb);
3142 ring_buffer_attach(event, rb);
3143 events = atomic_xchg(&rb->poll, 0);
3147 mutex_unlock(&event->mmap_mutex);
3149 poll_wait(file, &event->waitq, wait);
3154 static void perf_event_reset(struct perf_event *event)
3156 (void)perf_event_read(event);
3157 local64_set(&event->count, 0);
3158 perf_event_update_userpage(event);
3162 * Holding the top-level event's child_mutex means that any
3163 * descendant process that has inherited this event will block
3164 * in sync_child_event if it goes to exit, thus satisfying the
3165 * task existence requirements of perf_event_enable/disable.
3167 static void perf_event_for_each_child(struct perf_event *event,
3168 void (*func)(struct perf_event *))
3170 struct perf_event *child;
3172 WARN_ON_ONCE(event->ctx->parent_ctx);
3173 mutex_lock(&event->child_mutex);
3175 list_for_each_entry(child, &event->child_list, child_list)
3177 mutex_unlock(&event->child_mutex);
3180 static void perf_event_for_each(struct perf_event *event,
3181 void (*func)(struct perf_event *))
3183 struct perf_event_context *ctx = event->ctx;
3184 struct perf_event *sibling;
3186 WARN_ON_ONCE(ctx->parent_ctx);
3187 mutex_lock(&ctx->mutex);
3188 event = event->group_leader;
3190 perf_event_for_each_child(event, func);
3191 list_for_each_entry(sibling, &event->sibling_list, group_entry)
3192 perf_event_for_each_child(sibling, func);
3193 mutex_unlock(&ctx->mutex);
3196 static int perf_event_period(struct perf_event *event, u64 __user *arg)
3198 struct perf_event_context *ctx = event->ctx;
3202 if (!is_sampling_event(event))
3205 if (copy_from_user(&value, arg, sizeof(value)))
3211 raw_spin_lock_irq(&ctx->lock);
3212 if (event->attr.freq) {
3213 if (value > sysctl_perf_event_sample_rate) {
3218 event->attr.sample_freq = value;
3220 event->attr.sample_period = value;
3221 event->hw.sample_period = value;
3224 raw_spin_unlock_irq(&ctx->lock);
3229 static const struct file_operations perf_fops;
3231 static struct perf_event *perf_fget_light(int fd, int *fput_needed)
3235 file = fget_light(fd, fput_needed);
3237 return ERR_PTR(-EBADF);
3239 if (file->f_op != &perf_fops) {
3240 fput_light(file, *fput_needed);
3242 return ERR_PTR(-EBADF);
3245 return file->private_data;
3248 static int perf_event_set_output(struct perf_event *event,
3249 struct perf_event *output_event);
3250 static int perf_event_set_filter(struct perf_event *event, void __user *arg);
3252 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
3254 struct perf_event *event = file->private_data;
3255 void (*func)(struct perf_event *);
3259 case PERF_EVENT_IOC_ENABLE:
3260 func = perf_event_enable;
3262 case PERF_EVENT_IOC_DISABLE:
3263 func = perf_event_disable;
3265 case PERF_EVENT_IOC_RESET:
3266 func = perf_event_reset;
3269 case PERF_EVENT_IOC_REFRESH:
3270 return perf_event_refresh(event, arg);
3272 case PERF_EVENT_IOC_PERIOD:
3273 return perf_event_period(event, (u64 __user *)arg);
3275 case PERF_EVENT_IOC_SET_OUTPUT:
3277 struct perf_event *output_event = NULL;
3278 int fput_needed = 0;
3282 output_event = perf_fget_light(arg, &fput_needed);
3283 if (IS_ERR(output_event))
3284 return PTR_ERR(output_event);
3287 ret = perf_event_set_output(event, output_event);
3289 fput_light(output_event->filp, fput_needed);
3294 case PERF_EVENT_IOC_SET_FILTER:
3295 return perf_event_set_filter(event, (void __user *)arg);
3301 if (flags & PERF_IOC_FLAG_GROUP)
3302 perf_event_for_each(event, func);
3304 perf_event_for_each_child(event, func);
3309 int perf_event_task_enable(void)
3311 struct perf_event *event;
3313 mutex_lock(¤t->perf_event_mutex);
3314 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
3315 perf_event_for_each_child(event, perf_event_enable);
3316 mutex_unlock(¤t->perf_event_mutex);
3321 int perf_event_task_disable(void)
3323 struct perf_event *event;
3325 mutex_lock(¤t->perf_event_mutex);
3326 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
3327 perf_event_for_each_child(event, perf_event_disable);
3328 mutex_unlock(¤t->perf_event_mutex);
3333 static int perf_event_index(struct perf_event *event)
3335 if (event->hw.state & PERF_HES_STOPPED)
3338 if (event->state != PERF_EVENT_STATE_ACTIVE)
3341 return event->pmu->event_idx(event);
3344 static void calc_timer_values(struct perf_event *event,
3351 *now = perf_clock();
3352 ctx_time = event->shadow_ctx_time + *now;
3353 *enabled = ctx_time - event->tstamp_enabled;
3354 *running = ctx_time - event->tstamp_running;
3357 void __weak arch_perf_update_userpage(struct perf_event_mmap_page *userpg, u64 now)
3362 * Callers need to ensure there can be no nesting of this function, otherwise
3363 * the seqlock logic goes bad. We can not serialize this because the arch
3364 * code calls this from NMI context.
3366 void perf_event_update_userpage(struct perf_event *event)
3368 struct perf_event_mmap_page *userpg;
3369 struct ring_buffer *rb;
3370 u64 enabled, running, now;
3374 * compute total_time_enabled, total_time_running
3375 * based on snapshot values taken when the event
3376 * was last scheduled in.
3378 * we cannot simply called update_context_time()
3379 * because of locking issue as we can be called in
3382 calc_timer_values(event, &now, &enabled, &running);
3383 rb = rcu_dereference(event->rb);
3387 userpg = rb->user_page;
3390 * Disable preemption so as to not let the corresponding user-space
3391 * spin too long if we get preempted.
3396 userpg->index = perf_event_index(event);
3397 userpg->offset = perf_event_count(event);
3399 userpg->offset -= local64_read(&event->hw.prev_count);
3401 userpg->time_enabled = enabled +
3402 atomic64_read(&event->child_total_time_enabled);
3404 userpg->time_running = running +
3405 atomic64_read(&event->child_total_time_running);
3407 arch_perf_update_userpage(userpg, now);
3416 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
3418 struct perf_event *event = vma->vm_file->private_data;
3419 struct ring_buffer *rb;
3420 int ret = VM_FAULT_SIGBUS;
3422 if (vmf->flags & FAULT_FLAG_MKWRITE) {
3423 if (vmf->pgoff == 0)
3429 rb = rcu_dereference(event->rb);
3433 if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE))
3436 vmf->page = perf_mmap_to_page(rb, vmf->pgoff);
3440 get_page(vmf->page);
3441 vmf->page->mapping = vma->vm_file->f_mapping;
3442 vmf->page->index = vmf->pgoff;
3451 static void ring_buffer_attach(struct perf_event *event,
3452 struct ring_buffer *rb)
3454 unsigned long flags;
3456 if (!list_empty(&event->rb_entry))
3459 spin_lock_irqsave(&rb->event_lock, flags);
3460 if (!list_empty(&event->rb_entry))
3463 list_add(&event->rb_entry, &rb->event_list);
3465 spin_unlock_irqrestore(&rb->event_lock, flags);
3468 static void ring_buffer_detach(struct perf_event *event,
3469 struct ring_buffer *rb)
3471 unsigned long flags;
3473 if (list_empty(&event->rb_entry))
3476 spin_lock_irqsave(&rb->event_lock, flags);
3477 list_del_init(&event->rb_entry);
3478 wake_up_all(&event->waitq);
3479 spin_unlock_irqrestore(&rb->event_lock, flags);
3482 static void ring_buffer_wakeup(struct perf_event *event)
3484 struct ring_buffer *rb;
3487 rb = rcu_dereference(event->rb);
3491 list_for_each_entry_rcu(event, &rb->event_list, rb_entry)
3492 wake_up_all(&event->waitq);
3498 static void rb_free_rcu(struct rcu_head *rcu_head)
3500 struct ring_buffer *rb;
3502 rb = container_of(rcu_head, struct ring_buffer, rcu_head);
3506 static struct ring_buffer *ring_buffer_get(struct perf_event *event)
3508 struct ring_buffer *rb;
3511 rb = rcu_dereference(event->rb);
3513 if (!atomic_inc_not_zero(&rb->refcount))
3521 static void ring_buffer_put(struct ring_buffer *rb)
3523 struct perf_event *event, *n;
3524 unsigned long flags;
3526 if (!atomic_dec_and_test(&rb->refcount))
3529 spin_lock_irqsave(&rb->event_lock, flags);
3530 list_for_each_entry_safe(event, n, &rb->event_list, rb_entry) {
3531 list_del_init(&event->rb_entry);
3532 wake_up_all(&event->waitq);
3534 spin_unlock_irqrestore(&rb->event_lock, flags);
3536 call_rcu(&rb->rcu_head, rb_free_rcu);
3539 static void perf_mmap_open(struct vm_area_struct *vma)
3541 struct perf_event *event = vma->vm_file->private_data;
3543 atomic_inc(&event->mmap_count);
3546 static void perf_mmap_close(struct vm_area_struct *vma)
3548 struct perf_event *event = vma->vm_file->private_data;
3550 if (atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex)) {
3551 unsigned long size = perf_data_size(event->rb);
3552 struct user_struct *user = event->mmap_user;
3553 struct ring_buffer *rb = event->rb;
3555 atomic_long_sub((size >> PAGE_SHIFT) + 1, &user->locked_vm);
3556 vma->vm_mm->pinned_vm -= event->mmap_locked;
3557 rcu_assign_pointer(event->rb, NULL);
3558 ring_buffer_detach(event, rb);
3559 mutex_unlock(&event->mmap_mutex);
3561 ring_buffer_put(rb);
3566 static const struct vm_operations_struct perf_mmap_vmops = {
3567 .open = perf_mmap_open,
3568 .close = perf_mmap_close,
3569 .fault = perf_mmap_fault,
3570 .page_mkwrite = perf_mmap_fault,
3573 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
3575 struct perf_event *event = file->private_data;
3576 unsigned long user_locked, user_lock_limit;
3577 struct user_struct *user = current_user();
3578 unsigned long locked, lock_limit;
3579 struct ring_buffer *rb;
3580 unsigned long vma_size;
3581 unsigned long nr_pages;
3582 long user_extra, extra;
3583 int ret = 0, flags = 0;
3586 * Don't allow mmap() of inherited per-task counters. This would
3587 * create a performance issue due to all children writing to the
3590 if (event->cpu == -1 && event->attr.inherit)
3593 if (!(vma->vm_flags & VM_SHARED))
3596 vma_size = vma->vm_end - vma->vm_start;
3597 nr_pages = (vma_size / PAGE_SIZE) - 1;
3600 * If we have rb pages ensure they're a power-of-two number, so we
3601 * can do bitmasks instead of modulo.
3603 if (nr_pages != 0 && !is_power_of_2(nr_pages))
3606 if (vma_size != PAGE_SIZE * (1 + nr_pages))
3609 if (vma->vm_pgoff != 0)
3612 WARN_ON_ONCE(event->ctx->parent_ctx);
3613 mutex_lock(&event->mmap_mutex);
3615 if (event->rb->nr_pages == nr_pages)
3616 atomic_inc(&event->rb->refcount);
3622 user_extra = nr_pages + 1;
3623 user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10);
3626 * Increase the limit linearly with more CPUs:
3628 user_lock_limit *= num_online_cpus();
3630 user_locked = atomic_long_read(&user->locked_vm) + user_extra;
3633 if (user_locked > user_lock_limit)
3634 extra = user_locked - user_lock_limit;
3636 lock_limit = rlimit(RLIMIT_MEMLOCK);
3637 lock_limit >>= PAGE_SHIFT;
3638 locked = vma->vm_mm->pinned_vm + extra;
3640 if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() &&
3641 !capable(CAP_IPC_LOCK)) {
3648 if (vma->vm_flags & VM_WRITE)
3649 flags |= RING_BUFFER_WRITABLE;
3651 rb = rb_alloc(nr_pages,
3652 event->attr.watermark ? event->attr.wakeup_watermark : 0,
3659 rcu_assign_pointer(event->rb, rb);
3661 atomic_long_add(user_extra, &user->locked_vm);
3662 event->mmap_locked = extra;
3663 event->mmap_user = get_current_user();
3664 vma->vm_mm->pinned_vm += event->mmap_locked;
3666 perf_event_update_userpage(event);
3670 atomic_inc(&event->mmap_count);
3671 mutex_unlock(&event->mmap_mutex);
3673 vma->vm_flags |= VM_RESERVED;
3674 vma->vm_ops = &perf_mmap_vmops;
3679 static int perf_fasync(int fd, struct file *filp, int on)
3681 struct inode *inode = filp->f_path.dentry->d_inode;
3682 struct perf_event *event = filp->private_data;
3685 mutex_lock(&inode->i_mutex);
3686 retval = fasync_helper(fd, filp, on, &event->fasync);
3687 mutex_unlock(&inode->i_mutex);
3695 static const struct file_operations perf_fops = {
3696 .llseek = no_llseek,
3697 .release = perf_release,
3700 .unlocked_ioctl = perf_ioctl,
3701 .compat_ioctl = perf_ioctl,
3703 .fasync = perf_fasync,
3709 * If there's data, ensure we set the poll() state and publish everything
3710 * to user-space before waking everybody up.
3713 void perf_event_wakeup(struct perf_event *event)
3715 ring_buffer_wakeup(event);
3717 if (event->pending_kill) {
3718 kill_fasync(&event->fasync, SIGIO, event->pending_kill);
3719 event->pending_kill = 0;
3723 static void perf_pending_event(struct irq_work *entry)
3725 struct perf_event *event = container_of(entry,
3726 struct perf_event, pending);
3728 if (event->pending_disable) {
3729 event->pending_disable = 0;
3730 __perf_event_disable(event);
3733 if (event->pending_wakeup) {
3734 event->pending_wakeup = 0;
3735 perf_event_wakeup(event);
3740 * We assume there is only KVM supporting the callbacks.
3741 * Later on, we might change it to a list if there is
3742 * another virtualization implementation supporting the callbacks.
3744 struct perf_guest_info_callbacks *perf_guest_cbs;
3746 int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
3748 perf_guest_cbs = cbs;
3751 EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks);
3753 int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
3755 perf_guest_cbs = NULL;
3758 EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks);
3761 perf_output_sample_regs(struct perf_output_handle *handle,
3762 struct pt_regs *regs, u64 mask)
3766 for_each_set_bit(bit, (const unsigned long *) &mask,
3767 sizeof(mask) * BITS_PER_BYTE) {
3770 val = perf_reg_value(regs, bit);
3771 perf_output_put(handle, val);
3775 static void perf_sample_regs_user(struct perf_regs_user *regs_user,
3776 struct pt_regs *regs)
3778 if (!user_mode(regs)) {
3780 regs = task_pt_regs(current);
3786 regs_user->regs = regs;
3787 regs_user->abi = perf_reg_abi(current);
3792 * Get remaining task size from user stack pointer.
3794 * It'd be better to take stack vma map and limit this more
3795 * precisly, but there's no way to get it safely under interrupt,
3796 * so using TASK_SIZE as limit.
3798 static u64 perf_ustack_task_size(struct pt_regs *regs)
3800 unsigned long addr = perf_user_stack_pointer(regs);
3802 if (!addr || addr >= TASK_SIZE)
3805 return TASK_SIZE - addr;
3809 perf_sample_ustack_size(u16 stack_size, u16 header_size,
3810 struct pt_regs *regs)
3814 /* No regs, no stack pointer, no dump. */
3819 * Check if we fit in with the requested stack size into the:
3821 * If we don't, we limit the size to the TASK_SIZE.
3823 * - remaining sample size
3824 * If we don't, we customize the stack size to
3825 * fit in to the remaining sample size.
3828 task_size = min((u64) USHRT_MAX, perf_ustack_task_size(regs));
3829 stack_size = min(stack_size, (u16) task_size);
3831 /* Current header size plus static size and dynamic size. */
3832 header_size += 2 * sizeof(u64);
3834 /* Do we fit in with the current stack dump size? */
3835 if ((u16) (header_size + stack_size) < header_size) {
3837 * If we overflow the maximum size for the sample,
3838 * we customize the stack dump size to fit in.
3840 stack_size = USHRT_MAX - header_size - sizeof(u64);
3841 stack_size = round_up(stack_size, sizeof(u64));
3848 perf_output_sample_ustack(struct perf_output_handle *handle, u64 dump_size,
3849 struct pt_regs *regs)
3851 /* Case of a kernel thread, nothing to dump */
3854 perf_output_put(handle, size);
3863 * - the size requested by user or the best one we can fit
3864 * in to the sample max size
3866 * - user stack dump data
3868 * - the actual dumped size
3872 perf_output_put(handle, dump_size);
3875 sp = perf_user_stack_pointer(regs);
3876 rem = __output_copy_user(handle, (void *) sp, dump_size);
3877 dyn_size = dump_size - rem;
3879 perf_output_skip(handle, rem);
3882 perf_output_put(handle, dyn_size);
3886 static void __perf_event_header__init_id(struct perf_event_header *header,
3887 struct perf_sample_data *data,
3888 struct perf_event *event)
3890 u64 sample_type = event->attr.sample_type;
3892 data->type = sample_type;
3893 header->size += event->id_header_size;
3895 if (sample_type & PERF_SAMPLE_TID) {
3896 /* namespace issues */
3897 data->tid_entry.pid = perf_event_pid(event, current);
3898 data->tid_entry.tid = perf_event_tid(event, current);
3901 if (sample_type & PERF_SAMPLE_TIME)
3902 data->time = perf_clock();
3904 if (sample_type & PERF_SAMPLE_ID)
3905 data->id = primary_event_id(event);
3907 if (sample_type & PERF_SAMPLE_STREAM_ID)
3908 data->stream_id = event->id;
3910 if (sample_type & PERF_SAMPLE_CPU) {
3911 data->cpu_entry.cpu = raw_smp_processor_id();
3912 data->cpu_entry.reserved = 0;
3916 void perf_event_header__init_id(struct perf_event_header *header,
3917 struct perf_sample_data *data,
3918 struct perf_event *event)
3920 if (event->attr.sample_id_all)
3921 __perf_event_header__init_id(header, data, event);
3924 static void __perf_event__output_id_sample(struct perf_output_handle *handle,
3925 struct perf_sample_data *data)
3927 u64 sample_type = data->type;
3929 if (sample_type & PERF_SAMPLE_TID)
3930 perf_output_put(handle, data->tid_entry);
3932 if (sample_type & PERF_SAMPLE_TIME)
3933 perf_output_put(handle, data->time);
3935 if (sample_type & PERF_SAMPLE_ID)
3936 perf_output_put(handle, data->id);
3938 if (sample_type & PERF_SAMPLE_STREAM_ID)
3939 perf_output_put(handle, data->stream_id);
3941 if (sample_type & PERF_SAMPLE_CPU)
3942 perf_output_put(handle, data->cpu_entry);
3945 void perf_event__output_id_sample(struct perf_event *event,
3946 struct perf_output_handle *handle,
3947 struct perf_sample_data *sample)
3949 if (event->attr.sample_id_all)
3950 __perf_event__output_id_sample(handle, sample);
3953 static void perf_output_read_one(struct perf_output_handle *handle,
3954 struct perf_event *event,
3955 u64 enabled, u64 running)
3957 u64 read_format = event->attr.read_format;
3961 values[n++] = perf_event_count(event);
3962 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
3963 values[n++] = enabled +
3964 atomic64_read(&event->child_total_time_enabled);
3966 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
3967 values[n++] = running +
3968 atomic64_read(&event->child_total_time_running);
3970 if (read_format & PERF_FORMAT_ID)
3971 values[n++] = primary_event_id(event);
3973 __output_copy(handle, values, n * sizeof(u64));
3977 * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
3979 static void perf_output_read_group(struct perf_output_handle *handle,
3980 struct perf_event *event,
3981 u64 enabled, u64 running)
3983 struct perf_event *leader = event->group_leader, *sub;
3984 u64 read_format = event->attr.read_format;
3988 values[n++] = 1 + leader->nr_siblings;
3990 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3991 values[n++] = enabled;
3993 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3994 values[n++] = running;
3996 if (leader != event)
3997 leader->pmu->read(leader);
3999 values[n++] = perf_event_count(leader);
4000 if (read_format & PERF_FORMAT_ID)
4001 values[n++] = primary_event_id(leader);
4003 __output_copy(handle, values, n * sizeof(u64));
4005 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
4009 sub->pmu->read(sub);
4011 values[n++] = perf_event_count(sub);
4012 if (read_format & PERF_FORMAT_ID)
4013 values[n++] = primary_event_id(sub);
4015 __output_copy(handle, values, n * sizeof(u64));
4019 #define PERF_FORMAT_TOTAL_TIMES (PERF_FORMAT_TOTAL_TIME_ENABLED|\
4020 PERF_FORMAT_TOTAL_TIME_RUNNING)
4022 static void perf_output_read(struct perf_output_handle *handle,
4023 struct perf_event *event)
4025 u64 enabled = 0, running = 0, now;
4026 u64 read_format = event->attr.read_format;
4029 * compute total_time_enabled, total_time_running
4030 * based on snapshot values taken when the event
4031 * was last scheduled in.
4033 * we cannot simply called update_context_time()
4034 * because of locking issue as we are called in
4037 if (read_format & PERF_FORMAT_TOTAL_TIMES)
4038 calc_timer_values(event, &now, &enabled, &running);
4040 if (event->attr.read_format & PERF_FORMAT_GROUP)
4041 perf_output_read_group(handle, event, enabled, running);
4043 perf_output_read_one(handle, event, enabled, running);
4046 void perf_output_sample(struct perf_output_handle *handle,
4047 struct perf_event_header *header,
4048 struct perf_sample_data *data,
4049 struct perf_event *event)
4051 u64 sample_type = data->type;
4053 perf_output_put(handle, *header);
4055 if (sample_type & PERF_SAMPLE_IP)
4056 perf_output_put(handle, data->ip);
4058 if (sample_type & PERF_SAMPLE_TID)
4059 perf_output_put(handle, data->tid_entry);
4061 if (sample_type & PERF_SAMPLE_TIME)
4062 perf_output_put(handle, data->time);
4064 if (sample_type & PERF_SAMPLE_ADDR)
4065 perf_output_put(handle, data->addr);
4067 if (sample_type & PERF_SAMPLE_ID)
4068 perf_output_put(handle, data->id);
4070 if (sample_type & PERF_SAMPLE_STREAM_ID)
4071 perf_output_put(handle, data->stream_id);
4073 if (sample_type & PERF_SAMPLE_CPU)
4074 perf_output_put(handle, data->cpu_entry);
4076 if (sample_type & PERF_SAMPLE_PERIOD)
4077 perf_output_put(handle, data->period);
4079 if (sample_type & PERF_SAMPLE_READ)
4080 perf_output_read(handle, event);
4082 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
4083 if (data->callchain) {
4086 if (data->callchain)
4087 size += data->callchain->nr;
4089 size *= sizeof(u64);
4091 __output_copy(handle, data->callchain, size);
4094 perf_output_put(handle, nr);
4098 if (sample_type & PERF_SAMPLE_RAW) {
4100 perf_output_put(handle, data->raw->size);
4101 __output_copy(handle, data->raw->data,
4108 .size = sizeof(u32),
4111 perf_output_put(handle, raw);
4115 if (!event->attr.watermark) {
4116 int wakeup_events = event->attr.wakeup_events;
4118 if (wakeup_events) {
4119 struct ring_buffer *rb = handle->rb;
4120 int events = local_inc_return(&rb->events);
4122 if (events >= wakeup_events) {
4123 local_sub(wakeup_events, &rb->events);
4124 local_inc(&rb->wakeup);
4129 if (sample_type & PERF_SAMPLE_BRANCH_STACK) {
4130 if (data->br_stack) {
4133 size = data->br_stack->nr
4134 * sizeof(struct perf_branch_entry);
4136 perf_output_put(handle, data->br_stack->nr);
4137 perf_output_copy(handle, data->br_stack->entries, size);
4140 * we always store at least the value of nr
4143 perf_output_put(handle, nr);
4147 if (sample_type & PERF_SAMPLE_REGS_USER) {
4148 u64 abi = data->regs_user.abi;
4151 * If there are no regs to dump, notice it through
4152 * first u64 being zero (PERF_SAMPLE_REGS_ABI_NONE).
4154 perf_output_put(handle, abi);
4157 u64 mask = event->attr.sample_regs_user;
4158 perf_output_sample_regs(handle,
4159 data->regs_user.regs,
4164 if (sample_type & PERF_SAMPLE_STACK_USER)
4165 perf_output_sample_ustack(handle,
4166 data->stack_user_size,
4167 data->regs_user.regs);
4170 void perf_prepare_sample(struct perf_event_header *header,
4171 struct perf_sample_data *data,
4172 struct perf_event *event,
4173 struct pt_regs *regs)
4175 u64 sample_type = event->attr.sample_type;
4177 header->type = PERF_RECORD_SAMPLE;
4178 header->size = sizeof(*header) + event->header_size;
4181 header->misc |= perf_misc_flags(regs);
4183 __perf_event_header__init_id(header, data, event);
4185 if (sample_type & PERF_SAMPLE_IP)
4186 data->ip = perf_instruction_pointer(regs);
4188 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
4191 data->callchain = perf_callchain(event, regs);
4193 if (data->callchain)
4194 size += data->callchain->nr;
4196 header->size += size * sizeof(u64);
4199 if (sample_type & PERF_SAMPLE_RAW) {
4200 int size = sizeof(u32);
4203 size += data->raw->size;
4205 size += sizeof(u32);
4207 WARN_ON_ONCE(size & (sizeof(u64)-1));
4208 header->size += size;
4211 if (sample_type & PERF_SAMPLE_BRANCH_STACK) {
4212 int size = sizeof(u64); /* nr */
4213 if (data->br_stack) {
4214 size += data->br_stack->nr
4215 * sizeof(struct perf_branch_entry);
4217 header->size += size;
4220 if (sample_type & PERF_SAMPLE_REGS_USER) {
4221 /* regs dump ABI info */
4222 int size = sizeof(u64);
4224 perf_sample_regs_user(&data->regs_user, regs);
4226 if (data->regs_user.regs) {
4227 u64 mask = event->attr.sample_regs_user;
4228 size += hweight64(mask) * sizeof(u64);
4231 header->size += size;
4234 if (sample_type & PERF_SAMPLE_STACK_USER) {
4236 * Either we need PERF_SAMPLE_STACK_USER bit to be allways
4237 * processed as the last one or have additional check added
4238 * in case new sample type is added, because we could eat
4239 * up the rest of the sample size.
4241 struct perf_regs_user *uregs = &data->regs_user;
4242 u16 stack_size = event->attr.sample_stack_user;
4243 u16 size = sizeof(u64);
4246 perf_sample_regs_user(uregs, regs);
4248 stack_size = perf_sample_ustack_size(stack_size, header->size,
4252 * If there is something to dump, add space for the dump
4253 * itself and for the field that tells the dynamic size,
4254 * which is how many have been actually dumped.
4257 size += sizeof(u64) + stack_size;
4259 data->stack_user_size = stack_size;
4260 header->size += size;
4264 static void perf_event_output(struct perf_event *event,
4265 struct perf_sample_data *data,
4266 struct pt_regs *regs)
4268 struct perf_output_handle handle;
4269 struct perf_event_header header;
4271 /* protect the callchain buffers */
4274 perf_prepare_sample(&header, data, event, regs);
4276 if (perf_output_begin(&handle, event, header.size))
4279 perf_output_sample(&handle, &header, data, event);
4281 perf_output_end(&handle);
4291 struct perf_read_event {
4292 struct perf_event_header header;
4299 perf_event_read_event(struct perf_event *event,
4300 struct task_struct *task)
4302 struct perf_output_handle handle;
4303 struct perf_sample_data sample;
4304 struct perf_read_event read_event = {
4306 .type = PERF_RECORD_READ,
4308 .size = sizeof(read_event) + event->read_size,
4310 .pid = perf_event_pid(event, task),
4311 .tid = perf_event_tid(event, task),
4315 perf_event_header__init_id(&read_event.header, &sample, event);
4316 ret = perf_output_begin(&handle, event, read_event.header.size);
4320 perf_output_put(&handle, read_event);
4321 perf_output_read(&handle, event);
4322 perf_event__output_id_sample(event, &handle, &sample);
4324 perf_output_end(&handle);
4328 * task tracking -- fork/exit
4330 * enabled by: attr.comm | attr.mmap | attr.mmap_data | attr.task
4333 struct perf_task_event {
4334 struct task_struct *task;
4335 struct perf_event_context *task_ctx;
4338 struct perf_event_header header;
4348 static void perf_event_task_output(struct perf_event *event,
4349 struct perf_task_event *task_event)
4351 struct perf_output_handle handle;
4352 struct perf_sample_data sample;
4353 struct task_struct *task = task_event->task;
4354 int ret, size = task_event->event_id.header.size;
4356 perf_event_header__init_id(&task_event->event_id.header, &sample, event);
4358 ret = perf_output_begin(&handle, event,
4359 task_event->event_id.header.size);
4363 task_event->event_id.pid = perf_event_pid(event, task);
4364 task_event->event_id.ppid = perf_event_pid(event, current);
4366 task_event->event_id.tid = perf_event_tid(event, task);
4367 task_event->event_id.ptid = perf_event_tid(event, current);
4369 perf_output_put(&handle, task_event->event_id);
4371 perf_event__output_id_sample(event, &handle, &sample);
4373 perf_output_end(&handle);
4375 task_event->event_id.header.size = size;
4378 static int perf_event_task_match(struct perf_event *event)
4380 if (event->state < PERF_EVENT_STATE_INACTIVE)
4383 if (!event_filter_match(event))
4386 if (event->attr.comm || event->attr.mmap ||
4387 event->attr.mmap_data || event->attr.task)
4393 static void perf_event_task_ctx(struct perf_event_context *ctx,
4394 struct perf_task_event *task_event)
4396 struct perf_event *event;
4398 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
4399 if (perf_event_task_match(event))
4400 perf_event_task_output(event, task_event);
4404 static void perf_event_task_event(struct perf_task_event *task_event)
4406 struct perf_cpu_context *cpuctx;
4407 struct perf_event_context *ctx;
4412 list_for_each_entry_rcu(pmu, &pmus, entry) {
4413 cpuctx = get_cpu_ptr(pmu->pmu_cpu_context);
4414 if (cpuctx->active_pmu != pmu)
4416 perf_event_task_ctx(&cpuctx->ctx, task_event);
4418 ctx = task_event->task_ctx;
4420 ctxn = pmu->task_ctx_nr;
4423 ctx = rcu_dereference(current->perf_event_ctxp[ctxn]);
4426 perf_event_task_ctx(ctx, task_event);
4428 put_cpu_ptr(pmu->pmu_cpu_context);
4433 static void perf_event_task(struct task_struct *task,
4434 struct perf_event_context *task_ctx,
4437 struct perf_task_event task_event;
4439 if (!atomic_read(&nr_comm_events) &&
4440 !atomic_read(&nr_mmap_events) &&
4441 !atomic_read(&nr_task_events))
4444 task_event = (struct perf_task_event){
4446 .task_ctx = task_ctx,
4449 .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT,
4451 .size = sizeof(task_event.event_id),
4457 .time = perf_clock(),
4461 perf_event_task_event(&task_event);
4464 void perf_event_fork(struct task_struct *task)
4466 perf_event_task(task, NULL, 1);
4473 struct perf_comm_event {
4474 struct task_struct *task;
4479 struct perf_event_header header;
4486 static void perf_event_comm_output(struct perf_event *event,
4487 struct perf_comm_event *comm_event)
4489 struct perf_output_handle handle;
4490 struct perf_sample_data sample;
4491 int size = comm_event->event_id.header.size;
4494 perf_event_header__init_id(&comm_event->event_id.header, &sample, event);
4495 ret = perf_output_begin(&handle, event,
4496 comm_event->event_id.header.size);
4501 comm_event->event_id.pid = perf_event_pid(event, comm_event->task);
4502 comm_event->event_id.tid = perf_event_tid(event, comm_event->task);
4504 perf_output_put(&handle, comm_event->event_id);
4505 __output_copy(&handle, comm_event->comm,
4506 comm_event->comm_size);
4508 perf_event__output_id_sample(event, &handle, &sample);
4510 perf_output_end(&handle);
4512 comm_event->event_id.header.size = size;
4515 static int perf_event_comm_match(struct perf_event *event)
4517 if (event->state < PERF_EVENT_STATE_INACTIVE)
4520 if (!event_filter_match(event))
4523 if (event->attr.comm)
4529 static void perf_event_comm_ctx(struct perf_event_context *ctx,
4530 struct perf_comm_event *comm_event)
4532 struct perf_event *event;
4534 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
4535 if (perf_event_comm_match(event))
4536 perf_event_comm_output(event, comm_event);
4540 static void perf_event_comm_event(struct perf_comm_event *comm_event)
4542 struct perf_cpu_context *cpuctx;
4543 struct perf_event_context *ctx;
4544 char comm[TASK_COMM_LEN];
4549 memset(comm, 0, sizeof(comm));
4550 strlcpy(comm, comm_event->task->comm, sizeof(comm));
4551 size = ALIGN(strlen(comm)+1, sizeof(u64));
4553 comm_event->comm = comm;
4554 comm_event->comm_size = size;
4556 comm_event->event_id.header.size = sizeof(comm_event->event_id) + size;
4558 list_for_each_entry_rcu(pmu, &pmus, entry) {
4559 cpuctx = get_cpu_ptr(pmu->pmu_cpu_context);
4560 if (cpuctx->active_pmu != pmu)
4562 perf_event_comm_ctx(&cpuctx->ctx, comm_event);
4564 ctxn = pmu->task_ctx_nr;
4568 ctx = rcu_dereference(current->perf_event_ctxp[ctxn]);
4570 perf_event_comm_ctx(ctx, comm_event);
4572 put_cpu_ptr(pmu->pmu_cpu_context);
4577 void perf_event_comm(struct task_struct *task)
4579 struct perf_comm_event comm_event;
4580 struct perf_event_context *ctx;
4583 for_each_task_context_nr(ctxn) {
4584 ctx = task->perf_event_ctxp[ctxn];
4588 perf_event_enable_on_exec(ctx);
4591 if (!atomic_read(&nr_comm_events))
4594 comm_event = (struct perf_comm_event){
4600 .type = PERF_RECORD_COMM,
4609 perf_event_comm_event(&comm_event);
4616 struct perf_mmap_event {
4617 struct vm_area_struct *vma;
4619 const char *file_name;
4623 struct perf_event_header header;
4633 static void perf_event_mmap_output(struct perf_event *event,
4634 struct perf_mmap_event *mmap_event)
4636 struct perf_output_handle handle;
4637 struct perf_sample_data sample;
4638 int size = mmap_event->event_id.header.size;
4641 perf_event_header__init_id(&mmap_event->event_id.header, &sample, event);
4642 ret = perf_output_begin(&handle, event,
4643 mmap_event->event_id.header.size);
4647 mmap_event->event_id.pid = perf_event_pid(event, current);
4648 mmap_event->event_id.tid = perf_event_tid(event, current);
4650 perf_output_put(&handle, mmap_event->event_id);
4651 __output_copy(&handle, mmap_event->file_name,
4652 mmap_event->file_size);
4654 perf_event__output_id_sample(event, &handle, &sample);
4656 perf_output_end(&handle);
4658 mmap_event->event_id.header.size = size;
4661 static int perf_event_mmap_match(struct perf_event *event,
4662 struct perf_mmap_event *mmap_event,
4665 if (event->state < PERF_EVENT_STATE_INACTIVE)
4668 if (!event_filter_match(event))
4671 if ((!executable && event->attr.mmap_data) ||
4672 (executable && event->attr.mmap))
4678 static void perf_event_mmap_ctx(struct perf_event_context *ctx,
4679 struct perf_mmap_event *mmap_event,
4682 struct perf_event *event;
4684 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
4685 if (perf_event_mmap_match(event, mmap_event, executable))
4686 perf_event_mmap_output(event, mmap_event);
4690 static void perf_event_mmap_event(struct perf_mmap_event *mmap_event)
4692 struct perf_cpu_context *cpuctx;
4693 struct perf_event_context *ctx;
4694 struct vm_area_struct *vma = mmap_event->vma;
4695 struct file *file = vma->vm_file;
4703 memset(tmp, 0, sizeof(tmp));
4707 * d_path works from the end of the rb backwards, so we
4708 * need to add enough zero bytes after the string to handle
4709 * the 64bit alignment we do later.
4711 buf = kzalloc(PATH_MAX + sizeof(u64), GFP_KERNEL);
4713 name = strncpy(tmp, "//enomem", sizeof(tmp));
4716 name = d_path(&file->f_path, buf, PATH_MAX);
4718 name = strncpy(tmp, "//toolong", sizeof(tmp));
4722 if (arch_vma_name(mmap_event->vma)) {
4723 name = strncpy(tmp, arch_vma_name(mmap_event->vma),
4729 name = strncpy(tmp, "[vdso]", sizeof(tmp));
4731 } else if (vma->vm_start <= vma->vm_mm->start_brk &&
4732 vma->vm_end >= vma->vm_mm->brk) {
4733 name = strncpy(tmp, "[heap]", sizeof(tmp));
4735 } else if (vma->vm_start <= vma->vm_mm->start_stack &&
4736 vma->vm_end >= vma->vm_mm->start_stack) {
4737 name = strncpy(tmp, "[stack]", sizeof(tmp));
4741 name = strncpy(tmp, "//anon", sizeof(tmp));
4746 size = ALIGN(strlen(name)+1, sizeof(u64));
4748 mmap_event->file_name = name;
4749 mmap_event->file_size = size;
4751 mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size;
4754 list_for_each_entry_rcu(pmu, &pmus, entry) {
4755 cpuctx = get_cpu_ptr(pmu->pmu_cpu_context);
4756 if (cpuctx->active_pmu != pmu)
4758 perf_event_mmap_ctx(&cpuctx->ctx, mmap_event,
4759 vma->vm_flags & VM_EXEC);
4761 ctxn = pmu->task_ctx_nr;
4765 ctx = rcu_dereference(current->perf_event_ctxp[ctxn]);
4767 perf_event_mmap_ctx(ctx, mmap_event,
4768 vma->vm_flags & VM_EXEC);
4771 put_cpu_ptr(pmu->pmu_cpu_context);
4778 void perf_event_mmap(struct vm_area_struct *vma)
4780 struct perf_mmap_event mmap_event;
4782 if (!atomic_read(&nr_mmap_events))
4785 mmap_event = (struct perf_mmap_event){
4791 .type = PERF_RECORD_MMAP,
4792 .misc = PERF_RECORD_MISC_USER,
4797 .start = vma->vm_start,
4798 .len = vma->vm_end - vma->vm_start,
4799 .pgoff = (u64)vma->vm_pgoff << PAGE_SHIFT,
4803 perf_event_mmap_event(&mmap_event);
4807 * IRQ throttle logging
4810 static void perf_log_throttle(struct perf_event *event, int enable)
4812 struct perf_output_handle handle;
4813 struct perf_sample_data sample;
4817 struct perf_event_header header;
4821 } throttle_event = {
4823 .type = PERF_RECORD_THROTTLE,
4825 .size = sizeof(throttle_event),
4827 .time = perf_clock(),
4828 .id = primary_event_id(event),
4829 .stream_id = event->id,
4833 throttle_event.header.type = PERF_RECORD_UNTHROTTLE;
4835 perf_event_header__init_id(&throttle_event.header, &sample, event);
4837 ret = perf_output_begin(&handle, event,
4838 throttle_event.header.size);
4842 perf_output_put(&handle, throttle_event);
4843 perf_event__output_id_sample(event, &handle, &sample);
4844 perf_output_end(&handle);
4848 * Generic event overflow handling, sampling.
4851 static int __perf_event_overflow(struct perf_event *event,
4852 int throttle, struct perf_sample_data *data,
4853 struct pt_regs *regs)
4855 int events = atomic_read(&event->event_limit);
4856 struct hw_perf_event *hwc = &event->hw;
4861 * Non-sampling counters might still use the PMI to fold short
4862 * hardware counters, ignore those.
4864 if (unlikely(!is_sampling_event(event)))
4867 seq = __this_cpu_read(perf_throttled_seq);
4868 if (seq != hwc->interrupts_seq) {
4869 hwc->interrupts_seq = seq;
4870 hwc->interrupts = 1;
4873 if (unlikely(throttle
4874 && hwc->interrupts >= max_samples_per_tick)) {
4875 __this_cpu_inc(perf_throttled_count);
4876 hwc->interrupts = MAX_INTERRUPTS;
4877 perf_log_throttle(event, 0);
4882 if (event->attr.freq) {
4883 u64 now = perf_clock();
4884 s64 delta = now - hwc->freq_time_stamp;
4886 hwc->freq_time_stamp = now;
4888 if (delta > 0 && delta < 2*TICK_NSEC)
4889 perf_adjust_period(event, delta, hwc->last_period, true);
4893 * XXX event_limit might not quite work as expected on inherited
4897 event->pending_kill = POLL_IN;
4898 if (events && atomic_dec_and_test(&event->event_limit)) {
4900 event->pending_kill = POLL_HUP;
4901 event->pending_disable = 1;
4902 irq_work_queue(&event->pending);
4905 if (event->overflow_handler)
4906 event->overflow_handler(event, data, regs);
4908 perf_event_output(event, data, regs);
4910 if (event->fasync && event->pending_kill) {
4911 event->pending_wakeup = 1;
4912 irq_work_queue(&event->pending);
4918 int perf_event_overflow(struct perf_event *event,
4919 struct perf_sample_data *data,
4920 struct pt_regs *regs)
4922 return __perf_event_overflow(event, 1, data, regs);
4926 * Generic software event infrastructure
4929 struct swevent_htable {
4930 struct swevent_hlist *swevent_hlist;
4931 struct mutex hlist_mutex;
4934 /* Recursion avoidance in each contexts */
4935 int recursion[PERF_NR_CONTEXTS];
4938 static DEFINE_PER_CPU(struct swevent_htable, swevent_htable);
4941 * We directly increment event->count and keep a second value in
4942 * event->hw.period_left to count intervals. This period event
4943 * is kept in the range [-sample_period, 0] so that we can use the
4947 static u64 perf_swevent_set_period(struct perf_event *event)
4949 struct hw_perf_event *hwc = &event->hw;
4950 u64 period = hwc->last_period;
4954 hwc->last_period = hwc->sample_period;
4957 old = val = local64_read(&hwc->period_left);
4961 nr = div64_u64(period + val, period);
4962 offset = nr * period;
4964 if (local64_cmpxchg(&hwc->period_left, old, val) != old)
4970 static void perf_swevent_overflow(struct perf_event *event, u64 overflow,
4971 struct perf_sample_data *data,
4972 struct pt_regs *regs)
4974 struct hw_perf_event *hwc = &event->hw;
4978 overflow = perf_swevent_set_period(event);
4980 if (hwc->interrupts == MAX_INTERRUPTS)
4983 for (; overflow; overflow--) {
4984 if (__perf_event_overflow(event, throttle,
4987 * We inhibit the overflow from happening when
4988 * hwc->interrupts == MAX_INTERRUPTS.
4996 static void perf_swevent_event(struct perf_event *event, u64 nr,
4997 struct perf_sample_data *data,
4998 struct pt_regs *regs)
5000 struct hw_perf_event *hwc = &event->hw;
5002 local64_add(nr, &event->count);
5007 if (!is_sampling_event(event))
5010 if ((event->attr.sample_type & PERF_SAMPLE_PERIOD) && !event->attr.freq) {
5012 return perf_swevent_overflow(event, 1, data, regs);
5014 data->period = event->hw.last_period;
5016 if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq)
5017 return perf_swevent_overflow(event, 1, data, regs);
5019 if (local64_add_negative(nr, &hwc->period_left))
5022 perf_swevent_overflow(event, 0, data, regs);
5025 static int perf_exclude_event(struct perf_event *event,
5026 struct pt_regs *regs)
5028 if (event->hw.state & PERF_HES_STOPPED)
5032 if (event->attr.exclude_user && user_mode(regs))
5035 if (event->attr.exclude_kernel && !user_mode(regs))
5042 static int perf_swevent_match(struct perf_event *event,
5043 enum perf_type_id type,
5045 struct perf_sample_data *data,
5046 struct pt_regs *regs)
5048 if (event->attr.type != type)
5051 if (event->attr.config != event_id)
5054 if (perf_exclude_event(event, regs))
5060 static inline u64 swevent_hash(u64 type, u32 event_id)
5062 u64 val = event_id | (type << 32);
5064 return hash_64(val, SWEVENT_HLIST_BITS);
5067 static inline struct hlist_head *
5068 __find_swevent_head(struct swevent_hlist *hlist, u64 type, u32 event_id)
5070 u64 hash = swevent_hash(type, event_id);
5072 return &hlist->heads[hash];
5075 /* For the read side: events when they trigger */
5076 static inline struct hlist_head *
5077 find_swevent_head_rcu(struct swevent_htable *swhash, u64 type, u32 event_id)
5079 struct swevent_hlist *hlist;
5081 hlist = rcu_dereference(swhash->swevent_hlist);
5085 return __find_swevent_head(hlist, type, event_id);
5088 /* For the event head insertion and removal in the hlist */
5089 static inline struct hlist_head *
5090 find_swevent_head(struct swevent_htable *swhash, struct perf_event *event)
5092 struct swevent_hlist *hlist;
5093 u32 event_id = event->attr.config;
5094 u64 type = event->attr.type;
5097 * Event scheduling is always serialized against hlist allocation
5098 * and release. Which makes the protected version suitable here.
5099 * The context lock guarantees that.
5101 hlist = rcu_dereference_protected(swhash->swevent_hlist,
5102 lockdep_is_held(&event->ctx->lock));
5106 return __find_swevent_head(hlist, type, event_id);
5109 static void do_perf_sw_event(enum perf_type_id type, u32 event_id,
5111 struct perf_sample_data *data,
5112 struct pt_regs *regs)
5114 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
5115 struct perf_event *event;
5116 struct hlist_node *node;
5117 struct hlist_head *head;
5120 head = find_swevent_head_rcu(swhash, type, event_id);
5124 hlist_for_each_entry_rcu(event, node, head, hlist_entry) {
5125 if (perf_swevent_match(event, type, event_id, data, regs))
5126 perf_swevent_event(event, nr, data, regs);
5132 int perf_swevent_get_recursion_context(void)
5134 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
5136 return get_recursion_context(swhash->recursion);
5138 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context);
5140 inline void perf_swevent_put_recursion_context(int rctx)
5142 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
5144 put_recursion_context(swhash->recursion, rctx);
5147 void __perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)
5149 struct perf_sample_data data;
5152 preempt_disable_notrace();
5153 rctx = perf_swevent_get_recursion_context();
5157 perf_sample_data_init(&data, addr, 0);
5159 do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, &data, regs);
5161 perf_swevent_put_recursion_context(rctx);
5162 preempt_enable_notrace();
5165 static void perf_swevent_read(struct perf_event *event)
5169 static int perf_swevent_add(struct perf_event *event, int flags)
5171 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
5172 struct hw_perf_event *hwc = &event->hw;
5173 struct hlist_head *head;
5175 if (is_sampling_event(event)) {
5176 hwc->last_period = hwc->sample_period;
5177 perf_swevent_set_period(event);
5180 hwc->state = !(flags & PERF_EF_START);
5182 head = find_swevent_head(swhash, event);
5183 if (WARN_ON_ONCE(!head))
5186 hlist_add_head_rcu(&event->hlist_entry, head);
5191 static void perf_swevent_del(struct perf_event *event, int flags)
5193 hlist_del_rcu(&event->hlist_entry);
5196 static void perf_swevent_start(struct perf_event *event, int flags)
5198 event->hw.state = 0;
5201 static void perf_swevent_stop(struct perf_event *event, int flags)
5203 event->hw.state = PERF_HES_STOPPED;
5206 /* Deref the hlist from the update side */
5207 static inline struct swevent_hlist *
5208 swevent_hlist_deref(struct swevent_htable *swhash)
5210 return rcu_dereference_protected(swhash->swevent_hlist,
5211 lockdep_is_held(&swhash->hlist_mutex));
5214 static void swevent_hlist_release(struct swevent_htable *swhash)
5216 struct swevent_hlist *hlist = swevent_hlist_deref(swhash);
5221 rcu_assign_pointer(swhash->swevent_hlist, NULL);
5222 kfree_rcu(hlist, rcu_head);
5225 static void swevent_hlist_put_cpu(struct perf_event *event, int cpu)
5227 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
5229 mutex_lock(&swhash->hlist_mutex);
5231 if (!--swhash->hlist_refcount)
5232 swevent_hlist_release(swhash);
5234 mutex_unlock(&swhash->hlist_mutex);
5237 static void swevent_hlist_put(struct perf_event *event)
5241 if (event->cpu != -1) {
5242 swevent_hlist_put_cpu(event, event->cpu);
5246 for_each_possible_cpu(cpu)
5247 swevent_hlist_put_cpu(event, cpu);
5250 static int swevent_hlist_get_cpu(struct perf_event *event, int cpu)
5252 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
5255 mutex_lock(&swhash->hlist_mutex);
5257 if (!swevent_hlist_deref(swhash) && cpu_online(cpu)) {
5258 struct swevent_hlist *hlist;
5260 hlist = kzalloc(sizeof(*hlist), GFP_KERNEL);
5265 rcu_assign_pointer(swhash->swevent_hlist, hlist);
5267 swhash->hlist_refcount++;
5269 mutex_unlock(&swhash->hlist_mutex);
5274 static int swevent_hlist_get(struct perf_event *event)
5277 int cpu, failed_cpu;
5279 if (event->cpu != -1)
5280 return swevent_hlist_get_cpu(event, event->cpu);
5283 for_each_possible_cpu(cpu) {
5284 err = swevent_hlist_get_cpu(event, cpu);
5294 for_each_possible_cpu(cpu) {
5295 if (cpu == failed_cpu)
5297 swevent_hlist_put_cpu(event, cpu);
5304 struct static_key perf_swevent_enabled[PERF_COUNT_SW_MAX];
5306 static void sw_perf_event_destroy(struct perf_event *event)
5308 u64 event_id = event->attr.config;
5310 WARN_ON(event->parent);
5312 static_key_slow_dec(&perf_swevent_enabled[event_id]);
5313 swevent_hlist_put(event);
5316 static int perf_swevent_init(struct perf_event *event)
5318 int event_id = event->attr.config;
5320 if (event->attr.type != PERF_TYPE_SOFTWARE)
5324 * no branch sampling for software events
5326 if (has_branch_stack(event))
5330 case PERF_COUNT_SW_CPU_CLOCK:
5331 case PERF_COUNT_SW_TASK_CLOCK:
5338 if (event_id >= PERF_COUNT_SW_MAX)
5341 if (!event->parent) {
5344 err = swevent_hlist_get(event);
5348 static_key_slow_inc(&perf_swevent_enabled[event_id]);
5349 event->destroy = sw_perf_event_destroy;
5355 static int perf_swevent_event_idx(struct perf_event *event)
5360 static struct pmu perf_swevent = {
5361 .task_ctx_nr = perf_sw_context,
5363 .event_init = perf_swevent_init,
5364 .add = perf_swevent_add,
5365 .del = perf_swevent_del,
5366 .start = perf_swevent_start,
5367 .stop = perf_swevent_stop,
5368 .read = perf_swevent_read,
5370 .event_idx = perf_swevent_event_idx,
5373 #ifdef CONFIG_EVENT_TRACING
5375 static int perf_tp_filter_match(struct perf_event *event,
5376 struct perf_sample_data *data)
5378 void *record = data->raw->data;
5380 if (likely(!event->filter) || filter_match_preds(event->filter, record))
5385 static int perf_tp_event_match(struct perf_event *event,
5386 struct perf_sample_data *data,
5387 struct pt_regs *regs)
5389 if (event->hw.state & PERF_HES_STOPPED)
5392 * All tracepoints are from kernel-space.
5394 if (event->attr.exclude_kernel)
5397 if (!perf_tp_filter_match(event, data))
5403 void perf_tp_event(u64 addr, u64 count, void *record, int entry_size,
5404 struct pt_regs *regs, struct hlist_head *head, int rctx,
5405 struct task_struct *task)
5407 struct perf_sample_data data;
5408 struct perf_event *event;
5409 struct hlist_node *node;
5411 struct perf_raw_record raw = {
5416 perf_sample_data_init(&data, addr, 0);
5419 hlist_for_each_entry_rcu(event, node, head, hlist_entry) {
5420 if (perf_tp_event_match(event, &data, regs))
5421 perf_swevent_event(event, count, &data, regs);
5425 * If we got specified a target task, also iterate its context and
5426 * deliver this event there too.
5428 if (task && task != current) {
5429 struct perf_event_context *ctx;
5430 struct trace_entry *entry = record;
5433 ctx = rcu_dereference(task->perf_event_ctxp[perf_sw_context]);
5437 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
5438 if (event->attr.type != PERF_TYPE_TRACEPOINT)
5440 if (event->attr.config != entry->type)
5442 if (perf_tp_event_match(event, &data, regs))
5443 perf_swevent_event(event, count, &data, regs);
5449 perf_swevent_put_recursion_context(rctx);
5451 EXPORT_SYMBOL_GPL(perf_tp_event);
5453 static void tp_perf_event_destroy(struct perf_event *event)
5455 perf_trace_destroy(event);
5458 static int perf_tp_event_init(struct perf_event *event)
5462 if (event->attr.type != PERF_TYPE_TRACEPOINT)
5466 * no branch sampling for tracepoint events
5468 if (has_branch_stack(event))
5471 err = perf_trace_init(event);
5475 event->destroy = tp_perf_event_destroy;
5480 static struct pmu perf_tracepoint = {
5481 .task_ctx_nr = perf_sw_context,
5483 .event_init = perf_tp_event_init,
5484 .add = perf_trace_add,
5485 .del = perf_trace_del,
5486 .start = perf_swevent_start,
5487 .stop = perf_swevent_stop,
5488 .read = perf_swevent_read,
5490 .event_idx = perf_swevent_event_idx,
5493 static inline void perf_tp_register(void)
5495 perf_pmu_register(&perf_tracepoint, "tracepoint", PERF_TYPE_TRACEPOINT);
5498 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
5503 if (event->attr.type != PERF_TYPE_TRACEPOINT)
5506 filter_str = strndup_user(arg, PAGE_SIZE);
5507 if (IS_ERR(filter_str))
5508 return PTR_ERR(filter_str);
5510 ret = ftrace_profile_set_filter(event, event->attr.config, filter_str);
5516 static void perf_event_free_filter(struct perf_event *event)
5518 ftrace_profile_free_filter(event);
5523 static inline void perf_tp_register(void)
5527 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
5532 static void perf_event_free_filter(struct perf_event *event)
5536 #endif /* CONFIG_EVENT_TRACING */
5538 #ifdef CONFIG_HAVE_HW_BREAKPOINT
5539 void perf_bp_event(struct perf_event *bp, void *data)
5541 struct perf_sample_data sample;
5542 struct pt_regs *regs = data;
5544 perf_sample_data_init(&sample, bp->attr.bp_addr, 0);
5546 if (!bp->hw.state && !perf_exclude_event(bp, regs))
5547 perf_swevent_event(bp, 1, &sample, regs);
5552 * hrtimer based swevent callback
5555 static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer)
5557 enum hrtimer_restart ret = HRTIMER_RESTART;
5558 struct perf_sample_data data;
5559 struct pt_regs *regs;
5560 struct perf_event *event;
5563 event = container_of(hrtimer, struct perf_event, hw.hrtimer);
5565 if (event->state != PERF_EVENT_STATE_ACTIVE)
5566 return HRTIMER_NORESTART;
5568 event->pmu->read(event);
5570 perf_sample_data_init(&data, 0, event->hw.last_period);
5571 regs = get_irq_regs();
5573 if (regs && !perf_exclude_event(event, regs)) {
5574 if (!(event->attr.exclude_idle && is_idle_task(current)))
5575 if (__perf_event_overflow(event, 1, &data, regs))
5576 ret = HRTIMER_NORESTART;
5579 period = max_t(u64, 10000, event->hw.sample_period);
5580 hrtimer_forward_now(hrtimer, ns_to_ktime(period));
5585 static void perf_swevent_start_hrtimer(struct perf_event *event)
5587 struct hw_perf_event *hwc = &event->hw;
5590 if (!is_sampling_event(event))
5593 period = local64_read(&hwc->period_left);
5598 local64_set(&hwc->period_left, 0);
5600 period = max_t(u64, 10000, hwc->sample_period);
5602 __hrtimer_start_range_ns(&hwc->hrtimer,
5603 ns_to_ktime(period), 0,
5604 HRTIMER_MODE_REL_PINNED, 0);
5607 static void perf_swevent_cancel_hrtimer(struct perf_event *event)
5609 struct hw_perf_event *hwc = &event->hw;
5611 if (is_sampling_event(event)) {
5612 ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer);
5613 local64_set(&hwc->period_left, ktime_to_ns(remaining));
5615 hrtimer_cancel(&hwc->hrtimer);
5619 static void perf_swevent_init_hrtimer(struct perf_event *event)
5621 struct hw_perf_event *hwc = &event->hw;
5623 if (!is_sampling_event(event))
5626 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
5627 hwc->hrtimer.function = perf_swevent_hrtimer;
5630 * Since hrtimers have a fixed rate, we can do a static freq->period
5631 * mapping and avoid the whole period adjust feedback stuff.
5633 if (event->attr.freq) {
5634 long freq = event->attr.sample_freq;
5636 event->attr.sample_period = NSEC_PER_SEC / freq;
5637 hwc->sample_period = event->attr.sample_period;
5638 local64_set(&hwc->period_left, hwc->sample_period);
5639 event->attr.freq = 0;
5644 * Software event: cpu wall time clock
5647 static void cpu_clock_event_update(struct perf_event *event)
5652 now = local_clock();
5653 prev = local64_xchg(&event->hw.prev_count, now);
5654 local64_add(now - prev, &event->count);
5657 static void cpu_clock_event_start(struct perf_event *event, int flags)
5659 local64_set(&event->hw.prev_count, local_clock());
5660 perf_swevent_start_hrtimer(event);
5663 static void cpu_clock_event_stop(struct perf_event *event, int flags)
5665 perf_swevent_cancel_hrtimer(event);
5666 cpu_clock_event_update(event);
5669 static int cpu_clock_event_add(struct perf_event *event, int flags)
5671 if (flags & PERF_EF_START)
5672 cpu_clock_event_start(event, flags);
5677 static void cpu_clock_event_del(struct perf_event *event, int flags)
5679 cpu_clock_event_stop(event, flags);
5682 static void cpu_clock_event_read(struct perf_event *event)
5684 cpu_clock_event_update(event);
5687 static int cpu_clock_event_init(struct perf_event *event)
5689 if (event->attr.type != PERF_TYPE_SOFTWARE)
5692 if (event->attr.config != PERF_COUNT_SW_CPU_CLOCK)
5696 * no branch sampling for software events
5698 if (has_branch_stack(event))
5701 perf_swevent_init_hrtimer(event);
5706 static struct pmu perf_cpu_clock = {
5707 .task_ctx_nr = perf_sw_context,
5709 .event_init = cpu_clock_event_init,
5710 .add = cpu_clock_event_add,
5711 .del = cpu_clock_event_del,
5712 .start = cpu_clock_event_start,
5713 .stop = cpu_clock_event_stop,
5714 .read = cpu_clock_event_read,
5716 .event_idx = perf_swevent_event_idx,
5720 * Software event: task time clock
5723 static void task_clock_event_update(struct perf_event *event, u64 now)
5728 prev = local64_xchg(&event->hw.prev_count, now);
5730 local64_add(delta, &event->count);
5733 static void task_clock_event_start(struct perf_event *event, int flags)
5735 local64_set(&event->hw.prev_count, event->ctx->time);
5736 perf_swevent_start_hrtimer(event);
5739 static void task_clock_event_stop(struct perf_event *event, int flags)
5741 perf_swevent_cancel_hrtimer(event);
5742 task_clock_event_update(event, event->ctx->time);
5745 static int task_clock_event_add(struct perf_event *event, int flags)
5747 if (flags & PERF_EF_START)
5748 task_clock_event_start(event, flags);
5753 static void task_clock_event_del(struct perf_event *event, int flags)
5755 task_clock_event_stop(event, PERF_EF_UPDATE);
5758 static void task_clock_event_read(struct perf_event *event)
5760 u64 now = perf_clock();
5761 u64 delta = now - event->ctx->timestamp;
5762 u64 time = event->ctx->time + delta;
5764 task_clock_event_update(event, time);
5767 static int task_clock_event_init(struct perf_event *event)
5769 if (event->attr.type != PERF_TYPE_SOFTWARE)
5772 if (event->attr.config != PERF_COUNT_SW_TASK_CLOCK)
5776 * no branch sampling for software events
5778 if (has_branch_stack(event))
5781 perf_swevent_init_hrtimer(event);
5786 static struct pmu perf_task_clock = {
5787 .task_ctx_nr = perf_sw_context,
5789 .event_init = task_clock_event_init,
5790 .add = task_clock_event_add,
5791 .del = task_clock_event_del,
5792 .start = task_clock_event_start,
5793 .stop = task_clock_event_stop,
5794 .read = task_clock_event_read,
5796 .event_idx = perf_swevent_event_idx,
5799 static void perf_pmu_nop_void(struct pmu *pmu)
5803 static int perf_pmu_nop_int(struct pmu *pmu)
5808 static void perf_pmu_start_txn(struct pmu *pmu)
5810 perf_pmu_disable(pmu);
5813 static int perf_pmu_commit_txn(struct pmu *pmu)
5815 perf_pmu_enable(pmu);
5819 static void perf_pmu_cancel_txn(struct pmu *pmu)
5821 perf_pmu_enable(pmu);
5824 static int perf_event_idx_default(struct perf_event *event)
5826 return event->hw.idx + 1;
5830 * Ensures all contexts with the same task_ctx_nr have the same
5831 * pmu_cpu_context too.
5833 static void *find_pmu_context(int ctxn)
5840 list_for_each_entry(pmu, &pmus, entry) {
5841 if (pmu->task_ctx_nr == ctxn)
5842 return pmu->pmu_cpu_context;
5848 static void update_pmu_context(struct pmu *pmu, struct pmu *old_pmu)
5852 for_each_possible_cpu(cpu) {
5853 struct perf_cpu_context *cpuctx;
5855 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
5857 if (cpuctx->active_pmu == old_pmu)
5858 cpuctx->active_pmu = pmu;
5862 static void free_pmu_context(struct pmu *pmu)
5866 mutex_lock(&pmus_lock);
5868 * Like a real lame refcount.
5870 list_for_each_entry(i, &pmus, entry) {
5871 if (i->pmu_cpu_context == pmu->pmu_cpu_context) {
5872 update_pmu_context(i, pmu);
5877 free_percpu(pmu->pmu_cpu_context);
5879 mutex_unlock(&pmus_lock);
5881 static struct idr pmu_idr;
5884 type_show(struct device *dev, struct device_attribute *attr, char *page)
5886 struct pmu *pmu = dev_get_drvdata(dev);
5888 return snprintf(page, PAGE_SIZE-1, "%d\n", pmu->type);
5891 static struct device_attribute pmu_dev_attrs[] = {
5896 static int pmu_bus_running;
5897 static struct bus_type pmu_bus = {
5898 .name = "event_source",
5899 .dev_attrs = pmu_dev_attrs,
5902 static void pmu_dev_release(struct device *dev)
5907 static int pmu_dev_alloc(struct pmu *pmu)
5911 pmu->dev = kzalloc(sizeof(struct device), GFP_KERNEL);
5915 pmu->dev->groups = pmu->attr_groups;
5916 device_initialize(pmu->dev);
5917 ret = dev_set_name(pmu->dev, "%s", pmu->name);
5921 dev_set_drvdata(pmu->dev, pmu);
5922 pmu->dev->bus = &pmu_bus;
5923 pmu->dev->release = pmu_dev_release;
5924 ret = device_add(pmu->dev);
5932 put_device(pmu->dev);
5936 static struct lock_class_key cpuctx_mutex;
5937 static struct lock_class_key cpuctx_lock;
5939 int perf_pmu_register(struct pmu *pmu, char *name, int type)
5943 mutex_lock(&pmus_lock);
5945 pmu->pmu_disable_count = alloc_percpu(int);
5946 if (!pmu->pmu_disable_count)
5955 int err = idr_pre_get(&pmu_idr, GFP_KERNEL);
5959 err = idr_get_new_above(&pmu_idr, pmu, PERF_TYPE_MAX, &type);
5967 if (pmu_bus_running) {
5968 ret = pmu_dev_alloc(pmu);
5974 pmu->pmu_cpu_context = find_pmu_context(pmu->task_ctx_nr);
5975 if (pmu->pmu_cpu_context)
5976 goto got_cpu_context;
5978 pmu->pmu_cpu_context = alloc_percpu(struct perf_cpu_context);
5979 if (!pmu->pmu_cpu_context)
5982 for_each_possible_cpu(cpu) {
5983 struct perf_cpu_context *cpuctx;
5985 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
5986 __perf_event_init_context(&cpuctx->ctx);
5987 lockdep_set_class(&cpuctx->ctx.mutex, &cpuctx_mutex);
5988 lockdep_set_class(&cpuctx->ctx.lock, &cpuctx_lock);
5989 cpuctx->ctx.type = cpu_context;
5990 cpuctx->ctx.pmu = pmu;
5991 cpuctx->jiffies_interval = 1;
5992 INIT_LIST_HEAD(&cpuctx->rotation_list);
5993 cpuctx->active_pmu = pmu;
5997 if (!pmu->start_txn) {
5998 if (pmu->pmu_enable) {
6000 * If we have pmu_enable/pmu_disable calls, install
6001 * transaction stubs that use that to try and batch
6002 * hardware accesses.
6004 pmu->start_txn = perf_pmu_start_txn;
6005 pmu->commit_txn = perf_pmu_commit_txn;
6006 pmu->cancel_txn = perf_pmu_cancel_txn;
6008 pmu->start_txn = perf_pmu_nop_void;
6009 pmu->commit_txn = perf_pmu_nop_int;
6010 pmu->cancel_txn = perf_pmu_nop_void;
6014 if (!pmu->pmu_enable) {
6015 pmu->pmu_enable = perf_pmu_nop_void;
6016 pmu->pmu_disable = perf_pmu_nop_void;
6019 if (!pmu->event_idx)
6020 pmu->event_idx = perf_event_idx_default;
6022 list_add_rcu(&pmu->entry, &pmus);
6025 mutex_unlock(&pmus_lock);
6030 device_del(pmu->dev);
6031 put_device(pmu->dev);
6034 if (pmu->type >= PERF_TYPE_MAX)
6035 idr_remove(&pmu_idr, pmu->type);
6038 free_percpu(pmu->pmu_disable_count);
6042 void perf_pmu_unregister(struct pmu *pmu)
6044 mutex_lock(&pmus_lock);
6045 list_del_rcu(&pmu->entry);
6046 mutex_unlock(&pmus_lock);
6049 * We dereference the pmu list under both SRCU and regular RCU, so
6050 * synchronize against both of those.
6052 synchronize_srcu(&pmus_srcu);
6055 free_percpu(pmu->pmu_disable_count);
6056 if (pmu->type >= PERF_TYPE_MAX)
6057 idr_remove(&pmu_idr, pmu->type);
6058 device_del(pmu->dev);
6059 put_device(pmu->dev);
6060 free_pmu_context(pmu);
6063 struct pmu *perf_init_event(struct perf_event *event)
6065 struct pmu *pmu = NULL;
6069 idx = srcu_read_lock(&pmus_srcu);
6072 pmu = idr_find(&pmu_idr, event->attr.type);
6076 ret = pmu->event_init(event);
6082 list_for_each_entry_rcu(pmu, &pmus, entry) {
6084 ret = pmu->event_init(event);
6088 if (ret != -ENOENT) {
6093 pmu = ERR_PTR(-ENOENT);
6095 srcu_read_unlock(&pmus_srcu, idx);
6101 * Allocate and initialize a event structure
6103 static struct perf_event *
6104 perf_event_alloc(struct perf_event_attr *attr, int cpu,
6105 struct task_struct *task,
6106 struct perf_event *group_leader,
6107 struct perf_event *parent_event,
6108 perf_overflow_handler_t overflow_handler,
6112 struct perf_event *event;
6113 struct hw_perf_event *hwc;
6116 if ((unsigned)cpu >= nr_cpu_ids) {
6117 if (!task || cpu != -1)
6118 return ERR_PTR(-EINVAL);
6121 event = kzalloc(sizeof(*event), GFP_KERNEL);
6123 return ERR_PTR(-ENOMEM);
6126 * Single events are their own group leaders, with an
6127 * empty sibling list:
6130 group_leader = event;
6132 mutex_init(&event->child_mutex);
6133 INIT_LIST_HEAD(&event->child_list);
6135 INIT_LIST_HEAD(&event->group_entry);
6136 INIT_LIST_HEAD(&event->event_entry);
6137 INIT_LIST_HEAD(&event->sibling_list);
6138 INIT_LIST_HEAD(&event->rb_entry);
6140 init_waitqueue_head(&event->waitq);
6141 init_irq_work(&event->pending, perf_pending_event);
6143 mutex_init(&event->mmap_mutex);
6146 event->attr = *attr;
6147 event->group_leader = group_leader;
6151 event->parent = parent_event;
6153 event->ns = get_pid_ns(current->nsproxy->pid_ns);
6154 event->id = atomic64_inc_return(&perf_event_id);
6156 event->state = PERF_EVENT_STATE_INACTIVE;
6159 event->attach_state = PERF_ATTACH_TASK;
6160 #ifdef CONFIG_HAVE_HW_BREAKPOINT
6162 * hw_breakpoint is a bit difficult here..
6164 if (attr->type == PERF_TYPE_BREAKPOINT)
6165 event->hw.bp_target = task;
6169 if (!overflow_handler && parent_event) {
6170 overflow_handler = parent_event->overflow_handler;
6171 context = parent_event->overflow_handler_context;
6174 event->overflow_handler = overflow_handler;
6175 event->overflow_handler_context = context;
6178 event->state = PERF_EVENT_STATE_OFF;
6183 hwc->sample_period = attr->sample_period;
6184 if (attr->freq && attr->sample_freq)
6185 hwc->sample_period = 1;
6186 hwc->last_period = hwc->sample_period;
6188 local64_set(&hwc->period_left, hwc->sample_period);
6191 * we currently do not support PERF_FORMAT_GROUP on inherited events
6193 if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP))
6196 pmu = perf_init_event(event);
6202 else if (IS_ERR(pmu))
6207 put_pid_ns(event->ns);
6209 return ERR_PTR(err);
6212 if (!event->parent) {
6213 if (event->attach_state & PERF_ATTACH_TASK)
6214 static_key_slow_inc(&perf_sched_events.key);
6215 if (event->attr.mmap || event->attr.mmap_data)
6216 atomic_inc(&nr_mmap_events);
6217 if (event->attr.comm)
6218 atomic_inc(&nr_comm_events);
6219 if (event->attr.task)
6220 atomic_inc(&nr_task_events);
6221 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) {
6222 err = get_callchain_buffers();
6225 return ERR_PTR(err);
6228 if (has_branch_stack(event)) {
6229 static_key_slow_inc(&perf_sched_events.key);
6230 if (!(event->attach_state & PERF_ATTACH_TASK))
6231 atomic_inc(&per_cpu(perf_branch_stack_events,
6239 static int perf_copy_attr(struct perf_event_attr __user *uattr,
6240 struct perf_event_attr *attr)
6245 if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0))
6249 * zero the full structure, so that a short copy will be nice.
6251 memset(attr, 0, sizeof(*attr));
6253 ret = get_user(size, &uattr->size);
6257 if (size > PAGE_SIZE) /* silly large */
6260 if (!size) /* abi compat */
6261 size = PERF_ATTR_SIZE_VER0;
6263 if (size < PERF_ATTR_SIZE_VER0)
6267 * If we're handed a bigger struct than we know of,
6268 * ensure all the unknown bits are 0 - i.e. new
6269 * user-space does not rely on any kernel feature
6270 * extensions we dont know about yet.
6272 if (size > sizeof(*attr)) {
6273 unsigned char __user *addr;
6274 unsigned char __user *end;
6277 addr = (void __user *)uattr + sizeof(*attr);
6278 end = (void __user *)uattr + size;
6280 for (; addr < end; addr++) {
6281 ret = get_user(val, addr);
6287 size = sizeof(*attr);
6290 ret = copy_from_user(attr, uattr, size);
6294 if (attr->__reserved_1)
6297 if (attr->sample_type & ~(PERF_SAMPLE_MAX-1))
6300 if (attr->read_format & ~(PERF_FORMAT_MAX-1))
6303 if (attr->sample_type & PERF_SAMPLE_BRANCH_STACK) {
6304 u64 mask = attr->branch_sample_type;
6306 /* only using defined bits */
6307 if (mask & ~(PERF_SAMPLE_BRANCH_MAX-1))
6310 /* at least one branch bit must be set */
6311 if (!(mask & ~PERF_SAMPLE_BRANCH_PLM_ALL))
6314 /* kernel level capture: check permissions */
6315 if ((mask & PERF_SAMPLE_BRANCH_PERM_PLM)
6316 && perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
6319 /* propagate priv level, when not set for branch */
6320 if (!(mask & PERF_SAMPLE_BRANCH_PLM_ALL)) {
6322 /* exclude_kernel checked on syscall entry */
6323 if (!attr->exclude_kernel)
6324 mask |= PERF_SAMPLE_BRANCH_KERNEL;
6326 if (!attr->exclude_user)
6327 mask |= PERF_SAMPLE_BRANCH_USER;
6329 if (!attr->exclude_hv)
6330 mask |= PERF_SAMPLE_BRANCH_HV;
6332 * adjust user setting (for HW filter setup)
6334 attr->branch_sample_type = mask;
6338 if (attr->sample_type & PERF_SAMPLE_REGS_USER) {
6339 ret = perf_reg_validate(attr->sample_regs_user);
6344 if (attr->sample_type & PERF_SAMPLE_STACK_USER) {
6345 if (!arch_perf_have_user_stack_dump())
6349 * We have __u32 type for the size, but so far
6350 * we can only use __u16 as maximum due to the
6351 * __u16 sample size limit.
6353 if (attr->sample_stack_user >= USHRT_MAX)
6355 else if (!IS_ALIGNED(attr->sample_stack_user, sizeof(u64)))
6363 put_user(sizeof(*attr), &uattr->size);
6369 perf_event_set_output(struct perf_event *event, struct perf_event *output_event)
6371 struct ring_buffer *rb = NULL, *old_rb = NULL;
6377 /* don't allow circular references */
6378 if (event == output_event)
6382 * Don't allow cross-cpu buffers
6384 if (output_event->cpu != event->cpu)
6388 * If its not a per-cpu rb, it must be the same task.
6390 if (output_event->cpu == -1 && output_event->ctx != event->ctx)
6394 mutex_lock(&event->mmap_mutex);
6395 /* Can't redirect output if we've got an active mmap() */
6396 if (atomic_read(&event->mmap_count))
6400 /* get the rb we want to redirect to */
6401 rb = ring_buffer_get(output_event);
6407 rcu_assign_pointer(event->rb, rb);
6409 ring_buffer_detach(event, old_rb);
6412 mutex_unlock(&event->mmap_mutex);
6415 ring_buffer_put(old_rb);
6421 * sys_perf_event_open - open a performance event, associate it to a task/cpu
6423 * @attr_uptr: event_id type attributes for monitoring/sampling
6426 * @group_fd: group leader event fd
6428 SYSCALL_DEFINE5(perf_event_open,
6429 struct perf_event_attr __user *, attr_uptr,
6430 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
6432 struct perf_event *group_leader = NULL, *output_event = NULL;
6433 struct perf_event *event, *sibling;
6434 struct perf_event_attr attr;
6435 struct perf_event_context *ctx;
6436 struct file *event_file = NULL;
6437 struct file *group_file = NULL;
6438 struct task_struct *task = NULL;
6442 int fput_needed = 0;
6445 /* for future expandability... */
6446 if (flags & ~PERF_FLAG_ALL)
6449 err = perf_copy_attr(attr_uptr, &attr);
6453 if (!attr.exclude_kernel) {
6454 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
6459 if (attr.sample_freq > sysctl_perf_event_sample_rate)
6464 * In cgroup mode, the pid argument is used to pass the fd
6465 * opened to the cgroup directory in cgroupfs. The cpu argument
6466 * designates the cpu on which to monitor threads from that
6469 if ((flags & PERF_FLAG_PID_CGROUP) && (pid == -1 || cpu == -1))
6472 event_fd = get_unused_fd_flags(O_RDWR);
6476 if (group_fd != -1) {
6477 group_leader = perf_fget_light(group_fd, &fput_needed);
6478 if (IS_ERR(group_leader)) {
6479 err = PTR_ERR(group_leader);
6482 group_file = group_leader->filp;
6483 if (flags & PERF_FLAG_FD_OUTPUT)
6484 output_event = group_leader;
6485 if (flags & PERF_FLAG_FD_NO_GROUP)
6486 group_leader = NULL;
6489 if (pid != -1 && !(flags & PERF_FLAG_PID_CGROUP)) {
6490 task = find_lively_task_by_vpid(pid);
6492 err = PTR_ERR(task);
6499 event = perf_event_alloc(&attr, cpu, task, group_leader, NULL,
6501 if (IS_ERR(event)) {
6502 err = PTR_ERR(event);
6506 if (flags & PERF_FLAG_PID_CGROUP) {
6507 err = perf_cgroup_connect(pid, event, &attr, group_leader);
6512 * - that has cgroup constraint on event->cpu
6513 * - that may need work on context switch
6515 atomic_inc(&per_cpu(perf_cgroup_events, event->cpu));
6516 static_key_slow_inc(&perf_sched_events.key);
6520 * Special case software events and allow them to be part of
6521 * any hardware group.
6526 (is_software_event(event) != is_software_event(group_leader))) {
6527 if (is_software_event(event)) {
6529 * If event and group_leader are not both a software
6530 * event, and event is, then group leader is not.
6532 * Allow the addition of software events to !software
6533 * groups, this is safe because software events never
6536 pmu = group_leader->pmu;
6537 } else if (is_software_event(group_leader) &&
6538 (group_leader->group_flags & PERF_GROUP_SOFTWARE)) {
6540 * In case the group is a pure software group, and we
6541 * try to add a hardware event, move the whole group to
6542 * the hardware context.
6549 * Get the target context (task or percpu):
6551 ctx = find_get_context(pmu, task, event->cpu);
6558 put_task_struct(task);
6563 * Look up the group leader (we will attach this event to it):
6569 * Do not allow a recursive hierarchy (this new sibling
6570 * becoming part of another group-sibling):
6572 if (group_leader->group_leader != group_leader)
6575 * Do not allow to attach to a group in a different
6576 * task or CPU context:
6579 if (group_leader->ctx->type != ctx->type)
6582 if (group_leader->ctx != ctx)
6587 * Only a group leader can be exclusive or pinned
6589 if (attr.exclusive || attr.pinned)
6594 err = perf_event_set_output(event, output_event);
6599 event_file = anon_inode_getfile("[perf_event]", &perf_fops, event, O_RDWR);
6600 if (IS_ERR(event_file)) {
6601 err = PTR_ERR(event_file);
6606 struct perf_event_context *gctx = group_leader->ctx;
6608 mutex_lock(&gctx->mutex);
6609 perf_remove_from_context(group_leader);
6610 list_for_each_entry(sibling, &group_leader->sibling_list,
6612 perf_remove_from_context(sibling);
6615 mutex_unlock(&gctx->mutex);
6619 event->filp = event_file;
6620 WARN_ON_ONCE(ctx->parent_ctx);
6621 mutex_lock(&ctx->mutex);
6625 perf_install_in_context(ctx, group_leader, event->cpu);
6627 list_for_each_entry(sibling, &group_leader->sibling_list,
6629 perf_install_in_context(ctx, sibling, event->cpu);
6634 perf_install_in_context(ctx, event, event->cpu);
6636 perf_unpin_context(ctx);
6637 mutex_unlock(&ctx->mutex);
6641 event->owner = current;
6643 mutex_lock(¤t->perf_event_mutex);
6644 list_add_tail(&event->owner_entry, ¤t->perf_event_list);
6645 mutex_unlock(¤t->perf_event_mutex);
6648 * Precalculate sample_data sizes
6650 perf_event__header_size(event);
6651 perf_event__id_header_size(event);
6654 * Drop the reference on the group_event after placing the
6655 * new event on the sibling_list. This ensures destruction
6656 * of the group leader will find the pointer to itself in
6657 * perf_group_detach().
6659 fput_light(group_file, fput_needed);
6660 fd_install(event_fd, event_file);
6664 perf_unpin_context(ctx);
6671 put_task_struct(task);
6673 fput_light(group_file, fput_needed);
6675 put_unused_fd(event_fd);
6680 * perf_event_create_kernel_counter
6682 * @attr: attributes of the counter to create
6683 * @cpu: cpu in which the counter is bound
6684 * @task: task to profile (NULL for percpu)
6687 perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu,
6688 struct task_struct *task,
6689 perf_overflow_handler_t overflow_handler,
6692 struct perf_event_context *ctx;
6693 struct perf_event *event;
6697 * Get the target context (task or percpu):
6700 event = perf_event_alloc(attr, cpu, task, NULL, NULL,
6701 overflow_handler, context);
6702 if (IS_ERR(event)) {
6703 err = PTR_ERR(event);
6707 ctx = find_get_context(event->pmu, task, cpu);
6714 WARN_ON_ONCE(ctx->parent_ctx);
6715 mutex_lock(&ctx->mutex);
6716 perf_install_in_context(ctx, event, cpu);
6718 perf_unpin_context(ctx);
6719 mutex_unlock(&ctx->mutex);
6726 return ERR_PTR(err);
6728 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter);
6730 void perf_pmu_migrate_context(struct pmu *pmu, int src_cpu, int dst_cpu)
6732 struct perf_event_context *src_ctx;
6733 struct perf_event_context *dst_ctx;
6734 struct perf_event *event, *tmp;
6737 src_ctx = &per_cpu_ptr(pmu->pmu_cpu_context, src_cpu)->ctx;
6738 dst_ctx = &per_cpu_ptr(pmu->pmu_cpu_context, dst_cpu)->ctx;
6740 mutex_lock(&src_ctx->mutex);
6741 list_for_each_entry_safe(event, tmp, &src_ctx->event_list,
6743 perf_remove_from_context(event);
6745 list_add(&event->event_entry, &events);
6747 mutex_unlock(&src_ctx->mutex);
6751 mutex_lock(&dst_ctx->mutex);
6752 list_for_each_entry_safe(event, tmp, &events, event_entry) {
6753 list_del(&event->event_entry);
6754 if (event->state >= PERF_EVENT_STATE_OFF)
6755 event->state = PERF_EVENT_STATE_INACTIVE;
6756 perf_install_in_context(dst_ctx, event, dst_cpu);
6759 mutex_unlock(&dst_ctx->mutex);
6761 EXPORT_SYMBOL_GPL(perf_pmu_migrate_context);
6763 static void sync_child_event(struct perf_event *child_event,
6764 struct task_struct *child)
6766 struct perf_event *parent_event = child_event->parent;
6769 if (child_event->attr.inherit_stat)
6770 perf_event_read_event(child_event, child);
6772 child_val = perf_event_count(child_event);
6775 * Add back the child's count to the parent's count:
6777 atomic64_add(child_val, &parent_event->child_count);
6778 atomic64_add(child_event->total_time_enabled,
6779 &parent_event->child_total_time_enabled);
6780 atomic64_add(child_event->total_time_running,
6781 &parent_event->child_total_time_running);
6784 * Remove this event from the parent's list
6786 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
6787 mutex_lock(&parent_event->child_mutex);
6788 list_del_init(&child_event->child_list);
6789 mutex_unlock(&parent_event->child_mutex);
6792 * Release the parent event, if this was the last
6795 fput(parent_event->filp);
6799 __perf_event_exit_task(struct perf_event *child_event,
6800 struct perf_event_context *child_ctx,
6801 struct task_struct *child)
6803 if (child_event->parent) {
6804 raw_spin_lock_irq(&child_ctx->lock);
6805 perf_group_detach(child_event);
6806 raw_spin_unlock_irq(&child_ctx->lock);
6809 perf_remove_from_context(child_event);
6812 * It can happen that the parent exits first, and has events
6813 * that are still around due to the child reference. These
6814 * events need to be zapped.
6816 if (child_event->parent) {
6817 sync_child_event(child_event, child);
6818 free_event(child_event);
6822 static void perf_event_exit_task_context(struct task_struct *child, int ctxn)
6824 struct perf_event *child_event, *tmp;
6825 struct perf_event_context *child_ctx;
6826 unsigned long flags;
6828 if (likely(!child->perf_event_ctxp[ctxn])) {
6829 perf_event_task(child, NULL, 0);
6833 local_irq_save(flags);
6835 * We can't reschedule here because interrupts are disabled,
6836 * and either child is current or it is a task that can't be
6837 * scheduled, so we are now safe from rescheduling changing
6840 child_ctx = rcu_dereference_raw(child->perf_event_ctxp[ctxn]);
6843 * Take the context lock here so that if find_get_context is
6844 * reading child->perf_event_ctxp, we wait until it has
6845 * incremented the context's refcount before we do put_ctx below.
6847 raw_spin_lock(&child_ctx->lock);
6848 task_ctx_sched_out(child_ctx);
6849 child->perf_event_ctxp[ctxn] = NULL;
6851 * If this context is a clone; unclone it so it can't get
6852 * swapped to another process while we're removing all
6853 * the events from it.
6855 unclone_ctx(child_ctx);
6856 update_context_time(child_ctx);
6857 raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
6860 * Report the task dead after unscheduling the events so that we
6861 * won't get any samples after PERF_RECORD_EXIT. We can however still
6862 * get a few PERF_RECORD_READ events.
6864 perf_event_task(child, child_ctx, 0);
6867 * We can recurse on the same lock type through:
6869 * __perf_event_exit_task()
6870 * sync_child_event()
6871 * fput(parent_event->filp)
6873 * mutex_lock(&ctx->mutex)
6875 * But since its the parent context it won't be the same instance.
6877 mutex_lock(&child_ctx->mutex);
6880 list_for_each_entry_safe(child_event, tmp, &child_ctx->pinned_groups,
6882 __perf_event_exit_task(child_event, child_ctx, child);
6884 list_for_each_entry_safe(child_event, tmp, &child_ctx->flexible_groups,
6886 __perf_event_exit_task(child_event, child_ctx, child);
6889 * If the last event was a group event, it will have appended all
6890 * its siblings to the list, but we obtained 'tmp' before that which
6891 * will still point to the list head terminating the iteration.
6893 if (!list_empty(&child_ctx->pinned_groups) ||
6894 !list_empty(&child_ctx->flexible_groups))
6897 mutex_unlock(&child_ctx->mutex);
6903 * When a child task exits, feed back event values to parent events.
6905 void perf_event_exit_task(struct task_struct *child)
6907 struct perf_event *event, *tmp;
6910 mutex_lock(&child->perf_event_mutex);
6911 list_for_each_entry_safe(event, tmp, &child->perf_event_list,
6913 list_del_init(&event->owner_entry);
6916 * Ensure the list deletion is visible before we clear
6917 * the owner, closes a race against perf_release() where
6918 * we need to serialize on the owner->perf_event_mutex.
6921 event->owner = NULL;
6923 mutex_unlock(&child->perf_event_mutex);
6925 for_each_task_context_nr(ctxn)
6926 perf_event_exit_task_context(child, ctxn);
6929 static void perf_free_event(struct perf_event *event,
6930 struct perf_event_context *ctx)
6932 struct perf_event *parent = event->parent;
6934 if (WARN_ON_ONCE(!parent))
6937 mutex_lock(&parent->child_mutex);
6938 list_del_init(&event->child_list);
6939 mutex_unlock(&parent->child_mutex);
6943 perf_group_detach(event);
6944 list_del_event(event, ctx);
6949 * free an unexposed, unused context as created by inheritance by
6950 * perf_event_init_task below, used by fork() in case of fail.
6952 void perf_event_free_task(struct task_struct *task)
6954 struct perf_event_context *ctx;
6955 struct perf_event *event, *tmp;
6958 for_each_task_context_nr(ctxn) {
6959 ctx = task->perf_event_ctxp[ctxn];
6963 mutex_lock(&ctx->mutex);
6965 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups,
6967 perf_free_event(event, ctx);
6969 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups,
6971 perf_free_event(event, ctx);
6973 if (!list_empty(&ctx->pinned_groups) ||
6974 !list_empty(&ctx->flexible_groups))
6977 mutex_unlock(&ctx->mutex);
6983 void perf_event_delayed_put(struct task_struct *task)
6987 for_each_task_context_nr(ctxn)
6988 WARN_ON_ONCE(task->perf_event_ctxp[ctxn]);
6992 * inherit a event from parent task to child task:
6994 static struct perf_event *
6995 inherit_event(struct perf_event *parent_event,
6996 struct task_struct *parent,
6997 struct perf_event_context *parent_ctx,
6998 struct task_struct *child,
6999 struct perf_event *group_leader,
7000 struct perf_event_context *child_ctx)
7002 struct perf_event *child_event;
7003 unsigned long flags;
7006 * Instead of creating recursive hierarchies of events,
7007 * we link inherited events back to the original parent,
7008 * which has a filp for sure, which we use as the reference
7011 if (parent_event->parent)
7012 parent_event = parent_event->parent;
7014 child_event = perf_event_alloc(&parent_event->attr,
7017 group_leader, parent_event,
7019 if (IS_ERR(child_event))
7024 * Make the child state follow the state of the parent event,
7025 * not its attr.disabled bit. We hold the parent's mutex,
7026 * so we won't race with perf_event_{en, dis}able_family.
7028 if (parent_event->state >= PERF_EVENT_STATE_INACTIVE)
7029 child_event->state = PERF_EVENT_STATE_INACTIVE;
7031 child_event->state = PERF_EVENT_STATE_OFF;
7033 if (parent_event->attr.freq) {
7034 u64 sample_period = parent_event->hw.sample_period;
7035 struct hw_perf_event *hwc = &child_event->hw;
7037 hwc->sample_period = sample_period;
7038 hwc->last_period = sample_period;
7040 local64_set(&hwc->period_left, sample_period);
7043 child_event->ctx = child_ctx;
7044 child_event->overflow_handler = parent_event->overflow_handler;
7045 child_event->overflow_handler_context
7046 = parent_event->overflow_handler_context;
7049 * Precalculate sample_data sizes
7051 perf_event__header_size(child_event);
7052 perf_event__id_header_size(child_event);
7055 * Link it up in the child's context:
7057 raw_spin_lock_irqsave(&child_ctx->lock, flags);
7058 add_event_to_ctx(child_event, child_ctx);
7059 raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
7062 * Get a reference to the parent filp - we will fput it
7063 * when the child event exits. This is safe to do because
7064 * we are in the parent and we know that the filp still
7065 * exists and has a nonzero count:
7067 atomic_long_inc(&parent_event->filp->f_count);
7070 * Link this into the parent event's child list
7072 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
7073 mutex_lock(&parent_event->child_mutex);
7074 list_add_tail(&child_event->child_list, &parent_event->child_list);
7075 mutex_unlock(&parent_event->child_mutex);
7080 static int inherit_group(struct perf_event *parent_event,
7081 struct task_struct *parent,
7082 struct perf_event_context *parent_ctx,
7083 struct task_struct *child,
7084 struct perf_event_context *child_ctx)
7086 struct perf_event *leader;
7087 struct perf_event *sub;
7088 struct perf_event *child_ctr;
7090 leader = inherit_event(parent_event, parent, parent_ctx,
7091 child, NULL, child_ctx);
7093 return PTR_ERR(leader);
7094 list_for_each_entry(sub, &parent_event->sibling_list, group_entry) {
7095 child_ctr = inherit_event(sub, parent, parent_ctx,
7096 child, leader, child_ctx);
7097 if (IS_ERR(child_ctr))
7098 return PTR_ERR(child_ctr);
7104 inherit_task_group(struct perf_event *event, struct task_struct *parent,
7105 struct perf_event_context *parent_ctx,
7106 struct task_struct *child, int ctxn,
7110 struct perf_event_context *child_ctx;
7112 if (!event->attr.inherit) {
7117 child_ctx = child->perf_event_ctxp[ctxn];
7120 * This is executed from the parent task context, so
7121 * inherit events that have been marked for cloning.
7122 * First allocate and initialize a context for the
7126 child_ctx = alloc_perf_context(event->pmu, child);
7130 child->perf_event_ctxp[ctxn] = child_ctx;
7133 ret = inherit_group(event, parent, parent_ctx,
7143 * Initialize the perf_event context in task_struct
7145 int perf_event_init_context(struct task_struct *child, int ctxn)
7147 struct perf_event_context *child_ctx, *parent_ctx;
7148 struct perf_event_context *cloned_ctx;
7149 struct perf_event *event;
7150 struct task_struct *parent = current;
7151 int inherited_all = 1;
7152 unsigned long flags;
7155 if (likely(!parent->perf_event_ctxp[ctxn]))
7159 * If the parent's context is a clone, pin it so it won't get
7162 parent_ctx = perf_pin_task_context(parent, ctxn);
7165 * No need to check if parent_ctx != NULL here; since we saw
7166 * it non-NULL earlier, the only reason for it to become NULL
7167 * is if we exit, and since we're currently in the middle of
7168 * a fork we can't be exiting at the same time.
7172 * Lock the parent list. No need to lock the child - not PID
7173 * hashed yet and not running, so nobody can access it.
7175 mutex_lock(&parent_ctx->mutex);
7178 * We dont have to disable NMIs - we are only looking at
7179 * the list, not manipulating it:
7181 list_for_each_entry(event, &parent_ctx->pinned_groups, group_entry) {
7182 ret = inherit_task_group(event, parent, parent_ctx,
7183 child, ctxn, &inherited_all);
7189 * We can't hold ctx->lock when iterating the ->flexible_group list due
7190 * to allocations, but we need to prevent rotation because
7191 * rotate_ctx() will change the list from interrupt context.
7193 raw_spin_lock_irqsave(&parent_ctx->lock, flags);
7194 parent_ctx->rotate_disable = 1;
7195 raw_spin_unlock_irqrestore(&parent_ctx->lock, flags);
7197 list_for_each_entry(event, &parent_ctx->flexible_groups, group_entry) {
7198 ret = inherit_task_group(event, parent, parent_ctx,
7199 child, ctxn, &inherited_all);
7204 raw_spin_lock_irqsave(&parent_ctx->lock, flags);
7205 parent_ctx->rotate_disable = 0;
7207 child_ctx = child->perf_event_ctxp[ctxn];
7209 if (child_ctx && inherited_all) {
7211 * Mark the child context as a clone of the parent
7212 * context, or of whatever the parent is a clone of.
7214 * Note that if the parent is a clone, the holding of
7215 * parent_ctx->lock avoids it from being uncloned.
7217 cloned_ctx = parent_ctx->parent_ctx;
7219 child_ctx->parent_ctx = cloned_ctx;
7220 child_ctx->parent_gen = parent_ctx->parent_gen;
7222 child_ctx->parent_ctx = parent_ctx;
7223 child_ctx->parent_gen = parent_ctx->generation;
7225 get_ctx(child_ctx->parent_ctx);
7228 raw_spin_unlock_irqrestore(&parent_ctx->lock, flags);
7229 mutex_unlock(&parent_ctx->mutex);
7231 perf_unpin_context(parent_ctx);
7232 put_ctx(parent_ctx);
7238 * Initialize the perf_event context in task_struct
7240 int perf_event_init_task(struct task_struct *child)
7244 memset(child->perf_event_ctxp, 0, sizeof(child->perf_event_ctxp));
7245 mutex_init(&child->perf_event_mutex);
7246 INIT_LIST_HEAD(&child->perf_event_list);
7248 for_each_task_context_nr(ctxn) {
7249 ret = perf_event_init_context(child, ctxn);
7257 static void __init perf_event_init_all_cpus(void)
7259 struct swevent_htable *swhash;
7262 for_each_possible_cpu(cpu) {
7263 swhash = &per_cpu(swevent_htable, cpu);
7264 mutex_init(&swhash->hlist_mutex);
7265 INIT_LIST_HEAD(&per_cpu(rotation_list, cpu));
7269 static void __cpuinit perf_event_init_cpu(int cpu)
7271 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
7273 mutex_lock(&swhash->hlist_mutex);
7274 if (swhash->hlist_refcount > 0) {
7275 struct swevent_hlist *hlist;
7277 hlist = kzalloc_node(sizeof(*hlist), GFP_KERNEL, cpu_to_node(cpu));
7279 rcu_assign_pointer(swhash->swevent_hlist, hlist);
7281 mutex_unlock(&swhash->hlist_mutex);
7284 #if defined CONFIG_HOTPLUG_CPU || defined CONFIG_KEXEC
7285 static void perf_pmu_rotate_stop(struct pmu *pmu)
7287 struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
7289 WARN_ON(!irqs_disabled());
7291 list_del_init(&cpuctx->rotation_list);
7294 static void __perf_event_exit_context(void *__info)
7296 struct perf_event_context *ctx = __info;
7297 struct perf_event *event, *tmp;
7299 perf_pmu_rotate_stop(ctx->pmu);
7301 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry)
7302 __perf_remove_from_context(event);
7303 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups, group_entry)
7304 __perf_remove_from_context(event);
7307 static void perf_event_exit_cpu_context(int cpu)
7309 struct perf_event_context *ctx;
7313 idx = srcu_read_lock(&pmus_srcu);
7314 list_for_each_entry_rcu(pmu, &pmus, entry) {
7315 ctx = &per_cpu_ptr(pmu->pmu_cpu_context, cpu)->ctx;
7317 mutex_lock(&ctx->mutex);
7318 smp_call_function_single(cpu, __perf_event_exit_context, ctx, 1);
7319 mutex_unlock(&ctx->mutex);
7321 srcu_read_unlock(&pmus_srcu, idx);
7324 static void perf_event_exit_cpu(int cpu)
7326 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
7328 mutex_lock(&swhash->hlist_mutex);
7329 swevent_hlist_release(swhash);
7330 mutex_unlock(&swhash->hlist_mutex);
7332 perf_event_exit_cpu_context(cpu);
7335 static inline void perf_event_exit_cpu(int cpu) { }
7339 perf_reboot(struct notifier_block *notifier, unsigned long val, void *v)
7343 for_each_online_cpu(cpu)
7344 perf_event_exit_cpu(cpu);
7350 * Run the perf reboot notifier at the very last possible moment so that
7351 * the generic watchdog code runs as long as possible.
7353 static struct notifier_block perf_reboot_notifier = {
7354 .notifier_call = perf_reboot,
7355 .priority = INT_MIN,
7358 static int __cpuinit
7359 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
7361 unsigned int cpu = (long)hcpu;
7363 switch (action & ~CPU_TASKS_FROZEN) {
7365 case CPU_UP_PREPARE:
7366 case CPU_DOWN_FAILED:
7367 perf_event_init_cpu(cpu);
7370 case CPU_UP_CANCELED:
7371 case CPU_DOWN_PREPARE:
7372 perf_event_exit_cpu(cpu);
7382 void __init perf_event_init(void)
7388 perf_event_init_all_cpus();
7389 init_srcu_struct(&pmus_srcu);
7390 perf_pmu_register(&perf_swevent, "software", PERF_TYPE_SOFTWARE);
7391 perf_pmu_register(&perf_cpu_clock, NULL, -1);
7392 perf_pmu_register(&perf_task_clock, NULL, -1);
7394 perf_cpu_notifier(perf_cpu_notify);
7395 register_reboot_notifier(&perf_reboot_notifier);
7397 ret = init_hw_breakpoint();
7398 WARN(ret, "hw_breakpoint initialization failed with: %d", ret);
7400 /* do not patch jump label more than once per second */
7401 jump_label_rate_limit(&perf_sched_events, HZ);
7404 * Build time assertion that we keep the data_head at the intended
7405 * location. IOW, validation we got the __reserved[] size right.
7407 BUILD_BUG_ON((offsetof(struct perf_event_mmap_page, data_head))
7411 static int __init perf_event_sysfs_init(void)
7416 mutex_lock(&pmus_lock);
7418 ret = bus_register(&pmu_bus);
7422 list_for_each_entry(pmu, &pmus, entry) {
7423 if (!pmu->name || pmu->type < 0)
7426 ret = pmu_dev_alloc(pmu);
7427 WARN(ret, "Failed to register pmu: %s, reason %d\n", pmu->name, ret);
7429 pmu_bus_running = 1;
7433 mutex_unlock(&pmus_lock);
7437 device_initcall(perf_event_sysfs_init);
7439 #ifdef CONFIG_CGROUP_PERF
7440 static struct cgroup_subsys_state *perf_cgroup_create(struct cgroup *cont)
7442 struct perf_cgroup *jc;
7444 jc = kzalloc(sizeof(*jc), GFP_KERNEL);
7446 return ERR_PTR(-ENOMEM);
7448 jc->info = alloc_percpu(struct perf_cgroup_info);
7451 return ERR_PTR(-ENOMEM);
7457 static void perf_cgroup_destroy(struct cgroup *cont)
7459 struct perf_cgroup *jc;
7460 jc = container_of(cgroup_subsys_state(cont, perf_subsys_id),
7461 struct perf_cgroup, css);
7462 free_percpu(jc->info);
7466 static int __perf_cgroup_move(void *info)
7468 struct task_struct *task = info;
7469 perf_cgroup_switch(task, PERF_CGROUP_SWOUT | PERF_CGROUP_SWIN);
7473 static void perf_cgroup_attach(struct cgroup *cgrp, struct cgroup_taskset *tset)
7475 struct task_struct *task;
7477 cgroup_taskset_for_each(task, cgrp, tset)
7478 task_function_call(task, __perf_cgroup_move, task);
7481 static void perf_cgroup_exit(struct cgroup *cgrp, struct cgroup *old_cgrp,
7482 struct task_struct *task)
7485 * cgroup_exit() is called in the copy_process() failure path.
7486 * Ignore this case since the task hasn't ran yet, this avoids
7487 * trying to poke a half freed task state from generic code.
7489 if (!(task->flags & PF_EXITING))
7492 task_function_call(task, __perf_cgroup_move, task);
7495 struct cgroup_subsys perf_subsys = {
7496 .name = "perf_event",
7497 .subsys_id = perf_subsys_id,
7498 .create = perf_cgroup_create,
7499 .destroy = perf_cgroup_destroy,
7500 .exit = perf_cgroup_exit,
7501 .attach = perf_cgroup_attach,
7503 #endif /* CONFIG_CGROUP_PERF */