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);
375 if (cpuctx->unique_pmu != pmu)
376 continue; /* ensure we process each cpuctx once */
379 * perf_cgroup_events says at least one
380 * context on this CPU has cgroup events.
382 * ctx->nr_cgroups reports the number of cgroup
383 * events for a context.
385 if (cpuctx->ctx.nr_cgroups > 0) {
386 perf_ctx_lock(cpuctx, cpuctx->task_ctx);
387 perf_pmu_disable(cpuctx->ctx.pmu);
389 if (mode & PERF_CGROUP_SWOUT) {
390 cpu_ctx_sched_out(cpuctx, EVENT_ALL);
392 * must not be done before ctxswout due
393 * to event_filter_match() in event_sched_out()
398 if (mode & PERF_CGROUP_SWIN) {
399 WARN_ON_ONCE(cpuctx->cgrp);
401 * set cgrp before ctxsw in to allow
402 * event_filter_match() to not have to pass
405 cpuctx->cgrp = perf_cgroup_from_task(task);
406 cpu_ctx_sched_in(cpuctx, EVENT_ALL, task);
408 perf_pmu_enable(cpuctx->ctx.pmu);
409 perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
415 local_irq_restore(flags);
418 static inline void perf_cgroup_sched_out(struct task_struct *task,
419 struct task_struct *next)
421 struct perf_cgroup *cgrp1;
422 struct perf_cgroup *cgrp2 = NULL;
425 * we come here when we know perf_cgroup_events > 0
427 cgrp1 = perf_cgroup_from_task(task);
430 * next is NULL when called from perf_event_enable_on_exec()
431 * that will systematically cause a cgroup_switch()
434 cgrp2 = perf_cgroup_from_task(next);
437 * only schedule out current cgroup events if we know
438 * that we are switching to a different cgroup. Otherwise,
439 * do no touch the cgroup events.
442 perf_cgroup_switch(task, PERF_CGROUP_SWOUT);
445 static inline void perf_cgroup_sched_in(struct task_struct *prev,
446 struct task_struct *task)
448 struct perf_cgroup *cgrp1;
449 struct perf_cgroup *cgrp2 = NULL;
452 * we come here when we know perf_cgroup_events > 0
454 cgrp1 = perf_cgroup_from_task(task);
456 /* prev can never be NULL */
457 cgrp2 = perf_cgroup_from_task(prev);
460 * only need to schedule in cgroup events if we are changing
461 * cgroup during ctxsw. Cgroup events were not scheduled
462 * out of ctxsw out if that was not the case.
465 perf_cgroup_switch(task, PERF_CGROUP_SWIN);
468 static inline int perf_cgroup_connect(int fd, struct perf_event *event,
469 struct perf_event_attr *attr,
470 struct perf_event *group_leader)
472 struct perf_cgroup *cgrp;
473 struct cgroup_subsys_state *css;
474 struct fd f = fdget(fd);
480 css = cgroup_css_from_dir(f.file, perf_subsys_id);
486 cgrp = container_of(css, struct perf_cgroup, css);
489 /* must be done before we fput() the file */
490 if (!perf_tryget_cgroup(event)) {
497 * all events in a group must monitor
498 * the same cgroup because a task belongs
499 * to only one perf cgroup at a time
501 if (group_leader && group_leader->cgrp != cgrp) {
502 perf_detach_cgroup(event);
511 perf_cgroup_set_shadow_time(struct perf_event *event, u64 now)
513 struct perf_cgroup_info *t;
514 t = per_cpu_ptr(event->cgrp->info, event->cpu);
515 event->shadow_ctx_time = now - t->timestamp;
519 perf_cgroup_defer_enabled(struct perf_event *event)
522 * when the current task's perf cgroup does not match
523 * the event's, we need to remember to call the
524 * perf_mark_enable() function the first time a task with
525 * a matching perf cgroup is scheduled in.
527 if (is_cgroup_event(event) && !perf_cgroup_match(event))
528 event->cgrp_defer_enabled = 1;
532 perf_cgroup_mark_enabled(struct perf_event *event,
533 struct perf_event_context *ctx)
535 struct perf_event *sub;
536 u64 tstamp = perf_event_time(event);
538 if (!event->cgrp_defer_enabled)
541 event->cgrp_defer_enabled = 0;
543 event->tstamp_enabled = tstamp - event->total_time_enabled;
544 list_for_each_entry(sub, &event->sibling_list, group_entry) {
545 if (sub->state >= PERF_EVENT_STATE_INACTIVE) {
546 sub->tstamp_enabled = tstamp - sub->total_time_enabled;
547 sub->cgrp_defer_enabled = 0;
551 #else /* !CONFIG_CGROUP_PERF */
554 perf_cgroup_match(struct perf_event *event)
559 static inline void perf_detach_cgroup(struct perf_event *event)
562 static inline int is_cgroup_event(struct perf_event *event)
567 static inline u64 perf_cgroup_event_cgrp_time(struct perf_event *event)
572 static inline void update_cgrp_time_from_event(struct perf_event *event)
576 static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx)
580 static inline void perf_cgroup_sched_out(struct task_struct *task,
581 struct task_struct *next)
585 static inline void perf_cgroup_sched_in(struct task_struct *prev,
586 struct task_struct *task)
590 static inline int perf_cgroup_connect(pid_t pid, struct perf_event *event,
591 struct perf_event_attr *attr,
592 struct perf_event *group_leader)
598 perf_cgroup_set_timestamp(struct task_struct *task,
599 struct perf_event_context *ctx)
604 perf_cgroup_switch(struct task_struct *task, struct task_struct *next)
609 perf_cgroup_set_shadow_time(struct perf_event *event, u64 now)
613 static inline u64 perf_cgroup_event_time(struct perf_event *event)
619 perf_cgroup_defer_enabled(struct perf_event *event)
624 perf_cgroup_mark_enabled(struct perf_event *event,
625 struct perf_event_context *ctx)
630 void perf_pmu_disable(struct pmu *pmu)
632 int *count = this_cpu_ptr(pmu->pmu_disable_count);
634 pmu->pmu_disable(pmu);
637 void perf_pmu_enable(struct pmu *pmu)
639 int *count = this_cpu_ptr(pmu->pmu_disable_count);
641 pmu->pmu_enable(pmu);
644 static DEFINE_PER_CPU(struct list_head, rotation_list);
647 * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
648 * because they're strictly cpu affine and rotate_start is called with IRQs
649 * disabled, while rotate_context is called from IRQ context.
651 static void perf_pmu_rotate_start(struct pmu *pmu)
653 struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
654 struct list_head *head = &__get_cpu_var(rotation_list);
656 WARN_ON(!irqs_disabled());
658 if (list_empty(&cpuctx->rotation_list))
659 list_add(&cpuctx->rotation_list, head);
662 static void get_ctx(struct perf_event_context *ctx)
664 WARN_ON(!atomic_inc_not_zero(&ctx->refcount));
667 static void put_ctx(struct perf_event_context *ctx)
669 if (atomic_dec_and_test(&ctx->refcount)) {
671 put_ctx(ctx->parent_ctx);
673 put_task_struct(ctx->task);
674 kfree_rcu(ctx, rcu_head);
678 static void unclone_ctx(struct perf_event_context *ctx)
680 if (ctx->parent_ctx) {
681 put_ctx(ctx->parent_ctx);
682 ctx->parent_ctx = NULL;
686 static u32 perf_event_pid(struct perf_event *event, struct task_struct *p)
689 * only top level events have the pid namespace they were created in
692 event = event->parent;
694 return task_tgid_nr_ns(p, event->ns);
697 static u32 perf_event_tid(struct perf_event *event, struct task_struct *p)
700 * only top level events have the pid namespace they were created in
703 event = event->parent;
705 return task_pid_nr_ns(p, event->ns);
709 * If we inherit events we want to return the parent event id
712 static u64 primary_event_id(struct perf_event *event)
717 id = event->parent->id;
723 * Get the perf_event_context for a task and lock it.
724 * This has to cope with with the fact that until it is locked,
725 * the context could get moved to another task.
727 static struct perf_event_context *
728 perf_lock_task_context(struct task_struct *task, int ctxn, unsigned long *flags)
730 struct perf_event_context *ctx;
734 ctx = rcu_dereference(task->perf_event_ctxp[ctxn]);
737 * If this context is a clone of another, it might
738 * get swapped for another underneath us by
739 * perf_event_task_sched_out, though the
740 * rcu_read_lock() protects us from any context
741 * getting freed. Lock the context and check if it
742 * got swapped before we could get the lock, and retry
743 * if so. If we locked the right context, then it
744 * can't get swapped on us any more.
746 raw_spin_lock_irqsave(&ctx->lock, *flags);
747 if (ctx != rcu_dereference(task->perf_event_ctxp[ctxn])) {
748 raw_spin_unlock_irqrestore(&ctx->lock, *flags);
752 if (!atomic_inc_not_zero(&ctx->refcount)) {
753 raw_spin_unlock_irqrestore(&ctx->lock, *flags);
762 * Get the context for a task and increment its pin_count so it
763 * can't get swapped to another task. This also increments its
764 * reference count so that the context can't get freed.
766 static struct perf_event_context *
767 perf_pin_task_context(struct task_struct *task, int ctxn)
769 struct perf_event_context *ctx;
772 ctx = perf_lock_task_context(task, ctxn, &flags);
775 raw_spin_unlock_irqrestore(&ctx->lock, flags);
780 static void perf_unpin_context(struct perf_event_context *ctx)
784 raw_spin_lock_irqsave(&ctx->lock, flags);
786 raw_spin_unlock_irqrestore(&ctx->lock, flags);
790 * Update the record of the current time in a context.
792 static void update_context_time(struct perf_event_context *ctx)
794 u64 now = perf_clock();
796 ctx->time += now - ctx->timestamp;
797 ctx->timestamp = now;
800 static u64 perf_event_time(struct perf_event *event)
802 struct perf_event_context *ctx = event->ctx;
804 if (is_cgroup_event(event))
805 return perf_cgroup_event_time(event);
807 return ctx ? ctx->time : 0;
811 * Update the total_time_enabled and total_time_running fields for a event.
812 * The caller of this function needs to hold the ctx->lock.
814 static void update_event_times(struct perf_event *event)
816 struct perf_event_context *ctx = event->ctx;
819 if (event->state < PERF_EVENT_STATE_INACTIVE ||
820 event->group_leader->state < PERF_EVENT_STATE_INACTIVE)
823 * in cgroup mode, time_enabled represents
824 * the time the event was enabled AND active
825 * tasks were in the monitored cgroup. This is
826 * independent of the activity of the context as
827 * there may be a mix of cgroup and non-cgroup events.
829 * That is why we treat cgroup events differently
832 if (is_cgroup_event(event))
833 run_end = perf_cgroup_event_time(event);
834 else if (ctx->is_active)
837 run_end = event->tstamp_stopped;
839 event->total_time_enabled = run_end - event->tstamp_enabled;
841 if (event->state == PERF_EVENT_STATE_INACTIVE)
842 run_end = event->tstamp_stopped;
844 run_end = perf_event_time(event);
846 event->total_time_running = run_end - event->tstamp_running;
851 * Update total_time_enabled and total_time_running for all events in a group.
853 static void update_group_times(struct perf_event *leader)
855 struct perf_event *event;
857 update_event_times(leader);
858 list_for_each_entry(event, &leader->sibling_list, group_entry)
859 update_event_times(event);
862 static struct list_head *
863 ctx_group_list(struct perf_event *event, struct perf_event_context *ctx)
865 if (event->attr.pinned)
866 return &ctx->pinned_groups;
868 return &ctx->flexible_groups;
872 * Add a event from the lists for its context.
873 * Must be called with ctx->mutex and ctx->lock held.
876 list_add_event(struct perf_event *event, struct perf_event_context *ctx)
878 WARN_ON_ONCE(event->attach_state & PERF_ATTACH_CONTEXT);
879 event->attach_state |= PERF_ATTACH_CONTEXT;
882 * If we're a stand alone event or group leader, we go to the context
883 * list, group events are kept attached to the group so that
884 * perf_group_detach can, at all times, locate all siblings.
886 if (event->group_leader == event) {
887 struct list_head *list;
889 if (is_software_event(event))
890 event->group_flags |= PERF_GROUP_SOFTWARE;
892 list = ctx_group_list(event, ctx);
893 list_add_tail(&event->group_entry, list);
896 if (is_cgroup_event(event))
899 if (has_branch_stack(event))
900 ctx->nr_branch_stack++;
902 list_add_rcu(&event->event_entry, &ctx->event_list);
904 perf_pmu_rotate_start(ctx->pmu);
906 if (event->attr.inherit_stat)
911 * Initialize event state based on the perf_event_attr::disabled.
913 static inline void perf_event__state_init(struct perf_event *event)
915 event->state = event->attr.disabled ? PERF_EVENT_STATE_OFF :
916 PERF_EVENT_STATE_INACTIVE;
920 * Called at perf_event creation and when events are attached/detached from a
923 static void perf_event__read_size(struct perf_event *event)
925 int entry = sizeof(u64); /* value */
929 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
932 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
935 if (event->attr.read_format & PERF_FORMAT_ID)
936 entry += sizeof(u64);
938 if (event->attr.read_format & PERF_FORMAT_GROUP) {
939 nr += event->group_leader->nr_siblings;
944 event->read_size = size;
947 static void perf_event__header_size(struct perf_event *event)
949 struct perf_sample_data *data;
950 u64 sample_type = event->attr.sample_type;
953 perf_event__read_size(event);
955 if (sample_type & PERF_SAMPLE_IP)
956 size += sizeof(data->ip);
958 if (sample_type & PERF_SAMPLE_ADDR)
959 size += sizeof(data->addr);
961 if (sample_type & PERF_SAMPLE_PERIOD)
962 size += sizeof(data->period);
964 if (sample_type & PERF_SAMPLE_READ)
965 size += event->read_size;
967 event->header_size = size;
970 static void perf_event__id_header_size(struct perf_event *event)
972 struct perf_sample_data *data;
973 u64 sample_type = event->attr.sample_type;
976 if (sample_type & PERF_SAMPLE_TID)
977 size += sizeof(data->tid_entry);
979 if (sample_type & PERF_SAMPLE_TIME)
980 size += sizeof(data->time);
982 if (sample_type & PERF_SAMPLE_ID)
983 size += sizeof(data->id);
985 if (sample_type & PERF_SAMPLE_STREAM_ID)
986 size += sizeof(data->stream_id);
988 if (sample_type & PERF_SAMPLE_CPU)
989 size += sizeof(data->cpu_entry);
991 event->id_header_size = size;
994 static void perf_group_attach(struct perf_event *event)
996 struct perf_event *group_leader = event->group_leader, *pos;
999 * We can have double attach due to group movement in perf_event_open.
1001 if (event->attach_state & PERF_ATTACH_GROUP)
1004 event->attach_state |= PERF_ATTACH_GROUP;
1006 if (group_leader == event)
1009 if (group_leader->group_flags & PERF_GROUP_SOFTWARE &&
1010 !is_software_event(event))
1011 group_leader->group_flags &= ~PERF_GROUP_SOFTWARE;
1013 list_add_tail(&event->group_entry, &group_leader->sibling_list);
1014 group_leader->nr_siblings++;
1016 perf_event__header_size(group_leader);
1018 list_for_each_entry(pos, &group_leader->sibling_list, group_entry)
1019 perf_event__header_size(pos);
1023 * Remove a event from the lists for its context.
1024 * Must be called with ctx->mutex and ctx->lock held.
1027 list_del_event(struct perf_event *event, struct perf_event_context *ctx)
1029 struct perf_cpu_context *cpuctx;
1031 * We can have double detach due to exit/hot-unplug + close.
1033 if (!(event->attach_state & PERF_ATTACH_CONTEXT))
1036 event->attach_state &= ~PERF_ATTACH_CONTEXT;
1038 if (is_cgroup_event(event)) {
1040 cpuctx = __get_cpu_context(ctx);
1042 * if there are no more cgroup events
1043 * then cler cgrp to avoid stale pointer
1044 * in update_cgrp_time_from_cpuctx()
1046 if (!ctx->nr_cgroups)
1047 cpuctx->cgrp = NULL;
1050 if (has_branch_stack(event))
1051 ctx->nr_branch_stack--;
1054 if (event->attr.inherit_stat)
1057 list_del_rcu(&event->event_entry);
1059 if (event->group_leader == event)
1060 list_del_init(&event->group_entry);
1062 update_group_times(event);
1065 * If event was in error state, then keep it
1066 * that way, otherwise bogus counts will be
1067 * returned on read(). The only way to get out
1068 * of error state is by explicit re-enabling
1071 if (event->state > PERF_EVENT_STATE_OFF)
1072 event->state = PERF_EVENT_STATE_OFF;
1075 static void perf_group_detach(struct perf_event *event)
1077 struct perf_event *sibling, *tmp;
1078 struct list_head *list = NULL;
1081 * We can have double detach due to exit/hot-unplug + close.
1083 if (!(event->attach_state & PERF_ATTACH_GROUP))
1086 event->attach_state &= ~PERF_ATTACH_GROUP;
1089 * If this is a sibling, remove it from its group.
1091 if (event->group_leader != event) {
1092 list_del_init(&event->group_entry);
1093 event->group_leader->nr_siblings--;
1097 if (!list_empty(&event->group_entry))
1098 list = &event->group_entry;
1101 * If this was a group event with sibling events then
1102 * upgrade the siblings to singleton events by adding them
1103 * to whatever list we are on.
1105 list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) {
1107 list_move_tail(&sibling->group_entry, list);
1108 sibling->group_leader = sibling;
1110 /* Inherit group flags from the previous leader */
1111 sibling->group_flags = event->group_flags;
1115 perf_event__header_size(event->group_leader);
1117 list_for_each_entry(tmp, &event->group_leader->sibling_list, group_entry)
1118 perf_event__header_size(tmp);
1122 event_filter_match(struct perf_event *event)
1124 return (event->cpu == -1 || event->cpu == smp_processor_id())
1125 && perf_cgroup_match(event);
1129 event_sched_out(struct perf_event *event,
1130 struct perf_cpu_context *cpuctx,
1131 struct perf_event_context *ctx)
1133 u64 tstamp = perf_event_time(event);
1136 * An event which could not be activated because of
1137 * filter mismatch still needs to have its timings
1138 * maintained, otherwise bogus information is return
1139 * via read() for time_enabled, time_running:
1141 if (event->state == PERF_EVENT_STATE_INACTIVE
1142 && !event_filter_match(event)) {
1143 delta = tstamp - event->tstamp_stopped;
1144 event->tstamp_running += delta;
1145 event->tstamp_stopped = tstamp;
1148 if (event->state != PERF_EVENT_STATE_ACTIVE)
1151 event->state = PERF_EVENT_STATE_INACTIVE;
1152 if (event->pending_disable) {
1153 event->pending_disable = 0;
1154 event->state = PERF_EVENT_STATE_OFF;
1156 event->tstamp_stopped = tstamp;
1157 event->pmu->del(event, 0);
1160 if (!is_software_event(event))
1161 cpuctx->active_oncpu--;
1163 if (event->attr.freq && event->attr.sample_freq)
1165 if (event->attr.exclusive || !cpuctx->active_oncpu)
1166 cpuctx->exclusive = 0;
1170 group_sched_out(struct perf_event *group_event,
1171 struct perf_cpu_context *cpuctx,
1172 struct perf_event_context *ctx)
1174 struct perf_event *event;
1175 int state = group_event->state;
1177 event_sched_out(group_event, cpuctx, ctx);
1180 * Schedule out siblings (if any):
1182 list_for_each_entry(event, &group_event->sibling_list, group_entry)
1183 event_sched_out(event, cpuctx, ctx);
1185 if (state == PERF_EVENT_STATE_ACTIVE && group_event->attr.exclusive)
1186 cpuctx->exclusive = 0;
1190 * Cross CPU call to remove a performance event
1192 * We disable the event on the hardware level first. After that we
1193 * remove it from the context list.
1195 static int __perf_remove_from_context(void *info)
1197 struct perf_event *event = info;
1198 struct perf_event_context *ctx = event->ctx;
1199 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1201 raw_spin_lock(&ctx->lock);
1202 event_sched_out(event, cpuctx, ctx);
1203 list_del_event(event, ctx);
1204 if (!ctx->nr_events && cpuctx->task_ctx == ctx) {
1206 cpuctx->task_ctx = NULL;
1208 raw_spin_unlock(&ctx->lock);
1215 * Remove the event from a task's (or a CPU's) list of events.
1217 * CPU events are removed with a smp call. For task events we only
1218 * call when the task is on a CPU.
1220 * If event->ctx is a cloned context, callers must make sure that
1221 * every task struct that event->ctx->task could possibly point to
1222 * remains valid. This is OK when called from perf_release since
1223 * that only calls us on the top-level context, which can't be a clone.
1224 * When called from perf_event_exit_task, it's OK because the
1225 * context has been detached from its task.
1227 static void perf_remove_from_context(struct perf_event *event)
1229 struct perf_event_context *ctx = event->ctx;
1230 struct task_struct *task = ctx->task;
1232 lockdep_assert_held(&ctx->mutex);
1236 * Per cpu events are removed via an smp call and
1237 * the removal is always successful.
1239 cpu_function_call(event->cpu, __perf_remove_from_context, event);
1244 if (!task_function_call(task, __perf_remove_from_context, event))
1247 raw_spin_lock_irq(&ctx->lock);
1249 * If we failed to find a running task, but find the context active now
1250 * that we've acquired the ctx->lock, retry.
1252 if (ctx->is_active) {
1253 raw_spin_unlock_irq(&ctx->lock);
1258 * Since the task isn't running, its safe to remove the event, us
1259 * holding the ctx->lock ensures the task won't get scheduled in.
1261 list_del_event(event, ctx);
1262 raw_spin_unlock_irq(&ctx->lock);
1266 * Cross CPU call to disable a performance event
1268 int __perf_event_disable(void *info)
1270 struct perf_event *event = info;
1271 struct perf_event_context *ctx = event->ctx;
1272 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1275 * If this is a per-task event, need to check whether this
1276 * event's task is the current task on this cpu.
1278 * Can trigger due to concurrent perf_event_context_sched_out()
1279 * flipping contexts around.
1281 if (ctx->task && cpuctx->task_ctx != ctx)
1284 raw_spin_lock(&ctx->lock);
1287 * If the event is on, turn it off.
1288 * If it is in error state, leave it in error state.
1290 if (event->state >= PERF_EVENT_STATE_INACTIVE) {
1291 update_context_time(ctx);
1292 update_cgrp_time_from_event(event);
1293 update_group_times(event);
1294 if (event == event->group_leader)
1295 group_sched_out(event, cpuctx, ctx);
1297 event_sched_out(event, cpuctx, ctx);
1298 event->state = PERF_EVENT_STATE_OFF;
1301 raw_spin_unlock(&ctx->lock);
1309 * If event->ctx is a cloned context, callers must make sure that
1310 * every task struct that event->ctx->task could possibly point to
1311 * remains valid. This condition is satisifed when called through
1312 * perf_event_for_each_child or perf_event_for_each because they
1313 * hold the top-level event's child_mutex, so any descendant that
1314 * goes to exit will block in sync_child_event.
1315 * When called from perf_pending_event it's OK because event->ctx
1316 * is the current context on this CPU and preemption is disabled,
1317 * hence we can't get into perf_event_task_sched_out for this context.
1319 void perf_event_disable(struct perf_event *event)
1321 struct perf_event_context *ctx = event->ctx;
1322 struct task_struct *task = ctx->task;
1326 * Disable the event on the cpu that it's on
1328 cpu_function_call(event->cpu, __perf_event_disable, event);
1333 if (!task_function_call(task, __perf_event_disable, event))
1336 raw_spin_lock_irq(&ctx->lock);
1338 * If the event is still active, we need to retry the cross-call.
1340 if (event->state == PERF_EVENT_STATE_ACTIVE) {
1341 raw_spin_unlock_irq(&ctx->lock);
1343 * Reload the task pointer, it might have been changed by
1344 * a concurrent perf_event_context_sched_out().
1351 * Since we have the lock this context can't be scheduled
1352 * in, so we can change the state safely.
1354 if (event->state == PERF_EVENT_STATE_INACTIVE) {
1355 update_group_times(event);
1356 event->state = PERF_EVENT_STATE_OFF;
1358 raw_spin_unlock_irq(&ctx->lock);
1360 EXPORT_SYMBOL_GPL(perf_event_disable);
1362 static void perf_set_shadow_time(struct perf_event *event,
1363 struct perf_event_context *ctx,
1367 * use the correct time source for the time snapshot
1369 * We could get by without this by leveraging the
1370 * fact that to get to this function, the caller
1371 * has most likely already called update_context_time()
1372 * and update_cgrp_time_xx() and thus both timestamp
1373 * are identical (or very close). Given that tstamp is,
1374 * already adjusted for cgroup, we could say that:
1375 * tstamp - ctx->timestamp
1377 * tstamp - cgrp->timestamp.
1379 * Then, in perf_output_read(), the calculation would
1380 * work with no changes because:
1381 * - event is guaranteed scheduled in
1382 * - no scheduled out in between
1383 * - thus the timestamp would be the same
1385 * But this is a bit hairy.
1387 * So instead, we have an explicit cgroup call to remain
1388 * within the time time source all along. We believe it
1389 * is cleaner and simpler to understand.
1391 if (is_cgroup_event(event))
1392 perf_cgroup_set_shadow_time(event, tstamp);
1394 event->shadow_ctx_time = tstamp - ctx->timestamp;
1397 #define MAX_INTERRUPTS (~0ULL)
1399 static void perf_log_throttle(struct perf_event *event, int enable);
1402 event_sched_in(struct perf_event *event,
1403 struct perf_cpu_context *cpuctx,
1404 struct perf_event_context *ctx)
1406 u64 tstamp = perf_event_time(event);
1408 if (event->state <= PERF_EVENT_STATE_OFF)
1411 event->state = PERF_EVENT_STATE_ACTIVE;
1412 event->oncpu = smp_processor_id();
1415 * Unthrottle events, since we scheduled we might have missed several
1416 * ticks already, also for a heavily scheduling task there is little
1417 * guarantee it'll get a tick in a timely manner.
1419 if (unlikely(event->hw.interrupts == MAX_INTERRUPTS)) {
1420 perf_log_throttle(event, 1);
1421 event->hw.interrupts = 0;
1425 * The new state must be visible before we turn it on in the hardware:
1429 if (event->pmu->add(event, PERF_EF_START)) {
1430 event->state = PERF_EVENT_STATE_INACTIVE;
1435 event->tstamp_running += tstamp - event->tstamp_stopped;
1437 perf_set_shadow_time(event, ctx, tstamp);
1439 if (!is_software_event(event))
1440 cpuctx->active_oncpu++;
1442 if (event->attr.freq && event->attr.sample_freq)
1445 if (event->attr.exclusive)
1446 cpuctx->exclusive = 1;
1452 group_sched_in(struct perf_event *group_event,
1453 struct perf_cpu_context *cpuctx,
1454 struct perf_event_context *ctx)
1456 struct perf_event *event, *partial_group = NULL;
1457 struct pmu *pmu = group_event->pmu;
1458 u64 now = ctx->time;
1459 bool simulate = false;
1461 if (group_event->state == PERF_EVENT_STATE_OFF)
1464 pmu->start_txn(pmu);
1466 if (event_sched_in(group_event, cpuctx, ctx)) {
1467 pmu->cancel_txn(pmu);
1472 * Schedule in siblings as one group (if any):
1474 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
1475 if (event_sched_in(event, cpuctx, ctx)) {
1476 partial_group = event;
1481 if (!pmu->commit_txn(pmu))
1486 * Groups can be scheduled in as one unit only, so undo any
1487 * partial group before returning:
1488 * The events up to the failed event are scheduled out normally,
1489 * tstamp_stopped will be updated.
1491 * The failed events and the remaining siblings need to have
1492 * their timings updated as if they had gone thru event_sched_in()
1493 * and event_sched_out(). This is required to get consistent timings
1494 * across the group. This also takes care of the case where the group
1495 * could never be scheduled by ensuring tstamp_stopped is set to mark
1496 * the time the event was actually stopped, such that time delta
1497 * calculation in update_event_times() is correct.
1499 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
1500 if (event == partial_group)
1504 event->tstamp_running += now - event->tstamp_stopped;
1505 event->tstamp_stopped = now;
1507 event_sched_out(event, cpuctx, ctx);
1510 event_sched_out(group_event, cpuctx, ctx);
1512 pmu->cancel_txn(pmu);
1518 * Work out whether we can put this event group on the CPU now.
1520 static int group_can_go_on(struct perf_event *event,
1521 struct perf_cpu_context *cpuctx,
1525 * Groups consisting entirely of software events can always go on.
1527 if (event->group_flags & PERF_GROUP_SOFTWARE)
1530 * If an exclusive group is already on, no other hardware
1533 if (cpuctx->exclusive)
1536 * If this group is exclusive and there are already
1537 * events on the CPU, it can't go on.
1539 if (event->attr.exclusive && cpuctx->active_oncpu)
1542 * Otherwise, try to add it if all previous groups were able
1548 static void add_event_to_ctx(struct perf_event *event,
1549 struct perf_event_context *ctx)
1551 u64 tstamp = perf_event_time(event);
1553 list_add_event(event, ctx);
1554 perf_group_attach(event);
1555 event->tstamp_enabled = tstamp;
1556 event->tstamp_running = tstamp;
1557 event->tstamp_stopped = tstamp;
1560 static void task_ctx_sched_out(struct perf_event_context *ctx);
1562 ctx_sched_in(struct perf_event_context *ctx,
1563 struct perf_cpu_context *cpuctx,
1564 enum event_type_t event_type,
1565 struct task_struct *task);
1567 static void perf_event_sched_in(struct perf_cpu_context *cpuctx,
1568 struct perf_event_context *ctx,
1569 struct task_struct *task)
1571 cpu_ctx_sched_in(cpuctx, EVENT_PINNED, task);
1573 ctx_sched_in(ctx, cpuctx, EVENT_PINNED, task);
1574 cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE, task);
1576 ctx_sched_in(ctx, cpuctx, EVENT_FLEXIBLE, task);
1580 * Cross CPU call to install and enable a performance event
1582 * Must be called with ctx->mutex held
1584 static int __perf_install_in_context(void *info)
1586 struct perf_event *event = info;
1587 struct perf_event_context *ctx = event->ctx;
1588 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1589 struct perf_event_context *task_ctx = cpuctx->task_ctx;
1590 struct task_struct *task = current;
1592 perf_ctx_lock(cpuctx, task_ctx);
1593 perf_pmu_disable(cpuctx->ctx.pmu);
1596 * If there was an active task_ctx schedule it out.
1599 task_ctx_sched_out(task_ctx);
1602 * If the context we're installing events in is not the
1603 * active task_ctx, flip them.
1605 if (ctx->task && task_ctx != ctx) {
1607 raw_spin_unlock(&task_ctx->lock);
1608 raw_spin_lock(&ctx->lock);
1613 cpuctx->task_ctx = task_ctx;
1614 task = task_ctx->task;
1617 cpu_ctx_sched_out(cpuctx, EVENT_ALL);
1619 update_context_time(ctx);
1621 * update cgrp time only if current cgrp
1622 * matches event->cgrp. Must be done before
1623 * calling add_event_to_ctx()
1625 update_cgrp_time_from_event(event);
1627 add_event_to_ctx(event, ctx);
1630 * Schedule everything back in
1632 perf_event_sched_in(cpuctx, task_ctx, task);
1634 perf_pmu_enable(cpuctx->ctx.pmu);
1635 perf_ctx_unlock(cpuctx, task_ctx);
1641 * Attach a performance event to a context
1643 * First we add the event to the list with the hardware enable bit
1644 * in event->hw_config cleared.
1646 * If the event is attached to a task which is on a CPU we use a smp
1647 * call to enable it in the task context. The task might have been
1648 * scheduled away, but we check this in the smp call again.
1651 perf_install_in_context(struct perf_event_context *ctx,
1652 struct perf_event *event,
1655 struct task_struct *task = ctx->task;
1657 lockdep_assert_held(&ctx->mutex);
1660 if (event->cpu != -1)
1665 * Per cpu events are installed via an smp call and
1666 * the install is always successful.
1668 cpu_function_call(cpu, __perf_install_in_context, event);
1673 if (!task_function_call(task, __perf_install_in_context, event))
1676 raw_spin_lock_irq(&ctx->lock);
1678 * If we failed to find a running task, but find the context active now
1679 * that we've acquired the ctx->lock, retry.
1681 if (ctx->is_active) {
1682 raw_spin_unlock_irq(&ctx->lock);
1687 * Since the task isn't running, its safe to add the event, us holding
1688 * the ctx->lock ensures the task won't get scheduled in.
1690 add_event_to_ctx(event, ctx);
1691 raw_spin_unlock_irq(&ctx->lock);
1695 * Put a event into inactive state and update time fields.
1696 * Enabling the leader of a group effectively enables all
1697 * the group members that aren't explicitly disabled, so we
1698 * have to update their ->tstamp_enabled also.
1699 * Note: this works for group members as well as group leaders
1700 * since the non-leader members' sibling_lists will be empty.
1702 static void __perf_event_mark_enabled(struct perf_event *event)
1704 struct perf_event *sub;
1705 u64 tstamp = perf_event_time(event);
1707 event->state = PERF_EVENT_STATE_INACTIVE;
1708 event->tstamp_enabled = tstamp - event->total_time_enabled;
1709 list_for_each_entry(sub, &event->sibling_list, group_entry) {
1710 if (sub->state >= PERF_EVENT_STATE_INACTIVE)
1711 sub->tstamp_enabled = tstamp - sub->total_time_enabled;
1716 * Cross CPU call to enable a performance event
1718 static int __perf_event_enable(void *info)
1720 struct perf_event *event = info;
1721 struct perf_event_context *ctx = event->ctx;
1722 struct perf_event *leader = event->group_leader;
1723 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1726 if (WARN_ON_ONCE(!ctx->is_active))
1729 raw_spin_lock(&ctx->lock);
1730 update_context_time(ctx);
1732 if (event->state >= PERF_EVENT_STATE_INACTIVE)
1736 * set current task's cgroup time reference point
1738 perf_cgroup_set_timestamp(current, ctx);
1740 __perf_event_mark_enabled(event);
1742 if (!event_filter_match(event)) {
1743 if (is_cgroup_event(event))
1744 perf_cgroup_defer_enabled(event);
1749 * If the event is in a group and isn't the group leader,
1750 * then don't put it on unless the group is on.
1752 if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE)
1755 if (!group_can_go_on(event, cpuctx, 1)) {
1758 if (event == leader)
1759 err = group_sched_in(event, cpuctx, ctx);
1761 err = event_sched_in(event, cpuctx, ctx);
1766 * If this event can't go on and it's part of a
1767 * group, then the whole group has to come off.
1769 if (leader != event)
1770 group_sched_out(leader, cpuctx, ctx);
1771 if (leader->attr.pinned) {
1772 update_group_times(leader);
1773 leader->state = PERF_EVENT_STATE_ERROR;
1778 raw_spin_unlock(&ctx->lock);
1786 * If event->ctx is a cloned context, callers must make sure that
1787 * every task struct that event->ctx->task could possibly point to
1788 * remains valid. This condition is satisfied when called through
1789 * perf_event_for_each_child or perf_event_for_each as described
1790 * for perf_event_disable.
1792 void perf_event_enable(struct perf_event *event)
1794 struct perf_event_context *ctx = event->ctx;
1795 struct task_struct *task = ctx->task;
1799 * Enable the event on the cpu that it's on
1801 cpu_function_call(event->cpu, __perf_event_enable, event);
1805 raw_spin_lock_irq(&ctx->lock);
1806 if (event->state >= PERF_EVENT_STATE_INACTIVE)
1810 * If the event is in error state, clear that first.
1811 * That way, if we see the event in error state below, we
1812 * know that it has gone back into error state, as distinct
1813 * from the task having been scheduled away before the
1814 * cross-call arrived.
1816 if (event->state == PERF_EVENT_STATE_ERROR)
1817 event->state = PERF_EVENT_STATE_OFF;
1820 if (!ctx->is_active) {
1821 __perf_event_mark_enabled(event);
1825 raw_spin_unlock_irq(&ctx->lock);
1827 if (!task_function_call(task, __perf_event_enable, event))
1830 raw_spin_lock_irq(&ctx->lock);
1833 * If the context is active and the event is still off,
1834 * we need to retry the cross-call.
1836 if (ctx->is_active && event->state == PERF_EVENT_STATE_OFF) {
1838 * task could have been flipped by a concurrent
1839 * perf_event_context_sched_out()
1846 raw_spin_unlock_irq(&ctx->lock);
1848 EXPORT_SYMBOL_GPL(perf_event_enable);
1850 int perf_event_refresh(struct perf_event *event, int refresh)
1853 * not supported on inherited events
1855 if (event->attr.inherit || !is_sampling_event(event))
1858 atomic_add(refresh, &event->event_limit);
1859 perf_event_enable(event);
1863 EXPORT_SYMBOL_GPL(perf_event_refresh);
1865 static void ctx_sched_out(struct perf_event_context *ctx,
1866 struct perf_cpu_context *cpuctx,
1867 enum event_type_t event_type)
1869 struct perf_event *event;
1870 int is_active = ctx->is_active;
1872 ctx->is_active &= ~event_type;
1873 if (likely(!ctx->nr_events))
1876 update_context_time(ctx);
1877 update_cgrp_time_from_cpuctx(cpuctx);
1878 if (!ctx->nr_active)
1881 perf_pmu_disable(ctx->pmu);
1882 if ((is_active & EVENT_PINNED) && (event_type & EVENT_PINNED)) {
1883 list_for_each_entry(event, &ctx->pinned_groups, group_entry)
1884 group_sched_out(event, cpuctx, ctx);
1887 if ((is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE)) {
1888 list_for_each_entry(event, &ctx->flexible_groups, group_entry)
1889 group_sched_out(event, cpuctx, ctx);
1891 perf_pmu_enable(ctx->pmu);
1895 * Test whether two contexts are equivalent, i.e. whether they
1896 * have both been cloned from the same version of the same context
1897 * and they both have the same number of enabled events.
1898 * If the number of enabled events is the same, then the set
1899 * of enabled events should be the same, because these are both
1900 * inherited contexts, therefore we can't access individual events
1901 * in them directly with an fd; we can only enable/disable all
1902 * events via prctl, or enable/disable all events in a family
1903 * via ioctl, which will have the same effect on both contexts.
1905 static int context_equiv(struct perf_event_context *ctx1,
1906 struct perf_event_context *ctx2)
1908 return ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx
1909 && ctx1->parent_gen == ctx2->parent_gen
1910 && !ctx1->pin_count && !ctx2->pin_count;
1913 static void __perf_event_sync_stat(struct perf_event *event,
1914 struct perf_event *next_event)
1918 if (!event->attr.inherit_stat)
1922 * Update the event value, we cannot use perf_event_read()
1923 * because we're in the middle of a context switch and have IRQs
1924 * disabled, which upsets smp_call_function_single(), however
1925 * we know the event must be on the current CPU, therefore we
1926 * don't need to use it.
1928 switch (event->state) {
1929 case PERF_EVENT_STATE_ACTIVE:
1930 event->pmu->read(event);
1933 case PERF_EVENT_STATE_INACTIVE:
1934 update_event_times(event);
1942 * In order to keep per-task stats reliable we need to flip the event
1943 * values when we flip the contexts.
1945 value = local64_read(&next_event->count);
1946 value = local64_xchg(&event->count, value);
1947 local64_set(&next_event->count, value);
1949 swap(event->total_time_enabled, next_event->total_time_enabled);
1950 swap(event->total_time_running, next_event->total_time_running);
1953 * Since we swizzled the values, update the user visible data too.
1955 perf_event_update_userpage(event);
1956 perf_event_update_userpage(next_event);
1959 #define list_next_entry(pos, member) \
1960 list_entry(pos->member.next, typeof(*pos), member)
1962 static void perf_event_sync_stat(struct perf_event_context *ctx,
1963 struct perf_event_context *next_ctx)
1965 struct perf_event *event, *next_event;
1970 update_context_time(ctx);
1972 event = list_first_entry(&ctx->event_list,
1973 struct perf_event, event_entry);
1975 next_event = list_first_entry(&next_ctx->event_list,
1976 struct perf_event, event_entry);
1978 while (&event->event_entry != &ctx->event_list &&
1979 &next_event->event_entry != &next_ctx->event_list) {
1981 __perf_event_sync_stat(event, next_event);
1983 event = list_next_entry(event, event_entry);
1984 next_event = list_next_entry(next_event, event_entry);
1988 static void perf_event_context_sched_out(struct task_struct *task, int ctxn,
1989 struct task_struct *next)
1991 struct perf_event_context *ctx = task->perf_event_ctxp[ctxn];
1992 struct perf_event_context *next_ctx;
1993 struct perf_event_context *parent;
1994 struct perf_cpu_context *cpuctx;
2000 cpuctx = __get_cpu_context(ctx);
2001 if (!cpuctx->task_ctx)
2005 parent = rcu_dereference(ctx->parent_ctx);
2006 next_ctx = next->perf_event_ctxp[ctxn];
2007 if (parent && next_ctx &&
2008 rcu_dereference(next_ctx->parent_ctx) == parent) {
2010 * Looks like the two contexts are clones, so we might be
2011 * able to optimize the context switch. We lock both
2012 * contexts and check that they are clones under the
2013 * lock (including re-checking that neither has been
2014 * uncloned in the meantime). It doesn't matter which
2015 * order we take the locks because no other cpu could
2016 * be trying to lock both of these tasks.
2018 raw_spin_lock(&ctx->lock);
2019 raw_spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING);
2020 if (context_equiv(ctx, next_ctx)) {
2022 * XXX do we need a memory barrier of sorts
2023 * wrt to rcu_dereference() of perf_event_ctxp
2025 task->perf_event_ctxp[ctxn] = next_ctx;
2026 next->perf_event_ctxp[ctxn] = ctx;
2028 next_ctx->task = task;
2031 perf_event_sync_stat(ctx, next_ctx);
2033 raw_spin_unlock(&next_ctx->lock);
2034 raw_spin_unlock(&ctx->lock);
2039 raw_spin_lock(&ctx->lock);
2040 ctx_sched_out(ctx, cpuctx, EVENT_ALL);
2041 cpuctx->task_ctx = NULL;
2042 raw_spin_unlock(&ctx->lock);
2046 #define for_each_task_context_nr(ctxn) \
2047 for ((ctxn) = 0; (ctxn) < perf_nr_task_contexts; (ctxn)++)
2050 * Called from scheduler to remove the events of the current task,
2051 * with interrupts disabled.
2053 * We stop each event and update the event value in event->count.
2055 * This does not protect us against NMI, but disable()
2056 * sets the disabled bit in the control field of event _before_
2057 * accessing the event control register. If a NMI hits, then it will
2058 * not restart the event.
2060 void __perf_event_task_sched_out(struct task_struct *task,
2061 struct task_struct *next)
2065 for_each_task_context_nr(ctxn)
2066 perf_event_context_sched_out(task, ctxn, next);
2069 * if cgroup events exist on this CPU, then we need
2070 * to check if we have to switch out PMU state.
2071 * cgroup event are system-wide mode only
2073 if (atomic_read(&__get_cpu_var(perf_cgroup_events)))
2074 perf_cgroup_sched_out(task, next);
2077 static void task_ctx_sched_out(struct perf_event_context *ctx)
2079 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
2081 if (!cpuctx->task_ctx)
2084 if (WARN_ON_ONCE(ctx != cpuctx->task_ctx))
2087 ctx_sched_out(ctx, cpuctx, EVENT_ALL);
2088 cpuctx->task_ctx = NULL;
2092 * Called with IRQs disabled
2094 static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
2095 enum event_type_t event_type)
2097 ctx_sched_out(&cpuctx->ctx, cpuctx, event_type);
2101 ctx_pinned_sched_in(struct perf_event_context *ctx,
2102 struct perf_cpu_context *cpuctx)
2104 struct perf_event *event;
2106 list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
2107 if (event->state <= PERF_EVENT_STATE_OFF)
2109 if (!event_filter_match(event))
2112 /* may need to reset tstamp_enabled */
2113 if (is_cgroup_event(event))
2114 perf_cgroup_mark_enabled(event, ctx);
2116 if (group_can_go_on(event, cpuctx, 1))
2117 group_sched_in(event, cpuctx, ctx);
2120 * If this pinned group hasn't been scheduled,
2121 * put it in error state.
2123 if (event->state == PERF_EVENT_STATE_INACTIVE) {
2124 update_group_times(event);
2125 event->state = PERF_EVENT_STATE_ERROR;
2131 ctx_flexible_sched_in(struct perf_event_context *ctx,
2132 struct perf_cpu_context *cpuctx)
2134 struct perf_event *event;
2137 list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
2138 /* Ignore events in OFF or ERROR state */
2139 if (event->state <= PERF_EVENT_STATE_OFF)
2142 * Listen to the 'cpu' scheduling filter constraint
2145 if (!event_filter_match(event))
2148 /* may need to reset tstamp_enabled */
2149 if (is_cgroup_event(event))
2150 perf_cgroup_mark_enabled(event, ctx);
2152 if (group_can_go_on(event, cpuctx, can_add_hw)) {
2153 if (group_sched_in(event, cpuctx, ctx))
2160 ctx_sched_in(struct perf_event_context *ctx,
2161 struct perf_cpu_context *cpuctx,
2162 enum event_type_t event_type,
2163 struct task_struct *task)
2166 int is_active = ctx->is_active;
2168 ctx->is_active |= event_type;
2169 if (likely(!ctx->nr_events))
2173 ctx->timestamp = now;
2174 perf_cgroup_set_timestamp(task, ctx);
2176 * First go through the list and put on any pinned groups
2177 * in order to give them the best chance of going on.
2179 if (!(is_active & EVENT_PINNED) && (event_type & EVENT_PINNED))
2180 ctx_pinned_sched_in(ctx, cpuctx);
2182 /* Then walk through the lower prio flexible groups */
2183 if (!(is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE))
2184 ctx_flexible_sched_in(ctx, cpuctx);
2187 static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
2188 enum event_type_t event_type,
2189 struct task_struct *task)
2191 struct perf_event_context *ctx = &cpuctx->ctx;
2193 ctx_sched_in(ctx, cpuctx, event_type, task);
2196 static void perf_event_context_sched_in(struct perf_event_context *ctx,
2197 struct task_struct *task)
2199 struct perf_cpu_context *cpuctx;
2201 cpuctx = __get_cpu_context(ctx);
2202 if (cpuctx->task_ctx == ctx)
2205 perf_ctx_lock(cpuctx, ctx);
2206 perf_pmu_disable(ctx->pmu);
2208 * We want to keep the following priority order:
2209 * cpu pinned (that don't need to move), task pinned,
2210 * cpu flexible, task flexible.
2212 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
2215 cpuctx->task_ctx = ctx;
2217 perf_event_sched_in(cpuctx, cpuctx->task_ctx, task);
2219 perf_pmu_enable(ctx->pmu);
2220 perf_ctx_unlock(cpuctx, ctx);
2223 * Since these rotations are per-cpu, we need to ensure the
2224 * cpu-context we got scheduled on is actually rotating.
2226 perf_pmu_rotate_start(ctx->pmu);
2230 * When sampling the branck stack in system-wide, it may be necessary
2231 * to flush the stack on context switch. This happens when the branch
2232 * stack does not tag its entries with the pid of the current task.
2233 * Otherwise it becomes impossible to associate a branch entry with a
2234 * task. This ambiguity is more likely to appear when the branch stack
2235 * supports priv level filtering and the user sets it to monitor only
2236 * at the user level (which could be a useful measurement in system-wide
2237 * mode). In that case, the risk is high of having a branch stack with
2238 * branch from multiple tasks. Flushing may mean dropping the existing
2239 * entries or stashing them somewhere in the PMU specific code layer.
2241 * This function provides the context switch callback to the lower code
2242 * layer. It is invoked ONLY when there is at least one system-wide context
2243 * with at least one active event using taken branch sampling.
2245 static void perf_branch_stack_sched_in(struct task_struct *prev,
2246 struct task_struct *task)
2248 struct perf_cpu_context *cpuctx;
2250 unsigned long flags;
2252 /* no need to flush branch stack if not changing task */
2256 local_irq_save(flags);
2260 list_for_each_entry_rcu(pmu, &pmus, entry) {
2261 cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
2264 * check if the context has at least one
2265 * event using PERF_SAMPLE_BRANCH_STACK
2267 if (cpuctx->ctx.nr_branch_stack > 0
2268 && pmu->flush_branch_stack) {
2270 pmu = cpuctx->ctx.pmu;
2272 perf_ctx_lock(cpuctx, cpuctx->task_ctx);
2274 perf_pmu_disable(pmu);
2276 pmu->flush_branch_stack();
2278 perf_pmu_enable(pmu);
2280 perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
2286 local_irq_restore(flags);
2290 * Called from scheduler to add the events of the current task
2291 * with interrupts disabled.
2293 * We restore the event value and then enable it.
2295 * This does not protect us against NMI, but enable()
2296 * sets the enabled bit in the control field of event _before_
2297 * accessing the event control register. If a NMI hits, then it will
2298 * keep the event running.
2300 void __perf_event_task_sched_in(struct task_struct *prev,
2301 struct task_struct *task)
2303 struct perf_event_context *ctx;
2306 for_each_task_context_nr(ctxn) {
2307 ctx = task->perf_event_ctxp[ctxn];
2311 perf_event_context_sched_in(ctx, task);
2314 * if cgroup events exist on this CPU, then we need
2315 * to check if we have to switch in PMU state.
2316 * cgroup event are system-wide mode only
2318 if (atomic_read(&__get_cpu_var(perf_cgroup_events)))
2319 perf_cgroup_sched_in(prev, task);
2321 /* check for system-wide branch_stack events */
2322 if (atomic_read(&__get_cpu_var(perf_branch_stack_events)))
2323 perf_branch_stack_sched_in(prev, task);
2326 static u64 perf_calculate_period(struct perf_event *event, u64 nsec, u64 count)
2328 u64 frequency = event->attr.sample_freq;
2329 u64 sec = NSEC_PER_SEC;
2330 u64 divisor, dividend;
2332 int count_fls, nsec_fls, frequency_fls, sec_fls;
2334 count_fls = fls64(count);
2335 nsec_fls = fls64(nsec);
2336 frequency_fls = fls64(frequency);
2340 * We got @count in @nsec, with a target of sample_freq HZ
2341 * the target period becomes:
2344 * period = -------------------
2345 * @nsec * sample_freq
2350 * Reduce accuracy by one bit such that @a and @b converge
2351 * to a similar magnitude.
2353 #define REDUCE_FLS(a, b) \
2355 if (a##_fls > b##_fls) { \
2365 * Reduce accuracy until either term fits in a u64, then proceed with
2366 * the other, so that finally we can do a u64/u64 division.
2368 while (count_fls + sec_fls > 64 && nsec_fls + frequency_fls > 64) {
2369 REDUCE_FLS(nsec, frequency);
2370 REDUCE_FLS(sec, count);
2373 if (count_fls + sec_fls > 64) {
2374 divisor = nsec * frequency;
2376 while (count_fls + sec_fls > 64) {
2377 REDUCE_FLS(count, sec);
2381 dividend = count * sec;
2383 dividend = count * sec;
2385 while (nsec_fls + frequency_fls > 64) {
2386 REDUCE_FLS(nsec, frequency);
2390 divisor = nsec * frequency;
2396 return div64_u64(dividend, divisor);
2399 static DEFINE_PER_CPU(int, perf_throttled_count);
2400 static DEFINE_PER_CPU(u64, perf_throttled_seq);
2402 static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count, bool disable)
2404 struct hw_perf_event *hwc = &event->hw;
2405 s64 period, sample_period;
2408 period = perf_calculate_period(event, nsec, count);
2410 delta = (s64)(period - hwc->sample_period);
2411 delta = (delta + 7) / 8; /* low pass filter */
2413 sample_period = hwc->sample_period + delta;
2418 hwc->sample_period = sample_period;
2420 if (local64_read(&hwc->period_left) > 8*sample_period) {
2422 event->pmu->stop(event, PERF_EF_UPDATE);
2424 local64_set(&hwc->period_left, 0);
2427 event->pmu->start(event, PERF_EF_RELOAD);
2432 * combine freq adjustment with unthrottling to avoid two passes over the
2433 * events. At the same time, make sure, having freq events does not change
2434 * the rate of unthrottling as that would introduce bias.
2436 static void perf_adjust_freq_unthr_context(struct perf_event_context *ctx,
2439 struct perf_event *event;
2440 struct hw_perf_event *hwc;
2441 u64 now, period = TICK_NSEC;
2445 * only need to iterate over all events iff:
2446 * - context have events in frequency mode (needs freq adjust)
2447 * - there are events to unthrottle on this cpu
2449 if (!(ctx->nr_freq || needs_unthr))
2452 raw_spin_lock(&ctx->lock);
2453 perf_pmu_disable(ctx->pmu);
2455 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
2456 if (event->state != PERF_EVENT_STATE_ACTIVE)
2459 if (!event_filter_match(event))
2464 if (needs_unthr && hwc->interrupts == MAX_INTERRUPTS) {
2465 hwc->interrupts = 0;
2466 perf_log_throttle(event, 1);
2467 event->pmu->start(event, 0);
2470 if (!event->attr.freq || !event->attr.sample_freq)
2474 * stop the event and update event->count
2476 event->pmu->stop(event, PERF_EF_UPDATE);
2478 now = local64_read(&event->count);
2479 delta = now - hwc->freq_count_stamp;
2480 hwc->freq_count_stamp = now;
2484 * reload only if value has changed
2485 * we have stopped the event so tell that
2486 * to perf_adjust_period() to avoid stopping it
2490 perf_adjust_period(event, period, delta, false);
2492 event->pmu->start(event, delta > 0 ? PERF_EF_RELOAD : 0);
2495 perf_pmu_enable(ctx->pmu);
2496 raw_spin_unlock(&ctx->lock);
2500 * Round-robin a context's events:
2502 static void rotate_ctx(struct perf_event_context *ctx)
2505 * Rotate the first entry last of non-pinned groups. Rotation might be
2506 * disabled by the inheritance code.
2508 if (!ctx->rotate_disable)
2509 list_rotate_left(&ctx->flexible_groups);
2513 * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
2514 * because they're strictly cpu affine and rotate_start is called with IRQs
2515 * disabled, while rotate_context is called from IRQ context.
2517 static void perf_rotate_context(struct perf_cpu_context *cpuctx)
2519 struct perf_event_context *ctx = NULL;
2520 int rotate = 0, remove = 1;
2522 if (cpuctx->ctx.nr_events) {
2524 if (cpuctx->ctx.nr_events != cpuctx->ctx.nr_active)
2528 ctx = cpuctx->task_ctx;
2529 if (ctx && ctx->nr_events) {
2531 if (ctx->nr_events != ctx->nr_active)
2538 perf_ctx_lock(cpuctx, cpuctx->task_ctx);
2539 perf_pmu_disable(cpuctx->ctx.pmu);
2541 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
2543 ctx_sched_out(ctx, cpuctx, EVENT_FLEXIBLE);
2545 rotate_ctx(&cpuctx->ctx);
2549 perf_event_sched_in(cpuctx, ctx, current);
2551 perf_pmu_enable(cpuctx->ctx.pmu);
2552 perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
2555 list_del_init(&cpuctx->rotation_list);
2558 void perf_event_task_tick(void)
2560 struct list_head *head = &__get_cpu_var(rotation_list);
2561 struct perf_cpu_context *cpuctx, *tmp;
2562 struct perf_event_context *ctx;
2565 WARN_ON(!irqs_disabled());
2567 __this_cpu_inc(perf_throttled_seq);
2568 throttled = __this_cpu_xchg(perf_throttled_count, 0);
2570 list_for_each_entry_safe(cpuctx, tmp, head, rotation_list) {
2572 perf_adjust_freq_unthr_context(ctx, throttled);
2574 ctx = cpuctx->task_ctx;
2576 perf_adjust_freq_unthr_context(ctx, throttled);
2578 if (cpuctx->jiffies_interval == 1 ||
2579 !(jiffies % cpuctx->jiffies_interval))
2580 perf_rotate_context(cpuctx);
2584 static int event_enable_on_exec(struct perf_event *event,
2585 struct perf_event_context *ctx)
2587 if (!event->attr.enable_on_exec)
2590 event->attr.enable_on_exec = 0;
2591 if (event->state >= PERF_EVENT_STATE_INACTIVE)
2594 __perf_event_mark_enabled(event);
2600 * Enable all of a task's events that have been marked enable-on-exec.
2601 * This expects task == current.
2603 static void perf_event_enable_on_exec(struct perf_event_context *ctx)
2605 struct perf_event *event;
2606 unsigned long flags;
2610 local_irq_save(flags);
2611 if (!ctx || !ctx->nr_events)
2615 * We must ctxsw out cgroup events to avoid conflict
2616 * when invoking perf_task_event_sched_in() later on
2617 * in this function. Otherwise we end up trying to
2618 * ctxswin cgroup events which are already scheduled
2621 perf_cgroup_sched_out(current, NULL);
2623 raw_spin_lock(&ctx->lock);
2624 task_ctx_sched_out(ctx);
2626 list_for_each_entry(event, &ctx->event_list, event_entry) {
2627 ret = event_enable_on_exec(event, ctx);
2633 * Unclone this context if we enabled any event.
2638 raw_spin_unlock(&ctx->lock);
2641 * Also calls ctxswin for cgroup events, if any:
2643 perf_event_context_sched_in(ctx, ctx->task);
2645 local_irq_restore(flags);
2649 * Cross CPU call to read the hardware event
2651 static void __perf_event_read(void *info)
2653 struct perf_event *event = info;
2654 struct perf_event_context *ctx = event->ctx;
2655 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
2658 * If this is a task context, we need to check whether it is
2659 * the current task context of this cpu. If not it has been
2660 * scheduled out before the smp call arrived. In that case
2661 * event->count would have been updated to a recent sample
2662 * when the event was scheduled out.
2664 if (ctx->task && cpuctx->task_ctx != ctx)
2667 raw_spin_lock(&ctx->lock);
2668 if (ctx->is_active) {
2669 update_context_time(ctx);
2670 update_cgrp_time_from_event(event);
2672 update_event_times(event);
2673 if (event->state == PERF_EVENT_STATE_ACTIVE)
2674 event->pmu->read(event);
2675 raw_spin_unlock(&ctx->lock);
2678 static inline u64 perf_event_count(struct perf_event *event)
2680 return local64_read(&event->count) + atomic64_read(&event->child_count);
2683 static u64 perf_event_read(struct perf_event *event)
2686 * If event is enabled and currently active on a CPU, update the
2687 * value in the event structure:
2689 if (event->state == PERF_EVENT_STATE_ACTIVE) {
2690 smp_call_function_single(event->oncpu,
2691 __perf_event_read, event, 1);
2692 } else if (event->state == PERF_EVENT_STATE_INACTIVE) {
2693 struct perf_event_context *ctx = event->ctx;
2694 unsigned long flags;
2696 raw_spin_lock_irqsave(&ctx->lock, flags);
2698 * may read while context is not active
2699 * (e.g., thread is blocked), in that case
2700 * we cannot update context time
2702 if (ctx->is_active) {
2703 update_context_time(ctx);
2704 update_cgrp_time_from_event(event);
2706 update_event_times(event);
2707 raw_spin_unlock_irqrestore(&ctx->lock, flags);
2710 return perf_event_count(event);
2714 * Initialize the perf_event context in a task_struct:
2716 static void __perf_event_init_context(struct perf_event_context *ctx)
2718 raw_spin_lock_init(&ctx->lock);
2719 mutex_init(&ctx->mutex);
2720 INIT_LIST_HEAD(&ctx->pinned_groups);
2721 INIT_LIST_HEAD(&ctx->flexible_groups);
2722 INIT_LIST_HEAD(&ctx->event_list);
2723 atomic_set(&ctx->refcount, 1);
2726 static struct perf_event_context *
2727 alloc_perf_context(struct pmu *pmu, struct task_struct *task)
2729 struct perf_event_context *ctx;
2731 ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL);
2735 __perf_event_init_context(ctx);
2738 get_task_struct(task);
2745 static struct task_struct *
2746 find_lively_task_by_vpid(pid_t vpid)
2748 struct task_struct *task;
2755 task = find_task_by_vpid(vpid);
2757 get_task_struct(task);
2761 return ERR_PTR(-ESRCH);
2763 /* Reuse ptrace permission checks for now. */
2765 if (!ptrace_may_access(task, PTRACE_MODE_READ))
2770 put_task_struct(task);
2771 return ERR_PTR(err);
2776 * Returns a matching context with refcount and pincount.
2778 static struct perf_event_context *
2779 find_get_context(struct pmu *pmu, struct task_struct *task, int cpu)
2781 struct perf_event_context *ctx;
2782 struct perf_cpu_context *cpuctx;
2783 unsigned long flags;
2787 /* Must be root to operate on a CPU event: */
2788 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
2789 return ERR_PTR(-EACCES);
2792 * We could be clever and allow to attach a event to an
2793 * offline CPU and activate it when the CPU comes up, but
2796 if (!cpu_online(cpu))
2797 return ERR_PTR(-ENODEV);
2799 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
2808 ctxn = pmu->task_ctx_nr;
2813 ctx = perf_lock_task_context(task, ctxn, &flags);
2817 raw_spin_unlock_irqrestore(&ctx->lock, flags);
2819 ctx = alloc_perf_context(pmu, task);
2825 mutex_lock(&task->perf_event_mutex);
2827 * If it has already passed perf_event_exit_task().
2828 * we must see PF_EXITING, it takes this mutex too.
2830 if (task->flags & PF_EXITING)
2832 else if (task->perf_event_ctxp[ctxn])
2837 rcu_assign_pointer(task->perf_event_ctxp[ctxn], ctx);
2839 mutex_unlock(&task->perf_event_mutex);
2841 if (unlikely(err)) {
2853 return ERR_PTR(err);
2856 static void perf_event_free_filter(struct perf_event *event);
2858 static void free_event_rcu(struct rcu_head *head)
2860 struct perf_event *event;
2862 event = container_of(head, struct perf_event, rcu_head);
2864 put_pid_ns(event->ns);
2865 perf_event_free_filter(event);
2869 static void ring_buffer_put(struct ring_buffer *rb);
2871 static void free_event(struct perf_event *event)
2873 irq_work_sync(&event->pending);
2875 if (!event->parent) {
2876 if (event->attach_state & PERF_ATTACH_TASK)
2877 static_key_slow_dec_deferred(&perf_sched_events);
2878 if (event->attr.mmap || event->attr.mmap_data)
2879 atomic_dec(&nr_mmap_events);
2880 if (event->attr.comm)
2881 atomic_dec(&nr_comm_events);
2882 if (event->attr.task)
2883 atomic_dec(&nr_task_events);
2884 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN)
2885 put_callchain_buffers();
2886 if (is_cgroup_event(event)) {
2887 atomic_dec(&per_cpu(perf_cgroup_events, event->cpu));
2888 static_key_slow_dec_deferred(&perf_sched_events);
2891 if (has_branch_stack(event)) {
2892 static_key_slow_dec_deferred(&perf_sched_events);
2893 /* is system-wide event */
2894 if (!(event->attach_state & PERF_ATTACH_TASK))
2895 atomic_dec(&per_cpu(perf_branch_stack_events,
2901 ring_buffer_put(event->rb);
2905 if (is_cgroup_event(event))
2906 perf_detach_cgroup(event);
2909 event->destroy(event);
2912 put_ctx(event->ctx);
2914 call_rcu(&event->rcu_head, free_event_rcu);
2917 int perf_event_release_kernel(struct perf_event *event)
2919 struct perf_event_context *ctx = event->ctx;
2921 WARN_ON_ONCE(ctx->parent_ctx);
2923 * There are two ways this annotation is useful:
2925 * 1) there is a lock recursion from perf_event_exit_task
2926 * see the comment there.
2928 * 2) there is a lock-inversion with mmap_sem through
2929 * perf_event_read_group(), which takes faults while
2930 * holding ctx->mutex, however this is called after
2931 * the last filedesc died, so there is no possibility
2932 * to trigger the AB-BA case.
2934 mutex_lock_nested(&ctx->mutex, SINGLE_DEPTH_NESTING);
2935 raw_spin_lock_irq(&ctx->lock);
2936 perf_group_detach(event);
2937 raw_spin_unlock_irq(&ctx->lock);
2938 perf_remove_from_context(event);
2939 mutex_unlock(&ctx->mutex);
2945 EXPORT_SYMBOL_GPL(perf_event_release_kernel);
2948 * Called when the last reference to the file is gone.
2950 static void put_event(struct perf_event *event)
2952 struct task_struct *owner;
2954 if (!atomic_long_dec_and_test(&event->refcount))
2958 owner = ACCESS_ONCE(event->owner);
2960 * Matches the smp_wmb() in perf_event_exit_task(). If we observe
2961 * !owner it means the list deletion is complete and we can indeed
2962 * free this event, otherwise we need to serialize on
2963 * owner->perf_event_mutex.
2965 smp_read_barrier_depends();
2968 * Since delayed_put_task_struct() also drops the last
2969 * task reference we can safely take a new reference
2970 * while holding the rcu_read_lock().
2972 get_task_struct(owner);
2977 mutex_lock(&owner->perf_event_mutex);
2979 * We have to re-check the event->owner field, if it is cleared
2980 * we raced with perf_event_exit_task(), acquiring the mutex
2981 * ensured they're done, and we can proceed with freeing the
2985 list_del_init(&event->owner_entry);
2986 mutex_unlock(&owner->perf_event_mutex);
2987 put_task_struct(owner);
2990 perf_event_release_kernel(event);
2993 static int perf_release(struct inode *inode, struct file *file)
2995 put_event(file->private_data);
2999 u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running)
3001 struct perf_event *child;
3007 mutex_lock(&event->child_mutex);
3008 total += perf_event_read(event);
3009 *enabled += event->total_time_enabled +
3010 atomic64_read(&event->child_total_time_enabled);
3011 *running += event->total_time_running +
3012 atomic64_read(&event->child_total_time_running);
3014 list_for_each_entry(child, &event->child_list, child_list) {
3015 total += perf_event_read(child);
3016 *enabled += child->total_time_enabled;
3017 *running += child->total_time_running;
3019 mutex_unlock(&event->child_mutex);
3023 EXPORT_SYMBOL_GPL(perf_event_read_value);
3025 static int perf_event_read_group(struct perf_event *event,
3026 u64 read_format, char __user *buf)
3028 struct perf_event *leader = event->group_leader, *sub;
3029 int n = 0, size = 0, ret = -EFAULT;
3030 struct perf_event_context *ctx = leader->ctx;
3032 u64 count, enabled, running;
3034 mutex_lock(&ctx->mutex);
3035 count = perf_event_read_value(leader, &enabled, &running);
3037 values[n++] = 1 + leader->nr_siblings;
3038 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3039 values[n++] = enabled;
3040 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3041 values[n++] = running;
3042 values[n++] = count;
3043 if (read_format & PERF_FORMAT_ID)
3044 values[n++] = primary_event_id(leader);
3046 size = n * sizeof(u64);
3048 if (copy_to_user(buf, values, size))
3053 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
3056 values[n++] = perf_event_read_value(sub, &enabled, &running);
3057 if (read_format & PERF_FORMAT_ID)
3058 values[n++] = primary_event_id(sub);
3060 size = n * sizeof(u64);
3062 if (copy_to_user(buf + ret, values, size)) {
3070 mutex_unlock(&ctx->mutex);
3075 static int perf_event_read_one(struct perf_event *event,
3076 u64 read_format, char __user *buf)
3078 u64 enabled, running;
3082 values[n++] = perf_event_read_value(event, &enabled, &running);
3083 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3084 values[n++] = enabled;
3085 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3086 values[n++] = running;
3087 if (read_format & PERF_FORMAT_ID)
3088 values[n++] = primary_event_id(event);
3090 if (copy_to_user(buf, values, n * sizeof(u64)))
3093 return n * sizeof(u64);
3097 * Read the performance event - simple non blocking version for now
3100 perf_read_hw(struct perf_event *event, char __user *buf, size_t count)
3102 u64 read_format = event->attr.read_format;
3106 * Return end-of-file for a read on a event that is in
3107 * error state (i.e. because it was pinned but it couldn't be
3108 * scheduled on to the CPU at some point).
3110 if (event->state == PERF_EVENT_STATE_ERROR)
3113 if (count < event->read_size)
3116 WARN_ON_ONCE(event->ctx->parent_ctx);
3117 if (read_format & PERF_FORMAT_GROUP)
3118 ret = perf_event_read_group(event, read_format, buf);
3120 ret = perf_event_read_one(event, read_format, buf);
3126 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
3128 struct perf_event *event = file->private_data;
3130 return perf_read_hw(event, buf, count);
3133 static unsigned int perf_poll(struct file *file, poll_table *wait)
3135 struct perf_event *event = file->private_data;
3136 struct ring_buffer *rb;
3137 unsigned int events = POLL_HUP;
3140 * Race between perf_event_set_output() and perf_poll(): perf_poll()
3141 * grabs the rb reference but perf_event_set_output() overrides it.
3142 * Here is the timeline for two threads T1, T2:
3143 * t0: T1, rb = rcu_dereference(event->rb)
3144 * t1: T2, old_rb = event->rb
3145 * t2: T2, event->rb = new rb
3146 * t3: T2, ring_buffer_detach(old_rb)
3147 * t4: T1, ring_buffer_attach(rb1)
3148 * t5: T1, poll_wait(event->waitq)
3150 * To avoid this problem, we grab mmap_mutex in perf_poll()
3151 * thereby ensuring that the assignment of the new ring buffer
3152 * and the detachment of the old buffer appear atomic to perf_poll()
3154 mutex_lock(&event->mmap_mutex);
3157 rb = rcu_dereference(event->rb);
3159 ring_buffer_attach(event, rb);
3160 events = atomic_xchg(&rb->poll, 0);
3164 mutex_unlock(&event->mmap_mutex);
3166 poll_wait(file, &event->waitq, wait);
3171 static void perf_event_reset(struct perf_event *event)
3173 (void)perf_event_read(event);
3174 local64_set(&event->count, 0);
3175 perf_event_update_userpage(event);
3179 * Holding the top-level event's child_mutex means that any
3180 * descendant process that has inherited this event will block
3181 * in sync_child_event if it goes to exit, thus satisfying the
3182 * task existence requirements of perf_event_enable/disable.
3184 static void perf_event_for_each_child(struct perf_event *event,
3185 void (*func)(struct perf_event *))
3187 struct perf_event *child;
3189 WARN_ON_ONCE(event->ctx->parent_ctx);
3190 mutex_lock(&event->child_mutex);
3192 list_for_each_entry(child, &event->child_list, child_list)
3194 mutex_unlock(&event->child_mutex);
3197 static void perf_event_for_each(struct perf_event *event,
3198 void (*func)(struct perf_event *))
3200 struct perf_event_context *ctx = event->ctx;
3201 struct perf_event *sibling;
3203 WARN_ON_ONCE(ctx->parent_ctx);
3204 mutex_lock(&ctx->mutex);
3205 event = event->group_leader;
3207 perf_event_for_each_child(event, func);
3208 list_for_each_entry(sibling, &event->sibling_list, group_entry)
3209 perf_event_for_each_child(sibling, func);
3210 mutex_unlock(&ctx->mutex);
3213 static int perf_event_period(struct perf_event *event, u64 __user *arg)
3215 struct perf_event_context *ctx = event->ctx;
3219 if (!is_sampling_event(event))
3222 if (copy_from_user(&value, arg, sizeof(value)))
3228 raw_spin_lock_irq(&ctx->lock);
3229 if (event->attr.freq) {
3230 if (value > sysctl_perf_event_sample_rate) {
3235 event->attr.sample_freq = value;
3237 event->attr.sample_period = value;
3238 event->hw.sample_period = value;
3241 raw_spin_unlock_irq(&ctx->lock);
3246 static const struct file_operations perf_fops;
3248 static inline int perf_fget_light(int fd, struct fd *p)
3250 struct fd f = fdget(fd);
3254 if (f.file->f_op != &perf_fops) {
3262 static int perf_event_set_output(struct perf_event *event,
3263 struct perf_event *output_event);
3264 static int perf_event_set_filter(struct perf_event *event, void __user *arg);
3266 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
3268 struct perf_event *event = file->private_data;
3269 void (*func)(struct perf_event *);
3273 case PERF_EVENT_IOC_ENABLE:
3274 func = perf_event_enable;
3276 case PERF_EVENT_IOC_DISABLE:
3277 func = perf_event_disable;
3279 case PERF_EVENT_IOC_RESET:
3280 func = perf_event_reset;
3283 case PERF_EVENT_IOC_REFRESH:
3284 return perf_event_refresh(event, arg);
3286 case PERF_EVENT_IOC_PERIOD:
3287 return perf_event_period(event, (u64 __user *)arg);
3289 case PERF_EVENT_IOC_SET_OUTPUT:
3293 struct perf_event *output_event;
3295 ret = perf_fget_light(arg, &output);
3298 output_event = output.file->private_data;
3299 ret = perf_event_set_output(event, output_event);
3302 ret = perf_event_set_output(event, NULL);
3307 case PERF_EVENT_IOC_SET_FILTER:
3308 return perf_event_set_filter(event, (void __user *)arg);
3314 if (flags & PERF_IOC_FLAG_GROUP)
3315 perf_event_for_each(event, func);
3317 perf_event_for_each_child(event, func);
3322 int perf_event_task_enable(void)
3324 struct perf_event *event;
3326 mutex_lock(¤t->perf_event_mutex);
3327 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
3328 perf_event_for_each_child(event, perf_event_enable);
3329 mutex_unlock(¤t->perf_event_mutex);
3334 int perf_event_task_disable(void)
3336 struct perf_event *event;
3338 mutex_lock(¤t->perf_event_mutex);
3339 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
3340 perf_event_for_each_child(event, perf_event_disable);
3341 mutex_unlock(¤t->perf_event_mutex);
3346 static int perf_event_index(struct perf_event *event)
3348 if (event->hw.state & PERF_HES_STOPPED)
3351 if (event->state != PERF_EVENT_STATE_ACTIVE)
3354 return event->pmu->event_idx(event);
3357 static void calc_timer_values(struct perf_event *event,
3364 *now = perf_clock();
3365 ctx_time = event->shadow_ctx_time + *now;
3366 *enabled = ctx_time - event->tstamp_enabled;
3367 *running = ctx_time - event->tstamp_running;
3370 void __weak arch_perf_update_userpage(struct perf_event_mmap_page *userpg, u64 now)
3375 * Callers need to ensure there can be no nesting of this function, otherwise
3376 * the seqlock logic goes bad. We can not serialize this because the arch
3377 * code calls this from NMI context.
3379 void perf_event_update_userpage(struct perf_event *event)
3381 struct perf_event_mmap_page *userpg;
3382 struct ring_buffer *rb;
3383 u64 enabled, running, now;
3387 * compute total_time_enabled, total_time_running
3388 * based on snapshot values taken when the event
3389 * was last scheduled in.
3391 * we cannot simply called update_context_time()
3392 * because of locking issue as we can be called in
3395 calc_timer_values(event, &now, &enabled, &running);
3396 rb = rcu_dereference(event->rb);
3400 userpg = rb->user_page;
3403 * Disable preemption so as to not let the corresponding user-space
3404 * spin too long if we get preempted.
3409 userpg->index = perf_event_index(event);
3410 userpg->offset = perf_event_count(event);
3412 userpg->offset -= local64_read(&event->hw.prev_count);
3414 userpg->time_enabled = enabled +
3415 atomic64_read(&event->child_total_time_enabled);
3417 userpg->time_running = running +
3418 atomic64_read(&event->child_total_time_running);
3420 arch_perf_update_userpage(userpg, now);
3429 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
3431 struct perf_event *event = vma->vm_file->private_data;
3432 struct ring_buffer *rb;
3433 int ret = VM_FAULT_SIGBUS;
3435 if (vmf->flags & FAULT_FLAG_MKWRITE) {
3436 if (vmf->pgoff == 0)
3442 rb = rcu_dereference(event->rb);
3446 if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE))
3449 vmf->page = perf_mmap_to_page(rb, vmf->pgoff);
3453 get_page(vmf->page);
3454 vmf->page->mapping = vma->vm_file->f_mapping;
3455 vmf->page->index = vmf->pgoff;
3464 static void ring_buffer_attach(struct perf_event *event,
3465 struct ring_buffer *rb)
3467 unsigned long flags;
3469 if (!list_empty(&event->rb_entry))
3472 spin_lock_irqsave(&rb->event_lock, flags);
3473 if (!list_empty(&event->rb_entry))
3476 list_add(&event->rb_entry, &rb->event_list);
3478 spin_unlock_irqrestore(&rb->event_lock, flags);
3481 static void ring_buffer_detach(struct perf_event *event,
3482 struct ring_buffer *rb)
3484 unsigned long flags;
3486 if (list_empty(&event->rb_entry))
3489 spin_lock_irqsave(&rb->event_lock, flags);
3490 list_del_init(&event->rb_entry);
3491 wake_up_all(&event->waitq);
3492 spin_unlock_irqrestore(&rb->event_lock, flags);
3495 static void ring_buffer_wakeup(struct perf_event *event)
3497 struct ring_buffer *rb;
3500 rb = rcu_dereference(event->rb);
3504 list_for_each_entry_rcu(event, &rb->event_list, rb_entry)
3505 wake_up_all(&event->waitq);
3511 static void rb_free_rcu(struct rcu_head *rcu_head)
3513 struct ring_buffer *rb;
3515 rb = container_of(rcu_head, struct ring_buffer, rcu_head);
3519 static struct ring_buffer *ring_buffer_get(struct perf_event *event)
3521 struct ring_buffer *rb;
3524 rb = rcu_dereference(event->rb);
3526 if (!atomic_inc_not_zero(&rb->refcount))
3534 static void ring_buffer_put(struct ring_buffer *rb)
3536 struct perf_event *event, *n;
3537 unsigned long flags;
3539 if (!atomic_dec_and_test(&rb->refcount))
3542 spin_lock_irqsave(&rb->event_lock, flags);
3543 list_for_each_entry_safe(event, n, &rb->event_list, rb_entry) {
3544 list_del_init(&event->rb_entry);
3545 wake_up_all(&event->waitq);
3547 spin_unlock_irqrestore(&rb->event_lock, flags);
3549 call_rcu(&rb->rcu_head, rb_free_rcu);
3552 static void perf_mmap_open(struct vm_area_struct *vma)
3554 struct perf_event *event = vma->vm_file->private_data;
3556 atomic_inc(&event->mmap_count);
3559 static void perf_mmap_close(struct vm_area_struct *vma)
3561 struct perf_event *event = vma->vm_file->private_data;
3563 if (atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex)) {
3564 unsigned long size = perf_data_size(event->rb);
3565 struct user_struct *user = event->mmap_user;
3566 struct ring_buffer *rb = event->rb;
3568 atomic_long_sub((size >> PAGE_SHIFT) + 1, &user->locked_vm);
3569 vma->vm_mm->pinned_vm -= event->mmap_locked;
3570 rcu_assign_pointer(event->rb, NULL);
3571 ring_buffer_detach(event, rb);
3572 mutex_unlock(&event->mmap_mutex);
3574 ring_buffer_put(rb);
3579 static const struct vm_operations_struct perf_mmap_vmops = {
3580 .open = perf_mmap_open,
3581 .close = perf_mmap_close,
3582 .fault = perf_mmap_fault,
3583 .page_mkwrite = perf_mmap_fault,
3586 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
3588 struct perf_event *event = file->private_data;
3589 unsigned long user_locked, user_lock_limit;
3590 struct user_struct *user = current_user();
3591 unsigned long locked, lock_limit;
3592 struct ring_buffer *rb;
3593 unsigned long vma_size;
3594 unsigned long nr_pages;
3595 long user_extra, extra;
3596 int ret = 0, flags = 0;
3599 * Don't allow mmap() of inherited per-task counters. This would
3600 * create a performance issue due to all children writing to the
3603 if (event->cpu == -1 && event->attr.inherit)
3606 if (!(vma->vm_flags & VM_SHARED))
3609 vma_size = vma->vm_end - vma->vm_start;
3610 nr_pages = (vma_size / PAGE_SIZE) - 1;
3613 * If we have rb pages ensure they're a power-of-two number, so we
3614 * can do bitmasks instead of modulo.
3616 if (nr_pages != 0 && !is_power_of_2(nr_pages))
3619 if (vma_size != PAGE_SIZE * (1 + nr_pages))
3622 if (vma->vm_pgoff != 0)
3625 WARN_ON_ONCE(event->ctx->parent_ctx);
3626 mutex_lock(&event->mmap_mutex);
3628 if (event->rb->nr_pages == nr_pages)
3629 atomic_inc(&event->rb->refcount);
3635 user_extra = nr_pages + 1;
3636 user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10);
3639 * Increase the limit linearly with more CPUs:
3641 user_lock_limit *= num_online_cpus();
3643 user_locked = atomic_long_read(&user->locked_vm) + user_extra;
3646 if (user_locked > user_lock_limit)
3647 extra = user_locked - user_lock_limit;
3649 lock_limit = rlimit(RLIMIT_MEMLOCK);
3650 lock_limit >>= PAGE_SHIFT;
3651 locked = vma->vm_mm->pinned_vm + extra;
3653 if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() &&
3654 !capable(CAP_IPC_LOCK)) {
3661 if (vma->vm_flags & VM_WRITE)
3662 flags |= RING_BUFFER_WRITABLE;
3664 rb = rb_alloc(nr_pages,
3665 event->attr.watermark ? event->attr.wakeup_watermark : 0,
3672 rcu_assign_pointer(event->rb, rb);
3674 atomic_long_add(user_extra, &user->locked_vm);
3675 event->mmap_locked = extra;
3676 event->mmap_user = get_current_user();
3677 vma->vm_mm->pinned_vm += event->mmap_locked;
3679 perf_event_update_userpage(event);
3683 atomic_inc(&event->mmap_count);
3684 mutex_unlock(&event->mmap_mutex);
3686 vma->vm_flags |= VM_DONTEXPAND | VM_DONTDUMP;
3687 vma->vm_ops = &perf_mmap_vmops;
3692 static int perf_fasync(int fd, struct file *filp, int on)
3694 struct inode *inode = file_inode(filp);
3695 struct perf_event *event = filp->private_data;
3698 mutex_lock(&inode->i_mutex);
3699 retval = fasync_helper(fd, filp, on, &event->fasync);
3700 mutex_unlock(&inode->i_mutex);
3708 static const struct file_operations perf_fops = {
3709 .llseek = no_llseek,
3710 .release = perf_release,
3713 .unlocked_ioctl = perf_ioctl,
3714 .compat_ioctl = perf_ioctl,
3716 .fasync = perf_fasync,
3722 * If there's data, ensure we set the poll() state and publish everything
3723 * to user-space before waking everybody up.
3726 void perf_event_wakeup(struct perf_event *event)
3728 ring_buffer_wakeup(event);
3730 if (event->pending_kill) {
3731 kill_fasync(&event->fasync, SIGIO, event->pending_kill);
3732 event->pending_kill = 0;
3736 static void perf_pending_event(struct irq_work *entry)
3738 struct perf_event *event = container_of(entry,
3739 struct perf_event, pending);
3741 if (event->pending_disable) {
3742 event->pending_disable = 0;
3743 __perf_event_disable(event);
3746 if (event->pending_wakeup) {
3747 event->pending_wakeup = 0;
3748 perf_event_wakeup(event);
3753 * We assume there is only KVM supporting the callbacks.
3754 * Later on, we might change it to a list if there is
3755 * another virtualization implementation supporting the callbacks.
3757 struct perf_guest_info_callbacks *perf_guest_cbs;
3759 int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
3761 perf_guest_cbs = cbs;
3764 EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks);
3766 int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
3768 perf_guest_cbs = NULL;
3771 EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks);
3774 perf_output_sample_regs(struct perf_output_handle *handle,
3775 struct pt_regs *regs, u64 mask)
3779 for_each_set_bit(bit, (const unsigned long *) &mask,
3780 sizeof(mask) * BITS_PER_BYTE) {
3783 val = perf_reg_value(regs, bit);
3784 perf_output_put(handle, val);
3788 static void perf_sample_regs_user(struct perf_regs_user *regs_user,
3789 struct pt_regs *regs)
3791 if (!user_mode(regs)) {
3793 regs = task_pt_regs(current);
3799 regs_user->regs = regs;
3800 regs_user->abi = perf_reg_abi(current);
3805 * Get remaining task size from user stack pointer.
3807 * It'd be better to take stack vma map and limit this more
3808 * precisly, but there's no way to get it safely under interrupt,
3809 * so using TASK_SIZE as limit.
3811 static u64 perf_ustack_task_size(struct pt_regs *regs)
3813 unsigned long addr = perf_user_stack_pointer(regs);
3815 if (!addr || addr >= TASK_SIZE)
3818 return TASK_SIZE - addr;
3822 perf_sample_ustack_size(u16 stack_size, u16 header_size,
3823 struct pt_regs *regs)
3827 /* No regs, no stack pointer, no dump. */
3832 * Check if we fit in with the requested stack size into the:
3834 * If we don't, we limit the size to the TASK_SIZE.
3836 * - remaining sample size
3837 * If we don't, we customize the stack size to
3838 * fit in to the remaining sample size.
3841 task_size = min((u64) USHRT_MAX, perf_ustack_task_size(regs));
3842 stack_size = min(stack_size, (u16) task_size);
3844 /* Current header size plus static size and dynamic size. */
3845 header_size += 2 * sizeof(u64);
3847 /* Do we fit in with the current stack dump size? */
3848 if ((u16) (header_size + stack_size) < header_size) {
3850 * If we overflow the maximum size for the sample,
3851 * we customize the stack dump size to fit in.
3853 stack_size = USHRT_MAX - header_size - sizeof(u64);
3854 stack_size = round_up(stack_size, sizeof(u64));
3861 perf_output_sample_ustack(struct perf_output_handle *handle, u64 dump_size,
3862 struct pt_regs *regs)
3864 /* Case of a kernel thread, nothing to dump */
3867 perf_output_put(handle, size);
3876 * - the size requested by user or the best one we can fit
3877 * in to the sample max size
3879 * - user stack dump data
3881 * - the actual dumped size
3885 perf_output_put(handle, dump_size);
3888 sp = perf_user_stack_pointer(regs);
3889 rem = __output_copy_user(handle, (void *) sp, dump_size);
3890 dyn_size = dump_size - rem;
3892 perf_output_skip(handle, rem);
3895 perf_output_put(handle, dyn_size);
3899 static void __perf_event_header__init_id(struct perf_event_header *header,
3900 struct perf_sample_data *data,
3901 struct perf_event *event)
3903 u64 sample_type = event->attr.sample_type;
3905 data->type = sample_type;
3906 header->size += event->id_header_size;
3908 if (sample_type & PERF_SAMPLE_TID) {
3909 /* namespace issues */
3910 data->tid_entry.pid = perf_event_pid(event, current);
3911 data->tid_entry.tid = perf_event_tid(event, current);
3914 if (sample_type & PERF_SAMPLE_TIME)
3915 data->time = perf_clock();
3917 if (sample_type & PERF_SAMPLE_ID)
3918 data->id = primary_event_id(event);
3920 if (sample_type & PERF_SAMPLE_STREAM_ID)
3921 data->stream_id = event->id;
3923 if (sample_type & PERF_SAMPLE_CPU) {
3924 data->cpu_entry.cpu = raw_smp_processor_id();
3925 data->cpu_entry.reserved = 0;
3929 void perf_event_header__init_id(struct perf_event_header *header,
3930 struct perf_sample_data *data,
3931 struct perf_event *event)
3933 if (event->attr.sample_id_all)
3934 __perf_event_header__init_id(header, data, event);
3937 static void __perf_event__output_id_sample(struct perf_output_handle *handle,
3938 struct perf_sample_data *data)
3940 u64 sample_type = data->type;
3942 if (sample_type & PERF_SAMPLE_TID)
3943 perf_output_put(handle, data->tid_entry);
3945 if (sample_type & PERF_SAMPLE_TIME)
3946 perf_output_put(handle, data->time);
3948 if (sample_type & PERF_SAMPLE_ID)
3949 perf_output_put(handle, data->id);
3951 if (sample_type & PERF_SAMPLE_STREAM_ID)
3952 perf_output_put(handle, data->stream_id);
3954 if (sample_type & PERF_SAMPLE_CPU)
3955 perf_output_put(handle, data->cpu_entry);
3958 void perf_event__output_id_sample(struct perf_event *event,
3959 struct perf_output_handle *handle,
3960 struct perf_sample_data *sample)
3962 if (event->attr.sample_id_all)
3963 __perf_event__output_id_sample(handle, sample);
3966 static void perf_output_read_one(struct perf_output_handle *handle,
3967 struct perf_event *event,
3968 u64 enabled, u64 running)
3970 u64 read_format = event->attr.read_format;
3974 values[n++] = perf_event_count(event);
3975 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
3976 values[n++] = enabled +
3977 atomic64_read(&event->child_total_time_enabled);
3979 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
3980 values[n++] = running +
3981 atomic64_read(&event->child_total_time_running);
3983 if (read_format & PERF_FORMAT_ID)
3984 values[n++] = primary_event_id(event);
3986 __output_copy(handle, values, n * sizeof(u64));
3990 * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
3992 static void perf_output_read_group(struct perf_output_handle *handle,
3993 struct perf_event *event,
3994 u64 enabled, u64 running)
3996 struct perf_event *leader = event->group_leader, *sub;
3997 u64 read_format = event->attr.read_format;
4001 values[n++] = 1 + leader->nr_siblings;
4003 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
4004 values[n++] = enabled;
4006 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
4007 values[n++] = running;
4009 if (leader != event)
4010 leader->pmu->read(leader);
4012 values[n++] = perf_event_count(leader);
4013 if (read_format & PERF_FORMAT_ID)
4014 values[n++] = primary_event_id(leader);
4016 __output_copy(handle, values, n * sizeof(u64));
4018 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
4022 sub->pmu->read(sub);
4024 values[n++] = perf_event_count(sub);
4025 if (read_format & PERF_FORMAT_ID)
4026 values[n++] = primary_event_id(sub);
4028 __output_copy(handle, values, n * sizeof(u64));
4032 #define PERF_FORMAT_TOTAL_TIMES (PERF_FORMAT_TOTAL_TIME_ENABLED|\
4033 PERF_FORMAT_TOTAL_TIME_RUNNING)
4035 static void perf_output_read(struct perf_output_handle *handle,
4036 struct perf_event *event)
4038 u64 enabled = 0, running = 0, now;
4039 u64 read_format = event->attr.read_format;
4042 * compute total_time_enabled, total_time_running
4043 * based on snapshot values taken when the event
4044 * was last scheduled in.
4046 * we cannot simply called update_context_time()
4047 * because of locking issue as we are called in
4050 if (read_format & PERF_FORMAT_TOTAL_TIMES)
4051 calc_timer_values(event, &now, &enabled, &running);
4053 if (event->attr.read_format & PERF_FORMAT_GROUP)
4054 perf_output_read_group(handle, event, enabled, running);
4056 perf_output_read_one(handle, event, enabled, running);
4059 void perf_output_sample(struct perf_output_handle *handle,
4060 struct perf_event_header *header,
4061 struct perf_sample_data *data,
4062 struct perf_event *event)
4064 u64 sample_type = data->type;
4066 perf_output_put(handle, *header);
4068 if (sample_type & PERF_SAMPLE_IP)
4069 perf_output_put(handle, data->ip);
4071 if (sample_type & PERF_SAMPLE_TID)
4072 perf_output_put(handle, data->tid_entry);
4074 if (sample_type & PERF_SAMPLE_TIME)
4075 perf_output_put(handle, data->time);
4077 if (sample_type & PERF_SAMPLE_ADDR)
4078 perf_output_put(handle, data->addr);
4080 if (sample_type & PERF_SAMPLE_ID)
4081 perf_output_put(handle, data->id);
4083 if (sample_type & PERF_SAMPLE_STREAM_ID)
4084 perf_output_put(handle, data->stream_id);
4086 if (sample_type & PERF_SAMPLE_CPU)
4087 perf_output_put(handle, data->cpu_entry);
4089 if (sample_type & PERF_SAMPLE_PERIOD)
4090 perf_output_put(handle, data->period);
4092 if (sample_type & PERF_SAMPLE_READ)
4093 perf_output_read(handle, event);
4095 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
4096 if (data->callchain) {
4099 if (data->callchain)
4100 size += data->callchain->nr;
4102 size *= sizeof(u64);
4104 __output_copy(handle, data->callchain, size);
4107 perf_output_put(handle, nr);
4111 if (sample_type & PERF_SAMPLE_RAW) {
4113 perf_output_put(handle, data->raw->size);
4114 __output_copy(handle, data->raw->data,
4121 .size = sizeof(u32),
4124 perf_output_put(handle, raw);
4128 if (!event->attr.watermark) {
4129 int wakeup_events = event->attr.wakeup_events;
4131 if (wakeup_events) {
4132 struct ring_buffer *rb = handle->rb;
4133 int events = local_inc_return(&rb->events);
4135 if (events >= wakeup_events) {
4136 local_sub(wakeup_events, &rb->events);
4137 local_inc(&rb->wakeup);
4142 if (sample_type & PERF_SAMPLE_BRANCH_STACK) {
4143 if (data->br_stack) {
4146 size = data->br_stack->nr
4147 * sizeof(struct perf_branch_entry);
4149 perf_output_put(handle, data->br_stack->nr);
4150 perf_output_copy(handle, data->br_stack->entries, size);
4153 * we always store at least the value of nr
4156 perf_output_put(handle, nr);
4160 if (sample_type & PERF_SAMPLE_REGS_USER) {
4161 u64 abi = data->regs_user.abi;
4164 * If there are no regs to dump, notice it through
4165 * first u64 being zero (PERF_SAMPLE_REGS_ABI_NONE).
4167 perf_output_put(handle, abi);
4170 u64 mask = event->attr.sample_regs_user;
4171 perf_output_sample_regs(handle,
4172 data->regs_user.regs,
4177 if (sample_type & PERF_SAMPLE_STACK_USER)
4178 perf_output_sample_ustack(handle,
4179 data->stack_user_size,
4180 data->regs_user.regs);
4183 void perf_prepare_sample(struct perf_event_header *header,
4184 struct perf_sample_data *data,
4185 struct perf_event *event,
4186 struct pt_regs *regs)
4188 u64 sample_type = event->attr.sample_type;
4190 header->type = PERF_RECORD_SAMPLE;
4191 header->size = sizeof(*header) + event->header_size;
4194 header->misc |= perf_misc_flags(regs);
4196 __perf_event_header__init_id(header, data, event);
4198 if (sample_type & PERF_SAMPLE_IP)
4199 data->ip = perf_instruction_pointer(regs);
4201 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
4204 data->callchain = perf_callchain(event, regs);
4206 if (data->callchain)
4207 size += data->callchain->nr;
4209 header->size += size * sizeof(u64);
4212 if (sample_type & PERF_SAMPLE_RAW) {
4213 int size = sizeof(u32);
4216 size += data->raw->size;
4218 size += sizeof(u32);
4220 WARN_ON_ONCE(size & (sizeof(u64)-1));
4221 header->size += size;
4224 if (sample_type & PERF_SAMPLE_BRANCH_STACK) {
4225 int size = sizeof(u64); /* nr */
4226 if (data->br_stack) {
4227 size += data->br_stack->nr
4228 * sizeof(struct perf_branch_entry);
4230 header->size += size;
4233 if (sample_type & PERF_SAMPLE_REGS_USER) {
4234 /* regs dump ABI info */
4235 int size = sizeof(u64);
4237 perf_sample_regs_user(&data->regs_user, regs);
4239 if (data->regs_user.regs) {
4240 u64 mask = event->attr.sample_regs_user;
4241 size += hweight64(mask) * sizeof(u64);
4244 header->size += size;
4247 if (sample_type & PERF_SAMPLE_STACK_USER) {
4249 * Either we need PERF_SAMPLE_STACK_USER bit to be allways
4250 * processed as the last one or have additional check added
4251 * in case new sample type is added, because we could eat
4252 * up the rest of the sample size.
4254 struct perf_regs_user *uregs = &data->regs_user;
4255 u16 stack_size = event->attr.sample_stack_user;
4256 u16 size = sizeof(u64);
4259 perf_sample_regs_user(uregs, regs);
4261 stack_size = perf_sample_ustack_size(stack_size, header->size,
4265 * If there is something to dump, add space for the dump
4266 * itself and for the field that tells the dynamic size,
4267 * which is how many have been actually dumped.
4270 size += sizeof(u64) + stack_size;
4272 data->stack_user_size = stack_size;
4273 header->size += size;
4277 static void perf_event_output(struct perf_event *event,
4278 struct perf_sample_data *data,
4279 struct pt_regs *regs)
4281 struct perf_output_handle handle;
4282 struct perf_event_header header;
4284 /* protect the callchain buffers */
4287 perf_prepare_sample(&header, data, event, regs);
4289 if (perf_output_begin(&handle, event, header.size))
4292 perf_output_sample(&handle, &header, data, event);
4294 perf_output_end(&handle);
4304 struct perf_read_event {
4305 struct perf_event_header header;
4312 perf_event_read_event(struct perf_event *event,
4313 struct task_struct *task)
4315 struct perf_output_handle handle;
4316 struct perf_sample_data sample;
4317 struct perf_read_event read_event = {
4319 .type = PERF_RECORD_READ,
4321 .size = sizeof(read_event) + event->read_size,
4323 .pid = perf_event_pid(event, task),
4324 .tid = perf_event_tid(event, task),
4328 perf_event_header__init_id(&read_event.header, &sample, event);
4329 ret = perf_output_begin(&handle, event, read_event.header.size);
4333 perf_output_put(&handle, read_event);
4334 perf_output_read(&handle, event);
4335 perf_event__output_id_sample(event, &handle, &sample);
4337 perf_output_end(&handle);
4341 * task tracking -- fork/exit
4343 * enabled by: attr.comm | attr.mmap | attr.mmap_data | attr.task
4346 struct perf_task_event {
4347 struct task_struct *task;
4348 struct perf_event_context *task_ctx;
4351 struct perf_event_header header;
4361 static void perf_event_task_output(struct perf_event *event,
4362 struct perf_task_event *task_event)
4364 struct perf_output_handle handle;
4365 struct perf_sample_data sample;
4366 struct task_struct *task = task_event->task;
4367 int ret, size = task_event->event_id.header.size;
4369 perf_event_header__init_id(&task_event->event_id.header, &sample, event);
4371 ret = perf_output_begin(&handle, event,
4372 task_event->event_id.header.size);
4376 task_event->event_id.pid = perf_event_pid(event, task);
4377 task_event->event_id.ppid = perf_event_pid(event, current);
4379 task_event->event_id.tid = perf_event_tid(event, task);
4380 task_event->event_id.ptid = perf_event_tid(event, current);
4382 perf_output_put(&handle, task_event->event_id);
4384 perf_event__output_id_sample(event, &handle, &sample);
4386 perf_output_end(&handle);
4388 task_event->event_id.header.size = size;
4391 static int perf_event_task_match(struct perf_event *event)
4393 if (event->state < PERF_EVENT_STATE_INACTIVE)
4396 if (!event_filter_match(event))
4399 if (event->attr.comm || event->attr.mmap ||
4400 event->attr.mmap_data || event->attr.task)
4406 static void perf_event_task_ctx(struct perf_event_context *ctx,
4407 struct perf_task_event *task_event)
4409 struct perf_event *event;
4411 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
4412 if (perf_event_task_match(event))
4413 perf_event_task_output(event, task_event);
4417 static void perf_event_task_event(struct perf_task_event *task_event)
4419 struct perf_cpu_context *cpuctx;
4420 struct perf_event_context *ctx;
4425 list_for_each_entry_rcu(pmu, &pmus, entry) {
4426 cpuctx = get_cpu_ptr(pmu->pmu_cpu_context);
4427 if (cpuctx->unique_pmu != pmu)
4429 perf_event_task_ctx(&cpuctx->ctx, task_event);
4431 ctx = task_event->task_ctx;
4433 ctxn = pmu->task_ctx_nr;
4436 ctx = rcu_dereference(current->perf_event_ctxp[ctxn]);
4438 perf_event_task_ctx(ctx, task_event);
4441 put_cpu_ptr(pmu->pmu_cpu_context);
4443 if (task_event->task_ctx)
4444 perf_event_task_ctx(task_event->task_ctx, task_event);
4449 static void perf_event_task(struct task_struct *task,
4450 struct perf_event_context *task_ctx,
4453 struct perf_task_event task_event;
4455 if (!atomic_read(&nr_comm_events) &&
4456 !atomic_read(&nr_mmap_events) &&
4457 !atomic_read(&nr_task_events))
4460 task_event = (struct perf_task_event){
4462 .task_ctx = task_ctx,
4465 .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT,
4467 .size = sizeof(task_event.event_id),
4473 .time = perf_clock(),
4477 perf_event_task_event(&task_event);
4480 void perf_event_fork(struct task_struct *task)
4482 perf_event_task(task, NULL, 1);
4489 struct perf_comm_event {
4490 struct task_struct *task;
4495 struct perf_event_header header;
4502 static void perf_event_comm_output(struct perf_event *event,
4503 struct perf_comm_event *comm_event)
4505 struct perf_output_handle handle;
4506 struct perf_sample_data sample;
4507 int size = comm_event->event_id.header.size;
4510 perf_event_header__init_id(&comm_event->event_id.header, &sample, event);
4511 ret = perf_output_begin(&handle, event,
4512 comm_event->event_id.header.size);
4517 comm_event->event_id.pid = perf_event_pid(event, comm_event->task);
4518 comm_event->event_id.tid = perf_event_tid(event, comm_event->task);
4520 perf_output_put(&handle, comm_event->event_id);
4521 __output_copy(&handle, comm_event->comm,
4522 comm_event->comm_size);
4524 perf_event__output_id_sample(event, &handle, &sample);
4526 perf_output_end(&handle);
4528 comm_event->event_id.header.size = size;
4531 static int perf_event_comm_match(struct perf_event *event)
4533 if (event->state < PERF_EVENT_STATE_INACTIVE)
4536 if (!event_filter_match(event))
4539 if (event->attr.comm)
4545 static void perf_event_comm_ctx(struct perf_event_context *ctx,
4546 struct perf_comm_event *comm_event)
4548 struct perf_event *event;
4550 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
4551 if (perf_event_comm_match(event))
4552 perf_event_comm_output(event, comm_event);
4556 static void perf_event_comm_event(struct perf_comm_event *comm_event)
4558 struct perf_cpu_context *cpuctx;
4559 struct perf_event_context *ctx;
4560 char comm[TASK_COMM_LEN];
4565 memset(comm, 0, sizeof(comm));
4566 strlcpy(comm, comm_event->task->comm, sizeof(comm));
4567 size = ALIGN(strlen(comm)+1, sizeof(u64));
4569 comm_event->comm = comm;
4570 comm_event->comm_size = size;
4572 comm_event->event_id.header.size = sizeof(comm_event->event_id) + size;
4574 list_for_each_entry_rcu(pmu, &pmus, entry) {
4575 cpuctx = get_cpu_ptr(pmu->pmu_cpu_context);
4576 if (cpuctx->unique_pmu != pmu)
4578 perf_event_comm_ctx(&cpuctx->ctx, comm_event);
4580 ctxn = pmu->task_ctx_nr;
4584 ctx = rcu_dereference(current->perf_event_ctxp[ctxn]);
4586 perf_event_comm_ctx(ctx, comm_event);
4588 put_cpu_ptr(pmu->pmu_cpu_context);
4593 void perf_event_comm(struct task_struct *task)
4595 struct perf_comm_event comm_event;
4596 struct perf_event_context *ctx;
4599 for_each_task_context_nr(ctxn) {
4600 ctx = task->perf_event_ctxp[ctxn];
4604 perf_event_enable_on_exec(ctx);
4607 if (!atomic_read(&nr_comm_events))
4610 comm_event = (struct perf_comm_event){
4616 .type = PERF_RECORD_COMM,
4625 perf_event_comm_event(&comm_event);
4632 struct perf_mmap_event {
4633 struct vm_area_struct *vma;
4635 const char *file_name;
4639 struct perf_event_header header;
4649 static void perf_event_mmap_output(struct perf_event *event,
4650 struct perf_mmap_event *mmap_event)
4652 struct perf_output_handle handle;
4653 struct perf_sample_data sample;
4654 int size = mmap_event->event_id.header.size;
4657 perf_event_header__init_id(&mmap_event->event_id.header, &sample, event);
4658 ret = perf_output_begin(&handle, event,
4659 mmap_event->event_id.header.size);
4663 mmap_event->event_id.pid = perf_event_pid(event, current);
4664 mmap_event->event_id.tid = perf_event_tid(event, current);
4666 perf_output_put(&handle, mmap_event->event_id);
4667 __output_copy(&handle, mmap_event->file_name,
4668 mmap_event->file_size);
4670 perf_event__output_id_sample(event, &handle, &sample);
4672 perf_output_end(&handle);
4674 mmap_event->event_id.header.size = size;
4677 static int perf_event_mmap_match(struct perf_event *event,
4678 struct perf_mmap_event *mmap_event,
4681 if (event->state < PERF_EVENT_STATE_INACTIVE)
4684 if (!event_filter_match(event))
4687 if ((!executable && event->attr.mmap_data) ||
4688 (executable && event->attr.mmap))
4694 static void perf_event_mmap_ctx(struct perf_event_context *ctx,
4695 struct perf_mmap_event *mmap_event,
4698 struct perf_event *event;
4700 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
4701 if (perf_event_mmap_match(event, mmap_event, executable))
4702 perf_event_mmap_output(event, mmap_event);
4706 static void perf_event_mmap_event(struct perf_mmap_event *mmap_event)
4708 struct perf_cpu_context *cpuctx;
4709 struct perf_event_context *ctx;
4710 struct vm_area_struct *vma = mmap_event->vma;
4711 struct file *file = vma->vm_file;
4719 memset(tmp, 0, sizeof(tmp));
4723 * d_path works from the end of the rb backwards, so we
4724 * need to add enough zero bytes after the string to handle
4725 * the 64bit alignment we do later.
4727 buf = kzalloc(PATH_MAX + sizeof(u64), GFP_KERNEL);
4729 name = strncpy(tmp, "//enomem", sizeof(tmp));
4732 name = d_path(&file->f_path, buf, PATH_MAX);
4734 name = strncpy(tmp, "//toolong", sizeof(tmp));
4738 if (arch_vma_name(mmap_event->vma)) {
4739 name = strncpy(tmp, arch_vma_name(mmap_event->vma),
4741 tmp[sizeof(tmp) - 1] = '\0';
4746 name = strncpy(tmp, "[vdso]", sizeof(tmp));
4748 } else if (vma->vm_start <= vma->vm_mm->start_brk &&
4749 vma->vm_end >= vma->vm_mm->brk) {
4750 name = strncpy(tmp, "[heap]", sizeof(tmp));
4752 } else if (vma->vm_start <= vma->vm_mm->start_stack &&
4753 vma->vm_end >= vma->vm_mm->start_stack) {
4754 name = strncpy(tmp, "[stack]", sizeof(tmp));
4758 name = strncpy(tmp, "//anon", sizeof(tmp));
4763 size = ALIGN(strlen(name)+1, sizeof(u64));
4765 mmap_event->file_name = name;
4766 mmap_event->file_size = size;
4768 mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size;
4771 list_for_each_entry_rcu(pmu, &pmus, entry) {
4772 cpuctx = get_cpu_ptr(pmu->pmu_cpu_context);
4773 if (cpuctx->unique_pmu != pmu)
4775 perf_event_mmap_ctx(&cpuctx->ctx, mmap_event,
4776 vma->vm_flags & VM_EXEC);
4778 ctxn = pmu->task_ctx_nr;
4782 ctx = rcu_dereference(current->perf_event_ctxp[ctxn]);
4784 perf_event_mmap_ctx(ctx, mmap_event,
4785 vma->vm_flags & VM_EXEC);
4788 put_cpu_ptr(pmu->pmu_cpu_context);
4795 void perf_event_mmap(struct vm_area_struct *vma)
4797 struct perf_mmap_event mmap_event;
4799 if (!atomic_read(&nr_mmap_events))
4802 mmap_event = (struct perf_mmap_event){
4808 .type = PERF_RECORD_MMAP,
4809 .misc = PERF_RECORD_MISC_USER,
4814 .start = vma->vm_start,
4815 .len = vma->vm_end - vma->vm_start,
4816 .pgoff = (u64)vma->vm_pgoff << PAGE_SHIFT,
4820 perf_event_mmap_event(&mmap_event);
4824 * IRQ throttle logging
4827 static void perf_log_throttle(struct perf_event *event, int enable)
4829 struct perf_output_handle handle;
4830 struct perf_sample_data sample;
4834 struct perf_event_header header;
4838 } throttle_event = {
4840 .type = PERF_RECORD_THROTTLE,
4842 .size = sizeof(throttle_event),
4844 .time = perf_clock(),
4845 .id = primary_event_id(event),
4846 .stream_id = event->id,
4850 throttle_event.header.type = PERF_RECORD_UNTHROTTLE;
4852 perf_event_header__init_id(&throttle_event.header, &sample, event);
4854 ret = perf_output_begin(&handle, event,
4855 throttle_event.header.size);
4859 perf_output_put(&handle, throttle_event);
4860 perf_event__output_id_sample(event, &handle, &sample);
4861 perf_output_end(&handle);
4865 * Generic event overflow handling, sampling.
4868 static int __perf_event_overflow(struct perf_event *event,
4869 int throttle, struct perf_sample_data *data,
4870 struct pt_regs *regs)
4872 int events = atomic_read(&event->event_limit);
4873 struct hw_perf_event *hwc = &event->hw;
4878 * Non-sampling counters might still use the PMI to fold short
4879 * hardware counters, ignore those.
4881 if (unlikely(!is_sampling_event(event)))
4884 seq = __this_cpu_read(perf_throttled_seq);
4885 if (seq != hwc->interrupts_seq) {
4886 hwc->interrupts_seq = seq;
4887 hwc->interrupts = 1;
4890 if (unlikely(throttle
4891 && hwc->interrupts >= max_samples_per_tick)) {
4892 __this_cpu_inc(perf_throttled_count);
4893 hwc->interrupts = MAX_INTERRUPTS;
4894 perf_log_throttle(event, 0);
4899 if (event->attr.freq) {
4900 u64 now = perf_clock();
4901 s64 delta = now - hwc->freq_time_stamp;
4903 hwc->freq_time_stamp = now;
4905 if (delta > 0 && delta < 2*TICK_NSEC)
4906 perf_adjust_period(event, delta, hwc->last_period, true);
4910 * XXX event_limit might not quite work as expected on inherited
4914 event->pending_kill = POLL_IN;
4915 if (events && atomic_dec_and_test(&event->event_limit)) {
4917 event->pending_kill = POLL_HUP;
4918 event->pending_disable = 1;
4919 irq_work_queue(&event->pending);
4922 if (event->overflow_handler)
4923 event->overflow_handler(event, data, regs);
4925 perf_event_output(event, data, regs);
4927 if (event->fasync && event->pending_kill) {
4928 event->pending_wakeup = 1;
4929 irq_work_queue(&event->pending);
4935 int perf_event_overflow(struct perf_event *event,
4936 struct perf_sample_data *data,
4937 struct pt_regs *regs)
4939 return __perf_event_overflow(event, 1, data, regs);
4943 * Generic software event infrastructure
4946 struct swevent_htable {
4947 struct swevent_hlist *swevent_hlist;
4948 struct mutex hlist_mutex;
4951 /* Recursion avoidance in each contexts */
4952 int recursion[PERF_NR_CONTEXTS];
4955 static DEFINE_PER_CPU(struct swevent_htable, swevent_htable);
4958 * We directly increment event->count and keep a second value in
4959 * event->hw.period_left to count intervals. This period event
4960 * is kept in the range [-sample_period, 0] so that we can use the
4964 static u64 perf_swevent_set_period(struct perf_event *event)
4966 struct hw_perf_event *hwc = &event->hw;
4967 u64 period = hwc->last_period;
4971 hwc->last_period = hwc->sample_period;
4974 old = val = local64_read(&hwc->period_left);
4978 nr = div64_u64(period + val, period);
4979 offset = nr * period;
4981 if (local64_cmpxchg(&hwc->period_left, old, val) != old)
4987 static void perf_swevent_overflow(struct perf_event *event, u64 overflow,
4988 struct perf_sample_data *data,
4989 struct pt_regs *regs)
4991 struct hw_perf_event *hwc = &event->hw;
4995 overflow = perf_swevent_set_period(event);
4997 if (hwc->interrupts == MAX_INTERRUPTS)
5000 for (; overflow; overflow--) {
5001 if (__perf_event_overflow(event, throttle,
5004 * We inhibit the overflow from happening when
5005 * hwc->interrupts == MAX_INTERRUPTS.
5013 static void perf_swevent_event(struct perf_event *event, u64 nr,
5014 struct perf_sample_data *data,
5015 struct pt_regs *regs)
5017 struct hw_perf_event *hwc = &event->hw;
5019 local64_add(nr, &event->count);
5024 if (!is_sampling_event(event))
5027 if ((event->attr.sample_type & PERF_SAMPLE_PERIOD) && !event->attr.freq) {
5029 return perf_swevent_overflow(event, 1, data, regs);
5031 data->period = event->hw.last_period;
5033 if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq)
5034 return perf_swevent_overflow(event, 1, data, regs);
5036 if (local64_add_negative(nr, &hwc->period_left))
5039 perf_swevent_overflow(event, 0, data, regs);
5042 static int perf_exclude_event(struct perf_event *event,
5043 struct pt_regs *regs)
5045 if (event->hw.state & PERF_HES_STOPPED)
5049 if (event->attr.exclude_user && user_mode(regs))
5052 if (event->attr.exclude_kernel && !user_mode(regs))
5059 static int perf_swevent_match(struct perf_event *event,
5060 enum perf_type_id type,
5062 struct perf_sample_data *data,
5063 struct pt_regs *regs)
5065 if (event->attr.type != type)
5068 if (event->attr.config != event_id)
5071 if (perf_exclude_event(event, regs))
5077 static inline u64 swevent_hash(u64 type, u32 event_id)
5079 u64 val = event_id | (type << 32);
5081 return hash_64(val, SWEVENT_HLIST_BITS);
5084 static inline struct hlist_head *
5085 __find_swevent_head(struct swevent_hlist *hlist, u64 type, u32 event_id)
5087 u64 hash = swevent_hash(type, event_id);
5089 return &hlist->heads[hash];
5092 /* For the read side: events when they trigger */
5093 static inline struct hlist_head *
5094 find_swevent_head_rcu(struct swevent_htable *swhash, u64 type, u32 event_id)
5096 struct swevent_hlist *hlist;
5098 hlist = rcu_dereference(swhash->swevent_hlist);
5102 return __find_swevent_head(hlist, type, event_id);
5105 /* For the event head insertion and removal in the hlist */
5106 static inline struct hlist_head *
5107 find_swevent_head(struct swevent_htable *swhash, struct perf_event *event)
5109 struct swevent_hlist *hlist;
5110 u32 event_id = event->attr.config;
5111 u64 type = event->attr.type;
5114 * Event scheduling is always serialized against hlist allocation
5115 * and release. Which makes the protected version suitable here.
5116 * The context lock guarantees that.
5118 hlist = rcu_dereference_protected(swhash->swevent_hlist,
5119 lockdep_is_held(&event->ctx->lock));
5123 return __find_swevent_head(hlist, type, event_id);
5126 static void do_perf_sw_event(enum perf_type_id type, u32 event_id,
5128 struct perf_sample_data *data,
5129 struct pt_regs *regs)
5131 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
5132 struct perf_event *event;
5133 struct hlist_head *head;
5136 head = find_swevent_head_rcu(swhash, type, event_id);
5140 hlist_for_each_entry_rcu(event, head, hlist_entry) {
5141 if (perf_swevent_match(event, type, event_id, data, regs))
5142 perf_swevent_event(event, nr, data, regs);
5148 int perf_swevent_get_recursion_context(void)
5150 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
5152 return get_recursion_context(swhash->recursion);
5154 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context);
5156 inline void perf_swevent_put_recursion_context(int rctx)
5158 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
5160 put_recursion_context(swhash->recursion, rctx);
5163 void __perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)
5165 struct perf_sample_data data;
5168 preempt_disable_notrace();
5169 rctx = perf_swevent_get_recursion_context();
5173 perf_sample_data_init(&data, addr, 0);
5175 do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, &data, regs);
5177 perf_swevent_put_recursion_context(rctx);
5178 preempt_enable_notrace();
5181 static void perf_swevent_read(struct perf_event *event)
5185 static int perf_swevent_add(struct perf_event *event, int flags)
5187 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
5188 struct hw_perf_event *hwc = &event->hw;
5189 struct hlist_head *head;
5191 if (is_sampling_event(event)) {
5192 hwc->last_period = hwc->sample_period;
5193 perf_swevent_set_period(event);
5196 hwc->state = !(flags & PERF_EF_START);
5198 head = find_swevent_head(swhash, event);
5199 if (WARN_ON_ONCE(!head))
5202 hlist_add_head_rcu(&event->hlist_entry, head);
5207 static void perf_swevent_del(struct perf_event *event, int flags)
5209 hlist_del_rcu(&event->hlist_entry);
5212 static void perf_swevent_start(struct perf_event *event, int flags)
5214 event->hw.state = 0;
5217 static void perf_swevent_stop(struct perf_event *event, int flags)
5219 event->hw.state = PERF_HES_STOPPED;
5222 /* Deref the hlist from the update side */
5223 static inline struct swevent_hlist *
5224 swevent_hlist_deref(struct swevent_htable *swhash)
5226 return rcu_dereference_protected(swhash->swevent_hlist,
5227 lockdep_is_held(&swhash->hlist_mutex));
5230 static void swevent_hlist_release(struct swevent_htable *swhash)
5232 struct swevent_hlist *hlist = swevent_hlist_deref(swhash);
5237 rcu_assign_pointer(swhash->swevent_hlist, NULL);
5238 kfree_rcu(hlist, rcu_head);
5241 static void swevent_hlist_put_cpu(struct perf_event *event, int cpu)
5243 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
5245 mutex_lock(&swhash->hlist_mutex);
5247 if (!--swhash->hlist_refcount)
5248 swevent_hlist_release(swhash);
5250 mutex_unlock(&swhash->hlist_mutex);
5253 static void swevent_hlist_put(struct perf_event *event)
5257 if (event->cpu != -1) {
5258 swevent_hlist_put_cpu(event, event->cpu);
5262 for_each_possible_cpu(cpu)
5263 swevent_hlist_put_cpu(event, cpu);
5266 static int swevent_hlist_get_cpu(struct perf_event *event, int cpu)
5268 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
5271 mutex_lock(&swhash->hlist_mutex);
5273 if (!swevent_hlist_deref(swhash) && cpu_online(cpu)) {
5274 struct swevent_hlist *hlist;
5276 hlist = kzalloc(sizeof(*hlist), GFP_KERNEL);
5281 rcu_assign_pointer(swhash->swevent_hlist, hlist);
5283 swhash->hlist_refcount++;
5285 mutex_unlock(&swhash->hlist_mutex);
5290 static int swevent_hlist_get(struct perf_event *event)
5293 int cpu, failed_cpu;
5295 if (event->cpu != -1)
5296 return swevent_hlist_get_cpu(event, event->cpu);
5299 for_each_possible_cpu(cpu) {
5300 err = swevent_hlist_get_cpu(event, cpu);
5310 for_each_possible_cpu(cpu) {
5311 if (cpu == failed_cpu)
5313 swevent_hlist_put_cpu(event, cpu);
5320 struct static_key perf_swevent_enabled[PERF_COUNT_SW_MAX];
5322 static void sw_perf_event_destroy(struct perf_event *event)
5324 u64 event_id = event->attr.config;
5326 WARN_ON(event->parent);
5328 static_key_slow_dec(&perf_swevent_enabled[event_id]);
5329 swevent_hlist_put(event);
5332 static int perf_swevent_init(struct perf_event *event)
5334 u64 event_id = event->attr.config;
5336 if (event->attr.type != PERF_TYPE_SOFTWARE)
5340 * no branch sampling for software events
5342 if (has_branch_stack(event))
5346 case PERF_COUNT_SW_CPU_CLOCK:
5347 case PERF_COUNT_SW_TASK_CLOCK:
5354 if (event_id >= PERF_COUNT_SW_MAX)
5357 if (!event->parent) {
5360 err = swevent_hlist_get(event);
5364 static_key_slow_inc(&perf_swevent_enabled[event_id]);
5365 event->destroy = sw_perf_event_destroy;
5371 static int perf_swevent_event_idx(struct perf_event *event)
5376 static struct pmu perf_swevent = {
5377 .task_ctx_nr = perf_sw_context,
5379 .event_init = perf_swevent_init,
5380 .add = perf_swevent_add,
5381 .del = perf_swevent_del,
5382 .start = perf_swevent_start,
5383 .stop = perf_swevent_stop,
5384 .read = perf_swevent_read,
5386 .event_idx = perf_swevent_event_idx,
5389 #ifdef CONFIG_EVENT_TRACING
5391 static int perf_tp_filter_match(struct perf_event *event,
5392 struct perf_sample_data *data)
5394 void *record = data->raw->data;
5396 if (likely(!event->filter) || filter_match_preds(event->filter, record))
5401 static int perf_tp_event_match(struct perf_event *event,
5402 struct perf_sample_data *data,
5403 struct pt_regs *regs)
5405 if (event->hw.state & PERF_HES_STOPPED)
5408 * All tracepoints are from kernel-space.
5410 if (event->attr.exclude_kernel)
5413 if (!perf_tp_filter_match(event, data))
5419 void perf_tp_event(u64 addr, u64 count, void *record, int entry_size,
5420 struct pt_regs *regs, struct hlist_head *head, int rctx,
5421 struct task_struct *task)
5423 struct perf_sample_data data;
5424 struct perf_event *event;
5426 struct perf_raw_record raw = {
5431 perf_sample_data_init(&data, addr, 0);
5434 hlist_for_each_entry_rcu(event, head, hlist_entry) {
5435 if (perf_tp_event_match(event, &data, regs))
5436 perf_swevent_event(event, count, &data, regs);
5440 * If we got specified a target task, also iterate its context and
5441 * deliver this event there too.
5443 if (task && task != current) {
5444 struct perf_event_context *ctx;
5445 struct trace_entry *entry = record;
5448 ctx = rcu_dereference(task->perf_event_ctxp[perf_sw_context]);
5452 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
5453 if (event->attr.type != PERF_TYPE_TRACEPOINT)
5455 if (event->attr.config != entry->type)
5457 if (perf_tp_event_match(event, &data, regs))
5458 perf_swevent_event(event, count, &data, regs);
5464 perf_swevent_put_recursion_context(rctx);
5466 EXPORT_SYMBOL_GPL(perf_tp_event);
5468 static void tp_perf_event_destroy(struct perf_event *event)
5470 perf_trace_destroy(event);
5473 static int perf_tp_event_init(struct perf_event *event)
5477 if (event->attr.type != PERF_TYPE_TRACEPOINT)
5481 * no branch sampling for tracepoint events
5483 if (has_branch_stack(event))
5486 err = perf_trace_init(event);
5490 event->destroy = tp_perf_event_destroy;
5495 static struct pmu perf_tracepoint = {
5496 .task_ctx_nr = perf_sw_context,
5498 .event_init = perf_tp_event_init,
5499 .add = perf_trace_add,
5500 .del = perf_trace_del,
5501 .start = perf_swevent_start,
5502 .stop = perf_swevent_stop,
5503 .read = perf_swevent_read,
5505 .event_idx = perf_swevent_event_idx,
5508 static inline void perf_tp_register(void)
5510 perf_pmu_register(&perf_tracepoint, "tracepoint", PERF_TYPE_TRACEPOINT);
5513 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
5518 if (event->attr.type != PERF_TYPE_TRACEPOINT)
5521 filter_str = strndup_user(arg, PAGE_SIZE);
5522 if (IS_ERR(filter_str))
5523 return PTR_ERR(filter_str);
5525 ret = ftrace_profile_set_filter(event, event->attr.config, filter_str);
5531 static void perf_event_free_filter(struct perf_event *event)
5533 ftrace_profile_free_filter(event);
5538 static inline void perf_tp_register(void)
5542 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
5547 static void perf_event_free_filter(struct perf_event *event)
5551 #endif /* CONFIG_EVENT_TRACING */
5553 #ifdef CONFIG_HAVE_HW_BREAKPOINT
5554 void perf_bp_event(struct perf_event *bp, void *data)
5556 struct perf_sample_data sample;
5557 struct pt_regs *regs = data;
5559 perf_sample_data_init(&sample, bp->attr.bp_addr, 0);
5561 if (!bp->hw.state && !perf_exclude_event(bp, regs))
5562 perf_swevent_event(bp, 1, &sample, regs);
5567 * hrtimer based swevent callback
5570 static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer)
5572 enum hrtimer_restart ret = HRTIMER_RESTART;
5573 struct perf_sample_data data;
5574 struct pt_regs *regs;
5575 struct perf_event *event;
5578 event = container_of(hrtimer, struct perf_event, hw.hrtimer);
5580 if (event->state != PERF_EVENT_STATE_ACTIVE)
5581 return HRTIMER_NORESTART;
5583 event->pmu->read(event);
5585 perf_sample_data_init(&data, 0, event->hw.last_period);
5586 regs = get_irq_regs();
5588 if (regs && !perf_exclude_event(event, regs)) {
5589 if (!(event->attr.exclude_idle && is_idle_task(current)))
5590 if (__perf_event_overflow(event, 1, &data, regs))
5591 ret = HRTIMER_NORESTART;
5594 period = max_t(u64, 10000, event->hw.sample_period);
5595 hrtimer_forward_now(hrtimer, ns_to_ktime(period));
5600 static void perf_swevent_start_hrtimer(struct perf_event *event)
5602 struct hw_perf_event *hwc = &event->hw;
5605 if (!is_sampling_event(event))
5608 period = local64_read(&hwc->period_left);
5613 local64_set(&hwc->period_left, 0);
5615 period = max_t(u64, 10000, hwc->sample_period);
5617 __hrtimer_start_range_ns(&hwc->hrtimer,
5618 ns_to_ktime(period), 0,
5619 HRTIMER_MODE_REL_PINNED, 0);
5622 static void perf_swevent_cancel_hrtimer(struct perf_event *event)
5624 struct hw_perf_event *hwc = &event->hw;
5626 if (is_sampling_event(event)) {
5627 ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer);
5628 local64_set(&hwc->period_left, ktime_to_ns(remaining));
5630 hrtimer_cancel(&hwc->hrtimer);
5634 static void perf_swevent_init_hrtimer(struct perf_event *event)
5636 struct hw_perf_event *hwc = &event->hw;
5638 if (!is_sampling_event(event))
5641 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
5642 hwc->hrtimer.function = perf_swevent_hrtimer;
5645 * Since hrtimers have a fixed rate, we can do a static freq->period
5646 * mapping and avoid the whole period adjust feedback stuff.
5648 if (event->attr.freq) {
5649 long freq = event->attr.sample_freq;
5651 event->attr.sample_period = NSEC_PER_SEC / freq;
5652 hwc->sample_period = event->attr.sample_period;
5653 local64_set(&hwc->period_left, hwc->sample_period);
5654 hwc->last_period = hwc->sample_period;
5655 event->attr.freq = 0;
5660 * Software event: cpu wall time clock
5663 static void cpu_clock_event_update(struct perf_event *event)
5668 now = local_clock();
5669 prev = local64_xchg(&event->hw.prev_count, now);
5670 local64_add(now - prev, &event->count);
5673 static void cpu_clock_event_start(struct perf_event *event, int flags)
5675 local64_set(&event->hw.prev_count, local_clock());
5676 perf_swevent_start_hrtimer(event);
5679 static void cpu_clock_event_stop(struct perf_event *event, int flags)
5681 perf_swevent_cancel_hrtimer(event);
5682 cpu_clock_event_update(event);
5685 static int cpu_clock_event_add(struct perf_event *event, int flags)
5687 if (flags & PERF_EF_START)
5688 cpu_clock_event_start(event, flags);
5693 static void cpu_clock_event_del(struct perf_event *event, int flags)
5695 cpu_clock_event_stop(event, flags);
5698 static void cpu_clock_event_read(struct perf_event *event)
5700 cpu_clock_event_update(event);
5703 static int cpu_clock_event_init(struct perf_event *event)
5705 if (event->attr.type != PERF_TYPE_SOFTWARE)
5708 if (event->attr.config != PERF_COUNT_SW_CPU_CLOCK)
5712 * no branch sampling for software events
5714 if (has_branch_stack(event))
5717 perf_swevent_init_hrtimer(event);
5722 static struct pmu perf_cpu_clock = {
5723 .task_ctx_nr = perf_sw_context,
5725 .event_init = cpu_clock_event_init,
5726 .add = cpu_clock_event_add,
5727 .del = cpu_clock_event_del,
5728 .start = cpu_clock_event_start,
5729 .stop = cpu_clock_event_stop,
5730 .read = cpu_clock_event_read,
5732 .event_idx = perf_swevent_event_idx,
5736 * Software event: task time clock
5739 static void task_clock_event_update(struct perf_event *event, u64 now)
5744 prev = local64_xchg(&event->hw.prev_count, now);
5746 local64_add(delta, &event->count);
5749 static void task_clock_event_start(struct perf_event *event, int flags)
5751 local64_set(&event->hw.prev_count, event->ctx->time);
5752 perf_swevent_start_hrtimer(event);
5755 static void task_clock_event_stop(struct perf_event *event, int flags)
5757 perf_swevent_cancel_hrtimer(event);
5758 task_clock_event_update(event, event->ctx->time);
5761 static int task_clock_event_add(struct perf_event *event, int flags)
5763 if (flags & PERF_EF_START)
5764 task_clock_event_start(event, flags);
5769 static void task_clock_event_del(struct perf_event *event, int flags)
5771 task_clock_event_stop(event, PERF_EF_UPDATE);
5774 static void task_clock_event_read(struct perf_event *event)
5776 u64 now = perf_clock();
5777 u64 delta = now - event->ctx->timestamp;
5778 u64 time = event->ctx->time + delta;
5780 task_clock_event_update(event, time);
5783 static int task_clock_event_init(struct perf_event *event)
5785 if (event->attr.type != PERF_TYPE_SOFTWARE)
5788 if (event->attr.config != PERF_COUNT_SW_TASK_CLOCK)
5792 * no branch sampling for software events
5794 if (has_branch_stack(event))
5797 perf_swevent_init_hrtimer(event);
5802 static struct pmu perf_task_clock = {
5803 .task_ctx_nr = perf_sw_context,
5805 .event_init = task_clock_event_init,
5806 .add = task_clock_event_add,
5807 .del = task_clock_event_del,
5808 .start = task_clock_event_start,
5809 .stop = task_clock_event_stop,
5810 .read = task_clock_event_read,
5812 .event_idx = perf_swevent_event_idx,
5815 static void perf_pmu_nop_void(struct pmu *pmu)
5819 static int perf_pmu_nop_int(struct pmu *pmu)
5824 static void perf_pmu_start_txn(struct pmu *pmu)
5826 perf_pmu_disable(pmu);
5829 static int perf_pmu_commit_txn(struct pmu *pmu)
5831 perf_pmu_enable(pmu);
5835 static void perf_pmu_cancel_txn(struct pmu *pmu)
5837 perf_pmu_enable(pmu);
5840 static int perf_event_idx_default(struct perf_event *event)
5842 return event->hw.idx + 1;
5846 * Ensures all contexts with the same task_ctx_nr have the same
5847 * pmu_cpu_context too.
5849 static void *find_pmu_context(int ctxn)
5856 list_for_each_entry(pmu, &pmus, entry) {
5857 if (pmu->task_ctx_nr == ctxn)
5858 return pmu->pmu_cpu_context;
5864 static void update_pmu_context(struct pmu *pmu, struct pmu *old_pmu)
5868 for_each_possible_cpu(cpu) {
5869 struct perf_cpu_context *cpuctx;
5871 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
5873 if (cpuctx->unique_pmu == old_pmu)
5874 cpuctx->unique_pmu = pmu;
5878 static void free_pmu_context(struct pmu *pmu)
5882 mutex_lock(&pmus_lock);
5884 * Like a real lame refcount.
5886 list_for_each_entry(i, &pmus, entry) {
5887 if (i->pmu_cpu_context == pmu->pmu_cpu_context) {
5888 update_pmu_context(i, pmu);
5893 free_percpu(pmu->pmu_cpu_context);
5895 mutex_unlock(&pmus_lock);
5897 static struct idr pmu_idr;
5900 type_show(struct device *dev, struct device_attribute *attr, char *page)
5902 struct pmu *pmu = dev_get_drvdata(dev);
5904 return snprintf(page, PAGE_SIZE-1, "%d\n", pmu->type);
5907 static struct device_attribute pmu_dev_attrs[] = {
5912 static int pmu_bus_running;
5913 static struct bus_type pmu_bus = {
5914 .name = "event_source",
5915 .dev_attrs = pmu_dev_attrs,
5918 static void pmu_dev_release(struct device *dev)
5923 static int pmu_dev_alloc(struct pmu *pmu)
5927 pmu->dev = kzalloc(sizeof(struct device), GFP_KERNEL);
5931 pmu->dev->groups = pmu->attr_groups;
5932 device_initialize(pmu->dev);
5933 ret = dev_set_name(pmu->dev, "%s", pmu->name);
5937 dev_set_drvdata(pmu->dev, pmu);
5938 pmu->dev->bus = &pmu_bus;
5939 pmu->dev->release = pmu_dev_release;
5940 ret = device_add(pmu->dev);
5948 put_device(pmu->dev);
5952 static struct lock_class_key cpuctx_mutex;
5953 static struct lock_class_key cpuctx_lock;
5955 int perf_pmu_register(struct pmu *pmu, char *name, int type)
5959 mutex_lock(&pmus_lock);
5961 pmu->pmu_disable_count = alloc_percpu(int);
5962 if (!pmu->pmu_disable_count)
5971 type = idr_alloc(&pmu_idr, pmu, PERF_TYPE_MAX, 0, GFP_KERNEL);
5979 if (pmu_bus_running) {
5980 ret = pmu_dev_alloc(pmu);
5986 pmu->pmu_cpu_context = find_pmu_context(pmu->task_ctx_nr);
5987 if (pmu->pmu_cpu_context)
5988 goto got_cpu_context;
5991 pmu->pmu_cpu_context = alloc_percpu(struct perf_cpu_context);
5992 if (!pmu->pmu_cpu_context)
5995 for_each_possible_cpu(cpu) {
5996 struct perf_cpu_context *cpuctx;
5998 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
5999 __perf_event_init_context(&cpuctx->ctx);
6000 lockdep_set_class(&cpuctx->ctx.mutex, &cpuctx_mutex);
6001 lockdep_set_class(&cpuctx->ctx.lock, &cpuctx_lock);
6002 cpuctx->ctx.type = cpu_context;
6003 cpuctx->ctx.pmu = pmu;
6004 cpuctx->jiffies_interval = 1;
6005 INIT_LIST_HEAD(&cpuctx->rotation_list);
6006 cpuctx->unique_pmu = pmu;
6010 if (!pmu->start_txn) {
6011 if (pmu->pmu_enable) {
6013 * If we have pmu_enable/pmu_disable calls, install
6014 * transaction stubs that use that to try and batch
6015 * hardware accesses.
6017 pmu->start_txn = perf_pmu_start_txn;
6018 pmu->commit_txn = perf_pmu_commit_txn;
6019 pmu->cancel_txn = perf_pmu_cancel_txn;
6021 pmu->start_txn = perf_pmu_nop_void;
6022 pmu->commit_txn = perf_pmu_nop_int;
6023 pmu->cancel_txn = perf_pmu_nop_void;
6027 if (!pmu->pmu_enable) {
6028 pmu->pmu_enable = perf_pmu_nop_void;
6029 pmu->pmu_disable = perf_pmu_nop_void;
6032 if (!pmu->event_idx)
6033 pmu->event_idx = perf_event_idx_default;
6035 list_add_rcu(&pmu->entry, &pmus);
6038 mutex_unlock(&pmus_lock);
6043 device_del(pmu->dev);
6044 put_device(pmu->dev);
6047 if (pmu->type >= PERF_TYPE_MAX)
6048 idr_remove(&pmu_idr, pmu->type);
6051 free_percpu(pmu->pmu_disable_count);
6055 void perf_pmu_unregister(struct pmu *pmu)
6057 mutex_lock(&pmus_lock);
6058 list_del_rcu(&pmu->entry);
6059 mutex_unlock(&pmus_lock);
6062 * We dereference the pmu list under both SRCU and regular RCU, so
6063 * synchronize against both of those.
6065 synchronize_srcu(&pmus_srcu);
6068 free_percpu(pmu->pmu_disable_count);
6069 if (pmu->type >= PERF_TYPE_MAX)
6070 idr_remove(&pmu_idr, pmu->type);
6071 device_del(pmu->dev);
6072 put_device(pmu->dev);
6073 free_pmu_context(pmu);
6076 struct pmu *perf_init_event(struct perf_event *event)
6078 struct pmu *pmu = NULL;
6082 idx = srcu_read_lock(&pmus_srcu);
6085 pmu = idr_find(&pmu_idr, event->attr.type);
6089 ret = pmu->event_init(event);
6095 list_for_each_entry_rcu(pmu, &pmus, entry) {
6097 ret = pmu->event_init(event);
6101 if (ret != -ENOENT) {
6106 pmu = ERR_PTR(-ENOENT);
6108 srcu_read_unlock(&pmus_srcu, idx);
6114 * Allocate and initialize a event structure
6116 static struct perf_event *
6117 perf_event_alloc(struct perf_event_attr *attr, int cpu,
6118 struct task_struct *task,
6119 struct perf_event *group_leader,
6120 struct perf_event *parent_event,
6121 perf_overflow_handler_t overflow_handler,
6125 struct perf_event *event;
6126 struct hw_perf_event *hwc;
6129 if ((unsigned)cpu >= nr_cpu_ids) {
6130 if (!task || cpu != -1)
6131 return ERR_PTR(-EINVAL);
6134 event = kzalloc(sizeof(*event), GFP_KERNEL);
6136 return ERR_PTR(-ENOMEM);
6139 * Single events are their own group leaders, with an
6140 * empty sibling list:
6143 group_leader = event;
6145 mutex_init(&event->child_mutex);
6146 INIT_LIST_HEAD(&event->child_list);
6148 INIT_LIST_HEAD(&event->group_entry);
6149 INIT_LIST_HEAD(&event->event_entry);
6150 INIT_LIST_HEAD(&event->sibling_list);
6151 INIT_LIST_HEAD(&event->rb_entry);
6153 init_waitqueue_head(&event->waitq);
6154 init_irq_work(&event->pending, perf_pending_event);
6156 mutex_init(&event->mmap_mutex);
6158 atomic_long_set(&event->refcount, 1);
6160 event->attr = *attr;
6161 event->group_leader = group_leader;
6165 event->parent = parent_event;
6167 event->ns = get_pid_ns(task_active_pid_ns(current));
6168 event->id = atomic64_inc_return(&perf_event_id);
6170 event->state = PERF_EVENT_STATE_INACTIVE;
6173 event->attach_state = PERF_ATTACH_TASK;
6175 if (attr->type == PERF_TYPE_TRACEPOINT)
6176 event->hw.tp_target = task;
6177 #ifdef CONFIG_HAVE_HW_BREAKPOINT
6179 * hw_breakpoint is a bit difficult here..
6181 else if (attr->type == PERF_TYPE_BREAKPOINT)
6182 event->hw.bp_target = task;
6186 if (!overflow_handler && parent_event) {
6187 overflow_handler = parent_event->overflow_handler;
6188 context = parent_event->overflow_handler_context;
6191 event->overflow_handler = overflow_handler;
6192 event->overflow_handler_context = context;
6194 perf_event__state_init(event);
6199 hwc->sample_period = attr->sample_period;
6200 if (attr->freq && attr->sample_freq)
6201 hwc->sample_period = 1;
6202 hwc->last_period = hwc->sample_period;
6204 local64_set(&hwc->period_left, hwc->sample_period);
6207 * we currently do not support PERF_FORMAT_GROUP on inherited events
6209 if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP))
6212 pmu = perf_init_event(event);
6218 else if (IS_ERR(pmu))
6223 put_pid_ns(event->ns);
6225 return ERR_PTR(err);
6228 if (!event->parent) {
6229 if (event->attach_state & PERF_ATTACH_TASK)
6230 static_key_slow_inc(&perf_sched_events.key);
6231 if (event->attr.mmap || event->attr.mmap_data)
6232 atomic_inc(&nr_mmap_events);
6233 if (event->attr.comm)
6234 atomic_inc(&nr_comm_events);
6235 if (event->attr.task)
6236 atomic_inc(&nr_task_events);
6237 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) {
6238 err = get_callchain_buffers();
6241 return ERR_PTR(err);
6244 if (has_branch_stack(event)) {
6245 static_key_slow_inc(&perf_sched_events.key);
6246 if (!(event->attach_state & PERF_ATTACH_TASK))
6247 atomic_inc(&per_cpu(perf_branch_stack_events,
6255 static int perf_copy_attr(struct perf_event_attr __user *uattr,
6256 struct perf_event_attr *attr)
6261 if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0))
6265 * zero the full structure, so that a short copy will be nice.
6267 memset(attr, 0, sizeof(*attr));
6269 ret = get_user(size, &uattr->size);
6273 if (size > PAGE_SIZE) /* silly large */
6276 if (!size) /* abi compat */
6277 size = PERF_ATTR_SIZE_VER0;
6279 if (size < PERF_ATTR_SIZE_VER0)
6283 * If we're handed a bigger struct than we know of,
6284 * ensure all the unknown bits are 0 - i.e. new
6285 * user-space does not rely on any kernel feature
6286 * extensions we dont know about yet.
6288 if (size > sizeof(*attr)) {
6289 unsigned char __user *addr;
6290 unsigned char __user *end;
6293 addr = (void __user *)uattr + sizeof(*attr);
6294 end = (void __user *)uattr + size;
6296 for (; addr < end; addr++) {
6297 ret = get_user(val, addr);
6303 size = sizeof(*attr);
6306 ret = copy_from_user(attr, uattr, size);
6310 if (attr->__reserved_1)
6313 if (attr->sample_type & ~(PERF_SAMPLE_MAX-1))
6316 if (attr->read_format & ~(PERF_FORMAT_MAX-1))
6319 if (attr->sample_type & PERF_SAMPLE_BRANCH_STACK) {
6320 u64 mask = attr->branch_sample_type;
6322 /* only using defined bits */
6323 if (mask & ~(PERF_SAMPLE_BRANCH_MAX-1))
6326 /* at least one branch bit must be set */
6327 if (!(mask & ~PERF_SAMPLE_BRANCH_PLM_ALL))
6330 /* kernel level capture: check permissions */
6331 if ((mask & PERF_SAMPLE_BRANCH_PERM_PLM)
6332 && perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
6335 /* propagate priv level, when not set for branch */
6336 if (!(mask & PERF_SAMPLE_BRANCH_PLM_ALL)) {
6338 /* exclude_kernel checked on syscall entry */
6339 if (!attr->exclude_kernel)
6340 mask |= PERF_SAMPLE_BRANCH_KERNEL;
6342 if (!attr->exclude_user)
6343 mask |= PERF_SAMPLE_BRANCH_USER;
6345 if (!attr->exclude_hv)
6346 mask |= PERF_SAMPLE_BRANCH_HV;
6348 * adjust user setting (for HW filter setup)
6350 attr->branch_sample_type = mask;
6354 if (attr->sample_type & PERF_SAMPLE_REGS_USER) {
6355 ret = perf_reg_validate(attr->sample_regs_user);
6360 if (attr->sample_type & PERF_SAMPLE_STACK_USER) {
6361 if (!arch_perf_have_user_stack_dump())
6365 * We have __u32 type for the size, but so far
6366 * we can only use __u16 as maximum due to the
6367 * __u16 sample size limit.
6369 if (attr->sample_stack_user >= USHRT_MAX)
6371 else if (!IS_ALIGNED(attr->sample_stack_user, sizeof(u64)))
6379 put_user(sizeof(*attr), &uattr->size);
6385 perf_event_set_output(struct perf_event *event, struct perf_event *output_event)
6387 struct ring_buffer *rb = NULL, *old_rb = NULL;
6393 /* don't allow circular references */
6394 if (event == output_event)
6398 * Don't allow cross-cpu buffers
6400 if (output_event->cpu != event->cpu)
6404 * If its not a per-cpu rb, it must be the same task.
6406 if (output_event->cpu == -1 && output_event->ctx != event->ctx)
6410 mutex_lock(&event->mmap_mutex);
6411 /* Can't redirect output if we've got an active mmap() */
6412 if (atomic_read(&event->mmap_count))
6416 /* get the rb we want to redirect to */
6417 rb = ring_buffer_get(output_event);
6423 rcu_assign_pointer(event->rb, rb);
6425 ring_buffer_detach(event, old_rb);
6428 mutex_unlock(&event->mmap_mutex);
6431 ring_buffer_put(old_rb);
6437 * sys_perf_event_open - open a performance event, associate it to a task/cpu
6439 * @attr_uptr: event_id type attributes for monitoring/sampling
6442 * @group_fd: group leader event fd
6444 SYSCALL_DEFINE5(perf_event_open,
6445 struct perf_event_attr __user *, attr_uptr,
6446 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
6448 struct perf_event *group_leader = NULL, *output_event = NULL;
6449 struct perf_event *event, *sibling;
6450 struct perf_event_attr attr;
6451 struct perf_event_context *ctx;
6452 struct file *event_file = NULL;
6453 struct fd group = {NULL, 0};
6454 struct task_struct *task = NULL;
6460 /* for future expandability... */
6461 if (flags & ~PERF_FLAG_ALL)
6464 err = perf_copy_attr(attr_uptr, &attr);
6468 if (!attr.exclude_kernel) {
6469 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
6474 if (attr.sample_freq > sysctl_perf_event_sample_rate)
6479 * In cgroup mode, the pid argument is used to pass the fd
6480 * opened to the cgroup directory in cgroupfs. The cpu argument
6481 * designates the cpu on which to monitor threads from that
6484 if ((flags & PERF_FLAG_PID_CGROUP) && (pid == -1 || cpu == -1))
6487 event_fd = get_unused_fd();
6491 if (group_fd != -1) {
6492 err = perf_fget_light(group_fd, &group);
6495 group_leader = group.file->private_data;
6496 if (flags & PERF_FLAG_FD_OUTPUT)
6497 output_event = group_leader;
6498 if (flags & PERF_FLAG_FD_NO_GROUP)
6499 group_leader = NULL;
6502 if (pid != -1 && !(flags & PERF_FLAG_PID_CGROUP)) {
6503 task = find_lively_task_by_vpid(pid);
6505 err = PTR_ERR(task);
6512 event = perf_event_alloc(&attr, cpu, task, group_leader, NULL,
6514 if (IS_ERR(event)) {
6515 err = PTR_ERR(event);
6519 if (flags & PERF_FLAG_PID_CGROUP) {
6520 err = perf_cgroup_connect(pid, event, &attr, group_leader);
6525 * - that has cgroup constraint on event->cpu
6526 * - that may need work on context switch
6528 atomic_inc(&per_cpu(perf_cgroup_events, event->cpu));
6529 static_key_slow_inc(&perf_sched_events.key);
6533 * Special case software events and allow them to be part of
6534 * any hardware group.
6539 (is_software_event(event) != is_software_event(group_leader))) {
6540 if (is_software_event(event)) {
6542 * If event and group_leader are not both a software
6543 * event, and event is, then group leader is not.
6545 * Allow the addition of software events to !software
6546 * groups, this is safe because software events never
6549 pmu = group_leader->pmu;
6550 } else if (is_software_event(group_leader) &&
6551 (group_leader->group_flags & PERF_GROUP_SOFTWARE)) {
6553 * In case the group is a pure software group, and we
6554 * try to add a hardware event, move the whole group to
6555 * the hardware context.
6562 * Get the target context (task or percpu):
6564 ctx = find_get_context(pmu, task, event->cpu);
6571 put_task_struct(task);
6576 * Look up the group leader (we will attach this event to it):
6582 * Do not allow a recursive hierarchy (this new sibling
6583 * becoming part of another group-sibling):
6585 if (group_leader->group_leader != group_leader)
6588 * Do not allow to attach to a group in a different
6589 * task or CPU context:
6592 if (group_leader->ctx->type != ctx->type)
6595 if (group_leader->ctx != ctx)
6600 * Only a group leader can be exclusive or pinned
6602 if (attr.exclusive || attr.pinned)
6607 err = perf_event_set_output(event, output_event);
6612 event_file = anon_inode_getfile("[perf_event]", &perf_fops, event, O_RDWR);
6613 if (IS_ERR(event_file)) {
6614 err = PTR_ERR(event_file);
6619 struct perf_event_context *gctx = group_leader->ctx;
6621 mutex_lock(&gctx->mutex);
6622 perf_remove_from_context(group_leader);
6625 * Removing from the context ends up with disabled
6626 * event. What we want here is event in the initial
6627 * startup state, ready to be add into new context.
6629 perf_event__state_init(group_leader);
6630 list_for_each_entry(sibling, &group_leader->sibling_list,
6632 perf_remove_from_context(sibling);
6633 perf_event__state_init(sibling);
6636 mutex_unlock(&gctx->mutex);
6640 WARN_ON_ONCE(ctx->parent_ctx);
6641 mutex_lock(&ctx->mutex);
6645 perf_install_in_context(ctx, group_leader, event->cpu);
6647 list_for_each_entry(sibling, &group_leader->sibling_list,
6649 perf_install_in_context(ctx, sibling, event->cpu);
6654 perf_install_in_context(ctx, event, event->cpu);
6656 perf_unpin_context(ctx);
6657 mutex_unlock(&ctx->mutex);
6661 event->owner = current;
6663 mutex_lock(¤t->perf_event_mutex);
6664 list_add_tail(&event->owner_entry, ¤t->perf_event_list);
6665 mutex_unlock(¤t->perf_event_mutex);
6668 * Precalculate sample_data sizes
6670 perf_event__header_size(event);
6671 perf_event__id_header_size(event);
6674 * Drop the reference on the group_event after placing the
6675 * new event on the sibling_list. This ensures destruction
6676 * of the group leader will find the pointer to itself in
6677 * perf_group_detach().
6680 fd_install(event_fd, event_file);
6684 perf_unpin_context(ctx);
6691 put_task_struct(task);
6695 put_unused_fd(event_fd);
6700 * perf_event_create_kernel_counter
6702 * @attr: attributes of the counter to create
6703 * @cpu: cpu in which the counter is bound
6704 * @task: task to profile (NULL for percpu)
6707 perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu,
6708 struct task_struct *task,
6709 perf_overflow_handler_t overflow_handler,
6712 struct perf_event_context *ctx;
6713 struct perf_event *event;
6717 * Get the target context (task or percpu):
6720 event = perf_event_alloc(attr, cpu, task, NULL, NULL,
6721 overflow_handler, context);
6722 if (IS_ERR(event)) {
6723 err = PTR_ERR(event);
6727 ctx = find_get_context(event->pmu, task, cpu);
6733 WARN_ON_ONCE(ctx->parent_ctx);
6734 mutex_lock(&ctx->mutex);
6735 perf_install_in_context(ctx, event, cpu);
6737 perf_unpin_context(ctx);
6738 mutex_unlock(&ctx->mutex);
6745 return ERR_PTR(err);
6747 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter);
6749 void perf_pmu_migrate_context(struct pmu *pmu, int src_cpu, int dst_cpu)
6751 struct perf_event_context *src_ctx;
6752 struct perf_event_context *dst_ctx;
6753 struct perf_event *event, *tmp;
6756 src_ctx = &per_cpu_ptr(pmu->pmu_cpu_context, src_cpu)->ctx;
6757 dst_ctx = &per_cpu_ptr(pmu->pmu_cpu_context, dst_cpu)->ctx;
6759 mutex_lock(&src_ctx->mutex);
6760 list_for_each_entry_safe(event, tmp, &src_ctx->event_list,
6762 perf_remove_from_context(event);
6764 list_add(&event->event_entry, &events);
6766 mutex_unlock(&src_ctx->mutex);
6770 mutex_lock(&dst_ctx->mutex);
6771 list_for_each_entry_safe(event, tmp, &events, event_entry) {
6772 list_del(&event->event_entry);
6773 if (event->state >= PERF_EVENT_STATE_OFF)
6774 event->state = PERF_EVENT_STATE_INACTIVE;
6775 perf_install_in_context(dst_ctx, event, dst_cpu);
6778 mutex_unlock(&dst_ctx->mutex);
6780 EXPORT_SYMBOL_GPL(perf_pmu_migrate_context);
6782 static void sync_child_event(struct perf_event *child_event,
6783 struct task_struct *child)
6785 struct perf_event *parent_event = child_event->parent;
6788 if (child_event->attr.inherit_stat)
6789 perf_event_read_event(child_event, child);
6791 child_val = perf_event_count(child_event);
6794 * Add back the child's count to the parent's count:
6796 atomic64_add(child_val, &parent_event->child_count);
6797 atomic64_add(child_event->total_time_enabled,
6798 &parent_event->child_total_time_enabled);
6799 atomic64_add(child_event->total_time_running,
6800 &parent_event->child_total_time_running);
6803 * Remove this event from the parent's list
6805 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
6806 mutex_lock(&parent_event->child_mutex);
6807 list_del_init(&child_event->child_list);
6808 mutex_unlock(&parent_event->child_mutex);
6811 * Release the parent event, if this was the last
6814 put_event(parent_event);
6818 __perf_event_exit_task(struct perf_event *child_event,
6819 struct perf_event_context *child_ctx,
6820 struct task_struct *child)
6822 if (child_event->parent) {
6823 raw_spin_lock_irq(&child_ctx->lock);
6824 perf_group_detach(child_event);
6825 raw_spin_unlock_irq(&child_ctx->lock);
6828 perf_remove_from_context(child_event);
6831 * It can happen that the parent exits first, and has events
6832 * that are still around due to the child reference. These
6833 * events need to be zapped.
6835 if (child_event->parent) {
6836 sync_child_event(child_event, child);
6837 free_event(child_event);
6841 static void perf_event_exit_task_context(struct task_struct *child, int ctxn)
6843 struct perf_event *child_event, *tmp;
6844 struct perf_event_context *child_ctx;
6845 unsigned long flags;
6847 if (likely(!child->perf_event_ctxp[ctxn])) {
6848 perf_event_task(child, NULL, 0);
6852 local_irq_save(flags);
6854 * We can't reschedule here because interrupts are disabled,
6855 * and either child is current or it is a task that can't be
6856 * scheduled, so we are now safe from rescheduling changing
6859 child_ctx = rcu_dereference_raw(child->perf_event_ctxp[ctxn]);
6862 * Take the context lock here so that if find_get_context is
6863 * reading child->perf_event_ctxp, we wait until it has
6864 * incremented the context's refcount before we do put_ctx below.
6866 raw_spin_lock(&child_ctx->lock);
6867 task_ctx_sched_out(child_ctx);
6868 child->perf_event_ctxp[ctxn] = NULL;
6870 * If this context is a clone; unclone it so it can't get
6871 * swapped to another process while we're removing all
6872 * the events from it.
6874 unclone_ctx(child_ctx);
6875 update_context_time(child_ctx);
6876 raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
6879 * Report the task dead after unscheduling the events so that we
6880 * won't get any samples after PERF_RECORD_EXIT. We can however still
6881 * get a few PERF_RECORD_READ events.
6883 perf_event_task(child, child_ctx, 0);
6886 * We can recurse on the same lock type through:
6888 * __perf_event_exit_task()
6889 * sync_child_event()
6891 * mutex_lock(&ctx->mutex)
6893 * But since its the parent context it won't be the same instance.
6895 mutex_lock(&child_ctx->mutex);
6898 list_for_each_entry_safe(child_event, tmp, &child_ctx->pinned_groups,
6900 __perf_event_exit_task(child_event, child_ctx, child);
6902 list_for_each_entry_safe(child_event, tmp, &child_ctx->flexible_groups,
6904 __perf_event_exit_task(child_event, child_ctx, child);
6907 * If the last event was a group event, it will have appended all
6908 * its siblings to the list, but we obtained 'tmp' before that which
6909 * will still point to the list head terminating the iteration.
6911 if (!list_empty(&child_ctx->pinned_groups) ||
6912 !list_empty(&child_ctx->flexible_groups))
6915 mutex_unlock(&child_ctx->mutex);
6921 * When a child task exits, feed back event values to parent events.
6923 void perf_event_exit_task(struct task_struct *child)
6925 struct perf_event *event, *tmp;
6928 mutex_lock(&child->perf_event_mutex);
6929 list_for_each_entry_safe(event, tmp, &child->perf_event_list,
6931 list_del_init(&event->owner_entry);
6934 * Ensure the list deletion is visible before we clear
6935 * the owner, closes a race against perf_release() where
6936 * we need to serialize on the owner->perf_event_mutex.
6939 event->owner = NULL;
6941 mutex_unlock(&child->perf_event_mutex);
6943 for_each_task_context_nr(ctxn)
6944 perf_event_exit_task_context(child, ctxn);
6947 static void perf_free_event(struct perf_event *event,
6948 struct perf_event_context *ctx)
6950 struct perf_event *parent = event->parent;
6952 if (WARN_ON_ONCE(!parent))
6955 mutex_lock(&parent->child_mutex);
6956 list_del_init(&event->child_list);
6957 mutex_unlock(&parent->child_mutex);
6961 perf_group_detach(event);
6962 list_del_event(event, ctx);
6967 * free an unexposed, unused context as created by inheritance by
6968 * perf_event_init_task below, used by fork() in case of fail.
6970 void perf_event_free_task(struct task_struct *task)
6972 struct perf_event_context *ctx;
6973 struct perf_event *event, *tmp;
6976 for_each_task_context_nr(ctxn) {
6977 ctx = task->perf_event_ctxp[ctxn];
6981 mutex_lock(&ctx->mutex);
6983 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups,
6985 perf_free_event(event, ctx);
6987 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups,
6989 perf_free_event(event, ctx);
6991 if (!list_empty(&ctx->pinned_groups) ||
6992 !list_empty(&ctx->flexible_groups))
6995 mutex_unlock(&ctx->mutex);
7001 void perf_event_delayed_put(struct task_struct *task)
7005 for_each_task_context_nr(ctxn)
7006 WARN_ON_ONCE(task->perf_event_ctxp[ctxn]);
7010 * inherit a event from parent task to child task:
7012 static struct perf_event *
7013 inherit_event(struct perf_event *parent_event,
7014 struct task_struct *parent,
7015 struct perf_event_context *parent_ctx,
7016 struct task_struct *child,
7017 struct perf_event *group_leader,
7018 struct perf_event_context *child_ctx)
7020 struct perf_event *child_event;
7021 unsigned long flags;
7024 * Instead of creating recursive hierarchies of events,
7025 * we link inherited events back to the original parent,
7026 * which has a filp for sure, which we use as the reference
7029 if (parent_event->parent)
7030 parent_event = parent_event->parent;
7032 child_event = perf_event_alloc(&parent_event->attr,
7035 group_leader, parent_event,
7037 if (IS_ERR(child_event))
7040 if (!atomic_long_inc_not_zero(&parent_event->refcount)) {
7041 free_event(child_event);
7048 * Make the child state follow the state of the parent event,
7049 * not its attr.disabled bit. We hold the parent's mutex,
7050 * so we won't race with perf_event_{en, dis}able_family.
7052 if (parent_event->state >= PERF_EVENT_STATE_INACTIVE)
7053 child_event->state = PERF_EVENT_STATE_INACTIVE;
7055 child_event->state = PERF_EVENT_STATE_OFF;
7057 if (parent_event->attr.freq) {
7058 u64 sample_period = parent_event->hw.sample_period;
7059 struct hw_perf_event *hwc = &child_event->hw;
7061 hwc->sample_period = sample_period;
7062 hwc->last_period = sample_period;
7064 local64_set(&hwc->period_left, sample_period);
7067 child_event->ctx = child_ctx;
7068 child_event->overflow_handler = parent_event->overflow_handler;
7069 child_event->overflow_handler_context
7070 = parent_event->overflow_handler_context;
7073 * Precalculate sample_data sizes
7075 perf_event__header_size(child_event);
7076 perf_event__id_header_size(child_event);
7079 * Link it up in the child's context:
7081 raw_spin_lock_irqsave(&child_ctx->lock, flags);
7082 add_event_to_ctx(child_event, child_ctx);
7083 raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
7086 * Link this into the parent event's child list
7088 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
7089 mutex_lock(&parent_event->child_mutex);
7090 list_add_tail(&child_event->child_list, &parent_event->child_list);
7091 mutex_unlock(&parent_event->child_mutex);
7096 static int inherit_group(struct perf_event *parent_event,
7097 struct task_struct *parent,
7098 struct perf_event_context *parent_ctx,
7099 struct task_struct *child,
7100 struct perf_event_context *child_ctx)
7102 struct perf_event *leader;
7103 struct perf_event *sub;
7104 struct perf_event *child_ctr;
7106 leader = inherit_event(parent_event, parent, parent_ctx,
7107 child, NULL, child_ctx);
7109 return PTR_ERR(leader);
7110 list_for_each_entry(sub, &parent_event->sibling_list, group_entry) {
7111 child_ctr = inherit_event(sub, parent, parent_ctx,
7112 child, leader, child_ctx);
7113 if (IS_ERR(child_ctr))
7114 return PTR_ERR(child_ctr);
7120 inherit_task_group(struct perf_event *event, struct task_struct *parent,
7121 struct perf_event_context *parent_ctx,
7122 struct task_struct *child, int ctxn,
7126 struct perf_event_context *child_ctx;
7128 if (!event->attr.inherit) {
7133 child_ctx = child->perf_event_ctxp[ctxn];
7136 * This is executed from the parent task context, so
7137 * inherit events that have been marked for cloning.
7138 * First allocate and initialize a context for the
7142 child_ctx = alloc_perf_context(event->pmu, child);
7146 child->perf_event_ctxp[ctxn] = child_ctx;
7149 ret = inherit_group(event, parent, parent_ctx,
7159 * Initialize the perf_event context in task_struct
7161 int perf_event_init_context(struct task_struct *child, int ctxn)
7163 struct perf_event_context *child_ctx, *parent_ctx;
7164 struct perf_event_context *cloned_ctx;
7165 struct perf_event *event;
7166 struct task_struct *parent = current;
7167 int inherited_all = 1;
7168 unsigned long flags;
7171 if (likely(!parent->perf_event_ctxp[ctxn]))
7175 * If the parent's context is a clone, pin it so it won't get
7178 parent_ctx = perf_pin_task_context(parent, ctxn);
7181 * No need to check if parent_ctx != NULL here; since we saw
7182 * it non-NULL earlier, the only reason for it to become NULL
7183 * is if we exit, and since we're currently in the middle of
7184 * a fork we can't be exiting at the same time.
7188 * Lock the parent list. No need to lock the child - not PID
7189 * hashed yet and not running, so nobody can access it.
7191 mutex_lock(&parent_ctx->mutex);
7194 * We dont have to disable NMIs - we are only looking at
7195 * the list, not manipulating it:
7197 list_for_each_entry(event, &parent_ctx->pinned_groups, group_entry) {
7198 ret = inherit_task_group(event, parent, parent_ctx,
7199 child, ctxn, &inherited_all);
7205 * We can't hold ctx->lock when iterating the ->flexible_group list due
7206 * to allocations, but we need to prevent rotation because
7207 * rotate_ctx() will change the list from interrupt context.
7209 raw_spin_lock_irqsave(&parent_ctx->lock, flags);
7210 parent_ctx->rotate_disable = 1;
7211 raw_spin_unlock_irqrestore(&parent_ctx->lock, flags);
7213 list_for_each_entry(event, &parent_ctx->flexible_groups, group_entry) {
7214 ret = inherit_task_group(event, parent, parent_ctx,
7215 child, ctxn, &inherited_all);
7220 raw_spin_lock_irqsave(&parent_ctx->lock, flags);
7221 parent_ctx->rotate_disable = 0;
7223 child_ctx = child->perf_event_ctxp[ctxn];
7225 if (child_ctx && inherited_all) {
7227 * Mark the child context as a clone of the parent
7228 * context, or of whatever the parent is a clone of.
7230 * Note that if the parent is a clone, the holding of
7231 * parent_ctx->lock avoids it from being uncloned.
7233 cloned_ctx = parent_ctx->parent_ctx;
7235 child_ctx->parent_ctx = cloned_ctx;
7236 child_ctx->parent_gen = parent_ctx->parent_gen;
7238 child_ctx->parent_ctx = parent_ctx;
7239 child_ctx->parent_gen = parent_ctx->generation;
7241 get_ctx(child_ctx->parent_ctx);
7244 raw_spin_unlock_irqrestore(&parent_ctx->lock, flags);
7245 mutex_unlock(&parent_ctx->mutex);
7247 perf_unpin_context(parent_ctx);
7248 put_ctx(parent_ctx);
7254 * Initialize the perf_event context in task_struct
7256 int perf_event_init_task(struct task_struct *child)
7260 memset(child->perf_event_ctxp, 0, sizeof(child->perf_event_ctxp));
7261 mutex_init(&child->perf_event_mutex);
7262 INIT_LIST_HEAD(&child->perf_event_list);
7264 for_each_task_context_nr(ctxn) {
7265 ret = perf_event_init_context(child, ctxn);
7273 static void __init perf_event_init_all_cpus(void)
7275 struct swevent_htable *swhash;
7278 for_each_possible_cpu(cpu) {
7279 swhash = &per_cpu(swevent_htable, cpu);
7280 mutex_init(&swhash->hlist_mutex);
7281 INIT_LIST_HEAD(&per_cpu(rotation_list, cpu));
7285 static void __cpuinit perf_event_init_cpu(int cpu)
7287 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
7289 mutex_lock(&swhash->hlist_mutex);
7290 if (swhash->hlist_refcount > 0) {
7291 struct swevent_hlist *hlist;
7293 hlist = kzalloc_node(sizeof(*hlist), GFP_KERNEL, cpu_to_node(cpu));
7295 rcu_assign_pointer(swhash->swevent_hlist, hlist);
7297 mutex_unlock(&swhash->hlist_mutex);
7300 #if defined CONFIG_HOTPLUG_CPU || defined CONFIG_KEXEC
7301 static void perf_pmu_rotate_stop(struct pmu *pmu)
7303 struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
7305 WARN_ON(!irqs_disabled());
7307 list_del_init(&cpuctx->rotation_list);
7310 static void __perf_event_exit_context(void *__info)
7312 struct perf_event_context *ctx = __info;
7313 struct perf_event *event, *tmp;
7315 perf_pmu_rotate_stop(ctx->pmu);
7317 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry)
7318 __perf_remove_from_context(event);
7319 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups, group_entry)
7320 __perf_remove_from_context(event);
7323 static void perf_event_exit_cpu_context(int cpu)
7325 struct perf_event_context *ctx;
7329 idx = srcu_read_lock(&pmus_srcu);
7330 list_for_each_entry_rcu(pmu, &pmus, entry) {
7331 ctx = &per_cpu_ptr(pmu->pmu_cpu_context, cpu)->ctx;
7333 mutex_lock(&ctx->mutex);
7334 smp_call_function_single(cpu, __perf_event_exit_context, ctx, 1);
7335 mutex_unlock(&ctx->mutex);
7337 srcu_read_unlock(&pmus_srcu, idx);
7340 static void perf_event_exit_cpu(int cpu)
7342 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
7344 mutex_lock(&swhash->hlist_mutex);
7345 swevent_hlist_release(swhash);
7346 mutex_unlock(&swhash->hlist_mutex);
7348 perf_event_exit_cpu_context(cpu);
7351 static inline void perf_event_exit_cpu(int cpu) { }
7355 perf_reboot(struct notifier_block *notifier, unsigned long val, void *v)
7359 for_each_online_cpu(cpu)
7360 perf_event_exit_cpu(cpu);
7366 * Run the perf reboot notifier at the very last possible moment so that
7367 * the generic watchdog code runs as long as possible.
7369 static struct notifier_block perf_reboot_notifier = {
7370 .notifier_call = perf_reboot,
7371 .priority = INT_MIN,
7374 static int __cpuinit
7375 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
7377 unsigned int cpu = (long)hcpu;
7379 switch (action & ~CPU_TASKS_FROZEN) {
7381 case CPU_UP_PREPARE:
7382 case CPU_DOWN_FAILED:
7383 perf_event_init_cpu(cpu);
7386 case CPU_UP_CANCELED:
7387 case CPU_DOWN_PREPARE:
7388 perf_event_exit_cpu(cpu);
7398 void __init perf_event_init(void)
7404 perf_event_init_all_cpus();
7405 init_srcu_struct(&pmus_srcu);
7406 perf_pmu_register(&perf_swevent, "software", PERF_TYPE_SOFTWARE);
7407 perf_pmu_register(&perf_cpu_clock, NULL, -1);
7408 perf_pmu_register(&perf_task_clock, NULL, -1);
7410 perf_cpu_notifier(perf_cpu_notify);
7411 register_reboot_notifier(&perf_reboot_notifier);
7413 ret = init_hw_breakpoint();
7414 WARN(ret, "hw_breakpoint initialization failed with: %d", ret);
7416 /* do not patch jump label more than once per second */
7417 jump_label_rate_limit(&perf_sched_events, HZ);
7420 * Build time assertion that we keep the data_head at the intended
7421 * location. IOW, validation we got the __reserved[] size right.
7423 BUILD_BUG_ON((offsetof(struct perf_event_mmap_page, data_head))
7427 static int __init perf_event_sysfs_init(void)
7432 mutex_lock(&pmus_lock);
7434 ret = bus_register(&pmu_bus);
7438 list_for_each_entry(pmu, &pmus, entry) {
7439 if (!pmu->name || pmu->type < 0)
7442 ret = pmu_dev_alloc(pmu);
7443 WARN(ret, "Failed to register pmu: %s, reason %d\n", pmu->name, ret);
7445 pmu_bus_running = 1;
7449 mutex_unlock(&pmus_lock);
7453 device_initcall(perf_event_sysfs_init);
7455 #ifdef CONFIG_CGROUP_PERF
7456 static struct cgroup_subsys_state *perf_cgroup_css_alloc(struct cgroup *cont)
7458 struct perf_cgroup *jc;
7460 jc = kzalloc(sizeof(*jc), GFP_KERNEL);
7462 return ERR_PTR(-ENOMEM);
7464 jc->info = alloc_percpu(struct perf_cgroup_info);
7467 return ERR_PTR(-ENOMEM);
7473 static void perf_cgroup_css_free(struct cgroup *cont)
7475 struct perf_cgroup *jc;
7476 jc = container_of(cgroup_subsys_state(cont, perf_subsys_id),
7477 struct perf_cgroup, css);
7478 free_percpu(jc->info);
7482 static int __perf_cgroup_move(void *info)
7484 struct task_struct *task = info;
7485 perf_cgroup_switch(task, PERF_CGROUP_SWOUT | PERF_CGROUP_SWIN);
7489 static void perf_cgroup_attach(struct cgroup *cgrp, struct cgroup_taskset *tset)
7491 struct task_struct *task;
7493 cgroup_taskset_for_each(task, cgrp, tset)
7494 task_function_call(task, __perf_cgroup_move, task);
7497 static void perf_cgroup_exit(struct cgroup *cgrp, struct cgroup *old_cgrp,
7498 struct task_struct *task)
7501 * cgroup_exit() is called in the copy_process() failure path.
7502 * Ignore this case since the task hasn't ran yet, this avoids
7503 * trying to poke a half freed task state from generic code.
7505 if (!(task->flags & PF_EXITING))
7508 task_function_call(task, __perf_cgroup_move, task);
7511 struct cgroup_subsys perf_subsys = {
7512 .name = "perf_event",
7513 .subsys_id = perf_subsys_id,
7514 .css_alloc = perf_cgroup_css_alloc,
7515 .css_free = perf_cgroup_css_free,
7516 .exit = perf_cgroup_exit,
7517 .attach = perf_cgroup_attach,
7520 * perf_event cgroup doesn't handle nesting correctly.
7521 * ctx->nr_cgroups adjustments should be propagated through the
7522 * cgroup hierarchy. Fix it and remove the following.
7524 .broken_hierarchy = true,
7526 #endif /* CONFIG_CGROUP_PERF */