2 * Performance events core code:
4 * Copyright (C) 2008 Thomas Gleixner <tglx@linutronix.de>
5 * Copyright (C) 2008-2011 Red Hat, Inc., Ingo Molnar
6 * Copyright (C) 2008-2011 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
7 * Copyright © 2009 Paul Mackerras, IBM Corp. <paulus@au1.ibm.com>
9 * For licensing details see kernel-base/COPYING
14 #include <linux/cpu.h>
15 #include <linux/smp.h>
16 #include <linux/idr.h>
17 #include <linux/file.h>
18 #include <linux/poll.h>
19 #include <linux/slab.h>
20 #include <linux/hash.h>
21 #include <linux/tick.h>
22 #include <linux/sysfs.h>
23 #include <linux/dcache.h>
24 #include <linux/percpu.h>
25 #include <linux/ptrace.h>
26 #include <linux/reboot.h>
27 #include <linux/vmstat.h>
28 #include <linux/device.h>
29 #include <linux/export.h>
30 #include <linux/vmalloc.h>
31 #include <linux/hardirq.h>
32 #include <linux/rculist.h>
33 #include <linux/uaccess.h>
34 #include <linux/syscalls.h>
35 #include <linux/anon_inodes.h>
36 #include <linux/kernel_stat.h>
37 #include <linux/perf_event.h>
38 #include <linux/ftrace_event.h>
39 #include <linux/hw_breakpoint.h>
40 #include <linux/mm_types.h>
41 #include <linux/cgroup.h>
45 #include <asm/irq_regs.h>
47 struct remote_function_call {
48 struct task_struct *p;
49 int (*func)(void *info);
54 static void remote_function(void *data)
56 struct remote_function_call *tfc = data;
57 struct task_struct *p = tfc->p;
61 if (task_cpu(p) != smp_processor_id() || !task_curr(p))
65 tfc->ret = tfc->func(tfc->info);
69 * task_function_call - call a function on the cpu on which a task runs
70 * @p: the task to evaluate
71 * @func: the function to be called
72 * @info: the function call argument
74 * Calls the function @func when the task is currently running. This might
75 * be on the current CPU, which just calls the function directly
77 * returns: @func return value, or
78 * -ESRCH - when the process isn't running
79 * -EAGAIN - when the process moved away
82 task_function_call(struct task_struct *p, int (*func) (void *info), void *info)
84 struct remote_function_call data = {
88 .ret = -ESRCH, /* No such (running) process */
92 smp_call_function_single(task_cpu(p), remote_function, &data, 1);
98 * cpu_function_call - call a function on the cpu
99 * @func: the function to be called
100 * @info: the function call argument
102 * Calls the function @func on the remote cpu.
104 * returns: @func return value or -ENXIO when the cpu is offline
106 static int cpu_function_call(int cpu, int (*func) (void *info), void *info)
108 struct remote_function_call data = {
112 .ret = -ENXIO, /* No such CPU */
115 smp_call_function_single(cpu, remote_function, &data, 1);
120 #define PERF_FLAG_ALL (PERF_FLAG_FD_NO_GROUP |\
121 PERF_FLAG_FD_OUTPUT |\
122 PERF_FLAG_PID_CGROUP |\
123 PERF_FLAG_FD_CLOEXEC)
126 * branch priv levels that need permission checks
128 #define PERF_SAMPLE_BRANCH_PERM_PLM \
129 (PERF_SAMPLE_BRANCH_KERNEL |\
130 PERF_SAMPLE_BRANCH_HV)
133 EVENT_FLEXIBLE = 0x1,
135 EVENT_ALL = EVENT_FLEXIBLE | EVENT_PINNED,
139 * perf_sched_events : >0 events exist
140 * perf_cgroup_events: >0 per-cpu cgroup events exist on this cpu
142 struct static_key_deferred perf_sched_events __read_mostly;
143 static DEFINE_PER_CPU(atomic_t, perf_cgroup_events);
144 static DEFINE_PER_CPU(atomic_t, perf_branch_stack_events);
146 static atomic_t nr_mmap_events __read_mostly;
147 static atomic_t nr_comm_events __read_mostly;
148 static atomic_t nr_task_events __read_mostly;
149 static atomic_t nr_freq_events __read_mostly;
151 static LIST_HEAD(pmus);
152 static DEFINE_MUTEX(pmus_lock);
153 static struct srcu_struct pmus_srcu;
156 * perf event paranoia level:
157 * -1 - not paranoid at all
158 * 0 - disallow raw tracepoint access for unpriv
159 * 1 - disallow cpu events for unpriv
160 * 2 - disallow kernel profiling for unpriv
162 int sysctl_perf_event_paranoid __read_mostly = 1;
164 /* Minimum for 512 kiB + 1 user control page */
165 int sysctl_perf_event_mlock __read_mostly = 512 + (PAGE_SIZE / 1024); /* 'free' kiB per user */
168 * max perf event sample rate
170 #define DEFAULT_MAX_SAMPLE_RATE 100000
171 #define DEFAULT_SAMPLE_PERIOD_NS (NSEC_PER_SEC / DEFAULT_MAX_SAMPLE_RATE)
172 #define DEFAULT_CPU_TIME_MAX_PERCENT 25
174 int sysctl_perf_event_sample_rate __read_mostly = DEFAULT_MAX_SAMPLE_RATE;
176 static int max_samples_per_tick __read_mostly = DIV_ROUND_UP(DEFAULT_MAX_SAMPLE_RATE, HZ);
177 static int perf_sample_period_ns __read_mostly = DEFAULT_SAMPLE_PERIOD_NS;
179 static int perf_sample_allowed_ns __read_mostly =
180 DEFAULT_SAMPLE_PERIOD_NS * DEFAULT_CPU_TIME_MAX_PERCENT / 100;
182 void update_perf_cpu_limits(void)
184 u64 tmp = perf_sample_period_ns;
186 tmp *= sysctl_perf_cpu_time_max_percent;
188 ACCESS_ONCE(perf_sample_allowed_ns) = tmp;
191 static int perf_rotate_context(struct perf_cpu_context *cpuctx);
193 int perf_proc_update_handler(struct ctl_table *table, int write,
194 void __user *buffer, size_t *lenp,
197 int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
202 max_samples_per_tick = DIV_ROUND_UP(sysctl_perf_event_sample_rate, HZ);
203 perf_sample_period_ns = NSEC_PER_SEC / sysctl_perf_event_sample_rate;
204 update_perf_cpu_limits();
209 int sysctl_perf_cpu_time_max_percent __read_mostly = DEFAULT_CPU_TIME_MAX_PERCENT;
211 int perf_cpu_time_max_percent_handler(struct ctl_table *table, int write,
212 void __user *buffer, size_t *lenp,
215 int ret = proc_dointvec(table, write, buffer, lenp, ppos);
220 update_perf_cpu_limits();
226 * perf samples are done in some very critical code paths (NMIs).
227 * If they take too much CPU time, the system can lock up and not
228 * get any real work done. This will drop the sample rate when
229 * we detect that events are taking too long.
231 #define NR_ACCUMULATED_SAMPLES 128
232 static DEFINE_PER_CPU(u64, running_sample_length);
234 void perf_sample_event_took(u64 sample_len_ns)
236 u64 avg_local_sample_len;
237 u64 local_samples_len;
238 u64 allowed_ns = ACCESS_ONCE(perf_sample_allowed_ns);
243 /* decay the counter by 1 average sample */
244 local_samples_len = __get_cpu_var(running_sample_length);
245 local_samples_len -= local_samples_len/NR_ACCUMULATED_SAMPLES;
246 local_samples_len += sample_len_ns;
247 __get_cpu_var(running_sample_length) = local_samples_len;
250 * note: this will be biased artifically low until we have
251 * seen NR_ACCUMULATED_SAMPLES. Doing it this way keeps us
252 * from having to maintain a count.
254 avg_local_sample_len = local_samples_len/NR_ACCUMULATED_SAMPLES;
256 if (avg_local_sample_len <= allowed_ns)
259 if (max_samples_per_tick <= 1)
262 max_samples_per_tick = DIV_ROUND_UP(max_samples_per_tick, 2);
263 sysctl_perf_event_sample_rate = max_samples_per_tick * HZ;
264 perf_sample_period_ns = NSEC_PER_SEC / sysctl_perf_event_sample_rate;
266 printk_ratelimited(KERN_WARNING
267 "perf samples too long (%lld > %lld), lowering "
268 "kernel.perf_event_max_sample_rate to %d\n",
269 avg_local_sample_len, allowed_ns,
270 sysctl_perf_event_sample_rate);
272 update_perf_cpu_limits();
275 static atomic64_t perf_event_id;
277 static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
278 enum event_type_t event_type);
280 static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
281 enum event_type_t event_type,
282 struct task_struct *task);
284 static void update_context_time(struct perf_event_context *ctx);
285 static u64 perf_event_time(struct perf_event *event);
287 void __weak perf_event_print_debug(void) { }
289 extern __weak const char *perf_pmu_name(void)
294 static inline u64 perf_clock(void)
296 return local_clock();
299 static inline struct perf_cpu_context *
300 __get_cpu_context(struct perf_event_context *ctx)
302 return this_cpu_ptr(ctx->pmu->pmu_cpu_context);
305 static void perf_ctx_lock(struct perf_cpu_context *cpuctx,
306 struct perf_event_context *ctx)
308 raw_spin_lock(&cpuctx->ctx.lock);
310 raw_spin_lock(&ctx->lock);
313 static void perf_ctx_unlock(struct perf_cpu_context *cpuctx,
314 struct perf_event_context *ctx)
317 raw_spin_unlock(&ctx->lock);
318 raw_spin_unlock(&cpuctx->ctx.lock);
321 #ifdef CONFIG_CGROUP_PERF
324 * perf_cgroup_info keeps track of time_enabled for a cgroup.
325 * This is a per-cpu dynamically allocated data structure.
327 struct perf_cgroup_info {
333 struct cgroup_subsys_state css;
334 struct perf_cgroup_info __percpu *info;
338 * Must ensure cgroup is pinned (css_get) before calling
339 * this function. In other words, we cannot call this function
340 * if there is no cgroup event for the current CPU context.
342 static inline struct perf_cgroup *
343 perf_cgroup_from_task(struct task_struct *task)
345 return container_of(task_css(task, perf_event_cgrp_id),
346 struct perf_cgroup, css);
350 perf_cgroup_match(struct perf_event *event)
352 struct perf_event_context *ctx = event->ctx;
353 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
355 /* @event doesn't care about cgroup */
359 /* wants specific cgroup scope but @cpuctx isn't associated with any */
364 * Cgroup scoping is recursive. An event enabled for a cgroup is
365 * also enabled for all its descendant cgroups. If @cpuctx's
366 * cgroup is a descendant of @event's (the test covers identity
367 * case), it's a match.
369 return cgroup_is_descendant(cpuctx->cgrp->css.cgroup,
370 event->cgrp->css.cgroup);
373 static inline void perf_put_cgroup(struct perf_event *event)
375 css_put(&event->cgrp->css);
378 static inline void perf_detach_cgroup(struct perf_event *event)
380 perf_put_cgroup(event);
384 static inline int is_cgroup_event(struct perf_event *event)
386 return event->cgrp != NULL;
389 static inline u64 perf_cgroup_event_time(struct perf_event *event)
391 struct perf_cgroup_info *t;
393 t = per_cpu_ptr(event->cgrp->info, event->cpu);
397 static inline void __update_cgrp_time(struct perf_cgroup *cgrp)
399 struct perf_cgroup_info *info;
404 info = this_cpu_ptr(cgrp->info);
406 info->time += now - info->timestamp;
407 info->timestamp = now;
410 static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx)
412 struct perf_cgroup *cgrp_out = cpuctx->cgrp;
414 __update_cgrp_time(cgrp_out);
417 static inline void update_cgrp_time_from_event(struct perf_event *event)
419 struct perf_cgroup *cgrp;
422 * ensure we access cgroup data only when needed and
423 * when we know the cgroup is pinned (css_get)
425 if (!is_cgroup_event(event))
428 cgrp = perf_cgroup_from_task(current);
430 * Do not update time when cgroup is not active
432 if (cgrp == event->cgrp)
433 __update_cgrp_time(event->cgrp);
437 perf_cgroup_set_timestamp(struct task_struct *task,
438 struct perf_event_context *ctx)
440 struct perf_cgroup *cgrp;
441 struct perf_cgroup_info *info;
444 * ctx->lock held by caller
445 * ensure we do not access cgroup data
446 * unless we have the cgroup pinned (css_get)
448 if (!task || !ctx->nr_cgroups)
451 cgrp = perf_cgroup_from_task(task);
452 info = this_cpu_ptr(cgrp->info);
453 info->timestamp = ctx->timestamp;
456 #define PERF_CGROUP_SWOUT 0x1 /* cgroup switch out every event */
457 #define PERF_CGROUP_SWIN 0x2 /* cgroup switch in events based on task */
460 * reschedule events based on the cgroup constraint of task.
462 * mode SWOUT : schedule out everything
463 * mode SWIN : schedule in based on cgroup for next
465 void perf_cgroup_switch(struct task_struct *task, int mode)
467 struct perf_cpu_context *cpuctx;
472 * disable interrupts to avoid geting nr_cgroup
473 * changes via __perf_event_disable(). Also
476 local_irq_save(flags);
479 * we reschedule only in the presence of cgroup
480 * constrained events.
484 list_for_each_entry_rcu(pmu, &pmus, entry) {
485 cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
486 if (cpuctx->unique_pmu != pmu)
487 continue; /* ensure we process each cpuctx once */
490 * perf_cgroup_events says at least one
491 * context on this CPU has cgroup events.
493 * ctx->nr_cgroups reports the number of cgroup
494 * events for a context.
496 if (cpuctx->ctx.nr_cgroups > 0) {
497 perf_ctx_lock(cpuctx, cpuctx->task_ctx);
498 perf_pmu_disable(cpuctx->ctx.pmu);
500 if (mode & PERF_CGROUP_SWOUT) {
501 cpu_ctx_sched_out(cpuctx, EVENT_ALL);
503 * must not be done before ctxswout due
504 * to event_filter_match() in event_sched_out()
509 if (mode & PERF_CGROUP_SWIN) {
510 WARN_ON_ONCE(cpuctx->cgrp);
512 * set cgrp before ctxsw in to allow
513 * event_filter_match() to not have to pass
516 cpuctx->cgrp = perf_cgroup_from_task(task);
517 cpu_ctx_sched_in(cpuctx, EVENT_ALL, task);
519 perf_pmu_enable(cpuctx->ctx.pmu);
520 perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
526 local_irq_restore(flags);
529 static inline void perf_cgroup_sched_out(struct task_struct *task,
530 struct task_struct *next)
532 struct perf_cgroup *cgrp1;
533 struct perf_cgroup *cgrp2 = NULL;
536 * we come here when we know perf_cgroup_events > 0
538 cgrp1 = perf_cgroup_from_task(task);
541 * next is NULL when called from perf_event_enable_on_exec()
542 * that will systematically cause a cgroup_switch()
545 cgrp2 = perf_cgroup_from_task(next);
548 * only schedule out current cgroup events if we know
549 * that we are switching to a different cgroup. Otherwise,
550 * do no touch the cgroup events.
553 perf_cgroup_switch(task, PERF_CGROUP_SWOUT);
556 static inline void perf_cgroup_sched_in(struct task_struct *prev,
557 struct task_struct *task)
559 struct perf_cgroup *cgrp1;
560 struct perf_cgroup *cgrp2 = NULL;
563 * we come here when we know perf_cgroup_events > 0
565 cgrp1 = perf_cgroup_from_task(task);
567 /* prev can never be NULL */
568 cgrp2 = perf_cgroup_from_task(prev);
571 * only need to schedule in cgroup events if we are changing
572 * cgroup during ctxsw. Cgroup events were not scheduled
573 * out of ctxsw out if that was not the case.
576 perf_cgroup_switch(task, PERF_CGROUP_SWIN);
579 static inline int perf_cgroup_connect(int fd, struct perf_event *event,
580 struct perf_event_attr *attr,
581 struct perf_event *group_leader)
583 struct perf_cgroup *cgrp;
584 struct cgroup_subsys_state *css;
585 struct fd f = fdget(fd);
591 css = css_tryget_from_dir(f.file->f_dentry, &perf_event_cgrp_subsys);
597 cgrp = container_of(css, struct perf_cgroup, css);
601 * all events in a group must monitor
602 * the same cgroup because a task belongs
603 * to only one perf cgroup at a time
605 if (group_leader && group_leader->cgrp != cgrp) {
606 perf_detach_cgroup(event);
615 perf_cgroup_set_shadow_time(struct perf_event *event, u64 now)
617 struct perf_cgroup_info *t;
618 t = per_cpu_ptr(event->cgrp->info, event->cpu);
619 event->shadow_ctx_time = now - t->timestamp;
623 perf_cgroup_defer_enabled(struct perf_event *event)
626 * when the current task's perf cgroup does not match
627 * the event's, we need to remember to call the
628 * perf_mark_enable() function the first time a task with
629 * a matching perf cgroup is scheduled in.
631 if (is_cgroup_event(event) && !perf_cgroup_match(event))
632 event->cgrp_defer_enabled = 1;
636 perf_cgroup_mark_enabled(struct perf_event *event,
637 struct perf_event_context *ctx)
639 struct perf_event *sub;
640 u64 tstamp = perf_event_time(event);
642 if (!event->cgrp_defer_enabled)
645 event->cgrp_defer_enabled = 0;
647 event->tstamp_enabled = tstamp - event->total_time_enabled;
648 list_for_each_entry(sub, &event->sibling_list, group_entry) {
649 if (sub->state >= PERF_EVENT_STATE_INACTIVE) {
650 sub->tstamp_enabled = tstamp - sub->total_time_enabled;
651 sub->cgrp_defer_enabled = 0;
655 #else /* !CONFIG_CGROUP_PERF */
658 perf_cgroup_match(struct perf_event *event)
663 static inline void perf_detach_cgroup(struct perf_event *event)
666 static inline int is_cgroup_event(struct perf_event *event)
671 static inline u64 perf_cgroup_event_cgrp_time(struct perf_event *event)
676 static inline void update_cgrp_time_from_event(struct perf_event *event)
680 static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx)
684 static inline void perf_cgroup_sched_out(struct task_struct *task,
685 struct task_struct *next)
689 static inline void perf_cgroup_sched_in(struct task_struct *prev,
690 struct task_struct *task)
694 static inline int perf_cgroup_connect(pid_t pid, struct perf_event *event,
695 struct perf_event_attr *attr,
696 struct perf_event *group_leader)
702 perf_cgroup_set_timestamp(struct task_struct *task,
703 struct perf_event_context *ctx)
708 perf_cgroup_switch(struct task_struct *task, struct task_struct *next)
713 perf_cgroup_set_shadow_time(struct perf_event *event, u64 now)
717 static inline u64 perf_cgroup_event_time(struct perf_event *event)
723 perf_cgroup_defer_enabled(struct perf_event *event)
728 perf_cgroup_mark_enabled(struct perf_event *event,
729 struct perf_event_context *ctx)
735 * set default to be dependent on timer tick just
738 #define PERF_CPU_HRTIMER (1000 / HZ)
740 * function must be called with interrupts disbled
742 static enum hrtimer_restart perf_cpu_hrtimer_handler(struct hrtimer *hr)
744 struct perf_cpu_context *cpuctx;
745 enum hrtimer_restart ret = HRTIMER_NORESTART;
748 WARN_ON(!irqs_disabled());
750 cpuctx = container_of(hr, struct perf_cpu_context, hrtimer);
752 rotations = perf_rotate_context(cpuctx);
755 * arm timer if needed
758 hrtimer_forward_now(hr, cpuctx->hrtimer_interval);
759 ret = HRTIMER_RESTART;
765 /* CPU is going down */
766 void perf_cpu_hrtimer_cancel(int cpu)
768 struct perf_cpu_context *cpuctx;
772 if (WARN_ON(cpu != smp_processor_id()))
775 local_irq_save(flags);
779 list_for_each_entry_rcu(pmu, &pmus, entry) {
780 cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
782 if (pmu->task_ctx_nr == perf_sw_context)
785 hrtimer_cancel(&cpuctx->hrtimer);
790 local_irq_restore(flags);
793 static void __perf_cpu_hrtimer_init(struct perf_cpu_context *cpuctx, int cpu)
795 struct hrtimer *hr = &cpuctx->hrtimer;
796 struct pmu *pmu = cpuctx->ctx.pmu;
799 /* no multiplexing needed for SW PMU */
800 if (pmu->task_ctx_nr == perf_sw_context)
804 * check default is sane, if not set then force to
805 * default interval (1/tick)
807 timer = pmu->hrtimer_interval_ms;
809 timer = pmu->hrtimer_interval_ms = PERF_CPU_HRTIMER;
811 cpuctx->hrtimer_interval = ns_to_ktime(NSEC_PER_MSEC * timer);
813 hrtimer_init(hr, CLOCK_MONOTONIC, HRTIMER_MODE_REL_PINNED);
814 hr->function = perf_cpu_hrtimer_handler;
817 static void perf_cpu_hrtimer_restart(struct perf_cpu_context *cpuctx)
819 struct hrtimer *hr = &cpuctx->hrtimer;
820 struct pmu *pmu = cpuctx->ctx.pmu;
823 if (pmu->task_ctx_nr == perf_sw_context)
826 if (hrtimer_active(hr))
829 if (!hrtimer_callback_running(hr))
830 __hrtimer_start_range_ns(hr, cpuctx->hrtimer_interval,
831 0, HRTIMER_MODE_REL_PINNED, 0);
834 void perf_pmu_disable(struct pmu *pmu)
836 int *count = this_cpu_ptr(pmu->pmu_disable_count);
838 pmu->pmu_disable(pmu);
841 void perf_pmu_enable(struct pmu *pmu)
843 int *count = this_cpu_ptr(pmu->pmu_disable_count);
845 pmu->pmu_enable(pmu);
848 static DEFINE_PER_CPU(struct list_head, rotation_list);
851 * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
852 * because they're strictly cpu affine and rotate_start is called with IRQs
853 * disabled, while rotate_context is called from IRQ context.
855 static void perf_pmu_rotate_start(struct pmu *pmu)
857 struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
858 struct list_head *head = &__get_cpu_var(rotation_list);
860 WARN_ON(!irqs_disabled());
862 if (list_empty(&cpuctx->rotation_list))
863 list_add(&cpuctx->rotation_list, head);
866 static void get_ctx(struct perf_event_context *ctx)
868 WARN_ON(!atomic_inc_not_zero(&ctx->refcount));
871 static void put_ctx(struct perf_event_context *ctx)
873 if (atomic_dec_and_test(&ctx->refcount)) {
875 put_ctx(ctx->parent_ctx);
877 put_task_struct(ctx->task);
878 kfree_rcu(ctx, rcu_head);
882 static void unclone_ctx(struct perf_event_context *ctx)
884 if (ctx->parent_ctx) {
885 put_ctx(ctx->parent_ctx);
886 ctx->parent_ctx = NULL;
891 static u32 perf_event_pid(struct perf_event *event, struct task_struct *p)
894 * only top level events have the pid namespace they were created in
897 event = event->parent;
899 return task_tgid_nr_ns(p, event->ns);
902 static u32 perf_event_tid(struct perf_event *event, struct task_struct *p)
905 * only top level events have the pid namespace they were created in
908 event = event->parent;
910 return task_pid_nr_ns(p, event->ns);
914 * If we inherit events we want to return the parent event id
917 static u64 primary_event_id(struct perf_event *event)
922 id = event->parent->id;
928 * Get the perf_event_context for a task and lock it.
929 * This has to cope with with the fact that until it is locked,
930 * the context could get moved to another task.
932 static struct perf_event_context *
933 perf_lock_task_context(struct task_struct *task, int ctxn, unsigned long *flags)
935 struct perf_event_context *ctx;
939 * One of the few rules of preemptible RCU is that one cannot do
940 * rcu_read_unlock() while holding a scheduler (or nested) lock when
941 * part of the read side critical section was preemptible -- see
942 * rcu_read_unlock_special().
944 * Since ctx->lock nests under rq->lock we must ensure the entire read
945 * side critical section is non-preemptible.
949 ctx = rcu_dereference(task->perf_event_ctxp[ctxn]);
952 * If this context is a clone of another, it might
953 * get swapped for another underneath us by
954 * perf_event_task_sched_out, though the
955 * rcu_read_lock() protects us from any context
956 * getting freed. Lock the context and check if it
957 * got swapped before we could get the lock, and retry
958 * if so. If we locked the right context, then it
959 * can't get swapped on us any more.
961 raw_spin_lock_irqsave(&ctx->lock, *flags);
962 if (ctx != rcu_dereference(task->perf_event_ctxp[ctxn])) {
963 raw_spin_unlock_irqrestore(&ctx->lock, *flags);
969 if (!atomic_inc_not_zero(&ctx->refcount)) {
970 raw_spin_unlock_irqrestore(&ctx->lock, *flags);
980 * Get the context for a task and increment its pin_count so it
981 * can't get swapped to another task. This also increments its
982 * reference count so that the context can't get freed.
984 static struct perf_event_context *
985 perf_pin_task_context(struct task_struct *task, int ctxn)
987 struct perf_event_context *ctx;
990 ctx = perf_lock_task_context(task, ctxn, &flags);
993 raw_spin_unlock_irqrestore(&ctx->lock, flags);
998 static void perf_unpin_context(struct perf_event_context *ctx)
1000 unsigned long flags;
1002 raw_spin_lock_irqsave(&ctx->lock, flags);
1004 raw_spin_unlock_irqrestore(&ctx->lock, flags);
1008 * Update the record of the current time in a context.
1010 static void update_context_time(struct perf_event_context *ctx)
1012 u64 now = perf_clock();
1014 ctx->time += now - ctx->timestamp;
1015 ctx->timestamp = now;
1018 static u64 perf_event_time(struct perf_event *event)
1020 struct perf_event_context *ctx = event->ctx;
1022 if (is_cgroup_event(event))
1023 return perf_cgroup_event_time(event);
1025 return ctx ? ctx->time : 0;
1029 * Update the total_time_enabled and total_time_running fields for a event.
1030 * The caller of this function needs to hold the ctx->lock.
1032 static void update_event_times(struct perf_event *event)
1034 struct perf_event_context *ctx = event->ctx;
1037 if (event->state < PERF_EVENT_STATE_INACTIVE ||
1038 event->group_leader->state < PERF_EVENT_STATE_INACTIVE)
1041 * in cgroup mode, time_enabled represents
1042 * the time the event was enabled AND active
1043 * tasks were in the monitored cgroup. This is
1044 * independent of the activity of the context as
1045 * there may be a mix of cgroup and non-cgroup events.
1047 * That is why we treat cgroup events differently
1050 if (is_cgroup_event(event))
1051 run_end = perf_cgroup_event_time(event);
1052 else if (ctx->is_active)
1053 run_end = ctx->time;
1055 run_end = event->tstamp_stopped;
1057 event->total_time_enabled = run_end - event->tstamp_enabled;
1059 if (event->state == PERF_EVENT_STATE_INACTIVE)
1060 run_end = event->tstamp_stopped;
1062 run_end = perf_event_time(event);
1064 event->total_time_running = run_end - event->tstamp_running;
1069 * Update total_time_enabled and total_time_running for all events in a group.
1071 static void update_group_times(struct perf_event *leader)
1073 struct perf_event *event;
1075 update_event_times(leader);
1076 list_for_each_entry(event, &leader->sibling_list, group_entry)
1077 update_event_times(event);
1080 static struct list_head *
1081 ctx_group_list(struct perf_event *event, struct perf_event_context *ctx)
1083 if (event->attr.pinned)
1084 return &ctx->pinned_groups;
1086 return &ctx->flexible_groups;
1090 * Add a event from the lists for its context.
1091 * Must be called with ctx->mutex and ctx->lock held.
1094 list_add_event(struct perf_event *event, struct perf_event_context *ctx)
1096 WARN_ON_ONCE(event->attach_state & PERF_ATTACH_CONTEXT);
1097 event->attach_state |= PERF_ATTACH_CONTEXT;
1100 * If we're a stand alone event or group leader, we go to the context
1101 * list, group events are kept attached to the group so that
1102 * perf_group_detach can, at all times, locate all siblings.
1104 if (event->group_leader == event) {
1105 struct list_head *list;
1107 if (is_software_event(event))
1108 event->group_flags |= PERF_GROUP_SOFTWARE;
1110 list = ctx_group_list(event, ctx);
1111 list_add_tail(&event->group_entry, list);
1114 if (is_cgroup_event(event))
1117 if (has_branch_stack(event))
1118 ctx->nr_branch_stack++;
1120 list_add_rcu(&event->event_entry, &ctx->event_list);
1121 if (!ctx->nr_events)
1122 perf_pmu_rotate_start(ctx->pmu);
1124 if (event->attr.inherit_stat)
1131 * Initialize event state based on the perf_event_attr::disabled.
1133 static inline void perf_event__state_init(struct perf_event *event)
1135 event->state = event->attr.disabled ? PERF_EVENT_STATE_OFF :
1136 PERF_EVENT_STATE_INACTIVE;
1140 * Called at perf_event creation and when events are attached/detached from a
1143 static void perf_event__read_size(struct perf_event *event)
1145 int entry = sizeof(u64); /* value */
1149 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1150 size += sizeof(u64);
1152 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1153 size += sizeof(u64);
1155 if (event->attr.read_format & PERF_FORMAT_ID)
1156 entry += sizeof(u64);
1158 if (event->attr.read_format & PERF_FORMAT_GROUP) {
1159 nr += event->group_leader->nr_siblings;
1160 size += sizeof(u64);
1164 event->read_size = size;
1167 static void perf_event__header_size(struct perf_event *event)
1169 struct perf_sample_data *data;
1170 u64 sample_type = event->attr.sample_type;
1173 perf_event__read_size(event);
1175 if (sample_type & PERF_SAMPLE_IP)
1176 size += sizeof(data->ip);
1178 if (sample_type & PERF_SAMPLE_ADDR)
1179 size += sizeof(data->addr);
1181 if (sample_type & PERF_SAMPLE_PERIOD)
1182 size += sizeof(data->period);
1184 if (sample_type & PERF_SAMPLE_WEIGHT)
1185 size += sizeof(data->weight);
1187 if (sample_type & PERF_SAMPLE_READ)
1188 size += event->read_size;
1190 if (sample_type & PERF_SAMPLE_DATA_SRC)
1191 size += sizeof(data->data_src.val);
1193 if (sample_type & PERF_SAMPLE_TRANSACTION)
1194 size += sizeof(data->txn);
1196 event->header_size = size;
1199 static void perf_event__id_header_size(struct perf_event *event)
1201 struct perf_sample_data *data;
1202 u64 sample_type = event->attr.sample_type;
1205 if (sample_type & PERF_SAMPLE_TID)
1206 size += sizeof(data->tid_entry);
1208 if (sample_type & PERF_SAMPLE_TIME)
1209 size += sizeof(data->time);
1211 if (sample_type & PERF_SAMPLE_IDENTIFIER)
1212 size += sizeof(data->id);
1214 if (sample_type & PERF_SAMPLE_ID)
1215 size += sizeof(data->id);
1217 if (sample_type & PERF_SAMPLE_STREAM_ID)
1218 size += sizeof(data->stream_id);
1220 if (sample_type & PERF_SAMPLE_CPU)
1221 size += sizeof(data->cpu_entry);
1223 event->id_header_size = size;
1226 static void perf_group_attach(struct perf_event *event)
1228 struct perf_event *group_leader = event->group_leader, *pos;
1231 * We can have double attach due to group movement in perf_event_open.
1233 if (event->attach_state & PERF_ATTACH_GROUP)
1236 event->attach_state |= PERF_ATTACH_GROUP;
1238 if (group_leader == event)
1241 if (group_leader->group_flags & PERF_GROUP_SOFTWARE &&
1242 !is_software_event(event))
1243 group_leader->group_flags &= ~PERF_GROUP_SOFTWARE;
1245 list_add_tail(&event->group_entry, &group_leader->sibling_list);
1246 group_leader->nr_siblings++;
1248 perf_event__header_size(group_leader);
1250 list_for_each_entry(pos, &group_leader->sibling_list, group_entry)
1251 perf_event__header_size(pos);
1255 * Remove a event from the lists for its context.
1256 * Must be called with ctx->mutex and ctx->lock held.
1259 list_del_event(struct perf_event *event, struct perf_event_context *ctx)
1261 struct perf_cpu_context *cpuctx;
1263 * We can have double detach due to exit/hot-unplug + close.
1265 if (!(event->attach_state & PERF_ATTACH_CONTEXT))
1268 event->attach_state &= ~PERF_ATTACH_CONTEXT;
1270 if (is_cgroup_event(event)) {
1272 cpuctx = __get_cpu_context(ctx);
1274 * if there are no more cgroup events
1275 * then cler cgrp to avoid stale pointer
1276 * in update_cgrp_time_from_cpuctx()
1278 if (!ctx->nr_cgroups)
1279 cpuctx->cgrp = NULL;
1282 if (has_branch_stack(event))
1283 ctx->nr_branch_stack--;
1286 if (event->attr.inherit_stat)
1289 list_del_rcu(&event->event_entry);
1291 if (event->group_leader == event)
1292 list_del_init(&event->group_entry);
1294 update_group_times(event);
1297 * If event was in error state, then keep it
1298 * that way, otherwise bogus counts will be
1299 * returned on read(). The only way to get out
1300 * of error state is by explicit re-enabling
1303 if (event->state > PERF_EVENT_STATE_OFF)
1304 event->state = PERF_EVENT_STATE_OFF;
1309 static void perf_group_detach(struct perf_event *event)
1311 struct perf_event *sibling, *tmp;
1312 struct list_head *list = NULL;
1315 * We can have double detach due to exit/hot-unplug + close.
1317 if (!(event->attach_state & PERF_ATTACH_GROUP))
1320 event->attach_state &= ~PERF_ATTACH_GROUP;
1323 * If this is a sibling, remove it from its group.
1325 if (event->group_leader != event) {
1326 list_del_init(&event->group_entry);
1327 event->group_leader->nr_siblings--;
1331 if (!list_empty(&event->group_entry))
1332 list = &event->group_entry;
1335 * If this was a group event with sibling events then
1336 * upgrade the siblings to singleton events by adding them
1337 * to whatever list we are on.
1339 list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) {
1341 list_move_tail(&sibling->group_entry, list);
1342 sibling->group_leader = sibling;
1344 /* Inherit group flags from the previous leader */
1345 sibling->group_flags = event->group_flags;
1349 perf_event__header_size(event->group_leader);
1351 list_for_each_entry(tmp, &event->group_leader->sibling_list, group_entry)
1352 perf_event__header_size(tmp);
1356 event_filter_match(struct perf_event *event)
1358 return (event->cpu == -1 || event->cpu == smp_processor_id())
1359 && perf_cgroup_match(event);
1363 event_sched_out(struct perf_event *event,
1364 struct perf_cpu_context *cpuctx,
1365 struct perf_event_context *ctx)
1367 u64 tstamp = perf_event_time(event);
1370 * An event which could not be activated because of
1371 * filter mismatch still needs to have its timings
1372 * maintained, otherwise bogus information is return
1373 * via read() for time_enabled, time_running:
1375 if (event->state == PERF_EVENT_STATE_INACTIVE
1376 && !event_filter_match(event)) {
1377 delta = tstamp - event->tstamp_stopped;
1378 event->tstamp_running += delta;
1379 event->tstamp_stopped = tstamp;
1382 if (event->state != PERF_EVENT_STATE_ACTIVE)
1385 perf_pmu_disable(event->pmu);
1387 event->state = PERF_EVENT_STATE_INACTIVE;
1388 if (event->pending_disable) {
1389 event->pending_disable = 0;
1390 event->state = PERF_EVENT_STATE_OFF;
1392 event->tstamp_stopped = tstamp;
1393 event->pmu->del(event, 0);
1396 if (!is_software_event(event))
1397 cpuctx->active_oncpu--;
1399 if (event->attr.freq && event->attr.sample_freq)
1401 if (event->attr.exclusive || !cpuctx->active_oncpu)
1402 cpuctx->exclusive = 0;
1404 perf_pmu_enable(event->pmu);
1408 group_sched_out(struct perf_event *group_event,
1409 struct perf_cpu_context *cpuctx,
1410 struct perf_event_context *ctx)
1412 struct perf_event *event;
1413 int state = group_event->state;
1415 event_sched_out(group_event, cpuctx, ctx);
1418 * Schedule out siblings (if any):
1420 list_for_each_entry(event, &group_event->sibling_list, group_entry)
1421 event_sched_out(event, cpuctx, ctx);
1423 if (state == PERF_EVENT_STATE_ACTIVE && group_event->attr.exclusive)
1424 cpuctx->exclusive = 0;
1428 * Cross CPU call to remove a performance event
1430 * We disable the event on the hardware level first. After that we
1431 * remove it from the context list.
1433 static int __perf_remove_from_context(void *info)
1435 struct perf_event *event = info;
1436 struct perf_event_context *ctx = event->ctx;
1437 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1439 raw_spin_lock(&ctx->lock);
1440 event_sched_out(event, cpuctx, ctx);
1441 list_del_event(event, ctx);
1442 if (!ctx->nr_events && cpuctx->task_ctx == ctx) {
1444 cpuctx->task_ctx = NULL;
1446 raw_spin_unlock(&ctx->lock);
1453 * Remove the event from a task's (or a CPU's) list of events.
1455 * CPU events are removed with a smp call. For task events we only
1456 * call when the task is on a CPU.
1458 * If event->ctx is a cloned context, callers must make sure that
1459 * every task struct that event->ctx->task could possibly point to
1460 * remains valid. This is OK when called from perf_release since
1461 * that only calls us on the top-level context, which can't be a clone.
1462 * When called from perf_event_exit_task, it's OK because the
1463 * context has been detached from its task.
1465 static void perf_remove_from_context(struct perf_event *event)
1467 struct perf_event_context *ctx = event->ctx;
1468 struct task_struct *task = ctx->task;
1470 lockdep_assert_held(&ctx->mutex);
1474 * Per cpu events are removed via an smp call and
1475 * the removal is always successful.
1477 cpu_function_call(event->cpu, __perf_remove_from_context, event);
1482 if (!task_function_call(task, __perf_remove_from_context, event))
1485 raw_spin_lock_irq(&ctx->lock);
1487 * If we failed to find a running task, but find the context active now
1488 * that we've acquired the ctx->lock, retry.
1490 if (ctx->is_active) {
1491 raw_spin_unlock_irq(&ctx->lock);
1496 * Since the task isn't running, its safe to remove the event, us
1497 * holding the ctx->lock ensures the task won't get scheduled in.
1499 list_del_event(event, ctx);
1500 raw_spin_unlock_irq(&ctx->lock);
1504 * Cross CPU call to disable a performance event
1506 int __perf_event_disable(void *info)
1508 struct perf_event *event = info;
1509 struct perf_event_context *ctx = event->ctx;
1510 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1513 * If this is a per-task event, need to check whether this
1514 * event's task is the current task on this cpu.
1516 * Can trigger due to concurrent perf_event_context_sched_out()
1517 * flipping contexts around.
1519 if (ctx->task && cpuctx->task_ctx != ctx)
1522 raw_spin_lock(&ctx->lock);
1525 * If the event is on, turn it off.
1526 * If it is in error state, leave it in error state.
1528 if (event->state >= PERF_EVENT_STATE_INACTIVE) {
1529 update_context_time(ctx);
1530 update_cgrp_time_from_event(event);
1531 update_group_times(event);
1532 if (event == event->group_leader)
1533 group_sched_out(event, cpuctx, ctx);
1535 event_sched_out(event, cpuctx, ctx);
1536 event->state = PERF_EVENT_STATE_OFF;
1539 raw_spin_unlock(&ctx->lock);
1547 * If event->ctx is a cloned context, callers must make sure that
1548 * every task struct that event->ctx->task could possibly point to
1549 * remains valid. This condition is satisifed when called through
1550 * perf_event_for_each_child or perf_event_for_each because they
1551 * hold the top-level event's child_mutex, so any descendant that
1552 * goes to exit will block in sync_child_event.
1553 * When called from perf_pending_event it's OK because event->ctx
1554 * is the current context on this CPU and preemption is disabled,
1555 * hence we can't get into perf_event_task_sched_out for this context.
1557 void perf_event_disable(struct perf_event *event)
1559 struct perf_event_context *ctx = event->ctx;
1560 struct task_struct *task = ctx->task;
1564 * Disable the event on the cpu that it's on
1566 cpu_function_call(event->cpu, __perf_event_disable, event);
1571 if (!task_function_call(task, __perf_event_disable, event))
1574 raw_spin_lock_irq(&ctx->lock);
1576 * If the event is still active, we need to retry the cross-call.
1578 if (event->state == PERF_EVENT_STATE_ACTIVE) {
1579 raw_spin_unlock_irq(&ctx->lock);
1581 * Reload the task pointer, it might have been changed by
1582 * a concurrent perf_event_context_sched_out().
1589 * Since we have the lock this context can't be scheduled
1590 * in, so we can change the state safely.
1592 if (event->state == PERF_EVENT_STATE_INACTIVE) {
1593 update_group_times(event);
1594 event->state = PERF_EVENT_STATE_OFF;
1596 raw_spin_unlock_irq(&ctx->lock);
1598 EXPORT_SYMBOL_GPL(perf_event_disable);
1600 static void perf_set_shadow_time(struct perf_event *event,
1601 struct perf_event_context *ctx,
1605 * use the correct time source for the time snapshot
1607 * We could get by without this by leveraging the
1608 * fact that to get to this function, the caller
1609 * has most likely already called update_context_time()
1610 * and update_cgrp_time_xx() and thus both timestamp
1611 * are identical (or very close). Given that tstamp is,
1612 * already adjusted for cgroup, we could say that:
1613 * tstamp - ctx->timestamp
1615 * tstamp - cgrp->timestamp.
1617 * Then, in perf_output_read(), the calculation would
1618 * work with no changes because:
1619 * - event is guaranteed scheduled in
1620 * - no scheduled out in between
1621 * - thus the timestamp would be the same
1623 * But this is a bit hairy.
1625 * So instead, we have an explicit cgroup call to remain
1626 * within the time time source all along. We believe it
1627 * is cleaner and simpler to understand.
1629 if (is_cgroup_event(event))
1630 perf_cgroup_set_shadow_time(event, tstamp);
1632 event->shadow_ctx_time = tstamp - ctx->timestamp;
1635 #define MAX_INTERRUPTS (~0ULL)
1637 static void perf_log_throttle(struct perf_event *event, int enable);
1640 event_sched_in(struct perf_event *event,
1641 struct perf_cpu_context *cpuctx,
1642 struct perf_event_context *ctx)
1644 u64 tstamp = perf_event_time(event);
1647 if (event->state <= PERF_EVENT_STATE_OFF)
1650 event->state = PERF_EVENT_STATE_ACTIVE;
1651 event->oncpu = smp_processor_id();
1654 * Unthrottle events, since we scheduled we might have missed several
1655 * ticks already, also for a heavily scheduling task there is little
1656 * guarantee it'll get a tick in a timely manner.
1658 if (unlikely(event->hw.interrupts == MAX_INTERRUPTS)) {
1659 perf_log_throttle(event, 1);
1660 event->hw.interrupts = 0;
1664 * The new state must be visible before we turn it on in the hardware:
1668 perf_pmu_disable(event->pmu);
1670 if (event->pmu->add(event, PERF_EF_START)) {
1671 event->state = PERF_EVENT_STATE_INACTIVE;
1677 event->tstamp_running += tstamp - event->tstamp_stopped;
1679 perf_set_shadow_time(event, ctx, tstamp);
1681 if (!is_software_event(event))
1682 cpuctx->active_oncpu++;
1684 if (event->attr.freq && event->attr.sample_freq)
1687 if (event->attr.exclusive)
1688 cpuctx->exclusive = 1;
1691 perf_pmu_enable(event->pmu);
1697 group_sched_in(struct perf_event *group_event,
1698 struct perf_cpu_context *cpuctx,
1699 struct perf_event_context *ctx)
1701 struct perf_event *event, *partial_group = NULL;
1702 struct pmu *pmu = group_event->pmu;
1703 u64 now = ctx->time;
1704 bool simulate = false;
1706 if (group_event->state == PERF_EVENT_STATE_OFF)
1709 pmu->start_txn(pmu);
1711 if (event_sched_in(group_event, cpuctx, ctx)) {
1712 pmu->cancel_txn(pmu);
1713 perf_cpu_hrtimer_restart(cpuctx);
1718 * Schedule in siblings as one group (if any):
1720 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
1721 if (event_sched_in(event, cpuctx, ctx)) {
1722 partial_group = event;
1727 if (!pmu->commit_txn(pmu))
1732 * Groups can be scheduled in as one unit only, so undo any
1733 * partial group before returning:
1734 * The events up to the failed event are scheduled out normally,
1735 * tstamp_stopped will be updated.
1737 * The failed events and the remaining siblings need to have
1738 * their timings updated as if they had gone thru event_sched_in()
1739 * and event_sched_out(). This is required to get consistent timings
1740 * across the group. This also takes care of the case where the group
1741 * could never be scheduled by ensuring tstamp_stopped is set to mark
1742 * the time the event was actually stopped, such that time delta
1743 * calculation in update_event_times() is correct.
1745 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
1746 if (event == partial_group)
1750 event->tstamp_running += now - event->tstamp_stopped;
1751 event->tstamp_stopped = now;
1753 event_sched_out(event, cpuctx, ctx);
1756 event_sched_out(group_event, cpuctx, ctx);
1758 pmu->cancel_txn(pmu);
1760 perf_cpu_hrtimer_restart(cpuctx);
1766 * Work out whether we can put this event group on the CPU now.
1768 static int group_can_go_on(struct perf_event *event,
1769 struct perf_cpu_context *cpuctx,
1773 * Groups consisting entirely of software events can always go on.
1775 if (event->group_flags & PERF_GROUP_SOFTWARE)
1778 * If an exclusive group is already on, no other hardware
1781 if (cpuctx->exclusive)
1784 * If this group is exclusive and there are already
1785 * events on the CPU, it can't go on.
1787 if (event->attr.exclusive && cpuctx->active_oncpu)
1790 * Otherwise, try to add it if all previous groups were able
1796 static void add_event_to_ctx(struct perf_event *event,
1797 struct perf_event_context *ctx)
1799 u64 tstamp = perf_event_time(event);
1801 list_add_event(event, ctx);
1802 perf_group_attach(event);
1803 event->tstamp_enabled = tstamp;
1804 event->tstamp_running = tstamp;
1805 event->tstamp_stopped = tstamp;
1808 static void task_ctx_sched_out(struct perf_event_context *ctx);
1810 ctx_sched_in(struct perf_event_context *ctx,
1811 struct perf_cpu_context *cpuctx,
1812 enum event_type_t event_type,
1813 struct task_struct *task);
1815 static void perf_event_sched_in(struct perf_cpu_context *cpuctx,
1816 struct perf_event_context *ctx,
1817 struct task_struct *task)
1819 cpu_ctx_sched_in(cpuctx, EVENT_PINNED, task);
1821 ctx_sched_in(ctx, cpuctx, EVENT_PINNED, task);
1822 cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE, task);
1824 ctx_sched_in(ctx, cpuctx, EVENT_FLEXIBLE, task);
1828 * Cross CPU call to install and enable a performance event
1830 * Must be called with ctx->mutex held
1832 static int __perf_install_in_context(void *info)
1834 struct perf_event *event = info;
1835 struct perf_event_context *ctx = event->ctx;
1836 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1837 struct perf_event_context *task_ctx = cpuctx->task_ctx;
1838 struct task_struct *task = current;
1840 perf_ctx_lock(cpuctx, task_ctx);
1841 perf_pmu_disable(cpuctx->ctx.pmu);
1844 * If there was an active task_ctx schedule it out.
1847 task_ctx_sched_out(task_ctx);
1850 * If the context we're installing events in is not the
1851 * active task_ctx, flip them.
1853 if (ctx->task && task_ctx != ctx) {
1855 raw_spin_unlock(&task_ctx->lock);
1856 raw_spin_lock(&ctx->lock);
1861 cpuctx->task_ctx = task_ctx;
1862 task = task_ctx->task;
1865 cpu_ctx_sched_out(cpuctx, EVENT_ALL);
1867 update_context_time(ctx);
1869 * update cgrp time only if current cgrp
1870 * matches event->cgrp. Must be done before
1871 * calling add_event_to_ctx()
1873 update_cgrp_time_from_event(event);
1875 add_event_to_ctx(event, ctx);
1878 * Schedule everything back in
1880 perf_event_sched_in(cpuctx, task_ctx, task);
1882 perf_pmu_enable(cpuctx->ctx.pmu);
1883 perf_ctx_unlock(cpuctx, task_ctx);
1889 * Attach a performance event to a context
1891 * First we add the event to the list with the hardware enable bit
1892 * in event->hw_config cleared.
1894 * If the event is attached to a task which is on a CPU we use a smp
1895 * call to enable it in the task context. The task might have been
1896 * scheduled away, but we check this in the smp call again.
1899 perf_install_in_context(struct perf_event_context *ctx,
1900 struct perf_event *event,
1903 struct task_struct *task = ctx->task;
1905 lockdep_assert_held(&ctx->mutex);
1908 if (event->cpu != -1)
1913 * Per cpu events are installed via an smp call and
1914 * the install is always successful.
1916 cpu_function_call(cpu, __perf_install_in_context, event);
1921 if (!task_function_call(task, __perf_install_in_context, event))
1924 raw_spin_lock_irq(&ctx->lock);
1926 * If we failed to find a running task, but find the context active now
1927 * that we've acquired the ctx->lock, retry.
1929 if (ctx->is_active) {
1930 raw_spin_unlock_irq(&ctx->lock);
1935 * Since the task isn't running, its safe to add the event, us holding
1936 * the ctx->lock ensures the task won't get scheduled in.
1938 add_event_to_ctx(event, ctx);
1939 raw_spin_unlock_irq(&ctx->lock);
1943 * Put a event into inactive state and update time fields.
1944 * Enabling the leader of a group effectively enables all
1945 * the group members that aren't explicitly disabled, so we
1946 * have to update their ->tstamp_enabled also.
1947 * Note: this works for group members as well as group leaders
1948 * since the non-leader members' sibling_lists will be empty.
1950 static void __perf_event_mark_enabled(struct perf_event *event)
1952 struct perf_event *sub;
1953 u64 tstamp = perf_event_time(event);
1955 event->state = PERF_EVENT_STATE_INACTIVE;
1956 event->tstamp_enabled = tstamp - event->total_time_enabled;
1957 list_for_each_entry(sub, &event->sibling_list, group_entry) {
1958 if (sub->state >= PERF_EVENT_STATE_INACTIVE)
1959 sub->tstamp_enabled = tstamp - sub->total_time_enabled;
1964 * Cross CPU call to enable a performance event
1966 static int __perf_event_enable(void *info)
1968 struct perf_event *event = info;
1969 struct perf_event_context *ctx = event->ctx;
1970 struct perf_event *leader = event->group_leader;
1971 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1975 * There's a time window between 'ctx->is_active' check
1976 * in perf_event_enable function and this place having:
1978 * - ctx->lock unlocked
1980 * where the task could be killed and 'ctx' deactivated
1981 * by perf_event_exit_task.
1983 if (!ctx->is_active)
1986 raw_spin_lock(&ctx->lock);
1987 update_context_time(ctx);
1989 if (event->state >= PERF_EVENT_STATE_INACTIVE)
1993 * set current task's cgroup time reference point
1995 perf_cgroup_set_timestamp(current, ctx);
1997 __perf_event_mark_enabled(event);
1999 if (!event_filter_match(event)) {
2000 if (is_cgroup_event(event))
2001 perf_cgroup_defer_enabled(event);
2006 * If the event is in a group and isn't the group leader,
2007 * then don't put it on unless the group is on.
2009 if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE)
2012 if (!group_can_go_on(event, cpuctx, 1)) {
2015 if (event == leader)
2016 err = group_sched_in(event, cpuctx, ctx);
2018 err = event_sched_in(event, cpuctx, ctx);
2023 * If this event can't go on and it's part of a
2024 * group, then the whole group has to come off.
2026 if (leader != event) {
2027 group_sched_out(leader, cpuctx, ctx);
2028 perf_cpu_hrtimer_restart(cpuctx);
2030 if (leader->attr.pinned) {
2031 update_group_times(leader);
2032 leader->state = PERF_EVENT_STATE_ERROR;
2037 raw_spin_unlock(&ctx->lock);
2045 * If event->ctx is a cloned context, callers must make sure that
2046 * every task struct that event->ctx->task could possibly point to
2047 * remains valid. This condition is satisfied when called through
2048 * perf_event_for_each_child or perf_event_for_each as described
2049 * for perf_event_disable.
2051 void perf_event_enable(struct perf_event *event)
2053 struct perf_event_context *ctx = event->ctx;
2054 struct task_struct *task = ctx->task;
2058 * Enable the event on the cpu that it's on
2060 cpu_function_call(event->cpu, __perf_event_enable, event);
2064 raw_spin_lock_irq(&ctx->lock);
2065 if (event->state >= PERF_EVENT_STATE_INACTIVE)
2069 * If the event is in error state, clear that first.
2070 * That way, if we see the event in error state below, we
2071 * know that it has gone back into error state, as distinct
2072 * from the task having been scheduled away before the
2073 * cross-call arrived.
2075 if (event->state == PERF_EVENT_STATE_ERROR)
2076 event->state = PERF_EVENT_STATE_OFF;
2079 if (!ctx->is_active) {
2080 __perf_event_mark_enabled(event);
2084 raw_spin_unlock_irq(&ctx->lock);
2086 if (!task_function_call(task, __perf_event_enable, event))
2089 raw_spin_lock_irq(&ctx->lock);
2092 * If the context is active and the event is still off,
2093 * we need to retry the cross-call.
2095 if (ctx->is_active && event->state == PERF_EVENT_STATE_OFF) {
2097 * task could have been flipped by a concurrent
2098 * perf_event_context_sched_out()
2105 raw_spin_unlock_irq(&ctx->lock);
2107 EXPORT_SYMBOL_GPL(perf_event_enable);
2109 int perf_event_refresh(struct perf_event *event, int refresh)
2112 * not supported on inherited events
2114 if (event->attr.inherit || !is_sampling_event(event))
2117 atomic_add(refresh, &event->event_limit);
2118 perf_event_enable(event);
2122 EXPORT_SYMBOL_GPL(perf_event_refresh);
2124 static void ctx_sched_out(struct perf_event_context *ctx,
2125 struct perf_cpu_context *cpuctx,
2126 enum event_type_t event_type)
2128 struct perf_event *event;
2129 int is_active = ctx->is_active;
2131 ctx->is_active &= ~event_type;
2132 if (likely(!ctx->nr_events))
2135 update_context_time(ctx);
2136 update_cgrp_time_from_cpuctx(cpuctx);
2137 if (!ctx->nr_active)
2140 perf_pmu_disable(ctx->pmu);
2141 if ((is_active & EVENT_PINNED) && (event_type & EVENT_PINNED)) {
2142 list_for_each_entry(event, &ctx->pinned_groups, group_entry)
2143 group_sched_out(event, cpuctx, ctx);
2146 if ((is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE)) {
2147 list_for_each_entry(event, &ctx->flexible_groups, group_entry)
2148 group_sched_out(event, cpuctx, ctx);
2150 perf_pmu_enable(ctx->pmu);
2154 * Test whether two contexts are equivalent, i.e. whether they have both been
2155 * cloned from the same version of the same context.
2157 * Equivalence is measured using a generation number in the context that is
2158 * incremented on each modification to it; see unclone_ctx(), list_add_event()
2159 * and list_del_event().
2161 static int context_equiv(struct perf_event_context *ctx1,
2162 struct perf_event_context *ctx2)
2164 /* Pinning disables the swap optimization */
2165 if (ctx1->pin_count || ctx2->pin_count)
2168 /* If ctx1 is the parent of ctx2 */
2169 if (ctx1 == ctx2->parent_ctx && ctx1->generation == ctx2->parent_gen)
2172 /* If ctx2 is the parent of ctx1 */
2173 if (ctx1->parent_ctx == ctx2 && ctx1->parent_gen == ctx2->generation)
2177 * If ctx1 and ctx2 have the same parent; we flatten the parent
2178 * hierarchy, see perf_event_init_context().
2180 if (ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx &&
2181 ctx1->parent_gen == ctx2->parent_gen)
2188 static void __perf_event_sync_stat(struct perf_event *event,
2189 struct perf_event *next_event)
2193 if (!event->attr.inherit_stat)
2197 * Update the event value, we cannot use perf_event_read()
2198 * because we're in the middle of a context switch and have IRQs
2199 * disabled, which upsets smp_call_function_single(), however
2200 * we know the event must be on the current CPU, therefore we
2201 * don't need to use it.
2203 switch (event->state) {
2204 case PERF_EVENT_STATE_ACTIVE:
2205 event->pmu->read(event);
2208 case PERF_EVENT_STATE_INACTIVE:
2209 update_event_times(event);
2217 * In order to keep per-task stats reliable we need to flip the event
2218 * values when we flip the contexts.
2220 value = local64_read(&next_event->count);
2221 value = local64_xchg(&event->count, value);
2222 local64_set(&next_event->count, value);
2224 swap(event->total_time_enabled, next_event->total_time_enabled);
2225 swap(event->total_time_running, next_event->total_time_running);
2228 * Since we swizzled the values, update the user visible data too.
2230 perf_event_update_userpage(event);
2231 perf_event_update_userpage(next_event);
2234 static void perf_event_sync_stat(struct perf_event_context *ctx,
2235 struct perf_event_context *next_ctx)
2237 struct perf_event *event, *next_event;
2242 update_context_time(ctx);
2244 event = list_first_entry(&ctx->event_list,
2245 struct perf_event, event_entry);
2247 next_event = list_first_entry(&next_ctx->event_list,
2248 struct perf_event, event_entry);
2250 while (&event->event_entry != &ctx->event_list &&
2251 &next_event->event_entry != &next_ctx->event_list) {
2253 __perf_event_sync_stat(event, next_event);
2255 event = list_next_entry(event, event_entry);
2256 next_event = list_next_entry(next_event, event_entry);
2260 static void perf_event_context_sched_out(struct task_struct *task, int ctxn,
2261 struct task_struct *next)
2263 struct perf_event_context *ctx = task->perf_event_ctxp[ctxn];
2264 struct perf_event_context *next_ctx;
2265 struct perf_event_context *parent, *next_parent;
2266 struct perf_cpu_context *cpuctx;
2272 cpuctx = __get_cpu_context(ctx);
2273 if (!cpuctx->task_ctx)
2277 next_ctx = next->perf_event_ctxp[ctxn];
2281 parent = rcu_dereference(ctx->parent_ctx);
2282 next_parent = rcu_dereference(next_ctx->parent_ctx);
2284 /* If neither context have a parent context; they cannot be clones. */
2285 if (!parent && !next_parent)
2288 if (next_parent == ctx || next_ctx == parent || next_parent == parent) {
2290 * Looks like the two contexts are clones, so we might be
2291 * able to optimize the context switch. We lock both
2292 * contexts and check that they are clones under the
2293 * lock (including re-checking that neither has been
2294 * uncloned in the meantime). It doesn't matter which
2295 * order we take the locks because no other cpu could
2296 * be trying to lock both of these tasks.
2298 raw_spin_lock(&ctx->lock);
2299 raw_spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING);
2300 if (context_equiv(ctx, next_ctx)) {
2302 * XXX do we need a memory barrier of sorts
2303 * wrt to rcu_dereference() of perf_event_ctxp
2305 task->perf_event_ctxp[ctxn] = next_ctx;
2306 next->perf_event_ctxp[ctxn] = ctx;
2308 next_ctx->task = task;
2311 perf_event_sync_stat(ctx, next_ctx);
2313 raw_spin_unlock(&next_ctx->lock);
2314 raw_spin_unlock(&ctx->lock);
2320 raw_spin_lock(&ctx->lock);
2321 ctx_sched_out(ctx, cpuctx, EVENT_ALL);
2322 cpuctx->task_ctx = NULL;
2323 raw_spin_unlock(&ctx->lock);
2327 #define for_each_task_context_nr(ctxn) \
2328 for ((ctxn) = 0; (ctxn) < perf_nr_task_contexts; (ctxn)++)
2331 * Called from scheduler to remove the events of the current task,
2332 * with interrupts disabled.
2334 * We stop each event and update the event value in event->count.
2336 * This does not protect us against NMI, but disable()
2337 * sets the disabled bit in the control field of event _before_
2338 * accessing the event control register. If a NMI hits, then it will
2339 * not restart the event.
2341 void __perf_event_task_sched_out(struct task_struct *task,
2342 struct task_struct *next)
2346 for_each_task_context_nr(ctxn)
2347 perf_event_context_sched_out(task, ctxn, next);
2350 * if cgroup events exist on this CPU, then we need
2351 * to check if we have to switch out PMU state.
2352 * cgroup event are system-wide mode only
2354 if (atomic_read(&__get_cpu_var(perf_cgroup_events)))
2355 perf_cgroup_sched_out(task, next);
2358 static void task_ctx_sched_out(struct perf_event_context *ctx)
2360 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
2362 if (!cpuctx->task_ctx)
2365 if (WARN_ON_ONCE(ctx != cpuctx->task_ctx))
2368 ctx_sched_out(ctx, cpuctx, EVENT_ALL);
2369 cpuctx->task_ctx = NULL;
2373 * Called with IRQs disabled
2375 static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
2376 enum event_type_t event_type)
2378 ctx_sched_out(&cpuctx->ctx, cpuctx, event_type);
2382 ctx_pinned_sched_in(struct perf_event_context *ctx,
2383 struct perf_cpu_context *cpuctx)
2385 struct perf_event *event;
2387 list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
2388 if (event->state <= PERF_EVENT_STATE_OFF)
2390 if (!event_filter_match(event))
2393 /* may need to reset tstamp_enabled */
2394 if (is_cgroup_event(event))
2395 perf_cgroup_mark_enabled(event, ctx);
2397 if (group_can_go_on(event, cpuctx, 1))
2398 group_sched_in(event, cpuctx, ctx);
2401 * If this pinned group hasn't been scheduled,
2402 * put it in error state.
2404 if (event->state == PERF_EVENT_STATE_INACTIVE) {
2405 update_group_times(event);
2406 event->state = PERF_EVENT_STATE_ERROR;
2412 ctx_flexible_sched_in(struct perf_event_context *ctx,
2413 struct perf_cpu_context *cpuctx)
2415 struct perf_event *event;
2418 list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
2419 /* Ignore events in OFF or ERROR state */
2420 if (event->state <= PERF_EVENT_STATE_OFF)
2423 * Listen to the 'cpu' scheduling filter constraint
2426 if (!event_filter_match(event))
2429 /* may need to reset tstamp_enabled */
2430 if (is_cgroup_event(event))
2431 perf_cgroup_mark_enabled(event, ctx);
2433 if (group_can_go_on(event, cpuctx, can_add_hw)) {
2434 if (group_sched_in(event, cpuctx, ctx))
2441 ctx_sched_in(struct perf_event_context *ctx,
2442 struct perf_cpu_context *cpuctx,
2443 enum event_type_t event_type,
2444 struct task_struct *task)
2447 int is_active = ctx->is_active;
2449 ctx->is_active |= event_type;
2450 if (likely(!ctx->nr_events))
2454 ctx->timestamp = now;
2455 perf_cgroup_set_timestamp(task, ctx);
2457 * First go through the list and put on any pinned groups
2458 * in order to give them the best chance of going on.
2460 if (!(is_active & EVENT_PINNED) && (event_type & EVENT_PINNED))
2461 ctx_pinned_sched_in(ctx, cpuctx);
2463 /* Then walk through the lower prio flexible groups */
2464 if (!(is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE))
2465 ctx_flexible_sched_in(ctx, cpuctx);
2468 static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
2469 enum event_type_t event_type,
2470 struct task_struct *task)
2472 struct perf_event_context *ctx = &cpuctx->ctx;
2474 ctx_sched_in(ctx, cpuctx, event_type, task);
2477 static void perf_event_context_sched_in(struct perf_event_context *ctx,
2478 struct task_struct *task)
2480 struct perf_cpu_context *cpuctx;
2482 cpuctx = __get_cpu_context(ctx);
2483 if (cpuctx->task_ctx == ctx)
2486 perf_ctx_lock(cpuctx, ctx);
2487 perf_pmu_disable(ctx->pmu);
2489 * We want to keep the following priority order:
2490 * cpu pinned (that don't need to move), task pinned,
2491 * cpu flexible, task flexible.
2493 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
2496 cpuctx->task_ctx = ctx;
2498 perf_event_sched_in(cpuctx, cpuctx->task_ctx, task);
2500 perf_pmu_enable(ctx->pmu);
2501 perf_ctx_unlock(cpuctx, ctx);
2504 * Since these rotations are per-cpu, we need to ensure the
2505 * cpu-context we got scheduled on is actually rotating.
2507 perf_pmu_rotate_start(ctx->pmu);
2511 * When sampling the branck stack in system-wide, it may be necessary
2512 * to flush the stack on context switch. This happens when the branch
2513 * stack does not tag its entries with the pid of the current task.
2514 * Otherwise it becomes impossible to associate a branch entry with a
2515 * task. This ambiguity is more likely to appear when the branch stack
2516 * supports priv level filtering and the user sets it to monitor only
2517 * at the user level (which could be a useful measurement in system-wide
2518 * mode). In that case, the risk is high of having a branch stack with
2519 * branch from multiple tasks. Flushing may mean dropping the existing
2520 * entries or stashing them somewhere in the PMU specific code layer.
2522 * This function provides the context switch callback to the lower code
2523 * layer. It is invoked ONLY when there is at least one system-wide context
2524 * with at least one active event using taken branch sampling.
2526 static void perf_branch_stack_sched_in(struct task_struct *prev,
2527 struct task_struct *task)
2529 struct perf_cpu_context *cpuctx;
2531 unsigned long flags;
2533 /* no need to flush branch stack if not changing task */
2537 local_irq_save(flags);
2541 list_for_each_entry_rcu(pmu, &pmus, entry) {
2542 cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
2545 * check if the context has at least one
2546 * event using PERF_SAMPLE_BRANCH_STACK
2548 if (cpuctx->ctx.nr_branch_stack > 0
2549 && pmu->flush_branch_stack) {
2551 pmu = cpuctx->ctx.pmu;
2553 perf_ctx_lock(cpuctx, cpuctx->task_ctx);
2555 perf_pmu_disable(pmu);
2557 pmu->flush_branch_stack();
2559 perf_pmu_enable(pmu);
2561 perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
2567 local_irq_restore(flags);
2571 * Called from scheduler to add the events of the current task
2572 * with interrupts disabled.
2574 * We restore the event value and then enable it.
2576 * This does not protect us against NMI, but enable()
2577 * sets the enabled bit in the control field of event _before_
2578 * accessing the event control register. If a NMI hits, then it will
2579 * keep the event running.
2581 void __perf_event_task_sched_in(struct task_struct *prev,
2582 struct task_struct *task)
2584 struct perf_event_context *ctx;
2587 for_each_task_context_nr(ctxn) {
2588 ctx = task->perf_event_ctxp[ctxn];
2592 perf_event_context_sched_in(ctx, task);
2595 * if cgroup events exist on this CPU, then we need
2596 * to check if we have to switch in PMU state.
2597 * cgroup event are system-wide mode only
2599 if (atomic_read(&__get_cpu_var(perf_cgroup_events)))
2600 perf_cgroup_sched_in(prev, task);
2602 /* check for system-wide branch_stack events */
2603 if (atomic_read(&__get_cpu_var(perf_branch_stack_events)))
2604 perf_branch_stack_sched_in(prev, task);
2607 static u64 perf_calculate_period(struct perf_event *event, u64 nsec, u64 count)
2609 u64 frequency = event->attr.sample_freq;
2610 u64 sec = NSEC_PER_SEC;
2611 u64 divisor, dividend;
2613 int count_fls, nsec_fls, frequency_fls, sec_fls;
2615 count_fls = fls64(count);
2616 nsec_fls = fls64(nsec);
2617 frequency_fls = fls64(frequency);
2621 * We got @count in @nsec, with a target of sample_freq HZ
2622 * the target period becomes:
2625 * period = -------------------
2626 * @nsec * sample_freq
2631 * Reduce accuracy by one bit such that @a and @b converge
2632 * to a similar magnitude.
2634 #define REDUCE_FLS(a, b) \
2636 if (a##_fls > b##_fls) { \
2646 * Reduce accuracy until either term fits in a u64, then proceed with
2647 * the other, so that finally we can do a u64/u64 division.
2649 while (count_fls + sec_fls > 64 && nsec_fls + frequency_fls > 64) {
2650 REDUCE_FLS(nsec, frequency);
2651 REDUCE_FLS(sec, count);
2654 if (count_fls + sec_fls > 64) {
2655 divisor = nsec * frequency;
2657 while (count_fls + sec_fls > 64) {
2658 REDUCE_FLS(count, sec);
2662 dividend = count * sec;
2664 dividend = count * sec;
2666 while (nsec_fls + frequency_fls > 64) {
2667 REDUCE_FLS(nsec, frequency);
2671 divisor = nsec * frequency;
2677 return div64_u64(dividend, divisor);
2680 static DEFINE_PER_CPU(int, perf_throttled_count);
2681 static DEFINE_PER_CPU(u64, perf_throttled_seq);
2683 static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count, bool disable)
2685 struct hw_perf_event *hwc = &event->hw;
2686 s64 period, sample_period;
2689 period = perf_calculate_period(event, nsec, count);
2691 delta = (s64)(period - hwc->sample_period);
2692 delta = (delta + 7) / 8; /* low pass filter */
2694 sample_period = hwc->sample_period + delta;
2699 hwc->sample_period = sample_period;
2701 if (local64_read(&hwc->period_left) > 8*sample_period) {
2703 event->pmu->stop(event, PERF_EF_UPDATE);
2705 local64_set(&hwc->period_left, 0);
2708 event->pmu->start(event, PERF_EF_RELOAD);
2713 * combine freq adjustment with unthrottling to avoid two passes over the
2714 * events. At the same time, make sure, having freq events does not change
2715 * the rate of unthrottling as that would introduce bias.
2717 static void perf_adjust_freq_unthr_context(struct perf_event_context *ctx,
2720 struct perf_event *event;
2721 struct hw_perf_event *hwc;
2722 u64 now, period = TICK_NSEC;
2726 * only need to iterate over all events iff:
2727 * - context have events in frequency mode (needs freq adjust)
2728 * - there are events to unthrottle on this cpu
2730 if (!(ctx->nr_freq || needs_unthr))
2733 raw_spin_lock(&ctx->lock);
2734 perf_pmu_disable(ctx->pmu);
2736 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
2737 if (event->state != PERF_EVENT_STATE_ACTIVE)
2740 if (!event_filter_match(event))
2743 perf_pmu_disable(event->pmu);
2747 if (hwc->interrupts == MAX_INTERRUPTS) {
2748 hwc->interrupts = 0;
2749 perf_log_throttle(event, 1);
2750 event->pmu->start(event, 0);
2753 if (!event->attr.freq || !event->attr.sample_freq)
2757 * stop the event and update event->count
2759 event->pmu->stop(event, PERF_EF_UPDATE);
2761 now = local64_read(&event->count);
2762 delta = now - hwc->freq_count_stamp;
2763 hwc->freq_count_stamp = now;
2767 * reload only if value has changed
2768 * we have stopped the event so tell that
2769 * to perf_adjust_period() to avoid stopping it
2773 perf_adjust_period(event, period, delta, false);
2775 event->pmu->start(event, delta > 0 ? PERF_EF_RELOAD : 0);
2777 perf_pmu_enable(event->pmu);
2780 perf_pmu_enable(ctx->pmu);
2781 raw_spin_unlock(&ctx->lock);
2785 * Round-robin a context's events:
2787 static void rotate_ctx(struct perf_event_context *ctx)
2790 * Rotate the first entry last of non-pinned groups. Rotation might be
2791 * disabled by the inheritance code.
2793 if (!ctx->rotate_disable)
2794 list_rotate_left(&ctx->flexible_groups);
2798 * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
2799 * because they're strictly cpu affine and rotate_start is called with IRQs
2800 * disabled, while rotate_context is called from IRQ context.
2802 static int perf_rotate_context(struct perf_cpu_context *cpuctx)
2804 struct perf_event_context *ctx = NULL;
2805 int rotate = 0, remove = 1;
2807 if (cpuctx->ctx.nr_events) {
2809 if (cpuctx->ctx.nr_events != cpuctx->ctx.nr_active)
2813 ctx = cpuctx->task_ctx;
2814 if (ctx && ctx->nr_events) {
2816 if (ctx->nr_events != ctx->nr_active)
2823 perf_ctx_lock(cpuctx, cpuctx->task_ctx);
2824 perf_pmu_disable(cpuctx->ctx.pmu);
2826 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
2828 ctx_sched_out(ctx, cpuctx, EVENT_FLEXIBLE);
2830 rotate_ctx(&cpuctx->ctx);
2834 perf_event_sched_in(cpuctx, ctx, current);
2836 perf_pmu_enable(cpuctx->ctx.pmu);
2837 perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
2840 list_del_init(&cpuctx->rotation_list);
2845 #ifdef CONFIG_NO_HZ_FULL
2846 bool perf_event_can_stop_tick(void)
2848 if (atomic_read(&nr_freq_events) ||
2849 __this_cpu_read(perf_throttled_count))
2856 void perf_event_task_tick(void)
2858 struct list_head *head = &__get_cpu_var(rotation_list);
2859 struct perf_cpu_context *cpuctx, *tmp;
2860 struct perf_event_context *ctx;
2863 WARN_ON(!irqs_disabled());
2865 __this_cpu_inc(perf_throttled_seq);
2866 throttled = __this_cpu_xchg(perf_throttled_count, 0);
2868 list_for_each_entry_safe(cpuctx, tmp, head, rotation_list) {
2870 perf_adjust_freq_unthr_context(ctx, throttled);
2872 ctx = cpuctx->task_ctx;
2874 perf_adjust_freq_unthr_context(ctx, throttled);
2878 static int event_enable_on_exec(struct perf_event *event,
2879 struct perf_event_context *ctx)
2881 if (!event->attr.enable_on_exec)
2884 event->attr.enable_on_exec = 0;
2885 if (event->state >= PERF_EVENT_STATE_INACTIVE)
2888 __perf_event_mark_enabled(event);
2894 * Enable all of a task's events that have been marked enable-on-exec.
2895 * This expects task == current.
2897 static void perf_event_enable_on_exec(struct perf_event_context *ctx)
2899 struct perf_event *event;
2900 unsigned long flags;
2904 local_irq_save(flags);
2905 if (!ctx || !ctx->nr_events)
2909 * We must ctxsw out cgroup events to avoid conflict
2910 * when invoking perf_task_event_sched_in() later on
2911 * in this function. Otherwise we end up trying to
2912 * ctxswin cgroup events which are already scheduled
2915 perf_cgroup_sched_out(current, NULL);
2917 raw_spin_lock(&ctx->lock);
2918 task_ctx_sched_out(ctx);
2920 list_for_each_entry(event, &ctx->event_list, event_entry) {
2921 ret = event_enable_on_exec(event, ctx);
2927 * Unclone this context if we enabled any event.
2932 raw_spin_unlock(&ctx->lock);
2935 * Also calls ctxswin for cgroup events, if any:
2937 perf_event_context_sched_in(ctx, ctx->task);
2939 local_irq_restore(flags);
2943 * Cross CPU call to read the hardware event
2945 static void __perf_event_read(void *info)
2947 struct perf_event *event = info;
2948 struct perf_event_context *ctx = event->ctx;
2949 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
2952 * If this is a task context, we need to check whether it is
2953 * the current task context of this cpu. If not it has been
2954 * scheduled out before the smp call arrived. In that case
2955 * event->count would have been updated to a recent sample
2956 * when the event was scheduled out.
2958 if (ctx->task && cpuctx->task_ctx != ctx)
2961 raw_spin_lock(&ctx->lock);
2962 if (ctx->is_active) {
2963 update_context_time(ctx);
2964 update_cgrp_time_from_event(event);
2966 update_event_times(event);
2967 if (event->state == PERF_EVENT_STATE_ACTIVE)
2968 event->pmu->read(event);
2969 raw_spin_unlock(&ctx->lock);
2972 static inline u64 perf_event_count(struct perf_event *event)
2974 return local64_read(&event->count) + atomic64_read(&event->child_count);
2977 static u64 perf_event_read(struct perf_event *event)
2980 * If event is enabled and currently active on a CPU, update the
2981 * value in the event structure:
2983 if (event->state == PERF_EVENT_STATE_ACTIVE) {
2984 smp_call_function_single(event->oncpu,
2985 __perf_event_read, event, 1);
2986 } else if (event->state == PERF_EVENT_STATE_INACTIVE) {
2987 struct perf_event_context *ctx = event->ctx;
2988 unsigned long flags;
2990 raw_spin_lock_irqsave(&ctx->lock, flags);
2992 * may read while context is not active
2993 * (e.g., thread is blocked), in that case
2994 * we cannot update context time
2996 if (ctx->is_active) {
2997 update_context_time(ctx);
2998 update_cgrp_time_from_event(event);
3000 update_event_times(event);
3001 raw_spin_unlock_irqrestore(&ctx->lock, flags);
3004 return perf_event_count(event);
3008 * Initialize the perf_event context in a task_struct:
3010 static void __perf_event_init_context(struct perf_event_context *ctx)
3012 raw_spin_lock_init(&ctx->lock);
3013 mutex_init(&ctx->mutex);
3014 INIT_LIST_HEAD(&ctx->pinned_groups);
3015 INIT_LIST_HEAD(&ctx->flexible_groups);
3016 INIT_LIST_HEAD(&ctx->event_list);
3017 atomic_set(&ctx->refcount, 1);
3020 static struct perf_event_context *
3021 alloc_perf_context(struct pmu *pmu, struct task_struct *task)
3023 struct perf_event_context *ctx;
3025 ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL);
3029 __perf_event_init_context(ctx);
3032 get_task_struct(task);
3039 static struct task_struct *
3040 find_lively_task_by_vpid(pid_t vpid)
3042 struct task_struct *task;
3049 task = find_task_by_vpid(vpid);
3051 get_task_struct(task);
3055 return ERR_PTR(-ESRCH);
3057 /* Reuse ptrace permission checks for now. */
3059 if (!ptrace_may_access(task, PTRACE_MODE_READ))
3064 put_task_struct(task);
3065 return ERR_PTR(err);
3070 * Returns a matching context with refcount and pincount.
3072 static struct perf_event_context *
3073 find_get_context(struct pmu *pmu, struct task_struct *task, int cpu)
3075 struct perf_event_context *ctx;
3076 struct perf_cpu_context *cpuctx;
3077 unsigned long flags;
3081 /* Must be root to operate on a CPU event: */
3082 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
3083 return ERR_PTR(-EACCES);
3086 * We could be clever and allow to attach a event to an
3087 * offline CPU and activate it when the CPU comes up, but
3090 if (!cpu_online(cpu))
3091 return ERR_PTR(-ENODEV);
3093 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
3102 ctxn = pmu->task_ctx_nr;
3107 ctx = perf_lock_task_context(task, ctxn, &flags);
3111 raw_spin_unlock_irqrestore(&ctx->lock, flags);
3113 ctx = alloc_perf_context(pmu, task);
3119 mutex_lock(&task->perf_event_mutex);
3121 * If it has already passed perf_event_exit_task().
3122 * we must see PF_EXITING, it takes this mutex too.
3124 if (task->flags & PF_EXITING)
3126 else if (task->perf_event_ctxp[ctxn])
3131 rcu_assign_pointer(task->perf_event_ctxp[ctxn], ctx);
3133 mutex_unlock(&task->perf_event_mutex);
3135 if (unlikely(err)) {
3147 return ERR_PTR(err);
3150 static void perf_event_free_filter(struct perf_event *event);
3152 static void free_event_rcu(struct rcu_head *head)
3154 struct perf_event *event;
3156 event = container_of(head, struct perf_event, rcu_head);
3158 put_pid_ns(event->ns);
3159 perf_event_free_filter(event);
3163 static void ring_buffer_put(struct ring_buffer *rb);
3164 static void ring_buffer_detach(struct perf_event *event, struct ring_buffer *rb);
3166 static void unaccount_event_cpu(struct perf_event *event, int cpu)
3171 if (has_branch_stack(event)) {
3172 if (!(event->attach_state & PERF_ATTACH_TASK))
3173 atomic_dec(&per_cpu(perf_branch_stack_events, cpu));
3175 if (is_cgroup_event(event))
3176 atomic_dec(&per_cpu(perf_cgroup_events, cpu));
3179 static void unaccount_event(struct perf_event *event)
3184 if (event->attach_state & PERF_ATTACH_TASK)
3185 static_key_slow_dec_deferred(&perf_sched_events);
3186 if (event->attr.mmap || event->attr.mmap_data)
3187 atomic_dec(&nr_mmap_events);
3188 if (event->attr.comm)
3189 atomic_dec(&nr_comm_events);
3190 if (event->attr.task)
3191 atomic_dec(&nr_task_events);
3192 if (event->attr.freq)
3193 atomic_dec(&nr_freq_events);
3194 if (is_cgroup_event(event))
3195 static_key_slow_dec_deferred(&perf_sched_events);
3196 if (has_branch_stack(event))
3197 static_key_slow_dec_deferred(&perf_sched_events);
3199 unaccount_event_cpu(event, event->cpu);
3202 static void __free_event(struct perf_event *event)
3204 if (!event->parent) {
3205 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN)
3206 put_callchain_buffers();
3210 event->destroy(event);
3213 put_ctx(event->ctx);
3215 call_rcu(&event->rcu_head, free_event_rcu);
3217 static void free_event(struct perf_event *event)
3219 irq_work_sync(&event->pending);
3221 unaccount_event(event);
3224 struct ring_buffer *rb;
3227 * Can happen when we close an event with re-directed output.
3229 * Since we have a 0 refcount, perf_mmap_close() will skip
3230 * over us; possibly making our ring_buffer_put() the last.
3232 mutex_lock(&event->mmap_mutex);
3235 rcu_assign_pointer(event->rb, NULL);
3236 ring_buffer_detach(event, rb);
3237 ring_buffer_put(rb); /* could be last */
3239 mutex_unlock(&event->mmap_mutex);
3242 if (is_cgroup_event(event))
3243 perf_detach_cgroup(event);
3246 __free_event(event);
3249 int perf_event_release_kernel(struct perf_event *event)
3251 struct perf_event_context *ctx = event->ctx;
3253 WARN_ON_ONCE(ctx->parent_ctx);
3255 * There are two ways this annotation is useful:
3257 * 1) there is a lock recursion from perf_event_exit_task
3258 * see the comment there.
3260 * 2) there is a lock-inversion with mmap_sem through
3261 * perf_event_read_group(), which takes faults while
3262 * holding ctx->mutex, however this is called after
3263 * the last filedesc died, so there is no possibility
3264 * to trigger the AB-BA case.
3266 mutex_lock_nested(&ctx->mutex, SINGLE_DEPTH_NESTING);
3267 raw_spin_lock_irq(&ctx->lock);
3268 perf_group_detach(event);
3269 raw_spin_unlock_irq(&ctx->lock);
3270 perf_remove_from_context(event);
3271 mutex_unlock(&ctx->mutex);
3277 EXPORT_SYMBOL_GPL(perf_event_release_kernel);
3280 * Called when the last reference to the file is gone.
3282 static void put_event(struct perf_event *event)
3284 struct task_struct *owner;
3286 if (!atomic_long_dec_and_test(&event->refcount))
3290 owner = ACCESS_ONCE(event->owner);
3292 * Matches the smp_wmb() in perf_event_exit_task(). If we observe
3293 * !owner it means the list deletion is complete and we can indeed
3294 * free this event, otherwise we need to serialize on
3295 * owner->perf_event_mutex.
3297 smp_read_barrier_depends();
3300 * Since delayed_put_task_struct() also drops the last
3301 * task reference we can safely take a new reference
3302 * while holding the rcu_read_lock().
3304 get_task_struct(owner);
3309 mutex_lock(&owner->perf_event_mutex);
3311 * We have to re-check the event->owner field, if it is cleared
3312 * we raced with perf_event_exit_task(), acquiring the mutex
3313 * ensured they're done, and we can proceed with freeing the
3317 list_del_init(&event->owner_entry);
3318 mutex_unlock(&owner->perf_event_mutex);
3319 put_task_struct(owner);
3322 perf_event_release_kernel(event);
3325 static int perf_release(struct inode *inode, struct file *file)
3327 put_event(file->private_data);
3331 u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running)
3333 struct perf_event *child;
3339 mutex_lock(&event->child_mutex);
3340 total += perf_event_read(event);
3341 *enabled += event->total_time_enabled +
3342 atomic64_read(&event->child_total_time_enabled);
3343 *running += event->total_time_running +
3344 atomic64_read(&event->child_total_time_running);
3346 list_for_each_entry(child, &event->child_list, child_list) {
3347 total += perf_event_read(child);
3348 *enabled += child->total_time_enabled;
3349 *running += child->total_time_running;
3351 mutex_unlock(&event->child_mutex);
3355 EXPORT_SYMBOL_GPL(perf_event_read_value);
3357 static int perf_event_read_group(struct perf_event *event,
3358 u64 read_format, char __user *buf)
3360 struct perf_event *leader = event->group_leader, *sub;
3361 int n = 0, size = 0, ret = -EFAULT;
3362 struct perf_event_context *ctx = leader->ctx;
3364 u64 count, enabled, running;
3366 mutex_lock(&ctx->mutex);
3367 count = perf_event_read_value(leader, &enabled, &running);
3369 values[n++] = 1 + leader->nr_siblings;
3370 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3371 values[n++] = enabled;
3372 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3373 values[n++] = running;
3374 values[n++] = count;
3375 if (read_format & PERF_FORMAT_ID)
3376 values[n++] = primary_event_id(leader);
3378 size = n * sizeof(u64);
3380 if (copy_to_user(buf, values, size))
3385 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
3388 values[n++] = perf_event_read_value(sub, &enabled, &running);
3389 if (read_format & PERF_FORMAT_ID)
3390 values[n++] = primary_event_id(sub);
3392 size = n * sizeof(u64);
3394 if (copy_to_user(buf + ret, values, size)) {
3402 mutex_unlock(&ctx->mutex);
3407 static int perf_event_read_one(struct perf_event *event,
3408 u64 read_format, char __user *buf)
3410 u64 enabled, running;
3414 values[n++] = perf_event_read_value(event, &enabled, &running);
3415 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3416 values[n++] = enabled;
3417 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3418 values[n++] = running;
3419 if (read_format & PERF_FORMAT_ID)
3420 values[n++] = primary_event_id(event);
3422 if (copy_to_user(buf, values, n * sizeof(u64)))
3425 return n * sizeof(u64);
3429 * Read the performance event - simple non blocking version for now
3432 perf_read_hw(struct perf_event *event, char __user *buf, size_t count)
3434 u64 read_format = event->attr.read_format;
3438 * Return end-of-file for a read on a event that is in
3439 * error state (i.e. because it was pinned but it couldn't be
3440 * scheduled on to the CPU at some point).
3442 if (event->state == PERF_EVENT_STATE_ERROR)
3445 if (count < event->read_size)
3448 WARN_ON_ONCE(event->ctx->parent_ctx);
3449 if (read_format & PERF_FORMAT_GROUP)
3450 ret = perf_event_read_group(event, read_format, buf);
3452 ret = perf_event_read_one(event, read_format, buf);
3458 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
3460 struct perf_event *event = file->private_data;
3462 return perf_read_hw(event, buf, count);
3465 static unsigned int perf_poll(struct file *file, poll_table *wait)
3467 struct perf_event *event = file->private_data;
3468 struct ring_buffer *rb;
3469 unsigned int events = POLL_HUP;
3472 * Pin the event->rb by taking event->mmap_mutex; otherwise
3473 * perf_event_set_output() can swizzle our rb and make us miss wakeups.
3475 mutex_lock(&event->mmap_mutex);
3478 events = atomic_xchg(&rb->poll, 0);
3479 mutex_unlock(&event->mmap_mutex);
3481 poll_wait(file, &event->waitq, wait);
3486 static void perf_event_reset(struct perf_event *event)
3488 (void)perf_event_read(event);
3489 local64_set(&event->count, 0);
3490 perf_event_update_userpage(event);
3494 * Holding the top-level event's child_mutex means that any
3495 * descendant process that has inherited this event will block
3496 * in sync_child_event if it goes to exit, thus satisfying the
3497 * task existence requirements of perf_event_enable/disable.
3499 static void perf_event_for_each_child(struct perf_event *event,
3500 void (*func)(struct perf_event *))
3502 struct perf_event *child;
3504 WARN_ON_ONCE(event->ctx->parent_ctx);
3505 mutex_lock(&event->child_mutex);
3507 list_for_each_entry(child, &event->child_list, child_list)
3509 mutex_unlock(&event->child_mutex);
3512 static void perf_event_for_each(struct perf_event *event,
3513 void (*func)(struct perf_event *))
3515 struct perf_event_context *ctx = event->ctx;
3516 struct perf_event *sibling;
3518 WARN_ON_ONCE(ctx->parent_ctx);
3519 mutex_lock(&ctx->mutex);
3520 event = event->group_leader;
3522 perf_event_for_each_child(event, func);
3523 list_for_each_entry(sibling, &event->sibling_list, group_entry)
3524 perf_event_for_each_child(sibling, func);
3525 mutex_unlock(&ctx->mutex);
3528 static int perf_event_period(struct perf_event *event, u64 __user *arg)
3530 struct perf_event_context *ctx = event->ctx;
3531 int ret = 0, active;
3534 if (!is_sampling_event(event))
3537 if (copy_from_user(&value, arg, sizeof(value)))
3543 raw_spin_lock_irq(&ctx->lock);
3544 if (event->attr.freq) {
3545 if (value > sysctl_perf_event_sample_rate) {
3550 event->attr.sample_freq = value;
3552 event->attr.sample_period = value;
3553 event->hw.sample_period = value;
3556 active = (event->state == PERF_EVENT_STATE_ACTIVE);
3558 perf_pmu_disable(ctx->pmu);
3559 event->pmu->stop(event, PERF_EF_UPDATE);
3562 local64_set(&event->hw.period_left, 0);
3565 event->pmu->start(event, PERF_EF_RELOAD);
3566 perf_pmu_enable(ctx->pmu);
3570 raw_spin_unlock_irq(&ctx->lock);
3575 static const struct file_operations perf_fops;
3577 static inline int perf_fget_light(int fd, struct fd *p)
3579 struct fd f = fdget(fd);
3583 if (f.file->f_op != &perf_fops) {
3591 static int perf_event_set_output(struct perf_event *event,
3592 struct perf_event *output_event);
3593 static int perf_event_set_filter(struct perf_event *event, void __user *arg);
3595 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
3597 struct perf_event *event = file->private_data;
3598 void (*func)(struct perf_event *);
3602 case PERF_EVENT_IOC_ENABLE:
3603 func = perf_event_enable;
3605 case PERF_EVENT_IOC_DISABLE:
3606 func = perf_event_disable;
3608 case PERF_EVENT_IOC_RESET:
3609 func = perf_event_reset;
3612 case PERF_EVENT_IOC_REFRESH:
3613 return perf_event_refresh(event, arg);
3615 case PERF_EVENT_IOC_PERIOD:
3616 return perf_event_period(event, (u64 __user *)arg);
3618 case PERF_EVENT_IOC_ID:
3620 u64 id = primary_event_id(event);
3622 if (copy_to_user((void __user *)arg, &id, sizeof(id)))
3627 case PERF_EVENT_IOC_SET_OUTPUT:
3631 struct perf_event *output_event;
3633 ret = perf_fget_light(arg, &output);
3636 output_event = output.file->private_data;
3637 ret = perf_event_set_output(event, output_event);
3640 ret = perf_event_set_output(event, NULL);
3645 case PERF_EVENT_IOC_SET_FILTER:
3646 return perf_event_set_filter(event, (void __user *)arg);
3652 if (flags & PERF_IOC_FLAG_GROUP)
3653 perf_event_for_each(event, func);
3655 perf_event_for_each_child(event, func);
3660 int perf_event_task_enable(void)
3662 struct perf_event *event;
3664 mutex_lock(¤t->perf_event_mutex);
3665 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
3666 perf_event_for_each_child(event, perf_event_enable);
3667 mutex_unlock(¤t->perf_event_mutex);
3672 int perf_event_task_disable(void)
3674 struct perf_event *event;
3676 mutex_lock(¤t->perf_event_mutex);
3677 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
3678 perf_event_for_each_child(event, perf_event_disable);
3679 mutex_unlock(¤t->perf_event_mutex);
3684 static int perf_event_index(struct perf_event *event)
3686 if (event->hw.state & PERF_HES_STOPPED)
3689 if (event->state != PERF_EVENT_STATE_ACTIVE)
3692 return event->pmu->event_idx(event);
3695 static void calc_timer_values(struct perf_event *event,
3702 *now = perf_clock();
3703 ctx_time = event->shadow_ctx_time + *now;
3704 *enabled = ctx_time - event->tstamp_enabled;
3705 *running = ctx_time - event->tstamp_running;
3708 static void perf_event_init_userpage(struct perf_event *event)
3710 struct perf_event_mmap_page *userpg;
3711 struct ring_buffer *rb;
3714 rb = rcu_dereference(event->rb);
3718 userpg = rb->user_page;
3720 /* Allow new userspace to detect that bit 0 is deprecated */
3721 userpg->cap_bit0_is_deprecated = 1;
3722 userpg->size = offsetof(struct perf_event_mmap_page, __reserved);
3728 void __weak arch_perf_update_userpage(struct perf_event_mmap_page *userpg, u64 now)
3733 * Callers need to ensure there can be no nesting of this function, otherwise
3734 * the seqlock logic goes bad. We can not serialize this because the arch
3735 * code calls this from NMI context.
3737 void perf_event_update_userpage(struct perf_event *event)
3739 struct perf_event_mmap_page *userpg;
3740 struct ring_buffer *rb;
3741 u64 enabled, running, now;
3744 rb = rcu_dereference(event->rb);
3749 * compute total_time_enabled, total_time_running
3750 * based on snapshot values taken when the event
3751 * was last scheduled in.
3753 * we cannot simply called update_context_time()
3754 * because of locking issue as we can be called in
3757 calc_timer_values(event, &now, &enabled, &running);
3759 userpg = rb->user_page;
3761 * Disable preemption so as to not let the corresponding user-space
3762 * spin too long if we get preempted.
3767 userpg->index = perf_event_index(event);
3768 userpg->offset = perf_event_count(event);
3770 userpg->offset -= local64_read(&event->hw.prev_count);
3772 userpg->time_enabled = enabled +
3773 atomic64_read(&event->child_total_time_enabled);
3775 userpg->time_running = running +
3776 atomic64_read(&event->child_total_time_running);
3778 arch_perf_update_userpage(userpg, now);
3787 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
3789 struct perf_event *event = vma->vm_file->private_data;
3790 struct ring_buffer *rb;
3791 int ret = VM_FAULT_SIGBUS;
3793 if (vmf->flags & FAULT_FLAG_MKWRITE) {
3794 if (vmf->pgoff == 0)
3800 rb = rcu_dereference(event->rb);
3804 if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE))
3807 vmf->page = perf_mmap_to_page(rb, vmf->pgoff);
3811 get_page(vmf->page);
3812 vmf->page->mapping = vma->vm_file->f_mapping;
3813 vmf->page->index = vmf->pgoff;
3822 static void ring_buffer_attach(struct perf_event *event,
3823 struct ring_buffer *rb)
3825 unsigned long flags;
3827 if (!list_empty(&event->rb_entry))
3830 spin_lock_irqsave(&rb->event_lock, flags);
3831 if (list_empty(&event->rb_entry))
3832 list_add(&event->rb_entry, &rb->event_list);
3833 spin_unlock_irqrestore(&rb->event_lock, flags);
3836 static void ring_buffer_detach(struct perf_event *event, struct ring_buffer *rb)
3838 unsigned long flags;
3840 if (list_empty(&event->rb_entry))
3843 spin_lock_irqsave(&rb->event_lock, flags);
3844 list_del_init(&event->rb_entry);
3845 wake_up_all(&event->waitq);
3846 spin_unlock_irqrestore(&rb->event_lock, flags);
3849 static void ring_buffer_wakeup(struct perf_event *event)
3851 struct ring_buffer *rb;
3854 rb = rcu_dereference(event->rb);
3856 list_for_each_entry_rcu(event, &rb->event_list, rb_entry)
3857 wake_up_all(&event->waitq);
3862 static void rb_free_rcu(struct rcu_head *rcu_head)
3864 struct ring_buffer *rb;
3866 rb = container_of(rcu_head, struct ring_buffer, rcu_head);
3870 static struct ring_buffer *ring_buffer_get(struct perf_event *event)
3872 struct ring_buffer *rb;
3875 rb = rcu_dereference(event->rb);
3877 if (!atomic_inc_not_zero(&rb->refcount))
3885 static void ring_buffer_put(struct ring_buffer *rb)
3887 if (!atomic_dec_and_test(&rb->refcount))
3890 WARN_ON_ONCE(!list_empty(&rb->event_list));
3892 call_rcu(&rb->rcu_head, rb_free_rcu);
3895 static void perf_mmap_open(struct vm_area_struct *vma)
3897 struct perf_event *event = vma->vm_file->private_data;
3899 atomic_inc(&event->mmap_count);
3900 atomic_inc(&event->rb->mmap_count);
3904 * A buffer can be mmap()ed multiple times; either directly through the same
3905 * event, or through other events by use of perf_event_set_output().
3907 * In order to undo the VM accounting done by perf_mmap() we need to destroy
3908 * the buffer here, where we still have a VM context. This means we need
3909 * to detach all events redirecting to us.
3911 static void perf_mmap_close(struct vm_area_struct *vma)
3913 struct perf_event *event = vma->vm_file->private_data;
3915 struct ring_buffer *rb = event->rb;
3916 struct user_struct *mmap_user = rb->mmap_user;
3917 int mmap_locked = rb->mmap_locked;
3918 unsigned long size = perf_data_size(rb);
3920 atomic_dec(&rb->mmap_count);
3922 if (!atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex))
3925 /* Detach current event from the buffer. */
3926 rcu_assign_pointer(event->rb, NULL);
3927 ring_buffer_detach(event, rb);
3928 mutex_unlock(&event->mmap_mutex);
3930 /* If there's still other mmap()s of this buffer, we're done. */
3931 if (atomic_read(&rb->mmap_count)) {
3932 ring_buffer_put(rb); /* can't be last */
3937 * No other mmap()s, detach from all other events that might redirect
3938 * into the now unreachable buffer. Somewhat complicated by the
3939 * fact that rb::event_lock otherwise nests inside mmap_mutex.
3943 list_for_each_entry_rcu(event, &rb->event_list, rb_entry) {
3944 if (!atomic_long_inc_not_zero(&event->refcount)) {
3946 * This event is en-route to free_event() which will
3947 * detach it and remove it from the list.
3953 mutex_lock(&event->mmap_mutex);
3955 * Check we didn't race with perf_event_set_output() which can
3956 * swizzle the rb from under us while we were waiting to
3957 * acquire mmap_mutex.
3959 * If we find a different rb; ignore this event, a next
3960 * iteration will no longer find it on the list. We have to
3961 * still restart the iteration to make sure we're not now
3962 * iterating the wrong list.
3964 if (event->rb == rb) {
3965 rcu_assign_pointer(event->rb, NULL);
3966 ring_buffer_detach(event, rb);
3967 ring_buffer_put(rb); /* can't be last, we still have one */
3969 mutex_unlock(&event->mmap_mutex);
3973 * Restart the iteration; either we're on the wrong list or
3974 * destroyed its integrity by doing a deletion.
3981 * It could be there's still a few 0-ref events on the list; they'll
3982 * get cleaned up by free_event() -- they'll also still have their
3983 * ref on the rb and will free it whenever they are done with it.
3985 * Aside from that, this buffer is 'fully' detached and unmapped,
3986 * undo the VM accounting.
3989 atomic_long_sub((size >> PAGE_SHIFT) + 1, &mmap_user->locked_vm);
3990 vma->vm_mm->pinned_vm -= mmap_locked;
3991 free_uid(mmap_user);
3993 ring_buffer_put(rb); /* could be last */
3996 static const struct vm_operations_struct perf_mmap_vmops = {
3997 .open = perf_mmap_open,
3998 .close = perf_mmap_close,
3999 .fault = perf_mmap_fault,
4000 .page_mkwrite = perf_mmap_fault,
4003 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
4005 struct perf_event *event = file->private_data;
4006 unsigned long user_locked, user_lock_limit;
4007 struct user_struct *user = current_user();
4008 unsigned long locked, lock_limit;
4009 struct ring_buffer *rb;
4010 unsigned long vma_size;
4011 unsigned long nr_pages;
4012 long user_extra, extra;
4013 int ret = 0, flags = 0;
4016 * Don't allow mmap() of inherited per-task counters. This would
4017 * create a performance issue due to all children writing to the
4020 if (event->cpu == -1 && event->attr.inherit)
4023 if (!(vma->vm_flags & VM_SHARED))
4026 vma_size = vma->vm_end - vma->vm_start;
4027 nr_pages = (vma_size / PAGE_SIZE) - 1;
4030 * If we have rb pages ensure they're a power-of-two number, so we
4031 * can do bitmasks instead of modulo.
4033 if (nr_pages != 0 && !is_power_of_2(nr_pages))
4036 if (vma_size != PAGE_SIZE * (1 + nr_pages))
4039 if (vma->vm_pgoff != 0)
4042 WARN_ON_ONCE(event->ctx->parent_ctx);
4044 mutex_lock(&event->mmap_mutex);
4046 if (event->rb->nr_pages != nr_pages) {
4051 if (!atomic_inc_not_zero(&event->rb->mmap_count)) {
4053 * Raced against perf_mmap_close() through
4054 * perf_event_set_output(). Try again, hope for better
4057 mutex_unlock(&event->mmap_mutex);
4064 user_extra = nr_pages + 1;
4065 user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10);
4068 * Increase the limit linearly with more CPUs:
4070 user_lock_limit *= num_online_cpus();
4072 user_locked = atomic_long_read(&user->locked_vm) + user_extra;
4075 if (user_locked > user_lock_limit)
4076 extra = user_locked - user_lock_limit;
4078 lock_limit = rlimit(RLIMIT_MEMLOCK);
4079 lock_limit >>= PAGE_SHIFT;
4080 locked = vma->vm_mm->pinned_vm + extra;
4082 if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() &&
4083 !capable(CAP_IPC_LOCK)) {
4090 if (vma->vm_flags & VM_WRITE)
4091 flags |= RING_BUFFER_WRITABLE;
4093 rb = rb_alloc(nr_pages,
4094 event->attr.watermark ? event->attr.wakeup_watermark : 0,
4102 atomic_set(&rb->mmap_count, 1);
4103 rb->mmap_locked = extra;
4104 rb->mmap_user = get_current_user();
4106 atomic_long_add(user_extra, &user->locked_vm);
4107 vma->vm_mm->pinned_vm += extra;
4109 ring_buffer_attach(event, rb);
4110 rcu_assign_pointer(event->rb, rb);
4112 perf_event_init_userpage(event);
4113 perf_event_update_userpage(event);
4117 atomic_inc(&event->mmap_count);
4118 mutex_unlock(&event->mmap_mutex);
4121 * Since pinned accounting is per vm we cannot allow fork() to copy our
4124 vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND | VM_DONTDUMP;
4125 vma->vm_ops = &perf_mmap_vmops;
4130 static int perf_fasync(int fd, struct file *filp, int on)
4132 struct inode *inode = file_inode(filp);
4133 struct perf_event *event = filp->private_data;
4136 mutex_lock(&inode->i_mutex);
4137 retval = fasync_helper(fd, filp, on, &event->fasync);
4138 mutex_unlock(&inode->i_mutex);
4146 static const struct file_operations perf_fops = {
4147 .llseek = no_llseek,
4148 .release = perf_release,
4151 .unlocked_ioctl = perf_ioctl,
4152 .compat_ioctl = perf_ioctl,
4154 .fasync = perf_fasync,
4160 * If there's data, ensure we set the poll() state and publish everything
4161 * to user-space before waking everybody up.
4164 void perf_event_wakeup(struct perf_event *event)
4166 ring_buffer_wakeup(event);
4168 if (event->pending_kill) {
4169 kill_fasync(&event->fasync, SIGIO, event->pending_kill);
4170 event->pending_kill = 0;
4174 static void perf_pending_event(struct irq_work *entry)
4176 struct perf_event *event = container_of(entry,
4177 struct perf_event, pending);
4179 if (event->pending_disable) {
4180 event->pending_disable = 0;
4181 __perf_event_disable(event);
4184 if (event->pending_wakeup) {
4185 event->pending_wakeup = 0;
4186 perf_event_wakeup(event);
4191 * We assume there is only KVM supporting the callbacks.
4192 * Later on, we might change it to a list if there is
4193 * another virtualization implementation supporting the callbacks.
4195 struct perf_guest_info_callbacks *perf_guest_cbs;
4197 int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
4199 perf_guest_cbs = cbs;
4202 EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks);
4204 int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
4206 perf_guest_cbs = NULL;
4209 EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks);
4212 perf_output_sample_regs(struct perf_output_handle *handle,
4213 struct pt_regs *regs, u64 mask)
4217 for_each_set_bit(bit, (const unsigned long *) &mask,
4218 sizeof(mask) * BITS_PER_BYTE) {
4221 val = perf_reg_value(regs, bit);
4222 perf_output_put(handle, val);
4226 static void perf_sample_regs_user(struct perf_regs_user *regs_user,
4227 struct pt_regs *regs)
4229 if (!user_mode(regs)) {
4231 regs = task_pt_regs(current);
4237 regs_user->regs = regs;
4238 regs_user->abi = perf_reg_abi(current);
4243 * Get remaining task size from user stack pointer.
4245 * It'd be better to take stack vma map and limit this more
4246 * precisly, but there's no way to get it safely under interrupt,
4247 * so using TASK_SIZE as limit.
4249 static u64 perf_ustack_task_size(struct pt_regs *regs)
4251 unsigned long addr = perf_user_stack_pointer(regs);
4253 if (!addr || addr >= TASK_SIZE)
4256 return TASK_SIZE - addr;
4260 perf_sample_ustack_size(u16 stack_size, u16 header_size,
4261 struct pt_regs *regs)
4265 /* No regs, no stack pointer, no dump. */
4270 * Check if we fit in with the requested stack size into the:
4272 * If we don't, we limit the size to the TASK_SIZE.
4274 * - remaining sample size
4275 * If we don't, we customize the stack size to
4276 * fit in to the remaining sample size.
4279 task_size = min((u64) USHRT_MAX, perf_ustack_task_size(regs));
4280 stack_size = min(stack_size, (u16) task_size);
4282 /* Current header size plus static size and dynamic size. */
4283 header_size += 2 * sizeof(u64);
4285 /* Do we fit in with the current stack dump size? */
4286 if ((u16) (header_size + stack_size) < header_size) {
4288 * If we overflow the maximum size for the sample,
4289 * we customize the stack dump size to fit in.
4291 stack_size = USHRT_MAX - header_size - sizeof(u64);
4292 stack_size = round_up(stack_size, sizeof(u64));
4299 perf_output_sample_ustack(struct perf_output_handle *handle, u64 dump_size,
4300 struct pt_regs *regs)
4302 /* Case of a kernel thread, nothing to dump */
4305 perf_output_put(handle, size);
4314 * - the size requested by user or the best one we can fit
4315 * in to the sample max size
4317 * - user stack dump data
4319 * - the actual dumped size
4323 perf_output_put(handle, dump_size);
4326 sp = perf_user_stack_pointer(regs);
4327 rem = __output_copy_user(handle, (void *) sp, dump_size);
4328 dyn_size = dump_size - rem;
4330 perf_output_skip(handle, rem);
4333 perf_output_put(handle, dyn_size);
4337 static void __perf_event_header__init_id(struct perf_event_header *header,
4338 struct perf_sample_data *data,
4339 struct perf_event *event)
4341 u64 sample_type = event->attr.sample_type;
4343 data->type = sample_type;
4344 header->size += event->id_header_size;
4346 if (sample_type & PERF_SAMPLE_TID) {
4347 /* namespace issues */
4348 data->tid_entry.pid = perf_event_pid(event, current);
4349 data->tid_entry.tid = perf_event_tid(event, current);
4352 if (sample_type & PERF_SAMPLE_TIME)
4353 data->time = perf_clock();
4355 if (sample_type & (PERF_SAMPLE_ID | PERF_SAMPLE_IDENTIFIER))
4356 data->id = primary_event_id(event);
4358 if (sample_type & PERF_SAMPLE_STREAM_ID)
4359 data->stream_id = event->id;
4361 if (sample_type & PERF_SAMPLE_CPU) {
4362 data->cpu_entry.cpu = raw_smp_processor_id();
4363 data->cpu_entry.reserved = 0;
4367 void perf_event_header__init_id(struct perf_event_header *header,
4368 struct perf_sample_data *data,
4369 struct perf_event *event)
4371 if (event->attr.sample_id_all)
4372 __perf_event_header__init_id(header, data, event);
4375 static void __perf_event__output_id_sample(struct perf_output_handle *handle,
4376 struct perf_sample_data *data)
4378 u64 sample_type = data->type;
4380 if (sample_type & PERF_SAMPLE_TID)
4381 perf_output_put(handle, data->tid_entry);
4383 if (sample_type & PERF_SAMPLE_TIME)
4384 perf_output_put(handle, data->time);
4386 if (sample_type & PERF_SAMPLE_ID)
4387 perf_output_put(handle, data->id);
4389 if (sample_type & PERF_SAMPLE_STREAM_ID)
4390 perf_output_put(handle, data->stream_id);
4392 if (sample_type & PERF_SAMPLE_CPU)
4393 perf_output_put(handle, data->cpu_entry);
4395 if (sample_type & PERF_SAMPLE_IDENTIFIER)
4396 perf_output_put(handle, data->id);
4399 void perf_event__output_id_sample(struct perf_event *event,
4400 struct perf_output_handle *handle,
4401 struct perf_sample_data *sample)
4403 if (event->attr.sample_id_all)
4404 __perf_event__output_id_sample(handle, sample);
4407 static void perf_output_read_one(struct perf_output_handle *handle,
4408 struct perf_event *event,
4409 u64 enabled, u64 running)
4411 u64 read_format = event->attr.read_format;
4415 values[n++] = perf_event_count(event);
4416 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
4417 values[n++] = enabled +
4418 atomic64_read(&event->child_total_time_enabled);
4420 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
4421 values[n++] = running +
4422 atomic64_read(&event->child_total_time_running);
4424 if (read_format & PERF_FORMAT_ID)
4425 values[n++] = primary_event_id(event);
4427 __output_copy(handle, values, n * sizeof(u64));
4431 * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
4433 static void perf_output_read_group(struct perf_output_handle *handle,
4434 struct perf_event *event,
4435 u64 enabled, u64 running)
4437 struct perf_event *leader = event->group_leader, *sub;
4438 u64 read_format = event->attr.read_format;
4442 values[n++] = 1 + leader->nr_siblings;
4444 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
4445 values[n++] = enabled;
4447 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
4448 values[n++] = running;
4450 if (leader != event)
4451 leader->pmu->read(leader);
4453 values[n++] = perf_event_count(leader);
4454 if (read_format & PERF_FORMAT_ID)
4455 values[n++] = primary_event_id(leader);
4457 __output_copy(handle, values, n * sizeof(u64));
4459 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
4462 if ((sub != event) &&
4463 (sub->state == PERF_EVENT_STATE_ACTIVE))
4464 sub->pmu->read(sub);
4466 values[n++] = perf_event_count(sub);
4467 if (read_format & PERF_FORMAT_ID)
4468 values[n++] = primary_event_id(sub);
4470 __output_copy(handle, values, n * sizeof(u64));
4474 #define PERF_FORMAT_TOTAL_TIMES (PERF_FORMAT_TOTAL_TIME_ENABLED|\
4475 PERF_FORMAT_TOTAL_TIME_RUNNING)
4477 static void perf_output_read(struct perf_output_handle *handle,
4478 struct perf_event *event)
4480 u64 enabled = 0, running = 0, now;
4481 u64 read_format = event->attr.read_format;
4484 * compute total_time_enabled, total_time_running
4485 * based on snapshot values taken when the event
4486 * was last scheduled in.
4488 * we cannot simply called update_context_time()
4489 * because of locking issue as we are called in
4492 if (read_format & PERF_FORMAT_TOTAL_TIMES)
4493 calc_timer_values(event, &now, &enabled, &running);
4495 if (event->attr.read_format & PERF_FORMAT_GROUP)
4496 perf_output_read_group(handle, event, enabled, running);
4498 perf_output_read_one(handle, event, enabled, running);
4501 void perf_output_sample(struct perf_output_handle *handle,
4502 struct perf_event_header *header,
4503 struct perf_sample_data *data,
4504 struct perf_event *event)
4506 u64 sample_type = data->type;
4508 perf_output_put(handle, *header);
4510 if (sample_type & PERF_SAMPLE_IDENTIFIER)
4511 perf_output_put(handle, data->id);
4513 if (sample_type & PERF_SAMPLE_IP)
4514 perf_output_put(handle, data->ip);
4516 if (sample_type & PERF_SAMPLE_TID)
4517 perf_output_put(handle, data->tid_entry);
4519 if (sample_type & PERF_SAMPLE_TIME)
4520 perf_output_put(handle, data->time);
4522 if (sample_type & PERF_SAMPLE_ADDR)
4523 perf_output_put(handle, data->addr);
4525 if (sample_type & PERF_SAMPLE_ID)
4526 perf_output_put(handle, data->id);
4528 if (sample_type & PERF_SAMPLE_STREAM_ID)
4529 perf_output_put(handle, data->stream_id);
4531 if (sample_type & PERF_SAMPLE_CPU)
4532 perf_output_put(handle, data->cpu_entry);
4534 if (sample_type & PERF_SAMPLE_PERIOD)
4535 perf_output_put(handle, data->period);
4537 if (sample_type & PERF_SAMPLE_READ)
4538 perf_output_read(handle, event);
4540 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
4541 if (data->callchain) {
4544 if (data->callchain)
4545 size += data->callchain->nr;
4547 size *= sizeof(u64);
4549 __output_copy(handle, data->callchain, size);
4552 perf_output_put(handle, nr);
4556 if (sample_type & PERF_SAMPLE_RAW) {
4558 perf_output_put(handle, data->raw->size);
4559 __output_copy(handle, data->raw->data,
4566 .size = sizeof(u32),
4569 perf_output_put(handle, raw);
4573 if (sample_type & PERF_SAMPLE_BRANCH_STACK) {
4574 if (data->br_stack) {
4577 size = data->br_stack->nr
4578 * sizeof(struct perf_branch_entry);
4580 perf_output_put(handle, data->br_stack->nr);
4581 perf_output_copy(handle, data->br_stack->entries, size);
4584 * we always store at least the value of nr
4587 perf_output_put(handle, nr);
4591 if (sample_type & PERF_SAMPLE_REGS_USER) {
4592 u64 abi = data->regs_user.abi;
4595 * If there are no regs to dump, notice it through
4596 * first u64 being zero (PERF_SAMPLE_REGS_ABI_NONE).
4598 perf_output_put(handle, abi);
4601 u64 mask = event->attr.sample_regs_user;
4602 perf_output_sample_regs(handle,
4603 data->regs_user.regs,
4608 if (sample_type & PERF_SAMPLE_STACK_USER) {
4609 perf_output_sample_ustack(handle,
4610 data->stack_user_size,
4611 data->regs_user.regs);
4614 if (sample_type & PERF_SAMPLE_WEIGHT)
4615 perf_output_put(handle, data->weight);
4617 if (sample_type & PERF_SAMPLE_DATA_SRC)
4618 perf_output_put(handle, data->data_src.val);
4620 if (sample_type & PERF_SAMPLE_TRANSACTION)
4621 perf_output_put(handle, data->txn);
4623 if (!event->attr.watermark) {
4624 int wakeup_events = event->attr.wakeup_events;
4626 if (wakeup_events) {
4627 struct ring_buffer *rb = handle->rb;
4628 int events = local_inc_return(&rb->events);
4630 if (events >= wakeup_events) {
4631 local_sub(wakeup_events, &rb->events);
4632 local_inc(&rb->wakeup);
4638 void perf_prepare_sample(struct perf_event_header *header,
4639 struct perf_sample_data *data,
4640 struct perf_event *event,
4641 struct pt_regs *regs)
4643 u64 sample_type = event->attr.sample_type;
4645 header->type = PERF_RECORD_SAMPLE;
4646 header->size = sizeof(*header) + event->header_size;
4649 header->misc |= perf_misc_flags(regs);
4651 __perf_event_header__init_id(header, data, event);
4653 if (sample_type & PERF_SAMPLE_IP)
4654 data->ip = perf_instruction_pointer(regs);
4656 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
4659 data->callchain = perf_callchain(event, regs);
4661 if (data->callchain)
4662 size += data->callchain->nr;
4664 header->size += size * sizeof(u64);
4667 if (sample_type & PERF_SAMPLE_RAW) {
4668 int size = sizeof(u32);
4671 size += data->raw->size;
4673 size += sizeof(u32);
4675 WARN_ON_ONCE(size & (sizeof(u64)-1));
4676 header->size += size;
4679 if (sample_type & PERF_SAMPLE_BRANCH_STACK) {
4680 int size = sizeof(u64); /* nr */
4681 if (data->br_stack) {
4682 size += data->br_stack->nr
4683 * sizeof(struct perf_branch_entry);
4685 header->size += size;
4688 if (sample_type & PERF_SAMPLE_REGS_USER) {
4689 /* regs dump ABI info */
4690 int size = sizeof(u64);
4692 perf_sample_regs_user(&data->regs_user, regs);
4694 if (data->regs_user.regs) {
4695 u64 mask = event->attr.sample_regs_user;
4696 size += hweight64(mask) * sizeof(u64);
4699 header->size += size;
4702 if (sample_type & PERF_SAMPLE_STACK_USER) {
4704 * Either we need PERF_SAMPLE_STACK_USER bit to be allways
4705 * processed as the last one or have additional check added
4706 * in case new sample type is added, because we could eat
4707 * up the rest of the sample size.
4709 struct perf_regs_user *uregs = &data->regs_user;
4710 u16 stack_size = event->attr.sample_stack_user;
4711 u16 size = sizeof(u64);
4714 perf_sample_regs_user(uregs, regs);
4716 stack_size = perf_sample_ustack_size(stack_size, header->size,
4720 * If there is something to dump, add space for the dump
4721 * itself and for the field that tells the dynamic size,
4722 * which is how many have been actually dumped.
4725 size += sizeof(u64) + stack_size;
4727 data->stack_user_size = stack_size;
4728 header->size += size;
4732 static void perf_event_output(struct perf_event *event,
4733 struct perf_sample_data *data,
4734 struct pt_regs *regs)
4736 struct perf_output_handle handle;
4737 struct perf_event_header header;
4739 /* protect the callchain buffers */
4742 perf_prepare_sample(&header, data, event, regs);
4744 if (perf_output_begin(&handle, event, header.size))
4747 perf_output_sample(&handle, &header, data, event);
4749 perf_output_end(&handle);
4759 struct perf_read_event {
4760 struct perf_event_header header;
4767 perf_event_read_event(struct perf_event *event,
4768 struct task_struct *task)
4770 struct perf_output_handle handle;
4771 struct perf_sample_data sample;
4772 struct perf_read_event read_event = {
4774 .type = PERF_RECORD_READ,
4776 .size = sizeof(read_event) + event->read_size,
4778 .pid = perf_event_pid(event, task),
4779 .tid = perf_event_tid(event, task),
4783 perf_event_header__init_id(&read_event.header, &sample, event);
4784 ret = perf_output_begin(&handle, event, read_event.header.size);
4788 perf_output_put(&handle, read_event);
4789 perf_output_read(&handle, event);
4790 perf_event__output_id_sample(event, &handle, &sample);
4792 perf_output_end(&handle);
4795 typedef void (perf_event_aux_output_cb)(struct perf_event *event, void *data);
4798 perf_event_aux_ctx(struct perf_event_context *ctx,
4799 perf_event_aux_output_cb output,
4802 struct perf_event *event;
4804 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
4805 if (event->state < PERF_EVENT_STATE_INACTIVE)
4807 if (!event_filter_match(event))
4809 output(event, data);
4814 perf_event_aux(perf_event_aux_output_cb output, void *data,
4815 struct perf_event_context *task_ctx)
4817 struct perf_cpu_context *cpuctx;
4818 struct perf_event_context *ctx;
4823 list_for_each_entry_rcu(pmu, &pmus, entry) {
4824 cpuctx = get_cpu_ptr(pmu->pmu_cpu_context);
4825 if (cpuctx->unique_pmu != pmu)
4827 perf_event_aux_ctx(&cpuctx->ctx, output, data);
4830 ctxn = pmu->task_ctx_nr;
4833 ctx = rcu_dereference(current->perf_event_ctxp[ctxn]);
4835 perf_event_aux_ctx(ctx, output, data);
4837 put_cpu_ptr(pmu->pmu_cpu_context);
4842 perf_event_aux_ctx(task_ctx, output, data);
4849 * task tracking -- fork/exit
4851 * enabled by: attr.comm | attr.mmap | attr.mmap2 | attr.mmap_data | attr.task
4854 struct perf_task_event {
4855 struct task_struct *task;
4856 struct perf_event_context *task_ctx;
4859 struct perf_event_header header;
4869 static int perf_event_task_match(struct perf_event *event)
4871 return event->attr.comm || event->attr.mmap ||
4872 event->attr.mmap2 || event->attr.mmap_data ||
4876 static void perf_event_task_output(struct perf_event *event,
4879 struct perf_task_event *task_event = data;
4880 struct perf_output_handle handle;
4881 struct perf_sample_data sample;
4882 struct task_struct *task = task_event->task;
4883 int ret, size = task_event->event_id.header.size;
4885 if (!perf_event_task_match(event))
4888 perf_event_header__init_id(&task_event->event_id.header, &sample, event);
4890 ret = perf_output_begin(&handle, event,
4891 task_event->event_id.header.size);
4895 task_event->event_id.pid = perf_event_pid(event, task);
4896 task_event->event_id.ppid = perf_event_pid(event, current);
4898 task_event->event_id.tid = perf_event_tid(event, task);
4899 task_event->event_id.ptid = perf_event_tid(event, current);
4901 perf_output_put(&handle, task_event->event_id);
4903 perf_event__output_id_sample(event, &handle, &sample);
4905 perf_output_end(&handle);
4907 task_event->event_id.header.size = size;
4910 static void perf_event_task(struct task_struct *task,
4911 struct perf_event_context *task_ctx,
4914 struct perf_task_event task_event;
4916 if (!atomic_read(&nr_comm_events) &&
4917 !atomic_read(&nr_mmap_events) &&
4918 !atomic_read(&nr_task_events))
4921 task_event = (struct perf_task_event){
4923 .task_ctx = task_ctx,
4926 .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT,
4928 .size = sizeof(task_event.event_id),
4934 .time = perf_clock(),
4938 perf_event_aux(perf_event_task_output,
4943 void perf_event_fork(struct task_struct *task)
4945 perf_event_task(task, NULL, 1);
4952 struct perf_comm_event {
4953 struct task_struct *task;
4958 struct perf_event_header header;
4965 static int perf_event_comm_match(struct perf_event *event)
4967 return event->attr.comm;
4970 static void perf_event_comm_output(struct perf_event *event,
4973 struct perf_comm_event *comm_event = data;
4974 struct perf_output_handle handle;
4975 struct perf_sample_data sample;
4976 int size = comm_event->event_id.header.size;
4979 if (!perf_event_comm_match(event))
4982 perf_event_header__init_id(&comm_event->event_id.header, &sample, event);
4983 ret = perf_output_begin(&handle, event,
4984 comm_event->event_id.header.size);
4989 comm_event->event_id.pid = perf_event_pid(event, comm_event->task);
4990 comm_event->event_id.tid = perf_event_tid(event, comm_event->task);
4992 perf_output_put(&handle, comm_event->event_id);
4993 __output_copy(&handle, comm_event->comm,
4994 comm_event->comm_size);
4996 perf_event__output_id_sample(event, &handle, &sample);
4998 perf_output_end(&handle);
5000 comm_event->event_id.header.size = size;
5003 static void perf_event_comm_event(struct perf_comm_event *comm_event)
5005 char comm[TASK_COMM_LEN];
5008 memset(comm, 0, sizeof(comm));
5009 strlcpy(comm, comm_event->task->comm, sizeof(comm));
5010 size = ALIGN(strlen(comm)+1, sizeof(u64));
5012 comm_event->comm = comm;
5013 comm_event->comm_size = size;
5015 comm_event->event_id.header.size = sizeof(comm_event->event_id) + size;
5017 perf_event_aux(perf_event_comm_output,
5022 void perf_event_comm(struct task_struct *task)
5024 struct perf_comm_event comm_event;
5025 struct perf_event_context *ctx;
5029 for_each_task_context_nr(ctxn) {
5030 ctx = task->perf_event_ctxp[ctxn];
5034 perf_event_enable_on_exec(ctx);
5038 if (!atomic_read(&nr_comm_events))
5041 comm_event = (struct perf_comm_event){
5047 .type = PERF_RECORD_COMM,
5056 perf_event_comm_event(&comm_event);
5063 struct perf_mmap_event {
5064 struct vm_area_struct *vma;
5066 const char *file_name;
5073 struct perf_event_header header;
5083 static int perf_event_mmap_match(struct perf_event *event,
5086 struct perf_mmap_event *mmap_event = data;
5087 struct vm_area_struct *vma = mmap_event->vma;
5088 int executable = vma->vm_flags & VM_EXEC;
5090 return (!executable && event->attr.mmap_data) ||
5091 (executable && (event->attr.mmap || event->attr.mmap2));
5094 static void perf_event_mmap_output(struct perf_event *event,
5097 struct perf_mmap_event *mmap_event = data;
5098 struct perf_output_handle handle;
5099 struct perf_sample_data sample;
5100 int size = mmap_event->event_id.header.size;
5103 if (!perf_event_mmap_match(event, data))
5106 if (event->attr.mmap2) {
5107 mmap_event->event_id.header.type = PERF_RECORD_MMAP2;
5108 mmap_event->event_id.header.size += sizeof(mmap_event->maj);
5109 mmap_event->event_id.header.size += sizeof(mmap_event->min);
5110 mmap_event->event_id.header.size += sizeof(mmap_event->ino);
5111 mmap_event->event_id.header.size += sizeof(mmap_event->ino_generation);
5114 perf_event_header__init_id(&mmap_event->event_id.header, &sample, event);
5115 ret = perf_output_begin(&handle, event,
5116 mmap_event->event_id.header.size);
5120 mmap_event->event_id.pid = perf_event_pid(event, current);
5121 mmap_event->event_id.tid = perf_event_tid(event, current);
5123 perf_output_put(&handle, mmap_event->event_id);
5125 if (event->attr.mmap2) {
5126 perf_output_put(&handle, mmap_event->maj);
5127 perf_output_put(&handle, mmap_event->min);
5128 perf_output_put(&handle, mmap_event->ino);
5129 perf_output_put(&handle, mmap_event->ino_generation);
5132 __output_copy(&handle, mmap_event->file_name,
5133 mmap_event->file_size);
5135 perf_event__output_id_sample(event, &handle, &sample);
5137 perf_output_end(&handle);
5139 mmap_event->event_id.header.size = size;
5142 static void perf_event_mmap_event(struct perf_mmap_event *mmap_event)
5144 struct vm_area_struct *vma = mmap_event->vma;
5145 struct file *file = vma->vm_file;
5146 int maj = 0, min = 0;
5147 u64 ino = 0, gen = 0;
5154 struct inode *inode;
5157 buf = kmalloc(PATH_MAX, GFP_KERNEL);
5163 * d_path() works from the end of the rb backwards, so we
5164 * need to add enough zero bytes after the string to handle
5165 * the 64bit alignment we do later.
5167 name = d_path(&file->f_path, buf, PATH_MAX - sizeof(u64));
5172 inode = file_inode(vma->vm_file);
5173 dev = inode->i_sb->s_dev;
5175 gen = inode->i_generation;
5180 name = (char *)arch_vma_name(vma);
5184 if (vma->vm_start <= vma->vm_mm->start_brk &&
5185 vma->vm_end >= vma->vm_mm->brk) {
5189 if (vma->vm_start <= vma->vm_mm->start_stack &&
5190 vma->vm_end >= vma->vm_mm->start_stack) {
5200 strlcpy(tmp, name, sizeof(tmp));
5204 * Since our buffer works in 8 byte units we need to align our string
5205 * size to a multiple of 8. However, we must guarantee the tail end is
5206 * zero'd out to avoid leaking random bits to userspace.
5208 size = strlen(name)+1;
5209 while (!IS_ALIGNED(size, sizeof(u64)))
5210 name[size++] = '\0';
5212 mmap_event->file_name = name;
5213 mmap_event->file_size = size;
5214 mmap_event->maj = maj;
5215 mmap_event->min = min;
5216 mmap_event->ino = ino;
5217 mmap_event->ino_generation = gen;
5219 if (!(vma->vm_flags & VM_EXEC))
5220 mmap_event->event_id.header.misc |= PERF_RECORD_MISC_MMAP_DATA;
5222 mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size;
5224 perf_event_aux(perf_event_mmap_output,
5231 void perf_event_mmap(struct vm_area_struct *vma)
5233 struct perf_mmap_event mmap_event;
5235 if (!atomic_read(&nr_mmap_events))
5238 mmap_event = (struct perf_mmap_event){
5244 .type = PERF_RECORD_MMAP,
5245 .misc = PERF_RECORD_MISC_USER,
5250 .start = vma->vm_start,
5251 .len = vma->vm_end - vma->vm_start,
5252 .pgoff = (u64)vma->vm_pgoff << PAGE_SHIFT,
5254 /* .maj (attr_mmap2 only) */
5255 /* .min (attr_mmap2 only) */
5256 /* .ino (attr_mmap2 only) */
5257 /* .ino_generation (attr_mmap2 only) */
5260 perf_event_mmap_event(&mmap_event);
5264 * IRQ throttle logging
5267 static void perf_log_throttle(struct perf_event *event, int enable)
5269 struct perf_output_handle handle;
5270 struct perf_sample_data sample;
5274 struct perf_event_header header;
5278 } throttle_event = {
5280 .type = PERF_RECORD_THROTTLE,
5282 .size = sizeof(throttle_event),
5284 .time = perf_clock(),
5285 .id = primary_event_id(event),
5286 .stream_id = event->id,
5290 throttle_event.header.type = PERF_RECORD_UNTHROTTLE;
5292 perf_event_header__init_id(&throttle_event.header, &sample, event);
5294 ret = perf_output_begin(&handle, event,
5295 throttle_event.header.size);
5299 perf_output_put(&handle, throttle_event);
5300 perf_event__output_id_sample(event, &handle, &sample);
5301 perf_output_end(&handle);
5305 * Generic event overflow handling, sampling.
5308 static int __perf_event_overflow(struct perf_event *event,
5309 int throttle, struct perf_sample_data *data,
5310 struct pt_regs *regs)
5312 int events = atomic_read(&event->event_limit);
5313 struct hw_perf_event *hwc = &event->hw;
5318 * Non-sampling counters might still use the PMI to fold short
5319 * hardware counters, ignore those.
5321 if (unlikely(!is_sampling_event(event)))
5324 seq = __this_cpu_read(perf_throttled_seq);
5325 if (seq != hwc->interrupts_seq) {
5326 hwc->interrupts_seq = seq;
5327 hwc->interrupts = 1;
5330 if (unlikely(throttle
5331 && hwc->interrupts >= max_samples_per_tick)) {
5332 __this_cpu_inc(perf_throttled_count);
5333 hwc->interrupts = MAX_INTERRUPTS;
5334 perf_log_throttle(event, 0);
5335 tick_nohz_full_kick();
5340 if (event->attr.freq) {
5341 u64 now = perf_clock();
5342 s64 delta = now - hwc->freq_time_stamp;
5344 hwc->freq_time_stamp = now;
5346 if (delta > 0 && delta < 2*TICK_NSEC)
5347 perf_adjust_period(event, delta, hwc->last_period, true);
5351 * XXX event_limit might not quite work as expected on inherited
5355 event->pending_kill = POLL_IN;
5356 if (events && atomic_dec_and_test(&event->event_limit)) {
5358 event->pending_kill = POLL_HUP;
5359 event->pending_disable = 1;
5360 irq_work_queue(&event->pending);
5363 if (event->overflow_handler)
5364 event->overflow_handler(event, data, regs);
5366 perf_event_output(event, data, regs);
5368 if (event->fasync && event->pending_kill) {
5369 event->pending_wakeup = 1;
5370 irq_work_queue(&event->pending);
5376 int perf_event_overflow(struct perf_event *event,
5377 struct perf_sample_data *data,
5378 struct pt_regs *regs)
5380 return __perf_event_overflow(event, 1, data, regs);
5384 * Generic software event infrastructure
5387 struct swevent_htable {
5388 struct swevent_hlist *swevent_hlist;
5389 struct mutex hlist_mutex;
5392 /* Recursion avoidance in each contexts */
5393 int recursion[PERF_NR_CONTEXTS];
5396 static DEFINE_PER_CPU(struct swevent_htable, swevent_htable);
5399 * We directly increment event->count and keep a second value in
5400 * event->hw.period_left to count intervals. This period event
5401 * is kept in the range [-sample_period, 0] so that we can use the
5405 u64 perf_swevent_set_period(struct perf_event *event)
5407 struct hw_perf_event *hwc = &event->hw;
5408 u64 period = hwc->last_period;
5412 hwc->last_period = hwc->sample_period;
5415 old = val = local64_read(&hwc->period_left);
5419 nr = div64_u64(period + val, period);
5420 offset = nr * period;
5422 if (local64_cmpxchg(&hwc->period_left, old, val) != old)
5428 static void perf_swevent_overflow(struct perf_event *event, u64 overflow,
5429 struct perf_sample_data *data,
5430 struct pt_regs *regs)
5432 struct hw_perf_event *hwc = &event->hw;
5436 overflow = perf_swevent_set_period(event);
5438 if (hwc->interrupts == MAX_INTERRUPTS)
5441 for (; overflow; overflow--) {
5442 if (__perf_event_overflow(event, throttle,
5445 * We inhibit the overflow from happening when
5446 * hwc->interrupts == MAX_INTERRUPTS.
5454 static void perf_swevent_event(struct perf_event *event, u64 nr,
5455 struct perf_sample_data *data,
5456 struct pt_regs *regs)
5458 struct hw_perf_event *hwc = &event->hw;
5460 local64_add(nr, &event->count);
5465 if (!is_sampling_event(event))
5468 if ((event->attr.sample_type & PERF_SAMPLE_PERIOD) && !event->attr.freq) {
5470 return perf_swevent_overflow(event, 1, data, regs);
5472 data->period = event->hw.last_period;
5474 if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq)
5475 return perf_swevent_overflow(event, 1, data, regs);
5477 if (local64_add_negative(nr, &hwc->period_left))
5480 perf_swevent_overflow(event, 0, data, regs);
5483 static int perf_exclude_event(struct perf_event *event,
5484 struct pt_regs *regs)
5486 if (event->hw.state & PERF_HES_STOPPED)
5490 if (event->attr.exclude_user && user_mode(regs))
5493 if (event->attr.exclude_kernel && !user_mode(regs))
5500 static int perf_swevent_match(struct perf_event *event,
5501 enum perf_type_id type,
5503 struct perf_sample_data *data,
5504 struct pt_regs *regs)
5506 if (event->attr.type != type)
5509 if (event->attr.config != event_id)
5512 if (perf_exclude_event(event, regs))
5518 static inline u64 swevent_hash(u64 type, u32 event_id)
5520 u64 val = event_id | (type << 32);
5522 return hash_64(val, SWEVENT_HLIST_BITS);
5525 static inline struct hlist_head *
5526 __find_swevent_head(struct swevent_hlist *hlist, u64 type, u32 event_id)
5528 u64 hash = swevent_hash(type, event_id);
5530 return &hlist->heads[hash];
5533 /* For the read side: events when they trigger */
5534 static inline struct hlist_head *
5535 find_swevent_head_rcu(struct swevent_htable *swhash, u64 type, u32 event_id)
5537 struct swevent_hlist *hlist;
5539 hlist = rcu_dereference(swhash->swevent_hlist);
5543 return __find_swevent_head(hlist, type, event_id);
5546 /* For the event head insertion and removal in the hlist */
5547 static inline struct hlist_head *
5548 find_swevent_head(struct swevent_htable *swhash, struct perf_event *event)
5550 struct swevent_hlist *hlist;
5551 u32 event_id = event->attr.config;
5552 u64 type = event->attr.type;
5555 * Event scheduling is always serialized against hlist allocation
5556 * and release. Which makes the protected version suitable here.
5557 * The context lock guarantees that.
5559 hlist = rcu_dereference_protected(swhash->swevent_hlist,
5560 lockdep_is_held(&event->ctx->lock));
5564 return __find_swevent_head(hlist, type, event_id);
5567 static void do_perf_sw_event(enum perf_type_id type, u32 event_id,
5569 struct perf_sample_data *data,
5570 struct pt_regs *regs)
5572 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
5573 struct perf_event *event;
5574 struct hlist_head *head;
5577 head = find_swevent_head_rcu(swhash, type, event_id);
5581 hlist_for_each_entry_rcu(event, head, hlist_entry) {
5582 if (perf_swevent_match(event, type, event_id, data, regs))
5583 perf_swevent_event(event, nr, data, regs);
5589 int perf_swevent_get_recursion_context(void)
5591 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
5593 return get_recursion_context(swhash->recursion);
5595 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context);
5597 inline void perf_swevent_put_recursion_context(int rctx)
5599 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
5601 put_recursion_context(swhash->recursion, rctx);
5604 void __perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)
5606 struct perf_sample_data data;
5609 preempt_disable_notrace();
5610 rctx = perf_swevent_get_recursion_context();
5614 perf_sample_data_init(&data, addr, 0);
5616 do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, &data, regs);
5618 perf_swevent_put_recursion_context(rctx);
5619 preempt_enable_notrace();
5622 static void perf_swevent_read(struct perf_event *event)
5626 static int perf_swevent_add(struct perf_event *event, int flags)
5628 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
5629 struct hw_perf_event *hwc = &event->hw;
5630 struct hlist_head *head;
5632 if (is_sampling_event(event)) {
5633 hwc->last_period = hwc->sample_period;
5634 perf_swevent_set_period(event);
5637 hwc->state = !(flags & PERF_EF_START);
5639 head = find_swevent_head(swhash, event);
5640 if (WARN_ON_ONCE(!head))
5643 hlist_add_head_rcu(&event->hlist_entry, head);
5648 static void perf_swevent_del(struct perf_event *event, int flags)
5650 hlist_del_rcu(&event->hlist_entry);
5653 static void perf_swevent_start(struct perf_event *event, int flags)
5655 event->hw.state = 0;
5658 static void perf_swevent_stop(struct perf_event *event, int flags)
5660 event->hw.state = PERF_HES_STOPPED;
5663 /* Deref the hlist from the update side */
5664 static inline struct swevent_hlist *
5665 swevent_hlist_deref(struct swevent_htable *swhash)
5667 return rcu_dereference_protected(swhash->swevent_hlist,
5668 lockdep_is_held(&swhash->hlist_mutex));
5671 static void swevent_hlist_release(struct swevent_htable *swhash)
5673 struct swevent_hlist *hlist = swevent_hlist_deref(swhash);
5678 rcu_assign_pointer(swhash->swevent_hlist, NULL);
5679 kfree_rcu(hlist, rcu_head);
5682 static void swevent_hlist_put_cpu(struct perf_event *event, int cpu)
5684 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
5686 mutex_lock(&swhash->hlist_mutex);
5688 if (!--swhash->hlist_refcount)
5689 swevent_hlist_release(swhash);
5691 mutex_unlock(&swhash->hlist_mutex);
5694 static void swevent_hlist_put(struct perf_event *event)
5698 for_each_possible_cpu(cpu)
5699 swevent_hlist_put_cpu(event, cpu);
5702 static int swevent_hlist_get_cpu(struct perf_event *event, int cpu)
5704 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
5707 mutex_lock(&swhash->hlist_mutex);
5709 if (!swevent_hlist_deref(swhash) && cpu_online(cpu)) {
5710 struct swevent_hlist *hlist;
5712 hlist = kzalloc(sizeof(*hlist), GFP_KERNEL);
5717 rcu_assign_pointer(swhash->swevent_hlist, hlist);
5719 swhash->hlist_refcount++;
5721 mutex_unlock(&swhash->hlist_mutex);
5726 static int swevent_hlist_get(struct perf_event *event)
5729 int cpu, failed_cpu;
5732 for_each_possible_cpu(cpu) {
5733 err = swevent_hlist_get_cpu(event, cpu);
5743 for_each_possible_cpu(cpu) {
5744 if (cpu == failed_cpu)
5746 swevent_hlist_put_cpu(event, cpu);
5753 struct static_key perf_swevent_enabled[PERF_COUNT_SW_MAX];
5755 static void sw_perf_event_destroy(struct perf_event *event)
5757 u64 event_id = event->attr.config;
5759 WARN_ON(event->parent);
5761 static_key_slow_dec(&perf_swevent_enabled[event_id]);
5762 swevent_hlist_put(event);
5765 static int perf_swevent_init(struct perf_event *event)
5767 u64 event_id = event->attr.config;
5769 if (event->attr.type != PERF_TYPE_SOFTWARE)
5773 * no branch sampling for software events
5775 if (has_branch_stack(event))
5779 case PERF_COUNT_SW_CPU_CLOCK:
5780 case PERF_COUNT_SW_TASK_CLOCK:
5787 if (event_id >= PERF_COUNT_SW_MAX)
5790 if (!event->parent) {
5793 err = swevent_hlist_get(event);
5797 static_key_slow_inc(&perf_swevent_enabled[event_id]);
5798 event->destroy = sw_perf_event_destroy;
5804 static int perf_swevent_event_idx(struct perf_event *event)
5809 static struct pmu perf_swevent = {
5810 .task_ctx_nr = perf_sw_context,
5812 .event_init = perf_swevent_init,
5813 .add = perf_swevent_add,
5814 .del = perf_swevent_del,
5815 .start = perf_swevent_start,
5816 .stop = perf_swevent_stop,
5817 .read = perf_swevent_read,
5819 .event_idx = perf_swevent_event_idx,
5822 #ifdef CONFIG_EVENT_TRACING
5824 static int perf_tp_filter_match(struct perf_event *event,
5825 struct perf_sample_data *data)
5827 void *record = data->raw->data;
5829 if (likely(!event->filter) || filter_match_preds(event->filter, record))
5834 static int perf_tp_event_match(struct perf_event *event,
5835 struct perf_sample_data *data,
5836 struct pt_regs *regs)
5838 if (event->hw.state & PERF_HES_STOPPED)
5841 * All tracepoints are from kernel-space.
5843 if (event->attr.exclude_kernel)
5846 if (!perf_tp_filter_match(event, data))
5852 void perf_tp_event(u64 addr, u64 count, void *record, int entry_size,
5853 struct pt_regs *regs, struct hlist_head *head, int rctx,
5854 struct task_struct *task)
5856 struct perf_sample_data data;
5857 struct perf_event *event;
5859 struct perf_raw_record raw = {
5864 perf_sample_data_init(&data, addr, 0);
5867 hlist_for_each_entry_rcu(event, head, hlist_entry) {
5868 if (perf_tp_event_match(event, &data, regs))
5869 perf_swevent_event(event, count, &data, regs);
5873 * If we got specified a target task, also iterate its context and
5874 * deliver this event there too.
5876 if (task && task != current) {
5877 struct perf_event_context *ctx;
5878 struct trace_entry *entry = record;
5881 ctx = rcu_dereference(task->perf_event_ctxp[perf_sw_context]);
5885 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
5886 if (event->attr.type != PERF_TYPE_TRACEPOINT)
5888 if (event->attr.config != entry->type)
5890 if (perf_tp_event_match(event, &data, regs))
5891 perf_swevent_event(event, count, &data, regs);
5897 perf_swevent_put_recursion_context(rctx);
5899 EXPORT_SYMBOL_GPL(perf_tp_event);
5901 static void tp_perf_event_destroy(struct perf_event *event)
5903 perf_trace_destroy(event);
5906 static int perf_tp_event_init(struct perf_event *event)
5910 if (event->attr.type != PERF_TYPE_TRACEPOINT)
5914 * no branch sampling for tracepoint events
5916 if (has_branch_stack(event))
5919 err = perf_trace_init(event);
5923 event->destroy = tp_perf_event_destroy;
5928 static struct pmu perf_tracepoint = {
5929 .task_ctx_nr = perf_sw_context,
5931 .event_init = perf_tp_event_init,
5932 .add = perf_trace_add,
5933 .del = perf_trace_del,
5934 .start = perf_swevent_start,
5935 .stop = perf_swevent_stop,
5936 .read = perf_swevent_read,
5938 .event_idx = perf_swevent_event_idx,
5941 static inline void perf_tp_register(void)
5943 perf_pmu_register(&perf_tracepoint, "tracepoint", PERF_TYPE_TRACEPOINT);
5946 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
5951 if (event->attr.type != PERF_TYPE_TRACEPOINT)
5954 filter_str = strndup_user(arg, PAGE_SIZE);
5955 if (IS_ERR(filter_str))
5956 return PTR_ERR(filter_str);
5958 ret = ftrace_profile_set_filter(event, event->attr.config, filter_str);
5964 static void perf_event_free_filter(struct perf_event *event)
5966 ftrace_profile_free_filter(event);
5971 static inline void perf_tp_register(void)
5975 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
5980 static void perf_event_free_filter(struct perf_event *event)
5984 #endif /* CONFIG_EVENT_TRACING */
5986 #ifdef CONFIG_HAVE_HW_BREAKPOINT
5987 void perf_bp_event(struct perf_event *bp, void *data)
5989 struct perf_sample_data sample;
5990 struct pt_regs *regs = data;
5992 perf_sample_data_init(&sample, bp->attr.bp_addr, 0);
5994 if (!bp->hw.state && !perf_exclude_event(bp, regs))
5995 perf_swevent_event(bp, 1, &sample, regs);
6000 * hrtimer based swevent callback
6003 static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer)
6005 enum hrtimer_restart ret = HRTIMER_RESTART;
6006 struct perf_sample_data data;
6007 struct pt_regs *regs;
6008 struct perf_event *event;
6011 event = container_of(hrtimer, struct perf_event, hw.hrtimer);
6013 if (event->state != PERF_EVENT_STATE_ACTIVE)
6014 return HRTIMER_NORESTART;
6016 event->pmu->read(event);
6018 perf_sample_data_init(&data, 0, event->hw.last_period);
6019 regs = get_irq_regs();
6021 if (regs && !perf_exclude_event(event, regs)) {
6022 if (!(event->attr.exclude_idle && is_idle_task(current)))
6023 if (__perf_event_overflow(event, 1, &data, regs))
6024 ret = HRTIMER_NORESTART;
6027 period = max_t(u64, 10000, event->hw.sample_period);
6028 hrtimer_forward_now(hrtimer, ns_to_ktime(period));
6033 static void perf_swevent_start_hrtimer(struct perf_event *event)
6035 struct hw_perf_event *hwc = &event->hw;
6038 if (!is_sampling_event(event))
6041 period = local64_read(&hwc->period_left);
6046 local64_set(&hwc->period_left, 0);
6048 period = max_t(u64, 10000, hwc->sample_period);
6050 __hrtimer_start_range_ns(&hwc->hrtimer,
6051 ns_to_ktime(period), 0,
6052 HRTIMER_MODE_REL_PINNED, 0);
6055 static void perf_swevent_cancel_hrtimer(struct perf_event *event)
6057 struct hw_perf_event *hwc = &event->hw;
6059 if (is_sampling_event(event)) {
6060 ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer);
6061 local64_set(&hwc->period_left, ktime_to_ns(remaining));
6063 hrtimer_cancel(&hwc->hrtimer);
6067 static void perf_swevent_init_hrtimer(struct perf_event *event)
6069 struct hw_perf_event *hwc = &event->hw;
6071 if (!is_sampling_event(event))
6074 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
6075 hwc->hrtimer.function = perf_swevent_hrtimer;
6078 * Since hrtimers have a fixed rate, we can do a static freq->period
6079 * mapping and avoid the whole period adjust feedback stuff.
6081 if (event->attr.freq) {
6082 long freq = event->attr.sample_freq;
6084 event->attr.sample_period = NSEC_PER_SEC / freq;
6085 hwc->sample_period = event->attr.sample_period;
6086 local64_set(&hwc->period_left, hwc->sample_period);
6087 hwc->last_period = hwc->sample_period;
6088 event->attr.freq = 0;
6093 * Software event: cpu wall time clock
6096 static void cpu_clock_event_update(struct perf_event *event)
6101 now = local_clock();
6102 prev = local64_xchg(&event->hw.prev_count, now);
6103 local64_add(now - prev, &event->count);
6106 static void cpu_clock_event_start(struct perf_event *event, int flags)
6108 local64_set(&event->hw.prev_count, local_clock());
6109 perf_swevent_start_hrtimer(event);
6112 static void cpu_clock_event_stop(struct perf_event *event, int flags)
6114 perf_swevent_cancel_hrtimer(event);
6115 cpu_clock_event_update(event);
6118 static int cpu_clock_event_add(struct perf_event *event, int flags)
6120 if (flags & PERF_EF_START)
6121 cpu_clock_event_start(event, flags);
6126 static void cpu_clock_event_del(struct perf_event *event, int flags)
6128 cpu_clock_event_stop(event, flags);
6131 static void cpu_clock_event_read(struct perf_event *event)
6133 cpu_clock_event_update(event);
6136 static int cpu_clock_event_init(struct perf_event *event)
6138 if (event->attr.type != PERF_TYPE_SOFTWARE)
6141 if (event->attr.config != PERF_COUNT_SW_CPU_CLOCK)
6145 * no branch sampling for software events
6147 if (has_branch_stack(event))
6150 perf_swevent_init_hrtimer(event);
6155 static struct pmu perf_cpu_clock = {
6156 .task_ctx_nr = perf_sw_context,
6158 .event_init = cpu_clock_event_init,
6159 .add = cpu_clock_event_add,
6160 .del = cpu_clock_event_del,
6161 .start = cpu_clock_event_start,
6162 .stop = cpu_clock_event_stop,
6163 .read = cpu_clock_event_read,
6165 .event_idx = perf_swevent_event_idx,
6169 * Software event: task time clock
6172 static void task_clock_event_update(struct perf_event *event, u64 now)
6177 prev = local64_xchg(&event->hw.prev_count, now);
6179 local64_add(delta, &event->count);
6182 static void task_clock_event_start(struct perf_event *event, int flags)
6184 local64_set(&event->hw.prev_count, event->ctx->time);
6185 perf_swevent_start_hrtimer(event);
6188 static void task_clock_event_stop(struct perf_event *event, int flags)
6190 perf_swevent_cancel_hrtimer(event);
6191 task_clock_event_update(event, event->ctx->time);
6194 static int task_clock_event_add(struct perf_event *event, int flags)
6196 if (flags & PERF_EF_START)
6197 task_clock_event_start(event, flags);
6202 static void task_clock_event_del(struct perf_event *event, int flags)
6204 task_clock_event_stop(event, PERF_EF_UPDATE);
6207 static void task_clock_event_read(struct perf_event *event)
6209 u64 now = perf_clock();
6210 u64 delta = now - event->ctx->timestamp;
6211 u64 time = event->ctx->time + delta;
6213 task_clock_event_update(event, time);
6216 static int task_clock_event_init(struct perf_event *event)
6218 if (event->attr.type != PERF_TYPE_SOFTWARE)
6221 if (event->attr.config != PERF_COUNT_SW_TASK_CLOCK)
6225 * no branch sampling for software events
6227 if (has_branch_stack(event))
6230 perf_swevent_init_hrtimer(event);
6235 static struct pmu perf_task_clock = {
6236 .task_ctx_nr = perf_sw_context,
6238 .event_init = task_clock_event_init,
6239 .add = task_clock_event_add,
6240 .del = task_clock_event_del,
6241 .start = task_clock_event_start,
6242 .stop = task_clock_event_stop,
6243 .read = task_clock_event_read,
6245 .event_idx = perf_swevent_event_idx,
6248 static void perf_pmu_nop_void(struct pmu *pmu)
6252 static int perf_pmu_nop_int(struct pmu *pmu)
6257 static void perf_pmu_start_txn(struct pmu *pmu)
6259 perf_pmu_disable(pmu);
6262 static int perf_pmu_commit_txn(struct pmu *pmu)
6264 perf_pmu_enable(pmu);
6268 static void perf_pmu_cancel_txn(struct pmu *pmu)
6270 perf_pmu_enable(pmu);
6273 static int perf_event_idx_default(struct perf_event *event)
6275 return event->hw.idx + 1;
6279 * Ensures all contexts with the same task_ctx_nr have the same
6280 * pmu_cpu_context too.
6282 static void *find_pmu_context(int ctxn)
6289 list_for_each_entry(pmu, &pmus, entry) {
6290 if (pmu->task_ctx_nr == ctxn)
6291 return pmu->pmu_cpu_context;
6297 static void update_pmu_context(struct pmu *pmu, struct pmu *old_pmu)
6301 for_each_possible_cpu(cpu) {
6302 struct perf_cpu_context *cpuctx;
6304 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
6306 if (cpuctx->unique_pmu == old_pmu)
6307 cpuctx->unique_pmu = pmu;
6311 static void free_pmu_context(struct pmu *pmu)
6315 mutex_lock(&pmus_lock);
6317 * Like a real lame refcount.
6319 list_for_each_entry(i, &pmus, entry) {
6320 if (i->pmu_cpu_context == pmu->pmu_cpu_context) {
6321 update_pmu_context(i, pmu);
6326 free_percpu(pmu->pmu_cpu_context);
6328 mutex_unlock(&pmus_lock);
6330 static struct idr pmu_idr;
6333 type_show(struct device *dev, struct device_attribute *attr, char *page)
6335 struct pmu *pmu = dev_get_drvdata(dev);
6337 return snprintf(page, PAGE_SIZE-1, "%d\n", pmu->type);
6339 static DEVICE_ATTR_RO(type);
6342 perf_event_mux_interval_ms_show(struct device *dev,
6343 struct device_attribute *attr,
6346 struct pmu *pmu = dev_get_drvdata(dev);
6348 return snprintf(page, PAGE_SIZE-1, "%d\n", pmu->hrtimer_interval_ms);
6352 perf_event_mux_interval_ms_store(struct device *dev,
6353 struct device_attribute *attr,
6354 const char *buf, size_t count)
6356 struct pmu *pmu = dev_get_drvdata(dev);
6357 int timer, cpu, ret;
6359 ret = kstrtoint(buf, 0, &timer);
6366 /* same value, noting to do */
6367 if (timer == pmu->hrtimer_interval_ms)
6370 pmu->hrtimer_interval_ms = timer;
6372 /* update all cpuctx for this PMU */
6373 for_each_possible_cpu(cpu) {
6374 struct perf_cpu_context *cpuctx;
6375 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
6376 cpuctx->hrtimer_interval = ns_to_ktime(NSEC_PER_MSEC * timer);
6378 if (hrtimer_active(&cpuctx->hrtimer))
6379 hrtimer_forward_now(&cpuctx->hrtimer, cpuctx->hrtimer_interval);
6384 static DEVICE_ATTR_RW(perf_event_mux_interval_ms);
6386 static struct attribute *pmu_dev_attrs[] = {
6387 &dev_attr_type.attr,
6388 &dev_attr_perf_event_mux_interval_ms.attr,
6391 ATTRIBUTE_GROUPS(pmu_dev);
6393 static int pmu_bus_running;
6394 static struct bus_type pmu_bus = {
6395 .name = "event_source",
6396 .dev_groups = pmu_dev_groups,
6399 static void pmu_dev_release(struct device *dev)
6404 static int pmu_dev_alloc(struct pmu *pmu)
6408 pmu->dev = kzalloc(sizeof(struct device), GFP_KERNEL);
6412 pmu->dev->groups = pmu->attr_groups;
6413 device_initialize(pmu->dev);
6414 ret = dev_set_name(pmu->dev, "%s", pmu->name);
6418 dev_set_drvdata(pmu->dev, pmu);
6419 pmu->dev->bus = &pmu_bus;
6420 pmu->dev->release = pmu_dev_release;
6421 ret = device_add(pmu->dev);
6429 put_device(pmu->dev);
6433 static struct lock_class_key cpuctx_mutex;
6434 static struct lock_class_key cpuctx_lock;
6436 int perf_pmu_register(struct pmu *pmu, const char *name, int type)
6440 mutex_lock(&pmus_lock);
6442 pmu->pmu_disable_count = alloc_percpu(int);
6443 if (!pmu->pmu_disable_count)
6452 type = idr_alloc(&pmu_idr, pmu, PERF_TYPE_MAX, 0, GFP_KERNEL);
6460 if (pmu_bus_running) {
6461 ret = pmu_dev_alloc(pmu);
6467 pmu->pmu_cpu_context = find_pmu_context(pmu->task_ctx_nr);
6468 if (pmu->pmu_cpu_context)
6469 goto got_cpu_context;
6472 pmu->pmu_cpu_context = alloc_percpu(struct perf_cpu_context);
6473 if (!pmu->pmu_cpu_context)
6476 for_each_possible_cpu(cpu) {
6477 struct perf_cpu_context *cpuctx;
6479 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
6480 __perf_event_init_context(&cpuctx->ctx);
6481 lockdep_set_class(&cpuctx->ctx.mutex, &cpuctx_mutex);
6482 lockdep_set_class(&cpuctx->ctx.lock, &cpuctx_lock);
6483 cpuctx->ctx.type = cpu_context;
6484 cpuctx->ctx.pmu = pmu;
6486 __perf_cpu_hrtimer_init(cpuctx, cpu);
6488 INIT_LIST_HEAD(&cpuctx->rotation_list);
6489 cpuctx->unique_pmu = pmu;
6493 if (!pmu->start_txn) {
6494 if (pmu->pmu_enable) {
6496 * If we have pmu_enable/pmu_disable calls, install
6497 * transaction stubs that use that to try and batch
6498 * hardware accesses.
6500 pmu->start_txn = perf_pmu_start_txn;
6501 pmu->commit_txn = perf_pmu_commit_txn;
6502 pmu->cancel_txn = perf_pmu_cancel_txn;
6504 pmu->start_txn = perf_pmu_nop_void;
6505 pmu->commit_txn = perf_pmu_nop_int;
6506 pmu->cancel_txn = perf_pmu_nop_void;
6510 if (!pmu->pmu_enable) {
6511 pmu->pmu_enable = perf_pmu_nop_void;
6512 pmu->pmu_disable = perf_pmu_nop_void;
6515 if (!pmu->event_idx)
6516 pmu->event_idx = perf_event_idx_default;
6518 list_add_rcu(&pmu->entry, &pmus);
6521 mutex_unlock(&pmus_lock);
6526 device_del(pmu->dev);
6527 put_device(pmu->dev);
6530 if (pmu->type >= PERF_TYPE_MAX)
6531 idr_remove(&pmu_idr, pmu->type);
6534 free_percpu(pmu->pmu_disable_count);
6538 void perf_pmu_unregister(struct pmu *pmu)
6540 mutex_lock(&pmus_lock);
6541 list_del_rcu(&pmu->entry);
6542 mutex_unlock(&pmus_lock);
6545 * We dereference the pmu list under both SRCU and regular RCU, so
6546 * synchronize against both of those.
6548 synchronize_srcu(&pmus_srcu);
6551 free_percpu(pmu->pmu_disable_count);
6552 if (pmu->type >= PERF_TYPE_MAX)
6553 idr_remove(&pmu_idr, pmu->type);
6554 device_del(pmu->dev);
6555 put_device(pmu->dev);
6556 free_pmu_context(pmu);
6559 struct pmu *perf_init_event(struct perf_event *event)
6561 struct pmu *pmu = NULL;
6565 idx = srcu_read_lock(&pmus_srcu);
6568 pmu = idr_find(&pmu_idr, event->attr.type);
6572 ret = pmu->event_init(event);
6578 list_for_each_entry_rcu(pmu, &pmus, entry) {
6580 ret = pmu->event_init(event);
6584 if (ret != -ENOENT) {
6589 pmu = ERR_PTR(-ENOENT);
6591 srcu_read_unlock(&pmus_srcu, idx);
6596 static void account_event_cpu(struct perf_event *event, int cpu)
6601 if (has_branch_stack(event)) {
6602 if (!(event->attach_state & PERF_ATTACH_TASK))
6603 atomic_inc(&per_cpu(perf_branch_stack_events, cpu));
6605 if (is_cgroup_event(event))
6606 atomic_inc(&per_cpu(perf_cgroup_events, cpu));
6609 static void account_event(struct perf_event *event)
6614 if (event->attach_state & PERF_ATTACH_TASK)
6615 static_key_slow_inc(&perf_sched_events.key);
6616 if (event->attr.mmap || event->attr.mmap_data)
6617 atomic_inc(&nr_mmap_events);
6618 if (event->attr.comm)
6619 atomic_inc(&nr_comm_events);
6620 if (event->attr.task)
6621 atomic_inc(&nr_task_events);
6622 if (event->attr.freq) {
6623 if (atomic_inc_return(&nr_freq_events) == 1)
6624 tick_nohz_full_kick_all();
6626 if (has_branch_stack(event))
6627 static_key_slow_inc(&perf_sched_events.key);
6628 if (is_cgroup_event(event))
6629 static_key_slow_inc(&perf_sched_events.key);
6631 account_event_cpu(event, event->cpu);
6635 * Allocate and initialize a event structure
6637 static struct perf_event *
6638 perf_event_alloc(struct perf_event_attr *attr, int cpu,
6639 struct task_struct *task,
6640 struct perf_event *group_leader,
6641 struct perf_event *parent_event,
6642 perf_overflow_handler_t overflow_handler,
6646 struct perf_event *event;
6647 struct hw_perf_event *hwc;
6650 if ((unsigned)cpu >= nr_cpu_ids) {
6651 if (!task || cpu != -1)
6652 return ERR_PTR(-EINVAL);
6655 event = kzalloc(sizeof(*event), GFP_KERNEL);
6657 return ERR_PTR(-ENOMEM);
6660 * Single events are their own group leaders, with an
6661 * empty sibling list:
6664 group_leader = event;
6666 mutex_init(&event->child_mutex);
6667 INIT_LIST_HEAD(&event->child_list);
6669 INIT_LIST_HEAD(&event->group_entry);
6670 INIT_LIST_HEAD(&event->event_entry);
6671 INIT_LIST_HEAD(&event->sibling_list);
6672 INIT_LIST_HEAD(&event->rb_entry);
6673 INIT_LIST_HEAD(&event->active_entry);
6674 INIT_HLIST_NODE(&event->hlist_entry);
6677 init_waitqueue_head(&event->waitq);
6678 init_irq_work(&event->pending, perf_pending_event);
6680 mutex_init(&event->mmap_mutex);
6682 atomic_long_set(&event->refcount, 1);
6684 event->attr = *attr;
6685 event->group_leader = group_leader;
6689 event->parent = parent_event;
6691 event->ns = get_pid_ns(task_active_pid_ns(current));
6692 event->id = atomic64_inc_return(&perf_event_id);
6694 event->state = PERF_EVENT_STATE_INACTIVE;
6697 event->attach_state = PERF_ATTACH_TASK;
6699 if (attr->type == PERF_TYPE_TRACEPOINT)
6700 event->hw.tp_target = task;
6701 #ifdef CONFIG_HAVE_HW_BREAKPOINT
6703 * hw_breakpoint is a bit difficult here..
6705 else if (attr->type == PERF_TYPE_BREAKPOINT)
6706 event->hw.bp_target = task;
6710 if (!overflow_handler && parent_event) {
6711 overflow_handler = parent_event->overflow_handler;
6712 context = parent_event->overflow_handler_context;
6715 event->overflow_handler = overflow_handler;
6716 event->overflow_handler_context = context;
6718 perf_event__state_init(event);
6723 hwc->sample_period = attr->sample_period;
6724 if (attr->freq && attr->sample_freq)
6725 hwc->sample_period = 1;
6726 hwc->last_period = hwc->sample_period;
6728 local64_set(&hwc->period_left, hwc->sample_period);
6731 * we currently do not support PERF_FORMAT_GROUP on inherited events
6733 if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP))
6736 pmu = perf_init_event(event);
6739 else if (IS_ERR(pmu)) {
6744 if (!event->parent) {
6745 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) {
6746 err = get_callchain_buffers();
6756 event->destroy(event);
6759 put_pid_ns(event->ns);
6762 return ERR_PTR(err);
6765 static int perf_copy_attr(struct perf_event_attr __user *uattr,
6766 struct perf_event_attr *attr)
6771 if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0))
6775 * zero the full structure, so that a short copy will be nice.
6777 memset(attr, 0, sizeof(*attr));
6779 ret = get_user(size, &uattr->size);
6783 if (size > PAGE_SIZE) /* silly large */
6786 if (!size) /* abi compat */
6787 size = PERF_ATTR_SIZE_VER0;
6789 if (size < PERF_ATTR_SIZE_VER0)
6793 * If we're handed a bigger struct than we know of,
6794 * ensure all the unknown bits are 0 - i.e. new
6795 * user-space does not rely on any kernel feature
6796 * extensions we dont know about yet.
6798 if (size > sizeof(*attr)) {
6799 unsigned char __user *addr;
6800 unsigned char __user *end;
6803 addr = (void __user *)uattr + sizeof(*attr);
6804 end = (void __user *)uattr + size;
6806 for (; addr < end; addr++) {
6807 ret = get_user(val, addr);
6813 size = sizeof(*attr);
6816 ret = copy_from_user(attr, uattr, size);
6820 /* disabled for now */
6824 if (attr->__reserved_1)
6827 if (attr->sample_type & ~(PERF_SAMPLE_MAX-1))
6830 if (attr->read_format & ~(PERF_FORMAT_MAX-1))
6833 if (attr->sample_type & PERF_SAMPLE_BRANCH_STACK) {
6834 u64 mask = attr->branch_sample_type;
6836 /* only using defined bits */
6837 if (mask & ~(PERF_SAMPLE_BRANCH_MAX-1))
6840 /* at least one branch bit must be set */
6841 if (!(mask & ~PERF_SAMPLE_BRANCH_PLM_ALL))
6844 /* propagate priv level, when not set for branch */
6845 if (!(mask & PERF_SAMPLE_BRANCH_PLM_ALL)) {
6847 /* exclude_kernel checked on syscall entry */
6848 if (!attr->exclude_kernel)
6849 mask |= PERF_SAMPLE_BRANCH_KERNEL;
6851 if (!attr->exclude_user)
6852 mask |= PERF_SAMPLE_BRANCH_USER;
6854 if (!attr->exclude_hv)
6855 mask |= PERF_SAMPLE_BRANCH_HV;
6857 * adjust user setting (for HW filter setup)
6859 attr->branch_sample_type = mask;
6861 /* privileged levels capture (kernel, hv): check permissions */
6862 if ((mask & PERF_SAMPLE_BRANCH_PERM_PLM)
6863 && perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
6867 if (attr->sample_type & PERF_SAMPLE_REGS_USER) {
6868 ret = perf_reg_validate(attr->sample_regs_user);
6873 if (attr->sample_type & PERF_SAMPLE_STACK_USER) {
6874 if (!arch_perf_have_user_stack_dump())
6878 * We have __u32 type for the size, but so far
6879 * we can only use __u16 as maximum due to the
6880 * __u16 sample size limit.
6882 if (attr->sample_stack_user >= USHRT_MAX)
6884 else if (!IS_ALIGNED(attr->sample_stack_user, sizeof(u64)))
6892 put_user(sizeof(*attr), &uattr->size);
6898 perf_event_set_output(struct perf_event *event, struct perf_event *output_event)
6900 struct ring_buffer *rb = NULL, *old_rb = NULL;
6906 /* don't allow circular references */
6907 if (event == output_event)
6911 * Don't allow cross-cpu buffers
6913 if (output_event->cpu != event->cpu)
6917 * If its not a per-cpu rb, it must be the same task.
6919 if (output_event->cpu == -1 && output_event->ctx != event->ctx)
6923 mutex_lock(&event->mmap_mutex);
6924 /* Can't redirect output if we've got an active mmap() */
6925 if (atomic_read(&event->mmap_count))
6931 /* get the rb we want to redirect to */
6932 rb = ring_buffer_get(output_event);
6938 ring_buffer_detach(event, old_rb);
6941 ring_buffer_attach(event, rb);
6943 rcu_assign_pointer(event->rb, rb);
6946 ring_buffer_put(old_rb);
6948 * Since we detached before setting the new rb, so that we
6949 * could attach the new rb, we could have missed a wakeup.
6952 wake_up_all(&event->waitq);
6957 mutex_unlock(&event->mmap_mutex);
6964 * sys_perf_event_open - open a performance event, associate it to a task/cpu
6966 * @attr_uptr: event_id type attributes for monitoring/sampling
6969 * @group_fd: group leader event fd
6971 SYSCALL_DEFINE5(perf_event_open,
6972 struct perf_event_attr __user *, attr_uptr,
6973 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
6975 struct perf_event *group_leader = NULL, *output_event = NULL;
6976 struct perf_event *event, *sibling;
6977 struct perf_event_attr attr;
6978 struct perf_event_context *ctx;
6979 struct file *event_file = NULL;
6980 struct fd group = {NULL, 0};
6981 struct task_struct *task = NULL;
6986 int f_flags = O_RDWR;
6988 /* for future expandability... */
6989 if (flags & ~PERF_FLAG_ALL)
6992 err = perf_copy_attr(attr_uptr, &attr);
6996 if (!attr.exclude_kernel) {
6997 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
7002 if (attr.sample_freq > sysctl_perf_event_sample_rate)
7007 * In cgroup mode, the pid argument is used to pass the fd
7008 * opened to the cgroup directory in cgroupfs. The cpu argument
7009 * designates the cpu on which to monitor threads from that
7012 if ((flags & PERF_FLAG_PID_CGROUP) && (pid == -1 || cpu == -1))
7015 if (flags & PERF_FLAG_FD_CLOEXEC)
7016 f_flags |= O_CLOEXEC;
7018 event_fd = get_unused_fd_flags(f_flags);
7022 if (group_fd != -1) {
7023 err = perf_fget_light(group_fd, &group);
7026 group_leader = group.file->private_data;
7027 if (flags & PERF_FLAG_FD_OUTPUT)
7028 output_event = group_leader;
7029 if (flags & PERF_FLAG_FD_NO_GROUP)
7030 group_leader = NULL;
7033 if (pid != -1 && !(flags & PERF_FLAG_PID_CGROUP)) {
7034 task = find_lively_task_by_vpid(pid);
7036 err = PTR_ERR(task);
7043 event = perf_event_alloc(&attr, cpu, task, group_leader, NULL,
7045 if (IS_ERR(event)) {
7046 err = PTR_ERR(event);
7050 if (flags & PERF_FLAG_PID_CGROUP) {
7051 err = perf_cgroup_connect(pid, event, &attr, group_leader);
7053 __free_event(event);
7058 account_event(event);
7061 * Special case software events and allow them to be part of
7062 * any hardware group.
7067 (is_software_event(event) != is_software_event(group_leader))) {
7068 if (is_software_event(event)) {
7070 * If event and group_leader are not both a software
7071 * event, and event is, then group leader is not.
7073 * Allow the addition of software events to !software
7074 * groups, this is safe because software events never
7077 pmu = group_leader->pmu;
7078 } else if (is_software_event(group_leader) &&
7079 (group_leader->group_flags & PERF_GROUP_SOFTWARE)) {
7081 * In case the group is a pure software group, and we
7082 * try to add a hardware event, move the whole group to
7083 * the hardware context.
7090 * Get the target context (task or percpu):
7092 ctx = find_get_context(pmu, task, event->cpu);
7099 put_task_struct(task);
7104 * Look up the group leader (we will attach this event to it):
7110 * Do not allow a recursive hierarchy (this new sibling
7111 * becoming part of another group-sibling):
7113 if (group_leader->group_leader != group_leader)
7116 * Do not allow to attach to a group in a different
7117 * task or CPU context:
7120 if (group_leader->ctx->type != ctx->type)
7123 if (group_leader->ctx != ctx)
7128 * Only a group leader can be exclusive or pinned
7130 if (attr.exclusive || attr.pinned)
7135 err = perf_event_set_output(event, output_event);
7140 event_file = anon_inode_getfile("[perf_event]", &perf_fops, event,
7142 if (IS_ERR(event_file)) {
7143 err = PTR_ERR(event_file);
7148 struct perf_event_context *gctx = group_leader->ctx;
7150 mutex_lock(&gctx->mutex);
7151 perf_remove_from_context(group_leader);
7154 * Removing from the context ends up with disabled
7155 * event. What we want here is event in the initial
7156 * startup state, ready to be add into new context.
7158 perf_event__state_init(group_leader);
7159 list_for_each_entry(sibling, &group_leader->sibling_list,
7161 perf_remove_from_context(sibling);
7162 perf_event__state_init(sibling);
7165 mutex_unlock(&gctx->mutex);
7169 WARN_ON_ONCE(ctx->parent_ctx);
7170 mutex_lock(&ctx->mutex);
7174 perf_install_in_context(ctx, group_leader, event->cpu);
7176 list_for_each_entry(sibling, &group_leader->sibling_list,
7178 perf_install_in_context(ctx, sibling, event->cpu);
7183 perf_install_in_context(ctx, event, event->cpu);
7184 perf_unpin_context(ctx);
7185 mutex_unlock(&ctx->mutex);
7189 event->owner = current;
7191 mutex_lock(¤t->perf_event_mutex);
7192 list_add_tail(&event->owner_entry, ¤t->perf_event_list);
7193 mutex_unlock(¤t->perf_event_mutex);
7196 * Precalculate sample_data sizes
7198 perf_event__header_size(event);
7199 perf_event__id_header_size(event);
7202 * Drop the reference on the group_event after placing the
7203 * new event on the sibling_list. This ensures destruction
7204 * of the group leader will find the pointer to itself in
7205 * perf_group_detach().
7208 fd_install(event_fd, event_file);
7212 perf_unpin_context(ctx);
7219 put_task_struct(task);
7223 put_unused_fd(event_fd);
7228 * perf_event_create_kernel_counter
7230 * @attr: attributes of the counter to create
7231 * @cpu: cpu in which the counter is bound
7232 * @task: task to profile (NULL for percpu)
7235 perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu,
7236 struct task_struct *task,
7237 perf_overflow_handler_t overflow_handler,
7240 struct perf_event_context *ctx;
7241 struct perf_event *event;
7245 * Get the target context (task or percpu):
7248 event = perf_event_alloc(attr, cpu, task, NULL, NULL,
7249 overflow_handler, context);
7250 if (IS_ERR(event)) {
7251 err = PTR_ERR(event);
7255 account_event(event);
7257 ctx = find_get_context(event->pmu, task, cpu);
7263 WARN_ON_ONCE(ctx->parent_ctx);
7264 mutex_lock(&ctx->mutex);
7265 perf_install_in_context(ctx, event, cpu);
7266 perf_unpin_context(ctx);
7267 mutex_unlock(&ctx->mutex);
7274 return ERR_PTR(err);
7276 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter);
7278 void perf_pmu_migrate_context(struct pmu *pmu, int src_cpu, int dst_cpu)
7280 struct perf_event_context *src_ctx;
7281 struct perf_event_context *dst_ctx;
7282 struct perf_event *event, *tmp;
7285 src_ctx = &per_cpu_ptr(pmu->pmu_cpu_context, src_cpu)->ctx;
7286 dst_ctx = &per_cpu_ptr(pmu->pmu_cpu_context, dst_cpu)->ctx;
7288 mutex_lock(&src_ctx->mutex);
7289 list_for_each_entry_safe(event, tmp, &src_ctx->event_list,
7291 perf_remove_from_context(event);
7292 unaccount_event_cpu(event, src_cpu);
7294 list_add(&event->migrate_entry, &events);
7296 mutex_unlock(&src_ctx->mutex);
7300 mutex_lock(&dst_ctx->mutex);
7301 list_for_each_entry_safe(event, tmp, &events, migrate_entry) {
7302 list_del(&event->migrate_entry);
7303 if (event->state >= PERF_EVENT_STATE_OFF)
7304 event->state = PERF_EVENT_STATE_INACTIVE;
7305 account_event_cpu(event, dst_cpu);
7306 perf_install_in_context(dst_ctx, event, dst_cpu);
7309 mutex_unlock(&dst_ctx->mutex);
7311 EXPORT_SYMBOL_GPL(perf_pmu_migrate_context);
7313 static void sync_child_event(struct perf_event *child_event,
7314 struct task_struct *child)
7316 struct perf_event *parent_event = child_event->parent;
7319 if (child_event->attr.inherit_stat)
7320 perf_event_read_event(child_event, child);
7322 child_val = perf_event_count(child_event);
7325 * Add back the child's count to the parent's count:
7327 atomic64_add(child_val, &parent_event->child_count);
7328 atomic64_add(child_event->total_time_enabled,
7329 &parent_event->child_total_time_enabled);
7330 atomic64_add(child_event->total_time_running,
7331 &parent_event->child_total_time_running);
7334 * Remove this event from the parent's list
7336 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
7337 mutex_lock(&parent_event->child_mutex);
7338 list_del_init(&child_event->child_list);
7339 mutex_unlock(&parent_event->child_mutex);
7342 * Release the parent event, if this was the last
7345 put_event(parent_event);
7349 __perf_event_exit_task(struct perf_event *child_event,
7350 struct perf_event_context *child_ctx,
7351 struct task_struct *child)
7353 if (child_event->parent) {
7354 raw_spin_lock_irq(&child_ctx->lock);
7355 perf_group_detach(child_event);
7356 raw_spin_unlock_irq(&child_ctx->lock);
7359 perf_remove_from_context(child_event);
7362 * It can happen that the parent exits first, and has events
7363 * that are still around due to the child reference. These
7364 * events need to be zapped.
7366 if (child_event->parent) {
7367 sync_child_event(child_event, child);
7368 free_event(child_event);
7372 static void perf_event_exit_task_context(struct task_struct *child, int ctxn)
7374 struct perf_event *child_event, *tmp;
7375 struct perf_event_context *child_ctx;
7376 unsigned long flags;
7378 if (likely(!child->perf_event_ctxp[ctxn])) {
7379 perf_event_task(child, NULL, 0);
7383 local_irq_save(flags);
7385 * We can't reschedule here because interrupts are disabled,
7386 * and either child is current or it is a task that can't be
7387 * scheduled, so we are now safe from rescheduling changing
7390 child_ctx = rcu_dereference_raw(child->perf_event_ctxp[ctxn]);
7393 * Take the context lock here so that if find_get_context is
7394 * reading child->perf_event_ctxp, we wait until it has
7395 * incremented the context's refcount before we do put_ctx below.
7397 raw_spin_lock(&child_ctx->lock);
7398 task_ctx_sched_out(child_ctx);
7399 child->perf_event_ctxp[ctxn] = NULL;
7401 * If this context is a clone; unclone it so it can't get
7402 * swapped to another process while we're removing all
7403 * the events from it.
7405 unclone_ctx(child_ctx);
7406 update_context_time(child_ctx);
7407 raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
7410 * Report the task dead after unscheduling the events so that we
7411 * won't get any samples after PERF_RECORD_EXIT. We can however still
7412 * get a few PERF_RECORD_READ events.
7414 perf_event_task(child, child_ctx, 0);
7417 * We can recurse on the same lock type through:
7419 * __perf_event_exit_task()
7420 * sync_child_event()
7422 * mutex_lock(&ctx->mutex)
7424 * But since its the parent context it won't be the same instance.
7426 mutex_lock(&child_ctx->mutex);
7429 list_for_each_entry_safe(child_event, tmp, &child_ctx->pinned_groups,
7431 __perf_event_exit_task(child_event, child_ctx, child);
7433 list_for_each_entry_safe(child_event, tmp, &child_ctx->flexible_groups,
7435 __perf_event_exit_task(child_event, child_ctx, child);
7438 * If the last event was a group event, it will have appended all
7439 * its siblings to the list, but we obtained 'tmp' before that which
7440 * will still point to the list head terminating the iteration.
7442 if (!list_empty(&child_ctx->pinned_groups) ||
7443 !list_empty(&child_ctx->flexible_groups))
7446 mutex_unlock(&child_ctx->mutex);
7452 * When a child task exits, feed back event values to parent events.
7454 void perf_event_exit_task(struct task_struct *child)
7456 struct perf_event *event, *tmp;
7459 mutex_lock(&child->perf_event_mutex);
7460 list_for_each_entry_safe(event, tmp, &child->perf_event_list,
7462 list_del_init(&event->owner_entry);
7465 * Ensure the list deletion is visible before we clear
7466 * the owner, closes a race against perf_release() where
7467 * we need to serialize on the owner->perf_event_mutex.
7470 event->owner = NULL;
7472 mutex_unlock(&child->perf_event_mutex);
7474 for_each_task_context_nr(ctxn)
7475 perf_event_exit_task_context(child, ctxn);
7478 static void perf_free_event(struct perf_event *event,
7479 struct perf_event_context *ctx)
7481 struct perf_event *parent = event->parent;
7483 if (WARN_ON_ONCE(!parent))
7486 mutex_lock(&parent->child_mutex);
7487 list_del_init(&event->child_list);
7488 mutex_unlock(&parent->child_mutex);
7492 perf_group_detach(event);
7493 list_del_event(event, ctx);
7498 * free an unexposed, unused context as created by inheritance by
7499 * perf_event_init_task below, used by fork() in case of fail.
7501 void perf_event_free_task(struct task_struct *task)
7503 struct perf_event_context *ctx;
7504 struct perf_event *event, *tmp;
7507 for_each_task_context_nr(ctxn) {
7508 ctx = task->perf_event_ctxp[ctxn];
7512 mutex_lock(&ctx->mutex);
7514 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups,
7516 perf_free_event(event, ctx);
7518 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups,
7520 perf_free_event(event, ctx);
7522 if (!list_empty(&ctx->pinned_groups) ||
7523 !list_empty(&ctx->flexible_groups))
7526 mutex_unlock(&ctx->mutex);
7532 void perf_event_delayed_put(struct task_struct *task)
7536 for_each_task_context_nr(ctxn)
7537 WARN_ON_ONCE(task->perf_event_ctxp[ctxn]);
7541 * inherit a event from parent task to child task:
7543 static struct perf_event *
7544 inherit_event(struct perf_event *parent_event,
7545 struct task_struct *parent,
7546 struct perf_event_context *parent_ctx,
7547 struct task_struct *child,
7548 struct perf_event *group_leader,
7549 struct perf_event_context *child_ctx)
7551 struct perf_event *child_event;
7552 unsigned long flags;
7555 * Instead of creating recursive hierarchies of events,
7556 * we link inherited events back to the original parent,
7557 * which has a filp for sure, which we use as the reference
7560 if (parent_event->parent)
7561 parent_event = parent_event->parent;
7563 child_event = perf_event_alloc(&parent_event->attr,
7566 group_leader, parent_event,
7568 if (IS_ERR(child_event))
7571 if (!atomic_long_inc_not_zero(&parent_event->refcount)) {
7572 free_event(child_event);
7579 * Make the child state follow the state of the parent event,
7580 * not its attr.disabled bit. We hold the parent's mutex,
7581 * so we won't race with perf_event_{en, dis}able_family.
7583 if (parent_event->state >= PERF_EVENT_STATE_INACTIVE)
7584 child_event->state = PERF_EVENT_STATE_INACTIVE;
7586 child_event->state = PERF_EVENT_STATE_OFF;
7588 if (parent_event->attr.freq) {
7589 u64 sample_period = parent_event->hw.sample_period;
7590 struct hw_perf_event *hwc = &child_event->hw;
7592 hwc->sample_period = sample_period;
7593 hwc->last_period = sample_period;
7595 local64_set(&hwc->period_left, sample_period);
7598 child_event->ctx = child_ctx;
7599 child_event->overflow_handler = parent_event->overflow_handler;
7600 child_event->overflow_handler_context
7601 = parent_event->overflow_handler_context;
7604 * Precalculate sample_data sizes
7606 perf_event__header_size(child_event);
7607 perf_event__id_header_size(child_event);
7610 * Link it up in the child's context:
7612 raw_spin_lock_irqsave(&child_ctx->lock, flags);
7613 add_event_to_ctx(child_event, child_ctx);
7614 raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
7617 * Link this into the parent event's child list
7619 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
7620 mutex_lock(&parent_event->child_mutex);
7621 list_add_tail(&child_event->child_list, &parent_event->child_list);
7622 mutex_unlock(&parent_event->child_mutex);
7627 static int inherit_group(struct perf_event *parent_event,
7628 struct task_struct *parent,
7629 struct perf_event_context *parent_ctx,
7630 struct task_struct *child,
7631 struct perf_event_context *child_ctx)
7633 struct perf_event *leader;
7634 struct perf_event *sub;
7635 struct perf_event *child_ctr;
7637 leader = inherit_event(parent_event, parent, parent_ctx,
7638 child, NULL, child_ctx);
7640 return PTR_ERR(leader);
7641 list_for_each_entry(sub, &parent_event->sibling_list, group_entry) {
7642 child_ctr = inherit_event(sub, parent, parent_ctx,
7643 child, leader, child_ctx);
7644 if (IS_ERR(child_ctr))
7645 return PTR_ERR(child_ctr);
7651 inherit_task_group(struct perf_event *event, struct task_struct *parent,
7652 struct perf_event_context *parent_ctx,
7653 struct task_struct *child, int ctxn,
7657 struct perf_event_context *child_ctx;
7659 if (!event->attr.inherit) {
7664 child_ctx = child->perf_event_ctxp[ctxn];
7667 * This is executed from the parent task context, so
7668 * inherit events that have been marked for cloning.
7669 * First allocate and initialize a context for the
7673 child_ctx = alloc_perf_context(parent_ctx->pmu, child);
7677 child->perf_event_ctxp[ctxn] = child_ctx;
7680 ret = inherit_group(event, parent, parent_ctx,
7690 * Initialize the perf_event context in task_struct
7692 int perf_event_init_context(struct task_struct *child, int ctxn)
7694 struct perf_event_context *child_ctx, *parent_ctx;
7695 struct perf_event_context *cloned_ctx;
7696 struct perf_event *event;
7697 struct task_struct *parent = current;
7698 int inherited_all = 1;
7699 unsigned long flags;
7702 if (likely(!parent->perf_event_ctxp[ctxn]))
7706 * If the parent's context is a clone, pin it so it won't get
7709 parent_ctx = perf_pin_task_context(parent, ctxn);
7712 * No need to check if parent_ctx != NULL here; since we saw
7713 * it non-NULL earlier, the only reason for it to become NULL
7714 * is if we exit, and since we're currently in the middle of
7715 * a fork we can't be exiting at the same time.
7719 * Lock the parent list. No need to lock the child - not PID
7720 * hashed yet and not running, so nobody can access it.
7722 mutex_lock(&parent_ctx->mutex);
7725 * We dont have to disable NMIs - we are only looking at
7726 * the list, not manipulating it:
7728 list_for_each_entry(event, &parent_ctx->pinned_groups, group_entry) {
7729 ret = inherit_task_group(event, parent, parent_ctx,
7730 child, ctxn, &inherited_all);
7736 * We can't hold ctx->lock when iterating the ->flexible_group list due
7737 * to allocations, but we need to prevent rotation because
7738 * rotate_ctx() will change the list from interrupt context.
7740 raw_spin_lock_irqsave(&parent_ctx->lock, flags);
7741 parent_ctx->rotate_disable = 1;
7742 raw_spin_unlock_irqrestore(&parent_ctx->lock, flags);
7744 list_for_each_entry(event, &parent_ctx->flexible_groups, group_entry) {
7745 ret = inherit_task_group(event, parent, parent_ctx,
7746 child, ctxn, &inherited_all);
7751 raw_spin_lock_irqsave(&parent_ctx->lock, flags);
7752 parent_ctx->rotate_disable = 0;
7754 child_ctx = child->perf_event_ctxp[ctxn];
7756 if (child_ctx && inherited_all) {
7758 * Mark the child context as a clone of the parent
7759 * context, or of whatever the parent is a clone of.
7761 * Note that if the parent is a clone, the holding of
7762 * parent_ctx->lock avoids it from being uncloned.
7764 cloned_ctx = parent_ctx->parent_ctx;
7766 child_ctx->parent_ctx = cloned_ctx;
7767 child_ctx->parent_gen = parent_ctx->parent_gen;
7769 child_ctx->parent_ctx = parent_ctx;
7770 child_ctx->parent_gen = parent_ctx->generation;
7772 get_ctx(child_ctx->parent_ctx);
7775 raw_spin_unlock_irqrestore(&parent_ctx->lock, flags);
7776 mutex_unlock(&parent_ctx->mutex);
7778 perf_unpin_context(parent_ctx);
7779 put_ctx(parent_ctx);
7785 * Initialize the perf_event context in task_struct
7787 int perf_event_init_task(struct task_struct *child)
7791 memset(child->perf_event_ctxp, 0, sizeof(child->perf_event_ctxp));
7792 mutex_init(&child->perf_event_mutex);
7793 INIT_LIST_HEAD(&child->perf_event_list);
7795 for_each_task_context_nr(ctxn) {
7796 ret = perf_event_init_context(child, ctxn);
7804 static void __init perf_event_init_all_cpus(void)
7806 struct swevent_htable *swhash;
7809 for_each_possible_cpu(cpu) {
7810 swhash = &per_cpu(swevent_htable, cpu);
7811 mutex_init(&swhash->hlist_mutex);
7812 INIT_LIST_HEAD(&per_cpu(rotation_list, cpu));
7816 static void perf_event_init_cpu(int cpu)
7818 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
7820 mutex_lock(&swhash->hlist_mutex);
7821 if (swhash->hlist_refcount > 0) {
7822 struct swevent_hlist *hlist;
7824 hlist = kzalloc_node(sizeof(*hlist), GFP_KERNEL, cpu_to_node(cpu));
7826 rcu_assign_pointer(swhash->swevent_hlist, hlist);
7828 mutex_unlock(&swhash->hlist_mutex);
7831 #if defined CONFIG_HOTPLUG_CPU || defined CONFIG_KEXEC
7832 static void perf_pmu_rotate_stop(struct pmu *pmu)
7834 struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
7836 WARN_ON(!irqs_disabled());
7838 list_del_init(&cpuctx->rotation_list);
7841 static void __perf_event_exit_context(void *__info)
7843 struct perf_event_context *ctx = __info;
7844 struct perf_event *event, *tmp;
7846 perf_pmu_rotate_stop(ctx->pmu);
7848 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry)
7849 __perf_remove_from_context(event);
7850 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups, group_entry)
7851 __perf_remove_from_context(event);
7854 static void perf_event_exit_cpu_context(int cpu)
7856 struct perf_event_context *ctx;
7860 idx = srcu_read_lock(&pmus_srcu);
7861 list_for_each_entry_rcu(pmu, &pmus, entry) {
7862 ctx = &per_cpu_ptr(pmu->pmu_cpu_context, cpu)->ctx;
7864 mutex_lock(&ctx->mutex);
7865 smp_call_function_single(cpu, __perf_event_exit_context, ctx, 1);
7866 mutex_unlock(&ctx->mutex);
7868 srcu_read_unlock(&pmus_srcu, idx);
7871 static void perf_event_exit_cpu(int cpu)
7873 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
7875 mutex_lock(&swhash->hlist_mutex);
7876 swevent_hlist_release(swhash);
7877 mutex_unlock(&swhash->hlist_mutex);
7879 perf_event_exit_cpu_context(cpu);
7882 static inline void perf_event_exit_cpu(int cpu) { }
7886 perf_reboot(struct notifier_block *notifier, unsigned long val, void *v)
7890 for_each_online_cpu(cpu)
7891 perf_event_exit_cpu(cpu);
7897 * Run the perf reboot notifier at the very last possible moment so that
7898 * the generic watchdog code runs as long as possible.
7900 static struct notifier_block perf_reboot_notifier = {
7901 .notifier_call = perf_reboot,
7902 .priority = INT_MIN,
7906 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
7908 unsigned int cpu = (long)hcpu;
7910 switch (action & ~CPU_TASKS_FROZEN) {
7912 case CPU_UP_PREPARE:
7913 case CPU_DOWN_FAILED:
7914 perf_event_init_cpu(cpu);
7917 case CPU_UP_CANCELED:
7918 case CPU_DOWN_PREPARE:
7919 perf_event_exit_cpu(cpu);
7928 void __init perf_event_init(void)
7934 perf_event_init_all_cpus();
7935 init_srcu_struct(&pmus_srcu);
7936 perf_pmu_register(&perf_swevent, "software", PERF_TYPE_SOFTWARE);
7937 perf_pmu_register(&perf_cpu_clock, NULL, -1);
7938 perf_pmu_register(&perf_task_clock, NULL, -1);
7940 perf_cpu_notifier(perf_cpu_notify);
7941 register_reboot_notifier(&perf_reboot_notifier);
7943 ret = init_hw_breakpoint();
7944 WARN(ret, "hw_breakpoint initialization failed with: %d", ret);
7946 /* do not patch jump label more than once per second */
7947 jump_label_rate_limit(&perf_sched_events, HZ);
7950 * Build time assertion that we keep the data_head at the intended
7951 * location. IOW, validation we got the __reserved[] size right.
7953 BUILD_BUG_ON((offsetof(struct perf_event_mmap_page, data_head))
7957 static int __init perf_event_sysfs_init(void)
7962 mutex_lock(&pmus_lock);
7964 ret = bus_register(&pmu_bus);
7968 list_for_each_entry(pmu, &pmus, entry) {
7969 if (!pmu->name || pmu->type < 0)
7972 ret = pmu_dev_alloc(pmu);
7973 WARN(ret, "Failed to register pmu: %s, reason %d\n", pmu->name, ret);
7975 pmu_bus_running = 1;
7979 mutex_unlock(&pmus_lock);
7983 device_initcall(perf_event_sysfs_init);
7985 #ifdef CONFIG_CGROUP_PERF
7986 static struct cgroup_subsys_state *
7987 perf_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
7989 struct perf_cgroup *jc;
7991 jc = kzalloc(sizeof(*jc), GFP_KERNEL);
7993 return ERR_PTR(-ENOMEM);
7995 jc->info = alloc_percpu(struct perf_cgroup_info);
7998 return ERR_PTR(-ENOMEM);
8004 static void perf_cgroup_css_free(struct cgroup_subsys_state *css)
8006 struct perf_cgroup *jc = container_of(css, struct perf_cgroup, css);
8008 free_percpu(jc->info);
8012 static int __perf_cgroup_move(void *info)
8014 struct task_struct *task = info;
8015 perf_cgroup_switch(task, PERF_CGROUP_SWOUT | PERF_CGROUP_SWIN);
8019 static void perf_cgroup_attach(struct cgroup_subsys_state *css,
8020 struct cgroup_taskset *tset)
8022 struct task_struct *task;
8024 cgroup_taskset_for_each(task, tset)
8025 task_function_call(task, __perf_cgroup_move, task);
8028 static void perf_cgroup_exit(struct cgroup_subsys_state *css,
8029 struct cgroup_subsys_state *old_css,
8030 struct task_struct *task)
8033 * cgroup_exit() is called in the copy_process() failure path.
8034 * Ignore this case since the task hasn't ran yet, this avoids
8035 * trying to poke a half freed task state from generic code.
8037 if (!(task->flags & PF_EXITING))
8040 task_function_call(task, __perf_cgroup_move, task);
8043 struct cgroup_subsys perf_event_cgrp_subsys = {
8044 .css_alloc = perf_cgroup_css_alloc,
8045 .css_free = perf_cgroup_css_free,
8046 .exit = perf_cgroup_exit,
8047 .attach = perf_cgroup_attach,
8049 #endif /* CONFIG_CGROUP_PERF */