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/cgroup.h>
38 #include <linux/perf_event.h>
39 #include <linux/trace_events.h>
40 #include <linux/hw_breakpoint.h>
41 #include <linux/mm_types.h>
42 #include <linux/module.h>
43 #include <linux/mman.h>
44 #include <linux/compat.h>
45 #include <linux/bpf.h>
46 #include <linux/filter.h>
50 #include <asm/irq_regs.h>
52 static struct workqueue_struct *perf_wq;
54 typedef int (*remote_function_f)(void *);
56 struct remote_function_call {
57 struct task_struct *p;
58 remote_function_f func;
63 static void remote_function(void *data)
65 struct remote_function_call *tfc = data;
66 struct task_struct *p = tfc->p;
70 if (task_cpu(p) != smp_processor_id() || !task_curr(p))
74 tfc->ret = tfc->func(tfc->info);
78 * task_function_call - call a function on the cpu on which a task runs
79 * @p: the task to evaluate
80 * @func: the function to be called
81 * @info: the function call argument
83 * Calls the function @func when the task is currently running. This might
84 * be on the current CPU, which just calls the function directly
86 * returns: @func return value, or
87 * -ESRCH - when the process isn't running
88 * -EAGAIN - when the process moved away
91 task_function_call(struct task_struct *p, remote_function_f func, void *info)
93 struct remote_function_call data = {
97 .ret = -ESRCH, /* No such (running) process */
101 smp_call_function_single(task_cpu(p), remote_function, &data, 1);
107 * cpu_function_call - call a function on the cpu
108 * @func: the function to be called
109 * @info: the function call argument
111 * Calls the function @func on the remote cpu.
113 * returns: @func return value or -ENXIO when the cpu is offline
115 static int cpu_function_call(int cpu, remote_function_f func, void *info)
117 struct remote_function_call data = {
121 .ret = -ENXIO, /* No such CPU */
124 smp_call_function_single(cpu, remote_function, &data, 1);
129 #define EVENT_OWNER_KERNEL ((void *) -1)
131 static bool is_kernel_event(struct perf_event *event)
133 return event->owner == EVENT_OWNER_KERNEL;
136 #define PERF_FLAG_ALL (PERF_FLAG_FD_NO_GROUP |\
137 PERF_FLAG_FD_OUTPUT |\
138 PERF_FLAG_PID_CGROUP |\
139 PERF_FLAG_FD_CLOEXEC)
142 * branch priv levels that need permission checks
144 #define PERF_SAMPLE_BRANCH_PERM_PLM \
145 (PERF_SAMPLE_BRANCH_KERNEL |\
146 PERF_SAMPLE_BRANCH_HV)
149 EVENT_FLEXIBLE = 0x1,
151 EVENT_ALL = EVENT_FLEXIBLE | EVENT_PINNED,
155 * perf_sched_events : >0 events exist
156 * perf_cgroup_events: >0 per-cpu cgroup events exist on this cpu
158 struct static_key_deferred perf_sched_events __read_mostly;
159 static DEFINE_PER_CPU(atomic_t, perf_cgroup_events);
160 static DEFINE_PER_CPU(int, perf_sched_cb_usages);
162 static atomic_t nr_mmap_events __read_mostly;
163 static atomic_t nr_comm_events __read_mostly;
164 static atomic_t nr_task_events __read_mostly;
165 static atomic_t nr_freq_events __read_mostly;
167 static LIST_HEAD(pmus);
168 static DEFINE_MUTEX(pmus_lock);
169 static struct srcu_struct pmus_srcu;
172 * perf event paranoia level:
173 * -1 - not paranoid at all
174 * 0 - disallow raw tracepoint access for unpriv
175 * 1 - disallow cpu events for unpriv
176 * 2 - disallow kernel profiling for unpriv
178 int sysctl_perf_event_paranoid __read_mostly = 1;
180 /* Minimum for 512 kiB + 1 user control page */
181 int sysctl_perf_event_mlock __read_mostly = 512 + (PAGE_SIZE / 1024); /* 'free' kiB per user */
184 * max perf event sample rate
186 #define DEFAULT_MAX_SAMPLE_RATE 100000
187 #define DEFAULT_SAMPLE_PERIOD_NS (NSEC_PER_SEC / DEFAULT_MAX_SAMPLE_RATE)
188 #define DEFAULT_CPU_TIME_MAX_PERCENT 25
190 int sysctl_perf_event_sample_rate __read_mostly = DEFAULT_MAX_SAMPLE_RATE;
192 static int max_samples_per_tick __read_mostly = DIV_ROUND_UP(DEFAULT_MAX_SAMPLE_RATE, HZ);
193 static int perf_sample_period_ns __read_mostly = DEFAULT_SAMPLE_PERIOD_NS;
195 static int perf_sample_allowed_ns __read_mostly =
196 DEFAULT_SAMPLE_PERIOD_NS * DEFAULT_CPU_TIME_MAX_PERCENT / 100;
198 void update_perf_cpu_limits(void)
200 u64 tmp = perf_sample_period_ns;
202 tmp *= sysctl_perf_cpu_time_max_percent;
204 ACCESS_ONCE(perf_sample_allowed_ns) = tmp;
207 static int perf_rotate_context(struct perf_cpu_context *cpuctx);
209 int perf_proc_update_handler(struct ctl_table *table, int write,
210 void __user *buffer, size_t *lenp,
213 int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
218 max_samples_per_tick = DIV_ROUND_UP(sysctl_perf_event_sample_rate, HZ);
219 perf_sample_period_ns = NSEC_PER_SEC / sysctl_perf_event_sample_rate;
220 update_perf_cpu_limits();
225 int sysctl_perf_cpu_time_max_percent __read_mostly = DEFAULT_CPU_TIME_MAX_PERCENT;
227 int perf_cpu_time_max_percent_handler(struct ctl_table *table, int write,
228 void __user *buffer, size_t *lenp,
231 int ret = proc_dointvec(table, write, buffer, lenp, ppos);
236 update_perf_cpu_limits();
242 * perf samples are done in some very critical code paths (NMIs).
243 * If they take too much CPU time, the system can lock up and not
244 * get any real work done. This will drop the sample rate when
245 * we detect that events are taking too long.
247 #define NR_ACCUMULATED_SAMPLES 128
248 static DEFINE_PER_CPU(u64, running_sample_length);
250 static void perf_duration_warn(struct irq_work *w)
252 u64 allowed_ns = ACCESS_ONCE(perf_sample_allowed_ns);
253 u64 avg_local_sample_len;
254 u64 local_samples_len;
256 local_samples_len = __this_cpu_read(running_sample_length);
257 avg_local_sample_len = local_samples_len/NR_ACCUMULATED_SAMPLES;
259 printk_ratelimited(KERN_WARNING
260 "perf interrupt took too long (%lld > %lld), lowering "
261 "kernel.perf_event_max_sample_rate to %d\n",
262 avg_local_sample_len, allowed_ns >> 1,
263 sysctl_perf_event_sample_rate);
266 static DEFINE_IRQ_WORK(perf_duration_work, perf_duration_warn);
268 void perf_sample_event_took(u64 sample_len_ns)
270 u64 allowed_ns = ACCESS_ONCE(perf_sample_allowed_ns);
271 u64 avg_local_sample_len;
272 u64 local_samples_len;
277 /* decay the counter by 1 average sample */
278 local_samples_len = __this_cpu_read(running_sample_length);
279 local_samples_len -= local_samples_len/NR_ACCUMULATED_SAMPLES;
280 local_samples_len += sample_len_ns;
281 __this_cpu_write(running_sample_length, local_samples_len);
284 * note: this will be biased artifically low until we have
285 * seen NR_ACCUMULATED_SAMPLES. Doing it this way keeps us
286 * from having to maintain a count.
288 avg_local_sample_len = local_samples_len/NR_ACCUMULATED_SAMPLES;
290 if (avg_local_sample_len <= allowed_ns)
293 if (max_samples_per_tick <= 1)
296 max_samples_per_tick = DIV_ROUND_UP(max_samples_per_tick, 2);
297 sysctl_perf_event_sample_rate = max_samples_per_tick * HZ;
298 perf_sample_period_ns = NSEC_PER_SEC / sysctl_perf_event_sample_rate;
300 update_perf_cpu_limits();
302 if (!irq_work_queue(&perf_duration_work)) {
303 early_printk("perf interrupt took too long (%lld > %lld), lowering "
304 "kernel.perf_event_max_sample_rate to %d\n",
305 avg_local_sample_len, allowed_ns >> 1,
306 sysctl_perf_event_sample_rate);
310 static atomic64_t perf_event_id;
312 static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
313 enum event_type_t event_type);
315 static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
316 enum event_type_t event_type,
317 struct task_struct *task);
319 static void update_context_time(struct perf_event_context *ctx);
320 static u64 perf_event_time(struct perf_event *event);
322 void __weak perf_event_print_debug(void) { }
324 extern __weak const char *perf_pmu_name(void)
329 static inline u64 perf_clock(void)
331 return local_clock();
334 static inline u64 perf_event_clock(struct perf_event *event)
336 return event->clock();
339 static inline struct perf_cpu_context *
340 __get_cpu_context(struct perf_event_context *ctx)
342 return this_cpu_ptr(ctx->pmu->pmu_cpu_context);
345 static void perf_ctx_lock(struct perf_cpu_context *cpuctx,
346 struct perf_event_context *ctx)
348 raw_spin_lock(&cpuctx->ctx.lock);
350 raw_spin_lock(&ctx->lock);
353 static void perf_ctx_unlock(struct perf_cpu_context *cpuctx,
354 struct perf_event_context *ctx)
357 raw_spin_unlock(&ctx->lock);
358 raw_spin_unlock(&cpuctx->ctx.lock);
361 #ifdef CONFIG_CGROUP_PERF
364 perf_cgroup_match(struct perf_event *event)
366 struct perf_event_context *ctx = event->ctx;
367 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
369 /* @event doesn't care about cgroup */
373 /* wants specific cgroup scope but @cpuctx isn't associated with any */
378 * Cgroup scoping is recursive. An event enabled for a cgroup is
379 * also enabled for all its descendant cgroups. If @cpuctx's
380 * cgroup is a descendant of @event's (the test covers identity
381 * case), it's a match.
383 return cgroup_is_descendant(cpuctx->cgrp->css.cgroup,
384 event->cgrp->css.cgroup);
387 static inline void perf_detach_cgroup(struct perf_event *event)
389 css_put(&event->cgrp->css);
393 static inline int is_cgroup_event(struct perf_event *event)
395 return event->cgrp != NULL;
398 static inline u64 perf_cgroup_event_time(struct perf_event *event)
400 struct perf_cgroup_info *t;
402 t = per_cpu_ptr(event->cgrp->info, event->cpu);
406 static inline void __update_cgrp_time(struct perf_cgroup *cgrp)
408 struct perf_cgroup_info *info;
413 info = this_cpu_ptr(cgrp->info);
415 info->time += now - info->timestamp;
416 info->timestamp = now;
419 static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx)
421 struct perf_cgroup *cgrp_out = cpuctx->cgrp;
423 __update_cgrp_time(cgrp_out);
426 static inline void update_cgrp_time_from_event(struct perf_event *event)
428 struct perf_cgroup *cgrp;
431 * ensure we access cgroup data only when needed and
432 * when we know the cgroup is pinned (css_get)
434 if (!is_cgroup_event(event))
437 cgrp = perf_cgroup_from_task(current);
439 * Do not update time when cgroup is not active
441 if (cgrp == event->cgrp)
442 __update_cgrp_time(event->cgrp);
446 perf_cgroup_set_timestamp(struct task_struct *task,
447 struct perf_event_context *ctx)
449 struct perf_cgroup *cgrp;
450 struct perf_cgroup_info *info;
453 * ctx->lock held by caller
454 * ensure we do not access cgroup data
455 * unless we have the cgroup pinned (css_get)
457 if (!task || !ctx->nr_cgroups)
460 cgrp = perf_cgroup_from_task(task);
461 info = this_cpu_ptr(cgrp->info);
462 info->timestamp = ctx->timestamp;
465 #define PERF_CGROUP_SWOUT 0x1 /* cgroup switch out every event */
466 #define PERF_CGROUP_SWIN 0x2 /* cgroup switch in events based on task */
469 * reschedule events based on the cgroup constraint of task.
471 * mode SWOUT : schedule out everything
472 * mode SWIN : schedule in based on cgroup for next
474 void perf_cgroup_switch(struct task_struct *task, int mode)
476 struct perf_cpu_context *cpuctx;
481 * disable interrupts to avoid geting nr_cgroup
482 * changes via __perf_event_disable(). Also
485 local_irq_save(flags);
488 * we reschedule only in the presence of cgroup
489 * constrained events.
493 list_for_each_entry_rcu(pmu, &pmus, entry) {
494 cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
495 if (cpuctx->unique_pmu != pmu)
496 continue; /* ensure we process each cpuctx once */
499 * perf_cgroup_events says at least one
500 * context on this CPU has cgroup events.
502 * ctx->nr_cgroups reports the number of cgroup
503 * events for a context.
505 if (cpuctx->ctx.nr_cgroups > 0) {
506 perf_ctx_lock(cpuctx, cpuctx->task_ctx);
507 perf_pmu_disable(cpuctx->ctx.pmu);
509 if (mode & PERF_CGROUP_SWOUT) {
510 cpu_ctx_sched_out(cpuctx, EVENT_ALL);
512 * must not be done before ctxswout due
513 * to event_filter_match() in event_sched_out()
518 if (mode & PERF_CGROUP_SWIN) {
519 WARN_ON_ONCE(cpuctx->cgrp);
521 * set cgrp before ctxsw in to allow
522 * event_filter_match() to not have to pass
525 cpuctx->cgrp = perf_cgroup_from_task(task);
526 cpu_ctx_sched_in(cpuctx, EVENT_ALL, task);
528 perf_pmu_enable(cpuctx->ctx.pmu);
529 perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
535 local_irq_restore(flags);
538 static inline void perf_cgroup_sched_out(struct task_struct *task,
539 struct task_struct *next)
541 struct perf_cgroup *cgrp1;
542 struct perf_cgroup *cgrp2 = NULL;
545 * we come here when we know perf_cgroup_events > 0
547 cgrp1 = perf_cgroup_from_task(task);
550 * next is NULL when called from perf_event_enable_on_exec()
551 * that will systematically cause a cgroup_switch()
554 cgrp2 = perf_cgroup_from_task(next);
557 * only schedule out current cgroup events if we know
558 * that we are switching to a different cgroup. Otherwise,
559 * do no touch the cgroup events.
562 perf_cgroup_switch(task, PERF_CGROUP_SWOUT);
565 static inline void perf_cgroup_sched_in(struct task_struct *prev,
566 struct task_struct *task)
568 struct perf_cgroup *cgrp1;
569 struct perf_cgroup *cgrp2 = NULL;
572 * we come here when we know perf_cgroup_events > 0
574 cgrp1 = perf_cgroup_from_task(task);
576 /* prev can never be NULL */
577 cgrp2 = perf_cgroup_from_task(prev);
580 * only need to schedule in cgroup events if we are changing
581 * cgroup during ctxsw. Cgroup events were not scheduled
582 * out of ctxsw out if that was not the case.
585 perf_cgroup_switch(task, PERF_CGROUP_SWIN);
588 static inline int perf_cgroup_connect(int fd, struct perf_event *event,
589 struct perf_event_attr *attr,
590 struct perf_event *group_leader)
592 struct perf_cgroup *cgrp;
593 struct cgroup_subsys_state *css;
594 struct fd f = fdget(fd);
600 css = css_tryget_online_from_dir(f.file->f_path.dentry,
601 &perf_event_cgrp_subsys);
607 cgrp = container_of(css, struct perf_cgroup, css);
611 * all events in a group must monitor
612 * the same cgroup because a task belongs
613 * to only one perf cgroup at a time
615 if (group_leader && group_leader->cgrp != cgrp) {
616 perf_detach_cgroup(event);
625 perf_cgroup_set_shadow_time(struct perf_event *event, u64 now)
627 struct perf_cgroup_info *t;
628 t = per_cpu_ptr(event->cgrp->info, event->cpu);
629 event->shadow_ctx_time = now - t->timestamp;
633 perf_cgroup_defer_enabled(struct perf_event *event)
636 * when the current task's perf cgroup does not match
637 * the event's, we need to remember to call the
638 * perf_mark_enable() function the first time a task with
639 * a matching perf cgroup is scheduled in.
641 if (is_cgroup_event(event) && !perf_cgroup_match(event))
642 event->cgrp_defer_enabled = 1;
646 perf_cgroup_mark_enabled(struct perf_event *event,
647 struct perf_event_context *ctx)
649 struct perf_event *sub;
650 u64 tstamp = perf_event_time(event);
652 if (!event->cgrp_defer_enabled)
655 event->cgrp_defer_enabled = 0;
657 event->tstamp_enabled = tstamp - event->total_time_enabled;
658 list_for_each_entry(sub, &event->sibling_list, group_entry) {
659 if (sub->state >= PERF_EVENT_STATE_INACTIVE) {
660 sub->tstamp_enabled = tstamp - sub->total_time_enabled;
661 sub->cgrp_defer_enabled = 0;
665 #else /* !CONFIG_CGROUP_PERF */
668 perf_cgroup_match(struct perf_event *event)
673 static inline void perf_detach_cgroup(struct perf_event *event)
676 static inline int is_cgroup_event(struct perf_event *event)
681 static inline u64 perf_cgroup_event_cgrp_time(struct perf_event *event)
686 static inline void update_cgrp_time_from_event(struct perf_event *event)
690 static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx)
694 static inline void perf_cgroup_sched_out(struct task_struct *task,
695 struct task_struct *next)
699 static inline void perf_cgroup_sched_in(struct task_struct *prev,
700 struct task_struct *task)
704 static inline int perf_cgroup_connect(pid_t pid, struct perf_event *event,
705 struct perf_event_attr *attr,
706 struct perf_event *group_leader)
712 perf_cgroup_set_timestamp(struct task_struct *task,
713 struct perf_event_context *ctx)
718 perf_cgroup_switch(struct task_struct *task, struct task_struct *next)
723 perf_cgroup_set_shadow_time(struct perf_event *event, u64 now)
727 static inline u64 perf_cgroup_event_time(struct perf_event *event)
733 perf_cgroup_defer_enabled(struct perf_event *event)
738 perf_cgroup_mark_enabled(struct perf_event *event,
739 struct perf_event_context *ctx)
745 * set default to be dependent on timer tick just
748 #define PERF_CPU_HRTIMER (1000 / HZ)
750 * function must be called with interrupts disbled
752 static enum hrtimer_restart perf_mux_hrtimer_handler(struct hrtimer *hr)
754 struct perf_cpu_context *cpuctx;
757 WARN_ON(!irqs_disabled());
759 cpuctx = container_of(hr, struct perf_cpu_context, hrtimer);
760 rotations = perf_rotate_context(cpuctx);
762 raw_spin_lock(&cpuctx->hrtimer_lock);
764 hrtimer_forward_now(hr, cpuctx->hrtimer_interval);
766 cpuctx->hrtimer_active = 0;
767 raw_spin_unlock(&cpuctx->hrtimer_lock);
769 return rotations ? HRTIMER_RESTART : HRTIMER_NORESTART;
772 static void __perf_mux_hrtimer_init(struct perf_cpu_context *cpuctx, int cpu)
774 struct hrtimer *timer = &cpuctx->hrtimer;
775 struct pmu *pmu = cpuctx->ctx.pmu;
778 /* no multiplexing needed for SW PMU */
779 if (pmu->task_ctx_nr == perf_sw_context)
783 * check default is sane, if not set then force to
784 * default interval (1/tick)
786 interval = pmu->hrtimer_interval_ms;
788 interval = pmu->hrtimer_interval_ms = PERF_CPU_HRTIMER;
790 cpuctx->hrtimer_interval = ns_to_ktime(NSEC_PER_MSEC * interval);
792 raw_spin_lock_init(&cpuctx->hrtimer_lock);
793 hrtimer_init(timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
794 timer->function = perf_mux_hrtimer_handler;
797 static int perf_mux_hrtimer_restart(struct perf_cpu_context *cpuctx)
799 struct hrtimer *timer = &cpuctx->hrtimer;
800 struct pmu *pmu = cpuctx->ctx.pmu;
804 if (pmu->task_ctx_nr == perf_sw_context)
807 raw_spin_lock_irqsave(&cpuctx->hrtimer_lock, flags);
808 if (!cpuctx->hrtimer_active) {
809 cpuctx->hrtimer_active = 1;
810 hrtimer_forward_now(timer, cpuctx->hrtimer_interval);
811 hrtimer_start_expires(timer, HRTIMER_MODE_ABS_PINNED);
813 raw_spin_unlock_irqrestore(&cpuctx->hrtimer_lock, flags);
818 void perf_pmu_disable(struct pmu *pmu)
820 int *count = this_cpu_ptr(pmu->pmu_disable_count);
822 pmu->pmu_disable(pmu);
825 void perf_pmu_enable(struct pmu *pmu)
827 int *count = this_cpu_ptr(pmu->pmu_disable_count);
829 pmu->pmu_enable(pmu);
832 static DEFINE_PER_CPU(struct list_head, active_ctx_list);
835 * perf_event_ctx_activate(), perf_event_ctx_deactivate(), and
836 * perf_event_task_tick() are fully serialized because they're strictly cpu
837 * affine and perf_event_ctx{activate,deactivate} are called with IRQs
838 * disabled, while perf_event_task_tick is called from IRQ context.
840 static void perf_event_ctx_activate(struct perf_event_context *ctx)
842 struct list_head *head = this_cpu_ptr(&active_ctx_list);
844 WARN_ON(!irqs_disabled());
846 WARN_ON(!list_empty(&ctx->active_ctx_list));
848 list_add(&ctx->active_ctx_list, head);
851 static void perf_event_ctx_deactivate(struct perf_event_context *ctx)
853 WARN_ON(!irqs_disabled());
855 WARN_ON(list_empty(&ctx->active_ctx_list));
857 list_del_init(&ctx->active_ctx_list);
860 static void get_ctx(struct perf_event_context *ctx)
862 WARN_ON(!atomic_inc_not_zero(&ctx->refcount));
865 static void free_ctx(struct rcu_head *head)
867 struct perf_event_context *ctx;
869 ctx = container_of(head, struct perf_event_context, rcu_head);
870 kfree(ctx->task_ctx_data);
874 static void put_ctx(struct perf_event_context *ctx)
876 if (atomic_dec_and_test(&ctx->refcount)) {
878 put_ctx(ctx->parent_ctx);
880 put_task_struct(ctx->task);
881 call_rcu(&ctx->rcu_head, free_ctx);
886 * Because of perf_event::ctx migration in sys_perf_event_open::move_group and
887 * perf_pmu_migrate_context() we need some magic.
889 * Those places that change perf_event::ctx will hold both
890 * perf_event_ctx::mutex of the 'old' and 'new' ctx value.
892 * Lock ordering is by mutex address. There are two other sites where
893 * perf_event_context::mutex nests and those are:
895 * - perf_event_exit_task_context() [ child , 0 ]
896 * __perf_event_exit_task()
898 * put_event() [ parent, 1 ]
900 * - perf_event_init_context() [ parent, 0 ]
901 * inherit_task_group()
906 * perf_try_init_event() [ child , 1 ]
908 * While it appears there is an obvious deadlock here -- the parent and child
909 * nesting levels are inverted between the two. This is in fact safe because
910 * life-time rules separate them. That is an exiting task cannot fork, and a
911 * spawning task cannot (yet) exit.
913 * But remember that that these are parent<->child context relations, and
914 * migration does not affect children, therefore these two orderings should not
917 * The change in perf_event::ctx does not affect children (as claimed above)
918 * because the sys_perf_event_open() case will install a new event and break
919 * the ctx parent<->child relation, and perf_pmu_migrate_context() is only
920 * concerned with cpuctx and that doesn't have children.
922 * The places that change perf_event::ctx will issue:
924 * perf_remove_from_context();
926 * perf_install_in_context();
928 * to affect the change. The remove_from_context() + synchronize_rcu() should
929 * quiesce the event, after which we can install it in the new location. This
930 * means that only external vectors (perf_fops, prctl) can perturb the event
931 * while in transit. Therefore all such accessors should also acquire
932 * perf_event_context::mutex to serialize against this.
934 * However; because event->ctx can change while we're waiting to acquire
935 * ctx->mutex we must be careful and use the below perf_event_ctx_lock()
939 * task_struct::perf_event_mutex
940 * perf_event_context::mutex
941 * perf_event_context::lock
942 * perf_event::child_mutex;
943 * perf_event::mmap_mutex
946 static struct perf_event_context *
947 perf_event_ctx_lock_nested(struct perf_event *event, int nesting)
949 struct perf_event_context *ctx;
953 ctx = ACCESS_ONCE(event->ctx);
954 if (!atomic_inc_not_zero(&ctx->refcount)) {
960 mutex_lock_nested(&ctx->mutex, nesting);
961 if (event->ctx != ctx) {
962 mutex_unlock(&ctx->mutex);
970 static inline struct perf_event_context *
971 perf_event_ctx_lock(struct perf_event *event)
973 return perf_event_ctx_lock_nested(event, 0);
976 static void perf_event_ctx_unlock(struct perf_event *event,
977 struct perf_event_context *ctx)
979 mutex_unlock(&ctx->mutex);
984 * This must be done under the ctx->lock, such as to serialize against
985 * context_equiv(), therefore we cannot call put_ctx() since that might end up
986 * calling scheduler related locks and ctx->lock nests inside those.
988 static __must_check struct perf_event_context *
989 unclone_ctx(struct perf_event_context *ctx)
991 struct perf_event_context *parent_ctx = ctx->parent_ctx;
993 lockdep_assert_held(&ctx->lock);
996 ctx->parent_ctx = NULL;
1002 static u32 perf_event_pid(struct perf_event *event, struct task_struct *p)
1005 * only top level events have the pid namespace they were created in
1008 event = event->parent;
1010 return task_tgid_nr_ns(p, event->ns);
1013 static u32 perf_event_tid(struct perf_event *event, struct task_struct *p)
1016 * only top level events have the pid namespace they were created in
1019 event = event->parent;
1021 return task_pid_nr_ns(p, event->ns);
1025 * If we inherit events we want to return the parent event id
1028 static u64 primary_event_id(struct perf_event *event)
1033 id = event->parent->id;
1039 * Get the perf_event_context for a task and lock it.
1040 * This has to cope with with the fact that until it is locked,
1041 * the context could get moved to another task.
1043 static struct perf_event_context *
1044 perf_lock_task_context(struct task_struct *task, int ctxn, unsigned long *flags)
1046 struct perf_event_context *ctx;
1050 * One of the few rules of preemptible RCU is that one cannot do
1051 * rcu_read_unlock() while holding a scheduler (or nested) lock when
1052 * part of the read side critical section was preemptible -- see
1053 * rcu_read_unlock_special().
1055 * Since ctx->lock nests under rq->lock we must ensure the entire read
1056 * side critical section is non-preemptible.
1060 ctx = rcu_dereference(task->perf_event_ctxp[ctxn]);
1063 * If this context is a clone of another, it might
1064 * get swapped for another underneath us by
1065 * perf_event_task_sched_out, though the
1066 * rcu_read_lock() protects us from any context
1067 * getting freed. Lock the context and check if it
1068 * got swapped before we could get the lock, and retry
1069 * if so. If we locked the right context, then it
1070 * can't get swapped on us any more.
1072 raw_spin_lock_irqsave(&ctx->lock, *flags);
1073 if (ctx != rcu_dereference(task->perf_event_ctxp[ctxn])) {
1074 raw_spin_unlock_irqrestore(&ctx->lock, *flags);
1080 if (!atomic_inc_not_zero(&ctx->refcount)) {
1081 raw_spin_unlock_irqrestore(&ctx->lock, *flags);
1091 * Get the context for a task and increment its pin_count so it
1092 * can't get swapped to another task. This also increments its
1093 * reference count so that the context can't get freed.
1095 static struct perf_event_context *
1096 perf_pin_task_context(struct task_struct *task, int ctxn)
1098 struct perf_event_context *ctx;
1099 unsigned long flags;
1101 ctx = perf_lock_task_context(task, ctxn, &flags);
1104 raw_spin_unlock_irqrestore(&ctx->lock, flags);
1109 static void perf_unpin_context(struct perf_event_context *ctx)
1111 unsigned long flags;
1113 raw_spin_lock_irqsave(&ctx->lock, flags);
1115 raw_spin_unlock_irqrestore(&ctx->lock, flags);
1119 * Update the record of the current time in a context.
1121 static void update_context_time(struct perf_event_context *ctx)
1123 u64 now = perf_clock();
1125 ctx->time += now - ctx->timestamp;
1126 ctx->timestamp = now;
1129 static u64 perf_event_time(struct perf_event *event)
1131 struct perf_event_context *ctx = event->ctx;
1133 if (is_cgroup_event(event))
1134 return perf_cgroup_event_time(event);
1136 return ctx ? ctx->time : 0;
1140 * Update the total_time_enabled and total_time_running fields for a event.
1141 * The caller of this function needs to hold the ctx->lock.
1143 static void update_event_times(struct perf_event *event)
1145 struct perf_event_context *ctx = event->ctx;
1148 if (event->state < PERF_EVENT_STATE_INACTIVE ||
1149 event->group_leader->state < PERF_EVENT_STATE_INACTIVE)
1152 * in cgroup mode, time_enabled represents
1153 * the time the event was enabled AND active
1154 * tasks were in the monitored cgroup. This is
1155 * independent of the activity of the context as
1156 * there may be a mix of cgroup and non-cgroup events.
1158 * That is why we treat cgroup events differently
1161 if (is_cgroup_event(event))
1162 run_end = perf_cgroup_event_time(event);
1163 else if (ctx->is_active)
1164 run_end = ctx->time;
1166 run_end = event->tstamp_stopped;
1168 event->total_time_enabled = run_end - event->tstamp_enabled;
1170 if (event->state == PERF_EVENT_STATE_INACTIVE)
1171 run_end = event->tstamp_stopped;
1173 run_end = perf_event_time(event);
1175 event->total_time_running = run_end - event->tstamp_running;
1180 * Update total_time_enabled and total_time_running for all events in a group.
1182 static void update_group_times(struct perf_event *leader)
1184 struct perf_event *event;
1186 update_event_times(leader);
1187 list_for_each_entry(event, &leader->sibling_list, group_entry)
1188 update_event_times(event);
1191 static struct list_head *
1192 ctx_group_list(struct perf_event *event, struct perf_event_context *ctx)
1194 if (event->attr.pinned)
1195 return &ctx->pinned_groups;
1197 return &ctx->flexible_groups;
1201 * Add a event from the lists for its context.
1202 * Must be called with ctx->mutex and ctx->lock held.
1205 list_add_event(struct perf_event *event, struct perf_event_context *ctx)
1207 WARN_ON_ONCE(event->attach_state & PERF_ATTACH_CONTEXT);
1208 event->attach_state |= PERF_ATTACH_CONTEXT;
1211 * If we're a stand alone event or group leader, we go to the context
1212 * list, group events are kept attached to the group so that
1213 * perf_group_detach can, at all times, locate all siblings.
1215 if (event->group_leader == event) {
1216 struct list_head *list;
1218 if (is_software_event(event))
1219 event->group_flags |= PERF_GROUP_SOFTWARE;
1221 list = ctx_group_list(event, ctx);
1222 list_add_tail(&event->group_entry, list);
1225 if (is_cgroup_event(event))
1228 list_add_rcu(&event->event_entry, &ctx->event_list);
1230 if (event->attr.inherit_stat)
1237 * Initialize event state based on the perf_event_attr::disabled.
1239 static inline void perf_event__state_init(struct perf_event *event)
1241 event->state = event->attr.disabled ? PERF_EVENT_STATE_OFF :
1242 PERF_EVENT_STATE_INACTIVE;
1246 * Called at perf_event creation and when events are attached/detached from a
1249 static void perf_event__read_size(struct perf_event *event)
1251 int entry = sizeof(u64); /* value */
1255 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1256 size += sizeof(u64);
1258 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1259 size += sizeof(u64);
1261 if (event->attr.read_format & PERF_FORMAT_ID)
1262 entry += sizeof(u64);
1264 if (event->attr.read_format & PERF_FORMAT_GROUP) {
1265 nr += event->group_leader->nr_siblings;
1266 size += sizeof(u64);
1270 event->read_size = size;
1273 static void perf_event__header_size(struct perf_event *event)
1275 struct perf_sample_data *data;
1276 u64 sample_type = event->attr.sample_type;
1279 perf_event__read_size(event);
1281 if (sample_type & PERF_SAMPLE_IP)
1282 size += sizeof(data->ip);
1284 if (sample_type & PERF_SAMPLE_ADDR)
1285 size += sizeof(data->addr);
1287 if (sample_type & PERF_SAMPLE_PERIOD)
1288 size += sizeof(data->period);
1290 if (sample_type & PERF_SAMPLE_WEIGHT)
1291 size += sizeof(data->weight);
1293 if (sample_type & PERF_SAMPLE_READ)
1294 size += event->read_size;
1296 if (sample_type & PERF_SAMPLE_DATA_SRC)
1297 size += sizeof(data->data_src.val);
1299 if (sample_type & PERF_SAMPLE_TRANSACTION)
1300 size += sizeof(data->txn);
1302 event->header_size = size;
1305 static void perf_event__id_header_size(struct perf_event *event)
1307 struct perf_sample_data *data;
1308 u64 sample_type = event->attr.sample_type;
1311 if (sample_type & PERF_SAMPLE_TID)
1312 size += sizeof(data->tid_entry);
1314 if (sample_type & PERF_SAMPLE_TIME)
1315 size += sizeof(data->time);
1317 if (sample_type & PERF_SAMPLE_IDENTIFIER)
1318 size += sizeof(data->id);
1320 if (sample_type & PERF_SAMPLE_ID)
1321 size += sizeof(data->id);
1323 if (sample_type & PERF_SAMPLE_STREAM_ID)
1324 size += sizeof(data->stream_id);
1326 if (sample_type & PERF_SAMPLE_CPU)
1327 size += sizeof(data->cpu_entry);
1329 event->id_header_size = size;
1332 static void perf_group_attach(struct perf_event *event)
1334 struct perf_event *group_leader = event->group_leader, *pos;
1337 * We can have double attach due to group movement in perf_event_open.
1339 if (event->attach_state & PERF_ATTACH_GROUP)
1342 event->attach_state |= PERF_ATTACH_GROUP;
1344 if (group_leader == event)
1347 WARN_ON_ONCE(group_leader->ctx != event->ctx);
1349 if (group_leader->group_flags & PERF_GROUP_SOFTWARE &&
1350 !is_software_event(event))
1351 group_leader->group_flags &= ~PERF_GROUP_SOFTWARE;
1353 list_add_tail(&event->group_entry, &group_leader->sibling_list);
1354 group_leader->nr_siblings++;
1356 perf_event__header_size(group_leader);
1358 list_for_each_entry(pos, &group_leader->sibling_list, group_entry)
1359 perf_event__header_size(pos);
1363 * Remove a event from the lists for its context.
1364 * Must be called with ctx->mutex and ctx->lock held.
1367 list_del_event(struct perf_event *event, struct perf_event_context *ctx)
1369 struct perf_cpu_context *cpuctx;
1371 WARN_ON_ONCE(event->ctx != ctx);
1372 lockdep_assert_held(&ctx->lock);
1375 * We can have double detach due to exit/hot-unplug + close.
1377 if (!(event->attach_state & PERF_ATTACH_CONTEXT))
1380 event->attach_state &= ~PERF_ATTACH_CONTEXT;
1382 if (is_cgroup_event(event)) {
1384 cpuctx = __get_cpu_context(ctx);
1386 * if there are no more cgroup events
1387 * then cler cgrp to avoid stale pointer
1388 * in update_cgrp_time_from_cpuctx()
1390 if (!ctx->nr_cgroups)
1391 cpuctx->cgrp = NULL;
1395 if (event->attr.inherit_stat)
1398 list_del_rcu(&event->event_entry);
1400 if (event->group_leader == event)
1401 list_del_init(&event->group_entry);
1403 update_group_times(event);
1406 * If event was in error state, then keep it
1407 * that way, otherwise bogus counts will be
1408 * returned on read(). The only way to get out
1409 * of error state is by explicit re-enabling
1412 if (event->state > PERF_EVENT_STATE_OFF)
1413 event->state = PERF_EVENT_STATE_OFF;
1418 static void perf_group_detach(struct perf_event *event)
1420 struct perf_event *sibling, *tmp;
1421 struct list_head *list = NULL;
1424 * We can have double detach due to exit/hot-unplug + close.
1426 if (!(event->attach_state & PERF_ATTACH_GROUP))
1429 event->attach_state &= ~PERF_ATTACH_GROUP;
1432 * If this is a sibling, remove it from its group.
1434 if (event->group_leader != event) {
1435 list_del_init(&event->group_entry);
1436 event->group_leader->nr_siblings--;
1440 if (!list_empty(&event->group_entry))
1441 list = &event->group_entry;
1444 * If this was a group event with sibling events then
1445 * upgrade the siblings to singleton events by adding them
1446 * to whatever list we are on.
1448 list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) {
1450 list_move_tail(&sibling->group_entry, list);
1451 sibling->group_leader = sibling;
1453 /* Inherit group flags from the previous leader */
1454 sibling->group_flags = event->group_flags;
1456 WARN_ON_ONCE(sibling->ctx != event->ctx);
1460 perf_event__header_size(event->group_leader);
1462 list_for_each_entry(tmp, &event->group_leader->sibling_list, group_entry)
1463 perf_event__header_size(tmp);
1467 * User event without the task.
1469 static bool is_orphaned_event(struct perf_event *event)
1471 return event && !is_kernel_event(event) && !event->owner;
1475 * Event has a parent but parent's task finished and it's
1476 * alive only because of children holding refference.
1478 static bool is_orphaned_child(struct perf_event *event)
1480 return is_orphaned_event(event->parent);
1483 static void orphans_remove_work(struct work_struct *work);
1485 static void schedule_orphans_remove(struct perf_event_context *ctx)
1487 if (!ctx->task || ctx->orphans_remove_sched || !perf_wq)
1490 if (queue_delayed_work(perf_wq, &ctx->orphans_remove, 1)) {
1492 ctx->orphans_remove_sched = true;
1496 static int __init perf_workqueue_init(void)
1498 perf_wq = create_singlethread_workqueue("perf");
1499 WARN(!perf_wq, "failed to create perf workqueue\n");
1500 return perf_wq ? 0 : -1;
1503 core_initcall(perf_workqueue_init);
1505 static inline int pmu_filter_match(struct perf_event *event)
1507 struct pmu *pmu = event->pmu;
1508 return pmu->filter_match ? pmu->filter_match(event) : 1;
1512 event_filter_match(struct perf_event *event)
1514 return (event->cpu == -1 || event->cpu == smp_processor_id())
1515 && perf_cgroup_match(event) && pmu_filter_match(event);
1519 event_sched_out(struct perf_event *event,
1520 struct perf_cpu_context *cpuctx,
1521 struct perf_event_context *ctx)
1523 u64 tstamp = perf_event_time(event);
1526 WARN_ON_ONCE(event->ctx != ctx);
1527 lockdep_assert_held(&ctx->lock);
1530 * An event which could not be activated because of
1531 * filter mismatch still needs to have its timings
1532 * maintained, otherwise bogus information is return
1533 * via read() for time_enabled, time_running:
1535 if (event->state == PERF_EVENT_STATE_INACTIVE
1536 && !event_filter_match(event)) {
1537 delta = tstamp - event->tstamp_stopped;
1538 event->tstamp_running += delta;
1539 event->tstamp_stopped = tstamp;
1542 if (event->state != PERF_EVENT_STATE_ACTIVE)
1545 perf_pmu_disable(event->pmu);
1547 event->state = PERF_EVENT_STATE_INACTIVE;
1548 if (event->pending_disable) {
1549 event->pending_disable = 0;
1550 event->state = PERF_EVENT_STATE_OFF;
1552 event->tstamp_stopped = tstamp;
1553 event->pmu->del(event, 0);
1556 if (!is_software_event(event))
1557 cpuctx->active_oncpu--;
1558 if (!--ctx->nr_active)
1559 perf_event_ctx_deactivate(ctx);
1560 if (event->attr.freq && event->attr.sample_freq)
1562 if (event->attr.exclusive || !cpuctx->active_oncpu)
1563 cpuctx->exclusive = 0;
1565 if (is_orphaned_child(event))
1566 schedule_orphans_remove(ctx);
1568 perf_pmu_enable(event->pmu);
1572 group_sched_out(struct perf_event *group_event,
1573 struct perf_cpu_context *cpuctx,
1574 struct perf_event_context *ctx)
1576 struct perf_event *event;
1577 int state = group_event->state;
1579 event_sched_out(group_event, cpuctx, ctx);
1582 * Schedule out siblings (if any):
1584 list_for_each_entry(event, &group_event->sibling_list, group_entry)
1585 event_sched_out(event, cpuctx, ctx);
1587 if (state == PERF_EVENT_STATE_ACTIVE && group_event->attr.exclusive)
1588 cpuctx->exclusive = 0;
1591 struct remove_event {
1592 struct perf_event *event;
1597 * Cross CPU call to remove a performance event
1599 * We disable the event on the hardware level first. After that we
1600 * remove it from the context list.
1602 static int __perf_remove_from_context(void *info)
1604 struct remove_event *re = info;
1605 struct perf_event *event = re->event;
1606 struct perf_event_context *ctx = event->ctx;
1607 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1609 raw_spin_lock(&ctx->lock);
1610 event_sched_out(event, cpuctx, ctx);
1611 if (re->detach_group)
1612 perf_group_detach(event);
1613 list_del_event(event, ctx);
1614 if (!ctx->nr_events && cpuctx->task_ctx == ctx) {
1616 cpuctx->task_ctx = NULL;
1618 raw_spin_unlock(&ctx->lock);
1625 * Remove the event from a task's (or a CPU's) list of events.
1627 * CPU events are removed with a smp call. For task events we only
1628 * call when the task is on a CPU.
1630 * If event->ctx is a cloned context, callers must make sure that
1631 * every task struct that event->ctx->task could possibly point to
1632 * remains valid. This is OK when called from perf_release since
1633 * that only calls us on the top-level context, which can't be a clone.
1634 * When called from perf_event_exit_task, it's OK because the
1635 * context has been detached from its task.
1637 static void perf_remove_from_context(struct perf_event *event, bool detach_group)
1639 struct perf_event_context *ctx = event->ctx;
1640 struct task_struct *task = ctx->task;
1641 struct remove_event re = {
1643 .detach_group = detach_group,
1646 lockdep_assert_held(&ctx->mutex);
1650 * Per cpu events are removed via an smp call. The removal can
1651 * fail if the CPU is currently offline, but in that case we
1652 * already called __perf_remove_from_context from
1653 * perf_event_exit_cpu.
1655 cpu_function_call(event->cpu, __perf_remove_from_context, &re);
1660 if (!task_function_call(task, __perf_remove_from_context, &re))
1663 raw_spin_lock_irq(&ctx->lock);
1665 * If we failed to find a running task, but find the context active now
1666 * that we've acquired the ctx->lock, retry.
1668 if (ctx->is_active) {
1669 raw_spin_unlock_irq(&ctx->lock);
1671 * Reload the task pointer, it might have been changed by
1672 * a concurrent perf_event_context_sched_out().
1679 * Since the task isn't running, its safe to remove the event, us
1680 * holding the ctx->lock ensures the task won't get scheduled in.
1683 perf_group_detach(event);
1684 list_del_event(event, ctx);
1685 raw_spin_unlock_irq(&ctx->lock);
1689 * Cross CPU call to disable a performance event
1691 int __perf_event_disable(void *info)
1693 struct perf_event *event = info;
1694 struct perf_event_context *ctx = event->ctx;
1695 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1698 * If this is a per-task event, need to check whether this
1699 * event's task is the current task on this cpu.
1701 * Can trigger due to concurrent perf_event_context_sched_out()
1702 * flipping contexts around.
1704 if (ctx->task && cpuctx->task_ctx != ctx)
1707 raw_spin_lock(&ctx->lock);
1710 * If the event is on, turn it off.
1711 * If it is in error state, leave it in error state.
1713 if (event->state >= PERF_EVENT_STATE_INACTIVE) {
1714 update_context_time(ctx);
1715 update_cgrp_time_from_event(event);
1716 update_group_times(event);
1717 if (event == event->group_leader)
1718 group_sched_out(event, cpuctx, ctx);
1720 event_sched_out(event, cpuctx, ctx);
1721 event->state = PERF_EVENT_STATE_OFF;
1724 raw_spin_unlock(&ctx->lock);
1732 * If event->ctx is a cloned context, callers must make sure that
1733 * every task struct that event->ctx->task could possibly point to
1734 * remains valid. This condition is satisifed when called through
1735 * perf_event_for_each_child or perf_event_for_each because they
1736 * hold the top-level event's child_mutex, so any descendant that
1737 * goes to exit will block in sync_child_event.
1738 * When called from perf_pending_event it's OK because event->ctx
1739 * is the current context on this CPU and preemption is disabled,
1740 * hence we can't get into perf_event_task_sched_out for this context.
1742 static void _perf_event_disable(struct perf_event *event)
1744 struct perf_event_context *ctx = event->ctx;
1745 struct task_struct *task = ctx->task;
1749 * Disable the event on the cpu that it's on
1751 cpu_function_call(event->cpu, __perf_event_disable, event);
1756 if (!task_function_call(task, __perf_event_disable, event))
1759 raw_spin_lock_irq(&ctx->lock);
1761 * If the event is still active, we need to retry the cross-call.
1763 if (event->state == PERF_EVENT_STATE_ACTIVE) {
1764 raw_spin_unlock_irq(&ctx->lock);
1766 * Reload the task pointer, it might have been changed by
1767 * a concurrent perf_event_context_sched_out().
1774 * Since we have the lock this context can't be scheduled
1775 * in, so we can change the state safely.
1777 if (event->state == PERF_EVENT_STATE_INACTIVE) {
1778 update_group_times(event);
1779 event->state = PERF_EVENT_STATE_OFF;
1781 raw_spin_unlock_irq(&ctx->lock);
1785 * Strictly speaking kernel users cannot create groups and therefore this
1786 * interface does not need the perf_event_ctx_lock() magic.
1788 void perf_event_disable(struct perf_event *event)
1790 struct perf_event_context *ctx;
1792 ctx = perf_event_ctx_lock(event);
1793 _perf_event_disable(event);
1794 perf_event_ctx_unlock(event, ctx);
1796 EXPORT_SYMBOL_GPL(perf_event_disable);
1798 static void perf_set_shadow_time(struct perf_event *event,
1799 struct perf_event_context *ctx,
1803 * use the correct time source for the time snapshot
1805 * We could get by without this by leveraging the
1806 * fact that to get to this function, the caller
1807 * has most likely already called update_context_time()
1808 * and update_cgrp_time_xx() and thus both timestamp
1809 * are identical (or very close). Given that tstamp is,
1810 * already adjusted for cgroup, we could say that:
1811 * tstamp - ctx->timestamp
1813 * tstamp - cgrp->timestamp.
1815 * Then, in perf_output_read(), the calculation would
1816 * work with no changes because:
1817 * - event is guaranteed scheduled in
1818 * - no scheduled out in between
1819 * - thus the timestamp would be the same
1821 * But this is a bit hairy.
1823 * So instead, we have an explicit cgroup call to remain
1824 * within the time time source all along. We believe it
1825 * is cleaner and simpler to understand.
1827 if (is_cgroup_event(event))
1828 perf_cgroup_set_shadow_time(event, tstamp);
1830 event->shadow_ctx_time = tstamp - ctx->timestamp;
1833 #define MAX_INTERRUPTS (~0ULL)
1835 static void perf_log_throttle(struct perf_event *event, int enable);
1836 static void perf_log_itrace_start(struct perf_event *event);
1839 event_sched_in(struct perf_event *event,
1840 struct perf_cpu_context *cpuctx,
1841 struct perf_event_context *ctx)
1843 u64 tstamp = perf_event_time(event);
1846 lockdep_assert_held(&ctx->lock);
1848 if (event->state <= PERF_EVENT_STATE_OFF)
1851 event->state = PERF_EVENT_STATE_ACTIVE;
1852 event->oncpu = smp_processor_id();
1855 * Unthrottle events, since we scheduled we might have missed several
1856 * ticks already, also for a heavily scheduling task there is little
1857 * guarantee it'll get a tick in a timely manner.
1859 if (unlikely(event->hw.interrupts == MAX_INTERRUPTS)) {
1860 perf_log_throttle(event, 1);
1861 event->hw.interrupts = 0;
1865 * The new state must be visible before we turn it on in the hardware:
1869 perf_pmu_disable(event->pmu);
1871 event->tstamp_running += tstamp - event->tstamp_stopped;
1873 perf_set_shadow_time(event, ctx, tstamp);
1875 perf_log_itrace_start(event);
1877 if (event->pmu->add(event, PERF_EF_START)) {
1878 event->state = PERF_EVENT_STATE_INACTIVE;
1884 if (!is_software_event(event))
1885 cpuctx->active_oncpu++;
1886 if (!ctx->nr_active++)
1887 perf_event_ctx_activate(ctx);
1888 if (event->attr.freq && event->attr.sample_freq)
1891 if (event->attr.exclusive)
1892 cpuctx->exclusive = 1;
1894 if (is_orphaned_child(event))
1895 schedule_orphans_remove(ctx);
1898 perf_pmu_enable(event->pmu);
1904 group_sched_in(struct perf_event *group_event,
1905 struct perf_cpu_context *cpuctx,
1906 struct perf_event_context *ctx)
1908 struct perf_event *event, *partial_group = NULL;
1909 struct pmu *pmu = ctx->pmu;
1910 u64 now = ctx->time;
1911 bool simulate = false;
1913 if (group_event->state == PERF_EVENT_STATE_OFF)
1916 pmu->start_txn(pmu);
1918 if (event_sched_in(group_event, cpuctx, ctx)) {
1919 pmu->cancel_txn(pmu);
1920 perf_mux_hrtimer_restart(cpuctx);
1925 * Schedule in siblings as one group (if any):
1927 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
1928 if (event_sched_in(event, cpuctx, ctx)) {
1929 partial_group = event;
1934 if (!pmu->commit_txn(pmu))
1939 * Groups can be scheduled in as one unit only, so undo any
1940 * partial group before returning:
1941 * The events up to the failed event are scheduled out normally,
1942 * tstamp_stopped will be updated.
1944 * The failed events and the remaining siblings need to have
1945 * their timings updated as if they had gone thru event_sched_in()
1946 * and event_sched_out(). This is required to get consistent timings
1947 * across the group. This also takes care of the case where the group
1948 * could never be scheduled by ensuring tstamp_stopped is set to mark
1949 * the time the event was actually stopped, such that time delta
1950 * calculation in update_event_times() is correct.
1952 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
1953 if (event == partial_group)
1957 event->tstamp_running += now - event->tstamp_stopped;
1958 event->tstamp_stopped = now;
1960 event_sched_out(event, cpuctx, ctx);
1963 event_sched_out(group_event, cpuctx, ctx);
1965 pmu->cancel_txn(pmu);
1967 perf_mux_hrtimer_restart(cpuctx);
1973 * Work out whether we can put this event group on the CPU now.
1975 static int group_can_go_on(struct perf_event *event,
1976 struct perf_cpu_context *cpuctx,
1980 * Groups consisting entirely of software events can always go on.
1982 if (event->group_flags & PERF_GROUP_SOFTWARE)
1985 * If an exclusive group is already on, no other hardware
1988 if (cpuctx->exclusive)
1991 * If this group is exclusive and there are already
1992 * events on the CPU, it can't go on.
1994 if (event->attr.exclusive && cpuctx->active_oncpu)
1997 * Otherwise, try to add it if all previous groups were able
2003 static void add_event_to_ctx(struct perf_event *event,
2004 struct perf_event_context *ctx)
2006 u64 tstamp = perf_event_time(event);
2008 list_add_event(event, ctx);
2009 perf_group_attach(event);
2010 event->tstamp_enabled = tstamp;
2011 event->tstamp_running = tstamp;
2012 event->tstamp_stopped = tstamp;
2015 static void task_ctx_sched_out(struct perf_event_context *ctx);
2017 ctx_sched_in(struct perf_event_context *ctx,
2018 struct perf_cpu_context *cpuctx,
2019 enum event_type_t event_type,
2020 struct task_struct *task);
2022 static void perf_event_sched_in(struct perf_cpu_context *cpuctx,
2023 struct perf_event_context *ctx,
2024 struct task_struct *task)
2026 cpu_ctx_sched_in(cpuctx, EVENT_PINNED, task);
2028 ctx_sched_in(ctx, cpuctx, EVENT_PINNED, task);
2029 cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE, task);
2031 ctx_sched_in(ctx, cpuctx, EVENT_FLEXIBLE, task);
2035 * Cross CPU call to install and enable a performance event
2037 * Must be called with ctx->mutex held
2039 static int __perf_install_in_context(void *info)
2041 struct perf_event *event = info;
2042 struct perf_event_context *ctx = event->ctx;
2043 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
2044 struct perf_event_context *task_ctx = cpuctx->task_ctx;
2045 struct task_struct *task = current;
2047 perf_ctx_lock(cpuctx, task_ctx);
2048 perf_pmu_disable(cpuctx->ctx.pmu);
2051 * If there was an active task_ctx schedule it out.
2054 task_ctx_sched_out(task_ctx);
2057 * If the context we're installing events in is not the
2058 * active task_ctx, flip them.
2060 if (ctx->task && task_ctx != ctx) {
2062 raw_spin_unlock(&task_ctx->lock);
2063 raw_spin_lock(&ctx->lock);
2068 cpuctx->task_ctx = task_ctx;
2069 task = task_ctx->task;
2072 cpu_ctx_sched_out(cpuctx, EVENT_ALL);
2074 update_context_time(ctx);
2076 * update cgrp time only if current cgrp
2077 * matches event->cgrp. Must be done before
2078 * calling add_event_to_ctx()
2080 update_cgrp_time_from_event(event);
2082 add_event_to_ctx(event, ctx);
2085 * Schedule everything back in
2087 perf_event_sched_in(cpuctx, task_ctx, task);
2089 perf_pmu_enable(cpuctx->ctx.pmu);
2090 perf_ctx_unlock(cpuctx, task_ctx);
2096 * Attach a performance event to a context
2098 * First we add the event to the list with the hardware enable bit
2099 * in event->hw_config cleared.
2101 * If the event is attached to a task which is on a CPU we use a smp
2102 * call to enable it in the task context. The task might have been
2103 * scheduled away, but we check this in the smp call again.
2106 perf_install_in_context(struct perf_event_context *ctx,
2107 struct perf_event *event,
2110 struct task_struct *task = ctx->task;
2112 lockdep_assert_held(&ctx->mutex);
2115 if (event->cpu != -1)
2120 * Per cpu events are installed via an smp call and
2121 * the install is always successful.
2123 cpu_function_call(cpu, __perf_install_in_context, event);
2128 if (!task_function_call(task, __perf_install_in_context, event))
2131 raw_spin_lock_irq(&ctx->lock);
2133 * If we failed to find a running task, but find the context active now
2134 * that we've acquired the ctx->lock, retry.
2136 if (ctx->is_active) {
2137 raw_spin_unlock_irq(&ctx->lock);
2139 * Reload the task pointer, it might have been changed by
2140 * a concurrent perf_event_context_sched_out().
2147 * Since the task isn't running, its safe to add the event, us holding
2148 * the ctx->lock ensures the task won't get scheduled in.
2150 add_event_to_ctx(event, ctx);
2151 raw_spin_unlock_irq(&ctx->lock);
2155 * Put a event into inactive state and update time fields.
2156 * Enabling the leader of a group effectively enables all
2157 * the group members that aren't explicitly disabled, so we
2158 * have to update their ->tstamp_enabled also.
2159 * Note: this works for group members as well as group leaders
2160 * since the non-leader members' sibling_lists will be empty.
2162 static void __perf_event_mark_enabled(struct perf_event *event)
2164 struct perf_event *sub;
2165 u64 tstamp = perf_event_time(event);
2167 event->state = PERF_EVENT_STATE_INACTIVE;
2168 event->tstamp_enabled = tstamp - event->total_time_enabled;
2169 list_for_each_entry(sub, &event->sibling_list, group_entry) {
2170 if (sub->state >= PERF_EVENT_STATE_INACTIVE)
2171 sub->tstamp_enabled = tstamp - sub->total_time_enabled;
2176 * Cross CPU call to enable a performance event
2178 static int __perf_event_enable(void *info)
2180 struct perf_event *event = info;
2181 struct perf_event_context *ctx = event->ctx;
2182 struct perf_event *leader = event->group_leader;
2183 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
2187 * There's a time window between 'ctx->is_active' check
2188 * in perf_event_enable function and this place having:
2190 * - ctx->lock unlocked
2192 * where the task could be killed and 'ctx' deactivated
2193 * by perf_event_exit_task.
2195 if (!ctx->is_active)
2198 raw_spin_lock(&ctx->lock);
2199 update_context_time(ctx);
2201 if (event->state >= PERF_EVENT_STATE_INACTIVE)
2205 * set current task's cgroup time reference point
2207 perf_cgroup_set_timestamp(current, ctx);
2209 __perf_event_mark_enabled(event);
2211 if (!event_filter_match(event)) {
2212 if (is_cgroup_event(event))
2213 perf_cgroup_defer_enabled(event);
2218 * If the event is in a group and isn't the group leader,
2219 * then don't put it on unless the group is on.
2221 if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE)
2224 if (!group_can_go_on(event, cpuctx, 1)) {
2227 if (event == leader)
2228 err = group_sched_in(event, cpuctx, ctx);
2230 err = event_sched_in(event, cpuctx, ctx);
2235 * If this event can't go on and it's part of a
2236 * group, then the whole group has to come off.
2238 if (leader != event) {
2239 group_sched_out(leader, cpuctx, ctx);
2240 perf_mux_hrtimer_restart(cpuctx);
2242 if (leader->attr.pinned) {
2243 update_group_times(leader);
2244 leader->state = PERF_EVENT_STATE_ERROR;
2249 raw_spin_unlock(&ctx->lock);
2257 * If event->ctx is a cloned context, callers must make sure that
2258 * every task struct that event->ctx->task could possibly point to
2259 * remains valid. This condition is satisfied when called through
2260 * perf_event_for_each_child or perf_event_for_each as described
2261 * for perf_event_disable.
2263 static void _perf_event_enable(struct perf_event *event)
2265 struct perf_event_context *ctx = event->ctx;
2266 struct task_struct *task = ctx->task;
2270 * Enable the event on the cpu that it's on
2272 cpu_function_call(event->cpu, __perf_event_enable, event);
2276 raw_spin_lock_irq(&ctx->lock);
2277 if (event->state >= PERF_EVENT_STATE_INACTIVE)
2281 * If the event is in error state, clear that first.
2282 * That way, if we see the event in error state below, we
2283 * know that it has gone back into error state, as distinct
2284 * from the task having been scheduled away before the
2285 * cross-call arrived.
2287 if (event->state == PERF_EVENT_STATE_ERROR)
2288 event->state = PERF_EVENT_STATE_OFF;
2291 if (!ctx->is_active) {
2292 __perf_event_mark_enabled(event);
2296 raw_spin_unlock_irq(&ctx->lock);
2298 if (!task_function_call(task, __perf_event_enable, event))
2301 raw_spin_lock_irq(&ctx->lock);
2304 * If the context is active and the event is still off,
2305 * we need to retry the cross-call.
2307 if (ctx->is_active && event->state == PERF_EVENT_STATE_OFF) {
2309 * task could have been flipped by a concurrent
2310 * perf_event_context_sched_out()
2317 raw_spin_unlock_irq(&ctx->lock);
2321 * See perf_event_disable();
2323 void perf_event_enable(struct perf_event *event)
2325 struct perf_event_context *ctx;
2327 ctx = perf_event_ctx_lock(event);
2328 _perf_event_enable(event);
2329 perf_event_ctx_unlock(event, ctx);
2331 EXPORT_SYMBOL_GPL(perf_event_enable);
2333 static int _perf_event_refresh(struct perf_event *event, int refresh)
2336 * not supported on inherited events
2338 if (event->attr.inherit || !is_sampling_event(event))
2341 atomic_add(refresh, &event->event_limit);
2342 _perf_event_enable(event);
2348 * See perf_event_disable()
2350 int perf_event_refresh(struct perf_event *event, int refresh)
2352 struct perf_event_context *ctx;
2355 ctx = perf_event_ctx_lock(event);
2356 ret = _perf_event_refresh(event, refresh);
2357 perf_event_ctx_unlock(event, ctx);
2361 EXPORT_SYMBOL_GPL(perf_event_refresh);
2363 static void ctx_sched_out(struct perf_event_context *ctx,
2364 struct perf_cpu_context *cpuctx,
2365 enum event_type_t event_type)
2367 struct perf_event *event;
2368 int is_active = ctx->is_active;
2370 ctx->is_active &= ~event_type;
2371 if (likely(!ctx->nr_events))
2374 update_context_time(ctx);
2375 update_cgrp_time_from_cpuctx(cpuctx);
2376 if (!ctx->nr_active)
2379 perf_pmu_disable(ctx->pmu);
2380 if ((is_active & EVENT_PINNED) && (event_type & EVENT_PINNED)) {
2381 list_for_each_entry(event, &ctx->pinned_groups, group_entry)
2382 group_sched_out(event, cpuctx, ctx);
2385 if ((is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE)) {
2386 list_for_each_entry(event, &ctx->flexible_groups, group_entry)
2387 group_sched_out(event, cpuctx, ctx);
2389 perf_pmu_enable(ctx->pmu);
2393 * Test whether two contexts are equivalent, i.e. whether they have both been
2394 * cloned from the same version of the same context.
2396 * Equivalence is measured using a generation number in the context that is
2397 * incremented on each modification to it; see unclone_ctx(), list_add_event()
2398 * and list_del_event().
2400 static int context_equiv(struct perf_event_context *ctx1,
2401 struct perf_event_context *ctx2)
2403 lockdep_assert_held(&ctx1->lock);
2404 lockdep_assert_held(&ctx2->lock);
2406 /* Pinning disables the swap optimization */
2407 if (ctx1->pin_count || ctx2->pin_count)
2410 /* If ctx1 is the parent of ctx2 */
2411 if (ctx1 == ctx2->parent_ctx && ctx1->generation == ctx2->parent_gen)
2414 /* If ctx2 is the parent of ctx1 */
2415 if (ctx1->parent_ctx == ctx2 && ctx1->parent_gen == ctx2->generation)
2419 * If ctx1 and ctx2 have the same parent; we flatten the parent
2420 * hierarchy, see perf_event_init_context().
2422 if (ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx &&
2423 ctx1->parent_gen == ctx2->parent_gen)
2430 static void __perf_event_sync_stat(struct perf_event *event,
2431 struct perf_event *next_event)
2435 if (!event->attr.inherit_stat)
2439 * Update the event value, we cannot use perf_event_read()
2440 * because we're in the middle of a context switch and have IRQs
2441 * disabled, which upsets smp_call_function_single(), however
2442 * we know the event must be on the current CPU, therefore we
2443 * don't need to use it.
2445 switch (event->state) {
2446 case PERF_EVENT_STATE_ACTIVE:
2447 event->pmu->read(event);
2450 case PERF_EVENT_STATE_INACTIVE:
2451 update_event_times(event);
2459 * In order to keep per-task stats reliable we need to flip the event
2460 * values when we flip the contexts.
2462 value = local64_read(&next_event->count);
2463 value = local64_xchg(&event->count, value);
2464 local64_set(&next_event->count, value);
2466 swap(event->total_time_enabled, next_event->total_time_enabled);
2467 swap(event->total_time_running, next_event->total_time_running);
2470 * Since we swizzled the values, update the user visible data too.
2472 perf_event_update_userpage(event);
2473 perf_event_update_userpage(next_event);
2476 static void perf_event_sync_stat(struct perf_event_context *ctx,
2477 struct perf_event_context *next_ctx)
2479 struct perf_event *event, *next_event;
2484 update_context_time(ctx);
2486 event = list_first_entry(&ctx->event_list,
2487 struct perf_event, event_entry);
2489 next_event = list_first_entry(&next_ctx->event_list,
2490 struct perf_event, event_entry);
2492 while (&event->event_entry != &ctx->event_list &&
2493 &next_event->event_entry != &next_ctx->event_list) {
2495 __perf_event_sync_stat(event, next_event);
2497 event = list_next_entry(event, event_entry);
2498 next_event = list_next_entry(next_event, event_entry);
2502 static void perf_event_context_sched_out(struct task_struct *task, int ctxn,
2503 struct task_struct *next)
2505 struct perf_event_context *ctx = task->perf_event_ctxp[ctxn];
2506 struct perf_event_context *next_ctx;
2507 struct perf_event_context *parent, *next_parent;
2508 struct perf_cpu_context *cpuctx;
2514 cpuctx = __get_cpu_context(ctx);
2515 if (!cpuctx->task_ctx)
2519 next_ctx = next->perf_event_ctxp[ctxn];
2523 parent = rcu_dereference(ctx->parent_ctx);
2524 next_parent = rcu_dereference(next_ctx->parent_ctx);
2526 /* If neither context have a parent context; they cannot be clones. */
2527 if (!parent && !next_parent)
2530 if (next_parent == ctx || next_ctx == parent || next_parent == parent) {
2532 * Looks like the two contexts are clones, so we might be
2533 * able to optimize the context switch. We lock both
2534 * contexts and check that they are clones under the
2535 * lock (including re-checking that neither has been
2536 * uncloned in the meantime). It doesn't matter which
2537 * order we take the locks because no other cpu could
2538 * be trying to lock both of these tasks.
2540 raw_spin_lock(&ctx->lock);
2541 raw_spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING);
2542 if (context_equiv(ctx, next_ctx)) {
2544 * XXX do we need a memory barrier of sorts
2545 * wrt to rcu_dereference() of perf_event_ctxp
2547 task->perf_event_ctxp[ctxn] = next_ctx;
2548 next->perf_event_ctxp[ctxn] = ctx;
2550 next_ctx->task = task;
2552 swap(ctx->task_ctx_data, next_ctx->task_ctx_data);
2556 perf_event_sync_stat(ctx, next_ctx);
2558 raw_spin_unlock(&next_ctx->lock);
2559 raw_spin_unlock(&ctx->lock);
2565 raw_spin_lock(&ctx->lock);
2566 ctx_sched_out(ctx, cpuctx, EVENT_ALL);
2567 cpuctx->task_ctx = NULL;
2568 raw_spin_unlock(&ctx->lock);
2572 void perf_sched_cb_dec(struct pmu *pmu)
2574 this_cpu_dec(perf_sched_cb_usages);
2577 void perf_sched_cb_inc(struct pmu *pmu)
2579 this_cpu_inc(perf_sched_cb_usages);
2583 * This function provides the context switch callback to the lower code
2584 * layer. It is invoked ONLY when the context switch callback is enabled.
2586 static void perf_pmu_sched_task(struct task_struct *prev,
2587 struct task_struct *next,
2590 struct perf_cpu_context *cpuctx;
2592 unsigned long flags;
2597 local_irq_save(flags);
2601 list_for_each_entry_rcu(pmu, &pmus, entry) {
2602 if (pmu->sched_task) {
2603 cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
2605 perf_ctx_lock(cpuctx, cpuctx->task_ctx);
2607 perf_pmu_disable(pmu);
2609 pmu->sched_task(cpuctx->task_ctx, sched_in);
2611 perf_pmu_enable(pmu);
2613 perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
2619 local_irq_restore(flags);
2622 #define for_each_task_context_nr(ctxn) \
2623 for ((ctxn) = 0; (ctxn) < perf_nr_task_contexts; (ctxn)++)
2626 * Called from scheduler to remove the events of the current task,
2627 * with interrupts disabled.
2629 * We stop each event and update the event value in event->count.
2631 * This does not protect us against NMI, but disable()
2632 * sets the disabled bit in the control field of event _before_
2633 * accessing the event control register. If a NMI hits, then it will
2634 * not restart the event.
2636 void __perf_event_task_sched_out(struct task_struct *task,
2637 struct task_struct *next)
2641 if (__this_cpu_read(perf_sched_cb_usages))
2642 perf_pmu_sched_task(task, next, false);
2644 for_each_task_context_nr(ctxn)
2645 perf_event_context_sched_out(task, ctxn, next);
2648 * if cgroup events exist on this CPU, then we need
2649 * to check if we have to switch out PMU state.
2650 * cgroup event are system-wide mode only
2652 if (atomic_read(this_cpu_ptr(&perf_cgroup_events)))
2653 perf_cgroup_sched_out(task, next);
2656 static void task_ctx_sched_out(struct perf_event_context *ctx)
2658 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
2660 if (!cpuctx->task_ctx)
2663 if (WARN_ON_ONCE(ctx != cpuctx->task_ctx))
2666 ctx_sched_out(ctx, cpuctx, EVENT_ALL);
2667 cpuctx->task_ctx = NULL;
2671 * Called with IRQs disabled
2673 static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
2674 enum event_type_t event_type)
2676 ctx_sched_out(&cpuctx->ctx, cpuctx, event_type);
2680 ctx_pinned_sched_in(struct perf_event_context *ctx,
2681 struct perf_cpu_context *cpuctx)
2683 struct perf_event *event;
2685 list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
2686 if (event->state <= PERF_EVENT_STATE_OFF)
2688 if (!event_filter_match(event))
2691 /* may need to reset tstamp_enabled */
2692 if (is_cgroup_event(event))
2693 perf_cgroup_mark_enabled(event, ctx);
2695 if (group_can_go_on(event, cpuctx, 1))
2696 group_sched_in(event, cpuctx, ctx);
2699 * If this pinned group hasn't been scheduled,
2700 * put it in error state.
2702 if (event->state == PERF_EVENT_STATE_INACTIVE) {
2703 update_group_times(event);
2704 event->state = PERF_EVENT_STATE_ERROR;
2710 ctx_flexible_sched_in(struct perf_event_context *ctx,
2711 struct perf_cpu_context *cpuctx)
2713 struct perf_event *event;
2716 list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
2717 /* Ignore events in OFF or ERROR state */
2718 if (event->state <= PERF_EVENT_STATE_OFF)
2721 * Listen to the 'cpu' scheduling filter constraint
2724 if (!event_filter_match(event))
2727 /* may need to reset tstamp_enabled */
2728 if (is_cgroup_event(event))
2729 perf_cgroup_mark_enabled(event, ctx);
2731 if (group_can_go_on(event, cpuctx, can_add_hw)) {
2732 if (group_sched_in(event, cpuctx, ctx))
2739 ctx_sched_in(struct perf_event_context *ctx,
2740 struct perf_cpu_context *cpuctx,
2741 enum event_type_t event_type,
2742 struct task_struct *task)
2745 int is_active = ctx->is_active;
2747 ctx->is_active |= event_type;
2748 if (likely(!ctx->nr_events))
2752 ctx->timestamp = now;
2753 perf_cgroup_set_timestamp(task, ctx);
2755 * First go through the list and put on any pinned groups
2756 * in order to give them the best chance of going on.
2758 if (!(is_active & EVENT_PINNED) && (event_type & EVENT_PINNED))
2759 ctx_pinned_sched_in(ctx, cpuctx);
2761 /* Then walk through the lower prio flexible groups */
2762 if (!(is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE))
2763 ctx_flexible_sched_in(ctx, cpuctx);
2766 static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
2767 enum event_type_t event_type,
2768 struct task_struct *task)
2770 struct perf_event_context *ctx = &cpuctx->ctx;
2772 ctx_sched_in(ctx, cpuctx, event_type, task);
2775 static void perf_event_context_sched_in(struct perf_event_context *ctx,
2776 struct task_struct *task)
2778 struct perf_cpu_context *cpuctx;
2780 cpuctx = __get_cpu_context(ctx);
2781 if (cpuctx->task_ctx == ctx)
2784 perf_ctx_lock(cpuctx, ctx);
2785 perf_pmu_disable(ctx->pmu);
2787 * We want to keep the following priority order:
2788 * cpu pinned (that don't need to move), task pinned,
2789 * cpu flexible, task flexible.
2791 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
2794 cpuctx->task_ctx = ctx;
2796 perf_event_sched_in(cpuctx, cpuctx->task_ctx, task);
2798 perf_pmu_enable(ctx->pmu);
2799 perf_ctx_unlock(cpuctx, ctx);
2803 * Called from scheduler to add the events of the current task
2804 * with interrupts disabled.
2806 * We restore the event value and then enable it.
2808 * This does not protect us against NMI, but enable()
2809 * sets the enabled bit in the control field of event _before_
2810 * accessing the event control register. If a NMI hits, then it will
2811 * keep the event running.
2813 void __perf_event_task_sched_in(struct task_struct *prev,
2814 struct task_struct *task)
2816 struct perf_event_context *ctx;
2819 for_each_task_context_nr(ctxn) {
2820 ctx = task->perf_event_ctxp[ctxn];
2824 perf_event_context_sched_in(ctx, task);
2827 * if cgroup events exist on this CPU, then we need
2828 * to check if we have to switch in PMU state.
2829 * cgroup event are system-wide mode only
2831 if (atomic_read(this_cpu_ptr(&perf_cgroup_events)))
2832 perf_cgroup_sched_in(prev, task);
2834 if (__this_cpu_read(perf_sched_cb_usages))
2835 perf_pmu_sched_task(prev, task, true);
2838 static u64 perf_calculate_period(struct perf_event *event, u64 nsec, u64 count)
2840 u64 frequency = event->attr.sample_freq;
2841 u64 sec = NSEC_PER_SEC;
2842 u64 divisor, dividend;
2844 int count_fls, nsec_fls, frequency_fls, sec_fls;
2846 count_fls = fls64(count);
2847 nsec_fls = fls64(nsec);
2848 frequency_fls = fls64(frequency);
2852 * We got @count in @nsec, with a target of sample_freq HZ
2853 * the target period becomes:
2856 * period = -------------------
2857 * @nsec * sample_freq
2862 * Reduce accuracy by one bit such that @a and @b converge
2863 * to a similar magnitude.
2865 #define REDUCE_FLS(a, b) \
2867 if (a##_fls > b##_fls) { \
2877 * Reduce accuracy until either term fits in a u64, then proceed with
2878 * the other, so that finally we can do a u64/u64 division.
2880 while (count_fls + sec_fls > 64 && nsec_fls + frequency_fls > 64) {
2881 REDUCE_FLS(nsec, frequency);
2882 REDUCE_FLS(sec, count);
2885 if (count_fls + sec_fls > 64) {
2886 divisor = nsec * frequency;
2888 while (count_fls + sec_fls > 64) {
2889 REDUCE_FLS(count, sec);
2893 dividend = count * sec;
2895 dividend = count * sec;
2897 while (nsec_fls + frequency_fls > 64) {
2898 REDUCE_FLS(nsec, frequency);
2902 divisor = nsec * frequency;
2908 return div64_u64(dividend, divisor);
2911 static DEFINE_PER_CPU(int, perf_throttled_count);
2912 static DEFINE_PER_CPU(u64, perf_throttled_seq);
2914 static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count, bool disable)
2916 struct hw_perf_event *hwc = &event->hw;
2917 s64 period, sample_period;
2920 period = perf_calculate_period(event, nsec, count);
2922 delta = (s64)(period - hwc->sample_period);
2923 delta = (delta + 7) / 8; /* low pass filter */
2925 sample_period = hwc->sample_period + delta;
2930 hwc->sample_period = sample_period;
2932 if (local64_read(&hwc->period_left) > 8*sample_period) {
2934 event->pmu->stop(event, PERF_EF_UPDATE);
2936 local64_set(&hwc->period_left, 0);
2939 event->pmu->start(event, PERF_EF_RELOAD);
2944 * combine freq adjustment with unthrottling to avoid two passes over the
2945 * events. At the same time, make sure, having freq events does not change
2946 * the rate of unthrottling as that would introduce bias.
2948 static void perf_adjust_freq_unthr_context(struct perf_event_context *ctx,
2951 struct perf_event *event;
2952 struct hw_perf_event *hwc;
2953 u64 now, period = TICK_NSEC;
2957 * only need to iterate over all events iff:
2958 * - context have events in frequency mode (needs freq adjust)
2959 * - there are events to unthrottle on this cpu
2961 if (!(ctx->nr_freq || needs_unthr))
2964 raw_spin_lock(&ctx->lock);
2965 perf_pmu_disable(ctx->pmu);
2967 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
2968 if (event->state != PERF_EVENT_STATE_ACTIVE)
2971 if (!event_filter_match(event))
2974 perf_pmu_disable(event->pmu);
2978 if (hwc->interrupts == MAX_INTERRUPTS) {
2979 hwc->interrupts = 0;
2980 perf_log_throttle(event, 1);
2981 event->pmu->start(event, 0);
2984 if (!event->attr.freq || !event->attr.sample_freq)
2988 * stop the event and update event->count
2990 event->pmu->stop(event, PERF_EF_UPDATE);
2992 now = local64_read(&event->count);
2993 delta = now - hwc->freq_count_stamp;
2994 hwc->freq_count_stamp = now;
2998 * reload only if value has changed
2999 * we have stopped the event so tell that
3000 * to perf_adjust_period() to avoid stopping it
3004 perf_adjust_period(event, period, delta, false);
3006 event->pmu->start(event, delta > 0 ? PERF_EF_RELOAD : 0);
3008 perf_pmu_enable(event->pmu);
3011 perf_pmu_enable(ctx->pmu);
3012 raw_spin_unlock(&ctx->lock);
3016 * Round-robin a context's events:
3018 static void rotate_ctx(struct perf_event_context *ctx)
3021 * Rotate the first entry last of non-pinned groups. Rotation might be
3022 * disabled by the inheritance code.
3024 if (!ctx->rotate_disable)
3025 list_rotate_left(&ctx->flexible_groups);
3028 static int perf_rotate_context(struct perf_cpu_context *cpuctx)
3030 struct perf_event_context *ctx = NULL;
3033 if (cpuctx->ctx.nr_events) {
3034 if (cpuctx->ctx.nr_events != cpuctx->ctx.nr_active)
3038 ctx = cpuctx->task_ctx;
3039 if (ctx && ctx->nr_events) {
3040 if (ctx->nr_events != ctx->nr_active)
3047 perf_ctx_lock(cpuctx, cpuctx->task_ctx);
3048 perf_pmu_disable(cpuctx->ctx.pmu);
3050 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
3052 ctx_sched_out(ctx, cpuctx, EVENT_FLEXIBLE);
3054 rotate_ctx(&cpuctx->ctx);
3058 perf_event_sched_in(cpuctx, ctx, current);
3060 perf_pmu_enable(cpuctx->ctx.pmu);
3061 perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
3067 #ifdef CONFIG_NO_HZ_FULL
3068 bool perf_event_can_stop_tick(void)
3070 if (atomic_read(&nr_freq_events) ||
3071 __this_cpu_read(perf_throttled_count))
3078 void perf_event_task_tick(void)
3080 struct list_head *head = this_cpu_ptr(&active_ctx_list);
3081 struct perf_event_context *ctx, *tmp;
3084 WARN_ON(!irqs_disabled());
3086 __this_cpu_inc(perf_throttled_seq);
3087 throttled = __this_cpu_xchg(perf_throttled_count, 0);
3089 list_for_each_entry_safe(ctx, tmp, head, active_ctx_list)
3090 perf_adjust_freq_unthr_context(ctx, throttled);
3093 static int event_enable_on_exec(struct perf_event *event,
3094 struct perf_event_context *ctx)
3096 if (!event->attr.enable_on_exec)
3099 event->attr.enable_on_exec = 0;
3100 if (event->state >= PERF_EVENT_STATE_INACTIVE)
3103 __perf_event_mark_enabled(event);
3109 * Enable all of a task's events that have been marked enable-on-exec.
3110 * This expects task == current.
3112 static void perf_event_enable_on_exec(struct perf_event_context *ctx)
3114 struct perf_event_context *clone_ctx = NULL;
3115 struct perf_event *event;
3116 unsigned long flags;
3120 local_irq_save(flags);
3121 if (!ctx || !ctx->nr_events)
3125 * We must ctxsw out cgroup events to avoid conflict
3126 * when invoking perf_task_event_sched_in() later on
3127 * in this function. Otherwise we end up trying to
3128 * ctxswin cgroup events which are already scheduled
3131 perf_cgroup_sched_out(current, NULL);
3133 raw_spin_lock(&ctx->lock);
3134 task_ctx_sched_out(ctx);
3136 list_for_each_entry(event, &ctx->event_list, event_entry) {
3137 ret = event_enable_on_exec(event, ctx);
3143 * Unclone this context if we enabled any event.
3146 clone_ctx = unclone_ctx(ctx);
3148 raw_spin_unlock(&ctx->lock);
3151 * Also calls ctxswin for cgroup events, if any:
3153 perf_event_context_sched_in(ctx, ctx->task);
3155 local_irq_restore(flags);
3161 void perf_event_exec(void)
3163 struct perf_event_context *ctx;
3167 for_each_task_context_nr(ctxn) {
3168 ctx = current->perf_event_ctxp[ctxn];
3172 perf_event_enable_on_exec(ctx);
3178 * Cross CPU call to read the hardware event
3180 static void __perf_event_read(void *info)
3182 struct perf_event *event = info;
3183 struct perf_event_context *ctx = event->ctx;
3184 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
3187 * If this is a task context, we need to check whether it is
3188 * the current task context of this cpu. If not it has been
3189 * scheduled out before the smp call arrived. In that case
3190 * event->count would have been updated to a recent sample
3191 * when the event was scheduled out.
3193 if (ctx->task && cpuctx->task_ctx != ctx)
3196 raw_spin_lock(&ctx->lock);
3197 if (ctx->is_active) {
3198 update_context_time(ctx);
3199 update_cgrp_time_from_event(event);
3201 update_event_times(event);
3202 if (event->state == PERF_EVENT_STATE_ACTIVE)
3203 event->pmu->read(event);
3204 raw_spin_unlock(&ctx->lock);
3207 static inline u64 perf_event_count(struct perf_event *event)
3209 if (event->pmu->count)
3210 return event->pmu->count(event);
3212 return __perf_event_count(event);
3216 * NMI-safe method to read a local event, that is an event that
3218 * - either for the current task, or for this CPU
3219 * - does not have inherit set, for inherited task events
3220 * will not be local and we cannot read them atomically
3221 * - must not have a pmu::count method
3223 u64 perf_event_read_local(struct perf_event *event)
3225 unsigned long flags;
3229 * Disabling interrupts avoids all counter scheduling (context
3230 * switches, timer based rotation and IPIs).
3232 local_irq_save(flags);
3234 /* If this is a per-task event, it must be for current */
3235 WARN_ON_ONCE((event->attach_state & PERF_ATTACH_TASK) &&
3236 event->hw.target != current);
3238 /* If this is a per-CPU event, it must be for this CPU */
3239 WARN_ON_ONCE(!(event->attach_state & PERF_ATTACH_TASK) &&
3240 event->cpu != smp_processor_id());
3243 * It must not be an event with inherit set, we cannot read
3244 * all child counters from atomic context.
3246 WARN_ON_ONCE(event->attr.inherit);
3249 * It must not have a pmu::count method, those are not
3252 WARN_ON_ONCE(event->pmu->count);
3255 * If the event is currently on this CPU, its either a per-task event,
3256 * or local to this CPU. Furthermore it means its ACTIVE (otherwise
3259 if (event->oncpu == smp_processor_id())
3260 event->pmu->read(event);
3262 val = local64_read(&event->count);
3263 local_irq_restore(flags);
3268 static u64 perf_event_read(struct perf_event *event)
3271 * If event is enabled and currently active on a CPU, update the
3272 * value in the event structure:
3274 if (event->state == PERF_EVENT_STATE_ACTIVE) {
3275 smp_call_function_single(event->oncpu,
3276 __perf_event_read, event, 1);
3277 } else if (event->state == PERF_EVENT_STATE_INACTIVE) {
3278 struct perf_event_context *ctx = event->ctx;
3279 unsigned long flags;
3281 raw_spin_lock_irqsave(&ctx->lock, flags);
3283 * may read while context is not active
3284 * (e.g., thread is blocked), in that case
3285 * we cannot update context time
3287 if (ctx->is_active) {
3288 update_context_time(ctx);
3289 update_cgrp_time_from_event(event);
3291 update_event_times(event);
3292 raw_spin_unlock_irqrestore(&ctx->lock, flags);
3295 return perf_event_count(event);
3299 * Initialize the perf_event context in a task_struct:
3301 static void __perf_event_init_context(struct perf_event_context *ctx)
3303 raw_spin_lock_init(&ctx->lock);
3304 mutex_init(&ctx->mutex);
3305 INIT_LIST_HEAD(&ctx->active_ctx_list);
3306 INIT_LIST_HEAD(&ctx->pinned_groups);
3307 INIT_LIST_HEAD(&ctx->flexible_groups);
3308 INIT_LIST_HEAD(&ctx->event_list);
3309 atomic_set(&ctx->refcount, 1);
3310 INIT_DELAYED_WORK(&ctx->orphans_remove, orphans_remove_work);
3313 static struct perf_event_context *
3314 alloc_perf_context(struct pmu *pmu, struct task_struct *task)
3316 struct perf_event_context *ctx;
3318 ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL);
3322 __perf_event_init_context(ctx);
3325 get_task_struct(task);
3332 static struct task_struct *
3333 find_lively_task_by_vpid(pid_t vpid)
3335 struct task_struct *task;
3342 task = find_task_by_vpid(vpid);
3344 get_task_struct(task);
3348 return ERR_PTR(-ESRCH);
3350 /* Reuse ptrace permission checks for now. */
3352 if (!ptrace_may_access(task, PTRACE_MODE_READ))
3357 put_task_struct(task);
3358 return ERR_PTR(err);
3363 * Returns a matching context with refcount and pincount.
3365 static struct perf_event_context *
3366 find_get_context(struct pmu *pmu, struct task_struct *task,
3367 struct perf_event *event)
3369 struct perf_event_context *ctx, *clone_ctx = NULL;
3370 struct perf_cpu_context *cpuctx;
3371 void *task_ctx_data = NULL;
3372 unsigned long flags;
3374 int cpu = event->cpu;
3377 /* Must be root to operate on a CPU event: */
3378 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
3379 return ERR_PTR(-EACCES);
3382 * We could be clever and allow to attach a event to an
3383 * offline CPU and activate it when the CPU comes up, but
3386 if (!cpu_online(cpu))
3387 return ERR_PTR(-ENODEV);
3389 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
3398 ctxn = pmu->task_ctx_nr;
3402 if (event->attach_state & PERF_ATTACH_TASK_DATA) {
3403 task_ctx_data = kzalloc(pmu->task_ctx_size, GFP_KERNEL);
3404 if (!task_ctx_data) {
3411 ctx = perf_lock_task_context(task, ctxn, &flags);
3413 clone_ctx = unclone_ctx(ctx);
3416 if (task_ctx_data && !ctx->task_ctx_data) {
3417 ctx->task_ctx_data = task_ctx_data;
3418 task_ctx_data = NULL;
3420 raw_spin_unlock_irqrestore(&ctx->lock, flags);
3425 ctx = alloc_perf_context(pmu, task);
3430 if (task_ctx_data) {
3431 ctx->task_ctx_data = task_ctx_data;
3432 task_ctx_data = NULL;
3436 mutex_lock(&task->perf_event_mutex);
3438 * If it has already passed perf_event_exit_task().
3439 * we must see PF_EXITING, it takes this mutex too.
3441 if (task->flags & PF_EXITING)
3443 else if (task->perf_event_ctxp[ctxn])
3448 rcu_assign_pointer(task->perf_event_ctxp[ctxn], ctx);
3450 mutex_unlock(&task->perf_event_mutex);
3452 if (unlikely(err)) {
3461 kfree(task_ctx_data);
3465 kfree(task_ctx_data);
3466 return ERR_PTR(err);
3469 static void perf_event_free_filter(struct perf_event *event);
3470 static void perf_event_free_bpf_prog(struct perf_event *event);
3472 static void free_event_rcu(struct rcu_head *head)
3474 struct perf_event *event;
3476 event = container_of(head, struct perf_event, rcu_head);
3478 put_pid_ns(event->ns);
3479 perf_event_free_filter(event);
3483 static void ring_buffer_attach(struct perf_event *event,
3484 struct ring_buffer *rb);
3486 static void unaccount_event_cpu(struct perf_event *event, int cpu)
3491 if (is_cgroup_event(event))
3492 atomic_dec(&per_cpu(perf_cgroup_events, cpu));
3495 static void unaccount_event(struct perf_event *event)
3500 if (event->attach_state & PERF_ATTACH_TASK)
3501 static_key_slow_dec_deferred(&perf_sched_events);
3502 if (event->attr.mmap || event->attr.mmap_data)
3503 atomic_dec(&nr_mmap_events);
3504 if (event->attr.comm)
3505 atomic_dec(&nr_comm_events);
3506 if (event->attr.task)
3507 atomic_dec(&nr_task_events);
3508 if (event->attr.freq)
3509 atomic_dec(&nr_freq_events);
3510 if (is_cgroup_event(event))
3511 static_key_slow_dec_deferred(&perf_sched_events);
3512 if (has_branch_stack(event))
3513 static_key_slow_dec_deferred(&perf_sched_events);
3515 unaccount_event_cpu(event, event->cpu);
3519 * The following implement mutual exclusion of events on "exclusive" pmus
3520 * (PERF_PMU_CAP_EXCLUSIVE). Such pmus can only have one event scheduled
3521 * at a time, so we disallow creating events that might conflict, namely:
3523 * 1) cpu-wide events in the presence of per-task events,
3524 * 2) per-task events in the presence of cpu-wide events,
3525 * 3) two matching events on the same context.
3527 * The former two cases are handled in the allocation path (perf_event_alloc(),
3528 * __free_event()), the latter -- before the first perf_install_in_context().
3530 static int exclusive_event_init(struct perf_event *event)
3532 struct pmu *pmu = event->pmu;
3534 if (!(pmu->capabilities & PERF_PMU_CAP_EXCLUSIVE))
3538 * Prevent co-existence of per-task and cpu-wide events on the
3539 * same exclusive pmu.
3541 * Negative pmu::exclusive_cnt means there are cpu-wide
3542 * events on this "exclusive" pmu, positive means there are
3545 * Since this is called in perf_event_alloc() path, event::ctx
3546 * doesn't exist yet; it is, however, safe to use PERF_ATTACH_TASK
3547 * to mean "per-task event", because unlike other attach states it
3548 * never gets cleared.
3550 if (event->attach_state & PERF_ATTACH_TASK) {
3551 if (!atomic_inc_unless_negative(&pmu->exclusive_cnt))
3554 if (!atomic_dec_unless_positive(&pmu->exclusive_cnt))
3561 static void exclusive_event_destroy(struct perf_event *event)
3563 struct pmu *pmu = event->pmu;
3565 if (!(pmu->capabilities & PERF_PMU_CAP_EXCLUSIVE))
3568 /* see comment in exclusive_event_init() */
3569 if (event->attach_state & PERF_ATTACH_TASK)
3570 atomic_dec(&pmu->exclusive_cnt);
3572 atomic_inc(&pmu->exclusive_cnt);
3575 static bool exclusive_event_match(struct perf_event *e1, struct perf_event *e2)
3577 if ((e1->pmu->capabilities & PERF_PMU_CAP_EXCLUSIVE) &&
3578 (e1->cpu == e2->cpu ||
3585 /* Called under the same ctx::mutex as perf_install_in_context() */
3586 static bool exclusive_event_installable(struct perf_event *event,
3587 struct perf_event_context *ctx)
3589 struct perf_event *iter_event;
3590 struct pmu *pmu = event->pmu;
3592 if (!(pmu->capabilities & PERF_PMU_CAP_EXCLUSIVE))
3595 list_for_each_entry(iter_event, &ctx->event_list, event_entry) {
3596 if (exclusive_event_match(iter_event, event))
3603 static void __free_event(struct perf_event *event)
3605 if (!event->parent) {
3606 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN)
3607 put_callchain_buffers();
3610 perf_event_free_bpf_prog(event);
3613 event->destroy(event);
3616 put_ctx(event->ctx);
3619 exclusive_event_destroy(event);
3620 module_put(event->pmu->module);
3623 call_rcu(&event->rcu_head, free_event_rcu);
3626 static void _free_event(struct perf_event *event)
3628 irq_work_sync(&event->pending);
3630 unaccount_event(event);
3634 * Can happen when we close an event with re-directed output.
3636 * Since we have a 0 refcount, perf_mmap_close() will skip
3637 * over us; possibly making our ring_buffer_put() the last.
3639 mutex_lock(&event->mmap_mutex);
3640 ring_buffer_attach(event, NULL);
3641 mutex_unlock(&event->mmap_mutex);
3644 if (is_cgroup_event(event))
3645 perf_detach_cgroup(event);
3647 __free_event(event);
3651 * Used to free events which have a known refcount of 1, such as in error paths
3652 * where the event isn't exposed yet and inherited events.
3654 static void free_event(struct perf_event *event)
3656 if (WARN(atomic_long_cmpxchg(&event->refcount, 1, 0) != 1,
3657 "unexpected event refcount: %ld; ptr=%p\n",
3658 atomic_long_read(&event->refcount), event)) {
3659 /* leak to avoid use-after-free */
3667 * Remove user event from the owner task.
3669 static void perf_remove_from_owner(struct perf_event *event)
3671 struct task_struct *owner;
3674 owner = ACCESS_ONCE(event->owner);
3676 * Matches the smp_wmb() in perf_event_exit_task(). If we observe
3677 * !owner it means the list deletion is complete and we can indeed
3678 * free this event, otherwise we need to serialize on
3679 * owner->perf_event_mutex.
3681 smp_read_barrier_depends();
3684 * Since delayed_put_task_struct() also drops the last
3685 * task reference we can safely take a new reference
3686 * while holding the rcu_read_lock().
3688 get_task_struct(owner);
3694 * If we're here through perf_event_exit_task() we're already
3695 * holding ctx->mutex which would be an inversion wrt. the
3696 * normal lock order.
3698 * However we can safely take this lock because its the child
3701 mutex_lock_nested(&owner->perf_event_mutex, SINGLE_DEPTH_NESTING);
3704 * We have to re-check the event->owner field, if it is cleared
3705 * we raced with perf_event_exit_task(), acquiring the mutex
3706 * ensured they're done, and we can proceed with freeing the
3710 list_del_init(&event->owner_entry);
3711 mutex_unlock(&owner->perf_event_mutex);
3712 put_task_struct(owner);
3716 static void put_event(struct perf_event *event)
3718 struct perf_event_context *ctx;
3720 if (!atomic_long_dec_and_test(&event->refcount))
3723 if (!is_kernel_event(event))
3724 perf_remove_from_owner(event);
3727 * There are two ways this annotation is useful:
3729 * 1) there is a lock recursion from perf_event_exit_task
3730 * see the comment there.
3732 * 2) there is a lock-inversion with mmap_sem through
3733 * perf_event_read_group(), which takes faults while
3734 * holding ctx->mutex, however this is called after
3735 * the last filedesc died, so there is no possibility
3736 * to trigger the AB-BA case.
3738 ctx = perf_event_ctx_lock_nested(event, SINGLE_DEPTH_NESTING);
3739 WARN_ON_ONCE(ctx->parent_ctx);
3740 perf_remove_from_context(event, true);
3741 perf_event_ctx_unlock(event, ctx);
3746 int perf_event_release_kernel(struct perf_event *event)
3751 EXPORT_SYMBOL_GPL(perf_event_release_kernel);
3754 * Called when the last reference to the file is gone.
3756 static int perf_release(struct inode *inode, struct file *file)
3758 put_event(file->private_data);
3763 * Remove all orphanes events from the context.
3765 static void orphans_remove_work(struct work_struct *work)
3767 struct perf_event_context *ctx;
3768 struct perf_event *event, *tmp;
3770 ctx = container_of(work, struct perf_event_context,
3771 orphans_remove.work);
3773 mutex_lock(&ctx->mutex);
3774 list_for_each_entry_safe(event, tmp, &ctx->event_list, event_entry) {
3775 struct perf_event *parent_event = event->parent;
3777 if (!is_orphaned_child(event))
3780 perf_remove_from_context(event, true);
3782 mutex_lock(&parent_event->child_mutex);
3783 list_del_init(&event->child_list);
3784 mutex_unlock(&parent_event->child_mutex);
3787 put_event(parent_event);
3790 raw_spin_lock_irq(&ctx->lock);
3791 ctx->orphans_remove_sched = false;
3792 raw_spin_unlock_irq(&ctx->lock);
3793 mutex_unlock(&ctx->mutex);
3798 u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running)
3800 struct perf_event *child;
3806 mutex_lock(&event->child_mutex);
3807 total += perf_event_read(event);
3808 *enabled += event->total_time_enabled +
3809 atomic64_read(&event->child_total_time_enabled);
3810 *running += event->total_time_running +
3811 atomic64_read(&event->child_total_time_running);
3813 list_for_each_entry(child, &event->child_list, child_list) {
3814 total += perf_event_read(child);
3815 *enabled += child->total_time_enabled;
3816 *running += child->total_time_running;
3818 mutex_unlock(&event->child_mutex);
3822 EXPORT_SYMBOL_GPL(perf_event_read_value);
3824 static int perf_event_read_group(struct perf_event *event,
3825 u64 read_format, char __user *buf)
3827 struct perf_event *leader = event->group_leader, *sub;
3828 struct perf_event_context *ctx = leader->ctx;
3829 int n = 0, size = 0, ret;
3830 u64 count, enabled, running;
3833 lockdep_assert_held(&ctx->mutex);
3835 count = perf_event_read_value(leader, &enabled, &running);
3837 values[n++] = 1 + leader->nr_siblings;
3838 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3839 values[n++] = enabled;
3840 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3841 values[n++] = running;
3842 values[n++] = count;
3843 if (read_format & PERF_FORMAT_ID)
3844 values[n++] = primary_event_id(leader);
3846 size = n * sizeof(u64);
3848 if (copy_to_user(buf, values, size))
3853 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
3856 values[n++] = perf_event_read_value(sub, &enabled, &running);
3857 if (read_format & PERF_FORMAT_ID)
3858 values[n++] = primary_event_id(sub);
3860 size = n * sizeof(u64);
3862 if (copy_to_user(buf + ret, values, size)) {
3872 static int perf_event_read_one(struct perf_event *event,
3873 u64 read_format, char __user *buf)
3875 u64 enabled, running;
3879 values[n++] = perf_event_read_value(event, &enabled, &running);
3880 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3881 values[n++] = enabled;
3882 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3883 values[n++] = running;
3884 if (read_format & PERF_FORMAT_ID)
3885 values[n++] = primary_event_id(event);
3887 if (copy_to_user(buf, values, n * sizeof(u64)))
3890 return n * sizeof(u64);
3893 static bool is_event_hup(struct perf_event *event)
3897 if (event->state != PERF_EVENT_STATE_EXIT)
3900 mutex_lock(&event->child_mutex);
3901 no_children = list_empty(&event->child_list);
3902 mutex_unlock(&event->child_mutex);
3907 * Read the performance event - simple non blocking version for now
3910 perf_read_hw(struct perf_event *event, char __user *buf, size_t count)
3912 u64 read_format = event->attr.read_format;
3916 * Return end-of-file for a read on a event that is in
3917 * error state (i.e. because it was pinned but it couldn't be
3918 * scheduled on to the CPU at some point).
3920 if (event->state == PERF_EVENT_STATE_ERROR)
3923 if (count < event->read_size)
3926 WARN_ON_ONCE(event->ctx->parent_ctx);
3927 if (read_format & PERF_FORMAT_GROUP)
3928 ret = perf_event_read_group(event, read_format, buf);
3930 ret = perf_event_read_one(event, read_format, buf);
3936 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
3938 struct perf_event *event = file->private_data;
3939 struct perf_event_context *ctx;
3942 ctx = perf_event_ctx_lock(event);
3943 ret = perf_read_hw(event, buf, count);
3944 perf_event_ctx_unlock(event, ctx);
3949 static unsigned int perf_poll(struct file *file, poll_table *wait)
3951 struct perf_event *event = file->private_data;
3952 struct ring_buffer *rb;
3953 unsigned int events = POLLHUP;
3955 poll_wait(file, &event->waitq, wait);
3957 if (is_event_hup(event))
3961 * Pin the event->rb by taking event->mmap_mutex; otherwise
3962 * perf_event_set_output() can swizzle our rb and make us miss wakeups.
3964 mutex_lock(&event->mmap_mutex);
3967 events = atomic_xchg(&rb->poll, 0);
3968 mutex_unlock(&event->mmap_mutex);
3972 static void _perf_event_reset(struct perf_event *event)
3974 (void)perf_event_read(event);
3975 local64_set(&event->count, 0);
3976 perf_event_update_userpage(event);
3980 * Holding the top-level event's child_mutex means that any
3981 * descendant process that has inherited this event will block
3982 * in sync_child_event if it goes to exit, thus satisfying the
3983 * task existence requirements of perf_event_enable/disable.
3985 static void perf_event_for_each_child(struct perf_event *event,
3986 void (*func)(struct perf_event *))
3988 struct perf_event *child;
3990 WARN_ON_ONCE(event->ctx->parent_ctx);
3992 mutex_lock(&event->child_mutex);
3994 list_for_each_entry(child, &event->child_list, child_list)
3996 mutex_unlock(&event->child_mutex);
3999 static void perf_event_for_each(struct perf_event *event,
4000 void (*func)(struct perf_event *))
4002 struct perf_event_context *ctx = event->ctx;
4003 struct perf_event *sibling;
4005 lockdep_assert_held(&ctx->mutex);
4007 event = event->group_leader;
4009 perf_event_for_each_child(event, func);
4010 list_for_each_entry(sibling, &event->sibling_list, group_entry)
4011 perf_event_for_each_child(sibling, func);
4014 static int perf_event_period(struct perf_event *event, u64 __user *arg)
4016 struct perf_event_context *ctx = event->ctx;
4017 int ret = 0, active;
4020 if (!is_sampling_event(event))
4023 if (copy_from_user(&value, arg, sizeof(value)))
4029 raw_spin_lock_irq(&ctx->lock);
4030 if (event->attr.freq) {
4031 if (value > sysctl_perf_event_sample_rate) {
4036 event->attr.sample_freq = value;
4038 event->attr.sample_period = value;
4039 event->hw.sample_period = value;
4042 active = (event->state == PERF_EVENT_STATE_ACTIVE);
4044 perf_pmu_disable(ctx->pmu);
4045 event->pmu->stop(event, PERF_EF_UPDATE);
4048 local64_set(&event->hw.period_left, 0);
4051 event->pmu->start(event, PERF_EF_RELOAD);
4052 perf_pmu_enable(ctx->pmu);
4056 raw_spin_unlock_irq(&ctx->lock);
4061 static const struct file_operations perf_fops;
4063 static inline int perf_fget_light(int fd, struct fd *p)
4065 struct fd f = fdget(fd);
4069 if (f.file->f_op != &perf_fops) {
4077 static int perf_event_set_output(struct perf_event *event,
4078 struct perf_event *output_event);
4079 static int perf_event_set_filter(struct perf_event *event, void __user *arg);
4080 static int perf_event_set_bpf_prog(struct perf_event *event, u32 prog_fd);
4082 static long _perf_ioctl(struct perf_event *event, unsigned int cmd, unsigned long arg)
4084 void (*func)(struct perf_event *);
4088 case PERF_EVENT_IOC_ENABLE:
4089 func = _perf_event_enable;
4091 case PERF_EVENT_IOC_DISABLE:
4092 func = _perf_event_disable;
4094 case PERF_EVENT_IOC_RESET:
4095 func = _perf_event_reset;
4098 case PERF_EVENT_IOC_REFRESH:
4099 return _perf_event_refresh(event, arg);
4101 case PERF_EVENT_IOC_PERIOD:
4102 return perf_event_period(event, (u64 __user *)arg);
4104 case PERF_EVENT_IOC_ID:
4106 u64 id = primary_event_id(event);
4108 if (copy_to_user((void __user *)arg, &id, sizeof(id)))
4113 case PERF_EVENT_IOC_SET_OUTPUT:
4117 struct perf_event *output_event;
4119 ret = perf_fget_light(arg, &output);
4122 output_event = output.file->private_data;
4123 ret = perf_event_set_output(event, output_event);
4126 ret = perf_event_set_output(event, NULL);
4131 case PERF_EVENT_IOC_SET_FILTER:
4132 return perf_event_set_filter(event, (void __user *)arg);
4134 case PERF_EVENT_IOC_SET_BPF:
4135 return perf_event_set_bpf_prog(event, arg);
4141 if (flags & PERF_IOC_FLAG_GROUP)
4142 perf_event_for_each(event, func);
4144 perf_event_for_each_child(event, func);
4149 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
4151 struct perf_event *event = file->private_data;
4152 struct perf_event_context *ctx;
4155 ctx = perf_event_ctx_lock(event);
4156 ret = _perf_ioctl(event, cmd, arg);
4157 perf_event_ctx_unlock(event, ctx);
4162 #ifdef CONFIG_COMPAT
4163 static long perf_compat_ioctl(struct file *file, unsigned int cmd,
4166 switch (_IOC_NR(cmd)) {
4167 case _IOC_NR(PERF_EVENT_IOC_SET_FILTER):
4168 case _IOC_NR(PERF_EVENT_IOC_ID):
4169 /* Fix up pointer size (usually 4 -> 8 in 32-on-64-bit case */
4170 if (_IOC_SIZE(cmd) == sizeof(compat_uptr_t)) {
4171 cmd &= ~IOCSIZE_MASK;
4172 cmd |= sizeof(void *) << IOCSIZE_SHIFT;
4176 return perf_ioctl(file, cmd, arg);
4179 # define perf_compat_ioctl NULL
4182 int perf_event_task_enable(void)
4184 struct perf_event_context *ctx;
4185 struct perf_event *event;
4187 mutex_lock(¤t->perf_event_mutex);
4188 list_for_each_entry(event, ¤t->perf_event_list, owner_entry) {
4189 ctx = perf_event_ctx_lock(event);
4190 perf_event_for_each_child(event, _perf_event_enable);
4191 perf_event_ctx_unlock(event, ctx);
4193 mutex_unlock(¤t->perf_event_mutex);
4198 int perf_event_task_disable(void)
4200 struct perf_event_context *ctx;
4201 struct perf_event *event;
4203 mutex_lock(¤t->perf_event_mutex);
4204 list_for_each_entry(event, ¤t->perf_event_list, owner_entry) {
4205 ctx = perf_event_ctx_lock(event);
4206 perf_event_for_each_child(event, _perf_event_disable);
4207 perf_event_ctx_unlock(event, ctx);
4209 mutex_unlock(¤t->perf_event_mutex);
4214 static int perf_event_index(struct perf_event *event)
4216 if (event->hw.state & PERF_HES_STOPPED)
4219 if (event->state != PERF_EVENT_STATE_ACTIVE)
4222 return event->pmu->event_idx(event);
4225 static void calc_timer_values(struct perf_event *event,
4232 *now = perf_clock();
4233 ctx_time = event->shadow_ctx_time + *now;
4234 *enabled = ctx_time - event->tstamp_enabled;
4235 *running = ctx_time - event->tstamp_running;
4238 static void perf_event_init_userpage(struct perf_event *event)
4240 struct perf_event_mmap_page *userpg;
4241 struct ring_buffer *rb;
4244 rb = rcu_dereference(event->rb);
4248 userpg = rb->user_page;
4250 /* Allow new userspace to detect that bit 0 is deprecated */
4251 userpg->cap_bit0_is_deprecated = 1;
4252 userpg->size = offsetof(struct perf_event_mmap_page, __reserved);
4253 userpg->data_offset = PAGE_SIZE;
4254 userpg->data_size = perf_data_size(rb);
4260 void __weak arch_perf_update_userpage(
4261 struct perf_event *event, struct perf_event_mmap_page *userpg, u64 now)
4266 * Callers need to ensure there can be no nesting of this function, otherwise
4267 * the seqlock logic goes bad. We can not serialize this because the arch
4268 * code calls this from NMI context.
4270 void perf_event_update_userpage(struct perf_event *event)
4272 struct perf_event_mmap_page *userpg;
4273 struct ring_buffer *rb;
4274 u64 enabled, running, now;
4277 rb = rcu_dereference(event->rb);
4282 * compute total_time_enabled, total_time_running
4283 * based on snapshot values taken when the event
4284 * was last scheduled in.
4286 * we cannot simply called update_context_time()
4287 * because of locking issue as we can be called in
4290 calc_timer_values(event, &now, &enabled, &running);
4292 userpg = rb->user_page;
4294 * Disable preemption so as to not let the corresponding user-space
4295 * spin too long if we get preempted.
4300 userpg->index = perf_event_index(event);
4301 userpg->offset = perf_event_count(event);
4303 userpg->offset -= local64_read(&event->hw.prev_count);
4305 userpg->time_enabled = enabled +
4306 atomic64_read(&event->child_total_time_enabled);
4308 userpg->time_running = running +
4309 atomic64_read(&event->child_total_time_running);
4311 arch_perf_update_userpage(event, userpg, now);
4320 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
4322 struct perf_event *event = vma->vm_file->private_data;
4323 struct ring_buffer *rb;
4324 int ret = VM_FAULT_SIGBUS;
4326 if (vmf->flags & FAULT_FLAG_MKWRITE) {
4327 if (vmf->pgoff == 0)
4333 rb = rcu_dereference(event->rb);
4337 if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE))
4340 vmf->page = perf_mmap_to_page(rb, vmf->pgoff);
4344 get_page(vmf->page);
4345 vmf->page->mapping = vma->vm_file->f_mapping;
4346 vmf->page->index = vmf->pgoff;
4355 static void ring_buffer_attach(struct perf_event *event,
4356 struct ring_buffer *rb)
4358 struct ring_buffer *old_rb = NULL;
4359 unsigned long flags;
4363 * Should be impossible, we set this when removing
4364 * event->rb_entry and wait/clear when adding event->rb_entry.
4366 WARN_ON_ONCE(event->rcu_pending);
4369 spin_lock_irqsave(&old_rb->event_lock, flags);
4370 list_del_rcu(&event->rb_entry);
4371 spin_unlock_irqrestore(&old_rb->event_lock, flags);
4373 event->rcu_batches = get_state_synchronize_rcu();
4374 event->rcu_pending = 1;
4378 if (event->rcu_pending) {
4379 cond_synchronize_rcu(event->rcu_batches);
4380 event->rcu_pending = 0;
4383 spin_lock_irqsave(&rb->event_lock, flags);
4384 list_add_rcu(&event->rb_entry, &rb->event_list);
4385 spin_unlock_irqrestore(&rb->event_lock, flags);
4388 rcu_assign_pointer(event->rb, rb);
4391 ring_buffer_put(old_rb);
4393 * Since we detached before setting the new rb, so that we
4394 * could attach the new rb, we could have missed a wakeup.
4397 wake_up_all(&event->waitq);
4401 static void ring_buffer_wakeup(struct perf_event *event)
4403 struct ring_buffer *rb;
4406 rb = rcu_dereference(event->rb);
4408 list_for_each_entry_rcu(event, &rb->event_list, rb_entry)
4409 wake_up_all(&event->waitq);
4414 struct ring_buffer *ring_buffer_get(struct perf_event *event)
4416 struct ring_buffer *rb;
4419 rb = rcu_dereference(event->rb);
4421 if (!atomic_inc_not_zero(&rb->refcount))
4429 void ring_buffer_put(struct ring_buffer *rb)
4431 if (!atomic_dec_and_test(&rb->refcount))
4434 WARN_ON_ONCE(!list_empty(&rb->event_list));
4436 call_rcu(&rb->rcu_head, rb_free_rcu);
4439 static void perf_mmap_open(struct vm_area_struct *vma)
4441 struct perf_event *event = vma->vm_file->private_data;
4443 atomic_inc(&event->mmap_count);
4444 atomic_inc(&event->rb->mmap_count);
4447 atomic_inc(&event->rb->aux_mmap_count);
4449 if (event->pmu->event_mapped)
4450 event->pmu->event_mapped(event);
4454 * A buffer can be mmap()ed multiple times; either directly through the same
4455 * event, or through other events by use of perf_event_set_output().
4457 * In order to undo the VM accounting done by perf_mmap() we need to destroy
4458 * the buffer here, where we still have a VM context. This means we need
4459 * to detach all events redirecting to us.
4461 static void perf_mmap_close(struct vm_area_struct *vma)
4463 struct perf_event *event = vma->vm_file->private_data;
4465 struct ring_buffer *rb = ring_buffer_get(event);
4466 struct user_struct *mmap_user = rb->mmap_user;
4467 int mmap_locked = rb->mmap_locked;
4468 unsigned long size = perf_data_size(rb);
4470 if (event->pmu->event_unmapped)
4471 event->pmu->event_unmapped(event);
4474 * rb->aux_mmap_count will always drop before rb->mmap_count and
4475 * event->mmap_count, so it is ok to use event->mmap_mutex to
4476 * serialize with perf_mmap here.
4478 if (rb_has_aux(rb) && vma->vm_pgoff == rb->aux_pgoff &&
4479 atomic_dec_and_mutex_lock(&rb->aux_mmap_count, &event->mmap_mutex)) {
4480 atomic_long_sub(rb->aux_nr_pages, &mmap_user->locked_vm);
4481 vma->vm_mm->pinned_vm -= rb->aux_mmap_locked;
4484 mutex_unlock(&event->mmap_mutex);
4487 atomic_dec(&rb->mmap_count);
4489 if (!atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex))
4492 ring_buffer_attach(event, NULL);
4493 mutex_unlock(&event->mmap_mutex);
4495 /* If there's still other mmap()s of this buffer, we're done. */
4496 if (atomic_read(&rb->mmap_count))
4500 * No other mmap()s, detach from all other events that might redirect
4501 * into the now unreachable buffer. Somewhat complicated by the
4502 * fact that rb::event_lock otherwise nests inside mmap_mutex.
4506 list_for_each_entry_rcu(event, &rb->event_list, rb_entry) {
4507 if (!atomic_long_inc_not_zero(&event->refcount)) {
4509 * This event is en-route to free_event() which will
4510 * detach it and remove it from the list.
4516 mutex_lock(&event->mmap_mutex);
4518 * Check we didn't race with perf_event_set_output() which can
4519 * swizzle the rb from under us while we were waiting to
4520 * acquire mmap_mutex.
4522 * If we find a different rb; ignore this event, a next
4523 * iteration will no longer find it on the list. We have to
4524 * still restart the iteration to make sure we're not now
4525 * iterating the wrong list.
4527 if (event->rb == rb)
4528 ring_buffer_attach(event, NULL);
4530 mutex_unlock(&event->mmap_mutex);
4534 * Restart the iteration; either we're on the wrong list or
4535 * destroyed its integrity by doing a deletion.
4542 * It could be there's still a few 0-ref events on the list; they'll
4543 * get cleaned up by free_event() -- they'll also still have their
4544 * ref on the rb and will free it whenever they are done with it.
4546 * Aside from that, this buffer is 'fully' detached and unmapped,
4547 * undo the VM accounting.
4550 atomic_long_sub((size >> PAGE_SHIFT) + 1, &mmap_user->locked_vm);
4551 vma->vm_mm->pinned_vm -= mmap_locked;
4552 free_uid(mmap_user);
4555 ring_buffer_put(rb); /* could be last */
4558 static const struct vm_operations_struct perf_mmap_vmops = {
4559 .open = perf_mmap_open,
4560 .close = perf_mmap_close, /* non mergable */
4561 .fault = perf_mmap_fault,
4562 .page_mkwrite = perf_mmap_fault,
4565 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
4567 struct perf_event *event = file->private_data;
4568 unsigned long user_locked, user_lock_limit;
4569 struct user_struct *user = current_user();
4570 unsigned long locked, lock_limit;
4571 struct ring_buffer *rb = NULL;
4572 unsigned long vma_size;
4573 unsigned long nr_pages;
4574 long user_extra = 0, extra = 0;
4575 int ret = 0, flags = 0;
4578 * Don't allow mmap() of inherited per-task counters. This would
4579 * create a performance issue due to all children writing to the
4582 if (event->cpu == -1 && event->attr.inherit)
4585 if (!(vma->vm_flags & VM_SHARED))
4588 vma_size = vma->vm_end - vma->vm_start;
4590 if (vma->vm_pgoff == 0) {
4591 nr_pages = (vma_size / PAGE_SIZE) - 1;
4594 * AUX area mapping: if rb->aux_nr_pages != 0, it's already
4595 * mapped, all subsequent mappings should have the same size
4596 * and offset. Must be above the normal perf buffer.
4598 u64 aux_offset, aux_size;
4603 nr_pages = vma_size / PAGE_SIZE;
4605 mutex_lock(&event->mmap_mutex);
4612 aux_offset = ACCESS_ONCE(rb->user_page->aux_offset);
4613 aux_size = ACCESS_ONCE(rb->user_page->aux_size);
4615 if (aux_offset < perf_data_size(rb) + PAGE_SIZE)
4618 if (aux_offset != vma->vm_pgoff << PAGE_SHIFT)
4621 /* already mapped with a different offset */
4622 if (rb_has_aux(rb) && rb->aux_pgoff != vma->vm_pgoff)
4625 if (aux_size != vma_size || aux_size != nr_pages * PAGE_SIZE)
4628 /* already mapped with a different size */
4629 if (rb_has_aux(rb) && rb->aux_nr_pages != nr_pages)
4632 if (!is_power_of_2(nr_pages))
4635 if (!atomic_inc_not_zero(&rb->mmap_count))
4638 if (rb_has_aux(rb)) {
4639 atomic_inc(&rb->aux_mmap_count);
4644 atomic_set(&rb->aux_mmap_count, 1);
4645 user_extra = nr_pages;
4651 * If we have rb pages ensure they're a power-of-two number, so we
4652 * can do bitmasks instead of modulo.
4654 if (nr_pages != 0 && !is_power_of_2(nr_pages))
4657 if (vma_size != PAGE_SIZE * (1 + nr_pages))
4660 WARN_ON_ONCE(event->ctx->parent_ctx);
4662 mutex_lock(&event->mmap_mutex);
4664 if (event->rb->nr_pages != nr_pages) {
4669 if (!atomic_inc_not_zero(&event->rb->mmap_count)) {
4671 * Raced against perf_mmap_close() through
4672 * perf_event_set_output(). Try again, hope for better
4675 mutex_unlock(&event->mmap_mutex);
4682 user_extra = nr_pages + 1;
4685 user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10);
4688 * Increase the limit linearly with more CPUs:
4690 user_lock_limit *= num_online_cpus();
4692 user_locked = atomic_long_read(&user->locked_vm) + user_extra;
4694 if (user_locked > user_lock_limit)
4695 extra = user_locked - user_lock_limit;
4697 lock_limit = rlimit(RLIMIT_MEMLOCK);
4698 lock_limit >>= PAGE_SHIFT;
4699 locked = vma->vm_mm->pinned_vm + extra;
4701 if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() &&
4702 !capable(CAP_IPC_LOCK)) {
4707 WARN_ON(!rb && event->rb);
4709 if (vma->vm_flags & VM_WRITE)
4710 flags |= RING_BUFFER_WRITABLE;
4713 rb = rb_alloc(nr_pages,
4714 event->attr.watermark ? event->attr.wakeup_watermark : 0,
4722 atomic_set(&rb->mmap_count, 1);
4723 rb->mmap_user = get_current_user();
4724 rb->mmap_locked = extra;
4726 ring_buffer_attach(event, rb);
4728 perf_event_init_userpage(event);
4729 perf_event_update_userpage(event);
4731 ret = rb_alloc_aux(rb, event, vma->vm_pgoff, nr_pages,
4732 event->attr.aux_watermark, flags);
4734 rb->aux_mmap_locked = extra;
4739 atomic_long_add(user_extra, &user->locked_vm);
4740 vma->vm_mm->pinned_vm += extra;
4742 atomic_inc(&event->mmap_count);
4744 atomic_dec(&rb->mmap_count);
4747 mutex_unlock(&event->mmap_mutex);
4750 * Since pinned accounting is per vm we cannot allow fork() to copy our
4753 vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND | VM_DONTDUMP;
4754 vma->vm_ops = &perf_mmap_vmops;
4756 if (event->pmu->event_mapped)
4757 event->pmu->event_mapped(event);
4762 static int perf_fasync(int fd, struct file *filp, int on)
4764 struct inode *inode = file_inode(filp);
4765 struct perf_event *event = filp->private_data;
4768 mutex_lock(&inode->i_mutex);
4769 retval = fasync_helper(fd, filp, on, &event->fasync);
4770 mutex_unlock(&inode->i_mutex);
4778 static const struct file_operations perf_fops = {
4779 .llseek = no_llseek,
4780 .release = perf_release,
4783 .unlocked_ioctl = perf_ioctl,
4784 .compat_ioctl = perf_compat_ioctl,
4786 .fasync = perf_fasync,
4792 * If there's data, ensure we set the poll() state and publish everything
4793 * to user-space before waking everybody up.
4796 void perf_event_wakeup(struct perf_event *event)
4798 ring_buffer_wakeup(event);
4800 if (event->pending_kill) {
4801 kill_fasync(&event->fasync, SIGIO, event->pending_kill);
4802 event->pending_kill = 0;
4806 static void perf_pending_event(struct irq_work *entry)
4808 struct perf_event *event = container_of(entry,
4809 struct perf_event, pending);
4812 rctx = perf_swevent_get_recursion_context();
4814 * If we 'fail' here, that's OK, it means recursion is already disabled
4815 * and we won't recurse 'further'.
4818 if (event->pending_disable) {
4819 event->pending_disable = 0;
4820 __perf_event_disable(event);
4823 if (event->pending_wakeup) {
4824 event->pending_wakeup = 0;
4825 perf_event_wakeup(event);
4829 perf_swevent_put_recursion_context(rctx);
4833 * We assume there is only KVM supporting the callbacks.
4834 * Later on, we might change it to a list if there is
4835 * another virtualization implementation supporting the callbacks.
4837 struct perf_guest_info_callbacks *perf_guest_cbs;
4839 int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
4841 perf_guest_cbs = cbs;
4844 EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks);
4846 int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
4848 perf_guest_cbs = NULL;
4851 EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks);
4854 perf_output_sample_regs(struct perf_output_handle *handle,
4855 struct pt_regs *regs, u64 mask)
4859 for_each_set_bit(bit, (const unsigned long *) &mask,
4860 sizeof(mask) * BITS_PER_BYTE) {
4863 val = perf_reg_value(regs, bit);
4864 perf_output_put(handle, val);
4868 static void perf_sample_regs_user(struct perf_regs *regs_user,
4869 struct pt_regs *regs,
4870 struct pt_regs *regs_user_copy)
4872 if (user_mode(regs)) {
4873 regs_user->abi = perf_reg_abi(current);
4874 regs_user->regs = regs;
4875 } else if (current->mm) {
4876 perf_get_regs_user(regs_user, regs, regs_user_copy);
4878 regs_user->abi = PERF_SAMPLE_REGS_ABI_NONE;
4879 regs_user->regs = NULL;
4883 static void perf_sample_regs_intr(struct perf_regs *regs_intr,
4884 struct pt_regs *regs)
4886 regs_intr->regs = regs;
4887 regs_intr->abi = perf_reg_abi(current);
4892 * Get remaining task size from user stack pointer.
4894 * It'd be better to take stack vma map and limit this more
4895 * precisly, but there's no way to get it safely under interrupt,
4896 * so using TASK_SIZE as limit.
4898 static u64 perf_ustack_task_size(struct pt_regs *regs)
4900 unsigned long addr = perf_user_stack_pointer(regs);
4902 if (!addr || addr >= TASK_SIZE)
4905 return TASK_SIZE - addr;
4909 perf_sample_ustack_size(u16 stack_size, u16 header_size,
4910 struct pt_regs *regs)
4914 /* No regs, no stack pointer, no dump. */
4919 * Check if we fit in with the requested stack size into the:
4921 * If we don't, we limit the size to the TASK_SIZE.
4923 * - remaining sample size
4924 * If we don't, we customize the stack size to
4925 * fit in to the remaining sample size.
4928 task_size = min((u64) USHRT_MAX, perf_ustack_task_size(regs));
4929 stack_size = min(stack_size, (u16) task_size);
4931 /* Current header size plus static size and dynamic size. */
4932 header_size += 2 * sizeof(u64);
4934 /* Do we fit in with the current stack dump size? */
4935 if ((u16) (header_size + stack_size) < header_size) {
4937 * If we overflow the maximum size for the sample,
4938 * we customize the stack dump size to fit in.
4940 stack_size = USHRT_MAX - header_size - sizeof(u64);
4941 stack_size = round_up(stack_size, sizeof(u64));
4948 perf_output_sample_ustack(struct perf_output_handle *handle, u64 dump_size,
4949 struct pt_regs *regs)
4951 /* Case of a kernel thread, nothing to dump */
4954 perf_output_put(handle, size);
4963 * - the size requested by user or the best one we can fit
4964 * in to the sample max size
4966 * - user stack dump data
4968 * - the actual dumped size
4972 perf_output_put(handle, dump_size);
4975 sp = perf_user_stack_pointer(regs);
4976 rem = __output_copy_user(handle, (void *) sp, dump_size);
4977 dyn_size = dump_size - rem;
4979 perf_output_skip(handle, rem);
4982 perf_output_put(handle, dyn_size);
4986 static void __perf_event_header__init_id(struct perf_event_header *header,
4987 struct perf_sample_data *data,
4988 struct perf_event *event)
4990 u64 sample_type = event->attr.sample_type;
4992 data->type = sample_type;
4993 header->size += event->id_header_size;
4995 if (sample_type & PERF_SAMPLE_TID) {
4996 /* namespace issues */
4997 data->tid_entry.pid = perf_event_pid(event, current);
4998 data->tid_entry.tid = perf_event_tid(event, current);
5001 if (sample_type & PERF_SAMPLE_TIME)
5002 data->time = perf_event_clock(event);
5004 if (sample_type & (PERF_SAMPLE_ID | PERF_SAMPLE_IDENTIFIER))
5005 data->id = primary_event_id(event);
5007 if (sample_type & PERF_SAMPLE_STREAM_ID)
5008 data->stream_id = event->id;
5010 if (sample_type & PERF_SAMPLE_CPU) {
5011 data->cpu_entry.cpu = raw_smp_processor_id();
5012 data->cpu_entry.reserved = 0;
5016 void perf_event_header__init_id(struct perf_event_header *header,
5017 struct perf_sample_data *data,
5018 struct perf_event *event)
5020 if (event->attr.sample_id_all)
5021 __perf_event_header__init_id(header, data, event);
5024 static void __perf_event__output_id_sample(struct perf_output_handle *handle,
5025 struct perf_sample_data *data)
5027 u64 sample_type = data->type;
5029 if (sample_type & PERF_SAMPLE_TID)
5030 perf_output_put(handle, data->tid_entry);
5032 if (sample_type & PERF_SAMPLE_TIME)
5033 perf_output_put(handle, data->time);
5035 if (sample_type & PERF_SAMPLE_ID)
5036 perf_output_put(handle, data->id);
5038 if (sample_type & PERF_SAMPLE_STREAM_ID)
5039 perf_output_put(handle, data->stream_id);
5041 if (sample_type & PERF_SAMPLE_CPU)
5042 perf_output_put(handle, data->cpu_entry);
5044 if (sample_type & PERF_SAMPLE_IDENTIFIER)
5045 perf_output_put(handle, data->id);
5048 void perf_event__output_id_sample(struct perf_event *event,
5049 struct perf_output_handle *handle,
5050 struct perf_sample_data *sample)
5052 if (event->attr.sample_id_all)
5053 __perf_event__output_id_sample(handle, sample);
5056 static void perf_output_read_one(struct perf_output_handle *handle,
5057 struct perf_event *event,
5058 u64 enabled, u64 running)
5060 u64 read_format = event->attr.read_format;
5064 values[n++] = perf_event_count(event);
5065 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
5066 values[n++] = enabled +
5067 atomic64_read(&event->child_total_time_enabled);
5069 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
5070 values[n++] = running +
5071 atomic64_read(&event->child_total_time_running);
5073 if (read_format & PERF_FORMAT_ID)
5074 values[n++] = primary_event_id(event);
5076 __output_copy(handle, values, n * sizeof(u64));
5080 * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
5082 static void perf_output_read_group(struct perf_output_handle *handle,
5083 struct perf_event *event,
5084 u64 enabled, u64 running)
5086 struct perf_event *leader = event->group_leader, *sub;
5087 u64 read_format = event->attr.read_format;
5091 values[n++] = 1 + leader->nr_siblings;
5093 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
5094 values[n++] = enabled;
5096 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
5097 values[n++] = running;
5099 if (leader != event)
5100 leader->pmu->read(leader);
5102 values[n++] = perf_event_count(leader);
5103 if (read_format & PERF_FORMAT_ID)
5104 values[n++] = primary_event_id(leader);
5106 __output_copy(handle, values, n * sizeof(u64));
5108 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
5111 if ((sub != event) &&
5112 (sub->state == PERF_EVENT_STATE_ACTIVE))
5113 sub->pmu->read(sub);
5115 values[n++] = perf_event_count(sub);
5116 if (read_format & PERF_FORMAT_ID)
5117 values[n++] = primary_event_id(sub);
5119 __output_copy(handle, values, n * sizeof(u64));
5123 #define PERF_FORMAT_TOTAL_TIMES (PERF_FORMAT_TOTAL_TIME_ENABLED|\
5124 PERF_FORMAT_TOTAL_TIME_RUNNING)
5126 static void perf_output_read(struct perf_output_handle *handle,
5127 struct perf_event *event)
5129 u64 enabled = 0, running = 0, now;
5130 u64 read_format = event->attr.read_format;
5133 * compute total_time_enabled, total_time_running
5134 * based on snapshot values taken when the event
5135 * was last scheduled in.
5137 * we cannot simply called update_context_time()
5138 * because of locking issue as we are called in
5141 if (read_format & PERF_FORMAT_TOTAL_TIMES)
5142 calc_timer_values(event, &now, &enabled, &running);
5144 if (event->attr.read_format & PERF_FORMAT_GROUP)
5145 perf_output_read_group(handle, event, enabled, running);
5147 perf_output_read_one(handle, event, enabled, running);
5150 void perf_output_sample(struct perf_output_handle *handle,
5151 struct perf_event_header *header,
5152 struct perf_sample_data *data,
5153 struct perf_event *event)
5155 u64 sample_type = data->type;
5157 perf_output_put(handle, *header);
5159 if (sample_type & PERF_SAMPLE_IDENTIFIER)
5160 perf_output_put(handle, data->id);
5162 if (sample_type & PERF_SAMPLE_IP)
5163 perf_output_put(handle, data->ip);
5165 if (sample_type & PERF_SAMPLE_TID)
5166 perf_output_put(handle, data->tid_entry);
5168 if (sample_type & PERF_SAMPLE_TIME)
5169 perf_output_put(handle, data->time);
5171 if (sample_type & PERF_SAMPLE_ADDR)
5172 perf_output_put(handle, data->addr);
5174 if (sample_type & PERF_SAMPLE_ID)
5175 perf_output_put(handle, data->id);
5177 if (sample_type & PERF_SAMPLE_STREAM_ID)
5178 perf_output_put(handle, data->stream_id);
5180 if (sample_type & PERF_SAMPLE_CPU)
5181 perf_output_put(handle, data->cpu_entry);
5183 if (sample_type & PERF_SAMPLE_PERIOD)
5184 perf_output_put(handle, data->period);
5186 if (sample_type & PERF_SAMPLE_READ)
5187 perf_output_read(handle, event);
5189 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
5190 if (data->callchain) {
5193 if (data->callchain)
5194 size += data->callchain->nr;
5196 size *= sizeof(u64);
5198 __output_copy(handle, data->callchain, size);
5201 perf_output_put(handle, nr);
5205 if (sample_type & PERF_SAMPLE_RAW) {
5207 perf_output_put(handle, data->raw->size);
5208 __output_copy(handle, data->raw->data,
5215 .size = sizeof(u32),
5218 perf_output_put(handle, raw);
5222 if (sample_type & PERF_SAMPLE_BRANCH_STACK) {
5223 if (data->br_stack) {
5226 size = data->br_stack->nr
5227 * sizeof(struct perf_branch_entry);
5229 perf_output_put(handle, data->br_stack->nr);
5230 perf_output_copy(handle, data->br_stack->entries, size);
5233 * we always store at least the value of nr
5236 perf_output_put(handle, nr);
5240 if (sample_type & PERF_SAMPLE_REGS_USER) {
5241 u64 abi = data->regs_user.abi;
5244 * If there are no regs to dump, notice it through
5245 * first u64 being zero (PERF_SAMPLE_REGS_ABI_NONE).
5247 perf_output_put(handle, abi);
5250 u64 mask = event->attr.sample_regs_user;
5251 perf_output_sample_regs(handle,
5252 data->regs_user.regs,
5257 if (sample_type & PERF_SAMPLE_STACK_USER) {
5258 perf_output_sample_ustack(handle,
5259 data->stack_user_size,
5260 data->regs_user.regs);
5263 if (sample_type & PERF_SAMPLE_WEIGHT)
5264 perf_output_put(handle, data->weight);
5266 if (sample_type & PERF_SAMPLE_DATA_SRC)
5267 perf_output_put(handle, data->data_src.val);
5269 if (sample_type & PERF_SAMPLE_TRANSACTION)
5270 perf_output_put(handle, data->txn);
5272 if (sample_type & PERF_SAMPLE_REGS_INTR) {
5273 u64 abi = data->regs_intr.abi;
5275 * If there are no regs to dump, notice it through
5276 * first u64 being zero (PERF_SAMPLE_REGS_ABI_NONE).
5278 perf_output_put(handle, abi);
5281 u64 mask = event->attr.sample_regs_intr;
5283 perf_output_sample_regs(handle,
5284 data->regs_intr.regs,
5289 if (!event->attr.watermark) {
5290 int wakeup_events = event->attr.wakeup_events;
5292 if (wakeup_events) {
5293 struct ring_buffer *rb = handle->rb;
5294 int events = local_inc_return(&rb->events);
5296 if (events >= wakeup_events) {
5297 local_sub(wakeup_events, &rb->events);
5298 local_inc(&rb->wakeup);
5304 void perf_prepare_sample(struct perf_event_header *header,
5305 struct perf_sample_data *data,
5306 struct perf_event *event,
5307 struct pt_regs *regs)
5309 u64 sample_type = event->attr.sample_type;
5311 header->type = PERF_RECORD_SAMPLE;
5312 header->size = sizeof(*header) + event->header_size;
5315 header->misc |= perf_misc_flags(regs);
5317 __perf_event_header__init_id(header, data, event);
5319 if (sample_type & PERF_SAMPLE_IP)
5320 data->ip = perf_instruction_pointer(regs);
5322 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
5325 data->callchain = perf_callchain(event, regs);
5327 if (data->callchain)
5328 size += data->callchain->nr;
5330 header->size += size * sizeof(u64);
5333 if (sample_type & PERF_SAMPLE_RAW) {
5334 int size = sizeof(u32);
5337 size += data->raw->size;
5339 size += sizeof(u32);
5341 WARN_ON_ONCE(size & (sizeof(u64)-1));
5342 header->size += size;
5345 if (sample_type & PERF_SAMPLE_BRANCH_STACK) {
5346 int size = sizeof(u64); /* nr */
5347 if (data->br_stack) {
5348 size += data->br_stack->nr
5349 * sizeof(struct perf_branch_entry);
5351 header->size += size;
5354 if (sample_type & (PERF_SAMPLE_REGS_USER | PERF_SAMPLE_STACK_USER))
5355 perf_sample_regs_user(&data->regs_user, regs,
5356 &data->regs_user_copy);
5358 if (sample_type & PERF_SAMPLE_REGS_USER) {
5359 /* regs dump ABI info */
5360 int size = sizeof(u64);
5362 if (data->regs_user.regs) {
5363 u64 mask = event->attr.sample_regs_user;
5364 size += hweight64(mask) * sizeof(u64);
5367 header->size += size;
5370 if (sample_type & PERF_SAMPLE_STACK_USER) {
5372 * Either we need PERF_SAMPLE_STACK_USER bit to be allways
5373 * processed as the last one or have additional check added
5374 * in case new sample type is added, because we could eat
5375 * up the rest of the sample size.
5377 u16 stack_size = event->attr.sample_stack_user;
5378 u16 size = sizeof(u64);
5380 stack_size = perf_sample_ustack_size(stack_size, header->size,
5381 data->regs_user.regs);
5384 * If there is something to dump, add space for the dump
5385 * itself and for the field that tells the dynamic size,
5386 * which is how many have been actually dumped.
5389 size += sizeof(u64) + stack_size;
5391 data->stack_user_size = stack_size;
5392 header->size += size;
5395 if (sample_type & PERF_SAMPLE_REGS_INTR) {
5396 /* regs dump ABI info */
5397 int size = sizeof(u64);
5399 perf_sample_regs_intr(&data->regs_intr, regs);
5401 if (data->regs_intr.regs) {
5402 u64 mask = event->attr.sample_regs_intr;
5404 size += hweight64(mask) * sizeof(u64);
5407 header->size += size;
5411 void perf_event_output(struct perf_event *event,
5412 struct perf_sample_data *data,
5413 struct pt_regs *regs)
5415 struct perf_output_handle handle;
5416 struct perf_event_header header;
5418 /* protect the callchain buffers */
5421 perf_prepare_sample(&header, data, event, regs);
5423 if (perf_output_begin(&handle, event, header.size))
5426 perf_output_sample(&handle, &header, data, event);
5428 perf_output_end(&handle);
5438 struct perf_read_event {
5439 struct perf_event_header header;
5446 perf_event_read_event(struct perf_event *event,
5447 struct task_struct *task)
5449 struct perf_output_handle handle;
5450 struct perf_sample_data sample;
5451 struct perf_read_event read_event = {
5453 .type = PERF_RECORD_READ,
5455 .size = sizeof(read_event) + event->read_size,
5457 .pid = perf_event_pid(event, task),
5458 .tid = perf_event_tid(event, task),
5462 perf_event_header__init_id(&read_event.header, &sample, event);
5463 ret = perf_output_begin(&handle, event, read_event.header.size);
5467 perf_output_put(&handle, read_event);
5468 perf_output_read(&handle, event);
5469 perf_event__output_id_sample(event, &handle, &sample);
5471 perf_output_end(&handle);
5474 typedef void (perf_event_aux_output_cb)(struct perf_event *event, void *data);
5477 perf_event_aux_ctx(struct perf_event_context *ctx,
5478 perf_event_aux_output_cb output,
5481 struct perf_event *event;
5483 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
5484 if (event->state < PERF_EVENT_STATE_INACTIVE)
5486 if (!event_filter_match(event))
5488 output(event, data);
5493 perf_event_aux(perf_event_aux_output_cb output, void *data,
5494 struct perf_event_context *task_ctx)
5496 struct perf_cpu_context *cpuctx;
5497 struct perf_event_context *ctx;
5502 list_for_each_entry_rcu(pmu, &pmus, entry) {
5503 cpuctx = get_cpu_ptr(pmu->pmu_cpu_context);
5504 if (cpuctx->unique_pmu != pmu)
5506 perf_event_aux_ctx(&cpuctx->ctx, output, data);
5509 ctxn = pmu->task_ctx_nr;
5512 ctx = rcu_dereference(current->perf_event_ctxp[ctxn]);
5514 perf_event_aux_ctx(ctx, output, data);
5516 put_cpu_ptr(pmu->pmu_cpu_context);
5521 perf_event_aux_ctx(task_ctx, output, data);
5528 * task tracking -- fork/exit
5530 * enabled by: attr.comm | attr.mmap | attr.mmap2 | attr.mmap_data | attr.task
5533 struct perf_task_event {
5534 struct task_struct *task;
5535 struct perf_event_context *task_ctx;
5538 struct perf_event_header header;
5548 static int perf_event_task_match(struct perf_event *event)
5550 return event->attr.comm || event->attr.mmap ||
5551 event->attr.mmap2 || event->attr.mmap_data ||
5555 static void perf_event_task_output(struct perf_event *event,
5558 struct perf_task_event *task_event = data;
5559 struct perf_output_handle handle;
5560 struct perf_sample_data sample;
5561 struct task_struct *task = task_event->task;
5562 int ret, size = task_event->event_id.header.size;
5564 if (!perf_event_task_match(event))
5567 perf_event_header__init_id(&task_event->event_id.header, &sample, event);
5569 ret = perf_output_begin(&handle, event,
5570 task_event->event_id.header.size);
5574 task_event->event_id.pid = perf_event_pid(event, task);
5575 task_event->event_id.ppid = perf_event_pid(event, current);
5577 task_event->event_id.tid = perf_event_tid(event, task);
5578 task_event->event_id.ptid = perf_event_tid(event, current);
5580 task_event->event_id.time = perf_event_clock(event);
5582 perf_output_put(&handle, task_event->event_id);
5584 perf_event__output_id_sample(event, &handle, &sample);
5586 perf_output_end(&handle);
5588 task_event->event_id.header.size = size;
5591 static void perf_event_task(struct task_struct *task,
5592 struct perf_event_context *task_ctx,
5595 struct perf_task_event task_event;
5597 if (!atomic_read(&nr_comm_events) &&
5598 !atomic_read(&nr_mmap_events) &&
5599 !atomic_read(&nr_task_events))
5602 task_event = (struct perf_task_event){
5604 .task_ctx = task_ctx,
5607 .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT,
5609 .size = sizeof(task_event.event_id),
5619 perf_event_aux(perf_event_task_output,
5624 void perf_event_fork(struct task_struct *task)
5626 perf_event_task(task, NULL, 1);
5633 struct perf_comm_event {
5634 struct task_struct *task;
5639 struct perf_event_header header;
5646 static int perf_event_comm_match(struct perf_event *event)
5648 return event->attr.comm;
5651 static void perf_event_comm_output(struct perf_event *event,
5654 struct perf_comm_event *comm_event = data;
5655 struct perf_output_handle handle;
5656 struct perf_sample_data sample;
5657 int size = comm_event->event_id.header.size;
5660 if (!perf_event_comm_match(event))
5663 perf_event_header__init_id(&comm_event->event_id.header, &sample, event);
5664 ret = perf_output_begin(&handle, event,
5665 comm_event->event_id.header.size);
5670 comm_event->event_id.pid = perf_event_pid(event, comm_event->task);
5671 comm_event->event_id.tid = perf_event_tid(event, comm_event->task);
5673 perf_output_put(&handle, comm_event->event_id);
5674 __output_copy(&handle, comm_event->comm,
5675 comm_event->comm_size);
5677 perf_event__output_id_sample(event, &handle, &sample);
5679 perf_output_end(&handle);
5681 comm_event->event_id.header.size = size;
5684 static void perf_event_comm_event(struct perf_comm_event *comm_event)
5686 char comm[TASK_COMM_LEN];
5689 memset(comm, 0, sizeof(comm));
5690 strlcpy(comm, comm_event->task->comm, sizeof(comm));
5691 size = ALIGN(strlen(comm)+1, sizeof(u64));
5693 comm_event->comm = comm;
5694 comm_event->comm_size = size;
5696 comm_event->event_id.header.size = sizeof(comm_event->event_id) + size;
5698 perf_event_aux(perf_event_comm_output,
5703 void perf_event_comm(struct task_struct *task, bool exec)
5705 struct perf_comm_event comm_event;
5707 if (!atomic_read(&nr_comm_events))
5710 comm_event = (struct perf_comm_event){
5716 .type = PERF_RECORD_COMM,
5717 .misc = exec ? PERF_RECORD_MISC_COMM_EXEC : 0,
5725 perf_event_comm_event(&comm_event);
5732 struct perf_mmap_event {
5733 struct vm_area_struct *vma;
5735 const char *file_name;
5743 struct perf_event_header header;
5753 static int perf_event_mmap_match(struct perf_event *event,
5756 struct perf_mmap_event *mmap_event = data;
5757 struct vm_area_struct *vma = mmap_event->vma;
5758 int executable = vma->vm_flags & VM_EXEC;
5760 return (!executable && event->attr.mmap_data) ||
5761 (executable && (event->attr.mmap || event->attr.mmap2));
5764 static void perf_event_mmap_output(struct perf_event *event,
5767 struct perf_mmap_event *mmap_event = data;
5768 struct perf_output_handle handle;
5769 struct perf_sample_data sample;
5770 int size = mmap_event->event_id.header.size;
5773 if (!perf_event_mmap_match(event, data))
5776 if (event->attr.mmap2) {
5777 mmap_event->event_id.header.type = PERF_RECORD_MMAP2;
5778 mmap_event->event_id.header.size += sizeof(mmap_event->maj);
5779 mmap_event->event_id.header.size += sizeof(mmap_event->min);
5780 mmap_event->event_id.header.size += sizeof(mmap_event->ino);
5781 mmap_event->event_id.header.size += sizeof(mmap_event->ino_generation);
5782 mmap_event->event_id.header.size += sizeof(mmap_event->prot);
5783 mmap_event->event_id.header.size += sizeof(mmap_event->flags);
5786 perf_event_header__init_id(&mmap_event->event_id.header, &sample, event);
5787 ret = perf_output_begin(&handle, event,
5788 mmap_event->event_id.header.size);
5792 mmap_event->event_id.pid = perf_event_pid(event, current);
5793 mmap_event->event_id.tid = perf_event_tid(event, current);
5795 perf_output_put(&handle, mmap_event->event_id);
5797 if (event->attr.mmap2) {
5798 perf_output_put(&handle, mmap_event->maj);
5799 perf_output_put(&handle, mmap_event->min);
5800 perf_output_put(&handle, mmap_event->ino);
5801 perf_output_put(&handle, mmap_event->ino_generation);
5802 perf_output_put(&handle, mmap_event->prot);
5803 perf_output_put(&handle, mmap_event->flags);
5806 __output_copy(&handle, mmap_event->file_name,
5807 mmap_event->file_size);
5809 perf_event__output_id_sample(event, &handle, &sample);
5811 perf_output_end(&handle);
5813 mmap_event->event_id.header.size = size;
5816 static void perf_event_mmap_event(struct perf_mmap_event *mmap_event)
5818 struct vm_area_struct *vma = mmap_event->vma;
5819 struct file *file = vma->vm_file;
5820 int maj = 0, min = 0;
5821 u64 ino = 0, gen = 0;
5822 u32 prot = 0, flags = 0;
5829 struct inode *inode;
5832 buf = kmalloc(PATH_MAX, GFP_KERNEL);
5838 * d_path() works from the end of the rb backwards, so we
5839 * need to add enough zero bytes after the string to handle
5840 * the 64bit alignment we do later.
5842 name = file_path(file, buf, PATH_MAX - sizeof(u64));
5847 inode = file_inode(vma->vm_file);
5848 dev = inode->i_sb->s_dev;
5850 gen = inode->i_generation;
5854 if (vma->vm_flags & VM_READ)
5856 if (vma->vm_flags & VM_WRITE)
5858 if (vma->vm_flags & VM_EXEC)
5861 if (vma->vm_flags & VM_MAYSHARE)
5864 flags = MAP_PRIVATE;
5866 if (vma->vm_flags & VM_DENYWRITE)
5867 flags |= MAP_DENYWRITE;
5868 if (vma->vm_flags & VM_MAYEXEC)
5869 flags |= MAP_EXECUTABLE;
5870 if (vma->vm_flags & VM_LOCKED)
5871 flags |= MAP_LOCKED;
5872 if (vma->vm_flags & VM_HUGETLB)
5873 flags |= MAP_HUGETLB;
5877 if (vma->vm_ops && vma->vm_ops->name) {
5878 name = (char *) vma->vm_ops->name(vma);
5883 name = (char *)arch_vma_name(vma);
5887 if (vma->vm_start <= vma->vm_mm->start_brk &&
5888 vma->vm_end >= vma->vm_mm->brk) {
5892 if (vma->vm_start <= vma->vm_mm->start_stack &&
5893 vma->vm_end >= vma->vm_mm->start_stack) {
5903 strlcpy(tmp, name, sizeof(tmp));
5907 * Since our buffer works in 8 byte units we need to align our string
5908 * size to a multiple of 8. However, we must guarantee the tail end is
5909 * zero'd out to avoid leaking random bits to userspace.
5911 size = strlen(name)+1;
5912 while (!IS_ALIGNED(size, sizeof(u64)))
5913 name[size++] = '\0';
5915 mmap_event->file_name = name;
5916 mmap_event->file_size = size;
5917 mmap_event->maj = maj;
5918 mmap_event->min = min;
5919 mmap_event->ino = ino;
5920 mmap_event->ino_generation = gen;
5921 mmap_event->prot = prot;
5922 mmap_event->flags = flags;
5924 if (!(vma->vm_flags & VM_EXEC))
5925 mmap_event->event_id.header.misc |= PERF_RECORD_MISC_MMAP_DATA;
5927 mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size;
5929 perf_event_aux(perf_event_mmap_output,
5936 void perf_event_mmap(struct vm_area_struct *vma)
5938 struct perf_mmap_event mmap_event;
5940 if (!atomic_read(&nr_mmap_events))
5943 mmap_event = (struct perf_mmap_event){
5949 .type = PERF_RECORD_MMAP,
5950 .misc = PERF_RECORD_MISC_USER,
5955 .start = vma->vm_start,
5956 .len = vma->vm_end - vma->vm_start,
5957 .pgoff = (u64)vma->vm_pgoff << PAGE_SHIFT,
5959 /* .maj (attr_mmap2 only) */
5960 /* .min (attr_mmap2 only) */
5961 /* .ino (attr_mmap2 only) */
5962 /* .ino_generation (attr_mmap2 only) */
5963 /* .prot (attr_mmap2 only) */
5964 /* .flags (attr_mmap2 only) */
5967 perf_event_mmap_event(&mmap_event);
5970 void perf_event_aux_event(struct perf_event *event, unsigned long head,
5971 unsigned long size, u64 flags)
5973 struct perf_output_handle handle;
5974 struct perf_sample_data sample;
5975 struct perf_aux_event {
5976 struct perf_event_header header;
5982 .type = PERF_RECORD_AUX,
5984 .size = sizeof(rec),
5992 perf_event_header__init_id(&rec.header, &sample, event);
5993 ret = perf_output_begin(&handle, event, rec.header.size);
5998 perf_output_put(&handle, rec);
5999 perf_event__output_id_sample(event, &handle, &sample);
6001 perf_output_end(&handle);
6005 * Lost/dropped samples logging
6007 void perf_log_lost_samples(struct perf_event *event, u64 lost)
6009 struct perf_output_handle handle;
6010 struct perf_sample_data sample;
6014 struct perf_event_header header;
6016 } lost_samples_event = {
6018 .type = PERF_RECORD_LOST_SAMPLES,
6020 .size = sizeof(lost_samples_event),
6025 perf_event_header__init_id(&lost_samples_event.header, &sample, event);
6027 ret = perf_output_begin(&handle, event,
6028 lost_samples_event.header.size);
6032 perf_output_put(&handle, lost_samples_event);
6033 perf_event__output_id_sample(event, &handle, &sample);
6034 perf_output_end(&handle);
6038 * IRQ throttle logging
6041 static void perf_log_throttle(struct perf_event *event, int enable)
6043 struct perf_output_handle handle;
6044 struct perf_sample_data sample;
6048 struct perf_event_header header;
6052 } throttle_event = {
6054 .type = PERF_RECORD_THROTTLE,
6056 .size = sizeof(throttle_event),
6058 .time = perf_event_clock(event),
6059 .id = primary_event_id(event),
6060 .stream_id = event->id,
6064 throttle_event.header.type = PERF_RECORD_UNTHROTTLE;
6066 perf_event_header__init_id(&throttle_event.header, &sample, event);
6068 ret = perf_output_begin(&handle, event,
6069 throttle_event.header.size);
6073 perf_output_put(&handle, throttle_event);
6074 perf_event__output_id_sample(event, &handle, &sample);
6075 perf_output_end(&handle);
6078 static void perf_log_itrace_start(struct perf_event *event)
6080 struct perf_output_handle handle;
6081 struct perf_sample_data sample;
6082 struct perf_aux_event {
6083 struct perf_event_header header;
6090 event = event->parent;
6092 if (!(event->pmu->capabilities & PERF_PMU_CAP_ITRACE) ||
6093 event->hw.itrace_started)
6096 event->hw.itrace_started = 1;
6098 rec.header.type = PERF_RECORD_ITRACE_START;
6099 rec.header.misc = 0;
6100 rec.header.size = sizeof(rec);
6101 rec.pid = perf_event_pid(event, current);
6102 rec.tid = perf_event_tid(event, current);
6104 perf_event_header__init_id(&rec.header, &sample, event);
6105 ret = perf_output_begin(&handle, event, rec.header.size);
6110 perf_output_put(&handle, rec);
6111 perf_event__output_id_sample(event, &handle, &sample);
6113 perf_output_end(&handle);
6117 * Generic event overflow handling, sampling.
6120 static int __perf_event_overflow(struct perf_event *event,
6121 int throttle, struct perf_sample_data *data,
6122 struct pt_regs *regs)
6124 int events = atomic_read(&event->event_limit);
6125 struct hw_perf_event *hwc = &event->hw;
6130 * Non-sampling counters might still use the PMI to fold short
6131 * hardware counters, ignore those.
6133 if (unlikely(!is_sampling_event(event)))
6136 seq = __this_cpu_read(perf_throttled_seq);
6137 if (seq != hwc->interrupts_seq) {
6138 hwc->interrupts_seq = seq;
6139 hwc->interrupts = 1;
6142 if (unlikely(throttle
6143 && hwc->interrupts >= max_samples_per_tick)) {
6144 __this_cpu_inc(perf_throttled_count);
6145 hwc->interrupts = MAX_INTERRUPTS;
6146 perf_log_throttle(event, 0);
6147 tick_nohz_full_kick();
6152 if (event->attr.freq) {
6153 u64 now = perf_clock();
6154 s64 delta = now - hwc->freq_time_stamp;
6156 hwc->freq_time_stamp = now;
6158 if (delta > 0 && delta < 2*TICK_NSEC)
6159 perf_adjust_period(event, delta, hwc->last_period, true);
6163 * XXX event_limit might not quite work as expected on inherited
6167 event->pending_kill = POLL_IN;
6168 if (events && atomic_dec_and_test(&event->event_limit)) {
6170 event->pending_kill = POLL_HUP;
6171 event->pending_disable = 1;
6172 irq_work_queue(&event->pending);
6175 if (event->overflow_handler)
6176 event->overflow_handler(event, data, regs);
6178 perf_event_output(event, data, regs);
6180 if (event->fasync && event->pending_kill) {
6181 event->pending_wakeup = 1;
6182 irq_work_queue(&event->pending);
6188 int perf_event_overflow(struct perf_event *event,
6189 struct perf_sample_data *data,
6190 struct pt_regs *regs)
6192 return __perf_event_overflow(event, 1, data, regs);
6196 * Generic software event infrastructure
6199 struct swevent_htable {
6200 struct swevent_hlist *swevent_hlist;
6201 struct mutex hlist_mutex;
6204 /* Recursion avoidance in each contexts */
6205 int recursion[PERF_NR_CONTEXTS];
6207 /* Keeps track of cpu being initialized/exited */
6211 static DEFINE_PER_CPU(struct swevent_htable, swevent_htable);
6214 * We directly increment event->count and keep a second value in
6215 * event->hw.period_left to count intervals. This period event
6216 * is kept in the range [-sample_period, 0] so that we can use the
6220 u64 perf_swevent_set_period(struct perf_event *event)
6222 struct hw_perf_event *hwc = &event->hw;
6223 u64 period = hwc->last_period;
6227 hwc->last_period = hwc->sample_period;
6230 old = val = local64_read(&hwc->period_left);
6234 nr = div64_u64(period + val, period);
6235 offset = nr * period;
6237 if (local64_cmpxchg(&hwc->period_left, old, val) != old)
6243 static void perf_swevent_overflow(struct perf_event *event, u64 overflow,
6244 struct perf_sample_data *data,
6245 struct pt_regs *regs)
6247 struct hw_perf_event *hwc = &event->hw;
6251 overflow = perf_swevent_set_period(event);
6253 if (hwc->interrupts == MAX_INTERRUPTS)
6256 for (; overflow; overflow--) {
6257 if (__perf_event_overflow(event, throttle,
6260 * We inhibit the overflow from happening when
6261 * hwc->interrupts == MAX_INTERRUPTS.
6269 static void perf_swevent_event(struct perf_event *event, u64 nr,
6270 struct perf_sample_data *data,
6271 struct pt_regs *regs)
6273 struct hw_perf_event *hwc = &event->hw;
6275 local64_add(nr, &event->count);
6280 if (!is_sampling_event(event))
6283 if ((event->attr.sample_type & PERF_SAMPLE_PERIOD) && !event->attr.freq) {
6285 return perf_swevent_overflow(event, 1, data, regs);
6287 data->period = event->hw.last_period;
6289 if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq)
6290 return perf_swevent_overflow(event, 1, data, regs);
6292 if (local64_add_negative(nr, &hwc->period_left))
6295 perf_swevent_overflow(event, 0, data, regs);
6298 static int perf_exclude_event(struct perf_event *event,
6299 struct pt_regs *regs)
6301 if (event->hw.state & PERF_HES_STOPPED)
6305 if (event->attr.exclude_user && user_mode(regs))
6308 if (event->attr.exclude_kernel && !user_mode(regs))
6315 static int perf_swevent_match(struct perf_event *event,
6316 enum perf_type_id type,
6318 struct perf_sample_data *data,
6319 struct pt_regs *regs)
6321 if (event->attr.type != type)
6324 if (event->attr.config != event_id)
6327 if (perf_exclude_event(event, regs))
6333 static inline u64 swevent_hash(u64 type, u32 event_id)
6335 u64 val = event_id | (type << 32);
6337 return hash_64(val, SWEVENT_HLIST_BITS);
6340 static inline struct hlist_head *
6341 __find_swevent_head(struct swevent_hlist *hlist, u64 type, u32 event_id)
6343 u64 hash = swevent_hash(type, event_id);
6345 return &hlist->heads[hash];
6348 /* For the read side: events when they trigger */
6349 static inline struct hlist_head *
6350 find_swevent_head_rcu(struct swevent_htable *swhash, u64 type, u32 event_id)
6352 struct swevent_hlist *hlist;
6354 hlist = rcu_dereference(swhash->swevent_hlist);
6358 return __find_swevent_head(hlist, type, event_id);
6361 /* For the event head insertion and removal in the hlist */
6362 static inline struct hlist_head *
6363 find_swevent_head(struct swevent_htable *swhash, struct perf_event *event)
6365 struct swevent_hlist *hlist;
6366 u32 event_id = event->attr.config;
6367 u64 type = event->attr.type;
6370 * Event scheduling is always serialized against hlist allocation
6371 * and release. Which makes the protected version suitable here.
6372 * The context lock guarantees that.
6374 hlist = rcu_dereference_protected(swhash->swevent_hlist,
6375 lockdep_is_held(&event->ctx->lock));
6379 return __find_swevent_head(hlist, type, event_id);
6382 static void do_perf_sw_event(enum perf_type_id type, u32 event_id,
6384 struct perf_sample_data *data,
6385 struct pt_regs *regs)
6387 struct swevent_htable *swhash = this_cpu_ptr(&swevent_htable);
6388 struct perf_event *event;
6389 struct hlist_head *head;
6392 head = find_swevent_head_rcu(swhash, type, event_id);
6396 hlist_for_each_entry_rcu(event, head, hlist_entry) {
6397 if (perf_swevent_match(event, type, event_id, data, regs))
6398 perf_swevent_event(event, nr, data, regs);
6404 DEFINE_PER_CPU(struct pt_regs, __perf_regs[4]);
6406 int perf_swevent_get_recursion_context(void)
6408 struct swevent_htable *swhash = this_cpu_ptr(&swevent_htable);
6410 return get_recursion_context(swhash->recursion);
6412 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context);
6414 inline void perf_swevent_put_recursion_context(int rctx)
6416 struct swevent_htable *swhash = this_cpu_ptr(&swevent_htable);
6418 put_recursion_context(swhash->recursion, rctx);
6421 void ___perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)
6423 struct perf_sample_data data;
6425 if (WARN_ON_ONCE(!regs))
6428 perf_sample_data_init(&data, addr, 0);
6429 do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, &data, regs);
6432 void __perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)
6436 preempt_disable_notrace();
6437 rctx = perf_swevent_get_recursion_context();
6438 if (unlikely(rctx < 0))
6441 ___perf_sw_event(event_id, nr, regs, addr);
6443 perf_swevent_put_recursion_context(rctx);
6445 preempt_enable_notrace();
6448 static void perf_swevent_read(struct perf_event *event)
6452 static int perf_swevent_add(struct perf_event *event, int flags)
6454 struct swevent_htable *swhash = this_cpu_ptr(&swevent_htable);
6455 struct hw_perf_event *hwc = &event->hw;
6456 struct hlist_head *head;
6458 if (is_sampling_event(event)) {
6459 hwc->last_period = hwc->sample_period;
6460 perf_swevent_set_period(event);
6463 hwc->state = !(flags & PERF_EF_START);
6465 head = find_swevent_head(swhash, event);
6468 * We can race with cpu hotplug code. Do not
6469 * WARN if the cpu just got unplugged.
6471 WARN_ON_ONCE(swhash->online);
6475 hlist_add_head_rcu(&event->hlist_entry, head);
6476 perf_event_update_userpage(event);
6481 static void perf_swevent_del(struct perf_event *event, int flags)
6483 hlist_del_rcu(&event->hlist_entry);
6486 static void perf_swevent_start(struct perf_event *event, int flags)
6488 event->hw.state = 0;
6491 static void perf_swevent_stop(struct perf_event *event, int flags)
6493 event->hw.state = PERF_HES_STOPPED;
6496 /* Deref the hlist from the update side */
6497 static inline struct swevent_hlist *
6498 swevent_hlist_deref(struct swevent_htable *swhash)
6500 return rcu_dereference_protected(swhash->swevent_hlist,
6501 lockdep_is_held(&swhash->hlist_mutex));
6504 static void swevent_hlist_release(struct swevent_htable *swhash)
6506 struct swevent_hlist *hlist = swevent_hlist_deref(swhash);
6511 RCU_INIT_POINTER(swhash->swevent_hlist, NULL);
6512 kfree_rcu(hlist, rcu_head);
6515 static void swevent_hlist_put_cpu(struct perf_event *event, int cpu)
6517 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
6519 mutex_lock(&swhash->hlist_mutex);
6521 if (!--swhash->hlist_refcount)
6522 swevent_hlist_release(swhash);
6524 mutex_unlock(&swhash->hlist_mutex);
6527 static void swevent_hlist_put(struct perf_event *event)
6531 for_each_possible_cpu(cpu)
6532 swevent_hlist_put_cpu(event, cpu);
6535 static int swevent_hlist_get_cpu(struct perf_event *event, int cpu)
6537 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
6540 mutex_lock(&swhash->hlist_mutex);
6542 if (!swevent_hlist_deref(swhash) && cpu_online(cpu)) {
6543 struct swevent_hlist *hlist;
6545 hlist = kzalloc(sizeof(*hlist), GFP_KERNEL);
6550 rcu_assign_pointer(swhash->swevent_hlist, hlist);
6552 swhash->hlist_refcount++;
6554 mutex_unlock(&swhash->hlist_mutex);
6559 static int swevent_hlist_get(struct perf_event *event)
6562 int cpu, failed_cpu;
6565 for_each_possible_cpu(cpu) {
6566 err = swevent_hlist_get_cpu(event, cpu);
6576 for_each_possible_cpu(cpu) {
6577 if (cpu == failed_cpu)
6579 swevent_hlist_put_cpu(event, cpu);
6586 struct static_key perf_swevent_enabled[PERF_COUNT_SW_MAX];
6588 static void sw_perf_event_destroy(struct perf_event *event)
6590 u64 event_id = event->attr.config;
6592 WARN_ON(event->parent);
6594 static_key_slow_dec(&perf_swevent_enabled[event_id]);
6595 swevent_hlist_put(event);
6598 static int perf_swevent_init(struct perf_event *event)
6600 u64 event_id = event->attr.config;
6602 if (event->attr.type != PERF_TYPE_SOFTWARE)
6606 * no branch sampling for software events
6608 if (has_branch_stack(event))
6612 case PERF_COUNT_SW_CPU_CLOCK:
6613 case PERF_COUNT_SW_TASK_CLOCK:
6620 if (event_id >= PERF_COUNT_SW_MAX)
6623 if (!event->parent) {
6626 err = swevent_hlist_get(event);
6630 static_key_slow_inc(&perf_swevent_enabled[event_id]);
6631 event->destroy = sw_perf_event_destroy;
6637 static struct pmu perf_swevent = {
6638 .task_ctx_nr = perf_sw_context,
6640 .capabilities = PERF_PMU_CAP_NO_NMI,
6642 .event_init = perf_swevent_init,
6643 .add = perf_swevent_add,
6644 .del = perf_swevent_del,
6645 .start = perf_swevent_start,
6646 .stop = perf_swevent_stop,
6647 .read = perf_swevent_read,
6650 #ifdef CONFIG_EVENT_TRACING
6652 static int perf_tp_filter_match(struct perf_event *event,
6653 struct perf_sample_data *data)
6655 void *record = data->raw->data;
6657 if (likely(!event->filter) || filter_match_preds(event->filter, record))
6662 static int perf_tp_event_match(struct perf_event *event,
6663 struct perf_sample_data *data,
6664 struct pt_regs *regs)
6666 if (event->hw.state & PERF_HES_STOPPED)
6669 * All tracepoints are from kernel-space.
6671 if (event->attr.exclude_kernel)
6674 if (!perf_tp_filter_match(event, data))
6680 void perf_tp_event(u64 addr, u64 count, void *record, int entry_size,
6681 struct pt_regs *regs, struct hlist_head *head, int rctx,
6682 struct task_struct *task)
6684 struct perf_sample_data data;
6685 struct perf_event *event;
6687 struct perf_raw_record raw = {
6692 perf_sample_data_init(&data, addr, 0);
6695 hlist_for_each_entry_rcu(event, head, hlist_entry) {
6696 if (perf_tp_event_match(event, &data, regs))
6697 perf_swevent_event(event, count, &data, regs);
6701 * If we got specified a target task, also iterate its context and
6702 * deliver this event there too.
6704 if (task && task != current) {
6705 struct perf_event_context *ctx;
6706 struct trace_entry *entry = record;
6709 ctx = rcu_dereference(task->perf_event_ctxp[perf_sw_context]);
6713 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
6714 if (event->attr.type != PERF_TYPE_TRACEPOINT)
6716 if (event->attr.config != entry->type)
6718 if (perf_tp_event_match(event, &data, regs))
6719 perf_swevent_event(event, count, &data, regs);
6725 perf_swevent_put_recursion_context(rctx);
6727 EXPORT_SYMBOL_GPL(perf_tp_event);
6729 static void tp_perf_event_destroy(struct perf_event *event)
6731 perf_trace_destroy(event);
6734 static int perf_tp_event_init(struct perf_event *event)
6738 if (event->attr.type != PERF_TYPE_TRACEPOINT)
6742 * no branch sampling for tracepoint events
6744 if (has_branch_stack(event))
6747 err = perf_trace_init(event);
6751 event->destroy = tp_perf_event_destroy;
6756 static struct pmu perf_tracepoint = {
6757 .task_ctx_nr = perf_sw_context,
6759 .event_init = perf_tp_event_init,
6760 .add = perf_trace_add,
6761 .del = perf_trace_del,
6762 .start = perf_swevent_start,
6763 .stop = perf_swevent_stop,
6764 .read = perf_swevent_read,
6767 static inline void perf_tp_register(void)
6769 perf_pmu_register(&perf_tracepoint, "tracepoint", PERF_TYPE_TRACEPOINT);
6772 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
6777 if (event->attr.type != PERF_TYPE_TRACEPOINT)
6780 filter_str = strndup_user(arg, PAGE_SIZE);
6781 if (IS_ERR(filter_str))
6782 return PTR_ERR(filter_str);
6784 ret = ftrace_profile_set_filter(event, event->attr.config, filter_str);
6790 static void perf_event_free_filter(struct perf_event *event)
6792 ftrace_profile_free_filter(event);
6795 static int perf_event_set_bpf_prog(struct perf_event *event, u32 prog_fd)
6797 struct bpf_prog *prog;
6799 if (event->attr.type != PERF_TYPE_TRACEPOINT)
6802 if (event->tp_event->prog)
6805 if (!(event->tp_event->flags & TRACE_EVENT_FL_KPROBE))
6806 /* bpf programs can only be attached to kprobes */
6809 prog = bpf_prog_get(prog_fd);
6811 return PTR_ERR(prog);
6813 if (prog->type != BPF_PROG_TYPE_KPROBE) {
6814 /* valid fd, but invalid bpf program type */
6819 event->tp_event->prog = prog;
6824 static void perf_event_free_bpf_prog(struct perf_event *event)
6826 struct bpf_prog *prog;
6828 if (!event->tp_event)
6831 prog = event->tp_event->prog;
6833 event->tp_event->prog = NULL;
6840 static inline void perf_tp_register(void)
6844 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
6849 static void perf_event_free_filter(struct perf_event *event)
6853 static int perf_event_set_bpf_prog(struct perf_event *event, u32 prog_fd)
6858 static void perf_event_free_bpf_prog(struct perf_event *event)
6861 #endif /* CONFIG_EVENT_TRACING */
6863 #ifdef CONFIG_HAVE_HW_BREAKPOINT
6864 void perf_bp_event(struct perf_event *bp, void *data)
6866 struct perf_sample_data sample;
6867 struct pt_regs *regs = data;
6869 perf_sample_data_init(&sample, bp->attr.bp_addr, 0);
6871 if (!bp->hw.state && !perf_exclude_event(bp, regs))
6872 perf_swevent_event(bp, 1, &sample, regs);
6877 * hrtimer based swevent callback
6880 static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer)
6882 enum hrtimer_restart ret = HRTIMER_RESTART;
6883 struct perf_sample_data data;
6884 struct pt_regs *regs;
6885 struct perf_event *event;
6888 event = container_of(hrtimer, struct perf_event, hw.hrtimer);
6890 if (event->state != PERF_EVENT_STATE_ACTIVE)
6891 return HRTIMER_NORESTART;
6893 event->pmu->read(event);
6895 perf_sample_data_init(&data, 0, event->hw.last_period);
6896 regs = get_irq_regs();
6898 if (regs && !perf_exclude_event(event, regs)) {
6899 if (!(event->attr.exclude_idle && is_idle_task(current)))
6900 if (__perf_event_overflow(event, 1, &data, regs))
6901 ret = HRTIMER_NORESTART;
6904 period = max_t(u64, 10000, event->hw.sample_period);
6905 hrtimer_forward_now(hrtimer, ns_to_ktime(period));
6910 static void perf_swevent_start_hrtimer(struct perf_event *event)
6912 struct hw_perf_event *hwc = &event->hw;
6915 if (!is_sampling_event(event))
6918 period = local64_read(&hwc->period_left);
6923 local64_set(&hwc->period_left, 0);
6925 period = max_t(u64, 10000, hwc->sample_period);
6927 hrtimer_start(&hwc->hrtimer, ns_to_ktime(period),
6928 HRTIMER_MODE_REL_PINNED);
6931 static void perf_swevent_cancel_hrtimer(struct perf_event *event)
6933 struct hw_perf_event *hwc = &event->hw;
6935 if (is_sampling_event(event)) {
6936 ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer);
6937 local64_set(&hwc->period_left, ktime_to_ns(remaining));
6939 hrtimer_cancel(&hwc->hrtimer);
6943 static void perf_swevent_init_hrtimer(struct perf_event *event)
6945 struct hw_perf_event *hwc = &event->hw;
6947 if (!is_sampling_event(event))
6950 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
6951 hwc->hrtimer.function = perf_swevent_hrtimer;
6954 * Since hrtimers have a fixed rate, we can do a static freq->period
6955 * mapping and avoid the whole period adjust feedback stuff.
6957 if (event->attr.freq) {
6958 long freq = event->attr.sample_freq;
6960 event->attr.sample_period = NSEC_PER_SEC / freq;
6961 hwc->sample_period = event->attr.sample_period;
6962 local64_set(&hwc->period_left, hwc->sample_period);
6963 hwc->last_period = hwc->sample_period;
6964 event->attr.freq = 0;
6969 * Software event: cpu wall time clock
6972 static void cpu_clock_event_update(struct perf_event *event)
6977 now = local_clock();
6978 prev = local64_xchg(&event->hw.prev_count, now);
6979 local64_add(now - prev, &event->count);
6982 static void cpu_clock_event_start(struct perf_event *event, int flags)
6984 local64_set(&event->hw.prev_count, local_clock());
6985 perf_swevent_start_hrtimer(event);
6988 static void cpu_clock_event_stop(struct perf_event *event, int flags)
6990 perf_swevent_cancel_hrtimer(event);
6991 cpu_clock_event_update(event);
6994 static int cpu_clock_event_add(struct perf_event *event, int flags)
6996 if (flags & PERF_EF_START)
6997 cpu_clock_event_start(event, flags);
6998 perf_event_update_userpage(event);
7003 static void cpu_clock_event_del(struct perf_event *event, int flags)
7005 cpu_clock_event_stop(event, flags);
7008 static void cpu_clock_event_read(struct perf_event *event)
7010 cpu_clock_event_update(event);
7013 static int cpu_clock_event_init(struct perf_event *event)
7015 if (event->attr.type != PERF_TYPE_SOFTWARE)
7018 if (event->attr.config != PERF_COUNT_SW_CPU_CLOCK)
7022 * no branch sampling for software events
7024 if (has_branch_stack(event))
7027 perf_swevent_init_hrtimer(event);
7032 static struct pmu perf_cpu_clock = {
7033 .task_ctx_nr = perf_sw_context,
7035 .capabilities = PERF_PMU_CAP_NO_NMI,
7037 .event_init = cpu_clock_event_init,
7038 .add = cpu_clock_event_add,
7039 .del = cpu_clock_event_del,
7040 .start = cpu_clock_event_start,
7041 .stop = cpu_clock_event_stop,
7042 .read = cpu_clock_event_read,
7046 * Software event: task time clock
7049 static void task_clock_event_update(struct perf_event *event, u64 now)
7054 prev = local64_xchg(&event->hw.prev_count, now);
7056 local64_add(delta, &event->count);
7059 static void task_clock_event_start(struct perf_event *event, int flags)
7061 local64_set(&event->hw.prev_count, event->ctx->time);
7062 perf_swevent_start_hrtimer(event);
7065 static void task_clock_event_stop(struct perf_event *event, int flags)
7067 perf_swevent_cancel_hrtimer(event);
7068 task_clock_event_update(event, event->ctx->time);
7071 static int task_clock_event_add(struct perf_event *event, int flags)
7073 if (flags & PERF_EF_START)
7074 task_clock_event_start(event, flags);
7075 perf_event_update_userpage(event);
7080 static void task_clock_event_del(struct perf_event *event, int flags)
7082 task_clock_event_stop(event, PERF_EF_UPDATE);
7085 static void task_clock_event_read(struct perf_event *event)
7087 u64 now = perf_clock();
7088 u64 delta = now - event->ctx->timestamp;
7089 u64 time = event->ctx->time + delta;
7091 task_clock_event_update(event, time);
7094 static int task_clock_event_init(struct perf_event *event)
7096 if (event->attr.type != PERF_TYPE_SOFTWARE)
7099 if (event->attr.config != PERF_COUNT_SW_TASK_CLOCK)
7103 * no branch sampling for software events
7105 if (has_branch_stack(event))
7108 perf_swevent_init_hrtimer(event);
7113 static struct pmu perf_task_clock = {
7114 .task_ctx_nr = perf_sw_context,
7116 .capabilities = PERF_PMU_CAP_NO_NMI,
7118 .event_init = task_clock_event_init,
7119 .add = task_clock_event_add,
7120 .del = task_clock_event_del,
7121 .start = task_clock_event_start,
7122 .stop = task_clock_event_stop,
7123 .read = task_clock_event_read,
7126 static void perf_pmu_nop_void(struct pmu *pmu)
7130 static int perf_pmu_nop_int(struct pmu *pmu)
7135 static void perf_pmu_start_txn(struct pmu *pmu)
7137 perf_pmu_disable(pmu);
7140 static int perf_pmu_commit_txn(struct pmu *pmu)
7142 perf_pmu_enable(pmu);
7146 static void perf_pmu_cancel_txn(struct pmu *pmu)
7148 perf_pmu_enable(pmu);
7151 static int perf_event_idx_default(struct perf_event *event)
7157 * Ensures all contexts with the same task_ctx_nr have the same
7158 * pmu_cpu_context too.
7160 static struct perf_cpu_context __percpu *find_pmu_context(int ctxn)
7167 list_for_each_entry(pmu, &pmus, entry) {
7168 if (pmu->task_ctx_nr == ctxn)
7169 return pmu->pmu_cpu_context;
7175 static void update_pmu_context(struct pmu *pmu, struct pmu *old_pmu)
7179 for_each_possible_cpu(cpu) {
7180 struct perf_cpu_context *cpuctx;
7182 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
7184 if (cpuctx->unique_pmu == old_pmu)
7185 cpuctx->unique_pmu = pmu;
7189 static void free_pmu_context(struct pmu *pmu)
7193 mutex_lock(&pmus_lock);
7195 * Like a real lame refcount.
7197 list_for_each_entry(i, &pmus, entry) {
7198 if (i->pmu_cpu_context == pmu->pmu_cpu_context) {
7199 update_pmu_context(i, pmu);
7204 free_percpu(pmu->pmu_cpu_context);
7206 mutex_unlock(&pmus_lock);
7208 static struct idr pmu_idr;
7211 type_show(struct device *dev, struct device_attribute *attr, char *page)
7213 struct pmu *pmu = dev_get_drvdata(dev);
7215 return snprintf(page, PAGE_SIZE-1, "%d\n", pmu->type);
7217 static DEVICE_ATTR_RO(type);
7220 perf_event_mux_interval_ms_show(struct device *dev,
7221 struct device_attribute *attr,
7224 struct pmu *pmu = dev_get_drvdata(dev);
7226 return snprintf(page, PAGE_SIZE-1, "%d\n", pmu->hrtimer_interval_ms);
7229 static DEFINE_MUTEX(mux_interval_mutex);
7232 perf_event_mux_interval_ms_store(struct device *dev,
7233 struct device_attribute *attr,
7234 const char *buf, size_t count)
7236 struct pmu *pmu = dev_get_drvdata(dev);
7237 int timer, cpu, ret;
7239 ret = kstrtoint(buf, 0, &timer);
7246 /* same value, noting to do */
7247 if (timer == pmu->hrtimer_interval_ms)
7250 mutex_lock(&mux_interval_mutex);
7251 pmu->hrtimer_interval_ms = timer;
7253 /* update all cpuctx for this PMU */
7255 for_each_online_cpu(cpu) {
7256 struct perf_cpu_context *cpuctx;
7257 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
7258 cpuctx->hrtimer_interval = ns_to_ktime(NSEC_PER_MSEC * timer);
7260 cpu_function_call(cpu,
7261 (remote_function_f)perf_mux_hrtimer_restart, cpuctx);
7264 mutex_unlock(&mux_interval_mutex);
7268 static DEVICE_ATTR_RW(perf_event_mux_interval_ms);
7270 static struct attribute *pmu_dev_attrs[] = {
7271 &dev_attr_type.attr,
7272 &dev_attr_perf_event_mux_interval_ms.attr,
7275 ATTRIBUTE_GROUPS(pmu_dev);
7277 static int pmu_bus_running;
7278 static struct bus_type pmu_bus = {
7279 .name = "event_source",
7280 .dev_groups = pmu_dev_groups,
7283 static void pmu_dev_release(struct device *dev)
7288 static int pmu_dev_alloc(struct pmu *pmu)
7292 pmu->dev = kzalloc(sizeof(struct device), GFP_KERNEL);
7296 pmu->dev->groups = pmu->attr_groups;
7297 device_initialize(pmu->dev);
7298 ret = dev_set_name(pmu->dev, "%s", pmu->name);
7302 dev_set_drvdata(pmu->dev, pmu);
7303 pmu->dev->bus = &pmu_bus;
7304 pmu->dev->release = pmu_dev_release;
7305 ret = device_add(pmu->dev);
7313 put_device(pmu->dev);
7317 static struct lock_class_key cpuctx_mutex;
7318 static struct lock_class_key cpuctx_lock;
7320 int perf_pmu_register(struct pmu *pmu, const char *name, int type)
7324 mutex_lock(&pmus_lock);
7326 pmu->pmu_disable_count = alloc_percpu(int);
7327 if (!pmu->pmu_disable_count)
7336 type = idr_alloc(&pmu_idr, pmu, PERF_TYPE_MAX, 0, GFP_KERNEL);
7344 if (pmu_bus_running) {
7345 ret = pmu_dev_alloc(pmu);
7351 pmu->pmu_cpu_context = find_pmu_context(pmu->task_ctx_nr);
7352 if (pmu->pmu_cpu_context)
7353 goto got_cpu_context;
7356 pmu->pmu_cpu_context = alloc_percpu(struct perf_cpu_context);
7357 if (!pmu->pmu_cpu_context)
7360 for_each_possible_cpu(cpu) {
7361 struct perf_cpu_context *cpuctx;
7363 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
7364 __perf_event_init_context(&cpuctx->ctx);
7365 lockdep_set_class(&cpuctx->ctx.mutex, &cpuctx_mutex);
7366 lockdep_set_class(&cpuctx->ctx.lock, &cpuctx_lock);
7367 cpuctx->ctx.pmu = pmu;
7369 __perf_mux_hrtimer_init(cpuctx, cpu);
7371 cpuctx->unique_pmu = pmu;
7375 if (!pmu->start_txn) {
7376 if (pmu->pmu_enable) {
7378 * If we have pmu_enable/pmu_disable calls, install
7379 * transaction stubs that use that to try and batch
7380 * hardware accesses.
7382 pmu->start_txn = perf_pmu_start_txn;
7383 pmu->commit_txn = perf_pmu_commit_txn;
7384 pmu->cancel_txn = perf_pmu_cancel_txn;
7386 pmu->start_txn = perf_pmu_nop_void;
7387 pmu->commit_txn = perf_pmu_nop_int;
7388 pmu->cancel_txn = perf_pmu_nop_void;
7392 if (!pmu->pmu_enable) {
7393 pmu->pmu_enable = perf_pmu_nop_void;
7394 pmu->pmu_disable = perf_pmu_nop_void;
7397 if (!pmu->event_idx)
7398 pmu->event_idx = perf_event_idx_default;
7400 list_add_rcu(&pmu->entry, &pmus);
7401 atomic_set(&pmu->exclusive_cnt, 0);
7404 mutex_unlock(&pmus_lock);
7409 device_del(pmu->dev);
7410 put_device(pmu->dev);
7413 if (pmu->type >= PERF_TYPE_MAX)
7414 idr_remove(&pmu_idr, pmu->type);
7417 free_percpu(pmu->pmu_disable_count);
7420 EXPORT_SYMBOL_GPL(perf_pmu_register);
7422 void perf_pmu_unregister(struct pmu *pmu)
7424 mutex_lock(&pmus_lock);
7425 list_del_rcu(&pmu->entry);
7426 mutex_unlock(&pmus_lock);
7429 * We dereference the pmu list under both SRCU and regular RCU, so
7430 * synchronize against both of those.
7432 synchronize_srcu(&pmus_srcu);
7435 free_percpu(pmu->pmu_disable_count);
7436 if (pmu->type >= PERF_TYPE_MAX)
7437 idr_remove(&pmu_idr, pmu->type);
7438 device_del(pmu->dev);
7439 put_device(pmu->dev);
7440 free_pmu_context(pmu);
7442 EXPORT_SYMBOL_GPL(perf_pmu_unregister);
7444 static int perf_try_init_event(struct pmu *pmu, struct perf_event *event)
7446 struct perf_event_context *ctx = NULL;
7449 if (!try_module_get(pmu->module))
7452 if (event->group_leader != event) {
7454 * This ctx->mutex can nest when we're called through
7455 * inheritance. See the perf_event_ctx_lock_nested() comment.
7457 ctx = perf_event_ctx_lock_nested(event->group_leader,
7458 SINGLE_DEPTH_NESTING);
7463 ret = pmu->event_init(event);
7466 perf_event_ctx_unlock(event->group_leader, ctx);
7469 module_put(pmu->module);
7474 struct pmu *perf_init_event(struct perf_event *event)
7476 struct pmu *pmu = NULL;
7480 idx = srcu_read_lock(&pmus_srcu);
7483 pmu = idr_find(&pmu_idr, event->attr.type);
7486 ret = perf_try_init_event(pmu, event);
7492 list_for_each_entry_rcu(pmu, &pmus, entry) {
7493 ret = perf_try_init_event(pmu, event);
7497 if (ret != -ENOENT) {
7502 pmu = ERR_PTR(-ENOENT);
7504 srcu_read_unlock(&pmus_srcu, idx);
7509 static void account_event_cpu(struct perf_event *event, int cpu)
7514 if (is_cgroup_event(event))
7515 atomic_inc(&per_cpu(perf_cgroup_events, cpu));
7518 static void account_event(struct perf_event *event)
7523 if (event->attach_state & PERF_ATTACH_TASK)
7524 static_key_slow_inc(&perf_sched_events.key);
7525 if (event->attr.mmap || event->attr.mmap_data)
7526 atomic_inc(&nr_mmap_events);
7527 if (event->attr.comm)
7528 atomic_inc(&nr_comm_events);
7529 if (event->attr.task)
7530 atomic_inc(&nr_task_events);
7531 if (event->attr.freq) {
7532 if (atomic_inc_return(&nr_freq_events) == 1)
7533 tick_nohz_full_kick_all();
7535 if (has_branch_stack(event))
7536 static_key_slow_inc(&perf_sched_events.key);
7537 if (is_cgroup_event(event))
7538 static_key_slow_inc(&perf_sched_events.key);
7540 account_event_cpu(event, event->cpu);
7544 * Allocate and initialize a event structure
7546 static struct perf_event *
7547 perf_event_alloc(struct perf_event_attr *attr, int cpu,
7548 struct task_struct *task,
7549 struct perf_event *group_leader,
7550 struct perf_event *parent_event,
7551 perf_overflow_handler_t overflow_handler,
7552 void *context, int cgroup_fd)
7555 struct perf_event *event;
7556 struct hw_perf_event *hwc;
7559 if ((unsigned)cpu >= nr_cpu_ids) {
7560 if (!task || cpu != -1)
7561 return ERR_PTR(-EINVAL);
7564 event = kzalloc(sizeof(*event), GFP_KERNEL);
7566 return ERR_PTR(-ENOMEM);
7569 * Single events are their own group leaders, with an
7570 * empty sibling list:
7573 group_leader = event;
7575 mutex_init(&event->child_mutex);
7576 INIT_LIST_HEAD(&event->child_list);
7578 INIT_LIST_HEAD(&event->group_entry);
7579 INIT_LIST_HEAD(&event->event_entry);
7580 INIT_LIST_HEAD(&event->sibling_list);
7581 INIT_LIST_HEAD(&event->rb_entry);
7582 INIT_LIST_HEAD(&event->active_entry);
7583 INIT_HLIST_NODE(&event->hlist_entry);
7586 init_waitqueue_head(&event->waitq);
7587 init_irq_work(&event->pending, perf_pending_event);
7589 mutex_init(&event->mmap_mutex);
7591 atomic_long_set(&event->refcount, 1);
7593 event->attr = *attr;
7594 event->group_leader = group_leader;
7598 event->parent = parent_event;
7600 event->ns = get_pid_ns(task_active_pid_ns(current));
7601 event->id = atomic64_inc_return(&perf_event_id);
7603 event->state = PERF_EVENT_STATE_INACTIVE;
7606 event->attach_state = PERF_ATTACH_TASK;
7608 * XXX pmu::event_init needs to know what task to account to
7609 * and we cannot use the ctx information because we need the
7610 * pmu before we get a ctx.
7612 event->hw.target = task;
7615 event->clock = &local_clock;
7617 event->clock = parent_event->clock;
7619 if (!overflow_handler && parent_event) {
7620 overflow_handler = parent_event->overflow_handler;
7621 context = parent_event->overflow_handler_context;
7624 event->overflow_handler = overflow_handler;
7625 event->overflow_handler_context = context;
7627 perf_event__state_init(event);
7632 hwc->sample_period = attr->sample_period;
7633 if (attr->freq && attr->sample_freq)
7634 hwc->sample_period = 1;
7635 hwc->last_period = hwc->sample_period;
7637 local64_set(&hwc->period_left, hwc->sample_period);
7640 * we currently do not support PERF_FORMAT_GROUP on inherited events
7642 if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP))
7645 if (!has_branch_stack(event))
7646 event->attr.branch_sample_type = 0;
7648 if (cgroup_fd != -1) {
7649 err = perf_cgroup_connect(cgroup_fd, event, attr, group_leader);
7654 pmu = perf_init_event(event);
7657 else if (IS_ERR(pmu)) {
7662 err = exclusive_event_init(event);
7666 if (!event->parent) {
7667 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) {
7668 err = get_callchain_buffers();
7677 exclusive_event_destroy(event);
7681 event->destroy(event);
7682 module_put(pmu->module);
7684 if (is_cgroup_event(event))
7685 perf_detach_cgroup(event);
7687 put_pid_ns(event->ns);
7690 return ERR_PTR(err);
7693 static int perf_copy_attr(struct perf_event_attr __user *uattr,
7694 struct perf_event_attr *attr)
7699 if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0))
7703 * zero the full structure, so that a short copy will be nice.
7705 memset(attr, 0, sizeof(*attr));
7707 ret = get_user(size, &uattr->size);
7711 if (size > PAGE_SIZE) /* silly large */
7714 if (!size) /* abi compat */
7715 size = PERF_ATTR_SIZE_VER0;
7717 if (size < PERF_ATTR_SIZE_VER0)
7721 * If we're handed a bigger struct than we know of,
7722 * ensure all the unknown bits are 0 - i.e. new
7723 * user-space does not rely on any kernel feature
7724 * extensions we dont know about yet.
7726 if (size > sizeof(*attr)) {
7727 unsigned char __user *addr;
7728 unsigned char __user *end;
7731 addr = (void __user *)uattr + sizeof(*attr);
7732 end = (void __user *)uattr + size;
7734 for (; addr < end; addr++) {
7735 ret = get_user(val, addr);
7741 size = sizeof(*attr);
7744 ret = copy_from_user(attr, uattr, size);
7748 if (attr->__reserved_1)
7751 if (attr->sample_type & ~(PERF_SAMPLE_MAX-1))
7754 if (attr->read_format & ~(PERF_FORMAT_MAX-1))
7757 if (attr->sample_type & PERF_SAMPLE_BRANCH_STACK) {
7758 u64 mask = attr->branch_sample_type;
7760 /* only using defined bits */
7761 if (mask & ~(PERF_SAMPLE_BRANCH_MAX-1))
7764 /* at least one branch bit must be set */
7765 if (!(mask & ~PERF_SAMPLE_BRANCH_PLM_ALL))
7768 /* propagate priv level, when not set for branch */
7769 if (!(mask & PERF_SAMPLE_BRANCH_PLM_ALL)) {
7771 /* exclude_kernel checked on syscall entry */
7772 if (!attr->exclude_kernel)
7773 mask |= PERF_SAMPLE_BRANCH_KERNEL;
7775 if (!attr->exclude_user)
7776 mask |= PERF_SAMPLE_BRANCH_USER;
7778 if (!attr->exclude_hv)
7779 mask |= PERF_SAMPLE_BRANCH_HV;
7781 * adjust user setting (for HW filter setup)
7783 attr->branch_sample_type = mask;
7785 /* privileged levels capture (kernel, hv): check permissions */
7786 if ((mask & PERF_SAMPLE_BRANCH_PERM_PLM)
7787 && perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
7791 if (attr->sample_type & PERF_SAMPLE_REGS_USER) {
7792 ret = perf_reg_validate(attr->sample_regs_user);
7797 if (attr->sample_type & PERF_SAMPLE_STACK_USER) {
7798 if (!arch_perf_have_user_stack_dump())
7802 * We have __u32 type for the size, but so far
7803 * we can only use __u16 as maximum due to the
7804 * __u16 sample size limit.
7806 if (attr->sample_stack_user >= USHRT_MAX)
7808 else if (!IS_ALIGNED(attr->sample_stack_user, sizeof(u64)))
7812 if (attr->sample_type & PERF_SAMPLE_REGS_INTR)
7813 ret = perf_reg_validate(attr->sample_regs_intr);
7818 put_user(sizeof(*attr), &uattr->size);
7824 perf_event_set_output(struct perf_event *event, struct perf_event *output_event)
7826 struct ring_buffer *rb = NULL;
7832 /* don't allow circular references */
7833 if (event == output_event)
7837 * Don't allow cross-cpu buffers
7839 if (output_event->cpu != event->cpu)
7843 * If its not a per-cpu rb, it must be the same task.
7845 if (output_event->cpu == -1 && output_event->ctx != event->ctx)
7849 * Mixing clocks in the same buffer is trouble you don't need.
7851 if (output_event->clock != event->clock)
7855 * If both events generate aux data, they must be on the same PMU
7857 if (has_aux(event) && has_aux(output_event) &&
7858 event->pmu != output_event->pmu)
7862 mutex_lock(&event->mmap_mutex);
7863 /* Can't redirect output if we've got an active mmap() */
7864 if (atomic_read(&event->mmap_count))
7868 /* get the rb we want to redirect to */
7869 rb = ring_buffer_get(output_event);
7874 ring_buffer_attach(event, rb);
7878 mutex_unlock(&event->mmap_mutex);
7884 static void mutex_lock_double(struct mutex *a, struct mutex *b)
7890 mutex_lock_nested(b, SINGLE_DEPTH_NESTING);
7893 static int perf_event_set_clock(struct perf_event *event, clockid_t clk_id)
7895 bool nmi_safe = false;
7898 case CLOCK_MONOTONIC:
7899 event->clock = &ktime_get_mono_fast_ns;
7903 case CLOCK_MONOTONIC_RAW:
7904 event->clock = &ktime_get_raw_fast_ns;
7908 case CLOCK_REALTIME:
7909 event->clock = &ktime_get_real_ns;
7912 case CLOCK_BOOTTIME:
7913 event->clock = &ktime_get_boot_ns;
7917 event->clock = &ktime_get_tai_ns;
7924 if (!nmi_safe && !(event->pmu->capabilities & PERF_PMU_CAP_NO_NMI))
7931 * sys_perf_event_open - open a performance event, associate it to a task/cpu
7933 * @attr_uptr: event_id type attributes for monitoring/sampling
7936 * @group_fd: group leader event fd
7938 SYSCALL_DEFINE5(perf_event_open,
7939 struct perf_event_attr __user *, attr_uptr,
7940 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
7942 struct perf_event *group_leader = NULL, *output_event = NULL;
7943 struct perf_event *event, *sibling;
7944 struct perf_event_attr attr;
7945 struct perf_event_context *ctx, *uninitialized_var(gctx);
7946 struct file *event_file = NULL;
7947 struct fd group = {NULL, 0};
7948 struct task_struct *task = NULL;
7953 int f_flags = O_RDWR;
7956 /* for future expandability... */
7957 if (flags & ~PERF_FLAG_ALL)
7960 err = perf_copy_attr(attr_uptr, &attr);
7964 if (!attr.exclude_kernel) {
7965 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
7970 if (attr.sample_freq > sysctl_perf_event_sample_rate)
7973 if (attr.sample_period & (1ULL << 63))
7978 * In cgroup mode, the pid argument is used to pass the fd
7979 * opened to the cgroup directory in cgroupfs. The cpu argument
7980 * designates the cpu on which to monitor threads from that
7983 if ((flags & PERF_FLAG_PID_CGROUP) && (pid == -1 || cpu == -1))
7986 if (flags & PERF_FLAG_FD_CLOEXEC)
7987 f_flags |= O_CLOEXEC;
7989 event_fd = get_unused_fd_flags(f_flags);
7993 if (group_fd != -1) {
7994 err = perf_fget_light(group_fd, &group);
7997 group_leader = group.file->private_data;
7998 if (flags & PERF_FLAG_FD_OUTPUT)
7999 output_event = group_leader;
8000 if (flags & PERF_FLAG_FD_NO_GROUP)
8001 group_leader = NULL;
8004 if (pid != -1 && !(flags & PERF_FLAG_PID_CGROUP)) {
8005 task = find_lively_task_by_vpid(pid);
8007 err = PTR_ERR(task);
8012 if (task && group_leader &&
8013 group_leader->attr.inherit != attr.inherit) {
8020 if (flags & PERF_FLAG_PID_CGROUP)
8023 event = perf_event_alloc(&attr, cpu, task, group_leader, NULL,
8024 NULL, NULL, cgroup_fd);
8025 if (IS_ERR(event)) {
8026 err = PTR_ERR(event);
8030 if (is_sampling_event(event)) {
8031 if (event->pmu->capabilities & PERF_PMU_CAP_NO_INTERRUPT) {
8037 account_event(event);
8040 * Special case software events and allow them to be part of
8041 * any hardware group.
8045 if (attr.use_clockid) {
8046 err = perf_event_set_clock(event, attr.clockid);
8052 (is_software_event(event) != is_software_event(group_leader))) {
8053 if (is_software_event(event)) {
8055 * If event and group_leader are not both a software
8056 * event, and event is, then group leader is not.
8058 * Allow the addition of software events to !software
8059 * groups, this is safe because software events never
8062 pmu = group_leader->pmu;
8063 } else if (is_software_event(group_leader) &&
8064 (group_leader->group_flags & PERF_GROUP_SOFTWARE)) {
8066 * In case the group is a pure software group, and we
8067 * try to add a hardware event, move the whole group to
8068 * the hardware context.
8075 * Get the target context (task or percpu):
8077 ctx = find_get_context(pmu, task, event);
8083 if ((pmu->capabilities & PERF_PMU_CAP_EXCLUSIVE) && group_leader) {
8089 put_task_struct(task);
8094 * Look up the group leader (we will attach this event to it):
8100 * Do not allow a recursive hierarchy (this new sibling
8101 * becoming part of another group-sibling):
8103 if (group_leader->group_leader != group_leader)
8106 /* All events in a group should have the same clock */
8107 if (group_leader->clock != event->clock)
8111 * Do not allow to attach to a group in a different
8112 * task or CPU context:
8116 * Make sure we're both on the same task, or both
8119 if (group_leader->ctx->task != ctx->task)
8123 * Make sure we're both events for the same CPU;
8124 * grouping events for different CPUs is broken; since
8125 * you can never concurrently schedule them anyhow.
8127 if (group_leader->cpu != event->cpu)
8130 if (group_leader->ctx != ctx)
8135 * Only a group leader can be exclusive or pinned
8137 if (attr.exclusive || attr.pinned)
8142 err = perf_event_set_output(event, output_event);
8147 event_file = anon_inode_getfile("[perf_event]", &perf_fops, event,
8149 if (IS_ERR(event_file)) {
8150 err = PTR_ERR(event_file);
8155 gctx = group_leader->ctx;
8158 * See perf_event_ctx_lock() for comments on the details
8159 * of swizzling perf_event::ctx.
8161 mutex_lock_double(&gctx->mutex, &ctx->mutex);
8163 perf_remove_from_context(group_leader, false);
8165 list_for_each_entry(sibling, &group_leader->sibling_list,
8167 perf_remove_from_context(sibling, false);
8171 mutex_lock(&ctx->mutex);
8174 WARN_ON_ONCE(ctx->parent_ctx);
8178 * Wait for everybody to stop referencing the events through
8179 * the old lists, before installing it on new lists.
8184 * Install the group siblings before the group leader.
8186 * Because a group leader will try and install the entire group
8187 * (through the sibling list, which is still in-tact), we can
8188 * end up with siblings installed in the wrong context.
8190 * By installing siblings first we NO-OP because they're not
8191 * reachable through the group lists.
8193 list_for_each_entry(sibling, &group_leader->sibling_list,
8195 perf_event__state_init(sibling);
8196 perf_install_in_context(ctx, sibling, sibling->cpu);
8201 * Removing from the context ends up with disabled
8202 * event. What we want here is event in the initial
8203 * startup state, ready to be add into new context.
8205 perf_event__state_init(group_leader);
8206 perf_install_in_context(ctx, group_leader, group_leader->cpu);
8210 if (!exclusive_event_installable(event, ctx)) {
8212 mutex_unlock(&ctx->mutex);
8217 perf_install_in_context(ctx, event, event->cpu);
8218 perf_unpin_context(ctx);
8221 mutex_unlock(&gctx->mutex);
8224 mutex_unlock(&ctx->mutex);
8228 event->owner = current;
8230 mutex_lock(¤t->perf_event_mutex);
8231 list_add_tail(&event->owner_entry, ¤t->perf_event_list);
8232 mutex_unlock(¤t->perf_event_mutex);
8235 * Precalculate sample_data sizes
8237 perf_event__header_size(event);
8238 perf_event__id_header_size(event);
8241 * Drop the reference on the group_event after placing the
8242 * new event on the sibling_list. This ensures destruction
8243 * of the group leader will find the pointer to itself in
8244 * perf_group_detach().
8247 fd_install(event_fd, event_file);
8251 perf_unpin_context(ctx);
8259 put_task_struct(task);
8263 put_unused_fd(event_fd);
8268 * perf_event_create_kernel_counter
8270 * @attr: attributes of the counter to create
8271 * @cpu: cpu in which the counter is bound
8272 * @task: task to profile (NULL for percpu)
8275 perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu,
8276 struct task_struct *task,
8277 perf_overflow_handler_t overflow_handler,
8280 struct perf_event_context *ctx;
8281 struct perf_event *event;
8285 * Get the target context (task or percpu):
8288 event = perf_event_alloc(attr, cpu, task, NULL, NULL,
8289 overflow_handler, context, -1);
8290 if (IS_ERR(event)) {
8291 err = PTR_ERR(event);
8295 /* Mark owner so we could distinguish it from user events. */
8296 event->owner = EVENT_OWNER_KERNEL;
8298 account_event(event);
8300 ctx = find_get_context(event->pmu, task, event);
8306 WARN_ON_ONCE(ctx->parent_ctx);
8307 mutex_lock(&ctx->mutex);
8308 if (!exclusive_event_installable(event, ctx)) {
8309 mutex_unlock(&ctx->mutex);
8310 perf_unpin_context(ctx);
8316 perf_install_in_context(ctx, event, cpu);
8317 perf_unpin_context(ctx);
8318 mutex_unlock(&ctx->mutex);
8325 return ERR_PTR(err);
8327 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter);
8329 void perf_pmu_migrate_context(struct pmu *pmu, int src_cpu, int dst_cpu)
8331 struct perf_event_context *src_ctx;
8332 struct perf_event_context *dst_ctx;
8333 struct perf_event *event, *tmp;
8336 src_ctx = &per_cpu_ptr(pmu->pmu_cpu_context, src_cpu)->ctx;
8337 dst_ctx = &per_cpu_ptr(pmu->pmu_cpu_context, dst_cpu)->ctx;
8340 * See perf_event_ctx_lock() for comments on the details
8341 * of swizzling perf_event::ctx.
8343 mutex_lock_double(&src_ctx->mutex, &dst_ctx->mutex);
8344 list_for_each_entry_safe(event, tmp, &src_ctx->event_list,
8346 perf_remove_from_context(event, false);
8347 unaccount_event_cpu(event, src_cpu);
8349 list_add(&event->migrate_entry, &events);
8353 * Wait for the events to quiesce before re-instating them.
8358 * Re-instate events in 2 passes.
8360 * Skip over group leaders and only install siblings on this first
8361 * pass, siblings will not get enabled without a leader, however a
8362 * leader will enable its siblings, even if those are still on the old
8365 list_for_each_entry_safe(event, tmp, &events, migrate_entry) {
8366 if (event->group_leader == event)
8369 list_del(&event->migrate_entry);
8370 if (event->state >= PERF_EVENT_STATE_OFF)
8371 event->state = PERF_EVENT_STATE_INACTIVE;
8372 account_event_cpu(event, dst_cpu);
8373 perf_install_in_context(dst_ctx, event, dst_cpu);
8378 * Once all the siblings are setup properly, install the group leaders
8381 list_for_each_entry_safe(event, tmp, &events, migrate_entry) {
8382 list_del(&event->migrate_entry);
8383 if (event->state >= PERF_EVENT_STATE_OFF)
8384 event->state = PERF_EVENT_STATE_INACTIVE;
8385 account_event_cpu(event, dst_cpu);
8386 perf_install_in_context(dst_ctx, event, dst_cpu);
8389 mutex_unlock(&dst_ctx->mutex);
8390 mutex_unlock(&src_ctx->mutex);
8392 EXPORT_SYMBOL_GPL(perf_pmu_migrate_context);
8394 static void sync_child_event(struct perf_event *child_event,
8395 struct task_struct *child)
8397 struct perf_event *parent_event = child_event->parent;
8400 if (child_event->attr.inherit_stat)
8401 perf_event_read_event(child_event, child);
8403 child_val = perf_event_count(child_event);
8406 * Add back the child's count to the parent's count:
8408 atomic64_add(child_val, &parent_event->child_count);
8409 atomic64_add(child_event->total_time_enabled,
8410 &parent_event->child_total_time_enabled);
8411 atomic64_add(child_event->total_time_running,
8412 &parent_event->child_total_time_running);
8415 * Remove this event from the parent's list
8417 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
8418 mutex_lock(&parent_event->child_mutex);
8419 list_del_init(&child_event->child_list);
8420 mutex_unlock(&parent_event->child_mutex);
8423 * Make sure user/parent get notified, that we just
8426 perf_event_wakeup(parent_event);
8429 * Release the parent event, if this was the last
8432 put_event(parent_event);
8436 __perf_event_exit_task(struct perf_event *child_event,
8437 struct perf_event_context *child_ctx,
8438 struct task_struct *child)
8441 * Do not destroy the 'original' grouping; because of the context
8442 * switch optimization the original events could've ended up in a
8443 * random child task.
8445 * If we were to destroy the original group, all group related
8446 * operations would cease to function properly after this random
8449 * Do destroy all inherited groups, we don't care about those
8450 * and being thorough is better.
8452 perf_remove_from_context(child_event, !!child_event->parent);
8455 * It can happen that the parent exits first, and has events
8456 * that are still around due to the child reference. These
8457 * events need to be zapped.
8459 if (child_event->parent) {
8460 sync_child_event(child_event, child);
8461 free_event(child_event);
8463 child_event->state = PERF_EVENT_STATE_EXIT;
8464 perf_event_wakeup(child_event);
8468 static void perf_event_exit_task_context(struct task_struct *child, int ctxn)
8470 struct perf_event *child_event, *next;
8471 struct perf_event_context *child_ctx, *clone_ctx = NULL;
8472 unsigned long flags;
8474 if (likely(!child->perf_event_ctxp[ctxn])) {
8475 perf_event_task(child, NULL, 0);
8479 local_irq_save(flags);
8481 * We can't reschedule here because interrupts are disabled,
8482 * and either child is current or it is a task that can't be
8483 * scheduled, so we are now safe from rescheduling changing
8486 child_ctx = rcu_dereference_raw(child->perf_event_ctxp[ctxn]);
8489 * Take the context lock here so that if find_get_context is
8490 * reading child->perf_event_ctxp, we wait until it has
8491 * incremented the context's refcount before we do put_ctx below.
8493 raw_spin_lock(&child_ctx->lock);
8494 task_ctx_sched_out(child_ctx);
8495 child->perf_event_ctxp[ctxn] = NULL;
8498 * If this context is a clone; unclone it so it can't get
8499 * swapped to another process while we're removing all
8500 * the events from it.
8502 clone_ctx = unclone_ctx(child_ctx);
8503 update_context_time(child_ctx);
8504 raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
8510 * Report the task dead after unscheduling the events so that we
8511 * won't get any samples after PERF_RECORD_EXIT. We can however still
8512 * get a few PERF_RECORD_READ events.
8514 perf_event_task(child, child_ctx, 0);
8517 * We can recurse on the same lock type through:
8519 * __perf_event_exit_task()
8520 * sync_child_event()
8522 * mutex_lock(&ctx->mutex)
8524 * But since its the parent context it won't be the same instance.
8526 mutex_lock(&child_ctx->mutex);
8528 list_for_each_entry_safe(child_event, next, &child_ctx->event_list, event_entry)
8529 __perf_event_exit_task(child_event, child_ctx, child);
8531 mutex_unlock(&child_ctx->mutex);
8537 * When a child task exits, feed back event values to parent events.
8539 void perf_event_exit_task(struct task_struct *child)
8541 struct perf_event *event, *tmp;
8544 mutex_lock(&child->perf_event_mutex);
8545 list_for_each_entry_safe(event, tmp, &child->perf_event_list,
8547 list_del_init(&event->owner_entry);
8550 * Ensure the list deletion is visible before we clear
8551 * the owner, closes a race against perf_release() where
8552 * we need to serialize on the owner->perf_event_mutex.
8555 event->owner = NULL;
8557 mutex_unlock(&child->perf_event_mutex);
8559 for_each_task_context_nr(ctxn)
8560 perf_event_exit_task_context(child, ctxn);
8563 static void perf_free_event(struct perf_event *event,
8564 struct perf_event_context *ctx)
8566 struct perf_event *parent = event->parent;
8568 if (WARN_ON_ONCE(!parent))
8571 mutex_lock(&parent->child_mutex);
8572 list_del_init(&event->child_list);
8573 mutex_unlock(&parent->child_mutex);
8577 raw_spin_lock_irq(&ctx->lock);
8578 perf_group_detach(event);
8579 list_del_event(event, ctx);
8580 raw_spin_unlock_irq(&ctx->lock);
8585 * Free an unexposed, unused context as created by inheritance by
8586 * perf_event_init_task below, used by fork() in case of fail.
8588 * Not all locks are strictly required, but take them anyway to be nice and
8589 * help out with the lockdep assertions.
8591 void perf_event_free_task(struct task_struct *task)
8593 struct perf_event_context *ctx;
8594 struct perf_event *event, *tmp;
8597 for_each_task_context_nr(ctxn) {
8598 ctx = task->perf_event_ctxp[ctxn];
8602 mutex_lock(&ctx->mutex);
8604 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups,
8606 perf_free_event(event, ctx);
8608 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups,
8610 perf_free_event(event, ctx);
8612 if (!list_empty(&ctx->pinned_groups) ||
8613 !list_empty(&ctx->flexible_groups))
8616 mutex_unlock(&ctx->mutex);
8622 void perf_event_delayed_put(struct task_struct *task)
8626 for_each_task_context_nr(ctxn)
8627 WARN_ON_ONCE(task->perf_event_ctxp[ctxn]);
8630 struct perf_event *perf_event_get(unsigned int fd)
8634 struct perf_event *event;
8636 err = perf_fget_light(fd, &f);
8638 return ERR_PTR(err);
8640 event = f.file->private_data;
8641 atomic_long_inc(&event->refcount);
8647 const struct perf_event_attr *perf_event_attrs(struct perf_event *event)
8650 return ERR_PTR(-EINVAL);
8652 return &event->attr;
8656 * inherit a event from parent task to child task:
8658 static struct perf_event *
8659 inherit_event(struct perf_event *parent_event,
8660 struct task_struct *parent,
8661 struct perf_event_context *parent_ctx,
8662 struct task_struct *child,
8663 struct perf_event *group_leader,
8664 struct perf_event_context *child_ctx)
8666 enum perf_event_active_state parent_state = parent_event->state;
8667 struct perf_event *child_event;
8668 unsigned long flags;
8671 * Instead of creating recursive hierarchies of events,
8672 * we link inherited events back to the original parent,
8673 * which has a filp for sure, which we use as the reference
8676 if (parent_event->parent)
8677 parent_event = parent_event->parent;
8679 child_event = perf_event_alloc(&parent_event->attr,
8682 group_leader, parent_event,
8684 if (IS_ERR(child_event))
8687 if (is_orphaned_event(parent_event) ||
8688 !atomic_long_inc_not_zero(&parent_event->refcount)) {
8689 free_event(child_event);
8696 * Make the child state follow the state of the parent event,
8697 * not its attr.disabled bit. We hold the parent's mutex,
8698 * so we won't race with perf_event_{en, dis}able_family.
8700 if (parent_state >= PERF_EVENT_STATE_INACTIVE)
8701 child_event->state = PERF_EVENT_STATE_INACTIVE;
8703 child_event->state = PERF_EVENT_STATE_OFF;
8705 if (parent_event->attr.freq) {
8706 u64 sample_period = parent_event->hw.sample_period;
8707 struct hw_perf_event *hwc = &child_event->hw;
8709 hwc->sample_period = sample_period;
8710 hwc->last_period = sample_period;
8712 local64_set(&hwc->period_left, sample_period);
8715 child_event->ctx = child_ctx;
8716 child_event->overflow_handler = parent_event->overflow_handler;
8717 child_event->overflow_handler_context
8718 = parent_event->overflow_handler_context;
8721 * Precalculate sample_data sizes
8723 perf_event__header_size(child_event);
8724 perf_event__id_header_size(child_event);
8727 * Link it up in the child's context:
8729 raw_spin_lock_irqsave(&child_ctx->lock, flags);
8730 add_event_to_ctx(child_event, child_ctx);
8731 raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
8734 * Link this into the parent event's child list
8736 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
8737 mutex_lock(&parent_event->child_mutex);
8738 list_add_tail(&child_event->child_list, &parent_event->child_list);
8739 mutex_unlock(&parent_event->child_mutex);
8744 static int inherit_group(struct perf_event *parent_event,
8745 struct task_struct *parent,
8746 struct perf_event_context *parent_ctx,
8747 struct task_struct *child,
8748 struct perf_event_context *child_ctx)
8750 struct perf_event *leader;
8751 struct perf_event *sub;
8752 struct perf_event *child_ctr;
8754 leader = inherit_event(parent_event, parent, parent_ctx,
8755 child, NULL, child_ctx);
8757 return PTR_ERR(leader);
8758 list_for_each_entry(sub, &parent_event->sibling_list, group_entry) {
8759 child_ctr = inherit_event(sub, parent, parent_ctx,
8760 child, leader, child_ctx);
8761 if (IS_ERR(child_ctr))
8762 return PTR_ERR(child_ctr);
8768 inherit_task_group(struct perf_event *event, struct task_struct *parent,
8769 struct perf_event_context *parent_ctx,
8770 struct task_struct *child, int ctxn,
8774 struct perf_event_context *child_ctx;
8776 if (!event->attr.inherit) {
8781 child_ctx = child->perf_event_ctxp[ctxn];
8784 * This is executed from the parent task context, so
8785 * inherit events that have been marked for cloning.
8786 * First allocate and initialize a context for the
8790 child_ctx = alloc_perf_context(parent_ctx->pmu, child);
8794 child->perf_event_ctxp[ctxn] = child_ctx;
8797 ret = inherit_group(event, parent, parent_ctx,
8807 * Initialize the perf_event context in task_struct
8809 static int perf_event_init_context(struct task_struct *child, int ctxn)
8811 struct perf_event_context *child_ctx, *parent_ctx;
8812 struct perf_event_context *cloned_ctx;
8813 struct perf_event *event;
8814 struct task_struct *parent = current;
8815 int inherited_all = 1;
8816 unsigned long flags;
8819 if (likely(!parent->perf_event_ctxp[ctxn]))
8823 * If the parent's context is a clone, pin it so it won't get
8826 parent_ctx = perf_pin_task_context(parent, ctxn);
8831 * No need to check if parent_ctx != NULL here; since we saw
8832 * it non-NULL earlier, the only reason for it to become NULL
8833 * is if we exit, and since we're currently in the middle of
8834 * a fork we can't be exiting at the same time.
8838 * Lock the parent list. No need to lock the child - not PID
8839 * hashed yet and not running, so nobody can access it.
8841 mutex_lock(&parent_ctx->mutex);
8844 * We dont have to disable NMIs - we are only looking at
8845 * the list, not manipulating it:
8847 list_for_each_entry(event, &parent_ctx->pinned_groups, group_entry) {
8848 ret = inherit_task_group(event, parent, parent_ctx,
8849 child, ctxn, &inherited_all);
8855 * We can't hold ctx->lock when iterating the ->flexible_group list due
8856 * to allocations, but we need to prevent rotation because
8857 * rotate_ctx() will change the list from interrupt context.
8859 raw_spin_lock_irqsave(&parent_ctx->lock, flags);
8860 parent_ctx->rotate_disable = 1;
8861 raw_spin_unlock_irqrestore(&parent_ctx->lock, flags);
8863 list_for_each_entry(event, &parent_ctx->flexible_groups, group_entry) {
8864 ret = inherit_task_group(event, parent, parent_ctx,
8865 child, ctxn, &inherited_all);
8870 raw_spin_lock_irqsave(&parent_ctx->lock, flags);
8871 parent_ctx->rotate_disable = 0;
8873 child_ctx = child->perf_event_ctxp[ctxn];
8875 if (child_ctx && inherited_all) {
8877 * Mark the child context as a clone of the parent
8878 * context, or of whatever the parent is a clone of.
8880 * Note that if the parent is a clone, the holding of
8881 * parent_ctx->lock avoids it from being uncloned.
8883 cloned_ctx = parent_ctx->parent_ctx;
8885 child_ctx->parent_ctx = cloned_ctx;
8886 child_ctx->parent_gen = parent_ctx->parent_gen;
8888 child_ctx->parent_ctx = parent_ctx;
8889 child_ctx->parent_gen = parent_ctx->generation;
8891 get_ctx(child_ctx->parent_ctx);
8894 raw_spin_unlock_irqrestore(&parent_ctx->lock, flags);
8895 mutex_unlock(&parent_ctx->mutex);
8897 perf_unpin_context(parent_ctx);
8898 put_ctx(parent_ctx);
8904 * Initialize the perf_event context in task_struct
8906 int perf_event_init_task(struct task_struct *child)
8910 memset(child->perf_event_ctxp, 0, sizeof(child->perf_event_ctxp));
8911 mutex_init(&child->perf_event_mutex);
8912 INIT_LIST_HEAD(&child->perf_event_list);
8914 for_each_task_context_nr(ctxn) {
8915 ret = perf_event_init_context(child, ctxn);
8917 perf_event_free_task(child);
8925 static void __init perf_event_init_all_cpus(void)
8927 struct swevent_htable *swhash;
8930 for_each_possible_cpu(cpu) {
8931 swhash = &per_cpu(swevent_htable, cpu);
8932 mutex_init(&swhash->hlist_mutex);
8933 INIT_LIST_HEAD(&per_cpu(active_ctx_list, cpu));
8937 static void perf_event_init_cpu(int cpu)
8939 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
8941 mutex_lock(&swhash->hlist_mutex);
8942 swhash->online = true;
8943 if (swhash->hlist_refcount > 0) {
8944 struct swevent_hlist *hlist;
8946 hlist = kzalloc_node(sizeof(*hlist), GFP_KERNEL, cpu_to_node(cpu));
8948 rcu_assign_pointer(swhash->swevent_hlist, hlist);
8950 mutex_unlock(&swhash->hlist_mutex);
8953 #if defined CONFIG_HOTPLUG_CPU || defined CONFIG_KEXEC
8954 static void __perf_event_exit_context(void *__info)
8956 struct remove_event re = { .detach_group = true };
8957 struct perf_event_context *ctx = __info;
8960 list_for_each_entry_rcu(re.event, &ctx->event_list, event_entry)
8961 __perf_remove_from_context(&re);
8965 static void perf_event_exit_cpu_context(int cpu)
8967 struct perf_event_context *ctx;
8971 idx = srcu_read_lock(&pmus_srcu);
8972 list_for_each_entry_rcu(pmu, &pmus, entry) {
8973 ctx = &per_cpu_ptr(pmu->pmu_cpu_context, cpu)->ctx;
8975 mutex_lock(&ctx->mutex);
8976 smp_call_function_single(cpu, __perf_event_exit_context, ctx, 1);
8977 mutex_unlock(&ctx->mutex);
8979 srcu_read_unlock(&pmus_srcu, idx);
8982 static void perf_event_exit_cpu(int cpu)
8984 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
8986 perf_event_exit_cpu_context(cpu);
8988 mutex_lock(&swhash->hlist_mutex);
8989 swhash->online = false;
8990 swevent_hlist_release(swhash);
8991 mutex_unlock(&swhash->hlist_mutex);
8994 static inline void perf_event_exit_cpu(int cpu) { }
8998 perf_reboot(struct notifier_block *notifier, unsigned long val, void *v)
9002 for_each_online_cpu(cpu)
9003 perf_event_exit_cpu(cpu);
9009 * Run the perf reboot notifier at the very last possible moment so that
9010 * the generic watchdog code runs as long as possible.
9012 static struct notifier_block perf_reboot_notifier = {
9013 .notifier_call = perf_reboot,
9014 .priority = INT_MIN,
9018 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
9020 unsigned int cpu = (long)hcpu;
9022 switch (action & ~CPU_TASKS_FROZEN) {
9024 case CPU_UP_PREPARE:
9025 case CPU_DOWN_FAILED:
9026 perf_event_init_cpu(cpu);
9029 case CPU_UP_CANCELED:
9030 case CPU_DOWN_PREPARE:
9031 perf_event_exit_cpu(cpu);
9040 void __init perf_event_init(void)
9046 perf_event_init_all_cpus();
9047 init_srcu_struct(&pmus_srcu);
9048 perf_pmu_register(&perf_swevent, "software", PERF_TYPE_SOFTWARE);
9049 perf_pmu_register(&perf_cpu_clock, NULL, -1);
9050 perf_pmu_register(&perf_task_clock, NULL, -1);
9052 perf_cpu_notifier(perf_cpu_notify);
9053 register_reboot_notifier(&perf_reboot_notifier);
9055 ret = init_hw_breakpoint();
9056 WARN(ret, "hw_breakpoint initialization failed with: %d", ret);
9058 /* do not patch jump label more than once per second */
9059 jump_label_rate_limit(&perf_sched_events, HZ);
9062 * Build time assertion that we keep the data_head at the intended
9063 * location. IOW, validation we got the __reserved[] size right.
9065 BUILD_BUG_ON((offsetof(struct perf_event_mmap_page, data_head))
9069 ssize_t perf_event_sysfs_show(struct device *dev, struct device_attribute *attr,
9072 struct perf_pmu_events_attr *pmu_attr =
9073 container_of(attr, struct perf_pmu_events_attr, attr);
9075 if (pmu_attr->event_str)
9076 return sprintf(page, "%s\n", pmu_attr->event_str);
9081 static int __init perf_event_sysfs_init(void)
9086 mutex_lock(&pmus_lock);
9088 ret = bus_register(&pmu_bus);
9092 list_for_each_entry(pmu, &pmus, entry) {
9093 if (!pmu->name || pmu->type < 0)
9096 ret = pmu_dev_alloc(pmu);
9097 WARN(ret, "Failed to register pmu: %s, reason %d\n", pmu->name, ret);
9099 pmu_bus_running = 1;
9103 mutex_unlock(&pmus_lock);
9107 device_initcall(perf_event_sysfs_init);
9109 #ifdef CONFIG_CGROUP_PERF
9110 static struct cgroup_subsys_state *
9111 perf_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
9113 struct perf_cgroup *jc;
9115 jc = kzalloc(sizeof(*jc), GFP_KERNEL);
9117 return ERR_PTR(-ENOMEM);
9119 jc->info = alloc_percpu(struct perf_cgroup_info);
9122 return ERR_PTR(-ENOMEM);
9128 static void perf_cgroup_css_free(struct cgroup_subsys_state *css)
9130 struct perf_cgroup *jc = container_of(css, struct perf_cgroup, css);
9132 free_percpu(jc->info);
9136 static int __perf_cgroup_move(void *info)
9138 struct task_struct *task = info;
9139 perf_cgroup_switch(task, PERF_CGROUP_SWOUT | PERF_CGROUP_SWIN);
9143 static void perf_cgroup_attach(struct cgroup_subsys_state *css,
9144 struct cgroup_taskset *tset)
9146 struct task_struct *task;
9148 cgroup_taskset_for_each(task, tset)
9149 task_function_call(task, __perf_cgroup_move, task);
9152 static void perf_cgroup_exit(struct cgroup_subsys_state *css,
9153 struct cgroup_subsys_state *old_css,
9154 struct task_struct *task)
9157 * cgroup_exit() is called in the copy_process() failure path.
9158 * Ignore this case since the task hasn't ran yet, this avoids
9159 * trying to poke a half freed task state from generic code.
9161 if (!(task->flags & PF_EXITING))
9164 task_function_call(task, __perf_cgroup_move, task);
9167 struct cgroup_subsys perf_event_cgrp_subsys = {
9168 .css_alloc = perf_cgroup_css_alloc,
9169 .css_free = perf_cgroup_css_free,
9170 .exit = perf_cgroup_exit,
9171 .attach = perf_cgroup_attach,
9173 #endif /* CONFIG_CGROUP_PERF */