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);
3215 static u64 perf_event_read(struct perf_event *event)
3218 * If event is enabled and currently active on a CPU, update the
3219 * value in the event structure:
3221 if (event->state == PERF_EVENT_STATE_ACTIVE) {
3222 smp_call_function_single(event->oncpu,
3223 __perf_event_read, event, 1);
3224 } else if (event->state == PERF_EVENT_STATE_INACTIVE) {
3225 struct perf_event_context *ctx = event->ctx;
3226 unsigned long flags;
3228 raw_spin_lock_irqsave(&ctx->lock, flags);
3230 * may read while context is not active
3231 * (e.g., thread is blocked), in that case
3232 * we cannot update context time
3234 if (ctx->is_active) {
3235 update_context_time(ctx);
3236 update_cgrp_time_from_event(event);
3238 update_event_times(event);
3239 raw_spin_unlock_irqrestore(&ctx->lock, flags);
3242 return perf_event_count(event);
3246 * Initialize the perf_event context in a task_struct:
3248 static void __perf_event_init_context(struct perf_event_context *ctx)
3250 raw_spin_lock_init(&ctx->lock);
3251 mutex_init(&ctx->mutex);
3252 INIT_LIST_HEAD(&ctx->active_ctx_list);
3253 INIT_LIST_HEAD(&ctx->pinned_groups);
3254 INIT_LIST_HEAD(&ctx->flexible_groups);
3255 INIT_LIST_HEAD(&ctx->event_list);
3256 atomic_set(&ctx->refcount, 1);
3257 INIT_DELAYED_WORK(&ctx->orphans_remove, orphans_remove_work);
3260 static struct perf_event_context *
3261 alloc_perf_context(struct pmu *pmu, struct task_struct *task)
3263 struct perf_event_context *ctx;
3265 ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL);
3269 __perf_event_init_context(ctx);
3272 get_task_struct(task);
3279 static struct task_struct *
3280 find_lively_task_by_vpid(pid_t vpid)
3282 struct task_struct *task;
3289 task = find_task_by_vpid(vpid);
3291 get_task_struct(task);
3295 return ERR_PTR(-ESRCH);
3297 /* Reuse ptrace permission checks for now. */
3299 if (!ptrace_may_access(task, PTRACE_MODE_READ))
3304 put_task_struct(task);
3305 return ERR_PTR(err);
3310 * Returns a matching context with refcount and pincount.
3312 static struct perf_event_context *
3313 find_get_context(struct pmu *pmu, struct task_struct *task,
3314 struct perf_event *event)
3316 struct perf_event_context *ctx, *clone_ctx = NULL;
3317 struct perf_cpu_context *cpuctx;
3318 void *task_ctx_data = NULL;
3319 unsigned long flags;
3321 int cpu = event->cpu;
3324 /* Must be root to operate on a CPU event: */
3325 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
3326 return ERR_PTR(-EACCES);
3329 * We could be clever and allow to attach a event to an
3330 * offline CPU and activate it when the CPU comes up, but
3333 if (!cpu_online(cpu))
3334 return ERR_PTR(-ENODEV);
3336 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
3345 ctxn = pmu->task_ctx_nr;
3349 if (event->attach_state & PERF_ATTACH_TASK_DATA) {
3350 task_ctx_data = kzalloc(pmu->task_ctx_size, GFP_KERNEL);
3351 if (!task_ctx_data) {
3358 ctx = perf_lock_task_context(task, ctxn, &flags);
3360 clone_ctx = unclone_ctx(ctx);
3363 if (task_ctx_data && !ctx->task_ctx_data) {
3364 ctx->task_ctx_data = task_ctx_data;
3365 task_ctx_data = NULL;
3367 raw_spin_unlock_irqrestore(&ctx->lock, flags);
3372 ctx = alloc_perf_context(pmu, task);
3377 if (task_ctx_data) {
3378 ctx->task_ctx_data = task_ctx_data;
3379 task_ctx_data = NULL;
3383 mutex_lock(&task->perf_event_mutex);
3385 * If it has already passed perf_event_exit_task().
3386 * we must see PF_EXITING, it takes this mutex too.
3388 if (task->flags & PF_EXITING)
3390 else if (task->perf_event_ctxp[ctxn])
3395 rcu_assign_pointer(task->perf_event_ctxp[ctxn], ctx);
3397 mutex_unlock(&task->perf_event_mutex);
3399 if (unlikely(err)) {
3408 kfree(task_ctx_data);
3412 kfree(task_ctx_data);
3413 return ERR_PTR(err);
3416 static void perf_event_free_filter(struct perf_event *event);
3417 static void perf_event_free_bpf_prog(struct perf_event *event);
3419 static void free_event_rcu(struct rcu_head *head)
3421 struct perf_event *event;
3423 event = container_of(head, struct perf_event, rcu_head);
3425 put_pid_ns(event->ns);
3426 perf_event_free_filter(event);
3430 static void ring_buffer_attach(struct perf_event *event,
3431 struct ring_buffer *rb);
3433 static void unaccount_event_cpu(struct perf_event *event, int cpu)
3438 if (is_cgroup_event(event))
3439 atomic_dec(&per_cpu(perf_cgroup_events, cpu));
3442 static void unaccount_event(struct perf_event *event)
3447 if (event->attach_state & PERF_ATTACH_TASK)
3448 static_key_slow_dec_deferred(&perf_sched_events);
3449 if (event->attr.mmap || event->attr.mmap_data)
3450 atomic_dec(&nr_mmap_events);
3451 if (event->attr.comm)
3452 atomic_dec(&nr_comm_events);
3453 if (event->attr.task)
3454 atomic_dec(&nr_task_events);
3455 if (event->attr.freq)
3456 atomic_dec(&nr_freq_events);
3457 if (is_cgroup_event(event))
3458 static_key_slow_dec_deferred(&perf_sched_events);
3459 if (has_branch_stack(event))
3460 static_key_slow_dec_deferred(&perf_sched_events);
3462 unaccount_event_cpu(event, event->cpu);
3466 * The following implement mutual exclusion of events on "exclusive" pmus
3467 * (PERF_PMU_CAP_EXCLUSIVE). Such pmus can only have one event scheduled
3468 * at a time, so we disallow creating events that might conflict, namely:
3470 * 1) cpu-wide events in the presence of per-task events,
3471 * 2) per-task events in the presence of cpu-wide events,
3472 * 3) two matching events on the same context.
3474 * The former two cases are handled in the allocation path (perf_event_alloc(),
3475 * __free_event()), the latter -- before the first perf_install_in_context().
3477 static int exclusive_event_init(struct perf_event *event)
3479 struct pmu *pmu = event->pmu;
3481 if (!(pmu->capabilities & PERF_PMU_CAP_EXCLUSIVE))
3485 * Prevent co-existence of per-task and cpu-wide events on the
3486 * same exclusive pmu.
3488 * Negative pmu::exclusive_cnt means there are cpu-wide
3489 * events on this "exclusive" pmu, positive means there are
3492 * Since this is called in perf_event_alloc() path, event::ctx
3493 * doesn't exist yet; it is, however, safe to use PERF_ATTACH_TASK
3494 * to mean "per-task event", because unlike other attach states it
3495 * never gets cleared.
3497 if (event->attach_state & PERF_ATTACH_TASK) {
3498 if (!atomic_inc_unless_negative(&pmu->exclusive_cnt))
3501 if (!atomic_dec_unless_positive(&pmu->exclusive_cnt))
3508 static void exclusive_event_destroy(struct perf_event *event)
3510 struct pmu *pmu = event->pmu;
3512 if (!(pmu->capabilities & PERF_PMU_CAP_EXCLUSIVE))
3515 /* see comment in exclusive_event_init() */
3516 if (event->attach_state & PERF_ATTACH_TASK)
3517 atomic_dec(&pmu->exclusive_cnt);
3519 atomic_inc(&pmu->exclusive_cnt);
3522 static bool exclusive_event_match(struct perf_event *e1, struct perf_event *e2)
3524 if ((e1->pmu->capabilities & PERF_PMU_CAP_EXCLUSIVE) &&
3525 (e1->cpu == e2->cpu ||
3532 /* Called under the same ctx::mutex as perf_install_in_context() */
3533 static bool exclusive_event_installable(struct perf_event *event,
3534 struct perf_event_context *ctx)
3536 struct perf_event *iter_event;
3537 struct pmu *pmu = event->pmu;
3539 if (!(pmu->capabilities & PERF_PMU_CAP_EXCLUSIVE))
3542 list_for_each_entry(iter_event, &ctx->event_list, event_entry) {
3543 if (exclusive_event_match(iter_event, event))
3550 static void __free_event(struct perf_event *event)
3552 if (!event->parent) {
3553 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN)
3554 put_callchain_buffers();
3557 perf_event_free_bpf_prog(event);
3560 event->destroy(event);
3563 put_ctx(event->ctx);
3566 exclusive_event_destroy(event);
3567 module_put(event->pmu->module);
3570 call_rcu(&event->rcu_head, free_event_rcu);
3573 static void _free_event(struct perf_event *event)
3575 irq_work_sync(&event->pending);
3577 unaccount_event(event);
3581 * Can happen when we close an event with re-directed output.
3583 * Since we have a 0 refcount, perf_mmap_close() will skip
3584 * over us; possibly making our ring_buffer_put() the last.
3586 mutex_lock(&event->mmap_mutex);
3587 ring_buffer_attach(event, NULL);
3588 mutex_unlock(&event->mmap_mutex);
3591 if (is_cgroup_event(event))
3592 perf_detach_cgroup(event);
3594 __free_event(event);
3598 * Used to free events which have a known refcount of 1, such as in error paths
3599 * where the event isn't exposed yet and inherited events.
3601 static void free_event(struct perf_event *event)
3603 if (WARN(atomic_long_cmpxchg(&event->refcount, 1, 0) != 1,
3604 "unexpected event refcount: %ld; ptr=%p\n",
3605 atomic_long_read(&event->refcount), event)) {
3606 /* leak to avoid use-after-free */
3614 * Remove user event from the owner task.
3616 static void perf_remove_from_owner(struct perf_event *event)
3618 struct task_struct *owner;
3621 owner = ACCESS_ONCE(event->owner);
3623 * Matches the smp_wmb() in perf_event_exit_task(). If we observe
3624 * !owner it means the list deletion is complete and we can indeed
3625 * free this event, otherwise we need to serialize on
3626 * owner->perf_event_mutex.
3628 smp_read_barrier_depends();
3631 * Since delayed_put_task_struct() also drops the last
3632 * task reference we can safely take a new reference
3633 * while holding the rcu_read_lock().
3635 get_task_struct(owner);
3641 * If we're here through perf_event_exit_task() we're already
3642 * holding ctx->mutex which would be an inversion wrt. the
3643 * normal lock order.
3645 * However we can safely take this lock because its the child
3648 mutex_lock_nested(&owner->perf_event_mutex, SINGLE_DEPTH_NESTING);
3651 * We have to re-check the event->owner field, if it is cleared
3652 * we raced with perf_event_exit_task(), acquiring the mutex
3653 * ensured they're done, and we can proceed with freeing the
3657 list_del_init(&event->owner_entry);
3658 mutex_unlock(&owner->perf_event_mutex);
3659 put_task_struct(owner);
3663 static void put_event(struct perf_event *event)
3665 struct perf_event_context *ctx;
3667 if (!atomic_long_dec_and_test(&event->refcount))
3670 if (!is_kernel_event(event))
3671 perf_remove_from_owner(event);
3674 * There are two ways this annotation is useful:
3676 * 1) there is a lock recursion from perf_event_exit_task
3677 * see the comment there.
3679 * 2) there is a lock-inversion with mmap_sem through
3680 * perf_event_read_group(), which takes faults while
3681 * holding ctx->mutex, however this is called after
3682 * the last filedesc died, so there is no possibility
3683 * to trigger the AB-BA case.
3685 ctx = perf_event_ctx_lock_nested(event, SINGLE_DEPTH_NESTING);
3686 WARN_ON_ONCE(ctx->parent_ctx);
3687 perf_remove_from_context(event, true);
3688 perf_event_ctx_unlock(event, ctx);
3693 int perf_event_release_kernel(struct perf_event *event)
3698 EXPORT_SYMBOL_GPL(perf_event_release_kernel);
3701 * Called when the last reference to the file is gone.
3703 static int perf_release(struct inode *inode, struct file *file)
3705 put_event(file->private_data);
3710 * Remove all orphanes events from the context.
3712 static void orphans_remove_work(struct work_struct *work)
3714 struct perf_event_context *ctx;
3715 struct perf_event *event, *tmp;
3717 ctx = container_of(work, struct perf_event_context,
3718 orphans_remove.work);
3720 mutex_lock(&ctx->mutex);
3721 list_for_each_entry_safe(event, tmp, &ctx->event_list, event_entry) {
3722 struct perf_event *parent_event = event->parent;
3724 if (!is_orphaned_child(event))
3727 perf_remove_from_context(event, true);
3729 mutex_lock(&parent_event->child_mutex);
3730 list_del_init(&event->child_list);
3731 mutex_unlock(&parent_event->child_mutex);
3734 put_event(parent_event);
3737 raw_spin_lock_irq(&ctx->lock);
3738 ctx->orphans_remove_sched = false;
3739 raw_spin_unlock_irq(&ctx->lock);
3740 mutex_unlock(&ctx->mutex);
3745 u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running)
3747 struct perf_event *child;
3753 mutex_lock(&event->child_mutex);
3754 total += perf_event_read(event);
3755 *enabled += event->total_time_enabled +
3756 atomic64_read(&event->child_total_time_enabled);
3757 *running += event->total_time_running +
3758 atomic64_read(&event->child_total_time_running);
3760 list_for_each_entry(child, &event->child_list, child_list) {
3761 total += perf_event_read(child);
3762 *enabled += child->total_time_enabled;
3763 *running += child->total_time_running;
3765 mutex_unlock(&event->child_mutex);
3769 EXPORT_SYMBOL_GPL(perf_event_read_value);
3771 static int perf_event_read_group(struct perf_event *event,
3772 u64 read_format, char __user *buf)
3774 struct perf_event *leader = event->group_leader, *sub;
3775 struct perf_event_context *ctx = leader->ctx;
3776 int n = 0, size = 0, ret;
3777 u64 count, enabled, running;
3780 lockdep_assert_held(&ctx->mutex);
3782 count = perf_event_read_value(leader, &enabled, &running);
3784 values[n++] = 1 + leader->nr_siblings;
3785 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3786 values[n++] = enabled;
3787 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3788 values[n++] = running;
3789 values[n++] = count;
3790 if (read_format & PERF_FORMAT_ID)
3791 values[n++] = primary_event_id(leader);
3793 size = n * sizeof(u64);
3795 if (copy_to_user(buf, values, size))
3800 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
3803 values[n++] = perf_event_read_value(sub, &enabled, &running);
3804 if (read_format & PERF_FORMAT_ID)
3805 values[n++] = primary_event_id(sub);
3807 size = n * sizeof(u64);
3809 if (copy_to_user(buf + ret, values, size)) {
3819 static int perf_event_read_one(struct perf_event *event,
3820 u64 read_format, char __user *buf)
3822 u64 enabled, running;
3826 values[n++] = perf_event_read_value(event, &enabled, &running);
3827 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3828 values[n++] = enabled;
3829 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3830 values[n++] = running;
3831 if (read_format & PERF_FORMAT_ID)
3832 values[n++] = primary_event_id(event);
3834 if (copy_to_user(buf, values, n * sizeof(u64)))
3837 return n * sizeof(u64);
3840 static bool is_event_hup(struct perf_event *event)
3844 if (event->state != PERF_EVENT_STATE_EXIT)
3847 mutex_lock(&event->child_mutex);
3848 no_children = list_empty(&event->child_list);
3849 mutex_unlock(&event->child_mutex);
3854 * Read the performance event - simple non blocking version for now
3857 perf_read_hw(struct perf_event *event, char __user *buf, size_t count)
3859 u64 read_format = event->attr.read_format;
3863 * Return end-of-file for a read on a event that is in
3864 * error state (i.e. because it was pinned but it couldn't be
3865 * scheduled on to the CPU at some point).
3867 if (event->state == PERF_EVENT_STATE_ERROR)
3870 if (count < event->read_size)
3873 WARN_ON_ONCE(event->ctx->parent_ctx);
3874 if (read_format & PERF_FORMAT_GROUP)
3875 ret = perf_event_read_group(event, read_format, buf);
3877 ret = perf_event_read_one(event, read_format, buf);
3883 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
3885 struct perf_event *event = file->private_data;
3886 struct perf_event_context *ctx;
3889 ctx = perf_event_ctx_lock(event);
3890 ret = perf_read_hw(event, buf, count);
3891 perf_event_ctx_unlock(event, ctx);
3896 static unsigned int perf_poll(struct file *file, poll_table *wait)
3898 struct perf_event *event = file->private_data;
3899 struct ring_buffer *rb;
3900 unsigned int events = POLLHUP;
3902 poll_wait(file, &event->waitq, wait);
3904 if (is_event_hup(event))
3908 * Pin the event->rb by taking event->mmap_mutex; otherwise
3909 * perf_event_set_output() can swizzle our rb and make us miss wakeups.
3911 mutex_lock(&event->mmap_mutex);
3914 events = atomic_xchg(&rb->poll, 0);
3915 mutex_unlock(&event->mmap_mutex);
3919 static void _perf_event_reset(struct perf_event *event)
3921 (void)perf_event_read(event);
3922 local64_set(&event->count, 0);
3923 perf_event_update_userpage(event);
3927 * Holding the top-level event's child_mutex means that any
3928 * descendant process that has inherited this event will block
3929 * in sync_child_event if it goes to exit, thus satisfying the
3930 * task existence requirements of perf_event_enable/disable.
3932 static void perf_event_for_each_child(struct perf_event *event,
3933 void (*func)(struct perf_event *))
3935 struct perf_event *child;
3937 WARN_ON_ONCE(event->ctx->parent_ctx);
3939 mutex_lock(&event->child_mutex);
3941 list_for_each_entry(child, &event->child_list, child_list)
3943 mutex_unlock(&event->child_mutex);
3946 static void perf_event_for_each(struct perf_event *event,
3947 void (*func)(struct perf_event *))
3949 struct perf_event_context *ctx = event->ctx;
3950 struct perf_event *sibling;
3952 lockdep_assert_held(&ctx->mutex);
3954 event = event->group_leader;
3956 perf_event_for_each_child(event, func);
3957 list_for_each_entry(sibling, &event->sibling_list, group_entry)
3958 perf_event_for_each_child(sibling, func);
3961 static int perf_event_period(struct perf_event *event, u64 __user *arg)
3963 struct perf_event_context *ctx = event->ctx;
3964 int ret = 0, active;
3967 if (!is_sampling_event(event))
3970 if (copy_from_user(&value, arg, sizeof(value)))
3976 raw_spin_lock_irq(&ctx->lock);
3977 if (event->attr.freq) {
3978 if (value > sysctl_perf_event_sample_rate) {
3983 event->attr.sample_freq = value;
3985 event->attr.sample_period = value;
3986 event->hw.sample_period = value;
3989 active = (event->state == PERF_EVENT_STATE_ACTIVE);
3991 perf_pmu_disable(ctx->pmu);
3992 event->pmu->stop(event, PERF_EF_UPDATE);
3995 local64_set(&event->hw.period_left, 0);
3998 event->pmu->start(event, PERF_EF_RELOAD);
3999 perf_pmu_enable(ctx->pmu);
4003 raw_spin_unlock_irq(&ctx->lock);
4008 static const struct file_operations perf_fops;
4010 static inline int perf_fget_light(int fd, struct fd *p)
4012 struct fd f = fdget(fd);
4016 if (f.file->f_op != &perf_fops) {
4024 static int perf_event_set_output(struct perf_event *event,
4025 struct perf_event *output_event);
4026 static int perf_event_set_filter(struct perf_event *event, void __user *arg);
4027 static int perf_event_set_bpf_prog(struct perf_event *event, u32 prog_fd);
4029 static long _perf_ioctl(struct perf_event *event, unsigned int cmd, unsigned long arg)
4031 void (*func)(struct perf_event *);
4035 case PERF_EVENT_IOC_ENABLE:
4036 func = _perf_event_enable;
4038 case PERF_EVENT_IOC_DISABLE:
4039 func = _perf_event_disable;
4041 case PERF_EVENT_IOC_RESET:
4042 func = _perf_event_reset;
4045 case PERF_EVENT_IOC_REFRESH:
4046 return _perf_event_refresh(event, arg);
4048 case PERF_EVENT_IOC_PERIOD:
4049 return perf_event_period(event, (u64 __user *)arg);
4051 case PERF_EVENT_IOC_ID:
4053 u64 id = primary_event_id(event);
4055 if (copy_to_user((void __user *)arg, &id, sizeof(id)))
4060 case PERF_EVENT_IOC_SET_OUTPUT:
4064 struct perf_event *output_event;
4066 ret = perf_fget_light(arg, &output);
4069 output_event = output.file->private_data;
4070 ret = perf_event_set_output(event, output_event);
4073 ret = perf_event_set_output(event, NULL);
4078 case PERF_EVENT_IOC_SET_FILTER:
4079 return perf_event_set_filter(event, (void __user *)arg);
4081 case PERF_EVENT_IOC_SET_BPF:
4082 return perf_event_set_bpf_prog(event, arg);
4088 if (flags & PERF_IOC_FLAG_GROUP)
4089 perf_event_for_each(event, func);
4091 perf_event_for_each_child(event, func);
4096 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
4098 struct perf_event *event = file->private_data;
4099 struct perf_event_context *ctx;
4102 ctx = perf_event_ctx_lock(event);
4103 ret = _perf_ioctl(event, cmd, arg);
4104 perf_event_ctx_unlock(event, ctx);
4109 #ifdef CONFIG_COMPAT
4110 static long perf_compat_ioctl(struct file *file, unsigned int cmd,
4113 switch (_IOC_NR(cmd)) {
4114 case _IOC_NR(PERF_EVENT_IOC_SET_FILTER):
4115 case _IOC_NR(PERF_EVENT_IOC_ID):
4116 /* Fix up pointer size (usually 4 -> 8 in 32-on-64-bit case */
4117 if (_IOC_SIZE(cmd) == sizeof(compat_uptr_t)) {
4118 cmd &= ~IOCSIZE_MASK;
4119 cmd |= sizeof(void *) << IOCSIZE_SHIFT;
4123 return perf_ioctl(file, cmd, arg);
4126 # define perf_compat_ioctl NULL
4129 int perf_event_task_enable(void)
4131 struct perf_event_context *ctx;
4132 struct perf_event *event;
4134 mutex_lock(¤t->perf_event_mutex);
4135 list_for_each_entry(event, ¤t->perf_event_list, owner_entry) {
4136 ctx = perf_event_ctx_lock(event);
4137 perf_event_for_each_child(event, _perf_event_enable);
4138 perf_event_ctx_unlock(event, ctx);
4140 mutex_unlock(¤t->perf_event_mutex);
4145 int perf_event_task_disable(void)
4147 struct perf_event_context *ctx;
4148 struct perf_event *event;
4150 mutex_lock(¤t->perf_event_mutex);
4151 list_for_each_entry(event, ¤t->perf_event_list, owner_entry) {
4152 ctx = perf_event_ctx_lock(event);
4153 perf_event_for_each_child(event, _perf_event_disable);
4154 perf_event_ctx_unlock(event, ctx);
4156 mutex_unlock(¤t->perf_event_mutex);
4161 static int perf_event_index(struct perf_event *event)
4163 if (event->hw.state & PERF_HES_STOPPED)
4166 if (event->state != PERF_EVENT_STATE_ACTIVE)
4169 return event->pmu->event_idx(event);
4172 static void calc_timer_values(struct perf_event *event,
4179 *now = perf_clock();
4180 ctx_time = event->shadow_ctx_time + *now;
4181 *enabled = ctx_time - event->tstamp_enabled;
4182 *running = ctx_time - event->tstamp_running;
4185 static void perf_event_init_userpage(struct perf_event *event)
4187 struct perf_event_mmap_page *userpg;
4188 struct ring_buffer *rb;
4191 rb = rcu_dereference(event->rb);
4195 userpg = rb->user_page;
4197 /* Allow new userspace to detect that bit 0 is deprecated */
4198 userpg->cap_bit0_is_deprecated = 1;
4199 userpg->size = offsetof(struct perf_event_mmap_page, __reserved);
4200 userpg->data_offset = PAGE_SIZE;
4201 userpg->data_size = perf_data_size(rb);
4207 void __weak arch_perf_update_userpage(
4208 struct perf_event *event, struct perf_event_mmap_page *userpg, u64 now)
4213 * Callers need to ensure there can be no nesting of this function, otherwise
4214 * the seqlock logic goes bad. We can not serialize this because the arch
4215 * code calls this from NMI context.
4217 void perf_event_update_userpage(struct perf_event *event)
4219 struct perf_event_mmap_page *userpg;
4220 struct ring_buffer *rb;
4221 u64 enabled, running, now;
4224 rb = rcu_dereference(event->rb);
4229 * compute total_time_enabled, total_time_running
4230 * based on snapshot values taken when the event
4231 * was last scheduled in.
4233 * we cannot simply called update_context_time()
4234 * because of locking issue as we can be called in
4237 calc_timer_values(event, &now, &enabled, &running);
4239 userpg = rb->user_page;
4241 * Disable preemption so as to not let the corresponding user-space
4242 * spin too long if we get preempted.
4247 userpg->index = perf_event_index(event);
4248 userpg->offset = perf_event_count(event);
4250 userpg->offset -= local64_read(&event->hw.prev_count);
4252 userpg->time_enabled = enabled +
4253 atomic64_read(&event->child_total_time_enabled);
4255 userpg->time_running = running +
4256 atomic64_read(&event->child_total_time_running);
4258 arch_perf_update_userpage(event, userpg, now);
4267 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
4269 struct perf_event *event = vma->vm_file->private_data;
4270 struct ring_buffer *rb;
4271 int ret = VM_FAULT_SIGBUS;
4273 if (vmf->flags & FAULT_FLAG_MKWRITE) {
4274 if (vmf->pgoff == 0)
4280 rb = rcu_dereference(event->rb);
4284 if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE))
4287 vmf->page = perf_mmap_to_page(rb, vmf->pgoff);
4291 get_page(vmf->page);
4292 vmf->page->mapping = vma->vm_file->f_mapping;
4293 vmf->page->index = vmf->pgoff;
4302 static void ring_buffer_attach(struct perf_event *event,
4303 struct ring_buffer *rb)
4305 struct ring_buffer *old_rb = NULL;
4306 unsigned long flags;
4310 * Should be impossible, we set this when removing
4311 * event->rb_entry and wait/clear when adding event->rb_entry.
4313 WARN_ON_ONCE(event->rcu_pending);
4316 spin_lock_irqsave(&old_rb->event_lock, flags);
4317 list_del_rcu(&event->rb_entry);
4318 spin_unlock_irqrestore(&old_rb->event_lock, flags);
4320 event->rcu_batches = get_state_synchronize_rcu();
4321 event->rcu_pending = 1;
4325 if (event->rcu_pending) {
4326 cond_synchronize_rcu(event->rcu_batches);
4327 event->rcu_pending = 0;
4330 spin_lock_irqsave(&rb->event_lock, flags);
4331 list_add_rcu(&event->rb_entry, &rb->event_list);
4332 spin_unlock_irqrestore(&rb->event_lock, flags);
4335 rcu_assign_pointer(event->rb, rb);
4338 ring_buffer_put(old_rb);
4340 * Since we detached before setting the new rb, so that we
4341 * could attach the new rb, we could have missed a wakeup.
4344 wake_up_all(&event->waitq);
4348 static void ring_buffer_wakeup(struct perf_event *event)
4350 struct ring_buffer *rb;
4353 rb = rcu_dereference(event->rb);
4355 list_for_each_entry_rcu(event, &rb->event_list, rb_entry)
4356 wake_up_all(&event->waitq);
4361 struct ring_buffer *ring_buffer_get(struct perf_event *event)
4363 struct ring_buffer *rb;
4366 rb = rcu_dereference(event->rb);
4368 if (!atomic_inc_not_zero(&rb->refcount))
4376 void ring_buffer_put(struct ring_buffer *rb)
4378 if (!atomic_dec_and_test(&rb->refcount))
4381 WARN_ON_ONCE(!list_empty(&rb->event_list));
4383 call_rcu(&rb->rcu_head, rb_free_rcu);
4386 static void perf_mmap_open(struct vm_area_struct *vma)
4388 struct perf_event *event = vma->vm_file->private_data;
4390 atomic_inc(&event->mmap_count);
4391 atomic_inc(&event->rb->mmap_count);
4394 atomic_inc(&event->rb->aux_mmap_count);
4396 if (event->pmu->event_mapped)
4397 event->pmu->event_mapped(event);
4401 * A buffer can be mmap()ed multiple times; either directly through the same
4402 * event, or through other events by use of perf_event_set_output().
4404 * In order to undo the VM accounting done by perf_mmap() we need to destroy
4405 * the buffer here, where we still have a VM context. This means we need
4406 * to detach all events redirecting to us.
4408 static void perf_mmap_close(struct vm_area_struct *vma)
4410 struct perf_event *event = vma->vm_file->private_data;
4412 struct ring_buffer *rb = ring_buffer_get(event);
4413 struct user_struct *mmap_user = rb->mmap_user;
4414 int mmap_locked = rb->mmap_locked;
4415 unsigned long size = perf_data_size(rb);
4417 if (event->pmu->event_unmapped)
4418 event->pmu->event_unmapped(event);
4421 * rb->aux_mmap_count will always drop before rb->mmap_count and
4422 * event->mmap_count, so it is ok to use event->mmap_mutex to
4423 * serialize with perf_mmap here.
4425 if (rb_has_aux(rb) && vma->vm_pgoff == rb->aux_pgoff &&
4426 atomic_dec_and_mutex_lock(&rb->aux_mmap_count, &event->mmap_mutex)) {
4427 atomic_long_sub(rb->aux_nr_pages, &mmap_user->locked_vm);
4428 vma->vm_mm->pinned_vm -= rb->aux_mmap_locked;
4431 mutex_unlock(&event->mmap_mutex);
4434 atomic_dec(&rb->mmap_count);
4436 if (!atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex))
4439 ring_buffer_attach(event, NULL);
4440 mutex_unlock(&event->mmap_mutex);
4442 /* If there's still other mmap()s of this buffer, we're done. */
4443 if (atomic_read(&rb->mmap_count))
4447 * No other mmap()s, detach from all other events that might redirect
4448 * into the now unreachable buffer. Somewhat complicated by the
4449 * fact that rb::event_lock otherwise nests inside mmap_mutex.
4453 list_for_each_entry_rcu(event, &rb->event_list, rb_entry) {
4454 if (!atomic_long_inc_not_zero(&event->refcount)) {
4456 * This event is en-route to free_event() which will
4457 * detach it and remove it from the list.
4463 mutex_lock(&event->mmap_mutex);
4465 * Check we didn't race with perf_event_set_output() which can
4466 * swizzle the rb from under us while we were waiting to
4467 * acquire mmap_mutex.
4469 * If we find a different rb; ignore this event, a next
4470 * iteration will no longer find it on the list. We have to
4471 * still restart the iteration to make sure we're not now
4472 * iterating the wrong list.
4474 if (event->rb == rb)
4475 ring_buffer_attach(event, NULL);
4477 mutex_unlock(&event->mmap_mutex);
4481 * Restart the iteration; either we're on the wrong list or
4482 * destroyed its integrity by doing a deletion.
4489 * It could be there's still a few 0-ref events on the list; they'll
4490 * get cleaned up by free_event() -- they'll also still have their
4491 * ref on the rb and will free it whenever they are done with it.
4493 * Aside from that, this buffer is 'fully' detached and unmapped,
4494 * undo the VM accounting.
4497 atomic_long_sub((size >> PAGE_SHIFT) + 1, &mmap_user->locked_vm);
4498 vma->vm_mm->pinned_vm -= mmap_locked;
4499 free_uid(mmap_user);
4502 ring_buffer_put(rb); /* could be last */
4505 static const struct vm_operations_struct perf_mmap_vmops = {
4506 .open = perf_mmap_open,
4507 .close = perf_mmap_close, /* non mergable */
4508 .fault = perf_mmap_fault,
4509 .page_mkwrite = perf_mmap_fault,
4512 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
4514 struct perf_event *event = file->private_data;
4515 unsigned long user_locked, user_lock_limit;
4516 struct user_struct *user = current_user();
4517 unsigned long locked, lock_limit;
4518 struct ring_buffer *rb = NULL;
4519 unsigned long vma_size;
4520 unsigned long nr_pages;
4521 long user_extra = 0, extra = 0;
4522 int ret = 0, flags = 0;
4525 * Don't allow mmap() of inherited per-task counters. This would
4526 * create a performance issue due to all children writing to the
4529 if (event->cpu == -1 && event->attr.inherit)
4532 if (!(vma->vm_flags & VM_SHARED))
4535 vma_size = vma->vm_end - vma->vm_start;
4537 if (vma->vm_pgoff == 0) {
4538 nr_pages = (vma_size / PAGE_SIZE) - 1;
4541 * AUX area mapping: if rb->aux_nr_pages != 0, it's already
4542 * mapped, all subsequent mappings should have the same size
4543 * and offset. Must be above the normal perf buffer.
4545 u64 aux_offset, aux_size;
4550 nr_pages = vma_size / PAGE_SIZE;
4552 mutex_lock(&event->mmap_mutex);
4559 aux_offset = ACCESS_ONCE(rb->user_page->aux_offset);
4560 aux_size = ACCESS_ONCE(rb->user_page->aux_size);
4562 if (aux_offset < perf_data_size(rb) + PAGE_SIZE)
4565 if (aux_offset != vma->vm_pgoff << PAGE_SHIFT)
4568 /* already mapped with a different offset */
4569 if (rb_has_aux(rb) && rb->aux_pgoff != vma->vm_pgoff)
4572 if (aux_size != vma_size || aux_size != nr_pages * PAGE_SIZE)
4575 /* already mapped with a different size */
4576 if (rb_has_aux(rb) && rb->aux_nr_pages != nr_pages)
4579 if (!is_power_of_2(nr_pages))
4582 if (!atomic_inc_not_zero(&rb->mmap_count))
4585 if (rb_has_aux(rb)) {
4586 atomic_inc(&rb->aux_mmap_count);
4591 atomic_set(&rb->aux_mmap_count, 1);
4592 user_extra = nr_pages;
4598 * If we have rb pages ensure they're a power-of-two number, so we
4599 * can do bitmasks instead of modulo.
4601 if (nr_pages != 0 && !is_power_of_2(nr_pages))
4604 if (vma_size != PAGE_SIZE * (1 + nr_pages))
4607 WARN_ON_ONCE(event->ctx->parent_ctx);
4609 mutex_lock(&event->mmap_mutex);
4611 if (event->rb->nr_pages != nr_pages) {
4616 if (!atomic_inc_not_zero(&event->rb->mmap_count)) {
4618 * Raced against perf_mmap_close() through
4619 * perf_event_set_output(). Try again, hope for better
4622 mutex_unlock(&event->mmap_mutex);
4629 user_extra = nr_pages + 1;
4632 user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10);
4635 * Increase the limit linearly with more CPUs:
4637 user_lock_limit *= num_online_cpus();
4639 user_locked = atomic_long_read(&user->locked_vm) + user_extra;
4641 if (user_locked > user_lock_limit)
4642 extra = user_locked - user_lock_limit;
4644 lock_limit = rlimit(RLIMIT_MEMLOCK);
4645 lock_limit >>= PAGE_SHIFT;
4646 locked = vma->vm_mm->pinned_vm + extra;
4648 if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() &&
4649 !capable(CAP_IPC_LOCK)) {
4654 WARN_ON(!rb && event->rb);
4656 if (vma->vm_flags & VM_WRITE)
4657 flags |= RING_BUFFER_WRITABLE;
4660 rb = rb_alloc(nr_pages,
4661 event->attr.watermark ? event->attr.wakeup_watermark : 0,
4669 atomic_set(&rb->mmap_count, 1);
4670 rb->mmap_user = get_current_user();
4671 rb->mmap_locked = extra;
4673 ring_buffer_attach(event, rb);
4675 perf_event_init_userpage(event);
4676 perf_event_update_userpage(event);
4678 ret = rb_alloc_aux(rb, event, vma->vm_pgoff, nr_pages,
4679 event->attr.aux_watermark, flags);
4681 rb->aux_mmap_locked = extra;
4686 atomic_long_add(user_extra, &user->locked_vm);
4687 vma->vm_mm->pinned_vm += extra;
4689 atomic_inc(&event->mmap_count);
4691 atomic_dec(&rb->mmap_count);
4694 mutex_unlock(&event->mmap_mutex);
4697 * Since pinned accounting is per vm we cannot allow fork() to copy our
4700 vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND | VM_DONTDUMP;
4701 vma->vm_ops = &perf_mmap_vmops;
4703 if (event->pmu->event_mapped)
4704 event->pmu->event_mapped(event);
4709 static int perf_fasync(int fd, struct file *filp, int on)
4711 struct inode *inode = file_inode(filp);
4712 struct perf_event *event = filp->private_data;
4715 mutex_lock(&inode->i_mutex);
4716 retval = fasync_helper(fd, filp, on, &event->fasync);
4717 mutex_unlock(&inode->i_mutex);
4725 static const struct file_operations perf_fops = {
4726 .llseek = no_llseek,
4727 .release = perf_release,
4730 .unlocked_ioctl = perf_ioctl,
4731 .compat_ioctl = perf_compat_ioctl,
4733 .fasync = perf_fasync,
4739 * If there's data, ensure we set the poll() state and publish everything
4740 * to user-space before waking everybody up.
4743 void perf_event_wakeup(struct perf_event *event)
4745 ring_buffer_wakeup(event);
4747 if (event->pending_kill) {
4748 kill_fasync(&event->fasync, SIGIO, event->pending_kill);
4749 event->pending_kill = 0;
4753 static void perf_pending_event(struct irq_work *entry)
4755 struct perf_event *event = container_of(entry,
4756 struct perf_event, pending);
4759 rctx = perf_swevent_get_recursion_context();
4761 * If we 'fail' here, that's OK, it means recursion is already disabled
4762 * and we won't recurse 'further'.
4765 if (event->pending_disable) {
4766 event->pending_disable = 0;
4767 __perf_event_disable(event);
4770 if (event->pending_wakeup) {
4771 event->pending_wakeup = 0;
4772 perf_event_wakeup(event);
4776 perf_swevent_put_recursion_context(rctx);
4780 * We assume there is only KVM supporting the callbacks.
4781 * Later on, we might change it to a list if there is
4782 * another virtualization implementation supporting the callbacks.
4784 struct perf_guest_info_callbacks *perf_guest_cbs;
4786 int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
4788 perf_guest_cbs = cbs;
4791 EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks);
4793 int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
4795 perf_guest_cbs = NULL;
4798 EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks);
4801 perf_output_sample_regs(struct perf_output_handle *handle,
4802 struct pt_regs *regs, u64 mask)
4806 for_each_set_bit(bit, (const unsigned long *) &mask,
4807 sizeof(mask) * BITS_PER_BYTE) {
4810 val = perf_reg_value(regs, bit);
4811 perf_output_put(handle, val);
4815 static void perf_sample_regs_user(struct perf_regs *regs_user,
4816 struct pt_regs *regs,
4817 struct pt_regs *regs_user_copy)
4819 if (user_mode(regs)) {
4820 regs_user->abi = perf_reg_abi(current);
4821 regs_user->regs = regs;
4822 } else if (current->mm) {
4823 perf_get_regs_user(regs_user, regs, regs_user_copy);
4825 regs_user->abi = PERF_SAMPLE_REGS_ABI_NONE;
4826 regs_user->regs = NULL;
4830 static void perf_sample_regs_intr(struct perf_regs *regs_intr,
4831 struct pt_regs *regs)
4833 regs_intr->regs = regs;
4834 regs_intr->abi = perf_reg_abi(current);
4839 * Get remaining task size from user stack pointer.
4841 * It'd be better to take stack vma map and limit this more
4842 * precisly, but there's no way to get it safely under interrupt,
4843 * so using TASK_SIZE as limit.
4845 static u64 perf_ustack_task_size(struct pt_regs *regs)
4847 unsigned long addr = perf_user_stack_pointer(regs);
4849 if (!addr || addr >= TASK_SIZE)
4852 return TASK_SIZE - addr;
4856 perf_sample_ustack_size(u16 stack_size, u16 header_size,
4857 struct pt_regs *regs)
4861 /* No regs, no stack pointer, no dump. */
4866 * Check if we fit in with the requested stack size into the:
4868 * If we don't, we limit the size to the TASK_SIZE.
4870 * - remaining sample size
4871 * If we don't, we customize the stack size to
4872 * fit in to the remaining sample size.
4875 task_size = min((u64) USHRT_MAX, perf_ustack_task_size(regs));
4876 stack_size = min(stack_size, (u16) task_size);
4878 /* Current header size plus static size and dynamic size. */
4879 header_size += 2 * sizeof(u64);
4881 /* Do we fit in with the current stack dump size? */
4882 if ((u16) (header_size + stack_size) < header_size) {
4884 * If we overflow the maximum size for the sample,
4885 * we customize the stack dump size to fit in.
4887 stack_size = USHRT_MAX - header_size - sizeof(u64);
4888 stack_size = round_up(stack_size, sizeof(u64));
4895 perf_output_sample_ustack(struct perf_output_handle *handle, u64 dump_size,
4896 struct pt_regs *regs)
4898 /* Case of a kernel thread, nothing to dump */
4901 perf_output_put(handle, size);
4910 * - the size requested by user or the best one we can fit
4911 * in to the sample max size
4913 * - user stack dump data
4915 * - the actual dumped size
4919 perf_output_put(handle, dump_size);
4922 sp = perf_user_stack_pointer(regs);
4923 rem = __output_copy_user(handle, (void *) sp, dump_size);
4924 dyn_size = dump_size - rem;
4926 perf_output_skip(handle, rem);
4929 perf_output_put(handle, dyn_size);
4933 static void __perf_event_header__init_id(struct perf_event_header *header,
4934 struct perf_sample_data *data,
4935 struct perf_event *event)
4937 u64 sample_type = event->attr.sample_type;
4939 data->type = sample_type;
4940 header->size += event->id_header_size;
4942 if (sample_type & PERF_SAMPLE_TID) {
4943 /* namespace issues */
4944 data->tid_entry.pid = perf_event_pid(event, current);
4945 data->tid_entry.tid = perf_event_tid(event, current);
4948 if (sample_type & PERF_SAMPLE_TIME)
4949 data->time = perf_event_clock(event);
4951 if (sample_type & (PERF_SAMPLE_ID | PERF_SAMPLE_IDENTIFIER))
4952 data->id = primary_event_id(event);
4954 if (sample_type & PERF_SAMPLE_STREAM_ID)
4955 data->stream_id = event->id;
4957 if (sample_type & PERF_SAMPLE_CPU) {
4958 data->cpu_entry.cpu = raw_smp_processor_id();
4959 data->cpu_entry.reserved = 0;
4963 void perf_event_header__init_id(struct perf_event_header *header,
4964 struct perf_sample_data *data,
4965 struct perf_event *event)
4967 if (event->attr.sample_id_all)
4968 __perf_event_header__init_id(header, data, event);
4971 static void __perf_event__output_id_sample(struct perf_output_handle *handle,
4972 struct perf_sample_data *data)
4974 u64 sample_type = data->type;
4976 if (sample_type & PERF_SAMPLE_TID)
4977 perf_output_put(handle, data->tid_entry);
4979 if (sample_type & PERF_SAMPLE_TIME)
4980 perf_output_put(handle, data->time);
4982 if (sample_type & PERF_SAMPLE_ID)
4983 perf_output_put(handle, data->id);
4985 if (sample_type & PERF_SAMPLE_STREAM_ID)
4986 perf_output_put(handle, data->stream_id);
4988 if (sample_type & PERF_SAMPLE_CPU)
4989 perf_output_put(handle, data->cpu_entry);
4991 if (sample_type & PERF_SAMPLE_IDENTIFIER)
4992 perf_output_put(handle, data->id);
4995 void perf_event__output_id_sample(struct perf_event *event,
4996 struct perf_output_handle *handle,
4997 struct perf_sample_data *sample)
4999 if (event->attr.sample_id_all)
5000 __perf_event__output_id_sample(handle, sample);
5003 static void perf_output_read_one(struct perf_output_handle *handle,
5004 struct perf_event *event,
5005 u64 enabled, u64 running)
5007 u64 read_format = event->attr.read_format;
5011 values[n++] = perf_event_count(event);
5012 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
5013 values[n++] = enabled +
5014 atomic64_read(&event->child_total_time_enabled);
5016 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
5017 values[n++] = running +
5018 atomic64_read(&event->child_total_time_running);
5020 if (read_format & PERF_FORMAT_ID)
5021 values[n++] = primary_event_id(event);
5023 __output_copy(handle, values, n * sizeof(u64));
5027 * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
5029 static void perf_output_read_group(struct perf_output_handle *handle,
5030 struct perf_event *event,
5031 u64 enabled, u64 running)
5033 struct perf_event *leader = event->group_leader, *sub;
5034 u64 read_format = event->attr.read_format;
5038 values[n++] = 1 + leader->nr_siblings;
5040 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
5041 values[n++] = enabled;
5043 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
5044 values[n++] = running;
5046 if (leader != event)
5047 leader->pmu->read(leader);
5049 values[n++] = perf_event_count(leader);
5050 if (read_format & PERF_FORMAT_ID)
5051 values[n++] = primary_event_id(leader);
5053 __output_copy(handle, values, n * sizeof(u64));
5055 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
5058 if ((sub != event) &&
5059 (sub->state == PERF_EVENT_STATE_ACTIVE))
5060 sub->pmu->read(sub);
5062 values[n++] = perf_event_count(sub);
5063 if (read_format & PERF_FORMAT_ID)
5064 values[n++] = primary_event_id(sub);
5066 __output_copy(handle, values, n * sizeof(u64));
5070 #define PERF_FORMAT_TOTAL_TIMES (PERF_FORMAT_TOTAL_TIME_ENABLED|\
5071 PERF_FORMAT_TOTAL_TIME_RUNNING)
5073 static void perf_output_read(struct perf_output_handle *handle,
5074 struct perf_event *event)
5076 u64 enabled = 0, running = 0, now;
5077 u64 read_format = event->attr.read_format;
5080 * compute total_time_enabled, total_time_running
5081 * based on snapshot values taken when the event
5082 * was last scheduled in.
5084 * we cannot simply called update_context_time()
5085 * because of locking issue as we are called in
5088 if (read_format & PERF_FORMAT_TOTAL_TIMES)
5089 calc_timer_values(event, &now, &enabled, &running);
5091 if (event->attr.read_format & PERF_FORMAT_GROUP)
5092 perf_output_read_group(handle, event, enabled, running);
5094 perf_output_read_one(handle, event, enabled, running);
5097 void perf_output_sample(struct perf_output_handle *handle,
5098 struct perf_event_header *header,
5099 struct perf_sample_data *data,
5100 struct perf_event *event)
5102 u64 sample_type = data->type;
5104 perf_output_put(handle, *header);
5106 if (sample_type & PERF_SAMPLE_IDENTIFIER)
5107 perf_output_put(handle, data->id);
5109 if (sample_type & PERF_SAMPLE_IP)
5110 perf_output_put(handle, data->ip);
5112 if (sample_type & PERF_SAMPLE_TID)
5113 perf_output_put(handle, data->tid_entry);
5115 if (sample_type & PERF_SAMPLE_TIME)
5116 perf_output_put(handle, data->time);
5118 if (sample_type & PERF_SAMPLE_ADDR)
5119 perf_output_put(handle, data->addr);
5121 if (sample_type & PERF_SAMPLE_ID)
5122 perf_output_put(handle, data->id);
5124 if (sample_type & PERF_SAMPLE_STREAM_ID)
5125 perf_output_put(handle, data->stream_id);
5127 if (sample_type & PERF_SAMPLE_CPU)
5128 perf_output_put(handle, data->cpu_entry);
5130 if (sample_type & PERF_SAMPLE_PERIOD)
5131 perf_output_put(handle, data->period);
5133 if (sample_type & PERF_SAMPLE_READ)
5134 perf_output_read(handle, event);
5136 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
5137 if (data->callchain) {
5140 if (data->callchain)
5141 size += data->callchain->nr;
5143 size *= sizeof(u64);
5145 __output_copy(handle, data->callchain, size);
5148 perf_output_put(handle, nr);
5152 if (sample_type & PERF_SAMPLE_RAW) {
5154 perf_output_put(handle, data->raw->size);
5155 __output_copy(handle, data->raw->data,
5162 .size = sizeof(u32),
5165 perf_output_put(handle, raw);
5169 if (sample_type & PERF_SAMPLE_BRANCH_STACK) {
5170 if (data->br_stack) {
5173 size = data->br_stack->nr
5174 * sizeof(struct perf_branch_entry);
5176 perf_output_put(handle, data->br_stack->nr);
5177 perf_output_copy(handle, data->br_stack->entries, size);
5180 * we always store at least the value of nr
5183 perf_output_put(handle, nr);
5187 if (sample_type & PERF_SAMPLE_REGS_USER) {
5188 u64 abi = data->regs_user.abi;
5191 * If there are no regs to dump, notice it through
5192 * first u64 being zero (PERF_SAMPLE_REGS_ABI_NONE).
5194 perf_output_put(handle, abi);
5197 u64 mask = event->attr.sample_regs_user;
5198 perf_output_sample_regs(handle,
5199 data->regs_user.regs,
5204 if (sample_type & PERF_SAMPLE_STACK_USER) {
5205 perf_output_sample_ustack(handle,
5206 data->stack_user_size,
5207 data->regs_user.regs);
5210 if (sample_type & PERF_SAMPLE_WEIGHT)
5211 perf_output_put(handle, data->weight);
5213 if (sample_type & PERF_SAMPLE_DATA_SRC)
5214 perf_output_put(handle, data->data_src.val);
5216 if (sample_type & PERF_SAMPLE_TRANSACTION)
5217 perf_output_put(handle, data->txn);
5219 if (sample_type & PERF_SAMPLE_REGS_INTR) {
5220 u64 abi = data->regs_intr.abi;
5222 * If there are no regs to dump, notice it through
5223 * first u64 being zero (PERF_SAMPLE_REGS_ABI_NONE).
5225 perf_output_put(handle, abi);
5228 u64 mask = event->attr.sample_regs_intr;
5230 perf_output_sample_regs(handle,
5231 data->regs_intr.regs,
5236 if (!event->attr.watermark) {
5237 int wakeup_events = event->attr.wakeup_events;
5239 if (wakeup_events) {
5240 struct ring_buffer *rb = handle->rb;
5241 int events = local_inc_return(&rb->events);
5243 if (events >= wakeup_events) {
5244 local_sub(wakeup_events, &rb->events);
5245 local_inc(&rb->wakeup);
5251 void perf_prepare_sample(struct perf_event_header *header,
5252 struct perf_sample_data *data,
5253 struct perf_event *event,
5254 struct pt_regs *regs)
5256 u64 sample_type = event->attr.sample_type;
5258 header->type = PERF_RECORD_SAMPLE;
5259 header->size = sizeof(*header) + event->header_size;
5262 header->misc |= perf_misc_flags(regs);
5264 __perf_event_header__init_id(header, data, event);
5266 if (sample_type & PERF_SAMPLE_IP)
5267 data->ip = perf_instruction_pointer(regs);
5269 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
5272 data->callchain = perf_callchain(event, regs);
5274 if (data->callchain)
5275 size += data->callchain->nr;
5277 header->size += size * sizeof(u64);
5280 if (sample_type & PERF_SAMPLE_RAW) {
5281 int size = sizeof(u32);
5284 size += data->raw->size;
5286 size += sizeof(u32);
5288 WARN_ON_ONCE(size & (sizeof(u64)-1));
5289 header->size += size;
5292 if (sample_type & PERF_SAMPLE_BRANCH_STACK) {
5293 int size = sizeof(u64); /* nr */
5294 if (data->br_stack) {
5295 size += data->br_stack->nr
5296 * sizeof(struct perf_branch_entry);
5298 header->size += size;
5301 if (sample_type & (PERF_SAMPLE_REGS_USER | PERF_SAMPLE_STACK_USER))
5302 perf_sample_regs_user(&data->regs_user, regs,
5303 &data->regs_user_copy);
5305 if (sample_type & PERF_SAMPLE_REGS_USER) {
5306 /* regs dump ABI info */
5307 int size = sizeof(u64);
5309 if (data->regs_user.regs) {
5310 u64 mask = event->attr.sample_regs_user;
5311 size += hweight64(mask) * sizeof(u64);
5314 header->size += size;
5317 if (sample_type & PERF_SAMPLE_STACK_USER) {
5319 * Either we need PERF_SAMPLE_STACK_USER bit to be allways
5320 * processed as the last one or have additional check added
5321 * in case new sample type is added, because we could eat
5322 * up the rest of the sample size.
5324 u16 stack_size = event->attr.sample_stack_user;
5325 u16 size = sizeof(u64);
5327 stack_size = perf_sample_ustack_size(stack_size, header->size,
5328 data->regs_user.regs);
5331 * If there is something to dump, add space for the dump
5332 * itself and for the field that tells the dynamic size,
5333 * which is how many have been actually dumped.
5336 size += sizeof(u64) + stack_size;
5338 data->stack_user_size = stack_size;
5339 header->size += size;
5342 if (sample_type & PERF_SAMPLE_REGS_INTR) {
5343 /* regs dump ABI info */
5344 int size = sizeof(u64);
5346 perf_sample_regs_intr(&data->regs_intr, regs);
5348 if (data->regs_intr.regs) {
5349 u64 mask = event->attr.sample_regs_intr;
5351 size += hweight64(mask) * sizeof(u64);
5354 header->size += size;
5358 void perf_event_output(struct perf_event *event,
5359 struct perf_sample_data *data,
5360 struct pt_regs *regs)
5362 struct perf_output_handle handle;
5363 struct perf_event_header header;
5365 /* protect the callchain buffers */
5368 perf_prepare_sample(&header, data, event, regs);
5370 if (perf_output_begin(&handle, event, header.size))
5373 perf_output_sample(&handle, &header, data, event);
5375 perf_output_end(&handle);
5385 struct perf_read_event {
5386 struct perf_event_header header;
5393 perf_event_read_event(struct perf_event *event,
5394 struct task_struct *task)
5396 struct perf_output_handle handle;
5397 struct perf_sample_data sample;
5398 struct perf_read_event read_event = {
5400 .type = PERF_RECORD_READ,
5402 .size = sizeof(read_event) + event->read_size,
5404 .pid = perf_event_pid(event, task),
5405 .tid = perf_event_tid(event, task),
5409 perf_event_header__init_id(&read_event.header, &sample, event);
5410 ret = perf_output_begin(&handle, event, read_event.header.size);
5414 perf_output_put(&handle, read_event);
5415 perf_output_read(&handle, event);
5416 perf_event__output_id_sample(event, &handle, &sample);
5418 perf_output_end(&handle);
5421 typedef void (perf_event_aux_output_cb)(struct perf_event *event, void *data);
5424 perf_event_aux_ctx(struct perf_event_context *ctx,
5425 perf_event_aux_output_cb output,
5428 struct perf_event *event;
5430 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
5431 if (event->state < PERF_EVENT_STATE_INACTIVE)
5433 if (!event_filter_match(event))
5435 output(event, data);
5440 perf_event_aux(perf_event_aux_output_cb output, void *data,
5441 struct perf_event_context *task_ctx)
5443 struct perf_cpu_context *cpuctx;
5444 struct perf_event_context *ctx;
5449 list_for_each_entry_rcu(pmu, &pmus, entry) {
5450 cpuctx = get_cpu_ptr(pmu->pmu_cpu_context);
5451 if (cpuctx->unique_pmu != pmu)
5453 perf_event_aux_ctx(&cpuctx->ctx, output, data);
5456 ctxn = pmu->task_ctx_nr;
5459 ctx = rcu_dereference(current->perf_event_ctxp[ctxn]);
5461 perf_event_aux_ctx(ctx, output, data);
5463 put_cpu_ptr(pmu->pmu_cpu_context);
5468 perf_event_aux_ctx(task_ctx, output, data);
5475 * task tracking -- fork/exit
5477 * enabled by: attr.comm | attr.mmap | attr.mmap2 | attr.mmap_data | attr.task
5480 struct perf_task_event {
5481 struct task_struct *task;
5482 struct perf_event_context *task_ctx;
5485 struct perf_event_header header;
5495 static int perf_event_task_match(struct perf_event *event)
5497 return event->attr.comm || event->attr.mmap ||
5498 event->attr.mmap2 || event->attr.mmap_data ||
5502 static void perf_event_task_output(struct perf_event *event,
5505 struct perf_task_event *task_event = data;
5506 struct perf_output_handle handle;
5507 struct perf_sample_data sample;
5508 struct task_struct *task = task_event->task;
5509 int ret, size = task_event->event_id.header.size;
5511 if (!perf_event_task_match(event))
5514 perf_event_header__init_id(&task_event->event_id.header, &sample, event);
5516 ret = perf_output_begin(&handle, event,
5517 task_event->event_id.header.size);
5521 task_event->event_id.pid = perf_event_pid(event, task);
5522 task_event->event_id.ppid = perf_event_pid(event, current);
5524 task_event->event_id.tid = perf_event_tid(event, task);
5525 task_event->event_id.ptid = perf_event_tid(event, current);
5527 task_event->event_id.time = perf_event_clock(event);
5529 perf_output_put(&handle, task_event->event_id);
5531 perf_event__output_id_sample(event, &handle, &sample);
5533 perf_output_end(&handle);
5535 task_event->event_id.header.size = size;
5538 static void perf_event_task(struct task_struct *task,
5539 struct perf_event_context *task_ctx,
5542 struct perf_task_event task_event;
5544 if (!atomic_read(&nr_comm_events) &&
5545 !atomic_read(&nr_mmap_events) &&
5546 !atomic_read(&nr_task_events))
5549 task_event = (struct perf_task_event){
5551 .task_ctx = task_ctx,
5554 .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT,
5556 .size = sizeof(task_event.event_id),
5566 perf_event_aux(perf_event_task_output,
5571 void perf_event_fork(struct task_struct *task)
5573 perf_event_task(task, NULL, 1);
5580 struct perf_comm_event {
5581 struct task_struct *task;
5586 struct perf_event_header header;
5593 static int perf_event_comm_match(struct perf_event *event)
5595 return event->attr.comm;
5598 static void perf_event_comm_output(struct perf_event *event,
5601 struct perf_comm_event *comm_event = data;
5602 struct perf_output_handle handle;
5603 struct perf_sample_data sample;
5604 int size = comm_event->event_id.header.size;
5607 if (!perf_event_comm_match(event))
5610 perf_event_header__init_id(&comm_event->event_id.header, &sample, event);
5611 ret = perf_output_begin(&handle, event,
5612 comm_event->event_id.header.size);
5617 comm_event->event_id.pid = perf_event_pid(event, comm_event->task);
5618 comm_event->event_id.tid = perf_event_tid(event, comm_event->task);
5620 perf_output_put(&handle, comm_event->event_id);
5621 __output_copy(&handle, comm_event->comm,
5622 comm_event->comm_size);
5624 perf_event__output_id_sample(event, &handle, &sample);
5626 perf_output_end(&handle);
5628 comm_event->event_id.header.size = size;
5631 static void perf_event_comm_event(struct perf_comm_event *comm_event)
5633 char comm[TASK_COMM_LEN];
5636 memset(comm, 0, sizeof(comm));
5637 strlcpy(comm, comm_event->task->comm, sizeof(comm));
5638 size = ALIGN(strlen(comm)+1, sizeof(u64));
5640 comm_event->comm = comm;
5641 comm_event->comm_size = size;
5643 comm_event->event_id.header.size = sizeof(comm_event->event_id) + size;
5645 perf_event_aux(perf_event_comm_output,
5650 void perf_event_comm(struct task_struct *task, bool exec)
5652 struct perf_comm_event comm_event;
5654 if (!atomic_read(&nr_comm_events))
5657 comm_event = (struct perf_comm_event){
5663 .type = PERF_RECORD_COMM,
5664 .misc = exec ? PERF_RECORD_MISC_COMM_EXEC : 0,
5672 perf_event_comm_event(&comm_event);
5679 struct perf_mmap_event {
5680 struct vm_area_struct *vma;
5682 const char *file_name;
5690 struct perf_event_header header;
5700 static int perf_event_mmap_match(struct perf_event *event,
5703 struct perf_mmap_event *mmap_event = data;
5704 struct vm_area_struct *vma = mmap_event->vma;
5705 int executable = vma->vm_flags & VM_EXEC;
5707 return (!executable && event->attr.mmap_data) ||
5708 (executable && (event->attr.mmap || event->attr.mmap2));
5711 static void perf_event_mmap_output(struct perf_event *event,
5714 struct perf_mmap_event *mmap_event = data;
5715 struct perf_output_handle handle;
5716 struct perf_sample_data sample;
5717 int size = mmap_event->event_id.header.size;
5720 if (!perf_event_mmap_match(event, data))
5723 if (event->attr.mmap2) {
5724 mmap_event->event_id.header.type = PERF_RECORD_MMAP2;
5725 mmap_event->event_id.header.size += sizeof(mmap_event->maj);
5726 mmap_event->event_id.header.size += sizeof(mmap_event->min);
5727 mmap_event->event_id.header.size += sizeof(mmap_event->ino);
5728 mmap_event->event_id.header.size += sizeof(mmap_event->ino_generation);
5729 mmap_event->event_id.header.size += sizeof(mmap_event->prot);
5730 mmap_event->event_id.header.size += sizeof(mmap_event->flags);
5733 perf_event_header__init_id(&mmap_event->event_id.header, &sample, event);
5734 ret = perf_output_begin(&handle, event,
5735 mmap_event->event_id.header.size);
5739 mmap_event->event_id.pid = perf_event_pid(event, current);
5740 mmap_event->event_id.tid = perf_event_tid(event, current);
5742 perf_output_put(&handle, mmap_event->event_id);
5744 if (event->attr.mmap2) {
5745 perf_output_put(&handle, mmap_event->maj);
5746 perf_output_put(&handle, mmap_event->min);
5747 perf_output_put(&handle, mmap_event->ino);
5748 perf_output_put(&handle, mmap_event->ino_generation);
5749 perf_output_put(&handle, mmap_event->prot);
5750 perf_output_put(&handle, mmap_event->flags);
5753 __output_copy(&handle, mmap_event->file_name,
5754 mmap_event->file_size);
5756 perf_event__output_id_sample(event, &handle, &sample);
5758 perf_output_end(&handle);
5760 mmap_event->event_id.header.size = size;
5763 static void perf_event_mmap_event(struct perf_mmap_event *mmap_event)
5765 struct vm_area_struct *vma = mmap_event->vma;
5766 struct file *file = vma->vm_file;
5767 int maj = 0, min = 0;
5768 u64 ino = 0, gen = 0;
5769 u32 prot = 0, flags = 0;
5776 struct inode *inode;
5779 buf = kmalloc(PATH_MAX, GFP_KERNEL);
5785 * d_path() works from the end of the rb backwards, so we
5786 * need to add enough zero bytes after the string to handle
5787 * the 64bit alignment we do later.
5789 name = file_path(file, buf, PATH_MAX - sizeof(u64));
5794 inode = file_inode(vma->vm_file);
5795 dev = inode->i_sb->s_dev;
5797 gen = inode->i_generation;
5801 if (vma->vm_flags & VM_READ)
5803 if (vma->vm_flags & VM_WRITE)
5805 if (vma->vm_flags & VM_EXEC)
5808 if (vma->vm_flags & VM_MAYSHARE)
5811 flags = MAP_PRIVATE;
5813 if (vma->vm_flags & VM_DENYWRITE)
5814 flags |= MAP_DENYWRITE;
5815 if (vma->vm_flags & VM_MAYEXEC)
5816 flags |= MAP_EXECUTABLE;
5817 if (vma->vm_flags & VM_LOCKED)
5818 flags |= MAP_LOCKED;
5819 if (vma->vm_flags & VM_HUGETLB)
5820 flags |= MAP_HUGETLB;
5824 if (vma->vm_ops && vma->vm_ops->name) {
5825 name = (char *) vma->vm_ops->name(vma);
5830 name = (char *)arch_vma_name(vma);
5834 if (vma->vm_start <= vma->vm_mm->start_brk &&
5835 vma->vm_end >= vma->vm_mm->brk) {
5839 if (vma->vm_start <= vma->vm_mm->start_stack &&
5840 vma->vm_end >= vma->vm_mm->start_stack) {
5850 strlcpy(tmp, name, sizeof(tmp));
5854 * Since our buffer works in 8 byte units we need to align our string
5855 * size to a multiple of 8. However, we must guarantee the tail end is
5856 * zero'd out to avoid leaking random bits to userspace.
5858 size = strlen(name)+1;
5859 while (!IS_ALIGNED(size, sizeof(u64)))
5860 name[size++] = '\0';
5862 mmap_event->file_name = name;
5863 mmap_event->file_size = size;
5864 mmap_event->maj = maj;
5865 mmap_event->min = min;
5866 mmap_event->ino = ino;
5867 mmap_event->ino_generation = gen;
5868 mmap_event->prot = prot;
5869 mmap_event->flags = flags;
5871 if (!(vma->vm_flags & VM_EXEC))
5872 mmap_event->event_id.header.misc |= PERF_RECORD_MISC_MMAP_DATA;
5874 mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size;
5876 perf_event_aux(perf_event_mmap_output,
5883 void perf_event_mmap(struct vm_area_struct *vma)
5885 struct perf_mmap_event mmap_event;
5887 if (!atomic_read(&nr_mmap_events))
5890 mmap_event = (struct perf_mmap_event){
5896 .type = PERF_RECORD_MMAP,
5897 .misc = PERF_RECORD_MISC_USER,
5902 .start = vma->vm_start,
5903 .len = vma->vm_end - vma->vm_start,
5904 .pgoff = (u64)vma->vm_pgoff << PAGE_SHIFT,
5906 /* .maj (attr_mmap2 only) */
5907 /* .min (attr_mmap2 only) */
5908 /* .ino (attr_mmap2 only) */
5909 /* .ino_generation (attr_mmap2 only) */
5910 /* .prot (attr_mmap2 only) */
5911 /* .flags (attr_mmap2 only) */
5914 perf_event_mmap_event(&mmap_event);
5917 void perf_event_aux_event(struct perf_event *event, unsigned long head,
5918 unsigned long size, u64 flags)
5920 struct perf_output_handle handle;
5921 struct perf_sample_data sample;
5922 struct perf_aux_event {
5923 struct perf_event_header header;
5929 .type = PERF_RECORD_AUX,
5931 .size = sizeof(rec),
5939 perf_event_header__init_id(&rec.header, &sample, event);
5940 ret = perf_output_begin(&handle, event, rec.header.size);
5945 perf_output_put(&handle, rec);
5946 perf_event__output_id_sample(event, &handle, &sample);
5948 perf_output_end(&handle);
5952 * Lost/dropped samples logging
5954 void perf_log_lost_samples(struct perf_event *event, u64 lost)
5956 struct perf_output_handle handle;
5957 struct perf_sample_data sample;
5961 struct perf_event_header header;
5963 } lost_samples_event = {
5965 .type = PERF_RECORD_LOST_SAMPLES,
5967 .size = sizeof(lost_samples_event),
5972 perf_event_header__init_id(&lost_samples_event.header, &sample, event);
5974 ret = perf_output_begin(&handle, event,
5975 lost_samples_event.header.size);
5979 perf_output_put(&handle, lost_samples_event);
5980 perf_event__output_id_sample(event, &handle, &sample);
5981 perf_output_end(&handle);
5985 * IRQ throttle logging
5988 static void perf_log_throttle(struct perf_event *event, int enable)
5990 struct perf_output_handle handle;
5991 struct perf_sample_data sample;
5995 struct perf_event_header header;
5999 } throttle_event = {
6001 .type = PERF_RECORD_THROTTLE,
6003 .size = sizeof(throttle_event),
6005 .time = perf_event_clock(event),
6006 .id = primary_event_id(event),
6007 .stream_id = event->id,
6011 throttle_event.header.type = PERF_RECORD_UNTHROTTLE;
6013 perf_event_header__init_id(&throttle_event.header, &sample, event);
6015 ret = perf_output_begin(&handle, event,
6016 throttle_event.header.size);
6020 perf_output_put(&handle, throttle_event);
6021 perf_event__output_id_sample(event, &handle, &sample);
6022 perf_output_end(&handle);
6025 static void perf_log_itrace_start(struct perf_event *event)
6027 struct perf_output_handle handle;
6028 struct perf_sample_data sample;
6029 struct perf_aux_event {
6030 struct perf_event_header header;
6037 event = event->parent;
6039 if (!(event->pmu->capabilities & PERF_PMU_CAP_ITRACE) ||
6040 event->hw.itrace_started)
6043 event->hw.itrace_started = 1;
6045 rec.header.type = PERF_RECORD_ITRACE_START;
6046 rec.header.misc = 0;
6047 rec.header.size = sizeof(rec);
6048 rec.pid = perf_event_pid(event, current);
6049 rec.tid = perf_event_tid(event, current);
6051 perf_event_header__init_id(&rec.header, &sample, event);
6052 ret = perf_output_begin(&handle, event, rec.header.size);
6057 perf_output_put(&handle, rec);
6058 perf_event__output_id_sample(event, &handle, &sample);
6060 perf_output_end(&handle);
6064 * Generic event overflow handling, sampling.
6067 static int __perf_event_overflow(struct perf_event *event,
6068 int throttle, struct perf_sample_data *data,
6069 struct pt_regs *regs)
6071 int events = atomic_read(&event->event_limit);
6072 struct hw_perf_event *hwc = &event->hw;
6077 * Non-sampling counters might still use the PMI to fold short
6078 * hardware counters, ignore those.
6080 if (unlikely(!is_sampling_event(event)))
6083 seq = __this_cpu_read(perf_throttled_seq);
6084 if (seq != hwc->interrupts_seq) {
6085 hwc->interrupts_seq = seq;
6086 hwc->interrupts = 1;
6089 if (unlikely(throttle
6090 && hwc->interrupts >= max_samples_per_tick)) {
6091 __this_cpu_inc(perf_throttled_count);
6092 hwc->interrupts = MAX_INTERRUPTS;
6093 perf_log_throttle(event, 0);
6094 tick_nohz_full_kick();
6099 if (event->attr.freq) {
6100 u64 now = perf_clock();
6101 s64 delta = now - hwc->freq_time_stamp;
6103 hwc->freq_time_stamp = now;
6105 if (delta > 0 && delta < 2*TICK_NSEC)
6106 perf_adjust_period(event, delta, hwc->last_period, true);
6110 * XXX event_limit might not quite work as expected on inherited
6114 event->pending_kill = POLL_IN;
6115 if (events && atomic_dec_and_test(&event->event_limit)) {
6117 event->pending_kill = POLL_HUP;
6118 event->pending_disable = 1;
6119 irq_work_queue(&event->pending);
6122 if (event->overflow_handler)
6123 event->overflow_handler(event, data, regs);
6125 perf_event_output(event, data, regs);
6127 if (event->fasync && event->pending_kill) {
6128 event->pending_wakeup = 1;
6129 irq_work_queue(&event->pending);
6135 int perf_event_overflow(struct perf_event *event,
6136 struct perf_sample_data *data,
6137 struct pt_regs *regs)
6139 return __perf_event_overflow(event, 1, data, regs);
6143 * Generic software event infrastructure
6146 struct swevent_htable {
6147 struct swevent_hlist *swevent_hlist;
6148 struct mutex hlist_mutex;
6151 /* Recursion avoidance in each contexts */
6152 int recursion[PERF_NR_CONTEXTS];
6154 /* Keeps track of cpu being initialized/exited */
6158 static DEFINE_PER_CPU(struct swevent_htable, swevent_htable);
6161 * We directly increment event->count and keep a second value in
6162 * event->hw.period_left to count intervals. This period event
6163 * is kept in the range [-sample_period, 0] so that we can use the
6167 u64 perf_swevent_set_period(struct perf_event *event)
6169 struct hw_perf_event *hwc = &event->hw;
6170 u64 period = hwc->last_period;
6174 hwc->last_period = hwc->sample_period;
6177 old = val = local64_read(&hwc->period_left);
6181 nr = div64_u64(period + val, period);
6182 offset = nr * period;
6184 if (local64_cmpxchg(&hwc->period_left, old, val) != old)
6190 static void perf_swevent_overflow(struct perf_event *event, u64 overflow,
6191 struct perf_sample_data *data,
6192 struct pt_regs *regs)
6194 struct hw_perf_event *hwc = &event->hw;
6198 overflow = perf_swevent_set_period(event);
6200 if (hwc->interrupts == MAX_INTERRUPTS)
6203 for (; overflow; overflow--) {
6204 if (__perf_event_overflow(event, throttle,
6207 * We inhibit the overflow from happening when
6208 * hwc->interrupts == MAX_INTERRUPTS.
6216 static void perf_swevent_event(struct perf_event *event, u64 nr,
6217 struct perf_sample_data *data,
6218 struct pt_regs *regs)
6220 struct hw_perf_event *hwc = &event->hw;
6222 local64_add(nr, &event->count);
6227 if (!is_sampling_event(event))
6230 if ((event->attr.sample_type & PERF_SAMPLE_PERIOD) && !event->attr.freq) {
6232 return perf_swevent_overflow(event, 1, data, regs);
6234 data->period = event->hw.last_period;
6236 if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq)
6237 return perf_swevent_overflow(event, 1, data, regs);
6239 if (local64_add_negative(nr, &hwc->period_left))
6242 perf_swevent_overflow(event, 0, data, regs);
6245 static int perf_exclude_event(struct perf_event *event,
6246 struct pt_regs *regs)
6248 if (event->hw.state & PERF_HES_STOPPED)
6252 if (event->attr.exclude_user && user_mode(regs))
6255 if (event->attr.exclude_kernel && !user_mode(regs))
6262 static int perf_swevent_match(struct perf_event *event,
6263 enum perf_type_id type,
6265 struct perf_sample_data *data,
6266 struct pt_regs *regs)
6268 if (event->attr.type != type)
6271 if (event->attr.config != event_id)
6274 if (perf_exclude_event(event, regs))
6280 static inline u64 swevent_hash(u64 type, u32 event_id)
6282 u64 val = event_id | (type << 32);
6284 return hash_64(val, SWEVENT_HLIST_BITS);
6287 static inline struct hlist_head *
6288 __find_swevent_head(struct swevent_hlist *hlist, u64 type, u32 event_id)
6290 u64 hash = swevent_hash(type, event_id);
6292 return &hlist->heads[hash];
6295 /* For the read side: events when they trigger */
6296 static inline struct hlist_head *
6297 find_swevent_head_rcu(struct swevent_htable *swhash, u64 type, u32 event_id)
6299 struct swevent_hlist *hlist;
6301 hlist = rcu_dereference(swhash->swevent_hlist);
6305 return __find_swevent_head(hlist, type, event_id);
6308 /* For the event head insertion and removal in the hlist */
6309 static inline struct hlist_head *
6310 find_swevent_head(struct swevent_htable *swhash, struct perf_event *event)
6312 struct swevent_hlist *hlist;
6313 u32 event_id = event->attr.config;
6314 u64 type = event->attr.type;
6317 * Event scheduling is always serialized against hlist allocation
6318 * and release. Which makes the protected version suitable here.
6319 * The context lock guarantees that.
6321 hlist = rcu_dereference_protected(swhash->swevent_hlist,
6322 lockdep_is_held(&event->ctx->lock));
6326 return __find_swevent_head(hlist, type, event_id);
6329 static void do_perf_sw_event(enum perf_type_id type, u32 event_id,
6331 struct perf_sample_data *data,
6332 struct pt_regs *regs)
6334 struct swevent_htable *swhash = this_cpu_ptr(&swevent_htable);
6335 struct perf_event *event;
6336 struct hlist_head *head;
6339 head = find_swevent_head_rcu(swhash, type, event_id);
6343 hlist_for_each_entry_rcu(event, head, hlist_entry) {
6344 if (perf_swevent_match(event, type, event_id, data, regs))
6345 perf_swevent_event(event, nr, data, regs);
6351 DEFINE_PER_CPU(struct pt_regs, __perf_regs[4]);
6353 int perf_swevent_get_recursion_context(void)
6355 struct swevent_htable *swhash = this_cpu_ptr(&swevent_htable);
6357 return get_recursion_context(swhash->recursion);
6359 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context);
6361 inline void perf_swevent_put_recursion_context(int rctx)
6363 struct swevent_htable *swhash = this_cpu_ptr(&swevent_htable);
6365 put_recursion_context(swhash->recursion, rctx);
6368 void ___perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)
6370 struct perf_sample_data data;
6372 if (WARN_ON_ONCE(!regs))
6375 perf_sample_data_init(&data, addr, 0);
6376 do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, &data, regs);
6379 void __perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)
6383 preempt_disable_notrace();
6384 rctx = perf_swevent_get_recursion_context();
6385 if (unlikely(rctx < 0))
6388 ___perf_sw_event(event_id, nr, regs, addr);
6390 perf_swevent_put_recursion_context(rctx);
6392 preempt_enable_notrace();
6395 static void perf_swevent_read(struct perf_event *event)
6399 static int perf_swevent_add(struct perf_event *event, int flags)
6401 struct swevent_htable *swhash = this_cpu_ptr(&swevent_htable);
6402 struct hw_perf_event *hwc = &event->hw;
6403 struct hlist_head *head;
6405 if (is_sampling_event(event)) {
6406 hwc->last_period = hwc->sample_period;
6407 perf_swevent_set_period(event);
6410 hwc->state = !(flags & PERF_EF_START);
6412 head = find_swevent_head(swhash, event);
6415 * We can race with cpu hotplug code. Do not
6416 * WARN if the cpu just got unplugged.
6418 WARN_ON_ONCE(swhash->online);
6422 hlist_add_head_rcu(&event->hlist_entry, head);
6423 perf_event_update_userpage(event);
6428 static void perf_swevent_del(struct perf_event *event, int flags)
6430 hlist_del_rcu(&event->hlist_entry);
6433 static void perf_swevent_start(struct perf_event *event, int flags)
6435 event->hw.state = 0;
6438 static void perf_swevent_stop(struct perf_event *event, int flags)
6440 event->hw.state = PERF_HES_STOPPED;
6443 /* Deref the hlist from the update side */
6444 static inline struct swevent_hlist *
6445 swevent_hlist_deref(struct swevent_htable *swhash)
6447 return rcu_dereference_protected(swhash->swevent_hlist,
6448 lockdep_is_held(&swhash->hlist_mutex));
6451 static void swevent_hlist_release(struct swevent_htable *swhash)
6453 struct swevent_hlist *hlist = swevent_hlist_deref(swhash);
6458 RCU_INIT_POINTER(swhash->swevent_hlist, NULL);
6459 kfree_rcu(hlist, rcu_head);
6462 static void swevent_hlist_put_cpu(struct perf_event *event, int cpu)
6464 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
6466 mutex_lock(&swhash->hlist_mutex);
6468 if (!--swhash->hlist_refcount)
6469 swevent_hlist_release(swhash);
6471 mutex_unlock(&swhash->hlist_mutex);
6474 static void swevent_hlist_put(struct perf_event *event)
6478 for_each_possible_cpu(cpu)
6479 swevent_hlist_put_cpu(event, cpu);
6482 static int swevent_hlist_get_cpu(struct perf_event *event, int cpu)
6484 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
6487 mutex_lock(&swhash->hlist_mutex);
6489 if (!swevent_hlist_deref(swhash) && cpu_online(cpu)) {
6490 struct swevent_hlist *hlist;
6492 hlist = kzalloc(sizeof(*hlist), GFP_KERNEL);
6497 rcu_assign_pointer(swhash->swevent_hlist, hlist);
6499 swhash->hlist_refcount++;
6501 mutex_unlock(&swhash->hlist_mutex);
6506 static int swevent_hlist_get(struct perf_event *event)
6509 int cpu, failed_cpu;
6512 for_each_possible_cpu(cpu) {
6513 err = swevent_hlist_get_cpu(event, cpu);
6523 for_each_possible_cpu(cpu) {
6524 if (cpu == failed_cpu)
6526 swevent_hlist_put_cpu(event, cpu);
6533 struct static_key perf_swevent_enabled[PERF_COUNT_SW_MAX];
6535 static void sw_perf_event_destroy(struct perf_event *event)
6537 u64 event_id = event->attr.config;
6539 WARN_ON(event->parent);
6541 static_key_slow_dec(&perf_swevent_enabled[event_id]);
6542 swevent_hlist_put(event);
6545 static int perf_swevent_init(struct perf_event *event)
6547 u64 event_id = event->attr.config;
6549 if (event->attr.type != PERF_TYPE_SOFTWARE)
6553 * no branch sampling for software events
6555 if (has_branch_stack(event))
6559 case PERF_COUNT_SW_CPU_CLOCK:
6560 case PERF_COUNT_SW_TASK_CLOCK:
6567 if (event_id >= PERF_COUNT_SW_MAX)
6570 if (!event->parent) {
6573 err = swevent_hlist_get(event);
6577 static_key_slow_inc(&perf_swevent_enabled[event_id]);
6578 event->destroy = sw_perf_event_destroy;
6584 static struct pmu perf_swevent = {
6585 .task_ctx_nr = perf_sw_context,
6587 .capabilities = PERF_PMU_CAP_NO_NMI,
6589 .event_init = perf_swevent_init,
6590 .add = perf_swevent_add,
6591 .del = perf_swevent_del,
6592 .start = perf_swevent_start,
6593 .stop = perf_swevent_stop,
6594 .read = perf_swevent_read,
6597 #ifdef CONFIG_EVENT_TRACING
6599 static int perf_tp_filter_match(struct perf_event *event,
6600 struct perf_sample_data *data)
6602 void *record = data->raw->data;
6604 if (likely(!event->filter) || filter_match_preds(event->filter, record))
6609 static int perf_tp_event_match(struct perf_event *event,
6610 struct perf_sample_data *data,
6611 struct pt_regs *regs)
6613 if (event->hw.state & PERF_HES_STOPPED)
6616 * All tracepoints are from kernel-space.
6618 if (event->attr.exclude_kernel)
6621 if (!perf_tp_filter_match(event, data))
6627 void perf_tp_event(u64 addr, u64 count, void *record, int entry_size,
6628 struct pt_regs *regs, struct hlist_head *head, int rctx,
6629 struct task_struct *task)
6631 struct perf_sample_data data;
6632 struct perf_event *event;
6634 struct perf_raw_record raw = {
6639 perf_sample_data_init(&data, addr, 0);
6642 hlist_for_each_entry_rcu(event, head, hlist_entry) {
6643 if (perf_tp_event_match(event, &data, regs))
6644 perf_swevent_event(event, count, &data, regs);
6648 * If we got specified a target task, also iterate its context and
6649 * deliver this event there too.
6651 if (task && task != current) {
6652 struct perf_event_context *ctx;
6653 struct trace_entry *entry = record;
6656 ctx = rcu_dereference(task->perf_event_ctxp[perf_sw_context]);
6660 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
6661 if (event->attr.type != PERF_TYPE_TRACEPOINT)
6663 if (event->attr.config != entry->type)
6665 if (perf_tp_event_match(event, &data, regs))
6666 perf_swevent_event(event, count, &data, regs);
6672 perf_swevent_put_recursion_context(rctx);
6674 EXPORT_SYMBOL_GPL(perf_tp_event);
6676 static void tp_perf_event_destroy(struct perf_event *event)
6678 perf_trace_destroy(event);
6681 static int perf_tp_event_init(struct perf_event *event)
6685 if (event->attr.type != PERF_TYPE_TRACEPOINT)
6689 * no branch sampling for tracepoint events
6691 if (has_branch_stack(event))
6694 err = perf_trace_init(event);
6698 event->destroy = tp_perf_event_destroy;
6703 static struct pmu perf_tracepoint = {
6704 .task_ctx_nr = perf_sw_context,
6706 .event_init = perf_tp_event_init,
6707 .add = perf_trace_add,
6708 .del = perf_trace_del,
6709 .start = perf_swevent_start,
6710 .stop = perf_swevent_stop,
6711 .read = perf_swevent_read,
6714 static inline void perf_tp_register(void)
6716 perf_pmu_register(&perf_tracepoint, "tracepoint", PERF_TYPE_TRACEPOINT);
6719 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
6724 if (event->attr.type != PERF_TYPE_TRACEPOINT)
6727 filter_str = strndup_user(arg, PAGE_SIZE);
6728 if (IS_ERR(filter_str))
6729 return PTR_ERR(filter_str);
6731 ret = ftrace_profile_set_filter(event, event->attr.config, filter_str);
6737 static void perf_event_free_filter(struct perf_event *event)
6739 ftrace_profile_free_filter(event);
6742 static int perf_event_set_bpf_prog(struct perf_event *event, u32 prog_fd)
6744 struct bpf_prog *prog;
6746 if (event->attr.type != PERF_TYPE_TRACEPOINT)
6749 if (event->tp_event->prog)
6752 if (!(event->tp_event->flags & TRACE_EVENT_FL_KPROBE))
6753 /* bpf programs can only be attached to kprobes */
6756 prog = bpf_prog_get(prog_fd);
6758 return PTR_ERR(prog);
6760 if (prog->type != BPF_PROG_TYPE_KPROBE) {
6761 /* valid fd, but invalid bpf program type */
6766 event->tp_event->prog = prog;
6771 static void perf_event_free_bpf_prog(struct perf_event *event)
6773 struct bpf_prog *prog;
6775 if (!event->tp_event)
6778 prog = event->tp_event->prog;
6780 event->tp_event->prog = NULL;
6787 static inline void perf_tp_register(void)
6791 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
6796 static void perf_event_free_filter(struct perf_event *event)
6800 static int perf_event_set_bpf_prog(struct perf_event *event, u32 prog_fd)
6805 static void perf_event_free_bpf_prog(struct perf_event *event)
6808 #endif /* CONFIG_EVENT_TRACING */
6810 #ifdef CONFIG_HAVE_HW_BREAKPOINT
6811 void perf_bp_event(struct perf_event *bp, void *data)
6813 struct perf_sample_data sample;
6814 struct pt_regs *regs = data;
6816 perf_sample_data_init(&sample, bp->attr.bp_addr, 0);
6818 if (!bp->hw.state && !perf_exclude_event(bp, regs))
6819 perf_swevent_event(bp, 1, &sample, regs);
6824 * hrtimer based swevent callback
6827 static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer)
6829 enum hrtimer_restart ret = HRTIMER_RESTART;
6830 struct perf_sample_data data;
6831 struct pt_regs *regs;
6832 struct perf_event *event;
6835 event = container_of(hrtimer, struct perf_event, hw.hrtimer);
6837 if (event->state != PERF_EVENT_STATE_ACTIVE)
6838 return HRTIMER_NORESTART;
6840 event->pmu->read(event);
6842 perf_sample_data_init(&data, 0, event->hw.last_period);
6843 regs = get_irq_regs();
6845 if (regs && !perf_exclude_event(event, regs)) {
6846 if (!(event->attr.exclude_idle && is_idle_task(current)))
6847 if (__perf_event_overflow(event, 1, &data, regs))
6848 ret = HRTIMER_NORESTART;
6851 period = max_t(u64, 10000, event->hw.sample_period);
6852 hrtimer_forward_now(hrtimer, ns_to_ktime(period));
6857 static void perf_swevent_start_hrtimer(struct perf_event *event)
6859 struct hw_perf_event *hwc = &event->hw;
6862 if (!is_sampling_event(event))
6865 period = local64_read(&hwc->period_left);
6870 local64_set(&hwc->period_left, 0);
6872 period = max_t(u64, 10000, hwc->sample_period);
6874 hrtimer_start(&hwc->hrtimer, ns_to_ktime(period),
6875 HRTIMER_MODE_REL_PINNED);
6878 static void perf_swevent_cancel_hrtimer(struct perf_event *event)
6880 struct hw_perf_event *hwc = &event->hw;
6882 if (is_sampling_event(event)) {
6883 ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer);
6884 local64_set(&hwc->period_left, ktime_to_ns(remaining));
6886 hrtimer_cancel(&hwc->hrtimer);
6890 static void perf_swevent_init_hrtimer(struct perf_event *event)
6892 struct hw_perf_event *hwc = &event->hw;
6894 if (!is_sampling_event(event))
6897 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
6898 hwc->hrtimer.function = perf_swevent_hrtimer;
6901 * Since hrtimers have a fixed rate, we can do a static freq->period
6902 * mapping and avoid the whole period adjust feedback stuff.
6904 if (event->attr.freq) {
6905 long freq = event->attr.sample_freq;
6907 event->attr.sample_period = NSEC_PER_SEC / freq;
6908 hwc->sample_period = event->attr.sample_period;
6909 local64_set(&hwc->period_left, hwc->sample_period);
6910 hwc->last_period = hwc->sample_period;
6911 event->attr.freq = 0;
6916 * Software event: cpu wall time clock
6919 static void cpu_clock_event_update(struct perf_event *event)
6924 now = local_clock();
6925 prev = local64_xchg(&event->hw.prev_count, now);
6926 local64_add(now - prev, &event->count);
6929 static void cpu_clock_event_start(struct perf_event *event, int flags)
6931 local64_set(&event->hw.prev_count, local_clock());
6932 perf_swevent_start_hrtimer(event);
6935 static void cpu_clock_event_stop(struct perf_event *event, int flags)
6937 perf_swevent_cancel_hrtimer(event);
6938 cpu_clock_event_update(event);
6941 static int cpu_clock_event_add(struct perf_event *event, int flags)
6943 if (flags & PERF_EF_START)
6944 cpu_clock_event_start(event, flags);
6945 perf_event_update_userpage(event);
6950 static void cpu_clock_event_del(struct perf_event *event, int flags)
6952 cpu_clock_event_stop(event, flags);
6955 static void cpu_clock_event_read(struct perf_event *event)
6957 cpu_clock_event_update(event);
6960 static int cpu_clock_event_init(struct perf_event *event)
6962 if (event->attr.type != PERF_TYPE_SOFTWARE)
6965 if (event->attr.config != PERF_COUNT_SW_CPU_CLOCK)
6969 * no branch sampling for software events
6971 if (has_branch_stack(event))
6974 perf_swevent_init_hrtimer(event);
6979 static struct pmu perf_cpu_clock = {
6980 .task_ctx_nr = perf_sw_context,
6982 .capabilities = PERF_PMU_CAP_NO_NMI,
6984 .event_init = cpu_clock_event_init,
6985 .add = cpu_clock_event_add,
6986 .del = cpu_clock_event_del,
6987 .start = cpu_clock_event_start,
6988 .stop = cpu_clock_event_stop,
6989 .read = cpu_clock_event_read,
6993 * Software event: task time clock
6996 static void task_clock_event_update(struct perf_event *event, u64 now)
7001 prev = local64_xchg(&event->hw.prev_count, now);
7003 local64_add(delta, &event->count);
7006 static void task_clock_event_start(struct perf_event *event, int flags)
7008 local64_set(&event->hw.prev_count, event->ctx->time);
7009 perf_swevent_start_hrtimer(event);
7012 static void task_clock_event_stop(struct perf_event *event, int flags)
7014 perf_swevent_cancel_hrtimer(event);
7015 task_clock_event_update(event, event->ctx->time);
7018 static int task_clock_event_add(struct perf_event *event, int flags)
7020 if (flags & PERF_EF_START)
7021 task_clock_event_start(event, flags);
7022 perf_event_update_userpage(event);
7027 static void task_clock_event_del(struct perf_event *event, int flags)
7029 task_clock_event_stop(event, PERF_EF_UPDATE);
7032 static void task_clock_event_read(struct perf_event *event)
7034 u64 now = perf_clock();
7035 u64 delta = now - event->ctx->timestamp;
7036 u64 time = event->ctx->time + delta;
7038 task_clock_event_update(event, time);
7041 static int task_clock_event_init(struct perf_event *event)
7043 if (event->attr.type != PERF_TYPE_SOFTWARE)
7046 if (event->attr.config != PERF_COUNT_SW_TASK_CLOCK)
7050 * no branch sampling for software events
7052 if (has_branch_stack(event))
7055 perf_swevent_init_hrtimer(event);
7060 static struct pmu perf_task_clock = {
7061 .task_ctx_nr = perf_sw_context,
7063 .capabilities = PERF_PMU_CAP_NO_NMI,
7065 .event_init = task_clock_event_init,
7066 .add = task_clock_event_add,
7067 .del = task_clock_event_del,
7068 .start = task_clock_event_start,
7069 .stop = task_clock_event_stop,
7070 .read = task_clock_event_read,
7073 static void perf_pmu_nop_void(struct pmu *pmu)
7077 static int perf_pmu_nop_int(struct pmu *pmu)
7082 static void perf_pmu_start_txn(struct pmu *pmu)
7084 perf_pmu_disable(pmu);
7087 static int perf_pmu_commit_txn(struct pmu *pmu)
7089 perf_pmu_enable(pmu);
7093 static void perf_pmu_cancel_txn(struct pmu *pmu)
7095 perf_pmu_enable(pmu);
7098 static int perf_event_idx_default(struct perf_event *event)
7104 * Ensures all contexts with the same task_ctx_nr have the same
7105 * pmu_cpu_context too.
7107 static struct perf_cpu_context __percpu *find_pmu_context(int ctxn)
7114 list_for_each_entry(pmu, &pmus, entry) {
7115 if (pmu->task_ctx_nr == ctxn)
7116 return pmu->pmu_cpu_context;
7122 static void update_pmu_context(struct pmu *pmu, struct pmu *old_pmu)
7126 for_each_possible_cpu(cpu) {
7127 struct perf_cpu_context *cpuctx;
7129 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
7131 if (cpuctx->unique_pmu == old_pmu)
7132 cpuctx->unique_pmu = pmu;
7136 static void free_pmu_context(struct pmu *pmu)
7140 mutex_lock(&pmus_lock);
7142 * Like a real lame refcount.
7144 list_for_each_entry(i, &pmus, entry) {
7145 if (i->pmu_cpu_context == pmu->pmu_cpu_context) {
7146 update_pmu_context(i, pmu);
7151 free_percpu(pmu->pmu_cpu_context);
7153 mutex_unlock(&pmus_lock);
7155 static struct idr pmu_idr;
7158 type_show(struct device *dev, struct device_attribute *attr, char *page)
7160 struct pmu *pmu = dev_get_drvdata(dev);
7162 return snprintf(page, PAGE_SIZE-1, "%d\n", pmu->type);
7164 static DEVICE_ATTR_RO(type);
7167 perf_event_mux_interval_ms_show(struct device *dev,
7168 struct device_attribute *attr,
7171 struct pmu *pmu = dev_get_drvdata(dev);
7173 return snprintf(page, PAGE_SIZE-1, "%d\n", pmu->hrtimer_interval_ms);
7176 static DEFINE_MUTEX(mux_interval_mutex);
7179 perf_event_mux_interval_ms_store(struct device *dev,
7180 struct device_attribute *attr,
7181 const char *buf, size_t count)
7183 struct pmu *pmu = dev_get_drvdata(dev);
7184 int timer, cpu, ret;
7186 ret = kstrtoint(buf, 0, &timer);
7193 /* same value, noting to do */
7194 if (timer == pmu->hrtimer_interval_ms)
7197 mutex_lock(&mux_interval_mutex);
7198 pmu->hrtimer_interval_ms = timer;
7200 /* update all cpuctx for this PMU */
7202 for_each_online_cpu(cpu) {
7203 struct perf_cpu_context *cpuctx;
7204 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
7205 cpuctx->hrtimer_interval = ns_to_ktime(NSEC_PER_MSEC * timer);
7207 cpu_function_call(cpu,
7208 (remote_function_f)perf_mux_hrtimer_restart, cpuctx);
7211 mutex_unlock(&mux_interval_mutex);
7215 static DEVICE_ATTR_RW(perf_event_mux_interval_ms);
7217 static struct attribute *pmu_dev_attrs[] = {
7218 &dev_attr_type.attr,
7219 &dev_attr_perf_event_mux_interval_ms.attr,
7222 ATTRIBUTE_GROUPS(pmu_dev);
7224 static int pmu_bus_running;
7225 static struct bus_type pmu_bus = {
7226 .name = "event_source",
7227 .dev_groups = pmu_dev_groups,
7230 static void pmu_dev_release(struct device *dev)
7235 static int pmu_dev_alloc(struct pmu *pmu)
7239 pmu->dev = kzalloc(sizeof(struct device), GFP_KERNEL);
7243 pmu->dev->groups = pmu->attr_groups;
7244 device_initialize(pmu->dev);
7245 ret = dev_set_name(pmu->dev, "%s", pmu->name);
7249 dev_set_drvdata(pmu->dev, pmu);
7250 pmu->dev->bus = &pmu_bus;
7251 pmu->dev->release = pmu_dev_release;
7252 ret = device_add(pmu->dev);
7260 put_device(pmu->dev);
7264 static struct lock_class_key cpuctx_mutex;
7265 static struct lock_class_key cpuctx_lock;
7267 int perf_pmu_register(struct pmu *pmu, const char *name, int type)
7271 mutex_lock(&pmus_lock);
7273 pmu->pmu_disable_count = alloc_percpu(int);
7274 if (!pmu->pmu_disable_count)
7283 type = idr_alloc(&pmu_idr, pmu, PERF_TYPE_MAX, 0, GFP_KERNEL);
7291 if (pmu_bus_running) {
7292 ret = pmu_dev_alloc(pmu);
7298 pmu->pmu_cpu_context = find_pmu_context(pmu->task_ctx_nr);
7299 if (pmu->pmu_cpu_context)
7300 goto got_cpu_context;
7303 pmu->pmu_cpu_context = alloc_percpu(struct perf_cpu_context);
7304 if (!pmu->pmu_cpu_context)
7307 for_each_possible_cpu(cpu) {
7308 struct perf_cpu_context *cpuctx;
7310 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
7311 __perf_event_init_context(&cpuctx->ctx);
7312 lockdep_set_class(&cpuctx->ctx.mutex, &cpuctx_mutex);
7313 lockdep_set_class(&cpuctx->ctx.lock, &cpuctx_lock);
7314 cpuctx->ctx.pmu = pmu;
7316 __perf_mux_hrtimer_init(cpuctx, cpu);
7318 cpuctx->unique_pmu = pmu;
7322 if (!pmu->start_txn) {
7323 if (pmu->pmu_enable) {
7325 * If we have pmu_enable/pmu_disable calls, install
7326 * transaction stubs that use that to try and batch
7327 * hardware accesses.
7329 pmu->start_txn = perf_pmu_start_txn;
7330 pmu->commit_txn = perf_pmu_commit_txn;
7331 pmu->cancel_txn = perf_pmu_cancel_txn;
7333 pmu->start_txn = perf_pmu_nop_void;
7334 pmu->commit_txn = perf_pmu_nop_int;
7335 pmu->cancel_txn = perf_pmu_nop_void;
7339 if (!pmu->pmu_enable) {
7340 pmu->pmu_enable = perf_pmu_nop_void;
7341 pmu->pmu_disable = perf_pmu_nop_void;
7344 if (!pmu->event_idx)
7345 pmu->event_idx = perf_event_idx_default;
7347 list_add_rcu(&pmu->entry, &pmus);
7348 atomic_set(&pmu->exclusive_cnt, 0);
7351 mutex_unlock(&pmus_lock);
7356 device_del(pmu->dev);
7357 put_device(pmu->dev);
7360 if (pmu->type >= PERF_TYPE_MAX)
7361 idr_remove(&pmu_idr, pmu->type);
7364 free_percpu(pmu->pmu_disable_count);
7367 EXPORT_SYMBOL_GPL(perf_pmu_register);
7369 void perf_pmu_unregister(struct pmu *pmu)
7371 mutex_lock(&pmus_lock);
7372 list_del_rcu(&pmu->entry);
7373 mutex_unlock(&pmus_lock);
7376 * We dereference the pmu list under both SRCU and regular RCU, so
7377 * synchronize against both of those.
7379 synchronize_srcu(&pmus_srcu);
7382 free_percpu(pmu->pmu_disable_count);
7383 if (pmu->type >= PERF_TYPE_MAX)
7384 idr_remove(&pmu_idr, pmu->type);
7385 device_del(pmu->dev);
7386 put_device(pmu->dev);
7387 free_pmu_context(pmu);
7389 EXPORT_SYMBOL_GPL(perf_pmu_unregister);
7391 static int perf_try_init_event(struct pmu *pmu, struct perf_event *event)
7393 struct perf_event_context *ctx = NULL;
7396 if (!try_module_get(pmu->module))
7399 if (event->group_leader != event) {
7401 * This ctx->mutex can nest when we're called through
7402 * inheritance. See the perf_event_ctx_lock_nested() comment.
7404 ctx = perf_event_ctx_lock_nested(event->group_leader,
7405 SINGLE_DEPTH_NESTING);
7410 ret = pmu->event_init(event);
7413 perf_event_ctx_unlock(event->group_leader, ctx);
7416 module_put(pmu->module);
7421 struct pmu *perf_init_event(struct perf_event *event)
7423 struct pmu *pmu = NULL;
7427 idx = srcu_read_lock(&pmus_srcu);
7430 pmu = idr_find(&pmu_idr, event->attr.type);
7433 ret = perf_try_init_event(pmu, event);
7439 list_for_each_entry_rcu(pmu, &pmus, entry) {
7440 ret = perf_try_init_event(pmu, event);
7444 if (ret != -ENOENT) {
7449 pmu = ERR_PTR(-ENOENT);
7451 srcu_read_unlock(&pmus_srcu, idx);
7456 static void account_event_cpu(struct perf_event *event, int cpu)
7461 if (is_cgroup_event(event))
7462 atomic_inc(&per_cpu(perf_cgroup_events, cpu));
7465 static void account_event(struct perf_event *event)
7470 if (event->attach_state & PERF_ATTACH_TASK)
7471 static_key_slow_inc(&perf_sched_events.key);
7472 if (event->attr.mmap || event->attr.mmap_data)
7473 atomic_inc(&nr_mmap_events);
7474 if (event->attr.comm)
7475 atomic_inc(&nr_comm_events);
7476 if (event->attr.task)
7477 atomic_inc(&nr_task_events);
7478 if (event->attr.freq) {
7479 if (atomic_inc_return(&nr_freq_events) == 1)
7480 tick_nohz_full_kick_all();
7482 if (has_branch_stack(event))
7483 static_key_slow_inc(&perf_sched_events.key);
7484 if (is_cgroup_event(event))
7485 static_key_slow_inc(&perf_sched_events.key);
7487 account_event_cpu(event, event->cpu);
7491 * Allocate and initialize a event structure
7493 static struct perf_event *
7494 perf_event_alloc(struct perf_event_attr *attr, int cpu,
7495 struct task_struct *task,
7496 struct perf_event *group_leader,
7497 struct perf_event *parent_event,
7498 perf_overflow_handler_t overflow_handler,
7499 void *context, int cgroup_fd)
7502 struct perf_event *event;
7503 struct hw_perf_event *hwc;
7506 if ((unsigned)cpu >= nr_cpu_ids) {
7507 if (!task || cpu != -1)
7508 return ERR_PTR(-EINVAL);
7511 event = kzalloc(sizeof(*event), GFP_KERNEL);
7513 return ERR_PTR(-ENOMEM);
7516 * Single events are their own group leaders, with an
7517 * empty sibling list:
7520 group_leader = event;
7522 mutex_init(&event->child_mutex);
7523 INIT_LIST_HEAD(&event->child_list);
7525 INIT_LIST_HEAD(&event->group_entry);
7526 INIT_LIST_HEAD(&event->event_entry);
7527 INIT_LIST_HEAD(&event->sibling_list);
7528 INIT_LIST_HEAD(&event->rb_entry);
7529 INIT_LIST_HEAD(&event->active_entry);
7530 INIT_HLIST_NODE(&event->hlist_entry);
7533 init_waitqueue_head(&event->waitq);
7534 init_irq_work(&event->pending, perf_pending_event);
7536 mutex_init(&event->mmap_mutex);
7538 atomic_long_set(&event->refcount, 1);
7540 event->attr = *attr;
7541 event->group_leader = group_leader;
7545 event->parent = parent_event;
7547 event->ns = get_pid_ns(task_active_pid_ns(current));
7548 event->id = atomic64_inc_return(&perf_event_id);
7550 event->state = PERF_EVENT_STATE_INACTIVE;
7553 event->attach_state = PERF_ATTACH_TASK;
7555 * XXX pmu::event_init needs to know what task to account to
7556 * and we cannot use the ctx information because we need the
7557 * pmu before we get a ctx.
7559 event->hw.target = task;
7562 event->clock = &local_clock;
7564 event->clock = parent_event->clock;
7566 if (!overflow_handler && parent_event) {
7567 overflow_handler = parent_event->overflow_handler;
7568 context = parent_event->overflow_handler_context;
7571 event->overflow_handler = overflow_handler;
7572 event->overflow_handler_context = context;
7574 perf_event__state_init(event);
7579 hwc->sample_period = attr->sample_period;
7580 if (attr->freq && attr->sample_freq)
7581 hwc->sample_period = 1;
7582 hwc->last_period = hwc->sample_period;
7584 local64_set(&hwc->period_left, hwc->sample_period);
7587 * we currently do not support PERF_FORMAT_GROUP on inherited events
7589 if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP))
7592 if (!has_branch_stack(event))
7593 event->attr.branch_sample_type = 0;
7595 if (cgroup_fd != -1) {
7596 err = perf_cgroup_connect(cgroup_fd, event, attr, group_leader);
7601 pmu = perf_init_event(event);
7604 else if (IS_ERR(pmu)) {
7609 err = exclusive_event_init(event);
7613 if (!event->parent) {
7614 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) {
7615 err = get_callchain_buffers();
7624 exclusive_event_destroy(event);
7628 event->destroy(event);
7629 module_put(pmu->module);
7631 if (is_cgroup_event(event))
7632 perf_detach_cgroup(event);
7634 put_pid_ns(event->ns);
7637 return ERR_PTR(err);
7640 static int perf_copy_attr(struct perf_event_attr __user *uattr,
7641 struct perf_event_attr *attr)
7646 if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0))
7650 * zero the full structure, so that a short copy will be nice.
7652 memset(attr, 0, sizeof(*attr));
7654 ret = get_user(size, &uattr->size);
7658 if (size > PAGE_SIZE) /* silly large */
7661 if (!size) /* abi compat */
7662 size = PERF_ATTR_SIZE_VER0;
7664 if (size < PERF_ATTR_SIZE_VER0)
7668 * If we're handed a bigger struct than we know of,
7669 * ensure all the unknown bits are 0 - i.e. new
7670 * user-space does not rely on any kernel feature
7671 * extensions we dont know about yet.
7673 if (size > sizeof(*attr)) {
7674 unsigned char __user *addr;
7675 unsigned char __user *end;
7678 addr = (void __user *)uattr + sizeof(*attr);
7679 end = (void __user *)uattr + size;
7681 for (; addr < end; addr++) {
7682 ret = get_user(val, addr);
7688 size = sizeof(*attr);
7691 ret = copy_from_user(attr, uattr, size);
7695 if (attr->__reserved_1)
7698 if (attr->sample_type & ~(PERF_SAMPLE_MAX-1))
7701 if (attr->read_format & ~(PERF_FORMAT_MAX-1))
7704 if (attr->sample_type & PERF_SAMPLE_BRANCH_STACK) {
7705 u64 mask = attr->branch_sample_type;
7707 /* only using defined bits */
7708 if (mask & ~(PERF_SAMPLE_BRANCH_MAX-1))
7711 /* at least one branch bit must be set */
7712 if (!(mask & ~PERF_SAMPLE_BRANCH_PLM_ALL))
7715 /* propagate priv level, when not set for branch */
7716 if (!(mask & PERF_SAMPLE_BRANCH_PLM_ALL)) {
7718 /* exclude_kernel checked on syscall entry */
7719 if (!attr->exclude_kernel)
7720 mask |= PERF_SAMPLE_BRANCH_KERNEL;
7722 if (!attr->exclude_user)
7723 mask |= PERF_SAMPLE_BRANCH_USER;
7725 if (!attr->exclude_hv)
7726 mask |= PERF_SAMPLE_BRANCH_HV;
7728 * adjust user setting (for HW filter setup)
7730 attr->branch_sample_type = mask;
7732 /* privileged levels capture (kernel, hv): check permissions */
7733 if ((mask & PERF_SAMPLE_BRANCH_PERM_PLM)
7734 && perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
7738 if (attr->sample_type & PERF_SAMPLE_REGS_USER) {
7739 ret = perf_reg_validate(attr->sample_regs_user);
7744 if (attr->sample_type & PERF_SAMPLE_STACK_USER) {
7745 if (!arch_perf_have_user_stack_dump())
7749 * We have __u32 type for the size, but so far
7750 * we can only use __u16 as maximum due to the
7751 * __u16 sample size limit.
7753 if (attr->sample_stack_user >= USHRT_MAX)
7755 else if (!IS_ALIGNED(attr->sample_stack_user, sizeof(u64)))
7759 if (attr->sample_type & PERF_SAMPLE_REGS_INTR)
7760 ret = perf_reg_validate(attr->sample_regs_intr);
7765 put_user(sizeof(*attr), &uattr->size);
7771 perf_event_set_output(struct perf_event *event, struct perf_event *output_event)
7773 struct ring_buffer *rb = NULL;
7779 /* don't allow circular references */
7780 if (event == output_event)
7784 * Don't allow cross-cpu buffers
7786 if (output_event->cpu != event->cpu)
7790 * If its not a per-cpu rb, it must be the same task.
7792 if (output_event->cpu == -1 && output_event->ctx != event->ctx)
7796 * Mixing clocks in the same buffer is trouble you don't need.
7798 if (output_event->clock != event->clock)
7802 * If both events generate aux data, they must be on the same PMU
7804 if (has_aux(event) && has_aux(output_event) &&
7805 event->pmu != output_event->pmu)
7809 mutex_lock(&event->mmap_mutex);
7810 /* Can't redirect output if we've got an active mmap() */
7811 if (atomic_read(&event->mmap_count))
7815 /* get the rb we want to redirect to */
7816 rb = ring_buffer_get(output_event);
7821 ring_buffer_attach(event, rb);
7825 mutex_unlock(&event->mmap_mutex);
7831 static void mutex_lock_double(struct mutex *a, struct mutex *b)
7837 mutex_lock_nested(b, SINGLE_DEPTH_NESTING);
7840 static int perf_event_set_clock(struct perf_event *event, clockid_t clk_id)
7842 bool nmi_safe = false;
7845 case CLOCK_MONOTONIC:
7846 event->clock = &ktime_get_mono_fast_ns;
7850 case CLOCK_MONOTONIC_RAW:
7851 event->clock = &ktime_get_raw_fast_ns;
7855 case CLOCK_REALTIME:
7856 event->clock = &ktime_get_real_ns;
7859 case CLOCK_BOOTTIME:
7860 event->clock = &ktime_get_boot_ns;
7864 event->clock = &ktime_get_tai_ns;
7871 if (!nmi_safe && !(event->pmu->capabilities & PERF_PMU_CAP_NO_NMI))
7878 * sys_perf_event_open - open a performance event, associate it to a task/cpu
7880 * @attr_uptr: event_id type attributes for monitoring/sampling
7883 * @group_fd: group leader event fd
7885 SYSCALL_DEFINE5(perf_event_open,
7886 struct perf_event_attr __user *, attr_uptr,
7887 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
7889 struct perf_event *group_leader = NULL, *output_event = NULL;
7890 struct perf_event *event, *sibling;
7891 struct perf_event_attr attr;
7892 struct perf_event_context *ctx, *uninitialized_var(gctx);
7893 struct file *event_file = NULL;
7894 struct fd group = {NULL, 0};
7895 struct task_struct *task = NULL;
7900 int f_flags = O_RDWR;
7903 /* for future expandability... */
7904 if (flags & ~PERF_FLAG_ALL)
7907 err = perf_copy_attr(attr_uptr, &attr);
7911 if (!attr.exclude_kernel) {
7912 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
7917 if (attr.sample_freq > sysctl_perf_event_sample_rate)
7920 if (attr.sample_period & (1ULL << 63))
7925 * In cgroup mode, the pid argument is used to pass the fd
7926 * opened to the cgroup directory in cgroupfs. The cpu argument
7927 * designates the cpu on which to monitor threads from that
7930 if ((flags & PERF_FLAG_PID_CGROUP) && (pid == -1 || cpu == -1))
7933 if (flags & PERF_FLAG_FD_CLOEXEC)
7934 f_flags |= O_CLOEXEC;
7936 event_fd = get_unused_fd_flags(f_flags);
7940 if (group_fd != -1) {
7941 err = perf_fget_light(group_fd, &group);
7944 group_leader = group.file->private_data;
7945 if (flags & PERF_FLAG_FD_OUTPUT)
7946 output_event = group_leader;
7947 if (flags & PERF_FLAG_FD_NO_GROUP)
7948 group_leader = NULL;
7951 if (pid != -1 && !(flags & PERF_FLAG_PID_CGROUP)) {
7952 task = find_lively_task_by_vpid(pid);
7954 err = PTR_ERR(task);
7959 if (task && group_leader &&
7960 group_leader->attr.inherit != attr.inherit) {
7967 if (flags & PERF_FLAG_PID_CGROUP)
7970 event = perf_event_alloc(&attr, cpu, task, group_leader, NULL,
7971 NULL, NULL, cgroup_fd);
7972 if (IS_ERR(event)) {
7973 err = PTR_ERR(event);
7977 if (is_sampling_event(event)) {
7978 if (event->pmu->capabilities & PERF_PMU_CAP_NO_INTERRUPT) {
7984 account_event(event);
7987 * Special case software events and allow them to be part of
7988 * any hardware group.
7992 if (attr.use_clockid) {
7993 err = perf_event_set_clock(event, attr.clockid);
7999 (is_software_event(event) != is_software_event(group_leader))) {
8000 if (is_software_event(event)) {
8002 * If event and group_leader are not both a software
8003 * event, and event is, then group leader is not.
8005 * Allow the addition of software events to !software
8006 * groups, this is safe because software events never
8009 pmu = group_leader->pmu;
8010 } else if (is_software_event(group_leader) &&
8011 (group_leader->group_flags & PERF_GROUP_SOFTWARE)) {
8013 * In case the group is a pure software group, and we
8014 * try to add a hardware event, move the whole group to
8015 * the hardware context.
8022 * Get the target context (task or percpu):
8024 ctx = find_get_context(pmu, task, event);
8030 if ((pmu->capabilities & PERF_PMU_CAP_EXCLUSIVE) && group_leader) {
8036 put_task_struct(task);
8041 * Look up the group leader (we will attach this event to it):
8047 * Do not allow a recursive hierarchy (this new sibling
8048 * becoming part of another group-sibling):
8050 if (group_leader->group_leader != group_leader)
8053 /* All events in a group should have the same clock */
8054 if (group_leader->clock != event->clock)
8058 * Do not allow to attach to a group in a different
8059 * task or CPU context:
8063 * Make sure we're both on the same task, or both
8066 if (group_leader->ctx->task != ctx->task)
8070 * Make sure we're both events for the same CPU;
8071 * grouping events for different CPUs is broken; since
8072 * you can never concurrently schedule them anyhow.
8074 if (group_leader->cpu != event->cpu)
8077 if (group_leader->ctx != ctx)
8082 * Only a group leader can be exclusive or pinned
8084 if (attr.exclusive || attr.pinned)
8089 err = perf_event_set_output(event, output_event);
8094 event_file = anon_inode_getfile("[perf_event]", &perf_fops, event,
8096 if (IS_ERR(event_file)) {
8097 err = PTR_ERR(event_file);
8102 gctx = group_leader->ctx;
8105 * See perf_event_ctx_lock() for comments on the details
8106 * of swizzling perf_event::ctx.
8108 mutex_lock_double(&gctx->mutex, &ctx->mutex);
8110 perf_remove_from_context(group_leader, false);
8112 list_for_each_entry(sibling, &group_leader->sibling_list,
8114 perf_remove_from_context(sibling, false);
8118 mutex_lock(&ctx->mutex);
8121 WARN_ON_ONCE(ctx->parent_ctx);
8125 * Wait for everybody to stop referencing the events through
8126 * the old lists, before installing it on new lists.
8131 * Install the group siblings before the group leader.
8133 * Because a group leader will try and install the entire group
8134 * (through the sibling list, which is still in-tact), we can
8135 * end up with siblings installed in the wrong context.
8137 * By installing siblings first we NO-OP because they're not
8138 * reachable through the group lists.
8140 list_for_each_entry(sibling, &group_leader->sibling_list,
8142 perf_event__state_init(sibling);
8143 perf_install_in_context(ctx, sibling, sibling->cpu);
8148 * Removing from the context ends up with disabled
8149 * event. What we want here is event in the initial
8150 * startup state, ready to be add into new context.
8152 perf_event__state_init(group_leader);
8153 perf_install_in_context(ctx, group_leader, group_leader->cpu);
8157 if (!exclusive_event_installable(event, ctx)) {
8159 mutex_unlock(&ctx->mutex);
8164 perf_install_in_context(ctx, event, event->cpu);
8165 perf_unpin_context(ctx);
8168 mutex_unlock(&gctx->mutex);
8171 mutex_unlock(&ctx->mutex);
8175 event->owner = current;
8177 mutex_lock(¤t->perf_event_mutex);
8178 list_add_tail(&event->owner_entry, ¤t->perf_event_list);
8179 mutex_unlock(¤t->perf_event_mutex);
8182 * Precalculate sample_data sizes
8184 perf_event__header_size(event);
8185 perf_event__id_header_size(event);
8188 * Drop the reference on the group_event after placing the
8189 * new event on the sibling_list. This ensures destruction
8190 * of the group leader will find the pointer to itself in
8191 * perf_group_detach().
8194 fd_install(event_fd, event_file);
8198 perf_unpin_context(ctx);
8206 put_task_struct(task);
8210 put_unused_fd(event_fd);
8215 * perf_event_create_kernel_counter
8217 * @attr: attributes of the counter to create
8218 * @cpu: cpu in which the counter is bound
8219 * @task: task to profile (NULL for percpu)
8222 perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu,
8223 struct task_struct *task,
8224 perf_overflow_handler_t overflow_handler,
8227 struct perf_event_context *ctx;
8228 struct perf_event *event;
8232 * Get the target context (task or percpu):
8235 event = perf_event_alloc(attr, cpu, task, NULL, NULL,
8236 overflow_handler, context, -1);
8237 if (IS_ERR(event)) {
8238 err = PTR_ERR(event);
8242 /* Mark owner so we could distinguish it from user events. */
8243 event->owner = EVENT_OWNER_KERNEL;
8245 account_event(event);
8247 ctx = find_get_context(event->pmu, task, event);
8253 WARN_ON_ONCE(ctx->parent_ctx);
8254 mutex_lock(&ctx->mutex);
8255 if (!exclusive_event_installable(event, ctx)) {
8256 mutex_unlock(&ctx->mutex);
8257 perf_unpin_context(ctx);
8263 perf_install_in_context(ctx, event, cpu);
8264 perf_unpin_context(ctx);
8265 mutex_unlock(&ctx->mutex);
8272 return ERR_PTR(err);
8274 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter);
8276 void perf_pmu_migrate_context(struct pmu *pmu, int src_cpu, int dst_cpu)
8278 struct perf_event_context *src_ctx;
8279 struct perf_event_context *dst_ctx;
8280 struct perf_event *event, *tmp;
8283 src_ctx = &per_cpu_ptr(pmu->pmu_cpu_context, src_cpu)->ctx;
8284 dst_ctx = &per_cpu_ptr(pmu->pmu_cpu_context, dst_cpu)->ctx;
8287 * See perf_event_ctx_lock() for comments on the details
8288 * of swizzling perf_event::ctx.
8290 mutex_lock_double(&src_ctx->mutex, &dst_ctx->mutex);
8291 list_for_each_entry_safe(event, tmp, &src_ctx->event_list,
8293 perf_remove_from_context(event, false);
8294 unaccount_event_cpu(event, src_cpu);
8296 list_add(&event->migrate_entry, &events);
8300 * Wait for the events to quiesce before re-instating them.
8305 * Re-instate events in 2 passes.
8307 * Skip over group leaders and only install siblings on this first
8308 * pass, siblings will not get enabled without a leader, however a
8309 * leader will enable its siblings, even if those are still on the old
8312 list_for_each_entry_safe(event, tmp, &events, migrate_entry) {
8313 if (event->group_leader == event)
8316 list_del(&event->migrate_entry);
8317 if (event->state >= PERF_EVENT_STATE_OFF)
8318 event->state = PERF_EVENT_STATE_INACTIVE;
8319 account_event_cpu(event, dst_cpu);
8320 perf_install_in_context(dst_ctx, event, dst_cpu);
8325 * Once all the siblings are setup properly, install the group leaders
8328 list_for_each_entry_safe(event, tmp, &events, migrate_entry) {
8329 list_del(&event->migrate_entry);
8330 if (event->state >= PERF_EVENT_STATE_OFF)
8331 event->state = PERF_EVENT_STATE_INACTIVE;
8332 account_event_cpu(event, dst_cpu);
8333 perf_install_in_context(dst_ctx, event, dst_cpu);
8336 mutex_unlock(&dst_ctx->mutex);
8337 mutex_unlock(&src_ctx->mutex);
8339 EXPORT_SYMBOL_GPL(perf_pmu_migrate_context);
8341 static void sync_child_event(struct perf_event *child_event,
8342 struct task_struct *child)
8344 struct perf_event *parent_event = child_event->parent;
8347 if (child_event->attr.inherit_stat)
8348 perf_event_read_event(child_event, child);
8350 child_val = perf_event_count(child_event);
8353 * Add back the child's count to the parent's count:
8355 atomic64_add(child_val, &parent_event->child_count);
8356 atomic64_add(child_event->total_time_enabled,
8357 &parent_event->child_total_time_enabled);
8358 atomic64_add(child_event->total_time_running,
8359 &parent_event->child_total_time_running);
8362 * Remove this event from the parent's list
8364 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
8365 mutex_lock(&parent_event->child_mutex);
8366 list_del_init(&child_event->child_list);
8367 mutex_unlock(&parent_event->child_mutex);
8370 * Make sure user/parent get notified, that we just
8373 perf_event_wakeup(parent_event);
8376 * Release the parent event, if this was the last
8379 put_event(parent_event);
8383 __perf_event_exit_task(struct perf_event *child_event,
8384 struct perf_event_context *child_ctx,
8385 struct task_struct *child)
8388 * Do not destroy the 'original' grouping; because of the context
8389 * switch optimization the original events could've ended up in a
8390 * random child task.
8392 * If we were to destroy the original group, all group related
8393 * operations would cease to function properly after this random
8396 * Do destroy all inherited groups, we don't care about those
8397 * and being thorough is better.
8399 perf_remove_from_context(child_event, !!child_event->parent);
8402 * It can happen that the parent exits first, and has events
8403 * that are still around due to the child reference. These
8404 * events need to be zapped.
8406 if (child_event->parent) {
8407 sync_child_event(child_event, child);
8408 free_event(child_event);
8410 child_event->state = PERF_EVENT_STATE_EXIT;
8411 perf_event_wakeup(child_event);
8415 static void perf_event_exit_task_context(struct task_struct *child, int ctxn)
8417 struct perf_event *child_event, *next;
8418 struct perf_event_context *child_ctx, *clone_ctx = NULL;
8419 unsigned long flags;
8421 if (likely(!child->perf_event_ctxp[ctxn])) {
8422 perf_event_task(child, NULL, 0);
8426 local_irq_save(flags);
8428 * We can't reschedule here because interrupts are disabled,
8429 * and either child is current or it is a task that can't be
8430 * scheduled, so we are now safe from rescheduling changing
8433 child_ctx = rcu_dereference_raw(child->perf_event_ctxp[ctxn]);
8436 * Take the context lock here so that if find_get_context is
8437 * reading child->perf_event_ctxp, we wait until it has
8438 * incremented the context's refcount before we do put_ctx below.
8440 raw_spin_lock(&child_ctx->lock);
8441 task_ctx_sched_out(child_ctx);
8442 child->perf_event_ctxp[ctxn] = NULL;
8445 * If this context is a clone; unclone it so it can't get
8446 * swapped to another process while we're removing all
8447 * the events from it.
8449 clone_ctx = unclone_ctx(child_ctx);
8450 update_context_time(child_ctx);
8451 raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
8457 * Report the task dead after unscheduling the events so that we
8458 * won't get any samples after PERF_RECORD_EXIT. We can however still
8459 * get a few PERF_RECORD_READ events.
8461 perf_event_task(child, child_ctx, 0);
8464 * We can recurse on the same lock type through:
8466 * __perf_event_exit_task()
8467 * sync_child_event()
8469 * mutex_lock(&ctx->mutex)
8471 * But since its the parent context it won't be the same instance.
8473 mutex_lock(&child_ctx->mutex);
8475 list_for_each_entry_safe(child_event, next, &child_ctx->event_list, event_entry)
8476 __perf_event_exit_task(child_event, child_ctx, child);
8478 mutex_unlock(&child_ctx->mutex);
8484 * When a child task exits, feed back event values to parent events.
8486 void perf_event_exit_task(struct task_struct *child)
8488 struct perf_event *event, *tmp;
8491 mutex_lock(&child->perf_event_mutex);
8492 list_for_each_entry_safe(event, tmp, &child->perf_event_list,
8494 list_del_init(&event->owner_entry);
8497 * Ensure the list deletion is visible before we clear
8498 * the owner, closes a race against perf_release() where
8499 * we need to serialize on the owner->perf_event_mutex.
8502 event->owner = NULL;
8504 mutex_unlock(&child->perf_event_mutex);
8506 for_each_task_context_nr(ctxn)
8507 perf_event_exit_task_context(child, ctxn);
8510 static void perf_free_event(struct perf_event *event,
8511 struct perf_event_context *ctx)
8513 struct perf_event *parent = event->parent;
8515 if (WARN_ON_ONCE(!parent))
8518 mutex_lock(&parent->child_mutex);
8519 list_del_init(&event->child_list);
8520 mutex_unlock(&parent->child_mutex);
8524 raw_spin_lock_irq(&ctx->lock);
8525 perf_group_detach(event);
8526 list_del_event(event, ctx);
8527 raw_spin_unlock_irq(&ctx->lock);
8532 * Free an unexposed, unused context as created by inheritance by
8533 * perf_event_init_task below, used by fork() in case of fail.
8535 * Not all locks are strictly required, but take them anyway to be nice and
8536 * help out with the lockdep assertions.
8538 void perf_event_free_task(struct task_struct *task)
8540 struct perf_event_context *ctx;
8541 struct perf_event *event, *tmp;
8544 for_each_task_context_nr(ctxn) {
8545 ctx = task->perf_event_ctxp[ctxn];
8549 mutex_lock(&ctx->mutex);
8551 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups,
8553 perf_free_event(event, ctx);
8555 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups,
8557 perf_free_event(event, ctx);
8559 if (!list_empty(&ctx->pinned_groups) ||
8560 !list_empty(&ctx->flexible_groups))
8563 mutex_unlock(&ctx->mutex);
8569 void perf_event_delayed_put(struct task_struct *task)
8573 for_each_task_context_nr(ctxn)
8574 WARN_ON_ONCE(task->perf_event_ctxp[ctxn]);
8578 * inherit a event from parent task to child task:
8580 static struct perf_event *
8581 inherit_event(struct perf_event *parent_event,
8582 struct task_struct *parent,
8583 struct perf_event_context *parent_ctx,
8584 struct task_struct *child,
8585 struct perf_event *group_leader,
8586 struct perf_event_context *child_ctx)
8588 enum perf_event_active_state parent_state = parent_event->state;
8589 struct perf_event *child_event;
8590 unsigned long flags;
8593 * Instead of creating recursive hierarchies of events,
8594 * we link inherited events back to the original parent,
8595 * which has a filp for sure, which we use as the reference
8598 if (parent_event->parent)
8599 parent_event = parent_event->parent;
8601 child_event = perf_event_alloc(&parent_event->attr,
8604 group_leader, parent_event,
8606 if (IS_ERR(child_event))
8609 if (is_orphaned_event(parent_event) ||
8610 !atomic_long_inc_not_zero(&parent_event->refcount)) {
8611 free_event(child_event);
8618 * Make the child state follow the state of the parent event,
8619 * not its attr.disabled bit. We hold the parent's mutex,
8620 * so we won't race with perf_event_{en, dis}able_family.
8622 if (parent_state >= PERF_EVENT_STATE_INACTIVE)
8623 child_event->state = PERF_EVENT_STATE_INACTIVE;
8625 child_event->state = PERF_EVENT_STATE_OFF;
8627 if (parent_event->attr.freq) {
8628 u64 sample_period = parent_event->hw.sample_period;
8629 struct hw_perf_event *hwc = &child_event->hw;
8631 hwc->sample_period = sample_period;
8632 hwc->last_period = sample_period;
8634 local64_set(&hwc->period_left, sample_period);
8637 child_event->ctx = child_ctx;
8638 child_event->overflow_handler = parent_event->overflow_handler;
8639 child_event->overflow_handler_context
8640 = parent_event->overflow_handler_context;
8643 * Precalculate sample_data sizes
8645 perf_event__header_size(child_event);
8646 perf_event__id_header_size(child_event);
8649 * Link it up in the child's context:
8651 raw_spin_lock_irqsave(&child_ctx->lock, flags);
8652 add_event_to_ctx(child_event, child_ctx);
8653 raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
8656 * Link this into the parent event's child list
8658 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
8659 mutex_lock(&parent_event->child_mutex);
8660 list_add_tail(&child_event->child_list, &parent_event->child_list);
8661 mutex_unlock(&parent_event->child_mutex);
8666 static int inherit_group(struct perf_event *parent_event,
8667 struct task_struct *parent,
8668 struct perf_event_context *parent_ctx,
8669 struct task_struct *child,
8670 struct perf_event_context *child_ctx)
8672 struct perf_event *leader;
8673 struct perf_event *sub;
8674 struct perf_event *child_ctr;
8676 leader = inherit_event(parent_event, parent, parent_ctx,
8677 child, NULL, child_ctx);
8679 return PTR_ERR(leader);
8680 list_for_each_entry(sub, &parent_event->sibling_list, group_entry) {
8681 child_ctr = inherit_event(sub, parent, parent_ctx,
8682 child, leader, child_ctx);
8683 if (IS_ERR(child_ctr))
8684 return PTR_ERR(child_ctr);
8690 inherit_task_group(struct perf_event *event, struct task_struct *parent,
8691 struct perf_event_context *parent_ctx,
8692 struct task_struct *child, int ctxn,
8696 struct perf_event_context *child_ctx;
8698 if (!event->attr.inherit) {
8703 child_ctx = child->perf_event_ctxp[ctxn];
8706 * This is executed from the parent task context, so
8707 * inherit events that have been marked for cloning.
8708 * First allocate and initialize a context for the
8712 child_ctx = alloc_perf_context(parent_ctx->pmu, child);
8716 child->perf_event_ctxp[ctxn] = child_ctx;
8719 ret = inherit_group(event, parent, parent_ctx,
8729 * Initialize the perf_event context in task_struct
8731 static int perf_event_init_context(struct task_struct *child, int ctxn)
8733 struct perf_event_context *child_ctx, *parent_ctx;
8734 struct perf_event_context *cloned_ctx;
8735 struct perf_event *event;
8736 struct task_struct *parent = current;
8737 int inherited_all = 1;
8738 unsigned long flags;
8741 if (likely(!parent->perf_event_ctxp[ctxn]))
8745 * If the parent's context is a clone, pin it so it won't get
8748 parent_ctx = perf_pin_task_context(parent, ctxn);
8753 * No need to check if parent_ctx != NULL here; since we saw
8754 * it non-NULL earlier, the only reason for it to become NULL
8755 * is if we exit, and since we're currently in the middle of
8756 * a fork we can't be exiting at the same time.
8760 * Lock the parent list. No need to lock the child - not PID
8761 * hashed yet and not running, so nobody can access it.
8763 mutex_lock(&parent_ctx->mutex);
8766 * We dont have to disable NMIs - we are only looking at
8767 * the list, not manipulating it:
8769 list_for_each_entry(event, &parent_ctx->pinned_groups, group_entry) {
8770 ret = inherit_task_group(event, parent, parent_ctx,
8771 child, ctxn, &inherited_all);
8777 * We can't hold ctx->lock when iterating the ->flexible_group list due
8778 * to allocations, but we need to prevent rotation because
8779 * rotate_ctx() will change the list from interrupt context.
8781 raw_spin_lock_irqsave(&parent_ctx->lock, flags);
8782 parent_ctx->rotate_disable = 1;
8783 raw_spin_unlock_irqrestore(&parent_ctx->lock, flags);
8785 list_for_each_entry(event, &parent_ctx->flexible_groups, group_entry) {
8786 ret = inherit_task_group(event, parent, parent_ctx,
8787 child, ctxn, &inherited_all);
8792 raw_spin_lock_irqsave(&parent_ctx->lock, flags);
8793 parent_ctx->rotate_disable = 0;
8795 child_ctx = child->perf_event_ctxp[ctxn];
8797 if (child_ctx && inherited_all) {
8799 * Mark the child context as a clone of the parent
8800 * context, or of whatever the parent is a clone of.
8802 * Note that if the parent is a clone, the holding of
8803 * parent_ctx->lock avoids it from being uncloned.
8805 cloned_ctx = parent_ctx->parent_ctx;
8807 child_ctx->parent_ctx = cloned_ctx;
8808 child_ctx->parent_gen = parent_ctx->parent_gen;
8810 child_ctx->parent_ctx = parent_ctx;
8811 child_ctx->parent_gen = parent_ctx->generation;
8813 get_ctx(child_ctx->parent_ctx);
8816 raw_spin_unlock_irqrestore(&parent_ctx->lock, flags);
8817 mutex_unlock(&parent_ctx->mutex);
8819 perf_unpin_context(parent_ctx);
8820 put_ctx(parent_ctx);
8826 * Initialize the perf_event context in task_struct
8828 int perf_event_init_task(struct task_struct *child)
8832 memset(child->perf_event_ctxp, 0, sizeof(child->perf_event_ctxp));
8833 mutex_init(&child->perf_event_mutex);
8834 INIT_LIST_HEAD(&child->perf_event_list);
8836 for_each_task_context_nr(ctxn) {
8837 ret = perf_event_init_context(child, ctxn);
8839 perf_event_free_task(child);
8847 static void __init perf_event_init_all_cpus(void)
8849 struct swevent_htable *swhash;
8852 for_each_possible_cpu(cpu) {
8853 swhash = &per_cpu(swevent_htable, cpu);
8854 mutex_init(&swhash->hlist_mutex);
8855 INIT_LIST_HEAD(&per_cpu(active_ctx_list, cpu));
8859 static void perf_event_init_cpu(int cpu)
8861 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
8863 mutex_lock(&swhash->hlist_mutex);
8864 swhash->online = true;
8865 if (swhash->hlist_refcount > 0) {
8866 struct swevent_hlist *hlist;
8868 hlist = kzalloc_node(sizeof(*hlist), GFP_KERNEL, cpu_to_node(cpu));
8870 rcu_assign_pointer(swhash->swevent_hlist, hlist);
8872 mutex_unlock(&swhash->hlist_mutex);
8875 #if defined CONFIG_HOTPLUG_CPU || defined CONFIG_KEXEC
8876 static void __perf_event_exit_context(void *__info)
8878 struct remove_event re = { .detach_group = true };
8879 struct perf_event_context *ctx = __info;
8882 list_for_each_entry_rcu(re.event, &ctx->event_list, event_entry)
8883 __perf_remove_from_context(&re);
8887 static void perf_event_exit_cpu_context(int cpu)
8889 struct perf_event_context *ctx;
8893 idx = srcu_read_lock(&pmus_srcu);
8894 list_for_each_entry_rcu(pmu, &pmus, entry) {
8895 ctx = &per_cpu_ptr(pmu->pmu_cpu_context, cpu)->ctx;
8897 mutex_lock(&ctx->mutex);
8898 smp_call_function_single(cpu, __perf_event_exit_context, ctx, 1);
8899 mutex_unlock(&ctx->mutex);
8901 srcu_read_unlock(&pmus_srcu, idx);
8904 static void perf_event_exit_cpu(int cpu)
8906 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
8908 perf_event_exit_cpu_context(cpu);
8910 mutex_lock(&swhash->hlist_mutex);
8911 swhash->online = false;
8912 swevent_hlist_release(swhash);
8913 mutex_unlock(&swhash->hlist_mutex);
8916 static inline void perf_event_exit_cpu(int cpu) { }
8920 perf_reboot(struct notifier_block *notifier, unsigned long val, void *v)
8924 for_each_online_cpu(cpu)
8925 perf_event_exit_cpu(cpu);
8931 * Run the perf reboot notifier at the very last possible moment so that
8932 * the generic watchdog code runs as long as possible.
8934 static struct notifier_block perf_reboot_notifier = {
8935 .notifier_call = perf_reboot,
8936 .priority = INT_MIN,
8940 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
8942 unsigned int cpu = (long)hcpu;
8944 switch (action & ~CPU_TASKS_FROZEN) {
8946 case CPU_UP_PREPARE:
8947 case CPU_DOWN_FAILED:
8948 perf_event_init_cpu(cpu);
8951 case CPU_UP_CANCELED:
8952 case CPU_DOWN_PREPARE:
8953 perf_event_exit_cpu(cpu);
8962 void __init perf_event_init(void)
8968 perf_event_init_all_cpus();
8969 init_srcu_struct(&pmus_srcu);
8970 perf_pmu_register(&perf_swevent, "software", PERF_TYPE_SOFTWARE);
8971 perf_pmu_register(&perf_cpu_clock, NULL, -1);
8972 perf_pmu_register(&perf_task_clock, NULL, -1);
8974 perf_cpu_notifier(perf_cpu_notify);
8975 register_reboot_notifier(&perf_reboot_notifier);
8977 ret = init_hw_breakpoint();
8978 WARN(ret, "hw_breakpoint initialization failed with: %d", ret);
8980 /* do not patch jump label more than once per second */
8981 jump_label_rate_limit(&perf_sched_events, HZ);
8984 * Build time assertion that we keep the data_head at the intended
8985 * location. IOW, validation we got the __reserved[] size right.
8987 BUILD_BUG_ON((offsetof(struct perf_event_mmap_page, data_head))
8991 ssize_t perf_event_sysfs_show(struct device *dev, struct device_attribute *attr,
8994 struct perf_pmu_events_attr *pmu_attr =
8995 container_of(attr, struct perf_pmu_events_attr, attr);
8997 if (pmu_attr->event_str)
8998 return sprintf(page, "%s\n", pmu_attr->event_str);
9003 static int __init perf_event_sysfs_init(void)
9008 mutex_lock(&pmus_lock);
9010 ret = bus_register(&pmu_bus);
9014 list_for_each_entry(pmu, &pmus, entry) {
9015 if (!pmu->name || pmu->type < 0)
9018 ret = pmu_dev_alloc(pmu);
9019 WARN(ret, "Failed to register pmu: %s, reason %d\n", pmu->name, ret);
9021 pmu_bus_running = 1;
9025 mutex_unlock(&pmus_lock);
9029 device_initcall(perf_event_sysfs_init);
9031 #ifdef CONFIG_CGROUP_PERF
9032 static struct cgroup_subsys_state *
9033 perf_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
9035 struct perf_cgroup *jc;
9037 jc = kzalloc(sizeof(*jc), GFP_KERNEL);
9039 return ERR_PTR(-ENOMEM);
9041 jc->info = alloc_percpu(struct perf_cgroup_info);
9044 return ERR_PTR(-ENOMEM);
9050 static void perf_cgroup_css_free(struct cgroup_subsys_state *css)
9052 struct perf_cgroup *jc = container_of(css, struct perf_cgroup, css);
9054 free_percpu(jc->info);
9058 static int __perf_cgroup_move(void *info)
9060 struct task_struct *task = info;
9061 perf_cgroup_switch(task, PERF_CGROUP_SWOUT | PERF_CGROUP_SWIN);
9065 static void perf_cgroup_attach(struct cgroup_subsys_state *css,
9066 struct cgroup_taskset *tset)
9068 struct task_struct *task;
9070 cgroup_taskset_for_each(task, tset)
9071 task_function_call(task, __perf_cgroup_move, task);
9074 static void perf_cgroup_exit(struct cgroup_subsys_state *css,
9075 struct cgroup_subsys_state *old_css,
9076 struct task_struct *task)
9079 * cgroup_exit() is called in the copy_process() failure path.
9080 * Ignore this case since the task hasn't ran yet, this avoids
9081 * trying to poke a half freed task state from generic code.
9083 if (!(task->flags & PF_EXITING))
9086 task_function_call(task, __perf_cgroup_move, task);
9089 struct cgroup_subsys perf_event_cgrp_subsys = {
9090 .css_alloc = perf_cgroup_css_alloc,
9091 .css_free = perf_cgroup_css_free,
9092 .exit = perf_cgroup_exit,
9093 .attach = perf_cgroup_attach,
9095 #endif /* CONFIG_CGROUP_PERF */